]>
Commit | Line | Data |
---|---|---|
c8f0269d | 1 | /* Output routines for GCC for Renesas / SuperH SH. |
5b71da26 | 2 | Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001, 2002, |
2efd0aa6 | 3 | 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011 |
e938feb4 | 4 | Free Software Foundation, Inc. |
c5c76735 | 5 | Contributed by Steve Chamberlain (sac@cygnus.com). |
f676971a | 6 | Improved by Jim Wilson (wilson@cygnus.com). |
bc45ade3 | 7 | |
7ec022b2 | 8 | This file is part of GCC. |
c15c9075 | 9 | |
7ec022b2 | 10 | GCC is free software; you can redistribute it and/or modify |
c15c9075 | 11 | it under the terms of the GNU General Public License as published by |
2f83c7d6 | 12 | the Free Software Foundation; either version 3, or (at your option) |
c15c9075 RK |
13 | any later version. |
14 | ||
7ec022b2 | 15 | GCC is distributed in the hope that it will be useful, |
c15c9075 RK |
16 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
17 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
18 | GNU General Public License for more details. | |
19 | ||
20 | You should have received a copy of the GNU General Public License | |
2f83c7d6 NC |
21 | along with GCC; see the file COPYING3. If not see |
22 | <http://www.gnu.org/licenses/>. */ | |
bc45ade3 | 23 | |
bc45ade3 | 24 | #include "config.h" |
c5c76735 | 25 | #include "system.h" |
4977bab6 ZW |
26 | #include "coretypes.h" |
27 | #include "tm.h" | |
318881c0 | 28 | #include "insn-config.h" |
bc45ade3 | 29 | #include "rtl.h" |
bc45ade3 | 30 | #include "tree.h" |
bc45ade3 | 31 | #include "flags.h" |
bc45ade3 | 32 | #include "expr.h" |
e78d8e51 | 33 | #include "optabs.h" |
6f50eb9c | 34 | #include "reload.h" |
49ad7cfa | 35 | #include "function.h" |
8aa2a305 JW |
36 | #include "regs.h" |
37 | #include "hard-reg-set.h" | |
38 | #include "output.h" | |
956a5206 | 39 | #include "insn-attr.h" |
718f9c0f | 40 | #include "diagnostic-core.h" |
318881c0 | 41 | #include "recog.h" |
950a3816 | 42 | #include "integrate.h" |
a80b0574 | 43 | #include "dwarf2.h" |
318881c0 | 44 | #include "tm_p.h" |
672a6f42 NB |
45 | #include "target.h" |
46 | #include "target-def.h" | |
c1b92d09 | 47 | #include "langhooks.h" |
1aa03f38 | 48 | #include "basic-block.h" |
6fb5fa3c | 49 | #include "df.h" |
117dca74 | 50 | #include "cfglayout.h" |
bcc8cc82 | 51 | #include "intl.h" |
58565a33 | 52 | #include "sched-int.h" |
bd9a3465 | 53 | #include "params.h" |
8582b18a | 54 | #include "ggc.h" |
726a989a | 55 | #include "gimple.h" |
73a4d10b | 56 | #include "cfgloop.h" |
6dd3c0a5 | 57 | #include "alloc-pool.h" |
32a7ab3d | 58 | #include "tm-constrs.h" |
96e45421 | 59 | #include "opts.h" |
5fac2ed2 | 60 | |
bc45ade3 | 61 | |
1245df60 R |
62 | int code_for_indirect_jump_scratch = CODE_FOR_indirect_jump_scratch; |
63 | ||
00f8ff66 SC |
64 | #define MSW (TARGET_LITTLE_ENDIAN ? 1 : 0) |
65 | #define LSW (TARGET_LITTLE_ENDIAN ? 0 : 1) | |
66 | ||
fa5322fa AO |
67 | /* These are some macros to abstract register modes. */ |
68 | #define CONST_OK_FOR_ADD(size) \ | |
735cb76e | 69 | (TARGET_SHMEDIA ? CONST_OK_FOR_I10 (size) : CONST_OK_FOR_I08 (size)) |
fa5322fa AO |
70 | #define GEN_MOV (*(TARGET_SHMEDIA64 ? gen_movdi : gen_movsi)) |
71 | #define GEN_ADD3 (*(TARGET_SHMEDIA64 ? gen_adddi3 : gen_addsi3)) | |
72 | #define GEN_SUB3 (*(TARGET_SHMEDIA64 ? gen_subdi3 : gen_subsi3)) | |
73 | ||
561642fa AP |
74 | /* Used to simplify the logic below. Find the attributes wherever |
75 | they may be. */ | |
76 | #define SH_ATTRIBUTES(decl) \ | |
77 | (TYPE_P (decl)) ? TYPE_ATTRIBUTES (decl) \ | |
78 | : DECL_ATTRIBUTES (decl) \ | |
79 | ? (DECL_ATTRIBUTES (decl)) \ | |
80 | : TYPE_ATTRIBUTES (TREE_TYPE (decl)) | |
81 | ||
552ecbd9 AH |
82 | /* Set to 1 by expand_prologue() when the function is an interrupt handler. */ |
83 | int current_function_interrupt; | |
84 | ||
a6ab9fc0 R |
85 | tree sh_deferred_function_attributes; |
86 | tree *sh_deferred_function_attributes_tail = &sh_deferred_function_attributes; | |
16bea517 | 87 | |
6f317ef3 | 88 | /* Global variables for machine-dependent things. */ |
bc45ade3 | 89 | |
16bea517 JW |
90 | /* Which cpu are we scheduling for. */ |
91 | enum processor_type sh_cpu; | |
92 | ||
58565a33 SKG |
93 | /* Definitions used in ready queue reordering for first scheduling pass. */ |
94 | ||
95 | /* Reg weights arrays for modes SFmode and SImode, indexed by insn LUID. */ | |
96 | static short *regmode_weight[2]; | |
97 | ||
98 | /* Total SFmode and SImode weights of scheduled insns. */ | |
99 | static int curr_regmode_pressure[2]; | |
100 | ||
6fb5fa3c DB |
101 | /* Number of r0 life regions. */ |
102 | static int r0_life_regions; | |
103 | ||
58565a33 SKG |
104 | /* If true, skip cycles for Q -> R movement. */ |
105 | static int skip_cycles = 0; | |
106 | ||
107 | /* Cached value of can_issue_more. This is cached in sh_variable_issue hook | |
108 | and returned from sh_reorder2. */ | |
109 | static short cached_can_issue_more; | |
110 | ||
826159e5 KK |
111 | /* Unique number for UNSPEC_BBR pattern. */ |
112 | static unsigned int unspec_bbr_uid = 1; | |
113 | ||
bc45ade3 | 114 | /* Provides the class number of the smallest class containing |
16bea517 | 115 | reg number. */ |
bc45ade3 | 116 | |
fada1961 | 117 | enum reg_class regno_reg_class[FIRST_PSEUDO_REGISTER] = |
bc45ade3 | 118 | { |
8e87e161 | 119 | R0_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, |
bc45ade3 SC |
120 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, |
121 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
122 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
fa5322fa AO |
123 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, |
124 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
125 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
126 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
127 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
128 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
129 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
130 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
131 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
132 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
133 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
134 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
1a95a963 JW |
135 | FP0_REGS,FP_REGS, FP_REGS, FP_REGS, |
136 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
137 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
138 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
fa5322fa AO |
139 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, |
140 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
141 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
142 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
143 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
144 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
145 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
146 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
147 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
148 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
149 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
150 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
151 | TARGET_REGS, TARGET_REGS, TARGET_REGS, TARGET_REGS, | |
152 | TARGET_REGS, TARGET_REGS, TARGET_REGS, TARGET_REGS, | |
225e4f43 R |
153 | DF_REGS, DF_REGS, DF_REGS, DF_REGS, |
154 | DF_REGS, DF_REGS, DF_REGS, DF_REGS, | |
fa5322fa AO |
155 | NO_REGS, GENERAL_REGS, PR_REGS, T_REGS, |
156 | MAC_REGS, MAC_REGS, FPUL_REGS, FPSCR_REGS, | |
96a2347e | 157 | GENERAL_REGS, GENERAL_REGS, |
225e4f43 R |
158 | }; |
159 | ||
fa5322fa AO |
160 | char sh_register_names[FIRST_PSEUDO_REGISTER] \ |
161 | [MAX_REGISTER_NAME_LENGTH + 1] = SH_REGISTER_NAMES_INITIALIZER; | |
162 | ||
163 | char sh_additional_register_names[ADDREGNAMES_SIZE] \ | |
164 | [MAX_ADDITIONAL_REGISTER_NAME_LENGTH + 1] | |
165 | = SH_ADDITIONAL_REGISTER_NAMES_INITIALIZER; | |
bc45ade3 | 166 | |
225e4f43 R |
167 | int assembler_dialect; |
168 | ||
fe3ad572 SC |
169 | static bool shmedia_space_reserved_for_target_registers; |
170 | ||
96e45421 JM |
171 | static bool sh_handle_option (struct gcc_options *, struct gcc_options *, |
172 | const struct cl_decoded_option *, location_t); | |
cf277499 SB |
173 | static void split_branches (rtx); |
174 | static int branch_dest (rtx); | |
175 | static void force_into (rtx, rtx); | |
176 | static void print_slot (rtx); | |
177 | static rtx add_constant (rtx, enum machine_mode, rtx); | |
078c8b08 | 178 | static void dump_table (rtx, rtx); |
cf277499 SB |
179 | static int hi_const (rtx); |
180 | static int broken_move (rtx); | |
181 | static int mova_p (rtx); | |
182 | static rtx find_barrier (int, rtx, rtx); | |
183 | static int noncall_uses_reg (rtx, rtx, rtx *); | |
184 | static rtx gen_block_redirect (rtx, int, int); | |
185 | static void sh_reorg (void); | |
e7a8f681 | 186 | static void sh_option_override (void); |
7e4aae92 | 187 | static void sh_option_init_struct (struct gcc_options *); |
128dc8e2 | 188 | static void sh_option_default_params (void); |
21d05694 | 189 | static void output_stack_adjust (int, rtx, int, HARD_REG_SET *, bool); |
cf277499 SB |
190 | static rtx frame_insn (rtx); |
191 | static rtx push (int); | |
192 | static void pop (int); | |
193 | static void push_regs (HARD_REG_SET *, int); | |
194 | static int calc_live_regs (HARD_REG_SET *); | |
cf277499 | 195 | static HOST_WIDE_INT rounded_frame_size (int); |
7a296495 | 196 | static bool sh_frame_pointer_required (void); |
cf277499 | 197 | static rtx mark_constant_pool_use (rtx); |
cf277499 | 198 | static tree sh_handle_interrupt_handler_attribute (tree *, tree, tree, int, bool *); |
561642fa AP |
199 | static tree sh_handle_resbank_handler_attribute (tree *, tree, |
200 | tree, int, bool *); | |
201 | static tree sh2a_handle_function_vector_handler_attribute (tree *, tree, | |
202 | tree, int, bool *); | |
cf277499 SB |
203 | static tree sh_handle_sp_switch_attribute (tree *, tree, tree, int, bool *); |
204 | static tree sh_handle_trap_exit_attribute (tree *, tree, tree, int, bool *); | |
205 | static tree sh_handle_renesas_attribute (tree *, tree, tree, int, bool *); | |
9e94b957 NF |
206 | static void sh_print_operand (FILE *, rtx, int); |
207 | static void sh_print_operand_address (FILE *, rtx); | |
208 | static bool sh_print_operand_punct_valid_p (unsigned char code); | |
ef875e12 | 209 | static bool sh_asm_output_addr_const_extra (FILE *file, rtx x); |
cf277499 SB |
210 | static void sh_output_function_epilogue (FILE *, HOST_WIDE_INT); |
211 | static void sh_insert_attributes (tree, tree *); | |
8d932be3 | 212 | static const char *sh_check_pch_target_flags (int); |
e7a8f681 | 213 | static int sh_register_move_cost (enum machine_mode, reg_class_t, reg_class_t); |
cf277499 | 214 | static int sh_adjust_cost (rtx, rtx, rtx, int); |
cf277499 | 215 | static int sh_issue_rate (void); |
58565a33 SKG |
216 | static int sh_dfa_new_cycle (FILE *, int, rtx, int, int, int *sort_p); |
217 | static short find_set_regmode_weight (rtx, enum machine_mode); | |
218 | static short find_insn_regmode_weight (rtx, enum machine_mode); | |
c6d56443 | 219 | static void find_regmode_weight (basic_block, enum machine_mode); |
6fb5fa3c | 220 | static int find_r0_life_regions (basic_block); |
58565a33 SKG |
221 | static void sh_md_init_global (FILE *, int, int); |
222 | static void sh_md_finish_global (FILE *, int); | |
223 | static int rank_for_reorder (const void *, const void *); | |
224 | static void swap_reorder (rtx *, int); | |
225 | static void ready_reorder (rtx *, int); | |
226 | static short high_pressure (enum machine_mode); | |
227 | static int sh_reorder (FILE *, int, rtx *, int *, int); | |
228 | static int sh_reorder2 (FILE *, int, rtx *, int *, int); | |
229 | static void sh_md_init (FILE *, int, int); | |
230 | static int sh_variable_issue (FILE *, int, rtx, int); | |
f676971a | 231 | |
cf277499 SB |
232 | static bool sh_function_ok_for_sibcall (tree, tree); |
233 | ||
234 | static bool sh_cannot_modify_jumps_p (void); | |
a87cf97e | 235 | static reg_class_t sh_target_reg_class (void); |
fe3ad572 | 236 | static bool sh_optimize_target_register_callee_saved (bool); |
3101faab | 237 | static bool sh_ms_bitfield_layout_p (const_tree); |
cf277499 SB |
238 | |
239 | static void sh_init_builtins (void); | |
1cc24576 | 240 | static tree sh_builtin_decl (unsigned, bool); |
cf277499 | 241 | static void sh_media_init_builtins (void); |
1cc24576 | 242 | static tree sh_media_builtin_decl (unsigned, bool); |
cf277499 SB |
243 | static rtx sh_expand_builtin (tree, rtx, rtx, enum machine_mode, int); |
244 | static void sh_output_mi_thunk (FILE *, tree, HOST_WIDE_INT, HOST_WIDE_INT, tree); | |
245 | static void sh_file_start (void); | |
246 | static int flow_dependent_p (rtx, rtx); | |
7bc980e1 | 247 | static void flow_dependent_p_1 (rtx, const_rtx, void *); |
cf277499 SB |
248 | static int shiftcosts (rtx); |
249 | static int andcosts (rtx); | |
250 | static int addsubcosts (rtx); | |
251 | static int multcosts (rtx); | |
252 | static bool unspec_caller_rtx_p (rtx); | |
253 | static bool sh_cannot_copy_insn_p (rtx); | |
f40751dd JH |
254 | static bool sh_rtx_costs (rtx, int, int, int *, bool); |
255 | static int sh_address_cost (rtx, bool); | |
b48f503c KK |
256 | static int sh_pr_n_sets (void); |
257 | static rtx sh_allocate_initial_value (rtx); | |
486b10aa AS |
258 | static reg_class_t sh_preferred_reload_class (rtx, reg_class_t); |
259 | static reg_class_t sh_secondary_reload (bool, rtx, reg_class_t, | |
260 | enum machine_mode, | |
261 | struct secondary_reload_info *); | |
1bb99877 | 262 | static bool sh_legitimate_address_p (enum machine_mode, rtx, bool); |
506d7b68 | 263 | static rtx sh_legitimize_address (rtx, rtx, enum machine_mode); |
90833f7f | 264 | static rtx sh_delegitimize_address (rtx); |
fe3ad572 SC |
265 | static int shmedia_target_regs_stack_space (HARD_REG_SET *); |
266 | static int shmedia_reserve_space_for_target_registers_p (int, HARD_REG_SET *); | |
267 | static int shmedia_target_regs_stack_adjust (HARD_REG_SET *); | |
fada1961 R |
268 | static int scavenge_reg (HARD_REG_SET *s); |
269 | struct save_schedule_s; | |
270 | static struct save_entry_s *sh5_schedule_saves (HARD_REG_SET *, | |
271 | struct save_schedule_s *, int); | |
61f71b34 | 272 | |
cf277499 | 273 | static rtx sh_struct_value_rtx (tree, int); |
d64728cf | 274 | static rtx sh_function_value (const_tree, const_tree, bool); |
e7a8f681 | 275 | static bool sh_function_value_regno_p (const unsigned int); |
d64728cf | 276 | static rtx sh_libcall_value (enum machine_mode, const_rtx); |
586de218 | 277 | static bool sh_return_in_memory (const_tree, const_tree); |
cf277499 SB |
278 | static rtx sh_builtin_saveregs (void); |
279 | static void sh_setup_incoming_varargs (CUMULATIVE_ARGS *, enum machine_mode, tree, int *, int); | |
280 | static bool sh_strict_argument_naming (CUMULATIVE_ARGS *); | |
281 | static bool sh_pretend_outgoing_varargs_named (CUMULATIVE_ARGS *); | |
c35d187f | 282 | static tree sh_build_builtin_va_list (void); |
d7bd8aeb | 283 | static void sh_va_start (tree, rtx); |
726a989a | 284 | static tree sh_gimplify_va_arg_expr (tree, tree, gimple_seq *, gimple_seq *); |
d64728cf | 285 | static bool sh_promote_prototypes (const_tree); |
d055ce45 KK |
286 | static enum machine_mode sh_promote_function_mode (const_tree type, |
287 | enum machine_mode, | |
288 | int *punsignedp, | |
289 | const_tree funtype, | |
6e957ac8 | 290 | int for_return); |
8cd5a4e0 | 291 | static bool sh_pass_by_reference (CUMULATIVE_ARGS *, enum machine_mode, |
586de218 | 292 | const_tree, bool); |
6cdd5672 | 293 | static bool sh_callee_copies (CUMULATIVE_ARGS *, enum machine_mode, |
586de218 | 294 | const_tree, bool); |
78a52f11 RH |
295 | static int sh_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode, |
296 | tree, bool); | |
b64336b8 NF |
297 | static void sh_function_arg_advance (CUMULATIVE_ARGS *, enum machine_mode, |
298 | const_tree, bool); | |
299 | static rtx sh_function_arg (CUMULATIVE_ARGS *, enum machine_mode, | |
300 | const_tree, bool); | |
b0fbea13 | 301 | static bool sh_scalar_mode_supported_p (enum machine_mode); |
3101faab | 302 | static int sh_dwarf_calling_convention (const_tree); |
561642fa AP |
303 | static void sh_encode_section_info (tree, rtx, int); |
304 | static int sh2a_function_vector_p (tree); | |
994322ef RH |
305 | static void sh_trampoline_init (rtx, tree, rtx); |
306 | static rtx sh_trampoline_adjust_address (rtx); | |
5efd84c5 | 307 | static void sh_conditional_register_usage (void); |
5a82ecd9 ILT |
308 | \f |
309 | static const struct attribute_spec sh_attribute_table[] = | |
310 | { | |
62d784f7 KT |
311 | /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler, |
312 | affects_type_identity } */ | |
313 | { "interrupt_handler", 0, 0, true, false, false, | |
314 | sh_handle_interrupt_handler_attribute, false }, | |
315 | { "sp_switch", 1, 1, true, false, false, | |
316 | sh_handle_sp_switch_attribute, false }, | |
317 | { "trap_exit", 1, 1, true, false, false, | |
318 | sh_handle_trap_exit_attribute, false }, | |
319 | { "renesas", 0, 0, false, true, false, | |
320 | sh_handle_renesas_attribute, false }, | |
321 | { "trapa_handler", 0, 0, true, false, false, | |
322 | sh_handle_interrupt_handler_attribute, false }, | |
323 | { "nosave_low_regs", 0, 0, true, false, false, | |
324 | sh_handle_interrupt_handler_attribute, false }, | |
325 | { "resbank", 0, 0, true, false, false, | |
326 | sh_handle_resbank_handler_attribute, false }, | |
327 | { "function_vector", 1, 1, true, false, false, | |
328 | sh2a_handle_function_vector_handler_attribute, false }, | |
62d784f7 | 329 | { NULL, 0, 0, false, false, false, NULL, false } |
5a82ecd9 | 330 | }; |
3020190e JM |
331 | |
332 | /* Set default optimization options. */ | |
333 | static const struct default_options sh_option_optimization_table[] = | |
334 | { | |
335 | { OPT_LEVELS_1_PLUS, OPT_fomit_frame_pointer, NULL, 1 }, | |
336 | { OPT_LEVELS_1_PLUS_SPEED_ONLY, OPT_mdiv_, "inv:minlat", 1 }, | |
337 | { OPT_LEVELS_SIZE, OPT_mdiv_, SH_DIV_STR_FOR_SIZE, 1 }, | |
338 | { OPT_LEVELS_0_ONLY, OPT_mdiv_, "", 1 }, | |
339 | { OPT_LEVELS_SIZE, OPT_mcbranchdi, NULL, 0 }, | |
340 | /* We can't meaningfully test TARGET_SHMEDIA here, because -m | |
341 | options haven't been parsed yet, hence we'd read only the | |
342 | default. sh_target_reg_class will return NO_REGS if this is | |
343 | not SHMEDIA, so it's OK to always set | |
344 | flag_branch_target_load_optimize. */ | |
345 | { OPT_LEVELS_2_PLUS, OPT_fbranch_target_load_optimize, NULL, 1 }, | |
346 | { OPT_LEVELS_NONE, 0, NULL, 0 } | |
347 | }; | |
672a6f42 NB |
348 | \f |
349 | /* Initialize the GCC target structure. */ | |
91d231cb JM |
350 | #undef TARGET_ATTRIBUTE_TABLE |
351 | #define TARGET_ATTRIBUTE_TABLE sh_attribute_table | |
672a6f42 | 352 | |
301d03af RS |
353 | /* The next two are used for debug info when compiling with -gdwarf. */ |
354 | #undef TARGET_ASM_UNALIGNED_HI_OP | |
355 | #define TARGET_ASM_UNALIGNED_HI_OP "\t.uaword\t" | |
356 | #undef TARGET_ASM_UNALIGNED_SI_OP | |
357 | #define TARGET_ASM_UNALIGNED_SI_OP "\t.ualong\t" | |
358 | ||
c5387660 | 359 | /* These are NULLed out on non-SH5 in TARGET_OPTION_OVERRIDE. */ |
fa5322fa AO |
360 | #undef TARGET_ASM_UNALIGNED_DI_OP |
361 | #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaquad\t" | |
362 | #undef TARGET_ASM_ALIGNED_DI_OP | |
363 | #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t" | |
364 | ||
e7a8f681 N |
365 | #undef TARGET_OPTION_OVERRIDE |
366 | #define TARGET_OPTION_OVERRIDE sh_option_override | |
3020190e JM |
367 | #undef TARGET_OPTION_OPTIMIZATION_TABLE |
368 | #define TARGET_OPTION_OPTIMIZATION_TABLE sh_option_optimization_table | |
7e4aae92 JM |
369 | #undef TARGET_OPTION_INIT_STRUCT |
370 | #define TARGET_OPTION_INIT_STRUCT sh_option_init_struct | |
128dc8e2 JM |
371 | #undef TARGET_OPTION_DEFAULT_PARAMS |
372 | #define TARGET_OPTION_DEFAULT_PARAMS sh_option_default_params | |
e7a8f681 | 373 | |
9e94b957 NF |
374 | #undef TARGET_PRINT_OPERAND |
375 | #define TARGET_PRINT_OPERAND sh_print_operand | |
376 | #undef TARGET_PRINT_OPERAND_ADDRESS | |
377 | #define TARGET_PRINT_OPERAND_ADDRESS sh_print_operand_address | |
378 | #undef TARGET_PRINT_OPERAND_PUNCT_VALID_P | |
379 | #define TARGET_PRINT_OPERAND_PUNCT_VALID_P sh_print_operand_punct_valid_p | |
ef875e12 AS |
380 | #undef TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA |
381 | #define TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA sh_asm_output_addr_const_extra | |
382 | ||
08c148a8 NB |
383 | #undef TARGET_ASM_FUNCTION_EPILOGUE |
384 | #define TARGET_ASM_FUNCTION_EPILOGUE sh_output_function_epilogue | |
385 | ||
1aa03f38 R |
386 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
387 | #define TARGET_ASM_OUTPUT_MI_THUNK sh_output_mi_thunk | |
388 | ||
389 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
3101faab | 390 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true |
1aa03f38 | 391 | |
1bc7c5b6 ZW |
392 | #undef TARGET_ASM_FILE_START |
393 | #define TARGET_ASM_FILE_START sh_file_start | |
394 | #undef TARGET_ASM_FILE_START_FILE_DIRECTIVE | |
395 | #define TARGET_ASM_FILE_START_FILE_DIRECTIVE true | |
396 | ||
c0fb94d7 RS |
397 | #undef TARGET_DEFAULT_TARGET_FLAGS |
398 | #define TARGET_DEFAULT_TARGET_FLAGS TARGET_DEFAULT | |
399 | #undef TARGET_HANDLE_OPTION | |
400 | #define TARGET_HANDLE_OPTION sh_handle_option | |
401 | ||
e7a8f681 N |
402 | #undef TARGET_REGISTER_MOVE_COST |
403 | #define TARGET_REGISTER_MOVE_COST sh_register_move_cost | |
404 | ||
12a68f1f JM |
405 | #undef TARGET_INSERT_ATTRIBUTES |
406 | #define TARGET_INSERT_ATTRIBUTES sh_insert_attributes | |
407 | ||
c237e94a ZW |
408 | #undef TARGET_SCHED_ADJUST_COST |
409 | #define TARGET_SCHED_ADJUST_COST sh_adjust_cost | |
410 | ||
fae15c93 VM |
411 | #undef TARGET_SCHED_ISSUE_RATE |
412 | #define TARGET_SCHED_ISSUE_RATE sh_issue_rate | |
413 | ||
58565a33 SKG |
414 | /* The next 5 hooks have been implemented for reenabling sched1. With the |
415 | help of these macros we are limiting the movement of insns in sched1 to | |
f676971a | 416 | reduce the register pressure. The overall idea is to keep count of SImode |
58565a33 SKG |
417 | and SFmode regs required by already scheduled insns. When these counts |
418 | cross some threshold values; give priority to insns that free registers. | |
419 | The insn that frees registers is most likely to be the insn with lowest | |
f676971a | 420 | LUID (original insn order); but such an insn might be there in the stalled |
58565a33 SKG |
421 | queue (Q) instead of the ready queue (R). To solve this, we skip cycles |
422 | upto a max of 8 cycles so that such insns may move from Q -> R. | |
423 | ||
424 | The description of the hooks are as below: | |
425 | ||
426 | TARGET_SCHED_INIT_GLOBAL: Added a new target hook in the generic | |
427 | scheduler; it is called inside the sched_init function just after | |
428 | find_insn_reg_weights function call. It is used to calculate the SImode | |
1ae58c30 | 429 | and SFmode weights of insns of basic blocks; much similar to what |
f676971a | 430 | find_insn_reg_weights does. |
58565a33 SKG |
431 | TARGET_SCHED_FINISH_GLOBAL: Corresponding cleanup hook. |
432 | ||
433 | TARGET_SCHED_DFA_NEW_CYCLE: Skip cycles if high register pressure is | |
434 | indicated by TARGET_SCHED_REORDER2; doing this may move insns from | |
435 | (Q)->(R). | |
436 | ||
437 | TARGET_SCHED_REORDER: If the register pressure for SImode or SFmode is | |
438 | high; reorder the ready queue so that the insn with lowest LUID will be | |
439 | issued next. | |
440 | ||
441 | TARGET_SCHED_REORDER2: If the register pressure is high, indicate to | |
442 | TARGET_SCHED_DFA_NEW_CYCLE to skip cycles. | |
443 | ||
444 | TARGET_SCHED_VARIABLE_ISSUE: Cache the value of can_issue_more so that it | |
445 | can be returned from TARGET_SCHED_REORDER2. | |
446 | ||
447 | TARGET_SCHED_INIT: Reset the register pressure counting variables. */ | |
448 | ||
449 | #undef TARGET_SCHED_DFA_NEW_CYCLE | |
450 | #define TARGET_SCHED_DFA_NEW_CYCLE sh_dfa_new_cycle | |
451 | ||
452 | #undef TARGET_SCHED_INIT_GLOBAL | |
453 | #define TARGET_SCHED_INIT_GLOBAL sh_md_init_global | |
454 | ||
455 | #undef TARGET_SCHED_FINISH_GLOBAL | |
456 | #define TARGET_SCHED_FINISH_GLOBAL sh_md_finish_global | |
457 | ||
458 | #undef TARGET_SCHED_VARIABLE_ISSUE | |
459 | #define TARGET_SCHED_VARIABLE_ISSUE sh_variable_issue | |
460 | ||
461 | #undef TARGET_SCHED_REORDER | |
462 | #define TARGET_SCHED_REORDER sh_reorder | |
463 | ||
464 | #undef TARGET_SCHED_REORDER2 | |
465 | #define TARGET_SCHED_REORDER2 sh_reorder2 | |
466 | ||
467 | #undef TARGET_SCHED_INIT | |
468 | #define TARGET_SCHED_INIT sh_md_init | |
469 | ||
90833f7f KK |
470 | #undef TARGET_DELEGITIMIZE_ADDRESS |
471 | #define TARGET_DELEGITIMIZE_ADDRESS sh_delegitimize_address | |
472 | ||
506d7b68 PB |
473 | #undef TARGET_LEGITIMIZE_ADDRESS |
474 | #define TARGET_LEGITIMIZE_ADDRESS sh_legitimize_address | |
475 | ||
fa5322fa AO |
476 | #undef TARGET_CANNOT_MODIFY_JUMPS_P |
477 | #define TARGET_CANNOT_MODIFY_JUMPS_P sh_cannot_modify_jumps_p | |
fe3ad572 SC |
478 | #undef TARGET_BRANCH_TARGET_REGISTER_CLASS |
479 | #define TARGET_BRANCH_TARGET_REGISTER_CLASS sh_target_reg_class | |
480 | #undef TARGET_BRANCH_TARGET_REGISTER_CALLEE_SAVED | |
481 | #define TARGET_BRANCH_TARGET_REGISTER_CALLEE_SAVED \ | |
482 | sh_optimize_target_register_callee_saved | |
fa5322fa AO |
483 | |
484 | #undef TARGET_MS_BITFIELD_LAYOUT_P | |
485 | #define TARGET_MS_BITFIELD_LAYOUT_P sh_ms_bitfield_layout_p | |
486 | ||
c1b92d09 R |
487 | #undef TARGET_INIT_BUILTINS |
488 | #define TARGET_INIT_BUILTINS sh_init_builtins | |
1cc24576 KK |
489 | #undef TARGET_BUILTIN_DECL |
490 | #define TARGET_BUILTIN_DECL sh_builtin_decl | |
c1b92d09 R |
491 | #undef TARGET_EXPAND_BUILTIN |
492 | #define TARGET_EXPAND_BUILTIN sh_expand_builtin | |
493 | ||
4977bab6 ZW |
494 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL |
495 | #define TARGET_FUNCTION_OK_FOR_SIBCALL sh_function_ok_for_sibcall | |
496 | ||
9f3a9a08 KK |
497 | #undef TARGET_CANNOT_COPY_INSN_P |
498 | #define TARGET_CANNOT_COPY_INSN_P sh_cannot_copy_insn_p | |
3c50106f RH |
499 | #undef TARGET_RTX_COSTS |
500 | #define TARGET_RTX_COSTS sh_rtx_costs | |
dcefdf67 RH |
501 | #undef TARGET_ADDRESS_COST |
502 | #define TARGET_ADDRESS_COST sh_address_cost | |
b48f503c KK |
503 | #undef TARGET_ALLOCATE_INITIAL_VALUE |
504 | #define TARGET_ALLOCATE_INITIAL_VALUE sh_allocate_initial_value | |
3c50106f | 505 | |
18dbd950 RS |
506 | #undef TARGET_MACHINE_DEPENDENT_REORG |
507 | #define TARGET_MACHINE_DEPENDENT_REORG sh_reorg | |
508 | ||
63a9de54 CB |
509 | #undef TARGET_DWARF_REGISTER_SPAN |
510 | #define TARGET_DWARF_REGISTER_SPAN sh_dwarf_register_span | |
511 | ||
463f02cd KK |
512 | #ifdef HAVE_AS_TLS |
513 | #undef TARGET_HAVE_TLS | |
514 | #define TARGET_HAVE_TLS true | |
515 | #endif | |
516 | ||
61f71b34 DD |
517 | #undef TARGET_PROMOTE_PROTOTYPES |
518 | #define TARGET_PROMOTE_PROTOTYPES sh_promote_prototypes | |
d055ce45 KK |
519 | #undef TARGET_PROMOTE_FUNCTION_MODE |
520 | #define TARGET_PROMOTE_FUNCTION_MODE sh_promote_function_mode | |
61f71b34 | 521 | |
d64728cf AS |
522 | #undef TARGET_FUNCTION_VALUE |
523 | #define TARGET_FUNCTION_VALUE sh_function_value | |
e7a8f681 N |
524 | #undef TARGET_FUNCTION_VALUE_REGNO_P |
525 | #define TARGET_FUNCTION_VALUE_REGNO_P sh_function_value_regno_p | |
d64728cf AS |
526 | #undef TARGET_LIBCALL_VALUE |
527 | #define TARGET_LIBCALL_VALUE sh_libcall_value | |
61f71b34 DD |
528 | #undef TARGET_STRUCT_VALUE_RTX |
529 | #define TARGET_STRUCT_VALUE_RTX sh_struct_value_rtx | |
530 | #undef TARGET_RETURN_IN_MEMORY | |
531 | #define TARGET_RETURN_IN_MEMORY sh_return_in_memory | |
532 | ||
533 | #undef TARGET_EXPAND_BUILTIN_SAVEREGS | |
534 | #define TARGET_EXPAND_BUILTIN_SAVEREGS sh_builtin_saveregs | |
535 | #undef TARGET_SETUP_INCOMING_VARARGS | |
536 | #define TARGET_SETUP_INCOMING_VARARGS sh_setup_incoming_varargs | |
537 | #undef TARGET_STRICT_ARGUMENT_NAMING | |
538 | #define TARGET_STRICT_ARGUMENT_NAMING sh_strict_argument_naming | |
539 | #undef TARGET_PRETEND_OUTGOING_VARARGS_NAMED | |
540 | #define TARGET_PRETEND_OUTGOING_VARARGS_NAMED sh_pretend_outgoing_varargs_named | |
fe984136 RH |
541 | #undef TARGET_MUST_PASS_IN_STACK |
542 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
8cd5a4e0 RH |
543 | #undef TARGET_PASS_BY_REFERENCE |
544 | #define TARGET_PASS_BY_REFERENCE sh_pass_by_reference | |
6cdd5672 RH |
545 | #undef TARGET_CALLEE_COPIES |
546 | #define TARGET_CALLEE_COPIES sh_callee_copies | |
78a52f11 RH |
547 | #undef TARGET_ARG_PARTIAL_BYTES |
548 | #define TARGET_ARG_PARTIAL_BYTES sh_arg_partial_bytes | |
b64336b8 NF |
549 | #undef TARGET_FUNCTION_ARG |
550 | #define TARGET_FUNCTION_ARG sh_function_arg | |
551 | #undef TARGET_FUNCTION_ARG_ADVANCE | |
552 | #define TARGET_FUNCTION_ARG_ADVANCE sh_function_arg_advance | |
61f71b34 | 553 | |
c35d187f RH |
554 | #undef TARGET_BUILD_BUILTIN_VA_LIST |
555 | #define TARGET_BUILD_BUILTIN_VA_LIST sh_build_builtin_va_list | |
d7bd8aeb JJ |
556 | #undef TARGET_EXPAND_BUILTIN_VA_START |
557 | #define TARGET_EXPAND_BUILTIN_VA_START sh_va_start | |
5fac2ed2 RH |
558 | #undef TARGET_GIMPLIFY_VA_ARG_EXPR |
559 | #define TARGET_GIMPLIFY_VA_ARG_EXPR sh_gimplify_va_arg_expr | |
c35d187f | 560 | |
b0fbea13 KK |
561 | #undef TARGET_SCALAR_MODE_SUPPORTED_P |
562 | #define TARGET_SCALAR_MODE_SUPPORTED_P sh_scalar_mode_supported_p | |
f676971a EC |
563 | #undef TARGET_VECTOR_MODE_SUPPORTED_P |
564 | #define TARGET_VECTOR_MODE_SUPPORTED_P sh_vector_mode_supported_p | |
565 | ||
8d932be3 RS |
566 | #undef TARGET_CHECK_PCH_TARGET_FLAGS |
567 | #define TARGET_CHECK_PCH_TARGET_FLAGS sh_check_pch_target_flags | |
bcc8cc82 | 568 | |
a1c496cb EC |
569 | #undef TARGET_DWARF_CALLING_CONVENTION |
570 | #define TARGET_DWARF_CALLING_CONVENTION sh_dwarf_calling_convention | |
571 | ||
7a296495 CB |
572 | #undef TARGET_FRAME_POINTER_REQUIRED |
573 | #define TARGET_FRAME_POINTER_REQUIRED sh_frame_pointer_required | |
574 | ||
58565a33 SKG |
575 | /* Return regmode weight for insn. */ |
576 | #define INSN_REGMODE_WEIGHT(INSN, MODE) regmode_weight[((MODE) == SImode) ? 0 : 1][INSN_UID (INSN)] | |
577 | ||
578 | /* Return current register pressure for regmode. */ | |
579 | #define CURR_REGMODE_PRESSURE(MODE) curr_regmode_pressure[((MODE) == SImode) ? 0 : 1] | |
580 | ||
561642fa AP |
581 | #undef TARGET_ENCODE_SECTION_INFO |
582 | #define TARGET_ENCODE_SECTION_INFO sh_encode_section_info | |
583 | ||
8a99f6f9 R |
584 | #undef TARGET_SECONDARY_RELOAD |
585 | #define TARGET_SECONDARY_RELOAD sh_secondary_reload | |
586 | ||
486b10aa AS |
587 | #undef TARGET_PREFERRED_RELOAD_CLASS |
588 | #define TARGET_PREFERRED_RELOAD_CLASS sh_preferred_reload_class | |
589 | ||
5efd84c5 NF |
590 | #undef TARGET_CONDITIONAL_REGISTER_USAGE |
591 | #define TARGET_CONDITIONAL_REGISTER_USAGE sh_conditional_register_usage | |
592 | ||
1bb99877 KK |
593 | #undef TARGET_LEGITIMATE_ADDRESS_P |
594 | #define TARGET_LEGITIMATE_ADDRESS_P sh_legitimate_address_p | |
595 | ||
994322ef RH |
596 | #undef TARGET_TRAMPOLINE_INIT |
597 | #define TARGET_TRAMPOLINE_INIT sh_trampoline_init | |
598 | #undef TARGET_TRAMPOLINE_ADJUST_ADDRESS | |
599 | #define TARGET_TRAMPOLINE_ADJUST_ADDRESS sh_trampoline_adjust_address | |
600 | ||
561642fa AP |
601 | /* Machine-specific symbol_ref flags. */ |
602 | #define SYMBOL_FLAG_FUNCVEC_FUNCTION (SYMBOL_FLAG_MACH_DEP << 0) | |
603 | ||
f6897b10 | 604 | struct gcc_target targetm = TARGET_INITIALIZER; |
b9654711 | 605 | \f |
c0fb94d7 RS |
606 | /* Implement TARGET_HANDLE_OPTION. */ |
607 | ||
608 | static bool | |
96e45421 JM |
609 | sh_handle_option (struct gcc_options *opts, struct gcc_options *opts_set, |
610 | const struct cl_decoded_option *decoded, | |
611 | location_t loc ATTRIBUTE_UNUSED) | |
c0fb94d7 | 612 | { |
96e45421 JM |
613 | size_t code = decoded->opt_index; |
614 | ||
615 | gcc_assert (opts == &global_options); | |
616 | gcc_assert (opts_set == &global_options_set); | |
617 | ||
c0fb94d7 RS |
618 | switch (code) |
619 | { | |
620 | case OPT_m1: | |
621 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH1; | |
622 | return true; | |
623 | ||
624 | case OPT_m2: | |
625 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2; | |
626 | return true; | |
627 | ||
628 | case OPT_m2a: | |
629 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A; | |
630 | return true; | |
631 | ||
632 | case OPT_m2a_nofpu: | |
633 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_NOFPU; | |
634 | return true; | |
635 | ||
636 | case OPT_m2a_single: | |
637 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_SINGLE; | |
638 | return true; | |
639 | ||
640 | case OPT_m2a_single_only: | |
641 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_SINGLE_ONLY; | |
642 | return true; | |
643 | ||
644 | case OPT_m2e: | |
645 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2E; | |
646 | return true; | |
647 | ||
648 | case OPT_m3: | |
649 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH3; | |
650 | return true; | |
651 | ||
652 | case OPT_m3e: | |
653 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH3E; | |
654 | return true; | |
655 | ||
656 | case OPT_m4: | |
25e651ca AS |
657 | case OPT_m4_100: |
658 | case OPT_m4_200: | |
78d310c2 | 659 | case OPT_m4_300: |
c0fb94d7 RS |
660 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4; |
661 | return true; | |
662 | ||
663 | case OPT_m4_nofpu: | |
78d310c2 R |
664 | case OPT_m4_100_nofpu: |
665 | case OPT_m4_200_nofpu: | |
666 | case OPT_m4_300_nofpu: | |
667 | case OPT_m4_340: | |
25e651ca AS |
668 | case OPT_m4_400: |
669 | case OPT_m4_500: | |
c0fb94d7 RS |
670 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_NOFPU; |
671 | return true; | |
672 | ||
673 | case OPT_m4_single: | |
25e651ca AS |
674 | case OPT_m4_100_single: |
675 | case OPT_m4_200_single: | |
78d310c2 | 676 | case OPT_m4_300_single: |
c0fb94d7 RS |
677 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_SINGLE; |
678 | return true; | |
679 | ||
680 | case OPT_m4_single_only: | |
25e651ca AS |
681 | case OPT_m4_100_single_only: |
682 | case OPT_m4_200_single_only: | |
78d310c2 | 683 | case OPT_m4_300_single_only: |
c0fb94d7 RS |
684 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_SINGLE_ONLY; |
685 | return true; | |
686 | ||
687 | case OPT_m4a: | |
688 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A; | |
689 | return true; | |
690 | ||
691 | case OPT_m4a_nofpu: | |
692 | case OPT_m4al: | |
693 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_NOFPU; | |
694 | return true; | |
695 | ||
696 | case OPT_m4a_single: | |
697 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_SINGLE; | |
698 | return true; | |
699 | ||
700 | case OPT_m4a_single_only: | |
701 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_SINGLE_ONLY; | |
702 | return true; | |
703 | ||
704 | case OPT_m5_32media: | |
705 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_32MEDIA; | |
706 | return true; | |
707 | ||
708 | case OPT_m5_32media_nofpu: | |
709 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_32MEDIA_NOFPU; | |
710 | return true; | |
711 | ||
712 | case OPT_m5_64media: | |
713 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_64MEDIA; | |
714 | return true; | |
715 | ||
716 | case OPT_m5_64media_nofpu: | |
717 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_64MEDIA_NOFPU; | |
718 | return true; | |
719 | ||
720 | case OPT_m5_compact: | |
721 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_COMPACT; | |
722 | return true; | |
723 | ||
724 | case OPT_m5_compact_nofpu: | |
725 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_COMPACT_NOFPU; | |
726 | return true; | |
727 | ||
728 | default: | |
729 | return true; | |
730 | } | |
731 | } | |
732 | \f | |
7e4aae92 JM |
733 | /* Implement TARGET_OPTION_INIT_STRUCT. */ |
734 | static void | |
735 | sh_option_init_struct (struct gcc_options *opts) | |
736 | { | |
737 | /* We can't meaningfully test TARGET_SH2E / TARGET_IEEE | |
738 | here, so leave it to TARGET_OPTION_OVERRIDE to set | |
739 | flag_finite_math_only. We set it to 2 here so we know if the user | |
740 | explicitly requested this to be on or off. */ | |
741 | opts->x_flag_finite_math_only = 2; | |
742 | } | |
743 | ||
128dc8e2 JM |
744 | /* Implement TARGET_OPTION_DEFAULT_PARAMS. */ |
745 | static void | |
746 | sh_option_default_params (void) | |
747 | { | |
748 | set_default_param_value (PARAM_SIMULTANEOUS_PREFETCHES, 2); | |
bd9a3465 KK |
749 | } |
750 | ||
e7a8f681 N |
751 | /* Implement TARGET_OPTION_OVERRIDE macro. Validate and override |
752 | various options, and do some machine dependent initialization. */ | |
753 | static void | |
754 | sh_option_override (void) | |
bd9a3465 KK |
755 | { |
756 | int regno; | |
757 | ||
758 | SUBTARGET_OVERRIDE_OPTIONS; | |
3020190e JM |
759 | if (optimize > 1 && !optimize_size) |
760 | target_flags |= MASK_SAVE_ALL_TARGET_REGS; | |
bd9a3465 KK |
761 | if (flag_finite_math_only == 2) |
762 | flag_finite_math_only | |
763 | = !flag_signaling_nans && TARGET_SH2E && ! TARGET_IEEE; | |
764 | if (TARGET_SH2E && !flag_finite_math_only) | |
765 | target_flags |= MASK_IEEE; | |
766 | sh_cpu = PROCESSOR_SH1; | |
767 | assembler_dialect = 0; | |
768 | if (TARGET_SH2) | |
769 | sh_cpu = PROCESSOR_SH2; | |
770 | if (TARGET_SH2E) | |
771 | sh_cpu = PROCESSOR_SH2E; | |
772 | if (TARGET_SH2A) | |
7a61cf6f | 773 | sh_cpu = PROCESSOR_SH2A; |
bd9a3465 KK |
774 | if (TARGET_SH3) |
775 | sh_cpu = PROCESSOR_SH3; | |
776 | if (TARGET_SH3E) | |
777 | sh_cpu = PROCESSOR_SH3E; | |
778 | if (TARGET_SH4) | |
779 | { | |
780 | assembler_dialect = 1; | |
781 | sh_cpu = PROCESSOR_SH4; | |
782 | } | |
783 | if (TARGET_SH4A_ARCH) | |
784 | { | |
785 | assembler_dialect = 1; | |
786 | sh_cpu = PROCESSOR_SH4A; | |
787 | } | |
788 | if (TARGET_SH5) | |
789 | { | |
790 | sh_cpu = PROCESSOR_SH5; | |
791 | target_flags |= MASK_ALIGN_DOUBLE; | |
792 | if (TARGET_SHMEDIA_FPU) | |
793 | target_flags |= MASK_FMOVD; | |
794 | if (TARGET_SHMEDIA) | |
795 | { | |
796 | /* There are no delay slots on SHmedia. */ | |
797 | flag_delayed_branch = 0; | |
798 | /* Relaxation isn't yet supported for SHmedia */ | |
799 | target_flags &= ~MASK_RELAX; | |
800 | /* After reload, if conversion does little good but can cause | |
801 | ICEs: | |
802 | - find_if_block doesn't do anything for SH because we don't | |
803 | have conditional execution patterns. (We use conditional | |
804 | move patterns, which are handled differently, and only | |
805 | before reload). | |
806 | - find_cond_trap doesn't do anything for the SH because we | |
807 | don't have conditional traps. | |
808 | - find_if_case_1 uses redirect_edge_and_branch_force in | |
809 | the only path that does an optimization, and this causes | |
810 | an ICE when branch targets are in registers. | |
811 | - find_if_case_2 doesn't do anything for the SHmedia after | |
812 | reload except when it can redirect a tablejump - and | |
813 | that's rather rare. */ | |
814 | flag_if_conversion2 = 0; | |
815 | if (! strcmp (sh_div_str, "call")) | |
816 | sh_div_strategy = SH_DIV_CALL; | |
817 | else if (! strcmp (sh_div_str, "call2")) | |
818 | sh_div_strategy = SH_DIV_CALL2; | |
819 | if (! strcmp (sh_div_str, "fp") && TARGET_FPU_ANY) | |
820 | sh_div_strategy = SH_DIV_FP; | |
821 | else if (! strcmp (sh_div_str, "inv")) | |
822 | sh_div_strategy = SH_DIV_INV; | |
823 | else if (! strcmp (sh_div_str, "inv:minlat")) | |
824 | sh_div_strategy = SH_DIV_INV_MINLAT; | |
825 | else if (! strcmp (sh_div_str, "inv20u")) | |
826 | sh_div_strategy = SH_DIV_INV20U; | |
827 | else if (! strcmp (sh_div_str, "inv20l")) | |
828 | sh_div_strategy = SH_DIV_INV20L; | |
829 | else if (! strcmp (sh_div_str, "inv:call2")) | |
830 | sh_div_strategy = SH_DIV_INV_CALL2; | |
831 | else if (! strcmp (sh_div_str, "inv:call")) | |
832 | sh_div_strategy = SH_DIV_INV_CALL; | |
833 | else if (! strcmp (sh_div_str, "inv:fp")) | |
834 | { | |
835 | if (TARGET_FPU_ANY) | |
836 | sh_div_strategy = SH_DIV_INV_FP; | |
837 | else | |
838 | sh_div_strategy = SH_DIV_INV; | |
839 | } | |
840 | TARGET_CBRANCHDI4 = 0; | |
841 | /* Assembler CFI isn't yet fully supported for SHmedia. */ | |
842 | flag_dwarf2_cfi_asm = 0; | |
843 | } | |
844 | } | |
845 | else | |
846 | { | |
847 | /* Only the sh64-elf assembler fully supports .quad properly. */ | |
848 | targetm.asm_out.aligned_op.di = NULL; | |
849 | targetm.asm_out.unaligned_op.di = NULL; | |
850 | } | |
851 | if (TARGET_SH1) | |
852 | { | |
853 | if (! strcmp (sh_div_str, "call-div1")) | |
854 | sh_div_strategy = SH_DIV_CALL_DIV1; | |
855 | else if (! strcmp (sh_div_str, "call-fp") | |
856 | && (TARGET_FPU_DOUBLE | |
857 | || (TARGET_HARD_SH4 && TARGET_SH2E) | |
858 | || (TARGET_SHCOMPACT && TARGET_FPU_ANY))) | |
859 | sh_div_strategy = SH_DIV_CALL_FP; | |
860 | else if (! strcmp (sh_div_str, "call-table") && TARGET_SH2) | |
861 | sh_div_strategy = SH_DIV_CALL_TABLE; | |
862 | else | |
863 | /* Pick one that makes most sense for the target in general. | |
864 | It is not much good to use different functions depending | |
865 | on -Os, since then we'll end up with two different functions | |
866 | when some of the code is compiled for size, and some for | |
867 | speed. */ | |
868 | ||
869 | /* SH4 tends to emphasize speed. */ | |
870 | if (TARGET_HARD_SH4) | |
871 | sh_div_strategy = SH_DIV_CALL_TABLE; | |
872 | /* These have their own way of doing things. */ | |
873 | else if (TARGET_SH2A) | |
874 | sh_div_strategy = SH_DIV_INTRINSIC; | |
875 | /* ??? Should we use the integer SHmedia function instead? */ | |
876 | else if (TARGET_SHCOMPACT && TARGET_FPU_ANY) | |
877 | sh_div_strategy = SH_DIV_CALL_FP; | |
878 | /* SH1 .. SH3 cores often go into small-footprint systems, so | |
879 | default to the smallest implementation available. */ | |
880 | else if (TARGET_SH2) /* ??? EXPERIMENTAL */ | |
881 | sh_div_strategy = SH_DIV_CALL_TABLE; | |
882 | else | |
883 | sh_div_strategy = SH_DIV_CALL_DIV1; | |
884 | } | |
885 | if (!TARGET_SH1) | |
886 | TARGET_PRETEND_CMOVE = 0; | |
887 | if (sh_divsi3_libfunc[0]) | |
888 | ; /* User supplied - leave it alone. */ | |
889 | else if (TARGET_DIVIDE_CALL_FP) | |
890 | sh_divsi3_libfunc = "__sdivsi3_i4"; | |
891 | else if (TARGET_DIVIDE_CALL_TABLE) | |
892 | sh_divsi3_libfunc = "__sdivsi3_i4i"; | |
893 | else if (TARGET_SH5) | |
894 | sh_divsi3_libfunc = "__sdivsi3_1"; | |
895 | else | |
896 | sh_divsi3_libfunc = "__sdivsi3"; | |
897 | if (sh_branch_cost == -1) | |
898 | sh_branch_cost | |
899 | = TARGET_SH5 ? 1 : ! TARGET_SH2 || TARGET_HARD_SH4 ? 2 : 1; | |
900 | ||
901 | for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++) | |
902 | if (! VALID_REGISTER_P (regno)) | |
903 | sh_register_names[regno][0] = '\0'; | |
904 | ||
905 | for (regno = 0; regno < ADDREGNAMES_SIZE; regno++) | |
906 | if (! VALID_REGISTER_P (ADDREGNAMES_REGNO (regno))) | |
907 | sh_additional_register_names[regno][0] = '\0'; | |
908 | ||
7a296495 | 909 | flag_omit_frame_pointer = (PREFERRED_DEBUGGING_TYPE == DWARF2_DEBUG); |
bd9a3465 KK |
910 | |
911 | if ((flag_pic && ! TARGET_PREFERGOT) | |
912 | || (TARGET_SHMEDIA && !TARGET_PT_FIXED)) | |
913 | flag_no_function_cse = 1; | |
914 | ||
42db504c | 915 | if (targetm.small_register_classes_for_mode_p (VOIDmode)) \ |
bd9a3465 KK |
916 | { |
917 | /* Never run scheduling before reload, since that can | |
918 | break global alloc, and generates slower code anyway due | |
919 | to the pressure on R0. */ | |
920 | /* Enable sched1 for SH4 if the user explicitly requests. | |
921 | When sched1 is enabled, the ready queue will be reordered by | |
922 | the target hooks if pressure is high. We can not do this for | |
923 | PIC, SH3 and lower as they give spill failures for R0. */ | |
924 | if (!TARGET_HARD_SH4 || flag_pic) | |
925 | flag_schedule_insns = 0; | |
926 | /* ??? Current exception handling places basic block boundaries | |
927 | after call_insns. It causes the high pressure on R0 and gives | |
928 | spill failures for R0 in reload. See PR 22553 and the thread | |
929 | on gcc-patches | |
930 | <http://gcc.gnu.org/ml/gcc-patches/2005-10/msg00816.html>. */ | |
931 | else if (flag_exceptions) | |
932 | { | |
0a587f65 | 933 | if (flag_schedule_insns && global_options_set.x_flag_schedule_insns) |
bd9a3465 KK |
934 | warning (0, "ignoring -fschedule-insns because of exception handling bug"); |
935 | flag_schedule_insns = 0; | |
936 | } | |
0a587f65 JM |
937 | else if (flag_schedule_insns |
938 | && !global_options_set.x_flag_schedule_insns) | |
bd9a3465 KK |
939 | flag_schedule_insns = 0; |
940 | } | |
941 | ||
7a296495 CB |
942 | if ((target_flags_explicit & MASK_ACCUMULATE_OUTGOING_ARGS) == 0) |
943 | target_flags |= MASK_ACCUMULATE_OUTGOING_ARGS; | |
944 | ||
945 | /* Unwind info is not correct around the CFG unless either a frame | |
946 | pointer is present or M_A_O_A is set. Fixing this requires rewriting | |
947 | unwind info generation to be aware of the CFG and propagating states | |
948 | around edges. */ | |
949 | if ((flag_unwind_tables || flag_asynchronous_unwind_tables | |
950 | || flag_exceptions || flag_non_call_exceptions) | |
951 | && flag_omit_frame_pointer | |
952 | && !(target_flags & MASK_ACCUMULATE_OUTGOING_ARGS)) | |
953 | { | |
954 | if (target_flags_explicit & MASK_ACCUMULATE_OUTGOING_ARGS) | |
955 | warning (0, "unwind tables currently require either a frame pointer " | |
956 | "or -maccumulate-outgoing-args for correctness"); | |
957 | target_flags |= MASK_ACCUMULATE_OUTGOING_ARGS; | |
958 | } | |
959 | ||
91d1d521 KK |
960 | /* Unwinding with -freorder-blocks-and-partition does not work on this |
961 | architecture, because it requires far jumps to label crossing between | |
962 | hot/cold sections which are rejected on this architecture. */ | |
963 | if (flag_reorder_blocks_and_partition) | |
964 | { | |
965 | if (flag_exceptions) | |
966 | { | |
967 | inform (input_location, | |
968 | "-freorder-blocks-and-partition does not work with " | |
969 | "exceptions on this architecture"); | |
970 | flag_reorder_blocks_and_partition = 0; | |
971 | flag_reorder_blocks = 1; | |
972 | } | |
973 | else if (flag_unwind_tables) | |
974 | { | |
975 | inform (input_location, | |
976 | "-freorder-blocks-and-partition does not support unwind " | |
977 | "info on this architecture"); | |
978 | flag_reorder_blocks_and_partition = 0; | |
979 | flag_reorder_blocks = 1; | |
980 | } | |
981 | } | |
982 | ||
bd9a3465 KK |
983 | if (align_loops == 0) |
984 | align_loops = 1 << (TARGET_SH5 ? 3 : 2); | |
985 | if (align_jumps == 0) | |
986 | align_jumps = 1 << CACHE_LOG; | |
987 | else if (align_jumps < (TARGET_SHMEDIA ? 4 : 2)) | |
988 | align_jumps = TARGET_SHMEDIA ? 4 : 2; | |
989 | ||
990 | /* Allocation boundary (in *bytes*) for the code of a function. | |
991 | SH1: 32 bit alignment is faster, because instructions are always | |
992 | fetched as a pair from a longword boundary. | |
993 | SH2 .. SH5 : align to cache line start. */ | |
994 | if (align_functions == 0) | |
995 | align_functions | |
3020190e | 996 | = optimize_size ? FUNCTION_BOUNDARY/8 : (1 << CACHE_LOG); |
bd9a3465 KK |
997 | /* The linker relaxation code breaks when a function contains |
998 | alignments that are larger than that at the start of a | |
999 | compilation unit. */ | |
1000 | if (TARGET_RELAX) | |
1001 | { | |
1002 | int min_align | |
1003 | = align_loops > align_jumps ? align_loops : align_jumps; | |
1004 | ||
1005 | /* Also take possible .long constants / mova tables int account. */ | |
1006 | if (min_align < 4) | |
1007 | min_align = 4; | |
1008 | if (align_functions < min_align) | |
1009 | align_functions = min_align; | |
1010 | } | |
1011 | ||
1012 | if (sh_fixed_range_str) | |
1013 | sh_fix_range (sh_fixed_range_str); | |
0685e770 DD |
1014 | |
1015 | /* This target defaults to strict volatile bitfields. */ | |
1016 | if (flag_strict_volatile_bitfields < 0) | |
1017 | flag_strict_volatile_bitfields = 1; | |
bd9a3465 KK |
1018 | } |
1019 | \f | |
16bea517 | 1020 | /* Print the operand address in x to the stream. */ |
bc45ade3 | 1021 | |
9e94b957 NF |
1022 | static void |
1023 | sh_print_operand_address (FILE *stream, rtx x) | |
bc45ade3 SC |
1024 | { |
1025 | switch (GET_CODE (x)) | |
1026 | { | |
1027 | case REG: | |
225e4f43 R |
1028 | case SUBREG: |
1029 | fprintf (stream, "@%s", reg_names[true_regnum (x)]); | |
bc45ade3 | 1030 | break; |
8aa2a305 | 1031 | |
bc45ade3 SC |
1032 | case PLUS: |
1033 | { | |
1034 | rtx base = XEXP (x, 0); | |
1035 | rtx index = XEXP (x, 1); | |
1036 | ||
bc45ade3 SC |
1037 | switch (GET_CODE (index)) |
1038 | { | |
1039 | case CONST_INT: | |
c5b9ef02 | 1040 | fprintf (stream, "@(%d,%s)", (int) INTVAL (index), |
225e4f43 | 1041 | reg_names[true_regnum (base)]); |
bc45ade3 SC |
1042 | break; |
1043 | ||
1044 | case REG: | |
225e4f43 R |
1045 | case SUBREG: |
1046 | { | |
1047 | int base_num = true_regnum (base); | |
1048 | int index_num = true_regnum (index); | |
1049 | ||
1050 | fprintf (stream, "@(r0,%s)", | |
1051 | reg_names[MAX (base_num, index_num)]); | |
1052 | break; | |
1053 | } | |
bc45ade3 SC |
1054 | |
1055 | default: | |
f5b9e7c9 | 1056 | gcc_unreachable (); |
bc45ade3 SC |
1057 | } |
1058 | } | |
bc45ade3 | 1059 | break; |
8aa2a305 | 1060 | |
bc45ade3 | 1061 | case PRE_DEC: |
225e4f43 | 1062 | fprintf (stream, "@-%s", reg_names[true_regnum (XEXP (x, 0))]); |
bc45ade3 SC |
1063 | break; |
1064 | ||
1065 | case POST_INC: | |
225e4f43 | 1066 | fprintf (stream, "@%s+", reg_names[true_regnum (XEXP (x, 0))]); |
bc45ade3 SC |
1067 | break; |
1068 | ||
1069 | default: | |
b91455de | 1070 | x = mark_constant_pool_use (x); |
422be3c3 | 1071 | output_addr_const (stream, x); |
bc45ade3 SC |
1072 | break; |
1073 | } | |
1074 | } | |
1075 | ||
1076 | /* Print operand x (an rtx) in assembler syntax to file stream | |
1077 | according to modifier code. | |
1078 | ||
b9654711 | 1079 | '.' print a .s if insn needs delay slot |
1245df60 | 1080 | ',' print LOCAL_LABEL_PREFIX |
4408efce | 1081 | '@' print trap, rte or rts depending upon pragma interruptness |
8aa2a305 | 1082 | '#' output a nop if there is nothing to put in the delay slot |
fae778eb | 1083 | ''' print likelihood suffix (/u for unlikely). |
73a4d10b | 1084 | '>' print branch target if -fverbose-asm |
8aa2a305 | 1085 | 'O' print a constant without the # |
00f8ff66 | 1086 | 'R' print the LSW of a dp value - changes if in little endian |
00f8ff66 | 1087 | 'S' print the MSW of a dp value - changes if in little endian |
225e4f43 | 1088 | 'T' print the next word of a dp value - same as 'R' in big endian mode. |
31b6f0ae R |
1089 | 'M' SHMEDIA: print an `x' if `m' will print `base,index'. |
1090 | otherwise: print .b / .w / .l / .s / .d suffix if operand is a MEM. | |
c1b92d09 | 1091 | 'N' print 'r63' if the operand is (const_int 0). |
312209c6 | 1092 | 'd' print a V2SF reg as dN instead of fpN. |
fa5322fa | 1093 | 'm' print a pair `base,offset' or `base,index', for LD and ST. |
73a4d10b | 1094 | 'U' Likewise for {LD,ST}{HI,LO}. |
de6adfa2 N |
1095 | 'V' print the position of a single bit set. |
1096 | 'W' print the position of a single bit cleared. | |
9eb3a0dd | 1097 | 't' print a memory address which is a register. |
fa5322fa | 1098 | 'u' prints the lowest 16 bits of CONST_INT, as an unsigned value. |
225e4f43 | 1099 | 'o' output an operator. */ |
bc45ade3 | 1100 | |
9e94b957 NF |
1101 | static void |
1102 | sh_print_operand (FILE *stream, rtx x, int code) | |
bc45ade3 | 1103 | { |
73a4d10b R |
1104 | int regno; |
1105 | enum machine_mode mode; | |
1106 | ||
bc45ade3 SC |
1107 | switch (code) |
1108 | { | |
a6ab9fc0 R |
1109 | tree trapa_attr; |
1110 | ||
b9654711 | 1111 | case '.': |
79b2746a | 1112 | if (final_sequence |
c608a684 R |
1113 | && ! INSN_ANNULLED_BRANCH_P (XVECEXP (final_sequence, 0, 0)) |
1114 | && get_attr_length (XVECEXP (final_sequence, 0, 1))) | |
1245df60 R |
1115 | fprintf (stream, ASSEMBLER_DIALECT ? "/s" : ".s"); |
1116 | break; | |
1117 | case ',': | |
1118 | fprintf (stream, "%s", LOCAL_LABEL_PREFIX); | |
b9654711 | 1119 | break; |
d3ae8277 | 1120 | case '@': |
a6ab9fc0 R |
1121 | trapa_attr = lookup_attribute ("trap_exit", |
1122 | DECL_ATTRIBUTES (current_function_decl)); | |
1123 | if (trapa_attr) | |
1124 | fprintf (stream, "trapa #%ld", | |
1125 | (long) TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (trapa_attr)))); | |
e3ba8d11 | 1126 | else if (sh_cfun_interrupt_handler_p ()) |
561642fa AP |
1127 | { |
1128 | if (sh_cfun_resbank_handler_p ()) | |
1129 | fprintf (stream, "resbank\n"); | |
1130 | fprintf (stream, "rte"); | |
1131 | } | |
d3ae8277 | 1132 | else |
8e87e161 | 1133 | fprintf (stream, "rts"); |
d3ae8277 | 1134 | break; |
bc45ade3 | 1135 | case '#': |
16bea517 | 1136 | /* Output a nop if there's nothing in the delay slot. */ |
bc45ade3 | 1137 | if (dbr_sequence_length () == 0) |
8aa2a305 | 1138 | fprintf (stream, "\n\tnop"); |
bc45ade3 | 1139 | break; |
c8cc4417 R |
1140 | case '\'': |
1141 | { | |
1142 | rtx note = find_reg_note (current_output_insn, REG_BR_PROB, 0); | |
1143 | ||
1144 | if (note && INTVAL (XEXP (note, 0)) * 2 < REG_BR_PROB_BASE) | |
1145 | fputs ("/u", stream); | |
1146 | break; | |
1147 | } | |
73a4d10b R |
1148 | case '>': |
1149 | if (flag_verbose_asm && JUMP_LABEL (current_output_insn)) | |
1150 | { | |
1151 | fputs ("\t! target: ", stream); | |
1152 | output_addr_const (stream, JUMP_LABEL (current_output_insn)); | |
1153 | } | |
1154 | break; | |
b9654711 | 1155 | case 'O': |
b91455de | 1156 | x = mark_constant_pool_use (x); |
0d7e008e | 1157 | output_addr_const (stream, x); |
bc45ade3 | 1158 | break; |
4802062c R |
1159 | /* N.B.: %R / %S / %T adjust memory addresses by four. |
1160 | For SHMEDIA, that means they can be used to access the first and | |
1161 | second 32 bit part of a 64 bit (or larger) value that | |
1162 | might be held in floating point registers or memory. | |
1163 | While they can be used to access 64 bit parts of a larger value | |
1164 | held in general purpose registers, that won't work with memory - | |
1165 | neither for fp registers, since the frxx names are used. */ | |
bc45ade3 | 1166 | case 'R': |
4802062c R |
1167 | if (REG_P (x) || GET_CODE (x) == SUBREG) |
1168 | { | |
1169 | regno = true_regnum (x); | |
1170 | regno += FP_REGISTER_P (regno) ? 1 : LSW; | |
1171 | fputs (reg_names[regno], (stream)); | |
1172 | } | |
1173 | else if (MEM_P (x)) | |
1174 | { | |
1175 | x = adjust_address (x, SImode, 4 * LSW); | |
9e94b957 | 1176 | sh_print_operand_address (stream, XEXP (x, 0)); |
4802062c R |
1177 | } |
1178 | else | |
1179 | { | |
1180 | rtx sub = NULL_RTX; | |
1181 | ||
1182 | mode = GET_MODE (x); | |
1183 | if (mode == VOIDmode) | |
1184 | mode = DImode; | |
1185 | if (GET_MODE_SIZE (mode) >= 8) | |
1186 | sub = simplify_subreg (SImode, x, mode, 4 * LSW); | |
1187 | if (sub) | |
9e94b957 | 1188 | sh_print_operand (stream, sub, 0); |
4802062c R |
1189 | else |
1190 | output_operand_lossage ("invalid operand to %%R"); | |
1191 | } | |
8aa2a305 JW |
1192 | break; |
1193 | case 'S': | |
4802062c R |
1194 | if (REG_P (x) || GET_CODE (x) == SUBREG) |
1195 | { | |
1196 | regno = true_regnum (x); | |
1197 | regno += FP_REGISTER_P (regno) ? 0 : MSW; | |
1198 | fputs (reg_names[regno], (stream)); | |
1199 | } | |
1200 | else if (MEM_P (x)) | |
1201 | { | |
1202 | x = adjust_address (x, SImode, 4 * MSW); | |
9e94b957 | 1203 | sh_print_operand_address (stream, XEXP (x, 0)); |
4802062c R |
1204 | } |
1205 | else | |
1206 | { | |
1207 | rtx sub = NULL_RTX; | |
1208 | ||
1209 | mode = GET_MODE (x); | |
1210 | if (mode == VOIDmode) | |
1211 | mode = DImode; | |
1212 | if (GET_MODE_SIZE (mode) >= 8) | |
1213 | sub = simplify_subreg (SImode, x, mode, 4 * MSW); | |
1214 | if (sub) | |
9e94b957 | 1215 | sh_print_operand (stream, sub, 0); |
4802062c R |
1216 | else |
1217 | output_operand_lossage ("invalid operand to %%S"); | |
1218 | } | |
00f8ff66 SC |
1219 | break; |
1220 | case 'T': | |
16bea517 | 1221 | /* Next word of a double. */ |
bc45ade3 SC |
1222 | switch (GET_CODE (x)) |
1223 | { | |
1224 | case REG: | |
1225 | fputs (reg_names[REGNO (x) + 1], (stream)); | |
1226 | break; | |
1227 | case MEM: | |
225e4f43 R |
1228 | if (GET_CODE (XEXP (x, 0)) != PRE_DEC |
1229 | && GET_CODE (XEXP (x, 0)) != POST_INC) | |
b72f00af | 1230 | x = adjust_address (x, SImode, 4); |
9e94b957 | 1231 | sh_print_operand_address (stream, XEXP (x, 0)); |
bc45ade3 | 1232 | break; |
318881c0 KG |
1233 | default: |
1234 | break; | |
bc45ade3 SC |
1235 | } |
1236 | break; | |
9eb3a0dd N |
1237 | |
1238 | case 't': | |
f3536097 | 1239 | gcc_assert (MEM_P (x)); |
9eb3a0dd N |
1240 | x = XEXP (x, 0); |
1241 | switch (GET_CODE (x)) | |
1242 | { | |
1243 | case REG: | |
1244 | case SUBREG: | |
9e94b957 | 1245 | sh_print_operand (stream, x, 0); |
9eb3a0dd N |
1246 | break; |
1247 | default: | |
1248 | break; | |
1249 | } | |
1250 | break; | |
1251 | ||
225e4f43 R |
1252 | case 'o': |
1253 | switch (GET_CODE (x)) | |
1254 | { | |
1255 | case PLUS: fputs ("add", stream); break; | |
1256 | case MINUS: fputs ("sub", stream); break; | |
1257 | case MULT: fputs ("mul", stream); break; | |
1258 | case DIV: fputs ("div", stream); break; | |
c8cc4417 R |
1259 | case EQ: fputs ("eq", stream); break; |
1260 | case NE: fputs ("ne", stream); break; | |
1261 | case GT: case LT: fputs ("gt", stream); break; | |
1262 | case GE: case LE: fputs ("ge", stream); break; | |
1263 | case GTU: case LTU: fputs ("gtu", stream); break; | |
1264 | case GEU: case LEU: fputs ("geu", stream); break; | |
318881c0 KG |
1265 | default: |
1266 | break; | |
225e4f43 R |
1267 | } |
1268 | break; | |
fa5322fa | 1269 | case 'M': |
31b6f0ae R |
1270 | if (TARGET_SHMEDIA) |
1271 | { | |
f3536097 | 1272 | if (MEM_P (x) |
31b6f0ae | 1273 | && GET_CODE (XEXP (x, 0)) == PLUS |
f3536097 | 1274 | && (REG_P (XEXP (XEXP (x, 0), 1)) |
31b6f0ae R |
1275 | || GET_CODE (XEXP (XEXP (x, 0), 1)) == SUBREG)) |
1276 | fputc ('x', stream); | |
1277 | } | |
1278 | else | |
1279 | { | |
f3536097 | 1280 | if (MEM_P (x)) |
31b6f0ae R |
1281 | { |
1282 | switch (GET_MODE (x)) | |
1283 | { | |
1284 | case QImode: fputs (".b", stream); break; | |
1285 | case HImode: fputs (".w", stream); break; | |
1286 | case SImode: fputs (".l", stream); break; | |
1287 | case SFmode: fputs (".s", stream); break; | |
1288 | case DFmode: fputs (".d", stream); break; | |
1289 | default: gcc_unreachable (); | |
1290 | } | |
1291 | } | |
1292 | } | |
fa5322fa AO |
1293 | break; |
1294 | ||
1295 | case 'm': | |
f3536097 | 1296 | gcc_assert (MEM_P (x)); |
fa5322fa | 1297 | x = XEXP (x, 0); |
73a4d10b R |
1298 | /* Fall through. */ |
1299 | case 'U': | |
fa5322fa AO |
1300 | switch (GET_CODE (x)) |
1301 | { | |
1302 | case REG: | |
1303 | case SUBREG: | |
9e94b957 | 1304 | sh_print_operand (stream, x, 0); |
fa5322fa AO |
1305 | fputs (", 0", stream); |
1306 | break; | |
1307 | ||
1308 | case PLUS: | |
9e94b957 | 1309 | sh_print_operand (stream, XEXP (x, 0), 0); |
fa5322fa | 1310 | fputs (", ", stream); |
9e94b957 | 1311 | sh_print_operand (stream, XEXP (x, 1), 0); |
fa5322fa AO |
1312 | break; |
1313 | ||
1314 | default: | |
f5b9e7c9 | 1315 | gcc_unreachable (); |
fa5322fa AO |
1316 | } |
1317 | break; | |
1318 | ||
de6adfa2 N |
1319 | case 'V': |
1320 | { | |
1321 | int num = exact_log2 (INTVAL (x)); | |
1322 | gcc_assert (num >= 0); | |
1323 | fprintf (stream, "#%d", num); | |
1324 | } | |
1325 | break; | |
1326 | ||
1327 | case 'W': | |
1328 | { | |
1329 | int num = exact_log2 (~INTVAL (x)); | |
1330 | gcc_assert (num >= 0); | |
1331 | fprintf (stream, "#%d", num); | |
1332 | } | |
1333 | break; | |
1334 | ||
312209c6 | 1335 | case 'd': |
f3536097 | 1336 | gcc_assert (REG_P (x) && GET_MODE (x) == V2SFmode); |
312209c6 AO |
1337 | |
1338 | fprintf ((stream), "d%s", reg_names[REGNO (x)] + 1); | |
1339 | break; | |
f676971a | 1340 | |
c1b92d09 | 1341 | case 'N': |
0ac78517 | 1342 | if (x == CONST0_RTX (GET_MODE (x))) |
c1b92d09 R |
1343 | { |
1344 | fprintf ((stream), "r63"); | |
1345 | break; | |
1346 | } | |
1347 | goto default_output; | |
fa5322fa | 1348 | case 'u': |
f3536097 | 1349 | if (CONST_INT_P (x)) |
c8cc4417 | 1350 | { |
fa5322fa AO |
1351 | fprintf ((stream), "%u", (unsigned) INTVAL (x) & (0x10000 - 1)); |
1352 | break; | |
1353 | } | |
1354 | /* Fall through. */ | |
1355 | ||
c1b92d09 | 1356 | default_output: |
bc45ade3 | 1357 | default: |
73a4d10b R |
1358 | regno = 0; |
1359 | mode = GET_MODE (x); | |
1360 | ||
bc45ade3 SC |
1361 | switch (GET_CODE (x)) |
1362 | { | |
73a4d10b R |
1363 | case TRUNCATE: |
1364 | { | |
1365 | rtx inner = XEXP (x, 0); | |
1366 | int offset = 0; | |
1367 | enum machine_mode inner_mode; | |
1368 | ||
1369 | /* We might see SUBREGs with vector mode registers inside. */ | |
1370 | if (GET_CODE (inner) == SUBREG | |
1371 | && (GET_MODE_SIZE (GET_MODE (inner)) | |
1372 | == GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner)))) | |
1373 | && subreg_lowpart_p (inner)) | |
1374 | inner = SUBREG_REG (inner); | |
f3536097 | 1375 | if (CONST_INT_P (inner)) |
73a4d10b R |
1376 | { |
1377 | x = GEN_INT (trunc_int_for_mode (INTVAL (inner), GET_MODE (x))); | |
1378 | goto default_output; | |
1379 | } | |
1380 | inner_mode = GET_MODE (inner); | |
1381 | if (GET_CODE (inner) == SUBREG | |
1382 | && (GET_MODE_SIZE (GET_MODE (inner)) | |
1383 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner)))) | |
f3536097 | 1384 | && REG_P (SUBREG_REG (inner))) |
73a4d10b R |
1385 | { |
1386 | offset = subreg_regno_offset (REGNO (SUBREG_REG (inner)), | |
1387 | GET_MODE (SUBREG_REG (inner)), | |
1388 | SUBREG_BYTE (inner), | |
1389 | GET_MODE (inner)); | |
1390 | inner = SUBREG_REG (inner); | |
1391 | } | |
f3536097 | 1392 | if (!REG_P (inner) || GET_MODE_SIZE (inner_mode) > 8) |
73a4d10b R |
1393 | abort (); |
1394 | /* Floating point register pairs are always big endian; | |
30dc60c7 | 1395 | general purpose registers are 64 bit wide. */ |
73a4d10b R |
1396 | regno = REGNO (inner); |
1397 | regno = (HARD_REGNO_NREGS (regno, inner_mode) | |
1398 | - HARD_REGNO_NREGS (regno, mode)) | |
1399 | + offset; | |
1400 | x = inner; | |
1401 | goto reg; | |
1402 | } | |
1403 | case SIGN_EXTEND: | |
1404 | x = XEXP (x, 0); | |
1405 | goto reg; | |
fa5322fa AO |
1406 | /* FIXME: We need this on SHmedia32 because reload generates |
1407 | some sign-extended HI or QI loads into DImode registers | |
1408 | but, because Pmode is SImode, the address ends up with a | |
1409 | subreg:SI of the DImode register. Maybe reload should be | |
1410 | fixed so as to apply alter_subreg to such loads? */ | |
73a4d10b R |
1411 | case IF_THEN_ELSE: |
1412 | gcc_assert (trapping_target_operand (x, VOIDmode)); | |
1413 | x = XEXP (XEXP (x, 2), 0); | |
1414 | goto default_output; | |
fa5322fa | 1415 | case SUBREG: |
f5b9e7c9 | 1416 | gcc_assert (SUBREG_BYTE (x) == 0 |
f3536097 | 1417 | && REG_P (SUBREG_REG (x))); |
fa5322fa AO |
1418 | |
1419 | x = SUBREG_REG (x); | |
1420 | /* Fall through. */ | |
1421 | ||
73a4d10b | 1422 | reg: |
bc45ade3 | 1423 | case REG: |
73a4d10b R |
1424 | regno += REGNO (x); |
1425 | if (FP_REGISTER_P (regno) | |
1426 | && mode == V16SFmode) | |
1427 | fprintf ((stream), "mtrx%s", reg_names[regno] + 2); | |
fa5322fa | 1428 | else if (FP_REGISTER_P (REGNO (x)) |
73a4d10b R |
1429 | && mode == V4SFmode) |
1430 | fprintf ((stream), "fv%s", reg_names[regno] + 2); | |
f3536097 | 1431 | else if (REG_P (x) |
73a4d10b R |
1432 | && mode == V2SFmode) |
1433 | fprintf ((stream), "fp%s", reg_names[regno] + 2); | |
fa5322fa | 1434 | else if (FP_REGISTER_P (REGNO (x)) |
73a4d10b R |
1435 | && GET_MODE_SIZE (mode) > 4) |
1436 | fprintf ((stream), "d%s", reg_names[regno] + 1); | |
225e4f43 | 1437 | else |
73a4d10b | 1438 | fputs (reg_names[regno], (stream)); |
bc45ade3 | 1439 | break; |
fa5322fa | 1440 | |
bc45ade3 SC |
1441 | case MEM: |
1442 | output_address (XEXP (x, 0)); | |
1443 | break; | |
f676971a | 1444 | |
bc45ade3 | 1445 | default: |
fa5322fa AO |
1446 | if (TARGET_SH1) |
1447 | fputc ('#', stream); | |
bc45ade3 SC |
1448 | output_addr_const (stream, x); |
1449 | break; | |
bc45ade3 SC |
1450 | } |
1451 | break; | |
1452 | } | |
1453 | } | |
9e94b957 NF |
1454 | |
1455 | static bool | |
1456 | sh_print_operand_punct_valid_p (unsigned char code) | |
1457 | { | |
1458 | return (code == '.' || code == '#' || code == '@' || code == ',' | |
1459 | || code == '$' || code == '\'' || code == '>'); | |
1460 | } | |
ef875e12 AS |
1461 | |
1462 | /* Implement TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA. */ | |
1463 | ||
1464 | static bool | |
1465 | sh_asm_output_addr_const_extra (FILE *file, rtx x) | |
1466 | { | |
1467 | if (GET_CODE (x) == UNSPEC) | |
1468 | { | |
1469 | switch (XINT (x, 1)) | |
1470 | { | |
1471 | case UNSPEC_DATALABEL: | |
1472 | fputs ("datalabel ", file); | |
1473 | output_addr_const (file, XVECEXP (x, 0, 0)); | |
1474 | break; | |
1475 | case UNSPEC_PIC: | |
1476 | /* GLOBAL_OFFSET_TABLE or local symbols, no suffix. */ | |
1477 | output_addr_const (file, XVECEXP (x, 0, 0)); | |
1478 | break; | |
1479 | case UNSPEC_GOT: | |
1480 | output_addr_const (file, XVECEXP (x, 0, 0)); | |
1481 | fputs ("@GOT", file); | |
1482 | break; | |
1483 | case UNSPEC_GOTOFF: | |
1484 | output_addr_const (file, XVECEXP (x, 0, 0)); | |
1485 | fputs ("@GOTOFF", file); | |
1486 | break; | |
1487 | case UNSPEC_PLT: | |
1488 | output_addr_const (file, XVECEXP (x, 0, 0)); | |
1489 | fputs ("@PLT", file); | |
1490 | break; | |
1491 | case UNSPEC_GOTPLT: | |
1492 | output_addr_const (file, XVECEXP (x, 0, 0)); | |
1493 | fputs ("@GOTPLT", file); | |
1494 | break; | |
1495 | case UNSPEC_DTPOFF: | |
1496 | output_addr_const (file, XVECEXP (x, 0, 0)); | |
1497 | fputs ("@DTPOFF", file); | |
1498 | break; | |
1499 | case UNSPEC_GOTTPOFF: | |
1500 | output_addr_const (file, XVECEXP (x, 0, 0)); | |
1501 | fputs ("@GOTTPOFF", file); | |
1502 | break; | |
1503 | case UNSPEC_TPOFF: | |
1504 | output_addr_const (file, XVECEXP (x, 0, 0)); | |
1505 | fputs ("@TPOFF", file); | |
1506 | break; | |
1507 | case UNSPEC_CALLER: | |
1508 | { | |
1509 | char name[32]; | |
1510 | /* LPCS stands for Label for PIC Call Site. */ | |
1511 | targetm.asm_out.generate_internal_label (name, "LPCS", | |
1512 | INTVAL (XVECEXP (x, 0, 0))); | |
1513 | assemble_name (file, name); | |
1514 | } | |
1515 | break; | |
1516 | case UNSPEC_EXTRACT_S16: | |
1517 | case UNSPEC_EXTRACT_U16: | |
1518 | { | |
1519 | rtx val, shift; | |
1520 | ||
1521 | val = XVECEXP (x, 0, 0); | |
1522 | shift = XVECEXP (x, 0, 1); | |
1523 | fputc ('(', file); | |
1524 | if (shift != const0_rtx) | |
1525 | fputc ('(', file); | |
1526 | if (GET_CODE (val) == CONST | |
1527 | || GET_RTX_CLASS (GET_CODE (val)) != RTX_OBJ) | |
1528 | { | |
1529 | fputc ('(', file); | |
1530 | output_addr_const (file, val); | |
1531 | fputc (')', file); | |
1532 | } | |
1533 | else | |
1534 | output_addr_const (file, val); | |
1535 | if (shift != const0_rtx) | |
1536 | { | |
1537 | fputs (" >> ", file); | |
1538 | output_addr_const (file, shift); | |
1539 | fputc (')', file); | |
1540 | } | |
1541 | fputs (" & 65535)", file); | |
1542 | } | |
1543 | break; | |
1544 | case UNSPEC_SYMOFF: | |
1545 | output_addr_const (file, XVECEXP (x, 0, 0)); | |
1546 | fputc ('-', file); | |
1547 | if (GET_CODE (XVECEXP (x, 0, 1)) == CONST) | |
1548 | { | |
1549 | fputc ('(', file); | |
1550 | output_addr_const (file, XVECEXP (x, 0, 1)); | |
1551 | fputc (')', file); | |
1552 | } | |
1553 | else | |
1554 | output_addr_const (file, XVECEXP (x, 0, 1)); | |
1555 | break; | |
1556 | case UNSPEC_PCREL_SYMOFF: | |
1557 | output_addr_const (file, XVECEXP (x, 0, 0)); | |
1558 | fputs ("-(", file); | |
1559 | output_addr_const (file, XVECEXP (x, 0, 1)); | |
1560 | fputs ("-.)", file); | |
1561 | break; | |
1562 | default: | |
1563 | return false; | |
1564 | } | |
1565 | return true; | |
1566 | } | |
1567 | else | |
1568 | return false; | |
1569 | } | |
bc45ade3 | 1570 | \f |
561642fa AP |
1571 | |
1572 | /* Encode symbol attributes of a SYMBOL_REF into its | |
1573 | SYMBOL_REF_FLAGS. */ | |
1574 | static void | |
1575 | sh_encode_section_info (tree decl, rtx rtl, int first) | |
1576 | { | |
1577 | default_encode_section_info (decl, rtl, first); | |
1578 | ||
1579 | if (TREE_CODE (decl) == FUNCTION_DECL | |
1580 | && sh2a_function_vector_p (decl) && TARGET_SH2A) | |
1581 | SYMBOL_REF_FLAGS (XEXP (rtl, 0)) |= SYMBOL_FLAG_FUNCVEC_FUNCTION; | |
1582 | } | |
1583 | ||
0aa692fc R |
1584 | /* Like force_operand, but guarantees that VALUE ends up in TARGET. */ |
1585 | static void | |
cf277499 | 1586 | force_into (rtx value, rtx target) |
0aa692fc R |
1587 | { |
1588 | value = force_operand (value, target); | |
1589 | if (! rtx_equal_p (value, target)) | |
1590 | emit_insn (gen_move_insn (target, value)); | |
1591 | } | |
1592 | ||
0d7e008e SC |
1593 | /* Emit code to perform a block move. Choose the best method. |
1594 | ||
1595 | OPERANDS[0] is the destination. | |
1596 | OPERANDS[1] is the source. | |
1597 | OPERANDS[2] is the size. | |
1598 | OPERANDS[3] is the alignment safe to use. */ | |
1599 | ||
0d7e008e | 1600 | int |
cf277499 | 1601 | expand_block_move (rtx *operands) |
0d7e008e SC |
1602 | { |
1603 | int align = INTVAL (operands[3]); | |
f3536097 | 1604 | int constp = (CONST_INT_P (operands[2])); |
0d7e008e | 1605 | int bytes = (constp ? INTVAL (operands[2]) : 0); |
0d7e008e | 1606 | |
312209c6 AO |
1607 | if (! constp) |
1608 | return 0; | |
1609 | ||
1610 | /* If we could use mov.l to move words and dest is word-aligned, we | |
1611 | can use movua.l for loads and still generate a relatively short | |
1612 | and efficient sequence. */ | |
1613 | if (TARGET_SH4A_ARCH && align < 4 | |
1614 | && MEM_ALIGN (operands[0]) >= 32 | |
1615 | && can_move_by_pieces (bytes, 32)) | |
1616 | { | |
1617 | rtx dest = copy_rtx (operands[0]); | |
1618 | rtx src = copy_rtx (operands[1]); | |
1619 | /* We could use different pseudos for each copied word, but | |
1620 | since movua can only load into r0, it's kind of | |
1621 | pointless. */ | |
1622 | rtx temp = gen_reg_rtx (SImode); | |
1623 | rtx src_addr = copy_addr_to_reg (XEXP (src, 0)); | |
1624 | int copied = 0; | |
1625 | ||
1626 | while (copied + 4 <= bytes) | |
1627 | { | |
1628 | rtx to = adjust_address (dest, SImode, copied); | |
85144c5c RS |
1629 | rtx from = adjust_automodify_address (src, BLKmode, |
1630 | src_addr, copied); | |
312209c6 | 1631 | |
85144c5c | 1632 | set_mem_size (from, GEN_INT (4)); |
312209c6 AO |
1633 | emit_insn (gen_movua (temp, from)); |
1634 | emit_move_insn (src_addr, plus_constant (src_addr, 4)); | |
1635 | emit_move_insn (to, temp); | |
1636 | copied += 4; | |
1637 | } | |
1638 | ||
1639 | if (copied < bytes) | |
1640 | move_by_pieces (adjust_address (dest, BLKmode, copied), | |
1641 | adjust_automodify_address (src, BLKmode, | |
1642 | src_addr, copied), | |
1643 | bytes - copied, align, 0); | |
1644 | ||
1645 | return 1; | |
1646 | } | |
1647 | ||
8aa2a305 JW |
1648 | /* If it isn't a constant number of bytes, or if it doesn't have 4 byte |
1649 | alignment, or if it isn't a multiple of 4 bytes, then fail. */ | |
312209c6 | 1650 | if (align < 4 || (bytes % 4 != 0)) |
d3ae8277 SC |
1651 | return 0; |
1652 | ||
225e4f43 R |
1653 | if (TARGET_HARD_SH4) |
1654 | { | |
1655 | if (bytes < 12) | |
1656 | return 0; | |
1657 | else if (bytes == 12) | |
1658 | { | |
73a4d10b | 1659 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
c0d4e710 KH |
1660 | rtx r4 = gen_rtx_REG (SImode, 4); |
1661 | rtx r5 = gen_rtx_REG (SImode, 5); | |
225e4f43 | 1662 | |
73a4d10b | 1663 | function_symbol (func_addr_rtx, "__movmemSI12_i4", SFUNC_STATIC); |
0aa692fc R |
1664 | force_into (XEXP (operands[0], 0), r4); |
1665 | force_into (XEXP (operands[1], 0), r5); | |
225e4f43 R |
1666 | emit_insn (gen_block_move_real_i4 (func_addr_rtx)); |
1667 | return 1; | |
1668 | } | |
3020190e | 1669 | else if (! optimize_size) |
225e4f43 | 1670 | { |
73a4d10b R |
1671 | const char *entry_name; |
1672 | rtx func_addr_rtx = gen_reg_rtx (Pmode); | |
225e4f43 | 1673 | int dwords; |
c0d4e710 KH |
1674 | rtx r4 = gen_rtx_REG (SImode, 4); |
1675 | rtx r5 = gen_rtx_REG (SImode, 5); | |
1676 | rtx r6 = gen_rtx_REG (SImode, 6); | |
225e4f43 | 1677 | |
73a4d10b R |
1678 | entry_name = (bytes & 4 ? "__movmem_i4_odd" : "__movmem_i4_even"); |
1679 | function_symbol (func_addr_rtx, entry_name, SFUNC_STATIC); | |
0aa692fc R |
1680 | force_into (XEXP (operands[0], 0), r4); |
1681 | force_into (XEXP (operands[1], 0), r5); | |
225e4f43 R |
1682 | |
1683 | dwords = bytes >> 3; | |
1684 | emit_insn (gen_move_insn (r6, GEN_INT (dwords - 1))); | |
1685 | emit_insn (gen_block_lump_real_i4 (func_addr_rtx)); | |
1686 | return 1; | |
1687 | } | |
1688 | else | |
1689 | return 0; | |
1690 | } | |
8aa2a305 | 1691 | if (bytes < 64) |
0d7e008e SC |
1692 | { |
1693 | char entry[30]; | |
73a4d10b | 1694 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
c5c76735 JL |
1695 | rtx r4 = gen_rtx_REG (SImode, 4); |
1696 | rtx r5 = gen_rtx_REG (SImode, 5); | |
8aa2a305 | 1697 | |
70128ad9 | 1698 | sprintf (entry, "__movmemSI%d", bytes); |
73a4d10b | 1699 | function_symbol (func_addr_rtx, entry, SFUNC_STATIC); |
0aa692fc R |
1700 | force_into (XEXP (operands[0], 0), r4); |
1701 | force_into (XEXP (operands[1], 0), r5); | |
0d7e008e SC |
1702 | emit_insn (gen_block_move_real (func_addr_rtx)); |
1703 | return 1; | |
1704 | } | |
8aa2a305 JW |
1705 | |
1706 | /* This is the same number of bytes as a memcpy call, but to a different | |
1707 | less common function name, so this will occasionally use more space. */ | |
3020190e | 1708 | if (! optimize_size) |
0d7e008e | 1709 | { |
73a4d10b | 1710 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
8aa2a305 | 1711 | int final_switch, while_loop; |
c5c76735 JL |
1712 | rtx r4 = gen_rtx_REG (SImode, 4); |
1713 | rtx r5 = gen_rtx_REG (SImode, 5); | |
1714 | rtx r6 = gen_rtx_REG (SImode, 6); | |
0d7e008e | 1715 | |
73a4d10b | 1716 | function_symbol (func_addr_rtx, "__movmem", SFUNC_STATIC); |
0aa692fc R |
1717 | force_into (XEXP (operands[0], 0), r4); |
1718 | force_into (XEXP (operands[1], 0), r5); | |
0d7e008e | 1719 | |
8aa2a305 JW |
1720 | /* r6 controls the size of the move. 16 is decremented from it |
1721 | for each 64 bytes moved. Then the negative bit left over is used | |
1722 | as an index into a list of move instructions. e.g., a 72 byte move | |
1723 | would be set up with size(r6) = 14, for one iteration through the | |
1724 | big while loop, and a switch of -2 for the last part. */ | |
0d7e008e | 1725 | |
8aa2a305 JW |
1726 | final_switch = 16 - ((bytes / 4) % 16); |
1727 | while_loop = ((bytes / 4) / 16 - 1) * 16; | |
1728 | emit_insn (gen_move_insn (r6, GEN_INT (while_loop + final_switch))); | |
1729 | emit_insn (gen_block_lump_real (func_addr_rtx)); | |
1730 | return 1; | |
0d7e008e | 1731 | } |
0d7e008e | 1732 | |
d3ae8277 | 1733 | return 0; |
0d7e008e SC |
1734 | } |
1735 | ||
bc45ade3 | 1736 | /* Prepare operands for a move define_expand; specifically, one of the |
8aa2a305 | 1737 | operands must be in a register. */ |
bc45ade3 | 1738 | |
b9654711 | 1739 | int |
cf277499 | 1740 | prepare_move_operands (rtx operands[], enum machine_mode mode) |
bc45ade3 | 1741 | { |
463f02cd KK |
1742 | if ((mode == SImode || mode == DImode) |
1743 | && flag_pic | |
1744 | && ! ((mode == Pmode || mode == ptr_mode) | |
81f40b79 | 1745 | && tls_symbolic_operand (operands[1], Pmode) != TLS_MODEL_NONE)) |
1a66cd67 AO |
1746 | { |
1747 | rtx temp; | |
1748 | if (SYMBOLIC_CONST_P (operands[1])) | |
1749 | { | |
f3536097 | 1750 | if (MEM_P (operands[0])) |
1a66cd67 | 1751 | operands[1] = force_reg (Pmode, operands[1]); |
5615d8fd SC |
1752 | else if (TARGET_SHMEDIA |
1753 | && GET_CODE (operands[1]) == LABEL_REF | |
fa5322fa AO |
1754 | && target_reg_operand (operands[0], mode)) |
1755 | /* It's ok. */; | |
1a66cd67 AO |
1756 | else |
1757 | { | |
b3a13419 ILT |
1758 | temp = (!can_create_pseudo_p () |
1759 | ? operands[0] | |
1760 | : gen_reg_rtx (Pmode)); | |
fa5322fa | 1761 | operands[1] = legitimize_pic_address (operands[1], mode, temp); |
1a66cd67 AO |
1762 | } |
1763 | } | |
1764 | else if (GET_CODE (operands[1]) == CONST | |
1765 | && GET_CODE (XEXP (operands[1], 0)) == PLUS | |
1766 | && SYMBOLIC_CONST_P (XEXP (XEXP (operands[1], 0), 0))) | |
1767 | { | |
b3a13419 | 1768 | temp = !can_create_pseudo_p () ? operands[0] : gen_reg_rtx (Pmode); |
1a66cd67 | 1769 | temp = legitimize_pic_address (XEXP (XEXP (operands[1], 0), 0), |
fa5322fa AO |
1770 | mode, temp); |
1771 | operands[1] = expand_binop (mode, add_optab, temp, | |
1a66cd67 | 1772 | XEXP (XEXP (operands[1], 0), 1), |
b3a13419 ILT |
1773 | (!can_create_pseudo_p () |
1774 | ? temp | |
1775 | : gen_reg_rtx (Pmode)), | |
68c37dc9 | 1776 | 0, OPTAB_LIB_WIDEN); |
1a66cd67 AO |
1777 | } |
1778 | } | |
1779 | ||
e22856e8 JW |
1780 | if (! reload_in_progress && ! reload_completed) |
1781 | { | |
1782 | /* Copy the source to a register if both operands aren't registers. */ | |
1783 | if (! register_operand (operands[0], mode) | |
d9da94a1 | 1784 | && ! sh_register_operand (operands[1], mode)) |
e22856e8 JW |
1785 | operands[1] = copy_to_mode_reg (mode, operands[1]); |
1786 | ||
f3536097 | 1787 | if (MEM_P (operands[0]) && ! memory_operand (operands[0], mode)) |
266a2732 R |
1788 | { |
1789 | /* This is like change_address_1 (operands[0], mode, 0, 1) , | |
1790 | except that we can't use that function because it is static. */ | |
0a2aaacc KG |
1791 | rtx new_rtx = change_address (operands[0], mode, 0); |
1792 | MEM_COPY_ATTRIBUTES (new_rtx, operands[0]); | |
1793 | operands[0] = new_rtx; | |
266a2732 R |
1794 | } |
1795 | ||
e22856e8 JW |
1796 | /* This case can happen while generating code to move the result |
1797 | of a library call to the target. Reject `st r0,@(rX,rY)' because | |
1798 | reload will fail to find a spill register for rX, since r0 is already | |
1799 | being used for the source. */ | |
73a4d10b R |
1800 | else if (TARGET_SH1 |
1801 | && refers_to_regno_p (R0_REG, R0_REG + 1, operands[1], (rtx *)0) | |
f3536097 | 1802 | && MEM_P (operands[0]) |
e22856e8 | 1803 | && GET_CODE (XEXP (operands[0], 0)) == PLUS |
f3536097 | 1804 | && REG_P (XEXP (XEXP (operands[0], 0), 1))) |
e22856e8 JW |
1805 | operands[1] = copy_to_mode_reg (mode, operands[1]); |
1806 | } | |
8aa2a305 | 1807 | |
463f02cd KK |
1808 | if (mode == Pmode || mode == ptr_mode) |
1809 | { | |
6058b2fa | 1810 | rtx op0, op1, opc; |
463f02cd KK |
1811 | enum tls_model tls_kind; |
1812 | ||
1813 | op0 = operands[0]; | |
1814 | op1 = operands[1]; | |
6058b2fa KK |
1815 | if (GET_CODE (op1) == CONST |
1816 | && GET_CODE (XEXP (op1, 0)) == PLUS | |
81f40b79 ILT |
1817 | && (tls_symbolic_operand (XEXP (XEXP (op1, 0), 0), Pmode) |
1818 | != TLS_MODEL_NONE)) | |
6058b2fa KK |
1819 | { |
1820 | opc = XEXP (XEXP (op1, 0), 1); | |
1821 | op1 = XEXP (XEXP (op1, 0), 0); | |
1822 | } | |
1823 | else | |
1824 | opc = NULL_RTX; | |
1825 | ||
81f40b79 | 1826 | if ((tls_kind = tls_symbolic_operand (op1, Pmode)) != TLS_MODEL_NONE) |
463f02cd KK |
1827 | { |
1828 | rtx tga_op1, tga_ret, tmp, tmp2; | |
1829 | ||
463f02cd KK |
1830 | switch (tls_kind) |
1831 | { | |
1832 | case TLS_MODEL_GLOBAL_DYNAMIC: | |
1833 | tga_ret = gen_rtx_REG (Pmode, R0_REG); | |
13ecc9e0 | 1834 | emit_call_insn (gen_tls_global_dynamic (tga_ret, op1)); |
463f02cd KK |
1835 | op1 = tga_ret; |
1836 | break; | |
1837 | ||
1838 | case TLS_MODEL_LOCAL_DYNAMIC: | |
1839 | tga_ret = gen_rtx_REG (Pmode, R0_REG); | |
13ecc9e0 | 1840 | emit_call_insn (gen_tls_local_dynamic (tga_ret, op1)); |
463f02cd KK |
1841 | |
1842 | tmp = gen_reg_rtx (Pmode); | |
1843 | emit_move_insn (tmp, tga_ret); | |
1844 | ||
1845 | if (register_operand (op0, Pmode)) | |
1846 | tmp2 = op0; | |
1847 | else | |
1848 | tmp2 = gen_reg_rtx (Pmode); | |
1849 | ||
1850 | emit_insn (gen_symDTPOFF2reg (tmp2, op1, tmp)); | |
1851 | op1 = tmp2; | |
1852 | break; | |
1853 | ||
1854 | case TLS_MODEL_INITIAL_EXEC: | |
1855 | if (! flag_pic) | |
b741fa00 | 1856 | { |
b92c3264 KK |
1857 | /* Don't schedule insns for getting GOT address when |
1858 | the first scheduling is enabled, to avoid spill | |
1859 | failures for R0. */ | |
1860 | if (flag_schedule_insns) | |
1861 | emit_insn (gen_blockage ()); | |
b741fa00 | 1862 | emit_insn (gen_GOTaddr2picreg ()); |
c41c1387 | 1863 | emit_use (gen_rtx_REG (SImode, PIC_REG)); |
b92c3264 KK |
1864 | if (flag_schedule_insns) |
1865 | emit_insn (gen_blockage ()); | |
b741fa00 | 1866 | } |
b3a13419 | 1867 | tga_op1 = !can_create_pseudo_p () ? op0 : gen_reg_rtx (Pmode); |
463f02cd KK |
1868 | tmp = gen_sym2GOTTPOFF (op1); |
1869 | emit_insn (gen_tls_initial_exec (tga_op1, tmp)); | |
1870 | op1 = tga_op1; | |
1871 | break; | |
1872 | ||
1873 | case TLS_MODEL_LOCAL_EXEC: | |
1874 | tmp2 = gen_reg_rtx (Pmode); | |
1875 | emit_insn (gen_load_gbr (tmp2)); | |
1876 | tmp = gen_reg_rtx (Pmode); | |
1877 | emit_insn (gen_symTPOFF2reg (tmp, op1)); | |
463f02cd KK |
1878 | |
1879 | if (register_operand (op0, Pmode)) | |
1880 | op1 = op0; | |
1881 | else | |
1882 | op1 = gen_reg_rtx (Pmode); | |
1883 | ||
1884 | emit_insn (gen_addsi3 (op1, tmp, tmp2)); | |
1885 | break; | |
1886 | ||
1887 | default: | |
f5b9e7c9 | 1888 | gcc_unreachable (); |
463f02cd | 1889 | } |
6058b2fa KK |
1890 | if (opc) |
1891 | emit_insn (gen_addsi3 (op1, op1, force_reg (SImode, opc))); | |
463f02cd KK |
1892 | operands[1] = op1; |
1893 | } | |
1894 | } | |
1895 | ||
8aa2a305 JW |
1896 | return 0; |
1897 | } | |
1898 | ||
78d310c2 R |
1899 | enum rtx_code |
1900 | prepare_cbranch_operands (rtx *operands, enum machine_mode mode, | |
1901 | enum rtx_code comparison) | |
1902 | { | |
1903 | rtx op1; | |
1904 | rtx scratch = NULL_RTX; | |
1905 | ||
b757d36f | 1906 | if (comparison == LAST_AND_UNUSED_RTX_CODE) |
78d310c2 R |
1907 | comparison = GET_CODE (operands[0]); |
1908 | else | |
1909 | scratch = operands[4]; | |
f3536097 SZ |
1910 | if (CONST_INT_P (operands[1]) |
1911 | && !CONST_INT_P (operands[2])) | |
78d310c2 R |
1912 | { |
1913 | rtx tmp = operands[1]; | |
1914 | ||
1915 | operands[1] = operands[2]; | |
1916 | operands[2] = tmp; | |
1917 | comparison = swap_condition (comparison); | |
1918 | } | |
f3536097 | 1919 | if (CONST_INT_P (operands[2])) |
78d310c2 R |
1920 | { |
1921 | HOST_WIDE_INT val = INTVAL (operands[2]); | |
1922 | if ((val == -1 || val == -0x81) | |
1923 | && (comparison == GT || comparison == LE)) | |
1924 | { | |
1925 | comparison = (comparison == GT) ? GE : LT; | |
1926 | operands[2] = gen_int_mode (val + 1, mode); | |
1927 | } | |
1928 | else if ((val == 1 || val == 0x80) | |
1929 | && (comparison == GE || comparison == LT)) | |
1930 | { | |
1931 | comparison = (comparison == GE) ? GT : LE; | |
1932 | operands[2] = gen_int_mode (val - 1, mode); | |
1933 | } | |
1934 | else if (val == 1 && (comparison == GEU || comparison == LTU)) | |
1935 | { | |
1936 | comparison = (comparison == GEU) ? NE : EQ; | |
1937 | operands[2] = CONST0_RTX (mode); | |
1938 | } | |
1939 | else if (val == 0x80 && (comparison == GEU || comparison == LTU)) | |
1940 | { | |
1941 | comparison = (comparison == GEU) ? GTU : LEU; | |
1942 | operands[2] = gen_int_mode (val - 1, mode); | |
1943 | } | |
1944 | else if (val == 0 && (comparison == GTU || comparison == LEU)) | |
1945 | comparison = (comparison == GTU) ? NE : EQ; | |
1946 | else if (mode == SImode | |
1947 | && ((val == 0x7fffffff | |
1948 | && (comparison == GTU || comparison == LEU)) | |
1949 | || ((unsigned HOST_WIDE_INT) val | |
1950 | == (unsigned HOST_WIDE_INT) 0x7fffffff + 1 | |
1951 | && (comparison == GEU || comparison == LTU)))) | |
1952 | { | |
1953 | comparison = (comparison == GTU || comparison == GEU) ? LT : GE; | |
1954 | operands[2] = CONST0_RTX (mode); | |
1955 | } | |
1956 | } | |
1957 | op1 = operands[1]; | |
b3a13419 | 1958 | if (can_create_pseudo_p ()) |
78d310c2 R |
1959 | operands[1] = force_reg (mode, op1); |
1960 | /* When we are handling DImode comparisons, we want to keep constants so | |
1961 | that we can optimize the component comparisons; however, memory loads | |
1962 | are better issued as a whole so that they can be scheduled well. | |
1963 | SImode equality comparisons allow I08 constants, but only when they | |
1964 | compare r0. Hence, if operands[1] has to be loaded from somewhere else | |
1965 | into a register, that register might as well be r0, and we allow the | |
1966 | constant. If it is already in a register, this is likely to be | |
2f8e468b | 1967 | allocated to a different hard register, thus we load the constant into |
78d310c2 R |
1968 | a register unless it is zero. */ |
1969 | if (!REG_P (operands[2]) | |
f3536097 | 1970 | && (!CONST_INT_P (operands[2]) |
78d310c2 R |
1971 | || (mode == SImode && operands[2] != CONST0_RTX (SImode) |
1972 | && ((comparison != EQ && comparison != NE) | |
1973 | || (REG_P (op1) && REGNO (op1) != R0_REG) | |
32a7ab3d | 1974 | || !satisfies_constraint_I08 (operands[2]))))) |
78d310c2 R |
1975 | { |
1976 | if (scratch && GET_MODE (scratch) == mode) | |
1977 | { | |
1978 | emit_move_insn (scratch, operands[2]); | |
1979 | operands[2] = scratch; | |
1980 | } | |
b3a13419 | 1981 | else if (can_create_pseudo_p ()) |
78d310c2 R |
1982 | operands[2] = force_reg (mode, operands[2]); |
1983 | } | |
1984 | return comparison; | |
1985 | } | |
1986 | ||
1987 | void | |
1988 | expand_cbranchsi4 (rtx *operands, enum rtx_code comparison, int probability) | |
1989 | { | |
1990 | rtx (*branch_expander) (rtx) = gen_branch_true; | |
1991 | rtx jump; | |
1992 | ||
1993 | comparison = prepare_cbranch_operands (operands, SImode, comparison); | |
1994 | switch (comparison) | |
1995 | { | |
1996 | case NE: case LT: case LE: case LTU: case LEU: | |
1997 | comparison = reverse_condition (comparison); | |
1998 | branch_expander = gen_branch_false; | |
1999 | default: ; | |
2000 | } | |
2001 | emit_insn (gen_rtx_SET (VOIDmode, gen_rtx_REG (SImode, T_REG), | |
2002 | gen_rtx_fmt_ee (comparison, SImode, | |
2003 | operands[1], operands[2]))); | |
2004 | jump = emit_jump_insn (branch_expander (operands[3])); | |
2005 | if (probability >= 0) | |
b757d36f | 2006 | add_reg_note (jump, REG_BR_PROB, GEN_INT (probability)); |
78d310c2 R |
2007 | |
2008 | } | |
2009 | ||
2010 | /* ??? How should we distribute probabilities when more than one branch | |
2011 | is generated. So far we only have soem ad-hoc observations: | |
2012 | - If the operands are random, they are likely to differ in both parts. | |
2013 | - If comparing items in a hash chain, the operands are random or equal; | |
2014 | operation should be EQ or NE. | |
2015 | - If items are searched in an ordered tree from the root, we can expect | |
2016 | the highpart to be unequal about half of the time; operation should be | |
2f8e468b | 2017 | an inequality comparison, operands non-constant, and overall probability |
78d310c2 R |
2018 | about 50%. Likewise for quicksort. |
2019 | - Range checks will be often made against constants. Even if we assume for | |
2020 | simplicity an even distribution of the non-constant operand over a | |
2021 | sub-range here, the same probability could be generated with differently | |
2022 | wide sub-ranges - as long as the ratio of the part of the subrange that | |
2023 | is before the threshold to the part that comes after the threshold stays | |
2024 | the same. Thus, we can't really tell anything here; | |
2025 | assuming random distribution is at least simple. | |
2026 | */ | |
2027 | ||
2028 | bool | |
2029 | expand_cbranchdi4 (rtx *operands, enum rtx_code comparison) | |
2030 | { | |
2031 | enum rtx_code msw_taken, msw_skip, lsw_taken; | |
cb3d3842 | 2032 | rtx skip_label = NULL_RTX; |
78d310c2 R |
2033 | rtx op1h, op1l, op2h, op2l; |
2034 | int num_branches; | |
2035 | int prob, rev_prob; | |
2036 | int msw_taken_prob = -1, msw_skip_prob = -1, lsw_taken_prob = -1; | |
6f8eec29 | 2037 | rtx scratch = operands[4]; |
78d310c2 R |
2038 | |
2039 | comparison = prepare_cbranch_operands (operands, DImode, comparison); | |
2040 | op1h = gen_highpart_mode (SImode, DImode, operands[1]); | |
2041 | op2h = gen_highpart_mode (SImode, DImode, operands[2]); | |
2042 | op1l = gen_lowpart (SImode, operands[1]); | |
2043 | op2l = gen_lowpart (SImode, operands[2]); | |
b757d36f | 2044 | msw_taken = msw_skip = lsw_taken = LAST_AND_UNUSED_RTX_CODE; |
78d310c2 R |
2045 | prob = split_branch_probability; |
2046 | rev_prob = REG_BR_PROB_BASE - prob; | |
2047 | switch (comparison) | |
2048 | { | |
2049 | /* ??? Should we use the cmpeqdi_t pattern for equality comparisons? | |
2050 | That costs 1 cycle more when the first branch can be predicted taken, | |
2051 | but saves us mispredicts because only one branch needs prediction. | |
2052 | It also enables generating the cmpeqdi_t-1 pattern. */ | |
2053 | case EQ: | |
2054 | if (TARGET_CMPEQDI_T) | |
2055 | { | |
2056 | emit_insn (gen_cmpeqdi_t (operands[1], operands[2])); | |
2057 | emit_jump_insn (gen_branch_true (operands[3])); | |
2058 | return true; | |
2059 | } | |
2060 | msw_skip = NE; | |
2061 | lsw_taken = EQ; | |
2062 | if (prob >= 0) | |
2063 | { | |
2064 | /* If we had more precision, we'd use rev_prob - (rev_prob >> 32) . | |
2065 | */ | |
2066 | msw_skip_prob = rev_prob; | |
2067 | if (REG_BR_PROB_BASE <= 65535) | |
2068 | lsw_taken_prob = prob ? REG_BR_PROB_BASE : 0; | |
2069 | else | |
2070 | { | |
2071 | gcc_assert (HOST_BITS_PER_WIDEST_INT >= 64); | |
2072 | lsw_taken_prob | |
2073 | = (prob | |
2074 | ? (REG_BR_PROB_BASE | |
2075 | - ((HOST_WIDEST_INT) REG_BR_PROB_BASE * rev_prob | |
2076 | / ((HOST_WIDEST_INT) prob << 32))) | |
2077 | : 0); | |
2078 | } | |
2079 | } | |
2080 | break; | |
2081 | case NE: | |
2082 | if (TARGET_CMPEQDI_T) | |
2083 | { | |
2084 | emit_insn (gen_cmpeqdi_t (operands[1], operands[2])); | |
2085 | emit_jump_insn (gen_branch_false (operands[3])); | |
2086 | return true; | |
2087 | } | |
2088 | msw_taken = NE; | |
6f8eec29 | 2089 | msw_taken_prob = prob; |
78d310c2 R |
2090 | lsw_taken = NE; |
2091 | lsw_taken_prob = 0; | |
2092 | break; | |
2093 | case GTU: case GT: | |
2094 | msw_taken = comparison; | |
f3536097 | 2095 | if (CONST_INT_P (op2l) && INTVAL (op2l) == -1) |
78d310c2 R |
2096 | break; |
2097 | if (comparison != GTU || op2h != CONST0_RTX (SImode)) | |
2098 | msw_skip = swap_condition (msw_taken); | |
2099 | lsw_taken = GTU; | |
2100 | break; | |
2101 | case GEU: case GE: | |
2102 | if (op2l == CONST0_RTX (SImode)) | |
2103 | msw_taken = comparison; | |
2104 | else | |
2105 | { | |
2106 | msw_taken = comparison == GE ? GT : GTU; | |
2107 | msw_skip = swap_condition (msw_taken); | |
2108 | lsw_taken = GEU; | |
2109 | } | |
2110 | break; | |
2111 | case LTU: case LT: | |
2112 | msw_taken = comparison; | |
2113 | if (op2l == CONST0_RTX (SImode)) | |
2114 | break; | |
2115 | msw_skip = swap_condition (msw_taken); | |
2116 | lsw_taken = LTU; | |
2117 | break; | |
2118 | case LEU: case LE: | |
f3536097 | 2119 | if (CONST_INT_P (op2l) && INTVAL (op2l) == -1) |
78d310c2 R |
2120 | msw_taken = comparison; |
2121 | else | |
2122 | { | |
2123 | lsw_taken = LEU; | |
2124 | if (comparison == LE) | |
2125 | msw_taken = LT; | |
2126 | else if (op2h != CONST0_RTX (SImode)) | |
2127 | msw_taken = LTU; | |
2128 | else | |
2129 | break; | |
2130 | msw_skip = swap_condition (msw_taken); | |
2131 | } | |
2132 | break; | |
2133 | default: return false; | |
2134 | } | |
b757d36f KK |
2135 | num_branches = ((msw_taken != LAST_AND_UNUSED_RTX_CODE) |
2136 | + (msw_skip != LAST_AND_UNUSED_RTX_CODE) | |
2137 | + (lsw_taken != LAST_AND_UNUSED_RTX_CODE)); | |
78d310c2 R |
2138 | if (comparison != EQ && comparison != NE && num_branches > 1) |
2139 | { | |
2140 | if (!CONSTANT_P (operands[2]) | |
2141 | && prob >= (int) (REG_BR_PROB_BASE * 3 / 8U) | |
2142 | && prob <= (int) (REG_BR_PROB_BASE * 5 / 8U)) | |
2143 | { | |
2144 | msw_taken_prob = prob / 2U; | |
2145 | msw_skip_prob | |
2146 | = REG_BR_PROB_BASE * rev_prob / (REG_BR_PROB_BASE + rev_prob); | |
2147 | lsw_taken_prob = prob; | |
2148 | } | |
2149 | else | |
2150 | { | |
2151 | msw_taken_prob = prob; | |
2152 | msw_skip_prob = REG_BR_PROB_BASE; | |
2153 | /* ??? If we have a constant op2h, should we use that when | |
2154 | calculating lsw_taken_prob? */ | |
2155 | lsw_taken_prob = prob; | |
2156 | } | |
2157 | } | |
2158 | operands[1] = op1h; | |
2159 | operands[2] = op2h; | |
2160 | operands[4] = NULL_RTX; | |
6f8eec29 | 2161 | if (reload_completed |
99052692 KK |
2162 | && ! arith_reg_or_0_operand (op2h, SImode) |
2163 | && (true_regnum (op1h) || (comparison != EQ && comparison != NE)) | |
b757d36f KK |
2164 | && (msw_taken != LAST_AND_UNUSED_RTX_CODE |
2165 | || msw_skip != LAST_AND_UNUSED_RTX_CODE)) | |
6f8eec29 R |
2166 | { |
2167 | emit_move_insn (scratch, operands[2]); | |
2168 | operands[2] = scratch; | |
2169 | } | |
b757d36f | 2170 | if (msw_taken != LAST_AND_UNUSED_RTX_CODE) |
78d310c2 | 2171 | expand_cbranchsi4 (operands, msw_taken, msw_taken_prob); |
b757d36f | 2172 | if (msw_skip != LAST_AND_UNUSED_RTX_CODE) |
78d310c2 R |
2173 | { |
2174 | rtx taken_label = operands[3]; | |
2175 | ||
d60819f8 CB |
2176 | /* Operands were possibly modified, but msw_skip doesn't expect this. |
2177 | Always use the original ones. */ | |
b757d36f | 2178 | if (msw_taken != LAST_AND_UNUSED_RTX_CODE) |
d60819f8 CB |
2179 | { |
2180 | operands[1] = op1h; | |
2181 | operands[2] = op2h; | |
2182 | } | |
2183 | ||
78d310c2 R |
2184 | operands[3] = skip_label = gen_label_rtx (); |
2185 | expand_cbranchsi4 (operands, msw_skip, msw_skip_prob); | |
2186 | operands[3] = taken_label; | |
2187 | } | |
2188 | operands[1] = op1l; | |
2189 | operands[2] = op2l; | |
b757d36f | 2190 | if (lsw_taken != LAST_AND_UNUSED_RTX_CODE) |
6f8eec29 R |
2191 | { |
2192 | if (reload_completed | |
99052692 KK |
2193 | && ! arith_reg_or_0_operand (op2l, SImode) |
2194 | && (true_regnum (op1l) || (lsw_taken != EQ && lsw_taken != NE))) | |
2195 | { | |
2196 | emit_move_insn (scratch, operands[2]); | |
2197 | operands[2] = scratch; | |
2198 | } | |
6f8eec29 R |
2199 | expand_cbranchsi4 (operands, lsw_taken, lsw_taken_prob); |
2200 | } | |
b757d36f | 2201 | if (msw_skip != LAST_AND_UNUSED_RTX_CODE) |
78d310c2 R |
2202 | emit_label (skip_label); |
2203 | return true; | |
2204 | } | |
2205 | ||
f90b7a5a PB |
2206 | /* Emit INSN, possibly in a PARALLEL with an USE of fpscr for SH4. */ |
2207 | ||
2208 | static void | |
2209 | sh_emit_set_t_insn (rtx insn, enum machine_mode mode) | |
2210 | { | |
2211 | if ((TARGET_SH4 || TARGET_SH2A) && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
2212 | { | |
2213 | insn = gen_rtx_PARALLEL (VOIDmode, | |
2214 | gen_rtvec (2, insn, | |
2215 | gen_rtx_USE (VOIDmode, get_fpscr_rtx ()))); | |
2216 | (mode == SFmode ? emit_sf_insn : emit_df_insn) (insn); | |
2217 | } | |
2218 | else | |
2219 | emit_insn (insn); | |
2220 | } | |
2221 | ||
8aa2a305 | 2222 | /* Prepare the operands for an scc instruction; make sure that the |
f90b7a5a PB |
2223 | compare has been done and the result is in T_REG. */ |
2224 | void | |
2225 | sh_emit_scc_to_t (enum rtx_code code, rtx op0, rtx op1) | |
8aa2a305 | 2226 | { |
c5c76735 | 2227 | rtx t_reg = gen_rtx_REG (SImode, T_REG); |
8aa2a305 | 2228 | enum rtx_code oldcode = code; |
9374bd85 | 2229 | enum machine_mode mode; |
8aa2a305 JW |
2230 | |
2231 | /* First need a compare insn. */ | |
2232 | switch (code) | |
bc45ade3 | 2233 | { |
8aa2a305 JW |
2234 | case NE: |
2235 | /* It isn't possible to handle this case. */ | |
f5b9e7c9 | 2236 | gcc_unreachable (); |
8aa2a305 JW |
2237 | case LT: |
2238 | code = GT; | |
2239 | break; | |
2240 | case LE: | |
2241 | code = GE; | |
2242 | break; | |
2243 | case LTU: | |
2244 | code = GTU; | |
2245 | break; | |
2246 | case LEU: | |
2247 | code = GEU; | |
2248 | break; | |
10db12e0 R |
2249 | default: |
2250 | break; | |
bc45ade3 | 2251 | } |
8aa2a305 | 2252 | if (code != oldcode) |
b9654711 | 2253 | { |
f90b7a5a PB |
2254 | rtx tmp = op0; |
2255 | op0 = op1; | |
2256 | op1 = tmp; | |
0d7e008e | 2257 | } |
b9654711 | 2258 | |
f90b7a5a | 2259 | mode = GET_MODE (op0); |
9374bd85 | 2260 | if (mode == VOIDmode) |
f90b7a5a | 2261 | mode = GET_MODE (op1); |
9374bd85 | 2262 | |
f90b7a5a | 2263 | op0 = force_reg (mode, op0); |
1245df60 | 2264 | if ((code != EQ && code != NE |
f90b7a5a | 2265 | && (op1 != const0_rtx |
1245df60 | 2266 | || code == GTU || code == GEU || code == LTU || code == LEU)) |
f90b7a5a | 2267 | || (mode == DImode && op1 != const0_rtx) |
3a8699c7 | 2268 | || (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT)) |
f90b7a5a | 2269 | op1 = force_reg (mode, op1); |
0d7e008e | 2270 | |
f90b7a5a PB |
2271 | sh_emit_set_t_insn (gen_rtx_SET (VOIDmode, t_reg, |
2272 | gen_rtx_fmt_ee (code, SImode, op0, op1)), | |
2273 | mode); | |
2274 | } | |
0d7e008e | 2275 | |
f90b7a5a PB |
2276 | rtx |
2277 | sh_emit_cheap_store_flag (enum machine_mode mode, enum rtx_code code, | |
2278 | rtx op0, rtx op1) | |
2279 | { | |
2280 | rtx target = gen_reg_rtx (SImode); | |
2281 | rtx tmp; | |
2282 | ||
2283 | gcc_assert (TARGET_SHMEDIA); | |
2284 | switch (code) | |
2285 | { | |
2286 | case EQ: | |
2287 | case GT: | |
2288 | case LT: | |
2289 | case UNORDERED: | |
2290 | case GTU: | |
2291 | case LTU: | |
2292 | tmp = gen_rtx_fmt_ee (code, SImode, op0, op1); | |
2293 | emit_insn (gen_cstore4_media (target, tmp, op0, op1)); | |
2294 | code = NE; | |
2295 | break; | |
2296 | ||
2297 | case NE: | |
2298 | case GE: | |
2299 | case LE: | |
2300 | case ORDERED: | |
2301 | case GEU: | |
2302 | case LEU: | |
2303 | tmp = gen_rtx_fmt_ee (reverse_condition (code), mode, op0, op1); | |
2304 | emit_insn (gen_cstore4_media (target, tmp, op0, op1)); | |
2305 | code = EQ; | |
2306 | break; | |
2307 | ||
2308 | case UNEQ: | |
2309 | case UNGE: | |
2310 | case UNGT: | |
2311 | case UNLE: | |
2312 | case UNLT: | |
2313 | case LTGT: | |
2314 | return NULL_RTX; | |
2315 | ||
2316 | default: | |
2317 | gcc_unreachable (); | |
2318 | } | |
2319 | ||
2320 | if (mode == DImode) | |
2321 | { | |
2322 | rtx t2 = gen_reg_rtx (DImode); | |
2323 | emit_insn (gen_extendsidi2 (t2, target)); | |
2324 | target = t2; | |
2325 | } | |
2326 | ||
2327 | return gen_rtx_fmt_ee (code, VOIDmode, target, const0_rtx); | |
bc45ade3 SC |
2328 | } |
2329 | ||
8aa2a305 JW |
2330 | /* Called from the md file, set up the operands of a compare instruction. */ |
2331 | ||
2332 | void | |
f90b7a5a | 2333 | sh_emit_compare_and_branch (rtx *operands, enum machine_mode mode) |
bc45ade3 | 2334 | { |
f90b7a5a PB |
2335 | enum rtx_code code = GET_CODE (operands[0]); |
2336 | enum rtx_code branch_code; | |
2337 | rtx op0 = operands[1]; | |
2338 | rtx op1 = operands[2]; | |
2339 | rtx insn, tem; | |
2340 | bool need_ccmpeq = false; | |
2341 | ||
2342 | if (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
bc45ade3 | 2343 | { |
f90b7a5a PB |
2344 | op0 = force_reg (mode, op0); |
2345 | op1 = force_reg (mode, op1); | |
bc45ade3 | 2346 | } |
f90b7a5a | 2347 | else |
1245df60 | 2348 | { |
f90b7a5a PB |
2349 | if (code != EQ || mode == DImode) |
2350 | { | |
2351 | /* Force args into regs, since we can't use constants here. */ | |
2352 | op0 = force_reg (mode, op0); | |
2353 | if (op1 != const0_rtx || code == GTU || code == GEU) | |
2354 | op1 = force_reg (mode, op1); | |
2355 | } | |
1245df60 | 2356 | } |
f90b7a5a PB |
2357 | |
2358 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
2359 | { | |
2360 | if (code == LT | |
2361 | || (code == LE && TARGET_IEEE && TARGET_SH2E) | |
2362 | || (code == GE && !(TARGET_IEEE && TARGET_SH2E))) | |
2363 | { | |
2364 | tem = op0, op0 = op1, op1 = tem; | |
2365 | code = swap_condition (code); | |
2366 | } | |
2367 | ||
2368 | /* GE becomes fcmp/gt+fcmp/eq, for SH2E and TARGET_IEEE only. */ | |
2369 | if (code == GE) | |
2370 | { | |
2371 | gcc_assert (TARGET_IEEE && TARGET_SH2E); | |
2372 | need_ccmpeq = true; | |
2373 | code = GT; | |
2374 | } | |
2375 | ||
2376 | /* Now we can have EQ, NE, GT, LE. NE and LE are then transformed | |
2377 | to EQ/GT respectively. */ | |
2378 | gcc_assert (code == EQ || code == GT || code == NE || code == LE); | |
2379 | } | |
2380 | ||
2381 | switch (code) | |
2382 | { | |
2383 | case EQ: | |
2384 | case GT: | |
2385 | case GE: | |
2386 | case GTU: | |
2387 | case GEU: | |
2388 | branch_code = code; | |
2389 | break; | |
2390 | case NE: | |
2391 | case LT: | |
2392 | case LE: | |
2393 | case LTU: | |
2394 | case LEU: | |
2395 | branch_code = reverse_condition (code); | |
2396 | break; | |
2397 | default: | |
2398 | gcc_unreachable (); | |
2399 | } | |
2400 | ||
2401 | insn = gen_rtx_SET (VOIDmode, | |
2402 | gen_rtx_REG (SImode, T_REG), | |
2403 | gen_rtx_fmt_ee (branch_code, SImode, op0, op1)); | |
2404 | ||
2405 | sh_emit_set_t_insn (insn, mode); | |
2406 | if (need_ccmpeq) | |
2407 | sh_emit_set_t_insn (gen_ieee_ccmpeqsf_t (op0, op1), mode); | |
2408 | ||
2409 | if (branch_code == code) | |
2410 | emit_jump_insn (gen_branch_true (operands[3])); | |
1245df60 | 2411 | else |
f90b7a5a PB |
2412 | emit_jump_insn (gen_branch_false (operands[3])); |
2413 | } | |
2414 | ||
2415 | void | |
2416 | sh_emit_compare_and_set (rtx *operands, enum machine_mode mode) | |
2417 | { | |
2418 | enum rtx_code code = GET_CODE (operands[1]); | |
2419 | rtx op0 = operands[2]; | |
2420 | rtx op1 = operands[3]; | |
2421 | rtx lab = NULL_RTX; | |
2422 | bool invert = false; | |
2423 | rtx tem; | |
2424 | ||
2425 | op0 = force_reg (mode, op0); | |
2426 | if ((code != EQ && code != NE | |
2427 | && (op1 != const0_rtx | |
2428 | || code == GTU || code == GEU || code == LTU || code == LEU)) | |
2429 | || (mode == DImode && op1 != const0_rtx) | |
2430 | || (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT)) | |
2431 | op1 = force_reg (mode, op1); | |
2432 | ||
2433 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
225e4f43 | 2434 | { |
f90b7a5a PB |
2435 | if (code == LT || code == LE) |
2436 | { | |
2437 | code = swap_condition (code); | |
2438 | tem = op0, op0 = op1, op1 = tem; | |
2439 | } | |
2440 | if (code == GE) | |
2441 | { | |
2442 | if (TARGET_IEEE) | |
2443 | { | |
2444 | lab = gen_label_rtx (); | |
2445 | sh_emit_scc_to_t (EQ, op0, op1); | |
2446 | emit_jump_insn (gen_branch_true (lab)); | |
2447 | code = GT; | |
2448 | } | |
2449 | else | |
2450 | { | |
2451 | code = LT; | |
2452 | invert = true; | |
2453 | } | |
2454 | } | |
225e4f43 | 2455 | } |
f90b7a5a PB |
2456 | |
2457 | if (code == NE) | |
2458 | { | |
2459 | code = EQ; | |
2460 | invert = true; | |
2461 | } | |
2462 | ||
2463 | sh_emit_scc_to_t (code, op0, op1); | |
2464 | if (lab) | |
2465 | emit_label (lab); | |
2466 | if (invert) | |
2467 | emit_insn (gen_movnegt (operands[0])); | |
225e4f43 | 2468 | else |
f90b7a5a | 2469 | emit_move_insn (operands[0], gen_rtx_REG (SImode, T_REG)); |
bc45ade3 SC |
2470 | } |
2471 | \f | |
16bea517 | 2472 | /* Functions to output assembly code. */ |
bc45ade3 | 2473 | |
b9654711 | 2474 | /* Return a sequence of instructions to perform DI or DF move. |
bc45ade3 | 2475 | |
b9654711 | 2476 | Since the SH cannot move a DI or DF in one instruction, we have |
16bea517 | 2477 | to take care when we see overlapping source and dest registers. */ |
0d7e008e | 2478 | |
318881c0 | 2479 | const char * |
cf277499 SB |
2480 | output_movedouble (rtx insn ATTRIBUTE_UNUSED, rtx operands[], |
2481 | enum machine_mode mode) | |
bc45ade3 | 2482 | { |
b9654711 SC |
2483 | rtx dst = operands[0]; |
2484 | rtx src = operands[1]; | |
b9654711 | 2485 | |
f3536097 | 2486 | if (MEM_P (dst) |
16bea517 | 2487 | && GET_CODE (XEXP (dst, 0)) == PRE_DEC) |
8aa2a305 JW |
2488 | return "mov.l %T1,%0\n\tmov.l %1,%0"; |
2489 | ||
b9654711 SC |
2490 | if (register_operand (dst, mode) |
2491 | && register_operand (src, mode)) | |
bc45ade3 | 2492 | { |
b9654711 | 2493 | if (REGNO (src) == MACH_REG) |
00f8ff66 | 2494 | return "sts mach,%S0\n\tsts macl,%R0"; |
bc45ade3 | 2495 | |
8aa2a305 JW |
2496 | /* When mov.d r1,r2 do r2->r3 then r1->r2; |
2497 | when mov.d r1,r0 do r1->r0 then r2->r1. */ | |
b9654711 SC |
2498 | |
2499 | if (REGNO (src) + 1 == REGNO (dst)) | |
5325c0fa | 2500 | return "mov %T1,%T0\n\tmov %1,%0"; |
b9654711 | 2501 | else |
5325c0fa | 2502 | return "mov %1,%0\n\tmov %T1,%T0"; |
b9654711 | 2503 | } |
f3536097 | 2504 | else if (CONST_INT_P (src)) |
bc45ade3 | 2505 | { |
8aa2a305 JW |
2506 | if (INTVAL (src) < 0) |
2507 | output_asm_insn ("mov #-1,%S0", operands); | |
bc45ade3 | 2508 | else |
8aa2a305 | 2509 | output_asm_insn ("mov #0,%S0", operands); |
bc45ade3 | 2510 | |
8aa2a305 | 2511 | return "mov %1,%R0"; |
0d7e008e | 2512 | } |
f3536097 | 2513 | else if (MEM_P (src)) |
bc45ade3 | 2514 | { |
8aa2a305 | 2515 | int ptrreg = -1; |
b9654711 SC |
2516 | int dreg = REGNO (dst); |
2517 | rtx inside = XEXP (src, 0); | |
bc45ade3 | 2518 | |
f5b9e7c9 | 2519 | switch (GET_CODE (inside)) |
bc45ade3 | 2520 | { |
f5b9e7c9 NS |
2521 | case REG: |
2522 | ptrreg = REGNO (inside); | |
2523 | break; | |
2524 | ||
2525 | case SUBREG: | |
2526 | ptrreg = subreg_regno (inside); | |
2527 | break; | |
2528 | ||
2529 | case PLUS: | |
8aa2a305 JW |
2530 | ptrreg = REGNO (XEXP (inside, 0)); |
2531 | /* ??? A r0+REG address shouldn't be possible here, because it isn't | |
2532 | an offsettable address. Unfortunately, offsettable addresses use | |
2533 | QImode to check the offset, and a QImode offsettable address | |
2534 | requires r0 for the other operand, which is not currently | |
2535 | supported, so we can't use the 'o' constraint. | |
2536 | Thus we must check for and handle r0+REG addresses here. | |
2537 | We punt for now, since this is likely very rare. */ | |
f3536097 | 2538 | gcc_assert (!REG_P (XEXP (inside, 1))); |
f5b9e7c9 NS |
2539 | break; |
2540 | ||
2541 | case LABEL_REF: | |
2542 | return "mov.l %1,%0\n\tmov.l %1+4,%T0"; | |
2543 | case POST_INC: | |
2544 | return "mov.l %1,%0\n\tmov.l %1,%T0"; | |
2545 | default: | |
2546 | gcc_unreachable (); | |
bc45ade3 | 2547 | } |
bc45ade3 | 2548 | |
8aa2a305 JW |
2549 | /* Work out the safe way to copy. Copy into the second half first. */ |
2550 | if (dreg == ptrreg) | |
2551 | return "mov.l %T1,%T0\n\tmov.l %1,%0"; | |
bc45ade3 SC |
2552 | } |
2553 | ||
00f8ff66 | 2554 | return "mov.l %1,%0\n\tmov.l %T1,%T0"; |
bc45ade3 SC |
2555 | } |
2556 | ||
8aa2a305 JW |
2557 | /* Print an instruction which would have gone into a delay slot after |
2558 | another instruction, but couldn't because the other instruction expanded | |
2559 | into a sequence where putting the slot insn at the end wouldn't work. */ | |
0d7e008e | 2560 | |
8aa2a305 | 2561 | static void |
cf277499 | 2562 | print_slot (rtx insn) |
8aa2a305 | 2563 | { |
c9d691e9 | 2564 | final_scan_insn (XVECEXP (insn, 0, 1), asm_out_file, optimize, 1, NULL); |
b9654711 | 2565 | |
8aa2a305 JW |
2566 | INSN_DELETED_P (XVECEXP (insn, 0, 1)) = 1; |
2567 | } | |
0d7e008e | 2568 | |
318881c0 | 2569 | const char * |
cf277499 | 2570 | output_far_jump (rtx insn, rtx op) |
0d7e008e | 2571 | { |
0a2aaacc | 2572 | struct { rtx lab, reg, op; } this_jmp; |
950a3816 | 2573 | rtx braf_base_lab = NULL_RTX; |
318881c0 | 2574 | const char *jump; |
1245df60 | 2575 | int far; |
9d98a694 | 2576 | int offset = branch_dest (insn) - INSN_ADDRESSES (INSN_UID (insn)); |
10f4f635 | 2577 | rtx prev; |
0d7e008e | 2578 | |
0a2aaacc | 2579 | this_jmp.lab = gen_label_rtx (); |
b9654711 | 2580 | |
1a953d0f R |
2581 | if (TARGET_SH2 |
2582 | && offset >= -32764 | |
2583 | && offset - get_attr_length (insn) <= 32766) | |
1245df60 R |
2584 | { |
2585 | far = 0; | |
e6dfd05f | 2586 | jump = "mov.w %O0,%1; braf %1"; |
1245df60 R |
2587 | } |
2588 | else | |
2589 | { | |
2590 | far = 1; | |
1a66cd67 | 2591 | if (flag_pic) |
e6dfd05f AO |
2592 | { |
2593 | if (TARGET_SH2) | |
2594 | jump = "mov.l %O0,%1; braf %1"; | |
2595 | else | |
2596 | jump = "mov.l r0,@-r15; mova %O0,r0; mov.l @r0,%1; add r0,%1; mov.l @r15+,r0; jmp @%1"; | |
2597 | } | |
1a66cd67 | 2598 | else |
e6dfd05f | 2599 | jump = "mov.l %O0,%1; jmp @%1"; |
1245df60 R |
2600 | } |
2601 | /* If we have a scratch register available, use it. */ | |
f3536097 | 2602 | if (NONJUMP_INSN_P ((prev = prev_nonnote_insn (insn))) |
10f4f635 | 2603 | && INSN_CODE (prev) == CODE_FOR_indirect_jump_scratch) |
1245df60 | 2604 | { |
0a2aaacc KG |
2605 | this_jmp.reg = SET_DEST (XVECEXP (PATTERN (prev), 0, 0)); |
2606 | if (REGNO (this_jmp.reg) == R0_REG && flag_pic && ! TARGET_SH2) | |
e6dfd05f | 2607 | jump = "mov.l r1,@-r15; mova %O0,r0; mov.l @r0,r1; add r1,r0; mov.l @r15+,r1; jmp @%1"; |
0a2aaacc | 2608 | output_asm_insn (jump, &this_jmp.lab); |
1245df60 R |
2609 | if (dbr_sequence_length ()) |
2610 | print_slot (final_sequence); | |
2611 | else | |
2612 | output_asm_insn ("nop", 0); | |
2613 | } | |
2614 | else | |
2615 | { | |
2616 | /* Output the delay slot insn first if any. */ | |
2617 | if (dbr_sequence_length ()) | |
2618 | print_slot (final_sequence); | |
2619 | ||
0a2aaacc | 2620 | this_jmp.reg = gen_rtx_REG (SImode, 13); |
fa5322fa AO |
2621 | /* We must keep the stack aligned to 8-byte boundaries on SH5. |
2622 | Fortunately, MACL is fixed and call-clobbered, and we never | |
2623 | need its value across jumps, so save r13 in it instead of in | |
2624 | the stack. */ | |
2625 | if (TARGET_SH5) | |
2626 | output_asm_insn ("lds r13, macl", 0); | |
2627 | else | |
2628 | output_asm_insn ("mov.l r13,@-r15", 0); | |
0a2aaacc | 2629 | output_asm_insn (jump, &this_jmp.lab); |
fa5322fa AO |
2630 | if (TARGET_SH5) |
2631 | output_asm_insn ("sts macl, r13", 0); | |
2632 | else | |
2633 | output_asm_insn ("mov.l @r15+,r13", 0); | |
1245df60 | 2634 | } |
e6dfd05f AO |
2635 | if (far && flag_pic && TARGET_SH2) |
2636 | { | |
2637 | braf_base_lab = gen_label_rtx (); | |
4977bab6 | 2638 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
e6dfd05f AO |
2639 | CODE_LABEL_NUMBER (braf_base_lab)); |
2640 | } | |
1245df60 R |
2641 | if (far) |
2642 | output_asm_insn (".align 2", 0); | |
0a2aaacc KG |
2643 | (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (this_jmp.lab)); |
2644 | this_jmp.op = op; | |
1a66cd67 | 2645 | if (far && flag_pic) |
e6dfd05f AO |
2646 | { |
2647 | if (TARGET_SH2) | |
0a2aaacc KG |
2648 | this_jmp.lab = braf_base_lab; |
2649 | output_asm_insn (".long %O2-%O0", &this_jmp.lab); | |
e6dfd05f | 2650 | } |
1a66cd67 | 2651 | else |
0a2aaacc | 2652 | output_asm_insn (far ? ".long %O2" : ".word %O2-%O0", &this_jmp.lab); |
0d7e008e SC |
2653 | return ""; |
2654 | } | |
bc45ade3 | 2655 | |
8aa2a305 JW |
2656 | /* Local label counter, used for constants in the pool and inside |
2657 | pattern branches. */ | |
2658 | ||
2659 | static int lf = 100; | |
2660 | ||
2661 | /* Output code for ordinary branches. */ | |
2662 | ||
318881c0 | 2663 | const char * |
cf277499 | 2664 | output_branch (int logic, rtx insn, rtx *operands) |
bc45ade3 | 2665 | { |
afbc5d86 | 2666 | switch (get_attr_length (insn)) |
bc45ade3 | 2667 | { |
33f7f353 JR |
2668 | case 6: |
2669 | /* This can happen if filling the delay slot has caused a forward | |
2670 | branch to exceed its range (we could reverse it, but only | |
2671 | when we know we won't overextend other branches; this should | |
2672 | best be handled by relaxation). | |
2673 | It can also happen when other condbranches hoist delay slot insn | |
1245df60 R |
2674 | from their destination, thus leading to code size increase. |
2675 | But the branch will still be in the range -4092..+4098 bytes. */ | |
bc45ade3 | 2676 | |
33f7f353 | 2677 | if (! TARGET_RELAX) |
1245df60 | 2678 | { |
33f7f353 JR |
2679 | int label = lf++; |
2680 | /* The call to print_slot will clobber the operands. */ | |
2681 | rtx op0 = operands[0]; | |
f676971a | 2682 | |
33f7f353 JR |
2683 | /* If the instruction in the delay slot is annulled (true), then |
2684 | there is no delay slot where we can put it now. The only safe | |
2685 | place for it is after the label. final will do that by default. */ | |
f676971a | 2686 | |
33f7f353 | 2687 | if (final_sequence |
bee9efbc KK |
2688 | && ! INSN_ANNULLED_BRANCH_P (XVECEXP (final_sequence, 0, 0)) |
2689 | && get_attr_length (XVECEXP (final_sequence, 0, 1))) | |
33f7f353 | 2690 | { |
afbc5d86 | 2691 | asm_fprintf (asm_out_file, "\tb%s%ss\t%LLF%d\n", logic ? "f" : "t", |
33f7f353 JR |
2692 | ASSEMBLER_DIALECT ? "/" : ".", label); |
2693 | print_slot (final_sequence); | |
2694 | } | |
2695 | else | |
afbc5d86 | 2696 | asm_fprintf (asm_out_file, "\tb%s\t%LLF%d\n", logic ? "f" : "t", label); |
f676971a | 2697 | |
afbc5d86 R |
2698 | output_asm_insn ("bra\t%l0", &op0); |
2699 | fprintf (asm_out_file, "\tnop\n"); | |
832a3292 | 2700 | (*targetm.asm_out.internal_label) (asm_out_file, "LF", label); |
f676971a | 2701 | |
33f7f353 | 2702 | return ""; |
1245df60 | 2703 | } |
33f7f353 JR |
2704 | /* When relaxing, handle this like a short branch. The linker |
2705 | will fix it up if it still doesn't fit after relaxation. */ | |
2706 | case 2: | |
2707 | return logic ? "bt%.\t%l0" : "bf%.\t%l0"; | |
3a8699c7 AO |
2708 | |
2709 | /* These are for SH2e, in which we have to account for the | |
2710 | extra nop because of the hardware bug in annulled branches. */ | |
2711 | case 8: | |
2712 | if (! TARGET_RELAX) | |
2713 | { | |
2714 | int label = lf++; | |
2715 | ||
f5b9e7c9 NS |
2716 | gcc_assert (!final_sequence |
2717 | || !(INSN_ANNULLED_BRANCH_P | |
2718 | (XVECEXP (final_sequence, 0, 0)))); | |
3a8699c7 AO |
2719 | asm_fprintf (asm_out_file, "b%s%ss\t%LLF%d\n", |
2720 | logic ? "f" : "t", | |
2721 | ASSEMBLER_DIALECT ? "/" : ".", label); | |
2722 | fprintf (asm_out_file, "\tnop\n"); | |
2723 | output_asm_insn ("bra\t%l0", operands); | |
2724 | fprintf (asm_out_file, "\tnop\n"); | |
2725 | (*targetm.asm_out.internal_label) (asm_out_file, "LF", label); | |
2726 | ||
2727 | return ""; | |
2728 | } | |
2729 | /* When relaxing, fall through. */ | |
2730 | case 4: | |
2731 | { | |
2732 | char buffer[10]; | |
f676971a | 2733 | |
3a8699c7 AO |
2734 | sprintf (buffer, "b%s%ss\t%%l0", |
2735 | logic ? "t" : "f", | |
2736 | ASSEMBLER_DIALECT ? "/" : "."); | |
2737 | output_asm_insn (buffer, &operands[0]); | |
2738 | return "nop"; | |
2739 | } | |
2740 | ||
33f7f353 | 2741 | default: |
afbc5d86 R |
2742 | /* There should be no longer branches now - that would |
2743 | indicate that something has destroyed the branches set | |
2744 | up in machine_dependent_reorg. */ | |
f5b9e7c9 | 2745 | gcc_unreachable (); |
1245df60 | 2746 | } |
1245df60 | 2747 | } |
bc45ade3 | 2748 | |
0a2aaacc | 2749 | /* Output a code sequence for INSN using TEMPL with OPERANDS; but before, |
78d310c2 R |
2750 | fill in operands 9 as a label to the successor insn. |
2751 | We try to use jump threading where possible. | |
2752 | IF CODE matches the comparison in the IF_THEN_ELSE of a following jump, | |
2753 | we assume the jump is taken. I.e. EQ means follow jmp and bf, NE means | |
2754 | follow jmp and bt, if the address is in range. */ | |
318881c0 | 2755 | const char * |
0a2aaacc | 2756 | output_branchy_insn (enum rtx_code code, const char *templ, |
cf277499 | 2757 | rtx insn, rtx *operands) |
1245df60 R |
2758 | { |
2759 | rtx next_insn = NEXT_INSN (insn); | |
79b2746a | 2760 | |
f3536097 | 2761 | if (next_insn && JUMP_P (next_insn) && condjump_p (next_insn)) |
1245df60 R |
2762 | { |
2763 | rtx src = SET_SRC (PATTERN (next_insn)); | |
2764 | if (GET_CODE (src) == IF_THEN_ELSE && GET_CODE (XEXP (src, 0)) != code) | |
2765 | { | |
2766 | /* Following branch not taken */ | |
2767 | operands[9] = gen_label_rtx (); | |
2768 | emit_label_after (operands[9], next_insn); | |
9d98a694 AO |
2769 | INSN_ADDRESSES_NEW (operands[9], |
2770 | INSN_ADDRESSES (INSN_UID (next_insn)) | |
2771 | + get_attr_length (next_insn)); | |
0a2aaacc | 2772 | return templ; |
1245df60 R |
2773 | } |
2774 | else | |
2775 | { | |
33f7f353 | 2776 | int offset = (branch_dest (next_insn) |
9d98a694 | 2777 | - INSN_ADDRESSES (INSN_UID (next_insn)) + 4); |
33f7f353 | 2778 | if (offset >= -252 && offset <= 258) |
1245df60 R |
2779 | { |
2780 | if (GET_CODE (src) == IF_THEN_ELSE) | |
2781 | /* branch_true */ | |
2782 | src = XEXP (src, 1); | |
2783 | operands[9] = src; | |
0a2aaacc | 2784 | return templ; |
1245df60 R |
2785 | } |
2786 | } | |
bc45ade3 | 2787 | } |
1245df60 R |
2788 | operands[9] = gen_label_rtx (); |
2789 | emit_label_after (operands[9], insn); | |
9d98a694 AO |
2790 | INSN_ADDRESSES_NEW (operands[9], |
2791 | INSN_ADDRESSES (INSN_UID (insn)) | |
2792 | + get_attr_length (insn)); | |
0a2aaacc | 2793 | return templ; |
1245df60 | 2794 | } |
e4fa6b06 | 2795 | |
318881c0 | 2796 | const char * |
cf277499 | 2797 | output_ieee_ccmpeq (rtx insn, rtx *operands) |
1245df60 | 2798 | { |
9fd0c419 R |
2799 | return output_branchy_insn (NE, "bt\t%l9\n\tfcmp/eq\t%1,%0", |
2800 | insn, operands); | |
bc45ade3 | 2801 | } |
8e87e161 | 2802 | \f |
1bc7c5b6 | 2803 | /* Output the start of the assembler file. */ |
8aa2a305 | 2804 | |
1bc7c5b6 | 2805 | static void |
cf277499 | 2806 | sh_file_start (void) |
bc45ade3 | 2807 | { |
1bc7c5b6 ZW |
2808 | default_file_start (); |
2809 | ||
2810 | if (TARGET_ELF) | |
2811 | /* We need to show the text section with the proper | |
2812 | attributes as in TEXT_SECTION_ASM_OP, before dwarf2out | |
0fe7abcc | 2813 | emits it without attributes in TEXT_SECTION_ASM_OP, else GAS |
1bc7c5b6 ZW |
2814 | will complain. We can teach GAS specifically about the |
2815 | default attributes for our choice of text section, but | |
2816 | then we would have to change GAS again if/when we change | |
2817 | the text section name. */ | |
2818 | fprintf (asm_out_file, "%s\n", TEXT_SECTION_ASM_OP); | |
2819 | else | |
2820 | /* Switch to the data section so that the coffsem symbol | |
2821 | isn't in the text section. */ | |
d6b5193b | 2822 | switch_to_section (data_section); |
b9654711 | 2823 | |
8aa2a305 | 2824 | if (TARGET_LITTLE_ENDIAN) |
1bc7c5b6 | 2825 | fputs ("\t.little\n", asm_out_file); |
fa5322fa | 2826 | |
1bc7c5b6 ZW |
2827 | if (!TARGET_ELF) |
2828 | { | |
2829 | if (TARGET_SHCOMPACT) | |
2830 | fputs ("\t.mode\tSHcompact\n", asm_out_file); | |
2831 | else if (TARGET_SHMEDIA) | |
2832 | fprintf (asm_out_file, "\t.mode\tSHmedia\n\t.abi\t%i\n", | |
2833 | TARGET_SHMEDIA64 ? 64 : 32); | |
2834 | } | |
bc45ade3 | 2835 | } |
0d7e008e | 2836 | \f |
9f3a9a08 KK |
2837 | /* Check if PAT includes UNSPEC_CALLER unspec pattern. */ |
2838 | ||
2839 | static bool | |
cf277499 | 2840 | unspec_caller_rtx_p (rtx pat) |
9f3a9a08 | 2841 | { |
dc3ba671 RS |
2842 | rtx base, offset; |
2843 | int i; | |
2844 | ||
2845 | split_const (pat, &base, &offset); | |
2846 | if (GET_CODE (base) == UNSPEC) | |
9f3a9a08 | 2847 | { |
dc3ba671 | 2848 | if (XINT (base, 1) == UNSPEC_CALLER) |
9f3a9a08 | 2849 | return true; |
dc3ba671 RS |
2850 | for (i = 0; i < XVECLEN (base, 0); i++) |
2851 | if (unspec_caller_rtx_p (XVECEXP (base, 0, i))) | |
2852 | return true; | |
9f3a9a08 | 2853 | } |
9f3a9a08 KK |
2854 | return false; |
2855 | } | |
2856 | ||
2857 | /* Indicate that INSN cannot be duplicated. This is true for insn | |
569b7f6a | 2858 | that generates a unique label. */ |
9f3a9a08 KK |
2859 | |
2860 | static bool | |
cf277499 | 2861 | sh_cannot_copy_insn_p (rtx insn) |
9f3a9a08 KK |
2862 | { |
2863 | rtx pat; | |
2864 | ||
2865 | if (!reload_completed || !flag_pic) | |
2866 | return false; | |
2867 | ||
f3536097 | 2868 | if (!NONJUMP_INSN_P (insn)) |
9f3a9a08 KK |
2869 | return false; |
2870 | if (asm_noperands (insn) >= 0) | |
2871 | return false; | |
2872 | ||
2873 | pat = PATTERN (insn); | |
2874 | if (GET_CODE (pat) != SET) | |
2875 | return false; | |
2876 | pat = SET_SRC (pat); | |
2877 | ||
2878 | if (unspec_caller_rtx_p (pat)) | |
2879 | return true; | |
2880 | ||
2881 | return false; | |
2882 | } | |
2883 | \f | |
16bea517 | 2884 | /* Actual number of instructions used to make a shift by N. */ |
0b5826ac | 2885 | static const char ashiftrt_insns[] = |
16bea517 JW |
2886 | { 0,1,2,3,4,5,8,8,8,8,8,8,8,8,8,8,2,3,4,5,8,8,8,8,8,8,8,8,8,8,8,2}; |
2887 | ||
2888 | /* Left shift and logical right shift are the same. */ | |
0b5826ac | 2889 | static const char shift_insns[] = |
16bea517 | 2890 | { 0,1,1,2,2,3,3,4,1,2,2,3,3,4,3,3,1,2,2,3,3,4,3,3,2,3,3,4,4,4,3,3}; |
8aa2a305 | 2891 | |
16bea517 JW |
2892 | /* Individual shift amounts needed to get the above length sequences. |
2893 | One bit right shifts clobber the T bit, so when possible, put one bit | |
2894 | shifts in the middle of the sequence, so the ends are eligible for | |
2895 | branch delay slots. */ | |
0139adca | 2896 | static const short shift_amounts[32][5] = { |
16bea517 JW |
2897 | {0}, {1}, {2}, {2, 1}, |
2898 | {2, 2}, {2, 1, 2}, {2, 2, 2}, {2, 2, 1, 2}, | |
2899 | {8}, {8, 1}, {8, 2}, {8, 1, 2}, | |
2900 | {8, 2, 2}, {8, 2, 1, 2}, {8, -2, 8}, {8, -1, 8}, | |
2901 | {16}, {16, 1}, {16, 2}, {16, 1, 2}, | |
2902 | {16, 2, 2}, {16, 2, 1, 2}, {16, -2, 8}, {16, -1, 8}, | |
2903 | {16, 8}, {16, 1, 8}, {16, 8, 2}, {16, 8, 1, 2}, | |
1d3534c0 | 2904 | {16, 8, 2, 2}, {16, -1, -2, 16}, {16, -2, 16}, {16, -1, 16}}; |
16bea517 | 2905 | |
8d481241 JW |
2906 | /* Likewise, but for shift amounts < 16, up to three highmost bits |
2907 | might be clobbered. This is typically used when combined with some | |
2908 | kind of sign or zero extension. */ | |
f676971a | 2909 | |
0b5826ac | 2910 | static const char ext_shift_insns[] = |
8d481241 JW |
2911 | { 0,1,1,2,2,3,2,2,1,2,2,3,3,3,2,2,1,2,2,3,3,4,3,3,2,3,3,4,4,4,3,3}; |
2912 | ||
0b5826ac | 2913 | static const short ext_shift_amounts[32][4] = { |
8d481241 JW |
2914 | {0}, {1}, {2}, {2, 1}, |
2915 | {2, 2}, {2, 1, 2}, {8, -2}, {8, -1}, | |
2916 | {8}, {8, 1}, {8, 2}, {8, 1, 2}, | |
2917 | {8, 2, 2}, {16, -2, -1}, {16, -2}, {16, -1}, | |
2918 | {16}, {16, 1}, {16, 2}, {16, 1, 2}, | |
2919 | {16, 2, 2}, {16, 2, 1, 2}, {16, -2, 8}, {16, -1, 8}, | |
2920 | {16, 8}, {16, 1, 8}, {16, 8, 2}, {16, 8, 1, 2}, | |
2921 | {16, 8, 2, 2}, {16, -1, -2, 16}, {16, -2, 16}, {16, -1, 16}}; | |
2922 | ||
ae9d19c0 JR |
2923 | /* Assuming we have a value that has been sign-extended by at least one bit, |
2924 | can we use the ext_shift_amounts with the last shift turned to an arithmetic shift | |
2925 | to shift it by N without data loss, and quicker than by other means? */ | |
2926 | #define EXT_SHIFT_SIGNED(n) (((n) | 8) == 15) | |
2927 | ||
16bea517 JW |
2928 | /* This is used in length attributes in sh.md to help compute the length |
2929 | of arbitrary constant shift instructions. */ | |
bc45ade3 | 2930 | |
16bea517 | 2931 | int |
cf277499 | 2932 | shift_insns_rtx (rtx insn) |
16bea517 JW |
2933 | { |
2934 | rtx set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
6c34a092 | 2935 | int shift_count = INTVAL (XEXP (set_src, 1)) & 31; |
16bea517 | 2936 | enum rtx_code shift_code = GET_CODE (set_src); |
00f8ff66 | 2937 | |
16bea517 JW |
2938 | switch (shift_code) |
2939 | { | |
2940 | case ASHIFTRT: | |
2941 | return ashiftrt_insns[shift_count]; | |
2942 | case LSHIFTRT: | |
2943 | case ASHIFT: | |
2944 | return shift_insns[shift_count]; | |
2945 | default: | |
f5b9e7c9 | 2946 | gcc_unreachable (); |
16bea517 JW |
2947 | } |
2948 | } | |
2949 | ||
16bea517 | 2950 | /* Return the cost of a shift. */ |
bc45ade3 | 2951 | |
3c50106f | 2952 | static inline int |
cf277499 | 2953 | shiftcosts (rtx x) |
bc45ade3 | 2954 | { |
3d422cb1 | 2955 | int value; |
8aa2a305 | 2956 | |
fa5322fa AO |
2957 | if (TARGET_SHMEDIA) |
2958 | return 1; | |
2959 | ||
c762ab6e R |
2960 | if (GET_MODE_SIZE (GET_MODE (x)) > UNITS_PER_WORD) |
2961 | { | |
2962 | if (GET_MODE (x) == DImode | |
f3536097 | 2963 | && CONST_INT_P (XEXP (x, 1)) |
c762ab6e R |
2964 | && INTVAL (XEXP (x, 1)) == 1) |
2965 | return 2; | |
2966 | ||
2967 | /* Everything else is invalid, because there is no pattern for it. */ | |
cb2eb96f | 2968 | return MAX_COST; |
c762ab6e | 2969 | } |
16bea517 | 2970 | /* If shift by a non constant, then this will be expensive. */ |
f3536097 | 2971 | if (!CONST_INT_P (XEXP (x, 1))) |
1245df60 | 2972 | return SH_DYNAMIC_SHIFT_COST; |
bc45ade3 | 2973 | |
6c34a092 PB |
2974 | /* Otherwise, return the true cost in instructions. Cope with out of range |
2975 | shift counts more or less arbitrarily. */ | |
2976 | value = INTVAL (XEXP (x, 1)) & 31; | |
3d422cb1 | 2977 | |
8aa2a305 | 2978 | if (GET_CODE (x) == ASHIFTRT) |
49b6d06b JW |
2979 | { |
2980 | int cost = ashiftrt_insns[value]; | |
2981 | /* If SH3, then we put the constant in a reg and use shad. */ | |
1245df60 R |
2982 | if (cost > 1 + SH_DYNAMIC_SHIFT_COST) |
2983 | cost = 1 + SH_DYNAMIC_SHIFT_COST; | |
49b6d06b JW |
2984 | return cost; |
2985 | } | |
8aa2a305 JW |
2986 | else |
2987 | return shift_insns[value]; | |
0d7e008e | 2988 | } |
b9654711 | 2989 | |
8aa2a305 JW |
2990 | /* Return the cost of an AND operation. */ |
2991 | ||
3c50106f | 2992 | static inline int |
cf277499 | 2993 | andcosts (rtx x) |
0d7e008e SC |
2994 | { |
2995 | int i; | |
8aa2a305 | 2996 | |
a6f71af5 | 2997 | /* Anding with a register is a single cycle and instruction. */ |
f3536097 | 2998 | if (!CONST_INT_P (XEXP (x, 1))) |
a6f71af5 | 2999 | return 1; |
8aa2a305 JW |
3000 | |
3001 | i = INTVAL (XEXP (x, 1)); | |
fa5322fa AO |
3002 | |
3003 | if (TARGET_SHMEDIA) | |
3004 | { | |
32a7ab3d KK |
3005 | if (satisfies_constraint_I10 (XEXP (x, 1)) |
3006 | || satisfies_constraint_J16 (XEXP (x, 1))) | |
fa5322fa AO |
3007 | return 1; |
3008 | else | |
f40751dd | 3009 | return 1 + rtx_cost (XEXP (x, 1), AND, !optimize_size); |
fa5322fa AO |
3010 | } |
3011 | ||
a6f71af5 | 3012 | /* These constants are single cycle extu.[bw] instructions. */ |
0d7e008e | 3013 | if (i == 0xff || i == 0xffff) |
a6f71af5 | 3014 | return 1; |
735cb76e | 3015 | /* Constants that can be used in an and immediate instruction in a single |
a6f71af5 | 3016 | cycle, but this requires r0, so make it a little more expensive. */ |
735cb76e | 3017 | if (CONST_OK_FOR_K08 (i)) |
0d7e008e | 3018 | return 2; |
a6f71af5 JW |
3019 | /* Constants that can be loaded with a mov immediate and an and. |
3020 | This case is probably unnecessary. */ | |
735cb76e | 3021 | if (CONST_OK_FOR_I08 (i)) |
a6f71af5 JW |
3022 | return 2; |
3023 | /* Any other constants requires a 2 cycle pc-relative load plus an and. | |
3024 | This case is probably unnecessary. */ | |
3025 | return 3; | |
0d7e008e | 3026 | } |
d3ae8277 | 3027 | |
630c79be BS |
3028 | /* Return the cost of an addition or a subtraction. */ |
3029 | ||
3c50106f | 3030 | static inline int |
cf277499 | 3031 | addsubcosts (rtx x) |
630c79be BS |
3032 | { |
3033 | /* Adding a register is a single cycle insn. */ | |
f3536097 | 3034 | if (REG_P (XEXP (x, 1)) |
50ceefc2 | 3035 | || GET_CODE (XEXP (x, 1)) == SUBREG) |
630c79be BS |
3036 | return 1; |
3037 | ||
3038 | /* Likewise for small constants. */ | |
f3536097 | 3039 | if (CONST_INT_P (XEXP (x, 1)) |
fa5322fa | 3040 | && CONST_OK_FOR_ADD (INTVAL (XEXP (x, 1)))) |
630c79be BS |
3041 | return 1; |
3042 | ||
fa5322fa AO |
3043 | if (TARGET_SHMEDIA) |
3044 | switch (GET_CODE (XEXP (x, 1))) | |
3045 | { | |
3046 | case CONST: | |
3047 | case LABEL_REF: | |
3048 | case SYMBOL_REF: | |
3049 | return TARGET_SHMEDIA64 ? 5 : 3; | |
3050 | ||
3051 | case CONST_INT: | |
735cb76e | 3052 | if (CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1)))) |
fa5322fa | 3053 | return 2; |
735cb76e | 3054 | else if (CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1)) >> 16)) |
fa5322fa | 3055 | return 3; |
735cb76e | 3056 | else if (CONST_OK_FOR_I16 ((INTVAL (XEXP (x, 1)) >> 16) >> 16)) |
fa5322fa AO |
3057 | return 4; |
3058 | ||
3059 | /* Fall through. */ | |
3060 | default: | |
832a3292 | 3061 | return 5; |
fa5322fa AO |
3062 | } |
3063 | ||
630c79be BS |
3064 | /* Any other constant requires a 2 cycle pc-relative load plus an |
3065 | addition. */ | |
3066 | return 3; | |
3067 | } | |
3068 | ||
16bea517 | 3069 | /* Return the cost of a multiply. */ |
3c50106f | 3070 | static inline int |
cf277499 | 3071 | multcosts (rtx x ATTRIBUTE_UNUSED) |
0d7e008e | 3072 | { |
7dd2f19b RS |
3073 | if (sh_multcost >= 0) |
3074 | return sh_multcost; | |
fa5322fa | 3075 | if (TARGET_SHMEDIA) |
73a4d10b R |
3076 | /* ??? We have a mul insn, but it has a latency of three, and doesn't |
3077 | accept constants. Ideally, we would use a cost of one or two and | |
3078 | add the cost of the operand, but disregard the latter when inside loops | |
3079 | and loop invariant code motion is still to follow. | |
3080 | Using a multiply first and splitting it later if it's a loss | |
3081 | doesn't work because of different sign / zero extension semantics | |
3082 | of multiplies vs. shifts. */ | |
3020190e | 3083 | return optimize_size ? 2 : 3; |
fa5322fa | 3084 | |
0d7e008e | 3085 | if (TARGET_SH2) |
d3ae8277 SC |
3086 | { |
3087 | /* We have a mul insn, so we can never take more than the mul and the | |
a7771f78 | 3088 | read of the mac reg, but count more because of the latency and extra |
16bea517 | 3089 | reg usage. */ |
3020190e | 3090 | if (optimize_size) |
8e87e161 | 3091 | return 2; |
a7771f78 | 3092 | return 3; |
d3ae8277 SC |
3093 | } |
3094 | ||
a7771f78 | 3095 | /* If we're aiming at small code, then just count the number of |
16bea517 | 3096 | insns in a multiply call sequence. */ |
3020190e | 3097 | if (optimize_size) |
8aa2a305 | 3098 | return 5; |
d3ae8277 | 3099 | |
16bea517 | 3100 | /* Otherwise count all the insns in the routine we'd be calling too. */ |
d3ae8277 | 3101 | return 20; |
0d7e008e | 3102 | } |
b9654711 | 3103 | |
3c50106f RH |
3104 | /* Compute a (partial) cost for rtx X. Return true if the complete |
3105 | cost has been computed, and false if subexpressions should be | |
3106 | scanned. In either case, *TOTAL contains the cost result. */ | |
3107 | ||
3108 | static bool | |
f40751dd JH |
3109 | sh_rtx_costs (rtx x, int code, int outer_code, int *total, |
3110 | bool speed ATTRIBUTE_UNUSED) | |
3c50106f RH |
3111 | { |
3112 | switch (code) | |
3113 | { | |
3114 | case CONST_INT: | |
3115 | if (TARGET_SHMEDIA) | |
3116 | { | |
3117 | if (INTVAL (x) == 0) | |
3118 | *total = 0; | |
3119 | else if (outer_code == AND && and_operand ((x), DImode)) | |
3120 | *total = 0; | |
3121 | else if ((outer_code == IOR || outer_code == XOR | |
3122 | || outer_code == PLUS) | |
735cb76e | 3123 | && CONST_OK_FOR_I10 (INTVAL (x))) |
3c50106f | 3124 | *total = 0; |
735cb76e | 3125 | else if (CONST_OK_FOR_I16 (INTVAL (x))) |
3c50106f | 3126 | *total = COSTS_N_INSNS (outer_code != SET); |
735cb76e | 3127 | else if (CONST_OK_FOR_I16 (INTVAL (x) >> 16)) |
73a4d10b | 3128 | *total = COSTS_N_INSNS ((outer_code != SET) + 1); |
735cb76e | 3129 | else if (CONST_OK_FOR_I16 ((INTVAL (x) >> 16) >> 16)) |
42201282 | 3130 | *total = COSTS_N_INSNS ((outer_code != SET) + 2); |
3c50106f | 3131 | else |
42201282 | 3132 | *total = COSTS_N_INSNS ((outer_code != SET) + 3); |
3c50106f RH |
3133 | return true; |
3134 | } | |
735cb76e | 3135 | if (CONST_OK_FOR_I08 (INTVAL (x))) |
3c50106f RH |
3136 | *total = 0; |
3137 | else if ((outer_code == AND || outer_code == IOR || outer_code == XOR) | |
735cb76e | 3138 | && CONST_OK_FOR_K08 (INTVAL (x))) |
3c50106f | 3139 | *total = 1; |
78d310c2 | 3140 | /* prepare_cmp_insn will force costly constants int registers before |
35912544 | 3141 | the cbranch[sd]i4 patterns can see them, so preserve potentially |
78d310c2 R |
3142 | interesting ones not covered by I08 above. */ |
3143 | else if (outer_code == COMPARE | |
3144 | && ((unsigned HOST_WIDE_INT) INTVAL (x) | |
3145 | == (unsigned HOST_WIDE_INT) 0x7fffffff + 1 | |
3146 | || INTVAL (x) == 0x7fffffff | |
3147 | || INTVAL (x) == 0x80 || INTVAL (x) == -0x81)) | |
3148 | *total = 1; | |
3c50106f RH |
3149 | else |
3150 | *total = 8; | |
3151 | return true; | |
3152 | ||
3153 | case CONST: | |
3154 | case LABEL_REF: | |
3155 | case SYMBOL_REF: | |
3156 | if (TARGET_SHMEDIA64) | |
3157 | *total = COSTS_N_INSNS (4); | |
3158 | else if (TARGET_SHMEDIA32) | |
3159 | *total = COSTS_N_INSNS (2); | |
3160 | else | |
3161 | *total = 5; | |
3162 | return true; | |
3163 | ||
3164 | case CONST_DOUBLE: | |
3165 | if (TARGET_SHMEDIA) | |
3166 | *total = COSTS_N_INSNS (4); | |
78d310c2 | 3167 | /* prepare_cmp_insn will force costly constants int registers before |
35912544 | 3168 | the cbranchdi4 pattern can see them, so preserve potentially |
78d310c2 R |
3169 | interesting ones. */ |
3170 | else if (outer_code == COMPARE && GET_MODE (x) == DImode) | |
3171 | *total = 1; | |
3c50106f RH |
3172 | else |
3173 | *total = 10; | |
3174 | return true; | |
73a4d10b R |
3175 | case CONST_VECTOR: |
3176 | if (x == CONST0_RTX (GET_MODE (x))) | |
3177 | *total = 0; | |
3178 | else if (sh_1el_vec (x, VOIDmode)) | |
3179 | *total = outer_code != SET; | |
3180 | if (sh_rep_vec (x, VOIDmode)) | |
3181 | *total = ((GET_MODE_UNIT_SIZE (GET_MODE (x)) + 3) / 4 | |
3182 | + (outer_code != SET)); | |
3183 | *total = COSTS_N_INSNS (3) + (outer_code != SET); | |
3184 | return true; | |
3c50106f RH |
3185 | |
3186 | case PLUS: | |
73a4d10b | 3187 | case MINUS: |
3c50106f RH |
3188 | *total = COSTS_N_INSNS (addsubcosts (x)); |
3189 | return true; | |
3190 | ||
3191 | case AND: | |
3192 | *total = COSTS_N_INSNS (andcosts (x)); | |
3193 | return true; | |
3194 | ||
3195 | case MULT: | |
3196 | *total = COSTS_N_INSNS (multcosts (x)); | |
3197 | return true; | |
3198 | ||
3199 | case ASHIFT: | |
3200 | case ASHIFTRT: | |
3201 | case LSHIFTRT: | |
3202 | *total = COSTS_N_INSNS (shiftcosts (x)); | |
3203 | return true; | |
3204 | ||
3205 | case DIV: | |
3206 | case UDIV: | |
3207 | case MOD: | |
3208 | case UMOD: | |
3209 | *total = COSTS_N_INSNS (20); | |
3210 | return true; | |
3211 | ||
73a4d10b R |
3212 | case PARALLEL: |
3213 | if (sh_1el_vec (x, VOIDmode)) | |
3214 | *total = outer_code != SET; | |
3215 | if (sh_rep_vec (x, VOIDmode)) | |
3216 | *total = ((GET_MODE_UNIT_SIZE (GET_MODE (x)) + 3) / 4 | |
3217 | + (outer_code != SET)); | |
3218 | *total = COSTS_N_INSNS (3) + (outer_code != SET); | |
3219 | return true; | |
3220 | ||
3c50106f RH |
3221 | case FLOAT: |
3222 | case FIX: | |
3223 | *total = 100; | |
3224 | return true; | |
3225 | ||
3226 | default: | |
3227 | return false; | |
3228 | } | |
3229 | } | |
3230 | ||
dcefdf67 RH |
3231 | /* Compute the cost of an address. For the SH, all valid addresses are |
3232 | the same cost. Use a slightly higher cost for reg + reg addressing, | |
3233 | since it increases pressure on r0. */ | |
3234 | ||
3235 | static int | |
f40751dd JH |
3236 | sh_address_cost (rtx X, |
3237 | bool speed ATTRIBUTE_UNUSED) | |
dcefdf67 RH |
3238 | { |
3239 | return (GET_CODE (X) == PLUS | |
3240 | && ! CONSTANT_P (XEXP (X, 1)) | |
3241 | && ! TARGET_SHMEDIA ? 1 : 0); | |
3242 | } | |
3c50106f | 3243 | |
16bea517 | 3244 | /* Code to expand a shift. */ |
b9654711 | 3245 | |
0d7e008e | 3246 | void |
cf277499 | 3247 | gen_ashift (int type, int n, rtx reg) |
0d7e008e | 3248 | { |
16bea517 JW |
3249 | /* Negative values here come from the shift_amounts array. */ |
3250 | if (n < 0) | |
3251 | { | |
3252 | if (type == ASHIFT) | |
3253 | type = LSHIFTRT; | |
3254 | else | |
3255 | type = ASHIFT; | |
3256 | n = -n; | |
3257 | } | |
3258 | ||
0d7e008e | 3259 | switch (type) |
bc45ade3 | 3260 | { |
0d7e008e SC |
3261 | case ASHIFTRT: |
3262 | emit_insn (gen_ashrsi3_k (reg, reg, GEN_INT (n))); | |
3263 | break; | |
3264 | case LSHIFTRT: | |
16bea517 JW |
3265 | if (n == 1) |
3266 | emit_insn (gen_lshrsi3_m (reg, reg, GEN_INT (n))); | |
3267 | else | |
3268 | emit_insn (gen_lshrsi3_k (reg, reg, GEN_INT (n))); | |
0d7e008e SC |
3269 | break; |
3270 | case ASHIFT: | |
7e2fda6e | 3271 | emit_insn (gen_ashlsi3_std (reg, reg, GEN_INT (n))); |
0d7e008e | 3272 | break; |
bc45ade3 | 3273 | } |
bc45ade3 | 3274 | } |
bc45ade3 | 3275 | |
8d481241 JW |
3276 | /* Same for HImode */ |
3277 | ||
3278 | void | |
cf277499 | 3279 | gen_ashift_hi (int type, int n, rtx reg) |
8d481241 JW |
3280 | { |
3281 | /* Negative values here come from the shift_amounts array. */ | |
3282 | if (n < 0) | |
3283 | { | |
3284 | if (type == ASHIFT) | |
3285 | type = LSHIFTRT; | |
3286 | else | |
3287 | type = ASHIFT; | |
3288 | n = -n; | |
3289 | } | |
3290 | ||
3291 | switch (type) | |
3292 | { | |
3293 | case ASHIFTRT: | |
8d481241 | 3294 | case LSHIFTRT: |
d0c42859 R |
3295 | /* We don't have HImode right shift operations because using the |
3296 | ordinary 32 bit shift instructions for that doesn't generate proper | |
3297 | zero/sign extension. | |
3298 | gen_ashift_hi is only called in contexts where we know that the | |
3299 | sign extension works out correctly. */ | |
97d6fd65 | 3300 | { |
ddef6bc7 | 3301 | int offset = 0; |
97d6fd65 R |
3302 | if (GET_CODE (reg) == SUBREG) |
3303 | { | |
ddef6bc7 | 3304 | offset = SUBREG_BYTE (reg); |
97d6fd65 R |
3305 | reg = SUBREG_REG (reg); |
3306 | } | |
ddef6bc7 | 3307 | gen_ashift (type, n, gen_rtx_SUBREG (SImode, reg, offset)); |
97d6fd65 R |
3308 | break; |
3309 | } | |
8d481241 JW |
3310 | case ASHIFT: |
3311 | emit_insn (gen_ashlhi3_k (reg, reg, GEN_INT (n))); | |
3312 | break; | |
3313 | } | |
3314 | } | |
3315 | ||
8aa2a305 JW |
3316 | /* Output RTL to split a constant shift into its component SH constant |
3317 | shift instructions. */ | |
f676971a | 3318 | |
318881c0 | 3319 | void |
cf277499 | 3320 | gen_shifty_op (int code, rtx *operands) |
bc45ade3 | 3321 | { |
16bea517 | 3322 | int value = INTVAL (operands[2]); |
8aa2a305 | 3323 | int max, i; |
00f8ff66 | 3324 | |
cff3d762 | 3325 | /* Truncate the shift count in case it is out of bounds. */ |
6c34a092 | 3326 | value = value & 31; |
f676971a | 3327 | |
8aa2a305 | 3328 | if (value == 31) |
16bea517 | 3329 | { |
8aa2a305 | 3330 | if (code == LSHIFTRT) |
0d7e008e | 3331 | { |
8aa2a305 JW |
3332 | emit_insn (gen_rotlsi3_1 (operands[0], operands[0])); |
3333 | emit_insn (gen_movt (operands[0])); | |
3334 | return; | |
16bea517 | 3335 | } |
8aa2a305 | 3336 | else if (code == ASHIFT) |
16bea517 | 3337 | { |
8aa2a305 JW |
3338 | /* There is a two instruction sequence for 31 bit left shifts, |
3339 | but it requires r0. */ | |
f3536097 | 3340 | if (REG_P (operands[0]) && REGNO (operands[0]) == 0) |
0d7e008e | 3341 | { |
8aa2a305 JW |
3342 | emit_insn (gen_andsi3 (operands[0], operands[0], const1_rtx)); |
3343 | emit_insn (gen_rotlsi3_31 (operands[0], operands[0])); | |
3344 | return; | |
0d7e008e | 3345 | } |
16bea517 | 3346 | } |
8aa2a305 | 3347 | } |
c17f53a8 JW |
3348 | else if (value == 0) |
3349 | { | |
73a4d10b R |
3350 | /* This can happen even when optimizing, if there were subregs before |
3351 | reload. Don't output a nop here, as this is never optimized away; | |
3352 | use a no-op move instead. */ | |
3353 | emit_insn (gen_rtx_SET (VOIDmode, operands[0], operands[0])); | |
c17f53a8 JW |
3354 | return; |
3355 | } | |
a9f71ad8 | 3356 | |
8aa2a305 JW |
3357 | max = shift_insns[value]; |
3358 | for (i = 0; i < max; i++) | |
3359 | gen_ashift (code, shift_amounts[value][i], operands[0]); | |
3360 | } | |
f676971a | 3361 | |
8d481241 JW |
3362 | /* Same as above, but optimized for values where the topmost bits don't |
3363 | matter. */ | |
3364 | ||
318881c0 | 3365 | void |
cf277499 | 3366 | gen_shifty_hi_op (int code, rtx *operands) |
8d481241 JW |
3367 | { |
3368 | int value = INTVAL (operands[2]); | |
3369 | int max, i; | |
cf277499 | 3370 | void (*gen_fun) (int, int, rtx); |
8d481241 JW |
3371 | |
3372 | /* This operation is used by and_shl for SImode values with a few | |
3373 | high bits known to be cleared. */ | |
3374 | value &= 31; | |
3375 | if (value == 0) | |
3376 | { | |
3377 | emit_insn (gen_nop ()); | |
3378 | return; | |
3379 | } | |
3380 | ||
3381 | gen_fun = GET_MODE (operands[0]) == HImode ? gen_ashift_hi : gen_ashift; | |
3382 | if (code == ASHIFT) | |
3383 | { | |
3384 | max = ext_shift_insns[value]; | |
3385 | for (i = 0; i < max; i++) | |
3386 | gen_fun (code, ext_shift_amounts[value][i], operands[0]); | |
3387 | } | |
3388 | else | |
3389 | /* When shifting right, emit the shifts in reverse order, so that | |
3390 | solitary negative values come first. */ | |
3391 | for (i = ext_shift_insns[value] - 1; i >= 0; i--) | |
3392 | gen_fun (code, ext_shift_amounts[value][i], operands[0]); | |
3393 | } | |
8aa2a305 JW |
3394 | |
3395 | /* Output RTL for an arithmetic right shift. */ | |
3396 | ||
3397 | /* ??? Rewrite to use super-optimizer sequences. */ | |
3398 | ||
3399 | int | |
cf277499 | 3400 | expand_ashiftrt (rtx *operands) |
8aa2a305 JW |
3401 | { |
3402 | rtx wrk; | |
3403 | char func[18]; | |
8aa2a305 JW |
3404 | int value; |
3405 | ||
49b6d06b | 3406 | if (TARGET_SH3) |
20b04867 | 3407 | { |
f3536097 | 3408 | if (!CONST_INT_P (operands[2])) |
49b6d06b JW |
3409 | { |
3410 | rtx count = copy_to_mode_reg (SImode, operands[2]); | |
3411 | emit_insn (gen_negsi2 (count, count)); | |
3412 | emit_insn (gen_ashrsi3_d (operands[0], operands[1], count)); | |
3413 | return 1; | |
3414 | } | |
1245df60 R |
3415 | else if (ashiftrt_insns[INTVAL (operands[2]) & 31] |
3416 | > 1 + SH_DYNAMIC_SHIFT_COST) | |
49b6d06b | 3417 | { |
1245df60 R |
3418 | rtx count |
3419 | = force_reg (SImode, GEN_INT (- (INTVAL (operands[2]) & 31))); | |
49b6d06b JW |
3420 | emit_insn (gen_ashrsi3_d (operands[0], operands[1], count)); |
3421 | return 1; | |
3422 | } | |
20b04867 | 3423 | } |
f3536097 | 3424 | if (!CONST_INT_P (operands[2])) |
8aa2a305 JW |
3425 | return 0; |
3426 | ||
1245df60 | 3427 | value = INTVAL (operands[2]) & 31; |
8aa2a305 JW |
3428 | |
3429 | if (value == 31) | |
3430 | { | |
73a4d10b R |
3431 | /* If we are called from abs expansion, arrange things so that we |
3432 | we can use a single MT instruction that doesn't clobber the source, | |
3433 | if LICM can hoist out the load of the constant zero. */ | |
3434 | if (currently_expanding_to_rtl) | |
3435 | { | |
3436 | emit_insn (gen_cmpgtsi_t (force_reg (SImode, CONST0_RTX (SImode)), | |
3437 | operands[1])); | |
3438 | emit_insn (gen_mov_neg_si_t (operands[0])); | |
3439 | return 1; | |
3440 | } | |
8aa2a305 JW |
3441 | emit_insn (gen_ashrsi2_31 (operands[0], operands[1])); |
3442 | return 1; | |
3443 | } | |
3444 | else if (value >= 16 && value <= 19) | |
3445 | { | |
3446 | wrk = gen_reg_rtx (SImode); | |
3447 | emit_insn (gen_ashrsi2_16 (wrk, operands[1])); | |
3448 | value -= 16; | |
3449 | while (value--) | |
3450 | gen_ashift (ASHIFTRT, 1, wrk); | |
3451 | emit_move_insn (operands[0], wrk); | |
3452 | return 1; | |
a9f71ad8 | 3453 | } |
8aa2a305 JW |
3454 | /* Expand a short sequence inline, longer call a magic routine. */ |
3455 | else if (value <= 5) | |
3456 | { | |
3457 | wrk = gen_reg_rtx (SImode); | |
3458 | emit_move_insn (wrk, operands[1]); | |
3459 | while (value--) | |
3460 | gen_ashift (ASHIFTRT, 1, wrk); | |
3461 | emit_move_insn (operands[0], wrk); | |
3462 | return 1; | |
3463 | } | |
3464 | ||
3465 | wrk = gen_reg_rtx (Pmode); | |
3466 | ||
3467 | /* Load the value into an arg reg and call a helper. */ | |
c5c76735 | 3468 | emit_move_insn (gen_rtx_REG (SImode, 4), operands[1]); |
8aa2a305 | 3469 | sprintf (func, "__ashiftrt_r4_%d", value); |
73a4d10b | 3470 | function_symbol (wrk, func, SFUNC_STATIC); |
8aa2a305 | 3471 | emit_insn (gen_ashrsi3_n (GEN_INT (value), wrk)); |
c5c76735 | 3472 | emit_move_insn (operands[0], gen_rtx_REG (SImode, 4)); |
8aa2a305 | 3473 | return 1; |
bc45ade3 | 3474 | } |
8d481241 | 3475 | |
318881c0 | 3476 | int |
cf277499 | 3477 | sh_dynamicalize_shift_p (rtx count) |
1245df60 | 3478 | { |
6c34a092 | 3479 | return shift_insns[INTVAL (count) & 31] > 1 + SH_DYNAMIC_SHIFT_COST; |
1245df60 R |
3480 | } |
3481 | ||
8d481241 JW |
3482 | /* Try to find a good way to implement the combiner pattern |
3483 | [(set (match_operand:SI 0 "register_operand" "r") | |
3484 | (and:SI (ashift:SI (match_operand:SI 1 "register_operand" "r") | |
3485 | (match_operand:SI 2 "const_int_operand" "n")) | |
3486 | (match_operand:SI 3 "const_int_operand" "n"))) . | |
3487 | LEFT_RTX is operand 2 in the above pattern, and MASK_RTX is operand 3. | |
3488 | return 0 for simple right / left or left/right shift combination. | |
3489 | return 1 for a combination of shifts with zero_extend. | |
3490 | return 2 for a combination of shifts with an AND that needs r0. | |
3491 | return 3 for a combination of shifts with an AND that needs an extra | |
3492 | scratch register, when the three highmost bits of the AND mask are clear. | |
3493 | return 4 for a combination of shifts with an AND that needs an extra | |
3494 | scratch register, when any of the three highmost bits of the AND mask | |
3495 | is set. | |
3496 | If ATTRP is set, store an initial right shift width in ATTRP[0], | |
3497 | and the instruction length in ATTRP[1] . These values are not valid | |
3498 | when returning 0. | |
3499 | When ATTRP is set and returning 1, ATTRP[2] gets set to the index into | |
3500 | shift_amounts for the last shift value that is to be used before the | |
3501 | sign extend. */ | |
3502 | int | |
cf277499 | 3503 | shl_and_kind (rtx left_rtx, rtx mask_rtx, int *attrp) |
8d481241 JW |
3504 | { |
3505 | unsigned HOST_WIDE_INT mask, lsb, mask2, lsb2; | |
3506 | int left = INTVAL (left_rtx), right; | |
3507 | int best = 0; | |
3508 | int cost, best_cost = 10000; | |
3509 | int best_right = 0, best_len = 0; | |
3510 | int i; | |
3511 | int can_ext; | |
3512 | ||
3513 | if (left < 0 || left > 31) | |
3514 | return 0; | |
f3536097 | 3515 | if (CONST_INT_P (mask_rtx)) |
8d481241 JW |
3516 | mask = (unsigned HOST_WIDE_INT) INTVAL (mask_rtx) >> left; |
3517 | else | |
3518 | mask = (unsigned HOST_WIDE_INT) GET_MODE_MASK (SImode) >> left; | |
832a3292 | 3519 | /* Can this be expressed as a right shift / left shift pair? */ |
8d481241 JW |
3520 | lsb = ((mask ^ (mask - 1)) >> 1) + 1; |
3521 | right = exact_log2 (lsb); | |
3522 | mask2 = ~(mask + lsb - 1); | |
3523 | lsb2 = ((mask2 ^ (mask2 - 1)) >> 1) + 1; | |
3524 | /* mask has no zeroes but trailing zeroes <==> ! mask2 */ | |
3525 | if (! mask2) | |
3526 | best_cost = shift_insns[right] + shift_insns[right + left]; | |
3527 | /* mask has no trailing zeroes <==> ! right */ | |
3528 | else if (! right && mask2 == ~(lsb2 - 1)) | |
3529 | { | |
3530 | int late_right = exact_log2 (lsb2); | |
3531 | best_cost = shift_insns[left + late_right] + shift_insns[late_right]; | |
3532 | } | |
832a3292 | 3533 | /* Try to use zero extend. */ |
8d481241 JW |
3534 | if (mask2 == ~(lsb2 - 1)) |
3535 | { | |
3536 | int width, first; | |
3537 | ||
3538 | for (width = 8; width <= 16; width += 8) | |
3539 | { | |
832a3292 KH |
3540 | /* Can we zero-extend right away? */ |
3541 | if (lsb2 == (unsigned HOST_WIDE_INT) 1 << width) | |
8d481241 JW |
3542 | { |
3543 | cost | |
3544 | = 1 + ext_shift_insns[right] + ext_shift_insns[left + right]; | |
3545 | if (cost < best_cost) | |
3546 | { | |
3547 | best = 1; | |
3548 | best_cost = cost; | |
3549 | best_right = right; | |
3550 | best_len = cost; | |
3551 | if (attrp) | |
3552 | attrp[2] = -1; | |
3553 | } | |
3554 | continue; | |
3555 | } | |
3556 | /* ??? Could try to put zero extend into initial right shift, | |
6f317ef3 | 3557 | or even shift a bit left before the right shift. */ |
8d481241 JW |
3558 | /* Determine value of first part of left shift, to get to the |
3559 | zero extend cut-off point. */ | |
3560 | first = width - exact_log2 (lsb2) + right; | |
3561 | if (first >= 0 && right + left - first >= 0) | |
3562 | { | |
3563 | cost = ext_shift_insns[right] + ext_shift_insns[first] + 1 | |
3564 | + ext_shift_insns[right + left - first]; | |
3565 | if (cost < best_cost) | |
3566 | { | |
3567 | best = 1; | |
3568 | best_cost = cost; | |
3569 | best_right = right; | |
3570 | best_len = cost; | |
3571 | if (attrp) | |
3572 | attrp[2] = first; | |
832a3292 | 3573 | } |
8d481241 JW |
3574 | } |
3575 | } | |
3576 | } | |
3577 | /* Try to use r0 AND pattern */ | |
3578 | for (i = 0; i <= 2; i++) | |
3579 | { | |
3580 | if (i > right) | |
3581 | break; | |
735cb76e | 3582 | if (! CONST_OK_FOR_K08 (mask >> i)) |
8d481241 JW |
3583 | continue; |
3584 | cost = (i != 0) + 2 + ext_shift_insns[left + i]; | |
3585 | if (cost < best_cost) | |
3586 | { | |
3587 | best = 2; | |
3588 | best_cost = cost; | |
3589 | best_right = i; | |
3590 | best_len = cost - 1; | |
3591 | } | |
3592 | } | |
3593 | /* Try to use a scratch register to hold the AND operand. */ | |
832a3292 | 3594 | can_ext = ((mask << left) & ((unsigned HOST_WIDE_INT) 3 << 30)) == 0; |
8d481241 JW |
3595 | for (i = 0; i <= 2; i++) |
3596 | { | |
3597 | if (i > right) | |
3598 | break; | |
735cb76e | 3599 | cost = (i != 0) + (CONST_OK_FOR_I08 (mask >> i) ? 2 : 3) |
6ab911bb | 3600 | + (can_ext ? ext_shift_insns : shift_insns)[left + i]; |
8d481241 JW |
3601 | if (cost < best_cost) |
3602 | { | |
3603 | best = 4 - can_ext; | |
3604 | best_cost = cost; | |
3605 | best_right = i; | |
735cb76e | 3606 | best_len = cost - 1 - ! CONST_OK_FOR_I08 (mask >> i); |
8d481241 JW |
3607 | } |
3608 | } | |
3609 | ||
3610 | if (attrp) | |
3611 | { | |
3612 | attrp[0] = best_right; | |
3613 | attrp[1] = best_len; | |
3614 | } | |
3615 | return best; | |
3616 | } | |
3617 | ||
3618 | /* This is used in length attributes of the unnamed instructions | |
3619 | corresponding to shl_and_kind return values of 1 and 2. */ | |
3620 | int | |
cf277499 | 3621 | shl_and_length (rtx insn) |
8d481241 JW |
3622 | { |
3623 | rtx set_src, left_rtx, mask_rtx; | |
3624 | int attributes[3]; | |
3625 | ||
3626 | set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
3627 | left_rtx = XEXP (XEXP (set_src, 0), 1); | |
3628 | mask_rtx = XEXP (set_src, 1); | |
3629 | shl_and_kind (left_rtx, mask_rtx, attributes); | |
3630 | return attributes[1]; | |
3631 | } | |
3632 | ||
3633 | /* This is used in length attribute of the and_shl_scratch instruction. */ | |
3634 | ||
3635 | int | |
cf277499 | 3636 | shl_and_scr_length (rtx insn) |
8d481241 JW |
3637 | { |
3638 | rtx set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
6c34a092 | 3639 | int len = shift_insns[INTVAL (XEXP (set_src, 1)) & 31]; |
8d481241 | 3640 | rtx op = XEXP (set_src, 0); |
6c34a092 | 3641 | len += shift_insns[INTVAL (XEXP (op, 1)) & 31] + 1; |
8d481241 | 3642 | op = XEXP (XEXP (op, 0), 0); |
6c34a092 | 3643 | return len + shift_insns[INTVAL (XEXP (op, 1)) & 31]; |
8d481241 JW |
3644 | } |
3645 | ||
8d481241 JW |
3646 | /* Generate rtl for instructions for which shl_and_kind advised a particular |
3647 | method of generating them, i.e. returned zero. */ | |
3648 | ||
3649 | int | |
cf277499 | 3650 | gen_shl_and (rtx dest, rtx left_rtx, rtx mask_rtx, rtx source) |
8d481241 JW |
3651 | { |
3652 | int attributes[3]; | |
3653 | unsigned HOST_WIDE_INT mask; | |
3654 | int kind = shl_and_kind (left_rtx, mask_rtx, attributes); | |
3655 | int right, total_shift; | |
832a3292 | 3656 | void (*shift_gen_fun) (int, rtx *) = gen_shifty_hi_op; |
8d481241 JW |
3657 | |
3658 | right = attributes[0]; | |
3659 | total_shift = INTVAL (left_rtx) + right; | |
3660 | mask = (unsigned HOST_WIDE_INT) INTVAL (mask_rtx) >> total_shift; | |
3661 | switch (kind) | |
3662 | { | |
3663 | default: | |
3664 | return -1; | |
3665 | case 1: | |
3666 | { | |
3667 | int first = attributes[2]; | |
3668 | rtx operands[3]; | |
3669 | ||
3670 | if (first < 0) | |
3671 | { | |
7174c937 | 3672 | emit_insn ((mask << right) <= 0xff |
832a3292 KH |
3673 | ? gen_zero_extendqisi2 (dest, |
3674 | gen_lowpart (QImode, source)) | |
3675 | : gen_zero_extendhisi2 (dest, | |
3676 | gen_lowpart (HImode, source))); | |
8d481241 JW |
3677 | source = dest; |
3678 | } | |
3679 | if (source != dest) | |
3680 | emit_insn (gen_movsi (dest, source)); | |
3681 | operands[0] = dest; | |
3682 | if (right) | |
3683 | { | |
3684 | operands[2] = GEN_INT (right); | |
3685 | gen_shifty_hi_op (LSHIFTRT, operands); | |
3686 | } | |
3687 | if (first > 0) | |
3688 | { | |
3689 | operands[2] = GEN_INT (first); | |
3690 | gen_shifty_hi_op (ASHIFT, operands); | |
3691 | total_shift -= first; | |
3692 | mask <<= first; | |
3693 | } | |
3694 | if (first >= 0) | |
85af47b9 | 3695 | emit_insn (mask <= 0xff |
832a3292 KH |
3696 | ? gen_zero_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
3697 | : gen_zero_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
8d481241 JW |
3698 | if (total_shift > 0) |
3699 | { | |
3700 | operands[2] = GEN_INT (total_shift); | |
3701 | gen_shifty_hi_op (ASHIFT, operands); | |
3702 | } | |
3703 | break; | |
3704 | } | |
3705 | case 4: | |
3706 | shift_gen_fun = gen_shifty_op; | |
8d481241 | 3707 | case 3: |
24c50999 JR |
3708 | /* If the topmost bit that matters is set, set the topmost bits |
3709 | that don't matter. This way, we might be able to get a shorter | |
3710 | signed constant. */ | |
832a3292 KH |
3711 | if (mask & ((HOST_WIDE_INT) 1 << (31 - total_shift))) |
3712 | mask |= (HOST_WIDE_INT) ~0 << (31 - total_shift); | |
61fb6bac | 3713 | case 2: |
8d481241 JW |
3714 | /* Don't expand fine-grained when combining, because that will |
3715 | make the pattern fail. */ | |
4586b4ca | 3716 | if (currently_expanding_to_rtl |
8d481241 JW |
3717 | || reload_in_progress || reload_completed) |
3718 | { | |
3719 | rtx operands[3]; | |
f676971a | 3720 | |
61fb6bac R |
3721 | /* Cases 3 and 4 should be handled by this split |
3722 | only while combining */ | |
f5b9e7c9 | 3723 | gcc_assert (kind <= 2); |
8d481241 JW |
3724 | if (right) |
3725 | { | |
3726 | emit_insn (gen_lshrsi3 (dest, source, GEN_INT (right))); | |
3727 | source = dest; | |
3728 | } | |
3729 | emit_insn (gen_andsi3 (dest, source, GEN_INT (mask))); | |
afad3d2c JW |
3730 | if (total_shift) |
3731 | { | |
3732 | operands[0] = dest; | |
3733 | operands[1] = dest; | |
3734 | operands[2] = GEN_INT (total_shift); | |
3735 | shift_gen_fun (ASHIFT, operands); | |
3736 | } | |
8d481241 JW |
3737 | break; |
3738 | } | |
3739 | else | |
3740 | { | |
3741 | int neg = 0; | |
3742 | if (kind != 4 && total_shift < 16) | |
3743 | { | |
3744 | neg = -ext_shift_amounts[total_shift][1]; | |
3745 | if (neg > 0) | |
3746 | neg -= ext_shift_amounts[total_shift][2]; | |
3747 | else | |
3748 | neg = 0; | |
3749 | } | |
3750 | emit_insn (gen_and_shl_scratch (dest, source, | |
3751 | GEN_INT (right), | |
3752 | GEN_INT (mask), | |
3753 | GEN_INT (total_shift + neg), | |
3754 | GEN_INT (neg))); | |
3755 | emit_insn (gen_movsi (dest, dest)); | |
3756 | break; | |
3757 | } | |
3758 | } | |
3759 | return 0; | |
3760 | } | |
3761 | ||
3762 | /* Try to find a good way to implement the combiner pattern | |
3763 | [(set (match_operand:SI 0 "register_operand" "=r") | |
3764 | (sign_extract:SI (ashift:SI (match_operand:SI 1 "register_operand" "r") | |
3765 | (match_operand:SI 2 "const_int_operand" "n") | |
3766 | (match_operand:SI 3 "const_int_operand" "n") | |
3767 | (const_int 0))) | |
4773afa4 | 3768 | (clobber (reg:SI T_REG))] |
8d481241 JW |
3769 | LEFT_RTX is operand 2 in the above pattern, and SIZE_RTX is operand 3. |
3770 | return 0 for simple left / right shift combination. | |
3771 | return 1 for left shift / 8 bit sign extend / left shift. | |
3772 | return 2 for left shift / 16 bit sign extend / left shift. | |
3773 | return 3 for left shift / 8 bit sign extend / shift / sign extend. | |
3774 | return 4 for left shift / 16 bit sign extend / shift / sign extend. | |
3775 | return 5 for left shift / 16 bit sign extend / right shift | |
3776 | return 6 for < 8 bit sign extend / left shift. | |
3777 | return 7 for < 8 bit sign extend / left shift / single right shift. | |
3778 | If COSTP is nonzero, assign the calculated cost to *COSTP. */ | |
3779 | ||
3780 | int | |
cf277499 | 3781 | shl_sext_kind (rtx left_rtx, rtx size_rtx, int *costp) |
8d481241 JW |
3782 | { |
3783 | int left, size, insize, ext; | |
3a8699c7 | 3784 | int cost = 0, best_cost; |
8d481241 JW |
3785 | int kind; |
3786 | ||
3787 | left = INTVAL (left_rtx); | |
3788 | size = INTVAL (size_rtx); | |
3789 | insize = size - left; | |
f5b9e7c9 | 3790 | gcc_assert (insize > 0); |
8d481241 JW |
3791 | /* Default to left / right shift. */ |
3792 | kind = 0; | |
3793 | best_cost = shift_insns[32 - insize] + ashiftrt_insns[32 - size]; | |
3794 | if (size <= 16) | |
3795 | { | |
3796 | /* 16 bit shift / sign extend / 16 bit shift */ | |
3797 | cost = shift_insns[16 - insize] + 1 + ashiftrt_insns[16 - size]; | |
afad3d2c JW |
3798 | /* If ashiftrt_insns[16 - size] is 8, this choice will be overridden |
3799 | below, by alternative 3 or something even better. */ | |
8d481241 JW |
3800 | if (cost < best_cost) |
3801 | { | |
3802 | kind = 5; | |
3803 | best_cost = cost; | |
3804 | } | |
3805 | } | |
3806 | /* Try a plain sign extend between two shifts. */ | |
3807 | for (ext = 16; ext >= insize; ext -= 8) | |
3808 | { | |
3809 | if (ext <= size) | |
3810 | { | |
3811 | cost = ext_shift_insns[ext - insize] + 1 + shift_insns[size - ext]; | |
3812 | if (cost < best_cost) | |
3813 | { | |
11f9ed1a | 3814 | kind = ext / (unsigned) 8; |
8d481241 JW |
3815 | best_cost = cost; |
3816 | } | |
3817 | } | |
ae9d19c0 JR |
3818 | /* Check if we can do a sloppy shift with a final signed shift |
3819 | restoring the sign. */ | |
3820 | if (EXT_SHIFT_SIGNED (size - ext)) | |
3821 | cost = ext_shift_insns[ext - insize] + ext_shift_insns[size - ext] + 1; | |
3822 | /* If not, maybe it's still cheaper to do the second shift sloppy, | |
3823 | and do a final sign extend? */ | |
3824 | else if (size <= 16) | |
3825 | cost = ext_shift_insns[ext - insize] + 1 | |
3826 | + ext_shift_insns[size > ext ? size - ext : ext - size] + 1; | |
3827 | else | |
3828 | continue; | |
3829 | if (cost < best_cost) | |
8d481241 | 3830 | { |
11f9ed1a | 3831 | kind = ext / (unsigned) 8 + 2; |
ae9d19c0 | 3832 | best_cost = cost; |
8d481241 JW |
3833 | } |
3834 | } | |
3835 | /* Check if we can sign extend in r0 */ | |
3836 | if (insize < 8) | |
3837 | { | |
3838 | cost = 3 + shift_insns[left]; | |
3839 | if (cost < best_cost) | |
3840 | { | |
3841 | kind = 6; | |
3842 | best_cost = cost; | |
3843 | } | |
3844 | /* Try the same with a final signed shift. */ | |
3845 | if (left < 31) | |
3846 | { | |
3847 | cost = 3 + ext_shift_insns[left + 1] + 1; | |
3848 | if (cost < best_cost) | |
3849 | { | |
3850 | kind = 7; | |
3851 | best_cost = cost; | |
3852 | } | |
3853 | } | |
3854 | } | |
3855 | if (TARGET_SH3) | |
3856 | { | |
3857 | /* Try to use a dynamic shift. */ | |
1245df60 | 3858 | cost = shift_insns[32 - insize] + 1 + SH_DYNAMIC_SHIFT_COST; |
8d481241 JW |
3859 | if (cost < best_cost) |
3860 | { | |
3861 | kind = 0; | |
3862 | best_cost = cost; | |
3863 | } | |
3864 | } | |
3865 | if (costp) | |
3866 | *costp = cost; | |
3867 | return kind; | |
3868 | } | |
3869 | ||
3870 | /* Function to be used in the length attribute of the instructions | |
3871 | implementing this pattern. */ | |
3872 | ||
3873 | int | |
cf277499 | 3874 | shl_sext_length (rtx insn) |
8d481241 JW |
3875 | { |
3876 | rtx set_src, left_rtx, size_rtx; | |
3877 | int cost; | |
3878 | ||
3879 | set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
3880 | left_rtx = XEXP (XEXP (set_src, 0), 1); | |
3881 | size_rtx = XEXP (set_src, 1); | |
3882 | shl_sext_kind (left_rtx, size_rtx, &cost); | |
3883 | return cost; | |
3884 | } | |
3885 | ||
3886 | /* Generate rtl for this pattern */ | |
3887 | ||
3888 | int | |
cf277499 | 3889 | gen_shl_sext (rtx dest, rtx left_rtx, rtx size_rtx, rtx source) |
8d481241 JW |
3890 | { |
3891 | int kind; | |
d00d338c | 3892 | int left, size, insize, cost; |
8d481241 JW |
3893 | rtx operands[3]; |
3894 | ||
d00d338c | 3895 | kind = shl_sext_kind (left_rtx, size_rtx, &cost); |
8d481241 JW |
3896 | left = INTVAL (left_rtx); |
3897 | size = INTVAL (size_rtx); | |
3898 | insize = size - left; | |
3899 | switch (kind) | |
3900 | { | |
3901 | case 1: | |
3902 | case 2: | |
3903 | case 3: | |
3904 | case 4: | |
3905 | { | |
3906 | int ext = kind & 1 ? 8 : 16; | |
3907 | int shift2 = size - ext; | |
3908 | ||
3909 | /* Don't expand fine-grained when combining, because that will | |
3910 | make the pattern fail. */ | |
4586b4ca | 3911 | if (! currently_expanding_to_rtl |
8d481241 JW |
3912 | && ! reload_in_progress && ! reload_completed) |
3913 | { | |
3914 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
3915 | emit_insn (gen_movsi (dest, source)); | |
3916 | break; | |
3917 | } | |
3918 | if (dest != source) | |
3919 | emit_insn (gen_movsi (dest, source)); | |
3920 | operands[0] = dest; | |
afad3d2c JW |
3921 | if (ext - insize) |
3922 | { | |
3923 | operands[2] = GEN_INT (ext - insize); | |
3924 | gen_shifty_hi_op (ASHIFT, operands); | |
3925 | } | |
8d481241 | 3926 | emit_insn (kind & 1 |
832a3292 KH |
3927 | ? gen_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
3928 | : gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
8d481241 JW |
3929 | if (kind <= 2) |
3930 | { | |
afad3d2c JW |
3931 | if (shift2) |
3932 | { | |
3933 | operands[2] = GEN_INT (shift2); | |
3934 | gen_shifty_op (ASHIFT, operands); | |
3935 | } | |
8d481241 JW |
3936 | } |
3937 | else | |
3938 | { | |
afad3d2c | 3939 | if (shift2 > 0) |
8d481241 | 3940 | { |
ae9d19c0 JR |
3941 | if (EXT_SHIFT_SIGNED (shift2)) |
3942 | { | |
3943 | operands[2] = GEN_INT (shift2 + 1); | |
3944 | gen_shifty_op (ASHIFT, operands); | |
a556fd39 | 3945 | operands[2] = const1_rtx; |
ae9d19c0 JR |
3946 | gen_shifty_op (ASHIFTRT, operands); |
3947 | break; | |
3948 | } | |
8d481241 JW |
3949 | operands[2] = GEN_INT (shift2); |
3950 | gen_shifty_hi_op (ASHIFT, operands); | |
3951 | } | |
afad3d2c | 3952 | else if (shift2) |
8d481241 JW |
3953 | { |
3954 | operands[2] = GEN_INT (-shift2); | |
3955 | gen_shifty_hi_op (LSHIFTRT, operands); | |
3956 | } | |
3957 | emit_insn (size <= 8 | |
afad3d2c JW |
3958 | ? gen_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
3959 | : gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
8d481241 JW |
3960 | } |
3961 | break; | |
3962 | } | |
3963 | case 5: | |
afad3d2c JW |
3964 | { |
3965 | int i = 16 - size; | |
4586b4ca | 3966 | if (! currently_expanding_to_rtl |
913d8e13 R |
3967 | && ! reload_in_progress && ! reload_completed) |
3968 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
3969 | else | |
3970 | { | |
3971 | operands[0] = dest; | |
3972 | operands[2] = GEN_INT (16 - insize); | |
3973 | gen_shifty_hi_op (ASHIFT, operands); | |
3974 | emit_insn (gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
3975 | } | |
afad3d2c JW |
3976 | /* Don't use gen_ashrsi3 because it generates new pseudos. */ |
3977 | while (--i >= 0) | |
3978 | gen_ashift (ASHIFTRT, 1, dest); | |
3979 | break; | |
3980 | } | |
8d481241 JW |
3981 | case 6: |
3982 | case 7: | |
3983 | /* Don't expand fine-grained when combining, because that will | |
3984 | make the pattern fail. */ | |
4586b4ca | 3985 | if (! currently_expanding_to_rtl |
8d481241 JW |
3986 | && ! reload_in_progress && ! reload_completed) |
3987 | { | |
3988 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
3989 | emit_insn (gen_movsi (dest, source)); | |
3990 | break; | |
3991 | } | |
3992 | emit_insn (gen_andsi3 (dest, source, GEN_INT ((1 << insize) - 1))); | |
3993 | emit_insn (gen_xorsi3 (dest, dest, GEN_INT (1 << (insize - 1)))); | |
3994 | emit_insn (gen_addsi3 (dest, dest, GEN_INT (-1 << (insize - 1)))); | |
3995 | operands[0] = dest; | |
3996 | operands[2] = kind == 7 ? GEN_INT (left + 1) : left_rtx; | |
3997 | gen_shifty_op (ASHIFT, operands); | |
3998 | if (kind == 7) | |
a556fd39 | 3999 | emit_insn (gen_ashrsi3_k (dest, dest, const1_rtx)); |
8d481241 JW |
4000 | break; |
4001 | default: | |
4002 | return -1; | |
4003 | } | |
4004 | return 0; | |
4005 | } | |
fa5322fa AO |
4006 | |
4007 | /* Prefix a symbol_ref name with "datalabel". */ | |
4008 | ||
4009 | rtx | |
cf277499 | 4010 | gen_datalabel_ref (rtx sym) |
fa5322fa | 4011 | { |
73a4d10b R |
4012 | const char *str; |
4013 | ||
fa5322fa AO |
4014 | if (GET_CODE (sym) == LABEL_REF) |
4015 | return gen_rtx_CONST (GET_MODE (sym), | |
4016 | gen_rtx_UNSPEC (GET_MODE (sym), | |
4017 | gen_rtvec (1, sym), | |
4018 | UNSPEC_DATALABEL)); | |
f676971a | 4019 | |
f5b9e7c9 | 4020 | gcc_assert (GET_CODE (sym) == SYMBOL_REF); |
fa5322fa | 4021 | |
73a4d10b R |
4022 | str = XSTR (sym, 0); |
4023 | /* Share all SYMBOL_REF strings with the same value - that is important | |
4024 | for cse. */ | |
4025 | str = IDENTIFIER_POINTER (get_identifier (str)); | |
4026 | XSTR (sym, 0) = str; | |
4027 | ||
fa5322fa AO |
4028 | return sym; |
4029 | } | |
4030 | ||
8aa2a305 | 4031 | \f |
6dd3c0a5 SB |
4032 | static alloc_pool label_ref_list_pool; |
4033 | ||
4034 | typedef struct label_ref_list_d | |
4035 | { | |
4036 | rtx label; | |
4037 | struct label_ref_list_d *next; | |
4038 | } *label_ref_list_t; | |
4039 | ||
8aa2a305 JW |
4040 | /* The SH cannot load a large constant into a register, constants have to |
4041 | come from a pc relative load. The reference of a pc relative load | |
0fa2e4df | 4042 | instruction must be less than 1k in front of the instruction. This |
8aa2a305 JW |
4043 | means that we often have to dump a constant inside a function, and |
4044 | generate code to branch around it. | |
bc45ade3 | 4045 | |
8aa2a305 JW |
4046 | It is important to minimize this, since the branches will slow things |
4047 | down and make things bigger. | |
4048 | ||
4049 | Worst case code looks like: | |
4050 | ||
4051 | mov.l L1,rn | |
4052 | bra L2 | |
4053 | nop | |
4054 | align | |
4055 | L1: .long value | |
4056 | L2: | |
4057 | .. | |
4058 | ||
4059 | mov.l L3,rn | |
4060 | bra L4 | |
4061 | nop | |
4062 | align | |
4063 | L3: .long value | |
4064 | L4: | |
4065 | .. | |
4066 | ||
4067 | We fix this by performing a scan before scheduling, which notices which | |
4068 | instructions need to have their operands fetched from the constant table | |
4069 | and builds the table. | |
4070 | ||
4071 | The algorithm is: | |
4072 | ||
4073 | scan, find an instruction which needs a pcrel move. Look forward, find the | |
4074 | last barrier which is within MAX_COUNT bytes of the requirement. | |
4075 | If there isn't one, make one. Process all the instructions between | |
4076 | the find and the barrier. | |
4077 | ||
4078 | In the above example, we can tell that L3 is within 1k of L1, so | |
4079 | the first move can be shrunk from the 3 insn+constant sequence into | |
4080 | just 1 insn, and the constant moved to L3 to make: | |
4081 | ||
4082 | mov.l L1,rn | |
4083 | .. | |
4084 | mov.l L3,rn | |
4085 | bra L4 | |
4086 | nop | |
4087 | align | |
4088 | L3:.long value | |
4089 | L4:.long value | |
4090 | ||
4091 | Then the second move becomes the target for the shortening process. */ | |
4092 | ||
4093 | typedef struct | |
4094 | { | |
4095 | rtx value; /* Value in table. */ | |
4096 | rtx label; /* Label of value. */ | |
6dd3c0a5 | 4097 | label_ref_list_t wend; /* End of window. */ |
8aa2a305 | 4098 | enum machine_mode mode; /* Mode of value. */ |
3503150c RS |
4099 | |
4100 | /* True if this constant is accessed as part of a post-increment | |
4101 | sequence. Note that HImode constants are never accessed in this way. */ | |
4102 | bool part_of_sequence_p; | |
8aa2a305 JW |
4103 | } pool_node; |
4104 | ||
4105 | /* The maximum number of constants that can fit into one pool, since | |
73a4d10b R |
4106 | constants in the range 0..510 are at least 2 bytes long, and in the |
4107 | range from there to 1018 at least 4 bytes. */ | |
8aa2a305 | 4108 | |
73a4d10b | 4109 | #define MAX_POOL_SIZE 372 |
8aa2a305 JW |
4110 | static pool_node pool_vector[MAX_POOL_SIZE]; |
4111 | static int pool_size; | |
b91455de KK |
4112 | static rtx pool_window_label; |
4113 | static int pool_window_last; | |
8aa2a305 | 4114 | |
4241ecb0 R |
4115 | static int max_labelno_before_reorg; |
4116 | ||
8aa2a305 JW |
4117 | /* ??? If we need a constant in HImode which is the truncated value of a |
4118 | constant we need in SImode, we could combine the two entries thus saving | |
4119 | two bytes. Is this common enough to be worth the effort of implementing | |
4120 | it? */ | |
4121 | ||
4122 | /* ??? This stuff should be done at the same time that we shorten branches. | |
4123 | As it is now, we must assume that all branches are the maximum size, and | |
4124 | this causes us to almost always output constant pools sooner than | |
4125 | necessary. */ | |
4126 | ||
4127 | /* Add a constant to the pool and return its label. */ | |
4128 | ||
4129 | static rtx | |
cf277499 | 4130 | add_constant (rtx x, enum machine_mode mode, rtx last_value) |
0d7e008e SC |
4131 | { |
4132 | int i; | |
0a2aaacc | 4133 | rtx lab, new_rtx; |
6dd3c0a5 | 4134 | label_ref_list_t ref, newref; |
8aa2a305 JW |
4135 | |
4136 | /* First see if we've already got it. */ | |
4137 | for (i = 0; i < pool_size; i++) | |
0d7e008e | 4138 | { |
8aa2a305 JW |
4139 | if (x->code == pool_vector[i].value->code |
4140 | && mode == pool_vector[i].mode) | |
0d7e008e | 4141 | { |
8aa2a305 JW |
4142 | if (x->code == CODE_LABEL) |
4143 | { | |
4144 | if (XINT (x, 3) != XINT (pool_vector[i].value, 3)) | |
4145 | continue; | |
4146 | } | |
4147 | if (rtx_equal_p (x, pool_vector[i].value)) | |
225e4f43 | 4148 | { |
0a2aaacc | 4149 | lab = new_rtx = 0; |
225e4f43 R |
4150 | if (! last_value |
4151 | || ! i | |
4152 | || ! rtx_equal_p (last_value, pool_vector[i-1].value)) | |
4153 | { | |
0a2aaacc KG |
4154 | new_rtx = gen_label_rtx (); |
4155 | LABEL_REFS (new_rtx) = pool_vector[i].label; | |
4156 | pool_vector[i].label = lab = new_rtx; | |
225e4f43 | 4157 | } |
b91455de KK |
4158 | if (lab && pool_window_label) |
4159 | { | |
6dd3c0a5 SB |
4160 | newref = (label_ref_list_t) pool_alloc (label_ref_list_pool); |
4161 | newref->label = pool_window_label; | |
b91455de | 4162 | ref = pool_vector[pool_window_last].wend; |
6dd3c0a5 | 4163 | newref->next = ref; |
b91455de KK |
4164 | pool_vector[pool_window_last].wend = newref; |
4165 | } | |
0a2aaacc KG |
4166 | if (new_rtx) |
4167 | pool_window_label = new_rtx; | |
b91455de | 4168 | pool_window_last = i; |
225e4f43 R |
4169 | return lab; |
4170 | } | |
0d7e008e | 4171 | } |
0d7e008e | 4172 | } |
b9654711 | 4173 | |
8aa2a305 JW |
4174 | /* Need a new one. */ |
4175 | pool_vector[pool_size].value = x; | |
225e4f43 | 4176 | if (last_value && rtx_equal_p (last_value, pool_vector[pool_size - 1].value)) |
3503150c RS |
4177 | { |
4178 | lab = 0; | |
4179 | pool_vector[pool_size - 1].part_of_sequence_p = true; | |
4180 | } | |
225e4f43 R |
4181 | else |
4182 | lab = gen_label_rtx (); | |
8aa2a305 JW |
4183 | pool_vector[pool_size].mode = mode; |
4184 | pool_vector[pool_size].label = lab; | |
6dd3c0a5 | 4185 | pool_vector[pool_size].wend = NULL; |
3503150c | 4186 | pool_vector[pool_size].part_of_sequence_p = (lab == 0); |
b91455de KK |
4187 | if (lab && pool_window_label) |
4188 | { | |
6dd3c0a5 SB |
4189 | newref = (label_ref_list_t) pool_alloc (label_ref_list_pool); |
4190 | newref->label = pool_window_label; | |
b91455de | 4191 | ref = pool_vector[pool_window_last].wend; |
6dd3c0a5 | 4192 | newref->next = ref; |
b91455de KK |
4193 | pool_vector[pool_window_last].wend = newref; |
4194 | } | |
4195 | if (lab) | |
4196 | pool_window_label = lab; | |
4197 | pool_window_last = pool_size; | |
8aa2a305 JW |
4198 | pool_size++; |
4199 | return lab; | |
0d7e008e | 4200 | } |
16bea517 | 4201 | |
078c8b08 R |
4202 | /* Output the literal table. START, if nonzero, is the first instruction |
4203 | this table is needed for, and also indicates that there is at least one | |
4204 | casesi_worker_2 instruction; We have to emit the operand3 labels from | |
4205 | these insns at a 4-byte aligned position. BARRIER is the barrier | |
4206 | after which we are to place the table. */ | |
b9654711 | 4207 | |
b9654711 | 4208 | static void |
078c8b08 | 4209 | dump_table (rtx start, rtx barrier) |
b9654711 | 4210 | { |
078c8b08 | 4211 | rtx scan = barrier; |
0d7e008e | 4212 | int i; |
0d7e008e | 4213 | int need_align = 1; |
6dd3c0a5 SB |
4214 | rtx lab; |
4215 | label_ref_list_t ref; | |
aa06e8f5 | 4216 | int have_df = 0; |
b9654711 | 4217 | |
16bea517 | 4218 | /* Do two passes, first time dump out the HI sized constants. */ |
b9654711 | 4219 | |
0d7e008e | 4220 | for (i = 0; i < pool_size; i++) |
b9654711 | 4221 | { |
8aa2a305 JW |
4222 | pool_node *p = &pool_vector[i]; |
4223 | ||
0d7e008e SC |
4224 | if (p->mode == HImode) |
4225 | { | |
4226 | if (need_align) | |
4227 | { | |
4228 | scan = emit_insn_after (gen_align_2 (), scan); | |
4229 | need_align = 0; | |
4230 | } | |
b91455de KK |
4231 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
4232 | scan = emit_label_after (lab, scan); | |
4233 | scan = emit_insn_after (gen_consttable_2 (p->value, const0_rtx), | |
4234 | scan); | |
6dd3c0a5 | 4235 | for (ref = p->wend; ref; ref = ref->next) |
b91455de | 4236 | { |
6dd3c0a5 | 4237 | lab = ref->label; |
b91455de KK |
4238 | scan = emit_insn_after (gen_consttable_window_end (lab), scan); |
4239 | } | |
0d7e008e | 4240 | } |
aa06e8f5 R |
4241 | else if (p->mode == DFmode) |
4242 | have_df = 1; | |
b9654711 | 4243 | } |
8aa2a305 | 4244 | |
0d7e008e | 4245 | need_align = 1; |
b9654711 | 4246 | |
078c8b08 R |
4247 | if (start) |
4248 | { | |
4249 | scan = emit_insn_after (gen_align_4 (), scan); | |
4250 | need_align = 0; | |
4251 | for (; start != barrier; start = NEXT_INSN (start)) | |
f3536097 | 4252 | if (NONJUMP_INSN_P (start) |
078c8b08 R |
4253 | && recog_memoized (start) == CODE_FOR_casesi_worker_2) |
4254 | { | |
4255 | rtx src = SET_SRC (XVECEXP (PATTERN (start), 0, 0)); | |
4256 | rtx lab = XEXP (XVECEXP (src, 0, 3), 0); | |
4257 | ||
4258 | scan = emit_label_after (lab, scan); | |
4259 | } | |
4260 | } | |
aa06e8f5 | 4261 | if (TARGET_FMOVD && TARGET_ALIGN_DOUBLE && have_df) |
fa5322fa AO |
4262 | { |
4263 | rtx align_insn = NULL_RTX; | |
4264 | ||
4265 | scan = emit_label_after (gen_label_rtx (), scan); | |
4266 | scan = emit_insn_after (gen_align_log (GEN_INT (3)), scan); | |
4267 | need_align = 0; | |
4268 | ||
4269 | for (i = 0; i < pool_size; i++) | |
4270 | { | |
4271 | pool_node *p = &pool_vector[i]; | |
4272 | ||
4273 | switch (p->mode) | |
4274 | { | |
4275 | case HImode: | |
4276 | break; | |
4277 | case SImode: | |
4278 | case SFmode: | |
3503150c | 4279 | if (align_insn && !p->part_of_sequence_p) |
fa5322fa AO |
4280 | { |
4281 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) | |
4282 | emit_label_before (lab, align_insn); | |
4283 | emit_insn_before (gen_consttable_4 (p->value, const0_rtx), | |
4284 | align_insn); | |
6dd3c0a5 | 4285 | for (ref = p->wend; ref; ref = ref->next) |
fa5322fa | 4286 | { |
6dd3c0a5 | 4287 | lab = ref->label; |
fa5322fa | 4288 | emit_insn_before (gen_consttable_window_end (lab), |
832a3292 | 4289 | align_insn); |
fa5322fa AO |
4290 | } |
4291 | delete_insn (align_insn); | |
4292 | align_insn = NULL_RTX; | |
4293 | continue; | |
4294 | } | |
4295 | else | |
4296 | { | |
4297 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) | |
4298 | scan = emit_label_after (lab, scan); | |
4299 | scan = emit_insn_after (gen_consttable_4 (p->value, | |
4300 | const0_rtx), scan); | |
4301 | need_align = ! need_align; | |
4302 | } | |
4303 | break; | |
4304 | case DFmode: | |
fa5322fa AO |
4305 | if (need_align) |
4306 | { | |
4307 | scan = emit_insn_after (gen_align_log (GEN_INT (3)), scan); | |
4308 | align_insn = scan; | |
4309 | need_align = 0; | |
4310 | } | |
aa06e8f5 | 4311 | case DImode: |
fa5322fa AO |
4312 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
4313 | scan = emit_label_after (lab, scan); | |
4314 | scan = emit_insn_after (gen_consttable_8 (p->value, const0_rtx), | |
4315 | scan); | |
4316 | break; | |
4317 | default: | |
f5b9e7c9 | 4318 | gcc_unreachable (); |
fa5322fa AO |
4319 | } |
4320 | ||
4321 | if (p->mode != HImode) | |
4322 | { | |
6dd3c0a5 | 4323 | for (ref = p->wend; ref; ref = ref->next) |
fa5322fa | 4324 | { |
6dd3c0a5 | 4325 | lab = ref->label; |
fa5322fa AO |
4326 | scan = emit_insn_after (gen_consttable_window_end (lab), |
4327 | scan); | |
4328 | } | |
4329 | } | |
4330 | } | |
4331 | ||
4332 | pool_size = 0; | |
4333 | } | |
f676971a | 4334 | |
0d7e008e | 4335 | for (i = 0; i < pool_size; i++) |
b9654711 | 4336 | { |
8aa2a305 | 4337 | pool_node *p = &pool_vector[i]; |
b9654711 | 4338 | |
0d7e008e | 4339 | switch (p->mode) |
b9654711 | 4340 | { |
0d7e008e SC |
4341 | case HImode: |
4342 | break; | |
4343 | case SImode: | |
e577c183 | 4344 | case SFmode: |
0d7e008e | 4345 | if (need_align) |
b9654711 | 4346 | { |
0d7e008e | 4347 | need_align = 0; |
d3ae8277 | 4348 | scan = emit_label_after (gen_label_rtx (), scan); |
0d7e008e | 4349 | scan = emit_insn_after (gen_align_4 (), scan); |
b9654711 | 4350 | } |
b91455de KK |
4351 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
4352 | scan = emit_label_after (lab, scan); | |
4353 | scan = emit_insn_after (gen_consttable_4 (p->value, const0_rtx), | |
4354 | scan); | |
0d7e008e | 4355 | break; |
e577c183 | 4356 | case DFmode: |
0d7e008e SC |
4357 | case DImode: |
4358 | if (need_align) | |
4359 | { | |
4360 | need_align = 0; | |
d3ae8277 | 4361 | scan = emit_label_after (gen_label_rtx (), scan); |
0d7e008e SC |
4362 | scan = emit_insn_after (gen_align_4 (), scan); |
4363 | } | |
b91455de KK |
4364 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
4365 | scan = emit_label_after (lab, scan); | |
4366 | scan = emit_insn_after (gen_consttable_8 (p->value, const0_rtx), | |
4367 | scan); | |
0d7e008e SC |
4368 | break; |
4369 | default: | |
f5b9e7c9 | 4370 | gcc_unreachable (); |
b9654711 | 4371 | } |
b91455de KK |
4372 | |
4373 | if (p->mode != HImode) | |
4374 | { | |
6dd3c0a5 | 4375 | for (ref = p->wend; ref; ref = ref->next) |
b91455de | 4376 | { |
6dd3c0a5 | 4377 | lab = ref->label; |
b91455de KK |
4378 | scan = emit_insn_after (gen_consttable_window_end (lab), scan); |
4379 | } | |
4380 | } | |
b9654711 | 4381 | } |
b9654711 | 4382 | |
0d7e008e SC |
4383 | scan = emit_insn_after (gen_consttable_end (), scan); |
4384 | scan = emit_barrier_after (scan); | |
4385 | pool_size = 0; | |
b91455de KK |
4386 | pool_window_label = NULL_RTX; |
4387 | pool_window_last = 0; | |
0d7e008e | 4388 | } |
b9654711 | 4389 | |
5e7a8ee0 | 4390 | /* Return nonzero if constant would be an ok source for a |
8aa2a305 | 4391 | mov.w instead of a mov.l. */ |
b9654711 | 4392 | |
16bea517 | 4393 | static int |
cf277499 | 4394 | hi_const (rtx src) |
0d7e008e | 4395 | { |
f3536097 | 4396 | return (CONST_INT_P (src) |
8aa2a305 JW |
4397 | && INTVAL (src) >= -32768 |
4398 | && INTVAL (src) <= 32767); | |
b9654711 SC |
4399 | } |
4400 | ||
4241ecb0 R |
4401 | #define MOVA_LABELREF(mova) XVECEXP (SET_SRC (PATTERN (mova)), 0, 0) |
4402 | ||
5e7a8ee0 | 4403 | /* Nonzero if the insn is a move instruction which needs to be fixed. */ |
8aa2a305 JW |
4404 | |
4405 | /* ??? For a DImode/DFmode moves, we don't need to fix it if each half of the | |
735cb76e R |
4406 | CONST_DOUBLE input value is CONST_OK_FOR_I08. For a SFmode move, we don't |
4407 | need to fix it if the input value is CONST_OK_FOR_I08. */ | |
8aa2a305 JW |
4408 | |
4409 | static int | |
cf277499 | 4410 | broken_move (rtx insn) |
b9654711 | 4411 | { |
f3536097 | 4412 | if (NONJUMP_INSN_P (insn)) |
e577c183 JR |
4413 | { |
4414 | rtx pat = PATTERN (insn); | |
4415 | if (GET_CODE (pat) == PARALLEL) | |
4416 | pat = XVECEXP (pat, 0, 0); | |
4417 | if (GET_CODE (pat) == SET | |
3b962ea0 | 4418 | /* We can load any 8-bit value if we don't care what the high |
e577c183 JR |
4419 | order bits end up as. */ |
4420 | && GET_MODE (SET_DEST (pat)) != QImode | |
43c05634 AO |
4421 | && (CONSTANT_P (SET_SRC (pat)) |
4422 | /* Match mova_const. */ | |
4423 | || (GET_CODE (SET_SRC (pat)) == UNSPEC | |
4424 | && XINT (SET_SRC (pat), 1) == UNSPEC_MOVA | |
4425 | && GET_CODE (XVECEXP (SET_SRC (pat), 0, 0)) == CONST)) | |
3a8699c7 | 4426 | && ! (TARGET_SH2E |
1245df60 | 4427 | && GET_CODE (SET_SRC (pat)) == CONST_DOUBLE |
e577c183 JR |
4428 | && (fp_zero_operand (SET_SRC (pat)) |
4429 | || fp_one_operand (SET_SRC (pat))) | |
7a61cf6f | 4430 | /* In general we don't know the current setting of fpscr, so disable fldi. |
18778292 R |
4431 | There is an exception if this was a register-register move |
4432 | before reload - and hence it was ascertained that we have | |
4433 | single precision setting - and in a post-reload optimization | |
4434 | we changed this to do a constant load. In that case | |
4435 | we don't have an r0 clobber, hence we must use fldi. */ | |
7a61cf6f | 4436 | && (TARGET_FMOVD |
18778292 R |
4437 | || (GET_CODE (XEXP (XVECEXP (PATTERN (insn), 0, 2), 0)) |
4438 | == SCRATCH)) | |
f3536097 | 4439 | && REG_P (SET_DEST (pat)) |
104ee20b | 4440 | && FP_REGISTER_P (REGNO (SET_DEST (pat)))) |
157371cf AO |
4441 | && ! (TARGET_SH2A |
4442 | && GET_MODE (SET_DEST (pat)) == SImode | |
e990551a JS |
4443 | && (satisfies_constraint_I20 (SET_SRC (pat)) |
4444 | || satisfies_constraint_I28 (SET_SRC (pat)))) | |
32a7ab3d | 4445 | && ! satisfies_constraint_I08 (SET_SRC (pat))) |
e577c183 JR |
4446 | return 1; |
4447 | } | |
d3ae8277 | 4448 | |
8aa2a305 | 4449 | return 0; |
b9654711 | 4450 | } |
b9654711 | 4451 | |
1245df60 | 4452 | static int |
cf277499 | 4453 | mova_p (rtx insn) |
1245df60 | 4454 | { |
f3536097 | 4455 | return (NONJUMP_INSN_P (insn) |
1245df60 R |
4456 | && GET_CODE (PATTERN (insn)) == SET |
4457 | && GET_CODE (SET_SRC (PATTERN (insn))) == UNSPEC | |
43c05634 AO |
4458 | && XINT (SET_SRC (PATTERN (insn)), 1) == UNSPEC_MOVA |
4459 | /* Don't match mova_const. */ | |
4241ecb0 | 4460 | && GET_CODE (MOVA_LABELREF (insn)) == LABEL_REF); |
1245df60 R |
4461 | } |
4462 | ||
078c8b08 R |
4463 | /* Fix up a mova from a switch that went out of range. */ |
4464 | static void | |
4465 | fixup_mova (rtx mova) | |
4466 | { | |
4241ecb0 | 4467 | PUT_MODE (XEXP (MOVA_LABELREF (mova), 0), QImode); |
078c8b08 R |
4468 | if (! flag_pic) |
4469 | { | |
4241ecb0 | 4470 | SET_SRC (PATTERN (mova)) = MOVA_LABELREF (mova); |
078c8b08 R |
4471 | INSN_CODE (mova) = -1; |
4472 | } | |
4473 | else | |
4474 | { | |
4475 | rtx worker = mova; | |
4476 | rtx lab = gen_label_rtx (); | |
dc3ba671 | 4477 | rtx wpat, wpat0, wpat1, wsrc, target, base, diff; |
078c8b08 R |
4478 | |
4479 | do | |
4480 | { | |
4481 | worker = NEXT_INSN (worker); | |
f5b9e7c9 | 4482 | gcc_assert (worker |
f3536097 SZ |
4483 | && !LABEL_P (worker) |
4484 | && !JUMP_P (worker)); | |
4485 | } while (NOTE_P (worker) | |
0bf72672 | 4486 | || recog_memoized (worker) != CODE_FOR_casesi_worker_1); |
078c8b08 R |
4487 | wpat = PATTERN (worker); |
4488 | wpat0 = XVECEXP (wpat, 0, 0); | |
4489 | wpat1 = XVECEXP (wpat, 0, 1); | |
4490 | wsrc = SET_SRC (wpat0); | |
4491 | PATTERN (worker) = (gen_casesi_worker_2 | |
4492 | (SET_DEST (wpat0), XVECEXP (wsrc, 0, 1), | |
4493 | XEXP (XVECEXP (wsrc, 0, 2), 0), lab, | |
4494 | XEXP (wpat1, 0))); | |
4495 | INSN_CODE (worker) = -1; | |
dc3ba671 RS |
4496 | target = XVECEXP (SET_SRC (PATTERN (mova)), 0, 0); |
4497 | base = gen_rtx_LABEL_REF (Pmode, lab); | |
4498 | diff = gen_rtx_UNSPEC (Pmode, gen_rtvec (2, target, base), UNSPEC_SYMOFF); | |
078c8b08 R |
4499 | SET_SRC (PATTERN (mova)) = gen_rtx_CONST (Pmode, diff); |
4500 | INSN_CODE (mova) = -1; | |
4501 | } | |
4502 | } | |
4503 | ||
4241ecb0 R |
4504 | /* NEW_MOVA is a mova we've just encountered while scanning forward. Update |
4505 | *num_mova, and check if the new mova is not nested within the first one. | |
4506 | return 0 if *first_mova was replaced, 1 if new_mova was replaced, | |
4507 | 2 if new_mova has been assigned to *first_mova, -1 otherwise.. */ | |
4508 | static int | |
4509 | untangle_mova (int *num_mova, rtx *first_mova, rtx new_mova) | |
4510 | { | |
c6d56443 KK |
4511 | int n_addr = 0; /* Initialization to shut up spurious warning. */ |
4512 | int f_target, n_target = 0; /* Likewise. */ | |
4241ecb0 R |
4513 | |
4514 | if (optimize) | |
4515 | { | |
77253e56 KK |
4516 | /* If NEW_MOVA has no address yet, it will be handled later. */ |
4517 | if (INSN_ADDRESSES_SIZE() <= (unsigned) INSN_UID (new_mova)) | |
4518 | return -1; | |
4519 | ||
4241ecb0 R |
4520 | n_addr = INSN_ADDRESSES (INSN_UID (new_mova)); |
4521 | n_target = INSN_ADDRESSES (INSN_UID (XEXP (MOVA_LABELREF (new_mova), 0))); | |
4522 | if (n_addr > n_target || n_addr + 1022 < n_target) | |
4523 | { | |
4524 | /* Change the mova into a load. | |
4525 | broken_move will then return true for it. */ | |
4526 | fixup_mova (new_mova); | |
4527 | return 1; | |
4528 | } | |
4529 | } | |
4530 | if (!(*num_mova)++) | |
4531 | { | |
4532 | *first_mova = new_mova; | |
4533 | return 2; | |
4534 | } | |
4535 | if (!optimize | |
4536 | || ((f_target | |
4537 | = INSN_ADDRESSES (INSN_UID (XEXP (MOVA_LABELREF (*first_mova), 0)))) | |
4538 | >= n_target)) | |
4539 | return -1; | |
4540 | ||
4541 | (*num_mova)--; | |
4542 | if (f_target - INSN_ADDRESSES (INSN_UID (*first_mova)) | |
4543 | > n_target - n_addr) | |
4544 | { | |
4545 | fixup_mova (*first_mova); | |
4546 | return 0; | |
4547 | } | |
4548 | else | |
4549 | { | |
4550 | fixup_mova (new_mova); | |
4551 | return 1; | |
4552 | } | |
4553 | } | |
4554 | ||
8aa2a305 JW |
4555 | /* Find the last barrier from insn FROM which is close enough to hold the |
4556 | constant pool. If we can't find one, then create one near the end of | |
4557 | the range. */ | |
b9654711 | 4558 | |
8aa2a305 | 4559 | static rtx |
cf277499 | 4560 | find_barrier (int num_mova, rtx mova, rtx from) |
b9654711 | 4561 | { |
0d7e008e SC |
4562 | int count_si = 0; |
4563 | int count_hi = 0; | |
4564 | int found_hi = 0; | |
4565 | int found_si = 0; | |
fa5322fa | 4566 | int found_di = 0; |
33f7f353 JR |
4567 | int hi_align = 2; |
4568 | int si_align = 2; | |
1245df60 | 4569 | int leading_mova = num_mova; |
3a8699c7 | 4570 | rtx barrier_before_mova = 0, found_barrier = 0, good_barrier = 0; |
c17f53a8 JW |
4571 | int si_limit; |
4572 | int hi_limit; | |
7a3cb3ab | 4573 | rtx orig = from; |
e938feb4 | 4574 | rtx last_got = NULL_RTX; |
fc9c984d | 4575 | rtx last_symoff = NULL_RTX; |
8aa2a305 JW |
4576 | |
4577 | /* For HImode: range is 510, add 4 because pc counts from address of | |
4578 | second instruction after this one, subtract 2 for the jump instruction | |
3adbde60 JW |
4579 | that we may need to emit before the table, subtract 2 for the instruction |
4580 | that fills the jump delay slot (in very rare cases, reorg will take an | |
4581 | instruction from after the constant pool or will leave the delay slot | |
4582 | empty). This gives 510. | |
8aa2a305 JW |
4583 | For SImode: range is 1020, add 4 because pc counts from address of |
4584 | second instruction after this one, subtract 2 in case pc is 2 byte | |
4585 | aligned, subtract 2 for the jump instruction that we may need to emit | |
3adbde60 JW |
4586 | before the table, subtract 2 for the instruction that fills the jump |
4587 | delay slot. This gives 1018. */ | |
c17f53a8 | 4588 | |
1245df60 | 4589 | /* The branch will always be shortened now that the reference address for |
956d6950 | 4590 | forward branches is the successor address, thus we need no longer make |
1245df60 | 4591 | adjustments to the [sh]i_limit for -O0. */ |
c17f53a8 | 4592 | |
1245df60 R |
4593 | si_limit = 1018; |
4594 | hi_limit = 510; | |
e4fa6b06 | 4595 | |
c17f53a8 | 4596 | while (from && count_si < si_limit && count_hi < hi_limit) |
0d7e008e | 4597 | { |
33f7f353 JR |
4598 | int inc = get_attr_length (from); |
4599 | int new_align = 1; | |
1245df60 | 4600 | |
4241ecb0 R |
4601 | /* If this is a label that existed at the time of the compute_alignments |
4602 | call, determine the alignment. N.B. When find_barrier recurses for | |
4603 | an out-of-reach mova, we might see labels at the start of previously | |
4604 | inserted constant tables. */ | |
f3536097 | 4605 | if (LABEL_P (from) |
4241ecb0 | 4606 | && CODE_LABEL_NUMBER (from) <= max_labelno_before_reorg) |
77008a44 R |
4607 | { |
4608 | if (optimize) | |
4609 | new_align = 1 << label_to_alignment (from); | |
f3536097 | 4610 | else if (BARRIER_P (prev_nonnote_insn (from))) |
77008a44 R |
4611 | new_align = 1 << barrier_align (from); |
4612 | else | |
4613 | new_align = 1; | |
4614 | inc = 0; | |
4615 | } | |
4241ecb0 R |
4616 | /* In case we are scanning a constant table because of recursion, check |
4617 | for explicit alignments. If the table is long, we might be forced | |
4618 | to emit the new table in front of it; the length of the alignment | |
4619 | might be the last straw. */ | |
f3536097 | 4620 | else if (NONJUMP_INSN_P (from) |
4241ecb0 R |
4621 | && GET_CODE (PATTERN (from)) == UNSPEC_VOLATILE |
4622 | && XINT (PATTERN (from), 1) == UNSPECV_ALIGN) | |
4623 | new_align = INTVAL (XVECEXP (PATTERN (from), 0, 0)); | |
4624 | /* When we find the end of a constant table, paste the new constant | |
4625 | at the end. That is better than putting it in front because | |
4626 | this way, we don't need extra alignment for adding a 4-byte-aligned | |
4627 | mov(a) label to a 2/4 or 8/4 byte aligned table. */ | |
f3536097 | 4628 | else if (NONJUMP_INSN_P (from) |
4241ecb0 R |
4629 | && GET_CODE (PATTERN (from)) == UNSPEC_VOLATILE |
4630 | && XINT (PATTERN (from), 1) == UNSPECV_CONST_END) | |
4631 | return from; | |
8aa2a305 | 4632 | |
f3536097 | 4633 | if (BARRIER_P (from)) |
1245df60 | 4634 | { |
6d4d15b8 | 4635 | rtx next; |
33f7f353 | 4636 | |
1245df60 | 4637 | found_barrier = from; |
33f7f353 | 4638 | |
956d6950 | 4639 | /* If we are at the end of the function, or in front of an alignment |
1245df60 R |
4640 | instruction, we need not insert an extra alignment. We prefer |
4641 | this kind of barrier. */ | |
33f7f353 | 4642 | if (barrier_align (from) > 2) |
1245df60 | 4643 | good_barrier = from; |
6d4d15b8 KK |
4644 | |
4645 | /* If we are at the end of a hot/cold block, dump the constants | |
4646 | here. */ | |
4647 | next = NEXT_INSN (from); | |
4648 | if (next | |
4649 | && NOTE_P (next) | |
4650 | && NOTE_KIND (next) == NOTE_INSN_SWITCH_TEXT_SECTIONS) | |
4651 | break; | |
1245df60 | 4652 | } |
8aa2a305 | 4653 | |
8aa2a305 | 4654 | if (broken_move (from)) |
0d7e008e | 4655 | { |
1245df60 R |
4656 | rtx pat, src, dst; |
4657 | enum machine_mode mode; | |
4658 | ||
4659 | pat = PATTERN (from); | |
4660 | if (GET_CODE (pat) == PARALLEL) | |
4661 | pat = XVECEXP (pat, 0, 0); | |
4662 | src = SET_SRC (pat); | |
4663 | dst = SET_DEST (pat); | |
4664 | mode = GET_MODE (dst); | |
c17f53a8 | 4665 | |
e938feb4 CB |
4666 | /* GOT pcrelat setting comes in pair of |
4667 | mova .L8,r0 | |
4668 | mov.l .L8,r12 | |
4669 | instructions. (plus add r0,r12). | |
4670 | Remember if we see one without the other. */ | |
4671 | if (GET_CODE (src) == UNSPEC && PIC_ADDR_P (XVECEXP (src, 0, 0))) | |
4672 | last_got = last_got ? NULL_RTX : from; | |
4673 | else if (PIC_ADDR_P (src)) | |
4674 | last_got = last_got ? NULL_RTX : from; | |
4675 | ||
c17f53a8 JW |
4676 | /* We must explicitly check the mode, because sometimes the |
4677 | front end will generate code to load unsigned constants into | |
4678 | HImode targets without properly sign extending them. */ | |
225e4f43 R |
4679 | if (mode == HImode |
4680 | || (mode == SImode && hi_const (src) && REGNO (dst) != FPUL_REG)) | |
00e94d65 | 4681 | { |
1245df60 | 4682 | found_hi += 2; |
00e94d65 RK |
4683 | /* We put the short constants before the long constants, so |
4684 | we must count the length of short constants in the range | |
4685 | for the long constants. */ | |
4686 | /* ??? This isn't optimal, but is easy to do. */ | |
1245df60 | 4687 | si_limit -= 2; |
00e94d65 | 4688 | } |
0d7e008e | 4689 | else |
1245df60 | 4690 | { |
fa5322fa AO |
4691 | /* We dump DF/DI constants before SF/SI ones, because |
4692 | the limit is the same, but the alignment requirements | |
4693 | are higher. We may waste up to 4 additional bytes | |
4694 | for alignment, and the DF/DI constant may have | |
24746a42 | 4695 | another SF/SI constant placed before it. */ |
fa5322fa AO |
4696 | if (TARGET_SHCOMPACT |
4697 | && ! found_di | |
4698 | && (mode == DFmode || mode == DImode)) | |
4699 | { | |
4700 | found_di = 1; | |
4701 | si_limit -= 8; | |
4702 | } | |
33f7f353 JR |
4703 | while (si_align > 2 && found_si + si_align - 2 > count_si) |
4704 | si_align >>= 1; | |
1245df60 R |
4705 | if (found_si > count_si) |
4706 | count_si = found_si; | |
4707 | found_si += GET_MODE_SIZE (mode); | |
4708 | if (num_mova) | |
4709 | si_limit -= GET_MODE_SIZE (mode); | |
4710 | } | |
0d7e008e | 4711 | } |
5325c0fa | 4712 | |
33f7f353 | 4713 | if (mova_p (from)) |
1245df60 | 4714 | { |
4241ecb0 | 4715 | switch (untangle_mova (&num_mova, &mova, from)) |
1245df60 | 4716 | { |
fc9c984d KK |
4717 | case 1: |
4718 | if (flag_pic) | |
4719 | { | |
4720 | rtx src = SET_SRC (PATTERN (from)); | |
4721 | if (GET_CODE (src) == CONST | |
4722 | && GET_CODE (XEXP (src, 0)) == UNSPEC | |
4723 | && XINT (XEXP (src, 0), 1) == UNSPEC_SYMOFF) | |
4724 | last_symoff = from; | |
4725 | } | |
4726 | break; | |
4241ecb0 R |
4727 | case 0: return find_barrier (0, 0, mova); |
4728 | case 2: | |
4729 | { | |
4730 | leading_mova = 0; | |
4731 | barrier_before_mova | |
4732 | = good_barrier ? good_barrier : found_barrier; | |
4733 | } | |
4734 | default: break; | |
1245df60 R |
4735 | } |
4736 | if (found_si > count_si) | |
4737 | count_si = found_si; | |
4738 | } | |
f3536097 | 4739 | else if (JUMP_TABLE_DATA_P (from)) |
1245df60 | 4740 | { |
4241ecb0 R |
4741 | if ((num_mova > 1 && GET_MODE (prev_nonnote_insn (from)) == VOIDmode) |
4742 | || (num_mova | |
4743 | && (prev_nonnote_insn (from) | |
4744 | == XEXP (MOVA_LABELREF (mova), 0)))) | |
1245df60 | 4745 | num_mova--; |
10f4f635 | 4746 | if (barrier_align (next_real_insn (from)) == align_jumps_log) |
1245df60 | 4747 | { |
38e01259 | 4748 | /* We have just passed the barrier in front of the |
a0798779 R |
4749 | ADDR_DIFF_VEC, which is stored in found_barrier. Since |
4750 | the ADDR_DIFF_VEC is accessed as data, just like our pool | |
4751 | constants, this is a good opportunity to accommodate what | |
4752 | we have gathered so far. | |
1245df60 R |
4753 | If we waited any longer, we could end up at a barrier in |
4754 | front of code, which gives worse cache usage for separated | |
4755 | instruction / data caches. */ | |
a0798779 | 4756 | good_barrier = found_barrier; |
1245df60 R |
4757 | break; |
4758 | } | |
a0798779 R |
4759 | else |
4760 | { | |
4761 | rtx body = PATTERN (from); | |
4762 | inc = XVECLEN (body, 1) * GET_MODE_SIZE (GET_MODE (body)); | |
4763 | } | |
1245df60 | 4764 | } |
50ceefc2 | 4765 | /* For the SH1, we generate alignments even after jumps-around-jumps. */ |
f3536097 | 4766 | else if (JUMP_P (from) |
50ceefc2 | 4767 | && ! TARGET_SH2 |
3020190e | 4768 | && ! optimize_size) |
50ceefc2 | 4769 | new_align = 4; |
5325c0fa | 4770 | |
9e3c399e CB |
4771 | /* There is a possibility that a bf is transformed into a bf/s by the |
4772 | delay slot scheduler. */ | |
4773 | if (JUMP_P (from) && !JUMP_TABLE_DATA_P (from) | |
4774 | && get_attr_type (from) == TYPE_CBRANCH | |
4775 | && GET_CODE (PATTERN (NEXT_INSN (PREV_INSN (from)))) != SEQUENCE) | |
4776 | inc += 2; | |
4777 | ||
0d7e008e | 4778 | if (found_si) |
33f7f353 | 4779 | { |
50ceefc2 | 4780 | count_si += inc; |
33f7f353 JR |
4781 | if (new_align > si_align) |
4782 | { | |
318881c0 | 4783 | si_limit -= (count_si - 1) & (new_align - si_align); |
33f7f353 JR |
4784 | si_align = new_align; |
4785 | } | |
318881c0 | 4786 | count_si = (count_si + new_align - 1) & -new_align; |
33f7f353 | 4787 | } |
0d7e008e | 4788 | if (found_hi) |
33f7f353 | 4789 | { |
50ceefc2 | 4790 | count_hi += inc; |
33f7f353 JR |
4791 | if (new_align > hi_align) |
4792 | { | |
318881c0 | 4793 | hi_limit -= (count_hi - 1) & (new_align - hi_align); |
33f7f353 JR |
4794 | hi_align = new_align; |
4795 | } | |
318881c0 | 4796 | count_hi = (count_hi + new_align - 1) & -new_align; |
33f7f353 | 4797 | } |
0d7e008e SC |
4798 | from = NEXT_INSN (from); |
4799 | } | |
4800 | ||
1245df60 | 4801 | if (num_mova) |
eda44c78 R |
4802 | { |
4803 | if (leading_mova) | |
4804 | { | |
4805 | /* Try as we might, the leading mova is out of range. Change | |
4806 | it into a load (which will become a pcload) and retry. */ | |
078c8b08 | 4807 | fixup_mova (mova); |
eda44c78 R |
4808 | return find_barrier (0, 0, mova); |
4809 | } | |
4810 | else | |
4811 | { | |
4812 | /* Insert the constant pool table before the mova instruction, | |
4813 | to prevent the mova label reference from going out of range. */ | |
4814 | from = mova; | |
4815 | good_barrier = found_barrier = barrier_before_mova; | |
4816 | } | |
4817 | } | |
5325c0fa | 4818 | |
1245df60 R |
4819 | if (found_barrier) |
4820 | { | |
33f7f353 | 4821 | if (good_barrier && next_real_insn (found_barrier)) |
1245df60 | 4822 | found_barrier = good_barrier; |
1245df60 R |
4823 | } |
4824 | else | |
b9654711 | 4825 | { |
8aa2a305 JW |
4826 | /* We didn't find a barrier in time to dump our stuff, |
4827 | so we'll make one. */ | |
0d7e008e | 4828 | rtx label = gen_label_rtx (); |
8aa2a305 | 4829 | |
fc9c984d KK |
4830 | /* Don't emit a constant table in the middle of insns for |
4831 | casesi_worker_2. This is a bit overkill but is enough | |
4832 | because casesi_worker_2 wouldn't appear so frequently. */ | |
4833 | if (last_symoff) | |
4834 | from = last_symoff; | |
4835 | ||
5dbcc9c0 RK |
4836 | /* If we exceeded the range, then we must back up over the last |
4837 | instruction we looked at. Otherwise, we just need to undo the | |
4838 | NEXT_INSN at the end of the loop. */ | |
7a3cb3ab KK |
4839 | if (PREV_INSN (from) != orig |
4840 | && (count_hi > hi_limit || count_si > si_limit)) | |
5dbcc9c0 RK |
4841 | from = PREV_INSN (PREV_INSN (from)); |
4842 | else | |
4843 | from = PREV_INSN (from); | |
4844 | ||
e938feb4 CB |
4845 | /* Don't emit a constant table int the middle of global pointer setting, |
4846 | since that that would move the addressing base GOT into another table. | |
4847 | We need the first mov instruction before the _GLOBAL_OFFSET_TABLE_ | |
4848 | in the pool anyway, so just move up the whole constant pool. */ | |
4849 | if (last_got) | |
4850 | from = PREV_INSN (last_got); | |
4851 | ||
bd55b817 KK |
4852 | /* Don't insert the constant pool table at the position which |
4853 | may be the landing pad. */ | |
4854 | if (flag_exceptions | |
4855 | && CALL_P (from) | |
4856 | && find_reg_note (from, REG_EH_REGION, NULL_RTX)) | |
4857 | from = PREV_INSN (from); | |
4858 | ||
8aa2a305 JW |
4859 | /* Walk back to be just before any jump or label. |
4860 | Putting it before a label reduces the number of times the branch | |
4861 | around the constant pool table will be hit. Putting it before | |
4862 | a jump makes it more likely that the bra delay slot will be | |
4863 | filled. */ | |
f3536097 SZ |
4864 | while (NOTE_P (from) || JUMP_P (from) |
4865 | || LABEL_P (from)) | |
8aa2a305 JW |
4866 | from = PREV_INSN (from); |
4867 | ||
5969b52d KK |
4868 | /* Make sure we do not split between a call and its corresponding |
4869 | CALL_ARG_LOCATION note. */ | |
4870 | if (CALL_P (from)) | |
4871 | { | |
4872 | rtx next = NEXT_INSN (from); | |
4873 | if (next && NOTE_P (next) | |
4874 | && NOTE_KIND (next) == NOTE_INSN_CALL_ARG_LOCATION) | |
4875 | from = next; | |
4876 | } | |
4877 | ||
0d7e008e SC |
4878 | from = emit_jump_insn_after (gen_jump (label), from); |
4879 | JUMP_LABEL (from) = label; | |
e081ed26 | 4880 | LABEL_NUSES (label) = 1; |
0d7e008e SC |
4881 | found_barrier = emit_barrier_after (from); |
4882 | emit_label_after (label, found_barrier); | |
b9654711 | 4883 | } |
b9654711 | 4884 | |
8aa2a305 | 4885 | return found_barrier; |
0d7e008e | 4886 | } |
b9654711 | 4887 | |
4787bce0 JW |
4888 | /* If the instruction INSN is implemented by a special function, and we can |
4889 | positively find the register that is used to call the sfunc, and this | |
4890 | register is not used anywhere else in this instruction - except as the | |
4891 | destination of a set, return this register; else, return 0. */ | |
1245df60 | 4892 | rtx |
cf277499 | 4893 | sfunc_uses_reg (rtx insn) |
4787bce0 JW |
4894 | { |
4895 | int i; | |
4896 | rtx pattern, part, reg_part, reg; | |
4897 | ||
f3536097 | 4898 | if (!NONJUMP_INSN_P (insn)) |
4787bce0 JW |
4899 | return 0; |
4900 | pattern = PATTERN (insn); | |
4901 | if (GET_CODE (pattern) != PARALLEL || get_attr_type (insn) != TYPE_SFUNC) | |
4902 | return 0; | |
4903 | ||
4904 | for (reg_part = 0, i = XVECLEN (pattern, 0) - 1; i >= 1; i--) | |
4905 | { | |
4906 | part = XVECEXP (pattern, 0, i); | |
1245df60 | 4907 | if (GET_CODE (part) == USE && GET_MODE (XEXP (part, 0)) == SImode) |
4787bce0 JW |
4908 | reg_part = part; |
4909 | } | |
4910 | if (! reg_part) | |
4911 | return 0; | |
4912 | reg = XEXP (reg_part, 0); | |
4913 | for (i = XVECLEN (pattern, 0) - 1; i >= 0; i--) | |
4914 | { | |
4915 | part = XVECEXP (pattern, 0, i); | |
225e4f43 | 4916 | if (part == reg_part || GET_CODE (part) == CLOBBER) |
4787bce0 JW |
4917 | continue; |
4918 | if (reg_mentioned_p (reg, ((GET_CODE (part) == SET | |
f3536097 | 4919 | && REG_P (SET_DEST (part))) |
4787bce0 JW |
4920 | ? SET_SRC (part) : part))) |
4921 | return 0; | |
4922 | } | |
4923 | return reg; | |
4924 | } | |
4925 | ||
933c3ba3 JW |
4926 | /* See if the only way in which INSN uses REG is by calling it, or by |
4927 | setting it while calling it. Set *SET to a SET rtx if the register | |
4928 | is set by INSN. */ | |
4929 | ||
4930 | static int | |
cf277499 | 4931 | noncall_uses_reg (rtx reg, rtx insn, rtx *set) |
933c3ba3 | 4932 | { |
4787bce0 | 4933 | rtx pattern, reg2; |
933c3ba3 JW |
4934 | |
4935 | *set = NULL_RTX; | |
4936 | ||
4787bce0 JW |
4937 | reg2 = sfunc_uses_reg (insn); |
4938 | if (reg2 && REGNO (reg2) == REGNO (reg)) | |
4939 | { | |
4940 | pattern = single_set (insn); | |
4941 | if (pattern | |
f3536097 | 4942 | && REG_P (SET_DEST (pattern)) |
4787bce0 JW |
4943 | && REGNO (reg) == REGNO (SET_DEST (pattern))) |
4944 | *set = pattern; | |
4945 | return 0; | |
4946 | } | |
f3536097 | 4947 | if (!CALL_P (insn)) |
933c3ba3 JW |
4948 | { |
4949 | /* We don't use rtx_equal_p because we don't care if the mode is | |
4950 | different. */ | |
4951 | pattern = single_set (insn); | |
4952 | if (pattern | |
f3536097 | 4953 | && REG_P (SET_DEST (pattern)) |
933c3ba3 JW |
4954 | && REGNO (reg) == REGNO (SET_DEST (pattern))) |
4955 | { | |
4787bce0 JW |
4956 | rtx par, part; |
4957 | int i; | |
4958 | ||
933c3ba3 | 4959 | *set = pattern; |
4787bce0 JW |
4960 | par = PATTERN (insn); |
4961 | if (GET_CODE (par) == PARALLEL) | |
4962 | for (i = XVECLEN (par, 0) - 1; i >= 0; i--) | |
4963 | { | |
4964 | part = XVECEXP (par, 0, i); | |
4965 | if (GET_CODE (part) != SET && reg_mentioned_p (reg, part)) | |
4966 | return 1; | |
4967 | } | |
4968 | return reg_mentioned_p (reg, SET_SRC (pattern)); | |
933c3ba3 JW |
4969 | } |
4970 | ||
4971 | return 1; | |
4972 | } | |
4973 | ||
4974 | pattern = PATTERN (insn); | |
4975 | ||
4976 | if (GET_CODE (pattern) == PARALLEL) | |
4977 | { | |
4978 | int i; | |
4979 | ||
4980 | for (i = XVECLEN (pattern, 0) - 1; i >= 1; i--) | |
4981 | if (reg_mentioned_p (reg, XVECEXP (pattern, 0, i))) | |
4982 | return 1; | |
4983 | pattern = XVECEXP (pattern, 0, 0); | |
4984 | } | |
4985 | ||
4986 | if (GET_CODE (pattern) == SET) | |
4987 | { | |
4988 | if (reg_mentioned_p (reg, SET_DEST (pattern))) | |
4989 | { | |
4990 | /* We don't use rtx_equal_p, because we don't care if the | |
4991 | mode is different. */ | |
f3536097 | 4992 | if (!REG_P (SET_DEST (pattern)) |
933c3ba3 JW |
4993 | || REGNO (reg) != REGNO (SET_DEST (pattern))) |
4994 | return 1; | |
4995 | ||
4996 | *set = pattern; | |
4997 | } | |
4998 | ||
4999 | pattern = SET_SRC (pattern); | |
5000 | } | |
5001 | ||
5002 | if (GET_CODE (pattern) != CALL | |
f3536097 | 5003 | || !MEM_P (XEXP (pattern, 0)) |
933c3ba3 JW |
5004 | || ! rtx_equal_p (reg, XEXP (XEXP (pattern, 0), 0))) |
5005 | return 1; | |
5006 | ||
5007 | return 0; | |
5008 | } | |
5009 | ||
1245df60 R |
5010 | /* Given a X, a pattern of an insn or a part of it, return a mask of used |
5011 | general registers. Bits 0..15 mean that the respective registers | |
5012 | are used as inputs in the instruction. Bits 16..31 mean that the | |
5013 | registers 0..15, respectively, are used as outputs, or are clobbered. | |
5014 | IS_DEST should be set to 16 if X is the destination of a SET, else to 0. */ | |
5015 | int | |
cf277499 | 5016 | regs_used (rtx x, int is_dest) |
1245df60 R |
5017 | { |
5018 | enum rtx_code code; | |
6f7d635c | 5019 | const char *fmt; |
1245df60 R |
5020 | int i, used = 0; |
5021 | ||
5022 | if (! x) | |
5023 | return used; | |
5024 | code = GET_CODE (x); | |
5025 | switch (code) | |
5026 | { | |
5027 | case REG: | |
5028 | if (REGNO (x) < 16) | |
5029 | return (((1 << HARD_REGNO_NREGS (0, GET_MODE (x))) - 1) | |
5030 | << (REGNO (x) + is_dest)); | |
5031 | return 0; | |
5032 | case SUBREG: | |
5033 | { | |
5034 | rtx y = SUBREG_REG (x); | |
f676971a | 5035 | |
f3536097 | 5036 | if (!REG_P (y)) |
1245df60 R |
5037 | break; |
5038 | if (REGNO (y) < 16) | |
5039 | return (((1 << HARD_REGNO_NREGS (0, GET_MODE (x))) - 1) | |
ddef6bc7 JJ |
5040 | << (REGNO (y) + |
5041 | subreg_regno_offset (REGNO (y), | |
5042 | GET_MODE (y), | |
5043 | SUBREG_BYTE (x), | |
5044 | GET_MODE (x)) + is_dest)); | |
1245df60 R |
5045 | return 0; |
5046 | } | |
5047 | case SET: | |
5048 | return regs_used (SET_SRC (x), 0) | regs_used (SET_DEST (x), 16); | |
5049 | case RETURN: | |
5050 | /* If there was a return value, it must have been indicated with USE. */ | |
5051 | return 0x00ffff00; | |
5052 | case CLOBBER: | |
5053 | is_dest = 1; | |
5054 | break; | |
5055 | case MEM: | |
5056 | is_dest = 0; | |
5057 | break; | |
5058 | case CALL: | |
5059 | used |= 0x00ff00f0; | |
5060 | break; | |
318881c0 KG |
5061 | default: |
5062 | break; | |
1245df60 R |
5063 | } |
5064 | ||
5065 | fmt = GET_RTX_FORMAT (code); | |
5066 | ||
5067 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
5068 | { | |
5069 | if (fmt[i] == 'E') | |
5070 | { | |
5071 | register int j; | |
5072 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
5073 | used |= regs_used (XVECEXP (x, i, j), is_dest); | |
5074 | } | |
5075 | else if (fmt[i] == 'e') | |
5076 | used |= regs_used (XEXP (x, i), is_dest); | |
5077 | } | |
5078 | return used; | |
5079 | } | |
5080 | ||
5081 | /* Create an instruction that prevents redirection of a conditional branch | |
956d6950 | 5082 | to the destination of the JUMP with address ADDR. |
1245df60 R |
5083 | If the branch needs to be implemented as an indirect jump, try to find |
5084 | a scratch register for it. | |
5085 | If NEED_BLOCK is 0, don't do anything unless we need a scratch register. | |
5086 | If any preceding insn that doesn't fit into a delay slot is good enough, | |
5087 | pass 1. Pass 2 if a definite blocking insn is needed. | |
5088 | -1 is used internally to avoid deep recursion. | |
5089 | If a blocking instruction is made or recognized, return it. */ | |
f676971a | 5090 | |
1245df60 | 5091 | static rtx |
cf277499 | 5092 | gen_block_redirect (rtx jump, int addr, int need_block) |
1245df60 R |
5093 | { |
5094 | int dead = 0; | |
5095 | rtx prev = prev_nonnote_insn (jump); | |
5096 | rtx dest; | |
5097 | ||
5098 | /* First, check if we already have an instruction that satisfies our need. */ | |
f3536097 | 5099 | if (prev && NONJUMP_INSN_P (prev) && ! INSN_DELETED_P (prev)) |
1245df60 R |
5100 | { |
5101 | if (INSN_CODE (prev) == CODE_FOR_indirect_jump_scratch) | |
5102 | return prev; | |
5103 | if (GET_CODE (PATTERN (prev)) == USE | |
5104 | || GET_CODE (PATTERN (prev)) == CLOBBER | |
5105 | || get_attr_in_delay_slot (prev) == IN_DELAY_SLOT_YES) | |
5106 | prev = jump; | |
5107 | else if ((need_block &= ~1) < 0) | |
5108 | return prev; | |
5109 | else if (recog_memoized (prev) == CODE_FOR_block_branch_redirect) | |
5110 | need_block = 0; | |
5111 | } | |
f9654065 R |
5112 | if (GET_CODE (PATTERN (jump)) == RETURN) |
5113 | { | |
5114 | if (! need_block) | |
5115 | return prev; | |
5116 | /* Reorg even does nasty things with return insns that cause branches | |
5117 | to go out of range - see find_end_label and callers. */ | |
a556fd39 | 5118 | return emit_insn_before (gen_block_branch_redirect (const0_rtx) , jump); |
f9654065 | 5119 | } |
1245df60 R |
5120 | /* We can't use JUMP_LABEL here because it might be undefined |
5121 | when not optimizing. */ | |
5122 | dest = XEXP (SET_SRC (PATTERN (jump)), 0); | |
5123 | /* If the branch is out of range, try to find a scratch register for it. */ | |
5124 | if (optimize | |
11f9ed1a KG |
5125 | && (INSN_ADDRESSES (INSN_UID (dest)) - addr + (unsigned) 4092 |
5126 | > 4092 + 4098)) | |
1245df60 R |
5127 | { |
5128 | rtx scan; | |
5129 | /* Don't look for the stack pointer as a scratch register, | |
956d6950 | 5130 | it would cause trouble if an interrupt occurred. */ |
0a2aaacc | 5131 | unsigned attempt = 0x7fff, used; |
1245df60 | 5132 | int jump_left = flag_expensive_optimizations + 1; |
f676971a | 5133 | |
1245df60 R |
5134 | /* It is likely that the most recent eligible instruction is wanted for |
5135 | the delay slot. Therefore, find out which registers it uses, and | |
5136 | try to avoid using them. */ | |
f676971a | 5137 | |
318881c0 | 5138 | for (scan = jump; (scan = PREV_INSN (scan)); ) |
1245df60 R |
5139 | { |
5140 | enum rtx_code code; | |
5141 | ||
5142 | if (INSN_DELETED_P (scan)) | |
5143 | continue; | |
5144 | code = GET_CODE (scan); | |
5145 | if (code == CODE_LABEL || code == JUMP_INSN) | |
5146 | break; | |
5147 | if (code == INSN | |
5148 | && GET_CODE (PATTERN (scan)) != USE | |
5149 | && GET_CODE (PATTERN (scan)) != CLOBBER | |
5150 | && get_attr_in_delay_slot (scan) == IN_DELAY_SLOT_YES) | |
5151 | { | |
0a2aaacc | 5152 | attempt &= ~regs_used (PATTERN (scan), 0); |
1245df60 R |
5153 | break; |
5154 | } | |
5155 | } | |
eda44c78 R |
5156 | for (used = dead = 0, scan = JUMP_LABEL (jump); |
5157 | (scan = NEXT_INSN (scan)); ) | |
1245df60 R |
5158 | { |
5159 | enum rtx_code code; | |
5160 | ||
5161 | if (INSN_DELETED_P (scan)) | |
5162 | continue; | |
5163 | code = GET_CODE (scan); | |
ec8e098d | 5164 | if (INSN_P (scan)) |
1245df60 R |
5165 | { |
5166 | used |= regs_used (PATTERN (scan), 0); | |
5167 | if (code == CALL_INSN) | |
5168 | used |= regs_used (CALL_INSN_FUNCTION_USAGE (scan), 0); | |
5169 | dead |= (used >> 16) & ~used; | |
0a2aaacc | 5170 | if (dead & attempt) |
1245df60 | 5171 | { |
0a2aaacc | 5172 | dead &= attempt; |
1245df60 R |
5173 | break; |
5174 | } | |
5175 | if (code == JUMP_INSN) | |
eda44c78 R |
5176 | { |
5177 | if (jump_left-- && simplejump_p (scan)) | |
5178 | scan = JUMP_LABEL (scan); | |
5179 | else | |
5180 | break; | |
5181 | } | |
1245df60 R |
5182 | } |
5183 | } | |
5184 | /* Mask out the stack pointer again, in case it was | |
5185 | the only 'free' register we have found. */ | |
5186 | dead &= 0x7fff; | |
5187 | } | |
5188 | /* If the immediate destination is still in range, check for possible | |
5189 | threading with a jump beyond the delay slot insn. | |
5190 | Don't check if we are called recursively; the jump has been or will be | |
956d6950 | 5191 | checked in a different invocation then. */ |
f676971a | 5192 | |
1245df60 R |
5193 | else if (optimize && need_block >= 0) |
5194 | { | |
5195 | rtx next = next_active_insn (next_active_insn (dest)); | |
f3536097 | 5196 | if (next && JUMP_P (next) |
1245df60 | 5197 | && GET_CODE (PATTERN (next)) == SET |
ddf8ec7e | 5198 | && recog_memoized (next) == CODE_FOR_jump_compact) |
1245df60 R |
5199 | { |
5200 | dest = JUMP_LABEL (next); | |
5201 | if (dest | |
11f9ed1a | 5202 | && (INSN_ADDRESSES (INSN_UID (dest)) - addr + (unsigned) 4092 |
9d98a694 AO |
5203 | > 4092 + 4098)) |
5204 | gen_block_redirect (next, INSN_ADDRESSES (INSN_UID (next)), -1); | |
1245df60 R |
5205 | } |
5206 | } | |
5207 | ||
5208 | if (dead) | |
5209 | { | |
c5c76735 | 5210 | rtx reg = gen_rtx_REG (SImode, exact_log2 (dead & -dead)); |
1245df60 R |
5211 | |
5212 | /* It would be nice if we could convert the jump into an indirect | |
956d6950 | 5213 | jump / far branch right now, and thus exposing all constituent |
1245df60 R |
5214 | instructions to further optimization. However, reorg uses |
5215 | simplejump_p to determine if there is an unconditional jump where | |
5216 | it should try to schedule instructions from the target of the | |
5217 | branch; simplejump_p fails for indirect jumps even if they have | |
5218 | a JUMP_LABEL. */ | |
5219 | rtx insn = emit_insn_before (gen_indirect_jump_scratch | |
826159e5 KK |
5220 | (reg, GEN_INT (unspec_bbr_uid++)), |
5221 | jump); | |
10f4f635 R |
5222 | /* ??? We would like this to have the scope of the jump, but that |
5223 | scope will change when a delay slot insn of an inner scope is added. | |
5224 | Hence, after delay slot scheduling, we'll have to expect | |
5225 | NOTE_INSN_BLOCK_END notes between the indirect_jump_scratch and | |
5226 | the jump. */ | |
f676971a | 5227 | |
86143cb4 | 5228 | INSN_LOCATOR (insn) = INSN_LOCATOR (jump); |
1245df60 R |
5229 | INSN_CODE (insn) = CODE_FOR_indirect_jump_scratch; |
5230 | return insn; | |
5231 | } | |
5232 | else if (need_block) | |
5233 | /* We can't use JUMP_LABEL here because it might be undefined | |
5234 | when not optimizing. */ | |
5235 | return emit_insn_before (gen_block_branch_redirect | |
826159e5 KK |
5236 | (GEN_INT (unspec_bbr_uid++)), |
5237 | jump); | |
1245df60 R |
5238 | return prev; |
5239 | } | |
5240 | ||
5241 | #define CONDJUMP_MIN -252 | |
5242 | #define CONDJUMP_MAX 262 | |
5243 | struct far_branch | |
5244 | { | |
5245 | /* A label (to be placed) in front of the jump | |
5246 | that jumps to our ultimate destination. */ | |
5247 | rtx near_label; | |
5248 | /* Where we are going to insert it if we cannot move the jump any farther, | |
5249 | or the jump itself if we have picked up an existing jump. */ | |
5250 | rtx insert_place; | |
5251 | /* The ultimate destination. */ | |
5252 | rtx far_label; | |
5253 | struct far_branch *prev; | |
5254 | /* If the branch has already been created, its address; | |
5255 | else the address of its first prospective user. */ | |
5256 | int address; | |
5257 | }; | |
5258 | ||
cf277499 | 5259 | static void gen_far_branch (struct far_branch *); |
1245df60 | 5260 | enum mdep_reorg_phase_e mdep_reorg_phase; |
c5b9ef02 | 5261 | static void |
cf277499 | 5262 | gen_far_branch (struct far_branch *bp) |
1245df60 R |
5263 | { |
5264 | rtx insn = bp->insert_place; | |
5265 | rtx jump; | |
5266 | rtx label = gen_label_rtx (); | |
f5b9e7c9 | 5267 | int ok; |
1245df60 R |
5268 | |
5269 | emit_label_after (label, insn); | |
5270 | if (bp->far_label) | |
5271 | { | |
5272 | jump = emit_jump_insn_after (gen_jump (bp->far_label), insn); | |
5273 | LABEL_NUSES (bp->far_label)++; | |
5274 | } | |
5275 | else | |
5276 | jump = emit_jump_insn_after (gen_return (), insn); | |
225e4f43 R |
5277 | /* Emit a barrier so that reorg knows that any following instructions |
5278 | are not reachable via a fall-through path. | |
fae778eb | 5279 | But don't do this when not optimizing, since we wouldn't suppress the |
225e4f43 R |
5280 | alignment for the barrier then, and could end up with out-of-range |
5281 | pc-relative loads. */ | |
5282 | if (optimize) | |
5283 | emit_barrier_after (jump); | |
1245df60 R |
5284 | emit_label_after (bp->near_label, insn); |
5285 | JUMP_LABEL (jump) = bp->far_label; | |
f5b9e7c9 NS |
5286 | ok = invert_jump (insn, label, 1); |
5287 | gcc_assert (ok); | |
5288 | ||
f9654065 R |
5289 | /* If we are branching around a jump (rather than a return), prevent |
5290 | reorg from using an insn from the jump target as the delay slot insn - | |
5291 | when reorg did this, it pessimized code (we rather hide the delay slot) | |
5292 | and it could cause branches to go out of range. */ | |
5293 | if (bp->far_label) | |
5294 | (emit_insn_after | |
5295 | (gen_stuff_delay_slot | |
826159e5 | 5296 | (GEN_INT (unspec_bbr_uid++), |
f9654065 R |
5297 | GEN_INT (recog_memoized (insn) == CODE_FOR_branch_false)), |
5298 | insn)); | |
1245df60 R |
5299 | /* Prevent reorg from undoing our splits. */ |
5300 | gen_block_redirect (jump, bp->address += 2, 2); | |
5301 | } | |
5302 | ||
1245df60 R |
5303 | /* Fix up ADDR_DIFF_VECs. */ |
5304 | void | |
cf277499 | 5305 | fixup_addr_diff_vecs (rtx first) |
1245df60 R |
5306 | { |
5307 | rtx insn; | |
33f7f353 | 5308 | |
1245df60 R |
5309 | for (insn = first; insn; insn = NEXT_INSN (insn)) |
5310 | { | |
eb3881bf | 5311 | rtx vec_lab, pat, prev, prevpat, x, braf_label; |
1245df60 | 5312 | |
f3536097 | 5313 | if (!JUMP_P (insn) |
1245df60 R |
5314 | || GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC) |
5315 | continue; | |
5316 | pat = PATTERN (insn); | |
33f7f353 | 5317 | vec_lab = XEXP (XEXP (pat, 0), 0); |
1245df60 | 5318 | |
33f7f353 JR |
5319 | /* Search the matching casesi_jump_2. */ |
5320 | for (prev = vec_lab; ; prev = PREV_INSN (prev)) | |
1245df60 | 5321 | { |
f3536097 | 5322 | if (!JUMP_P (prev)) |
33f7f353 JR |
5323 | continue; |
5324 | prevpat = PATTERN (prev); | |
5325 | if (GET_CODE (prevpat) != PARALLEL || XVECLEN (prevpat, 0) != 2) | |
5326 | continue; | |
5327 | x = XVECEXP (prevpat, 0, 1); | |
5328 | if (GET_CODE (x) != USE) | |
5329 | continue; | |
5330 | x = XEXP (x, 0); | |
5331 | if (GET_CODE (x) == LABEL_REF && XEXP (x, 0) == vec_lab) | |
5332 | break; | |
1245df60 | 5333 | } |
157371cf AO |
5334 | /* FIXME: This is a bug in the optimizer, but it seems harmless |
5335 | to just avoid panicing. */ | |
5336 | if (!prev) | |
5337 | continue; | |
eb3881bf R |
5338 | |
5339 | /* Emit the reference label of the braf where it belongs, right after | |
5340 | the casesi_jump_2 (i.e. braf). */ | |
5341 | braf_label = XEXP (XEXP (SET_SRC (XVECEXP (prevpat, 0, 0)), 1), 0); | |
5342 | emit_label_after (braf_label, prev); | |
5343 | ||
33f7f353 JR |
5344 | /* Fix up the ADDR_DIF_VEC to be relative |
5345 | to the reference address of the braf. */ | |
eb3881bf | 5346 | XEXP (XEXP (pat, 0), 0) = braf_label; |
1245df60 | 5347 | } |
1245df60 R |
5348 | } |
5349 | ||
33f7f353 JR |
5350 | /* BARRIER_OR_LABEL is either a BARRIER or a CODE_LABEL immediately following |
5351 | a barrier. Return the base 2 logarithm of the desired alignment. */ | |
1245df60 | 5352 | int |
cf277499 | 5353 | barrier_align (rtx barrier_or_label) |
1245df60 | 5354 | { |
33f7f353 | 5355 | rtx next = next_real_insn (barrier_or_label), pat, prev; |
3a8699c7 | 5356 | int slot, credit, jump_to_next = 0; |
f676971a | 5357 | |
33f7f353 | 5358 | if (! next) |
1245df60 | 5359 | return 0; |
1245df60 | 5360 | |
33f7f353 | 5361 | pat = PATTERN (next); |
1245df60 | 5362 | |
33f7f353 JR |
5363 | if (GET_CODE (pat) == ADDR_DIFF_VEC) |
5364 | return 2; | |
1245df60 | 5365 | |
4773afa4 | 5366 | if (GET_CODE (pat) == UNSPEC_VOLATILE && XINT (pat, 1) == UNSPECV_ALIGN) |
33f7f353 JR |
5367 | /* This is a barrier in front of a constant table. */ |
5368 | return 0; | |
5369 | ||
5370 | prev = prev_real_insn (barrier_or_label); | |
5371 | if (GET_CODE (PATTERN (prev)) == ADDR_DIFF_VEC) | |
1245df60 | 5372 | { |
33f7f353 JR |
5373 | pat = PATTERN (prev); |
5374 | /* If this is a very small table, we want to keep the alignment after | |
5375 | the table to the minimum for proper code alignment. */ | |
3020190e | 5376 | return ((optimize_size |
d042370e | 5377 | || ((unsigned) XVECLEN (pat, 1) * GET_MODE_SIZE (GET_MODE (pat)) |
832a3292 | 5378 | <= (unsigned) 1 << (CACHE_LOG - 2))) |
10f4f635 | 5379 | ? 1 << TARGET_SHMEDIA : align_jumps_log); |
1245df60 | 5380 | } |
33f7f353 | 5381 | |
3020190e | 5382 | if (optimize_size) |
33f7f353 JR |
5383 | return 0; |
5384 | ||
4d070fd3 | 5385 | if (! TARGET_SH2 || ! optimize) |
10f4f635 | 5386 | return align_jumps_log; |
33f7f353 | 5387 | |
225e4f43 R |
5388 | /* When fixing up pcloads, a constant table might be inserted just before |
5389 | the basic block that ends with the barrier. Thus, we can't trust the | |
5390 | instruction lengths before that. */ | |
5391 | if (mdep_reorg_phase > SH_FIXUP_PCLOAD) | |
1245df60 | 5392 | { |
225e4f43 R |
5393 | /* Check if there is an immediately preceding branch to the insn beyond |
5394 | the barrier. We must weight the cost of discarding useful information | |
5395 | from the current cache line when executing this branch and there is | |
5396 | an alignment, against that of fetching unneeded insn in front of the | |
5397 | branch target when there is no alignment. */ | |
5398 | ||
f676971a EC |
5399 | /* There are two delay_slot cases to consider. One is the simple case |
5400 | where the preceding branch is to the insn beyond the barrier (simple | |
5401 | delay slot filling), and the other is where the preceding branch has | |
5402 | a delay slot that is a duplicate of the insn after the barrier | |
5403 | (fill_eager_delay_slots) and the branch is to the insn after the insn | |
b1b41fe4 CP |
5404 | after the barrier. */ |
5405 | ||
225e4f43 R |
5406 | /* PREV is presumed to be the JUMP_INSN for the barrier under |
5407 | investigation. Skip to the insn before it. */ | |
5408 | prev = prev_real_insn (prev); | |
5409 | ||
c5b9ef02 | 5410 | for (slot = 2, credit = (1 << (CACHE_LOG - 2)) + 2; |
f3536097 | 5411 | credit >= 0 && prev && NONJUMP_INSN_P (prev); |
225e4f43 R |
5412 | prev = prev_real_insn (prev)) |
5413 | { | |
b1b41fe4 | 5414 | jump_to_next = 0; |
225e4f43 R |
5415 | if (GET_CODE (PATTERN (prev)) == USE |
5416 | || GET_CODE (PATTERN (prev)) == CLOBBER) | |
5417 | continue; | |
5418 | if (GET_CODE (PATTERN (prev)) == SEQUENCE) | |
b1b41fe4 CP |
5419 | { |
5420 | prev = XVECEXP (PATTERN (prev), 0, 1); | |
f676971a | 5421 | if (INSN_UID (prev) == INSN_UID (next)) |
b1b41fe4 CP |
5422 | { |
5423 | /* Delay slot was filled with insn at jump target. */ | |
5424 | jump_to_next = 1; | |
5425 | continue; | |
5426 | } | |
5427 | } | |
5428 | ||
225e4f43 R |
5429 | if (slot && |
5430 | get_attr_in_delay_slot (prev) == IN_DELAY_SLOT_YES) | |
5431 | slot = 0; | |
5432 | credit -= get_attr_length (prev); | |
5433 | } | |
5434 | if (prev | |
f3536097 | 5435 | && JUMP_P (prev) |
2b9d17c7 AH |
5436 | && JUMP_LABEL (prev)) |
5437 | { | |
5438 | rtx x; | |
300c5453 | 5439 | if (jump_to_next |
2b9d17c7 | 5440 | || next_real_insn (JUMP_LABEL (prev)) == next |
1594c6bc AO |
5441 | /* If relax_delay_slots() decides NEXT was redundant |
5442 | with some previous instruction, it will have | |
5443 | redirected PREV's jump to the following insn. */ | |
5d291213 | 5444 | || JUMP_LABEL (prev) == next_nonnote_insn (next) |
300c5453 AO |
5445 | /* There is no upper bound on redundant instructions |
5446 | that might have been skipped, but we must not put an | |
2b9d17c7 | 5447 | alignment where none had been before. */ |
f676971a EC |
5448 | || (x = (NEXT_INSN (NEXT_INSN (PREV_INSN (prev)))), |
5449 | (INSN_P (x) | |
2b9d17c7 | 5450 | && (INSN_CODE (x) == CODE_FOR_block_branch_redirect |
9959c8c7 R |
5451 | || INSN_CODE (x) == CODE_FOR_indirect_jump_scratch |
5452 | || INSN_CODE (x) == CODE_FOR_stuff_delay_slot)))) | |
2b9d17c7 AH |
5453 | { |
5454 | rtx pat = PATTERN (prev); | |
5455 | if (GET_CODE (pat) == PARALLEL) | |
300c5453 | 5456 | pat = XVECEXP (pat, 0, 0); |
2b9d17c7 AH |
5457 | if (credit - slot >= (GET_CODE (SET_SRC (pat)) == PC ? 2 : 0)) |
5458 | return 0; | |
5459 | } | |
300c5453 | 5460 | } |
33f7f353 | 5461 | } |
f676971a | 5462 | |
10f4f635 | 5463 | return align_jumps_log; |
1245df60 R |
5464 | } |
5465 | ||
2148624a R |
5466 | /* If we are inside a phony loop, almost any kind of label can turn up as the |
5467 | first one in the loop. Aligning a braf label causes incorrect switch | |
5468 | destination addresses; we can detect braf labels because they are | |
5469 | followed by a BARRIER. | |
5470 | Applying loop alignment to small constant or switch tables is a waste | |
5471 | of space, so we suppress this too. */ | |
5472 | int | |
cf277499 | 5473 | sh_loop_align (rtx label) |
2148624a R |
5474 | { |
5475 | rtx next = label; | |
5476 | ||
5477 | do | |
5478 | next = next_nonnote_insn (next); | |
f3536097 | 5479 | while (next && LABEL_P (next)); |
2148624a R |
5480 | |
5481 | if (! next | |
2c3c49de | 5482 | || ! INSN_P (next) |
2148624a R |
5483 | || GET_CODE (PATTERN (next)) == ADDR_DIFF_VEC |
5484 | || recog_memoized (next) == CODE_FOR_consttable_2) | |
5485 | return 0; | |
fa5322fa | 5486 | |
10f4f635 | 5487 | return align_loops_log; |
2148624a R |
5488 | } |
5489 | ||
18dbd950 | 5490 | /* Do a final pass over the function, just before delayed branch |
933c3ba3 | 5491 | scheduling. */ |
b9654711 | 5492 | |
18dbd950 | 5493 | static void |
cf277499 | 5494 | sh_reorg (void) |
0d7e008e | 5495 | { |
18dbd950 | 5496 | rtx first, insn, mova = NULL_RTX; |
1245df60 | 5497 | int num_mova; |
c5c76735 JL |
5498 | rtx r0_rtx = gen_rtx_REG (Pmode, 0); |
5499 | rtx r0_inc_rtx = gen_rtx_POST_INC (Pmode, r0_rtx); | |
8aa2a305 | 5500 | |
18dbd950 | 5501 | first = get_insns (); |
4241ecb0 | 5502 | max_labelno_before_reorg = max_label_num (); |
18dbd950 | 5503 | |
cb51ecd2 AO |
5504 | /* We must split call insns before introducing `mova's. If we're |
5505 | optimizing, they'll have already been split. Otherwise, make | |
5506 | sure we don't split them too late. */ | |
5507 | if (! optimize) | |
6f862f2f | 5508 | split_all_insns_noflow (); |
cb51ecd2 | 5509 | |
fa5322fa AO |
5510 | if (TARGET_SHMEDIA) |
5511 | return; | |
5512 | ||
933c3ba3 JW |
5513 | /* If relaxing, generate pseudo-ops to associate function calls with |
5514 | the symbols they call. It does no harm to not generate these | |
5515 | pseudo-ops. However, when we can generate them, it enables to | |
5516 | linker to potentially relax the jsr to a bsr, and eliminate the | |
5517 | register load and, possibly, the constant pool entry. */ | |
5518 | ||
1245df60 | 5519 | mdep_reorg_phase = SH_INSERT_USES_LABELS; |
933c3ba3 JW |
5520 | if (TARGET_RELAX) |
5521 | { | |
cf7c4aa6 HPN |
5522 | /* Remove all REG_LABEL_OPERAND notes. We want to use them for our |
5523 | own purposes. This works because none of the remaining passes | |
933c3ba3 JW |
5524 | need to look at them. |
5525 | ||
5526 | ??? But it may break in the future. We should use a machine | |
5527 | dependent REG_NOTE, or some other approach entirely. */ | |
5528 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
5529 | { | |
2c3c49de | 5530 | if (INSN_P (insn)) |
933c3ba3 JW |
5531 | { |
5532 | rtx note; | |
5533 | ||
cf7c4aa6 HPN |
5534 | while ((note = find_reg_note (insn, REG_LABEL_OPERAND, |
5535 | NULL_RTX)) != 0) | |
933c3ba3 JW |
5536 | remove_note (insn, note); |
5537 | } | |
5538 | } | |
5539 | ||
5540 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
5541 | { | |
5542 | rtx pattern, reg, link, set, scan, dies, label; | |
5543 | int rescan = 0, foundinsn = 0; | |
5544 | ||
f3536097 | 5545 | if (CALL_P (insn)) |
4787bce0 JW |
5546 | { |
5547 | pattern = PATTERN (insn); | |
933c3ba3 | 5548 | |
4787bce0 JW |
5549 | if (GET_CODE (pattern) == PARALLEL) |
5550 | pattern = XVECEXP (pattern, 0, 0); | |
5551 | if (GET_CODE (pattern) == SET) | |
5552 | pattern = SET_SRC (pattern); | |
933c3ba3 | 5553 | |
4787bce0 | 5554 | if (GET_CODE (pattern) != CALL |
f3536097 | 5555 | || !MEM_P (XEXP (pattern, 0))) |
4787bce0 | 5556 | continue; |
933c3ba3 | 5557 | |
4787bce0 JW |
5558 | reg = XEXP (XEXP (pattern, 0), 0); |
5559 | } | |
5560 | else | |
5561 | { | |
5562 | reg = sfunc_uses_reg (insn); | |
5563 | if (! reg) | |
5564 | continue; | |
5565 | } | |
933c3ba3 | 5566 | |
f3536097 | 5567 | if (!REG_P (reg)) |
933c3ba3 JW |
5568 | continue; |
5569 | ||
6fb5fa3c DB |
5570 | /* Try scanning backward to find where the register is set. */ |
5571 | link = NULL; | |
5572 | for (scan = PREV_INSN (insn); | |
f3536097 | 5573 | scan && !LABEL_P (scan); |
6fb5fa3c | 5574 | scan = PREV_INSN (scan)) |
933c3ba3 | 5575 | { |
6fb5fa3c DB |
5576 | if (! INSN_P (scan)) |
5577 | continue; | |
933c3ba3 | 5578 | |
6fb5fa3c DB |
5579 | if (! reg_mentioned_p (reg, scan)) |
5580 | continue; | |
933c3ba3 | 5581 | |
6fb5fa3c DB |
5582 | if (noncall_uses_reg (reg, scan, &set)) |
5583 | break; | |
933c3ba3 | 5584 | |
6fb5fa3c DB |
5585 | if (set) |
5586 | { | |
5587 | link = scan; | |
5588 | break; | |
933c3ba3 JW |
5589 | } |
5590 | } | |
5591 | ||
5592 | if (! link) | |
5593 | continue; | |
5594 | ||
5595 | /* The register is set at LINK. */ | |
5596 | ||
5597 | /* We can only optimize the function call if the register is | |
5598 | being set to a symbol. In theory, we could sometimes | |
5599 | optimize calls to a constant location, but the assembler | |
5600 | and linker do not support that at present. */ | |
5601 | if (GET_CODE (SET_SRC (set)) != SYMBOL_REF | |
5602 | && GET_CODE (SET_SRC (set)) != LABEL_REF) | |
5603 | continue; | |
5604 | ||
5605 | /* Scan forward from LINK to the place where REG dies, and | |
5606 | make sure that the only insns which use REG are | |
5607 | themselves function calls. */ | |
5608 | ||
c1e8e6bd JW |
5609 | /* ??? This doesn't work for call targets that were allocated |
5610 | by reload, since there may not be a REG_DEAD note for the | |
5611 | register. */ | |
5612 | ||
933c3ba3 JW |
5613 | dies = NULL_RTX; |
5614 | for (scan = NEXT_INSN (link); scan; scan = NEXT_INSN (scan)) | |
5615 | { | |
5616 | rtx scanset; | |
5617 | ||
c1e8e6bd JW |
5618 | /* Don't try to trace forward past a CODE_LABEL if we haven't |
5619 | seen INSN yet. Ordinarily, we will only find the setting insn | |
6fb5fa3c | 5620 | if it is in the same basic block. However, |
c1e8e6bd JW |
5621 | cross-jumping can insert code labels in between the load and |
5622 | the call, and can result in situations where a single call | |
5623 | insn may have two targets depending on where we came from. */ | |
5624 | ||
f3536097 | 5625 | if (LABEL_P (scan) && ! foundinsn) |
c1e8e6bd JW |
5626 | break; |
5627 | ||
2c3c49de | 5628 | if (! INSN_P (scan)) |
933c3ba3 JW |
5629 | continue; |
5630 | ||
5631 | /* Don't try to trace forward past a JUMP. To optimize | |
5632 | safely, we would have to check that all the | |
c1e8e6bd | 5633 | instructions at the jump destination did not use REG. */ |
933c3ba3 | 5634 | |
f3536097 | 5635 | if (JUMP_P (scan)) |
933c3ba3 JW |
5636 | break; |
5637 | ||
5638 | if (! reg_mentioned_p (reg, scan)) | |
5639 | continue; | |
5640 | ||
5641 | if (noncall_uses_reg (reg, scan, &scanset)) | |
5642 | break; | |
5643 | ||
5644 | if (scan == insn) | |
5645 | foundinsn = 1; | |
5646 | ||
4787bce0 | 5647 | if (scan != insn |
f3536097 | 5648 | && (CALL_P (scan) || sfunc_uses_reg (scan))) |
933c3ba3 JW |
5649 | { |
5650 | /* There is a function call to this register other | |
5651 | than the one we are checking. If we optimize | |
5652 | this call, we need to rescan again below. */ | |
5653 | rescan = 1; | |
5654 | } | |
5655 | ||
5656 | /* ??? We shouldn't have to worry about SCANSET here. | |
5657 | We should just be able to check for a REG_DEAD note | |
5658 | on a function call. However, the REG_DEAD notes are | |
5659 | apparently not dependable around libcalls; c-torture | |
5660 | execute/920501-2 is a test case. If SCANSET is set, | |
5661 | then this insn sets the register, so it must have | |
5662 | died earlier. Unfortunately, this will only handle | |
5663 | the cases in which the register is, in fact, set in a | |
5664 | later insn. */ | |
5665 | ||
5666 | /* ??? We shouldn't have to use FOUNDINSN here. | |
6fb5fa3c DB |
5667 | This dates back to when we used LOG_LINKS to find |
5668 | the most recent insn which sets the register. */ | |
933c3ba3 JW |
5669 | |
5670 | if (foundinsn | |
5671 | && (scanset | |
5672 | || find_reg_note (scan, REG_DEAD, reg))) | |
5673 | { | |
5674 | dies = scan; | |
5675 | break; | |
5676 | } | |
5677 | } | |
5678 | ||
5679 | if (! dies) | |
5680 | { | |
5681 | /* Either there was a branch, or some insn used REG | |
5682 | other than as a function call address. */ | |
5683 | continue; | |
5684 | } | |
5685 | ||
cf7c4aa6 HPN |
5686 | /* Create a code label, and put it in a REG_LABEL_OPERAND note |
5687 | on the insn which sets the register, and on each call insn | |
5688 | which uses the register. In final_prescan_insn we look for | |
5689 | the REG_LABEL_OPERAND notes, and output the appropriate label | |
933c3ba3 JW |
5690 | or pseudo-op. */ |
5691 | ||
5692 | label = gen_label_rtx (); | |
65c5f2a6 ILT |
5693 | add_reg_note (link, REG_LABEL_OPERAND, label); |
5694 | add_reg_note (insn, REG_LABEL_OPERAND, label); | |
933c3ba3 JW |
5695 | if (rescan) |
5696 | { | |
5697 | scan = link; | |
5698 | do | |
5699 | { | |
4787bce0 JW |
5700 | rtx reg2; |
5701 | ||
933c3ba3 JW |
5702 | scan = NEXT_INSN (scan); |
5703 | if (scan != insn | |
f3536097 | 5704 | && ((CALL_P (scan) |
4787bce0 JW |
5705 | && reg_mentioned_p (reg, scan)) |
5706 | || ((reg2 = sfunc_uses_reg (scan)) | |
5707 | && REGNO (reg2) == REGNO (reg)))) | |
65c5f2a6 | 5708 | add_reg_note (scan, REG_LABEL_OPERAND, label); |
933c3ba3 JW |
5709 | } |
5710 | while (scan != dies); | |
5711 | } | |
5712 | } | |
5713 | } | |
5714 | ||
33f7f353 JR |
5715 | if (TARGET_SH2) |
5716 | fixup_addr_diff_vecs (first); | |
1245df60 R |
5717 | |
5718 | if (optimize) | |
5719 | { | |
1245df60 R |
5720 | mdep_reorg_phase = SH_SHORTEN_BRANCHES0; |
5721 | shorten_branches (first); | |
5722 | } | |
6dd3c0a5 | 5723 | |
933c3ba3 JW |
5724 | /* Scan the function looking for move instructions which have to be |
5725 | changed to pc-relative loads and insert the literal tables. */ | |
6dd3c0a5 SB |
5726 | label_ref_list_pool = create_alloc_pool ("label references list", |
5727 | sizeof (struct label_ref_list_d), | |
5728 | 30); | |
1245df60 R |
5729 | mdep_reorg_phase = SH_FIXUP_PCLOAD; |
5730 | for (insn = first, num_mova = 0; insn; insn = NEXT_INSN (insn)) | |
0d7e008e | 5731 | { |
1245df60 R |
5732 | if (mova_p (insn)) |
5733 | { | |
078c8b08 R |
5734 | /* ??? basic block reordering can move a switch table dispatch |
5735 | below the switch table. Check if that has happened. | |
5736 | We only have the addresses available when optimizing; but then, | |
5737 | this check shouldn't be needed when not optimizing. */ | |
4241ecb0 | 5738 | if (!untangle_mova (&num_mova, &mova, insn)) |
078c8b08 | 5739 | { |
4241ecb0 R |
5740 | insn = mova; |
5741 | num_mova = 0; | |
078c8b08 | 5742 | } |
1245df60 | 5743 | } |
f3536097 | 5744 | else if (JUMP_P (insn) |
1245df60 | 5745 | && GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC |
4241ecb0 R |
5746 | && num_mova |
5747 | /* ??? loop invariant motion can also move a mova out of a | |
5748 | loop. Since loop does this code motion anyway, maybe we | |
5749 | should wrap UNSPEC_MOVA into a CONST, so that reload can | |
5750 | move it back. */ | |
5751 | && ((num_mova > 1 | |
5752 | && GET_MODE (prev_nonnote_insn (insn)) == VOIDmode) | |
5753 | || (prev_nonnote_insn (insn) | |
5754 | == XEXP (MOVA_LABELREF (mova), 0)))) | |
1245df60 R |
5755 | { |
5756 | rtx scan; | |
5757 | int total; | |
5758 | ||
5759 | num_mova--; | |
5760 | ||
5761 | /* Some code might have been inserted between the mova and | |
5762 | its ADDR_DIFF_VEC. Check if the mova is still in range. */ | |
5763 | for (scan = mova, total = 0; scan != insn; scan = NEXT_INSN (scan)) | |
33f7f353 | 5764 | total += get_attr_length (scan); |
1245df60 R |
5765 | |
5766 | /* range of mova is 1020, add 4 because pc counts from address of | |
5767 | second instruction after this one, subtract 2 in case pc is 2 | |
5768 | byte aligned. Possible alignment needed for the ADDR_DIFF_VEC | |
956d6950 | 5769 | cancels out with alignment effects of the mova itself. */ |
1245df60 R |
5770 | if (total > 1022) |
5771 | { | |
5772 | /* Change the mova into a load, and restart scanning | |
5773 | there. broken_move will then return true for mova. */ | |
078c8b08 | 5774 | fixup_mova (mova); |
1245df60 R |
5775 | insn = mova; |
5776 | } | |
5777 | } | |
078c8b08 | 5778 | if (broken_move (insn) |
f3536097 | 5779 | || (NONJUMP_INSN_P (insn) |
078c8b08 | 5780 | && recog_memoized (insn) == CODE_FOR_casesi_worker_2)) |
0d7e008e | 5781 | { |
0d7e008e | 5782 | rtx scan; |
8aa2a305 JW |
5783 | /* Scan ahead looking for a barrier to stick the constant table |
5784 | behind. */ | |
1245df60 | 5785 | rtx barrier = find_barrier (num_mova, mova, insn); |
3a8699c7 | 5786 | rtx last_float_move = NULL_RTX, last_float = 0, *last_float_addr = NULL; |
078c8b08 | 5787 | int need_aligned_label = 0; |
b9654711 | 5788 | |
1245df60 R |
5789 | if (num_mova && ! mova_p (mova)) |
5790 | { | |
5791 | /* find_barrier had to change the first mova into a | |
5792 | pcload; thus, we have to start with this new pcload. */ | |
5793 | insn = mova; | |
5794 | num_mova = 0; | |
5795 | } | |
16bea517 | 5796 | /* Now find all the moves between the points and modify them. */ |
0d7e008e SC |
5797 | for (scan = insn; scan != barrier; scan = NEXT_INSN (scan)) |
5798 | { | |
f3536097 | 5799 | if (LABEL_P (scan)) |
1245df60 | 5800 | last_float = 0; |
f3536097 | 5801 | if (NONJUMP_INSN_P (scan) |
078c8b08 R |
5802 | && recog_memoized (scan) == CODE_FOR_casesi_worker_2) |
5803 | need_aligned_label = 1; | |
0d7e008e SC |
5804 | if (broken_move (scan)) |
5805 | { | |
e577c183 JR |
5806 | rtx *patp = &PATTERN (scan), pat = *patp; |
5807 | rtx src, dst; | |
0d7e008e | 5808 | rtx lab; |
0d7e008e | 5809 | rtx newsrc; |
e577c183 JR |
5810 | enum machine_mode mode; |
5811 | ||
5812 | if (GET_CODE (pat) == PARALLEL) | |
5813 | patp = &XVECEXP (pat, 0, 0), pat = *patp; | |
5814 | src = SET_SRC (pat); | |
5815 | dst = SET_DEST (pat); | |
5816 | mode = GET_MODE (dst); | |
0d7e008e | 5817 | |
225e4f43 R |
5818 | if (mode == SImode && hi_const (src) |
5819 | && REGNO (dst) != FPUL_REG) | |
0d7e008e | 5820 | { |
23ed92ca | 5821 | int offset = 0; |
8aa2a305 | 5822 | |
0d7e008e | 5823 | mode = HImode; |
d3ae8277 | 5824 | while (GET_CODE (dst) == SUBREG) |
23ed92ca | 5825 | { |
ddef6bc7 JJ |
5826 | offset += subreg_regno_offset (REGNO (SUBREG_REG (dst)), |
5827 | GET_MODE (SUBREG_REG (dst)), | |
5828 | SUBREG_BYTE (dst), | |
5829 | GET_MODE (dst)); | |
23ed92ca JW |
5830 | dst = SUBREG_REG (dst); |
5831 | } | |
c5c76735 | 5832 | dst = gen_rtx_REG (HImode, REGNO (dst) + offset); |
0d7e008e | 5833 | } |
f3536097 | 5834 | if (REG_P (dst) && FP_ANY_REGISTER_P (REGNO (dst))) |
1245df60 | 5835 | { |
b8794099 | 5836 | /* This must be an insn that clobbers r0. */ |
0f2b41b2 AO |
5837 | rtx *clobberp = &XVECEXP (PATTERN (scan), 0, |
5838 | XVECLEN (PATTERN (scan), 0) | |
5839 | - 1); | |
5840 | rtx clobber = *clobberp; | |
b8794099 | 5841 | |
f5b9e7c9 NS |
5842 | gcc_assert (GET_CODE (clobber) == CLOBBER |
5843 | && rtx_equal_p (XEXP (clobber, 0), r0_rtx)); | |
b8794099 | 5844 | |
1245df60 R |
5845 | if (last_float |
5846 | && reg_set_between_p (r0_rtx, last_float_move, scan)) | |
5847 | last_float = 0; | |
3503150c RS |
5848 | if (last_float |
5849 | && TARGET_SHCOMPACT | |
5850 | && GET_MODE_SIZE (mode) != 4 | |
5851 | && GET_MODE_SIZE (GET_MODE (last_float)) == 4) | |
5852 | last_float = 0; | |
1245df60 R |
5853 | lab = add_constant (src, mode, last_float); |
5854 | if (lab) | |
5855 | emit_insn_before (gen_mova (lab), scan); | |
5856 | else | |
b8794099 R |
5857 | { |
5858 | /* There will be a REG_UNUSED note for r0 on | |
5859 | LAST_FLOAT_MOVE; we have to change it to REG_INC, | |
5860 | lest reorg:mark_target_live_regs will not | |
5861 | consider r0 to be used, and we end up with delay | |
475f26dd | 5862 | slot insn in front of SCAN that clobbers r0. */ |
b8794099 R |
5863 | rtx note |
5864 | = find_regno_note (last_float_move, REG_UNUSED, 0); | |
5865 | ||
ab1672a3 JW |
5866 | /* If we are not optimizing, then there may not be |
5867 | a note. */ | |
5868 | if (note) | |
81f40b79 | 5869 | PUT_REG_NOTE_KIND (note, REG_INC); |
b8794099 R |
5870 | |
5871 | *last_float_addr = r0_inc_rtx; | |
5872 | } | |
1245df60 R |
5873 | last_float_move = scan; |
5874 | last_float = src; | |
57d38024 | 5875 | newsrc = gen_const_mem (mode, |
318881c0 | 5876 | (((TARGET_SH4 && ! TARGET_FMOVD) |
225e4f43 | 5877 | || REGNO (dst) == FPUL_REG) |
1245df60 R |
5878 | ? r0_inc_rtx |
5879 | : r0_rtx)); | |
5880 | last_float_addr = &XEXP (newsrc, 0); | |
b8794099 R |
5881 | |
5882 | /* Remove the clobber of r0. */ | |
0f2b41b2 AO |
5883 | *clobberp = gen_rtx_CLOBBER (GET_MODE (clobber), |
5884 | gen_rtx_SCRATCH (Pmode)); | |
1245df60 | 5885 | } |
1a66cd67 | 5886 | /* This is a mova needing a label. Create it. */ |
43c05634 AO |
5887 | else if (GET_CODE (src) == UNSPEC |
5888 | && XINT (src, 1) == UNSPEC_MOVA | |
5889 | && GET_CODE (XVECEXP (src, 0, 0)) == CONST) | |
1a66cd67 | 5890 | { |
43c05634 | 5891 | lab = add_constant (XVECEXP (src, 0, 0), mode, 0); |
1a66cd67 | 5892 | newsrc = gen_rtx_LABEL_REF (VOIDmode, lab); |
e69d1422 | 5893 | newsrc = gen_rtx_UNSPEC (SImode, |
4773afa4 AO |
5894 | gen_rtvec (1, newsrc), |
5895 | UNSPEC_MOVA); | |
1a66cd67 | 5896 | } |
1245df60 R |
5897 | else |
5898 | { | |
5899 | lab = add_constant (src, mode, 0); | |
542a8afa RH |
5900 | newsrc = gen_rtx_LABEL_REF (VOIDmode, lab); |
5901 | newsrc = gen_const_mem (mode, newsrc); | |
1245df60 | 5902 | } |
c5c76735 | 5903 | *patp = gen_rtx_SET (VOIDmode, dst, newsrc); |
e577c183 | 5904 | INSN_CODE (scan) = -1; |
0d7e008e SC |
5905 | } |
5906 | } | |
078c8b08 | 5907 | dump_table (need_aligned_label ? insn : 0, barrier); |
1245df60 | 5908 | insn = barrier; |
0d7e008e SC |
5909 | } |
5910 | } | |
6dd3c0a5 | 5911 | free_alloc_pool (label_ref_list_pool); |
4241ecb0 R |
5912 | for (insn = first; insn; insn = NEXT_INSN (insn)) |
5913 | PUT_MODE (insn, VOIDmode); | |
5914 | ||
1245df60 | 5915 | mdep_reorg_phase = SH_SHORTEN_BRANCHES1; |
9d98a694 | 5916 | INSN_ADDRESSES_FREE (); |
1245df60 R |
5917 | split_branches (first); |
5918 | ||
5919 | /* The INSN_REFERENCES_ARE_DELAYED in sh.h is problematic because it | |
a920aefe | 5920 | also has an effect on the register that holds the address of the sfunc. |
1245df60 R |
5921 | Insert an extra dummy insn in front of each sfunc that pretends to |
5922 | use this register. */ | |
5923 | if (flag_delayed_branch) | |
5924 | { | |
5925 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
5926 | { | |
5927 | rtx reg = sfunc_uses_reg (insn); | |
5928 | ||
5929 | if (! reg) | |
5930 | continue; | |
5931 | emit_insn_before (gen_use_sfunc_addr (reg), insn); | |
5932 | } | |
5933 | } | |
225e4f43 R |
5934 | #if 0 |
5935 | /* fpscr is not actually a user variable, but we pretend it is for the | |
5936 | sake of the previous optimization passes, since we want it handled like | |
9f09b1f2 | 5937 | one. However, we don't have any debugging information for it, so turn |
225e4f43 R |
5938 | it into a non-user variable now. */ |
5939 | if (TARGET_SH4) | |
5940 | REG_USERVAR_P (get_fpscr_rtx ()) = 0; | |
5941 | #endif | |
1245df60 R |
5942 | mdep_reorg_phase = SH_AFTER_MDEP_REORG; |
5943 | } | |
5944 | ||
5945 | int | |
cf277499 | 5946 | get_dest_uid (rtx label, int max_uid) |
1245df60 R |
5947 | { |
5948 | rtx dest = next_real_insn (label); | |
5949 | int dest_uid; | |
5950 | if (! dest) | |
5951 | /* This can happen for an undefined label. */ | |
5952 | return 0; | |
5953 | dest_uid = INSN_UID (dest); | |
5954 | /* If this is a newly created branch redirection blocking instruction, | |
5955 | we cannot index the branch_uid or insn_addresses arrays with its | |
5956 | uid. But then, we won't need to, because the actual destination is | |
5957 | the following branch. */ | |
5958 | while (dest_uid >= max_uid) | |
5959 | { | |
5960 | dest = NEXT_INSN (dest); | |
5961 | dest_uid = INSN_UID (dest); | |
5962 | } | |
f3536097 | 5963 | if (JUMP_P (dest) && GET_CODE (PATTERN (dest)) == RETURN) |
1245df60 R |
5964 | return 0; |
5965 | return dest_uid; | |
5966 | } | |
5967 | ||
5968 | /* Split condbranches that are out of range. Also add clobbers for | |
5969 | scratch registers that are needed in far jumps. | |
5970 | We do this before delay slot scheduling, so that it can take our | |
5971 | newly created instructions into account. It also allows us to | |
5972 | find branches with common targets more easily. */ | |
5973 | ||
5974 | static void | |
cf277499 | 5975 | split_branches (rtx first) |
1245df60 R |
5976 | { |
5977 | rtx insn; | |
5978 | struct far_branch **uid_branch, *far_branch_list = 0; | |
5979 | int max_uid = get_max_uid (); | |
f5b9e7c9 | 5980 | int ok; |
1245df60 R |
5981 | |
5982 | /* Find out which branches are out of range. */ | |
1245df60 R |
5983 | shorten_branches (first); |
5984 | ||
5985 | uid_branch = (struct far_branch **) alloca (max_uid * sizeof *uid_branch); | |
2e09e75a | 5986 | memset ((char *) uid_branch, 0, max_uid * sizeof *uid_branch); |
1245df60 R |
5987 | |
5988 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
2c3c49de | 5989 | if (! INSN_P (insn)) |
1245df60 R |
5990 | continue; |
5991 | else if (INSN_DELETED_P (insn)) | |
5992 | { | |
5993 | /* Shorten_branches would split this instruction again, | |
5994 | so transform it into a note. */ | |
a38e7aa5 | 5995 | SET_INSN_DELETED (insn); |
1245df60 | 5996 | } |
f3536097 | 5997 | else if (JUMP_P (insn) |
1245df60 R |
5998 | /* Don't mess with ADDR_DIFF_VEC */ |
5999 | && (GET_CODE (PATTERN (insn)) == SET | |
6000 | || GET_CODE (PATTERN (insn)) == RETURN)) | |
6001 | { | |
6002 | enum attr_type type = get_attr_type (insn); | |
6003 | if (type == TYPE_CBRANCH) | |
6004 | { | |
6005 | rtx next, beyond; | |
f676971a | 6006 | |
1245df60 R |
6007 | if (get_attr_length (insn) > 4) |
6008 | { | |
6009 | rtx src = SET_SRC (PATTERN (insn)); | |
1245df60 | 6010 | rtx olabel = XEXP (XEXP (src, 1), 0); |
9d98a694 | 6011 | int addr = INSN_ADDRESSES (INSN_UID (insn)); |
1245df60 R |
6012 | rtx label = 0; |
6013 | int dest_uid = get_dest_uid (olabel, max_uid); | |
6014 | struct far_branch *bp = uid_branch[dest_uid]; | |
f676971a | 6015 | |
1245df60 | 6016 | /* redirect_jump needs a valid JUMP_LABEL, and it might delete |
25938114 | 6017 | the label if the LABEL_NUSES count drops to zero. There is |
1245df60 R |
6018 | always a jump_optimize pass that sets these values, but it |
6019 | proceeds to delete unreferenced code, and then if not | |
956d6950 | 6020 | optimizing, to un-delete the deleted instructions, thus |
1245df60 R |
6021 | leaving labels with too low uses counts. */ |
6022 | if (! optimize) | |
6023 | { | |
6024 | JUMP_LABEL (insn) = olabel; | |
6025 | LABEL_NUSES (olabel)++; | |
6026 | } | |
6027 | if (! bp) | |
6028 | { | |
6029 | bp = (struct far_branch *) alloca (sizeof *bp); | |
6030 | uid_branch[dest_uid] = bp; | |
6031 | bp->prev = far_branch_list; | |
6032 | far_branch_list = bp; | |
6033 | bp->far_label | |
6034 | = XEXP (XEXP (SET_SRC (PATTERN (insn)), 1), 0); | |
6035 | LABEL_NUSES (bp->far_label)++; | |
6036 | } | |
6037 | else | |
6038 | { | |
6039 | label = bp->near_label; | |
6040 | if (! label && bp->address - addr >= CONDJUMP_MIN) | |
6041 | { | |
6042 | rtx block = bp->insert_place; | |
6043 | ||
6044 | if (GET_CODE (PATTERN (block)) == RETURN) | |
6045 | block = PREV_INSN (block); | |
6046 | else | |
6047 | block = gen_block_redirect (block, | |
6048 | bp->address, 2); | |
6049 | label = emit_label_after (gen_label_rtx (), | |
6050 | PREV_INSN (block)); | |
6051 | bp->near_label = label; | |
6052 | } | |
6053 | else if (label && ! NEXT_INSN (label)) | |
eda44c78 R |
6054 | { |
6055 | if (addr + 2 - bp->address <= CONDJUMP_MAX) | |
6056 | bp->insert_place = insn; | |
6057 | else | |
6058 | gen_far_branch (bp); | |
6059 | } | |
1245df60 R |
6060 | } |
6061 | if (! label | |
318881c0 | 6062 | || (NEXT_INSN (label) && bp->address - addr < CONDJUMP_MIN)) |
1245df60 R |
6063 | { |
6064 | bp->near_label = label = gen_label_rtx (); | |
6065 | bp->insert_place = insn; | |
6066 | bp->address = addr; | |
6067 | } | |
2acdc1fb | 6068 | ok = redirect_jump (insn, label, 0); |
f5b9e7c9 | 6069 | gcc_assert (ok); |
1245df60 R |
6070 | } |
6071 | else | |
6072 | { | |
6073 | /* get_attr_length (insn) == 2 */ | |
6074 | /* Check if we have a pattern where reorg wants to redirect | |
6075 | the branch to a label from an unconditional branch that | |
6076 | is too far away. */ | |
6077 | /* We can't use JUMP_LABEL here because it might be undefined | |
6078 | when not optimizing. */ | |
33f7f353 | 6079 | /* A syntax error might cause beyond to be NULL_RTX. */ |
1245df60 R |
6080 | beyond |
6081 | = next_active_insn (XEXP (XEXP (SET_SRC (PATTERN (insn)), 1), | |
6082 | 0)); | |
f676971a | 6083 | |
33f7f353 | 6084 | if (beyond |
f3536097 | 6085 | && (JUMP_P (beyond) |
b94544e4 | 6086 | || ((beyond = next_active_insn (beyond)) |
f3536097 | 6087 | && JUMP_P (beyond))) |
1245df60 | 6088 | && GET_CODE (PATTERN (beyond)) == SET |
ddf8ec7e | 6089 | && recog_memoized (beyond) == CODE_FOR_jump_compact |
9d98a694 AO |
6090 | && ((INSN_ADDRESSES |
6091 | (INSN_UID (XEXP (SET_SRC (PATTERN (beyond)), 0))) | |
11f9ed1a | 6092 | - INSN_ADDRESSES (INSN_UID (insn)) + (unsigned) 252) |
1245df60 R |
6093 | > 252 + 258 + 2)) |
6094 | gen_block_redirect (beyond, | |
9d98a694 | 6095 | INSN_ADDRESSES (INSN_UID (beyond)), 1); |
1245df60 | 6096 | } |
f676971a | 6097 | |
1245df60 R |
6098 | next = next_active_insn (insn); |
6099 | ||
a7d6f306 KK |
6100 | if (next |
6101 | && (JUMP_P (next) | |
6102 | || ((next = next_active_insn (next)) | |
6103 | && JUMP_P (next))) | |
1245df60 | 6104 | && GET_CODE (PATTERN (next)) == SET |
ddf8ec7e | 6105 | && recog_memoized (next) == CODE_FOR_jump_compact |
9d98a694 AO |
6106 | && ((INSN_ADDRESSES |
6107 | (INSN_UID (XEXP (SET_SRC (PATTERN (next)), 0))) | |
11f9ed1a | 6108 | - INSN_ADDRESSES (INSN_UID (insn)) + (unsigned) 252) |
1245df60 | 6109 | > 252 + 258 + 2)) |
9d98a694 | 6110 | gen_block_redirect (next, INSN_ADDRESSES (INSN_UID (next)), 1); |
1245df60 R |
6111 | } |
6112 | else if (type == TYPE_JUMP || type == TYPE_RETURN) | |
6113 | { | |
9d98a694 | 6114 | int addr = INSN_ADDRESSES (INSN_UID (insn)); |
1245df60 R |
6115 | rtx far_label = 0; |
6116 | int dest_uid = 0; | |
6117 | struct far_branch *bp; | |
6118 | ||
6119 | if (type == TYPE_JUMP) | |
6120 | { | |
6121 | far_label = XEXP (SET_SRC (PATTERN (insn)), 0); | |
6122 | dest_uid = get_dest_uid (far_label, max_uid); | |
6123 | if (! dest_uid) | |
6124 | { | |
6125 | /* Parse errors can lead to labels outside | |
6126 | the insn stream. */ | |
6127 | if (! NEXT_INSN (far_label)) | |
6128 | continue; | |
6129 | ||
6130 | if (! optimize) | |
6131 | { | |
6132 | JUMP_LABEL (insn) = far_label; | |
6133 | LABEL_NUSES (far_label)++; | |
6134 | } | |
8ceaac3c | 6135 | redirect_jump (insn, NULL_RTX, 1); |
1245df60 R |
6136 | far_label = 0; |
6137 | } | |
6138 | } | |
6139 | bp = uid_branch[dest_uid]; | |
6140 | if (! bp) | |
6141 | { | |
6142 | bp = (struct far_branch *) alloca (sizeof *bp); | |
6143 | uid_branch[dest_uid] = bp; | |
6144 | bp->prev = far_branch_list; | |
6145 | far_branch_list = bp; | |
6146 | bp->near_label = 0; | |
6147 | bp->far_label = far_label; | |
6148 | if (far_label) | |
6149 | LABEL_NUSES (far_label)++; | |
6150 | } | |
6151 | else if (bp->near_label && ! NEXT_INSN (bp->near_label)) | |
6152 | if (addr - bp->address <= CONDJUMP_MAX) | |
6153 | emit_label_after (bp->near_label, PREV_INSN (insn)); | |
6154 | else | |
6155 | { | |
6156 | gen_far_branch (bp); | |
6157 | bp->near_label = 0; | |
6158 | } | |
6159 | else | |
6160 | bp->near_label = 0; | |
6161 | bp->address = addr; | |
6162 | bp->insert_place = insn; | |
6163 | if (! far_label) | |
6164 | emit_insn_before (gen_block_branch_redirect (const0_rtx), insn); | |
6165 | else | |
6166 | gen_block_redirect (insn, addr, bp->near_label ? 2 : 0); | |
6167 | } | |
6168 | } | |
6169 | /* Generate all pending far branches, | |
6170 | and free our references to the far labels. */ | |
6171 | while (far_branch_list) | |
6172 | { | |
6173 | if (far_branch_list->near_label | |
6174 | && ! NEXT_INSN (far_branch_list->near_label)) | |
6175 | gen_far_branch (far_branch_list); | |
6176 | if (optimize | |
6177 | && far_branch_list->far_label | |
6178 | && ! --LABEL_NUSES (far_branch_list->far_label)) | |
6179 | delete_insn (far_branch_list->far_label); | |
6180 | far_branch_list = far_branch_list->prev; | |
6181 | } | |
a0798779 R |
6182 | |
6183 | /* Instruction length information is no longer valid due to the new | |
6184 | instructions that have been generated. */ | |
6185 | init_insn_lengths (); | |
b9654711 SC |
6186 | } |
6187 | ||
8aa2a305 | 6188 | /* Dump out instruction addresses, which is useful for debugging the |
933c3ba3 JW |
6189 | constant pool table stuff. |
6190 | ||
6191 | If relaxing, output the label and pseudo-ops used to link together | |
6192 | calls and the instruction which set the registers. */ | |
8aa2a305 | 6193 | |
8aa2a305 JW |
6194 | /* ??? The addresses printed by this routine for insns are nonsense for |
6195 | insns which are inside of a sequence where none of the inner insns have | |
6196 | variable length. This is because the second pass of shorten_branches | |
6197 | does not bother to update them. */ | |
0d7e008e | 6198 | |
8e87e161 | 6199 | void |
cf277499 SB |
6200 | final_prescan_insn (rtx insn, rtx *opvec ATTRIBUTE_UNUSED, |
6201 | int noperands ATTRIBUTE_UNUSED) | |
b9654711 | 6202 | { |
8aa2a305 | 6203 | if (TARGET_DUMPISIZE) |
9d98a694 | 6204 | fprintf (asm_out_file, "\n! at %04x\n", INSN_ADDRESSES (INSN_UID (insn))); |
933c3ba3 JW |
6205 | |
6206 | if (TARGET_RELAX) | |
6207 | { | |
6208 | rtx note; | |
6209 | ||
cf7c4aa6 | 6210 | note = find_reg_note (insn, REG_LABEL_OPERAND, NULL_RTX); |
933c3ba3 JW |
6211 | if (note) |
6212 | { | |
6213 | rtx pattern; | |
6214 | ||
6215 | pattern = PATTERN (insn); | |
4b7d2f07 R |
6216 | if (GET_CODE (pattern) == PARALLEL) |
6217 | pattern = XVECEXP (pattern, 0, 0); | |
f5b9e7c9 NS |
6218 | switch (GET_CODE (pattern)) |
6219 | { | |
f5b9e7c9 NS |
6220 | case SET: |
6221 | if (GET_CODE (SET_SRC (pattern)) != CALL | |
6222 | && get_attr_type (insn) != TYPE_SFUNC) | |
6223 | { | |
6224 | targetm.asm_out.internal_label | |
6225 | (asm_out_file, "L", CODE_LABEL_NUMBER (XEXP (note, 0))); | |
6226 | break; | |
6227 | } | |
6228 | /* else FALLTHROUGH */ | |
6229 | case CALL: | |
6230 | asm_fprintf (asm_out_file, "\t.uses %LL%d\n", | |
6231 | CODE_LABEL_NUMBER (XEXP (note, 0))); | |
6232 | break; | |
6233 | ||
6234 | default: | |
6235 | gcc_unreachable (); | |
6236 | } | |
933c3ba3 JW |
6237 | } |
6238 | } | |
0d7e008e | 6239 | } |
b9654711 | 6240 | |
8aa2a305 | 6241 | /* Dump out any constants accumulated in the final pass. These will |
38e01259 | 6242 | only be labels. */ |
b9654711 | 6243 | |
318881c0 | 6244 | const char * |
cf277499 | 6245 | output_jump_label_table (void) |
0d7e008e | 6246 | { |
8aa2a305 JW |
6247 | int i; |
6248 | ||
6249 | if (pool_size) | |
6250 | { | |
6251 | fprintf (asm_out_file, "\t.align 2\n"); | |
6252 | for (i = 0; i < pool_size; i++) | |
6253 | { | |
6254 | pool_node *p = &pool_vector[i]; | |
6255 | ||
4977bab6 | 6256 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
8aa2a305 JW |
6257 | CODE_LABEL_NUMBER (p->label)); |
6258 | output_asm_insn (".long %O0", &p->value); | |
6259 | } | |
6260 | pool_size = 0; | |
6261 | } | |
b9654711 | 6262 | |
8aa2a305 JW |
6263 | return ""; |
6264 | } | |
6265 | \f | |
6266 | /* A full frame looks like: | |
16bea517 JW |
6267 | |
6268 | arg-5 | |
6269 | arg-4 | |
6270 | [ if current_function_anonymous_args | |
6271 | arg-3 | |
6272 | arg-2 | |
6273 | arg-1 | |
6274 | arg-0 ] | |
6275 | saved-fp | |
6276 | saved-r10 | |
6277 | saved-r11 | |
6278 | saved-r12 | |
6279 | saved-pr | |
6280 | local-n | |
6281 | .. | |
6282 | local-1 | |
6283 | local-0 <- fp points here. */ | |
8e87e161 | 6284 | |
8aa2a305 JW |
6285 | /* Number of bytes pushed for anonymous args, used to pass information |
6286 | between expand_prologue and expand_epilogue. */ | |
6287 | ||
fada1961 R |
6288 | /* Adjust the stack by SIZE bytes. REG holds the rtl of the register to be |
6289 | adjusted. If epilogue_p is zero, this is for a prologue; otherwise, it's | |
726d4cb7 KK |
6290 | for an epilogue and a negative value means that it's for a sibcall |
6291 | epilogue. If LIVE_REGS_MASK is nonzero, it points to a HARD_REG_SET of | |
6292 | all the registers that are about to be restored, and hence dead. */ | |
8aa2a305 JW |
6293 | |
6294 | static void | |
fada1961 | 6295 | output_stack_adjust (int size, rtx reg, int epilogue_p, |
21d05694 | 6296 | HARD_REG_SET *live_regs_mask, bool frame_p) |
8aa2a305 | 6297 | { |
21d05694 | 6298 | rtx (*emit_fn) (rtx) = frame_p ? &frame_insn : &emit_insn; |
8aa2a305 JW |
6299 | if (size) |
6300 | { | |
fa5322fa AO |
6301 | HOST_WIDE_INT align = STACK_BOUNDARY / BITS_PER_UNIT; |
6302 | ||
157371cf AO |
6303 | /* This test is bogus, as output_stack_adjust is used to re-align the |
6304 | stack. */ | |
6305 | #if 0 | |
f5b9e7c9 | 6306 | gcc_assert (!(size % align)); |
157371cf | 6307 | #endif |
fa5322fa AO |
6308 | |
6309 | if (CONST_OK_FOR_ADD (size)) | |
2754d3c5 | 6310 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size))); |
f3cd5375 JW |
6311 | /* Try to do it with two partial adjustments; however, we must make |
6312 | sure that the stack is properly aligned at all times, in case | |
6f317ef3 | 6313 | an interrupt occurs between the two partial adjustments. */ |
fa5322fa AO |
6314 | else if (CONST_OK_FOR_ADD (size / 2 & -align) |
6315 | && CONST_OK_FOR_ADD (size - (size / 2 & -align))) | |
f3cd5375 | 6316 | { |
2754d3c5 R |
6317 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size / 2 & -align))); |
6318 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size - (size / 2 & -align)))); | |
f3cd5375 JW |
6319 | } |
6320 | else | |
8aa2a305 | 6321 | { |
f3cd5375 | 6322 | rtx const_reg; |
2754d3c5 | 6323 | rtx insn; |
fada1961 R |
6324 | int temp = epilogue_p ? 7 : (TARGET_SH5 ? 0 : 1); |
6325 | int i; | |
f3cd5375 JW |
6326 | |
6327 | /* If TEMP is invalid, we could temporarily save a general | |
6328 | register to MACL. However, there is currently no need | |
f5b9e7c9 | 6329 | to handle this case, so just die when we see it. */ |
726d4cb7 KK |
6330 | if (epilogue_p < 0 |
6331 | || current_function_interrupt | |
62164eb4 | 6332 | || ! call_really_used_regs[temp] || fixed_regs[temp]) |
fada1961 | 6333 | temp = -1; |
726d4cb7 KK |
6334 | if (temp < 0 && ! current_function_interrupt |
6335 | && (TARGET_SHMEDIA || epilogue_p >= 0)) | |
fada1961 R |
6336 | { |
6337 | HARD_REG_SET temps; | |
6338 | COPY_HARD_REG_SET (temps, call_used_reg_set); | |
6339 | AND_COMPL_HARD_REG_SET (temps, call_fixed_reg_set); | |
726d4cb7 | 6340 | if (epilogue_p > 0) |
fada1961 | 6341 | { |
726d4cb7 | 6342 | int nreg = 0; |
38173d38 | 6343 | if (crtl->return_rtx) |
726d4cb7 KK |
6344 | { |
6345 | enum machine_mode mode; | |
38173d38 | 6346 | mode = GET_MODE (crtl->return_rtx); |
726d4cb7 KK |
6347 | if (BASE_RETURN_VALUE_REG (mode) == FIRST_RET_REG) |
6348 | nreg = HARD_REGNO_NREGS (FIRST_RET_REG, mode); | |
6349 | } | |
6350 | for (i = 0; i < nreg; i++) | |
fada1961 | 6351 | CLEAR_HARD_REG_BIT (temps, FIRST_RET_REG + i); |
e3b5732b | 6352 | if (crtl->calls_eh_return) |
fada1961 R |
6353 | { |
6354 | CLEAR_HARD_REG_BIT (temps, EH_RETURN_STACKADJ_REGNO); | |
6355 | for (i = 0; i <= 3; i++) | |
6356 | CLEAR_HARD_REG_BIT (temps, EH_RETURN_DATA_REGNO (i)); | |
6357 | } | |
6358 | } | |
726d4cb7 KK |
6359 | if (TARGET_SHMEDIA && epilogue_p < 0) |
6360 | for (i = FIRST_TARGET_REG; i <= LAST_TARGET_REG; i++) | |
6361 | CLEAR_HARD_REG_BIT (temps, i); | |
6362 | if (epilogue_p <= 0) | |
fada1961 R |
6363 | { |
6364 | for (i = FIRST_PARM_REG; | |
6365 | i < FIRST_PARM_REG + NPARM_REGS (SImode); i++) | |
6366 | CLEAR_HARD_REG_BIT (temps, i); | |
6de9cd9a | 6367 | if (cfun->static_chain_decl != NULL) |
fada1961 R |
6368 | CLEAR_HARD_REG_BIT (temps, STATIC_CHAIN_REGNUM); |
6369 | } | |
6370 | temp = scavenge_reg (&temps); | |
6371 | } | |
6372 | if (temp < 0 && live_regs_mask) | |
eb996a4a KK |
6373 | { |
6374 | HARD_REG_SET temps; | |
6375 | ||
6376 | COPY_HARD_REG_SET (temps, *live_regs_mask); | |
6377 | CLEAR_HARD_REG_BIT (temps, REGNO (reg)); | |
6378 | temp = scavenge_reg (&temps); | |
6379 | } | |
f3cd5375 | 6380 | if (temp < 0) |
726d4cb7 | 6381 | { |
f5b9e7c9 NS |
6382 | rtx adj_reg, tmp_reg, mem; |
6383 | ||
726d4cb7 KK |
6384 | /* If we reached here, the most likely case is the (sibcall) |
6385 | epilogue for non SHmedia. Put a special push/pop sequence | |
6386 | for such case as the last resort. This looks lengthy but | |
f5b9e7c9 NS |
6387 | would not be problem because it seems to be very |
6388 | rare. */ | |
6389 | ||
6390 | gcc_assert (!TARGET_SHMEDIA && epilogue_p); | |
6391 | ||
6392 | ||
6393 | /* ??? There is still the slight possibility that r4 or | |
6394 | r5 have been reserved as fixed registers or assigned | |
6395 | as global registers, and they change during an | |
6396 | interrupt. There are possible ways to handle this: | |
6397 | ||
6398 | - If we are adjusting the frame pointer (r14), we can do | |
6399 | with a single temp register and an ordinary push / pop | |
6400 | on the stack. | |
6401 | - Grab any call-used or call-saved registers (i.e. not | |
6402 | fixed or globals) for the temps we need. We might | |
6403 | also grab r14 if we are adjusting the stack pointer. | |
6404 | If we can't find enough available registers, issue | |
6405 | a diagnostic and die - the user must have reserved | |
6406 | way too many registers. | |
6407 | But since all this is rather unlikely to happen and | |
6408 | would require extra testing, we just die if r4 / r5 | |
6409 | are not available. */ | |
6410 | gcc_assert (!fixed_regs[4] && !fixed_regs[5] | |
6411 | && !global_regs[4] && !global_regs[5]); | |
6412 | ||
6413 | adj_reg = gen_rtx_REG (GET_MODE (reg), 4); | |
6414 | tmp_reg = gen_rtx_REG (GET_MODE (reg), 5); | |
57d38024 | 6415 | emit_move_insn (gen_tmp_stack_mem (Pmode, reg), adj_reg); |
f5b9e7c9 NS |
6416 | emit_insn (GEN_MOV (adj_reg, GEN_INT (size))); |
6417 | emit_insn (GEN_ADD3 (adj_reg, adj_reg, reg)); | |
57d38024 | 6418 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_PRE_DEC (Pmode, adj_reg)); |
f5b9e7c9 | 6419 | emit_move_insn (mem, tmp_reg); |
57d38024 R |
6420 | emit_move_insn (tmp_reg, gen_tmp_stack_mem (Pmode, reg)); |
6421 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_PRE_DEC (Pmode, adj_reg)); | |
6422 | emit_move_insn (mem, tmp_reg); | |
6423 | emit_move_insn (reg, adj_reg); | |
6424 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_POST_INC (Pmode, reg)); | |
6425 | emit_move_insn (adj_reg, mem); | |
6426 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_POST_INC (Pmode, reg)); | |
6427 | emit_move_insn (tmp_reg, mem); | |
eb996a4a | 6428 | /* Tell flow the insns that pop r4/r5 aren't dead. */ |
c41c1387 RS |
6429 | emit_use (tmp_reg); |
6430 | emit_use (adj_reg); | |
57d38024 | 6431 | return; |
726d4cb7 | 6432 | } |
fa5322fa | 6433 | const_reg = gen_rtx_REG (GET_MODE (reg), temp); |
f3cd5375 JW |
6434 | |
6435 | /* If SIZE is negative, subtract the positive value. | |
6436 | This sometimes allows a constant pool entry to be shared | |
6437 | between prologue and epilogue code. */ | |
6438 | if (size < 0) | |
885c9a39 | 6439 | { |
fa5322fa | 6440 | emit_insn (GEN_MOV (const_reg, GEN_INT (-size))); |
2754d3c5 | 6441 | insn = emit_fn (GEN_SUB3 (reg, reg, const_reg)); |
885c9a39 JR |
6442 | } |
6443 | else | |
6444 | { | |
fa5322fa | 6445 | emit_insn (GEN_MOV (const_reg, GEN_INT (size))); |
2754d3c5 | 6446 | insn = emit_fn (GEN_ADD3 (reg, reg, const_reg)); |
885c9a39 | 6447 | } |
fada1961 | 6448 | if (! epilogue_p) |
b757d36f KK |
6449 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, |
6450 | gen_rtx_SET (VOIDmode, reg, | |
6451 | gen_rtx_PLUS (SImode, reg, | |
6452 | GEN_INT (size)))); | |
8aa2a305 | 6453 | } |
8aa2a305 JW |
6454 | } |
6455 | } | |
6456 | ||
2754d3c5 | 6457 | static rtx |
cf277499 | 6458 | frame_insn (rtx x) |
2754d3c5 R |
6459 | { |
6460 | x = emit_insn (x); | |
6461 | RTX_FRAME_RELATED_P (x) = 1; | |
6462 | return x; | |
6463 | } | |
6464 | ||
8aa2a305 JW |
6465 | /* Output RTL to push register RN onto the stack. */ |
6466 | ||
2754d3c5 | 6467 | static rtx |
cf277499 | 6468 | push (int rn) |
8aa2a305 JW |
6469 | { |
6470 | rtx x; | |
225e4f43 R |
6471 | if (rn == FPUL_REG) |
6472 | x = gen_push_fpul (); | |
7144b2d8 DD |
6473 | else if (rn == FPSCR_REG) |
6474 | x = gen_push_fpscr (); | |
157371cf | 6475 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && ! TARGET_FPU_SINGLE |
104ee20b | 6476 | && FP_OR_XD_REGISTER_P (rn)) |
225e4f43 | 6477 | { |
104ee20b | 6478 | if (FP_REGISTER_P (rn) && (rn - FIRST_FP_REG) & 1) |
d042370e | 6479 | return NULL_RTX; |
c5c76735 | 6480 | x = gen_push_4 (gen_rtx_REG (DFmode, rn)); |
225e4f43 | 6481 | } |
3a8699c7 | 6482 | else if (TARGET_SH2E && FP_REGISTER_P (rn)) |
c5c76735 | 6483 | x = gen_push_e (gen_rtx_REG (SFmode, rn)); |
1a95a963 | 6484 | else |
c5c76735 | 6485 | x = gen_push (gen_rtx_REG (SImode, rn)); |
1a95a963 | 6486 | |
2754d3c5 | 6487 | x = frame_insn (x); |
b757d36f | 6488 | add_reg_note (x, REG_INC, gen_rtx_REG (SImode, STACK_POINTER_REGNUM)); |
2754d3c5 | 6489 | return x; |
8aa2a305 JW |
6490 | } |
6491 | ||
6492 | /* Output RTL to pop register RN from the stack. */ | |
6493 | ||
6494 | static void | |
cf277499 | 6495 | pop (int rn) |
8aa2a305 JW |
6496 | { |
6497 | rtx x; | |
225e4f43 R |
6498 | if (rn == FPUL_REG) |
6499 | x = gen_pop_fpul (); | |
7144b2d8 DD |
6500 | else if (rn == FPSCR_REG) |
6501 | x = gen_pop_fpscr (); | |
157371cf | 6502 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && ! TARGET_FPU_SINGLE |
104ee20b | 6503 | && FP_OR_XD_REGISTER_P (rn)) |
225e4f43 | 6504 | { |
104ee20b | 6505 | if (FP_REGISTER_P (rn) && (rn - FIRST_FP_REG) & 1) |
225e4f43 | 6506 | return; |
c5c76735 | 6507 | x = gen_pop_4 (gen_rtx_REG (DFmode, rn)); |
225e4f43 | 6508 | } |
3a8699c7 | 6509 | else if (TARGET_SH2E && FP_REGISTER_P (rn)) |
c5c76735 | 6510 | x = gen_pop_e (gen_rtx_REG (SFmode, rn)); |
1a95a963 | 6511 | else |
c5c76735 | 6512 | x = gen_pop (gen_rtx_REG (SImode, rn)); |
f676971a | 6513 | |
1245df60 | 6514 | x = emit_insn (x); |
b757d36f | 6515 | add_reg_note (x, REG_INC, gen_rtx_REG (SImode, STACK_POINTER_REGNUM)); |
8aa2a305 JW |
6516 | } |
6517 | ||
1245df60 | 6518 | /* Generate code to push the regs specified in the mask. */ |
8aa2a305 JW |
6519 | |
6520 | static void | |
cf277499 | 6521 | push_regs (HARD_REG_SET *mask, int interrupt_handler) |
8aa2a305 | 6522 | { |
5c7cafa8 | 6523 | int i = interrupt_handler ? LAST_BANKED_REG + 1 : 0; |
ac55736a | 6524 | int skip_fpscr = 0; |
8aa2a305 | 6525 | |
1245df60 R |
6526 | /* Push PR last; this gives better latencies after the prologue, and |
6527 | candidates for the return delay slot when there are no general | |
6528 | registers pushed. */ | |
5c7cafa8 | 6529 | for (; i < FIRST_PSEUDO_REGISTER; i++) |
ac55736a R |
6530 | { |
6531 | /* If this is an interrupt handler, and the SZ bit varies, | |
6532 | and we have to push any floating point register, we need | |
6533 | to switch to the correct precision first. */ | |
6534 | if (i == FIRST_FP_REG && interrupt_handler && TARGET_FMOVD | |
56b138ae | 6535 | && hard_reg_set_intersect_p (*mask, reg_class_contents[DF_REGS])) |
ac55736a R |
6536 | { |
6537 | HARD_REG_SET unsaved; | |
6538 | ||
6539 | push (FPSCR_REG); | |
832a3292 | 6540 | COMPL_HARD_REG_SET (unsaved, *mask); |
ac55736a R |
6541 | fpscr_set_from_mem (NORMAL_MODE (FP_MODE), unsaved); |
6542 | skip_fpscr = 1; | |
6543 | } | |
6544 | if (i != PR_REG | |
6545 | && (i != FPSCR_REG || ! skip_fpscr) | |
6546 | && TEST_HARD_REG_BIT (*mask, i)) | |
561642fa AP |
6547 | { |
6548 | /* If the ISR has RESBANK attribute assigned, don't push any of | |
6549 | the following registers - R0-R14, MACH, MACL and GBR. */ | |
6550 | if (! (sh_cfun_resbank_handler_p () | |
6551 | && ((i >= FIRST_GENERAL_REG && i < LAST_GENERAL_REG) | |
6552 | || i == MACH_REG | |
6553 | || i == MACL_REG | |
6554 | || i == GBR_REG))) | |
6555 | push (i); | |
6556 | } | |
ac55736a | 6557 | } |
5c7cafa8 CB |
6558 | |
6559 | /* Push banked registers last to improve delay slot opportunities. */ | |
6560 | if (interrupt_handler) | |
91a19652 N |
6561 | { |
6562 | bool use_movml = false; | |
6563 | ||
6564 | if (TARGET_SH2A) | |
6565 | { | |
6566 | unsigned int count = 0; | |
6567 | ||
6568 | for (i = FIRST_BANKED_REG; i <= LAST_BANKED_REG; i++) | |
6569 | if (TEST_HARD_REG_BIT (*mask, i)) | |
6570 | count++; | |
6571 | else | |
6572 | break; | |
6573 | ||
6574 | /* Use movml when all banked registers are pushed. */ | |
6575 | if (count == LAST_BANKED_REG - FIRST_BANKED_REG + 1) | |
6576 | use_movml = true; | |
6577 | } | |
6578 | ||
6579 | if (use_movml) | |
6580 | { | |
6581 | rtx x, mem, reg, set; | |
6582 | rtx sp_reg = gen_rtx_REG (SImode, STACK_POINTER_REGNUM); | |
6583 | ||
6584 | /* We must avoid scheduling multiple store insn with another | |
6585 | insns. */ | |
6586 | emit_insn (gen_blockage ()); | |
6587 | x = gen_movml_push_banked (sp_reg); | |
6588 | x = frame_insn (x); | |
6589 | for (i = FIRST_BANKED_REG; i <= LAST_BANKED_REG; i++) | |
6590 | { | |
6591 | mem = gen_rtx_MEM (SImode, plus_constant (sp_reg, i * 4)); | |
6592 | reg = gen_rtx_REG (SImode, i); | |
6593 | add_reg_note (x, REG_CFA_OFFSET, gen_rtx_SET (SImode, mem, reg)); | |
6594 | } | |
6595 | ||
6596 | set = gen_rtx_SET (SImode, sp_reg, plus_constant (sp_reg, - 32)); | |
6597 | add_reg_note (x, REG_CFA_ADJUST_CFA, set); | |
6598 | emit_insn (gen_blockage ()); | |
6599 | } | |
6600 | else | |
6601 | for (i = FIRST_BANKED_REG; i <= LAST_BANKED_REG; i++) | |
6602 | if (TEST_HARD_REG_BIT (*mask, i)) | |
6603 | push (i); | |
6604 | } | |
5c7cafa8 | 6605 | |
561642fa AP |
6606 | /* Don't push PR register for an ISR with RESBANK attribute assigned. */ |
6607 | if (TEST_HARD_REG_BIT (*mask, PR_REG) && !sh_cfun_resbank_handler_p ()) | |
1245df60 | 6608 | push (PR_REG); |
8aa2a305 JW |
6609 | } |
6610 | ||
fe3ad572 SC |
6611 | /* Calculate how much extra space is needed to save all callee-saved |
6612 | target registers. | |
6613 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
6614 | ||
6615 | static int | |
6616 | shmedia_target_regs_stack_space (HARD_REG_SET *live_regs_mask) | |
6617 | { | |
6618 | int reg; | |
6619 | int stack_space = 0; | |
6620 | int interrupt_handler = sh_cfun_interrupt_handler_p (); | |
6621 | ||
6622 | for (reg = LAST_TARGET_REG; reg >= FIRST_TARGET_REG; reg--) | |
62164eb4 | 6623 | if ((! call_really_used_regs[reg] || interrupt_handler) |
fe3ad572 SC |
6624 | && ! TEST_HARD_REG_BIT (*live_regs_mask, reg)) |
6625 | /* Leave space to save this target register on the stack, | |
24746a42 | 6626 | in case target register allocation wants to use it. */ |
fe3ad572 SC |
6627 | stack_space += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); |
6628 | return stack_space; | |
6629 | } | |
f676971a | 6630 | |
fe3ad572 SC |
6631 | /* Decide whether we should reserve space for callee-save target registers, |
6632 | in case target register allocation wants to use them. REGS_SAVED is | |
6633 | the space, in bytes, that is already required for register saves. | |
6634 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
6635 | ||
6636 | static int | |
6637 | shmedia_reserve_space_for_target_registers_p (int regs_saved, | |
6638 | HARD_REG_SET *live_regs_mask) | |
6639 | { | |
6640 | if (optimize_size) | |
6641 | return 0; | |
6642 | return shmedia_target_regs_stack_space (live_regs_mask) <= regs_saved; | |
6643 | } | |
6644 | ||
6645 | /* Decide how much space to reserve for callee-save target registers | |
6646 | in case target register allocation wants to use them. | |
6647 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
6648 | ||
6649 | static int | |
6650 | shmedia_target_regs_stack_adjust (HARD_REG_SET *live_regs_mask) | |
6651 | { | |
6652 | if (shmedia_space_reserved_for_target_registers) | |
6653 | return shmedia_target_regs_stack_space (live_regs_mask); | |
6654 | else | |
6655 | return 0; | |
6656 | } | |
6657 | ||
8aa2a305 | 6658 | /* Work out the registers which need to be saved, both as a mask and a |
490dbac7 | 6659 | count of saved words. Return the count. |
8aa2a305 JW |
6660 | |
6661 | If doing a pragma interrupt function, then push all regs used by the | |
6662 | function, and if we call another function (we can tell by looking at PR), | |
6663 | make sure that all the regs it clobbers are safe too. */ | |
6664 | ||
490dbac7 | 6665 | static int |
cf277499 | 6666 | calc_live_regs (HARD_REG_SET *live_regs_mask) |
8aa2a305 | 6667 | { |
0ef2c525 | 6668 | unsigned int reg; |
1245df60 | 6669 | int count; |
a6ab9fc0 R |
6670 | tree attrs; |
6671 | bool interrupt_or_trapa_handler, trapa_handler, interrupt_handler; | |
6672 | bool nosave_low_regs; | |
fada1961 | 6673 | int pr_live, has_call; |
157131d7 | 6674 | |
a6ab9fc0 R |
6675 | attrs = DECL_ATTRIBUTES (current_function_decl); |
6676 | interrupt_or_trapa_handler = sh_cfun_interrupt_handler_p (); | |
6677 | trapa_handler = lookup_attribute ("trapa_handler", attrs) != NULL_TREE; | |
6678 | interrupt_handler = interrupt_or_trapa_handler && ! trapa_handler; | |
6679 | nosave_low_regs = lookup_attribute ("nosave_low_regs", attrs) != NULL_TREE; | |
8aa2a305 | 6680 | |
fada1961 | 6681 | CLEAR_HARD_REG_SET (*live_regs_mask); |
157371cf | 6682 | if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && interrupt_handler |
6fb5fa3c | 6683 | && df_regs_ever_live_p (FPSCR_REG)) |
c0fb94d7 | 6684 | target_flags &= ~MASK_FPU_SINGLE; |
225e4f43 | 6685 | /* If we can save a lot of saves by switching to double mode, do that. */ |
157371cf | 6686 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && TARGET_FPU_SINGLE) |
225e4f43 | 6687 | for (count = 0, reg = FIRST_FP_REG; reg <= LAST_FP_REG; reg += 2) |
6fb5fa3c | 6688 | if (df_regs_ever_live_p (reg) && df_regs_ever_live_p (reg+1) |
62164eb4 | 6689 | && (! call_really_used_regs[reg] |
a6ab9fc0 | 6690 | || interrupt_handler) |
225e4f43 R |
6691 | && ++count > 2) |
6692 | { | |
c0fb94d7 | 6693 | target_flags &= ~MASK_FPU_SINGLE; |
225e4f43 R |
6694 | break; |
6695 | } | |
9d7ed806 R |
6696 | /* PR_MEDIA_REG is a general purpose register, thus global_alloc already |
6697 | knows how to use it. That means the pseudo originally allocated for | |
6698 | the initial value can become the PR_MEDIA_REG hard register, as seen for | |
6699 | execute/20010122-1.c:test9. */ | |
6700 | if (TARGET_SHMEDIA) | |
7d73a2ba R |
6701 | /* ??? this function is called from initial_elimination_offset, hence we |
6702 | can't use the result of sh_media_register_for_return here. */ | |
97b3d0f0 | 6703 | pr_live = sh_pr_n_sets (); |
9d7ed806 R |
6704 | else |
6705 | { | |
6706 | rtx pr_initial = has_hard_reg_initial_val (Pmode, PR_REG); | |
6707 | pr_live = (pr_initial | |
f3536097 | 6708 | ? (!REG_P (pr_initial) |
29db2a1e | 6709 | || REGNO (pr_initial) != (PR_REG)) |
6fb5fa3c | 6710 | : df_regs_ever_live_p (PR_REG)); |
7d73a2ba R |
6711 | /* For Shcompact, if not optimizing, we end up with a memory reference |
6712 | using the return address pointer for __builtin_return_address even | |
6713 | though there is no actual need to put the PR register on the stack. */ | |
6fb5fa3c | 6714 | pr_live |= df_regs_ever_live_p (RETURN_ADDRESS_POINTER_REGNUM); |
9d7ed806 | 6715 | } |
fa5322fa AO |
6716 | /* Force PR to be live if the prologue has to call the SHmedia |
6717 | argument decoder or register saver. */ | |
6718 | if (TARGET_SHCOMPACT | |
38173d38 | 6719 | && ((crtl->args.info.call_cookie |
fa5322fa | 6720 | & ~ CALL_COOKIE_RET_TRAMP (1)) |
e3b5732b | 6721 | || crtl->saves_all_registers)) |
fa5322fa | 6722 | pr_live = 1; |
fada1961 | 6723 | has_call = TARGET_SHMEDIA ? ! leaf_function_p () : pr_live; |
0ef2c525 | 6724 | for (count = 0, reg = FIRST_PSEUDO_REGISTER; reg-- != 0; ) |
8aa2a305 | 6725 | { |
97b3d0f0 | 6726 | if (reg == (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG) |
3007d592 | 6727 | ? pr_live |
a6ab9fc0 | 6728 | : interrupt_handler |
1245df60 | 6729 | ? (/* Need to save all the regs ever live. */ |
6fb5fa3c | 6730 | (df_regs_ever_live_p (reg) |
62164eb4 R |
6731 | || (call_really_used_regs[reg] |
6732 | && (! fixed_regs[reg] || reg == MACH_REG || reg == MACL_REG | |
6733 | || reg == PIC_OFFSET_TABLE_REGNUM) | |
fada1961 | 6734 | && has_call) |
a6ab9fc0 R |
6735 | || (TARGET_SHMEDIA && has_call |
6736 | && REGISTER_NATURAL_MODE (reg) == SImode | |
fada1961 | 6737 | && (GENERAL_REGISTER_P (reg) || TARGET_REGISTER_P (reg)))) |
1245df60 R |
6738 | && reg != STACK_POINTER_REGNUM && reg != ARG_POINTER_REGNUM |
6739 | && reg != RETURN_ADDRESS_POINTER_REGNUM | |
7144b2d8 DD |
6740 | && reg != T_REG && reg != GBR_REG |
6741 | /* Push fpscr only on targets which have FPU */ | |
6742 | && (reg != FPSCR_REG || TARGET_FPU_ANY)) | |
1245df60 | 6743 | : (/* Only push those regs which are used and need to be saved. */ |
180bde4f RS |
6744 | (TARGET_SHCOMPACT |
6745 | && flag_pic | |
38173d38 | 6746 | && crtl->args.info.call_cookie |
0ef2c525 | 6747 | && reg == PIC_OFFSET_TABLE_REGNUM) |
6fb5fa3c | 6748 | || (df_regs_ever_live_p (reg) |
9fdcdc74 KK |
6749 | && ((!call_really_used_regs[reg] |
6750 | && !(reg != PIC_OFFSET_TABLE_REGNUM | |
6751 | && fixed_regs[reg] && call_used_regs[reg])) | |
a6ab9fc0 | 6752 | || (trapa_handler && reg == FPSCR_REG && TARGET_FPU_ANY))) |
e3b5732b | 6753 | || (crtl->calls_eh_return |
0ef2c525 KK |
6754 | && (reg == EH_RETURN_DATA_REGNO (0) |
6755 | || reg == EH_RETURN_DATA_REGNO (1) | |
6756 | || reg == EH_RETURN_DATA_REGNO (2) | |
6757 | || reg == EH_RETURN_DATA_REGNO (3))) | |
61f71b34 | 6758 | || ((reg == MACL_REG || reg == MACH_REG) |
6fb5fa3c | 6759 | && df_regs_ever_live_p (reg) |
61f71b34 DD |
6760 | && sh_cfun_attr_renesas_p ()) |
6761 | )) | |
8aa2a305 | 6762 | { |
490dbac7 | 6763 | SET_HARD_REG_BIT (*live_regs_mask, reg); |
fa5322fa AO |
6764 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); |
6765 | ||
157371cf | 6766 | if ((TARGET_SH4 || TARGET_SH2A_DOUBLE || TARGET_SH5) && TARGET_FMOVD |
fa5322fa | 6767 | && GET_MODE_CLASS (REGISTER_NATURAL_MODE (reg)) == MODE_FLOAT) |
eda44c78 | 6768 | { |
104ee20b | 6769 | if (FP_REGISTER_P (reg)) |
eda44c78 | 6770 | { |
6fb5fa3c | 6771 | if (! TARGET_FPU_SINGLE && ! df_regs_ever_live_p (reg ^ 1)) |
eda44c78 | 6772 | { |
490dbac7 | 6773 | SET_HARD_REG_BIT (*live_regs_mask, (reg ^ 1)); |
fa5322fa | 6774 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg ^ 1)); |
eda44c78 R |
6775 | } |
6776 | } | |
fa5322fa | 6777 | else if (XD_REGISTER_P (reg)) |
eda44c78 R |
6778 | { |
6779 | /* Must switch to double mode to access these registers. */ | |
c0fb94d7 | 6780 | target_flags &= ~MASK_FPU_SINGLE; |
eda44c78 R |
6781 | } |
6782 | } | |
8aa2a305 | 6783 | } |
a6ab9fc0 R |
6784 | if (nosave_low_regs && reg == R8_REG) |
6785 | break; | |
8aa2a305 | 6786 | } |
fe3ad572 SC |
6787 | /* If we have a target register optimization pass after prologue / epilogue |
6788 | threading, we need to assume all target registers will be live even if | |
6789 | they aren't now. */ | |
6790 | if (flag_branch_target_load_optimize2 | |
6791 | && TARGET_SAVE_ALL_TARGET_REGS | |
6792 | && shmedia_space_reserved_for_target_registers) | |
6793 | for (reg = LAST_TARGET_REG; reg >= FIRST_TARGET_REG; reg--) | |
62164eb4 | 6794 | if ((! call_really_used_regs[reg] || interrupt_handler) |
fe3ad572 SC |
6795 | && ! TEST_HARD_REG_BIT (*live_regs_mask, reg)) |
6796 | { | |
6797 | SET_HARD_REG_BIT (*live_regs_mask, reg); | |
6798 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); | |
6799 | } | |
fada1961 R |
6800 | /* If this is an interrupt handler, we don't have any call-clobbered |
6801 | registers we can conveniently use for target register save/restore. | |
6802 | Make sure we save at least one general purpose register when we need | |
6803 | to save target registers. */ | |
6804 | if (interrupt_handler | |
56b138ae RS |
6805 | && hard_reg_set_intersect_p (*live_regs_mask, |
6806 | reg_class_contents[TARGET_REGS]) | |
6807 | && ! hard_reg_set_intersect_p (*live_regs_mask, | |
6808 | reg_class_contents[GENERAL_REGS])) | |
fada1961 R |
6809 | { |
6810 | SET_HARD_REG_BIT (*live_regs_mask, R0_REG); | |
6811 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (R0_REG)); | |
6812 | } | |
8aa2a305 | 6813 | |
490dbac7 | 6814 | return count; |
8aa2a305 JW |
6815 | } |
6816 | ||
6817 | /* Code to generate prologue and epilogue sequences */ | |
b9654711 | 6818 | |
a93d1ba2 | 6819 | /* PUSHED is the number of bytes that are being pushed on the |
0ad913af R |
6820 | stack for register saves. Return the frame size, padded |
6821 | appropriately so that the stack stays properly aligned. */ | |
e5f8d4c0 | 6822 | static HOST_WIDE_INT |
cf277499 | 6823 | rounded_frame_size (int pushed) |
e5f8d4c0 R |
6824 | { |
6825 | HOST_WIDE_INT size = get_frame_size (); | |
6826 | HOST_WIDE_INT align = STACK_BOUNDARY / BITS_PER_UNIT; | |
6827 | ||
7a296495 CB |
6828 | if (ACCUMULATE_OUTGOING_ARGS) |
6829 | size += crtl->outgoing_args_size; | |
6830 | ||
c5b9ef02 | 6831 | return ((size + pushed + align - 1) & -align) - pushed; |
e5f8d4c0 R |
6832 | } |
6833 | ||
fa5322fa AO |
6834 | /* Choose a call-clobbered target-branch register that remains |
6835 | unchanged along the whole function. We set it up as the return | |
6836 | value in the prologue. */ | |
6837 | int | |
cf277499 | 6838 | sh_media_register_for_return (void) |
fa5322fa AO |
6839 | { |
6840 | int regno; | |
6841 | int tr0_used; | |
6842 | ||
6843 | if (! current_function_is_leaf) | |
6844 | return -1; | |
fada1961 R |
6845 | if (lookup_attribute ("interrupt_handler", |
6846 | DECL_ATTRIBUTES (current_function_decl))) | |
6847 | return -1; | |
73a4d10b R |
6848 | if (sh_cfun_interrupt_handler_p ()) |
6849 | return -1; | |
fa5322fa | 6850 | |
6fb5fa3c | 6851 | tr0_used = flag_pic && df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM); |
fa5322fa AO |
6852 | |
6853 | for (regno = FIRST_TARGET_REG + tr0_used; regno <= LAST_TARGET_REG; regno++) | |
6fb5fa3c | 6854 | if (call_really_used_regs[regno] && ! df_regs_ever_live_p (regno)) |
fa5322fa AO |
6855 | return regno; |
6856 | ||
6857 | return -1; | |
6858 | } | |
6859 | ||
fada1961 R |
6860 | /* The maximum registers we need to save are: |
6861 | - 62 general purpose registers (r15 is stack pointer, r63 is zero) | |
6862 | - 32 floating point registers (for each pair, we save none, | |
6863 | one single precision value, or a double precision value). | |
6864 | - 8 target registers | |
6865 | - add 1 entry for a delimiter. */ | |
6866 | #define MAX_SAVED_REGS (62+32+8) | |
6867 | ||
6868 | typedef struct save_entry_s | |
6869 | { | |
6870 | unsigned char reg; | |
6871 | unsigned char mode; | |
6872 | short offset; | |
6873 | } save_entry; | |
6874 | ||
6875 | #define MAX_TEMPS 4 | |
6876 | ||
6877 | /* There will be a delimiter entry with VOIDmode both at the start and the | |
6878 | end of a filled in schedule. The end delimiter has the offset of the | |
6879 | save with the smallest (i.e. most negative) offset. */ | |
6880 | typedef struct save_schedule_s | |
6881 | { | |
6882 | save_entry entries[MAX_SAVED_REGS + 2]; | |
6883 | int temps[MAX_TEMPS+1]; | |
6884 | } save_schedule; | |
6885 | ||
6886 | /* Fill in SCHEDULE according to LIVE_REGS_MASK. If RESTORE is nonzero, | |
6887 | use reverse order. Returns the last entry written to (not counting | |
6888 | the delimiter). OFFSET_BASE is a number to be added to all offset | |
6889 | entries. */ | |
f676971a | 6890 | |
fada1961 R |
6891 | static save_entry * |
6892 | sh5_schedule_saves (HARD_REG_SET *live_regs_mask, save_schedule *schedule, | |
6893 | int offset_base) | |
6894 | { | |
6895 | int align, i; | |
6896 | save_entry *entry = schedule->entries; | |
6897 | int tmpx = 0; | |
6898 | int offset; | |
6899 | ||
6900 | if (! current_function_interrupt) | |
6901 | for (i = FIRST_GENERAL_REG; tmpx < MAX_TEMPS && i <= LAST_GENERAL_REG; i++) | |
62164eb4 | 6902 | if (call_really_used_regs[i] && ! fixed_regs[i] && i != PR_MEDIA_REG |
fada1961 R |
6903 | && ! FUNCTION_ARG_REGNO_P (i) |
6904 | && i != FIRST_RET_REG | |
6de9cd9a | 6905 | && ! (cfun->static_chain_decl != NULL && i == STATIC_CHAIN_REGNUM) |
e3b5732b | 6906 | && ! (crtl->calls_eh_return |
fada1961 | 6907 | && (i == EH_RETURN_STACKADJ_REGNO |
f94cef0a KK |
6908 | || ((unsigned) i >= EH_RETURN_DATA_REGNO (0) |
6909 | && (unsigned) i <= EH_RETURN_DATA_REGNO (3))))) | |
fada1961 R |
6910 | schedule->temps[tmpx++] = i; |
6911 | entry->reg = -1; | |
6912 | entry->mode = VOIDmode; | |
6913 | entry->offset = offset_base; | |
6914 | entry++; | |
6915 | /* We loop twice: first, we save 8-byte aligned registers in the | |
6916 | higher addresses, that are known to be aligned. Then, we | |
6917 | proceed to saving 32-bit registers that don't need 8-byte | |
6918 | alignment. | |
6919 | If this is an interrupt function, all registers that need saving | |
6920 | need to be saved in full. moreover, we need to postpone saving | |
6921 | target registers till we have saved some general purpose registers | |
6922 | we can then use as scratch registers. */ | |
6923 | offset = offset_base; | |
6924 | for (align = 1; align >= 0; align--) | |
6925 | { | |
6926 | for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; i--) | |
6927 | if (TEST_HARD_REG_BIT (*live_regs_mask, i)) | |
6928 | { | |
6929 | enum machine_mode mode = REGISTER_NATURAL_MODE (i); | |
6930 | int reg = i; | |
6931 | ||
6932 | if (current_function_interrupt) | |
6933 | { | |
6934 | if (TARGET_REGISTER_P (i)) | |
6935 | continue; | |
6936 | if (GENERAL_REGISTER_P (i)) | |
6937 | mode = DImode; | |
6938 | } | |
6939 | if (mode == SFmode && (i % 2) == 1 | |
6940 | && ! TARGET_FPU_SINGLE && FP_REGISTER_P (i) | |
6941 | && (TEST_HARD_REG_BIT (*live_regs_mask, (i ^ 1)))) | |
6942 | { | |
6943 | mode = DFmode; | |
6944 | i--; | |
6945 | reg--; | |
6946 | } | |
6947 | ||
6948 | /* If we're doing the aligned pass and this is not aligned, | |
6949 | or we're doing the unaligned pass and this is aligned, | |
6950 | skip it. */ | |
6951 | if ((GET_MODE_SIZE (mode) % (STACK_BOUNDARY / BITS_PER_UNIT) == 0) | |
6952 | != align) | |
6953 | continue; | |
6954 | ||
6955 | if (current_function_interrupt | |
6956 | && GENERAL_REGISTER_P (i) | |
6957 | && tmpx < MAX_TEMPS) | |
6958 | schedule->temps[tmpx++] = i; | |
6959 | ||
6960 | offset -= GET_MODE_SIZE (mode); | |
6961 | entry->reg = i; | |
6962 | entry->mode = mode; | |
6963 | entry->offset = offset; | |
6964 | entry++; | |
6965 | } | |
6966 | if (align && current_function_interrupt) | |
6967 | for (i = LAST_TARGET_REG; i >= FIRST_TARGET_REG; i--) | |
6968 | if (TEST_HARD_REG_BIT (*live_regs_mask, i)) | |
6969 | { | |
6970 | offset -= GET_MODE_SIZE (DImode); | |
6971 | entry->reg = i; | |
6972 | entry->mode = DImode; | |
6973 | entry->offset = offset; | |
6974 | entry++; | |
6975 | } | |
6976 | } | |
6977 | entry->reg = -1; | |
6978 | entry->mode = VOIDmode; | |
6979 | entry->offset = offset; | |
6980 | schedule->temps[tmpx] = -1; | |
6981 | return entry - 1; | |
6982 | } | |
6983 | ||
b9654711 | 6984 | void |
cf277499 | 6985 | sh_expand_prologue (void) |
b9654711 | 6986 | { |
490dbac7 | 6987 | HARD_REG_SET live_regs_mask; |
40d2032b | 6988 | int d, i; |
fa5322fa | 6989 | int d_rounding = 0; |
225e4f43 | 6990 | int save_flags = target_flags; |
49616835 | 6991 | int pretend_args; |
e631262b | 6992 | int stack_usage; |
a6ab9fc0 R |
6993 | tree sp_switch_attr |
6994 | = lookup_attribute ("sp_switch", DECL_ATTRIBUTES (current_function_decl)); | |
b9654711 | 6995 | |
e3ba8d11 | 6996 | current_function_interrupt = sh_cfun_interrupt_handler_p (); |
552ecbd9 | 6997 | |
0d7e008e | 6998 | /* We have pretend args if we had an object sent partially in registers |
8aa2a305 | 6999 | and partially on the stack, e.g. a large structure. */ |
38173d38 | 7000 | pretend_args = crtl->args.pretend_args_size; |
49616835 JR |
7001 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl) |
7002 | && (NPARM_REGS(SImode) | |
38173d38 | 7003 | > crtl->args.info.arg_count[(int) SH_ARG_INT])) |
49616835 | 7004 | pretend_args = 0; |
803b9370 | 7005 | |
49616835 | 7006 | output_stack_adjust (-pretend_args |
38173d38 | 7007 | - crtl->args.info.stack_regs * 8, |
803b9370 | 7008 | stack_pointer_rtx, 0, NULL, true); |
e631262b | 7009 | stack_usage = pretend_args + crtl->args.info.stack_regs * 8; |
b9654711 | 7010 | |
38173d38 | 7011 | if (TARGET_SHCOMPACT && flag_pic && crtl->args.info.call_cookie) |
2b1fd83f AO |
7012 | /* We're going to use the PIC register to load the address of the |
7013 | incoming-argument decoder and/or of the return trampoline from | |
7014 | the GOT, so make sure the PIC register is preserved and | |
7015 | initialized. */ | |
6fb5fa3c | 7016 | df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true); |
2b1fd83f | 7017 | |
fa5322fa | 7018 | if (TARGET_SHCOMPACT |
38173d38 | 7019 | && (crtl->args.info.call_cookie & ~ CALL_COOKIE_RET_TRAMP(1))) |
fa5322fa AO |
7020 | { |
7021 | int reg; | |
7022 | ||
7023 | /* First, make all registers with incoming arguments that will | |
7024 | be pushed onto the stack live, so that register renaming | |
7025 | doesn't overwrite them. */ | |
7026 | for (reg = 0; reg < NPARM_REGS (SImode); reg++) | |
38173d38 | 7027 | if (CALL_COOKIE_STACKSEQ_GET (crtl->args.info.call_cookie) |
fa5322fa AO |
7028 | >= NPARM_REGS (SImode) - reg) |
7029 | for (; reg < NPARM_REGS (SImode); reg++) | |
7030 | emit_insn (gen_shcompact_preserve_incoming_args | |
7031 | (gen_rtx_REG (SImode, FIRST_PARM_REG + reg))); | |
7032 | else if (CALL_COOKIE_INT_REG_GET | |
38173d38 | 7033 | (crtl->args.info.call_cookie, reg) == 1) |
fa5322fa AO |
7034 | emit_insn (gen_shcompact_preserve_incoming_args |
7035 | (gen_rtx_REG (SImode, FIRST_PARM_REG + reg))); | |
7036 | ||
7037 | emit_move_insn (gen_rtx_REG (Pmode, MACL_REG), | |
7038 | stack_pointer_rtx); | |
7039 | emit_move_insn (gen_rtx_REG (SImode, R0_REG), | |
38173d38 | 7040 | GEN_INT (crtl->args.info.call_cookie)); |
fa5322fa AO |
7041 | emit_move_insn (gen_rtx_REG (SImode, MACH_REG), |
7042 | gen_rtx_REG (SImode, R0_REG)); | |
7043 | } | |
7044 | else if (TARGET_SHMEDIA) | |
7045 | { | |
7046 | int tr = sh_media_register_for_return (); | |
7047 | ||
7048 | if (tr >= 0) | |
6fb5fa3c DB |
7049 | emit_move_insn (gen_rtx_REG (DImode, tr), |
7050 | gen_rtx_REG (DImode, PR_MEDIA_REG)); | |
fa5322fa AO |
7051 | } |
7052 | ||
7ab56274 | 7053 | /* Emit the code for SETUP_VARARGS. */ |
e3b5732b | 7054 | if (cfun->stdarg) |
b9654711 | 7055 | { |
49616835 | 7056 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl)) |
1245df60 | 7057 | { |
1a95a963 JW |
7058 | /* Push arg regs as if they'd been provided by caller in stack. */ |
7059 | for (i = 0; i < NPARM_REGS(SImode); i++) | |
7060 | { | |
7061 | int rn = NPARM_REGS(SImode) + FIRST_PARM_REG - i - 1; | |
2754d3c5 | 7062 | |
f676971a | 7063 | if (i >= (NPARM_REGS(SImode) |
38173d38 | 7064 | - crtl->args.info.arg_count[(int) SH_ARG_INT] |
aec373f1 | 7065 | )) |
1a95a963 | 7066 | break; |
535b951f | 7067 | push (rn); |
e631262b | 7068 | stack_usage += GET_MODE_SIZE (SImode); |
1a95a963 | 7069 | } |
1245df60 | 7070 | } |
b9654711 | 7071 | } |
1a95a963 | 7072 | |
4408efce | 7073 | /* If we're supposed to switch stacks at function entry, do so now. */ |
a6ab9fc0 R |
7074 | if (sp_switch_attr) |
7075 | { | |
20e19db4 | 7076 | rtx lab, newsrc; |
a6ab9fc0 R |
7077 | /* The argument specifies a variable holding the address of the |
7078 | stack the interrupt function should switch to/from at entry/exit. */ | |
20e19db4 | 7079 | tree arg = TREE_VALUE ( TREE_VALUE (sp_switch_attr)); |
a6ab9fc0 | 7080 | const char *s |
20e19db4 | 7081 | = ggc_strdup (TREE_STRING_POINTER (arg)); |
a6ab9fc0 R |
7082 | rtx sp_switch = gen_rtx_SYMBOL_REF (Pmode, s); |
7083 | ||
20e19db4 DD |
7084 | lab = add_constant (sp_switch, SImode, 0); |
7085 | newsrc = gen_rtx_LABEL_REF (VOIDmode, lab); | |
7086 | newsrc = gen_const_mem (SImode, newsrc); | |
7087 | ||
7088 | emit_insn (gen_sp_switch_1 (newsrc)); | |
a6ab9fc0 | 7089 | } |
4408efce | 7090 | |
490dbac7 | 7091 | d = calc_live_regs (&live_regs_mask); |
225e4f43 R |
7092 | /* ??? Maybe we could save some switching if we can move a mode switch |
7093 | that already happens to be at the function start into the prologue. */ | |
ac55736a | 7094 | if (target_flags != save_flags && ! current_function_interrupt) |
225e4f43 | 7095 | emit_insn (gen_toggle_sz ()); |
f676971a | 7096 | |
fa5322fa AO |
7097 | if (TARGET_SH5) |
7098 | { | |
fada1961 R |
7099 | int offset_base, offset; |
7100 | rtx r0 = NULL_RTX; | |
fa5322fa AO |
7101 | int offset_in_r0 = -1; |
7102 | int sp_in_r0 = 0; | |
fe3ad572 SC |
7103 | int tregs_space = shmedia_target_regs_stack_adjust (&live_regs_mask); |
7104 | int total_size, save_size; | |
fada1961 R |
7105 | save_schedule schedule; |
7106 | save_entry *entry; | |
7107 | int *tmp_pnt; | |
7108 | ||
62164eb4 | 7109 | if (call_really_used_regs[R0_REG] && ! fixed_regs[R0_REG] |
fada1961 R |
7110 | && ! current_function_interrupt) |
7111 | r0 = gen_rtx_REG (Pmode, R0_REG); | |
fe3ad572 SC |
7112 | |
7113 | /* D is the actual number of bytes that we need for saving registers, | |
7114 | however, in initial_elimination_offset we have committed to using | |
7115 | an additional TREGS_SPACE amount of bytes - in order to keep both | |
7116 | addresses to arguments supplied by the caller and local variables | |
7117 | valid, we must keep this gap. Place it between the incoming | |
7118 | arguments and the actually saved registers in a bid to optimize | |
7119 | locality of reference. */ | |
7120 | total_size = d + tregs_space; | |
7121 | total_size += rounded_frame_size (total_size); | |
7122 | save_size = total_size - rounded_frame_size (d); | |
7123 | if (save_size % (STACK_BOUNDARY / BITS_PER_UNIT)) | |
fa5322fa | 7124 | d_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) |
fe3ad572 SC |
7125 | - save_size % (STACK_BOUNDARY / BITS_PER_UNIT)); |
7126 | ||
7127 | /* If adjusting the stack in a single step costs nothing extra, do so. | |
7128 | I.e. either if a single addi is enough, or we need a movi anyway, | |
7129 | and we don't exceed the maximum offset range (the test for the | |
7130 | latter is conservative for simplicity). */ | |
7131 | if (TARGET_SHMEDIA | |
7132 | && (CONST_OK_FOR_I10 (-total_size) | |
7133 | || (! CONST_OK_FOR_I10 (-(save_size + d_rounding)) | |
7134 | && total_size <= 2044))) | |
7135 | d_rounding = total_size - save_size; | |
fa5322fa | 7136 | |
fada1961 | 7137 | offset_base = d + d_rounding; |
fe3ad572 SC |
7138 | |
7139 | output_stack_adjust (-(save_size + d_rounding), stack_pointer_rtx, | |
21d05694 | 7140 | 0, NULL, true); |
e631262b | 7141 | stack_usage += save_size + d_rounding; |
fa5322fa | 7142 | |
fada1961 R |
7143 | sh5_schedule_saves (&live_regs_mask, &schedule, offset_base); |
7144 | tmp_pnt = schedule.temps; | |
7145 | for (entry = &schedule.entries[1]; entry->mode != VOIDmode; entry++) | |
7146 | { | |
81f40b79 | 7147 | enum machine_mode mode = (enum machine_mode) entry->mode; |
0ef2c525 | 7148 | unsigned int reg = entry->reg; |
fada1961 | 7149 | rtx reg_rtx, mem_rtx, pre_dec = NULL_RTX; |
455786b6 | 7150 | rtx orig_reg_rtx; |
fa5322fa | 7151 | |
fada1961 | 7152 | offset = entry->offset; |
fa5322fa | 7153 | |
fada1961 | 7154 | reg_rtx = gen_rtx_REG (mode, reg); |
fa5322fa | 7155 | |
57d38024 R |
7156 | mem_rtx = gen_frame_mem (mode, |
7157 | gen_rtx_PLUS (Pmode, | |
7158 | stack_pointer_rtx, | |
7159 | GEN_INT (offset))); | |
fa5322fa | 7160 | |
6de7a512 PB |
7161 | if (!memory_address_p (mode, XEXP (mem_rtx, 0))) |
7162 | { | |
7163 | gcc_assert (r0); | |
7164 | mem_rtx = NULL_RTX; | |
7165 | } | |
fa5322fa | 7166 | |
6de7a512 PB |
7167 | if (HAVE_PRE_DECREMENT |
7168 | && (offset_in_r0 - offset == GET_MODE_SIZE (mode) | |
7169 | || mem_rtx == NULL_RTX | |
7170 | || reg == PR_REG || SPECIAL_REGISTER_P (reg))) | |
7171 | { | |
7172 | pre_dec = gen_frame_mem (mode, gen_rtx_PRE_DEC (Pmode, r0)); | |
fa5322fa | 7173 | |
6de7a512 | 7174 | if (!memory_address_p (mode, XEXP (pre_dec, 0))) |
fada1961 | 7175 | pre_dec = NULL_RTX; |
6de7a512 PB |
7176 | else |
7177 | { | |
7178 | mem_rtx = NULL_RTX; | |
7179 | offset += GET_MODE_SIZE (mode); | |
7180 | } | |
7181 | } | |
fa5322fa | 7182 | |
fada1961 R |
7183 | if (mem_rtx != NULL_RTX) |
7184 | goto addr_ok; | |
fa5322fa | 7185 | |
fada1961 R |
7186 | if (offset_in_r0 == -1) |
7187 | { | |
7188 | emit_move_insn (r0, GEN_INT (offset)); | |
7189 | offset_in_r0 = offset; | |
7190 | } | |
7191 | else if (offset != offset_in_r0) | |
7192 | { | |
7193 | emit_move_insn (r0, | |
7194 | gen_rtx_PLUS | |
7195 | (Pmode, r0, | |
7196 | GEN_INT (offset - offset_in_r0))); | |
7197 | offset_in_r0 += offset - offset_in_r0; | |
7198 | } | |
f676971a | 7199 | |
fada1961 R |
7200 | if (pre_dec != NULL_RTX) |
7201 | { | |
7202 | if (! sp_in_r0) | |
fa5322fa AO |
7203 | { |
7204 | emit_move_insn (r0, | |
7205 | gen_rtx_PLUS | |
fada1961 R |
7206 | (Pmode, r0, stack_pointer_rtx)); |
7207 | sp_in_r0 = 1; | |
fa5322fa | 7208 | } |
fa5322fa | 7209 | |
fada1961 R |
7210 | offset -= GET_MODE_SIZE (mode); |
7211 | offset_in_r0 -= GET_MODE_SIZE (mode); | |
fa5322fa | 7212 | |
fada1961 R |
7213 | mem_rtx = pre_dec; |
7214 | } | |
7215 | else if (sp_in_r0) | |
57d38024 | 7216 | mem_rtx = gen_frame_mem (mode, r0); |
fada1961 | 7217 | else |
57d38024 R |
7218 | mem_rtx = gen_frame_mem (mode, |
7219 | gen_rtx_PLUS (Pmode, | |
7220 | stack_pointer_rtx, | |
7221 | r0)); | |
fada1961 R |
7222 | |
7223 | /* We must not use an r0-based address for target-branch | |
7224 | registers or for special registers without pre-dec | |
7225 | memory addresses, since we store their values in r0 | |
7226 | first. */ | |
f5b9e7c9 NS |
7227 | gcc_assert (!TARGET_REGISTER_P (reg) |
7228 | && ((reg != PR_REG && !SPECIAL_REGISTER_P (reg)) | |
7229 | || mem_rtx == pre_dec)); | |
7230 | ||
fada1961 | 7231 | addr_ok: |
455786b6 | 7232 | orig_reg_rtx = reg_rtx; |
fada1961 R |
7233 | if (TARGET_REGISTER_P (reg) |
7234 | || ((reg == PR_REG || SPECIAL_REGISTER_P (reg)) | |
7235 | && mem_rtx != pre_dec)) | |
7236 | { | |
7237 | rtx tmp_reg = gen_rtx_REG (GET_MODE (reg_rtx), *tmp_pnt); | |
fa5322fa | 7238 | |
fada1961 | 7239 | emit_move_insn (tmp_reg, reg_rtx); |
fa5322fa | 7240 | |
fada1961 R |
7241 | if (REGNO (tmp_reg) == R0_REG) |
7242 | { | |
fa5322fa AO |
7243 | offset_in_r0 = -1; |
7244 | sp_in_r0 = 0; | |
f5b9e7c9 NS |
7245 | gcc_assert (!refers_to_regno_p |
7246 | (R0_REG, R0_REG+1, mem_rtx, (rtx *) 0)); | |
fa5322fa AO |
7247 | } |
7248 | ||
fada1961 R |
7249 | if (*++tmp_pnt <= 0) |
7250 | tmp_pnt = schedule.temps; | |
7251 | ||
7252 | reg_rtx = tmp_reg; | |
fa5322fa | 7253 | } |
fada1961 R |
7254 | { |
7255 | rtx insn; | |
7256 | ||
7257 | /* Mark as interesting for dwarf cfi generator */ | |
7258 | insn = emit_move_insn (mem_rtx, reg_rtx); | |
7259 | RTX_FRAME_RELATED_P (insn) = 1; | |
455786b6 KK |
7260 | /* If we use an intermediate register for the save, we can't |
7261 | describe this exactly in cfi as a copy of the to-be-saved | |
7262 | register into the temporary register and then the temporary | |
7263 | register on the stack, because the temporary register can | |
7264 | have a different natural size than the to-be-saved register. | |
7265 | Thus, we gloss over the intermediate copy and pretend we do | |
7266 | a direct save from the to-be-saved register. */ | |
7267 | if (REGNO (reg_rtx) != reg) | |
7268 | { | |
b757d36f | 7269 | rtx set; |
455786b6 KK |
7270 | |
7271 | set = gen_rtx_SET (VOIDmode, mem_rtx, orig_reg_rtx); | |
b757d36f | 7272 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, set); |
455786b6 | 7273 | } |
fada1961 | 7274 | |
f676971a | 7275 | if (TARGET_SHCOMPACT && (offset_in_r0 != -1)) |
fada1961 R |
7276 | { |
7277 | rtx reg_rtx = gen_rtx_REG (mode, reg); | |
b757d36f | 7278 | rtx set; |
57d38024 R |
7279 | rtx mem_rtx = gen_frame_mem (mode, |
7280 | gen_rtx_PLUS (Pmode, | |
7281 | stack_pointer_rtx, | |
7282 | GEN_INT (offset))); | |
fada1961 R |
7283 | |
7284 | set = gen_rtx_SET (VOIDmode, mem_rtx, reg_rtx); | |
b757d36f | 7285 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, set); |
fada1961 R |
7286 | } |
7287 | } | |
7288 | } | |
fa5322fa | 7289 | |
f5b9e7c9 | 7290 | gcc_assert (entry->offset == d_rounding); |
fa5322fa AO |
7291 | } |
7292 | else | |
e631262b KK |
7293 | { |
7294 | push_regs (&live_regs_mask, current_function_interrupt); | |
7295 | stack_usage += d; | |
7296 | } | |
1a66cd67 | 7297 | |
6fb5fa3c DB |
7298 | if (flag_pic && df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM)) |
7299 | emit_insn (gen_GOTaddr2picreg ()); | |
1a66cd67 | 7300 | |
fa5322fa AO |
7301 | if (SHMEDIA_REGS_STACK_ADJUST ()) |
7302 | { | |
fa5322fa AO |
7303 | /* This must NOT go through the PLT, otherwise mach and macl |
7304 | may be clobbered. */ | |
73a4d10b R |
7305 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
7306 | (TARGET_FPU_ANY | |
7307 | ? "__GCC_push_shmedia_regs" | |
7308 | : "__GCC_push_shmedia_regs_nofpu"), SFUNC_GOT); | |
fa5322fa AO |
7309 | emit_insn (gen_shmedia_save_restore_regs_compact |
7310 | (GEN_INT (-SHMEDIA_REGS_STACK_ADJUST ()))); | |
7311 | } | |
7312 | ||
ac55736a | 7313 | if (target_flags != save_flags && ! current_function_interrupt) |
6fb5fa3c | 7314 | emit_insn (gen_toggle_sz ()); |
1a95a963 | 7315 | |
225e4f43 R |
7316 | target_flags = save_flags; |
7317 | ||
fa5322fa | 7318 | output_stack_adjust (-rounded_frame_size (d) + d_rounding, |
21d05694 | 7319 | stack_pointer_rtx, 0, NULL, true); |
e631262b | 7320 | stack_usage += rounded_frame_size (d) - d_rounding; |
b9654711 SC |
7321 | |
7322 | if (frame_pointer_needed) | |
96a2347e | 7323 | frame_insn (GEN_MOV (hard_frame_pointer_rtx, stack_pointer_rtx)); |
fa5322fa | 7324 | |
fa5322fa | 7325 | if (TARGET_SHCOMPACT |
38173d38 | 7326 | && (crtl->args.info.call_cookie & ~ CALL_COOKIE_RET_TRAMP(1))) |
fa5322fa AO |
7327 | { |
7328 | /* This must NOT go through the PLT, otherwise mach and macl | |
7329 | may be clobbered. */ | |
73a4d10b R |
7330 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
7331 | "__GCC_shcompact_incoming_args", SFUNC_GOT); | |
fa5322fa AO |
7332 | emit_insn (gen_shcompact_incoming_args ()); |
7333 | } | |
e631262b KK |
7334 | |
7335 | if (flag_stack_usage) | |
7336 | current_function_static_stack_size = stack_usage; | |
b9654711 SC |
7337 | } |
7338 | ||
7339 | void | |
726d4cb7 | 7340 | sh_expand_epilogue (bool sibcall_p) |
b9654711 | 7341 | { |
490dbac7 | 7342 | HARD_REG_SET live_regs_mask; |
40d2032b | 7343 | int d, i; |
fa5322fa | 7344 | int d_rounding = 0; |
b9654711 | 7345 | |
225e4f43 | 7346 | int save_flags = target_flags; |
fe3ad572 | 7347 | int frame_size, save_size; |
ac55736a | 7348 | int fpscr_deferred = 0; |
726d4cb7 | 7349 | int e = sibcall_p ? -1 : 1; |
3d5a0820 | 7350 | |
490dbac7 | 7351 | d = calc_live_regs (&live_regs_mask); |
3d5a0820 | 7352 | |
fe3ad572 SC |
7353 | save_size = d; |
7354 | frame_size = rounded_frame_size (d); | |
7355 | ||
7356 | if (TARGET_SH5) | |
7357 | { | |
7358 | int tregs_space = shmedia_target_regs_stack_adjust (&live_regs_mask); | |
7359 | int total_size; | |
7360 | if (d % (STACK_BOUNDARY / BITS_PER_UNIT)) | |
7361 | d_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) | |
7362 | - d % (STACK_BOUNDARY / BITS_PER_UNIT)); | |
7363 | ||
7364 | total_size = d + tregs_space; | |
7365 | total_size += rounded_frame_size (total_size); | |
7366 | save_size = total_size - frame_size; | |
7367 | ||
7368 | /* If adjusting the stack in a single step costs nothing extra, do so. | |
7369 | I.e. either if a single addi is enough, or we need a movi anyway, | |
7370 | and we don't exceed the maximum offset range (the test for the | |
7371 | latter is conservative for simplicity). */ | |
7372 | if (TARGET_SHMEDIA | |
7373 | && ! frame_pointer_needed | |
7374 | && (CONST_OK_FOR_I10 (total_size) | |
7375 | || (! CONST_OK_FOR_I10 (save_size + d_rounding) | |
7376 | && total_size <= 2044))) | |
7377 | d_rounding = frame_size; | |
fa5322fa | 7378 | |
fe3ad572 SC |
7379 | frame_size -= d_rounding; |
7380 | } | |
16bea517 | 7381 | |
b9654711 | 7382 | if (frame_pointer_needed) |
46d81ffa | 7383 | { |
1964788a | 7384 | /* We must avoid scheduling the epilogue with previous basic blocks. |
47ea8d4e | 7385 | See PR/18032 and PR/40313. */ |
1964788a | 7386 | emit_insn (gen_blockage ()); |
96a2347e | 7387 | output_stack_adjust (frame_size, hard_frame_pointer_rtx, e, |
21d05694 | 7388 | &live_regs_mask, false); |
07f5b9aa JL |
7389 | |
7390 | /* We must avoid moving the stack pointer adjustment past code | |
7391 | which reads from the local frame, else an interrupt could | |
7392 | occur after the SP adjustment and clobber data in the local | |
7393 | frame. */ | |
7394 | emit_insn (gen_blockage ()); | |
96a2347e | 7395 | emit_insn (GEN_MOV (stack_pointer_rtx, hard_frame_pointer_rtx)); |
46d81ffa | 7396 | } |
3d5a0820 | 7397 | else if (frame_size) |
07f5b9aa JL |
7398 | { |
7399 | /* We must avoid moving the stack pointer adjustment past code | |
7400 | which reads from the local frame, else an interrupt could | |
7401 | occur after the SP adjustment and clobber data in the local | |
7402 | frame. */ | |
7403 | emit_insn (gen_blockage ()); | |
21d05694 DD |
7404 | output_stack_adjust (frame_size, stack_pointer_rtx, e, |
7405 | &live_regs_mask, false); | |
07f5b9aa | 7406 | } |
b9654711 | 7407 | |
fa5322fa AO |
7408 | if (SHMEDIA_REGS_STACK_ADJUST ()) |
7409 | { | |
73a4d10b R |
7410 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
7411 | (TARGET_FPU_ANY | |
7412 | ? "__GCC_pop_shmedia_regs" | |
7413 | : "__GCC_pop_shmedia_regs_nofpu"), SFUNC_GOT); | |
fa5322fa AO |
7414 | /* This must NOT go through the PLT, otherwise mach and macl |
7415 | may be clobbered. */ | |
7416 | emit_insn (gen_shmedia_save_restore_regs_compact | |
7417 | (GEN_INT (SHMEDIA_REGS_STACK_ADJUST ()))); | |
7418 | } | |
7419 | ||
16bea517 | 7420 | /* Pop all the registers. */ |
0d7e008e | 7421 | |
ac55736a | 7422 | if (target_flags != save_flags && ! current_function_interrupt) |
225e4f43 | 7423 | emit_insn (gen_toggle_sz ()); |
fa5322fa AO |
7424 | if (TARGET_SH5) |
7425 | { | |
fada1961 | 7426 | int offset_base, offset; |
fa5322fa AO |
7427 | int offset_in_r0 = -1; |
7428 | int sp_in_r0 = 0; | |
fa5322fa | 7429 | rtx r0 = gen_rtx_REG (Pmode, R0_REG); |
fada1961 R |
7430 | save_schedule schedule; |
7431 | save_entry *entry; | |
7432 | int *tmp_pnt; | |
7d73a2ba | 7433 | |
fada1961 R |
7434 | entry = sh5_schedule_saves (&live_regs_mask, &schedule, d_rounding); |
7435 | offset_base = -entry[1].offset + d_rounding; | |
7436 | tmp_pnt = schedule.temps; | |
7437 | for (; entry->mode != VOIDmode; entry--) | |
7438 | { | |
81f40b79 | 7439 | enum machine_mode mode = (enum machine_mode) entry->mode; |
fada1961 | 7440 | int reg = entry->reg; |
535b951f | 7441 | rtx reg_rtx, mem_rtx, post_inc = NULL_RTX; |
fa5322fa | 7442 | |
fada1961 R |
7443 | offset = offset_base + entry->offset; |
7444 | reg_rtx = gen_rtx_REG (mode, reg); | |
fa5322fa | 7445 | |
57d38024 R |
7446 | mem_rtx = gen_frame_mem (mode, |
7447 | gen_rtx_PLUS (Pmode, | |
7448 | stack_pointer_rtx, | |
7449 | GEN_INT (offset))); | |
fa5322fa | 7450 | |
6de7a512 PB |
7451 | if (!memory_address_p (mode, XEXP (mem_rtx, 0))) |
7452 | mem_rtx = NULL_RTX; | |
fa5322fa | 7453 | |
6de7a512 PB |
7454 | if (HAVE_POST_INCREMENT |
7455 | && (offset == offset_in_r0 | |
7456 | || (offset + GET_MODE_SIZE (mode) != d + d_rounding | |
7457 | && mem_rtx == NULL_RTX) | |
7458 | || reg == PR_REG || SPECIAL_REGISTER_P (reg))) | |
7459 | { | |
7460 | post_inc = gen_frame_mem (mode, gen_rtx_POST_INC (Pmode, r0)); | |
fa5322fa | 7461 | |
6de7a512 | 7462 | if (!memory_address_p (mode, XEXP (post_inc, 0))) |
fada1961 | 7463 | post_inc = NULL_RTX; |
6de7a512 | 7464 | else |
fada1961 | 7465 | mem_rtx = NULL_RTX; |
6de7a512 | 7466 | } |
f676971a | 7467 | |
fada1961 R |
7468 | if (mem_rtx != NULL_RTX) |
7469 | goto addr_ok; | |
fa5322fa | 7470 | |
fada1961 R |
7471 | if (offset_in_r0 == -1) |
7472 | { | |
7473 | emit_move_insn (r0, GEN_INT (offset)); | |
7474 | offset_in_r0 = offset; | |
7475 | } | |
7476 | else if (offset != offset_in_r0) | |
7477 | { | |
7478 | emit_move_insn (r0, | |
7479 | gen_rtx_PLUS | |
7480 | (Pmode, r0, | |
7481 | GEN_INT (offset - offset_in_r0))); | |
7482 | offset_in_r0 += offset - offset_in_r0; | |
7483 | } | |
f676971a | 7484 | |
fada1961 R |
7485 | if (post_inc != NULL_RTX) |
7486 | { | |
7487 | if (! sp_in_r0) | |
fa5322fa AO |
7488 | { |
7489 | emit_move_insn (r0, | |
7490 | gen_rtx_PLUS | |
fada1961 R |
7491 | (Pmode, r0, stack_pointer_rtx)); |
7492 | sp_in_r0 = 1; | |
fa5322fa | 7493 | } |
f676971a | 7494 | |
fada1961 | 7495 | mem_rtx = post_inc; |
fa5322fa | 7496 | |
fada1961 R |
7497 | offset_in_r0 += GET_MODE_SIZE (mode); |
7498 | } | |
7499 | else if (sp_in_r0) | |
57d38024 | 7500 | mem_rtx = gen_frame_mem (mode, r0); |
fada1961 | 7501 | else |
57d38024 R |
7502 | mem_rtx = gen_frame_mem (mode, |
7503 | gen_rtx_PLUS (Pmode, | |
7504 | stack_pointer_rtx, | |
7505 | r0)); | |
fa5322fa | 7506 | |
f5b9e7c9 NS |
7507 | gcc_assert ((reg != PR_REG && !SPECIAL_REGISTER_P (reg)) |
7508 | || mem_rtx == post_inc); | |
fa5322fa | 7509 | |
fada1961 R |
7510 | addr_ok: |
7511 | if ((reg == PR_REG || SPECIAL_REGISTER_P (reg)) | |
7512 | && mem_rtx != post_inc) | |
7513 | { | |
535b951f | 7514 | emit_move_insn (r0, mem_rtx); |
fada1961 R |
7515 | mem_rtx = r0; |
7516 | } | |
7517 | else if (TARGET_REGISTER_P (reg)) | |
7518 | { | |
7519 | rtx tmp_reg = gen_rtx_REG (mode, *tmp_pnt); | |
7520 | ||
7521 | /* Give the scheduler a bit of freedom by using up to | |
7522 | MAX_TEMPS registers in a round-robin fashion. */ | |
535b951f | 7523 | emit_move_insn (tmp_reg, mem_rtx); |
fada1961 R |
7524 | mem_rtx = tmp_reg; |
7525 | if (*++tmp_pnt < 0) | |
7526 | tmp_pnt = schedule.temps; | |
fa5322fa AO |
7527 | } |
7528 | ||
535b951f | 7529 | emit_move_insn (reg_rtx, mem_rtx); |
fada1961 R |
7530 | } |
7531 | ||
f5b9e7c9 | 7532 | gcc_assert (entry->offset + offset_base == d + d_rounding); |
fa5322fa | 7533 | } |
fe3ad572 | 7534 | else /* ! TARGET_SH5 */ |
b9654711 | 7535 | { |
5c7cafa8 CB |
7536 | int last_reg; |
7537 | ||
fe3ad572 | 7538 | save_size = 0; |
561642fa AP |
7539 | /* For an ISR with RESBANK attribute assigned, don't pop PR |
7540 | register. */ | |
7541 | if (TEST_HARD_REG_BIT (live_regs_mask, PR_REG) | |
7542 | && !sh_cfun_resbank_handler_p ()) | |
6fb5fa3c DB |
7543 | { |
7544 | if (!frame_pointer_needed) | |
7545 | emit_insn (gen_blockage ()); | |
7546 | pop (PR_REG); | |
7547 | } | |
5c7cafa8 | 7548 | |
6a62ca52 | 7549 | /* Banked registers are popped first to avoid being scheduled in the |
5c7cafa8 CB |
7550 | delay slot. RTE switches banks before the ds instruction. */ |
7551 | if (current_function_interrupt) | |
7552 | { | |
91a19652 N |
7553 | bool use_movml = false; |
7554 | ||
7555 | if (TARGET_SH2A) | |
7556 | { | |
7557 | unsigned int count = 0; | |
7558 | ||
7559 | for (i = FIRST_BANKED_REG; i <= LAST_BANKED_REG; i++) | |
7560 | if (TEST_HARD_REG_BIT (live_regs_mask, i)) | |
7561 | count++; | |
7562 | else | |
7563 | break; | |
7564 | ||
7565 | /* Use movml when all banked register are poped. */ | |
7566 | if (count == LAST_BANKED_REG - FIRST_BANKED_REG + 1) | |
7567 | use_movml = true; | |
7568 | } | |
7569 | ||
7570 | if (use_movml) | |
7571 | { | |
7572 | rtx sp_reg = gen_rtx_REG (SImode, STACK_POINTER_REGNUM); | |
7573 | ||
7574 | /* We must avoid scheduling multiple load insn with another | |
7575 | insns. */ | |
7576 | emit_insn (gen_blockage ()); | |
7577 | emit_insn (gen_movml_pop_banked (sp_reg)); | |
7578 | emit_insn (gen_blockage ()); | |
7579 | } | |
7580 | else | |
7581 | for (i = LAST_BANKED_REG; i >= FIRST_BANKED_REG; i--) | |
7582 | if (TEST_HARD_REG_BIT (live_regs_mask, i)) | |
7583 | pop (i); | |
5c7cafa8 CB |
7584 | |
7585 | last_reg = FIRST_PSEUDO_REGISTER - LAST_BANKED_REG - 1; | |
7586 | } | |
7587 | else | |
7588 | last_reg = FIRST_PSEUDO_REGISTER; | |
7589 | ||
7590 | for (i = 0; i < last_reg; i++) | |
fe3ad572 SC |
7591 | { |
7592 | int j = (FIRST_PSEUDO_REGISTER - 1) - i; | |
f676971a | 7593 | |
fe3ad572 | 7594 | if (j == FPSCR_REG && current_function_interrupt && TARGET_FMOVD |
56b138ae RS |
7595 | && hard_reg_set_intersect_p (live_regs_mask, |
7596 | reg_class_contents[DF_REGS])) | |
fe3ad572 | 7597 | fpscr_deferred = 1; |
561642fa AP |
7598 | /* For an ISR with RESBANK attribute assigned, don't pop |
7599 | following registers, R0-R14, MACH, MACL and GBR. */ | |
7600 | else if (j != PR_REG && TEST_HARD_REG_BIT (live_regs_mask, j) | |
7601 | && ! (sh_cfun_resbank_handler_p () | |
7602 | && ((j >= FIRST_GENERAL_REG | |
7603 | && j < LAST_GENERAL_REG) | |
7604 | || j == MACH_REG | |
7605 | || j == MACL_REG | |
7606 | || j == GBR_REG))) | |
fe3ad572 | 7607 | pop (j); |
5c7cafa8 | 7608 | |
fe3ad572 SC |
7609 | if (j == FIRST_FP_REG && fpscr_deferred) |
7610 | pop (FPSCR_REG); | |
fe3ad572 | 7611 | } |
b9654711 | 7612 | } |
ac55736a | 7613 | if (target_flags != save_flags && ! current_function_interrupt) |
225e4f43 R |
7614 | emit_insn (gen_toggle_sz ()); |
7615 | target_flags = save_flags; | |
b9654711 | 7616 | |
38173d38 | 7617 | output_stack_adjust (crtl->args.pretend_args_size |
fe3ad572 | 7618 | + save_size + d_rounding |
38173d38 | 7619 | + crtl->args.info.stack_regs * 8, |
21d05694 | 7620 | stack_pointer_rtx, e, NULL, false); |
4408efce | 7621 | |
e3b5732b | 7622 | if (crtl->calls_eh_return) |
4977bab6 ZW |
7623 | emit_insn (GEN_ADD3 (stack_pointer_rtx, stack_pointer_rtx, |
7624 | EH_RETURN_STACKADJ_RTX)); | |
7625 | ||
4408efce | 7626 | /* Switch back to the normal stack if necessary. */ |
a6ab9fc0 | 7627 | if (lookup_attribute ("sp_switch", DECL_ATTRIBUTES (current_function_decl))) |
4408efce | 7628 | emit_insn (gen_sp_switch_2 ()); |
89d4bc35 AO |
7629 | |
7630 | /* Tell flow the insn that pops PR isn't dead. */ | |
fa5322fa AO |
7631 | /* PR_REG will never be live in SHmedia mode, and we don't need to |
7632 | USE PR_MEDIA_REG, since it will be explicitly copied to TR0_REG | |
7633 | by the return pattern. */ | |
490dbac7 | 7634 | if (TEST_HARD_REG_BIT (live_regs_mask, PR_REG)) |
c41c1387 | 7635 | emit_use (gen_rtx_REG (SImode, PR_REG)); |
b9654711 SC |
7636 | } |
7637 | ||
afbc2905 R |
7638 | static int sh_need_epilogue_known = 0; |
7639 | ||
7640 | int | |
cf277499 | 7641 | sh_need_epilogue (void) |
afbc2905 R |
7642 | { |
7643 | if (! sh_need_epilogue_known) | |
7644 | { | |
7645 | rtx epilogue; | |
7646 | ||
7647 | start_sequence (); | |
726d4cb7 | 7648 | sh_expand_epilogue (0); |
2f937369 | 7649 | epilogue = get_insns (); |
afbc2905 | 7650 | end_sequence (); |
2f937369 | 7651 | sh_need_epilogue_known = (epilogue == NULL ? -1 : 1); |
afbc2905 R |
7652 | } |
7653 | return sh_need_epilogue_known > 0; | |
7654 | } | |
7655 | ||
4977bab6 ZW |
7656 | /* Emit code to change the current function's return address to RA. |
7657 | TEMP is available as a scratch register, if needed. */ | |
7658 | ||
7659 | void | |
cf277499 | 7660 | sh_set_return_address (rtx ra, rtx tmp) |
4977bab6 | 7661 | { |
490dbac7 | 7662 | HARD_REG_SET live_regs_mask; |
4977bab6 | 7663 | int d; |
4977bab6 ZW |
7664 | int pr_reg = TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG; |
7665 | int pr_offset; | |
7666 | ||
490dbac7 | 7667 | d = calc_live_regs (&live_regs_mask); |
4977bab6 ZW |
7668 | |
7669 | /* If pr_reg isn't life, we can set it (or the register given in | |
7670 | sh_media_register_for_return) directly. */ | |
490dbac7 | 7671 | if (! TEST_HARD_REG_BIT (live_regs_mask, pr_reg)) |
4977bab6 ZW |
7672 | { |
7673 | rtx rr; | |
7674 | ||
7675 | if (TARGET_SHMEDIA) | |
7676 | { | |
7677 | int rr_regno = sh_media_register_for_return (); | |
7678 | ||
7679 | if (rr_regno < 0) | |
7680 | rr_regno = pr_reg; | |
7681 | ||
7682 | rr = gen_rtx_REG (DImode, rr_regno); | |
7683 | } | |
7684 | else | |
7685 | rr = gen_rtx_REG (SImode, pr_reg); | |
7686 | ||
7687 | emit_insn (GEN_MOV (rr, ra)); | |
7688 | /* Tell flow the register for return isn't dead. */ | |
c41c1387 | 7689 | emit_use (rr); |
4977bab6 ZW |
7690 | return; |
7691 | } | |
7692 | ||
7693 | if (TARGET_SH5) | |
7694 | { | |
4977bab6 | 7695 | int offset; |
fada1961 R |
7696 | save_schedule schedule; |
7697 | save_entry *entry; | |
f676971a | 7698 | |
fada1961 R |
7699 | entry = sh5_schedule_saves (&live_regs_mask, &schedule, 0); |
7700 | offset = entry[1].offset; | |
7701 | for (; entry->mode != VOIDmode; entry--) | |
7702 | if (entry->reg == pr_reg) | |
7703 | goto found; | |
4977bab6 ZW |
7704 | |
7705 | /* We can't find pr register. */ | |
f5b9e7c9 | 7706 | gcc_unreachable (); |
4977bab6 ZW |
7707 | |
7708 | found: | |
fada1961 R |
7709 | offset = entry->offset - offset; |
7710 | pr_offset = (rounded_frame_size (d) + offset | |
4977bab6 ZW |
7711 | + SHMEDIA_REGS_STACK_ADJUST ()); |
7712 | } | |
7713 | else | |
fada1961 | 7714 | pr_offset = rounded_frame_size (d); |
4977bab6 ZW |
7715 | |
7716 | emit_insn (GEN_MOV (tmp, GEN_INT (pr_offset))); | |
7a296495 CB |
7717 | |
7718 | if (frame_pointer_needed) | |
7719 | emit_insn (GEN_ADD3 (tmp, tmp, hard_frame_pointer_rtx)); | |
7720 | else | |
7721 | emit_insn (GEN_ADD3 (tmp, tmp, stack_pointer_rtx)); | |
4977bab6 | 7722 | |
57d38024 | 7723 | tmp = gen_frame_mem (Pmode, tmp); |
4977bab6 | 7724 | emit_insn (GEN_MOV (tmp, ra)); |
7f9a14e3 KK |
7725 | /* Tell this store isn't dead. */ |
7726 | emit_use (tmp); | |
4977bab6 ZW |
7727 | } |
7728 | ||
8aa2a305 JW |
7729 | /* Clear variables at function end. */ |
7730 | ||
08c148a8 | 7731 | static void |
cf277499 SB |
7732 | sh_output_function_epilogue (FILE *file ATTRIBUTE_UNUSED, |
7733 | HOST_WIDE_INT size ATTRIBUTE_UNUSED) | |
8aa2a305 | 7734 | { |
afbc2905 | 7735 | sh_need_epilogue_known = 0; |
8aa2a305 JW |
7736 | } |
7737 | ||
61f71b34 | 7738 | static rtx |
cf277499 | 7739 | sh_builtin_saveregs (void) |
1a95a963 | 7740 | { |
1a95a963 | 7741 | /* First unnamed integer register. */ |
38173d38 | 7742 | int first_intreg = crtl->args.info.arg_count[(int) SH_ARG_INT]; |
1a95a963 JW |
7743 | /* Number of integer registers we need to save. */ |
7744 | int n_intregs = MAX (0, NPARM_REGS (SImode) - first_intreg); | |
7745 | /* First unnamed SFmode float reg */ | |
38173d38 | 7746 | int first_floatreg = crtl->args.info.arg_count[(int) SH_ARG_FLOAT]; |
1a95a963 JW |
7747 | /* Number of SFmode float regs to save. */ |
7748 | int n_floatregs = MAX (0, NPARM_REGS (SFmode) - first_floatreg); | |
318881c0 | 7749 | rtx regbuf, fpregs; |
ba4828e0 | 7750 | int bufsize, regno; |
4862826d | 7751 | alias_set_type alias_set; |
1a95a963 | 7752 | |
fa5322fa AO |
7753 | if (TARGET_SH5) |
7754 | { | |
7755 | if (n_intregs) | |
7756 | { | |
7757 | int pushregs = n_intregs; | |
7758 | ||
7759 | while (pushregs < NPARM_REGS (SImode) - 1 | |
7760 | && (CALL_COOKIE_INT_REG_GET | |
38173d38 | 7761 | (crtl->args.info.call_cookie, |
fa5322fa AO |
7762 | NPARM_REGS (SImode) - pushregs) |
7763 | == 1)) | |
7764 | { | |
38173d38 | 7765 | crtl->args.info.call_cookie |
fa5322fa AO |
7766 | &= ~ CALL_COOKIE_INT_REG (NPARM_REGS (SImode) |
7767 | - pushregs, 1); | |
7768 | pushregs++; | |
7769 | } | |
7770 | ||
7771 | if (pushregs == NPARM_REGS (SImode)) | |
38173d38 | 7772 | crtl->args.info.call_cookie |
fa5322fa AO |
7773 | |= (CALL_COOKIE_INT_REG (0, 1) |
7774 | | CALL_COOKIE_STACKSEQ (pushregs - 1)); | |
7775 | else | |
38173d38 | 7776 | crtl->args.info.call_cookie |
fa5322fa AO |
7777 | |= CALL_COOKIE_STACKSEQ (pushregs); |
7778 | ||
38173d38 | 7779 | crtl->args.pretend_args_size += 8 * n_intregs; |
fa5322fa AO |
7780 | } |
7781 | if (TARGET_SHCOMPACT) | |
7782 | return const0_rtx; | |
7783 | } | |
f676971a | 7784 | |
3a8699c7 | 7785 | if (! TARGET_SH2E && ! TARGET_SH4 && ! TARGET_SH5) |
fa5322fa AO |
7786 | { |
7787 | error ("__builtin_saveregs not supported by this subtarget"); | |
7788 | return const0_rtx; | |
7789 | } | |
7790 | ||
7791 | if (TARGET_SHMEDIA) | |
7792 | n_floatregs = 0; | |
7793 | ||
6f317ef3 | 7794 | /* Allocate block of memory for the regs. */ |
1a95a963 JW |
7795 | /* ??? If n_intregs + n_floatregs == 0, should we allocate at least 1 byte? |
7796 | Or can assign_stack_local accept a 0 SIZE argument? */ | |
7797 | bufsize = (n_intregs * UNITS_PER_WORD) + (n_floatregs * UNITS_PER_WORD); | |
7798 | ||
fa5322fa | 7799 | if (TARGET_SHMEDIA) |
57d38024 | 7800 | regbuf = gen_frame_mem (BLKmode, gen_rtx_REG (Pmode, ARG_POINTER_REGNUM)); |
f78ec36a R |
7801 | else if (n_floatregs & 1) |
7802 | { | |
7803 | rtx addr; | |
7804 | ||
7805 | regbuf = assign_stack_local (BLKmode, bufsize + UNITS_PER_WORD, 0); | |
7806 | addr = copy_to_mode_reg (Pmode, XEXP (regbuf, 0)); | |
7807 | emit_insn (gen_iorsi3 (addr, addr, GEN_INT (UNITS_PER_WORD))); | |
7808 | regbuf = change_address (regbuf, BLKmode, addr); | |
7809 | } | |
a11c004b R |
7810 | else if (STACK_BOUNDARY < 64 && TARGET_FPU_DOUBLE && n_floatregs) |
7811 | { | |
7812 | rtx addr, mask; | |
7813 | ||
7814 | regbuf = assign_stack_local (BLKmode, bufsize + UNITS_PER_WORD, 0); | |
7815 | addr = copy_to_mode_reg (Pmode, plus_constant (XEXP (regbuf, 0), 4)); | |
7816 | mask = copy_to_mode_reg (Pmode, GEN_INT (-8)); | |
7817 | emit_insn (gen_andsi3 (addr, addr, mask)); | |
7818 | regbuf = change_address (regbuf, BLKmode, addr); | |
7819 | } | |
fa5322fa | 7820 | else |
a11c004b | 7821 | regbuf = assign_stack_local (BLKmode, bufsize, TARGET_FPU_DOUBLE ? 64 : 0); |
514066a1 | 7822 | alias_set = get_varargs_alias_set (); |
ba4828e0 | 7823 | set_mem_alias_set (regbuf, alias_set); |
1a95a963 JW |
7824 | |
7825 | /* Save int args. | |
7826 | This is optimized to only save the regs that are necessary. Explicitly | |
7827 | named args need not be saved. */ | |
7828 | if (n_intregs > 0) | |
7829 | move_block_from_reg (BASE_ARG_REG (SImode) + first_intreg, | |
f4ef873c RK |
7830 | adjust_address (regbuf, BLKmode, |
7831 | n_floatregs * UNITS_PER_WORD), | |
c6b97fac | 7832 | n_intregs); |
1a95a963 | 7833 | |
fa5322fa AO |
7834 | if (TARGET_SHMEDIA) |
7835 | /* Return the address of the regbuf. */ | |
7836 | return XEXP (regbuf, 0); | |
7837 | ||
1a95a963 JW |
7838 | /* Save float args. |
7839 | This is optimized to only save the regs that are necessary. Explicitly | |
7840 | named args need not be saved. | |
7841 | We explicitly build a pointer to the buffer because it halves the insn | |
7842 | count when not optimizing (otherwise the pointer is built for each reg | |
1245df60 R |
7843 | saved). |
7844 | We emit the moves in reverse order so that we can use predecrement. */ | |
1a95a963 | 7845 | |
7d9c5aa1 R |
7846 | fpregs = copy_to_mode_reg (Pmode, |
7847 | plus_constant (XEXP (regbuf, 0), | |
7848 | n_floatregs * UNITS_PER_WORD)); | |
157371cf | 7849 | if (TARGET_SH4 || TARGET_SH2A_DOUBLE) |
225e4f43 | 7850 | { |
514066a1 | 7851 | rtx mem; |
225e4f43 R |
7852 | for (regno = NPARM_REGS (DFmode) - 2; regno >= first_floatreg; regno -= 2) |
7853 | { | |
7854 | emit_insn (gen_addsi3 (fpregs, fpregs, | |
7855 | GEN_INT (-2 * UNITS_PER_WORD))); | |
57d38024 | 7856 | mem = change_address (regbuf, DFmode, fpregs); |
f676971a | 7857 | emit_move_insn (mem, |
c0d4e710 | 7858 | gen_rtx_REG (DFmode, BASE_ARG_REG (DFmode) + regno)); |
225e4f43 R |
7859 | } |
7860 | regno = first_floatreg; | |
7861 | if (regno & 1) | |
7862 | { | |
832a3292 | 7863 | emit_insn (gen_addsi3 (fpregs, fpregs, GEN_INT (-UNITS_PER_WORD))); |
57d38024 | 7864 | mem = change_address (regbuf, SFmode, fpregs); |
514066a1 | 7865 | emit_move_insn (mem, |
c0d4e710 | 7866 | gen_rtx_REG (SFmode, BASE_ARG_REG (SFmode) + regno |
225e4f43 R |
7867 | - (TARGET_LITTLE_ENDIAN != 0))); |
7868 | } | |
7869 | } | |
7870 | else | |
1245df60 R |
7871 | for (regno = NPARM_REGS (SFmode) - 1; regno >= first_floatreg; regno--) |
7872 | { | |
514066a1 | 7873 | rtx mem; |
ba4828e0 | 7874 | |
832a3292 | 7875 | emit_insn (gen_addsi3 (fpregs, fpregs, GEN_INT (-UNITS_PER_WORD))); |
57d38024 | 7876 | mem = change_address (regbuf, SFmode, fpregs); |
514066a1 | 7877 | emit_move_insn (mem, |
c5c76735 | 7878 | gen_rtx_REG (SFmode, BASE_ARG_REG (SFmode) + regno)); |
1245df60 | 7879 | } |
1a95a963 JW |
7880 | |
7881 | /* Return the address of the regbuf. */ | |
7882 | return XEXP (regbuf, 0); | |
7883 | } | |
7884 | ||
514066a1 RH |
7885 | /* Define the `__builtin_va_list' type for the ABI. */ |
7886 | ||
c35d187f RH |
7887 | static tree |
7888 | sh_build_builtin_va_list (void) | |
514066a1 RH |
7889 | { |
7890 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
10575b5d | 7891 | tree record, type_decl; |
514066a1 | 7892 | |
61f71b34 DD |
7893 | if (TARGET_SH5 || (! TARGET_SH2E && ! TARGET_SH4) |
7894 | || TARGET_HITACHI || sh_cfun_attr_renesas_p ()) | |
514066a1 RH |
7895 | return ptr_type_node; |
7896 | ||
5d27ef94 | 7897 | record = (*lang_hooks.types.make_type) (RECORD_TYPE); |
10575b5d KK |
7898 | type_decl = build_decl (BUILTINS_LOCATION, |
7899 | TYPE_DECL, get_identifier ("__va_list_tag"), record); | |
514066a1 | 7900 | |
4c4bde29 AH |
7901 | f_next_o = build_decl (BUILTINS_LOCATION, |
7902 | FIELD_DECL, get_identifier ("__va_next_o"), | |
514066a1 | 7903 | ptr_type_node); |
4c4bde29 AH |
7904 | f_next_o_limit = build_decl (BUILTINS_LOCATION, |
7905 | FIELD_DECL, | |
514066a1 RH |
7906 | get_identifier ("__va_next_o_limit"), |
7907 | ptr_type_node); | |
4c4bde29 AH |
7908 | f_next_fp = build_decl (BUILTINS_LOCATION, |
7909 | FIELD_DECL, get_identifier ("__va_next_fp"), | |
514066a1 | 7910 | ptr_type_node); |
4c4bde29 AH |
7911 | f_next_fp_limit = build_decl (BUILTINS_LOCATION, |
7912 | FIELD_DECL, | |
514066a1 RH |
7913 | get_identifier ("__va_next_fp_limit"), |
7914 | ptr_type_node); | |
4c4bde29 AH |
7915 | f_next_stack = build_decl (BUILTINS_LOCATION, |
7916 | FIELD_DECL, get_identifier ("__va_next_stack"), | |
514066a1 RH |
7917 | ptr_type_node); |
7918 | ||
7919 | DECL_FIELD_CONTEXT (f_next_o) = record; | |
7920 | DECL_FIELD_CONTEXT (f_next_o_limit) = record; | |
7921 | DECL_FIELD_CONTEXT (f_next_fp) = record; | |
7922 | DECL_FIELD_CONTEXT (f_next_fp_limit) = record; | |
7923 | DECL_FIELD_CONTEXT (f_next_stack) = record; | |
7924 | ||
0fd2eac2 | 7925 | TYPE_STUB_DECL (record) = type_decl; |
10575b5d | 7926 | TYPE_NAME (record) = type_decl; |
514066a1 | 7927 | TYPE_FIELDS (record) = f_next_o; |
910ad8de NF |
7928 | DECL_CHAIN (f_next_o) = f_next_o_limit; |
7929 | DECL_CHAIN (f_next_o_limit) = f_next_fp; | |
7930 | DECL_CHAIN (f_next_fp) = f_next_fp_limit; | |
7931 | DECL_CHAIN (f_next_fp_limit) = f_next_stack; | |
514066a1 RH |
7932 | |
7933 | layout_type (record); | |
7934 | ||
7935 | return record; | |
7936 | } | |
7937 | ||
7938 | /* Implement `va_start' for varargs and stdarg. */ | |
7939 | ||
d7bd8aeb | 7940 | static void |
cf277499 | 7941 | sh_va_start (tree valist, rtx nextarg) |
514066a1 RH |
7942 | { |
7943 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
7944 | tree next_o, next_o_limit, next_fp, next_fp_limit, next_stack; | |
7945 | tree t, u; | |
7946 | int nfp, nint; | |
7947 | ||
fa5322fa AO |
7948 | if (TARGET_SH5) |
7949 | { | |
7950 | expand_builtin_saveregs (); | |
e5faf155 | 7951 | std_expand_builtin_va_start (valist, nextarg); |
fa5322fa AO |
7952 | return; |
7953 | } | |
7954 | ||
61f71b34 DD |
7955 | if ((! TARGET_SH2E && ! TARGET_SH4) |
7956 | || TARGET_HITACHI || sh_cfun_attr_renesas_p ()) | |
514066a1 | 7957 | { |
e5faf155 | 7958 | std_expand_builtin_va_start (valist, nextarg); |
514066a1 RH |
7959 | return; |
7960 | } | |
7961 | ||
7962 | f_next_o = TYPE_FIELDS (va_list_type_node); | |
910ad8de NF |
7963 | f_next_o_limit = DECL_CHAIN (f_next_o); |
7964 | f_next_fp = DECL_CHAIN (f_next_o_limit); | |
7965 | f_next_fp_limit = DECL_CHAIN (f_next_fp); | |
7966 | f_next_stack = DECL_CHAIN (f_next_fp_limit); | |
514066a1 | 7967 | |
47a25a46 | 7968 | next_o = build3 (COMPONENT_REF, TREE_TYPE (f_next_o), valist, f_next_o, |
44de5aeb | 7969 | NULL_TREE); |
47a25a46 RG |
7970 | next_o_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_o_limit), |
7971 | valist, f_next_o_limit, NULL_TREE); | |
7972 | next_fp = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp), valist, f_next_fp, | |
7973 | NULL_TREE); | |
7974 | next_fp_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp_limit), | |
7975 | valist, f_next_fp_limit, NULL_TREE); | |
7976 | next_stack = build3 (COMPONENT_REF, TREE_TYPE (f_next_stack), | |
7977 | valist, f_next_stack, NULL_TREE); | |
514066a1 RH |
7978 | |
7979 | /* Call __builtin_saveregs. */ | |
5be014d5 AP |
7980 | u = make_tree (sizetype, expand_builtin_saveregs ()); |
7981 | u = fold_convert (ptr_type_node, u); | |
726a989a | 7982 | t = build2 (MODIFY_EXPR, ptr_type_node, next_fp, u); |
514066a1 RH |
7983 | TREE_SIDE_EFFECTS (t) = 1; |
7984 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7985 | ||
38173d38 | 7986 | nfp = crtl->args.info.arg_count[SH_ARG_FLOAT]; |
514066a1 RH |
7987 | if (nfp < 8) |
7988 | nfp = 8 - nfp; | |
7989 | else | |
7990 | nfp = 0; | |
5be014d5 AP |
7991 | u = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, u, |
7992 | size_int (UNITS_PER_WORD * nfp)); | |
726a989a | 7993 | t = build2 (MODIFY_EXPR, ptr_type_node, next_fp_limit, u); |
514066a1 RH |
7994 | TREE_SIDE_EFFECTS (t) = 1; |
7995 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7996 | ||
726a989a | 7997 | t = build2 (MODIFY_EXPR, ptr_type_node, next_o, u); |
514066a1 RH |
7998 | TREE_SIDE_EFFECTS (t) = 1; |
7999 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
8000 | ||
38173d38 | 8001 | nint = crtl->args.info.arg_count[SH_ARG_INT]; |
514066a1 RH |
8002 | if (nint < 4) |
8003 | nint = 4 - nint; | |
8004 | else | |
8005 | nint = 0; | |
5be014d5 AP |
8006 | u = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, u, |
8007 | size_int (UNITS_PER_WORD * nint)); | |
726a989a | 8008 | t = build2 (MODIFY_EXPR, ptr_type_node, next_o_limit, u); |
514066a1 RH |
8009 | TREE_SIDE_EFFECTS (t) = 1; |
8010 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
8011 | ||
8012 | u = make_tree (ptr_type_node, nextarg); | |
726a989a | 8013 | t = build2 (MODIFY_EXPR, ptr_type_node, next_stack, u); |
514066a1 RH |
8014 | TREE_SIDE_EFFECTS (t) = 1; |
8015 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
8016 | } | |
8017 | ||
917f1b7e | 8018 | /* TYPE is a RECORD_TYPE. If there is only a single nonzero-sized |
21bdb106 JR |
8019 | member, return it. */ |
8020 | static tree | |
8021 | find_sole_member (tree type) | |
8022 | { | |
8023 | tree field, member = NULL_TREE; | |
8024 | ||
910ad8de | 8025 | for (field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field)) |
21bdb106 JR |
8026 | { |
8027 | if (TREE_CODE (field) != FIELD_DECL) | |
8028 | continue; | |
8029 | if (!DECL_SIZE (field)) | |
8030 | return NULL_TREE; | |
8031 | if (integer_zerop (DECL_SIZE (field))) | |
8032 | continue; | |
8033 | if (member) | |
8034 | return NULL_TREE; | |
8035 | member = field; | |
8036 | } | |
8037 | return member; | |
8038 | } | |
514066a1 RH |
8039 | /* Implement `va_arg'. */ |
8040 | ||
5fac2ed2 | 8041 | static tree |
726a989a RB |
8042 | sh_gimplify_va_arg_expr (tree valist, tree type, gimple_seq *pre_p, |
8043 | gimple_seq *post_p ATTRIBUTE_UNUSED) | |
514066a1 RH |
8044 | { |
8045 | HOST_WIDE_INT size, rsize; | |
318881c0 | 8046 | tree tmp, pptr_type_node; |
0ef2c525 | 8047 | tree addr, lab_over = NULL, result = NULL; |
03603192 | 8048 | int pass_by_ref = targetm.calls.must_pass_in_stack (TYPE_MODE (type), type); |
11160b58 | 8049 | tree eff_type; |
5fac2ed2 RH |
8050 | |
8051 | if (pass_by_ref) | |
8052 | type = build_pointer_type (type); | |
514066a1 RH |
8053 | |
8054 | size = int_size_in_bytes (type); | |
8055 | rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD; | |
8056 | pptr_type_node = build_pointer_type (ptr_type_node); | |
8057 | ||
61f71b34 DD |
8058 | if (! TARGET_SH5 && (TARGET_SH2E || TARGET_SH4) |
8059 | && ! (TARGET_HITACHI || sh_cfun_attr_renesas_p ())) | |
514066a1 RH |
8060 | { |
8061 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
8062 | tree next_o, next_o_limit, next_fp, next_fp_limit, next_stack; | |
8063 | int pass_as_float; | |
5fac2ed2 | 8064 | tree lab_false; |
21bdb106 | 8065 | tree member; |
514066a1 RH |
8066 | |
8067 | f_next_o = TYPE_FIELDS (va_list_type_node); | |
910ad8de NF |
8068 | f_next_o_limit = DECL_CHAIN (f_next_o); |
8069 | f_next_fp = DECL_CHAIN (f_next_o_limit); | |
8070 | f_next_fp_limit = DECL_CHAIN (f_next_fp); | |
8071 | f_next_stack = DECL_CHAIN (f_next_fp_limit); | |
514066a1 | 8072 | |
47a25a46 RG |
8073 | next_o = build3 (COMPONENT_REF, TREE_TYPE (f_next_o), valist, f_next_o, |
8074 | NULL_TREE); | |
8075 | next_o_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_o_limit), | |
8076 | valist, f_next_o_limit, NULL_TREE); | |
8077 | next_fp = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp), | |
8078 | valist, f_next_fp, NULL_TREE); | |
8079 | next_fp_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp_limit), | |
8080 | valist, f_next_fp_limit, NULL_TREE); | |
8081 | next_stack = build3 (COMPONENT_REF, TREE_TYPE (f_next_stack), | |
8082 | valist, f_next_stack, NULL_TREE); | |
514066a1 | 8083 | |
6d1cb95f R |
8084 | /* Structures with a single member with a distinct mode are passed |
8085 | like their member. This is relevant if the latter has a REAL_TYPE | |
8086 | or COMPLEX_TYPE type. */ | |
11160b58 R |
8087 | eff_type = type; |
8088 | while (TREE_CODE (eff_type) == RECORD_TYPE | |
8089 | && (member = find_sole_member (eff_type)) | |
21bdb106 JR |
8090 | && (TREE_CODE (TREE_TYPE (member)) == REAL_TYPE |
8091 | || TREE_CODE (TREE_TYPE (member)) == COMPLEX_TYPE | |
8092 | || TREE_CODE (TREE_TYPE (member)) == RECORD_TYPE)) | |
c26104e4 | 8093 | { |
21bdb106 | 8094 | tree field_type = TREE_TYPE (member); |
c26104e4 | 8095 | |
11160b58 R |
8096 | if (TYPE_MODE (eff_type) == TYPE_MODE (field_type)) |
8097 | eff_type = field_type; | |
c26104e4 | 8098 | else |
8a2209bf | 8099 | { |
11160b58 | 8100 | gcc_assert ((TYPE_ALIGN (eff_type) |
21bdb106 | 8101 | < GET_MODE_ALIGNMENT (TYPE_MODE (field_type))) |
11160b58 | 8102 | || (TYPE_ALIGN (eff_type) |
21bdb106 JR |
8103 | > GET_MODE_BITSIZE (TYPE_MODE (field_type)))); |
8104 | break; | |
8a2209bf | 8105 | } |
c26104e4 | 8106 | } |
5fac2ed2 | 8107 | |
bdf57c40 | 8108 | if (TARGET_SH4 || TARGET_SH2A_DOUBLE) |
514066a1 | 8109 | { |
11160b58 R |
8110 | pass_as_float = ((TREE_CODE (eff_type) == REAL_TYPE && size <= 8) |
8111 | || (TREE_CODE (eff_type) == COMPLEX_TYPE | |
8112 | && TREE_CODE (TREE_TYPE (eff_type)) == REAL_TYPE | |
514066a1 RH |
8113 | && size <= 16)); |
8114 | } | |
8115 | else | |
8116 | { | |
11160b58 | 8117 | pass_as_float = (TREE_CODE (eff_type) == REAL_TYPE && size == 4); |
514066a1 RH |
8118 | } |
8119 | ||
5fac2ed2 | 8120 | addr = create_tmp_var (pptr_type_node, NULL); |
4c4bde29 AH |
8121 | lab_false = create_artificial_label (UNKNOWN_LOCATION); |
8122 | lab_over = create_artificial_label (UNKNOWN_LOCATION); | |
514066a1 | 8123 | |
70f34814 | 8124 | valist = build_simple_mem_ref (addr); |
6d1cb95f | 8125 | |
514066a1 RH |
8126 | if (pass_as_float) |
8127 | { | |
a11c004b R |
8128 | tree next_fp_tmp = create_tmp_var (TREE_TYPE (f_next_fp), NULL); |
8129 | tree cmp; | |
11160b58 | 8130 | bool is_double = size == 8 && TREE_CODE (eff_type) == REAL_TYPE; |
f78ec36a | 8131 | |
8c54989a KK |
8132 | tmp = build1 (ADDR_EXPR, pptr_type_node, unshare_expr (next_fp)); |
8133 | gimplify_assign (unshare_expr (addr), tmp, pre_p); | |
a11c004b | 8134 | |
8c54989a | 8135 | gimplify_assign (unshare_expr (next_fp_tmp), valist, pre_p); |
a11c004b R |
8136 | tmp = next_fp_limit; |
8137 | if (size > 4 && !is_double) | |
8c54989a KK |
8138 | tmp = build2 (POINTER_PLUS_EXPR, TREE_TYPE (tmp), |
8139 | unshare_expr (tmp), size_int (4 - size)); | |
8140 | tmp = build2 (GE_EXPR, boolean_type_node, | |
8141 | unshare_expr (next_fp_tmp), unshare_expr (tmp)); | |
47a25a46 | 8142 | cmp = build3 (COND_EXPR, void_type_node, tmp, |
8c54989a KK |
8143 | build1 (GOTO_EXPR, void_type_node, |
8144 | unshare_expr (lab_false)), NULL_TREE); | |
a11c004b R |
8145 | if (!is_double) |
8146 | gimplify_and_add (cmp, pre_p); | |
514066a1 | 8147 | |
11160b58 R |
8148 | if (TYPE_ALIGN (eff_type) > BITS_PER_WORD |
8149 | || (is_double || size == 16)) | |
514066a1 | 8150 | { |
5be014d5 AP |
8151 | tmp = fold_convert (sizetype, next_fp_tmp); |
8152 | tmp = build2 (BIT_AND_EXPR, sizetype, tmp, | |
8153 | size_int (UNITS_PER_WORD)); | |
8154 | tmp = build2 (POINTER_PLUS_EXPR, ptr_type_node, | |
8c54989a KK |
8155 | unshare_expr (next_fp_tmp), tmp); |
8156 | gimplify_assign (unshare_expr (next_fp_tmp), tmp, pre_p); | |
514066a1 | 8157 | } |
a11c004b R |
8158 | if (is_double) |
8159 | gimplify_and_add (cmp, pre_p); | |
514066a1 | 8160 | |
6d1cb95f | 8161 | #ifdef FUNCTION_ARG_SCmode_WART |
11160b58 R |
8162 | if (TYPE_MODE (eff_type) == SCmode |
8163 | && TARGET_SH4 && TARGET_LITTLE_ENDIAN) | |
6d1cb95f | 8164 | { |
11160b58 | 8165 | tree subtype = TREE_TYPE (eff_type); |
5fac2ed2 | 8166 | tree real, imag; |
6d1cb95f | 8167 | |
a11c004b R |
8168 | imag |
8169 | = std_gimplify_va_arg_expr (next_fp_tmp, subtype, pre_p, NULL); | |
5fac2ed2 | 8170 | imag = get_initialized_tmp_var (imag, pre_p, NULL); |
6d1cb95f | 8171 | |
a11c004b R |
8172 | real |
8173 | = std_gimplify_va_arg_expr (next_fp_tmp, subtype, pre_p, NULL); | |
5fac2ed2 RH |
8174 | real = get_initialized_tmp_var (real, pre_p, NULL); |
8175 | ||
f2a945b1 KK |
8176 | result = build2 (COMPLEX_EXPR, eff_type, real, imag); |
8177 | if (type != eff_type) | |
8178 | result = build1 (VIEW_CONVERT_EXPR, type, result); | |
5fac2ed2 | 8179 | result = get_initialized_tmp_var (result, pre_p, NULL); |
6d1cb95f R |
8180 | } |
8181 | #endif /* FUNCTION_ARG_SCmode_WART */ | |
8182 | ||
8c54989a | 8183 | tmp = build1 (GOTO_EXPR, void_type_node, unshare_expr (lab_over)); |
5fac2ed2 RH |
8184 | gimplify_and_add (tmp, pre_p); |
8185 | ||
8c54989a | 8186 | tmp = build1 (LABEL_EXPR, void_type_node, unshare_expr (lab_false)); |
5fac2ed2 | 8187 | gimplify_and_add (tmp, pre_p); |
514066a1 | 8188 | |
8c54989a KK |
8189 | tmp = build1 (ADDR_EXPR, pptr_type_node, unshare_expr (next_stack)); |
8190 | gimplify_assign (unshare_expr (addr), tmp, pre_p); | |
8191 | gimplify_assign (unshare_expr (next_fp_tmp), | |
8192 | unshare_expr (valist), pre_p); | |
a11c004b | 8193 | |
8c54989a KK |
8194 | gimplify_assign (unshare_expr (valist), |
8195 | unshare_expr (next_fp_tmp), post_p); | |
a11c004b | 8196 | valist = next_fp_tmp; |
514066a1 RH |
8197 | } |
8198 | else | |
8199 | { | |
8c54989a KK |
8200 | tmp = build2 (POINTER_PLUS_EXPR, ptr_type_node, |
8201 | unshare_expr (next_o), size_int (rsize)); | |
8202 | tmp = build2 (GT_EXPR, boolean_type_node, tmp, | |
8203 | unshare_expr (next_o_limit)); | |
47a25a46 | 8204 | tmp = build3 (COND_EXPR, void_type_node, tmp, |
8c54989a KK |
8205 | build1 (GOTO_EXPR, void_type_node, |
8206 | unshare_expr (lab_false)), | |
8207 | NULL_TREE); | |
5fac2ed2 | 8208 | gimplify_and_add (tmp, pre_p); |
514066a1 | 8209 | |
8c54989a KK |
8210 | tmp = build1 (ADDR_EXPR, pptr_type_node, unshare_expr (next_o)); |
8211 | gimplify_assign (unshare_expr (addr), tmp, pre_p); | |
514066a1 | 8212 | |
8c54989a | 8213 | tmp = build1 (GOTO_EXPR, void_type_node, unshare_expr (lab_over)); |
5fac2ed2 RH |
8214 | gimplify_and_add (tmp, pre_p); |
8215 | ||
8c54989a | 8216 | tmp = build1 (LABEL_EXPR, void_type_node, unshare_expr (lab_false)); |
5fac2ed2 | 8217 | gimplify_and_add (tmp, pre_p); |
514066a1 | 8218 | |
bdf57c40 | 8219 | if (size > 4 && ! (TARGET_SH4 || TARGET_SH2A)) |
8c54989a KK |
8220 | gimplify_assign (unshare_expr (next_o), |
8221 | unshare_expr (next_o_limit), pre_p); | |
514066a1 | 8222 | |
8c54989a KK |
8223 | tmp = build1 (ADDR_EXPR, pptr_type_node, unshare_expr (next_stack)); |
8224 | gimplify_assign (unshare_expr (addr), tmp, pre_p); | |
514066a1 RH |
8225 | } |
8226 | ||
5fac2ed2 RH |
8227 | if (!result) |
8228 | { | |
8c54989a | 8229 | tmp = build1 (LABEL_EXPR, void_type_node, unshare_expr (lab_over)); |
5fac2ed2 RH |
8230 | gimplify_and_add (tmp, pre_p); |
8231 | } | |
514066a1 RH |
8232 | } |
8233 | ||
8234 | /* ??? In va-sh.h, there had been code to make values larger than | |
8235 | size 8 indirect. This does not match the FUNCTION_ARG macros. */ | |
8236 | ||
5fac2ed2 | 8237 | tmp = std_gimplify_va_arg_expr (valist, type, pre_p, NULL); |
6d1cb95f R |
8238 | if (result) |
8239 | { | |
726a989a | 8240 | gimplify_assign (result, tmp, pre_p); |
a1516d08 | 8241 | result = build1 (NOP_EXPR, TREE_TYPE (result), result); |
8c54989a | 8242 | tmp = build1 (LABEL_EXPR, void_type_node, unshare_expr (lab_over)); |
5fac2ed2 | 8243 | gimplify_and_add (tmp, pre_p); |
6d1cb95f R |
8244 | } |
8245 | else | |
5fac2ed2 | 8246 | result = tmp; |
6d1cb95f | 8247 | |
5e039e6b | 8248 | if (pass_by_ref) |
d6e9821f | 8249 | result = build_va_arg_indirect_ref (result); |
5fac2ed2 | 8250 | |
5e039e6b | 8251 | return result; |
514066a1 RH |
8252 | } |
8253 | ||
63a9de54 CB |
8254 | /* 64 bit floating points memory transfers are paired single precision loads |
8255 | or store. So DWARF information needs fixing in little endian (unless | |
8256 | PR=SZ=1 in FPSCR). */ | |
8257 | rtx | |
8258 | sh_dwarf_register_span (rtx reg) | |
8259 | { | |
8260 | unsigned regno = REGNO (reg); | |
8261 | ||
8262 | if (WORDS_BIG_ENDIAN || GET_MODE (reg) != DFmode) | |
8263 | return NULL_RTX; | |
8264 | ||
8265 | return | |
8266 | gen_rtx_PARALLEL (VOIDmode, | |
8267 | gen_rtvec (2, | |
8268 | gen_rtx_REG (SFmode, | |
8269 | DBX_REGISTER_NUMBER (regno+1)), | |
8270 | gen_rtx_REG (SFmode, | |
8271 | DBX_REGISTER_NUMBER (regno)))); | |
8272 | } | |
8273 | ||
6e957ac8 | 8274 | static enum machine_mode |
cde0f3fd | 8275 | sh_promote_function_mode (const_tree type, enum machine_mode mode, |
d8158ff1 | 8276 | int *punsignedp, const_tree funtype, |
8a575d5a | 8277 | int for_return) |
cde0f3fd PB |
8278 | { |
8279 | if (sh_promote_prototypes (funtype)) | |
d055ce45 | 8280 | return promote_mode (type, mode, punsignedp); |
cde0f3fd | 8281 | else |
8a575d5a PB |
8282 | return default_promote_function_mode (type, mode, punsignedp, funtype, |
8283 | for_return); | |
cde0f3fd PB |
8284 | } |
8285 | ||
d64728cf | 8286 | static bool |
586de218 | 8287 | sh_promote_prototypes (const_tree type) |
61f71b34 DD |
8288 | { |
8289 | if (TARGET_HITACHI) | |
8290 | return 0; | |
8291 | if (! type) | |
8292 | return 1; | |
8293 | return ! sh_attr_renesas_p (type); | |
8294 | } | |
8295 | ||
8cd5a4e0 RH |
8296 | /* Whether an argument must be passed by reference. On SHcompact, we |
8297 | pretend arguments wider than 32-bits that would have been passed in | |
8298 | registers are passed by reference, so that an SHmedia trampoline | |
8299 | loads them into the full 64-bits registers. */ | |
8300 | ||
8301 | static int | |
586de218 KG |
8302 | shcompact_byref (const CUMULATIVE_ARGS *cum, enum machine_mode mode, |
8303 | const_tree type, bool named) | |
8cd5a4e0 RH |
8304 | { |
8305 | unsigned HOST_WIDE_INT size; | |
8306 | ||
8307 | if (type) | |
8308 | size = int_size_in_bytes (type); | |
8309 | else | |
8310 | size = GET_MODE_SIZE (mode); | |
8311 | ||
8312 | if (cum->arg_count[SH_ARG_INT] < NPARM_REGS (SImode) | |
8313 | && (!named | |
8314 | || GET_SH_ARG_CLASS (mode) == SH_ARG_INT | |
8315 | || (GET_SH_ARG_CLASS (mode) == SH_ARG_FLOAT | |
8316 | && cum->arg_count[SH_ARG_FLOAT] >= NPARM_REGS (SFmode))) | |
8317 | && size > 4 | |
8318 | && !SHCOMPACT_FORCE_ON_STACK (mode, type) | |
8319 | && !SH5_WOULD_BE_PARTIAL_NREGS (*cum, mode, type, named)) | |
8320 | return size; | |
8321 | else | |
8322 | return 0; | |
8323 | } | |
8324 | ||
8325 | static bool | |
8326 | sh_pass_by_reference (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
586de218 | 8327 | const_tree type, bool named) |
8cd5a4e0 RH |
8328 | { |
8329 | if (targetm.calls.must_pass_in_stack (mode, type)) | |
8330 | return true; | |
8331 | ||
03603192 JR |
8332 | /* ??? std_gimplify_va_arg_expr passes NULL for cum. That function |
8333 | wants to know about pass-by-reference semantics for incoming | |
8334 | arguments. */ | |
8335 | if (! cum) | |
8336 | return false; | |
8337 | ||
8cd5a4e0 RH |
8338 | if (TARGET_SHCOMPACT) |
8339 | { | |
8340 | cum->byref = shcompact_byref (cum, mode, type, named); | |
8341 | return cum->byref != 0; | |
8342 | } | |
8343 | ||
8344 | return false; | |
8345 | } | |
8346 | ||
6cdd5672 RH |
8347 | static bool |
8348 | sh_callee_copies (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
586de218 | 8349 | const_tree type, bool named ATTRIBUTE_UNUSED) |
6cdd5672 RH |
8350 | { |
8351 | /* ??? How can it possibly be correct to return true only on the | |
8352 | caller side of the equation? Is there someplace else in the | |
8353 | sh backend that's magically producing the copies? */ | |
8354 | return (cum->outgoing | |
8355 | && ((mode == BLKmode ? TYPE_ALIGN (type) : GET_MODE_ALIGNMENT (mode)) | |
8356 | % SH_MIN_ALIGN_FOR_CALLEE_COPY == 0)); | |
8357 | } | |
8358 | ||
78a52f11 RH |
8359 | static int |
8360 | sh_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
8361 | tree type, bool named ATTRIBUTE_UNUSED) | |
8362 | { | |
8363 | int words = 0; | |
8364 | ||
8365 | if (!TARGET_SH5 | |
8366 | && PASS_IN_REG_P (*cum, mode, type) | |
8367 | && !(TARGET_SH4 || TARGET_SH2A_DOUBLE) | |
8368 | && (ROUND_REG (*cum, mode) | |
8369 | + (mode != BLKmode | |
8370 | ? ROUND_ADVANCE (GET_MODE_SIZE (mode)) | |
8371 | : ROUND_ADVANCE (int_size_in_bytes (type))) | |
8372 | > NPARM_REGS (mode))) | |
8373 | words = NPARM_REGS (mode) - ROUND_REG (*cum, mode); | |
8374 | ||
8375 | else if (!TARGET_SHCOMPACT | |
8376 | && SH5_WOULD_BE_PARTIAL_NREGS (*cum, mode, type, named)) | |
8377 | words = NPARM_REGS (SImode) - cum->arg_count[SH_ARG_INT]; | |
8378 | ||
8379 | return words * UNITS_PER_WORD; | |
8380 | } | |
8381 | ||
8382 | ||
61f71b34 DD |
8383 | /* Define where to put the arguments to a function. |
8384 | Value is zero to push the argument on the stack, | |
8385 | or a hard register in which to store the argument. | |
8386 | ||
8387 | MODE is the argument's machine mode. | |
8388 | TYPE is the data type of the argument (as a tree). | |
8389 | This is null for libcalls where that information may | |
8390 | not be available. | |
8391 | CUM is a variable of type CUMULATIVE_ARGS which gives info about | |
8392 | the preceding args and about the function being called. | |
8393 | NAMED is nonzero if this argument is a named parameter | |
8394 | (otherwise it is an extra parameter matching an ellipsis). | |
8395 | ||
8396 | On SH the first args are normally in registers | |
8397 | and the rest are pushed. Any arg that starts within the first | |
8398 | NPARM_REGS words is at least partially passed in a register unless | |
8399 | its data type forbids. */ | |
8400 | ||
b64336b8 | 8401 | static rtx |
cf277499 | 8402 | sh_function_arg (CUMULATIVE_ARGS *ca, enum machine_mode mode, |
b64336b8 | 8403 | const_tree type, bool named) |
61f71b34 DD |
8404 | { |
8405 | if (! TARGET_SH5 && mode == VOIDmode) | |
8406 | return GEN_INT (ca->renesas_abi ? 1 : 0); | |
8407 | ||
8408 | if (! TARGET_SH5 | |
8409 | && PASS_IN_REG_P (*ca, mode, type) | |
8410 | && (named || ! (TARGET_HITACHI || ca->renesas_abi))) | |
8411 | { | |
8412 | int regno; | |
8413 | ||
8414 | if (mode == SCmode && TARGET_SH4 && TARGET_LITTLE_ENDIAN | |
8415 | && (! FUNCTION_ARG_SCmode_WART || (ROUND_REG (*ca, mode) & 1))) | |
8416 | { | |
8417 | rtx r1 = gen_rtx_EXPR_LIST (VOIDmode, | |
8418 | gen_rtx_REG (SFmode, | |
8419 | BASE_ARG_REG (mode) | |
8420 | + (ROUND_REG (*ca, mode) ^ 1)), | |
8421 | const0_rtx); | |
832a3292 KH |
8422 | rtx r2 = gen_rtx_EXPR_LIST (VOIDmode, |
8423 | gen_rtx_REG (SFmode, | |
8424 | BASE_ARG_REG (mode) | |
8425 | + ((ROUND_REG (*ca, mode) + 1) ^ 1)), | |
8426 | GEN_INT (4)); | |
61f71b34 DD |
8427 | return gen_rtx_PARALLEL(SCmode, gen_rtvec(2, r1, r2)); |
8428 | } | |
8429 | ||
8430 | /* If the alignment of a DF value causes an SF register to be | |
8431 | skipped, we will use that skipped register for the next SF | |
8432 | value. */ | |
8433 | if ((TARGET_HITACHI || ca->renesas_abi) | |
8434 | && ca->free_single_fp_reg | |
8435 | && mode == SFmode) | |
8436 | return gen_rtx_REG (mode, ca->free_single_fp_reg); | |
8437 | ||
8438 | regno = (BASE_ARG_REG (mode) + ROUND_REG (*ca, mode)) | |
8439 | ^ (mode == SFmode && TARGET_SH4 | |
8440 | && TARGET_LITTLE_ENDIAN != 0 | |
8441 | && ! TARGET_HITACHI && ! ca->renesas_abi); | |
8442 | return gen_rtx_REG (mode, regno); | |
8443 | ||
8444 | } | |
f676971a | 8445 | |
61f71b34 DD |
8446 | if (TARGET_SH5) |
8447 | { | |
8448 | if (mode == VOIDmode && TARGET_SHCOMPACT) | |
8449 | return GEN_INT (ca->call_cookie); | |
8450 | ||
8451 | /* The following test assumes unnamed arguments are promoted to | |
8452 | DFmode. */ | |
8453 | if (mode == SFmode && ca->free_single_fp_reg) | |
8454 | return SH5_PROTOTYPED_FLOAT_ARG (*ca, mode, ca->free_single_fp_reg); | |
8455 | ||
8456 | if ((GET_SH_ARG_CLASS (mode) == SH_ARG_FLOAT) | |
8457 | && (named || ! ca->prototype_p) | |
8458 | && ca->arg_count[(int) SH_ARG_FLOAT] < NPARM_REGS (SFmode)) | |
8459 | { | |
8460 | if (! ca->prototype_p && TARGET_SHMEDIA) | |
8461 | return SH5_PROTOTYPELESS_FLOAT_ARG (*ca, mode); | |
8462 | ||
8463 | return SH5_PROTOTYPED_FLOAT_ARG (*ca, mode, | |
8464 | FIRST_FP_PARM_REG | |
8465 | + ca->arg_count[(int) SH_ARG_FLOAT]); | |
8466 | } | |
8467 | ||
8468 | if (ca->arg_count[(int) SH_ARG_INT] < NPARM_REGS (SImode) | |
8469 | && (! TARGET_SHCOMPACT | |
8470 | || (! SHCOMPACT_FORCE_ON_STACK (mode, type) | |
8471 | && ! SH5_WOULD_BE_PARTIAL_NREGS (*ca, mode, | |
8472 | type, named)))) | |
8473 | { | |
8474 | return gen_rtx_REG (mode, (FIRST_PARM_REG | |
8475 | + ca->arg_count[(int) SH_ARG_INT])); | |
8476 | } | |
8477 | ||
8478 | return 0; | |
8479 | } | |
8480 | ||
8481 | return 0; | |
8482 | } | |
f676971a | 8483 | |
61f71b34 DD |
8484 | /* Update the data in CUM to advance over an argument |
8485 | of mode MODE and data type TYPE. | |
8486 | (TYPE is null for libcalls where that information may not be | |
8487 | available.) */ | |
8488 | ||
b64336b8 | 8489 | static void |
cf277499 | 8490 | sh_function_arg_advance (CUMULATIVE_ARGS *ca, enum machine_mode mode, |
b64336b8 | 8491 | const_tree type, bool named) |
61f71b34 | 8492 | { |
832a3292 KH |
8493 | if (ca->force_mem) |
8494 | ca->force_mem = 0; | |
8495 | else if (TARGET_SH5) | |
8496 | { | |
b64336b8 NF |
8497 | const_tree type2 = (ca->byref && type |
8498 | ? TREE_TYPE (type) | |
8499 | : type); | |
832a3292 KH |
8500 | enum machine_mode mode2 = (ca->byref && type |
8501 | ? TYPE_MODE (type2) | |
8502 | : mode); | |
8503 | int dwords = ((ca->byref | |
8504 | ? ca->byref | |
8505 | : mode2 == BLKmode | |
8506 | ? int_size_in_bytes (type2) | |
8507 | : GET_MODE_SIZE (mode2)) + 7) / 8; | |
8508 | int numregs = MIN (dwords, NPARM_REGS (SImode) | |
8509 | - ca->arg_count[(int) SH_ARG_INT]); | |
8510 | ||
8511 | if (numregs) | |
8512 | { | |
8513 | ca->arg_count[(int) SH_ARG_INT] += numregs; | |
8514 | if (TARGET_SHCOMPACT | |
8515 | && SHCOMPACT_FORCE_ON_STACK (mode2, type2)) | |
8516 | { | |
8517 | ca->call_cookie | |
8518 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
8519 | - numregs, 1); | |
8520 | /* N.B. We want this also for outgoing. */ | |
8521 | ca->stack_regs += numregs; | |
8522 | } | |
8523 | else if (ca->byref) | |
8524 | { | |
8525 | if (! ca->outgoing) | |
8526 | ca->stack_regs += numregs; | |
8527 | ca->byref_regs += numregs; | |
8528 | ca->byref = 0; | |
8529 | do | |
8530 | ca->call_cookie | |
8531 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
8532 | - numregs, 2); | |
8533 | while (--numregs); | |
8534 | ca->call_cookie | |
8535 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
8536 | - 1, 1); | |
8537 | } | |
8538 | else if (dwords > numregs) | |
8539 | { | |
8540 | int pushregs = numregs; | |
8541 | ||
8542 | if (TARGET_SHCOMPACT) | |
8543 | ca->stack_regs += numregs; | |
8544 | while (pushregs < NPARM_REGS (SImode) - 1 | |
8545 | && (CALL_COOKIE_INT_REG_GET | |
8546 | (ca->call_cookie, | |
8547 | NPARM_REGS (SImode) - pushregs) | |
8548 | == 1)) | |
8549 | { | |
8550 | ca->call_cookie | |
8551 | &= ~ CALL_COOKIE_INT_REG (NPARM_REGS (SImode) | |
8552 | - pushregs, 1); | |
8553 | pushregs++; | |
8554 | } | |
8555 | if (numregs == NPARM_REGS (SImode)) | |
8556 | ca->call_cookie | |
8557 | |= CALL_COOKIE_INT_REG (0, 1) | |
8558 | | CALL_COOKIE_STACKSEQ (numregs - 1); | |
8559 | else | |
8560 | ca->call_cookie | |
8561 | |= CALL_COOKIE_STACKSEQ (numregs); | |
8562 | } | |
8563 | } | |
8564 | if (GET_SH_ARG_CLASS (mode2) == SH_ARG_FLOAT | |
8565 | && (named || ! ca->prototype_p)) | |
8566 | { | |
8567 | if (mode2 == SFmode && ca->free_single_fp_reg) | |
8568 | ca->free_single_fp_reg = 0; | |
8569 | else if (ca->arg_count[(int) SH_ARG_FLOAT] | |
8570 | < NPARM_REGS (SFmode)) | |
8571 | { | |
8572 | int numfpregs | |
8573 | = MIN ((GET_MODE_SIZE (mode2) + 7) / 8 * 2, | |
8574 | NPARM_REGS (SFmode) | |
8575 | - ca->arg_count[(int) SH_ARG_FLOAT]); | |
8576 | ||
8577 | ca->arg_count[(int) SH_ARG_FLOAT] += numfpregs; | |
8578 | ||
8579 | if (TARGET_SHCOMPACT && ! ca->prototype_p) | |
8580 | { | |
8581 | if (ca->outgoing && numregs > 0) | |
8582 | do | |
8583 | { | |
8584 | ca->call_cookie | |
8585 | |= (CALL_COOKIE_INT_REG | |
8586 | (ca->arg_count[(int) SH_ARG_INT] | |
8587 | - numregs + ((numfpregs - 2) / 2), | |
8588 | 4 + (ca->arg_count[(int) SH_ARG_FLOAT] | |
8589 | - numfpregs) / 2)); | |
8590 | } | |
8591 | while (numfpregs -= 2); | |
8592 | } | |
8593 | else if (mode2 == SFmode && (named) | |
8594 | && (ca->arg_count[(int) SH_ARG_FLOAT] | |
8595 | < NPARM_REGS (SFmode))) | |
8596 | ca->free_single_fp_reg | |
8597 | = FIRST_FP_PARM_REG - numfpregs | |
8598 | + ca->arg_count[(int) SH_ARG_FLOAT] + 1; | |
8599 | } | |
8600 | } | |
8601 | return; | |
8602 | } | |
8603 | ||
8604 | if ((TARGET_HITACHI || ca->renesas_abi) && TARGET_FPU_DOUBLE) | |
8605 | { | |
8606 | /* Note that we've used the skipped register. */ | |
8607 | if (mode == SFmode && ca->free_single_fp_reg) | |
8608 | { | |
8609 | ca->free_single_fp_reg = 0; | |
8610 | return; | |
8611 | } | |
8612 | /* When we have a DF after an SF, there's an SF register that get | |
8613 | skipped in order to align the DF value. We note this skipped | |
8614 | register, because the next SF value will use it, and not the | |
8615 | SF that follows the DF. */ | |
8616 | if (mode == DFmode | |
8617 | && ROUND_REG (*ca, DFmode) != ROUND_REG (*ca, SFmode)) | |
8618 | { | |
8619 | ca->free_single_fp_reg = (ROUND_REG (*ca, SFmode) | |
8620 | + BASE_ARG_REG (mode)); | |
8621 | } | |
8622 | } | |
8623 | ||
157371cf | 8624 | if (! ((TARGET_SH4 || TARGET_SH2A) || ca->renesas_abi) |
832a3292 KH |
8625 | || PASS_IN_REG_P (*ca, mode, type)) |
8626 | (ca->arg_count[(int) GET_SH_ARG_CLASS (mode)] | |
8627 | = (ROUND_REG (*ca, mode) | |
8628 | + (mode == BLKmode | |
8629 | ? ROUND_ADVANCE (int_size_in_bytes (type)) | |
8630 | : ROUND_ADVANCE (GET_MODE_SIZE (mode))))); | |
61f71b34 DD |
8631 | } |
8632 | ||
61f71b34 DD |
8633 | /* The Renesas calling convention doesn't quite fit into this scheme since |
8634 | the address is passed like an invisible argument, but one that is always | |
8635 | passed in memory. */ | |
8636 | static rtx | |
cf277499 | 8637 | sh_struct_value_rtx (tree fndecl, int incoming ATTRIBUTE_UNUSED) |
61f71b34 DD |
8638 | { |
8639 | if (TARGET_HITACHI || sh_attr_renesas_p (fndecl)) | |
8640 | return 0; | |
8641 | return gen_rtx_REG (Pmode, 2); | |
8642 | } | |
8643 | ||
d64728cf AS |
8644 | /* Worker function for TARGET_FUNCTION_VALUE. |
8645 | ||
8646 | For the SH, this is like LIBCALL_VALUE, except that we must change the | |
8647 | mode like PROMOTE_MODE does. | |
8648 | ??? PROMOTE_MODE is ignored for non-scalar types. The set of types | |
8649 | tested here has to be kept in sync with the one in explow.c:promote_mode. | |
8650 | */ | |
8651 | ||
8652 | static rtx | |
8653 | sh_function_value (const_tree valtype, | |
8654 | const_tree fn_decl_or_type, | |
8655 | bool outgoing ATTRIBUTE_UNUSED) | |
8656 | { | |
8657 | if (fn_decl_or_type | |
8658 | && !DECL_P (fn_decl_or_type)) | |
8659 | fn_decl_or_type = NULL; | |
8660 | ||
8661 | return gen_rtx_REG ( | |
8662 | ((GET_MODE_CLASS (TYPE_MODE (valtype)) == MODE_INT | |
8663 | && GET_MODE_SIZE (TYPE_MODE (valtype)) < 4 | |
8664 | && (TREE_CODE (valtype) == INTEGER_TYPE | |
8665 | || TREE_CODE (valtype) == ENUMERAL_TYPE | |
8666 | || TREE_CODE (valtype) == BOOLEAN_TYPE | |
8667 | || TREE_CODE (valtype) == REAL_TYPE | |
8668 | || TREE_CODE (valtype) == OFFSET_TYPE)) | |
8669 | && sh_promote_prototypes (fn_decl_or_type) | |
8670 | ? (TARGET_SHMEDIA64 ? DImode : SImode) : TYPE_MODE (valtype)), | |
8671 | BASE_RETURN_VALUE_REG (TYPE_MODE (valtype))); | |
8672 | } | |
8673 | ||
8674 | /* Worker function for TARGET_LIBCALL_VALUE. */ | |
8675 | ||
8676 | static rtx | |
8677 | sh_libcall_value (enum machine_mode mode, const_rtx fun ATTRIBUTE_UNUSED) | |
8678 | { | |
8679 | return gen_rtx_REG (mode, BASE_RETURN_VALUE_REG (mode)); | |
8680 | } | |
8681 | ||
e7a8f681 | 8682 | /* Return true if N is a possible register number of function value. */ |
d64728cf | 8683 | |
e7a8f681 | 8684 | static bool |
d64728cf AS |
8685 | sh_function_value_regno_p (const unsigned int regno) |
8686 | { | |
8687 | return ((regno) == FIRST_RET_REG | |
8688 | || (TARGET_SH2E && (regno) == FIRST_FP_RET_REG) | |
8689 | || (TARGET_SHMEDIA_FPU && (regno) == FIRST_FP_RET_REG)); | |
8690 | } | |
8691 | ||
bd5bd7ac KH |
8692 | /* Worker function for TARGET_RETURN_IN_MEMORY. */ |
8693 | ||
61f71b34 | 8694 | static bool |
586de218 | 8695 | sh_return_in_memory (const_tree type, const_tree fndecl) |
61f71b34 DD |
8696 | { |
8697 | if (TARGET_SH5) | |
8698 | { | |
8699 | if (TYPE_MODE (type) == BLKmode) | |
8700 | return ((unsigned HOST_WIDE_INT) int_size_in_bytes (type)) > 8; | |
8701 | else | |
8702 | return GET_MODE_SIZE (TYPE_MODE (type)) > 8; | |
8703 | } | |
8704 | else | |
8705 | { | |
8706 | return (TYPE_MODE (type) == BLKmode | |
8707 | || ((TARGET_HITACHI || sh_attr_renesas_p (fndecl)) | |
8708 | && TREE_CODE (type) == RECORD_TYPE)); | |
8709 | } | |
8710 | } | |
8711 | ||
8712 | /* We actually emit the code in sh_expand_prologue. We used to use | |
8713 | a static variable to flag that we need to emit this code, but that | |
8714 | doesn't when inlining, when functions are deferred and then emitted | |
8715 | later. Fortunately, we already have two flags that are part of struct | |
8716 | function that tell if a function uses varargs or stdarg. */ | |
8717 | static void | |
49616835 JR |
8718 | sh_setup_incoming_varargs (CUMULATIVE_ARGS *ca, |
8719 | enum machine_mode mode, | |
8720 | tree type, | |
8721 | int *pretend_arg_size, | |
cf277499 | 8722 | int second_time ATTRIBUTE_UNUSED) |
61f71b34 | 8723 | { |
e3b5732b | 8724 | gcc_assert (cfun->stdarg); |
49616835 JR |
8725 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl)) |
8726 | { | |
8727 | int named_parm_regs, anon_parm_regs; | |
8728 | ||
8729 | named_parm_regs = (ROUND_REG (*ca, mode) | |
8730 | + (mode == BLKmode | |
8731 | ? ROUND_ADVANCE (int_size_in_bytes (type)) | |
8732 | : ROUND_ADVANCE (GET_MODE_SIZE (mode)))); | |
8733 | anon_parm_regs = NPARM_REGS (SImode) - named_parm_regs; | |
8734 | if (anon_parm_regs > 0) | |
8735 | *pretend_arg_size = anon_parm_regs * 4; | |
8736 | } | |
61f71b34 DD |
8737 | } |
8738 | ||
8739 | static bool | |
cf277499 | 8740 | sh_strict_argument_naming (CUMULATIVE_ARGS *ca ATTRIBUTE_UNUSED) |
61f71b34 DD |
8741 | { |
8742 | return TARGET_SH5; | |
8743 | } | |
8744 | ||
8745 | static bool | |
cf277499 | 8746 | sh_pretend_outgoing_varargs_named (CUMULATIVE_ARGS *ca) |
61f71b34 DD |
8747 | { |
8748 | return ! (TARGET_HITACHI || ca->renesas_abi) && ! TARGET_SH5; | |
8749 | } | |
8750 | ||
8751 | ||
0d7e008e SC |
8752 | /* Define the offset between two registers, one to be eliminated, and |
8753 | the other its replacement, at the start of a routine. */ | |
8754 | ||
8755 | int | |
cf277499 | 8756 | initial_elimination_offset (int from, int to) |
0d7e008e SC |
8757 | { |
8758 | int regs_saved; | |
fa5322fa | 8759 | int regs_saved_rounding = 0; |
0d7e008e | 8760 | int total_saved_regs_space; |
e5f8d4c0 | 8761 | int total_auto_space; |
1245df60 | 8762 | int save_flags = target_flags; |
fa5322fa | 8763 | int copy_flags; |
490dbac7 | 8764 | HARD_REG_SET live_regs_mask; |
fe3ad572 SC |
8765 | |
8766 | shmedia_space_reserved_for_target_registers = false; | |
490dbac7 | 8767 | regs_saved = calc_live_regs (&live_regs_mask); |
fa5322fa | 8768 | regs_saved += SHMEDIA_REGS_STACK_ADJUST (); |
fe3ad572 SC |
8769 | |
8770 | if (shmedia_reserve_space_for_target_registers_p (regs_saved, &live_regs_mask)) | |
8771 | { | |
8772 | shmedia_space_reserved_for_target_registers = true; | |
8773 | regs_saved += shmedia_target_regs_stack_adjust (&live_regs_mask); | |
8774 | } | |
8775 | ||
fa5322fa AO |
8776 | if (TARGET_SH5 && regs_saved % (STACK_BOUNDARY / BITS_PER_UNIT)) |
8777 | regs_saved_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) | |
8778 | - regs_saved % (STACK_BOUNDARY / BITS_PER_UNIT)); | |
8e87e161 | 8779 | |
fa5322fa AO |
8780 | total_auto_space = rounded_frame_size (regs_saved) - regs_saved_rounding; |
8781 | copy_flags = target_flags; | |
1245df60 | 8782 | target_flags = save_flags; |
1a95a963 | 8783 | |
fa5322fa | 8784 | total_saved_regs_space = regs_saved + regs_saved_rounding; |
b9654711 | 8785 | |
96a2347e | 8786 | if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) |
fa5322fa | 8787 | return total_saved_regs_space + total_auto_space |
38173d38 | 8788 | + crtl->args.info.byref_regs * 8; |
8aa2a305 | 8789 | |
0d7e008e | 8790 | if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
fa5322fa | 8791 | return total_saved_regs_space + total_auto_space |
38173d38 | 8792 | + crtl->args.info.byref_regs * 8; |
8aa2a305 JW |
8793 | |
8794 | /* Initial gap between fp and sp is 0. */ | |
96a2347e | 8795 | if (from == HARD_FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
8aa2a305 JW |
8796 | return 0; |
8797 | ||
96a2347e KK |
8798 | if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
8799 | return rounded_frame_size (0); | |
8800 | ||
8801 | if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
8802 | return rounded_frame_size (0); | |
8803 | ||
f5b9e7c9 | 8804 | gcc_assert (from == RETURN_ADDRESS_POINTER_REGNUM |
96a2347e KK |
8805 | && (to == HARD_FRAME_POINTER_REGNUM |
8806 | || to == STACK_POINTER_REGNUM)); | |
f5b9e7c9 | 8807 | if (TARGET_SH5) |
d042370e | 8808 | { |
f5b9e7c9 NS |
8809 | int n = total_saved_regs_space; |
8810 | int pr_reg = TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG; | |
8811 | save_schedule schedule; | |
8812 | save_entry *entry; | |
8813 | ||
8814 | n += total_auto_space; | |
8815 | ||
8816 | /* If it wasn't saved, there's not much we can do. */ | |
8817 | if (! TEST_HARD_REG_BIT (live_regs_mask, pr_reg)) | |
8818 | return n; | |
8819 | ||
8820 | target_flags = copy_flags; | |
8821 | ||
8822 | sh5_schedule_saves (&live_regs_mask, &schedule, n); | |
8823 | for (entry = &schedule.entries[1]; entry->mode != VOIDmode; entry++) | |
8824 | if (entry->reg == pr_reg) | |
8825 | { | |
8826 | target_flags = save_flags; | |
8827 | return entry->offset; | |
8828 | } | |
8829 | gcc_unreachable (); | |
d042370e | 8830 | } |
f5b9e7c9 NS |
8831 | else |
8832 | return total_auto_space; | |
0d7e008e | 8833 | } |
3217af3e KK |
8834 | |
8835 | /* Parse the -mfixed-range= option string. */ | |
8836 | void | |
8837 | sh_fix_range (const char *const_str) | |
8838 | { | |
8839 | int i, first, last; | |
8840 | char *str, *dash, *comma; | |
8841 | ||
8842 | /* str must be of the form REG1'-'REG2{,REG1'-'REG} where REG1 and | |
8843 | REG2 are either register names or register numbers. The effect | |
8844 | of this option is to mark the registers in the range from REG1 to | |
8845 | REG2 as ``fixed'' so they won't be used by the compiler. */ | |
8846 | ||
8847 | i = strlen (const_str); | |
8848 | str = (char *) alloca (i + 1); | |
8849 | memcpy (str, const_str, i + 1); | |
8850 | ||
8851 | while (1) | |
8852 | { | |
8853 | dash = strchr (str, '-'); | |
8854 | if (!dash) | |
8855 | { | |
8856 | warning (0, "value of -mfixed-range must have form REG1-REG2"); | |
8857 | return; | |
8858 | } | |
8859 | *dash = '\0'; | |
8860 | comma = strchr (dash + 1, ','); | |
8861 | if (comma) | |
8862 | *comma = '\0'; | |
8863 | ||
8864 | first = decode_reg_name (str); | |
8865 | if (first < 0) | |
8866 | { | |
8867 | warning (0, "unknown register name: %s", str); | |
8868 | return; | |
8869 | } | |
8870 | ||
8871 | last = decode_reg_name (dash + 1); | |
8872 | if (last < 0) | |
8873 | { | |
8874 | warning (0, "unknown register name: %s", dash + 1); | |
8875 | return; | |
8876 | } | |
8877 | ||
8878 | *dash = '-'; | |
8879 | ||
8880 | if (first > last) | |
8881 | { | |
8882 | warning (0, "%s-%s is an empty range", str, dash + 1); | |
8883 | return; | |
8884 | } | |
8885 | ||
8886 | for (i = first; i <= last; ++i) | |
8887 | fixed_regs[i] = call_used_regs[i] = 1; | |
8888 | ||
8889 | if (!comma) | |
8890 | break; | |
8891 | ||
8892 | *comma = ','; | |
8893 | str = comma + 1; | |
8894 | } | |
8895 | } | |
8aa2a305 | 8896 | \f |
a6ab9fc0 | 8897 | /* Insert any deferred function attributes from earlier pragmas. */ |
12a68f1f | 8898 | static void |
cf277499 | 8899 | sh_insert_attributes (tree node, tree *attributes) |
157131d7 | 8900 | { |
a6ab9fc0 R |
8901 | tree attrs; |
8902 | ||
8903 | if (TREE_CODE (node) != FUNCTION_DECL) | |
157131d7 SC |
8904 | return; |
8905 | ||
8906 | /* We are only interested in fields. */ | |
6615c446 | 8907 | if (!DECL_P (node)) |
157131d7 SC |
8908 | return; |
8909 | ||
a6ab9fc0 R |
8910 | /* Append the attributes to the deferred attributes. */ |
8911 | *sh_deferred_function_attributes_tail = *attributes; | |
8912 | attrs = sh_deferred_function_attributes; | |
8913 | if (!attrs) | |
8914 | return; | |
8915 | ||
8916 | /* Some attributes imply or require the interrupt attribute. */ | |
8917 | if (!lookup_attribute ("interrupt_handler", attrs) | |
8918 | && !lookup_attribute ("interrupt_handler", DECL_ATTRIBUTES (node))) | |
8919 | { | |
8920 | /* If we have a trapa_handler, but no interrupt_handler attribute, | |
8921 | insert an interrupt_handler attribute. */ | |
8922 | if (lookup_attribute ("trapa_handler", attrs) != NULL_TREE) | |
8923 | /* We can't use sh_pr_interrupt here because that's not in the | |
8924 | java frontend. */ | |
8925 | attrs | |
8926 | = tree_cons (get_identifier("interrupt_handler"), NULL_TREE, attrs); | |
561642fa AP |
8927 | /* However, for sp_switch, trap_exit, nosave_low_regs and resbank, |
8928 | if the interrupt attribute is missing, we ignore the attribute | |
8929 | and warn. */ | |
a6ab9fc0 R |
8930 | else if (lookup_attribute ("sp_switch", attrs) |
8931 | || lookup_attribute ("trap_exit", attrs) | |
561642fa AP |
8932 | || lookup_attribute ("nosave_low_regs", attrs) |
8933 | || lookup_attribute ("resbank", attrs)) | |
a6ab9fc0 R |
8934 | { |
8935 | tree *tail; | |
8936 | ||
8937 | for (tail = attributes; attrs; attrs = TREE_CHAIN (attrs)) | |
8938 | { | |
8939 | if (is_attribute_p ("sp_switch", TREE_PURPOSE (attrs)) | |
8940 | || is_attribute_p ("trap_exit", TREE_PURPOSE (attrs)) | |
561642fa AP |
8941 | || is_attribute_p ("nosave_low_regs", TREE_PURPOSE (attrs)) |
8942 | || is_attribute_p ("resbank", TREE_PURPOSE (attrs))) | |
a6ab9fc0 | 8943 | warning (OPT_Wattributes, |
29d08eba JM |
8944 | "%qE attribute only applies to interrupt functions", |
8945 | TREE_PURPOSE (attrs)); | |
a6ab9fc0 R |
8946 | else |
8947 | { | |
8948 | *tail = tree_cons (TREE_PURPOSE (attrs), NULL_TREE, | |
8949 | NULL_TREE); | |
8950 | tail = &TREE_CHAIN (*tail); | |
8951 | } | |
8952 | } | |
8953 | attrs = *attributes; | |
8954 | } | |
8955 | } | |
8956 | ||
8957 | /* Install the processed list. */ | |
8958 | *attributes = attrs; | |
8959 | ||
8960 | /* Clear deferred attributes. */ | |
8961 | sh_deferred_function_attributes = NULL_TREE; | |
8962 | sh_deferred_function_attributes_tail = &sh_deferred_function_attributes; | |
157131d7 SC |
8963 | |
8964 | return; | |
8965 | } | |
8966 | ||
91d231cb | 8967 | /* Supported attributes: |
4408efce JL |
8968 | |
8969 | interrupt_handler -- specifies this function is an interrupt handler. | |
8970 | ||
a6ab9fc0 R |
8971 | trapa_handler - like above, but don't save all registers. |
8972 | ||
4408efce JL |
8973 | sp_switch -- specifies an alternate stack for an interrupt handler |
8974 | to run on. | |
8975 | ||
956d6950 | 8976 | trap_exit -- use a trapa to exit an interrupt function instead of |
61f71b34 DD |
8977 | an rte instruction. |
8978 | ||
a6ab9fc0 R |
8979 | nosave_low_regs - don't save r0..r7 in an interrupt handler. |
8980 | This is useful on the SH3 and upwards, | |
8981 | which has a separate set of low regs for User and Supervisor modes. | |
8982 | This should only be used for the lowest level of interrupts. Higher levels | |
8983 | of interrupts must save the registers in case they themselves are | |
8984 | interrupted. | |
8985 | ||
61f71b34 DD |
8986 | renesas -- use Renesas calling/layout conventions (functions and |
8987 | structures). | |
8988 | ||
561642fa AP |
8989 | resbank -- In case of an ISR, use a register bank to save registers |
8990 | R0-R14, MACH, MACL, GBR and PR. This is useful only on SH2A targets. | |
61f71b34 | 8991 | */ |
4408efce | 8992 | |
561642fa AP |
8993 | /* Handle a 'resbank' attribute. */ |
8994 | static tree | |
8995 | sh_handle_resbank_handler_attribute (tree * node, tree name, | |
8996 | tree args ATTRIBUTE_UNUSED, | |
8997 | int flags ATTRIBUTE_UNUSED, | |
8998 | bool * no_add_attrs) | |
8999 | { | |
9000 | if (!TARGET_SH2A) | |
9001 | { | |
29d08eba JM |
9002 | warning (OPT_Wattributes, "%qE attribute is supported only for SH2A", |
9003 | name); | |
561642fa AP |
9004 | *no_add_attrs = true; |
9005 | } | |
9006 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
9007 | { | |
29d08eba JM |
9008 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
9009 | name); | |
561642fa AP |
9010 | *no_add_attrs = true; |
9011 | } | |
9012 | ||
9013 | return NULL_TREE; | |
9014 | } | |
9015 | ||
91d231cb JM |
9016 | /* Handle an "interrupt_handler" attribute; arguments as in |
9017 | struct attribute_spec.handler. */ | |
9018 | static tree | |
cf277499 | 9019 | sh_handle_interrupt_handler_attribute (tree *node, tree name, |
561642fa AP |
9020 | tree args ATTRIBUTE_UNUSED, |
9021 | int flags ATTRIBUTE_UNUSED, | |
9022 | bool *no_add_attrs) | |
91d231cb JM |
9023 | { |
9024 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
4408efce | 9025 | { |
29d08eba JM |
9026 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
9027 | name); | |
91d231cb | 9028 | *no_add_attrs = true; |
4408efce | 9029 | } |
58ab7171 R |
9030 | else if (TARGET_SHCOMPACT) |
9031 | { | |
9032 | error ("attribute interrupt_handler is not compatible with -m5-compact"); | |
9033 | *no_add_attrs = true; | |
9034 | } | |
4408efce | 9035 | |
91d231cb JM |
9036 | return NULL_TREE; |
9037 | } | |
9038 | ||
561642fa AP |
9039 | /* Handle an 'function_vector' attribute; arguments as in |
9040 | struct attribute_spec.handler. */ | |
9041 | static tree | |
9042 | sh2a_handle_function_vector_handler_attribute (tree * node, tree name, | |
9043 | tree args ATTRIBUTE_UNUSED, | |
9044 | int flags ATTRIBUTE_UNUSED, | |
9045 | bool * no_add_attrs) | |
9046 | { | |
9047 | if (!TARGET_SH2A) | |
9048 | { | |
29d08eba JM |
9049 | warning (OPT_Wattributes, "%qE attribute only applies to SH2A", |
9050 | name); | |
561642fa AP |
9051 | *no_add_attrs = true; |
9052 | } | |
9053 | else if (TREE_CODE (*node) != FUNCTION_DECL) | |
9054 | { | |
29d08eba JM |
9055 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
9056 | name); | |
561642fa AP |
9057 | *no_add_attrs = true; |
9058 | } | |
9059 | else if (TREE_CODE (TREE_VALUE (args)) != INTEGER_CST) | |
9060 | { | |
9061 | /* The argument must be a constant integer. */ | |
9062 | warning (OPT_Wattributes, | |
29d08eba JM |
9063 | "%qE attribute argument not an integer constant", |
9064 | name); | |
561642fa AP |
9065 | *no_add_attrs = true; |
9066 | } | |
9067 | else if (TREE_INT_CST_LOW (TREE_VALUE (args)) > 255) | |
9068 | { | |
9069 | /* The argument value must be between 0 to 255. */ | |
9070 | warning (OPT_Wattributes, | |
29d08eba JM |
9071 | "%qE attribute argument should be between 0 to 255", |
9072 | name); | |
561642fa AP |
9073 | *no_add_attrs = true; |
9074 | } | |
9075 | return NULL_TREE; | |
9076 | } | |
9077 | ||
9078 | /* Returns 1 if current function has been assigned the attribute | |
9079 | 'function_vector'. */ | |
9080 | int | |
9081 | sh2a_is_function_vector_call (rtx x) | |
9082 | { | |
9083 | if (GET_CODE (x) == SYMBOL_REF | |
9084 | && (SYMBOL_REF_FLAGS (x) & SYMBOL_FLAG_FUNCVEC_FUNCTION)) | |
9085 | { | |
9086 | tree tr = SYMBOL_REF_DECL (x); | |
9087 | ||
9088 | if (sh2a_function_vector_p (tr)) | |
9089 | return 1; | |
9090 | } | |
9091 | ||
9092 | return 0; | |
9093 | } | |
9094 | ||
9095 | /* Returns the function vector number, if the the attribute | |
9096 | 'function_vector' is assigned, otherwise returns zero. */ | |
9097 | int | |
9098 | sh2a_get_function_vector_number (rtx x) | |
9099 | { | |
9100 | int num; | |
9101 | tree list, t; | |
9102 | ||
9103 | if ((GET_CODE (x) == SYMBOL_REF) | |
9104 | && (SYMBOL_REF_FLAGS (x) & SYMBOL_FLAG_FUNCVEC_FUNCTION)) | |
9105 | { | |
9106 | t = SYMBOL_REF_DECL (x); | |
9107 | ||
9108 | if (TREE_CODE (t) != FUNCTION_DECL) | |
9109 | return 0; | |
9110 | ||
9111 | list = SH_ATTRIBUTES (t); | |
9112 | while (list) | |
9113 | { | |
9114 | if (is_attribute_p ("function_vector", TREE_PURPOSE (list))) | |
9115 | { | |
9116 | num = TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (list))); | |
9117 | return num; | |
9118 | } | |
9119 | ||
9120 | list = TREE_CHAIN (list); | |
9121 | } | |
9122 | ||
9123 | return 0; | |
9124 | } | |
9125 | else | |
9126 | return 0; | |
9127 | } | |
9128 | ||
91d231cb JM |
9129 | /* Handle an "sp_switch" attribute; arguments as in |
9130 | struct attribute_spec.handler. */ | |
9131 | static tree | |
cf277499 SB |
9132 | sh_handle_sp_switch_attribute (tree *node, tree name, tree args, |
9133 | int flags ATTRIBUTE_UNUSED, bool *no_add_attrs) | |
91d231cb JM |
9134 | { |
9135 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
9136 | { | |
29d08eba JM |
9137 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
9138 | name); | |
91d231cb JM |
9139 | *no_add_attrs = true; |
9140 | } | |
91d231cb JM |
9141 | else if (TREE_CODE (TREE_VALUE (args)) != STRING_CST) |
9142 | { | |
4408efce | 9143 | /* The argument must be a constant string. */ |
29d08eba JM |
9144 | warning (OPT_Wattributes, "%qE attribute argument not a string constant", |
9145 | name); | |
91d231cb JM |
9146 | *no_add_attrs = true; |
9147 | } | |
4408efce | 9148 | |
91d231cb JM |
9149 | return NULL_TREE; |
9150 | } | |
9151 | ||
9152 | /* Handle an "trap_exit" attribute; arguments as in | |
9153 | struct attribute_spec.handler. */ | |
9154 | static tree | |
cf277499 SB |
9155 | sh_handle_trap_exit_attribute (tree *node, tree name, tree args, |
9156 | int flags ATTRIBUTE_UNUSED, bool *no_add_attrs) | |
91d231cb JM |
9157 | { |
9158 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
9159 | { | |
29d08eba JM |
9160 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
9161 | name); | |
91d231cb JM |
9162 | *no_add_attrs = true; |
9163 | } | |
a6ab9fc0 R |
9164 | /* The argument specifies a trap number to be used in a trapa instruction |
9165 | at function exit (instead of an rte instruction). */ | |
91d231cb JM |
9166 | else if (TREE_CODE (TREE_VALUE (args)) != INTEGER_CST) |
9167 | { | |
4408efce | 9168 | /* The argument must be a constant integer. */ |
29d08eba JM |
9169 | warning (OPT_Wattributes, "%qE attribute argument not an " |
9170 | "integer constant", name); | |
91d231cb JM |
9171 | *no_add_attrs = true; |
9172 | } | |
318881c0 | 9173 | |
91d231cb | 9174 | return NULL_TREE; |
4408efce JL |
9175 | } |
9176 | ||
61f71b34 | 9177 | static tree |
cf277499 SB |
9178 | sh_handle_renesas_attribute (tree *node ATTRIBUTE_UNUSED, |
9179 | tree name ATTRIBUTE_UNUSED, | |
9180 | tree args ATTRIBUTE_UNUSED, | |
9181 | int flags ATTRIBUTE_UNUSED, | |
9182 | bool *no_add_attrs ATTRIBUTE_UNUSED) | |
61f71b34 DD |
9183 | { |
9184 | return NULL_TREE; | |
9185 | } | |
9186 | ||
9187 | /* True if __attribute__((renesas)) or -mrenesas. */ | |
9188 | int | |
586de218 | 9189 | sh_attr_renesas_p (const_tree td) |
61f71b34 DD |
9190 | { |
9191 | if (TARGET_HITACHI) | |
9192 | return 1; | |
9193 | if (td == 0) | |
9194 | return 0; | |
9195 | if (DECL_P (td)) | |
9196 | td = TREE_TYPE (td); | |
a11c004b R |
9197 | if (td == error_mark_node) |
9198 | return 0; | |
61f71b34 DD |
9199 | return (lookup_attribute ("renesas", TYPE_ATTRIBUTES (td)) |
9200 | != NULL_TREE); | |
9201 | } | |
9202 | ||
9203 | /* True if __attribute__((renesas)) or -mrenesas, for the current | |
9204 | function. */ | |
9205 | int | |
cf277499 | 9206 | sh_cfun_attr_renesas_p (void) |
61f71b34 DD |
9207 | { |
9208 | return sh_attr_renesas_p (current_function_decl); | |
9209 | } | |
9210 | ||
e3ba8d11 | 9211 | int |
cf277499 | 9212 | sh_cfun_interrupt_handler_p (void) |
e3ba8d11 R |
9213 | { |
9214 | return (lookup_attribute ("interrupt_handler", | |
9215 | DECL_ATTRIBUTES (current_function_decl)) | |
9216 | != NULL_TREE); | |
9217 | } | |
bcc8cc82 | 9218 | |
561642fa AP |
9219 | /* Returns 1 if FUNC has been assigned the attribute |
9220 | "function_vector". */ | |
9221 | int | |
9222 | sh2a_function_vector_p (tree func) | |
9223 | { | |
9224 | tree list; | |
9225 | if (TREE_CODE (func) != FUNCTION_DECL) | |
9226 | return 0; | |
9227 | ||
9228 | list = SH_ATTRIBUTES (func); | |
9229 | while (list) | |
9230 | { | |
9231 | if (is_attribute_p ("function_vector", TREE_PURPOSE (list))) | |
9232 | return 1; | |
9233 | ||
9234 | list = TREE_CHAIN (list); | |
9235 | } | |
9236 | return 0; | |
9237 | } | |
9238 | ||
9239 | /* Returns TRUE if given tree has the "resbank" attribute. */ | |
9240 | ||
9241 | int | |
9242 | sh_cfun_resbank_handler_p (void) | |
9243 | { | |
9244 | return ((lookup_attribute ("resbank", | |
9245 | DECL_ATTRIBUTES (current_function_decl)) | |
9246 | != NULL_TREE) | |
9247 | && (lookup_attribute ("interrupt_handler", | |
9248 | DECL_ATTRIBUTES (current_function_decl)) | |
9249 | != NULL_TREE) && TARGET_SH2A); | |
9250 | } | |
9251 | ||
8d932be3 | 9252 | /* Implement TARGET_CHECK_PCH_TARGET_FLAGS. */ |
73a4d10b | 9253 | |
8d932be3 RS |
9254 | static const char * |
9255 | sh_check_pch_target_flags (int old_flags) | |
9256 | { | |
c0fb94d7 RS |
9257 | if ((old_flags ^ target_flags) & (MASK_SH1 | MASK_SH2 | MASK_SH3 |
9258 | | MASK_SH_E | MASK_HARD_SH4 | |
9259 | | MASK_FPU_SINGLE | MASK_SH4)) | |
9260 | return _("created and used with different architectures / ABIs"); | |
9261 | if ((old_flags ^ target_flags) & MASK_HITACHI) | |
9262 | return _("created and used with different ABIs"); | |
9263 | if ((old_flags ^ target_flags) & MASK_LITTLE_ENDIAN) | |
9264 | return _("created and used with different endianness"); | |
bcc8cc82 | 9265 | return NULL; |
bcc8cc82 | 9266 | } |
0d7e008e | 9267 | \f |
8aa2a305 | 9268 | /* Predicates used by the templates. */ |
0d7e008e | 9269 | |
8aa2a305 JW |
9270 | /* Returns 1 if OP is MACL, MACH or PR. The input must be a REG rtx. |
9271 | Used only in general_movsrc_operand. */ | |
0d7e008e | 9272 | |
8aa2a305 | 9273 | int |
cf277499 | 9274 | system_reg_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
0d7e008e | 9275 | { |
8aa2a305 | 9276 | switch (REGNO (op)) |
0d7e008e | 9277 | { |
8aa2a305 JW |
9278 | case PR_REG: |
9279 | case MACL_REG: | |
9280 | case MACH_REG: | |
9281 | return 1; | |
0d7e008e | 9282 | } |
8aa2a305 | 9283 | return 0; |
0d7e008e | 9284 | } |
0d7e008e | 9285 | |
1a95a963 JW |
9286 | /* Nonzero if OP is a floating point value with value 0.0. */ |
9287 | ||
9288 | int | |
cf277499 | 9289 | fp_zero_operand (rtx op) |
d3ae8277 | 9290 | { |
1a95a963 | 9291 | REAL_VALUE_TYPE r; |
d3ae8277 | 9292 | |
1a95a963 JW |
9293 | if (GET_MODE (op) != SFmode) |
9294 | return 0; | |
9295 | ||
9296 | REAL_VALUE_FROM_CONST_DOUBLE (r, op); | |
e4fa6b06 | 9297 | return REAL_VALUES_EQUAL (r, dconst0) && ! REAL_VALUE_MINUS_ZERO (r); |
d3ae8277 SC |
9298 | } |
9299 | ||
1a95a963 | 9300 | /* Nonzero if OP is a floating point value with value 1.0. */ |
d3ae8277 SC |
9301 | |
9302 | int | |
cf277499 | 9303 | fp_one_operand (rtx op) |
d3ae8277 | 9304 | { |
1a95a963 JW |
9305 | REAL_VALUE_TYPE r; |
9306 | ||
9307 | if (GET_MODE (op) != SFmode) | |
9308 | return 0; | |
9309 | ||
9310 | REAL_VALUE_FROM_CONST_DOUBLE (r, op); | |
9311 | return REAL_VALUES_EQUAL (r, dconst1); | |
d3ae8277 | 9312 | } |
1245df60 | 9313 | |
7a61cf6f | 9314 | /* In general mode switching is used. If we are |
9f09b1f2 R |
9315 | compiling without -mfmovd, movsf_ie isn't taken into account for |
9316 | mode switching. We could check in machine_dependent_reorg for | |
9317 | cases where we know we are in single precision mode, but there is | |
9318 | interface to find that out during reload, so we must avoid | |
9319 | choosing an fldi alternative during reload and thus failing to | |
9320 | allocate a scratch register for the constant loading. */ | |
9321 | int | |
cf277499 | 9322 | fldi_ok (void) |
9f09b1f2 | 9323 | { |
7a61cf6f | 9324 | return 1; |
9f09b1f2 R |
9325 | } |
9326 | ||
225e4f43 | 9327 | int |
cf277499 | 9328 | tertiary_reload_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
225e4f43 R |
9329 | { |
9330 | enum rtx_code code = GET_CODE (op); | |
9331 | return code == MEM || (TARGET_SH4 && code == CONST_DOUBLE); | |
9332 | } | |
9333 | ||
463f02cd | 9334 | /* Return the TLS type for TLS symbols, 0 for otherwise. */ |
81f40b79 | 9335 | enum tls_model |
cf277499 | 9336 | tls_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
463f02cd | 9337 | { |
463f02cd | 9338 | if (GET_CODE (op) != SYMBOL_REF) |
81f40b79 | 9339 | return TLS_MODEL_NONE; |
675ff4c7 | 9340 | return SYMBOL_REF_TLS_MODEL (op); |
463f02cd | 9341 | } |
1245df60 | 9342 | \f |
33f7f353 | 9343 | /* Return the destination address of a branch. */ |
f676971a | 9344 | |
318881c0 | 9345 | static int |
cf277499 | 9346 | branch_dest (rtx branch) |
1245df60 | 9347 | { |
33f7f353 JR |
9348 | rtx dest = SET_SRC (PATTERN (branch)); |
9349 | int dest_uid; | |
1245df60 R |
9350 | |
9351 | if (GET_CODE (dest) == IF_THEN_ELSE) | |
9352 | dest = XEXP (dest, 1); | |
9353 | dest = XEXP (dest, 0); | |
9354 | dest_uid = INSN_UID (dest); | |
9d98a694 | 9355 | return INSN_ADDRESSES (dest_uid); |
1245df60 | 9356 | } |
a55e9d2b | 9357 | \f |
5e7a8ee0 | 9358 | /* Return nonzero if REG is not used after INSN. |
a55e9d2b | 9359 | We assume REG is a reload reg, and therefore does |
8b760293 | 9360 | not live past labels. It may live past calls or jumps though. */ |
a55e9d2b | 9361 | int |
cf277499 | 9362 | reg_unused_after (rtx reg, rtx insn) |
a55e9d2b | 9363 | { |
8783b15e | 9364 | enum rtx_code code; |
a55e9d2b RK |
9365 | rtx set; |
9366 | ||
9367 | /* If the reg is set by this instruction, then it is safe for our | |
9368 | case. Disregard the case where this is a store to memory, since | |
9369 | we are checking a register used in the store address. */ | |
9370 | set = single_set (insn); | |
f3536097 | 9371 | if (set && !MEM_P (SET_DEST (set)) |
a55e9d2b RK |
9372 | && reg_overlap_mentioned_p (reg, SET_DEST (set))) |
9373 | return 1; | |
9374 | ||
318881c0 | 9375 | while ((insn = NEXT_INSN (insn))) |
a55e9d2b | 9376 | { |
ec8e098d PB |
9377 | rtx set; |
9378 | if (!INSN_P (insn)) | |
9379 | continue; | |
9380 | ||
a55e9d2b | 9381 | code = GET_CODE (insn); |
8783b15e | 9382 | |
c8f6f18d RK |
9383 | #if 0 |
9384 | /* If this is a label that existed before reload, then the register | |
9385 | if dead here. However, if this is a label added by reorg, then | |
9386 | the register may still be live here. We can't tell the difference, | |
9387 | so we just ignore labels completely. */ | |
8783b15e | 9388 | if (code == CODE_LABEL) |
a55e9d2b | 9389 | return 1; |
c8f6f18d RK |
9390 | /* else */ |
9391 | #endif | |
a55e9d2b | 9392 | |
8b760293 JW |
9393 | if (code == JUMP_INSN) |
9394 | return 0; | |
9395 | ||
8783b15e RK |
9396 | /* If this is a sequence, we must handle them all at once. |
9397 | We could have for instance a call that sets the target register, | |
e03f5d43 | 9398 | and an insn in a delay slot that uses the register. In this case, |
8783b15e | 9399 | we must return 0. */ |
8b760293 | 9400 | else if (code == INSN && GET_CODE (PATTERN (insn)) == SEQUENCE) |
a55e9d2b | 9401 | { |
8783b15e RK |
9402 | int i; |
9403 | int retval = 0; | |
9404 | ||
9405 | for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++) | |
9406 | { | |
9407 | rtx this_insn = XVECEXP (PATTERN (insn), 0, i); | |
9408 | rtx set = single_set (this_insn); | |
9409 | ||
f3536097 | 9410 | if (CALL_P (this_insn)) |
8783b15e | 9411 | code = CALL_INSN; |
f3536097 | 9412 | else if (JUMP_P (this_insn)) |
38f35781 JW |
9413 | { |
9414 | if (INSN_ANNULLED_BRANCH_P (this_insn)) | |
9415 | return 0; | |
9416 | code = JUMP_INSN; | |
9417 | } | |
a55e9d2b | 9418 | |
8783b15e RK |
9419 | if (set && reg_overlap_mentioned_p (reg, SET_SRC (set))) |
9420 | return 0; | |
9421 | if (set && reg_overlap_mentioned_p (reg, SET_DEST (set))) | |
9422 | { | |
f3536097 | 9423 | if (!MEM_P (SET_DEST (set))) |
8783b15e RK |
9424 | retval = 1; |
9425 | else | |
9426 | return 0; | |
9427 | } | |
9428 | if (set == 0 | |
9429 | && reg_overlap_mentioned_p (reg, PATTERN (this_insn))) | |
9430 | return 0; | |
9431 | } | |
9432 | if (retval == 1) | |
9433 | return 1; | |
38f35781 JW |
9434 | else if (code == JUMP_INSN) |
9435 | return 0; | |
8783b15e | 9436 | } |
a55e9d2b | 9437 | |
ec8e098d PB |
9438 | set = single_set (insn); |
9439 | if (set && reg_overlap_mentioned_p (reg, SET_SRC (set))) | |
9440 | return 0; | |
9441 | if (set && reg_overlap_mentioned_p (reg, SET_DEST (set))) | |
f3536097 | 9442 | return !MEM_P (SET_DEST (set)); |
ec8e098d PB |
9443 | if (set == 0 && reg_overlap_mentioned_p (reg, PATTERN (insn))) |
9444 | return 0; | |
8783b15e | 9445 | |
62164eb4 | 9446 | if (code == CALL_INSN && call_really_used_regs[REGNO (reg)]) |
8783b15e | 9447 | return 1; |
a55e9d2b RK |
9448 | } |
9449 | return 1; | |
9450 | } | |
225e4f43 | 9451 | \f |
c5b9ef02 | 9452 | #include "ggc.h" |
225e4f43 | 9453 | |
e2500fed | 9454 | static GTY(()) rtx fpscr_rtx; |
225e4f43 | 9455 | rtx |
cf277499 | 9456 | get_fpscr_rtx (void) |
225e4f43 | 9457 | { |
225e4f43 R |
9458 | if (! fpscr_rtx) |
9459 | { | |
c0d4e710 | 9460 | fpscr_rtx = gen_rtx_REG (PSImode, FPSCR_REG); |
225e4f43 | 9461 | REG_USERVAR_P (fpscr_rtx) = 1; |
225e4f43 R |
9462 | mark_user_reg (fpscr_rtx); |
9463 | } | |
9464 | if (! reload_completed || mdep_reorg_phase != SH_AFTER_MDEP_REORG) | |
9465 | mark_user_reg (fpscr_rtx); | |
9466 | return fpscr_rtx; | |
9467 | } | |
9468 | ||
24c2fde2 RH |
9469 | static GTY(()) tree fpscr_values; |
9470 | ||
9471 | static void | |
9472 | emit_fpu_switch (rtx scratch, int index) | |
9473 | { | |
9474 | rtx dst, src; | |
9475 | ||
9476 | if (fpscr_values == NULL) | |
9477 | { | |
9478 | tree t; | |
9479 | ||
9480 | t = build_index_type (integer_one_node); | |
9481 | t = build_array_type (integer_type_node, t); | |
4c4bde29 AH |
9482 | t = build_decl (BUILTINS_LOCATION, |
9483 | VAR_DECL, get_identifier ("__fpscr_values"), t); | |
24c2fde2 RH |
9484 | DECL_ARTIFICIAL (t) = 1; |
9485 | DECL_IGNORED_P (t) = 1; | |
9486 | DECL_EXTERNAL (t) = 1; | |
9487 | TREE_STATIC (t) = 1; | |
a5f5c6be | 9488 | TREE_PUBLIC (t) = 1; |
24c2fde2 RH |
9489 | TREE_USED (t) = 1; |
9490 | ||
9491 | fpscr_values = t; | |
9492 | } | |
9493 | ||
9494 | src = DECL_RTL (fpscr_values); | |
b3a13419 | 9495 | if (!can_create_pseudo_p ()) |
24c2fde2 RH |
9496 | { |
9497 | emit_move_insn (scratch, XEXP (src, 0)); | |
9498 | if (index != 0) | |
9499 | emit_insn (gen_addsi3 (scratch, scratch, GEN_INT (index * 4))); | |
9500 | src = adjust_automodify_address (src, PSImode, scratch, index * 4); | |
9501 | } | |
9502 | else | |
9503 | src = adjust_address (src, PSImode, index * 4); | |
9504 | ||
9505 | dst = get_fpscr_rtx (); | |
9506 | emit_move_insn (dst, src); | |
9507 | } | |
9508 | ||
225e4f43 | 9509 | void |
cf277499 | 9510 | emit_sf_insn (rtx pat) |
225e4f43 | 9511 | { |
225e4f43 | 9512 | emit_insn (pat); |
225e4f43 R |
9513 | } |
9514 | ||
9515 | void | |
cf277499 | 9516 | emit_df_insn (rtx pat) |
225e4f43 | 9517 | { |
225e4f43 | 9518 | emit_insn (pat); |
225e4f43 R |
9519 | } |
9520 | ||
9521 | void | |
cf277499 | 9522 | expand_sf_unop (rtx (*fun) (rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
9523 | { |
9524 | emit_sf_insn ((*fun) (operands[0], operands[1], get_fpscr_rtx ())); | |
9525 | } | |
9526 | ||
9527 | void | |
cf277499 | 9528 | expand_sf_binop (rtx (*fun) (rtx, rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
9529 | { |
9530 | emit_sf_insn ((*fun) (operands[0], operands[1], operands[2], | |
9531 | get_fpscr_rtx ())); | |
9532 | } | |
9533 | ||
9534 | void | |
cf277499 | 9535 | expand_df_unop (rtx (*fun) (rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
9536 | { |
9537 | emit_df_insn ((*fun) (operands[0], operands[1], get_fpscr_rtx ())); | |
9538 | } | |
9539 | ||
9540 | void | |
cf277499 | 9541 | expand_df_binop (rtx (*fun) (rtx, rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
9542 | { |
9543 | emit_df_insn ((*fun) (operands[0], operands[1], operands[2], | |
832a3292 | 9544 | get_fpscr_rtx ())); |
225e4f43 | 9545 | } |
225e4f43 | 9546 | \f |
cf277499 | 9547 | static rtx get_free_reg (HARD_REG_SET); |
9f09b1f2 R |
9548 | |
9549 | /* This function returns a register to use to load the address to load | |
9550 | the fpscr from. Currently it always returns r1 or r7, but when we are | |
9551 | able to use pseudo registers after combine, or have a better mechanism | |
9552 | for choosing a register, it should be done here. */ | |
9553 | /* REGS_LIVE is the liveness information for the point for which we | |
9554 | need this allocation. In some bare-bones exit blocks, r1 is live at the | |
9555 | start. We can even have all of r0..r3 being live: | |
9556 | __complex__ long long f (double d) { if (d == 0) return 2; else return 3; } | |
9557 | INSN before which new insns are placed with will clobber the register | |
9558 | we return. If a basic block consists only of setting the return value | |
9559 | register to a pseudo and using that register, the return value is not | |
9560 | live before or after this block, yet we we'll insert our insns right in | |
9561 | the middle. */ | |
225e4f43 | 9562 | |
9f09b1f2 | 9563 | static rtx |
cf277499 | 9564 | get_free_reg (HARD_REG_SET regs_live) |
225e4f43 | 9565 | { |
9f09b1f2 R |
9566 | if (! TEST_HARD_REG_BIT (regs_live, 1)) |
9567 | return gen_rtx_REG (Pmode, 1); | |
225e4f43 | 9568 | |
42db504c | 9569 | /* Hard reg 1 is live; since this is a small register classes target, |
9f09b1f2 | 9570 | there shouldn't be anything but a jump before the function end. */ |
f5b9e7c9 NS |
9571 | gcc_assert (!TEST_HARD_REG_BIT (regs_live, 7)); |
9572 | return gen_rtx_REG (Pmode, 7); | |
9f09b1f2 | 9573 | } |
225e4f43 | 9574 | |
f676971a | 9575 | /* This function will set the fpscr from memory. |
9f09b1f2 R |
9576 | MODE is the mode we are setting it to. */ |
9577 | void | |
cf277499 | 9578 | fpscr_set_from_mem (int mode, HARD_REG_SET regs_live) |
9f09b1f2 | 9579 | { |
81f40b79 | 9580 | enum attr_fp_mode fp_mode = (enum attr_fp_mode) mode; |
24c2fde2 | 9581 | enum attr_fp_mode norm_mode = ACTUAL_NORMAL_MODE (FP_MODE); |
5e77c6f9 | 9582 | rtx addr_reg; |
9f09b1f2 | 9583 | |
b3a13419 | 9584 | addr_reg = !can_create_pseudo_p () ? get_free_reg (regs_live) : NULL_RTX; |
24c2fde2 | 9585 | emit_fpu_switch (addr_reg, fp_mode == norm_mode); |
225e4f43 | 9586 | } |
9342c0c4 R |
9587 | |
9588 | /* Is the given character a logical line separator for the assembler? */ | |
9589 | #ifndef IS_ASM_LOGICAL_LINE_SEPARATOR | |
980d8882 | 9590 | #define IS_ASM_LOGICAL_LINE_SEPARATOR(C, STR) ((C) == ';') |
9342c0c4 R |
9591 | #endif |
9592 | ||
9593 | int | |
cf277499 | 9594 | sh_insn_length_adjustment (rtx insn) |
9342c0c4 R |
9595 | { |
9596 | /* Instructions with unfilled delay slots take up an extra two bytes for | |
9597 | the nop in the delay slot. */ | |
f3536097 | 9598 | if (((NONJUMP_INSN_P (insn) |
c1b92d09 R |
9599 | && GET_CODE (PATTERN (insn)) != USE |
9600 | && GET_CODE (PATTERN (insn)) != CLOBBER) | |
f3536097 | 9601 | || CALL_P (insn) |
69b18163 | 9602 | || (JUMP_P (insn) && !JUMP_TABLE_DATA_P (insn))) |
9342c0c4 R |
9603 | && GET_CODE (PATTERN (NEXT_INSN (PREV_INSN (insn)))) != SEQUENCE |
9604 | && get_attr_needs_delay_slot (insn) == NEEDS_DELAY_SLOT_YES) | |
9605 | return 2; | |
9606 | ||
3a8699c7 AO |
9607 | /* SH2e has a bug that prevents the use of annulled branches, so if |
9608 | the delay slot is not filled, we'll have to put a NOP in it. */ | |
b757d36f | 9609 | if (sh_cpu_attr == CPU_SH2E |
69b18163 | 9610 | && JUMP_P (insn) && !JUMP_TABLE_DATA_P (insn) |
3a8699c7 AO |
9611 | && get_attr_type (insn) == TYPE_CBRANCH |
9612 | && GET_CODE (PATTERN (NEXT_INSN (PREV_INSN (insn)))) != SEQUENCE) | |
9613 | return 2; | |
9614 | ||
9342c0c4 | 9615 | /* sh-dsp parallel processing insn take four bytes instead of two. */ |
f676971a | 9616 | |
f3536097 | 9617 | if (NONJUMP_INSN_P (insn)) |
9342c0c4 R |
9618 | { |
9619 | int sum = 0; | |
9620 | rtx body = PATTERN (insn); | |
0a2aaacc | 9621 | const char *templ; |
8ceaac3c | 9622 | char c; |
9342c0c4 R |
9623 | int maybe_label = 1; |
9624 | ||
9625 | if (GET_CODE (body) == ASM_INPUT) | |
0a2aaacc | 9626 | templ = XSTR (body, 0); |
9342c0c4 | 9627 | else if (asm_noperands (body) >= 0) |
0a2aaacc | 9628 | templ |
343d66ad | 9629 | = decode_asm_operands (body, NULL, NULL, NULL, NULL, NULL); |
9342c0c4 R |
9630 | else |
9631 | return 0; | |
9632 | do | |
9633 | { | |
9634 | int ppi_adjust = 0; | |
9635 | ||
9636 | do | |
0a2aaacc | 9637 | c = *templ++; |
9342c0c4 R |
9638 | while (c == ' ' || c == '\t'); |
9639 | /* all sh-dsp parallel-processing insns start with p. | |
9640 | The only non-ppi sh insn starting with p is pref. | |
9641 | The only ppi starting with pr is prnd. */ | |
0a2aaacc | 9642 | if ((c == 'p' || c == 'P') && strncasecmp ("re", templ, 2)) |
9342c0c4 R |
9643 | ppi_adjust = 2; |
9644 | /* The repeat pseudo-insn expands two three insns, a total of | |
9645 | six bytes in size. */ | |
9646 | else if ((c == 'r' || c == 'R') | |
0a2aaacc | 9647 | && ! strncasecmp ("epeat", templ, 5)) |
9342c0c4 | 9648 | ppi_adjust = 4; |
980d8882 | 9649 | while (c && c != '\n' |
0a2aaacc | 9650 | && ! IS_ASM_LOGICAL_LINE_SEPARATOR (c, templ)) |
9342c0c4 R |
9651 | { |
9652 | /* If this is a label, it is obviously not a ppi insn. */ | |
9653 | if (c == ':' && maybe_label) | |
9654 | { | |
9655 | ppi_adjust = 0; | |
9656 | break; | |
9657 | } | |
9658 | else if (c == '\'' || c == '"') | |
9659 | maybe_label = 0; | |
0a2aaacc | 9660 | c = *templ++; |
9342c0c4 R |
9661 | } |
9662 | sum += ppi_adjust; | |
9663 | maybe_label = c != ':'; | |
9664 | } | |
9665 | while (c); | |
9666 | return sum; | |
9667 | } | |
9668 | return 0; | |
9669 | } | |
1a66cd67 | 9670 | \f |
cdeed45a KK |
9671 | /* Return TRUE for a valid displacement for the REG+disp addressing |
9672 | with MODE. */ | |
9673 | ||
9674 | /* ??? The SH2e does not have the REG+disp addressing mode when loading values | |
9675 | into the FRx registers. We implement this by setting the maximum offset | |
9676 | to zero when the value is SFmode. This also restricts loading of SFmode | |
9677 | values into the integer registers, but that can't be helped. */ | |
9678 | ||
9679 | /* The SH allows a displacement in a QI or HI amode, but only when the | |
9680 | other operand is R0. GCC doesn't handle this very well, so we forgot | |
9681 | all of that. | |
9682 | ||
9683 | A legitimate index for a QI or HI is 0, SI can be any number 0..63, | |
9684 | DI can be any number 0..60. */ | |
9685 | ||
9686 | bool | |
9687 | sh_legitimate_index_p (enum machine_mode mode, rtx op) | |
9688 | { | |
f3536097 | 9689 | if (CONST_INT_P (op)) |
cdeed45a KK |
9690 | { |
9691 | if (TARGET_SHMEDIA) | |
9692 | { | |
9693 | int size; | |
9694 | ||
9695 | /* Check if this the address of an unaligned load / store. */ | |
9696 | if (mode == VOIDmode) | |
9697 | return CONST_OK_FOR_I06 (INTVAL (op)); | |
9698 | ||
9699 | size = GET_MODE_SIZE (mode); | |
9700 | return (!(INTVAL (op) & (size - 1)) | |
9701 | && INTVAL (op) >= -512 * size | |
9702 | && INTVAL (op) < 512 * size); | |
9703 | } | |
9704 | ||
9705 | if (TARGET_SH2A) | |
9706 | { | |
9707 | if (GET_MODE_SIZE (mode) == 1 | |
9708 | && (unsigned) INTVAL (op) < 4096) | |
9709 | return true; | |
9710 | } | |
9711 | ||
9712 | if ((GET_MODE_SIZE (mode) == 4 | |
9713 | && (unsigned) INTVAL (op) < 64 | |
9714 | && !(INTVAL (op) & 3) | |
9715 | && !(TARGET_SH2E && mode == SFmode)) | |
9716 | || (GET_MODE_SIZE (mode) == 4 | |
9717 | && (unsigned) INTVAL (op) < 16383 | |
9718 | && !(INTVAL (op) & 3) && TARGET_SH2A)) | |
9719 | return true; | |
9720 | ||
9721 | if ((GET_MODE_SIZE (mode) == 8 | |
9722 | && (unsigned) INTVAL (op) < 60 | |
9723 | && !(INTVAL (op) & 3) | |
9724 | && !((TARGET_SH4 || TARGET_SH2A) && mode == DFmode)) | |
9725 | || ((GET_MODE_SIZE (mode)==8) | |
9726 | && (unsigned) INTVAL (op) < 8192 | |
9727 | && !(INTVAL (op) & (TARGET_SH2A_DOUBLE ? 7 : 3)) | |
9728 | && (TARGET_SH2A && mode == DFmode))) | |
9729 | return true; | |
9730 | } | |
9731 | ||
9732 | return false; | |
9733 | } | |
9734 | ||
9735 | /* Recognize an RTL expression that is a valid memory address for | |
9736 | an instruction. | |
9737 | The MODE argument is the machine mode for the MEM expression | |
9738 | that wants to use this address. | |
9739 | Allow REG | |
9740 | REG+disp | |
9741 | REG+r0 | |
9742 | REG++ | |
9743 | --REG */ | |
9744 | ||
1bb99877 | 9745 | static bool |
cdeed45a KK |
9746 | sh_legitimate_address_p (enum machine_mode mode, rtx x, bool strict) |
9747 | { | |
9748 | if (MAYBE_BASE_REGISTER_RTX_P (x, strict)) | |
9749 | return true; | |
9750 | else if ((GET_CODE (x) == POST_INC || GET_CODE (x) == PRE_DEC) | |
9751 | && ! TARGET_SHMEDIA | |
9752 | && MAYBE_BASE_REGISTER_RTX_P (XEXP (x, 0), strict)) | |
9753 | return true; | |
9754 | else if (GET_CODE (x) == PLUS | |
9755 | && (mode != PSImode || reload_completed)) | |
9756 | { | |
9757 | rtx xop0 = XEXP (x, 0); | |
9758 | rtx xop1 = XEXP (x, 1); | |
9759 | ||
9760 | if (GET_MODE_SIZE (mode) <= 8 | |
9761 | && MAYBE_BASE_REGISTER_RTX_P (xop0, strict) | |
9762 | && sh_legitimate_index_p (mode, xop1)) | |
9763 | return true; | |
9764 | ||
9765 | if ((ALLOW_INDEXED_ADDRESS || GET_MODE (x) == DImode | |
9766 | || ((xop0 == stack_pointer_rtx | |
9767 | || xop0 == hard_frame_pointer_rtx) | |
9768 | && REG_P (xop1) && REGNO (xop1) == R0_REG) | |
9769 | || ((xop1 == stack_pointer_rtx | |
9770 | || xop1 == hard_frame_pointer_rtx) | |
9771 | && REG_P (xop0) && REGNO (xop0) == R0_REG)) | |
9772 | && ((!TARGET_SHMEDIA && GET_MODE_SIZE (mode) <= 4) | |
9773 | || (TARGET_SHMEDIA && GET_MODE_SIZE (mode) <= 8) | |
9774 | || ((TARGET_SH4 || TARGET_SH2A_DOUBLE) | |
9775 | && TARGET_FMOVD && mode == DFmode))) | |
9776 | { | |
9777 | if (MAYBE_BASE_REGISTER_RTX_P (xop1, strict) | |
9778 | && MAYBE_INDEX_REGISTER_RTX_P (xop0, strict)) | |
9779 | return true; | |
9780 | if (MAYBE_INDEX_REGISTER_RTX_P (xop1, strict) | |
9781 | && MAYBE_BASE_REGISTER_RTX_P (xop0, strict)) | |
9782 | return true; | |
9783 | } | |
9784 | } | |
9785 | ||
9786 | return false; | |
9787 | } | |
9788 | \f | |
55eec660 AO |
9789 | /* Return TRUE if X references a SYMBOL_REF or LABEL_REF whose symbol |
9790 | isn't protected by a PIC unspec. */ | |
1a66cd67 | 9791 | int |
cf277499 | 9792 | nonpic_symbol_mentioned_p (rtx x) |
1a66cd67 AO |
9793 | { |
9794 | register const char *fmt; | |
9795 | register int i; | |
9796 | ||
fa5322fa AO |
9797 | if (GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == LABEL_REF |
9798 | || GET_CODE (x) == PC) | |
1a66cd67 AO |
9799 | return 1; |
9800 | ||
fa5322fa AO |
9801 | /* We don't want to look into the possible MEM location of a |
9802 | CONST_DOUBLE, since we're not going to use it, in general. */ | |
9803 | if (GET_CODE (x) == CONST_DOUBLE) | |
9804 | return 0; | |
9805 | ||
1a66cd67 | 9806 | if (GET_CODE (x) == UNSPEC |
4773afa4 AO |
9807 | && (XINT (x, 1) == UNSPEC_PIC |
9808 | || XINT (x, 1) == UNSPEC_GOT | |
9809 | || XINT (x, 1) == UNSPEC_GOTOFF | |
fa5322fa | 9810 | || XINT (x, 1) == UNSPEC_GOTPLT |
463f02cd KK |
9811 | || XINT (x, 1) == UNSPEC_GOTTPOFF |
9812 | || XINT (x, 1) == UNSPEC_DTPOFF | |
58f451f0 | 9813 | || XINT (x, 1) == UNSPEC_TPOFF |
dc3ba671 RS |
9814 | || XINT (x, 1) == UNSPEC_PLT |
9815 | || XINT (x, 1) == UNSPEC_SYMOFF | |
9816 | || XINT (x, 1) == UNSPEC_PCREL_SYMOFF)) | |
832a3292 | 9817 | return 0; |
1a66cd67 AO |
9818 | |
9819 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
9820 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
9821 | { | |
9822 | if (fmt[i] == 'E') | |
9823 | { | |
9824 | register int j; | |
9825 | ||
9826 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
9827 | if (nonpic_symbol_mentioned_p (XVECEXP (x, i, j))) | |
9828 | return 1; | |
9829 | } | |
9830 | else if (fmt[i] == 'e' && nonpic_symbol_mentioned_p (XEXP (x, i))) | |
9831 | return 1; | |
9832 | } | |
9833 | ||
9834 | return 0; | |
9835 | } | |
9836 | ||
9837 | /* Convert a non-PIC address in `orig' to a PIC address using @GOT or | |
6f317ef3 | 9838 | @GOTOFF in `reg'. */ |
1a66cd67 | 9839 | rtx |
cf277499 SB |
9840 | legitimize_pic_address (rtx orig, enum machine_mode mode ATTRIBUTE_UNUSED, |
9841 | rtx reg) | |
1a66cd67 | 9842 | { |
81f40b79 | 9843 | if (tls_symbolic_operand (orig, Pmode) != TLS_MODEL_NONE) |
463f02cd KK |
9844 | return orig; |
9845 | ||
1a66cd67 | 9846 | if (GET_CODE (orig) == LABEL_REF |
675ff4c7 | 9847 | || (GET_CODE (orig) == SYMBOL_REF && SYMBOL_REF_LOCAL_P (orig))) |
1a66cd67 AO |
9848 | { |
9849 | if (reg == 0) | |
9850 | reg = gen_reg_rtx (Pmode); | |
9851 | ||
9852 | emit_insn (gen_symGOTOFF2reg (reg, orig)); | |
9853 | return reg; | |
9854 | } | |
9855 | else if (GET_CODE (orig) == SYMBOL_REF) | |
9856 | { | |
9857 | if (reg == 0) | |
9858 | reg = gen_reg_rtx (Pmode); | |
9859 | ||
9860 | emit_insn (gen_symGOT2reg (reg, orig)); | |
9861 | return reg; | |
9862 | } | |
9863 | return orig; | |
9864 | } | |
b91455de | 9865 | |
506d7b68 PB |
9866 | /* Try machine-dependent ways of modifying an illegitimate address |
9867 | to be legitimate. If we find one, return the new, valid address. | |
9868 | Otherwise, return X. | |
9869 | ||
9870 | For the SH, if X is almost suitable for indexing, but the offset is | |
9871 | out of range, convert it into a normal form so that CSE has a chance | |
9872 | of reducing the number of address registers used. */ | |
9873 | ||
9874 | static rtx | |
9875 | sh_legitimize_address (rtx x, rtx oldx, enum machine_mode mode) | |
9876 | { | |
9877 | if (flag_pic) | |
9878 | x = legitimize_pic_address (oldx, mode, NULL_RTX); | |
9879 | ||
9880 | if (GET_CODE (x) == PLUS | |
9881 | && (GET_MODE_SIZE (mode) == 4 | |
9882 | || GET_MODE_SIZE (mode) == 8) | |
f3536097 | 9883 | && CONST_INT_P (XEXP (x, 1)) |
506d7b68 PB |
9884 | && BASE_REGISTER_RTX_P (XEXP (x, 0)) |
9885 | && ! TARGET_SHMEDIA | |
9886 | && ! ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && mode == DFmode) | |
9887 | && ! (TARGET_SH2E && mode == SFmode)) | |
9888 | { | |
9889 | rtx index_rtx = XEXP (x, 1); | |
9890 | HOST_WIDE_INT offset = INTVAL (index_rtx), offset_base; | |
9891 | rtx sum; | |
9892 | ||
9893 | /* On rare occasions, we might get an unaligned pointer | |
9894 | that is indexed in a way to give an aligned address. | |
9895 | Therefore, keep the lower two bits in offset_base. */ | |
9896 | /* Instead of offset_base 128..131 use 124..127, so that | |
9897 | simple add suffices. */ | |
9898 | if (offset > 127) | |
9899 | offset_base = ((offset + 4) & ~60) - 4; | |
9900 | else | |
9901 | offset_base = offset & ~60; | |
9902 | ||
9903 | /* Sometimes the normal form does not suit DImode. We | |
9904 | could avoid that by using smaller ranges, but that | |
9905 | would give less optimized code when SImode is | |
9906 | prevalent. */ | |
9907 | if (GET_MODE_SIZE (mode) + offset - offset_base <= 64) | |
9908 | { | |
9909 | sum = expand_binop (Pmode, add_optab, XEXP (x, 0), | |
9910 | GEN_INT (offset_base), NULL_RTX, 0, | |
9911 | OPTAB_LIB_WIDEN); | |
9912 | ||
9913 | return gen_rtx_PLUS (Pmode, sum, GEN_INT (offset - offset_base)); | |
9914 | } | |
9915 | } | |
9916 | ||
9917 | return x; | |
9918 | } | |
9919 | ||
6f50eb9c KK |
9920 | /* Attempt to replace *P, which is an address that needs reloading, with |
9921 | a valid memory address for an operand of mode MODE. | |
9922 | Like for sh_legitimize_address, for the SH we try to get a normal form | |
9923 | of the address. That will allow inheritance of the address reloads. */ | |
9924 | ||
9925 | bool | |
9926 | sh_legitimize_reload_address (rtx *p, enum machine_mode mode, int opnum, | |
9927 | int itype) | |
9928 | { | |
9929 | enum reload_type type = (enum reload_type) itype; | |
9930 | ||
9931 | if (GET_CODE (*p) == PLUS | |
9932 | && (GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) | |
9933 | && CONST_INT_P (XEXP (*p, 1)) | |
7140f0f9 | 9934 | && MAYBE_BASE_REGISTER_RTX_P (XEXP (*p, 0), true) |
6f50eb9c KK |
9935 | && ! TARGET_SHMEDIA |
9936 | && ! (TARGET_SH4 && mode == DFmode) | |
9937 | && ! (mode == PSImode && type == RELOAD_FOR_INPUT_ADDRESS) | |
9938 | && (ALLOW_INDEXED_ADDRESS | |
9939 | || XEXP (*p, 0) == stack_pointer_rtx | |
9940 | || XEXP (*p, 0) == hard_frame_pointer_rtx)) | |
9941 | { | |
9942 | rtx index_rtx = XEXP (*p, 1); | |
9943 | HOST_WIDE_INT offset = INTVAL (index_rtx), offset_base; | |
9944 | rtx sum; | |
9945 | ||
9946 | if (TARGET_SH2A && mode == DFmode && (offset & 0x7)) | |
9947 | { | |
9948 | push_reload (*p, NULL_RTX, p, NULL, | |
9949 | BASE_REG_CLASS, Pmode, VOIDmode, 0, 0, opnum, type); | |
9950 | goto win; | |
9951 | } | |
9952 | if (TARGET_SH2E && mode == SFmode) | |
9953 | { | |
9954 | *p = copy_rtx (*p); | |
9955 | push_reload (*p, NULL_RTX, p, NULL, | |
9956 | BASE_REG_CLASS, Pmode, VOIDmode, 0, 0, opnum, type); | |
9957 | goto win; | |
9958 | } | |
9959 | /* Instead of offset_base 128..131 use 124..127, so that | |
9960 | simple add suffices. */ | |
9961 | if (offset > 127) | |
9962 | offset_base = ((offset + 4) & ~60) - 4; | |
9963 | else | |
9964 | offset_base = offset & ~60; | |
9965 | /* Sometimes the normal form does not suit DImode. We could avoid | |
9966 | that by using smaller ranges, but that would give less optimized | |
9967 | code when SImode is prevalent. */ | |
7140f0f9 | 9968 | if (GET_MODE_SIZE (mode) + offset - offset_base <= 64) |
6f50eb9c KK |
9969 | { |
9970 | sum = gen_rtx_PLUS (Pmode, XEXP (*p, 0), GEN_INT (offset_base)); | |
9971 | *p = gen_rtx_PLUS (Pmode, sum, GEN_INT (offset - offset_base)); | |
9972 | push_reload (sum, NULL_RTX, &XEXP (*p, 0), NULL, | |
9973 | BASE_REG_CLASS, Pmode, VOIDmode, 0, 0, opnum, type); | |
9974 | goto win; | |
9975 | } | |
9976 | } | |
9977 | /* We must re-recognize what we created before. */ | |
9978 | else if (GET_CODE (*p) == PLUS | |
9979 | && (GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) | |
9980 | && GET_CODE (XEXP (*p, 0)) == PLUS | |
9981 | && CONST_INT_P (XEXP (XEXP (*p, 0), 1)) | |
7140f0f9 | 9982 | && MAYBE_BASE_REGISTER_RTX_P (XEXP (XEXP (*p, 0), 0), true) |
6f50eb9c KK |
9983 | && CONST_INT_P (XEXP (*p, 1)) |
9984 | && ! TARGET_SHMEDIA | |
9985 | && ! (TARGET_SH2E && mode == SFmode)) | |
9986 | { | |
9987 | /* Because this address is so complex, we know it must have | |
9988 | been created by LEGITIMIZE_RELOAD_ADDRESS before; thus, | |
9989 | it is already unshared, and needs no further unsharing. */ | |
9990 | push_reload (XEXP (*p, 0), NULL_RTX, &XEXP (*p, 0), NULL, | |
9991 | BASE_REG_CLASS, Pmode, VOIDmode, 0, 0, opnum, type); | |
9992 | goto win; | |
9993 | } | |
9994 | ||
9995 | return false; | |
9996 | ||
9997 | win: | |
9998 | return true; | |
9999 | } | |
10000 | ||
90833f7f KK |
10001 | /* In the name of slightly smaller debug output, and to cater to |
10002 | general assembler lossage, recognize various UNSPEC sequences | |
10003 | and turn them back into a direct symbol reference. */ | |
10004 | ||
10005 | static rtx | |
10006 | sh_delegitimize_address (rtx orig_x) | |
10007 | { | |
10008 | rtx x, y; | |
10009 | ||
10010 | orig_x = delegitimize_mem_from_attrs (orig_x); | |
10011 | ||
10012 | x = orig_x; | |
10013 | if (MEM_P (x)) | |
10014 | x = XEXP (x, 0); | |
10015 | if (GET_CODE (x) == CONST) | |
10016 | { | |
10017 | y = XEXP (x, 0); | |
10018 | if (GET_CODE (y) == UNSPEC) | |
10019 | { | |
10020 | if (XINT (y, 1) == UNSPEC_GOT | |
2efd0aa6 KK |
10021 | || XINT (y, 1) == UNSPEC_GOTOFF |
10022 | || XINT (y, 1) == UNSPEC_SYMOFF) | |
90833f7f | 10023 | return XVECEXP (y, 0, 0); |
2efd0aa6 KK |
10024 | else if (XINT (y, 1) == UNSPEC_PCREL_SYMOFF) |
10025 | { | |
10026 | if (GET_CODE (XVECEXP (y, 0, 0)) == CONST) | |
10027 | { | |
10028 | rtx symplt = XEXP (XVECEXP (y, 0, 0), 0); | |
10029 | ||
10030 | if (GET_CODE (symplt) == UNSPEC | |
10031 | && XINT (symplt, 1) == UNSPEC_PLT) | |
10032 | return XVECEXP (symplt, 0, 0); | |
10033 | } | |
10034 | } | |
90833f7f KK |
10035 | else if (TARGET_SHMEDIA |
10036 | && (XINT (y, 1) == UNSPEC_EXTRACT_S16 | |
10037 | || XINT (y, 1) == UNSPEC_EXTRACT_U16)) | |
10038 | { | |
10039 | rtx offset = XVECEXP (y, 0, 1); | |
10040 | ||
10041 | x = gen_rtx_PLUS (Pmode, XVECEXP (y, 0, 0), offset); | |
10042 | if (MEM_P (orig_x)) | |
10043 | x = replace_equiv_address_nv (orig_x, x); | |
10044 | return x; | |
10045 | } | |
10046 | } | |
10047 | } | |
10048 | ||
10049 | return orig_x; | |
10050 | } | |
10051 | ||
b91455de KK |
10052 | /* Mark the use of a constant in the literal table. If the constant |
10053 | has multiple labels, make it unique. */ | |
c1b92d09 | 10054 | static rtx |
cf277499 | 10055 | mark_constant_pool_use (rtx x) |
b91455de KK |
10056 | { |
10057 | rtx insn, lab, pattern; | |
10058 | ||
10059 | if (x == NULL) | |
10060 | return x; | |
10061 | ||
10062 | switch (GET_CODE (x)) | |
10063 | { | |
10064 | case LABEL_REF: | |
10065 | x = XEXP (x, 0); | |
10066 | case CODE_LABEL: | |
10067 | break; | |
10068 | default: | |
10069 | return x; | |
10070 | } | |
10071 | ||
10072 | /* Get the first label in the list of labels for the same constant | |
10073 | and delete another labels in the list. */ | |
10074 | lab = x; | |
10075 | for (insn = PREV_INSN (x); insn; insn = PREV_INSN (insn)) | |
10076 | { | |
f3536097 | 10077 | if (!LABEL_P (insn) |
b91455de KK |
10078 | || LABEL_REFS (insn) != NEXT_INSN (insn)) |
10079 | break; | |
10080 | lab = insn; | |
10081 | } | |
10082 | ||
10083 | for (insn = LABEL_REFS (lab); insn; insn = LABEL_REFS (insn)) | |
10084 | INSN_DELETED_P (insn) = 1; | |
10085 | ||
10086 | /* Mark constants in a window. */ | |
10087 | for (insn = NEXT_INSN (x); insn; insn = NEXT_INSN (insn)) | |
10088 | { | |
f3536097 | 10089 | if (!NONJUMP_INSN_P (insn)) |
b91455de KK |
10090 | continue; |
10091 | ||
10092 | pattern = PATTERN (insn); | |
10093 | if (GET_CODE (pattern) != UNSPEC_VOLATILE) | |
10094 | continue; | |
10095 | ||
10096 | switch (XINT (pattern, 1)) | |
10097 | { | |
10098 | case UNSPECV_CONST2: | |
10099 | case UNSPECV_CONST4: | |
10100 | case UNSPECV_CONST8: | |
10101 | XVECEXP (pattern, 0, 1) = const1_rtx; | |
10102 | break; | |
10103 | case UNSPECV_WINDOW_END: | |
10104 | if (XVECEXP (pattern, 0, 0) == x) | |
10105 | return lab; | |
10106 | break; | |
10107 | case UNSPECV_CONST_END: | |
10108 | return lab; | |
10109 | default: | |
10110 | break; | |
10111 | } | |
10112 | } | |
10113 | ||
10114 | return lab; | |
10115 | } | |
6e7b03e1 AH |
10116 | \f |
10117 | /* Return true if it's possible to redirect BRANCH1 to the destination | |
10118 | of an unconditional jump BRANCH2. We only want to do this if the | |
10119 | resulting branch will have a short displacement. */ | |
f676971a | 10120 | int |
cf277499 | 10121 | sh_can_redirect_branch (rtx branch1, rtx branch2) |
6e7b03e1 AH |
10122 | { |
10123 | if (flag_expensive_optimizations && simplejump_p (branch2)) | |
10124 | { | |
10125 | rtx dest = XEXP (SET_SRC (single_set (branch2)), 0); | |
10126 | rtx insn; | |
10127 | int distance; | |
f676971a EC |
10128 | |
10129 | for (distance = 0, insn = NEXT_INSN (branch1); | |
10130 | insn && distance < 256; | |
6e7b03e1 AH |
10131 | insn = PREV_INSN (insn)) |
10132 | { | |
f676971a | 10133 | if (insn == dest) |
6e7b03e1 AH |
10134 | return 1; |
10135 | else | |
10136 | distance += get_attr_length (insn); | |
10137 | } | |
f676971a EC |
10138 | for (distance = 0, insn = NEXT_INSN (branch1); |
10139 | insn && distance < 256; | |
6e7b03e1 AH |
10140 | insn = NEXT_INSN (insn)) |
10141 | { | |
f676971a | 10142 | if (insn == dest) |
6e7b03e1 AH |
10143 | return 1; |
10144 | else | |
10145 | distance += get_attr_length (insn); | |
10146 | } | |
10147 | } | |
10148 | return 0; | |
10149 | } | |
10150 | ||
5e7a8ee0 | 10151 | /* Return nonzero if register old_reg can be renamed to register new_reg. */ |
03b8ec29 | 10152 | int |
cf277499 SB |
10153 | sh_hard_regno_rename_ok (unsigned int old_reg ATTRIBUTE_UNUSED, |
10154 | unsigned int new_reg) | |
03b8ec29 | 10155 | { |
832a3292 KH |
10156 | /* Interrupt functions can only use registers that have already been |
10157 | saved by the prologue, even if they would normally be | |
10158 | call-clobbered. */ | |
03b8ec29 | 10159 | |
6fb5fa3c | 10160 | if (sh_cfun_interrupt_handler_p () && !df_regs_ever_live_p (new_reg)) |
832a3292 | 10161 | return 0; |
03b8ec29 | 10162 | |
832a3292 | 10163 | return 1; |
03b8ec29 DD |
10164 | } |
10165 | ||
95c0af87 | 10166 | /* Function to update the integer COST |
c237e94a ZW |
10167 | based on the relationship between INSN that is dependent on |
10168 | DEP_INSN through the dependence LINK. The default is to make no | |
10169 | adjustment to COST. This can be used for example to specify to | |
10170 | the scheduler that an output- or anti-dependence does not incur | |
95c0af87 JR |
10171 | the same cost as a data-dependence. The return value should be |
10172 | the new value for COST. */ | |
c237e94a | 10173 | static int |
cf277499 | 10174 | sh_adjust_cost (rtx insn, rtx link ATTRIBUTE_UNUSED, rtx dep_insn, int cost) |
c237e94a | 10175 | { |
c49439f1 | 10176 | rtx reg, use_pat; |
c237e94a | 10177 | |
2ad65b0e SC |
10178 | if (TARGET_SHMEDIA) |
10179 | { | |
10180 | /* On SHmedia, if the dependence is an anti-dependence or | |
24746a42 | 10181 | output-dependence, there is no cost. */ |
2ad65b0e | 10182 | if (REG_NOTE_KIND (link) != 0) |
73a4d10b R |
10183 | { |
10184 | /* However, dependencies between target register loads and | |
10185 | uses of the register in a subsequent block that are separated | |
10186 | by a conditional branch are not modelled - we have to do with | |
10187 | the anti-dependency between the target register load and the | |
10188 | conditional branch that ends the current block. */ | |
10189 | if (REG_NOTE_KIND (link) == REG_DEP_ANTI | |
10190 | && GET_CODE (PATTERN (dep_insn)) == SET | |
10191 | && (get_attr_type (dep_insn) == TYPE_PT_MEDIA | |
10192 | || get_attr_type (dep_insn) == TYPE_PTABS_MEDIA) | |
10193 | && get_attr_type (insn) == TYPE_CBRANCH_MEDIA) | |
10194 | { | |
10195 | int orig_cost = cost; | |
10196 | rtx note = find_reg_note (insn, REG_BR_PROB, 0); | |
10197 | rtx target = ((! note | |
10198 | || INTVAL (XEXP (note, 0)) * 2 < REG_BR_PROB_BASE) | |
10199 | ? insn : JUMP_LABEL (insn)); | |
10200 | /* On the likely path, the branch costs 1, on the unlikely path, | |
10201 | it costs 3. */ | |
10202 | cost--; | |
10203 | do | |
10204 | target = next_active_insn (target); | |
10205 | while (target && ! flow_dependent_p (target, dep_insn) | |
10206 | && --cost > 0); | |
10207 | /* If two branches are executed in immediate succession, with the | |
10208 | first branch properly predicted, this causes a stall at the | |
10209 | second branch, hence we won't need the target for the | |
10210 | second branch for two cycles after the launch of the first | |
10211 | branch. */ | |
10212 | if (cost > orig_cost - 2) | |
10213 | cost = orig_cost - 2; | |
10214 | } | |
10215 | else | |
10216 | cost = 0; | |
10217 | } | |
2ad65b0e | 10218 | |
73a4d10b R |
10219 | else if (get_attr_is_mac_media (insn) |
10220 | && get_attr_is_mac_media (dep_insn)) | |
10221 | cost = 1; | |
10222 | ||
10223 | else if (! reload_completed | |
10224 | && GET_CODE (PATTERN (insn)) == SET | |
10225 | && GET_CODE (SET_SRC (PATTERN (insn))) == FLOAT | |
10226 | && GET_CODE (PATTERN (dep_insn)) == SET | |
10227 | && fp_arith_reg_operand (SET_SRC (PATTERN (dep_insn)), VOIDmode) | |
10228 | && cost < 4) | |
10229 | cost = 4; | |
10230 | /* Schedule the ptabs for a casesi_jump_media in preference to stuff | |
10231 | that is needed at the target. */ | |
10232 | else if (get_attr_type (insn) == TYPE_JUMP_MEDIA | |
10233 | && ! flow_dependent_p (insn, dep_insn)) | |
10234 | cost--; | |
2ad65b0e | 10235 | } |
c49439f1 | 10236 | else if (REG_NOTE_KIND (link) == 0) |
c237e94a | 10237 | { |
78d310c2 R |
10238 | enum attr_type type; |
10239 | rtx dep_set; | |
c49439f1 R |
10240 | |
10241 | if (recog_memoized (insn) < 0 | |
10242 | || recog_memoized (dep_insn) < 0) | |
95c0af87 | 10243 | return cost; |
c49439f1 | 10244 | |
78d310c2 | 10245 | dep_set = single_set (dep_insn); |
c49439f1 | 10246 | |
78d310c2 R |
10247 | /* The latency that we specify in the scheduling description refers |
10248 | to the actual output, not to an auto-increment register; for that, | |
10249 | the latency is one. */ | |
10250 | if (dep_set && MEM_P (SET_SRC (dep_set)) && cost > 1) | |
10251 | { | |
10252 | rtx set = single_set (insn); | |
10253 | ||
10254 | if (set | |
10255 | && !reg_mentioned_p (SET_DEST (dep_set), SET_SRC (set)) | |
10256 | && (!MEM_P (SET_DEST (set)) | |
10257 | || !reg_mentioned_p (SET_DEST (dep_set), | |
10258 | XEXP (SET_DEST (set), 0)))) | |
10259 | cost = 1; | |
10260 | } | |
c237e94a ZW |
10261 | /* The only input for a call that is timing-critical is the |
10262 | function's address. */ | |
f3536097 | 10263 | if (CALL_P (insn)) |
c49439f1 R |
10264 | { |
10265 | rtx call = PATTERN (insn); | |
10266 | ||
10267 | if (GET_CODE (call) == PARALLEL) | |
10268 | call = XVECEXP (call, 0 ,0); | |
10269 | if (GET_CODE (call) == SET) | |
10270 | call = SET_SRC (call); | |
f3536097 | 10271 | if (GET_CODE (call) == CALL && MEM_P (XEXP (call, 0)) |
73a4d10b R |
10272 | /* sibcalli_thunk uses a symbol_ref in an unspec. */ |
10273 | && (GET_CODE (XEXP (XEXP (call, 0), 0)) == UNSPEC | |
10274 | || ! reg_set_p (XEXP (XEXP (call, 0), 0), dep_insn))) | |
78d310c2 | 10275 | cost -= TARGET_SH4_300 ? 3 : 6; |
c49439f1 | 10276 | } |
c237e94a ZW |
10277 | /* Likewise, the most timing critical input for an sfuncs call |
10278 | is the function address. However, sfuncs typically start | |
10279 | using their arguments pretty quickly. | |
78d310c2 R |
10280 | Assume a four cycle delay for SH4 before they are needed. |
10281 | Cached ST40-300 calls are quicker, so assume only a one | |
10282 | cycle delay there. | |
10283 | ??? Maybe we should encode the delays till input registers | |
10284 | are needed by sfuncs into the sfunc call insn. */ | |
c49439f1 R |
10285 | /* All sfunc calls are parallels with at least four components. |
10286 | Exploit this to avoid unnecessary calls to sfunc_uses_reg. */ | |
10287 | else if (GET_CODE (PATTERN (insn)) == PARALLEL | |
10288 | && XVECLEN (PATTERN (insn), 0) >= 4 | |
10289 | && (reg = sfunc_uses_reg (insn))) | |
10290 | { | |
10291 | if (! reg_set_p (reg, dep_insn)) | |
78d310c2 R |
10292 | cost -= TARGET_SH4_300 ? 1 : 4; |
10293 | } | |
10294 | if (TARGET_HARD_SH4 && !TARGET_SH4_300) | |
10295 | { | |
10296 | enum attr_type dep_type = get_attr_type (dep_insn); | |
10297 | ||
10298 | if (dep_type == TYPE_FLOAD || dep_type == TYPE_PCFLOAD) | |
10299 | cost--; | |
10300 | else if ((dep_type == TYPE_LOAD_SI || dep_type == TYPE_PCLOAD_SI) | |
10301 | && (type = get_attr_type (insn)) != TYPE_CALL | |
10302 | && type != TYPE_SFUNC) | |
10303 | cost--; | |
10304 | /* When the preceding instruction loads the shift amount of | |
10305 | the following SHAD/SHLD, the latency of the load is increased | |
10306 | by 1 cycle. */ | |
10307 | if (get_attr_type (insn) == TYPE_DYN_SHIFT | |
10308 | && get_attr_any_int_load (dep_insn) == ANY_INT_LOAD_YES | |
c6ac6e82 | 10309 | && reg_overlap_mentioned_p (SET_DEST (dep_set), |
78d310c2 R |
10310 | XEXP (SET_SRC (single_set (insn)), |
10311 | 1))) | |
10312 | cost++; | |
10313 | /* When an LS group instruction with a latency of less than | |
10314 | 3 cycles is followed by a double-precision floating-point | |
10315 | instruction, FIPR, or FTRV, the latency of the first | |
10316 | instruction is increased to 3 cycles. */ | |
10317 | else if (cost < 3 | |
10318 | && get_attr_insn_class (dep_insn) == INSN_CLASS_LS_GROUP | |
10319 | && get_attr_dfp_comp (insn) == DFP_COMP_YES) | |
10320 | cost = 3; | |
10321 | /* The lsw register of a double-precision computation is ready one | |
10322 | cycle earlier. */ | |
10323 | else if (reload_completed | |
10324 | && get_attr_dfp_comp (dep_insn) == DFP_COMP_YES | |
10325 | && (use_pat = single_set (insn)) | |
10326 | && ! regno_use_in (REGNO (SET_DEST (single_set (dep_insn))), | |
10327 | SET_SRC (use_pat))) | |
10328 | cost -= 1; | |
10329 | ||
10330 | if (get_attr_any_fp_comp (dep_insn) == ANY_FP_COMP_YES | |
10331 | && get_attr_late_fp_use (insn) == LATE_FP_USE_YES) | |
10332 | cost -= 1; | |
10333 | } | |
10334 | else if (TARGET_SH4_300) | |
10335 | { | |
10336 | /* Stores need their input register two cycles later. */ | |
10337 | if (dep_set && cost >= 1 | |
10338 | && ((type = get_attr_type (insn)) == TYPE_STORE | |
10339 | || type == TYPE_PSTORE | |
10340 | || type == TYPE_FSTORE || type == TYPE_MAC_MEM)) | |
10341 | { | |
10342 | rtx set = single_set (insn); | |
10343 | ||
10344 | if (!reg_mentioned_p (SET_SRC (set), XEXP (SET_DEST (set), 0)) | |
10345 | && rtx_equal_p (SET_SRC (set), SET_DEST (dep_set))) | |
10346 | { | |
10347 | cost -= 2; | |
10348 | /* But don't reduce the cost below 1 if the address depends | |
10349 | on a side effect of dep_insn. */ | |
10350 | if (cost < 1 | |
10351 | && modified_in_p (XEXP (SET_DEST (set), 0), dep_insn)) | |
10352 | cost = 1; | |
10353 | } | |
10354 | } | |
10355 | } | |
c49439f1 R |
10356 | } |
10357 | /* An anti-dependence penalty of two applies if the first insn is a double | |
10358 | precision fadd / fsub / fmul. */ | |
78d310c2 R |
10359 | else if (!TARGET_SH4_300 |
10360 | && REG_NOTE_KIND (link) == REG_DEP_ANTI | |
c49439f1 | 10361 | && recog_memoized (dep_insn) >= 0 |
78d310c2 R |
10362 | && (get_attr_type (dep_insn) == TYPE_DFP_ARITH |
10363 | || get_attr_type (dep_insn) == TYPE_DFP_MUL) | |
c49439f1 R |
10364 | /* A lot of alleged anti-flow dependences are fake, |
10365 | so check this one is real. */ | |
10366 | && flow_dependent_p (dep_insn, insn)) | |
c237e94a | 10367 | cost = 2; |
c49439f1 | 10368 | |
c237e94a ZW |
10369 | return cost; |
10370 | } | |
3007d592 | 10371 | |
c49439f1 R |
10372 | /* Check if INSN is flow-dependent on DEP_INSN. Can also be used to check |
10373 | if DEP_INSN is anti-flow dependent on INSN. */ | |
10374 | static int | |
cf277499 | 10375 | flow_dependent_p (rtx insn, rtx dep_insn) |
c49439f1 R |
10376 | { |
10377 | rtx tmp = PATTERN (insn); | |
10378 | ||
10379 | note_stores (PATTERN (dep_insn), flow_dependent_p_1, &tmp); | |
10380 | return tmp == NULL_RTX; | |
10381 | } | |
10382 | ||
10383 | /* A helper function for flow_dependent_p called through note_stores. */ | |
10384 | static void | |
7bc980e1 | 10385 | flow_dependent_p_1 (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) |
c49439f1 R |
10386 | { |
10387 | rtx * pinsn = (rtx *) data; | |
10388 | ||
10389 | if (*pinsn && reg_referenced_p (x, *pinsn)) | |
10390 | *pinsn = NULL_RTX; | |
10391 | } | |
10392 | ||
b48f503c | 10393 | /* For use by sh_allocate_initial_value. Note that sh.md contains some |
3007d592 R |
10394 | 'special function' patterns (type sfunc) that clobber pr, but that |
10395 | do not look like function calls to leaf_function_p. Hence we must | |
10396 | do this extra check. */ | |
b48f503c | 10397 | static int |
cf277499 | 10398 | sh_pr_n_sets (void) |
3007d592 | 10399 | { |
6fb5fa3c | 10400 | return DF_REG_DEF_COUNT (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG); |
3007d592 | 10401 | } |
fa5322fa | 10402 | |
b48f503c KK |
10403 | /* Return where to allocate pseudo for a given hard register initial |
10404 | value. */ | |
10405 | static rtx | |
10406 | sh_allocate_initial_value (rtx hard_reg) | |
10407 | { | |
10408 | rtx x; | |
10409 | ||
10410 | if (REGNO (hard_reg) == (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG)) | |
10411 | { | |
10412 | if (current_function_is_leaf | |
10413 | && ! sh_pr_n_sets () | |
10414 | && ! (TARGET_SHCOMPACT | |
38173d38 | 10415 | && ((crtl->args.info.call_cookie |
b48f503c | 10416 | & ~ CALL_COOKIE_RET_TRAMP (1)) |
e3b5732b | 10417 | || crtl->saves_all_registers))) |
b48f503c KK |
10418 | x = hard_reg; |
10419 | else | |
57d38024 | 10420 | x = gen_frame_mem (Pmode, return_address_pointer_rtx); |
b48f503c KK |
10421 | } |
10422 | else | |
10423 | x = NULL_RTX; | |
10424 | ||
10425 | return x; | |
10426 | } | |
10427 | ||
c49439f1 R |
10428 | /* This function returns "2" to indicate dual issue for the SH4 |
10429 | processor. To be used by the DFA pipeline description. */ | |
fae15c93 | 10430 | static int |
832a3292 | 10431 | sh_issue_rate (void) |
fae15c93 | 10432 | { |
c49439f1 R |
10433 | if (TARGET_SUPERSCALAR) |
10434 | return 2; | |
10435 | else | |
10436 | return 1; | |
fae15c93 VM |
10437 | } |
10438 | ||
58565a33 SKG |
10439 | /* Functions for ready queue reordering for sched1. */ |
10440 | ||
10441 | /* Get weight for mode for a set x. */ | |
10442 | static short | |
508ea1c5 | 10443 | find_set_regmode_weight (rtx x, enum machine_mode mode) |
58565a33 SKG |
10444 | { |
10445 | if (GET_CODE (x) == CLOBBER && register_operand (SET_DEST (x), mode)) | |
10446 | return 1; | |
10447 | if (GET_CODE (x) == SET && register_operand (SET_DEST (x), mode)) | |
10448 | { | |
f3536097 | 10449 | if (REG_P (SET_DEST (x))) |
58565a33 SKG |
10450 | { |
10451 | if (!reg_mentioned_p (SET_DEST (x), SET_SRC (x))) | |
10452 | return 1; | |
10453 | else | |
10454 | return 0; | |
10455 | } | |
10456 | return 1; | |
10457 | } | |
10458 | return 0; | |
10459 | } | |
10460 | ||
10461 | /* Get regmode weight for insn. */ | |
10462 | static short | |
508ea1c5 | 10463 | find_insn_regmode_weight (rtx insn, enum machine_mode mode) |
58565a33 SKG |
10464 | { |
10465 | short reg_weight = 0; | |
10466 | rtx x; | |
10467 | ||
10468 | /* Increment weight for each register born here. */ | |
10469 | x = PATTERN (insn); | |
10470 | reg_weight += find_set_regmode_weight (x, mode); | |
10471 | if (GET_CODE (x) == PARALLEL) | |
10472 | { | |
10473 | int j; | |
10474 | for (j = XVECLEN (x, 0) - 1; j >= 0; j--) | |
10475 | { | |
10476 | x = XVECEXP (PATTERN (insn), 0, j); | |
10477 | reg_weight += find_set_regmode_weight (x, mode); | |
10478 | } | |
10479 | } | |
10480 | /* Decrement weight for each register that dies here. */ | |
10481 | for (x = REG_NOTES (insn); x; x = XEXP (x, 1)) | |
10482 | { | |
10483 | if (REG_NOTE_KIND (x) == REG_DEAD || REG_NOTE_KIND (x) == REG_UNUSED) | |
10484 | { | |
10485 | rtx note = XEXP (x, 0); | |
f3536097 | 10486 | if (REG_P (note) && GET_MODE (note) == mode) |
58565a33 SKG |
10487 | reg_weight--; |
10488 | } | |
10489 | } | |
10490 | return reg_weight; | |
10491 | } | |
10492 | ||
10493 | /* Calculate regmode weights for all insns of a basic block. */ | |
10494 | static void | |
c6d56443 | 10495 | find_regmode_weight (basic_block b, enum machine_mode mode) |
58565a33 SKG |
10496 | { |
10497 | rtx insn, next_tail, head, tail; | |
10498 | ||
c6d56443 | 10499 | get_ebb_head_tail (b, b, &head, &tail); |
58565a33 SKG |
10500 | next_tail = NEXT_INSN (tail); |
10501 | ||
10502 | for (insn = head; insn != next_tail; insn = NEXT_INSN (insn)) | |
10503 | { | |
10504 | /* Handle register life information. */ | |
10505 | if (!INSN_P (insn)) | |
10506 | continue; | |
10507 | ||
10508 | if (mode == SFmode) | |
10509 | INSN_REGMODE_WEIGHT (insn, mode) = | |
10510 | find_insn_regmode_weight (insn, mode) + 2 * find_insn_regmode_weight (insn, DFmode); | |
10511 | else if (mode == SImode) | |
10512 | INSN_REGMODE_WEIGHT (insn, mode) = | |
10513 | find_insn_regmode_weight (insn, mode) + 2 * find_insn_regmode_weight (insn, DImode); | |
10514 | } | |
10515 | } | |
10516 | ||
10517 | /* Comparison function for ready queue sorting. */ | |
10518 | static int | |
508ea1c5 | 10519 | rank_for_reorder (const void *x, const void *y) |
58565a33 SKG |
10520 | { |
10521 | rtx tmp = *(const rtx *) y; | |
10522 | rtx tmp2 = *(const rtx *) x; | |
10523 | ||
10524 | /* The insn in a schedule group should be issued the first. */ | |
10525 | if (SCHED_GROUP_P (tmp) != SCHED_GROUP_P (tmp2)) | |
10526 | return SCHED_GROUP_P (tmp2) ? 1 : -1; | |
10527 | ||
f676971a | 10528 | /* If insns are equally good, sort by INSN_LUID (original insn order), This |
58565a33 SKG |
10529 | minimizes instruction movement, thus minimizing sched's effect on |
10530 | register pressure. */ | |
10531 | return INSN_LUID (tmp) - INSN_LUID (tmp2); | |
10532 | } | |
10533 | ||
10534 | /* Resort the array A in which only element at index N may be out of order. */ | |
10535 | static void | |
508ea1c5 | 10536 | swap_reorder (rtx *a, int n) |
58565a33 SKG |
10537 | { |
10538 | rtx insn = a[n - 1]; | |
10539 | int i = n - 2; | |
10540 | ||
10541 | while (i >= 0 && rank_for_reorder (a + i, &insn) >= 0) | |
10542 | { | |
10543 | a[i + 1] = a[i]; | |
10544 | i -= 1; | |
10545 | } | |
10546 | a[i + 1] = insn; | |
10547 | } | |
10548 | ||
832a3292 KH |
10549 | #define SCHED_REORDER(READY, N_READY) \ |
10550 | do \ | |
10551 | { \ | |
10552 | if ((N_READY) == 2) \ | |
10553 | swap_reorder (READY, N_READY); \ | |
10554 | else if ((N_READY) > 2) \ | |
10555 | qsort (READY, N_READY, sizeof (rtx), rank_for_reorder); \ | |
10556 | } \ | |
10557 | while (0) | |
58565a33 SKG |
10558 | |
10559 | /* Sort the ready list READY by ascending priority, using the SCHED_REORDER | |
10560 | macro. */ | |
10561 | static void | |
508ea1c5 | 10562 | ready_reorder (rtx *ready, int nready) |
58565a33 SKG |
10563 | { |
10564 | SCHED_REORDER (ready, nready); | |
10565 | } | |
10566 | ||
6fb5fa3c DB |
10567 | /* Count life regions of r0 for a block. */ |
10568 | static int | |
10569 | find_r0_life_regions (basic_block b) | |
10570 | { | |
10571 | rtx end, insn; | |
10572 | rtx pset; | |
10573 | rtx r0_reg; | |
10574 | int live; | |
10575 | int set; | |
10576 | int death = 0; | |
10577 | ||
89a95777 | 10578 | if (REGNO_REG_SET_P (df_get_live_in (b), R0_REG)) |
6fb5fa3c DB |
10579 | { |
10580 | set = 1; | |
10581 | live = 1; | |
10582 | } | |
10583 | else | |
10584 | { | |
10585 | set = 0; | |
10586 | live = 0; | |
10587 | } | |
10588 | ||
10589 | insn = BB_HEAD (b); | |
10590 | end = BB_END (b); | |
10591 | r0_reg = gen_rtx_REG (SImode, R0_REG); | |
10592 | while (1) | |
10593 | { | |
10594 | if (INSN_P (insn)) | |
10595 | { | |
10596 | if (find_regno_note (insn, REG_DEAD, R0_REG)) | |
10597 | { | |
10598 | death++; | |
10599 | live = 0; | |
10600 | } | |
10601 | if (!live | |
10602 | && (pset = single_set (insn)) | |
10603 | && reg_overlap_mentioned_p (r0_reg, SET_DEST (pset)) | |
10604 | && !find_regno_note (insn, REG_UNUSED, R0_REG)) | |
10605 | { | |
10606 | set++; | |
10607 | live = 1; | |
10608 | } | |
10609 | } | |
10610 | if (insn == end) | |
10611 | break; | |
10612 | insn = NEXT_INSN (insn); | |
10613 | } | |
10614 | return set - death; | |
10615 | } | |
10616 | ||
58565a33 SKG |
10617 | /* Calculate regmode weights for all insns of all basic block. */ |
10618 | static void | |
508ea1c5 KH |
10619 | sh_md_init_global (FILE *dump ATTRIBUTE_UNUSED, |
10620 | int verbose ATTRIBUTE_UNUSED, | |
10621 | int old_max_uid) | |
58565a33 SKG |
10622 | { |
10623 | basic_block b; | |
10624 | ||
10625 | regmode_weight[0] = (short *) xcalloc (old_max_uid, sizeof (short)); | |
10626 | regmode_weight[1] = (short *) xcalloc (old_max_uid, sizeof (short)); | |
6fb5fa3c | 10627 | r0_life_regions = 0; |
58565a33 SKG |
10628 | |
10629 | FOR_EACH_BB_REVERSE (b) | |
10630 | { | |
c6d56443 KK |
10631 | find_regmode_weight (b, SImode); |
10632 | find_regmode_weight (b, SFmode); | |
6fb5fa3c DB |
10633 | if (!reload_completed) |
10634 | r0_life_regions += find_r0_life_regions (b); | |
58565a33 SKG |
10635 | } |
10636 | ||
10637 | CURR_REGMODE_PRESSURE (SImode) = 0; | |
10638 | CURR_REGMODE_PRESSURE (SFmode) = 0; | |
10639 | ||
10640 | } | |
10641 | ||
10642 | /* Cleanup. */ | |
10643 | static void | |
508ea1c5 KH |
10644 | sh_md_finish_global (FILE *dump ATTRIBUTE_UNUSED, |
10645 | int verbose ATTRIBUTE_UNUSED) | |
58565a33 SKG |
10646 | { |
10647 | if (regmode_weight[0]) | |
10648 | { | |
10649 | free (regmode_weight[0]); | |
10650 | regmode_weight[0] = NULL; | |
10651 | } | |
10652 | if (regmode_weight[1]) | |
10653 | { | |
10654 | free (regmode_weight[1]); | |
10655 | regmode_weight[1] = NULL; | |
10656 | } | |
10657 | } | |
10658 | ||
b0fbea13 KK |
10659 | /* The scalar modes supported differs from the default version in TImode |
10660 | for 32-bit SHMEDIA. */ | |
10661 | static bool | |
10662 | sh_scalar_mode_supported_p (enum machine_mode mode) | |
10663 | { | |
10664 | if (TARGET_SHMEDIA32 && mode == TImode) | |
10665 | return false; | |
10666 | ||
10667 | return default_scalar_mode_supported_p (mode); | |
10668 | } | |
10669 | ||
58565a33 SKG |
10670 | /* Cache the can_issue_more so that we can return it from reorder2. Also, |
10671 | keep count of register pressures on SImode and SFmode. */ | |
10672 | static int | |
508ea1c5 KH |
10673 | sh_variable_issue (FILE *dump ATTRIBUTE_UNUSED, |
10674 | int sched_verbose ATTRIBUTE_UNUSED, | |
10675 | rtx insn, | |
10676 | int can_issue_more) | |
58565a33 SKG |
10677 | { |
10678 | if (GET_CODE (PATTERN (insn)) != USE | |
10679 | && GET_CODE (PATTERN (insn)) != CLOBBER) | |
10680 | cached_can_issue_more = can_issue_more - 1; | |
10681 | else | |
10682 | cached_can_issue_more = can_issue_more; | |
10683 | ||
10684 | if (reload_completed) | |
10685 | return cached_can_issue_more; | |
10686 | ||
10687 | CURR_REGMODE_PRESSURE (SImode) += INSN_REGMODE_WEIGHT (insn, SImode); | |
10688 | CURR_REGMODE_PRESSURE (SFmode) += INSN_REGMODE_WEIGHT (insn, SFmode); | |
10689 | ||
10690 | return cached_can_issue_more; | |
10691 | } | |
10692 | ||
10693 | static void | |
508ea1c5 KH |
10694 | sh_md_init (FILE *dump ATTRIBUTE_UNUSED, |
10695 | int verbose ATTRIBUTE_UNUSED, | |
10696 | int veclen ATTRIBUTE_UNUSED) | |
58565a33 SKG |
10697 | { |
10698 | CURR_REGMODE_PRESSURE (SImode) = 0; | |
10699 | CURR_REGMODE_PRESSURE (SFmode) = 0; | |
10700 | } | |
10701 | ||
10702 | /* Some magic numbers. */ | |
10703 | /* Pressure on register r0 can lead to spill failures. so avoid sched1 for | |
10704 | functions that already have high pressure on r0. */ | |
10705 | #define R0_MAX_LIFE_REGIONS 2 | |
4ee31f1e | 10706 | /* Register Pressure thresholds for SImode and SFmode registers. */ |
58565a33 SKG |
10707 | #define SIMODE_MAX_WEIGHT 5 |
10708 | #define SFMODE_MAX_WEIGHT 10 | |
10709 | ||
10710 | /* Return true if the pressure is high for MODE. */ | |
10711 | static short | |
508ea1c5 | 10712 | high_pressure (enum machine_mode mode) |
58565a33 SKG |
10713 | { |
10714 | /* Pressure on register r0 can lead to spill failures. so avoid sched1 for | |
10715 | functions that already have high pressure on r0. */ | |
6fb5fa3c DB |
10716 | if (r0_life_regions >= R0_MAX_LIFE_REGIONS) |
10717 | return 1; | |
58565a33 SKG |
10718 | |
10719 | if (mode == SFmode) | |
10720 | return (CURR_REGMODE_PRESSURE (SFmode) > SFMODE_MAX_WEIGHT); | |
10721 | else | |
10722 | return (CURR_REGMODE_PRESSURE (SImode) > SIMODE_MAX_WEIGHT); | |
10723 | } | |
10724 | ||
10725 | /* Reorder ready queue if register pressure is high. */ | |
10726 | static int | |
508ea1c5 KH |
10727 | sh_reorder (FILE *dump ATTRIBUTE_UNUSED, |
10728 | int sched_verbose ATTRIBUTE_UNUSED, | |
10729 | rtx *ready, | |
10730 | int *n_readyp, | |
10731 | int clock_var ATTRIBUTE_UNUSED) | |
58565a33 SKG |
10732 | { |
10733 | if (reload_completed) | |
10734 | return sh_issue_rate (); | |
10735 | ||
10736 | if (high_pressure (SFmode) || high_pressure (SImode)) | |
10737 | { | |
10738 | ready_reorder (ready, *n_readyp); | |
10739 | } | |
10740 | ||
10741 | return sh_issue_rate (); | |
10742 | } | |
10743 | ||
10744 | /* Skip cycles if the current register pressure is high. */ | |
f676971a | 10745 | static int |
508ea1c5 KH |
10746 | sh_reorder2 (FILE *dump ATTRIBUTE_UNUSED, |
10747 | int sched_verbose ATTRIBUTE_UNUSED, | |
10748 | rtx *ready ATTRIBUTE_UNUSED, | |
10749 | int *n_readyp ATTRIBUTE_UNUSED, | |
10750 | int clock_var ATTRIBUTE_UNUSED) | |
58565a33 SKG |
10751 | { |
10752 | if (reload_completed) | |
10753 | return cached_can_issue_more; | |
10754 | ||
f676971a | 10755 | if (high_pressure(SFmode) || high_pressure (SImode)) |
58565a33 SKG |
10756 | skip_cycles = 1; |
10757 | ||
10758 | return cached_can_issue_more; | |
10759 | } | |
10760 | ||
10761 | /* Skip cycles without sorting the ready queue. This will move insn from | |
10762 | Q->R. If this is the last cycle we are skipping; allow sorting of ready | |
f676971a | 10763 | queue by sh_reorder. */ |
58565a33 | 10764 | |
f676971a EC |
10765 | /* Generally, skipping these many cycles are sufficient for all insns to move |
10766 | from Q -> R. */ | |
10767 | #define MAX_SKIPS 8 | |
58565a33 SKG |
10768 | |
10769 | static int | |
508ea1c5 KH |
10770 | sh_dfa_new_cycle (FILE *sched_dump ATTRIBUTE_UNUSED, |
10771 | int sched_verbose ATTRIBUTE_UNUSED, | |
10772 | rtx insn ATTRIBUTE_UNUSED, | |
10773 | int last_clock_var, | |
10774 | int clock_var, | |
10775 | int *sort_p) | |
58565a33 SKG |
10776 | { |
10777 | if (reload_completed) | |
832a3292 | 10778 | return 0; |
58565a33 | 10779 | |
f676971a | 10780 | if (skip_cycles) |
58565a33 | 10781 | { |
832a3292 KH |
10782 | if ((clock_var - last_clock_var) < MAX_SKIPS) |
10783 | { | |
10784 | *sort_p = 0; | |
10785 | return 1; | |
10786 | } | |
10787 | /* If this is the last cycle we are skipping, allow reordering of R. */ | |
10788 | if ((clock_var - last_clock_var) == MAX_SKIPS) | |
10789 | { | |
10790 | *sort_p = 1; | |
10791 | return 1; | |
10792 | } | |
58565a33 | 10793 | } |
58565a33 SKG |
10794 | |
10795 | skip_cycles = 0; | |
10796 | ||
10797 | return 0; | |
10798 | } | |
10799 | ||
fa5322fa AO |
10800 | /* SHmedia requires registers for branches, so we can't generate new |
10801 | branches past reload. */ | |
10802 | static bool | |
cf277499 | 10803 | sh_cannot_modify_jumps_p (void) |
fa5322fa AO |
10804 | { |
10805 | return (TARGET_SHMEDIA && (reload_in_progress || reload_completed)); | |
10806 | } | |
10807 | ||
a87cf97e | 10808 | static reg_class_t |
fe3ad572 SC |
10809 | sh_target_reg_class (void) |
10810 | { | |
10811 | return TARGET_SHMEDIA ? TARGET_REGS : NO_REGS; | |
10812 | } | |
10813 | ||
10814 | static bool | |
10815 | sh_optimize_target_register_callee_saved (bool after_prologue_epilogue_gen) | |
10816 | { | |
73a4d10b | 10817 | HARD_REG_SET dummy; |
cb3d3842 | 10818 | #if 0 |
73a4d10b | 10819 | rtx insn; |
cb3d3842 | 10820 | #endif |
73a4d10b R |
10821 | |
10822 | if (! shmedia_space_reserved_for_target_registers) | |
10823 | return 0; | |
10824 | if (after_prologue_epilogue_gen && ! TARGET_SAVE_ALL_TARGET_REGS) | |
10825 | return 0; | |
10826 | if (calc_live_regs (&dummy) >= 6 * 8) | |
10827 | return 1; | |
73a4d10b | 10828 | return 0; |
fe3ad572 SC |
10829 | } |
10830 | ||
fa5322fa | 10831 | static bool |
3101faab | 10832 | sh_ms_bitfield_layout_p (const_tree record_type ATTRIBUTE_UNUSED) |
fa5322fa | 10833 | { |
61f71b34 | 10834 | return (TARGET_SH5 || TARGET_HITACHI || sh_attr_renesas_p (record_type)); |
fa5322fa | 10835 | } |
ca903bba | 10836 | \f |
f676971a | 10837 | /* |
ca903bba R |
10838 | On the SH1..SH4, the trampoline looks like |
10839 | 2 0002 D202 mov.l l2,r2 | |
10840 | 1 0000 D301 mov.l l1,r3 | |
10841 | 3 0004 422B jmp @r2 | |
10842 | 4 0006 0009 nop | |
10843 | 5 0008 00000000 l1: .long area | |
10844 | 6 000c 00000000 l2: .long function | |
10845 | ||
10846 | SH5 (compact) uses r1 instead of r3 for the static chain. */ | |
10847 | ||
10848 | ||
10849 | /* Emit RTL insns to initialize the variable parts of a trampoline. | |
10850 | FNADDR is an RTX for the address of the function's pure code. | |
10851 | CXT is an RTX for the static chain value for the function. */ | |
10852 | ||
994322ef RH |
10853 | static void |
10854 | sh_trampoline_init (rtx tramp_mem, tree fndecl, rtx cxt) | |
ca903bba | 10855 | { |
994322ef RH |
10856 | rtx fnaddr = XEXP (DECL_RTL (fndecl), 0); |
10857 | rtx tramp = force_reg (Pmode, XEXP (tramp_mem, 0)); | |
57d38024 | 10858 | |
ca903bba R |
10859 | if (TARGET_SHMEDIA64) |
10860 | { | |
10861 | rtx tramp_templ; | |
10862 | int fixed_len; | |
10863 | ||
10864 | rtx movi1 = GEN_INT (0xcc000010); | |
10865 | rtx shori1 = GEN_INT (0xc8000010); | |
10866 | rtx src, dst; | |
10867 | ||
10868 | /* The following trampoline works within a +- 128 KB range for cxt: | |
10869 | ptb/u cxt,tr1; movi fnaddr >> 48,r0; shori fnaddr >> 32,r0; | |
10870 | shori fnaddr >> 16,r0; shori fnaddr,r0; ptabs/l r0,tr0 | |
10871 | gettr tr1,r1; blink tr0,r63 */ | |
10872 | /* Address rounding makes it hard to compute the exact bounds of the | |
10873 | offset for this trampoline, but we have a rather generous offset | |
10874 | range, so frame_offset should do fine as an upper bound. */ | |
10875 | if (cxt == virtual_stack_vars_rtx && frame_offset < 0x20000) | |
10876 | { | |
10877 | /* ??? could optimize this trampoline initialization | |
10878 | by writing DImode words with two insns each. */ | |
10879 | rtx mask = force_reg (DImode, GEN_INT (0x3fffc00)); | |
10880 | rtx insn = gen_rtx_MINUS (DImode, cxt, tramp); | |
10881 | insn = gen_rtx_ASHIFT (DImode, insn, GEN_INT (10-2)); | |
10882 | insn = gen_rtx_AND (DImode, insn, mask); | |
10883 | /* Or in ptb/u .,tr1 pattern */ | |
10884 | insn = gen_rtx_IOR (DImode, insn, gen_int_mode (0xec000010, SImode)); | |
10885 | insn = force_operand (insn, NULL_RTX); | |
10886 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 10887 | emit_move_insn (change_address (tramp_mem, SImode, NULL_RTX), insn); |
ca903bba R |
10888 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (38)); |
10889 | insn = gen_rtx_AND (DImode, insn, mask); | |
10890 | insn = force_operand (gen_rtx_IOR (DImode, movi1, insn), NULL_RTX); | |
10891 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 10892 | emit_move_insn (adjust_address (tramp_mem, SImode, 4), insn); |
ca903bba R |
10893 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (22)); |
10894 | insn = gen_rtx_AND (DImode, insn, mask); | |
10895 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
10896 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 10897 | emit_move_insn (adjust_address (tramp_mem, SImode, 8), insn); |
ca903bba R |
10898 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (6)); |
10899 | insn = gen_rtx_AND (DImode, insn, mask); | |
10900 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
10901 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 10902 | emit_move_insn (adjust_address (tramp_mem, SImode, 12), insn); |
ca903bba R |
10903 | insn = gen_rtx_ASHIFT (DImode, fnaddr, GEN_INT (10)); |
10904 | insn = gen_rtx_AND (DImode, insn, mask); | |
10905 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
10906 | insn = gen_lowpart (SImode, insn); | |
57d38024 R |
10907 | emit_move_insn (adjust_address (tramp_mem, SImode, 16), insn); |
10908 | emit_move_insn (adjust_address (tramp_mem, SImode, 20), | |
ca903bba | 10909 | GEN_INT (0x6bf10600)); |
57d38024 | 10910 | emit_move_insn (adjust_address (tramp_mem, SImode, 24), |
ca903bba | 10911 | GEN_INT (0x4415fc10)); |
57d38024 | 10912 | emit_move_insn (adjust_address (tramp_mem, SImode, 28), |
ca903bba R |
10913 | GEN_INT (0x4401fff0)); |
10914 | emit_insn (gen_ic_invalidate_line (tramp)); | |
10915 | return; | |
10916 | } | |
10917 | tramp_templ = gen_rtx_SYMBOL_REF (Pmode,"__GCC_nested_trampoline"); | |
10918 | fixed_len = TRAMPOLINE_SIZE - 2 * GET_MODE_SIZE (Pmode); | |
10919 | ||
10920 | tramp_templ = gen_datalabel_ref (tramp_templ); | |
57d38024 R |
10921 | dst = tramp_mem; |
10922 | src = gen_const_mem (BLKmode, tramp_templ); | |
ca903bba R |
10923 | set_mem_align (dst, 256); |
10924 | set_mem_align (src, 64); | |
44bb111a | 10925 | emit_block_move (dst, src, GEN_INT (fixed_len), BLOCK_OP_NORMAL); |
ca903bba | 10926 | |
57d38024 R |
10927 | emit_move_insn (adjust_address (tramp_mem, Pmode, fixed_len), fnaddr); |
10928 | emit_move_insn (adjust_address (tramp_mem, Pmode, | |
10929 | fixed_len + GET_MODE_SIZE (Pmode)), | |
ca903bba R |
10930 | cxt); |
10931 | emit_insn (gen_ic_invalidate_line (tramp)); | |
10932 | return; | |
10933 | } | |
10934 | else if (TARGET_SHMEDIA) | |
10935 | { | |
10936 | /* movi fnaddr >> 16,r1; shori fnaddr,r1; ptabs/l r1,tr0 | |
10937 | movi cxt >> 16,r1; shori cxt,r1; blink tr0,r63 */ | |
10938 | rtx quad0 = gen_reg_rtx (DImode), cxtload = gen_reg_rtx (DImode); | |
10939 | rtx quad1 = gen_reg_rtx (DImode), quad2 = gen_reg_rtx (DImode); | |
10940 | /* movi 0,r1: 0xcc000010 shori 0,r1: c8000010 concatenated, | |
825db093 | 10941 | rotated 10 right, and higher 16 bit of every 32 selected. */ |
ca903bba R |
10942 | rtx movishori |
10943 | = force_reg (V2HImode, (simplify_gen_subreg | |
10944 | (V2HImode, GEN_INT (0x4330432), SImode, 0))); | |
10945 | rtx ptabs = force_reg (DImode, GEN_INT (0x6bf10600)); | |
10946 | rtx blink = force_reg (DImode, GEN_INT (0x4401fff0)); | |
10947 | ||
ca903bba R |
10948 | fnaddr = force_reg (SImode, fnaddr); |
10949 | cxt = force_reg (SImode, cxt); | |
10950 | emit_insn (gen_mshflo_w_x (gen_rtx_SUBREG (V4HImode, quad0, 0), | |
10951 | gen_rtx_SUBREG (V2HImode, fnaddr, 0), | |
10952 | movishori)); | |
f1e2239f | 10953 | emit_insn (gen_rotrdi3_mextr (quad0, quad0, |
ca903bba | 10954 | GEN_INT (TARGET_LITTLE_ENDIAN ? 24 : 56))); |
a556fd39 | 10955 | emit_insn (gen_ashldi3_media (quad0, quad0, const2_rtx)); |
57d38024 | 10956 | emit_move_insn (change_address (tramp_mem, DImode, NULL_RTX), quad0); |
ca903bba R |
10957 | emit_insn (gen_mshflo_w_x (gen_rtx_SUBREG (V4HImode, cxtload, 0), |
10958 | gen_rtx_SUBREG (V2HImode, cxt, 0), | |
10959 | movishori)); | |
f1e2239f | 10960 | emit_insn (gen_rotrdi3_mextr (cxtload, cxtload, |
ca903bba | 10961 | GEN_INT (TARGET_LITTLE_ENDIAN ? 24 : 56))); |
a556fd39 | 10962 | emit_insn (gen_ashldi3_media (cxtload, cxtload, const2_rtx)); |
ca903bba R |
10963 | if (TARGET_LITTLE_ENDIAN) |
10964 | { | |
10965 | emit_insn (gen_mshflo_l_di (quad1, ptabs, cxtload)); | |
10966 | emit_insn (gen_mextr4 (quad2, cxtload, blink)); | |
10967 | } | |
10968 | else | |
10969 | { | |
10970 | emit_insn (gen_mextr4 (quad1, cxtload, ptabs)); | |
10971 | emit_insn (gen_mshflo_l_di (quad2, blink, cxtload)); | |
10972 | } | |
57d38024 R |
10973 | emit_move_insn (adjust_address (tramp_mem, DImode, 8), quad1); |
10974 | emit_move_insn (adjust_address (tramp_mem, DImode, 16), quad2); | |
ca903bba R |
10975 | emit_insn (gen_ic_invalidate_line (tramp)); |
10976 | return; | |
10977 | } | |
10978 | else if (TARGET_SHCOMPACT) | |
10979 | { | |
10980 | emit_insn (gen_initialize_trampoline (tramp, cxt, fnaddr)); | |
10981 | return; | |
10982 | } | |
57d38024 | 10983 | emit_move_insn (change_address (tramp_mem, SImode, NULL_RTX), |
ca903bba R |
10984 | gen_int_mode (TARGET_LITTLE_ENDIAN ? 0xd301d202 : 0xd202d301, |
10985 | SImode)); | |
57d38024 | 10986 | emit_move_insn (adjust_address (tramp_mem, SImode, 4), |
ca903bba R |
10987 | gen_int_mode (TARGET_LITTLE_ENDIAN ? 0x0009422b : 0x422b0009, |
10988 | SImode)); | |
57d38024 R |
10989 | emit_move_insn (adjust_address (tramp_mem, SImode, 8), cxt); |
10990 | emit_move_insn (adjust_address (tramp_mem, SImode, 12), fnaddr); | |
ca903bba R |
10991 | if (TARGET_HARVARD) |
10992 | { | |
dc557046 | 10993 | if (!TARGET_INLINE_IC_INVALIDATE |
6f8eec29 | 10994 | || (!(TARGET_SH4A_ARCH || TARGET_SH4_300) && TARGET_USERMODE)) |
73a4d10b R |
10995 | emit_library_call (function_symbol (NULL, "__ic_invalidate", |
10996 | FUNCTION_ORDINARY), | |
b757d36f | 10997 | LCT_NORMAL, VOIDmode, 1, tramp, SImode); |
ca903bba R |
10998 | else |
10999 | emit_insn (gen_ic_invalidate_line (tramp)); | |
11000 | } | |
11001 | } | |
11002 | ||
994322ef RH |
11003 | /* On SH5, trampolines are SHmedia code, so add 1 to the address. */ |
11004 | ||
11005 | static rtx | |
11006 | sh_trampoline_adjust_address (rtx tramp) | |
11007 | { | |
11008 | if (TARGET_SHMEDIA) | |
11009 | tramp = expand_simple_binop (Pmode, PLUS, tramp, const1_rtx, | |
11010 | gen_reg_rtx (Pmode), 0, OPTAB_LIB_WIDEN); | |
11011 | return tramp; | |
11012 | } | |
11013 | ||
4977bab6 ZW |
11014 | /* FIXME: This is overly conservative. A SHcompact function that |
11015 | receives arguments ``by reference'' will have them stored in its | |
11016 | own stack frame, so it must not pass pointers or references to | |
11017 | these arguments to other functions by means of sibling calls. */ | |
73a4d10b R |
11018 | /* If PIC, we cannot make sibling calls to global functions |
11019 | because the PLT requires r12 to be live. */ | |
4977bab6 | 11020 | static bool |
cf277499 | 11021 | sh_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED) |
4977bab6 | 11022 | { |
73a4d10b | 11023 | return (1 |
4977bab6 | 11024 | && (! TARGET_SHCOMPACT |
38173d38 | 11025 | || crtl->args.info.stack_regs == 0) |
73a4d10b R |
11026 | && ! sh_cfun_interrupt_handler_p () |
11027 | && (! flag_pic | |
11028 | || (decl && ! TREE_PUBLIC (decl)) | |
11029 | || (decl && DECL_VISIBILITY (decl) != VISIBILITY_DEFAULT))); | |
4977bab6 | 11030 | } |
c1b92d09 R |
11031 | \f |
11032 | /* Machine specific built-in functions. */ | |
11033 | ||
11034 | struct builtin_description | |
11035 | { | |
11036 | const enum insn_code icode; | |
11037 | const char *const name; | |
11038 | int signature; | |
1cc24576 | 11039 | tree fndecl; |
c1b92d09 R |
11040 | }; |
11041 | ||
11042 | /* describe number and signedness of arguments; arg[0] == result | |
11043 | (1: unsigned, 2: signed, 4: don't care, 8: pointer 0: no argument */ | |
3b962ea0 | 11044 | /* 9: 64-bit pointer, 10: 32-bit pointer */ |
c1b92d09 R |
11045 | static const char signature_args[][4] = |
11046 | { | |
11047 | #define SH_BLTIN_V2SI2 0 | |
11048 | { 4, 4 }, | |
11049 | #define SH_BLTIN_V4HI2 1 | |
11050 | { 4, 4 }, | |
11051 | #define SH_BLTIN_V2SI3 2 | |
11052 | { 4, 4, 4 }, | |
11053 | #define SH_BLTIN_V4HI3 3 | |
11054 | { 4, 4, 4 }, | |
11055 | #define SH_BLTIN_V8QI3 4 | |
11056 | { 4, 4, 4 }, | |
11057 | #define SH_BLTIN_MAC_HISI 5 | |
11058 | { 1, 4, 4, 1 }, | |
11059 | #define SH_BLTIN_SH_HI 6 | |
11060 | { 4, 4, 1 }, | |
11061 | #define SH_BLTIN_SH_SI 7 | |
11062 | { 4, 4, 1 }, | |
11063 | #define SH_BLTIN_V4HI2V2SI 8 | |
11064 | { 4, 4, 4 }, | |
11065 | #define SH_BLTIN_V4HI2V8QI 9 | |
11066 | { 4, 4, 4 }, | |
11067 | #define SH_BLTIN_SISF 10 | |
11068 | { 4, 2 }, | |
11069 | #define SH_BLTIN_LDUA_L 11 | |
73a4d10b | 11070 | { 2, 10 }, |
c1b92d09 | 11071 | #define SH_BLTIN_LDUA_Q 12 |
73a4d10b | 11072 | { 1, 10 }, |
c1b92d09 | 11073 | #define SH_BLTIN_STUA_L 13 |
73a4d10b | 11074 | { 0, 10, 2 }, |
c1b92d09 | 11075 | #define SH_BLTIN_STUA_Q 14 |
73a4d10b R |
11076 | { 0, 10, 1 }, |
11077 | #define SH_BLTIN_LDUA_L64 15 | |
11078 | { 2, 9 }, | |
11079 | #define SH_BLTIN_LDUA_Q64 16 | |
11080 | { 1, 9 }, | |
11081 | #define SH_BLTIN_STUA_L64 17 | |
11082 | { 0, 9, 2 }, | |
11083 | #define SH_BLTIN_STUA_Q64 18 | |
11084 | { 0, 9, 1 }, | |
11085 | #define SH_BLTIN_NUM_SHARED_SIGNATURES 19 | |
11086 | #define SH_BLTIN_2 19 | |
11087 | #define SH_BLTIN_SU 19 | |
c1b92d09 | 11088 | { 1, 2 }, |
73a4d10b R |
11089 | #define SH_BLTIN_3 20 |
11090 | #define SH_BLTIN_SUS 20 | |
c1b92d09 | 11091 | { 2, 2, 1 }, |
73a4d10b | 11092 | #define SH_BLTIN_PSSV 21 |
c1b92d09 | 11093 | { 0, 8, 2, 2 }, |
73a4d10b R |
11094 | #define SH_BLTIN_XXUU 22 |
11095 | #define SH_BLTIN_UUUU 22 | |
c1b92d09 | 11096 | { 1, 1, 1, 1 }, |
73a4d10b | 11097 | #define SH_BLTIN_PV 23 |
c1b92d09 R |
11098 | { 0, 8 }, |
11099 | }; | |
24746a42 | 11100 | /* mcmv: operands considered unsigned. */ |
c1b92d09 | 11101 | /* mmulsum_wq, msad_ubq: result considered unsigned long long. */ |
24746a42 KH |
11102 | /* mperm: control value considered unsigned int. */ |
11103 | /* mshalds, mshard, mshards, mshlld, mshlrd: shift count is unsigned int. */ | |
c1b92d09 R |
11104 | /* mshards_q: returns signed short. */ |
11105 | /* nsb: takes long long arg, returns unsigned char. */ | |
1cc24576 KK |
11106 | static struct builtin_description bdesc[] = |
11107 | { | |
11108 | { CODE_FOR_absv2si2, "__builtin_absv2si2", SH_BLTIN_V2SI2, 0 }, | |
11109 | { CODE_FOR_absv4hi2, "__builtin_absv4hi2", SH_BLTIN_V4HI2, 0 }, | |
11110 | { CODE_FOR_addv2si3, "__builtin_addv2si3", SH_BLTIN_V2SI3, 0 }, | |
11111 | { CODE_FOR_addv4hi3, "__builtin_addv4hi3", SH_BLTIN_V4HI3, 0 }, | |
11112 | { CODE_FOR_ssaddv2si3,"__builtin_ssaddv2si3", SH_BLTIN_V2SI3, 0 }, | |
11113 | { CODE_FOR_usaddv8qi3,"__builtin_usaddv8qi3", SH_BLTIN_V8QI3, 0 }, | |
11114 | { CODE_FOR_ssaddv4hi3,"__builtin_ssaddv4hi3", SH_BLTIN_V4HI3, 0 }, | |
11115 | { CODE_FOR_alloco_i, "__builtin_sh_media_ALLOCO", SH_BLTIN_PV, 0 }, | |
11116 | { CODE_FOR_negcmpeqv8qi,"__builtin_sh_media_MCMPEQ_B", SH_BLTIN_V8QI3, 0 }, | |
11117 | { CODE_FOR_negcmpeqv2si,"__builtin_sh_media_MCMPEQ_L", SH_BLTIN_V2SI3, 0 }, | |
11118 | { CODE_FOR_negcmpeqv4hi,"__builtin_sh_media_MCMPEQ_W", SH_BLTIN_V4HI3, 0 }, | |
11119 | { CODE_FOR_negcmpgtuv8qi,"__builtin_sh_media_MCMPGT_UB", SH_BLTIN_V8QI3, 0 }, | |
11120 | { CODE_FOR_negcmpgtv2si,"__builtin_sh_media_MCMPGT_L", SH_BLTIN_V2SI3, 0 }, | |
11121 | { CODE_FOR_negcmpgtv4hi,"__builtin_sh_media_MCMPGT_W", SH_BLTIN_V4HI3, 0 }, | |
11122 | { CODE_FOR_mcmv, "__builtin_sh_media_MCMV", SH_BLTIN_UUUU, 0 }, | |
11123 | { CODE_FOR_mcnvs_lw, "__builtin_sh_media_MCNVS_LW", SH_BLTIN_3, 0 }, | |
11124 | { CODE_FOR_mcnvs_wb, "__builtin_sh_media_MCNVS_WB", SH_BLTIN_V4HI2V8QI, 0 }, | |
11125 | { CODE_FOR_mcnvs_wub, "__builtin_sh_media_MCNVS_WUB", SH_BLTIN_V4HI2V8QI, 0 }, | |
11126 | { CODE_FOR_mextr1, "__builtin_sh_media_MEXTR1", SH_BLTIN_V8QI3, 0 }, | |
11127 | { CODE_FOR_mextr2, "__builtin_sh_media_MEXTR2", SH_BLTIN_V8QI3, 0 }, | |
11128 | { CODE_FOR_mextr3, "__builtin_sh_media_MEXTR3", SH_BLTIN_V8QI3, 0 }, | |
11129 | { CODE_FOR_mextr4, "__builtin_sh_media_MEXTR4", SH_BLTIN_V8QI3, 0 }, | |
11130 | { CODE_FOR_mextr5, "__builtin_sh_media_MEXTR5", SH_BLTIN_V8QI3, 0 }, | |
11131 | { CODE_FOR_mextr6, "__builtin_sh_media_MEXTR6", SH_BLTIN_V8QI3, 0 }, | |
11132 | { CODE_FOR_mextr7, "__builtin_sh_media_MEXTR7", SH_BLTIN_V8QI3, 0 }, | |
11133 | { CODE_FOR_mmacfx_wl, "__builtin_sh_media_MMACFX_WL", SH_BLTIN_MAC_HISI, 0 }, | |
11134 | { CODE_FOR_mmacnfx_wl,"__builtin_sh_media_MMACNFX_WL", SH_BLTIN_MAC_HISI, 0 }, | |
11135 | { CODE_FOR_mulv2si3, "__builtin_mulv2si3", SH_BLTIN_V2SI3, 0 }, | |
11136 | { CODE_FOR_mulv4hi3, "__builtin_mulv4hi3", SH_BLTIN_V4HI3, 0 }, | |
11137 | { CODE_FOR_mmulfx_l, "__builtin_sh_media_MMULFX_L", SH_BLTIN_V2SI3, 0 }, | |
11138 | { CODE_FOR_mmulfx_w, "__builtin_sh_media_MMULFX_W", SH_BLTIN_V4HI3, 0 }, | |
11139 | { CODE_FOR_mmulfxrp_w,"__builtin_sh_media_MMULFXRP_W", SH_BLTIN_V4HI3, 0 }, | |
11140 | { CODE_FOR_mmulhi_wl, "__builtin_sh_media_MMULHI_WL", SH_BLTIN_V4HI2V2SI, 0 }, | |
11141 | { CODE_FOR_mmullo_wl, "__builtin_sh_media_MMULLO_WL", SH_BLTIN_V4HI2V2SI, 0 }, | |
11142 | { CODE_FOR_mmulsum_wq,"__builtin_sh_media_MMULSUM_WQ", SH_BLTIN_XXUU, 0 }, | |
11143 | { CODE_FOR_mperm_w, "__builtin_sh_media_MPERM_W", SH_BLTIN_SH_HI, 0 }, | |
11144 | { CODE_FOR_msad_ubq, "__builtin_sh_media_MSAD_UBQ", SH_BLTIN_XXUU, 0 }, | |
11145 | { CODE_FOR_mshalds_l, "__builtin_sh_media_MSHALDS_L", SH_BLTIN_SH_SI, 0 }, | |
11146 | { CODE_FOR_mshalds_w, "__builtin_sh_media_MSHALDS_W", SH_BLTIN_SH_HI, 0 }, | |
11147 | { CODE_FOR_ashrv2si3, "__builtin_ashrv2si3", SH_BLTIN_SH_SI, 0 }, | |
11148 | { CODE_FOR_ashrv4hi3, "__builtin_ashrv4hi3", SH_BLTIN_SH_HI, 0 }, | |
11149 | { CODE_FOR_mshards_q, "__builtin_sh_media_MSHARDS_Q", SH_BLTIN_SUS, 0 }, | |
11150 | { CODE_FOR_mshfhi_b, "__builtin_sh_media_MSHFHI_B", SH_BLTIN_V8QI3, 0 }, | |
11151 | { CODE_FOR_mshfhi_l, "__builtin_sh_media_MSHFHI_L", SH_BLTIN_V2SI3, 0 }, | |
11152 | { CODE_FOR_mshfhi_w, "__builtin_sh_media_MSHFHI_W", SH_BLTIN_V4HI3, 0 }, | |
11153 | { CODE_FOR_mshflo_b, "__builtin_sh_media_MSHFLO_B", SH_BLTIN_V8QI3, 0 }, | |
11154 | { CODE_FOR_mshflo_l, "__builtin_sh_media_MSHFLO_L", SH_BLTIN_V2SI3, 0 }, | |
11155 | { CODE_FOR_mshflo_w, "__builtin_sh_media_MSHFLO_W", SH_BLTIN_V4HI3, 0 }, | |
11156 | { CODE_FOR_ashlv2si3, "__builtin_ashlv2si3", SH_BLTIN_SH_SI, 0 }, | |
11157 | { CODE_FOR_ashlv4hi3, "__builtin_ashlv4hi3", SH_BLTIN_SH_HI, 0 }, | |
11158 | { CODE_FOR_lshrv2si3, "__builtin_lshrv2si3", SH_BLTIN_SH_SI, 0 }, | |
11159 | { CODE_FOR_lshrv4hi3, "__builtin_lshrv4hi3", SH_BLTIN_SH_HI, 0 }, | |
11160 | { CODE_FOR_subv2si3, "__builtin_subv2si3", SH_BLTIN_V2SI3, 0 }, | |
11161 | { CODE_FOR_subv4hi3, "__builtin_subv4hi3", SH_BLTIN_V4HI3, 0 }, | |
11162 | { CODE_FOR_sssubv2si3,"__builtin_sssubv2si3", SH_BLTIN_V2SI3, 0 }, | |
11163 | { CODE_FOR_ussubv8qi3,"__builtin_ussubv8qi3", SH_BLTIN_V8QI3, 0 }, | |
11164 | { CODE_FOR_sssubv4hi3,"__builtin_sssubv4hi3", SH_BLTIN_V4HI3, 0 }, | |
11165 | { CODE_FOR_fcosa_s, "__builtin_sh_media_FCOSA_S", SH_BLTIN_SISF, 0 }, | |
11166 | { CODE_FOR_fsina_s, "__builtin_sh_media_FSINA_S", SH_BLTIN_SISF, 0 }, | |
11167 | { CODE_FOR_fipr, "__builtin_sh_media_FIPR_S", SH_BLTIN_3, 0 }, | |
11168 | { CODE_FOR_ftrv, "__builtin_sh_media_FTRV_S", SH_BLTIN_3, 0 }, | |
11169 | { CODE_FOR_mac_media, "__builtin_sh_media_FMAC_S", SH_BLTIN_3, 0 }, | |
11170 | { CODE_FOR_sqrtdf2, "__builtin_sh_media_FSQRT_D", SH_BLTIN_2, 0 }, | |
11171 | { CODE_FOR_sqrtsf2, "__builtin_sh_media_FSQRT_S", SH_BLTIN_2, 0 }, | |
11172 | { CODE_FOR_fsrra_s, "__builtin_sh_media_FSRRA_S", SH_BLTIN_2, 0 }, | |
11173 | { CODE_FOR_ldhi_l, "__builtin_sh_media_LDHI_L", SH_BLTIN_LDUA_L, 0 }, | |
11174 | { CODE_FOR_ldhi_q, "__builtin_sh_media_LDHI_Q", SH_BLTIN_LDUA_Q, 0 }, | |
11175 | { CODE_FOR_ldlo_l, "__builtin_sh_media_LDLO_L", SH_BLTIN_LDUA_L, 0 }, | |
11176 | { CODE_FOR_ldlo_q, "__builtin_sh_media_LDLO_Q", SH_BLTIN_LDUA_Q, 0 }, | |
11177 | { CODE_FOR_sthi_l, "__builtin_sh_media_STHI_L", SH_BLTIN_STUA_L, 0 }, | |
11178 | { CODE_FOR_sthi_q, "__builtin_sh_media_STHI_Q", SH_BLTIN_STUA_Q, 0 }, | |
11179 | { CODE_FOR_stlo_l, "__builtin_sh_media_STLO_L", SH_BLTIN_STUA_L, 0 }, | |
11180 | { CODE_FOR_stlo_q, "__builtin_sh_media_STLO_Q", SH_BLTIN_STUA_Q, 0 }, | |
11181 | { CODE_FOR_ldhi_l64, "__builtin_sh_media_LDHI_L", SH_BLTIN_LDUA_L64, 0 }, | |
11182 | { CODE_FOR_ldhi_q64, "__builtin_sh_media_LDHI_Q", SH_BLTIN_LDUA_Q64, 0 }, | |
11183 | { CODE_FOR_ldlo_l64, "__builtin_sh_media_LDLO_L", SH_BLTIN_LDUA_L64, 0 }, | |
11184 | { CODE_FOR_ldlo_q64, "__builtin_sh_media_LDLO_Q", SH_BLTIN_LDUA_Q64, 0 }, | |
11185 | { CODE_FOR_sthi_l64, "__builtin_sh_media_STHI_L", SH_BLTIN_STUA_L64, 0 }, | |
11186 | { CODE_FOR_sthi_q64, "__builtin_sh_media_STHI_Q", SH_BLTIN_STUA_Q64, 0 }, | |
11187 | { CODE_FOR_stlo_l64, "__builtin_sh_media_STLO_L", SH_BLTIN_STUA_L64, 0 }, | |
11188 | { CODE_FOR_stlo_q64, "__builtin_sh_media_STLO_Q", SH_BLTIN_STUA_Q64, 0 }, | |
11189 | { CODE_FOR_nsb, "__builtin_sh_media_NSB", SH_BLTIN_SU, 0 }, | |
11190 | { CODE_FOR_byterev, "__builtin_sh_media_BYTEREV", SH_BLTIN_2, 0 }, | |
11191 | { CODE_FOR_prefetch, "__builtin_sh_media_PREFO", SH_BLTIN_PSSV, 0 }, | |
c1b92d09 R |
11192 | }; |
11193 | ||
11194 | static void | |
cf277499 | 11195 | sh_media_init_builtins (void) |
c1b92d09 R |
11196 | { |
11197 | tree shared[SH_BLTIN_NUM_SHARED_SIGNATURES]; | |
1cc24576 | 11198 | struct builtin_description *d; |
c1b92d09 | 11199 | |
a3acdc0c | 11200 | memset (shared, 0, sizeof shared); |
e97a46ce | 11201 | for (d = bdesc; d - bdesc < (int) ARRAY_SIZE (bdesc); d++) |
c1b92d09 | 11202 | { |
73a4d10b | 11203 | tree type, arg_type = 0; |
c1b92d09 R |
11204 | int signature = d->signature; |
11205 | int i; | |
11206 | ||
11207 | if (signature < SH_BLTIN_NUM_SHARED_SIGNATURES && shared[signature]) | |
11208 | type = shared[signature]; | |
11209 | else | |
11210 | { | |
11211 | int has_result = signature_args[signature][0] != 0; | |
11212 | ||
73a4d10b R |
11213 | if ((signature_args[signature][1] & 8) |
11214 | && (((signature_args[signature][1] & 1) && TARGET_SHMEDIA32) | |
11215 | || ((signature_args[signature][1] & 2) && TARGET_SHMEDIA64))) | |
c1b92d09 R |
11216 | continue; |
11217 | if (! TARGET_FPU_ANY | |
11218 | && FLOAT_MODE_P (insn_data[d->icode].operand[0].mode)) | |
11219 | continue; | |
11220 | type = void_list_node; | |
11221 | for (i = 3; ; i--) | |
11222 | { | |
11223 | int arg = signature_args[signature][i]; | |
11224 | int opno = i - 1 + has_result; | |
11225 | ||
73a4d10b | 11226 | if (arg & 8) |
c1b92d09 R |
11227 | arg_type = ptr_type_node; |
11228 | else if (arg) | |
73a4d10b R |
11229 | arg_type = (*lang_hooks.types.type_for_mode) |
11230 | (insn_data[d->icode].operand[opno].mode, | |
11231 | (arg & 1)); | |
c1b92d09 R |
11232 | else if (i) |
11233 | continue; | |
11234 | else | |
11235 | arg_type = void_type_node; | |
11236 | if (i == 0) | |
11237 | break; | |
11238 | type = tree_cons (NULL_TREE, arg_type, type); | |
11239 | } | |
11240 | type = build_function_type (arg_type, type); | |
11241 | if (signature < SH_BLTIN_NUM_SHARED_SIGNATURES) | |
11242 | shared[signature] = type; | |
11243 | } | |
1cc24576 KK |
11244 | d->fndecl = |
11245 | add_builtin_function (d->name, type, d - bdesc, BUILT_IN_MD, | |
11246 | NULL, NULL_TREE); | |
c1b92d09 R |
11247 | } |
11248 | } | |
11249 | ||
1cc24576 KK |
11250 | /* Returns the shmedia builtin decl for CODE. */ |
11251 | ||
11252 | static tree | |
11253 | sh_media_builtin_decl (unsigned code, bool initialize_p ATTRIBUTE_UNUSED) | |
11254 | { | |
11255 | if (code >= ARRAY_SIZE (bdesc)) | |
11256 | return error_mark_node; | |
11257 | ||
11258 | return bdesc[code].fndecl; | |
11259 | } | |
11260 | ||
f676971a EC |
11261 | /* Implements target hook vector_mode_supported_p. */ |
11262 | bool | |
11263 | sh_vector_mode_supported_p (enum machine_mode mode) | |
11264 | { | |
11265 | if (TARGET_FPU_ANY | |
11266 | && ((mode == V2SFmode) | |
11267 | || (mode == V4SFmode) | |
11268 | || (mode == V16SFmode))) | |
11269 | return true; | |
11270 | ||
11271 | else if (TARGET_SHMEDIA | |
11272 | && ((mode == V8QImode) | |
11273 | || (mode == V2HImode) | |
11274 | || (mode == V4HImode) | |
11275 | || (mode == V2SImode))) | |
11276 | return true; | |
11277 | ||
11278 | return false; | |
11279 | } | |
11280 | ||
7a296495 CB |
11281 | bool |
11282 | sh_frame_pointer_required (void) | |
11283 | { | |
11284 | /* If needed override this in other tm.h files to cope with various OS | |
11285 | lossage requiring a frame pointer. */ | |
11286 | if (SUBTARGET_FRAME_POINTER_REQUIRED) | |
11287 | return true; | |
11288 | ||
11289 | if (crtl->profile) | |
11290 | return true; | |
11291 | ||
11292 | return false; | |
11293 | } | |
11294 | ||
a1c496cb EC |
11295 | /* Implements target hook dwarf_calling_convention. Return an enum |
11296 | of dwarf_calling_convention. */ | |
11297 | int | |
3101faab | 11298 | sh_dwarf_calling_convention (const_tree func) |
a1c496cb EC |
11299 | { |
11300 | if (sh_attr_renesas_p (func)) | |
29832164 | 11301 | return DW_CC_GNU_renesas_sh; |
a1c496cb EC |
11302 | |
11303 | return DW_CC_normal; | |
11304 | } | |
11305 | ||
c1b92d09 | 11306 | static void |
cf277499 | 11307 | sh_init_builtins (void) |
c1b92d09 R |
11308 | { |
11309 | if (TARGET_SHMEDIA) | |
11310 | sh_media_init_builtins (); | |
11311 | } | |
11312 | ||
1cc24576 KK |
11313 | /* Returns the sh builtin decl for CODE. */ |
11314 | ||
11315 | static tree | |
11316 | sh_builtin_decl (unsigned code, bool initialize_p ATTRIBUTE_UNUSED) | |
11317 | { | |
11318 | if (TARGET_SHMEDIA) | |
11319 | return sh_media_builtin_decl (code, initialize_p); | |
11320 | ||
11321 | return error_mark_node; | |
11322 | } | |
11323 | ||
c1b92d09 R |
11324 | /* Expand an expression EXP that calls a built-in function, |
11325 | with result going to TARGET if that's convenient | |
11326 | (and in mode MODE if that's convenient). | |
11327 | SUBTARGET may be used as the target for computing one of EXP's operands. | |
11328 | IGNORE is nonzero if the value is to be ignored. */ | |
11329 | ||
11330 | static rtx | |
cf277499 SB |
11331 | sh_expand_builtin (tree exp, rtx target, rtx subtarget ATTRIBUTE_UNUSED, |
11332 | enum machine_mode mode ATTRIBUTE_UNUSED, int ignore) | |
c1b92d09 | 11333 | { |
5039610b | 11334 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); |
c1b92d09 R |
11335 | unsigned int fcode = DECL_FUNCTION_CODE (fndecl); |
11336 | const struct builtin_description *d = &bdesc[fcode]; | |
11337 | enum insn_code icode = d->icode; | |
11338 | int signature = d->signature; | |
11339 | enum machine_mode tmode = VOIDmode; | |
11340 | int nop = 0, i; | |
11341 | rtx op[4]; | |
73a4d10b | 11342 | rtx pat = 0; |
c1b92d09 R |
11343 | |
11344 | if (signature_args[signature][0]) | |
11345 | { | |
11346 | if (ignore) | |
11347 | return 0; | |
11348 | ||
11349 | tmode = insn_data[icode].operand[0].mode; | |
11350 | if (! target | |
11351 | || GET_MODE (target) != tmode | |
11352 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
11353 | target = gen_reg_rtx (tmode); | |
11354 | op[nop++] = target; | |
11355 | } | |
11356 | else | |
11357 | target = 0; | |
11358 | ||
11359 | for (i = 1; i <= 3; i++, nop++) | |
11360 | { | |
11361 | tree arg; | |
11362 | enum machine_mode opmode, argmode; | |
73a4d10b | 11363 | tree optype; |
c1b92d09 R |
11364 | |
11365 | if (! signature_args[signature][i]) | |
11366 | break; | |
5039610b | 11367 | arg = CALL_EXPR_ARG (exp, i - 1); |
3578cf63 R |
11368 | if (arg == error_mark_node) |
11369 | return const0_rtx; | |
73a4d10b R |
11370 | if (signature_args[signature][i] & 8) |
11371 | { | |
11372 | opmode = ptr_mode; | |
11373 | optype = ptr_type_node; | |
11374 | } | |
11375 | else | |
11376 | { | |
11377 | opmode = insn_data[icode].operand[nop].mode; | |
11378 | optype = (*lang_hooks.types.type_for_mode) (opmode, 0); | |
11379 | } | |
c1b92d09 R |
11380 | argmode = TYPE_MODE (TREE_TYPE (arg)); |
11381 | if (argmode != opmode) | |
73a4d10b | 11382 | arg = build1 (NOP_EXPR, optype, arg); |
b757d36f | 11383 | op[nop] = expand_expr (arg, NULL_RTX, opmode, EXPAND_NORMAL); |
c1b92d09 R |
11384 | if (! (*insn_data[icode].operand[nop].predicate) (op[nop], opmode)) |
11385 | op[nop] = copy_to_mode_reg (opmode, op[nop]); | |
11386 | } | |
11387 | ||
11388 | switch (nop) | |
11389 | { | |
11390 | case 1: | |
11391 | pat = (*insn_data[d->icode].genfun) (op[0]); | |
11392 | break; | |
11393 | case 2: | |
11394 | pat = (*insn_data[d->icode].genfun) (op[0], op[1]); | |
11395 | break; | |
11396 | case 3: | |
11397 | pat = (*insn_data[d->icode].genfun) (op[0], op[1], op[2]); | |
11398 | break; | |
11399 | case 4: | |
11400 | pat = (*insn_data[d->icode].genfun) (op[0], op[1], op[2], op[3]); | |
11401 | break; | |
95c0af87 | 11402 | default: |
f5b9e7c9 | 11403 | gcc_unreachable (); |
c1b92d09 R |
11404 | } |
11405 | if (! pat) | |
11406 | return 0; | |
11407 | emit_insn (pat); | |
11408 | return target; | |
11409 | } | |
0ac78517 R |
11410 | |
11411 | void | |
cf277499 | 11412 | sh_expand_unop_v2sf (enum rtx_code code, rtx op0, rtx op1) |
0ac78517 R |
11413 | { |
11414 | rtx sel0 = const0_rtx; | |
11415 | rtx sel1 = const1_rtx; | |
cf277499 | 11416 | rtx (*fn) (rtx, rtx, rtx, rtx, rtx) = gen_unary_sf_op; |
0ac78517 R |
11417 | rtx op = gen_rtx_fmt_e (code, SFmode, op1); |
11418 | ||
11419 | emit_insn ((*fn) (op0, op1, op, sel0, sel0)); | |
11420 | emit_insn ((*fn) (op0, op1, op, sel1, sel1)); | |
11421 | } | |
11422 | ||
11423 | void | |
cf277499 | 11424 | sh_expand_binop_v2sf (enum rtx_code code, rtx op0, rtx op1, rtx op2) |
0ac78517 | 11425 | { |
0ac78517 R |
11426 | rtx op = gen_rtx_fmt_ee (code, SFmode, op1, op2); |
11427 | ||
f8be3f56 KK |
11428 | emit_insn (gen_binary_sf_op0 (op0, op1, op2, op)); |
11429 | emit_insn (gen_binary_sf_op1 (op0, op1, op2, op)); | |
0ac78517 R |
11430 | } |
11431 | ||
347146a4 KK |
11432 | /* Return true if hard register REGNO can hold a value of machine-mode MODE. |
11433 | We can allow any mode in any general register. The special registers | |
11434 | only allow SImode. Don't allow any mode in the PR. | |
11435 | ||
11436 | We cannot hold DCmode values in the XD registers because alter_reg | |
11437 | handles subregs of them incorrectly. We could work around this by | |
11438 | spacing the XD registers like the DR registers, but this would require | |
11439 | additional memory in every compilation to hold larger register vectors. | |
11440 | We could hold SFmode / SCmode values in XD registers, but that | |
11441 | would require a tertiary reload when reloading from / to memory, | |
11442 | and a secondary reload to reload from / to general regs; that | |
11443 | seems to be a loosing proposition. | |
11444 | ||
11445 | We want to allow TImode FP regs so that when V4SFmode is loaded as TImode, | |
11446 | it won't be ferried through GP registers first. */ | |
11447 | ||
11448 | bool | |
11449 | sh_hard_regno_mode_ok (unsigned int regno, enum machine_mode mode) | |
11450 | { | |
11451 | if (SPECIAL_REGISTER_P (regno)) | |
11452 | return mode == SImode; | |
11453 | ||
11454 | if (regno == FPUL_REG) | |
11455 | return (mode == SImode || mode == SFmode); | |
11456 | ||
11457 | if (FP_REGISTER_P (regno) && mode == SFmode) | |
11458 | return true; | |
11459 | ||
11460 | if (mode == V2SFmode) | |
11461 | { | |
11462 | if (((FP_REGISTER_P (regno) && (regno - FIRST_FP_REG) % 2 == 0) | |
11463 | || GENERAL_REGISTER_P (regno))) | |
11464 | return true; | |
11465 | else | |
11466 | return false; | |
11467 | } | |
11468 | ||
11469 | if (mode == V4SFmode) | |
11470 | { | |
11471 | if ((FP_REGISTER_P (regno) && (regno - FIRST_FP_REG) % 4 == 0) | |
11472 | || GENERAL_REGISTER_P (regno)) | |
11473 | return true; | |
11474 | else | |
11475 | return false; | |
11476 | } | |
11477 | ||
11478 | if (mode == V16SFmode) | |
11479 | { | |
11480 | if (TARGET_SHMEDIA) | |
11481 | { | |
11482 | if (FP_REGISTER_P (regno) && (regno - FIRST_FP_REG) % 16 == 0) | |
11483 | return true; | |
11484 | else | |
11485 | return false; | |
11486 | } | |
11487 | else | |
11488 | return regno == FIRST_XD_REG; | |
11489 | } | |
11490 | ||
11491 | if (FP_REGISTER_P (regno)) | |
11492 | { | |
11493 | if (mode == SFmode | |
11494 | || mode == SImode | |
11495 | || ((TARGET_SH2E || TARGET_SHMEDIA) && mode == SCmode) | |
11496 | || ((((TARGET_SH4 || TARGET_SH2A_DOUBLE) && mode == DFmode) | |
11497 | || mode == DCmode | |
11498 | || (TARGET_SHMEDIA | |
11499 | && (mode == DFmode || mode == DImode | |
11500 | || mode == V2SFmode || mode == TImode))) | |
11501 | && ((regno - FIRST_FP_REG) & 1) == 0) | |
11502 | || ((TARGET_SH4 || TARGET_SHMEDIA) && mode == TImode | |
11503 | && ((regno - FIRST_FP_REG) & 3) == 0)) | |
11504 | return true; | |
11505 | else | |
11506 | return false; | |
11507 | } | |
11508 | ||
11509 | if (XD_REGISTER_P (regno)) | |
11510 | return mode == DFmode; | |
11511 | ||
11512 | if (TARGET_REGISTER_P (regno)) | |
11513 | return (mode == DImode || mode == SImode || mode == PDImode); | |
11514 | ||
11515 | if (regno == PR_REG) | |
11516 | return mode == SImode; | |
11517 | ||
11518 | if (regno == FPSCR_REG) | |
11519 | return mode == PSImode; | |
11520 | ||
11521 | /* FIXME. This works around PR target/37633 for -O0. */ | |
11522 | if (!optimize && TARGET_SHMEDIA32 && GET_MODE_SIZE (mode) > 4) | |
11523 | { | |
11524 | unsigned int n = GET_MODE_SIZE (mode) / 8; | |
11525 | ||
11526 | if (regno >= FIRST_GENERAL_REG + 10 - n + 1 | |
11527 | && regno <= FIRST_GENERAL_REG + 14) | |
11528 | return false; | |
11529 | } | |
11530 | ||
11531 | return true; | |
11532 | } | |
11533 | ||
cff9f8d5 AH |
11534 | /* Return the class of registers for which a mode change from FROM to TO |
11535 | is invalid. */ | |
b0c42aed | 11536 | bool |
cf277499 | 11537 | sh_cannot_change_mode_class (enum machine_mode from, enum machine_mode to, |
0a2aaacc | 11538 | enum reg_class rclass) |
cff9f8d5 | 11539 | { |
312209c6 AO |
11540 | /* We want to enable the use of SUBREGs as a means to |
11541 | VEC_SELECT a single element of a vector. */ | |
11542 | if (to == SFmode && VECTOR_MODE_P (from) && GET_MODE_INNER (from) == SFmode) | |
0a2aaacc | 11543 | return (reg_classes_intersect_p (GENERAL_REGS, rclass)); |
312209c6 | 11544 | |
cff9f8d5 AH |
11545 | if (GET_MODE_SIZE (from) != GET_MODE_SIZE (to)) |
11546 | { | |
832a3292 KH |
11547 | if (TARGET_LITTLE_ENDIAN) |
11548 | { | |
11549 | if (GET_MODE_SIZE (to) < 8 || GET_MODE_SIZE (from) < 8) | |
0a2aaacc | 11550 | return reg_classes_intersect_p (DF_REGS, rclass); |
832a3292 KH |
11551 | } |
11552 | else | |
11553 | { | |
11554 | if (GET_MODE_SIZE (from) < 8) | |
0a2aaacc | 11555 | return reg_classes_intersect_p (DF_HI_REGS, rclass); |
832a3292 | 11556 | } |
cff9f8d5 | 11557 | } |
b0c42aed | 11558 | return 0; |
cff9f8d5 AH |
11559 | } |
11560 | ||
42db504c SB |
11561 | /* Return true if registers in machine mode MODE will likely be |
11562 | allocated to registers in small register classes. */ | |
11563 | ||
9c265c4d | 11564 | bool |
42db504c SB |
11565 | sh_small_register_classes_for_mode_p (enum machine_mode mode ATTRIBUTE_UNUSED) |
11566 | { | |
11567 | return (! TARGET_SHMEDIA); | |
11568 | } | |
ea4210ef RS |
11569 | |
11570 | /* If ADDRESS refers to a CODE_LABEL, add NUSES to the number of times | |
11571 | that label is used. */ | |
11572 | ||
11573 | void | |
cf277499 | 11574 | sh_mark_label (rtx address, int nuses) |
ea4210ef RS |
11575 | { |
11576 | if (GOTOFF_P (address)) | |
11577 | { | |
11578 | /* Extract the label or symbol. */ | |
11579 | address = XEXP (address, 0); | |
11580 | if (GET_CODE (address) == PLUS) | |
11581 | address = XEXP (address, 0); | |
11582 | address = XVECEXP (address, 0, 0); | |
11583 | } | |
11584 | if (GET_CODE (address) == LABEL_REF | |
f3536097 | 11585 | && LABEL_P (XEXP (address, 0))) |
ea4210ef RS |
11586 | LABEL_NUSES (XEXP (address, 0)) += nuses; |
11587 | } | |
11588 | ||
95c0af87 JR |
11589 | /* Compute extra cost of moving data between one register class |
11590 | and another. */ | |
11591 | ||
ec555f32 R |
11592 | /* If SECONDARY*_RELOAD_CLASS says something about the src/dst pair, regclass |
11593 | uses this information. Hence, the general register <-> floating point | |
95c0af87 JR |
11594 | register information here is not used for SFmode. */ |
11595 | ||
e7a8f681 | 11596 | static int |
cf277499 | 11597 | sh_register_move_cost (enum machine_mode mode, |
e7a8f681 | 11598 | reg_class_t srcclass, reg_class_t dstclass) |
95c0af87 JR |
11599 | { |
11600 | if (dstclass == T_REGS || dstclass == PR_REGS) | |
11601 | return 10; | |
11602 | ||
b1dcf523 DD |
11603 | if (dstclass == MAC_REGS && srcclass == MAC_REGS) |
11604 | return 4; | |
11605 | ||
ec555f32 R |
11606 | if (mode == SImode && ! TARGET_SHMEDIA && TARGET_FMOVD |
11607 | && REGCLASS_HAS_FP_REG (srcclass) | |
11608 | && REGCLASS_HAS_FP_REG (dstclass)) | |
11609 | return 4; | |
11610 | ||
9c24200c KK |
11611 | if (REGCLASS_HAS_FP_REG (dstclass) && srcclass == T_REGS) |
11612 | return ((TARGET_HARD_SH4 && !optimize_size) ? 10 : 7); | |
11613 | ||
a60a9553 | 11614 | if ((REGCLASS_HAS_FP_REG (dstclass) && srcclass == MAC_REGS) |
832a3292 | 11615 | || (dstclass == MAC_REGS && REGCLASS_HAS_FP_REG (srcclass))) |
a60a9553 R |
11616 | return 9; |
11617 | ||
95c0af87 JR |
11618 | if ((REGCLASS_HAS_FP_REG (dstclass) |
11619 | && REGCLASS_HAS_GENERAL_REG (srcclass)) | |
11620 | || (REGCLASS_HAS_GENERAL_REG (dstclass) | |
11621 | && REGCLASS_HAS_FP_REG (srcclass))) | |
832a3292 KH |
11622 | return ((TARGET_SHMEDIA ? 4 : TARGET_FMOVD ? 8 : 12) |
11623 | * ((GET_MODE_SIZE (mode) + 7) / 8U)); | |
95c0af87 JR |
11624 | |
11625 | if ((dstclass == FPUL_REGS | |
11626 | && REGCLASS_HAS_GENERAL_REG (srcclass)) | |
11627 | || (srcclass == FPUL_REGS | |
11628 | && REGCLASS_HAS_GENERAL_REG (dstclass))) | |
11629 | return 5; | |
11630 | ||
11631 | if ((dstclass == FPUL_REGS | |
11632 | && (srcclass == PR_REGS || srcclass == MAC_REGS || srcclass == T_REGS)) | |
f676971a | 11633 | || (srcclass == FPUL_REGS |
95c0af87 JR |
11634 | && (dstclass == PR_REGS || dstclass == MAC_REGS))) |
11635 | return 7; | |
11636 | ||
11637 | if ((srcclass == TARGET_REGS && ! REGCLASS_HAS_GENERAL_REG (dstclass)) | |
11638 | || ((dstclass) == TARGET_REGS && ! REGCLASS_HAS_GENERAL_REG (srcclass))) | |
11639 | return 20; | |
11640 | ||
73a4d10b R |
11641 | /* ??? ptabs faults on (value & 0x3) == 0x3 */ |
11642 | if (TARGET_SHMEDIA | |
11643 | && ((srcclass) == TARGET_REGS || (srcclass) == SIBCALL_REGS)) | |
11644 | { | |
7dd2f19b RS |
11645 | if (sh_gettrcost >= 0) |
11646 | return sh_gettrcost; | |
73a4d10b R |
11647 | else if (!TARGET_PT_FIXED) |
11648 | return 100; | |
11649 | } | |
11650 | ||
95c0af87 JR |
11651 | if ((srcclass == FPSCR_REGS && ! REGCLASS_HAS_GENERAL_REG (dstclass)) |
11652 | || (dstclass == FPSCR_REGS && ! REGCLASS_HAS_GENERAL_REG (srcclass))) | |
11653 | return 4; | |
11654 | ||
ec555f32 R |
11655 | if (TARGET_SHMEDIA |
11656 | || (TARGET_FMOVD | |
11657 | && ! REGCLASS_HAS_GENERAL_REG (srcclass) | |
11658 | && ! REGCLASS_HAS_GENERAL_REG (dstclass))) | |
11659 | return 2 * ((GET_MODE_SIZE (mode) + 7) / 8U); | |
11660 | ||
11661 | return 2 * ((GET_MODE_SIZE (mode) + 3) / 4U); | |
95c0af87 JR |
11662 | } |
11663 | ||
cf277499 | 11664 | static rtx emit_load_ptr (rtx, rtx); |
1aa03f38 R |
11665 | |
11666 | static rtx | |
cf277499 | 11667 | emit_load_ptr (rtx reg, rtx addr) |
1aa03f38 | 11668 | { |
57d38024 | 11669 | rtx mem = gen_const_mem (ptr_mode, addr); |
1aa03f38 R |
11670 | |
11671 | if (Pmode != ptr_mode) | |
11672 | mem = gen_rtx_SIGN_EXTEND (Pmode, mem); | |
11673 | return emit_move_insn (reg, mem); | |
11674 | } | |
11675 | ||
73a4d10b | 11676 | static void |
cf277499 SB |
11677 | sh_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED, |
11678 | HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset, | |
11679 | tree function) | |
1aa03f38 R |
11680 | { |
11681 | CUMULATIVE_ARGS cum; | |
11682 | int structure_value_byref = 0; | |
0a2aaacc | 11683 | rtx this_rtx, this_value, sibcall, insns, funexp; |
1aa03f38 | 11684 | tree funtype = TREE_TYPE (function); |
735cb76e | 11685 | int simple_add = CONST_OK_FOR_ADD (delta); |
1aa03f38 R |
11686 | int did_load = 0; |
11687 | rtx scratch0, scratch1, scratch2; | |
73a4d10b | 11688 | unsigned i; |
1aa03f38 R |
11689 | |
11690 | reload_completed = 1; | |
fe3ad572 | 11691 | epilogue_completed = 1; |
1aa03f38 R |
11692 | current_function_uses_only_leaf_regs = 1; |
11693 | ||
2e040219 | 11694 | emit_note (NOTE_INSN_PROLOGUE_END); |
1aa03f38 R |
11695 | |
11696 | /* Find the "this" pointer. We have such a wide range of ABIs for the | |
11697 | SH that it's best to do this completely machine independently. | |
f676971a | 11698 | "this" is passed as first argument, unless a structure return pointer |
1aa03f38 | 11699 | comes first, in which case "this" comes second. */ |
0f6937fe | 11700 | INIT_CUMULATIVE_ARGS (cum, funtype, NULL_RTX, 0, 1); |
1aa03f38 | 11701 | #ifndef PCC_STATIC_STRUCT_RETURN |
61f71b34 | 11702 | if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function)) |
1aa03f38 R |
11703 | structure_value_byref = 1; |
11704 | #endif /* not PCC_STATIC_STRUCT_RETURN */ | |
61f71b34 | 11705 | if (structure_value_byref && sh_struct_value_rtx (function, 0) == 0) |
f676971a | 11706 | { |
1aa03f38 R |
11707 | tree ptype = build_pointer_type (TREE_TYPE (funtype)); |
11708 | ||
b64336b8 | 11709 | sh_function_arg_advance (&cum, Pmode, ptype, true); |
1aa03f38 | 11710 | } |
b64336b8 | 11711 | this_rtx = sh_function_arg (&cum, Pmode, ptr_type_node, true); |
1aa03f38 R |
11712 | |
11713 | /* For SHcompact, we only have r0 for a scratch register: r1 is the | |
11714 | static chain pointer (even if you can't have nested virtual functions | |
11715 | right now, someone might implement them sometime), and the rest of the | |
11716 | registers are used for argument passing, are callee-saved, or reserved. */ | |
73a4d10b R |
11717 | /* We need to check call_used_regs / fixed_regs in case -fcall_saved-reg / |
11718 | -ffixed-reg has been used. */ | |
11719 | if (! call_used_regs[0] || fixed_regs[0]) | |
11720 | error ("r0 needs to be available as a call-clobbered register"); | |
1aa03f38 R |
11721 | scratch0 = scratch1 = scratch2 = gen_rtx_REG (Pmode, 0); |
11722 | if (! TARGET_SH5) | |
11723 | { | |
73a4d10b R |
11724 | if (call_used_regs[1] && ! fixed_regs[1]) |
11725 | scratch1 = gen_rtx_REG (ptr_mode, 1); | |
1aa03f38 R |
11726 | /* N.B., if not TARGET_HITACHI, register 2 is used to pass the pointer |
11727 | pointing where to return struct values. */ | |
73a4d10b R |
11728 | if (call_used_regs[3] && ! fixed_regs[3]) |
11729 | scratch2 = gen_rtx_REG (Pmode, 3); | |
1aa03f38 R |
11730 | } |
11731 | else if (TARGET_SHMEDIA) | |
11732 | { | |
73a4d10b R |
11733 | for (i = FIRST_GENERAL_REG; i <= LAST_GENERAL_REG; i++) |
11734 | if (i != REGNO (scratch0) && | |
11735 | call_used_regs[i] && ! fixed_regs[i] && ! FUNCTION_ARG_REGNO_P (i)) | |
11736 | { | |
11737 | scratch1 = gen_rtx_REG (ptr_mode, i); | |
11738 | break; | |
11739 | } | |
11740 | if (scratch1 == scratch0) | |
d8a07487 | 11741 | error ("need a second call-clobbered general purpose register"); |
73a4d10b R |
11742 | for (i = FIRST_TARGET_REG; i <= LAST_TARGET_REG; i++) |
11743 | if (call_used_regs[i] && ! fixed_regs[i]) | |
11744 | { | |
11745 | scratch2 = gen_rtx_REG (Pmode, i); | |
11746 | break; | |
11747 | } | |
11748 | if (scratch2 == scratch0) | |
d8a07487 | 11749 | error ("need a call-clobbered target register"); |
1aa03f38 R |
11750 | } |
11751 | ||
0a2aaacc | 11752 | this_value = plus_constant (this_rtx, delta); |
1aa03f38 R |
11753 | if (vcall_offset |
11754 | && (simple_add || scratch0 != scratch1) | |
11755 | && strict_memory_address_p (ptr_mode, this_value)) | |
11756 | { | |
11757 | emit_load_ptr (scratch0, this_value); | |
11758 | did_load = 1; | |
11759 | } | |
11760 | ||
11761 | if (!delta) | |
11762 | ; /* Do nothing. */ | |
11763 | else if (simple_add) | |
0a2aaacc | 11764 | emit_move_insn (this_rtx, this_value); |
1aa03f38 R |
11765 | else |
11766 | { | |
11767 | emit_move_insn (scratch1, GEN_INT (delta)); | |
0a2aaacc | 11768 | emit_insn (gen_add2_insn (this_rtx, scratch1)); |
1aa03f38 R |
11769 | } |
11770 | ||
11771 | if (vcall_offset) | |
11772 | { | |
11773 | rtx offset_addr; | |
11774 | ||
11775 | if (!did_load) | |
0a2aaacc | 11776 | emit_load_ptr (scratch0, this_rtx); |
1aa03f38 R |
11777 | |
11778 | offset_addr = plus_constant (scratch0, vcall_offset); | |
11779 | if (strict_memory_address_p (ptr_mode, offset_addr)) | |
11780 | ; /* Do nothing. */ | |
73a4d10b | 11781 | else if (! TARGET_SH5 && scratch0 != scratch1) |
1aa03f38 R |
11782 | { |
11783 | /* scratch0 != scratch1, and we have indexed loads. Get better | |
11784 | schedule by loading the offset into r1 and using an indexed | |
11785 | load - then the load of r1 can issue before the load from | |
0a2aaacc | 11786 | (this_rtx + delta) finishes. */ |
1aa03f38 R |
11787 | emit_move_insn (scratch1, GEN_INT (vcall_offset)); |
11788 | offset_addr = gen_rtx_PLUS (Pmode, scratch0, scratch1); | |
11789 | } | |
735cb76e | 11790 | else if (CONST_OK_FOR_ADD (vcall_offset)) |
1aa03f38 R |
11791 | { |
11792 | emit_insn (gen_add2_insn (scratch0, GEN_INT (vcall_offset))); | |
11793 | offset_addr = scratch0; | |
11794 | } | |
11795 | else if (scratch0 != scratch1) | |
11796 | { | |
11797 | emit_move_insn (scratch1, GEN_INT (vcall_offset)); | |
11798 | emit_insn (gen_add2_insn (scratch0, scratch1)); | |
11799 | offset_addr = scratch0; | |
11800 | } | |
11801 | else | |
f5b9e7c9 | 11802 | gcc_unreachable (); /* FIXME */ |
1aa03f38 R |
11803 | emit_load_ptr (scratch0, offset_addr); |
11804 | ||
832a3292 | 11805 | if (Pmode != ptr_mode) |
1aa03f38 | 11806 | scratch0 = gen_rtx_TRUNCATE (ptr_mode, scratch0); |
0a2aaacc | 11807 | emit_insn (gen_add2_insn (this_rtx, scratch0)); |
1aa03f38 R |
11808 | } |
11809 | ||
11810 | /* Generate a tail call to the target function. */ | |
11811 | if (! TREE_USED (function)) | |
11812 | { | |
11813 | assemble_external (function); | |
11814 | TREE_USED (function) = 1; | |
11815 | } | |
11816 | funexp = XEXP (DECL_RTL (function), 0); | |
73a4d10b R |
11817 | /* If the function is overridden, so is the thunk, hence we don't |
11818 | need GOT addressing even if this is a public symbol. */ | |
11819 | #if 0 | |
11820 | if (TARGET_SH1 && ! flag_weak) | |
11821 | sibcall = gen_sibcalli_thunk (funexp, const0_rtx); | |
11822 | else | |
11823 | #endif | |
11824 | if (TARGET_SH2 && flag_pic) | |
11825 | { | |
11826 | sibcall = gen_sibcall_pcrel (funexp, const0_rtx); | |
11827 | XEXP (XVECEXP (sibcall, 0, 2), 0) = scratch2; | |
11828 | } | |
11829 | else | |
11830 | { | |
11831 | if (TARGET_SHMEDIA && flag_pic) | |
11832 | { | |
11833 | funexp = gen_sym2PIC (funexp); | |
11834 | PUT_MODE (funexp, Pmode); | |
11835 | } | |
11836 | emit_move_insn (scratch2, funexp); | |
11837 | funexp = gen_rtx_MEM (FUNCTION_MODE, scratch2); | |
11838 | sibcall = gen_sibcall (funexp, const0_rtx, NULL_RTX); | |
11839 | } | |
11840 | sibcall = emit_call_insn (sibcall); | |
1aa03f38 | 11841 | SIBLING_CALL_P (sibcall) = 1; |
0a2aaacc | 11842 | use_reg (&CALL_INSN_FUNCTION_USAGE (sibcall), this_rtx); |
1aa03f38 R |
11843 | emit_barrier (); |
11844 | ||
a2855205 | 11845 | /* Run just enough of rest_of_compilation to do scheduling and get |
1aa03f38 R |
11846 | the insns emitted. Note that use_thunk calls |
11847 | assemble_start_function and assemble_end_function. */ | |
a2855205 | 11848 | |
55e092c4 | 11849 | insn_locators_alloc (); |
1aa03f38 R |
11850 | insns = get_insns (); |
11851 | ||
6fb5fa3c DB |
11852 | if (optimize > 0) |
11853 | { | |
11854 | if (! cfun->cfg) | |
eb6cbf91 | 11855 | init_flow (cfun); |
6fb5fa3c DB |
11856 | split_all_insns_noflow (); |
11857 | } | |
1aa03f38 | 11858 | |
18dbd950 | 11859 | sh_reorg (); |
1aa03f38 R |
11860 | |
11861 | if (optimize > 0 && flag_delayed_branch) | |
10d22567 | 11862 | dbr_schedule (insns); |
384a8760 | 11863 | |
1aa03f38 R |
11864 | shorten_branches (insns); |
11865 | final_start_function (insns, file, 1); | |
c9d691e9 | 11866 | final (insns, file, 1); |
1aa03f38 R |
11867 | final_end_function (); |
11868 | ||
1aa03f38 | 11869 | reload_completed = 0; |
fe3ad572 | 11870 | epilogue_completed = 0; |
1aa03f38 R |
11871 | } |
11872 | ||
90534361 | 11873 | rtx |
73a4d10b | 11874 | function_symbol (rtx target, const char *name, enum sh_function_kind kind) |
90534361 | 11875 | { |
73a4d10b R |
11876 | rtx sym; |
11877 | ||
11878 | /* If this is not an ordinary function, the name usually comes from a | |
11879 | string literal or an sprintf buffer. Make sure we use the same | |
11880 | string consistently, so that cse will be able to unify address loads. */ | |
11881 | if (kind != FUNCTION_ORDINARY) | |
11882 | name = IDENTIFIER_POINTER (get_identifier (name)); | |
11883 | sym = gen_rtx_SYMBOL_REF (Pmode, name); | |
90534361 | 11884 | SYMBOL_REF_FLAGS (sym) = SYMBOL_FLAG_FUNCTION; |
73a4d10b R |
11885 | if (flag_pic) |
11886 | switch (kind) | |
11887 | { | |
11888 | case FUNCTION_ORDINARY: | |
11889 | break; | |
11890 | case SFUNC_GOT: | |
11891 | { | |
11892 | rtx reg = target ? target : gen_reg_rtx (Pmode); | |
11893 | ||
11894 | emit_insn (gen_symGOT2reg (reg, sym)); | |
11895 | sym = reg; | |
11896 | break; | |
11897 | } | |
11898 | case SFUNC_STATIC: | |
11899 | { | |
11900 | /* ??? To allow cse to work, we use GOTOFF relocations. | |
11901 | we could add combiner patterns to transform this into | |
11902 | straight pc-relative calls with sym2PIC / bsrf when | |
11903 | label load and function call are still 1:1 and in the | |
11904 | same basic block during combine. */ | |
11905 | rtx reg = target ? target : gen_reg_rtx (Pmode); | |
11906 | ||
11907 | emit_insn (gen_symGOTOFF2reg (reg, sym)); | |
11908 | sym = reg; | |
11909 | break; | |
11910 | } | |
11911 | } | |
11912 | if (target && sym != target) | |
11913 | { | |
11914 | emit_move_insn (target, sym); | |
11915 | return target; | |
11916 | } | |
90534361 R |
11917 | return sym; |
11918 | } | |
11919 | ||
fada1961 R |
11920 | /* Find the number of a general purpose register in S. */ |
11921 | static int | |
11922 | scavenge_reg (HARD_REG_SET *s) | |
11923 | { | |
11924 | int r; | |
11925 | for (r = FIRST_GENERAL_REG; r <= LAST_GENERAL_REG; r++) | |
11926 | if (TEST_HARD_REG_BIT (*s, r)) | |
11927 | return r; | |
11928 | return -1; | |
11929 | } | |
11930 | ||
11931 | rtx | |
11932 | sh_get_pr_initial_val (void) | |
11933 | { | |
7d73a2ba R |
11934 | rtx val; |
11935 | ||
b869f904 R |
11936 | /* ??? Unfortunately, get_hard_reg_initial_val doesn't always work for the |
11937 | PR register on SHcompact, because it might be clobbered by the prologue. | |
7d73a2ba | 11938 | We check first if that is known to be the case. */ |
b869f904 | 11939 | if (TARGET_SHCOMPACT |
38173d38 | 11940 | && ((crtl->args.info.call_cookie |
7d73a2ba | 11941 | & ~ CALL_COOKIE_RET_TRAMP (1)) |
e3b5732b | 11942 | || crtl->saves_all_registers)) |
57d38024 | 11943 | return gen_frame_mem (SImode, return_address_pointer_rtx); |
7d73a2ba R |
11944 | |
11945 | /* If we haven't finished rtl generation, there might be a nonlocal label | |
11946 | that we haven't seen yet. | |
b3a13419 ILT |
11947 | ??? get_hard_reg_initial_val fails if it is called after register |
11948 | allocation has started, unless it has been called before for the | |
11949 | same register. And even then, we end in trouble if we didn't use | |
11950 | the register in the same basic block before. So call | |
11951 | get_hard_reg_initial_val now and wrap it in an unspec if we might | |
11952 | need to replace it. */ | |
dce20bbc R |
11953 | /* ??? We also must do this for TARGET_SH1 in general, because otherwise |
11954 | combine can put the pseudo returned by get_hard_reg_initial_val into | |
11955 | instructions that need a general purpose registers, which will fail to | |
11956 | be recognized when the pseudo becomes allocated to PR. */ | |
7d73a2ba R |
11957 | val |
11958 | = get_hard_reg_initial_val (Pmode, TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG); | |
dce20bbc | 11959 | if (TARGET_SH1) |
7d73a2ba R |
11960 | return gen_rtx_UNSPEC (SImode, gen_rtvec (1, val), UNSPEC_RA); |
11961 | return val; | |
fada1961 R |
11962 | } |
11963 | ||
3db1b434 | 11964 | int |
f90b7a5a | 11965 | sh_expand_t_scc (rtx operands[]) |
3db1b434 | 11966 | { |
f90b7a5a PB |
11967 | enum rtx_code code = GET_CODE (operands[1]); |
11968 | rtx target = operands[0]; | |
11969 | rtx op0 = operands[2]; | |
11970 | rtx op1 = operands[3]; | |
3db1b434 R |
11971 | rtx result = target; |
11972 | HOST_WIDE_INT val; | |
11973 | ||
f3536097 SZ |
11974 | if (!REG_P (op0) || REGNO (op0) != T_REG |
11975 | || !CONST_INT_P (op1)) | |
3db1b434 | 11976 | return 0; |
f3536097 | 11977 | if (!REG_P (result)) |
3db1b434 | 11978 | result = gen_reg_rtx (SImode); |
f90b7a5a | 11979 | val = INTVAL (op1); |
3db1b434 R |
11980 | if ((code == EQ && val == 1) || (code == NE && val == 0)) |
11981 | emit_insn (gen_movt (result)); | |
f326a6cb N |
11982 | else if (TARGET_SH2A && ((code == EQ && val == 0) |
11983 | || (code == NE && val == 1))) | |
c185de40 | 11984 | emit_insn (gen_xorsi3_movrt (result)); |
3db1b434 R |
11985 | else if ((code == EQ && val == 0) || (code == NE && val == 1)) |
11986 | { | |
c41c1387 | 11987 | emit_clobber (result); |
3db1b434 | 11988 | emit_insn (gen_subc (result, result, result)); |
a556fd39 | 11989 | emit_insn (gen_addsi3 (result, result, const1_rtx)); |
3db1b434 R |
11990 | } |
11991 | else if (code == EQ || code == NE) | |
11992 | emit_insn (gen_move_insn (result, GEN_INT (code == NE))); | |
11993 | else | |
11994 | return 0; | |
11995 | if (result != target) | |
11996 | emit_move_insn (target, result); | |
11997 | return 1; | |
11998 | } | |
11999 | ||
07d7d2f4 R |
12000 | /* INSN is an sfunc; return the rtx that describes the address used. */ |
12001 | static rtx | |
12002 | extract_sfunc_addr (rtx insn) | |
12003 | { | |
12004 | rtx pattern, part = NULL_RTX; | |
12005 | int len, i; | |
12006 | ||
12007 | pattern = PATTERN (insn); | |
12008 | len = XVECLEN (pattern, 0); | |
12009 | for (i = 0; i < len; i++) | |
12010 | { | |
12011 | part = XVECEXP (pattern, 0, i); | |
12012 | if (GET_CODE (part) == USE && GET_MODE (XEXP (part, 0)) == Pmode | |
12013 | && GENERAL_REGISTER_P (true_regnum (XEXP (part, 0)))) | |
12014 | return XEXP (part, 0); | |
12015 | } | |
f5b9e7c9 NS |
12016 | gcc_assert (GET_CODE (XVECEXP (pattern, 0, 0)) == UNSPEC_VOLATILE); |
12017 | return XVECEXP (XVECEXP (pattern, 0, 0), 0, 1); | |
07d7d2f4 R |
12018 | } |
12019 | ||
12020 | /* Verify that the register in use_sfunc_addr still agrees with the address | |
12021 | used in the sfunc. This prevents fill_slots_from_thread from changing | |
12022 | use_sfunc_addr. | |
12023 | INSN is the use_sfunc_addr instruction, and REG is the register it | |
12024 | guards. */ | |
12025 | int | |
12026 | check_use_sfunc_addr (rtx insn, rtx reg) | |
12027 | { | |
12028 | /* Search for the sfunc. It should really come right after INSN. */ | |
12029 | while ((insn = NEXT_INSN (insn))) | |
12030 | { | |
f3536097 | 12031 | if (LABEL_P (insn) || JUMP_P (insn)) |
07d7d2f4 R |
12032 | break; |
12033 | if (! INSN_P (insn)) | |
12034 | continue; | |
f676971a | 12035 | |
07d7d2f4 R |
12036 | if (GET_CODE (PATTERN (insn)) == SEQUENCE) |
12037 | insn = XVECEXP (PATTERN (insn), 0, 0); | |
12038 | if (GET_CODE (PATTERN (insn)) != PARALLEL | |
12039 | || get_attr_type (insn) != TYPE_SFUNC) | |
12040 | continue; | |
12041 | return rtx_equal_p (extract_sfunc_addr (insn), reg); | |
12042 | } | |
f5b9e7c9 | 12043 | gcc_unreachable (); |
07d7d2f4 R |
12044 | } |
12045 | ||
312209c6 AO |
12046 | /* This function returns a constant rtx that represents pi / 2**15 in |
12047 | SFmode. it's used to scale SFmode angles, in radians, to a | |
12048 | fixed-point signed 16.16-bit fraction of a full circle, i.e., 2*pi | |
12049 | maps to 0x10000). */ | |
12050 | ||
12051 | static GTY(()) rtx sh_fsca_sf2int_rtx; | |
12052 | ||
12053 | rtx | |
12054 | sh_fsca_sf2int (void) | |
12055 | { | |
12056 | if (! sh_fsca_sf2int_rtx) | |
12057 | { | |
12058 | REAL_VALUE_TYPE rv; | |
12059 | ||
12060 | real_from_string (&rv, "10430.378350470453"); | |
12061 | sh_fsca_sf2int_rtx = const_double_from_real_value (rv, SFmode); | |
12062 | } | |
12063 | ||
12064 | return sh_fsca_sf2int_rtx; | |
12065 | } | |
f676971a | 12066 | |
312209c6 AO |
12067 | /* This function returns a constant rtx that represents pi / 2**15 in |
12068 | DFmode. it's used to scale DFmode angles, in radians, to a | |
12069 | fixed-point signed 16.16-bit fraction of a full circle, i.e., 2*pi | |
12070 | maps to 0x10000). */ | |
12071 | ||
12072 | static GTY(()) rtx sh_fsca_df2int_rtx; | |
12073 | ||
12074 | rtx | |
12075 | sh_fsca_df2int (void) | |
12076 | { | |
12077 | if (! sh_fsca_df2int_rtx) | |
12078 | { | |
12079 | REAL_VALUE_TYPE rv; | |
12080 | ||
12081 | real_from_string (&rv, "10430.378350470453"); | |
12082 | sh_fsca_df2int_rtx = const_double_from_real_value (rv, DFmode); | |
12083 | } | |
12084 | ||
12085 | return sh_fsca_df2int_rtx; | |
12086 | } | |
f676971a | 12087 | |
312209c6 AO |
12088 | /* This function returns a constant rtx that represents 2**15 / pi in |
12089 | SFmode. it's used to scale a fixed-point signed 16.16-bit fraction | |
12090 | of a full circle back to a SFmode value, i.e., 0x10000 maps to | |
12091 | 2*pi). */ | |
12092 | ||
12093 | static GTY(()) rtx sh_fsca_int2sf_rtx; | |
12094 | ||
12095 | rtx | |
12096 | sh_fsca_int2sf (void) | |
12097 | { | |
12098 | if (! sh_fsca_int2sf_rtx) | |
12099 | { | |
12100 | REAL_VALUE_TYPE rv; | |
12101 | ||
12102 | real_from_string (&rv, "9.587379924285257e-5"); | |
12103 | sh_fsca_int2sf_rtx = const_double_from_real_value (rv, SFmode); | |
12104 | } | |
12105 | ||
12106 | return sh_fsca_int2sf_rtx; | |
12107 | } | |
cc15e98f | 12108 | |
8b0add27 | 12109 | /* Initialize the CUMULATIVE_ARGS structure. */ |
cc15e98f NC |
12110 | |
12111 | void | |
12112 | sh_init_cumulative_args (CUMULATIVE_ARGS * pcum, | |
12113 | tree fntype, | |
12114 | rtx libname ATTRIBUTE_UNUSED, | |
12115 | tree fndecl, | |
12116 | signed int n_named_args, | |
12117 | enum machine_mode mode) | |
12118 | { | |
12119 | pcum->arg_count [(int) SH_ARG_FLOAT] = 0; | |
12120 | pcum->free_single_fp_reg = 0; | |
12121 | pcum->stack_regs = 0; | |
12122 | pcum->byref_regs = 0; | |
12123 | pcum->byref = 0; | |
12124 | pcum->outgoing = (n_named_args == -1) ? 0 : 1; | |
12125 | ||
12126 | /* XXX - Should we check TARGET_HITACHI here ??? */ | |
12127 | pcum->renesas_abi = sh_attr_renesas_p (fntype) ? 1 : 0; | |
12128 | ||
12129 | if (fntype) | |
12130 | { | |
12131 | pcum->force_mem = ((TARGET_HITACHI || pcum->renesas_abi) | |
12132 | && aggregate_value_p (TREE_TYPE (fntype), fndecl)); | |
f4da8dce | 12133 | pcum->prototype_p = prototype_p (fntype); |
cc15e98f NC |
12134 | pcum->arg_count [(int) SH_ARG_INT] |
12135 | = TARGET_SH5 && aggregate_value_p (TREE_TYPE (fntype), fndecl); | |
12136 | ||
12137 | pcum->call_cookie | |
12138 | = CALL_COOKIE_RET_TRAMP (TARGET_SHCOMPACT | |
12139 | && pcum->arg_count [(int) SH_ARG_INT] == 0 | |
12140 | && (TYPE_MODE (TREE_TYPE (fntype)) == BLKmode | |
12141 | ? int_size_in_bytes (TREE_TYPE (fntype)) | |
12142 | : GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (fntype)))) > 4 | |
12143 | && (BASE_RETURN_VALUE_REG (TYPE_MODE (TREE_TYPE (fntype))) | |
12144 | == FIRST_RET_REG)); | |
12145 | } | |
12146 | else | |
12147 | { | |
12148 | pcum->arg_count [(int) SH_ARG_INT] = 0; | |
12149 | pcum->prototype_p = FALSE; | |
12150 | if (mode != VOIDmode) | |
12151 | { | |
29832164 | 12152 | pcum->call_cookie = |
cc15e98f NC |
12153 | CALL_COOKIE_RET_TRAMP (TARGET_SHCOMPACT |
12154 | && GET_MODE_SIZE (mode) > 4 | |
12155 | && BASE_RETURN_VALUE_REG (mode) == FIRST_RET_REG); | |
12156 | ||
12157 | /* If the default ABI is the Renesas ABI then all library | |
12158 | calls must assume that the library will be using the | |
12159 | Renesas ABI. So if the function would return its result | |
12160 | in memory then we must force the address of this memory | |
12161 | block onto the stack. Ideally we would like to call | |
12162 | targetm.calls.return_in_memory() here but we do not have | |
8b0add27 | 12163 | the TYPE or the FNDECL available so we synthesize the |
cc15e98f NC |
12164 | contents of that function as best we can. */ |
12165 | pcum->force_mem = | |
c0fb94d7 | 12166 | (TARGET_DEFAULT & MASK_HITACHI) |
cc15e98f NC |
12167 | && (mode == BLKmode |
12168 | || (GET_MODE_SIZE (mode) > 4 | |
12169 | && !(mode == DFmode | |
12170 | && TARGET_FPU_DOUBLE))); | |
12171 | } | |
12172 | else | |
12173 | { | |
12174 | pcum->call_cookie = 0; | |
12175 | pcum->force_mem = FALSE; | |
12176 | } | |
12177 | } | |
12178 | } | |
12179 | ||
73a4d10b R |
12180 | /* Replace any occurrence of FROM(n) in X with TO(n). The function does |
12181 | not enter into CONST_DOUBLE for the replace. | |
12182 | ||
12183 | Note that copying is not done so X must not be shared unless all copies | |
12184 | are to be modified. | |
12185 | ||
12186 | This is like replace_rtx, except that we operate on N_REPLACEMENTS | |
30dc60c7 | 12187 | replacements simultaneously - FROM(n) is replacements[n*2] and to(n) is |
73a4d10b R |
12188 | replacements[n*2+1] - and that we take mode changes into account. |
12189 | ||
30dc60c7 | 12190 | If a replacement is ambiguous, return NULL_RTX. |
73a4d10b R |
12191 | |
12192 | If MODIFY is zero, don't modify any rtl in place, | |
12193 | just return zero or nonzero for failure / success. */ | |
12194 | ||
12195 | rtx | |
12196 | replace_n_hard_rtx (rtx x, rtx *replacements, int n_replacements, int modify) | |
12197 | { | |
12198 | int i, j; | |
12199 | const char *fmt; | |
12200 | ||
12201 | /* The following prevents loops occurrence when we change MEM in | |
12202 | CONST_DOUBLE onto the same CONST_DOUBLE. */ | |
12203 | if (x != 0 && GET_CODE (x) == CONST_DOUBLE) | |
12204 | return x; | |
12205 | ||
12206 | for (i = n_replacements - 1; i >= 0 ; i--) | |
12207 | if (x == replacements[i*2] && GET_MODE (x) == GET_MODE (replacements[i*2+1])) | |
12208 | return replacements[i*2+1]; | |
12209 | ||
12210 | /* Allow this function to make replacements in EXPR_LISTs. */ | |
12211 | if (x == 0) | |
12212 | return 0; | |
12213 | ||
12214 | if (GET_CODE (x) == SUBREG) | |
12215 | { | |
0a2aaacc | 12216 | rtx new_rtx = replace_n_hard_rtx (SUBREG_REG (x), replacements, |
73a4d10b R |
12217 | n_replacements, modify); |
12218 | ||
f3536097 | 12219 | if (CONST_INT_P (new_rtx)) |
73a4d10b | 12220 | { |
0a2aaacc | 12221 | x = simplify_subreg (GET_MODE (x), new_rtx, |
73a4d10b R |
12222 | GET_MODE (SUBREG_REG (x)), |
12223 | SUBREG_BYTE (x)); | |
12224 | if (! x) | |
12225 | abort (); | |
12226 | } | |
12227 | else if (modify) | |
0a2aaacc | 12228 | SUBREG_REG (x) = new_rtx; |
73a4d10b R |
12229 | |
12230 | return x; | |
12231 | } | |
f3536097 | 12232 | else if (REG_P (x)) |
73a4d10b R |
12233 | { |
12234 | unsigned regno = REGNO (x); | |
12235 | unsigned nregs = (regno < FIRST_PSEUDO_REGISTER | |
12236 | ? HARD_REGNO_NREGS (regno, GET_MODE (x)) : 1); | |
12237 | rtx result = NULL_RTX; | |
12238 | ||
12239 | for (i = n_replacements - 1; i >= 0; i--) | |
12240 | { | |
12241 | rtx from = replacements[i*2]; | |
12242 | rtx to = replacements[i*2+1]; | |
12243 | unsigned from_regno, from_nregs, to_regno, new_regno; | |
12244 | ||
f3536097 | 12245 | if (!REG_P (from)) |
73a4d10b R |
12246 | continue; |
12247 | from_regno = REGNO (from); | |
12248 | from_nregs = (from_regno < FIRST_PSEUDO_REGISTER | |
12249 | ? HARD_REGNO_NREGS (from_regno, GET_MODE (from)) : 1); | |
12250 | if (regno < from_regno + from_nregs && regno + nregs > from_regno) | |
12251 | { | |
12252 | if (regno < from_regno | |
12253 | || regno + nregs > from_regno + nregs | |
f3536097 | 12254 | || !REG_P (to) |
73a4d10b R |
12255 | || result) |
12256 | return NULL_RTX; | |
12257 | to_regno = REGNO (to); | |
12258 | if (to_regno < FIRST_PSEUDO_REGISTER) | |
12259 | { | |
12260 | new_regno = regno + to_regno - from_regno; | |
12261 | if ((unsigned) HARD_REGNO_NREGS (new_regno, GET_MODE (x)) | |
12262 | != nregs) | |
12263 | return NULL_RTX; | |
12264 | result = gen_rtx_REG (GET_MODE (x), new_regno); | |
12265 | } | |
12266 | else if (GET_MODE (x) <= GET_MODE (to)) | |
12267 | result = gen_lowpart_common (GET_MODE (x), to); | |
12268 | else | |
12269 | result = gen_lowpart_SUBREG (GET_MODE (x), to); | |
12270 | } | |
12271 | } | |
12272 | return result ? result : x; | |
12273 | } | |
12274 | else if (GET_CODE (x) == ZERO_EXTEND) | |
12275 | { | |
0a2aaacc | 12276 | rtx new_rtx = replace_n_hard_rtx (XEXP (x, 0), replacements, |
73a4d10b R |
12277 | n_replacements, modify); |
12278 | ||
f3536097 | 12279 | if (CONST_INT_P (new_rtx)) |
73a4d10b R |
12280 | { |
12281 | x = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x), | |
0a2aaacc | 12282 | new_rtx, GET_MODE (XEXP (x, 0))); |
73a4d10b R |
12283 | if (! x) |
12284 | abort (); | |
12285 | } | |
12286 | else if (modify) | |
0a2aaacc | 12287 | XEXP (x, 0) = new_rtx; |
73a4d10b R |
12288 | |
12289 | return x; | |
12290 | } | |
12291 | ||
12292 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
12293 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
12294 | { | |
0a2aaacc | 12295 | rtx new_rtx; |
73a4d10b R |
12296 | |
12297 | if (fmt[i] == 'e') | |
12298 | { | |
0a2aaacc | 12299 | new_rtx = replace_n_hard_rtx (XEXP (x, i), replacements, |
73a4d10b | 12300 | n_replacements, modify); |
0a2aaacc | 12301 | if (!new_rtx) |
73a4d10b R |
12302 | return NULL_RTX; |
12303 | if (modify) | |
0a2aaacc | 12304 | XEXP (x, i) = new_rtx; |
73a4d10b R |
12305 | } |
12306 | else if (fmt[i] == 'E') | |
12307 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
12308 | { | |
0a2aaacc | 12309 | new_rtx = replace_n_hard_rtx (XVECEXP (x, i, j), replacements, |
73a4d10b | 12310 | n_replacements, modify); |
0a2aaacc | 12311 | if (!new_rtx) |
73a4d10b R |
12312 | return NULL_RTX; |
12313 | if (modify) | |
0a2aaacc | 12314 | XVECEXP (x, i, j) = new_rtx; |
73a4d10b R |
12315 | } |
12316 | } | |
12317 | ||
12318 | return x; | |
12319 | } | |
12320 | ||
12321 | rtx | |
12322 | sh_gen_truncate (enum machine_mode mode, rtx x, int need_sign_ext) | |
12323 | { | |
12324 | enum rtx_code code = TRUNCATE; | |
12325 | ||
12326 | if (GET_CODE (x) == ZERO_EXTEND || GET_CODE (x) == SIGN_EXTEND) | |
12327 | { | |
12328 | rtx inner = XEXP (x, 0); | |
12329 | enum machine_mode inner_mode = GET_MODE (inner); | |
12330 | ||
12331 | if (inner_mode == mode) | |
12332 | return inner; | |
12333 | else if (GET_MODE_SIZE (inner_mode) >= GET_MODE_SIZE (mode)) | |
12334 | x = inner; | |
12335 | else if (GET_MODE_SIZE (inner_mode) < GET_MODE_SIZE (mode) | |
12336 | && (! need_sign_ext || GET_CODE (x) == SIGN_EXTEND)) | |
12337 | { | |
12338 | code = GET_CODE (x); | |
12339 | x = inner; | |
12340 | } | |
12341 | } | |
12342 | return gen_rtx_fmt_e (code, mode, x); | |
12343 | } | |
12344 | ||
12345 | /* called via for_each_rtx after reload, to clean up truncates of | |
12346 | registers that span multiple actual hard registers. */ | |
12347 | int | |
12348 | shmedia_cleanup_truncate (rtx *p, void *n_changes) | |
12349 | { | |
12350 | rtx x = *p, reg; | |
12351 | ||
12352 | if (GET_CODE (x) != TRUNCATE) | |
12353 | return 0; | |
12354 | reg = XEXP (x, 0); | |
f3536097 | 12355 | if (GET_MODE_SIZE (GET_MODE (reg)) > 8 && REG_P (reg)) |
73a4d10b R |
12356 | { |
12357 | enum machine_mode reg_mode = GET_MODE (reg); | |
12358 | XEXP (x, 0) = simplify_subreg (DImode, reg, reg_mode, | |
12359 | subreg_lowpart_offset (DImode, reg_mode)); | |
12360 | *(int*) n_changes += 1; | |
12361 | return -1; | |
12362 | } | |
12363 | return 0; | |
12364 | } | |
12365 | ||
12366 | /* Load and store depend on the highpart of the address. However, | |
12367 | set_attr_alternative does not give well-defined results before reload, | |
12368 | so we must look at the rtl ourselves to see if any of the feeding | |
12369 | registers is used in a memref. */ | |
12370 | ||
12371 | /* Called by sh_contains_memref_p via for_each_rtx. */ | |
12372 | static int | |
12373 | sh_contains_memref_p_1 (rtx *loc, void *data ATTRIBUTE_UNUSED) | |
12374 | { | |
f3536097 | 12375 | return (MEM_P (*loc)); |
73a4d10b R |
12376 | } |
12377 | ||
a4d05547 | 12378 | /* Return nonzero iff INSN contains a MEM. */ |
73a4d10b R |
12379 | int |
12380 | sh_contains_memref_p (rtx insn) | |
12381 | { | |
12382 | return for_each_rtx (&PATTERN (insn), &sh_contains_memref_p_1, NULL); | |
12383 | } | |
12384 | ||
5c7cafa8 CB |
12385 | /* Return nonzero iff INSN loads a banked register. */ |
12386 | int | |
12387 | sh_loads_bankedreg_p (rtx insn) | |
12388 | { | |
12389 | if (GET_CODE (PATTERN (insn)) == SET) | |
12390 | { | |
12391 | rtx op = SET_DEST (PATTERN(insn)); | |
12392 | if (REG_P (op) && BANKED_REGISTER_P (REGNO (op))) | |
12393 | return 1; | |
12394 | } | |
12395 | ||
12396 | return 0; | |
12397 | } | |
12398 | ||
73a4d10b R |
12399 | /* FNADDR is the MEM expression from a call expander. Return an address |
12400 | to use in an SHmedia insn pattern. */ | |
12401 | rtx | |
12402 | shmedia_prepare_call_address (rtx fnaddr, int is_sibcall) | |
12403 | { | |
12404 | int is_sym; | |
12405 | ||
12406 | fnaddr = XEXP (fnaddr, 0); | |
12407 | is_sym = GET_CODE (fnaddr) == SYMBOL_REF; | |
12408 | if (flag_pic && is_sym) | |
12409 | { | |
12410 | if (! SYMBOL_REF_LOCAL_P (fnaddr)) | |
12411 | { | |
12412 | rtx reg = gen_reg_rtx (Pmode); | |
12413 | ||
12414 | /* We must not use GOTPLT for sibcalls, because PIC_REG | |
12415 | must be restored before the PLT code gets to run. */ | |
12416 | if (is_sibcall) | |
12417 | emit_insn (gen_symGOT2reg (reg, fnaddr)); | |
12418 | else | |
12419 | emit_insn (gen_symGOTPLT2reg (reg, fnaddr)); | |
12420 | fnaddr = reg; | |
12421 | } | |
12422 | else | |
12423 | { | |
12424 | fnaddr = gen_sym2PIC (fnaddr); | |
12425 | PUT_MODE (fnaddr, Pmode); | |
12426 | } | |
12427 | } | |
12428 | /* If ptabs might trap, make this visible to the rest of the compiler. | |
12429 | We generally assume that symbols pertain to valid locations, but | |
12430 | it is possible to generate invalid symbols with asm or linker tricks. | |
30dc60c7 | 12431 | In a list of functions where each returns its successor, an invalid |
73a4d10b R |
12432 | symbol might denote an empty list. */ |
12433 | if (!TARGET_PT_FIXED | |
12434 | && (!is_sym || TARGET_INVALID_SYMBOLS) | |
12435 | && (!REG_P (fnaddr) || ! TARGET_REGISTER_P (REGNO (fnaddr)))) | |
12436 | { | |
12437 | rtx tr = gen_reg_rtx (PDImode); | |
12438 | ||
12439 | emit_insn (gen_ptabs (tr, fnaddr)); | |
12440 | fnaddr = tr; | |
12441 | } | |
12442 | else if (! target_reg_operand (fnaddr, Pmode)) | |
12443 | fnaddr = copy_to_mode_reg (Pmode, fnaddr); | |
12444 | return fnaddr; | |
12445 | } | |
12446 | ||
486b10aa AS |
12447 | /* Implement TARGET_PREFERRED_RELOAD_CLASS. */ |
12448 | ||
12449 | static reg_class_t | |
12450 | sh_preferred_reload_class (rtx x, reg_class_t rclass) | |
12451 | { | |
12452 | if (rclass == NO_REGS | |
12453 | && TARGET_SHMEDIA | |
12454 | && (CONST_DOUBLE_P (x) | |
12455 | || GET_CODE (x) == SYMBOL_REF | |
12456 | || PIC_ADDR_P (x))) | |
12457 | return GENERAL_REGS; | |
12458 | ||
12459 | return rclass; | |
12460 | } | |
12461 | ||
12462 | /* Implement TARGET_SECONDARY_RELOAD. */ | |
12463 | ||
12464 | static reg_class_t | |
a87cf97e | 12465 | sh_secondary_reload (bool in_p, rtx x, reg_class_t rclass_i, |
8a99f6f9 R |
12466 | enum machine_mode mode, secondary_reload_info *sri) |
12467 | { | |
a87cf97e JR |
12468 | enum reg_class rclass = (enum reg_class) rclass_i; |
12469 | ||
8a99f6f9 R |
12470 | if (in_p) |
12471 | { | |
0a2aaacc | 12472 | if (REGCLASS_HAS_FP_REG (rclass) |
8a99f6f9 R |
12473 | && ! TARGET_SHMEDIA |
12474 | && immediate_operand ((x), mode) | |
12475 | && ! ((fp_zero_operand (x) || fp_one_operand (x)) | |
12476 | && mode == SFmode && fldi_ok ())) | |
12477 | switch (mode) | |
12478 | { | |
12479 | case SFmode: | |
12480 | sri->icode = CODE_FOR_reload_insf__frn; | |
12481 | return NO_REGS; | |
12482 | case DFmode: | |
12483 | sri->icode = CODE_FOR_reload_indf__frn; | |
12484 | return NO_REGS; | |
12485 | case SImode: | |
12486 | /* ??? If we knew that we are in the appropriate mode - | |
12487 | single precision - we could use a reload pattern directly. */ | |
12488 | return FPUL_REGS; | |
12489 | default: | |
12490 | abort (); | |
12491 | } | |
0a2aaacc | 12492 | if (rclass == FPUL_REGS |
f3536097 | 12493 | && ((REG_P (x) |
8a99f6f9 R |
12494 | && (REGNO (x) == MACL_REG || REGNO (x) == MACH_REG |
12495 | || REGNO (x) == T_REG)) | |
12496 | || GET_CODE (x) == PLUS)) | |
12497 | return GENERAL_REGS; | |
0a2aaacc | 12498 | if (rclass == FPUL_REGS && immediate_operand (x, mode)) |
8a99f6f9 | 12499 | { |
d973ac5a | 12500 | if (satisfies_constraint_I08 (x) || fp_zero_operand (x)) |
8a99f6f9 | 12501 | return GENERAL_REGS; |
d973ac5a KK |
12502 | else if (mode == SFmode) |
12503 | return FP_REGS; | |
8a99f6f9 R |
12504 | sri->icode = CODE_FOR_reload_insi__i_fpul; |
12505 | return NO_REGS; | |
12506 | } | |
0a2aaacc | 12507 | if (rclass == FPSCR_REGS |
f3536097 SZ |
12508 | && ((REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER) |
12509 | || (MEM_P (x) && GET_CODE (XEXP (x, 0)) == PLUS))) | |
8a99f6f9 | 12510 | return GENERAL_REGS; |
0a2aaacc | 12511 | if (REGCLASS_HAS_FP_REG (rclass) |
8a99f6f9 R |
12512 | && TARGET_SHMEDIA |
12513 | && immediate_operand (x, mode) | |
12514 | && x != CONST0_RTX (GET_MODE (x)) | |
12515 | && GET_MODE (x) != V4SFmode) | |
12516 | return GENERAL_REGS; | |
12517 | if ((mode == QImode || mode == HImode) | |
12518 | && TARGET_SHMEDIA && inqhi_operand (x, mode)) | |
12519 | { | |
12520 | sri->icode = ((mode == QImode) | |
12521 | ? CODE_FOR_reload_inqi : CODE_FOR_reload_inhi); | |
12522 | return NO_REGS; | |
12523 | } | |
0a2aaacc | 12524 | if (TARGET_SHMEDIA && rclass == GENERAL_REGS |
dc3ba671 | 12525 | && (GET_CODE (x) == LABEL_REF || PIC_ADDR_P (x))) |
8a99f6f9 R |
12526 | return TARGET_REGS; |
12527 | } /* end of input-only processing. */ | |
12528 | ||
0a2aaacc | 12529 | if (((REGCLASS_HAS_FP_REG (rclass) |
f3536097 | 12530 | && (REG_P (x) |
8a99f6f9 R |
12531 | && (GENERAL_OR_AP_REGISTER_P (REGNO (x)) |
12532 | || (FP_REGISTER_P (REGNO (x)) && mode == SImode | |
12533 | && TARGET_FMOVD)))) | |
0a2aaacc | 12534 | || (REGCLASS_HAS_GENERAL_REG (rclass) |
f3536097 | 12535 | && REG_P (x) |
8a99f6f9 R |
12536 | && FP_REGISTER_P (REGNO (x)))) |
12537 | && ! TARGET_SHMEDIA | |
12538 | && (mode == SFmode || mode == SImode)) | |
12539 | return FPUL_REGS; | |
0a2aaacc KG |
12540 | if ((rclass == FPUL_REGS |
12541 | || (REGCLASS_HAS_FP_REG (rclass) | |
8a99f6f9 | 12542 | && ! TARGET_SHMEDIA && mode == SImode)) |
f3536097 SZ |
12543 | && (MEM_P (x) |
12544 | || (REG_P (x) | |
8a99f6f9 R |
12545 | && (REGNO (x) >= FIRST_PSEUDO_REGISTER |
12546 | || REGNO (x) == T_REG | |
12547 | || system_reg_operand (x, VOIDmode))))) | |
12548 | { | |
0a2aaacc | 12549 | if (rclass == FPUL_REGS) |
8a99f6f9 R |
12550 | return GENERAL_REGS; |
12551 | return FPUL_REGS; | |
12552 | } | |
0a2aaacc KG |
12553 | if ((rclass == TARGET_REGS |
12554 | || (TARGET_SHMEDIA && rclass == SIBCALL_REGS)) | |
32a7ab3d | 12555 | && !satisfies_constraint_Csy (x) |
f3536097 | 12556 | && (!REG_P (x) || ! GENERAL_REGISTER_P (REGNO (x)))) |
8a99f6f9 | 12557 | return GENERAL_REGS; |
0a2aaacc | 12558 | if ((rclass == MAC_REGS || rclass == PR_REGS) |
f3536097 | 12559 | && REG_P (x) && ! GENERAL_REGISTER_P (REGNO (x)) |
0a2aaacc | 12560 | && rclass != REGNO_REG_CLASS (REGNO (x))) |
8a99f6f9 | 12561 | return GENERAL_REGS; |
f3536097 | 12562 | if (rclass != GENERAL_REGS && REG_P (x) |
8a99f6f9 R |
12563 | && TARGET_REGISTER_P (REGNO (x))) |
12564 | return GENERAL_REGS; | |
12565 | return NO_REGS; | |
12566 | } | |
12567 | ||
5efd84c5 NF |
12568 | static void |
12569 | sh_conditional_register_usage (void) | |
12570 | { | |
12571 | int regno; | |
12572 | for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno ++) | |
12573 | if (! VALID_REGISTER_P (regno)) | |
12574 | fixed_regs[regno] = call_used_regs[regno] = 1; | |
12575 | /* R8 and R9 are call-clobbered on SH5, but not on earlier SH ABIs. */ | |
12576 | if (TARGET_SH5) | |
12577 | { | |
12578 | call_used_regs[FIRST_GENERAL_REG + 8] | |
12579 | = call_used_regs[FIRST_GENERAL_REG + 9] = 1; | |
12580 | call_really_used_regs[FIRST_GENERAL_REG + 8] | |
12581 | = call_really_used_regs[FIRST_GENERAL_REG + 9] = 1; | |
12582 | } | |
12583 | if (TARGET_SHMEDIA) | |
12584 | { | |
12585 | regno_reg_class[FIRST_GENERAL_REG] = GENERAL_REGS; | |
12586 | CLEAR_HARD_REG_SET (reg_class_contents[FP0_REGS]); | |
12587 | regno_reg_class[FIRST_FP_REG] = FP_REGS; | |
12588 | } | |
12589 | if (flag_pic) | |
12590 | { | |
12591 | fixed_regs[PIC_OFFSET_TABLE_REGNUM] = 1; | |
12592 | call_used_regs[PIC_OFFSET_TABLE_REGNUM] = 1; | |
12593 | } | |
12594 | /* Renesas saves and restores mac registers on call. */ | |
12595 | if (TARGET_HITACHI && ! TARGET_NOMACSAVE) | |
12596 | { | |
12597 | call_really_used_regs[MACH_REG] = 0; | |
12598 | call_really_used_regs[MACL_REG] = 0; | |
12599 | } | |
12600 | for (regno = FIRST_FP_REG + (TARGET_LITTLE_ENDIAN != 0); | |
12601 | regno <= LAST_FP_REG; regno += 2) | |
12602 | SET_HARD_REG_BIT (reg_class_contents[DF_HI_REGS], regno); | |
12603 | if (TARGET_SHMEDIA) | |
12604 | { | |
12605 | for (regno = FIRST_TARGET_REG; regno <= LAST_TARGET_REG; regno ++) | |
12606 | if (! fixed_regs[regno] && call_really_used_regs[regno]) | |
12607 | SET_HARD_REG_BIT (reg_class_contents[SIBCALL_REGS], regno); | |
12608 | } | |
12609 | else | |
12610 | for (regno = FIRST_GENERAL_REG; regno <= LAST_GENERAL_REG; regno++) | |
12611 | if (! fixed_regs[regno] && call_really_used_regs[regno]) | |
12612 | SET_HARD_REG_BIT (reg_class_contents[SIBCALL_REGS], regno); | |
12613 | } | |
12614 | ||
12615 | ||
73a4d10b R |
12616 | enum sh_divide_strategy_e sh_div_strategy = SH_DIV_STRATEGY_DEFAULT; |
12617 | ||
e2500fed | 12618 | #include "gt-sh.h" |