]>
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, |
e938feb4 CB |
3 | 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010 |
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" |
318881c0 KG |
40 | #include "toplev.h" |
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" |
5fac2ed2 | 59 | |
bc45ade3 | 60 | |
1245df60 R |
61 | int code_for_indirect_jump_scratch = CODE_FOR_indirect_jump_scratch; |
62 | ||
00f8ff66 SC |
63 | #define MSW (TARGET_LITTLE_ENDIAN ? 1 : 0) |
64 | #define LSW (TARGET_LITTLE_ENDIAN ? 0 : 1) | |
65 | ||
fa5322fa AO |
66 | /* These are some macros to abstract register modes. */ |
67 | #define CONST_OK_FOR_ADD(size) \ | |
735cb76e | 68 | (TARGET_SHMEDIA ? CONST_OK_FOR_I10 (size) : CONST_OK_FOR_I08 (size)) |
fa5322fa AO |
69 | #define GEN_MOV (*(TARGET_SHMEDIA64 ? gen_movdi : gen_movsi)) |
70 | #define GEN_ADD3 (*(TARGET_SHMEDIA64 ? gen_adddi3 : gen_addsi3)) | |
71 | #define GEN_SUB3 (*(TARGET_SHMEDIA64 ? gen_subdi3 : gen_subsi3)) | |
72 | ||
561642fa AP |
73 | /* Used to simplify the logic below. Find the attributes wherever |
74 | they may be. */ | |
75 | #define SH_ATTRIBUTES(decl) \ | |
76 | (TYPE_P (decl)) ? TYPE_ATTRIBUTES (decl) \ | |
77 | : DECL_ATTRIBUTES (decl) \ | |
78 | ? (DECL_ATTRIBUTES (decl)) \ | |
79 | : TYPE_ATTRIBUTES (TREE_TYPE (decl)) | |
80 | ||
552ecbd9 AH |
81 | /* Set to 1 by expand_prologue() when the function is an interrupt handler. */ |
82 | int current_function_interrupt; | |
83 | ||
a6ab9fc0 R |
84 | tree sh_deferred_function_attributes; |
85 | tree *sh_deferred_function_attributes_tail = &sh_deferred_function_attributes; | |
16bea517 | 86 | |
6f317ef3 | 87 | /* Global variables for machine-dependent things. */ |
bc45ade3 | 88 | |
16bea517 JW |
89 | /* Which cpu are we scheduling for. */ |
90 | enum processor_type sh_cpu; | |
91 | ||
58565a33 SKG |
92 | /* Definitions used in ready queue reordering for first scheduling pass. */ |
93 | ||
94 | /* Reg weights arrays for modes SFmode and SImode, indexed by insn LUID. */ | |
95 | static short *regmode_weight[2]; | |
96 | ||
97 | /* Total SFmode and SImode weights of scheduled insns. */ | |
98 | static int curr_regmode_pressure[2]; | |
99 | ||
6fb5fa3c DB |
100 | /* Number of r0 life regions. */ |
101 | static int r0_life_regions; | |
102 | ||
58565a33 SKG |
103 | /* If true, skip cycles for Q -> R movement. */ |
104 | static int skip_cycles = 0; | |
105 | ||
106 | /* Cached value of can_issue_more. This is cached in sh_variable_issue hook | |
107 | and returned from sh_reorder2. */ | |
108 | static short cached_can_issue_more; | |
109 | ||
826159e5 KK |
110 | /* Unique number for UNSPEC_BBR pattern. */ |
111 | static unsigned int unspec_bbr_uid = 1; | |
112 | ||
bc45ade3 | 113 | /* Provides the class number of the smallest class containing |
16bea517 | 114 | reg number. */ |
bc45ade3 | 115 | |
fada1961 | 116 | enum reg_class regno_reg_class[FIRST_PSEUDO_REGISTER] = |
bc45ade3 | 117 | { |
8e87e161 | 118 | R0_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, |
bc45ade3 SC |
119 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, |
120 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
121 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
fa5322fa AO |
122 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, |
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, | |
1a95a963 JW |
134 | FP0_REGS,FP_REGS, FP_REGS, FP_REGS, |
135 | FP_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, | |
fa5322fa AO |
138 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, |
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 | TARGET_REGS, TARGET_REGS, TARGET_REGS, TARGET_REGS, | |
151 | TARGET_REGS, TARGET_REGS, TARGET_REGS, TARGET_REGS, | |
225e4f43 R |
152 | DF_REGS, DF_REGS, DF_REGS, DF_REGS, |
153 | DF_REGS, DF_REGS, DF_REGS, DF_REGS, | |
fa5322fa AO |
154 | NO_REGS, GENERAL_REGS, PR_REGS, T_REGS, |
155 | MAC_REGS, MAC_REGS, FPUL_REGS, FPSCR_REGS, | |
96a2347e | 156 | GENERAL_REGS, GENERAL_REGS, |
225e4f43 R |
157 | }; |
158 | ||
fa5322fa AO |
159 | char sh_register_names[FIRST_PSEUDO_REGISTER] \ |
160 | [MAX_REGISTER_NAME_LENGTH + 1] = SH_REGISTER_NAMES_INITIALIZER; | |
161 | ||
162 | char sh_additional_register_names[ADDREGNAMES_SIZE] \ | |
163 | [MAX_ADDITIONAL_REGISTER_NAME_LENGTH + 1] | |
164 | = SH_ADDITIONAL_REGISTER_NAMES_INITIALIZER; | |
bc45ade3 | 165 | |
225e4f43 R |
166 | int assembler_dialect; |
167 | ||
fe3ad572 SC |
168 | static bool shmedia_space_reserved_for_target_registers; |
169 | ||
c0fb94d7 | 170 | static bool sh_handle_option (size_t, const char *, int); |
cf277499 SB |
171 | static void split_branches (rtx); |
172 | static int branch_dest (rtx); | |
173 | static void force_into (rtx, rtx); | |
174 | static void print_slot (rtx); | |
175 | static rtx add_constant (rtx, enum machine_mode, rtx); | |
078c8b08 | 176 | static void dump_table (rtx, rtx); |
cf277499 SB |
177 | static int hi_const (rtx); |
178 | static int broken_move (rtx); | |
179 | static int mova_p (rtx); | |
180 | static rtx find_barrier (int, rtx, rtx); | |
181 | static int noncall_uses_reg (rtx, rtx, rtx *); | |
182 | static rtx gen_block_redirect (rtx, int, int); | |
183 | static void sh_reorg (void); | |
21d05694 | 184 | static void output_stack_adjust (int, rtx, int, HARD_REG_SET *, bool); |
cf277499 SB |
185 | static rtx frame_insn (rtx); |
186 | static rtx push (int); | |
187 | static void pop (int); | |
188 | static void push_regs (HARD_REG_SET *, int); | |
189 | static int calc_live_regs (HARD_REG_SET *); | |
cf277499 | 190 | static HOST_WIDE_INT rounded_frame_size (int); |
7a296495 | 191 | static bool sh_frame_pointer_required (void); |
cf277499 | 192 | static rtx mark_constant_pool_use (rtx); |
cf277499 | 193 | static tree sh_handle_interrupt_handler_attribute (tree *, tree, tree, int, bool *); |
561642fa AP |
194 | static tree sh_handle_resbank_handler_attribute (tree *, tree, |
195 | tree, int, bool *); | |
196 | static tree sh2a_handle_function_vector_handler_attribute (tree *, tree, | |
197 | tree, int, bool *); | |
cf277499 SB |
198 | static tree sh_handle_sp_switch_attribute (tree *, tree, tree, int, bool *); |
199 | static tree sh_handle_trap_exit_attribute (tree *, tree, tree, int, bool *); | |
200 | static tree sh_handle_renesas_attribute (tree *, tree, tree, int, bool *); | |
9e94b957 NF |
201 | static void sh_print_operand (FILE *, rtx, int); |
202 | static void sh_print_operand_address (FILE *, rtx); | |
203 | static bool sh_print_operand_punct_valid_p (unsigned char code); | |
cf277499 SB |
204 | static void sh_output_function_epilogue (FILE *, HOST_WIDE_INT); |
205 | static void sh_insert_attributes (tree, tree *); | |
8d932be3 | 206 | static const char *sh_check_pch_target_flags (int); |
cf277499 | 207 | static int sh_adjust_cost (rtx, rtx, rtx, int); |
cf277499 | 208 | static int sh_issue_rate (void); |
58565a33 SKG |
209 | static int sh_dfa_new_cycle (FILE *, int, rtx, int, int, int *sort_p); |
210 | static short find_set_regmode_weight (rtx, enum machine_mode); | |
211 | static short find_insn_regmode_weight (rtx, enum machine_mode); | |
c6d56443 | 212 | static void find_regmode_weight (basic_block, enum machine_mode); |
6fb5fa3c | 213 | static int find_r0_life_regions (basic_block); |
58565a33 SKG |
214 | static void sh_md_init_global (FILE *, int, int); |
215 | static void sh_md_finish_global (FILE *, int); | |
216 | static int rank_for_reorder (const void *, const void *); | |
217 | static void swap_reorder (rtx *, int); | |
218 | static void ready_reorder (rtx *, int); | |
219 | static short high_pressure (enum machine_mode); | |
220 | static int sh_reorder (FILE *, int, rtx *, int *, int); | |
221 | static int sh_reorder2 (FILE *, int, rtx *, int *, int); | |
222 | static void sh_md_init (FILE *, int, int); | |
223 | static int sh_variable_issue (FILE *, int, rtx, int); | |
f676971a | 224 | |
cf277499 SB |
225 | static bool sh_function_ok_for_sibcall (tree, tree); |
226 | ||
227 | static bool sh_cannot_modify_jumps_p (void); | |
a87cf97e | 228 | static reg_class_t sh_target_reg_class (void); |
fe3ad572 | 229 | static bool sh_optimize_target_register_callee_saved (bool); |
3101faab | 230 | static bool sh_ms_bitfield_layout_p (const_tree); |
cf277499 SB |
231 | |
232 | static void sh_init_builtins (void); | |
1cc24576 | 233 | static tree sh_builtin_decl (unsigned, bool); |
cf277499 | 234 | static void sh_media_init_builtins (void); |
1cc24576 | 235 | static tree sh_media_builtin_decl (unsigned, bool); |
cf277499 SB |
236 | static rtx sh_expand_builtin (tree, rtx, rtx, enum machine_mode, int); |
237 | static void sh_output_mi_thunk (FILE *, tree, HOST_WIDE_INT, HOST_WIDE_INT, tree); | |
238 | static void sh_file_start (void); | |
239 | static int flow_dependent_p (rtx, rtx); | |
7bc980e1 | 240 | static void flow_dependent_p_1 (rtx, const_rtx, void *); |
cf277499 SB |
241 | static int shiftcosts (rtx); |
242 | static int andcosts (rtx); | |
243 | static int addsubcosts (rtx); | |
244 | static int multcosts (rtx); | |
245 | static bool unspec_caller_rtx_p (rtx); | |
246 | static bool sh_cannot_copy_insn_p (rtx); | |
f40751dd JH |
247 | static bool sh_rtx_costs (rtx, int, int, int *, bool); |
248 | static int sh_address_cost (rtx, bool); | |
b48f503c KK |
249 | static int sh_pr_n_sets (void); |
250 | static rtx sh_allocate_initial_value (rtx); | |
1bb99877 | 251 | static bool sh_legitimate_address_p (enum machine_mode, rtx, bool); |
506d7b68 | 252 | static rtx sh_legitimize_address (rtx, rtx, enum machine_mode); |
fe3ad572 SC |
253 | static int shmedia_target_regs_stack_space (HARD_REG_SET *); |
254 | static int shmedia_reserve_space_for_target_registers_p (int, HARD_REG_SET *); | |
255 | static int shmedia_target_regs_stack_adjust (HARD_REG_SET *); | |
fada1961 R |
256 | static int scavenge_reg (HARD_REG_SET *s); |
257 | struct save_schedule_s; | |
258 | static struct save_entry_s *sh5_schedule_saves (HARD_REG_SET *, | |
259 | struct save_schedule_s *, int); | |
61f71b34 | 260 | |
cf277499 | 261 | static rtx sh_struct_value_rtx (tree, int); |
d64728cf AS |
262 | static rtx sh_function_value (const_tree, const_tree, bool); |
263 | static rtx sh_libcall_value (enum machine_mode, const_rtx); | |
586de218 | 264 | static bool sh_return_in_memory (const_tree, const_tree); |
cf277499 SB |
265 | static rtx sh_builtin_saveregs (void); |
266 | static void sh_setup_incoming_varargs (CUMULATIVE_ARGS *, enum machine_mode, tree, int *, int); | |
267 | static bool sh_strict_argument_naming (CUMULATIVE_ARGS *); | |
268 | static bool sh_pretend_outgoing_varargs_named (CUMULATIVE_ARGS *); | |
c35d187f | 269 | static tree sh_build_builtin_va_list (void); |
d7bd8aeb | 270 | static void sh_va_start (tree, rtx); |
726a989a | 271 | static tree sh_gimplify_va_arg_expr (tree, tree, gimple_seq *, gimple_seq *); |
d64728cf | 272 | static bool sh_promote_prototypes (const_tree); |
d055ce45 KK |
273 | static enum machine_mode sh_promote_function_mode (const_tree type, |
274 | enum machine_mode, | |
275 | int *punsignedp, | |
276 | const_tree funtype, | |
6e957ac8 | 277 | int for_return); |
8cd5a4e0 | 278 | static bool sh_pass_by_reference (CUMULATIVE_ARGS *, enum machine_mode, |
586de218 | 279 | const_tree, bool); |
6cdd5672 | 280 | static bool sh_callee_copies (CUMULATIVE_ARGS *, enum machine_mode, |
586de218 | 281 | const_tree, bool); |
78a52f11 RH |
282 | static int sh_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode, |
283 | tree, bool); | |
b0fbea13 | 284 | static bool sh_scalar_mode_supported_p (enum machine_mode); |
3101faab | 285 | static int sh_dwarf_calling_convention (const_tree); |
561642fa AP |
286 | static void sh_encode_section_info (tree, rtx, int); |
287 | static int sh2a_function_vector_p (tree); | |
994322ef RH |
288 | static void sh_trampoline_init (rtx, tree, rtx); |
289 | static rtx sh_trampoline_adjust_address (rtx); | |
5a82ecd9 ILT |
290 | \f |
291 | static const struct attribute_spec sh_attribute_table[] = | |
292 | { | |
293 | /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */ | |
294 | { "interrupt_handler", 0, 0, true, false, false, sh_handle_interrupt_handler_attribute }, | |
295 | { "sp_switch", 1, 1, true, false, false, sh_handle_sp_switch_attribute }, | |
296 | { "trap_exit", 1, 1, true, false, false, sh_handle_trap_exit_attribute }, | |
297 | { "renesas", 0, 0, false, true, false, sh_handle_renesas_attribute }, | |
298 | { "trapa_handler", 0, 0, true, false, false, sh_handle_interrupt_handler_attribute }, | |
299 | { "nosave_low_regs", 0, 0, true, false, false, sh_handle_interrupt_handler_attribute }, | |
300 | { "resbank", 0, 0, true, false, false, sh_handle_resbank_handler_attribute }, | |
301 | { "function_vector", 1, 1, true, false, false, sh2a_handle_function_vector_handler_attribute }, | |
302 | #ifdef SYMBIAN | |
303 | /* Symbian support adds three new attributes: | |
304 | dllexport - for exporting a function/variable that will live in a dll | |
305 | dllimport - for importing a function/variable from a dll | |
61f71b34 | 306 | |
5a82ecd9 ILT |
307 | Microsoft allows multiple declspecs in one __declspec, separating |
308 | them with spaces. We do NOT support this. Instead, use __declspec | |
309 | multiple times. */ | |
310 | { "dllimport", 0, 0, true, false, false, sh_symbian_handle_dll_attribute }, | |
311 | { "dllexport", 0, 0, true, false, false, sh_symbian_handle_dll_attribute }, | |
312 | #endif | |
313 | { NULL, 0, 0, false, false, false, NULL } | |
314 | }; | |
672a6f42 NB |
315 | \f |
316 | /* Initialize the GCC target structure. */ | |
91d231cb JM |
317 | #undef TARGET_ATTRIBUTE_TABLE |
318 | #define TARGET_ATTRIBUTE_TABLE sh_attribute_table | |
672a6f42 | 319 | |
301d03af RS |
320 | /* The next two are used for debug info when compiling with -gdwarf. */ |
321 | #undef TARGET_ASM_UNALIGNED_HI_OP | |
322 | #define TARGET_ASM_UNALIGNED_HI_OP "\t.uaword\t" | |
323 | #undef TARGET_ASM_UNALIGNED_SI_OP | |
324 | #define TARGET_ASM_UNALIGNED_SI_OP "\t.ualong\t" | |
325 | ||
fa5322fa AO |
326 | /* These are NULLed out on non-SH5 in OVERRIDE_OPTIONS. */ |
327 | #undef TARGET_ASM_UNALIGNED_DI_OP | |
328 | #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaquad\t" | |
329 | #undef TARGET_ASM_ALIGNED_DI_OP | |
330 | #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t" | |
331 | ||
9e94b957 NF |
332 | #undef TARGET_PRINT_OPERAND |
333 | #define TARGET_PRINT_OPERAND sh_print_operand | |
334 | #undef TARGET_PRINT_OPERAND_ADDRESS | |
335 | #define TARGET_PRINT_OPERAND_ADDRESS sh_print_operand_address | |
336 | #undef TARGET_PRINT_OPERAND_PUNCT_VALID_P | |
337 | #define TARGET_PRINT_OPERAND_PUNCT_VALID_P sh_print_operand_punct_valid_p | |
338 | ||
08c148a8 NB |
339 | #undef TARGET_ASM_FUNCTION_EPILOGUE |
340 | #define TARGET_ASM_FUNCTION_EPILOGUE sh_output_function_epilogue | |
341 | ||
1aa03f38 R |
342 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
343 | #define TARGET_ASM_OUTPUT_MI_THUNK sh_output_mi_thunk | |
344 | ||
345 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
3101faab | 346 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true |
1aa03f38 | 347 | |
1bc7c5b6 ZW |
348 | #undef TARGET_ASM_FILE_START |
349 | #define TARGET_ASM_FILE_START sh_file_start | |
350 | #undef TARGET_ASM_FILE_START_FILE_DIRECTIVE | |
351 | #define TARGET_ASM_FILE_START_FILE_DIRECTIVE true | |
352 | ||
c0fb94d7 RS |
353 | #undef TARGET_DEFAULT_TARGET_FLAGS |
354 | #define TARGET_DEFAULT_TARGET_FLAGS TARGET_DEFAULT | |
355 | #undef TARGET_HANDLE_OPTION | |
356 | #define TARGET_HANDLE_OPTION sh_handle_option | |
357 | ||
12a68f1f JM |
358 | #undef TARGET_INSERT_ATTRIBUTES |
359 | #define TARGET_INSERT_ATTRIBUTES sh_insert_attributes | |
360 | ||
c237e94a ZW |
361 | #undef TARGET_SCHED_ADJUST_COST |
362 | #define TARGET_SCHED_ADJUST_COST sh_adjust_cost | |
363 | ||
fae15c93 VM |
364 | #undef TARGET_SCHED_ISSUE_RATE |
365 | #define TARGET_SCHED_ISSUE_RATE sh_issue_rate | |
366 | ||
58565a33 SKG |
367 | /* The next 5 hooks have been implemented for reenabling sched1. With the |
368 | help of these macros we are limiting the movement of insns in sched1 to | |
f676971a | 369 | reduce the register pressure. The overall idea is to keep count of SImode |
58565a33 SKG |
370 | and SFmode regs required by already scheduled insns. When these counts |
371 | cross some threshold values; give priority to insns that free registers. | |
372 | The insn that frees registers is most likely to be the insn with lowest | |
f676971a | 373 | LUID (original insn order); but such an insn might be there in the stalled |
58565a33 SKG |
374 | queue (Q) instead of the ready queue (R). To solve this, we skip cycles |
375 | upto a max of 8 cycles so that such insns may move from Q -> R. | |
376 | ||
377 | The description of the hooks are as below: | |
378 | ||
379 | TARGET_SCHED_INIT_GLOBAL: Added a new target hook in the generic | |
380 | scheduler; it is called inside the sched_init function just after | |
381 | find_insn_reg_weights function call. It is used to calculate the SImode | |
1ae58c30 | 382 | and SFmode weights of insns of basic blocks; much similar to what |
f676971a | 383 | find_insn_reg_weights does. |
58565a33 SKG |
384 | TARGET_SCHED_FINISH_GLOBAL: Corresponding cleanup hook. |
385 | ||
386 | TARGET_SCHED_DFA_NEW_CYCLE: Skip cycles if high register pressure is | |
387 | indicated by TARGET_SCHED_REORDER2; doing this may move insns from | |
388 | (Q)->(R). | |
389 | ||
390 | TARGET_SCHED_REORDER: If the register pressure for SImode or SFmode is | |
391 | high; reorder the ready queue so that the insn with lowest LUID will be | |
392 | issued next. | |
393 | ||
394 | TARGET_SCHED_REORDER2: If the register pressure is high, indicate to | |
395 | TARGET_SCHED_DFA_NEW_CYCLE to skip cycles. | |
396 | ||
397 | TARGET_SCHED_VARIABLE_ISSUE: Cache the value of can_issue_more so that it | |
398 | can be returned from TARGET_SCHED_REORDER2. | |
399 | ||
400 | TARGET_SCHED_INIT: Reset the register pressure counting variables. */ | |
401 | ||
402 | #undef TARGET_SCHED_DFA_NEW_CYCLE | |
403 | #define TARGET_SCHED_DFA_NEW_CYCLE sh_dfa_new_cycle | |
404 | ||
405 | #undef TARGET_SCHED_INIT_GLOBAL | |
406 | #define TARGET_SCHED_INIT_GLOBAL sh_md_init_global | |
407 | ||
408 | #undef TARGET_SCHED_FINISH_GLOBAL | |
409 | #define TARGET_SCHED_FINISH_GLOBAL sh_md_finish_global | |
410 | ||
411 | #undef TARGET_SCHED_VARIABLE_ISSUE | |
412 | #define TARGET_SCHED_VARIABLE_ISSUE sh_variable_issue | |
413 | ||
414 | #undef TARGET_SCHED_REORDER | |
415 | #define TARGET_SCHED_REORDER sh_reorder | |
416 | ||
417 | #undef TARGET_SCHED_REORDER2 | |
418 | #define TARGET_SCHED_REORDER2 sh_reorder2 | |
419 | ||
420 | #undef TARGET_SCHED_INIT | |
421 | #define TARGET_SCHED_INIT sh_md_init | |
422 | ||
506d7b68 PB |
423 | #undef TARGET_LEGITIMIZE_ADDRESS |
424 | #define TARGET_LEGITIMIZE_ADDRESS sh_legitimize_address | |
425 | ||
fa5322fa AO |
426 | #undef TARGET_CANNOT_MODIFY_JUMPS_P |
427 | #define TARGET_CANNOT_MODIFY_JUMPS_P sh_cannot_modify_jumps_p | |
fe3ad572 SC |
428 | #undef TARGET_BRANCH_TARGET_REGISTER_CLASS |
429 | #define TARGET_BRANCH_TARGET_REGISTER_CLASS sh_target_reg_class | |
430 | #undef TARGET_BRANCH_TARGET_REGISTER_CALLEE_SAVED | |
431 | #define TARGET_BRANCH_TARGET_REGISTER_CALLEE_SAVED \ | |
432 | sh_optimize_target_register_callee_saved | |
fa5322fa AO |
433 | |
434 | #undef TARGET_MS_BITFIELD_LAYOUT_P | |
435 | #define TARGET_MS_BITFIELD_LAYOUT_P sh_ms_bitfield_layout_p | |
436 | ||
c1b92d09 R |
437 | #undef TARGET_INIT_BUILTINS |
438 | #define TARGET_INIT_BUILTINS sh_init_builtins | |
1cc24576 KK |
439 | #undef TARGET_BUILTIN_DECL |
440 | #define TARGET_BUILTIN_DECL sh_builtin_decl | |
c1b92d09 R |
441 | #undef TARGET_EXPAND_BUILTIN |
442 | #define TARGET_EXPAND_BUILTIN sh_expand_builtin | |
443 | ||
4977bab6 ZW |
444 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL |
445 | #define TARGET_FUNCTION_OK_FOR_SIBCALL sh_function_ok_for_sibcall | |
446 | ||
9f3a9a08 KK |
447 | #undef TARGET_CANNOT_COPY_INSN_P |
448 | #define TARGET_CANNOT_COPY_INSN_P sh_cannot_copy_insn_p | |
3c50106f RH |
449 | #undef TARGET_RTX_COSTS |
450 | #define TARGET_RTX_COSTS sh_rtx_costs | |
dcefdf67 RH |
451 | #undef TARGET_ADDRESS_COST |
452 | #define TARGET_ADDRESS_COST sh_address_cost | |
b48f503c KK |
453 | #undef TARGET_ALLOCATE_INITIAL_VALUE |
454 | #define TARGET_ALLOCATE_INITIAL_VALUE sh_allocate_initial_value | |
3c50106f | 455 | |
18dbd950 RS |
456 | #undef TARGET_MACHINE_DEPENDENT_REORG |
457 | #define TARGET_MACHINE_DEPENDENT_REORG sh_reorg | |
458 | ||
63a9de54 CB |
459 | #undef TARGET_DWARF_REGISTER_SPAN |
460 | #define TARGET_DWARF_REGISTER_SPAN sh_dwarf_register_span | |
461 | ||
463f02cd KK |
462 | #ifdef HAVE_AS_TLS |
463 | #undef TARGET_HAVE_TLS | |
464 | #define TARGET_HAVE_TLS true | |
465 | #endif | |
466 | ||
61f71b34 DD |
467 | #undef TARGET_PROMOTE_PROTOTYPES |
468 | #define TARGET_PROMOTE_PROTOTYPES sh_promote_prototypes | |
d055ce45 KK |
469 | #undef TARGET_PROMOTE_FUNCTION_MODE |
470 | #define TARGET_PROMOTE_FUNCTION_MODE sh_promote_function_mode | |
61f71b34 | 471 | |
d64728cf AS |
472 | #undef TARGET_FUNCTION_VALUE |
473 | #define TARGET_FUNCTION_VALUE sh_function_value | |
474 | #undef TARGET_LIBCALL_VALUE | |
475 | #define TARGET_LIBCALL_VALUE sh_libcall_value | |
61f71b34 DD |
476 | #undef TARGET_STRUCT_VALUE_RTX |
477 | #define TARGET_STRUCT_VALUE_RTX sh_struct_value_rtx | |
478 | #undef TARGET_RETURN_IN_MEMORY | |
479 | #define TARGET_RETURN_IN_MEMORY sh_return_in_memory | |
480 | ||
481 | #undef TARGET_EXPAND_BUILTIN_SAVEREGS | |
482 | #define TARGET_EXPAND_BUILTIN_SAVEREGS sh_builtin_saveregs | |
483 | #undef TARGET_SETUP_INCOMING_VARARGS | |
484 | #define TARGET_SETUP_INCOMING_VARARGS sh_setup_incoming_varargs | |
485 | #undef TARGET_STRICT_ARGUMENT_NAMING | |
486 | #define TARGET_STRICT_ARGUMENT_NAMING sh_strict_argument_naming | |
487 | #undef TARGET_PRETEND_OUTGOING_VARARGS_NAMED | |
488 | #define TARGET_PRETEND_OUTGOING_VARARGS_NAMED sh_pretend_outgoing_varargs_named | |
fe984136 RH |
489 | #undef TARGET_MUST_PASS_IN_STACK |
490 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
8cd5a4e0 RH |
491 | #undef TARGET_PASS_BY_REFERENCE |
492 | #define TARGET_PASS_BY_REFERENCE sh_pass_by_reference | |
6cdd5672 RH |
493 | #undef TARGET_CALLEE_COPIES |
494 | #define TARGET_CALLEE_COPIES sh_callee_copies | |
78a52f11 RH |
495 | #undef TARGET_ARG_PARTIAL_BYTES |
496 | #define TARGET_ARG_PARTIAL_BYTES sh_arg_partial_bytes | |
61f71b34 | 497 | |
c35d187f RH |
498 | #undef TARGET_BUILD_BUILTIN_VA_LIST |
499 | #define TARGET_BUILD_BUILTIN_VA_LIST sh_build_builtin_va_list | |
d7bd8aeb JJ |
500 | #undef TARGET_EXPAND_BUILTIN_VA_START |
501 | #define TARGET_EXPAND_BUILTIN_VA_START sh_va_start | |
5fac2ed2 RH |
502 | #undef TARGET_GIMPLIFY_VA_ARG_EXPR |
503 | #define TARGET_GIMPLIFY_VA_ARG_EXPR sh_gimplify_va_arg_expr | |
c35d187f | 504 | |
b0fbea13 KK |
505 | #undef TARGET_SCALAR_MODE_SUPPORTED_P |
506 | #define TARGET_SCALAR_MODE_SUPPORTED_P sh_scalar_mode_supported_p | |
f676971a EC |
507 | #undef TARGET_VECTOR_MODE_SUPPORTED_P |
508 | #define TARGET_VECTOR_MODE_SUPPORTED_P sh_vector_mode_supported_p | |
509 | ||
8d932be3 RS |
510 | #undef TARGET_CHECK_PCH_TARGET_FLAGS |
511 | #define TARGET_CHECK_PCH_TARGET_FLAGS sh_check_pch_target_flags | |
bcc8cc82 | 512 | |
a1c496cb EC |
513 | #undef TARGET_DWARF_CALLING_CONVENTION |
514 | #define TARGET_DWARF_CALLING_CONVENTION sh_dwarf_calling_convention | |
515 | ||
7a296495 CB |
516 | #undef TARGET_FRAME_POINTER_REQUIRED |
517 | #define TARGET_FRAME_POINTER_REQUIRED sh_frame_pointer_required | |
518 | ||
58565a33 SKG |
519 | /* Return regmode weight for insn. */ |
520 | #define INSN_REGMODE_WEIGHT(INSN, MODE) regmode_weight[((MODE) == SImode) ? 0 : 1][INSN_UID (INSN)] | |
521 | ||
522 | /* Return current register pressure for regmode. */ | |
523 | #define CURR_REGMODE_PRESSURE(MODE) curr_regmode_pressure[((MODE) == SImode) ? 0 : 1] | |
524 | ||
561642fa AP |
525 | #undef TARGET_ENCODE_SECTION_INFO |
526 | #define TARGET_ENCODE_SECTION_INFO sh_encode_section_info | |
527 | ||
b55e3aad NC |
528 | #ifdef SYMBIAN |
529 | ||
530 | #undef TARGET_ENCODE_SECTION_INFO | |
531 | #define TARGET_ENCODE_SECTION_INFO sh_symbian_encode_section_info | |
532 | #undef TARGET_STRIP_NAME_ENCODING | |
533 | #define TARGET_STRIP_NAME_ENCODING sh_symbian_strip_name_encoding | |
534 | #undef TARGET_CXX_IMPORT_EXPORT_CLASS | |
3c8348d3 | 535 | #define TARGET_CXX_IMPORT_EXPORT_CLASS sh_symbian_import_export_class |
b55e3aad NC |
536 | |
537 | #endif /* SYMBIAN */ | |
538 | ||
8a99f6f9 R |
539 | #undef TARGET_SECONDARY_RELOAD |
540 | #define TARGET_SECONDARY_RELOAD sh_secondary_reload | |
541 | ||
1bb99877 KK |
542 | #undef TARGET_LEGITIMATE_ADDRESS_P |
543 | #define TARGET_LEGITIMATE_ADDRESS_P sh_legitimate_address_p | |
544 | ||
994322ef RH |
545 | #undef TARGET_TRAMPOLINE_INIT |
546 | #define TARGET_TRAMPOLINE_INIT sh_trampoline_init | |
547 | #undef TARGET_TRAMPOLINE_ADJUST_ADDRESS | |
548 | #define TARGET_TRAMPOLINE_ADJUST_ADDRESS sh_trampoline_adjust_address | |
549 | ||
561642fa AP |
550 | /* Machine-specific symbol_ref flags. */ |
551 | #define SYMBOL_FLAG_FUNCVEC_FUNCTION (SYMBOL_FLAG_MACH_DEP << 0) | |
552 | ||
f6897b10 | 553 | struct gcc_target targetm = TARGET_INITIALIZER; |
b9654711 | 554 | \f |
c0fb94d7 RS |
555 | /* Implement TARGET_HANDLE_OPTION. */ |
556 | ||
557 | static bool | |
558 | sh_handle_option (size_t code, const char *arg ATTRIBUTE_UNUSED, | |
559 | int value ATTRIBUTE_UNUSED) | |
560 | { | |
561 | switch (code) | |
562 | { | |
563 | case OPT_m1: | |
564 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH1; | |
565 | return true; | |
566 | ||
567 | case OPT_m2: | |
568 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2; | |
569 | return true; | |
570 | ||
571 | case OPT_m2a: | |
572 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A; | |
573 | return true; | |
574 | ||
575 | case OPT_m2a_nofpu: | |
576 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_NOFPU; | |
577 | return true; | |
578 | ||
579 | case OPT_m2a_single: | |
580 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_SINGLE; | |
581 | return true; | |
582 | ||
583 | case OPT_m2a_single_only: | |
584 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_SINGLE_ONLY; | |
585 | return true; | |
586 | ||
587 | case OPT_m2e: | |
588 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2E; | |
589 | return true; | |
590 | ||
591 | case OPT_m3: | |
592 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH3; | |
593 | return true; | |
594 | ||
595 | case OPT_m3e: | |
596 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH3E; | |
597 | return true; | |
598 | ||
599 | case OPT_m4: | |
25e651ca AS |
600 | case OPT_m4_100: |
601 | case OPT_m4_200: | |
78d310c2 | 602 | case OPT_m4_300: |
c0fb94d7 RS |
603 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4; |
604 | return true; | |
605 | ||
606 | case OPT_m4_nofpu: | |
78d310c2 R |
607 | case OPT_m4_100_nofpu: |
608 | case OPT_m4_200_nofpu: | |
609 | case OPT_m4_300_nofpu: | |
610 | case OPT_m4_340: | |
25e651ca AS |
611 | case OPT_m4_400: |
612 | case OPT_m4_500: | |
c0fb94d7 RS |
613 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_NOFPU; |
614 | return true; | |
615 | ||
616 | case OPT_m4_single: | |
25e651ca AS |
617 | case OPT_m4_100_single: |
618 | case OPT_m4_200_single: | |
78d310c2 | 619 | case OPT_m4_300_single: |
c0fb94d7 RS |
620 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_SINGLE; |
621 | return true; | |
622 | ||
623 | case OPT_m4_single_only: | |
25e651ca AS |
624 | case OPT_m4_100_single_only: |
625 | case OPT_m4_200_single_only: | |
78d310c2 | 626 | case OPT_m4_300_single_only: |
c0fb94d7 RS |
627 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_SINGLE_ONLY; |
628 | return true; | |
629 | ||
630 | case OPT_m4a: | |
631 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A; | |
632 | return true; | |
633 | ||
634 | case OPT_m4a_nofpu: | |
635 | case OPT_m4al: | |
636 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_NOFPU; | |
637 | return true; | |
638 | ||
639 | case OPT_m4a_single: | |
640 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_SINGLE; | |
641 | return true; | |
642 | ||
643 | case OPT_m4a_single_only: | |
644 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_SINGLE_ONLY; | |
645 | return true; | |
646 | ||
647 | case OPT_m5_32media: | |
648 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_32MEDIA; | |
649 | return true; | |
650 | ||
651 | case OPT_m5_32media_nofpu: | |
652 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_32MEDIA_NOFPU; | |
653 | return true; | |
654 | ||
655 | case OPT_m5_64media: | |
656 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_64MEDIA; | |
657 | return true; | |
658 | ||
659 | case OPT_m5_64media_nofpu: | |
660 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_64MEDIA_NOFPU; | |
661 | return true; | |
662 | ||
663 | case OPT_m5_compact: | |
664 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_COMPACT; | |
665 | return true; | |
666 | ||
667 | case OPT_m5_compact_nofpu: | |
668 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_COMPACT_NOFPU; | |
669 | return true; | |
670 | ||
671 | default: | |
672 | return true; | |
673 | } | |
674 | } | |
675 | \f | |
bd9a3465 KK |
676 | /* Set default optimization options. */ |
677 | void | |
678 | sh_optimization_options (int level ATTRIBUTE_UNUSED, int size ATTRIBUTE_UNUSED) | |
679 | { | |
680 | if (level) | |
681 | { | |
bd9a3465 KK |
682 | if (!size) |
683 | sh_div_str = "inv:minlat"; | |
684 | } | |
685 | if (size) | |
686 | { | |
687 | target_flags |= MASK_SMALLCODE; | |
688 | sh_div_str = SH_DIV_STR_FOR_SIZE ; | |
689 | } | |
690 | else | |
691 | TARGET_CBRANCHDI4 = 1; | |
692 | /* We can't meaningfully test TARGET_SHMEDIA here, because -m options | |
693 | haven't been parsed yet, hence we'd read only the default. | |
694 | sh_target_reg_class will return NO_REGS if this is not SHMEDIA, so | |
695 | it's OK to always set flag_branch_target_load_optimize. */ | |
696 | if (level > 1) | |
697 | { | |
698 | flag_branch_target_load_optimize = 1; | |
699 | if (!size) | |
700 | target_flags |= MASK_SAVE_ALL_TARGET_REGS; | |
701 | } | |
702 | /* Likewise, we can't meaningfully test TARGET_SH2E / TARGET_IEEE | |
703 | here, so leave it to OVERRIDE_OPTIONS to set | |
704 | flag_finite_math_only. We set it to 2 here so we know if the user | |
705 | explicitly requested this to be on or off. */ | |
706 | flag_finite_math_only = 2; | |
707 | /* If flag_schedule_insns is 1, we set it to 2 here so we know if | |
708 | the user explicitly requested this to be on or off. */ | |
709 | if (flag_schedule_insns > 0) | |
710 | flag_schedule_insns = 2; | |
711 | ||
712 | set_param_value ("simultaneous-prefetches", 2); | |
713 | } | |
714 | ||
715 | /* Implement OVERRIDE_OPTIONS macro. Validate and override various | |
716 | options, and do some machine dependent initialization. */ | |
717 | void | |
718 | sh_override_options (void) | |
719 | { | |
720 | int regno; | |
721 | ||
722 | SUBTARGET_OVERRIDE_OPTIONS; | |
723 | if (flag_finite_math_only == 2) | |
724 | flag_finite_math_only | |
725 | = !flag_signaling_nans && TARGET_SH2E && ! TARGET_IEEE; | |
726 | if (TARGET_SH2E && !flag_finite_math_only) | |
727 | target_flags |= MASK_IEEE; | |
728 | sh_cpu = PROCESSOR_SH1; | |
729 | assembler_dialect = 0; | |
730 | if (TARGET_SH2) | |
731 | sh_cpu = PROCESSOR_SH2; | |
732 | if (TARGET_SH2E) | |
733 | sh_cpu = PROCESSOR_SH2E; | |
734 | if (TARGET_SH2A) | |
7a61cf6f | 735 | sh_cpu = PROCESSOR_SH2A; |
bd9a3465 KK |
736 | if (TARGET_SH3) |
737 | sh_cpu = PROCESSOR_SH3; | |
738 | if (TARGET_SH3E) | |
739 | sh_cpu = PROCESSOR_SH3E; | |
740 | if (TARGET_SH4) | |
741 | { | |
742 | assembler_dialect = 1; | |
743 | sh_cpu = PROCESSOR_SH4; | |
744 | } | |
745 | if (TARGET_SH4A_ARCH) | |
746 | { | |
747 | assembler_dialect = 1; | |
748 | sh_cpu = PROCESSOR_SH4A; | |
749 | } | |
750 | if (TARGET_SH5) | |
751 | { | |
752 | sh_cpu = PROCESSOR_SH5; | |
753 | target_flags |= MASK_ALIGN_DOUBLE; | |
754 | if (TARGET_SHMEDIA_FPU) | |
755 | target_flags |= MASK_FMOVD; | |
756 | if (TARGET_SHMEDIA) | |
757 | { | |
758 | /* There are no delay slots on SHmedia. */ | |
759 | flag_delayed_branch = 0; | |
760 | /* Relaxation isn't yet supported for SHmedia */ | |
761 | target_flags &= ~MASK_RELAX; | |
762 | /* After reload, if conversion does little good but can cause | |
763 | ICEs: | |
764 | - find_if_block doesn't do anything for SH because we don't | |
765 | have conditional execution patterns. (We use conditional | |
766 | move patterns, which are handled differently, and only | |
767 | before reload). | |
768 | - find_cond_trap doesn't do anything for the SH because we | |
769 | don't have conditional traps. | |
770 | - find_if_case_1 uses redirect_edge_and_branch_force in | |
771 | the only path that does an optimization, and this causes | |
772 | an ICE when branch targets are in registers. | |
773 | - find_if_case_2 doesn't do anything for the SHmedia after | |
774 | reload except when it can redirect a tablejump - and | |
775 | that's rather rare. */ | |
776 | flag_if_conversion2 = 0; | |
777 | if (! strcmp (sh_div_str, "call")) | |
778 | sh_div_strategy = SH_DIV_CALL; | |
779 | else if (! strcmp (sh_div_str, "call2")) | |
780 | sh_div_strategy = SH_DIV_CALL2; | |
781 | if (! strcmp (sh_div_str, "fp") && TARGET_FPU_ANY) | |
782 | sh_div_strategy = SH_DIV_FP; | |
783 | else if (! strcmp (sh_div_str, "inv")) | |
784 | sh_div_strategy = SH_DIV_INV; | |
785 | else if (! strcmp (sh_div_str, "inv:minlat")) | |
786 | sh_div_strategy = SH_DIV_INV_MINLAT; | |
787 | else if (! strcmp (sh_div_str, "inv20u")) | |
788 | sh_div_strategy = SH_DIV_INV20U; | |
789 | else if (! strcmp (sh_div_str, "inv20l")) | |
790 | sh_div_strategy = SH_DIV_INV20L; | |
791 | else if (! strcmp (sh_div_str, "inv:call2")) | |
792 | sh_div_strategy = SH_DIV_INV_CALL2; | |
793 | else if (! strcmp (sh_div_str, "inv:call")) | |
794 | sh_div_strategy = SH_DIV_INV_CALL; | |
795 | else if (! strcmp (sh_div_str, "inv:fp")) | |
796 | { | |
797 | if (TARGET_FPU_ANY) | |
798 | sh_div_strategy = SH_DIV_INV_FP; | |
799 | else | |
800 | sh_div_strategy = SH_DIV_INV; | |
801 | } | |
802 | TARGET_CBRANCHDI4 = 0; | |
803 | /* Assembler CFI isn't yet fully supported for SHmedia. */ | |
804 | flag_dwarf2_cfi_asm = 0; | |
805 | } | |
806 | } | |
807 | else | |
808 | { | |
809 | /* Only the sh64-elf assembler fully supports .quad properly. */ | |
810 | targetm.asm_out.aligned_op.di = NULL; | |
811 | targetm.asm_out.unaligned_op.di = NULL; | |
812 | } | |
813 | if (TARGET_SH1) | |
814 | { | |
815 | if (! strcmp (sh_div_str, "call-div1")) | |
816 | sh_div_strategy = SH_DIV_CALL_DIV1; | |
817 | else if (! strcmp (sh_div_str, "call-fp") | |
818 | && (TARGET_FPU_DOUBLE | |
819 | || (TARGET_HARD_SH4 && TARGET_SH2E) | |
820 | || (TARGET_SHCOMPACT && TARGET_FPU_ANY))) | |
821 | sh_div_strategy = SH_DIV_CALL_FP; | |
822 | else if (! strcmp (sh_div_str, "call-table") && TARGET_SH2) | |
823 | sh_div_strategy = SH_DIV_CALL_TABLE; | |
824 | else | |
825 | /* Pick one that makes most sense for the target in general. | |
826 | It is not much good to use different functions depending | |
827 | on -Os, since then we'll end up with two different functions | |
828 | when some of the code is compiled for size, and some for | |
829 | speed. */ | |
830 | ||
831 | /* SH4 tends to emphasize speed. */ | |
832 | if (TARGET_HARD_SH4) | |
833 | sh_div_strategy = SH_DIV_CALL_TABLE; | |
834 | /* These have their own way of doing things. */ | |
835 | else if (TARGET_SH2A) | |
836 | sh_div_strategy = SH_DIV_INTRINSIC; | |
837 | /* ??? Should we use the integer SHmedia function instead? */ | |
838 | else if (TARGET_SHCOMPACT && TARGET_FPU_ANY) | |
839 | sh_div_strategy = SH_DIV_CALL_FP; | |
840 | /* SH1 .. SH3 cores often go into small-footprint systems, so | |
841 | default to the smallest implementation available. */ | |
842 | else if (TARGET_SH2) /* ??? EXPERIMENTAL */ | |
843 | sh_div_strategy = SH_DIV_CALL_TABLE; | |
844 | else | |
845 | sh_div_strategy = SH_DIV_CALL_DIV1; | |
846 | } | |
847 | if (!TARGET_SH1) | |
848 | TARGET_PRETEND_CMOVE = 0; | |
849 | if (sh_divsi3_libfunc[0]) | |
850 | ; /* User supplied - leave it alone. */ | |
851 | else if (TARGET_DIVIDE_CALL_FP) | |
852 | sh_divsi3_libfunc = "__sdivsi3_i4"; | |
853 | else if (TARGET_DIVIDE_CALL_TABLE) | |
854 | sh_divsi3_libfunc = "__sdivsi3_i4i"; | |
855 | else if (TARGET_SH5) | |
856 | sh_divsi3_libfunc = "__sdivsi3_1"; | |
857 | else | |
858 | sh_divsi3_libfunc = "__sdivsi3"; | |
859 | if (sh_branch_cost == -1) | |
860 | sh_branch_cost | |
861 | = TARGET_SH5 ? 1 : ! TARGET_SH2 || TARGET_HARD_SH4 ? 2 : 1; | |
862 | ||
863 | for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++) | |
864 | if (! VALID_REGISTER_P (regno)) | |
865 | sh_register_names[regno][0] = '\0'; | |
866 | ||
867 | for (regno = 0; regno < ADDREGNAMES_SIZE; regno++) | |
868 | if (! VALID_REGISTER_P (ADDREGNAMES_REGNO (regno))) | |
869 | sh_additional_register_names[regno][0] = '\0'; | |
870 | ||
7a296495 | 871 | flag_omit_frame_pointer = (PREFERRED_DEBUGGING_TYPE == DWARF2_DEBUG); |
bd9a3465 KK |
872 | |
873 | if ((flag_pic && ! TARGET_PREFERGOT) | |
874 | || (TARGET_SHMEDIA && !TARGET_PT_FIXED)) | |
875 | flag_no_function_cse = 1; | |
876 | ||
42db504c | 877 | if (targetm.small_register_classes_for_mode_p (VOIDmode)) \ |
bd9a3465 KK |
878 | { |
879 | /* Never run scheduling before reload, since that can | |
880 | break global alloc, and generates slower code anyway due | |
881 | to the pressure on R0. */ | |
882 | /* Enable sched1 for SH4 if the user explicitly requests. | |
883 | When sched1 is enabled, the ready queue will be reordered by | |
884 | the target hooks if pressure is high. We can not do this for | |
885 | PIC, SH3 and lower as they give spill failures for R0. */ | |
886 | if (!TARGET_HARD_SH4 || flag_pic) | |
887 | flag_schedule_insns = 0; | |
888 | /* ??? Current exception handling places basic block boundaries | |
889 | after call_insns. It causes the high pressure on R0 and gives | |
890 | spill failures for R0 in reload. See PR 22553 and the thread | |
891 | on gcc-patches | |
892 | <http://gcc.gnu.org/ml/gcc-patches/2005-10/msg00816.html>. */ | |
893 | else if (flag_exceptions) | |
894 | { | |
895 | if (flag_schedule_insns == 1) | |
896 | warning (0, "ignoring -fschedule-insns because of exception handling bug"); | |
897 | flag_schedule_insns = 0; | |
898 | } | |
899 | else if (flag_schedule_insns == 2) | |
900 | flag_schedule_insns = 0; | |
901 | } | |
902 | ||
7a296495 CB |
903 | if ((target_flags_explicit & MASK_ACCUMULATE_OUTGOING_ARGS) == 0) |
904 | target_flags |= MASK_ACCUMULATE_OUTGOING_ARGS; | |
905 | ||
906 | /* Unwind info is not correct around the CFG unless either a frame | |
907 | pointer is present or M_A_O_A is set. Fixing this requires rewriting | |
908 | unwind info generation to be aware of the CFG and propagating states | |
909 | around edges. */ | |
910 | if ((flag_unwind_tables || flag_asynchronous_unwind_tables | |
911 | || flag_exceptions || flag_non_call_exceptions) | |
912 | && flag_omit_frame_pointer | |
913 | && !(target_flags & MASK_ACCUMULATE_OUTGOING_ARGS)) | |
914 | { | |
915 | if (target_flags_explicit & MASK_ACCUMULATE_OUTGOING_ARGS) | |
916 | warning (0, "unwind tables currently require either a frame pointer " | |
917 | "or -maccumulate-outgoing-args for correctness"); | |
918 | target_flags |= MASK_ACCUMULATE_OUTGOING_ARGS; | |
919 | } | |
920 | ||
91d1d521 KK |
921 | /* Unwinding with -freorder-blocks-and-partition does not work on this |
922 | architecture, because it requires far jumps to label crossing between | |
923 | hot/cold sections which are rejected on this architecture. */ | |
924 | if (flag_reorder_blocks_and_partition) | |
925 | { | |
926 | if (flag_exceptions) | |
927 | { | |
928 | inform (input_location, | |
929 | "-freorder-blocks-and-partition does not work with " | |
930 | "exceptions on this architecture"); | |
931 | flag_reorder_blocks_and_partition = 0; | |
932 | flag_reorder_blocks = 1; | |
933 | } | |
934 | else if (flag_unwind_tables) | |
935 | { | |
936 | inform (input_location, | |
937 | "-freorder-blocks-and-partition does not support unwind " | |
938 | "info on this architecture"); | |
939 | flag_reorder_blocks_and_partition = 0; | |
940 | flag_reorder_blocks = 1; | |
941 | } | |
942 | } | |
943 | ||
bd9a3465 KK |
944 | if (align_loops == 0) |
945 | align_loops = 1 << (TARGET_SH5 ? 3 : 2); | |
946 | if (align_jumps == 0) | |
947 | align_jumps = 1 << CACHE_LOG; | |
948 | else if (align_jumps < (TARGET_SHMEDIA ? 4 : 2)) | |
949 | align_jumps = TARGET_SHMEDIA ? 4 : 2; | |
950 | ||
951 | /* Allocation boundary (in *bytes*) for the code of a function. | |
952 | SH1: 32 bit alignment is faster, because instructions are always | |
953 | fetched as a pair from a longword boundary. | |
954 | SH2 .. SH5 : align to cache line start. */ | |
955 | if (align_functions == 0) | |
956 | align_functions | |
957 | = TARGET_SMALLCODE ? FUNCTION_BOUNDARY/8 : (1 << CACHE_LOG); | |
958 | /* The linker relaxation code breaks when a function contains | |
959 | alignments that are larger than that at the start of a | |
960 | compilation unit. */ | |
961 | if (TARGET_RELAX) | |
962 | { | |
963 | int min_align | |
964 | = align_loops > align_jumps ? align_loops : align_jumps; | |
965 | ||
966 | /* Also take possible .long constants / mova tables int account. */ | |
967 | if (min_align < 4) | |
968 | min_align = 4; | |
969 | if (align_functions < min_align) | |
970 | align_functions = min_align; | |
971 | } | |
972 | ||
973 | if (sh_fixed_range_str) | |
974 | sh_fix_range (sh_fixed_range_str); | |
975 | } | |
976 | \f | |
16bea517 | 977 | /* Print the operand address in x to the stream. */ |
bc45ade3 | 978 | |
9e94b957 NF |
979 | static void |
980 | sh_print_operand_address (FILE *stream, rtx x) | |
bc45ade3 SC |
981 | { |
982 | switch (GET_CODE (x)) | |
983 | { | |
984 | case REG: | |
225e4f43 R |
985 | case SUBREG: |
986 | fprintf (stream, "@%s", reg_names[true_regnum (x)]); | |
bc45ade3 | 987 | break; |
8aa2a305 | 988 | |
bc45ade3 SC |
989 | case PLUS: |
990 | { | |
991 | rtx base = XEXP (x, 0); | |
992 | rtx index = XEXP (x, 1); | |
993 | ||
bc45ade3 SC |
994 | switch (GET_CODE (index)) |
995 | { | |
996 | case CONST_INT: | |
c5b9ef02 | 997 | fprintf (stream, "@(%d,%s)", (int) INTVAL (index), |
225e4f43 | 998 | reg_names[true_regnum (base)]); |
bc45ade3 SC |
999 | break; |
1000 | ||
1001 | case REG: | |
225e4f43 R |
1002 | case SUBREG: |
1003 | { | |
1004 | int base_num = true_regnum (base); | |
1005 | int index_num = true_regnum (index); | |
1006 | ||
1007 | fprintf (stream, "@(r0,%s)", | |
1008 | reg_names[MAX (base_num, index_num)]); | |
1009 | break; | |
1010 | } | |
bc45ade3 SC |
1011 | |
1012 | default: | |
f5b9e7c9 | 1013 | gcc_unreachable (); |
bc45ade3 SC |
1014 | } |
1015 | } | |
bc45ade3 | 1016 | break; |
8aa2a305 | 1017 | |
bc45ade3 | 1018 | case PRE_DEC: |
225e4f43 | 1019 | fprintf (stream, "@-%s", reg_names[true_regnum (XEXP (x, 0))]); |
bc45ade3 SC |
1020 | break; |
1021 | ||
1022 | case POST_INC: | |
225e4f43 | 1023 | fprintf (stream, "@%s+", reg_names[true_regnum (XEXP (x, 0))]); |
bc45ade3 SC |
1024 | break; |
1025 | ||
1026 | default: | |
b91455de | 1027 | x = mark_constant_pool_use (x); |
422be3c3 | 1028 | output_addr_const (stream, x); |
bc45ade3 SC |
1029 | break; |
1030 | } | |
1031 | } | |
1032 | ||
1033 | /* Print operand x (an rtx) in assembler syntax to file stream | |
1034 | according to modifier code. | |
1035 | ||
b9654711 | 1036 | '.' print a .s if insn needs delay slot |
1245df60 | 1037 | ',' print LOCAL_LABEL_PREFIX |
4408efce | 1038 | '@' print trap, rte or rts depending upon pragma interruptness |
8aa2a305 | 1039 | '#' output a nop if there is nothing to put in the delay slot |
fae778eb | 1040 | ''' print likelihood suffix (/u for unlikely). |
73a4d10b | 1041 | '>' print branch target if -fverbose-asm |
8aa2a305 | 1042 | 'O' print a constant without the # |
00f8ff66 | 1043 | 'R' print the LSW of a dp value - changes if in little endian |
00f8ff66 | 1044 | 'S' print the MSW of a dp value - changes if in little endian |
225e4f43 | 1045 | 'T' print the next word of a dp value - same as 'R' in big endian mode. |
31b6f0ae R |
1046 | 'M' SHMEDIA: print an `x' if `m' will print `base,index'. |
1047 | otherwise: print .b / .w / .l / .s / .d suffix if operand is a MEM. | |
c1b92d09 | 1048 | 'N' print 'r63' if the operand is (const_int 0). |
312209c6 | 1049 | 'd' print a V2SF reg as dN instead of fpN. |
fa5322fa | 1050 | 'm' print a pair `base,offset' or `base,index', for LD and ST. |
73a4d10b | 1051 | 'U' Likewise for {LD,ST}{HI,LO}. |
de6adfa2 N |
1052 | 'V' print the position of a single bit set. |
1053 | 'W' print the position of a single bit cleared. | |
9eb3a0dd | 1054 | 't' print a memory address which is a register. |
fa5322fa | 1055 | 'u' prints the lowest 16 bits of CONST_INT, as an unsigned value. |
225e4f43 | 1056 | 'o' output an operator. */ |
bc45ade3 | 1057 | |
9e94b957 NF |
1058 | static void |
1059 | sh_print_operand (FILE *stream, rtx x, int code) | |
bc45ade3 | 1060 | { |
73a4d10b R |
1061 | int regno; |
1062 | enum machine_mode mode; | |
1063 | ||
bc45ade3 SC |
1064 | switch (code) |
1065 | { | |
a6ab9fc0 R |
1066 | tree trapa_attr; |
1067 | ||
b9654711 | 1068 | case '.': |
79b2746a | 1069 | if (final_sequence |
c608a684 R |
1070 | && ! INSN_ANNULLED_BRANCH_P (XVECEXP (final_sequence, 0, 0)) |
1071 | && get_attr_length (XVECEXP (final_sequence, 0, 1))) | |
1245df60 R |
1072 | fprintf (stream, ASSEMBLER_DIALECT ? "/s" : ".s"); |
1073 | break; | |
1074 | case ',': | |
1075 | fprintf (stream, "%s", LOCAL_LABEL_PREFIX); | |
b9654711 | 1076 | break; |
d3ae8277 | 1077 | case '@': |
a6ab9fc0 R |
1078 | trapa_attr = lookup_attribute ("trap_exit", |
1079 | DECL_ATTRIBUTES (current_function_decl)); | |
1080 | if (trapa_attr) | |
1081 | fprintf (stream, "trapa #%ld", | |
1082 | (long) TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (trapa_attr)))); | |
e3ba8d11 | 1083 | else if (sh_cfun_interrupt_handler_p ()) |
561642fa AP |
1084 | { |
1085 | if (sh_cfun_resbank_handler_p ()) | |
1086 | fprintf (stream, "resbank\n"); | |
1087 | fprintf (stream, "rte"); | |
1088 | } | |
d3ae8277 | 1089 | else |
8e87e161 | 1090 | fprintf (stream, "rts"); |
d3ae8277 | 1091 | break; |
bc45ade3 | 1092 | case '#': |
16bea517 | 1093 | /* Output a nop if there's nothing in the delay slot. */ |
bc45ade3 | 1094 | if (dbr_sequence_length () == 0) |
8aa2a305 | 1095 | fprintf (stream, "\n\tnop"); |
bc45ade3 | 1096 | break; |
c8cc4417 R |
1097 | case '\'': |
1098 | { | |
1099 | rtx note = find_reg_note (current_output_insn, REG_BR_PROB, 0); | |
1100 | ||
1101 | if (note && INTVAL (XEXP (note, 0)) * 2 < REG_BR_PROB_BASE) | |
1102 | fputs ("/u", stream); | |
1103 | break; | |
1104 | } | |
73a4d10b R |
1105 | case '>': |
1106 | if (flag_verbose_asm && JUMP_LABEL (current_output_insn)) | |
1107 | { | |
1108 | fputs ("\t! target: ", stream); | |
1109 | output_addr_const (stream, JUMP_LABEL (current_output_insn)); | |
1110 | } | |
1111 | break; | |
b9654711 | 1112 | case 'O': |
b91455de | 1113 | x = mark_constant_pool_use (x); |
0d7e008e | 1114 | output_addr_const (stream, x); |
bc45ade3 | 1115 | break; |
4802062c R |
1116 | /* N.B.: %R / %S / %T adjust memory addresses by four. |
1117 | For SHMEDIA, that means they can be used to access the first and | |
1118 | second 32 bit part of a 64 bit (or larger) value that | |
1119 | might be held in floating point registers or memory. | |
1120 | While they can be used to access 64 bit parts of a larger value | |
1121 | held in general purpose registers, that won't work with memory - | |
1122 | neither for fp registers, since the frxx names are used. */ | |
bc45ade3 | 1123 | case 'R': |
4802062c R |
1124 | if (REG_P (x) || GET_CODE (x) == SUBREG) |
1125 | { | |
1126 | regno = true_regnum (x); | |
1127 | regno += FP_REGISTER_P (regno) ? 1 : LSW; | |
1128 | fputs (reg_names[regno], (stream)); | |
1129 | } | |
1130 | else if (MEM_P (x)) | |
1131 | { | |
1132 | x = adjust_address (x, SImode, 4 * LSW); | |
9e94b957 | 1133 | sh_print_operand_address (stream, XEXP (x, 0)); |
4802062c R |
1134 | } |
1135 | else | |
1136 | { | |
1137 | rtx sub = NULL_RTX; | |
1138 | ||
1139 | mode = GET_MODE (x); | |
1140 | if (mode == VOIDmode) | |
1141 | mode = DImode; | |
1142 | if (GET_MODE_SIZE (mode) >= 8) | |
1143 | sub = simplify_subreg (SImode, x, mode, 4 * LSW); | |
1144 | if (sub) | |
9e94b957 | 1145 | sh_print_operand (stream, sub, 0); |
4802062c R |
1146 | else |
1147 | output_operand_lossage ("invalid operand to %%R"); | |
1148 | } | |
8aa2a305 JW |
1149 | break; |
1150 | case 'S': | |
4802062c R |
1151 | if (REG_P (x) || GET_CODE (x) == SUBREG) |
1152 | { | |
1153 | regno = true_regnum (x); | |
1154 | regno += FP_REGISTER_P (regno) ? 0 : MSW; | |
1155 | fputs (reg_names[regno], (stream)); | |
1156 | } | |
1157 | else if (MEM_P (x)) | |
1158 | { | |
1159 | x = adjust_address (x, SImode, 4 * MSW); | |
9e94b957 | 1160 | sh_print_operand_address (stream, XEXP (x, 0)); |
4802062c R |
1161 | } |
1162 | else | |
1163 | { | |
1164 | rtx sub = NULL_RTX; | |
1165 | ||
1166 | mode = GET_MODE (x); | |
1167 | if (mode == VOIDmode) | |
1168 | mode = DImode; | |
1169 | if (GET_MODE_SIZE (mode) >= 8) | |
1170 | sub = simplify_subreg (SImode, x, mode, 4 * MSW); | |
1171 | if (sub) | |
9e94b957 | 1172 | sh_print_operand (stream, sub, 0); |
4802062c R |
1173 | else |
1174 | output_operand_lossage ("invalid operand to %%S"); | |
1175 | } | |
00f8ff66 SC |
1176 | break; |
1177 | case 'T': | |
16bea517 | 1178 | /* Next word of a double. */ |
bc45ade3 SC |
1179 | switch (GET_CODE (x)) |
1180 | { | |
1181 | case REG: | |
1182 | fputs (reg_names[REGNO (x) + 1], (stream)); | |
1183 | break; | |
1184 | case MEM: | |
225e4f43 R |
1185 | if (GET_CODE (XEXP (x, 0)) != PRE_DEC |
1186 | && GET_CODE (XEXP (x, 0)) != POST_INC) | |
b72f00af | 1187 | x = adjust_address (x, SImode, 4); |
9e94b957 | 1188 | sh_print_operand_address (stream, XEXP (x, 0)); |
bc45ade3 | 1189 | break; |
318881c0 KG |
1190 | default: |
1191 | break; | |
bc45ade3 SC |
1192 | } |
1193 | break; | |
9eb3a0dd N |
1194 | |
1195 | case 't': | |
f3536097 | 1196 | gcc_assert (MEM_P (x)); |
9eb3a0dd N |
1197 | x = XEXP (x, 0); |
1198 | switch (GET_CODE (x)) | |
1199 | { | |
1200 | case REG: | |
1201 | case SUBREG: | |
9e94b957 | 1202 | sh_print_operand (stream, x, 0); |
9eb3a0dd N |
1203 | break; |
1204 | default: | |
1205 | break; | |
1206 | } | |
1207 | break; | |
1208 | ||
225e4f43 R |
1209 | case 'o': |
1210 | switch (GET_CODE (x)) | |
1211 | { | |
1212 | case PLUS: fputs ("add", stream); break; | |
1213 | case MINUS: fputs ("sub", stream); break; | |
1214 | case MULT: fputs ("mul", stream); break; | |
1215 | case DIV: fputs ("div", stream); break; | |
c8cc4417 R |
1216 | case EQ: fputs ("eq", stream); break; |
1217 | case NE: fputs ("ne", stream); break; | |
1218 | case GT: case LT: fputs ("gt", stream); break; | |
1219 | case GE: case LE: fputs ("ge", stream); break; | |
1220 | case GTU: case LTU: fputs ("gtu", stream); break; | |
1221 | case GEU: case LEU: fputs ("geu", stream); break; | |
318881c0 KG |
1222 | default: |
1223 | break; | |
225e4f43 R |
1224 | } |
1225 | break; | |
fa5322fa | 1226 | case 'M': |
31b6f0ae R |
1227 | if (TARGET_SHMEDIA) |
1228 | { | |
f3536097 | 1229 | if (MEM_P (x) |
31b6f0ae | 1230 | && GET_CODE (XEXP (x, 0)) == PLUS |
f3536097 | 1231 | && (REG_P (XEXP (XEXP (x, 0), 1)) |
31b6f0ae R |
1232 | || GET_CODE (XEXP (XEXP (x, 0), 1)) == SUBREG)) |
1233 | fputc ('x', stream); | |
1234 | } | |
1235 | else | |
1236 | { | |
f3536097 | 1237 | if (MEM_P (x)) |
31b6f0ae R |
1238 | { |
1239 | switch (GET_MODE (x)) | |
1240 | { | |
1241 | case QImode: fputs (".b", stream); break; | |
1242 | case HImode: fputs (".w", stream); break; | |
1243 | case SImode: fputs (".l", stream); break; | |
1244 | case SFmode: fputs (".s", stream); break; | |
1245 | case DFmode: fputs (".d", stream); break; | |
1246 | default: gcc_unreachable (); | |
1247 | } | |
1248 | } | |
1249 | } | |
fa5322fa AO |
1250 | break; |
1251 | ||
1252 | case 'm': | |
f3536097 | 1253 | gcc_assert (MEM_P (x)); |
fa5322fa | 1254 | x = XEXP (x, 0); |
73a4d10b R |
1255 | /* Fall through. */ |
1256 | case 'U': | |
fa5322fa AO |
1257 | switch (GET_CODE (x)) |
1258 | { | |
1259 | case REG: | |
1260 | case SUBREG: | |
9e94b957 | 1261 | sh_print_operand (stream, x, 0); |
fa5322fa AO |
1262 | fputs (", 0", stream); |
1263 | break; | |
1264 | ||
1265 | case PLUS: | |
9e94b957 | 1266 | sh_print_operand (stream, XEXP (x, 0), 0); |
fa5322fa | 1267 | fputs (", ", stream); |
9e94b957 | 1268 | sh_print_operand (stream, XEXP (x, 1), 0); |
fa5322fa AO |
1269 | break; |
1270 | ||
1271 | default: | |
f5b9e7c9 | 1272 | gcc_unreachable (); |
fa5322fa AO |
1273 | } |
1274 | break; | |
1275 | ||
de6adfa2 N |
1276 | case 'V': |
1277 | { | |
1278 | int num = exact_log2 (INTVAL (x)); | |
1279 | gcc_assert (num >= 0); | |
1280 | fprintf (stream, "#%d", num); | |
1281 | } | |
1282 | break; | |
1283 | ||
1284 | case 'W': | |
1285 | { | |
1286 | int num = exact_log2 (~INTVAL (x)); | |
1287 | gcc_assert (num >= 0); | |
1288 | fprintf (stream, "#%d", num); | |
1289 | } | |
1290 | break; | |
1291 | ||
312209c6 | 1292 | case 'd': |
f3536097 | 1293 | gcc_assert (REG_P (x) && GET_MODE (x) == V2SFmode); |
312209c6 AO |
1294 | |
1295 | fprintf ((stream), "d%s", reg_names[REGNO (x)] + 1); | |
1296 | break; | |
f676971a | 1297 | |
c1b92d09 | 1298 | case 'N': |
0ac78517 | 1299 | if (x == CONST0_RTX (GET_MODE (x))) |
c1b92d09 R |
1300 | { |
1301 | fprintf ((stream), "r63"); | |
1302 | break; | |
1303 | } | |
1304 | goto default_output; | |
fa5322fa | 1305 | case 'u': |
f3536097 | 1306 | if (CONST_INT_P (x)) |
c8cc4417 | 1307 | { |
fa5322fa AO |
1308 | fprintf ((stream), "%u", (unsigned) INTVAL (x) & (0x10000 - 1)); |
1309 | break; | |
1310 | } | |
1311 | /* Fall through. */ | |
1312 | ||
c1b92d09 | 1313 | default_output: |
bc45ade3 | 1314 | default: |
73a4d10b R |
1315 | regno = 0; |
1316 | mode = GET_MODE (x); | |
1317 | ||
bc45ade3 SC |
1318 | switch (GET_CODE (x)) |
1319 | { | |
73a4d10b R |
1320 | case TRUNCATE: |
1321 | { | |
1322 | rtx inner = XEXP (x, 0); | |
1323 | int offset = 0; | |
1324 | enum machine_mode inner_mode; | |
1325 | ||
1326 | /* We might see SUBREGs with vector mode registers inside. */ | |
1327 | if (GET_CODE (inner) == SUBREG | |
1328 | && (GET_MODE_SIZE (GET_MODE (inner)) | |
1329 | == GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner)))) | |
1330 | && subreg_lowpart_p (inner)) | |
1331 | inner = SUBREG_REG (inner); | |
f3536097 | 1332 | if (CONST_INT_P (inner)) |
73a4d10b R |
1333 | { |
1334 | x = GEN_INT (trunc_int_for_mode (INTVAL (inner), GET_MODE (x))); | |
1335 | goto default_output; | |
1336 | } | |
1337 | inner_mode = GET_MODE (inner); | |
1338 | if (GET_CODE (inner) == SUBREG | |
1339 | && (GET_MODE_SIZE (GET_MODE (inner)) | |
1340 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner)))) | |
f3536097 | 1341 | && REG_P (SUBREG_REG (inner))) |
73a4d10b R |
1342 | { |
1343 | offset = subreg_regno_offset (REGNO (SUBREG_REG (inner)), | |
1344 | GET_MODE (SUBREG_REG (inner)), | |
1345 | SUBREG_BYTE (inner), | |
1346 | GET_MODE (inner)); | |
1347 | inner = SUBREG_REG (inner); | |
1348 | } | |
f3536097 | 1349 | if (!REG_P (inner) || GET_MODE_SIZE (inner_mode) > 8) |
73a4d10b R |
1350 | abort (); |
1351 | /* Floating point register pairs are always big endian; | |
30dc60c7 | 1352 | general purpose registers are 64 bit wide. */ |
73a4d10b R |
1353 | regno = REGNO (inner); |
1354 | regno = (HARD_REGNO_NREGS (regno, inner_mode) | |
1355 | - HARD_REGNO_NREGS (regno, mode)) | |
1356 | + offset; | |
1357 | x = inner; | |
1358 | goto reg; | |
1359 | } | |
1360 | case SIGN_EXTEND: | |
1361 | x = XEXP (x, 0); | |
1362 | goto reg; | |
fa5322fa AO |
1363 | /* FIXME: We need this on SHmedia32 because reload generates |
1364 | some sign-extended HI or QI loads into DImode registers | |
1365 | but, because Pmode is SImode, the address ends up with a | |
1366 | subreg:SI of the DImode register. Maybe reload should be | |
1367 | fixed so as to apply alter_subreg to such loads? */ | |
73a4d10b R |
1368 | case IF_THEN_ELSE: |
1369 | gcc_assert (trapping_target_operand (x, VOIDmode)); | |
1370 | x = XEXP (XEXP (x, 2), 0); | |
1371 | goto default_output; | |
fa5322fa | 1372 | case SUBREG: |
f5b9e7c9 | 1373 | gcc_assert (SUBREG_BYTE (x) == 0 |
f3536097 | 1374 | && REG_P (SUBREG_REG (x))); |
fa5322fa AO |
1375 | |
1376 | x = SUBREG_REG (x); | |
1377 | /* Fall through. */ | |
1378 | ||
73a4d10b | 1379 | reg: |
bc45ade3 | 1380 | case REG: |
73a4d10b R |
1381 | regno += REGNO (x); |
1382 | if (FP_REGISTER_P (regno) | |
1383 | && mode == V16SFmode) | |
1384 | fprintf ((stream), "mtrx%s", reg_names[regno] + 2); | |
fa5322fa | 1385 | else if (FP_REGISTER_P (REGNO (x)) |
73a4d10b R |
1386 | && mode == V4SFmode) |
1387 | fprintf ((stream), "fv%s", reg_names[regno] + 2); | |
f3536097 | 1388 | else if (REG_P (x) |
73a4d10b R |
1389 | && mode == V2SFmode) |
1390 | fprintf ((stream), "fp%s", reg_names[regno] + 2); | |
fa5322fa | 1391 | else if (FP_REGISTER_P (REGNO (x)) |
73a4d10b R |
1392 | && GET_MODE_SIZE (mode) > 4) |
1393 | fprintf ((stream), "d%s", reg_names[regno] + 1); | |
225e4f43 | 1394 | else |
73a4d10b | 1395 | fputs (reg_names[regno], (stream)); |
bc45ade3 | 1396 | break; |
fa5322fa | 1397 | |
bc45ade3 SC |
1398 | case MEM: |
1399 | output_address (XEXP (x, 0)); | |
1400 | break; | |
f676971a | 1401 | |
bc45ade3 | 1402 | default: |
fa5322fa AO |
1403 | if (TARGET_SH1) |
1404 | fputc ('#', stream); | |
bc45ade3 SC |
1405 | output_addr_const (stream, x); |
1406 | break; | |
bc45ade3 SC |
1407 | } |
1408 | break; | |
1409 | } | |
1410 | } | |
9e94b957 NF |
1411 | |
1412 | static bool | |
1413 | sh_print_operand_punct_valid_p (unsigned char code) | |
1414 | { | |
1415 | return (code == '.' || code == '#' || code == '@' || code == ',' | |
1416 | || code == '$' || code == '\'' || code == '>'); | |
1417 | } | |
bc45ade3 | 1418 | \f |
561642fa AP |
1419 | |
1420 | /* Encode symbol attributes of a SYMBOL_REF into its | |
1421 | SYMBOL_REF_FLAGS. */ | |
1422 | static void | |
1423 | sh_encode_section_info (tree decl, rtx rtl, int first) | |
1424 | { | |
1425 | default_encode_section_info (decl, rtl, first); | |
1426 | ||
1427 | if (TREE_CODE (decl) == FUNCTION_DECL | |
1428 | && sh2a_function_vector_p (decl) && TARGET_SH2A) | |
1429 | SYMBOL_REF_FLAGS (XEXP (rtl, 0)) |= SYMBOL_FLAG_FUNCVEC_FUNCTION; | |
1430 | } | |
1431 | ||
0aa692fc R |
1432 | /* Like force_operand, but guarantees that VALUE ends up in TARGET. */ |
1433 | static void | |
cf277499 | 1434 | force_into (rtx value, rtx target) |
0aa692fc R |
1435 | { |
1436 | value = force_operand (value, target); | |
1437 | if (! rtx_equal_p (value, target)) | |
1438 | emit_insn (gen_move_insn (target, value)); | |
1439 | } | |
1440 | ||
0d7e008e SC |
1441 | /* Emit code to perform a block move. Choose the best method. |
1442 | ||
1443 | OPERANDS[0] is the destination. | |
1444 | OPERANDS[1] is the source. | |
1445 | OPERANDS[2] is the size. | |
1446 | OPERANDS[3] is the alignment safe to use. */ | |
1447 | ||
0d7e008e | 1448 | int |
cf277499 | 1449 | expand_block_move (rtx *operands) |
0d7e008e SC |
1450 | { |
1451 | int align = INTVAL (operands[3]); | |
f3536097 | 1452 | int constp = (CONST_INT_P (operands[2])); |
0d7e008e | 1453 | int bytes = (constp ? INTVAL (operands[2]) : 0); |
0d7e008e | 1454 | |
312209c6 AO |
1455 | if (! constp) |
1456 | return 0; | |
1457 | ||
1458 | /* If we could use mov.l to move words and dest is word-aligned, we | |
1459 | can use movua.l for loads and still generate a relatively short | |
1460 | and efficient sequence. */ | |
1461 | if (TARGET_SH4A_ARCH && align < 4 | |
1462 | && MEM_ALIGN (operands[0]) >= 32 | |
1463 | && can_move_by_pieces (bytes, 32)) | |
1464 | { | |
1465 | rtx dest = copy_rtx (operands[0]); | |
1466 | rtx src = copy_rtx (operands[1]); | |
1467 | /* We could use different pseudos for each copied word, but | |
1468 | since movua can only load into r0, it's kind of | |
1469 | pointless. */ | |
1470 | rtx temp = gen_reg_rtx (SImode); | |
1471 | rtx src_addr = copy_addr_to_reg (XEXP (src, 0)); | |
1472 | int copied = 0; | |
1473 | ||
1474 | while (copied + 4 <= bytes) | |
1475 | { | |
1476 | rtx to = adjust_address (dest, SImode, copied); | |
85144c5c RS |
1477 | rtx from = adjust_automodify_address (src, BLKmode, |
1478 | src_addr, copied); | |
312209c6 | 1479 | |
85144c5c | 1480 | set_mem_size (from, GEN_INT (4)); |
312209c6 AO |
1481 | emit_insn (gen_movua (temp, from)); |
1482 | emit_move_insn (src_addr, plus_constant (src_addr, 4)); | |
1483 | emit_move_insn (to, temp); | |
1484 | copied += 4; | |
1485 | } | |
1486 | ||
1487 | if (copied < bytes) | |
1488 | move_by_pieces (adjust_address (dest, BLKmode, copied), | |
1489 | adjust_automodify_address (src, BLKmode, | |
1490 | src_addr, copied), | |
1491 | bytes - copied, align, 0); | |
1492 | ||
1493 | return 1; | |
1494 | } | |
1495 | ||
8aa2a305 JW |
1496 | /* If it isn't a constant number of bytes, or if it doesn't have 4 byte |
1497 | alignment, or if it isn't a multiple of 4 bytes, then fail. */ | |
312209c6 | 1498 | if (align < 4 || (bytes % 4 != 0)) |
d3ae8277 SC |
1499 | return 0; |
1500 | ||
225e4f43 R |
1501 | if (TARGET_HARD_SH4) |
1502 | { | |
1503 | if (bytes < 12) | |
1504 | return 0; | |
1505 | else if (bytes == 12) | |
1506 | { | |
73a4d10b | 1507 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
c0d4e710 KH |
1508 | rtx r4 = gen_rtx_REG (SImode, 4); |
1509 | rtx r5 = gen_rtx_REG (SImode, 5); | |
225e4f43 | 1510 | |
73a4d10b | 1511 | function_symbol (func_addr_rtx, "__movmemSI12_i4", SFUNC_STATIC); |
0aa692fc R |
1512 | force_into (XEXP (operands[0], 0), r4); |
1513 | force_into (XEXP (operands[1], 0), r5); | |
225e4f43 R |
1514 | emit_insn (gen_block_move_real_i4 (func_addr_rtx)); |
1515 | return 1; | |
1516 | } | |
1517 | else if (! TARGET_SMALLCODE) | |
1518 | { | |
73a4d10b R |
1519 | const char *entry_name; |
1520 | rtx func_addr_rtx = gen_reg_rtx (Pmode); | |
225e4f43 | 1521 | int dwords; |
c0d4e710 KH |
1522 | rtx r4 = gen_rtx_REG (SImode, 4); |
1523 | rtx r5 = gen_rtx_REG (SImode, 5); | |
1524 | rtx r6 = gen_rtx_REG (SImode, 6); | |
225e4f43 | 1525 | |
73a4d10b R |
1526 | entry_name = (bytes & 4 ? "__movmem_i4_odd" : "__movmem_i4_even"); |
1527 | function_symbol (func_addr_rtx, entry_name, SFUNC_STATIC); | |
0aa692fc R |
1528 | force_into (XEXP (operands[0], 0), r4); |
1529 | force_into (XEXP (operands[1], 0), r5); | |
225e4f43 R |
1530 | |
1531 | dwords = bytes >> 3; | |
1532 | emit_insn (gen_move_insn (r6, GEN_INT (dwords - 1))); | |
1533 | emit_insn (gen_block_lump_real_i4 (func_addr_rtx)); | |
1534 | return 1; | |
1535 | } | |
1536 | else | |
1537 | return 0; | |
1538 | } | |
8aa2a305 | 1539 | if (bytes < 64) |
0d7e008e SC |
1540 | { |
1541 | char entry[30]; | |
73a4d10b | 1542 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
c5c76735 JL |
1543 | rtx r4 = gen_rtx_REG (SImode, 4); |
1544 | rtx r5 = gen_rtx_REG (SImode, 5); | |
8aa2a305 | 1545 | |
70128ad9 | 1546 | sprintf (entry, "__movmemSI%d", bytes); |
73a4d10b | 1547 | function_symbol (func_addr_rtx, entry, SFUNC_STATIC); |
0aa692fc R |
1548 | force_into (XEXP (operands[0], 0), r4); |
1549 | force_into (XEXP (operands[1], 0), r5); | |
0d7e008e SC |
1550 | emit_insn (gen_block_move_real (func_addr_rtx)); |
1551 | return 1; | |
1552 | } | |
8aa2a305 JW |
1553 | |
1554 | /* This is the same number of bytes as a memcpy call, but to a different | |
1555 | less common function name, so this will occasionally use more space. */ | |
1556 | if (! TARGET_SMALLCODE) | |
0d7e008e | 1557 | { |
73a4d10b | 1558 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
8aa2a305 | 1559 | int final_switch, while_loop; |
c5c76735 JL |
1560 | rtx r4 = gen_rtx_REG (SImode, 4); |
1561 | rtx r5 = gen_rtx_REG (SImode, 5); | |
1562 | rtx r6 = gen_rtx_REG (SImode, 6); | |
0d7e008e | 1563 | |
73a4d10b | 1564 | function_symbol (func_addr_rtx, "__movmem", SFUNC_STATIC); |
0aa692fc R |
1565 | force_into (XEXP (operands[0], 0), r4); |
1566 | force_into (XEXP (operands[1], 0), r5); | |
0d7e008e | 1567 | |
8aa2a305 JW |
1568 | /* r6 controls the size of the move. 16 is decremented from it |
1569 | for each 64 bytes moved. Then the negative bit left over is used | |
1570 | as an index into a list of move instructions. e.g., a 72 byte move | |
1571 | would be set up with size(r6) = 14, for one iteration through the | |
1572 | big while loop, and a switch of -2 for the last part. */ | |
0d7e008e | 1573 | |
8aa2a305 JW |
1574 | final_switch = 16 - ((bytes / 4) % 16); |
1575 | while_loop = ((bytes / 4) / 16 - 1) * 16; | |
1576 | emit_insn (gen_move_insn (r6, GEN_INT (while_loop + final_switch))); | |
1577 | emit_insn (gen_block_lump_real (func_addr_rtx)); | |
1578 | return 1; | |
0d7e008e | 1579 | } |
0d7e008e | 1580 | |
d3ae8277 | 1581 | return 0; |
0d7e008e SC |
1582 | } |
1583 | ||
bc45ade3 | 1584 | /* Prepare operands for a move define_expand; specifically, one of the |
8aa2a305 | 1585 | operands must be in a register. */ |
bc45ade3 | 1586 | |
b9654711 | 1587 | int |
cf277499 | 1588 | prepare_move_operands (rtx operands[], enum machine_mode mode) |
bc45ade3 | 1589 | { |
463f02cd KK |
1590 | if ((mode == SImode || mode == DImode) |
1591 | && flag_pic | |
1592 | && ! ((mode == Pmode || mode == ptr_mode) | |
81f40b79 | 1593 | && tls_symbolic_operand (operands[1], Pmode) != TLS_MODEL_NONE)) |
1a66cd67 AO |
1594 | { |
1595 | rtx temp; | |
1596 | if (SYMBOLIC_CONST_P (operands[1])) | |
1597 | { | |
f3536097 | 1598 | if (MEM_P (operands[0])) |
1a66cd67 | 1599 | operands[1] = force_reg (Pmode, operands[1]); |
5615d8fd SC |
1600 | else if (TARGET_SHMEDIA |
1601 | && GET_CODE (operands[1]) == LABEL_REF | |
fa5322fa AO |
1602 | && target_reg_operand (operands[0], mode)) |
1603 | /* It's ok. */; | |
1a66cd67 AO |
1604 | else |
1605 | { | |
b3a13419 ILT |
1606 | temp = (!can_create_pseudo_p () |
1607 | ? operands[0] | |
1608 | : gen_reg_rtx (Pmode)); | |
fa5322fa | 1609 | operands[1] = legitimize_pic_address (operands[1], mode, temp); |
1a66cd67 AO |
1610 | } |
1611 | } | |
1612 | else if (GET_CODE (operands[1]) == CONST | |
1613 | && GET_CODE (XEXP (operands[1], 0)) == PLUS | |
1614 | && SYMBOLIC_CONST_P (XEXP (XEXP (operands[1], 0), 0))) | |
1615 | { | |
b3a13419 | 1616 | temp = !can_create_pseudo_p () ? operands[0] : gen_reg_rtx (Pmode); |
1a66cd67 | 1617 | temp = legitimize_pic_address (XEXP (XEXP (operands[1], 0), 0), |
fa5322fa AO |
1618 | mode, temp); |
1619 | operands[1] = expand_binop (mode, add_optab, temp, | |
1a66cd67 | 1620 | XEXP (XEXP (operands[1], 0), 1), |
b3a13419 ILT |
1621 | (!can_create_pseudo_p () |
1622 | ? temp | |
1623 | : gen_reg_rtx (Pmode)), | |
68c37dc9 | 1624 | 0, OPTAB_LIB_WIDEN); |
1a66cd67 AO |
1625 | } |
1626 | } | |
1627 | ||
e22856e8 JW |
1628 | if (! reload_in_progress && ! reload_completed) |
1629 | { | |
1630 | /* Copy the source to a register if both operands aren't registers. */ | |
1631 | if (! register_operand (operands[0], mode) | |
d9da94a1 | 1632 | && ! sh_register_operand (operands[1], mode)) |
e22856e8 JW |
1633 | operands[1] = copy_to_mode_reg (mode, operands[1]); |
1634 | ||
f3536097 | 1635 | if (MEM_P (operands[0]) && ! memory_operand (operands[0], mode)) |
266a2732 R |
1636 | { |
1637 | /* This is like change_address_1 (operands[0], mode, 0, 1) , | |
1638 | except that we can't use that function because it is static. */ | |
0a2aaacc KG |
1639 | rtx new_rtx = change_address (operands[0], mode, 0); |
1640 | MEM_COPY_ATTRIBUTES (new_rtx, operands[0]); | |
1641 | operands[0] = new_rtx; | |
266a2732 R |
1642 | } |
1643 | ||
e22856e8 JW |
1644 | /* This case can happen while generating code to move the result |
1645 | of a library call to the target. Reject `st r0,@(rX,rY)' because | |
1646 | reload will fail to find a spill register for rX, since r0 is already | |
1647 | being used for the source. */ | |
73a4d10b R |
1648 | else if (TARGET_SH1 |
1649 | && refers_to_regno_p (R0_REG, R0_REG + 1, operands[1], (rtx *)0) | |
f3536097 | 1650 | && MEM_P (operands[0]) |
e22856e8 | 1651 | && GET_CODE (XEXP (operands[0], 0)) == PLUS |
f3536097 | 1652 | && REG_P (XEXP (XEXP (operands[0], 0), 1))) |
e22856e8 JW |
1653 | operands[1] = copy_to_mode_reg (mode, operands[1]); |
1654 | } | |
8aa2a305 | 1655 | |
463f02cd KK |
1656 | if (mode == Pmode || mode == ptr_mode) |
1657 | { | |
6058b2fa | 1658 | rtx op0, op1, opc; |
463f02cd KK |
1659 | enum tls_model tls_kind; |
1660 | ||
1661 | op0 = operands[0]; | |
1662 | op1 = operands[1]; | |
6058b2fa KK |
1663 | if (GET_CODE (op1) == CONST |
1664 | && GET_CODE (XEXP (op1, 0)) == PLUS | |
81f40b79 ILT |
1665 | && (tls_symbolic_operand (XEXP (XEXP (op1, 0), 0), Pmode) |
1666 | != TLS_MODEL_NONE)) | |
6058b2fa KK |
1667 | { |
1668 | opc = XEXP (XEXP (op1, 0), 1); | |
1669 | op1 = XEXP (XEXP (op1, 0), 0); | |
1670 | } | |
1671 | else | |
1672 | opc = NULL_RTX; | |
1673 | ||
81f40b79 | 1674 | if ((tls_kind = tls_symbolic_operand (op1, Pmode)) != TLS_MODEL_NONE) |
463f02cd KK |
1675 | { |
1676 | rtx tga_op1, tga_ret, tmp, tmp2; | |
1677 | ||
463f02cd KK |
1678 | switch (tls_kind) |
1679 | { | |
1680 | case TLS_MODEL_GLOBAL_DYNAMIC: | |
1681 | tga_ret = gen_rtx_REG (Pmode, R0_REG); | |
13ecc9e0 | 1682 | emit_call_insn (gen_tls_global_dynamic (tga_ret, op1)); |
463f02cd KK |
1683 | op1 = tga_ret; |
1684 | break; | |
1685 | ||
1686 | case TLS_MODEL_LOCAL_DYNAMIC: | |
1687 | tga_ret = gen_rtx_REG (Pmode, R0_REG); | |
13ecc9e0 | 1688 | emit_call_insn (gen_tls_local_dynamic (tga_ret, op1)); |
463f02cd KK |
1689 | |
1690 | tmp = gen_reg_rtx (Pmode); | |
1691 | emit_move_insn (tmp, tga_ret); | |
1692 | ||
1693 | if (register_operand (op0, Pmode)) | |
1694 | tmp2 = op0; | |
1695 | else | |
1696 | tmp2 = gen_reg_rtx (Pmode); | |
1697 | ||
1698 | emit_insn (gen_symDTPOFF2reg (tmp2, op1, tmp)); | |
1699 | op1 = tmp2; | |
1700 | break; | |
1701 | ||
1702 | case TLS_MODEL_INITIAL_EXEC: | |
1703 | if (! flag_pic) | |
b741fa00 | 1704 | { |
b92c3264 KK |
1705 | /* Don't schedule insns for getting GOT address when |
1706 | the first scheduling is enabled, to avoid spill | |
1707 | failures for R0. */ | |
1708 | if (flag_schedule_insns) | |
1709 | emit_insn (gen_blockage ()); | |
b741fa00 | 1710 | emit_insn (gen_GOTaddr2picreg ()); |
c41c1387 | 1711 | emit_use (gen_rtx_REG (SImode, PIC_REG)); |
b92c3264 KK |
1712 | if (flag_schedule_insns) |
1713 | emit_insn (gen_blockage ()); | |
b741fa00 | 1714 | } |
b3a13419 | 1715 | tga_op1 = !can_create_pseudo_p () ? op0 : gen_reg_rtx (Pmode); |
463f02cd KK |
1716 | tmp = gen_sym2GOTTPOFF (op1); |
1717 | emit_insn (gen_tls_initial_exec (tga_op1, tmp)); | |
1718 | op1 = tga_op1; | |
1719 | break; | |
1720 | ||
1721 | case TLS_MODEL_LOCAL_EXEC: | |
1722 | tmp2 = gen_reg_rtx (Pmode); | |
1723 | emit_insn (gen_load_gbr (tmp2)); | |
1724 | tmp = gen_reg_rtx (Pmode); | |
1725 | emit_insn (gen_symTPOFF2reg (tmp, op1)); | |
463f02cd KK |
1726 | |
1727 | if (register_operand (op0, Pmode)) | |
1728 | op1 = op0; | |
1729 | else | |
1730 | op1 = gen_reg_rtx (Pmode); | |
1731 | ||
1732 | emit_insn (gen_addsi3 (op1, tmp, tmp2)); | |
1733 | break; | |
1734 | ||
1735 | default: | |
f5b9e7c9 | 1736 | gcc_unreachable (); |
463f02cd | 1737 | } |
6058b2fa KK |
1738 | if (opc) |
1739 | emit_insn (gen_addsi3 (op1, op1, force_reg (SImode, opc))); | |
463f02cd KK |
1740 | operands[1] = op1; |
1741 | } | |
1742 | } | |
1743 | ||
8aa2a305 JW |
1744 | return 0; |
1745 | } | |
1746 | ||
78d310c2 R |
1747 | enum rtx_code |
1748 | prepare_cbranch_operands (rtx *operands, enum machine_mode mode, | |
1749 | enum rtx_code comparison) | |
1750 | { | |
1751 | rtx op1; | |
1752 | rtx scratch = NULL_RTX; | |
1753 | ||
b757d36f | 1754 | if (comparison == LAST_AND_UNUSED_RTX_CODE) |
78d310c2 R |
1755 | comparison = GET_CODE (operands[0]); |
1756 | else | |
1757 | scratch = operands[4]; | |
f3536097 SZ |
1758 | if (CONST_INT_P (operands[1]) |
1759 | && !CONST_INT_P (operands[2])) | |
78d310c2 R |
1760 | { |
1761 | rtx tmp = operands[1]; | |
1762 | ||
1763 | operands[1] = operands[2]; | |
1764 | operands[2] = tmp; | |
1765 | comparison = swap_condition (comparison); | |
1766 | } | |
f3536097 | 1767 | if (CONST_INT_P (operands[2])) |
78d310c2 R |
1768 | { |
1769 | HOST_WIDE_INT val = INTVAL (operands[2]); | |
1770 | if ((val == -1 || val == -0x81) | |
1771 | && (comparison == GT || comparison == LE)) | |
1772 | { | |
1773 | comparison = (comparison == GT) ? GE : LT; | |
1774 | operands[2] = gen_int_mode (val + 1, mode); | |
1775 | } | |
1776 | else if ((val == 1 || val == 0x80) | |
1777 | && (comparison == GE || comparison == LT)) | |
1778 | { | |
1779 | comparison = (comparison == GE) ? GT : LE; | |
1780 | operands[2] = gen_int_mode (val - 1, mode); | |
1781 | } | |
1782 | else if (val == 1 && (comparison == GEU || comparison == LTU)) | |
1783 | { | |
1784 | comparison = (comparison == GEU) ? NE : EQ; | |
1785 | operands[2] = CONST0_RTX (mode); | |
1786 | } | |
1787 | else if (val == 0x80 && (comparison == GEU || comparison == LTU)) | |
1788 | { | |
1789 | comparison = (comparison == GEU) ? GTU : LEU; | |
1790 | operands[2] = gen_int_mode (val - 1, mode); | |
1791 | } | |
1792 | else if (val == 0 && (comparison == GTU || comparison == LEU)) | |
1793 | comparison = (comparison == GTU) ? NE : EQ; | |
1794 | else if (mode == SImode | |
1795 | && ((val == 0x7fffffff | |
1796 | && (comparison == GTU || comparison == LEU)) | |
1797 | || ((unsigned HOST_WIDE_INT) val | |
1798 | == (unsigned HOST_WIDE_INT) 0x7fffffff + 1 | |
1799 | && (comparison == GEU || comparison == LTU)))) | |
1800 | { | |
1801 | comparison = (comparison == GTU || comparison == GEU) ? LT : GE; | |
1802 | operands[2] = CONST0_RTX (mode); | |
1803 | } | |
1804 | } | |
1805 | op1 = operands[1]; | |
b3a13419 | 1806 | if (can_create_pseudo_p ()) |
78d310c2 R |
1807 | operands[1] = force_reg (mode, op1); |
1808 | /* When we are handling DImode comparisons, we want to keep constants so | |
1809 | that we can optimize the component comparisons; however, memory loads | |
1810 | are better issued as a whole so that they can be scheduled well. | |
1811 | SImode equality comparisons allow I08 constants, but only when they | |
1812 | compare r0. Hence, if operands[1] has to be loaded from somewhere else | |
1813 | into a register, that register might as well be r0, and we allow the | |
1814 | constant. If it is already in a register, this is likely to be | |
2f8e468b | 1815 | allocated to a different hard register, thus we load the constant into |
78d310c2 R |
1816 | a register unless it is zero. */ |
1817 | if (!REG_P (operands[2]) | |
f3536097 | 1818 | && (!CONST_INT_P (operands[2]) |
78d310c2 R |
1819 | || (mode == SImode && operands[2] != CONST0_RTX (SImode) |
1820 | && ((comparison != EQ && comparison != NE) | |
1821 | || (REG_P (op1) && REGNO (op1) != R0_REG) | |
32a7ab3d | 1822 | || !satisfies_constraint_I08 (operands[2]))))) |
78d310c2 R |
1823 | { |
1824 | if (scratch && GET_MODE (scratch) == mode) | |
1825 | { | |
1826 | emit_move_insn (scratch, operands[2]); | |
1827 | operands[2] = scratch; | |
1828 | } | |
b3a13419 | 1829 | else if (can_create_pseudo_p ()) |
78d310c2 R |
1830 | operands[2] = force_reg (mode, operands[2]); |
1831 | } | |
1832 | return comparison; | |
1833 | } | |
1834 | ||
1835 | void | |
1836 | expand_cbranchsi4 (rtx *operands, enum rtx_code comparison, int probability) | |
1837 | { | |
1838 | rtx (*branch_expander) (rtx) = gen_branch_true; | |
1839 | rtx jump; | |
1840 | ||
1841 | comparison = prepare_cbranch_operands (operands, SImode, comparison); | |
1842 | switch (comparison) | |
1843 | { | |
1844 | case NE: case LT: case LE: case LTU: case LEU: | |
1845 | comparison = reverse_condition (comparison); | |
1846 | branch_expander = gen_branch_false; | |
1847 | default: ; | |
1848 | } | |
1849 | emit_insn (gen_rtx_SET (VOIDmode, gen_rtx_REG (SImode, T_REG), | |
1850 | gen_rtx_fmt_ee (comparison, SImode, | |
1851 | operands[1], operands[2]))); | |
1852 | jump = emit_jump_insn (branch_expander (operands[3])); | |
1853 | if (probability >= 0) | |
b757d36f | 1854 | add_reg_note (jump, REG_BR_PROB, GEN_INT (probability)); |
78d310c2 R |
1855 | |
1856 | } | |
1857 | ||
1858 | /* ??? How should we distribute probabilities when more than one branch | |
1859 | is generated. So far we only have soem ad-hoc observations: | |
1860 | - If the operands are random, they are likely to differ in both parts. | |
1861 | - If comparing items in a hash chain, the operands are random or equal; | |
1862 | operation should be EQ or NE. | |
1863 | - If items are searched in an ordered tree from the root, we can expect | |
1864 | the highpart to be unequal about half of the time; operation should be | |
2f8e468b | 1865 | an inequality comparison, operands non-constant, and overall probability |
78d310c2 R |
1866 | about 50%. Likewise for quicksort. |
1867 | - Range checks will be often made against constants. Even if we assume for | |
1868 | simplicity an even distribution of the non-constant operand over a | |
1869 | sub-range here, the same probability could be generated with differently | |
1870 | wide sub-ranges - as long as the ratio of the part of the subrange that | |
1871 | is before the threshold to the part that comes after the threshold stays | |
1872 | the same. Thus, we can't really tell anything here; | |
1873 | assuming random distribution is at least simple. | |
1874 | */ | |
1875 | ||
1876 | bool | |
1877 | expand_cbranchdi4 (rtx *operands, enum rtx_code comparison) | |
1878 | { | |
1879 | enum rtx_code msw_taken, msw_skip, lsw_taken; | |
cb3d3842 | 1880 | rtx skip_label = NULL_RTX; |
78d310c2 R |
1881 | rtx op1h, op1l, op2h, op2l; |
1882 | int num_branches; | |
1883 | int prob, rev_prob; | |
1884 | int msw_taken_prob = -1, msw_skip_prob = -1, lsw_taken_prob = -1; | |
6f8eec29 | 1885 | rtx scratch = operands[4]; |
78d310c2 R |
1886 | |
1887 | comparison = prepare_cbranch_operands (operands, DImode, comparison); | |
1888 | op1h = gen_highpart_mode (SImode, DImode, operands[1]); | |
1889 | op2h = gen_highpart_mode (SImode, DImode, operands[2]); | |
1890 | op1l = gen_lowpart (SImode, operands[1]); | |
1891 | op2l = gen_lowpart (SImode, operands[2]); | |
b757d36f | 1892 | msw_taken = msw_skip = lsw_taken = LAST_AND_UNUSED_RTX_CODE; |
78d310c2 R |
1893 | prob = split_branch_probability; |
1894 | rev_prob = REG_BR_PROB_BASE - prob; | |
1895 | switch (comparison) | |
1896 | { | |
1897 | /* ??? Should we use the cmpeqdi_t pattern for equality comparisons? | |
1898 | That costs 1 cycle more when the first branch can be predicted taken, | |
1899 | but saves us mispredicts because only one branch needs prediction. | |
1900 | It also enables generating the cmpeqdi_t-1 pattern. */ | |
1901 | case EQ: | |
1902 | if (TARGET_CMPEQDI_T) | |
1903 | { | |
1904 | emit_insn (gen_cmpeqdi_t (operands[1], operands[2])); | |
1905 | emit_jump_insn (gen_branch_true (operands[3])); | |
1906 | return true; | |
1907 | } | |
1908 | msw_skip = NE; | |
1909 | lsw_taken = EQ; | |
1910 | if (prob >= 0) | |
1911 | { | |
1912 | /* If we had more precision, we'd use rev_prob - (rev_prob >> 32) . | |
1913 | */ | |
1914 | msw_skip_prob = rev_prob; | |
1915 | if (REG_BR_PROB_BASE <= 65535) | |
1916 | lsw_taken_prob = prob ? REG_BR_PROB_BASE : 0; | |
1917 | else | |
1918 | { | |
1919 | gcc_assert (HOST_BITS_PER_WIDEST_INT >= 64); | |
1920 | lsw_taken_prob | |
1921 | = (prob | |
1922 | ? (REG_BR_PROB_BASE | |
1923 | - ((HOST_WIDEST_INT) REG_BR_PROB_BASE * rev_prob | |
1924 | / ((HOST_WIDEST_INT) prob << 32))) | |
1925 | : 0); | |
1926 | } | |
1927 | } | |
1928 | break; | |
1929 | case NE: | |
1930 | if (TARGET_CMPEQDI_T) | |
1931 | { | |
1932 | emit_insn (gen_cmpeqdi_t (operands[1], operands[2])); | |
1933 | emit_jump_insn (gen_branch_false (operands[3])); | |
1934 | return true; | |
1935 | } | |
1936 | msw_taken = NE; | |
6f8eec29 | 1937 | msw_taken_prob = prob; |
78d310c2 R |
1938 | lsw_taken = NE; |
1939 | lsw_taken_prob = 0; | |
1940 | break; | |
1941 | case GTU: case GT: | |
1942 | msw_taken = comparison; | |
f3536097 | 1943 | if (CONST_INT_P (op2l) && INTVAL (op2l) == -1) |
78d310c2 R |
1944 | break; |
1945 | if (comparison != GTU || op2h != CONST0_RTX (SImode)) | |
1946 | msw_skip = swap_condition (msw_taken); | |
1947 | lsw_taken = GTU; | |
1948 | break; | |
1949 | case GEU: case GE: | |
1950 | if (op2l == CONST0_RTX (SImode)) | |
1951 | msw_taken = comparison; | |
1952 | else | |
1953 | { | |
1954 | msw_taken = comparison == GE ? GT : GTU; | |
1955 | msw_skip = swap_condition (msw_taken); | |
1956 | lsw_taken = GEU; | |
1957 | } | |
1958 | break; | |
1959 | case LTU: case LT: | |
1960 | msw_taken = comparison; | |
1961 | if (op2l == CONST0_RTX (SImode)) | |
1962 | break; | |
1963 | msw_skip = swap_condition (msw_taken); | |
1964 | lsw_taken = LTU; | |
1965 | break; | |
1966 | case LEU: case LE: | |
f3536097 | 1967 | if (CONST_INT_P (op2l) && INTVAL (op2l) == -1) |
78d310c2 R |
1968 | msw_taken = comparison; |
1969 | else | |
1970 | { | |
1971 | lsw_taken = LEU; | |
1972 | if (comparison == LE) | |
1973 | msw_taken = LT; | |
1974 | else if (op2h != CONST0_RTX (SImode)) | |
1975 | msw_taken = LTU; | |
1976 | else | |
1977 | break; | |
1978 | msw_skip = swap_condition (msw_taken); | |
1979 | } | |
1980 | break; | |
1981 | default: return false; | |
1982 | } | |
b757d36f KK |
1983 | num_branches = ((msw_taken != LAST_AND_UNUSED_RTX_CODE) |
1984 | + (msw_skip != LAST_AND_UNUSED_RTX_CODE) | |
1985 | + (lsw_taken != LAST_AND_UNUSED_RTX_CODE)); | |
78d310c2 R |
1986 | if (comparison != EQ && comparison != NE && num_branches > 1) |
1987 | { | |
1988 | if (!CONSTANT_P (operands[2]) | |
1989 | && prob >= (int) (REG_BR_PROB_BASE * 3 / 8U) | |
1990 | && prob <= (int) (REG_BR_PROB_BASE * 5 / 8U)) | |
1991 | { | |
1992 | msw_taken_prob = prob / 2U; | |
1993 | msw_skip_prob | |
1994 | = REG_BR_PROB_BASE * rev_prob / (REG_BR_PROB_BASE + rev_prob); | |
1995 | lsw_taken_prob = prob; | |
1996 | } | |
1997 | else | |
1998 | { | |
1999 | msw_taken_prob = prob; | |
2000 | msw_skip_prob = REG_BR_PROB_BASE; | |
2001 | /* ??? If we have a constant op2h, should we use that when | |
2002 | calculating lsw_taken_prob? */ | |
2003 | lsw_taken_prob = prob; | |
2004 | } | |
2005 | } | |
2006 | operands[1] = op1h; | |
2007 | operands[2] = op2h; | |
2008 | operands[4] = NULL_RTX; | |
6f8eec29 | 2009 | if (reload_completed |
99052692 KK |
2010 | && ! arith_reg_or_0_operand (op2h, SImode) |
2011 | && (true_regnum (op1h) || (comparison != EQ && comparison != NE)) | |
b757d36f KK |
2012 | && (msw_taken != LAST_AND_UNUSED_RTX_CODE |
2013 | || msw_skip != LAST_AND_UNUSED_RTX_CODE)) | |
6f8eec29 R |
2014 | { |
2015 | emit_move_insn (scratch, operands[2]); | |
2016 | operands[2] = scratch; | |
2017 | } | |
b757d36f | 2018 | if (msw_taken != LAST_AND_UNUSED_RTX_CODE) |
78d310c2 | 2019 | expand_cbranchsi4 (operands, msw_taken, msw_taken_prob); |
b757d36f | 2020 | if (msw_skip != LAST_AND_UNUSED_RTX_CODE) |
78d310c2 R |
2021 | { |
2022 | rtx taken_label = operands[3]; | |
2023 | ||
d60819f8 CB |
2024 | /* Operands were possibly modified, but msw_skip doesn't expect this. |
2025 | Always use the original ones. */ | |
b757d36f | 2026 | if (msw_taken != LAST_AND_UNUSED_RTX_CODE) |
d60819f8 CB |
2027 | { |
2028 | operands[1] = op1h; | |
2029 | operands[2] = op2h; | |
2030 | } | |
2031 | ||
78d310c2 R |
2032 | operands[3] = skip_label = gen_label_rtx (); |
2033 | expand_cbranchsi4 (operands, msw_skip, msw_skip_prob); | |
2034 | operands[3] = taken_label; | |
2035 | } | |
2036 | operands[1] = op1l; | |
2037 | operands[2] = op2l; | |
b757d36f | 2038 | if (lsw_taken != LAST_AND_UNUSED_RTX_CODE) |
6f8eec29 R |
2039 | { |
2040 | if (reload_completed | |
99052692 KK |
2041 | && ! arith_reg_or_0_operand (op2l, SImode) |
2042 | && (true_regnum (op1l) || (lsw_taken != EQ && lsw_taken != NE))) | |
2043 | { | |
2044 | emit_move_insn (scratch, operands[2]); | |
2045 | operands[2] = scratch; | |
2046 | } | |
6f8eec29 R |
2047 | expand_cbranchsi4 (operands, lsw_taken, lsw_taken_prob); |
2048 | } | |
b757d36f | 2049 | if (msw_skip != LAST_AND_UNUSED_RTX_CODE) |
78d310c2 R |
2050 | emit_label (skip_label); |
2051 | return true; | |
2052 | } | |
2053 | ||
f90b7a5a PB |
2054 | /* Emit INSN, possibly in a PARALLEL with an USE of fpscr for SH4. */ |
2055 | ||
2056 | static void | |
2057 | sh_emit_set_t_insn (rtx insn, enum machine_mode mode) | |
2058 | { | |
2059 | if ((TARGET_SH4 || TARGET_SH2A) && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
2060 | { | |
2061 | insn = gen_rtx_PARALLEL (VOIDmode, | |
2062 | gen_rtvec (2, insn, | |
2063 | gen_rtx_USE (VOIDmode, get_fpscr_rtx ()))); | |
2064 | (mode == SFmode ? emit_sf_insn : emit_df_insn) (insn); | |
2065 | } | |
2066 | else | |
2067 | emit_insn (insn); | |
2068 | } | |
2069 | ||
8aa2a305 | 2070 | /* Prepare the operands for an scc instruction; make sure that the |
f90b7a5a PB |
2071 | compare has been done and the result is in T_REG. */ |
2072 | void | |
2073 | sh_emit_scc_to_t (enum rtx_code code, rtx op0, rtx op1) | |
8aa2a305 | 2074 | { |
c5c76735 | 2075 | rtx t_reg = gen_rtx_REG (SImode, T_REG); |
8aa2a305 | 2076 | enum rtx_code oldcode = code; |
9374bd85 | 2077 | enum machine_mode mode; |
8aa2a305 JW |
2078 | |
2079 | /* First need a compare insn. */ | |
2080 | switch (code) | |
bc45ade3 | 2081 | { |
8aa2a305 JW |
2082 | case NE: |
2083 | /* It isn't possible to handle this case. */ | |
f5b9e7c9 | 2084 | gcc_unreachable (); |
8aa2a305 JW |
2085 | case LT: |
2086 | code = GT; | |
2087 | break; | |
2088 | case LE: | |
2089 | code = GE; | |
2090 | break; | |
2091 | case LTU: | |
2092 | code = GTU; | |
2093 | break; | |
2094 | case LEU: | |
2095 | code = GEU; | |
2096 | break; | |
10db12e0 R |
2097 | default: |
2098 | break; | |
bc45ade3 | 2099 | } |
8aa2a305 | 2100 | if (code != oldcode) |
b9654711 | 2101 | { |
f90b7a5a PB |
2102 | rtx tmp = op0; |
2103 | op0 = op1; | |
2104 | op1 = tmp; | |
0d7e008e | 2105 | } |
b9654711 | 2106 | |
f90b7a5a | 2107 | mode = GET_MODE (op0); |
9374bd85 | 2108 | if (mode == VOIDmode) |
f90b7a5a | 2109 | mode = GET_MODE (op1); |
9374bd85 | 2110 | |
f90b7a5a | 2111 | op0 = force_reg (mode, op0); |
1245df60 | 2112 | if ((code != EQ && code != NE |
f90b7a5a | 2113 | && (op1 != const0_rtx |
1245df60 | 2114 | || code == GTU || code == GEU || code == LTU || code == LEU)) |
f90b7a5a | 2115 | || (mode == DImode && op1 != const0_rtx) |
3a8699c7 | 2116 | || (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT)) |
f90b7a5a | 2117 | op1 = force_reg (mode, op1); |
0d7e008e | 2118 | |
f90b7a5a PB |
2119 | sh_emit_set_t_insn (gen_rtx_SET (VOIDmode, t_reg, |
2120 | gen_rtx_fmt_ee (code, SImode, op0, op1)), | |
2121 | mode); | |
2122 | } | |
0d7e008e | 2123 | |
f90b7a5a PB |
2124 | rtx |
2125 | sh_emit_cheap_store_flag (enum machine_mode mode, enum rtx_code code, | |
2126 | rtx op0, rtx op1) | |
2127 | { | |
2128 | rtx target = gen_reg_rtx (SImode); | |
2129 | rtx tmp; | |
2130 | ||
2131 | gcc_assert (TARGET_SHMEDIA); | |
2132 | switch (code) | |
2133 | { | |
2134 | case EQ: | |
2135 | case GT: | |
2136 | case LT: | |
2137 | case UNORDERED: | |
2138 | case GTU: | |
2139 | case LTU: | |
2140 | tmp = gen_rtx_fmt_ee (code, SImode, op0, op1); | |
2141 | emit_insn (gen_cstore4_media (target, tmp, op0, op1)); | |
2142 | code = NE; | |
2143 | break; | |
2144 | ||
2145 | case NE: | |
2146 | case GE: | |
2147 | case LE: | |
2148 | case ORDERED: | |
2149 | case GEU: | |
2150 | case LEU: | |
2151 | tmp = gen_rtx_fmt_ee (reverse_condition (code), mode, op0, op1); | |
2152 | emit_insn (gen_cstore4_media (target, tmp, op0, op1)); | |
2153 | code = EQ; | |
2154 | break; | |
2155 | ||
2156 | case UNEQ: | |
2157 | case UNGE: | |
2158 | case UNGT: | |
2159 | case UNLE: | |
2160 | case UNLT: | |
2161 | case LTGT: | |
2162 | return NULL_RTX; | |
2163 | ||
2164 | default: | |
2165 | gcc_unreachable (); | |
2166 | } | |
2167 | ||
2168 | if (mode == DImode) | |
2169 | { | |
2170 | rtx t2 = gen_reg_rtx (DImode); | |
2171 | emit_insn (gen_extendsidi2 (t2, target)); | |
2172 | target = t2; | |
2173 | } | |
2174 | ||
2175 | return gen_rtx_fmt_ee (code, VOIDmode, target, const0_rtx); | |
bc45ade3 SC |
2176 | } |
2177 | ||
8aa2a305 JW |
2178 | /* Called from the md file, set up the operands of a compare instruction. */ |
2179 | ||
2180 | void | |
f90b7a5a | 2181 | sh_emit_compare_and_branch (rtx *operands, enum machine_mode mode) |
bc45ade3 | 2182 | { |
f90b7a5a PB |
2183 | enum rtx_code code = GET_CODE (operands[0]); |
2184 | enum rtx_code branch_code; | |
2185 | rtx op0 = operands[1]; | |
2186 | rtx op1 = operands[2]; | |
2187 | rtx insn, tem; | |
2188 | bool need_ccmpeq = false; | |
2189 | ||
2190 | if (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
bc45ade3 | 2191 | { |
f90b7a5a PB |
2192 | op0 = force_reg (mode, op0); |
2193 | op1 = force_reg (mode, op1); | |
bc45ade3 | 2194 | } |
f90b7a5a | 2195 | else |
1245df60 | 2196 | { |
f90b7a5a PB |
2197 | if (code != EQ || mode == DImode) |
2198 | { | |
2199 | /* Force args into regs, since we can't use constants here. */ | |
2200 | op0 = force_reg (mode, op0); | |
2201 | if (op1 != const0_rtx || code == GTU || code == GEU) | |
2202 | op1 = force_reg (mode, op1); | |
2203 | } | |
1245df60 | 2204 | } |
f90b7a5a PB |
2205 | |
2206 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
2207 | { | |
2208 | if (code == LT | |
2209 | || (code == LE && TARGET_IEEE && TARGET_SH2E) | |
2210 | || (code == GE && !(TARGET_IEEE && TARGET_SH2E))) | |
2211 | { | |
2212 | tem = op0, op0 = op1, op1 = tem; | |
2213 | code = swap_condition (code); | |
2214 | } | |
2215 | ||
2216 | /* GE becomes fcmp/gt+fcmp/eq, for SH2E and TARGET_IEEE only. */ | |
2217 | if (code == GE) | |
2218 | { | |
2219 | gcc_assert (TARGET_IEEE && TARGET_SH2E); | |
2220 | need_ccmpeq = true; | |
2221 | code = GT; | |
2222 | } | |
2223 | ||
2224 | /* Now we can have EQ, NE, GT, LE. NE and LE are then transformed | |
2225 | to EQ/GT respectively. */ | |
2226 | gcc_assert (code == EQ || code == GT || code == NE || code == LE); | |
2227 | } | |
2228 | ||
2229 | switch (code) | |
2230 | { | |
2231 | case EQ: | |
2232 | case GT: | |
2233 | case GE: | |
2234 | case GTU: | |
2235 | case GEU: | |
2236 | branch_code = code; | |
2237 | break; | |
2238 | case NE: | |
2239 | case LT: | |
2240 | case LE: | |
2241 | case LTU: | |
2242 | case LEU: | |
2243 | branch_code = reverse_condition (code); | |
2244 | break; | |
2245 | default: | |
2246 | gcc_unreachable (); | |
2247 | } | |
2248 | ||
2249 | insn = gen_rtx_SET (VOIDmode, | |
2250 | gen_rtx_REG (SImode, T_REG), | |
2251 | gen_rtx_fmt_ee (branch_code, SImode, op0, op1)); | |
2252 | ||
2253 | sh_emit_set_t_insn (insn, mode); | |
2254 | if (need_ccmpeq) | |
2255 | sh_emit_set_t_insn (gen_ieee_ccmpeqsf_t (op0, op1), mode); | |
2256 | ||
2257 | if (branch_code == code) | |
2258 | emit_jump_insn (gen_branch_true (operands[3])); | |
1245df60 | 2259 | else |
f90b7a5a PB |
2260 | emit_jump_insn (gen_branch_false (operands[3])); |
2261 | } | |
2262 | ||
2263 | void | |
2264 | sh_emit_compare_and_set (rtx *operands, enum machine_mode mode) | |
2265 | { | |
2266 | enum rtx_code code = GET_CODE (operands[1]); | |
2267 | rtx op0 = operands[2]; | |
2268 | rtx op1 = operands[3]; | |
2269 | rtx lab = NULL_RTX; | |
2270 | bool invert = false; | |
2271 | rtx tem; | |
2272 | ||
2273 | op0 = force_reg (mode, op0); | |
2274 | if ((code != EQ && code != NE | |
2275 | && (op1 != const0_rtx | |
2276 | || code == GTU || code == GEU || code == LTU || code == LEU)) | |
2277 | || (mode == DImode && op1 != const0_rtx) | |
2278 | || (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT)) | |
2279 | op1 = force_reg (mode, op1); | |
2280 | ||
2281 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
225e4f43 | 2282 | { |
f90b7a5a PB |
2283 | if (code == LT || code == LE) |
2284 | { | |
2285 | code = swap_condition (code); | |
2286 | tem = op0, op0 = op1, op1 = tem; | |
2287 | } | |
2288 | if (code == GE) | |
2289 | { | |
2290 | if (TARGET_IEEE) | |
2291 | { | |
2292 | lab = gen_label_rtx (); | |
2293 | sh_emit_scc_to_t (EQ, op0, op1); | |
2294 | emit_jump_insn (gen_branch_true (lab)); | |
2295 | code = GT; | |
2296 | } | |
2297 | else | |
2298 | { | |
2299 | code = LT; | |
2300 | invert = true; | |
2301 | } | |
2302 | } | |
225e4f43 | 2303 | } |
f90b7a5a PB |
2304 | |
2305 | if (code == NE) | |
2306 | { | |
2307 | code = EQ; | |
2308 | invert = true; | |
2309 | } | |
2310 | ||
2311 | sh_emit_scc_to_t (code, op0, op1); | |
2312 | if (lab) | |
2313 | emit_label (lab); | |
2314 | if (invert) | |
2315 | emit_insn (gen_movnegt (operands[0])); | |
225e4f43 | 2316 | else |
f90b7a5a | 2317 | emit_move_insn (operands[0], gen_rtx_REG (SImode, T_REG)); |
bc45ade3 SC |
2318 | } |
2319 | \f | |
16bea517 | 2320 | /* Functions to output assembly code. */ |
bc45ade3 | 2321 | |
b9654711 | 2322 | /* Return a sequence of instructions to perform DI or DF move. |
bc45ade3 | 2323 | |
b9654711 | 2324 | Since the SH cannot move a DI or DF in one instruction, we have |
16bea517 | 2325 | to take care when we see overlapping source and dest registers. */ |
0d7e008e | 2326 | |
318881c0 | 2327 | const char * |
cf277499 SB |
2328 | output_movedouble (rtx insn ATTRIBUTE_UNUSED, rtx operands[], |
2329 | enum machine_mode mode) | |
bc45ade3 | 2330 | { |
b9654711 SC |
2331 | rtx dst = operands[0]; |
2332 | rtx src = operands[1]; | |
b9654711 | 2333 | |
f3536097 | 2334 | if (MEM_P (dst) |
16bea517 | 2335 | && GET_CODE (XEXP (dst, 0)) == PRE_DEC) |
8aa2a305 JW |
2336 | return "mov.l %T1,%0\n\tmov.l %1,%0"; |
2337 | ||
b9654711 SC |
2338 | if (register_operand (dst, mode) |
2339 | && register_operand (src, mode)) | |
bc45ade3 | 2340 | { |
b9654711 | 2341 | if (REGNO (src) == MACH_REG) |
00f8ff66 | 2342 | return "sts mach,%S0\n\tsts macl,%R0"; |
bc45ade3 | 2343 | |
8aa2a305 JW |
2344 | /* When mov.d r1,r2 do r2->r3 then r1->r2; |
2345 | when mov.d r1,r0 do r1->r0 then r2->r1. */ | |
b9654711 SC |
2346 | |
2347 | if (REGNO (src) + 1 == REGNO (dst)) | |
5325c0fa | 2348 | return "mov %T1,%T0\n\tmov %1,%0"; |
b9654711 | 2349 | else |
5325c0fa | 2350 | return "mov %1,%0\n\tmov %T1,%T0"; |
b9654711 | 2351 | } |
f3536097 | 2352 | else if (CONST_INT_P (src)) |
bc45ade3 | 2353 | { |
8aa2a305 JW |
2354 | if (INTVAL (src) < 0) |
2355 | output_asm_insn ("mov #-1,%S0", operands); | |
bc45ade3 | 2356 | else |
8aa2a305 | 2357 | output_asm_insn ("mov #0,%S0", operands); |
bc45ade3 | 2358 | |
8aa2a305 | 2359 | return "mov %1,%R0"; |
0d7e008e | 2360 | } |
f3536097 | 2361 | else if (MEM_P (src)) |
bc45ade3 | 2362 | { |
8aa2a305 | 2363 | int ptrreg = -1; |
b9654711 SC |
2364 | int dreg = REGNO (dst); |
2365 | rtx inside = XEXP (src, 0); | |
bc45ade3 | 2366 | |
f5b9e7c9 | 2367 | switch (GET_CODE (inside)) |
bc45ade3 | 2368 | { |
f5b9e7c9 NS |
2369 | case REG: |
2370 | ptrreg = REGNO (inside); | |
2371 | break; | |
2372 | ||
2373 | case SUBREG: | |
2374 | ptrreg = subreg_regno (inside); | |
2375 | break; | |
2376 | ||
2377 | case PLUS: | |
8aa2a305 JW |
2378 | ptrreg = REGNO (XEXP (inside, 0)); |
2379 | /* ??? A r0+REG address shouldn't be possible here, because it isn't | |
2380 | an offsettable address. Unfortunately, offsettable addresses use | |
2381 | QImode to check the offset, and a QImode offsettable address | |
2382 | requires r0 for the other operand, which is not currently | |
2383 | supported, so we can't use the 'o' constraint. | |
2384 | Thus we must check for and handle r0+REG addresses here. | |
2385 | We punt for now, since this is likely very rare. */ | |
f3536097 | 2386 | gcc_assert (!REG_P (XEXP (inside, 1))); |
f5b9e7c9 NS |
2387 | break; |
2388 | ||
2389 | case LABEL_REF: | |
2390 | return "mov.l %1,%0\n\tmov.l %1+4,%T0"; | |
2391 | case POST_INC: | |
2392 | return "mov.l %1,%0\n\tmov.l %1,%T0"; | |
2393 | default: | |
2394 | gcc_unreachable (); | |
bc45ade3 | 2395 | } |
bc45ade3 | 2396 | |
8aa2a305 JW |
2397 | /* Work out the safe way to copy. Copy into the second half first. */ |
2398 | if (dreg == ptrreg) | |
2399 | return "mov.l %T1,%T0\n\tmov.l %1,%0"; | |
bc45ade3 SC |
2400 | } |
2401 | ||
00f8ff66 | 2402 | return "mov.l %1,%0\n\tmov.l %T1,%T0"; |
bc45ade3 SC |
2403 | } |
2404 | ||
8aa2a305 JW |
2405 | /* Print an instruction which would have gone into a delay slot after |
2406 | another instruction, but couldn't because the other instruction expanded | |
2407 | into a sequence where putting the slot insn at the end wouldn't work. */ | |
0d7e008e | 2408 | |
8aa2a305 | 2409 | static void |
cf277499 | 2410 | print_slot (rtx insn) |
8aa2a305 | 2411 | { |
c9d691e9 | 2412 | final_scan_insn (XVECEXP (insn, 0, 1), asm_out_file, optimize, 1, NULL); |
b9654711 | 2413 | |
8aa2a305 JW |
2414 | INSN_DELETED_P (XVECEXP (insn, 0, 1)) = 1; |
2415 | } | |
0d7e008e | 2416 | |
318881c0 | 2417 | const char * |
cf277499 | 2418 | output_far_jump (rtx insn, rtx op) |
0d7e008e | 2419 | { |
0a2aaacc | 2420 | struct { rtx lab, reg, op; } this_jmp; |
950a3816 | 2421 | rtx braf_base_lab = NULL_RTX; |
318881c0 | 2422 | const char *jump; |
1245df60 | 2423 | int far; |
9d98a694 | 2424 | int offset = branch_dest (insn) - INSN_ADDRESSES (INSN_UID (insn)); |
10f4f635 | 2425 | rtx prev; |
0d7e008e | 2426 | |
0a2aaacc | 2427 | this_jmp.lab = gen_label_rtx (); |
b9654711 | 2428 | |
1a953d0f R |
2429 | if (TARGET_SH2 |
2430 | && offset >= -32764 | |
2431 | && offset - get_attr_length (insn) <= 32766) | |
1245df60 R |
2432 | { |
2433 | far = 0; | |
e6dfd05f | 2434 | jump = "mov.w %O0,%1; braf %1"; |
1245df60 R |
2435 | } |
2436 | else | |
2437 | { | |
2438 | far = 1; | |
1a66cd67 | 2439 | if (flag_pic) |
e6dfd05f AO |
2440 | { |
2441 | if (TARGET_SH2) | |
2442 | jump = "mov.l %O0,%1; braf %1"; | |
2443 | else | |
2444 | jump = "mov.l r0,@-r15; mova %O0,r0; mov.l @r0,%1; add r0,%1; mov.l @r15+,r0; jmp @%1"; | |
2445 | } | |
1a66cd67 | 2446 | else |
e6dfd05f | 2447 | jump = "mov.l %O0,%1; jmp @%1"; |
1245df60 R |
2448 | } |
2449 | /* If we have a scratch register available, use it. */ | |
f3536097 | 2450 | if (NONJUMP_INSN_P ((prev = prev_nonnote_insn (insn))) |
10f4f635 | 2451 | && INSN_CODE (prev) == CODE_FOR_indirect_jump_scratch) |
1245df60 | 2452 | { |
0a2aaacc KG |
2453 | this_jmp.reg = SET_DEST (XVECEXP (PATTERN (prev), 0, 0)); |
2454 | if (REGNO (this_jmp.reg) == R0_REG && flag_pic && ! TARGET_SH2) | |
e6dfd05f | 2455 | jump = "mov.l r1,@-r15; mova %O0,r0; mov.l @r0,r1; add r1,r0; mov.l @r15+,r1; jmp @%1"; |
0a2aaacc | 2456 | output_asm_insn (jump, &this_jmp.lab); |
1245df60 R |
2457 | if (dbr_sequence_length ()) |
2458 | print_slot (final_sequence); | |
2459 | else | |
2460 | output_asm_insn ("nop", 0); | |
2461 | } | |
2462 | else | |
2463 | { | |
2464 | /* Output the delay slot insn first if any. */ | |
2465 | if (dbr_sequence_length ()) | |
2466 | print_slot (final_sequence); | |
2467 | ||
0a2aaacc | 2468 | this_jmp.reg = gen_rtx_REG (SImode, 13); |
fa5322fa AO |
2469 | /* We must keep the stack aligned to 8-byte boundaries on SH5. |
2470 | Fortunately, MACL is fixed and call-clobbered, and we never | |
2471 | need its value across jumps, so save r13 in it instead of in | |
2472 | the stack. */ | |
2473 | if (TARGET_SH5) | |
2474 | output_asm_insn ("lds r13, macl", 0); | |
2475 | else | |
2476 | output_asm_insn ("mov.l r13,@-r15", 0); | |
0a2aaacc | 2477 | output_asm_insn (jump, &this_jmp.lab); |
fa5322fa AO |
2478 | if (TARGET_SH5) |
2479 | output_asm_insn ("sts macl, r13", 0); | |
2480 | else | |
2481 | output_asm_insn ("mov.l @r15+,r13", 0); | |
1245df60 | 2482 | } |
e6dfd05f AO |
2483 | if (far && flag_pic && TARGET_SH2) |
2484 | { | |
2485 | braf_base_lab = gen_label_rtx (); | |
4977bab6 | 2486 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
e6dfd05f AO |
2487 | CODE_LABEL_NUMBER (braf_base_lab)); |
2488 | } | |
1245df60 R |
2489 | if (far) |
2490 | output_asm_insn (".align 2", 0); | |
0a2aaacc KG |
2491 | (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (this_jmp.lab)); |
2492 | this_jmp.op = op; | |
1a66cd67 | 2493 | if (far && flag_pic) |
e6dfd05f AO |
2494 | { |
2495 | if (TARGET_SH2) | |
0a2aaacc KG |
2496 | this_jmp.lab = braf_base_lab; |
2497 | output_asm_insn (".long %O2-%O0", &this_jmp.lab); | |
e6dfd05f | 2498 | } |
1a66cd67 | 2499 | else |
0a2aaacc | 2500 | output_asm_insn (far ? ".long %O2" : ".word %O2-%O0", &this_jmp.lab); |
0d7e008e SC |
2501 | return ""; |
2502 | } | |
bc45ade3 | 2503 | |
8aa2a305 JW |
2504 | /* Local label counter, used for constants in the pool and inside |
2505 | pattern branches. */ | |
2506 | ||
2507 | static int lf = 100; | |
2508 | ||
2509 | /* Output code for ordinary branches. */ | |
2510 | ||
318881c0 | 2511 | const char * |
cf277499 | 2512 | output_branch (int logic, rtx insn, rtx *operands) |
bc45ade3 | 2513 | { |
afbc5d86 | 2514 | switch (get_attr_length (insn)) |
bc45ade3 | 2515 | { |
33f7f353 JR |
2516 | case 6: |
2517 | /* This can happen if filling the delay slot has caused a forward | |
2518 | branch to exceed its range (we could reverse it, but only | |
2519 | when we know we won't overextend other branches; this should | |
2520 | best be handled by relaxation). | |
2521 | It can also happen when other condbranches hoist delay slot insn | |
1245df60 R |
2522 | from their destination, thus leading to code size increase. |
2523 | But the branch will still be in the range -4092..+4098 bytes. */ | |
bc45ade3 | 2524 | |
33f7f353 | 2525 | if (! TARGET_RELAX) |
1245df60 | 2526 | { |
33f7f353 JR |
2527 | int label = lf++; |
2528 | /* The call to print_slot will clobber the operands. */ | |
2529 | rtx op0 = operands[0]; | |
f676971a | 2530 | |
33f7f353 JR |
2531 | /* If the instruction in the delay slot is annulled (true), then |
2532 | there is no delay slot where we can put it now. The only safe | |
2533 | place for it is after the label. final will do that by default. */ | |
f676971a | 2534 | |
33f7f353 | 2535 | if (final_sequence |
bee9efbc KK |
2536 | && ! INSN_ANNULLED_BRANCH_P (XVECEXP (final_sequence, 0, 0)) |
2537 | && get_attr_length (XVECEXP (final_sequence, 0, 1))) | |
33f7f353 | 2538 | { |
afbc5d86 | 2539 | asm_fprintf (asm_out_file, "\tb%s%ss\t%LLF%d\n", logic ? "f" : "t", |
33f7f353 JR |
2540 | ASSEMBLER_DIALECT ? "/" : ".", label); |
2541 | print_slot (final_sequence); | |
2542 | } | |
2543 | else | |
afbc5d86 | 2544 | asm_fprintf (asm_out_file, "\tb%s\t%LLF%d\n", logic ? "f" : "t", label); |
f676971a | 2545 | |
afbc5d86 R |
2546 | output_asm_insn ("bra\t%l0", &op0); |
2547 | fprintf (asm_out_file, "\tnop\n"); | |
832a3292 | 2548 | (*targetm.asm_out.internal_label) (asm_out_file, "LF", label); |
f676971a | 2549 | |
33f7f353 | 2550 | return ""; |
1245df60 | 2551 | } |
33f7f353 JR |
2552 | /* When relaxing, handle this like a short branch. The linker |
2553 | will fix it up if it still doesn't fit after relaxation. */ | |
2554 | case 2: | |
2555 | return logic ? "bt%.\t%l0" : "bf%.\t%l0"; | |
3a8699c7 AO |
2556 | |
2557 | /* These are for SH2e, in which we have to account for the | |
2558 | extra nop because of the hardware bug in annulled branches. */ | |
2559 | case 8: | |
2560 | if (! TARGET_RELAX) | |
2561 | { | |
2562 | int label = lf++; | |
2563 | ||
f5b9e7c9 NS |
2564 | gcc_assert (!final_sequence |
2565 | || !(INSN_ANNULLED_BRANCH_P | |
2566 | (XVECEXP (final_sequence, 0, 0)))); | |
3a8699c7 AO |
2567 | asm_fprintf (asm_out_file, "b%s%ss\t%LLF%d\n", |
2568 | logic ? "f" : "t", | |
2569 | ASSEMBLER_DIALECT ? "/" : ".", label); | |
2570 | fprintf (asm_out_file, "\tnop\n"); | |
2571 | output_asm_insn ("bra\t%l0", operands); | |
2572 | fprintf (asm_out_file, "\tnop\n"); | |
2573 | (*targetm.asm_out.internal_label) (asm_out_file, "LF", label); | |
2574 | ||
2575 | return ""; | |
2576 | } | |
2577 | /* When relaxing, fall through. */ | |
2578 | case 4: | |
2579 | { | |
2580 | char buffer[10]; | |
f676971a | 2581 | |
3a8699c7 AO |
2582 | sprintf (buffer, "b%s%ss\t%%l0", |
2583 | logic ? "t" : "f", | |
2584 | ASSEMBLER_DIALECT ? "/" : "."); | |
2585 | output_asm_insn (buffer, &operands[0]); | |
2586 | return "nop"; | |
2587 | } | |
2588 | ||
33f7f353 | 2589 | default: |
afbc5d86 R |
2590 | /* There should be no longer branches now - that would |
2591 | indicate that something has destroyed the branches set | |
2592 | up in machine_dependent_reorg. */ | |
f5b9e7c9 | 2593 | gcc_unreachable (); |
1245df60 | 2594 | } |
1245df60 | 2595 | } |
bc45ade3 | 2596 | |
0a2aaacc | 2597 | /* Output a code sequence for INSN using TEMPL with OPERANDS; but before, |
78d310c2 R |
2598 | fill in operands 9 as a label to the successor insn. |
2599 | We try to use jump threading where possible. | |
2600 | IF CODE matches the comparison in the IF_THEN_ELSE of a following jump, | |
2601 | we assume the jump is taken. I.e. EQ means follow jmp and bf, NE means | |
2602 | follow jmp and bt, if the address is in range. */ | |
318881c0 | 2603 | const char * |
0a2aaacc | 2604 | output_branchy_insn (enum rtx_code code, const char *templ, |
cf277499 | 2605 | rtx insn, rtx *operands) |
1245df60 R |
2606 | { |
2607 | rtx next_insn = NEXT_INSN (insn); | |
79b2746a | 2608 | |
f3536097 | 2609 | if (next_insn && JUMP_P (next_insn) && condjump_p (next_insn)) |
1245df60 R |
2610 | { |
2611 | rtx src = SET_SRC (PATTERN (next_insn)); | |
2612 | if (GET_CODE (src) == IF_THEN_ELSE && GET_CODE (XEXP (src, 0)) != code) | |
2613 | { | |
2614 | /* Following branch not taken */ | |
2615 | operands[9] = gen_label_rtx (); | |
2616 | emit_label_after (operands[9], next_insn); | |
9d98a694 AO |
2617 | INSN_ADDRESSES_NEW (operands[9], |
2618 | INSN_ADDRESSES (INSN_UID (next_insn)) | |
2619 | + get_attr_length (next_insn)); | |
0a2aaacc | 2620 | return templ; |
1245df60 R |
2621 | } |
2622 | else | |
2623 | { | |
33f7f353 | 2624 | int offset = (branch_dest (next_insn) |
9d98a694 | 2625 | - INSN_ADDRESSES (INSN_UID (next_insn)) + 4); |
33f7f353 | 2626 | if (offset >= -252 && offset <= 258) |
1245df60 R |
2627 | { |
2628 | if (GET_CODE (src) == IF_THEN_ELSE) | |
2629 | /* branch_true */ | |
2630 | src = XEXP (src, 1); | |
2631 | operands[9] = src; | |
0a2aaacc | 2632 | return templ; |
1245df60 R |
2633 | } |
2634 | } | |
bc45ade3 | 2635 | } |
1245df60 R |
2636 | operands[9] = gen_label_rtx (); |
2637 | emit_label_after (operands[9], insn); | |
9d98a694 AO |
2638 | INSN_ADDRESSES_NEW (operands[9], |
2639 | INSN_ADDRESSES (INSN_UID (insn)) | |
2640 | + get_attr_length (insn)); | |
0a2aaacc | 2641 | return templ; |
1245df60 | 2642 | } |
e4fa6b06 | 2643 | |
318881c0 | 2644 | const char * |
cf277499 | 2645 | output_ieee_ccmpeq (rtx insn, rtx *operands) |
1245df60 | 2646 | { |
9fd0c419 R |
2647 | return output_branchy_insn (NE, "bt\t%l9\n\tfcmp/eq\t%1,%0", |
2648 | insn, operands); | |
bc45ade3 | 2649 | } |
8e87e161 | 2650 | \f |
1bc7c5b6 | 2651 | /* Output the start of the assembler file. */ |
8aa2a305 | 2652 | |
1bc7c5b6 | 2653 | static void |
cf277499 | 2654 | sh_file_start (void) |
bc45ade3 | 2655 | { |
1bc7c5b6 ZW |
2656 | default_file_start (); |
2657 | ||
b55e3aad NC |
2658 | #ifdef SYMBIAN |
2659 | /* Declare the .directive section before it is used. */ | |
2660 | fputs ("\t.section .directive, \"SM\", @progbits, 1\n", asm_out_file); | |
2661 | fputs ("\t.asciz \"#<SYMEDIT>#\\n\"\n", asm_out_file); | |
2662 | #endif | |
f676971a | 2663 | |
1bc7c5b6 ZW |
2664 | if (TARGET_ELF) |
2665 | /* We need to show the text section with the proper | |
2666 | attributes as in TEXT_SECTION_ASM_OP, before dwarf2out | |
0fe7abcc | 2667 | emits it without attributes in TEXT_SECTION_ASM_OP, else GAS |
1bc7c5b6 ZW |
2668 | will complain. We can teach GAS specifically about the |
2669 | default attributes for our choice of text section, but | |
2670 | then we would have to change GAS again if/when we change | |
2671 | the text section name. */ | |
2672 | fprintf (asm_out_file, "%s\n", TEXT_SECTION_ASM_OP); | |
2673 | else | |
2674 | /* Switch to the data section so that the coffsem symbol | |
2675 | isn't in the text section. */ | |
d6b5193b | 2676 | switch_to_section (data_section); |
b9654711 | 2677 | |
8aa2a305 | 2678 | if (TARGET_LITTLE_ENDIAN) |
1bc7c5b6 | 2679 | fputs ("\t.little\n", asm_out_file); |
fa5322fa | 2680 | |
1bc7c5b6 ZW |
2681 | if (!TARGET_ELF) |
2682 | { | |
2683 | if (TARGET_SHCOMPACT) | |
2684 | fputs ("\t.mode\tSHcompact\n", asm_out_file); | |
2685 | else if (TARGET_SHMEDIA) | |
2686 | fprintf (asm_out_file, "\t.mode\tSHmedia\n\t.abi\t%i\n", | |
2687 | TARGET_SHMEDIA64 ? 64 : 32); | |
2688 | } | |
bc45ade3 | 2689 | } |
0d7e008e | 2690 | \f |
9f3a9a08 KK |
2691 | /* Check if PAT includes UNSPEC_CALLER unspec pattern. */ |
2692 | ||
2693 | static bool | |
cf277499 | 2694 | unspec_caller_rtx_p (rtx pat) |
9f3a9a08 | 2695 | { |
dc3ba671 RS |
2696 | rtx base, offset; |
2697 | int i; | |
2698 | ||
2699 | split_const (pat, &base, &offset); | |
2700 | if (GET_CODE (base) == UNSPEC) | |
9f3a9a08 | 2701 | { |
dc3ba671 | 2702 | if (XINT (base, 1) == UNSPEC_CALLER) |
9f3a9a08 | 2703 | return true; |
dc3ba671 RS |
2704 | for (i = 0; i < XVECLEN (base, 0); i++) |
2705 | if (unspec_caller_rtx_p (XVECEXP (base, 0, i))) | |
2706 | return true; | |
9f3a9a08 | 2707 | } |
9f3a9a08 KK |
2708 | return false; |
2709 | } | |
2710 | ||
2711 | /* Indicate that INSN cannot be duplicated. This is true for insn | |
569b7f6a | 2712 | that generates a unique label. */ |
9f3a9a08 KK |
2713 | |
2714 | static bool | |
cf277499 | 2715 | sh_cannot_copy_insn_p (rtx insn) |
9f3a9a08 KK |
2716 | { |
2717 | rtx pat; | |
2718 | ||
2719 | if (!reload_completed || !flag_pic) | |
2720 | return false; | |
2721 | ||
f3536097 | 2722 | if (!NONJUMP_INSN_P (insn)) |
9f3a9a08 KK |
2723 | return false; |
2724 | if (asm_noperands (insn) >= 0) | |
2725 | return false; | |
2726 | ||
2727 | pat = PATTERN (insn); | |
2728 | if (GET_CODE (pat) != SET) | |
2729 | return false; | |
2730 | pat = SET_SRC (pat); | |
2731 | ||
2732 | if (unspec_caller_rtx_p (pat)) | |
2733 | return true; | |
2734 | ||
2735 | return false; | |
2736 | } | |
2737 | \f | |
16bea517 | 2738 | /* Actual number of instructions used to make a shift by N. */ |
0b5826ac | 2739 | static const char ashiftrt_insns[] = |
16bea517 JW |
2740 | { 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}; |
2741 | ||
2742 | /* Left shift and logical right shift are the same. */ | |
0b5826ac | 2743 | static const char shift_insns[] = |
16bea517 | 2744 | { 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 | 2745 | |
16bea517 JW |
2746 | /* Individual shift amounts needed to get the above length sequences. |
2747 | One bit right shifts clobber the T bit, so when possible, put one bit | |
2748 | shifts in the middle of the sequence, so the ends are eligible for | |
2749 | branch delay slots. */ | |
0139adca | 2750 | static const short shift_amounts[32][5] = { |
16bea517 JW |
2751 | {0}, {1}, {2}, {2, 1}, |
2752 | {2, 2}, {2, 1, 2}, {2, 2, 2}, {2, 2, 1, 2}, | |
2753 | {8}, {8, 1}, {8, 2}, {8, 1, 2}, | |
2754 | {8, 2, 2}, {8, 2, 1, 2}, {8, -2, 8}, {8, -1, 8}, | |
2755 | {16}, {16, 1}, {16, 2}, {16, 1, 2}, | |
2756 | {16, 2, 2}, {16, 2, 1, 2}, {16, -2, 8}, {16, -1, 8}, | |
2757 | {16, 8}, {16, 1, 8}, {16, 8, 2}, {16, 8, 1, 2}, | |
1d3534c0 | 2758 | {16, 8, 2, 2}, {16, -1, -2, 16}, {16, -2, 16}, {16, -1, 16}}; |
16bea517 | 2759 | |
8d481241 JW |
2760 | /* Likewise, but for shift amounts < 16, up to three highmost bits |
2761 | might be clobbered. This is typically used when combined with some | |
2762 | kind of sign or zero extension. */ | |
f676971a | 2763 | |
0b5826ac | 2764 | static const char ext_shift_insns[] = |
8d481241 JW |
2765 | { 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}; |
2766 | ||
0b5826ac | 2767 | static const short ext_shift_amounts[32][4] = { |
8d481241 JW |
2768 | {0}, {1}, {2}, {2, 1}, |
2769 | {2, 2}, {2, 1, 2}, {8, -2}, {8, -1}, | |
2770 | {8}, {8, 1}, {8, 2}, {8, 1, 2}, | |
2771 | {8, 2, 2}, {16, -2, -1}, {16, -2}, {16, -1}, | |
2772 | {16}, {16, 1}, {16, 2}, {16, 1, 2}, | |
2773 | {16, 2, 2}, {16, 2, 1, 2}, {16, -2, 8}, {16, -1, 8}, | |
2774 | {16, 8}, {16, 1, 8}, {16, 8, 2}, {16, 8, 1, 2}, | |
2775 | {16, 8, 2, 2}, {16, -1, -2, 16}, {16, -2, 16}, {16, -1, 16}}; | |
2776 | ||
ae9d19c0 JR |
2777 | /* Assuming we have a value that has been sign-extended by at least one bit, |
2778 | can we use the ext_shift_amounts with the last shift turned to an arithmetic shift | |
2779 | to shift it by N without data loss, and quicker than by other means? */ | |
2780 | #define EXT_SHIFT_SIGNED(n) (((n) | 8) == 15) | |
2781 | ||
16bea517 JW |
2782 | /* This is used in length attributes in sh.md to help compute the length |
2783 | of arbitrary constant shift instructions. */ | |
bc45ade3 | 2784 | |
16bea517 | 2785 | int |
cf277499 | 2786 | shift_insns_rtx (rtx insn) |
16bea517 JW |
2787 | { |
2788 | rtx set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
6c34a092 | 2789 | int shift_count = INTVAL (XEXP (set_src, 1)) & 31; |
16bea517 | 2790 | enum rtx_code shift_code = GET_CODE (set_src); |
00f8ff66 | 2791 | |
16bea517 JW |
2792 | switch (shift_code) |
2793 | { | |
2794 | case ASHIFTRT: | |
2795 | return ashiftrt_insns[shift_count]; | |
2796 | case LSHIFTRT: | |
2797 | case ASHIFT: | |
2798 | return shift_insns[shift_count]; | |
2799 | default: | |
f5b9e7c9 | 2800 | gcc_unreachable (); |
16bea517 JW |
2801 | } |
2802 | } | |
2803 | ||
16bea517 | 2804 | /* Return the cost of a shift. */ |
bc45ade3 | 2805 | |
3c50106f | 2806 | static inline int |
cf277499 | 2807 | shiftcosts (rtx x) |
bc45ade3 | 2808 | { |
3d422cb1 | 2809 | int value; |
8aa2a305 | 2810 | |
fa5322fa AO |
2811 | if (TARGET_SHMEDIA) |
2812 | return 1; | |
2813 | ||
c762ab6e R |
2814 | if (GET_MODE_SIZE (GET_MODE (x)) > UNITS_PER_WORD) |
2815 | { | |
2816 | if (GET_MODE (x) == DImode | |
f3536097 | 2817 | && CONST_INT_P (XEXP (x, 1)) |
c762ab6e R |
2818 | && INTVAL (XEXP (x, 1)) == 1) |
2819 | return 2; | |
2820 | ||
2821 | /* Everything else is invalid, because there is no pattern for it. */ | |
cb2eb96f | 2822 | return MAX_COST; |
c762ab6e | 2823 | } |
16bea517 | 2824 | /* If shift by a non constant, then this will be expensive. */ |
f3536097 | 2825 | if (!CONST_INT_P (XEXP (x, 1))) |
1245df60 | 2826 | return SH_DYNAMIC_SHIFT_COST; |
bc45ade3 | 2827 | |
6c34a092 PB |
2828 | /* Otherwise, return the true cost in instructions. Cope with out of range |
2829 | shift counts more or less arbitrarily. */ | |
2830 | value = INTVAL (XEXP (x, 1)) & 31; | |
3d422cb1 | 2831 | |
8aa2a305 | 2832 | if (GET_CODE (x) == ASHIFTRT) |
49b6d06b JW |
2833 | { |
2834 | int cost = ashiftrt_insns[value]; | |
2835 | /* If SH3, then we put the constant in a reg and use shad. */ | |
1245df60 R |
2836 | if (cost > 1 + SH_DYNAMIC_SHIFT_COST) |
2837 | cost = 1 + SH_DYNAMIC_SHIFT_COST; | |
49b6d06b JW |
2838 | return cost; |
2839 | } | |
8aa2a305 JW |
2840 | else |
2841 | return shift_insns[value]; | |
0d7e008e | 2842 | } |
b9654711 | 2843 | |
8aa2a305 JW |
2844 | /* Return the cost of an AND operation. */ |
2845 | ||
3c50106f | 2846 | static inline int |
cf277499 | 2847 | andcosts (rtx x) |
0d7e008e SC |
2848 | { |
2849 | int i; | |
8aa2a305 | 2850 | |
a6f71af5 | 2851 | /* Anding with a register is a single cycle and instruction. */ |
f3536097 | 2852 | if (!CONST_INT_P (XEXP (x, 1))) |
a6f71af5 | 2853 | return 1; |
8aa2a305 JW |
2854 | |
2855 | i = INTVAL (XEXP (x, 1)); | |
fa5322fa AO |
2856 | |
2857 | if (TARGET_SHMEDIA) | |
2858 | { | |
32a7ab3d KK |
2859 | if (satisfies_constraint_I10 (XEXP (x, 1)) |
2860 | || satisfies_constraint_J16 (XEXP (x, 1))) | |
fa5322fa AO |
2861 | return 1; |
2862 | else | |
f40751dd | 2863 | return 1 + rtx_cost (XEXP (x, 1), AND, !optimize_size); |
fa5322fa AO |
2864 | } |
2865 | ||
a6f71af5 | 2866 | /* These constants are single cycle extu.[bw] instructions. */ |
0d7e008e | 2867 | if (i == 0xff || i == 0xffff) |
a6f71af5 | 2868 | return 1; |
735cb76e | 2869 | /* Constants that can be used in an and immediate instruction in a single |
a6f71af5 | 2870 | cycle, but this requires r0, so make it a little more expensive. */ |
735cb76e | 2871 | if (CONST_OK_FOR_K08 (i)) |
0d7e008e | 2872 | return 2; |
a6f71af5 JW |
2873 | /* Constants that can be loaded with a mov immediate and an and. |
2874 | This case is probably unnecessary. */ | |
735cb76e | 2875 | if (CONST_OK_FOR_I08 (i)) |
a6f71af5 JW |
2876 | return 2; |
2877 | /* Any other constants requires a 2 cycle pc-relative load plus an and. | |
2878 | This case is probably unnecessary. */ | |
2879 | return 3; | |
0d7e008e | 2880 | } |
d3ae8277 | 2881 | |
630c79be BS |
2882 | /* Return the cost of an addition or a subtraction. */ |
2883 | ||
3c50106f | 2884 | static inline int |
cf277499 | 2885 | addsubcosts (rtx x) |
630c79be BS |
2886 | { |
2887 | /* Adding a register is a single cycle insn. */ | |
f3536097 | 2888 | if (REG_P (XEXP (x, 1)) |
50ceefc2 | 2889 | || GET_CODE (XEXP (x, 1)) == SUBREG) |
630c79be BS |
2890 | return 1; |
2891 | ||
2892 | /* Likewise for small constants. */ | |
f3536097 | 2893 | if (CONST_INT_P (XEXP (x, 1)) |
fa5322fa | 2894 | && CONST_OK_FOR_ADD (INTVAL (XEXP (x, 1)))) |
630c79be BS |
2895 | return 1; |
2896 | ||
fa5322fa AO |
2897 | if (TARGET_SHMEDIA) |
2898 | switch (GET_CODE (XEXP (x, 1))) | |
2899 | { | |
2900 | case CONST: | |
2901 | case LABEL_REF: | |
2902 | case SYMBOL_REF: | |
2903 | return TARGET_SHMEDIA64 ? 5 : 3; | |
2904 | ||
2905 | case CONST_INT: | |
735cb76e | 2906 | if (CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1)))) |
fa5322fa | 2907 | return 2; |
735cb76e | 2908 | else if (CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1)) >> 16)) |
fa5322fa | 2909 | return 3; |
735cb76e | 2910 | else if (CONST_OK_FOR_I16 ((INTVAL (XEXP (x, 1)) >> 16) >> 16)) |
fa5322fa AO |
2911 | return 4; |
2912 | ||
2913 | /* Fall through. */ | |
2914 | default: | |
832a3292 | 2915 | return 5; |
fa5322fa AO |
2916 | } |
2917 | ||
630c79be BS |
2918 | /* Any other constant requires a 2 cycle pc-relative load plus an |
2919 | addition. */ | |
2920 | return 3; | |
2921 | } | |
2922 | ||
16bea517 | 2923 | /* Return the cost of a multiply. */ |
3c50106f | 2924 | static inline int |
cf277499 | 2925 | multcosts (rtx x ATTRIBUTE_UNUSED) |
0d7e008e | 2926 | { |
7dd2f19b RS |
2927 | if (sh_multcost >= 0) |
2928 | return sh_multcost; | |
fa5322fa | 2929 | if (TARGET_SHMEDIA) |
73a4d10b R |
2930 | /* ??? We have a mul insn, but it has a latency of three, and doesn't |
2931 | accept constants. Ideally, we would use a cost of one or two and | |
2932 | add the cost of the operand, but disregard the latter when inside loops | |
2933 | and loop invariant code motion is still to follow. | |
2934 | Using a multiply first and splitting it later if it's a loss | |
2935 | doesn't work because of different sign / zero extension semantics | |
2936 | of multiplies vs. shifts. */ | |
2937 | return TARGET_SMALLCODE ? 2 : 3; | |
fa5322fa | 2938 | |
0d7e008e | 2939 | if (TARGET_SH2) |
d3ae8277 SC |
2940 | { |
2941 | /* We have a mul insn, so we can never take more than the mul and the | |
a7771f78 | 2942 | read of the mac reg, but count more because of the latency and extra |
16bea517 | 2943 | reg usage. */ |
d3ae8277 | 2944 | if (TARGET_SMALLCODE) |
8e87e161 | 2945 | return 2; |
a7771f78 | 2946 | return 3; |
d3ae8277 SC |
2947 | } |
2948 | ||
a7771f78 | 2949 | /* If we're aiming at small code, then just count the number of |
16bea517 | 2950 | insns in a multiply call sequence. */ |
8e87e161 | 2951 | if (TARGET_SMALLCODE) |
8aa2a305 | 2952 | return 5; |
d3ae8277 | 2953 | |
16bea517 | 2954 | /* Otherwise count all the insns in the routine we'd be calling too. */ |
d3ae8277 | 2955 | return 20; |
0d7e008e | 2956 | } |
b9654711 | 2957 | |
3c50106f RH |
2958 | /* Compute a (partial) cost for rtx X. Return true if the complete |
2959 | cost has been computed, and false if subexpressions should be | |
2960 | scanned. In either case, *TOTAL contains the cost result. */ | |
2961 | ||
2962 | static bool | |
f40751dd JH |
2963 | sh_rtx_costs (rtx x, int code, int outer_code, int *total, |
2964 | bool speed ATTRIBUTE_UNUSED) | |
3c50106f RH |
2965 | { |
2966 | switch (code) | |
2967 | { | |
2968 | case CONST_INT: | |
2969 | if (TARGET_SHMEDIA) | |
2970 | { | |
2971 | if (INTVAL (x) == 0) | |
2972 | *total = 0; | |
2973 | else if (outer_code == AND && and_operand ((x), DImode)) | |
2974 | *total = 0; | |
2975 | else if ((outer_code == IOR || outer_code == XOR | |
2976 | || outer_code == PLUS) | |
735cb76e | 2977 | && CONST_OK_FOR_I10 (INTVAL (x))) |
3c50106f | 2978 | *total = 0; |
735cb76e | 2979 | else if (CONST_OK_FOR_I16 (INTVAL (x))) |
3c50106f | 2980 | *total = COSTS_N_INSNS (outer_code != SET); |
735cb76e | 2981 | else if (CONST_OK_FOR_I16 (INTVAL (x) >> 16)) |
73a4d10b | 2982 | *total = COSTS_N_INSNS ((outer_code != SET) + 1); |
735cb76e | 2983 | else if (CONST_OK_FOR_I16 ((INTVAL (x) >> 16) >> 16)) |
42201282 | 2984 | *total = COSTS_N_INSNS ((outer_code != SET) + 2); |
3c50106f | 2985 | else |
42201282 | 2986 | *total = COSTS_N_INSNS ((outer_code != SET) + 3); |
3c50106f RH |
2987 | return true; |
2988 | } | |
735cb76e | 2989 | if (CONST_OK_FOR_I08 (INTVAL (x))) |
3c50106f RH |
2990 | *total = 0; |
2991 | else if ((outer_code == AND || outer_code == IOR || outer_code == XOR) | |
735cb76e | 2992 | && CONST_OK_FOR_K08 (INTVAL (x))) |
3c50106f | 2993 | *total = 1; |
78d310c2 | 2994 | /* prepare_cmp_insn will force costly constants int registers before |
35912544 | 2995 | the cbranch[sd]i4 patterns can see them, so preserve potentially |
78d310c2 R |
2996 | interesting ones not covered by I08 above. */ |
2997 | else if (outer_code == COMPARE | |
2998 | && ((unsigned HOST_WIDE_INT) INTVAL (x) | |
2999 | == (unsigned HOST_WIDE_INT) 0x7fffffff + 1 | |
3000 | || INTVAL (x) == 0x7fffffff | |
3001 | || INTVAL (x) == 0x80 || INTVAL (x) == -0x81)) | |
3002 | *total = 1; | |
3c50106f RH |
3003 | else |
3004 | *total = 8; | |
3005 | return true; | |
3006 | ||
3007 | case CONST: | |
3008 | case LABEL_REF: | |
3009 | case SYMBOL_REF: | |
3010 | if (TARGET_SHMEDIA64) | |
3011 | *total = COSTS_N_INSNS (4); | |
3012 | else if (TARGET_SHMEDIA32) | |
3013 | *total = COSTS_N_INSNS (2); | |
3014 | else | |
3015 | *total = 5; | |
3016 | return true; | |
3017 | ||
3018 | case CONST_DOUBLE: | |
3019 | if (TARGET_SHMEDIA) | |
3020 | *total = COSTS_N_INSNS (4); | |
78d310c2 | 3021 | /* prepare_cmp_insn will force costly constants int registers before |
35912544 | 3022 | the cbranchdi4 pattern can see them, so preserve potentially |
78d310c2 R |
3023 | interesting ones. */ |
3024 | else if (outer_code == COMPARE && GET_MODE (x) == DImode) | |
3025 | *total = 1; | |
3c50106f RH |
3026 | else |
3027 | *total = 10; | |
3028 | return true; | |
73a4d10b R |
3029 | case CONST_VECTOR: |
3030 | if (x == CONST0_RTX (GET_MODE (x))) | |
3031 | *total = 0; | |
3032 | else if (sh_1el_vec (x, VOIDmode)) | |
3033 | *total = outer_code != SET; | |
3034 | if (sh_rep_vec (x, VOIDmode)) | |
3035 | *total = ((GET_MODE_UNIT_SIZE (GET_MODE (x)) + 3) / 4 | |
3036 | + (outer_code != SET)); | |
3037 | *total = COSTS_N_INSNS (3) + (outer_code != SET); | |
3038 | return true; | |
3c50106f RH |
3039 | |
3040 | case PLUS: | |
73a4d10b | 3041 | case MINUS: |
3c50106f RH |
3042 | *total = COSTS_N_INSNS (addsubcosts (x)); |
3043 | return true; | |
3044 | ||
3045 | case AND: | |
3046 | *total = COSTS_N_INSNS (andcosts (x)); | |
3047 | return true; | |
3048 | ||
3049 | case MULT: | |
3050 | *total = COSTS_N_INSNS (multcosts (x)); | |
3051 | return true; | |
3052 | ||
3053 | case ASHIFT: | |
3054 | case ASHIFTRT: | |
3055 | case LSHIFTRT: | |
3056 | *total = COSTS_N_INSNS (shiftcosts (x)); | |
3057 | return true; | |
3058 | ||
3059 | case DIV: | |
3060 | case UDIV: | |
3061 | case MOD: | |
3062 | case UMOD: | |
3063 | *total = COSTS_N_INSNS (20); | |
3064 | return true; | |
3065 | ||
73a4d10b R |
3066 | case PARALLEL: |
3067 | if (sh_1el_vec (x, VOIDmode)) | |
3068 | *total = outer_code != SET; | |
3069 | if (sh_rep_vec (x, VOIDmode)) | |
3070 | *total = ((GET_MODE_UNIT_SIZE (GET_MODE (x)) + 3) / 4 | |
3071 | + (outer_code != SET)); | |
3072 | *total = COSTS_N_INSNS (3) + (outer_code != SET); | |
3073 | return true; | |
3074 | ||
3c50106f RH |
3075 | case FLOAT: |
3076 | case FIX: | |
3077 | *total = 100; | |
3078 | return true; | |
3079 | ||
3080 | default: | |
3081 | return false; | |
3082 | } | |
3083 | } | |
3084 | ||
dcefdf67 RH |
3085 | /* Compute the cost of an address. For the SH, all valid addresses are |
3086 | the same cost. Use a slightly higher cost for reg + reg addressing, | |
3087 | since it increases pressure on r0. */ | |
3088 | ||
3089 | static int | |
f40751dd JH |
3090 | sh_address_cost (rtx X, |
3091 | bool speed ATTRIBUTE_UNUSED) | |
dcefdf67 RH |
3092 | { |
3093 | return (GET_CODE (X) == PLUS | |
3094 | && ! CONSTANT_P (XEXP (X, 1)) | |
3095 | && ! TARGET_SHMEDIA ? 1 : 0); | |
3096 | } | |
3c50106f | 3097 | |
16bea517 | 3098 | /* Code to expand a shift. */ |
b9654711 | 3099 | |
0d7e008e | 3100 | void |
cf277499 | 3101 | gen_ashift (int type, int n, rtx reg) |
0d7e008e | 3102 | { |
16bea517 JW |
3103 | /* Negative values here come from the shift_amounts array. */ |
3104 | if (n < 0) | |
3105 | { | |
3106 | if (type == ASHIFT) | |
3107 | type = LSHIFTRT; | |
3108 | else | |
3109 | type = ASHIFT; | |
3110 | n = -n; | |
3111 | } | |
3112 | ||
0d7e008e | 3113 | switch (type) |
bc45ade3 | 3114 | { |
0d7e008e SC |
3115 | case ASHIFTRT: |
3116 | emit_insn (gen_ashrsi3_k (reg, reg, GEN_INT (n))); | |
3117 | break; | |
3118 | case LSHIFTRT: | |
16bea517 JW |
3119 | if (n == 1) |
3120 | emit_insn (gen_lshrsi3_m (reg, reg, GEN_INT (n))); | |
3121 | else | |
3122 | emit_insn (gen_lshrsi3_k (reg, reg, GEN_INT (n))); | |
0d7e008e SC |
3123 | break; |
3124 | case ASHIFT: | |
7e2fda6e | 3125 | emit_insn (gen_ashlsi3_std (reg, reg, GEN_INT (n))); |
0d7e008e | 3126 | break; |
bc45ade3 | 3127 | } |
bc45ade3 | 3128 | } |
bc45ade3 | 3129 | |
8d481241 JW |
3130 | /* Same for HImode */ |
3131 | ||
3132 | void | |
cf277499 | 3133 | gen_ashift_hi (int type, int n, rtx reg) |
8d481241 JW |
3134 | { |
3135 | /* Negative values here come from the shift_amounts array. */ | |
3136 | if (n < 0) | |
3137 | { | |
3138 | if (type == ASHIFT) | |
3139 | type = LSHIFTRT; | |
3140 | else | |
3141 | type = ASHIFT; | |
3142 | n = -n; | |
3143 | } | |
3144 | ||
3145 | switch (type) | |
3146 | { | |
3147 | case ASHIFTRT: | |
8d481241 | 3148 | case LSHIFTRT: |
d0c42859 R |
3149 | /* We don't have HImode right shift operations because using the |
3150 | ordinary 32 bit shift instructions for that doesn't generate proper | |
3151 | zero/sign extension. | |
3152 | gen_ashift_hi is only called in contexts where we know that the | |
3153 | sign extension works out correctly. */ | |
97d6fd65 | 3154 | { |
ddef6bc7 | 3155 | int offset = 0; |
97d6fd65 R |
3156 | if (GET_CODE (reg) == SUBREG) |
3157 | { | |
ddef6bc7 | 3158 | offset = SUBREG_BYTE (reg); |
97d6fd65 R |
3159 | reg = SUBREG_REG (reg); |
3160 | } | |
ddef6bc7 | 3161 | gen_ashift (type, n, gen_rtx_SUBREG (SImode, reg, offset)); |
97d6fd65 R |
3162 | break; |
3163 | } | |
8d481241 JW |
3164 | case ASHIFT: |
3165 | emit_insn (gen_ashlhi3_k (reg, reg, GEN_INT (n))); | |
3166 | break; | |
3167 | } | |
3168 | } | |
3169 | ||
8aa2a305 JW |
3170 | /* Output RTL to split a constant shift into its component SH constant |
3171 | shift instructions. */ | |
f676971a | 3172 | |
318881c0 | 3173 | void |
cf277499 | 3174 | gen_shifty_op (int code, rtx *operands) |
bc45ade3 | 3175 | { |
16bea517 | 3176 | int value = INTVAL (operands[2]); |
8aa2a305 | 3177 | int max, i; |
00f8ff66 | 3178 | |
cff3d762 | 3179 | /* Truncate the shift count in case it is out of bounds. */ |
6c34a092 | 3180 | value = value & 31; |
f676971a | 3181 | |
8aa2a305 | 3182 | if (value == 31) |
16bea517 | 3183 | { |
8aa2a305 | 3184 | if (code == LSHIFTRT) |
0d7e008e | 3185 | { |
8aa2a305 JW |
3186 | emit_insn (gen_rotlsi3_1 (operands[0], operands[0])); |
3187 | emit_insn (gen_movt (operands[0])); | |
3188 | return; | |
16bea517 | 3189 | } |
8aa2a305 | 3190 | else if (code == ASHIFT) |
16bea517 | 3191 | { |
8aa2a305 JW |
3192 | /* There is a two instruction sequence for 31 bit left shifts, |
3193 | but it requires r0. */ | |
f3536097 | 3194 | if (REG_P (operands[0]) && REGNO (operands[0]) == 0) |
0d7e008e | 3195 | { |
8aa2a305 JW |
3196 | emit_insn (gen_andsi3 (operands[0], operands[0], const1_rtx)); |
3197 | emit_insn (gen_rotlsi3_31 (operands[0], operands[0])); | |
3198 | return; | |
0d7e008e | 3199 | } |
16bea517 | 3200 | } |
8aa2a305 | 3201 | } |
c17f53a8 JW |
3202 | else if (value == 0) |
3203 | { | |
73a4d10b R |
3204 | /* This can happen even when optimizing, if there were subregs before |
3205 | reload. Don't output a nop here, as this is never optimized away; | |
3206 | use a no-op move instead. */ | |
3207 | emit_insn (gen_rtx_SET (VOIDmode, operands[0], operands[0])); | |
c17f53a8 JW |
3208 | return; |
3209 | } | |
a9f71ad8 | 3210 | |
8aa2a305 JW |
3211 | max = shift_insns[value]; |
3212 | for (i = 0; i < max; i++) | |
3213 | gen_ashift (code, shift_amounts[value][i], operands[0]); | |
3214 | } | |
f676971a | 3215 | |
8d481241 JW |
3216 | /* Same as above, but optimized for values where the topmost bits don't |
3217 | matter. */ | |
3218 | ||
318881c0 | 3219 | void |
cf277499 | 3220 | gen_shifty_hi_op (int code, rtx *operands) |
8d481241 JW |
3221 | { |
3222 | int value = INTVAL (operands[2]); | |
3223 | int max, i; | |
cf277499 | 3224 | void (*gen_fun) (int, int, rtx); |
8d481241 JW |
3225 | |
3226 | /* This operation is used by and_shl for SImode values with a few | |
3227 | high bits known to be cleared. */ | |
3228 | value &= 31; | |
3229 | if (value == 0) | |
3230 | { | |
3231 | emit_insn (gen_nop ()); | |
3232 | return; | |
3233 | } | |
3234 | ||
3235 | gen_fun = GET_MODE (operands[0]) == HImode ? gen_ashift_hi : gen_ashift; | |
3236 | if (code == ASHIFT) | |
3237 | { | |
3238 | max = ext_shift_insns[value]; | |
3239 | for (i = 0; i < max; i++) | |
3240 | gen_fun (code, ext_shift_amounts[value][i], operands[0]); | |
3241 | } | |
3242 | else | |
3243 | /* When shifting right, emit the shifts in reverse order, so that | |
3244 | solitary negative values come first. */ | |
3245 | for (i = ext_shift_insns[value] - 1; i >= 0; i--) | |
3246 | gen_fun (code, ext_shift_amounts[value][i], operands[0]); | |
3247 | } | |
8aa2a305 JW |
3248 | |
3249 | /* Output RTL for an arithmetic right shift. */ | |
3250 | ||
3251 | /* ??? Rewrite to use super-optimizer sequences. */ | |
3252 | ||
3253 | int | |
cf277499 | 3254 | expand_ashiftrt (rtx *operands) |
8aa2a305 JW |
3255 | { |
3256 | rtx wrk; | |
3257 | char func[18]; | |
8aa2a305 JW |
3258 | int value; |
3259 | ||
49b6d06b | 3260 | if (TARGET_SH3) |
20b04867 | 3261 | { |
f3536097 | 3262 | if (!CONST_INT_P (operands[2])) |
49b6d06b JW |
3263 | { |
3264 | rtx count = copy_to_mode_reg (SImode, operands[2]); | |
3265 | emit_insn (gen_negsi2 (count, count)); | |
3266 | emit_insn (gen_ashrsi3_d (operands[0], operands[1], count)); | |
3267 | return 1; | |
3268 | } | |
1245df60 R |
3269 | else if (ashiftrt_insns[INTVAL (operands[2]) & 31] |
3270 | > 1 + SH_DYNAMIC_SHIFT_COST) | |
49b6d06b | 3271 | { |
1245df60 R |
3272 | rtx count |
3273 | = force_reg (SImode, GEN_INT (- (INTVAL (operands[2]) & 31))); | |
49b6d06b JW |
3274 | emit_insn (gen_ashrsi3_d (operands[0], operands[1], count)); |
3275 | return 1; | |
3276 | } | |
20b04867 | 3277 | } |
f3536097 | 3278 | if (!CONST_INT_P (operands[2])) |
8aa2a305 JW |
3279 | return 0; |
3280 | ||
1245df60 | 3281 | value = INTVAL (operands[2]) & 31; |
8aa2a305 JW |
3282 | |
3283 | if (value == 31) | |
3284 | { | |
73a4d10b R |
3285 | /* If we are called from abs expansion, arrange things so that we |
3286 | we can use a single MT instruction that doesn't clobber the source, | |
3287 | if LICM can hoist out the load of the constant zero. */ | |
3288 | if (currently_expanding_to_rtl) | |
3289 | { | |
3290 | emit_insn (gen_cmpgtsi_t (force_reg (SImode, CONST0_RTX (SImode)), | |
3291 | operands[1])); | |
3292 | emit_insn (gen_mov_neg_si_t (operands[0])); | |
3293 | return 1; | |
3294 | } | |
8aa2a305 JW |
3295 | emit_insn (gen_ashrsi2_31 (operands[0], operands[1])); |
3296 | return 1; | |
3297 | } | |
3298 | else if (value >= 16 && value <= 19) | |
3299 | { | |
3300 | wrk = gen_reg_rtx (SImode); | |
3301 | emit_insn (gen_ashrsi2_16 (wrk, operands[1])); | |
3302 | value -= 16; | |
3303 | while (value--) | |
3304 | gen_ashift (ASHIFTRT, 1, wrk); | |
3305 | emit_move_insn (operands[0], wrk); | |
3306 | return 1; | |
a9f71ad8 | 3307 | } |
8aa2a305 JW |
3308 | /* Expand a short sequence inline, longer call a magic routine. */ |
3309 | else if (value <= 5) | |
3310 | { | |
3311 | wrk = gen_reg_rtx (SImode); | |
3312 | emit_move_insn (wrk, operands[1]); | |
3313 | while (value--) | |
3314 | gen_ashift (ASHIFTRT, 1, wrk); | |
3315 | emit_move_insn (operands[0], wrk); | |
3316 | return 1; | |
3317 | } | |
3318 | ||
3319 | wrk = gen_reg_rtx (Pmode); | |
3320 | ||
3321 | /* Load the value into an arg reg and call a helper. */ | |
c5c76735 | 3322 | emit_move_insn (gen_rtx_REG (SImode, 4), operands[1]); |
8aa2a305 | 3323 | sprintf (func, "__ashiftrt_r4_%d", value); |
73a4d10b | 3324 | function_symbol (wrk, func, SFUNC_STATIC); |
8aa2a305 | 3325 | emit_insn (gen_ashrsi3_n (GEN_INT (value), wrk)); |
c5c76735 | 3326 | emit_move_insn (operands[0], gen_rtx_REG (SImode, 4)); |
8aa2a305 | 3327 | return 1; |
bc45ade3 | 3328 | } |
8d481241 | 3329 | |
318881c0 | 3330 | int |
cf277499 | 3331 | sh_dynamicalize_shift_p (rtx count) |
1245df60 | 3332 | { |
6c34a092 | 3333 | return shift_insns[INTVAL (count) & 31] > 1 + SH_DYNAMIC_SHIFT_COST; |
1245df60 R |
3334 | } |
3335 | ||
8d481241 JW |
3336 | /* Try to find a good way to implement the combiner pattern |
3337 | [(set (match_operand:SI 0 "register_operand" "r") | |
3338 | (and:SI (ashift:SI (match_operand:SI 1 "register_operand" "r") | |
3339 | (match_operand:SI 2 "const_int_operand" "n")) | |
3340 | (match_operand:SI 3 "const_int_operand" "n"))) . | |
3341 | LEFT_RTX is operand 2 in the above pattern, and MASK_RTX is operand 3. | |
3342 | return 0 for simple right / left or left/right shift combination. | |
3343 | return 1 for a combination of shifts with zero_extend. | |
3344 | return 2 for a combination of shifts with an AND that needs r0. | |
3345 | return 3 for a combination of shifts with an AND that needs an extra | |
3346 | scratch register, when the three highmost bits of the AND mask are clear. | |
3347 | return 4 for a combination of shifts with an AND that needs an extra | |
3348 | scratch register, when any of the three highmost bits of the AND mask | |
3349 | is set. | |
3350 | If ATTRP is set, store an initial right shift width in ATTRP[0], | |
3351 | and the instruction length in ATTRP[1] . These values are not valid | |
3352 | when returning 0. | |
3353 | When ATTRP is set and returning 1, ATTRP[2] gets set to the index into | |
3354 | shift_amounts for the last shift value that is to be used before the | |
3355 | sign extend. */ | |
3356 | int | |
cf277499 | 3357 | shl_and_kind (rtx left_rtx, rtx mask_rtx, int *attrp) |
8d481241 JW |
3358 | { |
3359 | unsigned HOST_WIDE_INT mask, lsb, mask2, lsb2; | |
3360 | int left = INTVAL (left_rtx), right; | |
3361 | int best = 0; | |
3362 | int cost, best_cost = 10000; | |
3363 | int best_right = 0, best_len = 0; | |
3364 | int i; | |
3365 | int can_ext; | |
3366 | ||
3367 | if (left < 0 || left > 31) | |
3368 | return 0; | |
f3536097 | 3369 | if (CONST_INT_P (mask_rtx)) |
8d481241 JW |
3370 | mask = (unsigned HOST_WIDE_INT) INTVAL (mask_rtx) >> left; |
3371 | else | |
3372 | mask = (unsigned HOST_WIDE_INT) GET_MODE_MASK (SImode) >> left; | |
832a3292 | 3373 | /* Can this be expressed as a right shift / left shift pair? */ |
8d481241 JW |
3374 | lsb = ((mask ^ (mask - 1)) >> 1) + 1; |
3375 | right = exact_log2 (lsb); | |
3376 | mask2 = ~(mask + lsb - 1); | |
3377 | lsb2 = ((mask2 ^ (mask2 - 1)) >> 1) + 1; | |
3378 | /* mask has no zeroes but trailing zeroes <==> ! mask2 */ | |
3379 | if (! mask2) | |
3380 | best_cost = shift_insns[right] + shift_insns[right + left]; | |
3381 | /* mask has no trailing zeroes <==> ! right */ | |
3382 | else if (! right && mask2 == ~(lsb2 - 1)) | |
3383 | { | |
3384 | int late_right = exact_log2 (lsb2); | |
3385 | best_cost = shift_insns[left + late_right] + shift_insns[late_right]; | |
3386 | } | |
832a3292 | 3387 | /* Try to use zero extend. */ |
8d481241 JW |
3388 | if (mask2 == ~(lsb2 - 1)) |
3389 | { | |
3390 | int width, first; | |
3391 | ||
3392 | for (width = 8; width <= 16; width += 8) | |
3393 | { | |
832a3292 KH |
3394 | /* Can we zero-extend right away? */ |
3395 | if (lsb2 == (unsigned HOST_WIDE_INT) 1 << width) | |
8d481241 JW |
3396 | { |
3397 | cost | |
3398 | = 1 + ext_shift_insns[right] + ext_shift_insns[left + right]; | |
3399 | if (cost < best_cost) | |
3400 | { | |
3401 | best = 1; | |
3402 | best_cost = cost; | |
3403 | best_right = right; | |
3404 | best_len = cost; | |
3405 | if (attrp) | |
3406 | attrp[2] = -1; | |
3407 | } | |
3408 | continue; | |
3409 | } | |
3410 | /* ??? Could try to put zero extend into initial right shift, | |
6f317ef3 | 3411 | or even shift a bit left before the right shift. */ |
8d481241 JW |
3412 | /* Determine value of first part of left shift, to get to the |
3413 | zero extend cut-off point. */ | |
3414 | first = width - exact_log2 (lsb2) + right; | |
3415 | if (first >= 0 && right + left - first >= 0) | |
3416 | { | |
3417 | cost = ext_shift_insns[right] + ext_shift_insns[first] + 1 | |
3418 | + ext_shift_insns[right + left - first]; | |
3419 | if (cost < best_cost) | |
3420 | { | |
3421 | best = 1; | |
3422 | best_cost = cost; | |
3423 | best_right = right; | |
3424 | best_len = cost; | |
3425 | if (attrp) | |
3426 | attrp[2] = first; | |
832a3292 | 3427 | } |
8d481241 JW |
3428 | } |
3429 | } | |
3430 | } | |
3431 | /* Try to use r0 AND pattern */ | |
3432 | for (i = 0; i <= 2; i++) | |
3433 | { | |
3434 | if (i > right) | |
3435 | break; | |
735cb76e | 3436 | if (! CONST_OK_FOR_K08 (mask >> i)) |
8d481241 JW |
3437 | continue; |
3438 | cost = (i != 0) + 2 + ext_shift_insns[left + i]; | |
3439 | if (cost < best_cost) | |
3440 | { | |
3441 | best = 2; | |
3442 | best_cost = cost; | |
3443 | best_right = i; | |
3444 | best_len = cost - 1; | |
3445 | } | |
3446 | } | |
3447 | /* Try to use a scratch register to hold the AND operand. */ | |
832a3292 | 3448 | can_ext = ((mask << left) & ((unsigned HOST_WIDE_INT) 3 << 30)) == 0; |
8d481241 JW |
3449 | for (i = 0; i <= 2; i++) |
3450 | { | |
3451 | if (i > right) | |
3452 | break; | |
735cb76e | 3453 | cost = (i != 0) + (CONST_OK_FOR_I08 (mask >> i) ? 2 : 3) |
6ab911bb | 3454 | + (can_ext ? ext_shift_insns : shift_insns)[left + i]; |
8d481241 JW |
3455 | if (cost < best_cost) |
3456 | { | |
3457 | best = 4 - can_ext; | |
3458 | best_cost = cost; | |
3459 | best_right = i; | |
735cb76e | 3460 | best_len = cost - 1 - ! CONST_OK_FOR_I08 (mask >> i); |
8d481241 JW |
3461 | } |
3462 | } | |
3463 | ||
3464 | if (attrp) | |
3465 | { | |
3466 | attrp[0] = best_right; | |
3467 | attrp[1] = best_len; | |
3468 | } | |
3469 | return best; | |
3470 | } | |
3471 | ||
3472 | /* This is used in length attributes of the unnamed instructions | |
3473 | corresponding to shl_and_kind return values of 1 and 2. */ | |
3474 | int | |
cf277499 | 3475 | shl_and_length (rtx insn) |
8d481241 JW |
3476 | { |
3477 | rtx set_src, left_rtx, mask_rtx; | |
3478 | int attributes[3]; | |
3479 | ||
3480 | set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
3481 | left_rtx = XEXP (XEXP (set_src, 0), 1); | |
3482 | mask_rtx = XEXP (set_src, 1); | |
3483 | shl_and_kind (left_rtx, mask_rtx, attributes); | |
3484 | return attributes[1]; | |
3485 | } | |
3486 | ||
3487 | /* This is used in length attribute of the and_shl_scratch instruction. */ | |
3488 | ||
3489 | int | |
cf277499 | 3490 | shl_and_scr_length (rtx insn) |
8d481241 JW |
3491 | { |
3492 | rtx set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
6c34a092 | 3493 | int len = shift_insns[INTVAL (XEXP (set_src, 1)) & 31]; |
8d481241 | 3494 | rtx op = XEXP (set_src, 0); |
6c34a092 | 3495 | len += shift_insns[INTVAL (XEXP (op, 1)) & 31] + 1; |
8d481241 | 3496 | op = XEXP (XEXP (op, 0), 0); |
6c34a092 | 3497 | return len + shift_insns[INTVAL (XEXP (op, 1)) & 31]; |
8d481241 JW |
3498 | } |
3499 | ||
8d481241 JW |
3500 | /* Generate rtl for instructions for which shl_and_kind advised a particular |
3501 | method of generating them, i.e. returned zero. */ | |
3502 | ||
3503 | int | |
cf277499 | 3504 | gen_shl_and (rtx dest, rtx left_rtx, rtx mask_rtx, rtx source) |
8d481241 JW |
3505 | { |
3506 | int attributes[3]; | |
3507 | unsigned HOST_WIDE_INT mask; | |
3508 | int kind = shl_and_kind (left_rtx, mask_rtx, attributes); | |
3509 | int right, total_shift; | |
832a3292 | 3510 | void (*shift_gen_fun) (int, rtx *) = gen_shifty_hi_op; |
8d481241 JW |
3511 | |
3512 | right = attributes[0]; | |
3513 | total_shift = INTVAL (left_rtx) + right; | |
3514 | mask = (unsigned HOST_WIDE_INT) INTVAL (mask_rtx) >> total_shift; | |
3515 | switch (kind) | |
3516 | { | |
3517 | default: | |
3518 | return -1; | |
3519 | case 1: | |
3520 | { | |
3521 | int first = attributes[2]; | |
3522 | rtx operands[3]; | |
3523 | ||
3524 | if (first < 0) | |
3525 | { | |
7174c937 | 3526 | emit_insn ((mask << right) <= 0xff |
832a3292 KH |
3527 | ? gen_zero_extendqisi2 (dest, |
3528 | gen_lowpart (QImode, source)) | |
3529 | : gen_zero_extendhisi2 (dest, | |
3530 | gen_lowpart (HImode, source))); | |
8d481241 JW |
3531 | source = dest; |
3532 | } | |
3533 | if (source != dest) | |
3534 | emit_insn (gen_movsi (dest, source)); | |
3535 | operands[0] = dest; | |
3536 | if (right) | |
3537 | { | |
3538 | operands[2] = GEN_INT (right); | |
3539 | gen_shifty_hi_op (LSHIFTRT, operands); | |
3540 | } | |
3541 | if (first > 0) | |
3542 | { | |
3543 | operands[2] = GEN_INT (first); | |
3544 | gen_shifty_hi_op (ASHIFT, operands); | |
3545 | total_shift -= first; | |
3546 | mask <<= first; | |
3547 | } | |
3548 | if (first >= 0) | |
85af47b9 | 3549 | emit_insn (mask <= 0xff |
832a3292 KH |
3550 | ? gen_zero_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
3551 | : gen_zero_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
8d481241 JW |
3552 | if (total_shift > 0) |
3553 | { | |
3554 | operands[2] = GEN_INT (total_shift); | |
3555 | gen_shifty_hi_op (ASHIFT, operands); | |
3556 | } | |
3557 | break; | |
3558 | } | |
3559 | case 4: | |
3560 | shift_gen_fun = gen_shifty_op; | |
8d481241 | 3561 | case 3: |
24c50999 JR |
3562 | /* If the topmost bit that matters is set, set the topmost bits |
3563 | that don't matter. This way, we might be able to get a shorter | |
3564 | signed constant. */ | |
832a3292 KH |
3565 | if (mask & ((HOST_WIDE_INT) 1 << (31 - total_shift))) |
3566 | mask |= (HOST_WIDE_INT) ~0 << (31 - total_shift); | |
61fb6bac | 3567 | case 2: |
8d481241 JW |
3568 | /* Don't expand fine-grained when combining, because that will |
3569 | make the pattern fail. */ | |
4586b4ca | 3570 | if (currently_expanding_to_rtl |
8d481241 JW |
3571 | || reload_in_progress || reload_completed) |
3572 | { | |
3573 | rtx operands[3]; | |
f676971a | 3574 | |
61fb6bac R |
3575 | /* Cases 3 and 4 should be handled by this split |
3576 | only while combining */ | |
f5b9e7c9 | 3577 | gcc_assert (kind <= 2); |
8d481241 JW |
3578 | if (right) |
3579 | { | |
3580 | emit_insn (gen_lshrsi3 (dest, source, GEN_INT (right))); | |
3581 | source = dest; | |
3582 | } | |
3583 | emit_insn (gen_andsi3 (dest, source, GEN_INT (mask))); | |
afad3d2c JW |
3584 | if (total_shift) |
3585 | { | |
3586 | operands[0] = dest; | |
3587 | operands[1] = dest; | |
3588 | operands[2] = GEN_INT (total_shift); | |
3589 | shift_gen_fun (ASHIFT, operands); | |
3590 | } | |
8d481241 JW |
3591 | break; |
3592 | } | |
3593 | else | |
3594 | { | |
3595 | int neg = 0; | |
3596 | if (kind != 4 && total_shift < 16) | |
3597 | { | |
3598 | neg = -ext_shift_amounts[total_shift][1]; | |
3599 | if (neg > 0) | |
3600 | neg -= ext_shift_amounts[total_shift][2]; | |
3601 | else | |
3602 | neg = 0; | |
3603 | } | |
3604 | emit_insn (gen_and_shl_scratch (dest, source, | |
3605 | GEN_INT (right), | |
3606 | GEN_INT (mask), | |
3607 | GEN_INT (total_shift + neg), | |
3608 | GEN_INT (neg))); | |
3609 | emit_insn (gen_movsi (dest, dest)); | |
3610 | break; | |
3611 | } | |
3612 | } | |
3613 | return 0; | |
3614 | } | |
3615 | ||
3616 | /* Try to find a good way to implement the combiner pattern | |
3617 | [(set (match_operand:SI 0 "register_operand" "=r") | |
3618 | (sign_extract:SI (ashift:SI (match_operand:SI 1 "register_operand" "r") | |
3619 | (match_operand:SI 2 "const_int_operand" "n") | |
3620 | (match_operand:SI 3 "const_int_operand" "n") | |
3621 | (const_int 0))) | |
4773afa4 | 3622 | (clobber (reg:SI T_REG))] |
8d481241 JW |
3623 | LEFT_RTX is operand 2 in the above pattern, and SIZE_RTX is operand 3. |
3624 | return 0 for simple left / right shift combination. | |
3625 | return 1 for left shift / 8 bit sign extend / left shift. | |
3626 | return 2 for left shift / 16 bit sign extend / left shift. | |
3627 | return 3 for left shift / 8 bit sign extend / shift / sign extend. | |
3628 | return 4 for left shift / 16 bit sign extend / shift / sign extend. | |
3629 | return 5 for left shift / 16 bit sign extend / right shift | |
3630 | return 6 for < 8 bit sign extend / left shift. | |
3631 | return 7 for < 8 bit sign extend / left shift / single right shift. | |
3632 | If COSTP is nonzero, assign the calculated cost to *COSTP. */ | |
3633 | ||
3634 | int | |
cf277499 | 3635 | shl_sext_kind (rtx left_rtx, rtx size_rtx, int *costp) |
8d481241 JW |
3636 | { |
3637 | int left, size, insize, ext; | |
3a8699c7 | 3638 | int cost = 0, best_cost; |
8d481241 JW |
3639 | int kind; |
3640 | ||
3641 | left = INTVAL (left_rtx); | |
3642 | size = INTVAL (size_rtx); | |
3643 | insize = size - left; | |
f5b9e7c9 | 3644 | gcc_assert (insize > 0); |
8d481241 JW |
3645 | /* Default to left / right shift. */ |
3646 | kind = 0; | |
3647 | best_cost = shift_insns[32 - insize] + ashiftrt_insns[32 - size]; | |
3648 | if (size <= 16) | |
3649 | { | |
3650 | /* 16 bit shift / sign extend / 16 bit shift */ | |
3651 | cost = shift_insns[16 - insize] + 1 + ashiftrt_insns[16 - size]; | |
afad3d2c JW |
3652 | /* If ashiftrt_insns[16 - size] is 8, this choice will be overridden |
3653 | below, by alternative 3 or something even better. */ | |
8d481241 JW |
3654 | if (cost < best_cost) |
3655 | { | |
3656 | kind = 5; | |
3657 | best_cost = cost; | |
3658 | } | |
3659 | } | |
3660 | /* Try a plain sign extend between two shifts. */ | |
3661 | for (ext = 16; ext >= insize; ext -= 8) | |
3662 | { | |
3663 | if (ext <= size) | |
3664 | { | |
3665 | cost = ext_shift_insns[ext - insize] + 1 + shift_insns[size - ext]; | |
3666 | if (cost < best_cost) | |
3667 | { | |
11f9ed1a | 3668 | kind = ext / (unsigned) 8; |
8d481241 JW |
3669 | best_cost = cost; |
3670 | } | |
3671 | } | |
ae9d19c0 JR |
3672 | /* Check if we can do a sloppy shift with a final signed shift |
3673 | restoring the sign. */ | |
3674 | if (EXT_SHIFT_SIGNED (size - ext)) | |
3675 | cost = ext_shift_insns[ext - insize] + ext_shift_insns[size - ext] + 1; | |
3676 | /* If not, maybe it's still cheaper to do the second shift sloppy, | |
3677 | and do a final sign extend? */ | |
3678 | else if (size <= 16) | |
3679 | cost = ext_shift_insns[ext - insize] + 1 | |
3680 | + ext_shift_insns[size > ext ? size - ext : ext - size] + 1; | |
3681 | else | |
3682 | continue; | |
3683 | if (cost < best_cost) | |
8d481241 | 3684 | { |
11f9ed1a | 3685 | kind = ext / (unsigned) 8 + 2; |
ae9d19c0 | 3686 | best_cost = cost; |
8d481241 JW |
3687 | } |
3688 | } | |
3689 | /* Check if we can sign extend in r0 */ | |
3690 | if (insize < 8) | |
3691 | { | |
3692 | cost = 3 + shift_insns[left]; | |
3693 | if (cost < best_cost) | |
3694 | { | |
3695 | kind = 6; | |
3696 | best_cost = cost; | |
3697 | } | |
3698 | /* Try the same with a final signed shift. */ | |
3699 | if (left < 31) | |
3700 | { | |
3701 | cost = 3 + ext_shift_insns[left + 1] + 1; | |
3702 | if (cost < best_cost) | |
3703 | { | |
3704 | kind = 7; | |
3705 | best_cost = cost; | |
3706 | } | |
3707 | } | |
3708 | } | |
3709 | if (TARGET_SH3) | |
3710 | { | |
3711 | /* Try to use a dynamic shift. */ | |
1245df60 | 3712 | cost = shift_insns[32 - insize] + 1 + SH_DYNAMIC_SHIFT_COST; |
8d481241 JW |
3713 | if (cost < best_cost) |
3714 | { | |
3715 | kind = 0; | |
3716 | best_cost = cost; | |
3717 | } | |
3718 | } | |
3719 | if (costp) | |
3720 | *costp = cost; | |
3721 | return kind; | |
3722 | } | |
3723 | ||
3724 | /* Function to be used in the length attribute of the instructions | |
3725 | implementing this pattern. */ | |
3726 | ||
3727 | int | |
cf277499 | 3728 | shl_sext_length (rtx insn) |
8d481241 JW |
3729 | { |
3730 | rtx set_src, left_rtx, size_rtx; | |
3731 | int cost; | |
3732 | ||
3733 | set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
3734 | left_rtx = XEXP (XEXP (set_src, 0), 1); | |
3735 | size_rtx = XEXP (set_src, 1); | |
3736 | shl_sext_kind (left_rtx, size_rtx, &cost); | |
3737 | return cost; | |
3738 | } | |
3739 | ||
3740 | /* Generate rtl for this pattern */ | |
3741 | ||
3742 | int | |
cf277499 | 3743 | gen_shl_sext (rtx dest, rtx left_rtx, rtx size_rtx, rtx source) |
8d481241 JW |
3744 | { |
3745 | int kind; | |
d00d338c | 3746 | int left, size, insize, cost; |
8d481241 JW |
3747 | rtx operands[3]; |
3748 | ||
d00d338c | 3749 | kind = shl_sext_kind (left_rtx, size_rtx, &cost); |
8d481241 JW |
3750 | left = INTVAL (left_rtx); |
3751 | size = INTVAL (size_rtx); | |
3752 | insize = size - left; | |
3753 | switch (kind) | |
3754 | { | |
3755 | case 1: | |
3756 | case 2: | |
3757 | case 3: | |
3758 | case 4: | |
3759 | { | |
3760 | int ext = kind & 1 ? 8 : 16; | |
3761 | int shift2 = size - ext; | |
3762 | ||
3763 | /* Don't expand fine-grained when combining, because that will | |
3764 | make the pattern fail. */ | |
4586b4ca | 3765 | if (! currently_expanding_to_rtl |
8d481241 JW |
3766 | && ! reload_in_progress && ! reload_completed) |
3767 | { | |
3768 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
3769 | emit_insn (gen_movsi (dest, source)); | |
3770 | break; | |
3771 | } | |
3772 | if (dest != source) | |
3773 | emit_insn (gen_movsi (dest, source)); | |
3774 | operands[0] = dest; | |
afad3d2c JW |
3775 | if (ext - insize) |
3776 | { | |
3777 | operands[2] = GEN_INT (ext - insize); | |
3778 | gen_shifty_hi_op (ASHIFT, operands); | |
3779 | } | |
8d481241 | 3780 | emit_insn (kind & 1 |
832a3292 KH |
3781 | ? gen_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
3782 | : gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
8d481241 JW |
3783 | if (kind <= 2) |
3784 | { | |
afad3d2c JW |
3785 | if (shift2) |
3786 | { | |
3787 | operands[2] = GEN_INT (shift2); | |
3788 | gen_shifty_op (ASHIFT, operands); | |
3789 | } | |
8d481241 JW |
3790 | } |
3791 | else | |
3792 | { | |
afad3d2c | 3793 | if (shift2 > 0) |
8d481241 | 3794 | { |
ae9d19c0 JR |
3795 | if (EXT_SHIFT_SIGNED (shift2)) |
3796 | { | |
3797 | operands[2] = GEN_INT (shift2 + 1); | |
3798 | gen_shifty_op (ASHIFT, operands); | |
a556fd39 | 3799 | operands[2] = const1_rtx; |
ae9d19c0 JR |
3800 | gen_shifty_op (ASHIFTRT, operands); |
3801 | break; | |
3802 | } | |
8d481241 JW |
3803 | operands[2] = GEN_INT (shift2); |
3804 | gen_shifty_hi_op (ASHIFT, operands); | |
3805 | } | |
afad3d2c | 3806 | else if (shift2) |
8d481241 JW |
3807 | { |
3808 | operands[2] = GEN_INT (-shift2); | |
3809 | gen_shifty_hi_op (LSHIFTRT, operands); | |
3810 | } | |
3811 | emit_insn (size <= 8 | |
afad3d2c JW |
3812 | ? gen_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
3813 | : gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
8d481241 JW |
3814 | } |
3815 | break; | |
3816 | } | |
3817 | case 5: | |
afad3d2c JW |
3818 | { |
3819 | int i = 16 - size; | |
4586b4ca | 3820 | if (! currently_expanding_to_rtl |
913d8e13 R |
3821 | && ! reload_in_progress && ! reload_completed) |
3822 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
3823 | else | |
3824 | { | |
3825 | operands[0] = dest; | |
3826 | operands[2] = GEN_INT (16 - insize); | |
3827 | gen_shifty_hi_op (ASHIFT, operands); | |
3828 | emit_insn (gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
3829 | } | |
afad3d2c JW |
3830 | /* Don't use gen_ashrsi3 because it generates new pseudos. */ |
3831 | while (--i >= 0) | |
3832 | gen_ashift (ASHIFTRT, 1, dest); | |
3833 | break; | |
3834 | } | |
8d481241 JW |
3835 | case 6: |
3836 | case 7: | |
3837 | /* Don't expand fine-grained when combining, because that will | |
3838 | make the pattern fail. */ | |
4586b4ca | 3839 | if (! currently_expanding_to_rtl |
8d481241 JW |
3840 | && ! reload_in_progress && ! reload_completed) |
3841 | { | |
3842 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
3843 | emit_insn (gen_movsi (dest, source)); | |
3844 | break; | |
3845 | } | |
3846 | emit_insn (gen_andsi3 (dest, source, GEN_INT ((1 << insize) - 1))); | |
3847 | emit_insn (gen_xorsi3 (dest, dest, GEN_INT (1 << (insize - 1)))); | |
3848 | emit_insn (gen_addsi3 (dest, dest, GEN_INT (-1 << (insize - 1)))); | |
3849 | operands[0] = dest; | |
3850 | operands[2] = kind == 7 ? GEN_INT (left + 1) : left_rtx; | |
3851 | gen_shifty_op (ASHIFT, operands); | |
3852 | if (kind == 7) | |
a556fd39 | 3853 | emit_insn (gen_ashrsi3_k (dest, dest, const1_rtx)); |
8d481241 JW |
3854 | break; |
3855 | default: | |
3856 | return -1; | |
3857 | } | |
3858 | return 0; | |
3859 | } | |
fa5322fa AO |
3860 | |
3861 | /* Prefix a symbol_ref name with "datalabel". */ | |
3862 | ||
3863 | rtx | |
cf277499 | 3864 | gen_datalabel_ref (rtx sym) |
fa5322fa | 3865 | { |
73a4d10b R |
3866 | const char *str; |
3867 | ||
fa5322fa AO |
3868 | if (GET_CODE (sym) == LABEL_REF) |
3869 | return gen_rtx_CONST (GET_MODE (sym), | |
3870 | gen_rtx_UNSPEC (GET_MODE (sym), | |
3871 | gen_rtvec (1, sym), | |
3872 | UNSPEC_DATALABEL)); | |
f676971a | 3873 | |
f5b9e7c9 | 3874 | gcc_assert (GET_CODE (sym) == SYMBOL_REF); |
fa5322fa | 3875 | |
73a4d10b R |
3876 | str = XSTR (sym, 0); |
3877 | /* Share all SYMBOL_REF strings with the same value - that is important | |
3878 | for cse. */ | |
3879 | str = IDENTIFIER_POINTER (get_identifier (str)); | |
3880 | XSTR (sym, 0) = str; | |
3881 | ||
fa5322fa AO |
3882 | return sym; |
3883 | } | |
3884 | ||
8aa2a305 | 3885 | \f |
6dd3c0a5 SB |
3886 | static alloc_pool label_ref_list_pool; |
3887 | ||
3888 | typedef struct label_ref_list_d | |
3889 | { | |
3890 | rtx label; | |
3891 | struct label_ref_list_d *next; | |
3892 | } *label_ref_list_t; | |
3893 | ||
8aa2a305 JW |
3894 | /* The SH cannot load a large constant into a register, constants have to |
3895 | come from a pc relative load. The reference of a pc relative load | |
0fa2e4df | 3896 | instruction must be less than 1k in front of the instruction. This |
8aa2a305 JW |
3897 | means that we often have to dump a constant inside a function, and |
3898 | generate code to branch around it. | |
bc45ade3 | 3899 | |
8aa2a305 JW |
3900 | It is important to minimize this, since the branches will slow things |
3901 | down and make things bigger. | |
3902 | ||
3903 | Worst case code looks like: | |
3904 | ||
3905 | mov.l L1,rn | |
3906 | bra L2 | |
3907 | nop | |
3908 | align | |
3909 | L1: .long value | |
3910 | L2: | |
3911 | .. | |
3912 | ||
3913 | mov.l L3,rn | |
3914 | bra L4 | |
3915 | nop | |
3916 | align | |
3917 | L3: .long value | |
3918 | L4: | |
3919 | .. | |
3920 | ||
3921 | We fix this by performing a scan before scheduling, which notices which | |
3922 | instructions need to have their operands fetched from the constant table | |
3923 | and builds the table. | |
3924 | ||
3925 | The algorithm is: | |
3926 | ||
3927 | scan, find an instruction which needs a pcrel move. Look forward, find the | |
3928 | last barrier which is within MAX_COUNT bytes of the requirement. | |
3929 | If there isn't one, make one. Process all the instructions between | |
3930 | the find and the barrier. | |
3931 | ||
3932 | In the above example, we can tell that L3 is within 1k of L1, so | |
3933 | the first move can be shrunk from the 3 insn+constant sequence into | |
3934 | just 1 insn, and the constant moved to L3 to make: | |
3935 | ||
3936 | mov.l L1,rn | |
3937 | .. | |
3938 | mov.l L3,rn | |
3939 | bra L4 | |
3940 | nop | |
3941 | align | |
3942 | L3:.long value | |
3943 | L4:.long value | |
3944 | ||
3945 | Then the second move becomes the target for the shortening process. */ | |
3946 | ||
3947 | typedef struct | |
3948 | { | |
3949 | rtx value; /* Value in table. */ | |
3950 | rtx label; /* Label of value. */ | |
6dd3c0a5 | 3951 | label_ref_list_t wend; /* End of window. */ |
8aa2a305 | 3952 | enum machine_mode mode; /* Mode of value. */ |
3503150c RS |
3953 | |
3954 | /* True if this constant is accessed as part of a post-increment | |
3955 | sequence. Note that HImode constants are never accessed in this way. */ | |
3956 | bool part_of_sequence_p; | |
8aa2a305 JW |
3957 | } pool_node; |
3958 | ||
3959 | /* The maximum number of constants that can fit into one pool, since | |
73a4d10b R |
3960 | constants in the range 0..510 are at least 2 bytes long, and in the |
3961 | range from there to 1018 at least 4 bytes. */ | |
8aa2a305 | 3962 | |
73a4d10b | 3963 | #define MAX_POOL_SIZE 372 |
8aa2a305 JW |
3964 | static pool_node pool_vector[MAX_POOL_SIZE]; |
3965 | static int pool_size; | |
b91455de KK |
3966 | static rtx pool_window_label; |
3967 | static int pool_window_last; | |
8aa2a305 | 3968 | |
4241ecb0 R |
3969 | static int max_labelno_before_reorg; |
3970 | ||
8aa2a305 JW |
3971 | /* ??? If we need a constant in HImode which is the truncated value of a |
3972 | constant we need in SImode, we could combine the two entries thus saving | |
3973 | two bytes. Is this common enough to be worth the effort of implementing | |
3974 | it? */ | |
3975 | ||
3976 | /* ??? This stuff should be done at the same time that we shorten branches. | |
3977 | As it is now, we must assume that all branches are the maximum size, and | |
3978 | this causes us to almost always output constant pools sooner than | |
3979 | necessary. */ | |
3980 | ||
3981 | /* Add a constant to the pool and return its label. */ | |
3982 | ||
3983 | static rtx | |
cf277499 | 3984 | add_constant (rtx x, enum machine_mode mode, rtx last_value) |
0d7e008e SC |
3985 | { |
3986 | int i; | |
0a2aaacc | 3987 | rtx lab, new_rtx; |
6dd3c0a5 | 3988 | label_ref_list_t ref, newref; |
8aa2a305 JW |
3989 | |
3990 | /* First see if we've already got it. */ | |
3991 | for (i = 0; i < pool_size; i++) | |
0d7e008e | 3992 | { |
8aa2a305 JW |
3993 | if (x->code == pool_vector[i].value->code |
3994 | && mode == pool_vector[i].mode) | |
0d7e008e | 3995 | { |
8aa2a305 JW |
3996 | if (x->code == CODE_LABEL) |
3997 | { | |
3998 | if (XINT (x, 3) != XINT (pool_vector[i].value, 3)) | |
3999 | continue; | |
4000 | } | |
4001 | if (rtx_equal_p (x, pool_vector[i].value)) | |
225e4f43 | 4002 | { |
0a2aaacc | 4003 | lab = new_rtx = 0; |
225e4f43 R |
4004 | if (! last_value |
4005 | || ! i | |
4006 | || ! rtx_equal_p (last_value, pool_vector[i-1].value)) | |
4007 | { | |
0a2aaacc KG |
4008 | new_rtx = gen_label_rtx (); |
4009 | LABEL_REFS (new_rtx) = pool_vector[i].label; | |
4010 | pool_vector[i].label = lab = new_rtx; | |
225e4f43 | 4011 | } |
b91455de KK |
4012 | if (lab && pool_window_label) |
4013 | { | |
6dd3c0a5 SB |
4014 | newref = (label_ref_list_t) pool_alloc (label_ref_list_pool); |
4015 | newref->label = pool_window_label; | |
b91455de | 4016 | ref = pool_vector[pool_window_last].wend; |
6dd3c0a5 | 4017 | newref->next = ref; |
b91455de KK |
4018 | pool_vector[pool_window_last].wend = newref; |
4019 | } | |
0a2aaacc KG |
4020 | if (new_rtx) |
4021 | pool_window_label = new_rtx; | |
b91455de | 4022 | pool_window_last = i; |
225e4f43 R |
4023 | return lab; |
4024 | } | |
0d7e008e | 4025 | } |
0d7e008e | 4026 | } |
b9654711 | 4027 | |
8aa2a305 JW |
4028 | /* Need a new one. */ |
4029 | pool_vector[pool_size].value = x; | |
225e4f43 | 4030 | if (last_value && rtx_equal_p (last_value, pool_vector[pool_size - 1].value)) |
3503150c RS |
4031 | { |
4032 | lab = 0; | |
4033 | pool_vector[pool_size - 1].part_of_sequence_p = true; | |
4034 | } | |
225e4f43 R |
4035 | else |
4036 | lab = gen_label_rtx (); | |
8aa2a305 JW |
4037 | pool_vector[pool_size].mode = mode; |
4038 | pool_vector[pool_size].label = lab; | |
6dd3c0a5 | 4039 | pool_vector[pool_size].wend = NULL; |
3503150c | 4040 | pool_vector[pool_size].part_of_sequence_p = (lab == 0); |
b91455de KK |
4041 | if (lab && pool_window_label) |
4042 | { | |
6dd3c0a5 SB |
4043 | newref = (label_ref_list_t) pool_alloc (label_ref_list_pool); |
4044 | newref->label = pool_window_label; | |
b91455de | 4045 | ref = pool_vector[pool_window_last].wend; |
6dd3c0a5 | 4046 | newref->next = ref; |
b91455de KK |
4047 | pool_vector[pool_window_last].wend = newref; |
4048 | } | |
4049 | if (lab) | |
4050 | pool_window_label = lab; | |
4051 | pool_window_last = pool_size; | |
8aa2a305 JW |
4052 | pool_size++; |
4053 | return lab; | |
0d7e008e | 4054 | } |
16bea517 | 4055 | |
078c8b08 R |
4056 | /* Output the literal table. START, if nonzero, is the first instruction |
4057 | this table is needed for, and also indicates that there is at least one | |
4058 | casesi_worker_2 instruction; We have to emit the operand3 labels from | |
4059 | these insns at a 4-byte aligned position. BARRIER is the barrier | |
4060 | after which we are to place the table. */ | |
b9654711 | 4061 | |
b9654711 | 4062 | static void |
078c8b08 | 4063 | dump_table (rtx start, rtx barrier) |
b9654711 | 4064 | { |
078c8b08 | 4065 | rtx scan = barrier; |
0d7e008e | 4066 | int i; |
0d7e008e | 4067 | int need_align = 1; |
6dd3c0a5 SB |
4068 | rtx lab; |
4069 | label_ref_list_t ref; | |
aa06e8f5 | 4070 | int have_df = 0; |
b9654711 | 4071 | |
16bea517 | 4072 | /* Do two passes, first time dump out the HI sized constants. */ |
b9654711 | 4073 | |
0d7e008e | 4074 | for (i = 0; i < pool_size; i++) |
b9654711 | 4075 | { |
8aa2a305 JW |
4076 | pool_node *p = &pool_vector[i]; |
4077 | ||
0d7e008e SC |
4078 | if (p->mode == HImode) |
4079 | { | |
4080 | if (need_align) | |
4081 | { | |
4082 | scan = emit_insn_after (gen_align_2 (), scan); | |
4083 | need_align = 0; | |
4084 | } | |
b91455de KK |
4085 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
4086 | scan = emit_label_after (lab, scan); | |
4087 | scan = emit_insn_after (gen_consttable_2 (p->value, const0_rtx), | |
4088 | scan); | |
6dd3c0a5 | 4089 | for (ref = p->wend; ref; ref = ref->next) |
b91455de | 4090 | { |
6dd3c0a5 | 4091 | lab = ref->label; |
b91455de KK |
4092 | scan = emit_insn_after (gen_consttable_window_end (lab), scan); |
4093 | } | |
0d7e008e | 4094 | } |
aa06e8f5 R |
4095 | else if (p->mode == DFmode) |
4096 | have_df = 1; | |
b9654711 | 4097 | } |
8aa2a305 | 4098 | |
0d7e008e | 4099 | need_align = 1; |
b9654711 | 4100 | |
078c8b08 R |
4101 | if (start) |
4102 | { | |
4103 | scan = emit_insn_after (gen_align_4 (), scan); | |
4104 | need_align = 0; | |
4105 | for (; start != barrier; start = NEXT_INSN (start)) | |
f3536097 | 4106 | if (NONJUMP_INSN_P (start) |
078c8b08 R |
4107 | && recog_memoized (start) == CODE_FOR_casesi_worker_2) |
4108 | { | |
4109 | rtx src = SET_SRC (XVECEXP (PATTERN (start), 0, 0)); | |
4110 | rtx lab = XEXP (XVECEXP (src, 0, 3), 0); | |
4111 | ||
4112 | scan = emit_label_after (lab, scan); | |
4113 | } | |
4114 | } | |
aa06e8f5 | 4115 | if (TARGET_FMOVD && TARGET_ALIGN_DOUBLE && have_df) |
fa5322fa AO |
4116 | { |
4117 | rtx align_insn = NULL_RTX; | |
4118 | ||
4119 | scan = emit_label_after (gen_label_rtx (), scan); | |
4120 | scan = emit_insn_after (gen_align_log (GEN_INT (3)), scan); | |
4121 | need_align = 0; | |
4122 | ||
4123 | for (i = 0; i < pool_size; i++) | |
4124 | { | |
4125 | pool_node *p = &pool_vector[i]; | |
4126 | ||
4127 | switch (p->mode) | |
4128 | { | |
4129 | case HImode: | |
4130 | break; | |
4131 | case SImode: | |
4132 | case SFmode: | |
3503150c | 4133 | if (align_insn && !p->part_of_sequence_p) |
fa5322fa AO |
4134 | { |
4135 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) | |
4136 | emit_label_before (lab, align_insn); | |
4137 | emit_insn_before (gen_consttable_4 (p->value, const0_rtx), | |
4138 | align_insn); | |
6dd3c0a5 | 4139 | for (ref = p->wend; ref; ref = ref->next) |
fa5322fa | 4140 | { |
6dd3c0a5 | 4141 | lab = ref->label; |
fa5322fa | 4142 | emit_insn_before (gen_consttable_window_end (lab), |
832a3292 | 4143 | align_insn); |
fa5322fa AO |
4144 | } |
4145 | delete_insn (align_insn); | |
4146 | align_insn = NULL_RTX; | |
4147 | continue; | |
4148 | } | |
4149 | else | |
4150 | { | |
4151 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) | |
4152 | scan = emit_label_after (lab, scan); | |
4153 | scan = emit_insn_after (gen_consttable_4 (p->value, | |
4154 | const0_rtx), scan); | |
4155 | need_align = ! need_align; | |
4156 | } | |
4157 | break; | |
4158 | case DFmode: | |
fa5322fa AO |
4159 | if (need_align) |
4160 | { | |
4161 | scan = emit_insn_after (gen_align_log (GEN_INT (3)), scan); | |
4162 | align_insn = scan; | |
4163 | need_align = 0; | |
4164 | } | |
aa06e8f5 | 4165 | case DImode: |
fa5322fa AO |
4166 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
4167 | scan = emit_label_after (lab, scan); | |
4168 | scan = emit_insn_after (gen_consttable_8 (p->value, const0_rtx), | |
4169 | scan); | |
4170 | break; | |
4171 | default: | |
f5b9e7c9 | 4172 | gcc_unreachable (); |
fa5322fa AO |
4173 | } |
4174 | ||
4175 | if (p->mode != HImode) | |
4176 | { | |
6dd3c0a5 | 4177 | for (ref = p->wend; ref; ref = ref->next) |
fa5322fa | 4178 | { |
6dd3c0a5 | 4179 | lab = ref->label; |
fa5322fa AO |
4180 | scan = emit_insn_after (gen_consttable_window_end (lab), |
4181 | scan); | |
4182 | } | |
4183 | } | |
4184 | } | |
4185 | ||
4186 | pool_size = 0; | |
4187 | } | |
f676971a | 4188 | |
0d7e008e | 4189 | for (i = 0; i < pool_size; i++) |
b9654711 | 4190 | { |
8aa2a305 | 4191 | pool_node *p = &pool_vector[i]; |
b9654711 | 4192 | |
0d7e008e | 4193 | switch (p->mode) |
b9654711 | 4194 | { |
0d7e008e SC |
4195 | case HImode: |
4196 | break; | |
4197 | case SImode: | |
e577c183 | 4198 | case SFmode: |
0d7e008e | 4199 | if (need_align) |
b9654711 | 4200 | { |
0d7e008e | 4201 | need_align = 0; |
d3ae8277 | 4202 | scan = emit_label_after (gen_label_rtx (), scan); |
0d7e008e | 4203 | scan = emit_insn_after (gen_align_4 (), scan); |
b9654711 | 4204 | } |
b91455de KK |
4205 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
4206 | scan = emit_label_after (lab, scan); | |
4207 | scan = emit_insn_after (gen_consttable_4 (p->value, const0_rtx), | |
4208 | scan); | |
0d7e008e | 4209 | break; |
e577c183 | 4210 | case DFmode: |
0d7e008e SC |
4211 | case DImode: |
4212 | if (need_align) | |
4213 | { | |
4214 | need_align = 0; | |
d3ae8277 | 4215 | scan = emit_label_after (gen_label_rtx (), scan); |
0d7e008e SC |
4216 | scan = emit_insn_after (gen_align_4 (), scan); |
4217 | } | |
b91455de KK |
4218 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
4219 | scan = emit_label_after (lab, scan); | |
4220 | scan = emit_insn_after (gen_consttable_8 (p->value, const0_rtx), | |
4221 | scan); | |
0d7e008e SC |
4222 | break; |
4223 | default: | |
f5b9e7c9 | 4224 | gcc_unreachable (); |
b9654711 | 4225 | } |
b91455de KK |
4226 | |
4227 | if (p->mode != HImode) | |
4228 | { | |
6dd3c0a5 | 4229 | for (ref = p->wend; ref; ref = ref->next) |
b91455de | 4230 | { |
6dd3c0a5 | 4231 | lab = ref->label; |
b91455de KK |
4232 | scan = emit_insn_after (gen_consttable_window_end (lab), scan); |
4233 | } | |
4234 | } | |
b9654711 | 4235 | } |
b9654711 | 4236 | |
0d7e008e SC |
4237 | scan = emit_insn_after (gen_consttable_end (), scan); |
4238 | scan = emit_barrier_after (scan); | |
4239 | pool_size = 0; | |
b91455de KK |
4240 | pool_window_label = NULL_RTX; |
4241 | pool_window_last = 0; | |
0d7e008e | 4242 | } |
b9654711 | 4243 | |
5e7a8ee0 | 4244 | /* Return nonzero if constant would be an ok source for a |
8aa2a305 | 4245 | mov.w instead of a mov.l. */ |
b9654711 | 4246 | |
16bea517 | 4247 | static int |
cf277499 | 4248 | hi_const (rtx src) |
0d7e008e | 4249 | { |
f3536097 | 4250 | return (CONST_INT_P (src) |
8aa2a305 JW |
4251 | && INTVAL (src) >= -32768 |
4252 | && INTVAL (src) <= 32767); | |
b9654711 SC |
4253 | } |
4254 | ||
4241ecb0 R |
4255 | #define MOVA_LABELREF(mova) XVECEXP (SET_SRC (PATTERN (mova)), 0, 0) |
4256 | ||
5e7a8ee0 | 4257 | /* Nonzero if the insn is a move instruction which needs to be fixed. */ |
8aa2a305 JW |
4258 | |
4259 | /* ??? For a DImode/DFmode moves, we don't need to fix it if each half of the | |
735cb76e R |
4260 | CONST_DOUBLE input value is CONST_OK_FOR_I08. For a SFmode move, we don't |
4261 | need to fix it if the input value is CONST_OK_FOR_I08. */ | |
8aa2a305 JW |
4262 | |
4263 | static int | |
cf277499 | 4264 | broken_move (rtx insn) |
b9654711 | 4265 | { |
f3536097 | 4266 | if (NONJUMP_INSN_P (insn)) |
e577c183 JR |
4267 | { |
4268 | rtx pat = PATTERN (insn); | |
4269 | if (GET_CODE (pat) == PARALLEL) | |
4270 | pat = XVECEXP (pat, 0, 0); | |
4271 | if (GET_CODE (pat) == SET | |
3b962ea0 | 4272 | /* We can load any 8-bit value if we don't care what the high |
e577c183 JR |
4273 | order bits end up as. */ |
4274 | && GET_MODE (SET_DEST (pat)) != QImode | |
43c05634 AO |
4275 | && (CONSTANT_P (SET_SRC (pat)) |
4276 | /* Match mova_const. */ | |
4277 | || (GET_CODE (SET_SRC (pat)) == UNSPEC | |
4278 | && XINT (SET_SRC (pat), 1) == UNSPEC_MOVA | |
4279 | && GET_CODE (XVECEXP (SET_SRC (pat), 0, 0)) == CONST)) | |
3a8699c7 | 4280 | && ! (TARGET_SH2E |
1245df60 | 4281 | && GET_CODE (SET_SRC (pat)) == CONST_DOUBLE |
e577c183 JR |
4282 | && (fp_zero_operand (SET_SRC (pat)) |
4283 | || fp_one_operand (SET_SRC (pat))) | |
7a61cf6f | 4284 | /* In general we don't know the current setting of fpscr, so disable fldi. |
18778292 R |
4285 | There is an exception if this was a register-register move |
4286 | before reload - and hence it was ascertained that we have | |
4287 | single precision setting - and in a post-reload optimization | |
4288 | we changed this to do a constant load. In that case | |
4289 | we don't have an r0 clobber, hence we must use fldi. */ | |
7a61cf6f | 4290 | && (TARGET_FMOVD |
18778292 R |
4291 | || (GET_CODE (XEXP (XVECEXP (PATTERN (insn), 0, 2), 0)) |
4292 | == SCRATCH)) | |
f3536097 | 4293 | && REG_P (SET_DEST (pat)) |
104ee20b | 4294 | && FP_REGISTER_P (REGNO (SET_DEST (pat)))) |
157371cf AO |
4295 | && ! (TARGET_SH2A |
4296 | && GET_MODE (SET_DEST (pat)) == SImode | |
e990551a JS |
4297 | && (satisfies_constraint_I20 (SET_SRC (pat)) |
4298 | || satisfies_constraint_I28 (SET_SRC (pat)))) | |
32a7ab3d | 4299 | && ! satisfies_constraint_I08 (SET_SRC (pat))) |
e577c183 JR |
4300 | return 1; |
4301 | } | |
d3ae8277 | 4302 | |
8aa2a305 | 4303 | return 0; |
b9654711 | 4304 | } |
b9654711 | 4305 | |
1245df60 | 4306 | static int |
cf277499 | 4307 | mova_p (rtx insn) |
1245df60 | 4308 | { |
f3536097 | 4309 | return (NONJUMP_INSN_P (insn) |
1245df60 R |
4310 | && GET_CODE (PATTERN (insn)) == SET |
4311 | && GET_CODE (SET_SRC (PATTERN (insn))) == UNSPEC | |
43c05634 AO |
4312 | && XINT (SET_SRC (PATTERN (insn)), 1) == UNSPEC_MOVA |
4313 | /* Don't match mova_const. */ | |
4241ecb0 | 4314 | && GET_CODE (MOVA_LABELREF (insn)) == LABEL_REF); |
1245df60 R |
4315 | } |
4316 | ||
078c8b08 R |
4317 | /* Fix up a mova from a switch that went out of range. */ |
4318 | static void | |
4319 | fixup_mova (rtx mova) | |
4320 | { | |
4241ecb0 | 4321 | PUT_MODE (XEXP (MOVA_LABELREF (mova), 0), QImode); |
078c8b08 R |
4322 | if (! flag_pic) |
4323 | { | |
4241ecb0 | 4324 | SET_SRC (PATTERN (mova)) = MOVA_LABELREF (mova); |
078c8b08 R |
4325 | INSN_CODE (mova) = -1; |
4326 | } | |
4327 | else | |
4328 | { | |
4329 | rtx worker = mova; | |
4330 | rtx lab = gen_label_rtx (); | |
dc3ba671 | 4331 | rtx wpat, wpat0, wpat1, wsrc, target, base, diff; |
078c8b08 R |
4332 | |
4333 | do | |
4334 | { | |
4335 | worker = NEXT_INSN (worker); | |
f5b9e7c9 | 4336 | gcc_assert (worker |
f3536097 SZ |
4337 | && !LABEL_P (worker) |
4338 | && !JUMP_P (worker)); | |
4339 | } while (NOTE_P (worker) | |
0bf72672 | 4340 | || recog_memoized (worker) != CODE_FOR_casesi_worker_1); |
078c8b08 R |
4341 | wpat = PATTERN (worker); |
4342 | wpat0 = XVECEXP (wpat, 0, 0); | |
4343 | wpat1 = XVECEXP (wpat, 0, 1); | |
4344 | wsrc = SET_SRC (wpat0); | |
4345 | PATTERN (worker) = (gen_casesi_worker_2 | |
4346 | (SET_DEST (wpat0), XVECEXP (wsrc, 0, 1), | |
4347 | XEXP (XVECEXP (wsrc, 0, 2), 0), lab, | |
4348 | XEXP (wpat1, 0))); | |
4349 | INSN_CODE (worker) = -1; | |
dc3ba671 RS |
4350 | target = XVECEXP (SET_SRC (PATTERN (mova)), 0, 0); |
4351 | base = gen_rtx_LABEL_REF (Pmode, lab); | |
4352 | diff = gen_rtx_UNSPEC (Pmode, gen_rtvec (2, target, base), UNSPEC_SYMOFF); | |
078c8b08 R |
4353 | SET_SRC (PATTERN (mova)) = gen_rtx_CONST (Pmode, diff); |
4354 | INSN_CODE (mova) = -1; | |
4355 | } | |
4356 | } | |
4357 | ||
4241ecb0 R |
4358 | /* NEW_MOVA is a mova we've just encountered while scanning forward. Update |
4359 | *num_mova, and check if the new mova is not nested within the first one. | |
4360 | return 0 if *first_mova was replaced, 1 if new_mova was replaced, | |
4361 | 2 if new_mova has been assigned to *first_mova, -1 otherwise.. */ | |
4362 | static int | |
4363 | untangle_mova (int *num_mova, rtx *first_mova, rtx new_mova) | |
4364 | { | |
c6d56443 KK |
4365 | int n_addr = 0; /* Initialization to shut up spurious warning. */ |
4366 | int f_target, n_target = 0; /* Likewise. */ | |
4241ecb0 R |
4367 | |
4368 | if (optimize) | |
4369 | { | |
77253e56 KK |
4370 | /* If NEW_MOVA has no address yet, it will be handled later. */ |
4371 | if (INSN_ADDRESSES_SIZE() <= (unsigned) INSN_UID (new_mova)) | |
4372 | return -1; | |
4373 | ||
4241ecb0 R |
4374 | n_addr = INSN_ADDRESSES (INSN_UID (new_mova)); |
4375 | n_target = INSN_ADDRESSES (INSN_UID (XEXP (MOVA_LABELREF (new_mova), 0))); | |
4376 | if (n_addr > n_target || n_addr + 1022 < n_target) | |
4377 | { | |
4378 | /* Change the mova into a load. | |
4379 | broken_move will then return true for it. */ | |
4380 | fixup_mova (new_mova); | |
4381 | return 1; | |
4382 | } | |
4383 | } | |
4384 | if (!(*num_mova)++) | |
4385 | { | |
4386 | *first_mova = new_mova; | |
4387 | return 2; | |
4388 | } | |
4389 | if (!optimize | |
4390 | || ((f_target | |
4391 | = INSN_ADDRESSES (INSN_UID (XEXP (MOVA_LABELREF (*first_mova), 0)))) | |
4392 | >= n_target)) | |
4393 | return -1; | |
4394 | ||
4395 | (*num_mova)--; | |
4396 | if (f_target - INSN_ADDRESSES (INSN_UID (*first_mova)) | |
4397 | > n_target - n_addr) | |
4398 | { | |
4399 | fixup_mova (*first_mova); | |
4400 | return 0; | |
4401 | } | |
4402 | else | |
4403 | { | |
4404 | fixup_mova (new_mova); | |
4405 | return 1; | |
4406 | } | |
4407 | } | |
4408 | ||
8aa2a305 JW |
4409 | /* Find the last barrier from insn FROM which is close enough to hold the |
4410 | constant pool. If we can't find one, then create one near the end of | |
4411 | the range. */ | |
b9654711 | 4412 | |
8aa2a305 | 4413 | static rtx |
cf277499 | 4414 | find_barrier (int num_mova, rtx mova, rtx from) |
b9654711 | 4415 | { |
0d7e008e SC |
4416 | int count_si = 0; |
4417 | int count_hi = 0; | |
4418 | int found_hi = 0; | |
4419 | int found_si = 0; | |
fa5322fa | 4420 | int found_di = 0; |
33f7f353 JR |
4421 | int hi_align = 2; |
4422 | int si_align = 2; | |
1245df60 | 4423 | int leading_mova = num_mova; |
3a8699c7 | 4424 | rtx barrier_before_mova = 0, found_barrier = 0, good_barrier = 0; |
c17f53a8 JW |
4425 | int si_limit; |
4426 | int hi_limit; | |
7a3cb3ab | 4427 | rtx orig = from; |
e938feb4 | 4428 | rtx last_got = NULL_RTX; |
fc9c984d | 4429 | rtx last_symoff = NULL_RTX; |
8aa2a305 JW |
4430 | |
4431 | /* For HImode: range is 510, add 4 because pc counts from address of | |
4432 | second instruction after this one, subtract 2 for the jump instruction | |
3adbde60 JW |
4433 | that we may need to emit before the table, subtract 2 for the instruction |
4434 | that fills the jump delay slot (in very rare cases, reorg will take an | |
4435 | instruction from after the constant pool or will leave the delay slot | |
4436 | empty). This gives 510. | |
8aa2a305 JW |
4437 | For SImode: range is 1020, add 4 because pc counts from address of |
4438 | second instruction after this one, subtract 2 in case pc is 2 byte | |
4439 | aligned, subtract 2 for the jump instruction that we may need to emit | |
3adbde60 JW |
4440 | before the table, subtract 2 for the instruction that fills the jump |
4441 | delay slot. This gives 1018. */ | |
c17f53a8 | 4442 | |
1245df60 | 4443 | /* The branch will always be shortened now that the reference address for |
956d6950 | 4444 | forward branches is the successor address, thus we need no longer make |
1245df60 | 4445 | adjustments to the [sh]i_limit for -O0. */ |
c17f53a8 | 4446 | |
1245df60 R |
4447 | si_limit = 1018; |
4448 | hi_limit = 510; | |
e4fa6b06 | 4449 | |
c17f53a8 | 4450 | while (from && count_si < si_limit && count_hi < hi_limit) |
0d7e008e | 4451 | { |
33f7f353 JR |
4452 | int inc = get_attr_length (from); |
4453 | int new_align = 1; | |
1245df60 | 4454 | |
4241ecb0 R |
4455 | /* If this is a label that existed at the time of the compute_alignments |
4456 | call, determine the alignment. N.B. When find_barrier recurses for | |
4457 | an out-of-reach mova, we might see labels at the start of previously | |
4458 | inserted constant tables. */ | |
f3536097 | 4459 | if (LABEL_P (from) |
4241ecb0 | 4460 | && CODE_LABEL_NUMBER (from) <= max_labelno_before_reorg) |
77008a44 R |
4461 | { |
4462 | if (optimize) | |
4463 | new_align = 1 << label_to_alignment (from); | |
f3536097 | 4464 | else if (BARRIER_P (prev_nonnote_insn (from))) |
77008a44 R |
4465 | new_align = 1 << barrier_align (from); |
4466 | else | |
4467 | new_align = 1; | |
4468 | inc = 0; | |
4469 | } | |
4241ecb0 R |
4470 | /* In case we are scanning a constant table because of recursion, check |
4471 | for explicit alignments. If the table is long, we might be forced | |
4472 | to emit the new table in front of it; the length of the alignment | |
4473 | might be the last straw. */ | |
f3536097 | 4474 | else if (NONJUMP_INSN_P (from) |
4241ecb0 R |
4475 | && GET_CODE (PATTERN (from)) == UNSPEC_VOLATILE |
4476 | && XINT (PATTERN (from), 1) == UNSPECV_ALIGN) | |
4477 | new_align = INTVAL (XVECEXP (PATTERN (from), 0, 0)); | |
4478 | /* When we find the end of a constant table, paste the new constant | |
4479 | at the end. That is better than putting it in front because | |
4480 | this way, we don't need extra alignment for adding a 4-byte-aligned | |
4481 | mov(a) label to a 2/4 or 8/4 byte aligned table. */ | |
f3536097 | 4482 | else if (NONJUMP_INSN_P (from) |
4241ecb0 R |
4483 | && GET_CODE (PATTERN (from)) == UNSPEC_VOLATILE |
4484 | && XINT (PATTERN (from), 1) == UNSPECV_CONST_END) | |
4485 | return from; | |
8aa2a305 | 4486 | |
f3536097 | 4487 | if (BARRIER_P (from)) |
1245df60 | 4488 | { |
6d4d15b8 | 4489 | rtx next; |
33f7f353 | 4490 | |
1245df60 | 4491 | found_barrier = from; |
33f7f353 | 4492 | |
956d6950 | 4493 | /* If we are at the end of the function, or in front of an alignment |
1245df60 R |
4494 | instruction, we need not insert an extra alignment. We prefer |
4495 | this kind of barrier. */ | |
33f7f353 | 4496 | if (barrier_align (from) > 2) |
1245df60 | 4497 | good_barrier = from; |
6d4d15b8 KK |
4498 | |
4499 | /* If we are at the end of a hot/cold block, dump the constants | |
4500 | here. */ | |
4501 | next = NEXT_INSN (from); | |
4502 | if (next | |
4503 | && NOTE_P (next) | |
4504 | && NOTE_KIND (next) == NOTE_INSN_SWITCH_TEXT_SECTIONS) | |
4505 | break; | |
1245df60 | 4506 | } |
8aa2a305 | 4507 | |
8aa2a305 | 4508 | if (broken_move (from)) |
0d7e008e | 4509 | { |
1245df60 R |
4510 | rtx pat, src, dst; |
4511 | enum machine_mode mode; | |
4512 | ||
4513 | pat = PATTERN (from); | |
4514 | if (GET_CODE (pat) == PARALLEL) | |
4515 | pat = XVECEXP (pat, 0, 0); | |
4516 | src = SET_SRC (pat); | |
4517 | dst = SET_DEST (pat); | |
4518 | mode = GET_MODE (dst); | |
c17f53a8 | 4519 | |
e938feb4 CB |
4520 | /* GOT pcrelat setting comes in pair of |
4521 | mova .L8,r0 | |
4522 | mov.l .L8,r12 | |
4523 | instructions. (plus add r0,r12). | |
4524 | Remember if we see one without the other. */ | |
4525 | if (GET_CODE (src) == UNSPEC && PIC_ADDR_P (XVECEXP (src, 0, 0))) | |
4526 | last_got = last_got ? NULL_RTX : from; | |
4527 | else if (PIC_ADDR_P (src)) | |
4528 | last_got = last_got ? NULL_RTX : from; | |
4529 | ||
c17f53a8 JW |
4530 | /* We must explicitly check the mode, because sometimes the |
4531 | front end will generate code to load unsigned constants into | |
4532 | HImode targets without properly sign extending them. */ | |
225e4f43 R |
4533 | if (mode == HImode |
4534 | || (mode == SImode && hi_const (src) && REGNO (dst) != FPUL_REG)) | |
00e94d65 | 4535 | { |
1245df60 | 4536 | found_hi += 2; |
00e94d65 RK |
4537 | /* We put the short constants before the long constants, so |
4538 | we must count the length of short constants in the range | |
4539 | for the long constants. */ | |
4540 | /* ??? This isn't optimal, but is easy to do. */ | |
1245df60 | 4541 | si_limit -= 2; |
00e94d65 | 4542 | } |
0d7e008e | 4543 | else |
1245df60 | 4544 | { |
fa5322fa AO |
4545 | /* We dump DF/DI constants before SF/SI ones, because |
4546 | the limit is the same, but the alignment requirements | |
4547 | are higher. We may waste up to 4 additional bytes | |
4548 | for alignment, and the DF/DI constant may have | |
24746a42 | 4549 | another SF/SI constant placed before it. */ |
fa5322fa AO |
4550 | if (TARGET_SHCOMPACT |
4551 | && ! found_di | |
4552 | && (mode == DFmode || mode == DImode)) | |
4553 | { | |
4554 | found_di = 1; | |
4555 | si_limit -= 8; | |
4556 | } | |
33f7f353 JR |
4557 | while (si_align > 2 && found_si + si_align - 2 > count_si) |
4558 | si_align >>= 1; | |
1245df60 R |
4559 | if (found_si > count_si) |
4560 | count_si = found_si; | |
4561 | found_si += GET_MODE_SIZE (mode); | |
4562 | if (num_mova) | |
4563 | si_limit -= GET_MODE_SIZE (mode); | |
4564 | } | |
0d7e008e | 4565 | } |
5325c0fa | 4566 | |
33f7f353 | 4567 | if (mova_p (from)) |
1245df60 | 4568 | { |
4241ecb0 | 4569 | switch (untangle_mova (&num_mova, &mova, from)) |
1245df60 | 4570 | { |
fc9c984d KK |
4571 | case 1: |
4572 | if (flag_pic) | |
4573 | { | |
4574 | rtx src = SET_SRC (PATTERN (from)); | |
4575 | if (GET_CODE (src) == CONST | |
4576 | && GET_CODE (XEXP (src, 0)) == UNSPEC | |
4577 | && XINT (XEXP (src, 0), 1) == UNSPEC_SYMOFF) | |
4578 | last_symoff = from; | |
4579 | } | |
4580 | break; | |
4241ecb0 R |
4581 | case 0: return find_barrier (0, 0, mova); |
4582 | case 2: | |
4583 | { | |
4584 | leading_mova = 0; | |
4585 | barrier_before_mova | |
4586 | = good_barrier ? good_barrier : found_barrier; | |
4587 | } | |
4588 | default: break; | |
1245df60 R |
4589 | } |
4590 | if (found_si > count_si) | |
4591 | count_si = found_si; | |
4592 | } | |
f3536097 | 4593 | else if (JUMP_TABLE_DATA_P (from)) |
1245df60 | 4594 | { |
4241ecb0 R |
4595 | if ((num_mova > 1 && GET_MODE (prev_nonnote_insn (from)) == VOIDmode) |
4596 | || (num_mova | |
4597 | && (prev_nonnote_insn (from) | |
4598 | == XEXP (MOVA_LABELREF (mova), 0)))) | |
1245df60 | 4599 | num_mova--; |
10f4f635 | 4600 | if (barrier_align (next_real_insn (from)) == align_jumps_log) |
1245df60 | 4601 | { |
38e01259 | 4602 | /* We have just passed the barrier in front of the |
a0798779 R |
4603 | ADDR_DIFF_VEC, which is stored in found_barrier. Since |
4604 | the ADDR_DIFF_VEC is accessed as data, just like our pool | |
4605 | constants, this is a good opportunity to accommodate what | |
4606 | we have gathered so far. | |
1245df60 R |
4607 | If we waited any longer, we could end up at a barrier in |
4608 | front of code, which gives worse cache usage for separated | |
4609 | instruction / data caches. */ | |
a0798779 | 4610 | good_barrier = found_barrier; |
1245df60 R |
4611 | break; |
4612 | } | |
a0798779 R |
4613 | else |
4614 | { | |
4615 | rtx body = PATTERN (from); | |
4616 | inc = XVECLEN (body, 1) * GET_MODE_SIZE (GET_MODE (body)); | |
4617 | } | |
1245df60 | 4618 | } |
50ceefc2 | 4619 | /* For the SH1, we generate alignments even after jumps-around-jumps. */ |
f3536097 | 4620 | else if (JUMP_P (from) |
50ceefc2 R |
4621 | && ! TARGET_SH2 |
4622 | && ! TARGET_SMALLCODE) | |
4623 | new_align = 4; | |
5325c0fa | 4624 | |
9e3c399e CB |
4625 | /* There is a possibility that a bf is transformed into a bf/s by the |
4626 | delay slot scheduler. */ | |
4627 | if (JUMP_P (from) && !JUMP_TABLE_DATA_P (from) | |
4628 | && get_attr_type (from) == TYPE_CBRANCH | |
4629 | && GET_CODE (PATTERN (NEXT_INSN (PREV_INSN (from)))) != SEQUENCE) | |
4630 | inc += 2; | |
4631 | ||
0d7e008e | 4632 | if (found_si) |
33f7f353 | 4633 | { |
50ceefc2 | 4634 | count_si += inc; |
33f7f353 JR |
4635 | if (new_align > si_align) |
4636 | { | |
318881c0 | 4637 | si_limit -= (count_si - 1) & (new_align - si_align); |
33f7f353 JR |
4638 | si_align = new_align; |
4639 | } | |
318881c0 | 4640 | count_si = (count_si + new_align - 1) & -new_align; |
33f7f353 | 4641 | } |
0d7e008e | 4642 | if (found_hi) |
33f7f353 | 4643 | { |
50ceefc2 | 4644 | count_hi += inc; |
33f7f353 JR |
4645 | if (new_align > hi_align) |
4646 | { | |
318881c0 | 4647 | hi_limit -= (count_hi - 1) & (new_align - hi_align); |
33f7f353 JR |
4648 | hi_align = new_align; |
4649 | } | |
318881c0 | 4650 | count_hi = (count_hi + new_align - 1) & -new_align; |
33f7f353 | 4651 | } |
0d7e008e SC |
4652 | from = NEXT_INSN (from); |
4653 | } | |
4654 | ||
1245df60 | 4655 | if (num_mova) |
eda44c78 R |
4656 | { |
4657 | if (leading_mova) | |
4658 | { | |
4659 | /* Try as we might, the leading mova is out of range. Change | |
4660 | it into a load (which will become a pcload) and retry. */ | |
078c8b08 | 4661 | fixup_mova (mova); |
eda44c78 R |
4662 | return find_barrier (0, 0, mova); |
4663 | } | |
4664 | else | |
4665 | { | |
4666 | /* Insert the constant pool table before the mova instruction, | |
4667 | to prevent the mova label reference from going out of range. */ | |
4668 | from = mova; | |
4669 | good_barrier = found_barrier = barrier_before_mova; | |
4670 | } | |
4671 | } | |
5325c0fa | 4672 | |
1245df60 R |
4673 | if (found_barrier) |
4674 | { | |
33f7f353 | 4675 | if (good_barrier && next_real_insn (found_barrier)) |
1245df60 | 4676 | found_barrier = good_barrier; |
1245df60 R |
4677 | } |
4678 | else | |
b9654711 | 4679 | { |
8aa2a305 JW |
4680 | /* We didn't find a barrier in time to dump our stuff, |
4681 | so we'll make one. */ | |
0d7e008e | 4682 | rtx label = gen_label_rtx (); |
8aa2a305 | 4683 | |
fc9c984d KK |
4684 | /* Don't emit a constant table in the middle of insns for |
4685 | casesi_worker_2. This is a bit overkill but is enough | |
4686 | because casesi_worker_2 wouldn't appear so frequently. */ | |
4687 | if (last_symoff) | |
4688 | from = last_symoff; | |
4689 | ||
5dbcc9c0 RK |
4690 | /* If we exceeded the range, then we must back up over the last |
4691 | instruction we looked at. Otherwise, we just need to undo the | |
4692 | NEXT_INSN at the end of the loop. */ | |
7a3cb3ab KK |
4693 | if (PREV_INSN (from) != orig |
4694 | && (count_hi > hi_limit || count_si > si_limit)) | |
5dbcc9c0 RK |
4695 | from = PREV_INSN (PREV_INSN (from)); |
4696 | else | |
4697 | from = PREV_INSN (from); | |
4698 | ||
e938feb4 CB |
4699 | /* Don't emit a constant table int the middle of global pointer setting, |
4700 | since that that would move the addressing base GOT into another table. | |
4701 | We need the first mov instruction before the _GLOBAL_OFFSET_TABLE_ | |
4702 | in the pool anyway, so just move up the whole constant pool. */ | |
4703 | if (last_got) | |
4704 | from = PREV_INSN (last_got); | |
4705 | ||
bd55b817 KK |
4706 | /* Don't insert the constant pool table at the position which |
4707 | may be the landing pad. */ | |
4708 | if (flag_exceptions | |
4709 | && CALL_P (from) | |
4710 | && find_reg_note (from, REG_EH_REGION, NULL_RTX)) | |
4711 | from = PREV_INSN (from); | |
4712 | ||
8aa2a305 JW |
4713 | /* Walk back to be just before any jump or label. |
4714 | Putting it before a label reduces the number of times the branch | |
4715 | around the constant pool table will be hit. Putting it before | |
4716 | a jump makes it more likely that the bra delay slot will be | |
4717 | filled. */ | |
f3536097 SZ |
4718 | while (NOTE_P (from) || JUMP_P (from) |
4719 | || LABEL_P (from)) | |
8aa2a305 JW |
4720 | from = PREV_INSN (from); |
4721 | ||
0d7e008e SC |
4722 | from = emit_jump_insn_after (gen_jump (label), from); |
4723 | JUMP_LABEL (from) = label; | |
e081ed26 | 4724 | LABEL_NUSES (label) = 1; |
0d7e008e SC |
4725 | found_barrier = emit_barrier_after (from); |
4726 | emit_label_after (label, found_barrier); | |
b9654711 | 4727 | } |
b9654711 | 4728 | |
8aa2a305 | 4729 | return found_barrier; |
0d7e008e | 4730 | } |
b9654711 | 4731 | |
4787bce0 JW |
4732 | /* If the instruction INSN is implemented by a special function, and we can |
4733 | positively find the register that is used to call the sfunc, and this | |
4734 | register is not used anywhere else in this instruction - except as the | |
4735 | destination of a set, return this register; else, return 0. */ | |
1245df60 | 4736 | rtx |
cf277499 | 4737 | sfunc_uses_reg (rtx insn) |
4787bce0 JW |
4738 | { |
4739 | int i; | |
4740 | rtx pattern, part, reg_part, reg; | |
4741 | ||
f3536097 | 4742 | if (!NONJUMP_INSN_P (insn)) |
4787bce0 JW |
4743 | return 0; |
4744 | pattern = PATTERN (insn); | |
4745 | if (GET_CODE (pattern) != PARALLEL || get_attr_type (insn) != TYPE_SFUNC) | |
4746 | return 0; | |
4747 | ||
4748 | for (reg_part = 0, i = XVECLEN (pattern, 0) - 1; i >= 1; i--) | |
4749 | { | |
4750 | part = XVECEXP (pattern, 0, i); | |
1245df60 | 4751 | if (GET_CODE (part) == USE && GET_MODE (XEXP (part, 0)) == SImode) |
4787bce0 JW |
4752 | reg_part = part; |
4753 | } | |
4754 | if (! reg_part) | |
4755 | return 0; | |
4756 | reg = XEXP (reg_part, 0); | |
4757 | for (i = XVECLEN (pattern, 0) - 1; i >= 0; i--) | |
4758 | { | |
4759 | part = XVECEXP (pattern, 0, i); | |
225e4f43 | 4760 | if (part == reg_part || GET_CODE (part) == CLOBBER) |
4787bce0 JW |
4761 | continue; |
4762 | if (reg_mentioned_p (reg, ((GET_CODE (part) == SET | |
f3536097 | 4763 | && REG_P (SET_DEST (part))) |
4787bce0 JW |
4764 | ? SET_SRC (part) : part))) |
4765 | return 0; | |
4766 | } | |
4767 | return reg; | |
4768 | } | |
4769 | ||
933c3ba3 JW |
4770 | /* See if the only way in which INSN uses REG is by calling it, or by |
4771 | setting it while calling it. Set *SET to a SET rtx if the register | |
4772 | is set by INSN. */ | |
4773 | ||
4774 | static int | |
cf277499 | 4775 | noncall_uses_reg (rtx reg, rtx insn, rtx *set) |
933c3ba3 | 4776 | { |
4787bce0 | 4777 | rtx pattern, reg2; |
933c3ba3 JW |
4778 | |
4779 | *set = NULL_RTX; | |
4780 | ||
4787bce0 JW |
4781 | reg2 = sfunc_uses_reg (insn); |
4782 | if (reg2 && REGNO (reg2) == REGNO (reg)) | |
4783 | { | |
4784 | pattern = single_set (insn); | |
4785 | if (pattern | |
f3536097 | 4786 | && REG_P (SET_DEST (pattern)) |
4787bce0 JW |
4787 | && REGNO (reg) == REGNO (SET_DEST (pattern))) |
4788 | *set = pattern; | |
4789 | return 0; | |
4790 | } | |
f3536097 | 4791 | if (!CALL_P (insn)) |
933c3ba3 JW |
4792 | { |
4793 | /* We don't use rtx_equal_p because we don't care if the mode is | |
4794 | different. */ | |
4795 | pattern = single_set (insn); | |
4796 | if (pattern | |
f3536097 | 4797 | && REG_P (SET_DEST (pattern)) |
933c3ba3 JW |
4798 | && REGNO (reg) == REGNO (SET_DEST (pattern))) |
4799 | { | |
4787bce0 JW |
4800 | rtx par, part; |
4801 | int i; | |
4802 | ||
933c3ba3 | 4803 | *set = pattern; |
4787bce0 JW |
4804 | par = PATTERN (insn); |
4805 | if (GET_CODE (par) == PARALLEL) | |
4806 | for (i = XVECLEN (par, 0) - 1; i >= 0; i--) | |
4807 | { | |
4808 | part = XVECEXP (par, 0, i); | |
4809 | if (GET_CODE (part) != SET && reg_mentioned_p (reg, part)) | |
4810 | return 1; | |
4811 | } | |
4812 | return reg_mentioned_p (reg, SET_SRC (pattern)); | |
933c3ba3 JW |
4813 | } |
4814 | ||
4815 | return 1; | |
4816 | } | |
4817 | ||
4818 | pattern = PATTERN (insn); | |
4819 | ||
4820 | if (GET_CODE (pattern) == PARALLEL) | |
4821 | { | |
4822 | int i; | |
4823 | ||
4824 | for (i = XVECLEN (pattern, 0) - 1; i >= 1; i--) | |
4825 | if (reg_mentioned_p (reg, XVECEXP (pattern, 0, i))) | |
4826 | return 1; | |
4827 | pattern = XVECEXP (pattern, 0, 0); | |
4828 | } | |
4829 | ||
4830 | if (GET_CODE (pattern) == SET) | |
4831 | { | |
4832 | if (reg_mentioned_p (reg, SET_DEST (pattern))) | |
4833 | { | |
4834 | /* We don't use rtx_equal_p, because we don't care if the | |
4835 | mode is different. */ | |
f3536097 | 4836 | if (!REG_P (SET_DEST (pattern)) |
933c3ba3 JW |
4837 | || REGNO (reg) != REGNO (SET_DEST (pattern))) |
4838 | return 1; | |
4839 | ||
4840 | *set = pattern; | |
4841 | } | |
4842 | ||
4843 | pattern = SET_SRC (pattern); | |
4844 | } | |
4845 | ||
4846 | if (GET_CODE (pattern) != CALL | |
f3536097 | 4847 | || !MEM_P (XEXP (pattern, 0)) |
933c3ba3 JW |
4848 | || ! rtx_equal_p (reg, XEXP (XEXP (pattern, 0), 0))) |
4849 | return 1; | |
4850 | ||
4851 | return 0; | |
4852 | } | |
4853 | ||
1245df60 R |
4854 | /* Given a X, a pattern of an insn or a part of it, return a mask of used |
4855 | general registers. Bits 0..15 mean that the respective registers | |
4856 | are used as inputs in the instruction. Bits 16..31 mean that the | |
4857 | registers 0..15, respectively, are used as outputs, or are clobbered. | |
4858 | IS_DEST should be set to 16 if X is the destination of a SET, else to 0. */ | |
4859 | int | |
cf277499 | 4860 | regs_used (rtx x, int is_dest) |
1245df60 R |
4861 | { |
4862 | enum rtx_code code; | |
6f7d635c | 4863 | const char *fmt; |
1245df60 R |
4864 | int i, used = 0; |
4865 | ||
4866 | if (! x) | |
4867 | return used; | |
4868 | code = GET_CODE (x); | |
4869 | switch (code) | |
4870 | { | |
4871 | case REG: | |
4872 | if (REGNO (x) < 16) | |
4873 | return (((1 << HARD_REGNO_NREGS (0, GET_MODE (x))) - 1) | |
4874 | << (REGNO (x) + is_dest)); | |
4875 | return 0; | |
4876 | case SUBREG: | |
4877 | { | |
4878 | rtx y = SUBREG_REG (x); | |
f676971a | 4879 | |
f3536097 | 4880 | if (!REG_P (y)) |
1245df60 R |
4881 | break; |
4882 | if (REGNO (y) < 16) | |
4883 | return (((1 << HARD_REGNO_NREGS (0, GET_MODE (x))) - 1) | |
ddef6bc7 JJ |
4884 | << (REGNO (y) + |
4885 | subreg_regno_offset (REGNO (y), | |
4886 | GET_MODE (y), | |
4887 | SUBREG_BYTE (x), | |
4888 | GET_MODE (x)) + is_dest)); | |
1245df60 R |
4889 | return 0; |
4890 | } | |
4891 | case SET: | |
4892 | return regs_used (SET_SRC (x), 0) | regs_used (SET_DEST (x), 16); | |
4893 | case RETURN: | |
4894 | /* If there was a return value, it must have been indicated with USE. */ | |
4895 | return 0x00ffff00; | |
4896 | case CLOBBER: | |
4897 | is_dest = 1; | |
4898 | break; | |
4899 | case MEM: | |
4900 | is_dest = 0; | |
4901 | break; | |
4902 | case CALL: | |
4903 | used |= 0x00ff00f0; | |
4904 | break; | |
318881c0 KG |
4905 | default: |
4906 | break; | |
1245df60 R |
4907 | } |
4908 | ||
4909 | fmt = GET_RTX_FORMAT (code); | |
4910 | ||
4911 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
4912 | { | |
4913 | if (fmt[i] == 'E') | |
4914 | { | |
4915 | register int j; | |
4916 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
4917 | used |= regs_used (XVECEXP (x, i, j), is_dest); | |
4918 | } | |
4919 | else if (fmt[i] == 'e') | |
4920 | used |= regs_used (XEXP (x, i), is_dest); | |
4921 | } | |
4922 | return used; | |
4923 | } | |
4924 | ||
4925 | /* Create an instruction that prevents redirection of a conditional branch | |
956d6950 | 4926 | to the destination of the JUMP with address ADDR. |
1245df60 R |
4927 | If the branch needs to be implemented as an indirect jump, try to find |
4928 | a scratch register for it. | |
4929 | If NEED_BLOCK is 0, don't do anything unless we need a scratch register. | |
4930 | If any preceding insn that doesn't fit into a delay slot is good enough, | |
4931 | pass 1. Pass 2 if a definite blocking insn is needed. | |
4932 | -1 is used internally to avoid deep recursion. | |
4933 | If a blocking instruction is made or recognized, return it. */ | |
f676971a | 4934 | |
1245df60 | 4935 | static rtx |
cf277499 | 4936 | gen_block_redirect (rtx jump, int addr, int need_block) |
1245df60 R |
4937 | { |
4938 | int dead = 0; | |
4939 | rtx prev = prev_nonnote_insn (jump); | |
4940 | rtx dest; | |
4941 | ||
4942 | /* First, check if we already have an instruction that satisfies our need. */ | |
f3536097 | 4943 | if (prev && NONJUMP_INSN_P (prev) && ! INSN_DELETED_P (prev)) |
1245df60 R |
4944 | { |
4945 | if (INSN_CODE (prev) == CODE_FOR_indirect_jump_scratch) | |
4946 | return prev; | |
4947 | if (GET_CODE (PATTERN (prev)) == USE | |
4948 | || GET_CODE (PATTERN (prev)) == CLOBBER | |
4949 | || get_attr_in_delay_slot (prev) == IN_DELAY_SLOT_YES) | |
4950 | prev = jump; | |
4951 | else if ((need_block &= ~1) < 0) | |
4952 | return prev; | |
4953 | else if (recog_memoized (prev) == CODE_FOR_block_branch_redirect) | |
4954 | need_block = 0; | |
4955 | } | |
f9654065 R |
4956 | if (GET_CODE (PATTERN (jump)) == RETURN) |
4957 | { | |
4958 | if (! need_block) | |
4959 | return prev; | |
4960 | /* Reorg even does nasty things with return insns that cause branches | |
4961 | to go out of range - see find_end_label and callers. */ | |
a556fd39 | 4962 | return emit_insn_before (gen_block_branch_redirect (const0_rtx) , jump); |
f9654065 | 4963 | } |
1245df60 R |
4964 | /* We can't use JUMP_LABEL here because it might be undefined |
4965 | when not optimizing. */ | |
4966 | dest = XEXP (SET_SRC (PATTERN (jump)), 0); | |
4967 | /* If the branch is out of range, try to find a scratch register for it. */ | |
4968 | if (optimize | |
11f9ed1a KG |
4969 | && (INSN_ADDRESSES (INSN_UID (dest)) - addr + (unsigned) 4092 |
4970 | > 4092 + 4098)) | |
1245df60 R |
4971 | { |
4972 | rtx scan; | |
4973 | /* Don't look for the stack pointer as a scratch register, | |
956d6950 | 4974 | it would cause trouble if an interrupt occurred. */ |
0a2aaacc | 4975 | unsigned attempt = 0x7fff, used; |
1245df60 | 4976 | int jump_left = flag_expensive_optimizations + 1; |
f676971a | 4977 | |
1245df60 R |
4978 | /* It is likely that the most recent eligible instruction is wanted for |
4979 | the delay slot. Therefore, find out which registers it uses, and | |
4980 | try to avoid using them. */ | |
f676971a | 4981 | |
318881c0 | 4982 | for (scan = jump; (scan = PREV_INSN (scan)); ) |
1245df60 R |
4983 | { |
4984 | enum rtx_code code; | |
4985 | ||
4986 | if (INSN_DELETED_P (scan)) | |
4987 | continue; | |
4988 | code = GET_CODE (scan); | |
4989 | if (code == CODE_LABEL || code == JUMP_INSN) | |
4990 | break; | |
4991 | if (code == INSN | |
4992 | && GET_CODE (PATTERN (scan)) != USE | |
4993 | && GET_CODE (PATTERN (scan)) != CLOBBER | |
4994 | && get_attr_in_delay_slot (scan) == IN_DELAY_SLOT_YES) | |
4995 | { | |
0a2aaacc | 4996 | attempt &= ~regs_used (PATTERN (scan), 0); |
1245df60 R |
4997 | break; |
4998 | } | |
4999 | } | |
eda44c78 R |
5000 | for (used = dead = 0, scan = JUMP_LABEL (jump); |
5001 | (scan = NEXT_INSN (scan)); ) | |
1245df60 R |
5002 | { |
5003 | enum rtx_code code; | |
5004 | ||
5005 | if (INSN_DELETED_P (scan)) | |
5006 | continue; | |
5007 | code = GET_CODE (scan); | |
ec8e098d | 5008 | if (INSN_P (scan)) |
1245df60 R |
5009 | { |
5010 | used |= regs_used (PATTERN (scan), 0); | |
5011 | if (code == CALL_INSN) | |
5012 | used |= regs_used (CALL_INSN_FUNCTION_USAGE (scan), 0); | |
5013 | dead |= (used >> 16) & ~used; | |
0a2aaacc | 5014 | if (dead & attempt) |
1245df60 | 5015 | { |
0a2aaacc | 5016 | dead &= attempt; |
1245df60 R |
5017 | break; |
5018 | } | |
5019 | if (code == JUMP_INSN) | |
eda44c78 R |
5020 | { |
5021 | if (jump_left-- && simplejump_p (scan)) | |
5022 | scan = JUMP_LABEL (scan); | |
5023 | else | |
5024 | break; | |
5025 | } | |
1245df60 R |
5026 | } |
5027 | } | |
5028 | /* Mask out the stack pointer again, in case it was | |
5029 | the only 'free' register we have found. */ | |
5030 | dead &= 0x7fff; | |
5031 | } | |
5032 | /* If the immediate destination is still in range, check for possible | |
5033 | threading with a jump beyond the delay slot insn. | |
5034 | Don't check if we are called recursively; the jump has been or will be | |
956d6950 | 5035 | checked in a different invocation then. */ |
f676971a | 5036 | |
1245df60 R |
5037 | else if (optimize && need_block >= 0) |
5038 | { | |
5039 | rtx next = next_active_insn (next_active_insn (dest)); | |
f3536097 | 5040 | if (next && JUMP_P (next) |
1245df60 | 5041 | && GET_CODE (PATTERN (next)) == SET |
ddf8ec7e | 5042 | && recog_memoized (next) == CODE_FOR_jump_compact) |
1245df60 R |
5043 | { |
5044 | dest = JUMP_LABEL (next); | |
5045 | if (dest | |
11f9ed1a | 5046 | && (INSN_ADDRESSES (INSN_UID (dest)) - addr + (unsigned) 4092 |
9d98a694 AO |
5047 | > 4092 + 4098)) |
5048 | gen_block_redirect (next, INSN_ADDRESSES (INSN_UID (next)), -1); | |
1245df60 R |
5049 | } |
5050 | } | |
5051 | ||
5052 | if (dead) | |
5053 | { | |
c5c76735 | 5054 | rtx reg = gen_rtx_REG (SImode, exact_log2 (dead & -dead)); |
1245df60 R |
5055 | |
5056 | /* It would be nice if we could convert the jump into an indirect | |
956d6950 | 5057 | jump / far branch right now, and thus exposing all constituent |
1245df60 R |
5058 | instructions to further optimization. However, reorg uses |
5059 | simplejump_p to determine if there is an unconditional jump where | |
5060 | it should try to schedule instructions from the target of the | |
5061 | branch; simplejump_p fails for indirect jumps even if they have | |
5062 | a JUMP_LABEL. */ | |
5063 | rtx insn = emit_insn_before (gen_indirect_jump_scratch | |
826159e5 KK |
5064 | (reg, GEN_INT (unspec_bbr_uid++)), |
5065 | jump); | |
10f4f635 R |
5066 | /* ??? We would like this to have the scope of the jump, but that |
5067 | scope will change when a delay slot insn of an inner scope is added. | |
5068 | Hence, after delay slot scheduling, we'll have to expect | |
5069 | NOTE_INSN_BLOCK_END notes between the indirect_jump_scratch and | |
5070 | the jump. */ | |
f676971a | 5071 | |
86143cb4 | 5072 | INSN_LOCATOR (insn) = INSN_LOCATOR (jump); |
1245df60 R |
5073 | INSN_CODE (insn) = CODE_FOR_indirect_jump_scratch; |
5074 | return insn; | |
5075 | } | |
5076 | else if (need_block) | |
5077 | /* We can't use JUMP_LABEL here because it might be undefined | |
5078 | when not optimizing. */ | |
5079 | return emit_insn_before (gen_block_branch_redirect | |
826159e5 KK |
5080 | (GEN_INT (unspec_bbr_uid++)), |
5081 | jump); | |
1245df60 R |
5082 | return prev; |
5083 | } | |
5084 | ||
5085 | #define CONDJUMP_MIN -252 | |
5086 | #define CONDJUMP_MAX 262 | |
5087 | struct far_branch | |
5088 | { | |
5089 | /* A label (to be placed) in front of the jump | |
5090 | that jumps to our ultimate destination. */ | |
5091 | rtx near_label; | |
5092 | /* Where we are going to insert it if we cannot move the jump any farther, | |
5093 | or the jump itself if we have picked up an existing jump. */ | |
5094 | rtx insert_place; | |
5095 | /* The ultimate destination. */ | |
5096 | rtx far_label; | |
5097 | struct far_branch *prev; | |
5098 | /* If the branch has already been created, its address; | |
5099 | else the address of its first prospective user. */ | |
5100 | int address; | |
5101 | }; | |
5102 | ||
cf277499 | 5103 | static void gen_far_branch (struct far_branch *); |
1245df60 | 5104 | enum mdep_reorg_phase_e mdep_reorg_phase; |
c5b9ef02 | 5105 | static void |
cf277499 | 5106 | gen_far_branch (struct far_branch *bp) |
1245df60 R |
5107 | { |
5108 | rtx insn = bp->insert_place; | |
5109 | rtx jump; | |
5110 | rtx label = gen_label_rtx (); | |
f5b9e7c9 | 5111 | int ok; |
1245df60 R |
5112 | |
5113 | emit_label_after (label, insn); | |
5114 | if (bp->far_label) | |
5115 | { | |
5116 | jump = emit_jump_insn_after (gen_jump (bp->far_label), insn); | |
5117 | LABEL_NUSES (bp->far_label)++; | |
5118 | } | |
5119 | else | |
5120 | jump = emit_jump_insn_after (gen_return (), insn); | |
225e4f43 R |
5121 | /* Emit a barrier so that reorg knows that any following instructions |
5122 | are not reachable via a fall-through path. | |
fae778eb | 5123 | But don't do this when not optimizing, since we wouldn't suppress the |
225e4f43 R |
5124 | alignment for the barrier then, and could end up with out-of-range |
5125 | pc-relative loads. */ | |
5126 | if (optimize) | |
5127 | emit_barrier_after (jump); | |
1245df60 R |
5128 | emit_label_after (bp->near_label, insn); |
5129 | JUMP_LABEL (jump) = bp->far_label; | |
f5b9e7c9 NS |
5130 | ok = invert_jump (insn, label, 1); |
5131 | gcc_assert (ok); | |
5132 | ||
f9654065 R |
5133 | /* If we are branching around a jump (rather than a return), prevent |
5134 | reorg from using an insn from the jump target as the delay slot insn - | |
5135 | when reorg did this, it pessimized code (we rather hide the delay slot) | |
5136 | and it could cause branches to go out of range. */ | |
5137 | if (bp->far_label) | |
5138 | (emit_insn_after | |
5139 | (gen_stuff_delay_slot | |
826159e5 | 5140 | (GEN_INT (unspec_bbr_uid++), |
f9654065 R |
5141 | GEN_INT (recog_memoized (insn) == CODE_FOR_branch_false)), |
5142 | insn)); | |
1245df60 R |
5143 | /* Prevent reorg from undoing our splits. */ |
5144 | gen_block_redirect (jump, bp->address += 2, 2); | |
5145 | } | |
5146 | ||
1245df60 R |
5147 | /* Fix up ADDR_DIFF_VECs. */ |
5148 | void | |
cf277499 | 5149 | fixup_addr_diff_vecs (rtx first) |
1245df60 R |
5150 | { |
5151 | rtx insn; | |
33f7f353 | 5152 | |
1245df60 R |
5153 | for (insn = first; insn; insn = NEXT_INSN (insn)) |
5154 | { | |
eb3881bf | 5155 | rtx vec_lab, pat, prev, prevpat, x, braf_label; |
1245df60 | 5156 | |
f3536097 | 5157 | if (!JUMP_P (insn) |
1245df60 R |
5158 | || GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC) |
5159 | continue; | |
5160 | pat = PATTERN (insn); | |
33f7f353 | 5161 | vec_lab = XEXP (XEXP (pat, 0), 0); |
1245df60 | 5162 | |
33f7f353 JR |
5163 | /* Search the matching casesi_jump_2. */ |
5164 | for (prev = vec_lab; ; prev = PREV_INSN (prev)) | |
1245df60 | 5165 | { |
f3536097 | 5166 | if (!JUMP_P (prev)) |
33f7f353 JR |
5167 | continue; |
5168 | prevpat = PATTERN (prev); | |
5169 | if (GET_CODE (prevpat) != PARALLEL || XVECLEN (prevpat, 0) != 2) | |
5170 | continue; | |
5171 | x = XVECEXP (prevpat, 0, 1); | |
5172 | if (GET_CODE (x) != USE) | |
5173 | continue; | |
5174 | x = XEXP (x, 0); | |
5175 | if (GET_CODE (x) == LABEL_REF && XEXP (x, 0) == vec_lab) | |
5176 | break; | |
1245df60 | 5177 | } |
157371cf AO |
5178 | /* FIXME: This is a bug in the optimizer, but it seems harmless |
5179 | to just avoid panicing. */ | |
5180 | if (!prev) | |
5181 | continue; | |
eb3881bf R |
5182 | |
5183 | /* Emit the reference label of the braf where it belongs, right after | |
5184 | the casesi_jump_2 (i.e. braf). */ | |
5185 | braf_label = XEXP (XEXP (SET_SRC (XVECEXP (prevpat, 0, 0)), 1), 0); | |
5186 | emit_label_after (braf_label, prev); | |
5187 | ||
33f7f353 JR |
5188 | /* Fix up the ADDR_DIF_VEC to be relative |
5189 | to the reference address of the braf. */ | |
eb3881bf | 5190 | XEXP (XEXP (pat, 0), 0) = braf_label; |
1245df60 | 5191 | } |
1245df60 R |
5192 | } |
5193 | ||
33f7f353 JR |
5194 | /* BARRIER_OR_LABEL is either a BARRIER or a CODE_LABEL immediately following |
5195 | a barrier. Return the base 2 logarithm of the desired alignment. */ | |
1245df60 | 5196 | int |
cf277499 | 5197 | barrier_align (rtx barrier_or_label) |
1245df60 | 5198 | { |
33f7f353 | 5199 | rtx next = next_real_insn (barrier_or_label), pat, prev; |
3a8699c7 | 5200 | int slot, credit, jump_to_next = 0; |
f676971a | 5201 | |
33f7f353 | 5202 | if (! next) |
1245df60 | 5203 | return 0; |
1245df60 | 5204 | |
33f7f353 | 5205 | pat = PATTERN (next); |
1245df60 | 5206 | |
33f7f353 JR |
5207 | if (GET_CODE (pat) == ADDR_DIFF_VEC) |
5208 | return 2; | |
1245df60 | 5209 | |
4773afa4 | 5210 | if (GET_CODE (pat) == UNSPEC_VOLATILE && XINT (pat, 1) == UNSPECV_ALIGN) |
33f7f353 JR |
5211 | /* This is a barrier in front of a constant table. */ |
5212 | return 0; | |
5213 | ||
5214 | prev = prev_real_insn (barrier_or_label); | |
5215 | if (GET_CODE (PATTERN (prev)) == ADDR_DIFF_VEC) | |
1245df60 | 5216 | { |
33f7f353 JR |
5217 | pat = PATTERN (prev); |
5218 | /* If this is a very small table, we want to keep the alignment after | |
5219 | the table to the minimum for proper code alignment. */ | |
5220 | return ((TARGET_SMALLCODE | |
d042370e | 5221 | || ((unsigned) XVECLEN (pat, 1) * GET_MODE_SIZE (GET_MODE (pat)) |
832a3292 | 5222 | <= (unsigned) 1 << (CACHE_LOG - 2))) |
10f4f635 | 5223 | ? 1 << TARGET_SHMEDIA : align_jumps_log); |
1245df60 | 5224 | } |
33f7f353 JR |
5225 | |
5226 | if (TARGET_SMALLCODE) | |
5227 | return 0; | |
5228 | ||
4d070fd3 | 5229 | if (! TARGET_SH2 || ! optimize) |
10f4f635 | 5230 | return align_jumps_log; |
33f7f353 | 5231 | |
225e4f43 R |
5232 | /* When fixing up pcloads, a constant table might be inserted just before |
5233 | the basic block that ends with the barrier. Thus, we can't trust the | |
5234 | instruction lengths before that. */ | |
5235 | if (mdep_reorg_phase > SH_FIXUP_PCLOAD) | |
1245df60 | 5236 | { |
225e4f43 R |
5237 | /* Check if there is an immediately preceding branch to the insn beyond |
5238 | the barrier. We must weight the cost of discarding useful information | |
5239 | from the current cache line when executing this branch and there is | |
5240 | an alignment, against that of fetching unneeded insn in front of the | |
5241 | branch target when there is no alignment. */ | |
5242 | ||
f676971a EC |
5243 | /* There are two delay_slot cases to consider. One is the simple case |
5244 | where the preceding branch is to the insn beyond the barrier (simple | |
5245 | delay slot filling), and the other is where the preceding branch has | |
5246 | a delay slot that is a duplicate of the insn after the barrier | |
5247 | (fill_eager_delay_slots) and the branch is to the insn after the insn | |
b1b41fe4 CP |
5248 | after the barrier. */ |
5249 | ||
225e4f43 R |
5250 | /* PREV is presumed to be the JUMP_INSN for the barrier under |
5251 | investigation. Skip to the insn before it. */ | |
5252 | prev = prev_real_insn (prev); | |
5253 | ||
c5b9ef02 | 5254 | for (slot = 2, credit = (1 << (CACHE_LOG - 2)) + 2; |
f3536097 | 5255 | credit >= 0 && prev && NONJUMP_INSN_P (prev); |
225e4f43 R |
5256 | prev = prev_real_insn (prev)) |
5257 | { | |
b1b41fe4 | 5258 | jump_to_next = 0; |
225e4f43 R |
5259 | if (GET_CODE (PATTERN (prev)) == USE |
5260 | || GET_CODE (PATTERN (prev)) == CLOBBER) | |
5261 | continue; | |
5262 | if (GET_CODE (PATTERN (prev)) == SEQUENCE) | |
b1b41fe4 CP |
5263 | { |
5264 | prev = XVECEXP (PATTERN (prev), 0, 1); | |
f676971a | 5265 | if (INSN_UID (prev) == INSN_UID (next)) |
b1b41fe4 CP |
5266 | { |
5267 | /* Delay slot was filled with insn at jump target. */ | |
5268 | jump_to_next = 1; | |
5269 | continue; | |
5270 | } | |
5271 | } | |
5272 | ||
225e4f43 R |
5273 | if (slot && |
5274 | get_attr_in_delay_slot (prev) == IN_DELAY_SLOT_YES) | |
5275 | slot = 0; | |
5276 | credit -= get_attr_length (prev); | |
5277 | } | |
5278 | if (prev | |
f3536097 | 5279 | && JUMP_P (prev) |
2b9d17c7 AH |
5280 | && JUMP_LABEL (prev)) |
5281 | { | |
5282 | rtx x; | |
300c5453 | 5283 | if (jump_to_next |
2b9d17c7 | 5284 | || next_real_insn (JUMP_LABEL (prev)) == next |
1594c6bc AO |
5285 | /* If relax_delay_slots() decides NEXT was redundant |
5286 | with some previous instruction, it will have | |
5287 | redirected PREV's jump to the following insn. */ | |
5d291213 | 5288 | || JUMP_LABEL (prev) == next_nonnote_insn (next) |
300c5453 AO |
5289 | /* There is no upper bound on redundant instructions |
5290 | that might have been skipped, but we must not put an | |
2b9d17c7 | 5291 | alignment where none had been before. */ |
f676971a EC |
5292 | || (x = (NEXT_INSN (NEXT_INSN (PREV_INSN (prev)))), |
5293 | (INSN_P (x) | |
2b9d17c7 | 5294 | && (INSN_CODE (x) == CODE_FOR_block_branch_redirect |
9959c8c7 R |
5295 | || INSN_CODE (x) == CODE_FOR_indirect_jump_scratch |
5296 | || INSN_CODE (x) == CODE_FOR_stuff_delay_slot)))) | |
2b9d17c7 AH |
5297 | { |
5298 | rtx pat = PATTERN (prev); | |
5299 | if (GET_CODE (pat) == PARALLEL) | |
300c5453 | 5300 | pat = XVECEXP (pat, 0, 0); |
2b9d17c7 AH |
5301 | if (credit - slot >= (GET_CODE (SET_SRC (pat)) == PC ? 2 : 0)) |
5302 | return 0; | |
5303 | } | |
300c5453 | 5304 | } |
33f7f353 | 5305 | } |
f676971a | 5306 | |
10f4f635 | 5307 | return align_jumps_log; |
1245df60 R |
5308 | } |
5309 | ||
2148624a R |
5310 | /* If we are inside a phony loop, almost any kind of label can turn up as the |
5311 | first one in the loop. Aligning a braf label causes incorrect switch | |
5312 | destination addresses; we can detect braf labels because they are | |
5313 | followed by a BARRIER. | |
5314 | Applying loop alignment to small constant or switch tables is a waste | |
5315 | of space, so we suppress this too. */ | |
5316 | int | |
cf277499 | 5317 | sh_loop_align (rtx label) |
2148624a R |
5318 | { |
5319 | rtx next = label; | |
5320 | ||
5321 | do | |
5322 | next = next_nonnote_insn (next); | |
f3536097 | 5323 | while (next && LABEL_P (next)); |
2148624a R |
5324 | |
5325 | if (! next | |
2c3c49de | 5326 | || ! INSN_P (next) |
2148624a R |
5327 | || GET_CODE (PATTERN (next)) == ADDR_DIFF_VEC |
5328 | || recog_memoized (next) == CODE_FOR_consttable_2) | |
5329 | return 0; | |
fa5322fa | 5330 | |
10f4f635 | 5331 | return align_loops_log; |
2148624a R |
5332 | } |
5333 | ||
18dbd950 | 5334 | /* Do a final pass over the function, just before delayed branch |
933c3ba3 | 5335 | scheduling. */ |
b9654711 | 5336 | |
18dbd950 | 5337 | static void |
cf277499 | 5338 | sh_reorg (void) |
0d7e008e | 5339 | { |
18dbd950 | 5340 | rtx first, insn, mova = NULL_RTX; |
1245df60 | 5341 | int num_mova; |
c5c76735 JL |
5342 | rtx r0_rtx = gen_rtx_REG (Pmode, 0); |
5343 | rtx r0_inc_rtx = gen_rtx_POST_INC (Pmode, r0_rtx); | |
8aa2a305 | 5344 | |
18dbd950 | 5345 | first = get_insns (); |
4241ecb0 | 5346 | max_labelno_before_reorg = max_label_num (); |
18dbd950 | 5347 | |
cb51ecd2 AO |
5348 | /* We must split call insns before introducing `mova's. If we're |
5349 | optimizing, they'll have already been split. Otherwise, make | |
5350 | sure we don't split them too late. */ | |
5351 | if (! optimize) | |
6f862f2f | 5352 | split_all_insns_noflow (); |
cb51ecd2 | 5353 | |
fa5322fa AO |
5354 | if (TARGET_SHMEDIA) |
5355 | return; | |
5356 | ||
933c3ba3 JW |
5357 | /* If relaxing, generate pseudo-ops to associate function calls with |
5358 | the symbols they call. It does no harm to not generate these | |
5359 | pseudo-ops. However, when we can generate them, it enables to | |
5360 | linker to potentially relax the jsr to a bsr, and eliminate the | |
5361 | register load and, possibly, the constant pool entry. */ | |
5362 | ||
1245df60 | 5363 | mdep_reorg_phase = SH_INSERT_USES_LABELS; |
933c3ba3 JW |
5364 | if (TARGET_RELAX) |
5365 | { | |
cf7c4aa6 HPN |
5366 | /* Remove all REG_LABEL_OPERAND notes. We want to use them for our |
5367 | own purposes. This works because none of the remaining passes | |
933c3ba3 JW |
5368 | need to look at them. |
5369 | ||
5370 | ??? But it may break in the future. We should use a machine | |
5371 | dependent REG_NOTE, or some other approach entirely. */ | |
5372 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
5373 | { | |
2c3c49de | 5374 | if (INSN_P (insn)) |
933c3ba3 JW |
5375 | { |
5376 | rtx note; | |
5377 | ||
cf7c4aa6 HPN |
5378 | while ((note = find_reg_note (insn, REG_LABEL_OPERAND, |
5379 | NULL_RTX)) != 0) | |
933c3ba3 JW |
5380 | remove_note (insn, note); |
5381 | } | |
5382 | } | |
5383 | ||
5384 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
5385 | { | |
5386 | rtx pattern, reg, link, set, scan, dies, label; | |
5387 | int rescan = 0, foundinsn = 0; | |
5388 | ||
f3536097 | 5389 | if (CALL_P (insn)) |
4787bce0 JW |
5390 | { |
5391 | pattern = PATTERN (insn); | |
933c3ba3 | 5392 | |
4787bce0 JW |
5393 | if (GET_CODE (pattern) == PARALLEL) |
5394 | pattern = XVECEXP (pattern, 0, 0); | |
5395 | if (GET_CODE (pattern) == SET) | |
5396 | pattern = SET_SRC (pattern); | |
933c3ba3 | 5397 | |
4787bce0 | 5398 | if (GET_CODE (pattern) != CALL |
f3536097 | 5399 | || !MEM_P (XEXP (pattern, 0))) |
4787bce0 | 5400 | continue; |
933c3ba3 | 5401 | |
4787bce0 JW |
5402 | reg = XEXP (XEXP (pattern, 0), 0); |
5403 | } | |
5404 | else | |
5405 | { | |
5406 | reg = sfunc_uses_reg (insn); | |
5407 | if (! reg) | |
5408 | continue; | |
5409 | } | |
933c3ba3 | 5410 | |
f3536097 | 5411 | if (!REG_P (reg)) |
933c3ba3 JW |
5412 | continue; |
5413 | ||
6fb5fa3c DB |
5414 | /* Try scanning backward to find where the register is set. */ |
5415 | link = NULL; | |
5416 | for (scan = PREV_INSN (insn); | |
f3536097 | 5417 | scan && !LABEL_P (scan); |
6fb5fa3c | 5418 | scan = PREV_INSN (scan)) |
933c3ba3 | 5419 | { |
6fb5fa3c DB |
5420 | if (! INSN_P (scan)) |
5421 | continue; | |
933c3ba3 | 5422 | |
6fb5fa3c DB |
5423 | if (! reg_mentioned_p (reg, scan)) |
5424 | continue; | |
933c3ba3 | 5425 | |
6fb5fa3c DB |
5426 | if (noncall_uses_reg (reg, scan, &set)) |
5427 | break; | |
933c3ba3 | 5428 | |
6fb5fa3c DB |
5429 | if (set) |
5430 | { | |
5431 | link = scan; | |
5432 | break; | |
933c3ba3 JW |
5433 | } |
5434 | } | |
5435 | ||
5436 | if (! link) | |
5437 | continue; | |
5438 | ||
5439 | /* The register is set at LINK. */ | |
5440 | ||
5441 | /* We can only optimize the function call if the register is | |
5442 | being set to a symbol. In theory, we could sometimes | |
5443 | optimize calls to a constant location, but the assembler | |
5444 | and linker do not support that at present. */ | |
5445 | if (GET_CODE (SET_SRC (set)) != SYMBOL_REF | |
5446 | && GET_CODE (SET_SRC (set)) != LABEL_REF) | |
5447 | continue; | |
5448 | ||
5449 | /* Scan forward from LINK to the place where REG dies, and | |
5450 | make sure that the only insns which use REG are | |
5451 | themselves function calls. */ | |
5452 | ||
c1e8e6bd JW |
5453 | /* ??? This doesn't work for call targets that were allocated |
5454 | by reload, since there may not be a REG_DEAD note for the | |
5455 | register. */ | |
5456 | ||
933c3ba3 JW |
5457 | dies = NULL_RTX; |
5458 | for (scan = NEXT_INSN (link); scan; scan = NEXT_INSN (scan)) | |
5459 | { | |
5460 | rtx scanset; | |
5461 | ||
c1e8e6bd JW |
5462 | /* Don't try to trace forward past a CODE_LABEL if we haven't |
5463 | seen INSN yet. Ordinarily, we will only find the setting insn | |
6fb5fa3c | 5464 | if it is in the same basic block. However, |
c1e8e6bd JW |
5465 | cross-jumping can insert code labels in between the load and |
5466 | the call, and can result in situations where a single call | |
5467 | insn may have two targets depending on where we came from. */ | |
5468 | ||
f3536097 | 5469 | if (LABEL_P (scan) && ! foundinsn) |
c1e8e6bd JW |
5470 | break; |
5471 | ||
2c3c49de | 5472 | if (! INSN_P (scan)) |
933c3ba3 JW |
5473 | continue; |
5474 | ||
5475 | /* Don't try to trace forward past a JUMP. To optimize | |
5476 | safely, we would have to check that all the | |
c1e8e6bd | 5477 | instructions at the jump destination did not use REG. */ |
933c3ba3 | 5478 | |
f3536097 | 5479 | if (JUMP_P (scan)) |
933c3ba3 JW |
5480 | break; |
5481 | ||
5482 | if (! reg_mentioned_p (reg, scan)) | |
5483 | continue; | |
5484 | ||
5485 | if (noncall_uses_reg (reg, scan, &scanset)) | |
5486 | break; | |
5487 | ||
5488 | if (scan == insn) | |
5489 | foundinsn = 1; | |
5490 | ||
4787bce0 | 5491 | if (scan != insn |
f3536097 | 5492 | && (CALL_P (scan) || sfunc_uses_reg (scan))) |
933c3ba3 JW |
5493 | { |
5494 | /* There is a function call to this register other | |
5495 | than the one we are checking. If we optimize | |
5496 | this call, we need to rescan again below. */ | |
5497 | rescan = 1; | |
5498 | } | |
5499 | ||
5500 | /* ??? We shouldn't have to worry about SCANSET here. | |
5501 | We should just be able to check for a REG_DEAD note | |
5502 | on a function call. However, the REG_DEAD notes are | |
5503 | apparently not dependable around libcalls; c-torture | |
5504 | execute/920501-2 is a test case. If SCANSET is set, | |
5505 | then this insn sets the register, so it must have | |
5506 | died earlier. Unfortunately, this will only handle | |
5507 | the cases in which the register is, in fact, set in a | |
5508 | later insn. */ | |
5509 | ||
5510 | /* ??? We shouldn't have to use FOUNDINSN here. | |
6fb5fa3c DB |
5511 | This dates back to when we used LOG_LINKS to find |
5512 | the most recent insn which sets the register. */ | |
933c3ba3 JW |
5513 | |
5514 | if (foundinsn | |
5515 | && (scanset | |
5516 | || find_reg_note (scan, REG_DEAD, reg))) | |
5517 | { | |
5518 | dies = scan; | |
5519 | break; | |
5520 | } | |
5521 | } | |
5522 | ||
5523 | if (! dies) | |
5524 | { | |
5525 | /* Either there was a branch, or some insn used REG | |
5526 | other than as a function call address. */ | |
5527 | continue; | |
5528 | } | |
5529 | ||
cf7c4aa6 HPN |
5530 | /* Create a code label, and put it in a REG_LABEL_OPERAND note |
5531 | on the insn which sets the register, and on each call insn | |
5532 | which uses the register. In final_prescan_insn we look for | |
5533 | the REG_LABEL_OPERAND notes, and output the appropriate label | |
933c3ba3 JW |
5534 | or pseudo-op. */ |
5535 | ||
5536 | label = gen_label_rtx (); | |
65c5f2a6 ILT |
5537 | add_reg_note (link, REG_LABEL_OPERAND, label); |
5538 | add_reg_note (insn, REG_LABEL_OPERAND, label); | |
933c3ba3 JW |
5539 | if (rescan) |
5540 | { | |
5541 | scan = link; | |
5542 | do | |
5543 | { | |
4787bce0 JW |
5544 | rtx reg2; |
5545 | ||
933c3ba3 JW |
5546 | scan = NEXT_INSN (scan); |
5547 | if (scan != insn | |
f3536097 | 5548 | && ((CALL_P (scan) |
4787bce0 JW |
5549 | && reg_mentioned_p (reg, scan)) |
5550 | || ((reg2 = sfunc_uses_reg (scan)) | |
5551 | && REGNO (reg2) == REGNO (reg)))) | |
65c5f2a6 | 5552 | add_reg_note (scan, REG_LABEL_OPERAND, label); |
933c3ba3 JW |
5553 | } |
5554 | while (scan != dies); | |
5555 | } | |
5556 | } | |
5557 | } | |
5558 | ||
33f7f353 JR |
5559 | if (TARGET_SH2) |
5560 | fixup_addr_diff_vecs (first); | |
1245df60 R |
5561 | |
5562 | if (optimize) | |
5563 | { | |
1245df60 R |
5564 | mdep_reorg_phase = SH_SHORTEN_BRANCHES0; |
5565 | shorten_branches (first); | |
5566 | } | |
6dd3c0a5 | 5567 | |
933c3ba3 JW |
5568 | /* Scan the function looking for move instructions which have to be |
5569 | changed to pc-relative loads and insert the literal tables. */ | |
6dd3c0a5 SB |
5570 | label_ref_list_pool = create_alloc_pool ("label references list", |
5571 | sizeof (struct label_ref_list_d), | |
5572 | 30); | |
1245df60 R |
5573 | mdep_reorg_phase = SH_FIXUP_PCLOAD; |
5574 | for (insn = first, num_mova = 0; insn; insn = NEXT_INSN (insn)) | |
0d7e008e | 5575 | { |
1245df60 R |
5576 | if (mova_p (insn)) |
5577 | { | |
078c8b08 R |
5578 | /* ??? basic block reordering can move a switch table dispatch |
5579 | below the switch table. Check if that has happened. | |
5580 | We only have the addresses available when optimizing; but then, | |
5581 | this check shouldn't be needed when not optimizing. */ | |
4241ecb0 | 5582 | if (!untangle_mova (&num_mova, &mova, insn)) |
078c8b08 | 5583 | { |
4241ecb0 R |
5584 | insn = mova; |
5585 | num_mova = 0; | |
078c8b08 | 5586 | } |
1245df60 | 5587 | } |
f3536097 | 5588 | else if (JUMP_P (insn) |
1245df60 | 5589 | && GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC |
4241ecb0 R |
5590 | && num_mova |
5591 | /* ??? loop invariant motion can also move a mova out of a | |
5592 | loop. Since loop does this code motion anyway, maybe we | |
5593 | should wrap UNSPEC_MOVA into a CONST, so that reload can | |
5594 | move it back. */ | |
5595 | && ((num_mova > 1 | |
5596 | && GET_MODE (prev_nonnote_insn (insn)) == VOIDmode) | |
5597 | || (prev_nonnote_insn (insn) | |
5598 | == XEXP (MOVA_LABELREF (mova), 0)))) | |
1245df60 R |
5599 | { |
5600 | rtx scan; | |
5601 | int total; | |
5602 | ||
5603 | num_mova--; | |
5604 | ||
5605 | /* Some code might have been inserted between the mova and | |
5606 | its ADDR_DIFF_VEC. Check if the mova is still in range. */ | |
5607 | for (scan = mova, total = 0; scan != insn; scan = NEXT_INSN (scan)) | |
33f7f353 | 5608 | total += get_attr_length (scan); |
1245df60 R |
5609 | |
5610 | /* range of mova is 1020, add 4 because pc counts from address of | |
5611 | second instruction after this one, subtract 2 in case pc is 2 | |
5612 | byte aligned. Possible alignment needed for the ADDR_DIFF_VEC | |
956d6950 | 5613 | cancels out with alignment effects of the mova itself. */ |
1245df60 R |
5614 | if (total > 1022) |
5615 | { | |
5616 | /* Change the mova into a load, and restart scanning | |
5617 | there. broken_move will then return true for mova. */ | |
078c8b08 | 5618 | fixup_mova (mova); |
1245df60 R |
5619 | insn = mova; |
5620 | } | |
5621 | } | |
078c8b08 | 5622 | if (broken_move (insn) |
f3536097 | 5623 | || (NONJUMP_INSN_P (insn) |
078c8b08 | 5624 | && recog_memoized (insn) == CODE_FOR_casesi_worker_2)) |
0d7e008e | 5625 | { |
0d7e008e | 5626 | rtx scan; |
8aa2a305 JW |
5627 | /* Scan ahead looking for a barrier to stick the constant table |
5628 | behind. */ | |
1245df60 | 5629 | rtx barrier = find_barrier (num_mova, mova, insn); |
3a8699c7 | 5630 | rtx last_float_move = NULL_RTX, last_float = 0, *last_float_addr = NULL; |
078c8b08 | 5631 | int need_aligned_label = 0; |
b9654711 | 5632 | |
1245df60 R |
5633 | if (num_mova && ! mova_p (mova)) |
5634 | { | |
5635 | /* find_barrier had to change the first mova into a | |
5636 | pcload; thus, we have to start with this new pcload. */ | |
5637 | insn = mova; | |
5638 | num_mova = 0; | |
5639 | } | |
16bea517 | 5640 | /* Now find all the moves between the points and modify them. */ |
0d7e008e SC |
5641 | for (scan = insn; scan != barrier; scan = NEXT_INSN (scan)) |
5642 | { | |
f3536097 | 5643 | if (LABEL_P (scan)) |
1245df60 | 5644 | last_float = 0; |
f3536097 | 5645 | if (NONJUMP_INSN_P (scan) |
078c8b08 R |
5646 | && recog_memoized (scan) == CODE_FOR_casesi_worker_2) |
5647 | need_aligned_label = 1; | |
0d7e008e SC |
5648 | if (broken_move (scan)) |
5649 | { | |
e577c183 JR |
5650 | rtx *patp = &PATTERN (scan), pat = *patp; |
5651 | rtx src, dst; | |
0d7e008e | 5652 | rtx lab; |
0d7e008e | 5653 | rtx newsrc; |
e577c183 JR |
5654 | enum machine_mode mode; |
5655 | ||
5656 | if (GET_CODE (pat) == PARALLEL) | |
5657 | patp = &XVECEXP (pat, 0, 0), pat = *patp; | |
5658 | src = SET_SRC (pat); | |
5659 | dst = SET_DEST (pat); | |
5660 | mode = GET_MODE (dst); | |
0d7e008e | 5661 | |
225e4f43 R |
5662 | if (mode == SImode && hi_const (src) |
5663 | && REGNO (dst) != FPUL_REG) | |
0d7e008e | 5664 | { |
23ed92ca | 5665 | int offset = 0; |
8aa2a305 | 5666 | |
0d7e008e | 5667 | mode = HImode; |
d3ae8277 | 5668 | while (GET_CODE (dst) == SUBREG) |
23ed92ca | 5669 | { |
ddef6bc7 JJ |
5670 | offset += subreg_regno_offset (REGNO (SUBREG_REG (dst)), |
5671 | GET_MODE (SUBREG_REG (dst)), | |
5672 | SUBREG_BYTE (dst), | |
5673 | GET_MODE (dst)); | |
23ed92ca JW |
5674 | dst = SUBREG_REG (dst); |
5675 | } | |
c5c76735 | 5676 | dst = gen_rtx_REG (HImode, REGNO (dst) + offset); |
0d7e008e | 5677 | } |
f3536097 | 5678 | if (REG_P (dst) && FP_ANY_REGISTER_P (REGNO (dst))) |
1245df60 | 5679 | { |
b8794099 | 5680 | /* This must be an insn that clobbers r0. */ |
0f2b41b2 AO |
5681 | rtx *clobberp = &XVECEXP (PATTERN (scan), 0, |
5682 | XVECLEN (PATTERN (scan), 0) | |
5683 | - 1); | |
5684 | rtx clobber = *clobberp; | |
b8794099 | 5685 | |
f5b9e7c9 NS |
5686 | gcc_assert (GET_CODE (clobber) == CLOBBER |
5687 | && rtx_equal_p (XEXP (clobber, 0), r0_rtx)); | |
b8794099 | 5688 | |
1245df60 R |
5689 | if (last_float |
5690 | && reg_set_between_p (r0_rtx, last_float_move, scan)) | |
5691 | last_float = 0; | |
3503150c RS |
5692 | if (last_float |
5693 | && TARGET_SHCOMPACT | |
5694 | && GET_MODE_SIZE (mode) != 4 | |
5695 | && GET_MODE_SIZE (GET_MODE (last_float)) == 4) | |
5696 | last_float = 0; | |
1245df60 R |
5697 | lab = add_constant (src, mode, last_float); |
5698 | if (lab) | |
5699 | emit_insn_before (gen_mova (lab), scan); | |
5700 | else | |
b8794099 R |
5701 | { |
5702 | /* There will be a REG_UNUSED note for r0 on | |
5703 | LAST_FLOAT_MOVE; we have to change it to REG_INC, | |
5704 | lest reorg:mark_target_live_regs will not | |
5705 | consider r0 to be used, and we end up with delay | |
475f26dd | 5706 | slot insn in front of SCAN that clobbers r0. */ |
b8794099 R |
5707 | rtx note |
5708 | = find_regno_note (last_float_move, REG_UNUSED, 0); | |
5709 | ||
ab1672a3 JW |
5710 | /* If we are not optimizing, then there may not be |
5711 | a note. */ | |
5712 | if (note) | |
81f40b79 | 5713 | PUT_REG_NOTE_KIND (note, REG_INC); |
b8794099 R |
5714 | |
5715 | *last_float_addr = r0_inc_rtx; | |
5716 | } | |
1245df60 R |
5717 | last_float_move = scan; |
5718 | last_float = src; | |
57d38024 | 5719 | newsrc = gen_const_mem (mode, |
318881c0 | 5720 | (((TARGET_SH4 && ! TARGET_FMOVD) |
225e4f43 | 5721 | || REGNO (dst) == FPUL_REG) |
1245df60 R |
5722 | ? r0_inc_rtx |
5723 | : r0_rtx)); | |
5724 | last_float_addr = &XEXP (newsrc, 0); | |
b8794099 R |
5725 | |
5726 | /* Remove the clobber of r0. */ | |
0f2b41b2 AO |
5727 | *clobberp = gen_rtx_CLOBBER (GET_MODE (clobber), |
5728 | gen_rtx_SCRATCH (Pmode)); | |
1245df60 | 5729 | } |
1a66cd67 | 5730 | /* This is a mova needing a label. Create it. */ |
43c05634 AO |
5731 | else if (GET_CODE (src) == UNSPEC |
5732 | && XINT (src, 1) == UNSPEC_MOVA | |
5733 | && GET_CODE (XVECEXP (src, 0, 0)) == CONST) | |
1a66cd67 | 5734 | { |
43c05634 | 5735 | lab = add_constant (XVECEXP (src, 0, 0), mode, 0); |
1a66cd67 | 5736 | newsrc = gen_rtx_LABEL_REF (VOIDmode, lab); |
e69d1422 | 5737 | newsrc = gen_rtx_UNSPEC (SImode, |
4773afa4 AO |
5738 | gen_rtvec (1, newsrc), |
5739 | UNSPEC_MOVA); | |
1a66cd67 | 5740 | } |
1245df60 R |
5741 | else |
5742 | { | |
5743 | lab = add_constant (src, mode, 0); | |
542a8afa RH |
5744 | newsrc = gen_rtx_LABEL_REF (VOIDmode, lab); |
5745 | newsrc = gen_const_mem (mode, newsrc); | |
1245df60 | 5746 | } |
c5c76735 | 5747 | *patp = gen_rtx_SET (VOIDmode, dst, newsrc); |
e577c183 | 5748 | INSN_CODE (scan) = -1; |
0d7e008e SC |
5749 | } |
5750 | } | |
078c8b08 | 5751 | dump_table (need_aligned_label ? insn : 0, barrier); |
1245df60 | 5752 | insn = barrier; |
0d7e008e SC |
5753 | } |
5754 | } | |
6dd3c0a5 | 5755 | free_alloc_pool (label_ref_list_pool); |
4241ecb0 R |
5756 | for (insn = first; insn; insn = NEXT_INSN (insn)) |
5757 | PUT_MODE (insn, VOIDmode); | |
5758 | ||
1245df60 | 5759 | mdep_reorg_phase = SH_SHORTEN_BRANCHES1; |
9d98a694 | 5760 | INSN_ADDRESSES_FREE (); |
1245df60 R |
5761 | split_branches (first); |
5762 | ||
5763 | /* The INSN_REFERENCES_ARE_DELAYED in sh.h is problematic because it | |
a920aefe | 5764 | also has an effect on the register that holds the address of the sfunc. |
1245df60 R |
5765 | Insert an extra dummy insn in front of each sfunc that pretends to |
5766 | use this register. */ | |
5767 | if (flag_delayed_branch) | |
5768 | { | |
5769 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
5770 | { | |
5771 | rtx reg = sfunc_uses_reg (insn); | |
5772 | ||
5773 | if (! reg) | |
5774 | continue; | |
5775 | emit_insn_before (gen_use_sfunc_addr (reg), insn); | |
5776 | } | |
5777 | } | |
225e4f43 R |
5778 | #if 0 |
5779 | /* fpscr is not actually a user variable, but we pretend it is for the | |
5780 | sake of the previous optimization passes, since we want it handled like | |
9f09b1f2 | 5781 | one. However, we don't have any debugging information for it, so turn |
225e4f43 R |
5782 | it into a non-user variable now. */ |
5783 | if (TARGET_SH4) | |
5784 | REG_USERVAR_P (get_fpscr_rtx ()) = 0; | |
5785 | #endif | |
1245df60 R |
5786 | mdep_reorg_phase = SH_AFTER_MDEP_REORG; |
5787 | } | |
5788 | ||
5789 | int | |
cf277499 | 5790 | get_dest_uid (rtx label, int max_uid) |
1245df60 R |
5791 | { |
5792 | rtx dest = next_real_insn (label); | |
5793 | int dest_uid; | |
5794 | if (! dest) | |
5795 | /* This can happen for an undefined label. */ | |
5796 | return 0; | |
5797 | dest_uid = INSN_UID (dest); | |
5798 | /* If this is a newly created branch redirection blocking instruction, | |
5799 | we cannot index the branch_uid or insn_addresses arrays with its | |
5800 | uid. But then, we won't need to, because the actual destination is | |
5801 | the following branch. */ | |
5802 | while (dest_uid >= max_uid) | |
5803 | { | |
5804 | dest = NEXT_INSN (dest); | |
5805 | dest_uid = INSN_UID (dest); | |
5806 | } | |
f3536097 | 5807 | if (JUMP_P (dest) && GET_CODE (PATTERN (dest)) == RETURN) |
1245df60 R |
5808 | return 0; |
5809 | return dest_uid; | |
5810 | } | |
5811 | ||
5812 | /* Split condbranches that are out of range. Also add clobbers for | |
5813 | scratch registers that are needed in far jumps. | |
5814 | We do this before delay slot scheduling, so that it can take our | |
5815 | newly created instructions into account. It also allows us to | |
5816 | find branches with common targets more easily. */ | |
5817 | ||
5818 | static void | |
cf277499 | 5819 | split_branches (rtx first) |
1245df60 R |
5820 | { |
5821 | rtx insn; | |
5822 | struct far_branch **uid_branch, *far_branch_list = 0; | |
5823 | int max_uid = get_max_uid (); | |
f5b9e7c9 | 5824 | int ok; |
1245df60 R |
5825 | |
5826 | /* Find out which branches are out of range. */ | |
1245df60 R |
5827 | shorten_branches (first); |
5828 | ||
5829 | uid_branch = (struct far_branch **) alloca (max_uid * sizeof *uid_branch); | |
2e09e75a | 5830 | memset ((char *) uid_branch, 0, max_uid * sizeof *uid_branch); |
1245df60 R |
5831 | |
5832 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
2c3c49de | 5833 | if (! INSN_P (insn)) |
1245df60 R |
5834 | continue; |
5835 | else if (INSN_DELETED_P (insn)) | |
5836 | { | |
5837 | /* Shorten_branches would split this instruction again, | |
5838 | so transform it into a note. */ | |
a38e7aa5 | 5839 | SET_INSN_DELETED (insn); |
1245df60 | 5840 | } |
f3536097 | 5841 | else if (JUMP_P (insn) |
1245df60 R |
5842 | /* Don't mess with ADDR_DIFF_VEC */ |
5843 | && (GET_CODE (PATTERN (insn)) == SET | |
5844 | || GET_CODE (PATTERN (insn)) == RETURN)) | |
5845 | { | |
5846 | enum attr_type type = get_attr_type (insn); | |
5847 | if (type == TYPE_CBRANCH) | |
5848 | { | |
5849 | rtx next, beyond; | |
f676971a | 5850 | |
1245df60 R |
5851 | if (get_attr_length (insn) > 4) |
5852 | { | |
5853 | rtx src = SET_SRC (PATTERN (insn)); | |
1245df60 | 5854 | rtx olabel = XEXP (XEXP (src, 1), 0); |
9d98a694 | 5855 | int addr = INSN_ADDRESSES (INSN_UID (insn)); |
1245df60 R |
5856 | rtx label = 0; |
5857 | int dest_uid = get_dest_uid (olabel, max_uid); | |
5858 | struct far_branch *bp = uid_branch[dest_uid]; | |
f676971a | 5859 | |
1245df60 | 5860 | /* redirect_jump needs a valid JUMP_LABEL, and it might delete |
25938114 | 5861 | the label if the LABEL_NUSES count drops to zero. There is |
1245df60 R |
5862 | always a jump_optimize pass that sets these values, but it |
5863 | proceeds to delete unreferenced code, and then if not | |
956d6950 | 5864 | optimizing, to un-delete the deleted instructions, thus |
1245df60 R |
5865 | leaving labels with too low uses counts. */ |
5866 | if (! optimize) | |
5867 | { | |
5868 | JUMP_LABEL (insn) = olabel; | |
5869 | LABEL_NUSES (olabel)++; | |
5870 | } | |
5871 | if (! bp) | |
5872 | { | |
5873 | bp = (struct far_branch *) alloca (sizeof *bp); | |
5874 | uid_branch[dest_uid] = bp; | |
5875 | bp->prev = far_branch_list; | |
5876 | far_branch_list = bp; | |
5877 | bp->far_label | |
5878 | = XEXP (XEXP (SET_SRC (PATTERN (insn)), 1), 0); | |
5879 | LABEL_NUSES (bp->far_label)++; | |
5880 | } | |
5881 | else | |
5882 | { | |
5883 | label = bp->near_label; | |
5884 | if (! label && bp->address - addr >= CONDJUMP_MIN) | |
5885 | { | |
5886 | rtx block = bp->insert_place; | |
5887 | ||
5888 | if (GET_CODE (PATTERN (block)) == RETURN) | |
5889 | block = PREV_INSN (block); | |
5890 | else | |
5891 | block = gen_block_redirect (block, | |
5892 | bp->address, 2); | |
5893 | label = emit_label_after (gen_label_rtx (), | |
5894 | PREV_INSN (block)); | |
5895 | bp->near_label = label; | |
5896 | } | |
5897 | else if (label && ! NEXT_INSN (label)) | |
eda44c78 R |
5898 | { |
5899 | if (addr + 2 - bp->address <= CONDJUMP_MAX) | |
5900 | bp->insert_place = insn; | |
5901 | else | |
5902 | gen_far_branch (bp); | |
5903 | } | |
1245df60 R |
5904 | } |
5905 | if (! label | |
318881c0 | 5906 | || (NEXT_INSN (label) && bp->address - addr < CONDJUMP_MIN)) |
1245df60 R |
5907 | { |
5908 | bp->near_label = label = gen_label_rtx (); | |
5909 | bp->insert_place = insn; | |
5910 | bp->address = addr; | |
5911 | } | |
2acdc1fb | 5912 | ok = redirect_jump (insn, label, 0); |
f5b9e7c9 | 5913 | gcc_assert (ok); |
1245df60 R |
5914 | } |
5915 | else | |
5916 | { | |
5917 | /* get_attr_length (insn) == 2 */ | |
5918 | /* Check if we have a pattern where reorg wants to redirect | |
5919 | the branch to a label from an unconditional branch that | |
5920 | is too far away. */ | |
5921 | /* We can't use JUMP_LABEL here because it might be undefined | |
5922 | when not optimizing. */ | |
33f7f353 | 5923 | /* A syntax error might cause beyond to be NULL_RTX. */ |
1245df60 R |
5924 | beyond |
5925 | = next_active_insn (XEXP (XEXP (SET_SRC (PATTERN (insn)), 1), | |
5926 | 0)); | |
f676971a | 5927 | |
33f7f353 | 5928 | if (beyond |
f3536097 | 5929 | && (JUMP_P (beyond) |
b94544e4 | 5930 | || ((beyond = next_active_insn (beyond)) |
f3536097 | 5931 | && JUMP_P (beyond))) |
1245df60 | 5932 | && GET_CODE (PATTERN (beyond)) == SET |
ddf8ec7e | 5933 | && recog_memoized (beyond) == CODE_FOR_jump_compact |
9d98a694 AO |
5934 | && ((INSN_ADDRESSES |
5935 | (INSN_UID (XEXP (SET_SRC (PATTERN (beyond)), 0))) | |
11f9ed1a | 5936 | - INSN_ADDRESSES (INSN_UID (insn)) + (unsigned) 252) |
1245df60 R |
5937 | > 252 + 258 + 2)) |
5938 | gen_block_redirect (beyond, | |
9d98a694 | 5939 | INSN_ADDRESSES (INSN_UID (beyond)), 1); |
1245df60 | 5940 | } |
f676971a | 5941 | |
1245df60 R |
5942 | next = next_active_insn (insn); |
5943 | ||
a7d6f306 KK |
5944 | if (next |
5945 | && (JUMP_P (next) | |
5946 | || ((next = next_active_insn (next)) | |
5947 | && JUMP_P (next))) | |
1245df60 | 5948 | && GET_CODE (PATTERN (next)) == SET |
ddf8ec7e | 5949 | && recog_memoized (next) == CODE_FOR_jump_compact |
9d98a694 AO |
5950 | && ((INSN_ADDRESSES |
5951 | (INSN_UID (XEXP (SET_SRC (PATTERN (next)), 0))) | |
11f9ed1a | 5952 | - INSN_ADDRESSES (INSN_UID (insn)) + (unsigned) 252) |
1245df60 | 5953 | > 252 + 258 + 2)) |
9d98a694 | 5954 | gen_block_redirect (next, INSN_ADDRESSES (INSN_UID (next)), 1); |
1245df60 R |
5955 | } |
5956 | else if (type == TYPE_JUMP || type == TYPE_RETURN) | |
5957 | { | |
9d98a694 | 5958 | int addr = INSN_ADDRESSES (INSN_UID (insn)); |
1245df60 R |
5959 | rtx far_label = 0; |
5960 | int dest_uid = 0; | |
5961 | struct far_branch *bp; | |
5962 | ||
5963 | if (type == TYPE_JUMP) | |
5964 | { | |
5965 | far_label = XEXP (SET_SRC (PATTERN (insn)), 0); | |
5966 | dest_uid = get_dest_uid (far_label, max_uid); | |
5967 | if (! dest_uid) | |
5968 | { | |
5969 | /* Parse errors can lead to labels outside | |
5970 | the insn stream. */ | |
5971 | if (! NEXT_INSN (far_label)) | |
5972 | continue; | |
5973 | ||
5974 | if (! optimize) | |
5975 | { | |
5976 | JUMP_LABEL (insn) = far_label; | |
5977 | LABEL_NUSES (far_label)++; | |
5978 | } | |
8ceaac3c | 5979 | redirect_jump (insn, NULL_RTX, 1); |
1245df60 R |
5980 | far_label = 0; |
5981 | } | |
5982 | } | |
5983 | bp = uid_branch[dest_uid]; | |
5984 | if (! bp) | |
5985 | { | |
5986 | bp = (struct far_branch *) alloca (sizeof *bp); | |
5987 | uid_branch[dest_uid] = bp; | |
5988 | bp->prev = far_branch_list; | |
5989 | far_branch_list = bp; | |
5990 | bp->near_label = 0; | |
5991 | bp->far_label = far_label; | |
5992 | if (far_label) | |
5993 | LABEL_NUSES (far_label)++; | |
5994 | } | |
5995 | else if (bp->near_label && ! NEXT_INSN (bp->near_label)) | |
5996 | if (addr - bp->address <= CONDJUMP_MAX) | |
5997 | emit_label_after (bp->near_label, PREV_INSN (insn)); | |
5998 | else | |
5999 | { | |
6000 | gen_far_branch (bp); | |
6001 | bp->near_label = 0; | |
6002 | } | |
6003 | else | |
6004 | bp->near_label = 0; | |
6005 | bp->address = addr; | |
6006 | bp->insert_place = insn; | |
6007 | if (! far_label) | |
6008 | emit_insn_before (gen_block_branch_redirect (const0_rtx), insn); | |
6009 | else | |
6010 | gen_block_redirect (insn, addr, bp->near_label ? 2 : 0); | |
6011 | } | |
6012 | } | |
6013 | /* Generate all pending far branches, | |
6014 | and free our references to the far labels. */ | |
6015 | while (far_branch_list) | |
6016 | { | |
6017 | if (far_branch_list->near_label | |
6018 | && ! NEXT_INSN (far_branch_list->near_label)) | |
6019 | gen_far_branch (far_branch_list); | |
6020 | if (optimize | |
6021 | && far_branch_list->far_label | |
6022 | && ! --LABEL_NUSES (far_branch_list->far_label)) | |
6023 | delete_insn (far_branch_list->far_label); | |
6024 | far_branch_list = far_branch_list->prev; | |
6025 | } | |
a0798779 R |
6026 | |
6027 | /* Instruction length information is no longer valid due to the new | |
6028 | instructions that have been generated. */ | |
6029 | init_insn_lengths (); | |
b9654711 SC |
6030 | } |
6031 | ||
8aa2a305 | 6032 | /* Dump out instruction addresses, which is useful for debugging the |
933c3ba3 JW |
6033 | constant pool table stuff. |
6034 | ||
6035 | If relaxing, output the label and pseudo-ops used to link together | |
6036 | calls and the instruction which set the registers. */ | |
8aa2a305 | 6037 | |
8aa2a305 JW |
6038 | /* ??? The addresses printed by this routine for insns are nonsense for |
6039 | insns which are inside of a sequence where none of the inner insns have | |
6040 | variable length. This is because the second pass of shorten_branches | |
6041 | does not bother to update them. */ | |
0d7e008e | 6042 | |
8e87e161 | 6043 | void |
cf277499 SB |
6044 | final_prescan_insn (rtx insn, rtx *opvec ATTRIBUTE_UNUSED, |
6045 | int noperands ATTRIBUTE_UNUSED) | |
b9654711 | 6046 | { |
8aa2a305 | 6047 | if (TARGET_DUMPISIZE) |
9d98a694 | 6048 | fprintf (asm_out_file, "\n! at %04x\n", INSN_ADDRESSES (INSN_UID (insn))); |
933c3ba3 JW |
6049 | |
6050 | if (TARGET_RELAX) | |
6051 | { | |
6052 | rtx note; | |
6053 | ||
cf7c4aa6 | 6054 | note = find_reg_note (insn, REG_LABEL_OPERAND, NULL_RTX); |
933c3ba3 JW |
6055 | if (note) |
6056 | { | |
6057 | rtx pattern; | |
6058 | ||
6059 | pattern = PATTERN (insn); | |
4b7d2f07 R |
6060 | if (GET_CODE (pattern) == PARALLEL) |
6061 | pattern = XVECEXP (pattern, 0, 0); | |
f5b9e7c9 NS |
6062 | switch (GET_CODE (pattern)) |
6063 | { | |
f5b9e7c9 NS |
6064 | case SET: |
6065 | if (GET_CODE (SET_SRC (pattern)) != CALL | |
6066 | && get_attr_type (insn) != TYPE_SFUNC) | |
6067 | { | |
6068 | targetm.asm_out.internal_label | |
6069 | (asm_out_file, "L", CODE_LABEL_NUMBER (XEXP (note, 0))); | |
6070 | break; | |
6071 | } | |
6072 | /* else FALLTHROUGH */ | |
6073 | case CALL: | |
6074 | asm_fprintf (asm_out_file, "\t.uses %LL%d\n", | |
6075 | CODE_LABEL_NUMBER (XEXP (note, 0))); | |
6076 | break; | |
6077 | ||
6078 | default: | |
6079 | gcc_unreachable (); | |
6080 | } | |
933c3ba3 JW |
6081 | } |
6082 | } | |
0d7e008e | 6083 | } |
b9654711 | 6084 | |
8aa2a305 | 6085 | /* Dump out any constants accumulated in the final pass. These will |
38e01259 | 6086 | only be labels. */ |
b9654711 | 6087 | |
318881c0 | 6088 | const char * |
cf277499 | 6089 | output_jump_label_table (void) |
0d7e008e | 6090 | { |
8aa2a305 JW |
6091 | int i; |
6092 | ||
6093 | if (pool_size) | |
6094 | { | |
6095 | fprintf (asm_out_file, "\t.align 2\n"); | |
6096 | for (i = 0; i < pool_size; i++) | |
6097 | { | |
6098 | pool_node *p = &pool_vector[i]; | |
6099 | ||
4977bab6 | 6100 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
8aa2a305 JW |
6101 | CODE_LABEL_NUMBER (p->label)); |
6102 | output_asm_insn (".long %O0", &p->value); | |
6103 | } | |
6104 | pool_size = 0; | |
6105 | } | |
b9654711 | 6106 | |
8aa2a305 JW |
6107 | return ""; |
6108 | } | |
6109 | \f | |
6110 | /* A full frame looks like: | |
16bea517 JW |
6111 | |
6112 | arg-5 | |
6113 | arg-4 | |
6114 | [ if current_function_anonymous_args | |
6115 | arg-3 | |
6116 | arg-2 | |
6117 | arg-1 | |
6118 | arg-0 ] | |
6119 | saved-fp | |
6120 | saved-r10 | |
6121 | saved-r11 | |
6122 | saved-r12 | |
6123 | saved-pr | |
6124 | local-n | |
6125 | .. | |
6126 | local-1 | |
6127 | local-0 <- fp points here. */ | |
8e87e161 | 6128 | |
8aa2a305 JW |
6129 | /* Number of bytes pushed for anonymous args, used to pass information |
6130 | between expand_prologue and expand_epilogue. */ | |
6131 | ||
fada1961 R |
6132 | /* Adjust the stack by SIZE bytes. REG holds the rtl of the register to be |
6133 | adjusted. If epilogue_p is zero, this is for a prologue; otherwise, it's | |
726d4cb7 KK |
6134 | for an epilogue and a negative value means that it's for a sibcall |
6135 | epilogue. If LIVE_REGS_MASK is nonzero, it points to a HARD_REG_SET of | |
6136 | all the registers that are about to be restored, and hence dead. */ | |
8aa2a305 JW |
6137 | |
6138 | static void | |
fada1961 | 6139 | output_stack_adjust (int size, rtx reg, int epilogue_p, |
21d05694 | 6140 | HARD_REG_SET *live_regs_mask, bool frame_p) |
8aa2a305 | 6141 | { |
21d05694 | 6142 | rtx (*emit_fn) (rtx) = frame_p ? &frame_insn : &emit_insn; |
8aa2a305 JW |
6143 | if (size) |
6144 | { | |
fa5322fa AO |
6145 | HOST_WIDE_INT align = STACK_BOUNDARY / BITS_PER_UNIT; |
6146 | ||
157371cf AO |
6147 | /* This test is bogus, as output_stack_adjust is used to re-align the |
6148 | stack. */ | |
6149 | #if 0 | |
f5b9e7c9 | 6150 | gcc_assert (!(size % align)); |
157371cf | 6151 | #endif |
fa5322fa AO |
6152 | |
6153 | if (CONST_OK_FOR_ADD (size)) | |
2754d3c5 | 6154 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size))); |
f3cd5375 JW |
6155 | /* Try to do it with two partial adjustments; however, we must make |
6156 | sure that the stack is properly aligned at all times, in case | |
6f317ef3 | 6157 | an interrupt occurs between the two partial adjustments. */ |
fa5322fa AO |
6158 | else if (CONST_OK_FOR_ADD (size / 2 & -align) |
6159 | && CONST_OK_FOR_ADD (size - (size / 2 & -align))) | |
f3cd5375 | 6160 | { |
2754d3c5 R |
6161 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size / 2 & -align))); |
6162 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size - (size / 2 & -align)))); | |
f3cd5375 JW |
6163 | } |
6164 | else | |
8aa2a305 | 6165 | { |
f3cd5375 | 6166 | rtx const_reg; |
2754d3c5 | 6167 | rtx insn; |
fada1961 R |
6168 | int temp = epilogue_p ? 7 : (TARGET_SH5 ? 0 : 1); |
6169 | int i; | |
f3cd5375 JW |
6170 | |
6171 | /* If TEMP is invalid, we could temporarily save a general | |
6172 | register to MACL. However, there is currently no need | |
f5b9e7c9 | 6173 | to handle this case, so just die when we see it. */ |
726d4cb7 KK |
6174 | if (epilogue_p < 0 |
6175 | || current_function_interrupt | |
62164eb4 | 6176 | || ! call_really_used_regs[temp] || fixed_regs[temp]) |
fada1961 | 6177 | temp = -1; |
726d4cb7 KK |
6178 | if (temp < 0 && ! current_function_interrupt |
6179 | && (TARGET_SHMEDIA || epilogue_p >= 0)) | |
fada1961 R |
6180 | { |
6181 | HARD_REG_SET temps; | |
6182 | COPY_HARD_REG_SET (temps, call_used_reg_set); | |
6183 | AND_COMPL_HARD_REG_SET (temps, call_fixed_reg_set); | |
726d4cb7 | 6184 | if (epilogue_p > 0) |
fada1961 | 6185 | { |
726d4cb7 | 6186 | int nreg = 0; |
38173d38 | 6187 | if (crtl->return_rtx) |
726d4cb7 KK |
6188 | { |
6189 | enum machine_mode mode; | |
38173d38 | 6190 | mode = GET_MODE (crtl->return_rtx); |
726d4cb7 KK |
6191 | if (BASE_RETURN_VALUE_REG (mode) == FIRST_RET_REG) |
6192 | nreg = HARD_REGNO_NREGS (FIRST_RET_REG, mode); | |
6193 | } | |
6194 | for (i = 0; i < nreg; i++) | |
fada1961 | 6195 | CLEAR_HARD_REG_BIT (temps, FIRST_RET_REG + i); |
e3b5732b | 6196 | if (crtl->calls_eh_return) |
fada1961 R |
6197 | { |
6198 | CLEAR_HARD_REG_BIT (temps, EH_RETURN_STACKADJ_REGNO); | |
6199 | for (i = 0; i <= 3; i++) | |
6200 | CLEAR_HARD_REG_BIT (temps, EH_RETURN_DATA_REGNO (i)); | |
6201 | } | |
6202 | } | |
726d4cb7 KK |
6203 | if (TARGET_SHMEDIA && epilogue_p < 0) |
6204 | for (i = FIRST_TARGET_REG; i <= LAST_TARGET_REG; i++) | |
6205 | CLEAR_HARD_REG_BIT (temps, i); | |
6206 | if (epilogue_p <= 0) | |
fada1961 R |
6207 | { |
6208 | for (i = FIRST_PARM_REG; | |
6209 | i < FIRST_PARM_REG + NPARM_REGS (SImode); i++) | |
6210 | CLEAR_HARD_REG_BIT (temps, i); | |
6de9cd9a | 6211 | if (cfun->static_chain_decl != NULL) |
fada1961 R |
6212 | CLEAR_HARD_REG_BIT (temps, STATIC_CHAIN_REGNUM); |
6213 | } | |
6214 | temp = scavenge_reg (&temps); | |
6215 | } | |
6216 | if (temp < 0 && live_regs_mask) | |
eb996a4a KK |
6217 | { |
6218 | HARD_REG_SET temps; | |
6219 | ||
6220 | COPY_HARD_REG_SET (temps, *live_regs_mask); | |
6221 | CLEAR_HARD_REG_BIT (temps, REGNO (reg)); | |
6222 | temp = scavenge_reg (&temps); | |
6223 | } | |
f3cd5375 | 6224 | if (temp < 0) |
726d4cb7 | 6225 | { |
f5b9e7c9 NS |
6226 | rtx adj_reg, tmp_reg, mem; |
6227 | ||
726d4cb7 KK |
6228 | /* If we reached here, the most likely case is the (sibcall) |
6229 | epilogue for non SHmedia. Put a special push/pop sequence | |
6230 | for such case as the last resort. This looks lengthy but | |
f5b9e7c9 NS |
6231 | would not be problem because it seems to be very |
6232 | rare. */ | |
6233 | ||
6234 | gcc_assert (!TARGET_SHMEDIA && epilogue_p); | |
6235 | ||
6236 | ||
6237 | /* ??? There is still the slight possibility that r4 or | |
6238 | r5 have been reserved as fixed registers or assigned | |
6239 | as global registers, and they change during an | |
6240 | interrupt. There are possible ways to handle this: | |
6241 | ||
6242 | - If we are adjusting the frame pointer (r14), we can do | |
6243 | with a single temp register and an ordinary push / pop | |
6244 | on the stack. | |
6245 | - Grab any call-used or call-saved registers (i.e. not | |
6246 | fixed or globals) for the temps we need. We might | |
6247 | also grab r14 if we are adjusting the stack pointer. | |
6248 | If we can't find enough available registers, issue | |
6249 | a diagnostic and die - the user must have reserved | |
6250 | way too many registers. | |
6251 | But since all this is rather unlikely to happen and | |
6252 | would require extra testing, we just die if r4 / r5 | |
6253 | are not available. */ | |
6254 | gcc_assert (!fixed_regs[4] && !fixed_regs[5] | |
6255 | && !global_regs[4] && !global_regs[5]); | |
6256 | ||
6257 | adj_reg = gen_rtx_REG (GET_MODE (reg), 4); | |
6258 | tmp_reg = gen_rtx_REG (GET_MODE (reg), 5); | |
57d38024 | 6259 | emit_move_insn (gen_tmp_stack_mem (Pmode, reg), adj_reg); |
f5b9e7c9 NS |
6260 | emit_insn (GEN_MOV (adj_reg, GEN_INT (size))); |
6261 | emit_insn (GEN_ADD3 (adj_reg, adj_reg, reg)); | |
57d38024 | 6262 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_PRE_DEC (Pmode, adj_reg)); |
f5b9e7c9 | 6263 | emit_move_insn (mem, tmp_reg); |
57d38024 R |
6264 | emit_move_insn (tmp_reg, gen_tmp_stack_mem (Pmode, reg)); |
6265 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_PRE_DEC (Pmode, adj_reg)); | |
6266 | emit_move_insn (mem, tmp_reg); | |
6267 | emit_move_insn (reg, adj_reg); | |
6268 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_POST_INC (Pmode, reg)); | |
6269 | emit_move_insn (adj_reg, mem); | |
6270 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_POST_INC (Pmode, reg)); | |
6271 | emit_move_insn (tmp_reg, mem); | |
eb996a4a | 6272 | /* Tell flow the insns that pop r4/r5 aren't dead. */ |
c41c1387 RS |
6273 | emit_use (tmp_reg); |
6274 | emit_use (adj_reg); | |
57d38024 | 6275 | return; |
726d4cb7 | 6276 | } |
fa5322fa | 6277 | const_reg = gen_rtx_REG (GET_MODE (reg), temp); |
f3cd5375 JW |
6278 | |
6279 | /* If SIZE is negative, subtract the positive value. | |
6280 | This sometimes allows a constant pool entry to be shared | |
6281 | between prologue and epilogue code. */ | |
6282 | if (size < 0) | |
885c9a39 | 6283 | { |
fa5322fa | 6284 | emit_insn (GEN_MOV (const_reg, GEN_INT (-size))); |
2754d3c5 | 6285 | insn = emit_fn (GEN_SUB3 (reg, reg, const_reg)); |
885c9a39 JR |
6286 | } |
6287 | else | |
6288 | { | |
fa5322fa | 6289 | emit_insn (GEN_MOV (const_reg, GEN_INT (size))); |
2754d3c5 | 6290 | insn = emit_fn (GEN_ADD3 (reg, reg, const_reg)); |
885c9a39 | 6291 | } |
fada1961 | 6292 | if (! epilogue_p) |
b757d36f KK |
6293 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, |
6294 | gen_rtx_SET (VOIDmode, reg, | |
6295 | gen_rtx_PLUS (SImode, reg, | |
6296 | GEN_INT (size)))); | |
8aa2a305 | 6297 | } |
8aa2a305 JW |
6298 | } |
6299 | } | |
6300 | ||
2754d3c5 | 6301 | static rtx |
cf277499 | 6302 | frame_insn (rtx x) |
2754d3c5 R |
6303 | { |
6304 | x = emit_insn (x); | |
6305 | RTX_FRAME_RELATED_P (x) = 1; | |
6306 | return x; | |
6307 | } | |
6308 | ||
8aa2a305 JW |
6309 | /* Output RTL to push register RN onto the stack. */ |
6310 | ||
2754d3c5 | 6311 | static rtx |
cf277499 | 6312 | push (int rn) |
8aa2a305 JW |
6313 | { |
6314 | rtx x; | |
225e4f43 R |
6315 | if (rn == FPUL_REG) |
6316 | x = gen_push_fpul (); | |
7144b2d8 DD |
6317 | else if (rn == FPSCR_REG) |
6318 | x = gen_push_fpscr (); | |
157371cf | 6319 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && ! TARGET_FPU_SINGLE |
104ee20b | 6320 | && FP_OR_XD_REGISTER_P (rn)) |
225e4f43 | 6321 | { |
104ee20b | 6322 | if (FP_REGISTER_P (rn) && (rn - FIRST_FP_REG) & 1) |
d042370e | 6323 | return NULL_RTX; |
c5c76735 | 6324 | x = gen_push_4 (gen_rtx_REG (DFmode, rn)); |
225e4f43 | 6325 | } |
3a8699c7 | 6326 | else if (TARGET_SH2E && FP_REGISTER_P (rn)) |
c5c76735 | 6327 | x = gen_push_e (gen_rtx_REG (SFmode, rn)); |
1a95a963 | 6328 | else |
c5c76735 | 6329 | x = gen_push (gen_rtx_REG (SImode, rn)); |
1a95a963 | 6330 | |
2754d3c5 | 6331 | x = frame_insn (x); |
b757d36f | 6332 | add_reg_note (x, REG_INC, gen_rtx_REG (SImode, STACK_POINTER_REGNUM)); |
2754d3c5 | 6333 | return x; |
8aa2a305 JW |
6334 | } |
6335 | ||
6336 | /* Output RTL to pop register RN from the stack. */ | |
6337 | ||
6338 | static void | |
cf277499 | 6339 | pop (int rn) |
8aa2a305 JW |
6340 | { |
6341 | rtx x; | |
225e4f43 R |
6342 | if (rn == FPUL_REG) |
6343 | x = gen_pop_fpul (); | |
7144b2d8 DD |
6344 | else if (rn == FPSCR_REG) |
6345 | x = gen_pop_fpscr (); | |
157371cf | 6346 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && ! TARGET_FPU_SINGLE |
104ee20b | 6347 | && FP_OR_XD_REGISTER_P (rn)) |
225e4f43 | 6348 | { |
104ee20b | 6349 | if (FP_REGISTER_P (rn) && (rn - FIRST_FP_REG) & 1) |
225e4f43 | 6350 | return; |
c5c76735 | 6351 | x = gen_pop_4 (gen_rtx_REG (DFmode, rn)); |
225e4f43 | 6352 | } |
3a8699c7 | 6353 | else if (TARGET_SH2E && FP_REGISTER_P (rn)) |
c5c76735 | 6354 | x = gen_pop_e (gen_rtx_REG (SFmode, rn)); |
1a95a963 | 6355 | else |
c5c76735 | 6356 | x = gen_pop (gen_rtx_REG (SImode, rn)); |
f676971a | 6357 | |
1245df60 | 6358 | x = emit_insn (x); |
b757d36f | 6359 | add_reg_note (x, REG_INC, gen_rtx_REG (SImode, STACK_POINTER_REGNUM)); |
8aa2a305 JW |
6360 | } |
6361 | ||
1245df60 | 6362 | /* Generate code to push the regs specified in the mask. */ |
8aa2a305 JW |
6363 | |
6364 | static void | |
cf277499 | 6365 | push_regs (HARD_REG_SET *mask, int interrupt_handler) |
8aa2a305 | 6366 | { |
5c7cafa8 | 6367 | int i = interrupt_handler ? LAST_BANKED_REG + 1 : 0; |
ac55736a | 6368 | int skip_fpscr = 0; |
8aa2a305 | 6369 | |
1245df60 R |
6370 | /* Push PR last; this gives better latencies after the prologue, and |
6371 | candidates for the return delay slot when there are no general | |
6372 | registers pushed. */ | |
5c7cafa8 | 6373 | for (; i < FIRST_PSEUDO_REGISTER; i++) |
ac55736a R |
6374 | { |
6375 | /* If this is an interrupt handler, and the SZ bit varies, | |
6376 | and we have to push any floating point register, we need | |
6377 | to switch to the correct precision first. */ | |
6378 | if (i == FIRST_FP_REG && interrupt_handler && TARGET_FMOVD | |
56b138ae | 6379 | && hard_reg_set_intersect_p (*mask, reg_class_contents[DF_REGS])) |
ac55736a R |
6380 | { |
6381 | HARD_REG_SET unsaved; | |
6382 | ||
6383 | push (FPSCR_REG); | |
832a3292 | 6384 | COMPL_HARD_REG_SET (unsaved, *mask); |
ac55736a R |
6385 | fpscr_set_from_mem (NORMAL_MODE (FP_MODE), unsaved); |
6386 | skip_fpscr = 1; | |
6387 | } | |
6388 | if (i != PR_REG | |
6389 | && (i != FPSCR_REG || ! skip_fpscr) | |
6390 | && TEST_HARD_REG_BIT (*mask, i)) | |
561642fa AP |
6391 | { |
6392 | /* If the ISR has RESBANK attribute assigned, don't push any of | |
6393 | the following registers - R0-R14, MACH, MACL and GBR. */ | |
6394 | if (! (sh_cfun_resbank_handler_p () | |
6395 | && ((i >= FIRST_GENERAL_REG && i < LAST_GENERAL_REG) | |
6396 | || i == MACH_REG | |
6397 | || i == MACL_REG | |
6398 | || i == GBR_REG))) | |
6399 | push (i); | |
6400 | } | |
ac55736a | 6401 | } |
5c7cafa8 CB |
6402 | |
6403 | /* Push banked registers last to improve delay slot opportunities. */ | |
6404 | if (interrupt_handler) | |
6405 | for (i = FIRST_BANKED_REG; i <= LAST_BANKED_REG; i++) | |
6406 | if (TEST_HARD_REG_BIT (*mask, i)) | |
6407 | push (i); | |
6408 | ||
561642fa AP |
6409 | /* Don't push PR register for an ISR with RESBANK attribute assigned. */ |
6410 | if (TEST_HARD_REG_BIT (*mask, PR_REG) && !sh_cfun_resbank_handler_p ()) | |
1245df60 | 6411 | push (PR_REG); |
8aa2a305 JW |
6412 | } |
6413 | ||
fe3ad572 SC |
6414 | /* Calculate how much extra space is needed to save all callee-saved |
6415 | target registers. | |
6416 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
6417 | ||
6418 | static int | |
6419 | shmedia_target_regs_stack_space (HARD_REG_SET *live_regs_mask) | |
6420 | { | |
6421 | int reg; | |
6422 | int stack_space = 0; | |
6423 | int interrupt_handler = sh_cfun_interrupt_handler_p (); | |
6424 | ||
6425 | for (reg = LAST_TARGET_REG; reg >= FIRST_TARGET_REG; reg--) | |
62164eb4 | 6426 | if ((! call_really_used_regs[reg] || interrupt_handler) |
fe3ad572 SC |
6427 | && ! TEST_HARD_REG_BIT (*live_regs_mask, reg)) |
6428 | /* Leave space to save this target register on the stack, | |
24746a42 | 6429 | in case target register allocation wants to use it. */ |
fe3ad572 SC |
6430 | stack_space += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); |
6431 | return stack_space; | |
6432 | } | |
f676971a | 6433 | |
fe3ad572 SC |
6434 | /* Decide whether we should reserve space for callee-save target registers, |
6435 | in case target register allocation wants to use them. REGS_SAVED is | |
6436 | the space, in bytes, that is already required for register saves. | |
6437 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
6438 | ||
6439 | static int | |
6440 | shmedia_reserve_space_for_target_registers_p (int regs_saved, | |
6441 | HARD_REG_SET *live_regs_mask) | |
6442 | { | |
6443 | if (optimize_size) | |
6444 | return 0; | |
6445 | return shmedia_target_regs_stack_space (live_regs_mask) <= regs_saved; | |
6446 | } | |
6447 | ||
6448 | /* Decide how much space to reserve for callee-save target registers | |
6449 | in case target register allocation wants to use them. | |
6450 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
6451 | ||
6452 | static int | |
6453 | shmedia_target_regs_stack_adjust (HARD_REG_SET *live_regs_mask) | |
6454 | { | |
6455 | if (shmedia_space_reserved_for_target_registers) | |
6456 | return shmedia_target_regs_stack_space (live_regs_mask); | |
6457 | else | |
6458 | return 0; | |
6459 | } | |
6460 | ||
8aa2a305 | 6461 | /* Work out the registers which need to be saved, both as a mask and a |
490dbac7 | 6462 | count of saved words. Return the count. |
8aa2a305 JW |
6463 | |
6464 | If doing a pragma interrupt function, then push all regs used by the | |
6465 | function, and if we call another function (we can tell by looking at PR), | |
6466 | make sure that all the regs it clobbers are safe too. */ | |
6467 | ||
490dbac7 | 6468 | static int |
cf277499 | 6469 | calc_live_regs (HARD_REG_SET *live_regs_mask) |
8aa2a305 | 6470 | { |
0ef2c525 | 6471 | unsigned int reg; |
1245df60 | 6472 | int count; |
a6ab9fc0 R |
6473 | tree attrs; |
6474 | bool interrupt_or_trapa_handler, trapa_handler, interrupt_handler; | |
6475 | bool nosave_low_regs; | |
fada1961 | 6476 | int pr_live, has_call; |
157131d7 | 6477 | |
a6ab9fc0 R |
6478 | attrs = DECL_ATTRIBUTES (current_function_decl); |
6479 | interrupt_or_trapa_handler = sh_cfun_interrupt_handler_p (); | |
6480 | trapa_handler = lookup_attribute ("trapa_handler", attrs) != NULL_TREE; | |
6481 | interrupt_handler = interrupt_or_trapa_handler && ! trapa_handler; | |
6482 | nosave_low_regs = lookup_attribute ("nosave_low_regs", attrs) != NULL_TREE; | |
8aa2a305 | 6483 | |
fada1961 | 6484 | CLEAR_HARD_REG_SET (*live_regs_mask); |
157371cf | 6485 | if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && interrupt_handler |
6fb5fa3c | 6486 | && df_regs_ever_live_p (FPSCR_REG)) |
c0fb94d7 | 6487 | target_flags &= ~MASK_FPU_SINGLE; |
225e4f43 | 6488 | /* If we can save a lot of saves by switching to double mode, do that. */ |
157371cf | 6489 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && TARGET_FPU_SINGLE) |
225e4f43 | 6490 | for (count = 0, reg = FIRST_FP_REG; reg <= LAST_FP_REG; reg += 2) |
6fb5fa3c | 6491 | if (df_regs_ever_live_p (reg) && df_regs_ever_live_p (reg+1) |
62164eb4 | 6492 | && (! call_really_used_regs[reg] |
a6ab9fc0 | 6493 | || interrupt_handler) |
225e4f43 R |
6494 | && ++count > 2) |
6495 | { | |
c0fb94d7 | 6496 | target_flags &= ~MASK_FPU_SINGLE; |
225e4f43 R |
6497 | break; |
6498 | } | |
9d7ed806 R |
6499 | /* PR_MEDIA_REG is a general purpose register, thus global_alloc already |
6500 | knows how to use it. That means the pseudo originally allocated for | |
6501 | the initial value can become the PR_MEDIA_REG hard register, as seen for | |
6502 | execute/20010122-1.c:test9. */ | |
6503 | if (TARGET_SHMEDIA) | |
7d73a2ba R |
6504 | /* ??? this function is called from initial_elimination_offset, hence we |
6505 | can't use the result of sh_media_register_for_return here. */ | |
97b3d0f0 | 6506 | pr_live = sh_pr_n_sets (); |
9d7ed806 R |
6507 | else |
6508 | { | |
6509 | rtx pr_initial = has_hard_reg_initial_val (Pmode, PR_REG); | |
6510 | pr_live = (pr_initial | |
f3536097 | 6511 | ? (!REG_P (pr_initial) |
29db2a1e | 6512 | || REGNO (pr_initial) != (PR_REG)) |
6fb5fa3c | 6513 | : df_regs_ever_live_p (PR_REG)); |
7d73a2ba R |
6514 | /* For Shcompact, if not optimizing, we end up with a memory reference |
6515 | using the return address pointer for __builtin_return_address even | |
6516 | though there is no actual need to put the PR register on the stack. */ | |
6fb5fa3c | 6517 | pr_live |= df_regs_ever_live_p (RETURN_ADDRESS_POINTER_REGNUM); |
9d7ed806 | 6518 | } |
fa5322fa AO |
6519 | /* Force PR to be live if the prologue has to call the SHmedia |
6520 | argument decoder or register saver. */ | |
6521 | if (TARGET_SHCOMPACT | |
38173d38 | 6522 | && ((crtl->args.info.call_cookie |
fa5322fa | 6523 | & ~ CALL_COOKIE_RET_TRAMP (1)) |
e3b5732b | 6524 | || crtl->saves_all_registers)) |
fa5322fa | 6525 | pr_live = 1; |
fada1961 | 6526 | has_call = TARGET_SHMEDIA ? ! leaf_function_p () : pr_live; |
0ef2c525 | 6527 | for (count = 0, reg = FIRST_PSEUDO_REGISTER; reg-- != 0; ) |
8aa2a305 | 6528 | { |
97b3d0f0 | 6529 | if (reg == (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG) |
3007d592 | 6530 | ? pr_live |
a6ab9fc0 | 6531 | : interrupt_handler |
1245df60 | 6532 | ? (/* Need to save all the regs ever live. */ |
6fb5fa3c | 6533 | (df_regs_ever_live_p (reg) |
62164eb4 R |
6534 | || (call_really_used_regs[reg] |
6535 | && (! fixed_regs[reg] || reg == MACH_REG || reg == MACL_REG | |
6536 | || reg == PIC_OFFSET_TABLE_REGNUM) | |
fada1961 | 6537 | && has_call) |
a6ab9fc0 R |
6538 | || (TARGET_SHMEDIA && has_call |
6539 | && REGISTER_NATURAL_MODE (reg) == SImode | |
fada1961 | 6540 | && (GENERAL_REGISTER_P (reg) || TARGET_REGISTER_P (reg)))) |
1245df60 R |
6541 | && reg != STACK_POINTER_REGNUM && reg != ARG_POINTER_REGNUM |
6542 | && reg != RETURN_ADDRESS_POINTER_REGNUM | |
7144b2d8 DD |
6543 | && reg != T_REG && reg != GBR_REG |
6544 | /* Push fpscr only on targets which have FPU */ | |
6545 | && (reg != FPSCR_REG || TARGET_FPU_ANY)) | |
1245df60 | 6546 | : (/* Only push those regs which are used and need to be saved. */ |
180bde4f RS |
6547 | (TARGET_SHCOMPACT |
6548 | && flag_pic | |
38173d38 | 6549 | && crtl->args.info.call_cookie |
0ef2c525 | 6550 | && reg == PIC_OFFSET_TABLE_REGNUM) |
6fb5fa3c | 6551 | || (df_regs_ever_live_p (reg) |
9fdcdc74 KK |
6552 | && ((!call_really_used_regs[reg] |
6553 | && !(reg != PIC_OFFSET_TABLE_REGNUM | |
6554 | && fixed_regs[reg] && call_used_regs[reg])) | |
a6ab9fc0 | 6555 | || (trapa_handler && reg == FPSCR_REG && TARGET_FPU_ANY))) |
e3b5732b | 6556 | || (crtl->calls_eh_return |
0ef2c525 KK |
6557 | && (reg == EH_RETURN_DATA_REGNO (0) |
6558 | || reg == EH_RETURN_DATA_REGNO (1) | |
6559 | || reg == EH_RETURN_DATA_REGNO (2) | |
6560 | || reg == EH_RETURN_DATA_REGNO (3))) | |
61f71b34 | 6561 | || ((reg == MACL_REG || reg == MACH_REG) |
6fb5fa3c | 6562 | && df_regs_ever_live_p (reg) |
61f71b34 DD |
6563 | && sh_cfun_attr_renesas_p ()) |
6564 | )) | |
8aa2a305 | 6565 | { |
490dbac7 | 6566 | SET_HARD_REG_BIT (*live_regs_mask, reg); |
fa5322fa AO |
6567 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); |
6568 | ||
157371cf | 6569 | if ((TARGET_SH4 || TARGET_SH2A_DOUBLE || TARGET_SH5) && TARGET_FMOVD |
fa5322fa | 6570 | && GET_MODE_CLASS (REGISTER_NATURAL_MODE (reg)) == MODE_FLOAT) |
eda44c78 | 6571 | { |
104ee20b | 6572 | if (FP_REGISTER_P (reg)) |
eda44c78 | 6573 | { |
6fb5fa3c | 6574 | if (! TARGET_FPU_SINGLE && ! df_regs_ever_live_p (reg ^ 1)) |
eda44c78 | 6575 | { |
490dbac7 | 6576 | SET_HARD_REG_BIT (*live_regs_mask, (reg ^ 1)); |
fa5322fa | 6577 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg ^ 1)); |
eda44c78 R |
6578 | } |
6579 | } | |
fa5322fa | 6580 | else if (XD_REGISTER_P (reg)) |
eda44c78 R |
6581 | { |
6582 | /* Must switch to double mode to access these registers. */ | |
c0fb94d7 | 6583 | target_flags &= ~MASK_FPU_SINGLE; |
eda44c78 R |
6584 | } |
6585 | } | |
8aa2a305 | 6586 | } |
a6ab9fc0 R |
6587 | if (nosave_low_regs && reg == R8_REG) |
6588 | break; | |
8aa2a305 | 6589 | } |
fe3ad572 SC |
6590 | /* If we have a target register optimization pass after prologue / epilogue |
6591 | threading, we need to assume all target registers will be live even if | |
6592 | they aren't now. */ | |
6593 | if (flag_branch_target_load_optimize2 | |
6594 | && TARGET_SAVE_ALL_TARGET_REGS | |
6595 | && shmedia_space_reserved_for_target_registers) | |
6596 | for (reg = LAST_TARGET_REG; reg >= FIRST_TARGET_REG; reg--) | |
62164eb4 | 6597 | if ((! call_really_used_regs[reg] || interrupt_handler) |
fe3ad572 SC |
6598 | && ! TEST_HARD_REG_BIT (*live_regs_mask, reg)) |
6599 | { | |
6600 | SET_HARD_REG_BIT (*live_regs_mask, reg); | |
6601 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); | |
6602 | } | |
fada1961 R |
6603 | /* If this is an interrupt handler, we don't have any call-clobbered |
6604 | registers we can conveniently use for target register save/restore. | |
6605 | Make sure we save at least one general purpose register when we need | |
6606 | to save target registers. */ | |
6607 | if (interrupt_handler | |
56b138ae RS |
6608 | && hard_reg_set_intersect_p (*live_regs_mask, |
6609 | reg_class_contents[TARGET_REGS]) | |
6610 | && ! hard_reg_set_intersect_p (*live_regs_mask, | |
6611 | reg_class_contents[GENERAL_REGS])) | |
fada1961 R |
6612 | { |
6613 | SET_HARD_REG_BIT (*live_regs_mask, R0_REG); | |
6614 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (R0_REG)); | |
6615 | } | |
8aa2a305 | 6616 | |
490dbac7 | 6617 | return count; |
8aa2a305 JW |
6618 | } |
6619 | ||
6620 | /* Code to generate prologue and epilogue sequences */ | |
b9654711 | 6621 | |
a93d1ba2 | 6622 | /* PUSHED is the number of bytes that are being pushed on the |
0ad913af R |
6623 | stack for register saves. Return the frame size, padded |
6624 | appropriately so that the stack stays properly aligned. */ | |
e5f8d4c0 | 6625 | static HOST_WIDE_INT |
cf277499 | 6626 | rounded_frame_size (int pushed) |
e5f8d4c0 R |
6627 | { |
6628 | HOST_WIDE_INT size = get_frame_size (); | |
6629 | HOST_WIDE_INT align = STACK_BOUNDARY / BITS_PER_UNIT; | |
6630 | ||
7a296495 CB |
6631 | if (ACCUMULATE_OUTGOING_ARGS) |
6632 | size += crtl->outgoing_args_size; | |
6633 | ||
c5b9ef02 | 6634 | return ((size + pushed + align - 1) & -align) - pushed; |
e5f8d4c0 R |
6635 | } |
6636 | ||
fa5322fa AO |
6637 | /* Choose a call-clobbered target-branch register that remains |
6638 | unchanged along the whole function. We set it up as the return | |
6639 | value in the prologue. */ | |
6640 | int | |
cf277499 | 6641 | sh_media_register_for_return (void) |
fa5322fa AO |
6642 | { |
6643 | int regno; | |
6644 | int tr0_used; | |
6645 | ||
6646 | if (! current_function_is_leaf) | |
6647 | return -1; | |
fada1961 R |
6648 | if (lookup_attribute ("interrupt_handler", |
6649 | DECL_ATTRIBUTES (current_function_decl))) | |
6650 | return -1; | |
73a4d10b R |
6651 | if (sh_cfun_interrupt_handler_p ()) |
6652 | return -1; | |
fa5322fa | 6653 | |
6fb5fa3c | 6654 | tr0_used = flag_pic && df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM); |
fa5322fa AO |
6655 | |
6656 | for (regno = FIRST_TARGET_REG + tr0_used; regno <= LAST_TARGET_REG; regno++) | |
6fb5fa3c | 6657 | if (call_really_used_regs[regno] && ! df_regs_ever_live_p (regno)) |
fa5322fa AO |
6658 | return regno; |
6659 | ||
6660 | return -1; | |
6661 | } | |
6662 | ||
fada1961 R |
6663 | /* The maximum registers we need to save are: |
6664 | - 62 general purpose registers (r15 is stack pointer, r63 is zero) | |
6665 | - 32 floating point registers (for each pair, we save none, | |
6666 | one single precision value, or a double precision value). | |
6667 | - 8 target registers | |
6668 | - add 1 entry for a delimiter. */ | |
6669 | #define MAX_SAVED_REGS (62+32+8) | |
6670 | ||
6671 | typedef struct save_entry_s | |
6672 | { | |
6673 | unsigned char reg; | |
6674 | unsigned char mode; | |
6675 | short offset; | |
6676 | } save_entry; | |
6677 | ||
6678 | #define MAX_TEMPS 4 | |
6679 | ||
6680 | /* There will be a delimiter entry with VOIDmode both at the start and the | |
6681 | end of a filled in schedule. The end delimiter has the offset of the | |
6682 | save with the smallest (i.e. most negative) offset. */ | |
6683 | typedef struct save_schedule_s | |
6684 | { | |
6685 | save_entry entries[MAX_SAVED_REGS + 2]; | |
6686 | int temps[MAX_TEMPS+1]; | |
6687 | } save_schedule; | |
6688 | ||
6689 | /* Fill in SCHEDULE according to LIVE_REGS_MASK. If RESTORE is nonzero, | |
6690 | use reverse order. Returns the last entry written to (not counting | |
6691 | the delimiter). OFFSET_BASE is a number to be added to all offset | |
6692 | entries. */ | |
f676971a | 6693 | |
fada1961 R |
6694 | static save_entry * |
6695 | sh5_schedule_saves (HARD_REG_SET *live_regs_mask, save_schedule *schedule, | |
6696 | int offset_base) | |
6697 | { | |
6698 | int align, i; | |
6699 | save_entry *entry = schedule->entries; | |
6700 | int tmpx = 0; | |
6701 | int offset; | |
6702 | ||
6703 | if (! current_function_interrupt) | |
6704 | for (i = FIRST_GENERAL_REG; tmpx < MAX_TEMPS && i <= LAST_GENERAL_REG; i++) | |
62164eb4 | 6705 | if (call_really_used_regs[i] && ! fixed_regs[i] && i != PR_MEDIA_REG |
fada1961 R |
6706 | && ! FUNCTION_ARG_REGNO_P (i) |
6707 | && i != FIRST_RET_REG | |
6de9cd9a | 6708 | && ! (cfun->static_chain_decl != NULL && i == STATIC_CHAIN_REGNUM) |
e3b5732b | 6709 | && ! (crtl->calls_eh_return |
fada1961 | 6710 | && (i == EH_RETURN_STACKADJ_REGNO |
f94cef0a KK |
6711 | || ((unsigned) i >= EH_RETURN_DATA_REGNO (0) |
6712 | && (unsigned) i <= EH_RETURN_DATA_REGNO (3))))) | |
fada1961 R |
6713 | schedule->temps[tmpx++] = i; |
6714 | entry->reg = -1; | |
6715 | entry->mode = VOIDmode; | |
6716 | entry->offset = offset_base; | |
6717 | entry++; | |
6718 | /* We loop twice: first, we save 8-byte aligned registers in the | |
6719 | higher addresses, that are known to be aligned. Then, we | |
6720 | proceed to saving 32-bit registers that don't need 8-byte | |
6721 | alignment. | |
6722 | If this is an interrupt function, all registers that need saving | |
6723 | need to be saved in full. moreover, we need to postpone saving | |
6724 | target registers till we have saved some general purpose registers | |
6725 | we can then use as scratch registers. */ | |
6726 | offset = offset_base; | |
6727 | for (align = 1; align >= 0; align--) | |
6728 | { | |
6729 | for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; i--) | |
6730 | if (TEST_HARD_REG_BIT (*live_regs_mask, i)) | |
6731 | { | |
6732 | enum machine_mode mode = REGISTER_NATURAL_MODE (i); | |
6733 | int reg = i; | |
6734 | ||
6735 | if (current_function_interrupt) | |
6736 | { | |
6737 | if (TARGET_REGISTER_P (i)) | |
6738 | continue; | |
6739 | if (GENERAL_REGISTER_P (i)) | |
6740 | mode = DImode; | |
6741 | } | |
6742 | if (mode == SFmode && (i % 2) == 1 | |
6743 | && ! TARGET_FPU_SINGLE && FP_REGISTER_P (i) | |
6744 | && (TEST_HARD_REG_BIT (*live_regs_mask, (i ^ 1)))) | |
6745 | { | |
6746 | mode = DFmode; | |
6747 | i--; | |
6748 | reg--; | |
6749 | } | |
6750 | ||
6751 | /* If we're doing the aligned pass and this is not aligned, | |
6752 | or we're doing the unaligned pass and this is aligned, | |
6753 | skip it. */ | |
6754 | if ((GET_MODE_SIZE (mode) % (STACK_BOUNDARY / BITS_PER_UNIT) == 0) | |
6755 | != align) | |
6756 | continue; | |
6757 | ||
6758 | if (current_function_interrupt | |
6759 | && GENERAL_REGISTER_P (i) | |
6760 | && tmpx < MAX_TEMPS) | |
6761 | schedule->temps[tmpx++] = i; | |
6762 | ||
6763 | offset -= GET_MODE_SIZE (mode); | |
6764 | entry->reg = i; | |
6765 | entry->mode = mode; | |
6766 | entry->offset = offset; | |
6767 | entry++; | |
6768 | } | |
6769 | if (align && current_function_interrupt) | |
6770 | for (i = LAST_TARGET_REG; i >= FIRST_TARGET_REG; i--) | |
6771 | if (TEST_HARD_REG_BIT (*live_regs_mask, i)) | |
6772 | { | |
6773 | offset -= GET_MODE_SIZE (DImode); | |
6774 | entry->reg = i; | |
6775 | entry->mode = DImode; | |
6776 | entry->offset = offset; | |
6777 | entry++; | |
6778 | } | |
6779 | } | |
6780 | entry->reg = -1; | |
6781 | entry->mode = VOIDmode; | |
6782 | entry->offset = offset; | |
6783 | schedule->temps[tmpx] = -1; | |
6784 | return entry - 1; | |
6785 | } | |
6786 | ||
b9654711 | 6787 | void |
cf277499 | 6788 | sh_expand_prologue (void) |
b9654711 | 6789 | { |
490dbac7 | 6790 | HARD_REG_SET live_regs_mask; |
40d2032b | 6791 | int d, i; |
fa5322fa | 6792 | int d_rounding = 0; |
225e4f43 | 6793 | int save_flags = target_flags; |
49616835 | 6794 | int pretend_args; |
a6ab9fc0 R |
6795 | tree sp_switch_attr |
6796 | = lookup_attribute ("sp_switch", DECL_ATTRIBUTES (current_function_decl)); | |
b9654711 | 6797 | |
e3ba8d11 | 6798 | current_function_interrupt = sh_cfun_interrupt_handler_p (); |
552ecbd9 | 6799 | |
0d7e008e | 6800 | /* We have pretend args if we had an object sent partially in registers |
8aa2a305 | 6801 | and partially on the stack, e.g. a large structure. */ |
38173d38 | 6802 | pretend_args = crtl->args.pretend_args_size; |
49616835 JR |
6803 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl) |
6804 | && (NPARM_REGS(SImode) | |
38173d38 | 6805 | > crtl->args.info.arg_count[(int) SH_ARG_INT])) |
49616835 | 6806 | pretend_args = 0; |
21d05694 | 6807 | /* Dwarf2 module doesn't expect frame related insns here. */ |
49616835 | 6808 | output_stack_adjust (-pretend_args |
38173d38 | 6809 | - crtl->args.info.stack_regs * 8, |
21d05694 | 6810 | stack_pointer_rtx, 0, NULL, false); |
b9654711 | 6811 | |
38173d38 | 6812 | if (TARGET_SHCOMPACT && flag_pic && crtl->args.info.call_cookie) |
2b1fd83f AO |
6813 | /* We're going to use the PIC register to load the address of the |
6814 | incoming-argument decoder and/or of the return trampoline from | |
6815 | the GOT, so make sure the PIC register is preserved and | |
6816 | initialized. */ | |
6fb5fa3c | 6817 | df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true); |
2b1fd83f | 6818 | |
fa5322fa | 6819 | if (TARGET_SHCOMPACT |
38173d38 | 6820 | && (crtl->args.info.call_cookie & ~ CALL_COOKIE_RET_TRAMP(1))) |
fa5322fa AO |
6821 | { |
6822 | int reg; | |
6823 | ||
6824 | /* First, make all registers with incoming arguments that will | |
6825 | be pushed onto the stack live, so that register renaming | |
6826 | doesn't overwrite them. */ | |
6827 | for (reg = 0; reg < NPARM_REGS (SImode); reg++) | |
38173d38 | 6828 | if (CALL_COOKIE_STACKSEQ_GET (crtl->args.info.call_cookie) |
fa5322fa AO |
6829 | >= NPARM_REGS (SImode) - reg) |
6830 | for (; reg < NPARM_REGS (SImode); reg++) | |
6831 | emit_insn (gen_shcompact_preserve_incoming_args | |
6832 | (gen_rtx_REG (SImode, FIRST_PARM_REG + reg))); | |
6833 | else if (CALL_COOKIE_INT_REG_GET | |
38173d38 | 6834 | (crtl->args.info.call_cookie, reg) == 1) |
fa5322fa AO |
6835 | emit_insn (gen_shcompact_preserve_incoming_args |
6836 | (gen_rtx_REG (SImode, FIRST_PARM_REG + reg))); | |
6837 | ||
6838 | emit_move_insn (gen_rtx_REG (Pmode, MACL_REG), | |
6839 | stack_pointer_rtx); | |
6840 | emit_move_insn (gen_rtx_REG (SImode, R0_REG), | |
38173d38 | 6841 | GEN_INT (crtl->args.info.call_cookie)); |
fa5322fa AO |
6842 | emit_move_insn (gen_rtx_REG (SImode, MACH_REG), |
6843 | gen_rtx_REG (SImode, R0_REG)); | |
6844 | } | |
6845 | else if (TARGET_SHMEDIA) | |
6846 | { | |
6847 | int tr = sh_media_register_for_return (); | |
6848 | ||
6849 | if (tr >= 0) | |
6fb5fa3c DB |
6850 | emit_move_insn (gen_rtx_REG (DImode, tr), |
6851 | gen_rtx_REG (DImode, PR_MEDIA_REG)); | |
fa5322fa AO |
6852 | } |
6853 | ||
7ab56274 | 6854 | /* Emit the code for SETUP_VARARGS. */ |
e3b5732b | 6855 | if (cfun->stdarg) |
b9654711 | 6856 | { |
49616835 | 6857 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl)) |
1245df60 | 6858 | { |
1a95a963 JW |
6859 | /* Push arg regs as if they'd been provided by caller in stack. */ |
6860 | for (i = 0; i < NPARM_REGS(SImode); i++) | |
6861 | { | |
6862 | int rn = NPARM_REGS(SImode) + FIRST_PARM_REG - i - 1; | |
2754d3c5 R |
6863 | rtx insn; |
6864 | ||
f676971a | 6865 | if (i >= (NPARM_REGS(SImode) |
38173d38 | 6866 | - crtl->args.info.arg_count[(int) SH_ARG_INT] |
aec373f1 | 6867 | )) |
1a95a963 | 6868 | break; |
2754d3c5 | 6869 | insn = push (rn); |
1a95a963 | 6870 | } |
1245df60 | 6871 | } |
b9654711 | 6872 | } |
1a95a963 | 6873 | |
4408efce | 6874 | /* If we're supposed to switch stacks at function entry, do so now. */ |
a6ab9fc0 R |
6875 | if (sp_switch_attr) |
6876 | { | |
20e19db4 | 6877 | rtx lab, newsrc; |
a6ab9fc0 R |
6878 | /* The argument specifies a variable holding the address of the |
6879 | stack the interrupt function should switch to/from at entry/exit. */ | |
20e19db4 | 6880 | tree arg = TREE_VALUE ( TREE_VALUE (sp_switch_attr)); |
a6ab9fc0 | 6881 | const char *s |
20e19db4 | 6882 | = ggc_strdup (TREE_STRING_POINTER (arg)); |
a6ab9fc0 R |
6883 | rtx sp_switch = gen_rtx_SYMBOL_REF (Pmode, s); |
6884 | ||
20e19db4 DD |
6885 | lab = add_constant (sp_switch, SImode, 0); |
6886 | newsrc = gen_rtx_LABEL_REF (VOIDmode, lab); | |
6887 | newsrc = gen_const_mem (SImode, newsrc); | |
6888 | ||
6889 | emit_insn (gen_sp_switch_1 (newsrc)); | |
a6ab9fc0 | 6890 | } |
4408efce | 6891 | |
490dbac7 | 6892 | d = calc_live_regs (&live_regs_mask); |
225e4f43 R |
6893 | /* ??? Maybe we could save some switching if we can move a mode switch |
6894 | that already happens to be at the function start into the prologue. */ | |
ac55736a | 6895 | if (target_flags != save_flags && ! current_function_interrupt) |
225e4f43 | 6896 | emit_insn (gen_toggle_sz ()); |
f676971a | 6897 | |
fa5322fa AO |
6898 | if (TARGET_SH5) |
6899 | { | |
fada1961 R |
6900 | int offset_base, offset; |
6901 | rtx r0 = NULL_RTX; | |
fa5322fa AO |
6902 | int offset_in_r0 = -1; |
6903 | int sp_in_r0 = 0; | |
fe3ad572 SC |
6904 | int tregs_space = shmedia_target_regs_stack_adjust (&live_regs_mask); |
6905 | int total_size, save_size; | |
fada1961 R |
6906 | save_schedule schedule; |
6907 | save_entry *entry; | |
6908 | int *tmp_pnt; | |
6909 | ||
62164eb4 | 6910 | if (call_really_used_regs[R0_REG] && ! fixed_regs[R0_REG] |
fada1961 R |
6911 | && ! current_function_interrupt) |
6912 | r0 = gen_rtx_REG (Pmode, R0_REG); | |
fe3ad572 SC |
6913 | |
6914 | /* D is the actual number of bytes that we need for saving registers, | |
6915 | however, in initial_elimination_offset we have committed to using | |
6916 | an additional TREGS_SPACE amount of bytes - in order to keep both | |
6917 | addresses to arguments supplied by the caller and local variables | |
6918 | valid, we must keep this gap. Place it between the incoming | |
6919 | arguments and the actually saved registers in a bid to optimize | |
6920 | locality of reference. */ | |
6921 | total_size = d + tregs_space; | |
6922 | total_size += rounded_frame_size (total_size); | |
6923 | save_size = total_size - rounded_frame_size (d); | |
6924 | if (save_size % (STACK_BOUNDARY / BITS_PER_UNIT)) | |
fa5322fa | 6925 | d_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) |
fe3ad572 SC |
6926 | - save_size % (STACK_BOUNDARY / BITS_PER_UNIT)); |
6927 | ||
6928 | /* If adjusting the stack in a single step costs nothing extra, do so. | |
6929 | I.e. either if a single addi is enough, or we need a movi anyway, | |
6930 | and we don't exceed the maximum offset range (the test for the | |
6931 | latter is conservative for simplicity). */ | |
6932 | if (TARGET_SHMEDIA | |
6933 | && (CONST_OK_FOR_I10 (-total_size) | |
6934 | || (! CONST_OK_FOR_I10 (-(save_size + d_rounding)) | |
6935 | && total_size <= 2044))) | |
6936 | d_rounding = total_size - save_size; | |
fa5322fa | 6937 | |
fada1961 | 6938 | offset_base = d + d_rounding; |
fe3ad572 SC |
6939 | |
6940 | output_stack_adjust (-(save_size + d_rounding), stack_pointer_rtx, | |
21d05694 | 6941 | 0, NULL, true); |
fa5322fa | 6942 | |
fada1961 R |
6943 | sh5_schedule_saves (&live_regs_mask, &schedule, offset_base); |
6944 | tmp_pnt = schedule.temps; | |
6945 | for (entry = &schedule.entries[1]; entry->mode != VOIDmode; entry++) | |
6946 | { | |
81f40b79 | 6947 | enum machine_mode mode = (enum machine_mode) entry->mode; |
0ef2c525 | 6948 | unsigned int reg = entry->reg; |
fada1961 | 6949 | rtx reg_rtx, mem_rtx, pre_dec = NULL_RTX; |
455786b6 | 6950 | rtx orig_reg_rtx; |
fa5322fa | 6951 | |
fada1961 | 6952 | offset = entry->offset; |
fa5322fa | 6953 | |
fada1961 | 6954 | reg_rtx = gen_rtx_REG (mode, reg); |
fa5322fa | 6955 | |
57d38024 R |
6956 | mem_rtx = gen_frame_mem (mode, |
6957 | gen_rtx_PLUS (Pmode, | |
6958 | stack_pointer_rtx, | |
6959 | GEN_INT (offset))); | |
fa5322fa | 6960 | |
6de7a512 PB |
6961 | if (!memory_address_p (mode, XEXP (mem_rtx, 0))) |
6962 | { | |
6963 | gcc_assert (r0); | |
6964 | mem_rtx = NULL_RTX; | |
6965 | } | |
fa5322fa | 6966 | |
6de7a512 PB |
6967 | if (HAVE_PRE_DECREMENT |
6968 | && (offset_in_r0 - offset == GET_MODE_SIZE (mode) | |
6969 | || mem_rtx == NULL_RTX | |
6970 | || reg == PR_REG || SPECIAL_REGISTER_P (reg))) | |
6971 | { | |
6972 | pre_dec = gen_frame_mem (mode, gen_rtx_PRE_DEC (Pmode, r0)); | |
fa5322fa | 6973 | |
6de7a512 | 6974 | if (!memory_address_p (mode, XEXP (pre_dec, 0))) |
fada1961 | 6975 | pre_dec = NULL_RTX; |
6de7a512 PB |
6976 | else |
6977 | { | |
6978 | mem_rtx = NULL_RTX; | |
6979 | offset += GET_MODE_SIZE (mode); | |
6980 | } | |
6981 | } | |
fa5322fa | 6982 | |
fada1961 R |
6983 | if (mem_rtx != NULL_RTX) |
6984 | goto addr_ok; | |
fa5322fa | 6985 | |
fada1961 R |
6986 | if (offset_in_r0 == -1) |
6987 | { | |
6988 | emit_move_insn (r0, GEN_INT (offset)); | |
6989 | offset_in_r0 = offset; | |
6990 | } | |
6991 | else if (offset != offset_in_r0) | |
6992 | { | |
6993 | emit_move_insn (r0, | |
6994 | gen_rtx_PLUS | |
6995 | (Pmode, r0, | |
6996 | GEN_INT (offset - offset_in_r0))); | |
6997 | offset_in_r0 += offset - offset_in_r0; | |
6998 | } | |
f676971a | 6999 | |
fada1961 R |
7000 | if (pre_dec != NULL_RTX) |
7001 | { | |
7002 | if (! sp_in_r0) | |
fa5322fa AO |
7003 | { |
7004 | emit_move_insn (r0, | |
7005 | gen_rtx_PLUS | |
fada1961 R |
7006 | (Pmode, r0, stack_pointer_rtx)); |
7007 | sp_in_r0 = 1; | |
fa5322fa | 7008 | } |
fa5322fa | 7009 | |
fada1961 R |
7010 | offset -= GET_MODE_SIZE (mode); |
7011 | offset_in_r0 -= GET_MODE_SIZE (mode); | |
fa5322fa | 7012 | |
fada1961 R |
7013 | mem_rtx = pre_dec; |
7014 | } | |
7015 | else if (sp_in_r0) | |
57d38024 | 7016 | mem_rtx = gen_frame_mem (mode, r0); |
fada1961 | 7017 | else |
57d38024 R |
7018 | mem_rtx = gen_frame_mem (mode, |
7019 | gen_rtx_PLUS (Pmode, | |
7020 | stack_pointer_rtx, | |
7021 | r0)); | |
fada1961 R |
7022 | |
7023 | /* We must not use an r0-based address for target-branch | |
7024 | registers or for special registers without pre-dec | |
7025 | memory addresses, since we store their values in r0 | |
7026 | first. */ | |
f5b9e7c9 NS |
7027 | gcc_assert (!TARGET_REGISTER_P (reg) |
7028 | && ((reg != PR_REG && !SPECIAL_REGISTER_P (reg)) | |
7029 | || mem_rtx == pre_dec)); | |
7030 | ||
fada1961 | 7031 | addr_ok: |
455786b6 | 7032 | orig_reg_rtx = reg_rtx; |
fada1961 R |
7033 | if (TARGET_REGISTER_P (reg) |
7034 | || ((reg == PR_REG || SPECIAL_REGISTER_P (reg)) | |
7035 | && mem_rtx != pre_dec)) | |
7036 | { | |
7037 | rtx tmp_reg = gen_rtx_REG (GET_MODE (reg_rtx), *tmp_pnt); | |
fa5322fa | 7038 | |
fada1961 | 7039 | emit_move_insn (tmp_reg, reg_rtx); |
fa5322fa | 7040 | |
fada1961 R |
7041 | if (REGNO (tmp_reg) == R0_REG) |
7042 | { | |
fa5322fa AO |
7043 | offset_in_r0 = -1; |
7044 | sp_in_r0 = 0; | |
f5b9e7c9 NS |
7045 | gcc_assert (!refers_to_regno_p |
7046 | (R0_REG, R0_REG+1, mem_rtx, (rtx *) 0)); | |
fa5322fa AO |
7047 | } |
7048 | ||
fada1961 R |
7049 | if (*++tmp_pnt <= 0) |
7050 | tmp_pnt = schedule.temps; | |
7051 | ||
7052 | reg_rtx = tmp_reg; | |
fa5322fa | 7053 | } |
fada1961 R |
7054 | { |
7055 | rtx insn; | |
7056 | ||
7057 | /* Mark as interesting for dwarf cfi generator */ | |
7058 | insn = emit_move_insn (mem_rtx, reg_rtx); | |
7059 | RTX_FRAME_RELATED_P (insn) = 1; | |
455786b6 KK |
7060 | /* If we use an intermediate register for the save, we can't |
7061 | describe this exactly in cfi as a copy of the to-be-saved | |
7062 | register into the temporary register and then the temporary | |
7063 | register on the stack, because the temporary register can | |
7064 | have a different natural size than the to-be-saved register. | |
7065 | Thus, we gloss over the intermediate copy and pretend we do | |
7066 | a direct save from the to-be-saved register. */ | |
7067 | if (REGNO (reg_rtx) != reg) | |
7068 | { | |
b757d36f | 7069 | rtx set; |
455786b6 KK |
7070 | |
7071 | set = gen_rtx_SET (VOIDmode, mem_rtx, orig_reg_rtx); | |
b757d36f | 7072 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, set); |
455786b6 | 7073 | } |
fada1961 | 7074 | |
f676971a | 7075 | if (TARGET_SHCOMPACT && (offset_in_r0 != -1)) |
fada1961 R |
7076 | { |
7077 | rtx reg_rtx = gen_rtx_REG (mode, reg); | |
b757d36f | 7078 | rtx set; |
57d38024 R |
7079 | rtx mem_rtx = gen_frame_mem (mode, |
7080 | gen_rtx_PLUS (Pmode, | |
7081 | stack_pointer_rtx, | |
7082 | GEN_INT (offset))); | |
fada1961 R |
7083 | |
7084 | set = gen_rtx_SET (VOIDmode, mem_rtx, reg_rtx); | |
b757d36f | 7085 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, set); |
fada1961 R |
7086 | } |
7087 | } | |
7088 | } | |
fa5322fa | 7089 | |
f5b9e7c9 | 7090 | gcc_assert (entry->offset == d_rounding); |
fa5322fa AO |
7091 | } |
7092 | else | |
ac55736a | 7093 | push_regs (&live_regs_mask, current_function_interrupt); |
1a66cd67 | 7094 | |
6fb5fa3c DB |
7095 | if (flag_pic && df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM)) |
7096 | emit_insn (gen_GOTaddr2picreg ()); | |
1a66cd67 | 7097 | |
fa5322fa AO |
7098 | if (SHMEDIA_REGS_STACK_ADJUST ()) |
7099 | { | |
fa5322fa AO |
7100 | /* This must NOT go through the PLT, otherwise mach and macl |
7101 | may be clobbered. */ | |
73a4d10b R |
7102 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
7103 | (TARGET_FPU_ANY | |
7104 | ? "__GCC_push_shmedia_regs" | |
7105 | : "__GCC_push_shmedia_regs_nofpu"), SFUNC_GOT); | |
fa5322fa AO |
7106 | emit_insn (gen_shmedia_save_restore_regs_compact |
7107 | (GEN_INT (-SHMEDIA_REGS_STACK_ADJUST ()))); | |
7108 | } | |
7109 | ||
ac55736a | 7110 | if (target_flags != save_flags && ! current_function_interrupt) |
6fb5fa3c | 7111 | emit_insn (gen_toggle_sz ()); |
1a95a963 | 7112 | |
225e4f43 R |
7113 | target_flags = save_flags; |
7114 | ||
fa5322fa | 7115 | output_stack_adjust (-rounded_frame_size (d) + d_rounding, |
21d05694 | 7116 | stack_pointer_rtx, 0, NULL, true); |
b9654711 SC |
7117 | |
7118 | if (frame_pointer_needed) | |
96a2347e | 7119 | frame_insn (GEN_MOV (hard_frame_pointer_rtx, stack_pointer_rtx)); |
fa5322fa | 7120 | |
fa5322fa | 7121 | if (TARGET_SHCOMPACT |
38173d38 | 7122 | && (crtl->args.info.call_cookie & ~ CALL_COOKIE_RET_TRAMP(1))) |
fa5322fa AO |
7123 | { |
7124 | /* This must NOT go through the PLT, otherwise mach and macl | |
7125 | may be clobbered. */ | |
73a4d10b R |
7126 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
7127 | "__GCC_shcompact_incoming_args", SFUNC_GOT); | |
fa5322fa AO |
7128 | emit_insn (gen_shcompact_incoming_args ()); |
7129 | } | |
b9654711 SC |
7130 | } |
7131 | ||
7132 | void | |
726d4cb7 | 7133 | sh_expand_epilogue (bool sibcall_p) |
b9654711 | 7134 | { |
490dbac7 | 7135 | HARD_REG_SET live_regs_mask; |
40d2032b | 7136 | int d, i; |
fa5322fa | 7137 | int d_rounding = 0; |
b9654711 | 7138 | |
225e4f43 | 7139 | int save_flags = target_flags; |
fe3ad572 | 7140 | int frame_size, save_size; |
ac55736a | 7141 | int fpscr_deferred = 0; |
726d4cb7 | 7142 | int e = sibcall_p ? -1 : 1; |
3d5a0820 | 7143 | |
490dbac7 | 7144 | d = calc_live_regs (&live_regs_mask); |
3d5a0820 | 7145 | |
fe3ad572 SC |
7146 | save_size = d; |
7147 | frame_size = rounded_frame_size (d); | |
7148 | ||
7149 | if (TARGET_SH5) | |
7150 | { | |
7151 | int tregs_space = shmedia_target_regs_stack_adjust (&live_regs_mask); | |
7152 | int total_size; | |
7153 | if (d % (STACK_BOUNDARY / BITS_PER_UNIT)) | |
7154 | d_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) | |
7155 | - d % (STACK_BOUNDARY / BITS_PER_UNIT)); | |
7156 | ||
7157 | total_size = d + tregs_space; | |
7158 | total_size += rounded_frame_size (total_size); | |
7159 | save_size = total_size - frame_size; | |
7160 | ||
7161 | /* If adjusting the stack in a single step costs nothing extra, do so. | |
7162 | I.e. either if a single addi is enough, or we need a movi anyway, | |
7163 | and we don't exceed the maximum offset range (the test for the | |
7164 | latter is conservative for simplicity). */ | |
7165 | if (TARGET_SHMEDIA | |
7166 | && ! frame_pointer_needed | |
7167 | && (CONST_OK_FOR_I10 (total_size) | |
7168 | || (! CONST_OK_FOR_I10 (save_size + d_rounding) | |
7169 | && total_size <= 2044))) | |
7170 | d_rounding = frame_size; | |
fa5322fa | 7171 | |
fe3ad572 SC |
7172 | frame_size -= d_rounding; |
7173 | } | |
16bea517 | 7174 | |
b9654711 | 7175 | if (frame_pointer_needed) |
46d81ffa | 7176 | { |
1964788a | 7177 | /* We must avoid scheduling the epilogue with previous basic blocks. |
47ea8d4e | 7178 | See PR/18032 and PR/40313. */ |
1964788a | 7179 | emit_insn (gen_blockage ()); |
96a2347e | 7180 | output_stack_adjust (frame_size, hard_frame_pointer_rtx, e, |
21d05694 | 7181 | &live_regs_mask, false); |
07f5b9aa JL |
7182 | |
7183 | /* We must avoid moving the stack pointer adjustment past code | |
7184 | which reads from the local frame, else an interrupt could | |
7185 | occur after the SP adjustment and clobber data in the local | |
7186 | frame. */ | |
7187 | emit_insn (gen_blockage ()); | |
96a2347e | 7188 | emit_insn (GEN_MOV (stack_pointer_rtx, hard_frame_pointer_rtx)); |
46d81ffa | 7189 | } |
3d5a0820 | 7190 | else if (frame_size) |
07f5b9aa JL |
7191 | { |
7192 | /* We must avoid moving the stack pointer adjustment past code | |
7193 | which reads from the local frame, else an interrupt could | |
7194 | occur after the SP adjustment and clobber data in the local | |
7195 | frame. */ | |
7196 | emit_insn (gen_blockage ()); | |
21d05694 DD |
7197 | output_stack_adjust (frame_size, stack_pointer_rtx, e, |
7198 | &live_regs_mask, false); | |
07f5b9aa | 7199 | } |
b9654711 | 7200 | |
fa5322fa AO |
7201 | if (SHMEDIA_REGS_STACK_ADJUST ()) |
7202 | { | |
73a4d10b R |
7203 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
7204 | (TARGET_FPU_ANY | |
7205 | ? "__GCC_pop_shmedia_regs" | |
7206 | : "__GCC_pop_shmedia_regs_nofpu"), SFUNC_GOT); | |
fa5322fa AO |
7207 | /* This must NOT go through the PLT, otherwise mach and macl |
7208 | may be clobbered. */ | |
7209 | emit_insn (gen_shmedia_save_restore_regs_compact | |
7210 | (GEN_INT (SHMEDIA_REGS_STACK_ADJUST ()))); | |
7211 | } | |
7212 | ||
16bea517 | 7213 | /* Pop all the registers. */ |
0d7e008e | 7214 | |
ac55736a | 7215 | if (target_flags != save_flags && ! current_function_interrupt) |
225e4f43 | 7216 | emit_insn (gen_toggle_sz ()); |
fa5322fa AO |
7217 | if (TARGET_SH5) |
7218 | { | |
fada1961 | 7219 | int offset_base, offset; |
fa5322fa AO |
7220 | int offset_in_r0 = -1; |
7221 | int sp_in_r0 = 0; | |
fa5322fa | 7222 | rtx r0 = gen_rtx_REG (Pmode, R0_REG); |
fada1961 R |
7223 | save_schedule schedule; |
7224 | save_entry *entry; | |
7225 | int *tmp_pnt; | |
7d73a2ba | 7226 | |
fada1961 R |
7227 | entry = sh5_schedule_saves (&live_regs_mask, &schedule, d_rounding); |
7228 | offset_base = -entry[1].offset + d_rounding; | |
7229 | tmp_pnt = schedule.temps; | |
7230 | for (; entry->mode != VOIDmode; entry--) | |
7231 | { | |
81f40b79 | 7232 | enum machine_mode mode = (enum machine_mode) entry->mode; |
fada1961 R |
7233 | int reg = entry->reg; |
7234 | rtx reg_rtx, mem_rtx, post_inc = NULL_RTX, insn; | |
fa5322fa | 7235 | |
fada1961 R |
7236 | offset = offset_base + entry->offset; |
7237 | reg_rtx = gen_rtx_REG (mode, reg); | |
fa5322fa | 7238 | |
57d38024 R |
7239 | mem_rtx = gen_frame_mem (mode, |
7240 | gen_rtx_PLUS (Pmode, | |
7241 | stack_pointer_rtx, | |
7242 | GEN_INT (offset))); | |
fa5322fa | 7243 | |
6de7a512 PB |
7244 | if (!memory_address_p (mode, XEXP (mem_rtx, 0))) |
7245 | mem_rtx = NULL_RTX; | |
fa5322fa | 7246 | |
6de7a512 PB |
7247 | if (HAVE_POST_INCREMENT |
7248 | && (offset == offset_in_r0 | |
7249 | || (offset + GET_MODE_SIZE (mode) != d + d_rounding | |
7250 | && mem_rtx == NULL_RTX) | |
7251 | || reg == PR_REG || SPECIAL_REGISTER_P (reg))) | |
7252 | { | |
7253 | post_inc = gen_frame_mem (mode, gen_rtx_POST_INC (Pmode, r0)); | |
fa5322fa | 7254 | |
6de7a512 | 7255 | if (!memory_address_p (mode, XEXP (post_inc, 0))) |
fada1961 | 7256 | post_inc = NULL_RTX; |
6de7a512 | 7257 | else |
fada1961 | 7258 | mem_rtx = NULL_RTX; |
6de7a512 | 7259 | } |
f676971a | 7260 | |
fada1961 R |
7261 | if (mem_rtx != NULL_RTX) |
7262 | goto addr_ok; | |
fa5322fa | 7263 | |
fada1961 R |
7264 | if (offset_in_r0 == -1) |
7265 | { | |
7266 | emit_move_insn (r0, GEN_INT (offset)); | |
7267 | offset_in_r0 = offset; | |
7268 | } | |
7269 | else if (offset != offset_in_r0) | |
7270 | { | |
7271 | emit_move_insn (r0, | |
7272 | gen_rtx_PLUS | |
7273 | (Pmode, r0, | |
7274 | GEN_INT (offset - offset_in_r0))); | |
7275 | offset_in_r0 += offset - offset_in_r0; | |
7276 | } | |
f676971a | 7277 | |
fada1961 R |
7278 | if (post_inc != NULL_RTX) |
7279 | { | |
7280 | if (! sp_in_r0) | |
fa5322fa AO |
7281 | { |
7282 | emit_move_insn (r0, | |
7283 | gen_rtx_PLUS | |
fada1961 R |
7284 | (Pmode, r0, stack_pointer_rtx)); |
7285 | sp_in_r0 = 1; | |
fa5322fa | 7286 | } |
f676971a | 7287 | |
fada1961 | 7288 | mem_rtx = post_inc; |
fa5322fa | 7289 | |
fada1961 R |
7290 | offset_in_r0 += GET_MODE_SIZE (mode); |
7291 | } | |
7292 | else if (sp_in_r0) | |
57d38024 | 7293 | mem_rtx = gen_frame_mem (mode, r0); |
fada1961 | 7294 | else |
57d38024 R |
7295 | mem_rtx = gen_frame_mem (mode, |
7296 | gen_rtx_PLUS (Pmode, | |
7297 | stack_pointer_rtx, | |
7298 | r0)); | |
fa5322fa | 7299 | |
f5b9e7c9 NS |
7300 | gcc_assert ((reg != PR_REG && !SPECIAL_REGISTER_P (reg)) |
7301 | || mem_rtx == post_inc); | |
fa5322fa | 7302 | |
fada1961 R |
7303 | addr_ok: |
7304 | if ((reg == PR_REG || SPECIAL_REGISTER_P (reg)) | |
7305 | && mem_rtx != post_inc) | |
7306 | { | |
7307 | insn = emit_move_insn (r0, mem_rtx); | |
7308 | mem_rtx = r0; | |
7309 | } | |
7310 | else if (TARGET_REGISTER_P (reg)) | |
7311 | { | |
7312 | rtx tmp_reg = gen_rtx_REG (mode, *tmp_pnt); | |
7313 | ||
7314 | /* Give the scheduler a bit of freedom by using up to | |
7315 | MAX_TEMPS registers in a round-robin fashion. */ | |
7316 | insn = emit_move_insn (tmp_reg, mem_rtx); | |
7317 | mem_rtx = tmp_reg; | |
7318 | if (*++tmp_pnt < 0) | |
7319 | tmp_pnt = schedule.temps; | |
fa5322fa AO |
7320 | } |
7321 | ||
fada1961 | 7322 | insn = emit_move_insn (reg_rtx, mem_rtx); |
fada1961 R |
7323 | } |
7324 | ||
f5b9e7c9 | 7325 | gcc_assert (entry->offset + offset_base == d + d_rounding); |
fa5322fa | 7326 | } |
fe3ad572 | 7327 | else /* ! TARGET_SH5 */ |
b9654711 | 7328 | { |
5c7cafa8 CB |
7329 | int last_reg; |
7330 | ||
fe3ad572 | 7331 | save_size = 0; |
561642fa AP |
7332 | /* For an ISR with RESBANK attribute assigned, don't pop PR |
7333 | register. */ | |
7334 | if (TEST_HARD_REG_BIT (live_regs_mask, PR_REG) | |
7335 | && !sh_cfun_resbank_handler_p ()) | |
6fb5fa3c DB |
7336 | { |
7337 | if (!frame_pointer_needed) | |
7338 | emit_insn (gen_blockage ()); | |
7339 | pop (PR_REG); | |
7340 | } | |
5c7cafa8 | 7341 | |
6a62ca52 | 7342 | /* Banked registers are popped first to avoid being scheduled in the |
5c7cafa8 CB |
7343 | delay slot. RTE switches banks before the ds instruction. */ |
7344 | if (current_function_interrupt) | |
7345 | { | |
6a62ca52 DD |
7346 | for (i = LAST_BANKED_REG; i >= FIRST_BANKED_REG; i--) |
7347 | if (TEST_HARD_REG_BIT (live_regs_mask, i)) | |
7348 | pop (i); | |
5c7cafa8 CB |
7349 | |
7350 | last_reg = FIRST_PSEUDO_REGISTER - LAST_BANKED_REG - 1; | |
7351 | } | |
7352 | else | |
7353 | last_reg = FIRST_PSEUDO_REGISTER; | |
7354 | ||
7355 | for (i = 0; i < last_reg; i++) | |
fe3ad572 SC |
7356 | { |
7357 | int j = (FIRST_PSEUDO_REGISTER - 1) - i; | |
f676971a | 7358 | |
fe3ad572 | 7359 | if (j == FPSCR_REG && current_function_interrupt && TARGET_FMOVD |
56b138ae RS |
7360 | && hard_reg_set_intersect_p (live_regs_mask, |
7361 | reg_class_contents[DF_REGS])) | |
fe3ad572 | 7362 | fpscr_deferred = 1; |
561642fa AP |
7363 | /* For an ISR with RESBANK attribute assigned, don't pop |
7364 | following registers, R0-R14, MACH, MACL and GBR. */ | |
7365 | else if (j != PR_REG && TEST_HARD_REG_BIT (live_regs_mask, j) | |
7366 | && ! (sh_cfun_resbank_handler_p () | |
7367 | && ((j >= FIRST_GENERAL_REG | |
7368 | && j < LAST_GENERAL_REG) | |
7369 | || j == MACH_REG | |
7370 | || j == MACL_REG | |
7371 | || j == GBR_REG))) | |
fe3ad572 | 7372 | pop (j); |
5c7cafa8 | 7373 | |
fe3ad572 SC |
7374 | if (j == FIRST_FP_REG && fpscr_deferred) |
7375 | pop (FPSCR_REG); | |
fe3ad572 | 7376 | } |
b9654711 | 7377 | } |
ac55736a | 7378 | if (target_flags != save_flags && ! current_function_interrupt) |
225e4f43 R |
7379 | emit_insn (gen_toggle_sz ()); |
7380 | target_flags = save_flags; | |
b9654711 | 7381 | |
38173d38 | 7382 | output_stack_adjust (crtl->args.pretend_args_size |
fe3ad572 | 7383 | + save_size + d_rounding |
38173d38 | 7384 | + crtl->args.info.stack_regs * 8, |
21d05694 | 7385 | stack_pointer_rtx, e, NULL, false); |
4408efce | 7386 | |
e3b5732b | 7387 | if (crtl->calls_eh_return) |
4977bab6 ZW |
7388 | emit_insn (GEN_ADD3 (stack_pointer_rtx, stack_pointer_rtx, |
7389 | EH_RETURN_STACKADJ_RTX)); | |
7390 | ||
4408efce | 7391 | /* Switch back to the normal stack if necessary. */ |
a6ab9fc0 | 7392 | if (lookup_attribute ("sp_switch", DECL_ATTRIBUTES (current_function_decl))) |
4408efce | 7393 | emit_insn (gen_sp_switch_2 ()); |
89d4bc35 AO |
7394 | |
7395 | /* Tell flow the insn that pops PR isn't dead. */ | |
fa5322fa AO |
7396 | /* PR_REG will never be live in SHmedia mode, and we don't need to |
7397 | USE PR_MEDIA_REG, since it will be explicitly copied to TR0_REG | |
7398 | by the return pattern. */ | |
490dbac7 | 7399 | if (TEST_HARD_REG_BIT (live_regs_mask, PR_REG)) |
c41c1387 | 7400 | emit_use (gen_rtx_REG (SImode, PR_REG)); |
b9654711 SC |
7401 | } |
7402 | ||
afbc2905 R |
7403 | static int sh_need_epilogue_known = 0; |
7404 | ||
7405 | int | |
cf277499 | 7406 | sh_need_epilogue (void) |
afbc2905 R |
7407 | { |
7408 | if (! sh_need_epilogue_known) | |
7409 | { | |
7410 | rtx epilogue; | |
7411 | ||
7412 | start_sequence (); | |
726d4cb7 | 7413 | sh_expand_epilogue (0); |
2f937369 | 7414 | epilogue = get_insns (); |
afbc2905 | 7415 | end_sequence (); |
2f937369 | 7416 | sh_need_epilogue_known = (epilogue == NULL ? -1 : 1); |
afbc2905 R |
7417 | } |
7418 | return sh_need_epilogue_known > 0; | |
7419 | } | |
7420 | ||
4977bab6 ZW |
7421 | /* Emit code to change the current function's return address to RA. |
7422 | TEMP is available as a scratch register, if needed. */ | |
7423 | ||
7424 | void | |
cf277499 | 7425 | sh_set_return_address (rtx ra, rtx tmp) |
4977bab6 | 7426 | { |
490dbac7 | 7427 | HARD_REG_SET live_regs_mask; |
4977bab6 | 7428 | int d; |
4977bab6 ZW |
7429 | int pr_reg = TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG; |
7430 | int pr_offset; | |
7431 | ||
490dbac7 | 7432 | d = calc_live_regs (&live_regs_mask); |
4977bab6 ZW |
7433 | |
7434 | /* If pr_reg isn't life, we can set it (or the register given in | |
7435 | sh_media_register_for_return) directly. */ | |
490dbac7 | 7436 | if (! TEST_HARD_REG_BIT (live_regs_mask, pr_reg)) |
4977bab6 ZW |
7437 | { |
7438 | rtx rr; | |
7439 | ||
7440 | if (TARGET_SHMEDIA) | |
7441 | { | |
7442 | int rr_regno = sh_media_register_for_return (); | |
7443 | ||
7444 | if (rr_regno < 0) | |
7445 | rr_regno = pr_reg; | |
7446 | ||
7447 | rr = gen_rtx_REG (DImode, rr_regno); | |
7448 | } | |
7449 | else | |
7450 | rr = gen_rtx_REG (SImode, pr_reg); | |
7451 | ||
7452 | emit_insn (GEN_MOV (rr, ra)); | |
7453 | /* Tell flow the register for return isn't dead. */ | |
c41c1387 | 7454 | emit_use (rr); |
4977bab6 ZW |
7455 | return; |
7456 | } | |
7457 | ||
7458 | if (TARGET_SH5) | |
7459 | { | |
4977bab6 | 7460 | int offset; |
fada1961 R |
7461 | save_schedule schedule; |
7462 | save_entry *entry; | |
f676971a | 7463 | |
fada1961 R |
7464 | entry = sh5_schedule_saves (&live_regs_mask, &schedule, 0); |
7465 | offset = entry[1].offset; | |
7466 | for (; entry->mode != VOIDmode; entry--) | |
7467 | if (entry->reg == pr_reg) | |
7468 | goto found; | |
4977bab6 ZW |
7469 | |
7470 | /* We can't find pr register. */ | |
f5b9e7c9 | 7471 | gcc_unreachable (); |
4977bab6 ZW |
7472 | |
7473 | found: | |
fada1961 R |
7474 | offset = entry->offset - offset; |
7475 | pr_offset = (rounded_frame_size (d) + offset | |
4977bab6 ZW |
7476 | + SHMEDIA_REGS_STACK_ADJUST ()); |
7477 | } | |
7478 | else | |
fada1961 | 7479 | pr_offset = rounded_frame_size (d); |
4977bab6 ZW |
7480 | |
7481 | emit_insn (GEN_MOV (tmp, GEN_INT (pr_offset))); | |
7a296495 CB |
7482 | |
7483 | if (frame_pointer_needed) | |
7484 | emit_insn (GEN_ADD3 (tmp, tmp, hard_frame_pointer_rtx)); | |
7485 | else | |
7486 | emit_insn (GEN_ADD3 (tmp, tmp, stack_pointer_rtx)); | |
4977bab6 | 7487 | |
57d38024 | 7488 | tmp = gen_frame_mem (Pmode, tmp); |
4977bab6 | 7489 | emit_insn (GEN_MOV (tmp, ra)); |
7f9a14e3 KK |
7490 | /* Tell this store isn't dead. */ |
7491 | emit_use (tmp); | |
4977bab6 ZW |
7492 | } |
7493 | ||
8aa2a305 JW |
7494 | /* Clear variables at function end. */ |
7495 | ||
08c148a8 | 7496 | static void |
cf277499 SB |
7497 | sh_output_function_epilogue (FILE *file ATTRIBUTE_UNUSED, |
7498 | HOST_WIDE_INT size ATTRIBUTE_UNUSED) | |
8aa2a305 | 7499 | { |
afbc2905 | 7500 | sh_need_epilogue_known = 0; |
8aa2a305 JW |
7501 | } |
7502 | ||
61f71b34 | 7503 | static rtx |
cf277499 | 7504 | sh_builtin_saveregs (void) |
1a95a963 | 7505 | { |
1a95a963 | 7506 | /* First unnamed integer register. */ |
38173d38 | 7507 | int first_intreg = crtl->args.info.arg_count[(int) SH_ARG_INT]; |
1a95a963 JW |
7508 | /* Number of integer registers we need to save. */ |
7509 | int n_intregs = MAX (0, NPARM_REGS (SImode) - first_intreg); | |
7510 | /* First unnamed SFmode float reg */ | |
38173d38 | 7511 | int first_floatreg = crtl->args.info.arg_count[(int) SH_ARG_FLOAT]; |
1a95a963 JW |
7512 | /* Number of SFmode float regs to save. */ |
7513 | int n_floatregs = MAX (0, NPARM_REGS (SFmode) - first_floatreg); | |
318881c0 | 7514 | rtx regbuf, fpregs; |
ba4828e0 | 7515 | int bufsize, regno; |
4862826d | 7516 | alias_set_type alias_set; |
1a95a963 | 7517 | |
fa5322fa AO |
7518 | if (TARGET_SH5) |
7519 | { | |
7520 | if (n_intregs) | |
7521 | { | |
7522 | int pushregs = n_intregs; | |
7523 | ||
7524 | while (pushregs < NPARM_REGS (SImode) - 1 | |
7525 | && (CALL_COOKIE_INT_REG_GET | |
38173d38 | 7526 | (crtl->args.info.call_cookie, |
fa5322fa AO |
7527 | NPARM_REGS (SImode) - pushregs) |
7528 | == 1)) | |
7529 | { | |
38173d38 | 7530 | crtl->args.info.call_cookie |
fa5322fa AO |
7531 | &= ~ CALL_COOKIE_INT_REG (NPARM_REGS (SImode) |
7532 | - pushregs, 1); | |
7533 | pushregs++; | |
7534 | } | |
7535 | ||
7536 | if (pushregs == NPARM_REGS (SImode)) | |
38173d38 | 7537 | crtl->args.info.call_cookie |
fa5322fa AO |
7538 | |= (CALL_COOKIE_INT_REG (0, 1) |
7539 | | CALL_COOKIE_STACKSEQ (pushregs - 1)); | |
7540 | else | |
38173d38 | 7541 | crtl->args.info.call_cookie |
fa5322fa AO |
7542 | |= CALL_COOKIE_STACKSEQ (pushregs); |
7543 | ||
38173d38 | 7544 | crtl->args.pretend_args_size += 8 * n_intregs; |
fa5322fa AO |
7545 | } |
7546 | if (TARGET_SHCOMPACT) | |
7547 | return const0_rtx; | |
7548 | } | |
f676971a | 7549 | |
3a8699c7 | 7550 | if (! TARGET_SH2E && ! TARGET_SH4 && ! TARGET_SH5) |
fa5322fa AO |
7551 | { |
7552 | error ("__builtin_saveregs not supported by this subtarget"); | |
7553 | return const0_rtx; | |
7554 | } | |
7555 | ||
7556 | if (TARGET_SHMEDIA) | |
7557 | n_floatregs = 0; | |
7558 | ||
6f317ef3 | 7559 | /* Allocate block of memory for the regs. */ |
1a95a963 JW |
7560 | /* ??? If n_intregs + n_floatregs == 0, should we allocate at least 1 byte? |
7561 | Or can assign_stack_local accept a 0 SIZE argument? */ | |
7562 | bufsize = (n_intregs * UNITS_PER_WORD) + (n_floatregs * UNITS_PER_WORD); | |
7563 | ||
fa5322fa | 7564 | if (TARGET_SHMEDIA) |
57d38024 | 7565 | regbuf = gen_frame_mem (BLKmode, gen_rtx_REG (Pmode, ARG_POINTER_REGNUM)); |
f78ec36a R |
7566 | else if (n_floatregs & 1) |
7567 | { | |
7568 | rtx addr; | |
7569 | ||
7570 | regbuf = assign_stack_local (BLKmode, bufsize + UNITS_PER_WORD, 0); | |
7571 | addr = copy_to_mode_reg (Pmode, XEXP (regbuf, 0)); | |
7572 | emit_insn (gen_iorsi3 (addr, addr, GEN_INT (UNITS_PER_WORD))); | |
7573 | regbuf = change_address (regbuf, BLKmode, addr); | |
7574 | } | |
a11c004b R |
7575 | else if (STACK_BOUNDARY < 64 && TARGET_FPU_DOUBLE && n_floatregs) |
7576 | { | |
7577 | rtx addr, mask; | |
7578 | ||
7579 | regbuf = assign_stack_local (BLKmode, bufsize + UNITS_PER_WORD, 0); | |
7580 | addr = copy_to_mode_reg (Pmode, plus_constant (XEXP (regbuf, 0), 4)); | |
7581 | mask = copy_to_mode_reg (Pmode, GEN_INT (-8)); | |
7582 | emit_insn (gen_andsi3 (addr, addr, mask)); | |
7583 | regbuf = change_address (regbuf, BLKmode, addr); | |
7584 | } | |
fa5322fa | 7585 | else |
a11c004b | 7586 | regbuf = assign_stack_local (BLKmode, bufsize, TARGET_FPU_DOUBLE ? 64 : 0); |
514066a1 | 7587 | alias_set = get_varargs_alias_set (); |
ba4828e0 | 7588 | set_mem_alias_set (regbuf, alias_set); |
1a95a963 JW |
7589 | |
7590 | /* Save int args. | |
7591 | This is optimized to only save the regs that are necessary. Explicitly | |
7592 | named args need not be saved. */ | |
7593 | if (n_intregs > 0) | |
7594 | move_block_from_reg (BASE_ARG_REG (SImode) + first_intreg, | |
f4ef873c RK |
7595 | adjust_address (regbuf, BLKmode, |
7596 | n_floatregs * UNITS_PER_WORD), | |
c6b97fac | 7597 | n_intregs); |
1a95a963 | 7598 | |
fa5322fa AO |
7599 | if (TARGET_SHMEDIA) |
7600 | /* Return the address of the regbuf. */ | |
7601 | return XEXP (regbuf, 0); | |
7602 | ||
1a95a963 JW |
7603 | /* Save float args. |
7604 | This is optimized to only save the regs that are necessary. Explicitly | |
7605 | named args need not be saved. | |
7606 | We explicitly build a pointer to the buffer because it halves the insn | |
7607 | count when not optimizing (otherwise the pointer is built for each reg | |
1245df60 R |
7608 | saved). |
7609 | We emit the moves in reverse order so that we can use predecrement. */ | |
1a95a963 | 7610 | |
7d9c5aa1 R |
7611 | fpregs = copy_to_mode_reg (Pmode, |
7612 | plus_constant (XEXP (regbuf, 0), | |
7613 | n_floatregs * UNITS_PER_WORD)); | |
157371cf | 7614 | if (TARGET_SH4 || TARGET_SH2A_DOUBLE) |
225e4f43 | 7615 | { |
514066a1 | 7616 | rtx mem; |
225e4f43 R |
7617 | for (regno = NPARM_REGS (DFmode) - 2; regno >= first_floatreg; regno -= 2) |
7618 | { | |
7619 | emit_insn (gen_addsi3 (fpregs, fpregs, | |
7620 | GEN_INT (-2 * UNITS_PER_WORD))); | |
57d38024 | 7621 | mem = change_address (regbuf, DFmode, fpregs); |
f676971a | 7622 | emit_move_insn (mem, |
c0d4e710 | 7623 | gen_rtx_REG (DFmode, BASE_ARG_REG (DFmode) + regno)); |
225e4f43 R |
7624 | } |
7625 | regno = first_floatreg; | |
7626 | if (regno & 1) | |
7627 | { | |
832a3292 | 7628 | emit_insn (gen_addsi3 (fpregs, fpregs, GEN_INT (-UNITS_PER_WORD))); |
57d38024 | 7629 | mem = change_address (regbuf, SFmode, fpregs); |
514066a1 | 7630 | emit_move_insn (mem, |
c0d4e710 | 7631 | gen_rtx_REG (SFmode, BASE_ARG_REG (SFmode) + regno |
225e4f43 R |
7632 | - (TARGET_LITTLE_ENDIAN != 0))); |
7633 | } | |
7634 | } | |
7635 | else | |
1245df60 R |
7636 | for (regno = NPARM_REGS (SFmode) - 1; regno >= first_floatreg; regno--) |
7637 | { | |
514066a1 | 7638 | rtx mem; |
ba4828e0 | 7639 | |
832a3292 | 7640 | emit_insn (gen_addsi3 (fpregs, fpregs, GEN_INT (-UNITS_PER_WORD))); |
57d38024 | 7641 | mem = change_address (regbuf, SFmode, fpregs); |
514066a1 | 7642 | emit_move_insn (mem, |
c5c76735 | 7643 | gen_rtx_REG (SFmode, BASE_ARG_REG (SFmode) + regno)); |
1245df60 | 7644 | } |
1a95a963 JW |
7645 | |
7646 | /* Return the address of the regbuf. */ | |
7647 | return XEXP (regbuf, 0); | |
7648 | } | |
7649 | ||
514066a1 RH |
7650 | /* Define the `__builtin_va_list' type for the ABI. */ |
7651 | ||
c35d187f RH |
7652 | static tree |
7653 | sh_build_builtin_va_list (void) | |
514066a1 RH |
7654 | { |
7655 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
10575b5d | 7656 | tree record, type_decl; |
514066a1 | 7657 | |
61f71b34 DD |
7658 | if (TARGET_SH5 || (! TARGET_SH2E && ! TARGET_SH4) |
7659 | || TARGET_HITACHI || sh_cfun_attr_renesas_p ()) | |
514066a1 RH |
7660 | return ptr_type_node; |
7661 | ||
5d27ef94 | 7662 | record = (*lang_hooks.types.make_type) (RECORD_TYPE); |
10575b5d KK |
7663 | type_decl = build_decl (BUILTINS_LOCATION, |
7664 | TYPE_DECL, get_identifier ("__va_list_tag"), record); | |
514066a1 | 7665 | |
4c4bde29 AH |
7666 | f_next_o = build_decl (BUILTINS_LOCATION, |
7667 | FIELD_DECL, get_identifier ("__va_next_o"), | |
514066a1 | 7668 | ptr_type_node); |
4c4bde29 AH |
7669 | f_next_o_limit = build_decl (BUILTINS_LOCATION, |
7670 | FIELD_DECL, | |
514066a1 RH |
7671 | get_identifier ("__va_next_o_limit"), |
7672 | ptr_type_node); | |
4c4bde29 AH |
7673 | f_next_fp = build_decl (BUILTINS_LOCATION, |
7674 | FIELD_DECL, get_identifier ("__va_next_fp"), | |
514066a1 | 7675 | ptr_type_node); |
4c4bde29 AH |
7676 | f_next_fp_limit = build_decl (BUILTINS_LOCATION, |
7677 | FIELD_DECL, | |
514066a1 RH |
7678 | get_identifier ("__va_next_fp_limit"), |
7679 | ptr_type_node); | |
4c4bde29 AH |
7680 | f_next_stack = build_decl (BUILTINS_LOCATION, |
7681 | FIELD_DECL, get_identifier ("__va_next_stack"), | |
514066a1 RH |
7682 | ptr_type_node); |
7683 | ||
7684 | DECL_FIELD_CONTEXT (f_next_o) = record; | |
7685 | DECL_FIELD_CONTEXT (f_next_o_limit) = record; | |
7686 | DECL_FIELD_CONTEXT (f_next_fp) = record; | |
7687 | DECL_FIELD_CONTEXT (f_next_fp_limit) = record; | |
7688 | DECL_FIELD_CONTEXT (f_next_stack) = record; | |
7689 | ||
10575b5d KK |
7690 | TREE_CHAIN (record) = type_decl; |
7691 | TYPE_NAME (record) = type_decl; | |
514066a1 RH |
7692 | TYPE_FIELDS (record) = f_next_o; |
7693 | TREE_CHAIN (f_next_o) = f_next_o_limit; | |
7694 | TREE_CHAIN (f_next_o_limit) = f_next_fp; | |
7695 | TREE_CHAIN (f_next_fp) = f_next_fp_limit; | |
7696 | TREE_CHAIN (f_next_fp_limit) = f_next_stack; | |
7697 | ||
7698 | layout_type (record); | |
7699 | ||
7700 | return record; | |
7701 | } | |
7702 | ||
7703 | /* Implement `va_start' for varargs and stdarg. */ | |
7704 | ||
d7bd8aeb | 7705 | static void |
cf277499 | 7706 | sh_va_start (tree valist, rtx nextarg) |
514066a1 RH |
7707 | { |
7708 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
7709 | tree next_o, next_o_limit, next_fp, next_fp_limit, next_stack; | |
7710 | tree t, u; | |
7711 | int nfp, nint; | |
7712 | ||
fa5322fa AO |
7713 | if (TARGET_SH5) |
7714 | { | |
7715 | expand_builtin_saveregs (); | |
e5faf155 | 7716 | std_expand_builtin_va_start (valist, nextarg); |
fa5322fa AO |
7717 | return; |
7718 | } | |
7719 | ||
61f71b34 DD |
7720 | if ((! TARGET_SH2E && ! TARGET_SH4) |
7721 | || TARGET_HITACHI || sh_cfun_attr_renesas_p ()) | |
514066a1 | 7722 | { |
e5faf155 | 7723 | std_expand_builtin_va_start (valist, nextarg); |
514066a1 RH |
7724 | return; |
7725 | } | |
7726 | ||
7727 | f_next_o = TYPE_FIELDS (va_list_type_node); | |
7728 | f_next_o_limit = TREE_CHAIN (f_next_o); | |
7729 | f_next_fp = TREE_CHAIN (f_next_o_limit); | |
7730 | f_next_fp_limit = TREE_CHAIN (f_next_fp); | |
7731 | f_next_stack = TREE_CHAIN (f_next_fp_limit); | |
7732 | ||
47a25a46 | 7733 | next_o = build3 (COMPONENT_REF, TREE_TYPE (f_next_o), valist, f_next_o, |
44de5aeb | 7734 | NULL_TREE); |
47a25a46 RG |
7735 | next_o_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_o_limit), |
7736 | valist, f_next_o_limit, NULL_TREE); | |
7737 | next_fp = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp), valist, f_next_fp, | |
7738 | NULL_TREE); | |
7739 | next_fp_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp_limit), | |
7740 | valist, f_next_fp_limit, NULL_TREE); | |
7741 | next_stack = build3 (COMPONENT_REF, TREE_TYPE (f_next_stack), | |
7742 | valist, f_next_stack, NULL_TREE); | |
514066a1 RH |
7743 | |
7744 | /* Call __builtin_saveregs. */ | |
5be014d5 AP |
7745 | u = make_tree (sizetype, expand_builtin_saveregs ()); |
7746 | u = fold_convert (ptr_type_node, u); | |
726a989a | 7747 | t = build2 (MODIFY_EXPR, ptr_type_node, next_fp, u); |
514066a1 RH |
7748 | TREE_SIDE_EFFECTS (t) = 1; |
7749 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7750 | ||
38173d38 | 7751 | nfp = crtl->args.info.arg_count[SH_ARG_FLOAT]; |
514066a1 RH |
7752 | if (nfp < 8) |
7753 | nfp = 8 - nfp; | |
7754 | else | |
7755 | nfp = 0; | |
5be014d5 AP |
7756 | u = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, u, |
7757 | size_int (UNITS_PER_WORD * nfp)); | |
726a989a | 7758 | t = build2 (MODIFY_EXPR, ptr_type_node, next_fp_limit, u); |
514066a1 RH |
7759 | TREE_SIDE_EFFECTS (t) = 1; |
7760 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7761 | ||
726a989a | 7762 | t = build2 (MODIFY_EXPR, ptr_type_node, next_o, u); |
514066a1 RH |
7763 | TREE_SIDE_EFFECTS (t) = 1; |
7764 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7765 | ||
38173d38 | 7766 | nint = crtl->args.info.arg_count[SH_ARG_INT]; |
514066a1 RH |
7767 | if (nint < 4) |
7768 | nint = 4 - nint; | |
7769 | else | |
7770 | nint = 0; | |
5be014d5 AP |
7771 | u = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, u, |
7772 | size_int (UNITS_PER_WORD * nint)); | |
726a989a | 7773 | t = build2 (MODIFY_EXPR, ptr_type_node, next_o_limit, u); |
514066a1 RH |
7774 | TREE_SIDE_EFFECTS (t) = 1; |
7775 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7776 | ||
7777 | u = make_tree (ptr_type_node, nextarg); | |
726a989a | 7778 | t = build2 (MODIFY_EXPR, ptr_type_node, next_stack, u); |
514066a1 RH |
7779 | TREE_SIDE_EFFECTS (t) = 1; |
7780 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7781 | } | |
7782 | ||
917f1b7e | 7783 | /* TYPE is a RECORD_TYPE. If there is only a single nonzero-sized |
21bdb106 JR |
7784 | member, return it. */ |
7785 | static tree | |
7786 | find_sole_member (tree type) | |
7787 | { | |
7788 | tree field, member = NULL_TREE; | |
7789 | ||
7790 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
7791 | { | |
7792 | if (TREE_CODE (field) != FIELD_DECL) | |
7793 | continue; | |
7794 | if (!DECL_SIZE (field)) | |
7795 | return NULL_TREE; | |
7796 | if (integer_zerop (DECL_SIZE (field))) | |
7797 | continue; | |
7798 | if (member) | |
7799 | return NULL_TREE; | |
7800 | member = field; | |
7801 | } | |
7802 | return member; | |
7803 | } | |
514066a1 RH |
7804 | /* Implement `va_arg'. */ |
7805 | ||
5fac2ed2 | 7806 | static tree |
726a989a RB |
7807 | sh_gimplify_va_arg_expr (tree valist, tree type, gimple_seq *pre_p, |
7808 | gimple_seq *post_p ATTRIBUTE_UNUSED) | |
514066a1 RH |
7809 | { |
7810 | HOST_WIDE_INT size, rsize; | |
318881c0 | 7811 | tree tmp, pptr_type_node; |
0ef2c525 | 7812 | tree addr, lab_over = NULL, result = NULL; |
03603192 | 7813 | int pass_by_ref = targetm.calls.must_pass_in_stack (TYPE_MODE (type), type); |
11160b58 | 7814 | tree eff_type; |
5fac2ed2 RH |
7815 | |
7816 | if (pass_by_ref) | |
7817 | type = build_pointer_type (type); | |
514066a1 RH |
7818 | |
7819 | size = int_size_in_bytes (type); | |
7820 | rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD; | |
7821 | pptr_type_node = build_pointer_type (ptr_type_node); | |
7822 | ||
61f71b34 DD |
7823 | if (! TARGET_SH5 && (TARGET_SH2E || TARGET_SH4) |
7824 | && ! (TARGET_HITACHI || sh_cfun_attr_renesas_p ())) | |
514066a1 RH |
7825 | { |
7826 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
7827 | tree next_o, next_o_limit, next_fp, next_fp_limit, next_stack; | |
7828 | int pass_as_float; | |
5fac2ed2 | 7829 | tree lab_false; |
21bdb106 | 7830 | tree member; |
514066a1 RH |
7831 | |
7832 | f_next_o = TYPE_FIELDS (va_list_type_node); | |
7833 | f_next_o_limit = TREE_CHAIN (f_next_o); | |
7834 | f_next_fp = TREE_CHAIN (f_next_o_limit); | |
7835 | f_next_fp_limit = TREE_CHAIN (f_next_fp); | |
7836 | f_next_stack = TREE_CHAIN (f_next_fp_limit); | |
7837 | ||
47a25a46 RG |
7838 | next_o = build3 (COMPONENT_REF, TREE_TYPE (f_next_o), valist, f_next_o, |
7839 | NULL_TREE); | |
7840 | next_o_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_o_limit), | |
7841 | valist, f_next_o_limit, NULL_TREE); | |
7842 | next_fp = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp), | |
7843 | valist, f_next_fp, NULL_TREE); | |
7844 | next_fp_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp_limit), | |
7845 | valist, f_next_fp_limit, NULL_TREE); | |
7846 | next_stack = build3 (COMPONENT_REF, TREE_TYPE (f_next_stack), | |
7847 | valist, f_next_stack, NULL_TREE); | |
514066a1 | 7848 | |
6d1cb95f R |
7849 | /* Structures with a single member with a distinct mode are passed |
7850 | like their member. This is relevant if the latter has a REAL_TYPE | |
7851 | or COMPLEX_TYPE type. */ | |
11160b58 R |
7852 | eff_type = type; |
7853 | while (TREE_CODE (eff_type) == RECORD_TYPE | |
7854 | && (member = find_sole_member (eff_type)) | |
21bdb106 JR |
7855 | && (TREE_CODE (TREE_TYPE (member)) == REAL_TYPE |
7856 | || TREE_CODE (TREE_TYPE (member)) == COMPLEX_TYPE | |
7857 | || TREE_CODE (TREE_TYPE (member)) == RECORD_TYPE)) | |
c26104e4 | 7858 | { |
21bdb106 | 7859 | tree field_type = TREE_TYPE (member); |
c26104e4 | 7860 | |
11160b58 R |
7861 | if (TYPE_MODE (eff_type) == TYPE_MODE (field_type)) |
7862 | eff_type = field_type; | |
c26104e4 | 7863 | else |
8a2209bf | 7864 | { |
11160b58 | 7865 | gcc_assert ((TYPE_ALIGN (eff_type) |
21bdb106 | 7866 | < GET_MODE_ALIGNMENT (TYPE_MODE (field_type))) |
11160b58 | 7867 | || (TYPE_ALIGN (eff_type) |
21bdb106 JR |
7868 | > GET_MODE_BITSIZE (TYPE_MODE (field_type)))); |
7869 | break; | |
8a2209bf | 7870 | } |
c26104e4 | 7871 | } |
5fac2ed2 | 7872 | |
bdf57c40 | 7873 | if (TARGET_SH4 || TARGET_SH2A_DOUBLE) |
514066a1 | 7874 | { |
11160b58 R |
7875 | pass_as_float = ((TREE_CODE (eff_type) == REAL_TYPE && size <= 8) |
7876 | || (TREE_CODE (eff_type) == COMPLEX_TYPE | |
7877 | && TREE_CODE (TREE_TYPE (eff_type)) == REAL_TYPE | |
514066a1 RH |
7878 | && size <= 16)); |
7879 | } | |
7880 | else | |
7881 | { | |
11160b58 | 7882 | pass_as_float = (TREE_CODE (eff_type) == REAL_TYPE && size == 4); |
514066a1 RH |
7883 | } |
7884 | ||
5fac2ed2 | 7885 | addr = create_tmp_var (pptr_type_node, NULL); |
4c4bde29 AH |
7886 | lab_false = create_artificial_label (UNKNOWN_LOCATION); |
7887 | lab_over = create_artificial_label (UNKNOWN_LOCATION); | |
514066a1 | 7888 | |
70f34814 | 7889 | valist = build_simple_mem_ref (addr); |
6d1cb95f | 7890 | |
514066a1 RH |
7891 | if (pass_as_float) |
7892 | { | |
a11c004b R |
7893 | tree next_fp_tmp = create_tmp_var (TREE_TYPE (f_next_fp), NULL); |
7894 | tree cmp; | |
11160b58 | 7895 | bool is_double = size == 8 && TREE_CODE (eff_type) == REAL_TYPE; |
f78ec36a | 7896 | |
8c54989a KK |
7897 | tmp = build1 (ADDR_EXPR, pptr_type_node, unshare_expr (next_fp)); |
7898 | gimplify_assign (unshare_expr (addr), tmp, pre_p); | |
a11c004b | 7899 | |
8c54989a | 7900 | gimplify_assign (unshare_expr (next_fp_tmp), valist, pre_p); |
a11c004b R |
7901 | tmp = next_fp_limit; |
7902 | if (size > 4 && !is_double) | |
8c54989a KK |
7903 | tmp = build2 (POINTER_PLUS_EXPR, TREE_TYPE (tmp), |
7904 | unshare_expr (tmp), size_int (4 - size)); | |
7905 | tmp = build2 (GE_EXPR, boolean_type_node, | |
7906 | unshare_expr (next_fp_tmp), unshare_expr (tmp)); | |
47a25a46 | 7907 | cmp = build3 (COND_EXPR, void_type_node, tmp, |
8c54989a KK |
7908 | build1 (GOTO_EXPR, void_type_node, |
7909 | unshare_expr (lab_false)), NULL_TREE); | |
a11c004b R |
7910 | if (!is_double) |
7911 | gimplify_and_add (cmp, pre_p); | |
514066a1 | 7912 | |
11160b58 R |
7913 | if (TYPE_ALIGN (eff_type) > BITS_PER_WORD |
7914 | || (is_double || size == 16)) | |
514066a1 | 7915 | { |
5be014d5 AP |
7916 | tmp = fold_convert (sizetype, next_fp_tmp); |
7917 | tmp = build2 (BIT_AND_EXPR, sizetype, tmp, | |
7918 | size_int (UNITS_PER_WORD)); | |
7919 | tmp = build2 (POINTER_PLUS_EXPR, ptr_type_node, | |
8c54989a KK |
7920 | unshare_expr (next_fp_tmp), tmp); |
7921 | gimplify_assign (unshare_expr (next_fp_tmp), tmp, pre_p); | |
514066a1 | 7922 | } |
a11c004b R |
7923 | if (is_double) |
7924 | gimplify_and_add (cmp, pre_p); | |
514066a1 | 7925 | |
6d1cb95f | 7926 | #ifdef FUNCTION_ARG_SCmode_WART |
11160b58 R |
7927 | if (TYPE_MODE (eff_type) == SCmode |
7928 | && TARGET_SH4 && TARGET_LITTLE_ENDIAN) | |
6d1cb95f | 7929 | { |
11160b58 | 7930 | tree subtype = TREE_TYPE (eff_type); |
5fac2ed2 | 7931 | tree real, imag; |
6d1cb95f | 7932 | |
a11c004b R |
7933 | imag |
7934 | = std_gimplify_va_arg_expr (next_fp_tmp, subtype, pre_p, NULL); | |
5fac2ed2 | 7935 | imag = get_initialized_tmp_var (imag, pre_p, NULL); |
6d1cb95f | 7936 | |
a11c004b R |
7937 | real |
7938 | = std_gimplify_va_arg_expr (next_fp_tmp, subtype, pre_p, NULL); | |
5fac2ed2 RH |
7939 | real = get_initialized_tmp_var (real, pre_p, NULL); |
7940 | ||
f2a945b1 KK |
7941 | result = build2 (COMPLEX_EXPR, eff_type, real, imag); |
7942 | if (type != eff_type) | |
7943 | result = build1 (VIEW_CONVERT_EXPR, type, result); | |
5fac2ed2 | 7944 | result = get_initialized_tmp_var (result, pre_p, NULL); |
6d1cb95f R |
7945 | } |
7946 | #endif /* FUNCTION_ARG_SCmode_WART */ | |
7947 | ||
8c54989a | 7948 | tmp = build1 (GOTO_EXPR, void_type_node, unshare_expr (lab_over)); |
5fac2ed2 RH |
7949 | gimplify_and_add (tmp, pre_p); |
7950 | ||
8c54989a | 7951 | tmp = build1 (LABEL_EXPR, void_type_node, unshare_expr (lab_false)); |
5fac2ed2 | 7952 | gimplify_and_add (tmp, pre_p); |
514066a1 | 7953 | |
8c54989a KK |
7954 | tmp = build1 (ADDR_EXPR, pptr_type_node, unshare_expr (next_stack)); |
7955 | gimplify_assign (unshare_expr (addr), tmp, pre_p); | |
7956 | gimplify_assign (unshare_expr (next_fp_tmp), | |
7957 | unshare_expr (valist), pre_p); | |
a11c004b | 7958 | |
8c54989a KK |
7959 | gimplify_assign (unshare_expr (valist), |
7960 | unshare_expr (next_fp_tmp), post_p); | |
a11c004b | 7961 | valist = next_fp_tmp; |
514066a1 RH |
7962 | } |
7963 | else | |
7964 | { | |
8c54989a KK |
7965 | tmp = build2 (POINTER_PLUS_EXPR, ptr_type_node, |
7966 | unshare_expr (next_o), size_int (rsize)); | |
7967 | tmp = build2 (GT_EXPR, boolean_type_node, tmp, | |
7968 | unshare_expr (next_o_limit)); | |
47a25a46 | 7969 | tmp = build3 (COND_EXPR, void_type_node, tmp, |
8c54989a KK |
7970 | build1 (GOTO_EXPR, void_type_node, |
7971 | unshare_expr (lab_false)), | |
7972 | NULL_TREE); | |
5fac2ed2 | 7973 | gimplify_and_add (tmp, pre_p); |
514066a1 | 7974 | |
8c54989a KK |
7975 | tmp = build1 (ADDR_EXPR, pptr_type_node, unshare_expr (next_o)); |
7976 | gimplify_assign (unshare_expr (addr), tmp, pre_p); | |
514066a1 | 7977 | |
8c54989a | 7978 | tmp = build1 (GOTO_EXPR, void_type_node, unshare_expr (lab_over)); |
5fac2ed2 RH |
7979 | gimplify_and_add (tmp, pre_p); |
7980 | ||
8c54989a | 7981 | tmp = build1 (LABEL_EXPR, void_type_node, unshare_expr (lab_false)); |
5fac2ed2 | 7982 | gimplify_and_add (tmp, pre_p); |
514066a1 | 7983 | |
bdf57c40 | 7984 | if (size > 4 && ! (TARGET_SH4 || TARGET_SH2A)) |
8c54989a KK |
7985 | gimplify_assign (unshare_expr (next_o), |
7986 | unshare_expr (next_o_limit), pre_p); | |
514066a1 | 7987 | |
8c54989a KK |
7988 | tmp = build1 (ADDR_EXPR, pptr_type_node, unshare_expr (next_stack)); |
7989 | gimplify_assign (unshare_expr (addr), tmp, pre_p); | |
514066a1 RH |
7990 | } |
7991 | ||
5fac2ed2 RH |
7992 | if (!result) |
7993 | { | |
8c54989a | 7994 | tmp = build1 (LABEL_EXPR, void_type_node, unshare_expr (lab_over)); |
5fac2ed2 RH |
7995 | gimplify_and_add (tmp, pre_p); |
7996 | } | |
514066a1 RH |
7997 | } |
7998 | ||
7999 | /* ??? In va-sh.h, there had been code to make values larger than | |
8000 | size 8 indirect. This does not match the FUNCTION_ARG macros. */ | |
8001 | ||
5fac2ed2 | 8002 | tmp = std_gimplify_va_arg_expr (valist, type, pre_p, NULL); |
6d1cb95f R |
8003 | if (result) |
8004 | { | |
726a989a | 8005 | gimplify_assign (result, tmp, pre_p); |
a1516d08 | 8006 | result = build1 (NOP_EXPR, TREE_TYPE (result), result); |
8c54989a | 8007 | tmp = build1 (LABEL_EXPR, void_type_node, unshare_expr (lab_over)); |
5fac2ed2 | 8008 | gimplify_and_add (tmp, pre_p); |
6d1cb95f R |
8009 | } |
8010 | else | |
5fac2ed2 | 8011 | result = tmp; |
6d1cb95f | 8012 | |
5e039e6b | 8013 | if (pass_by_ref) |
d6e9821f | 8014 | result = build_va_arg_indirect_ref (result); |
5fac2ed2 | 8015 | |
5e039e6b | 8016 | return result; |
514066a1 RH |
8017 | } |
8018 | ||
63a9de54 CB |
8019 | /* 64 bit floating points memory transfers are paired single precision loads |
8020 | or store. So DWARF information needs fixing in little endian (unless | |
8021 | PR=SZ=1 in FPSCR). */ | |
8022 | rtx | |
8023 | sh_dwarf_register_span (rtx reg) | |
8024 | { | |
8025 | unsigned regno = REGNO (reg); | |
8026 | ||
8027 | if (WORDS_BIG_ENDIAN || GET_MODE (reg) != DFmode) | |
8028 | return NULL_RTX; | |
8029 | ||
8030 | return | |
8031 | gen_rtx_PARALLEL (VOIDmode, | |
8032 | gen_rtvec (2, | |
8033 | gen_rtx_REG (SFmode, | |
8034 | DBX_REGISTER_NUMBER (regno+1)), | |
8035 | gen_rtx_REG (SFmode, | |
8036 | DBX_REGISTER_NUMBER (regno)))); | |
8037 | } | |
8038 | ||
6e957ac8 | 8039 | static enum machine_mode |
cde0f3fd | 8040 | sh_promote_function_mode (const_tree type, enum machine_mode mode, |
d8158ff1 KK |
8041 | int *punsignedp, const_tree funtype, |
8042 | int for_return ATTRIBUTE_UNUSED) | |
cde0f3fd PB |
8043 | { |
8044 | if (sh_promote_prototypes (funtype)) | |
d055ce45 | 8045 | return promote_mode (type, mode, punsignedp); |
cde0f3fd PB |
8046 | else |
8047 | return mode; | |
8048 | } | |
8049 | ||
d64728cf | 8050 | static bool |
586de218 | 8051 | sh_promote_prototypes (const_tree type) |
61f71b34 DD |
8052 | { |
8053 | if (TARGET_HITACHI) | |
8054 | return 0; | |
8055 | if (! type) | |
8056 | return 1; | |
8057 | return ! sh_attr_renesas_p (type); | |
8058 | } | |
8059 | ||
8cd5a4e0 RH |
8060 | /* Whether an argument must be passed by reference. On SHcompact, we |
8061 | pretend arguments wider than 32-bits that would have been passed in | |
8062 | registers are passed by reference, so that an SHmedia trampoline | |
8063 | loads them into the full 64-bits registers. */ | |
8064 | ||
8065 | static int | |
586de218 KG |
8066 | shcompact_byref (const CUMULATIVE_ARGS *cum, enum machine_mode mode, |
8067 | const_tree type, bool named) | |
8cd5a4e0 RH |
8068 | { |
8069 | unsigned HOST_WIDE_INT size; | |
8070 | ||
8071 | if (type) | |
8072 | size = int_size_in_bytes (type); | |
8073 | else | |
8074 | size = GET_MODE_SIZE (mode); | |
8075 | ||
8076 | if (cum->arg_count[SH_ARG_INT] < NPARM_REGS (SImode) | |
8077 | && (!named | |
8078 | || GET_SH_ARG_CLASS (mode) == SH_ARG_INT | |
8079 | || (GET_SH_ARG_CLASS (mode) == SH_ARG_FLOAT | |
8080 | && cum->arg_count[SH_ARG_FLOAT] >= NPARM_REGS (SFmode))) | |
8081 | && size > 4 | |
8082 | && !SHCOMPACT_FORCE_ON_STACK (mode, type) | |
8083 | && !SH5_WOULD_BE_PARTIAL_NREGS (*cum, mode, type, named)) | |
8084 | return size; | |
8085 | else | |
8086 | return 0; | |
8087 | } | |
8088 | ||
8089 | static bool | |
8090 | sh_pass_by_reference (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
586de218 | 8091 | const_tree type, bool named) |
8cd5a4e0 RH |
8092 | { |
8093 | if (targetm.calls.must_pass_in_stack (mode, type)) | |
8094 | return true; | |
8095 | ||
03603192 JR |
8096 | /* ??? std_gimplify_va_arg_expr passes NULL for cum. That function |
8097 | wants to know about pass-by-reference semantics for incoming | |
8098 | arguments. */ | |
8099 | if (! cum) | |
8100 | return false; | |
8101 | ||
8cd5a4e0 RH |
8102 | if (TARGET_SHCOMPACT) |
8103 | { | |
8104 | cum->byref = shcompact_byref (cum, mode, type, named); | |
8105 | return cum->byref != 0; | |
8106 | } | |
8107 | ||
8108 | return false; | |
8109 | } | |
8110 | ||
6cdd5672 RH |
8111 | static bool |
8112 | sh_callee_copies (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
586de218 | 8113 | const_tree type, bool named ATTRIBUTE_UNUSED) |
6cdd5672 RH |
8114 | { |
8115 | /* ??? How can it possibly be correct to return true only on the | |
8116 | caller side of the equation? Is there someplace else in the | |
8117 | sh backend that's magically producing the copies? */ | |
8118 | return (cum->outgoing | |
8119 | && ((mode == BLKmode ? TYPE_ALIGN (type) : GET_MODE_ALIGNMENT (mode)) | |
8120 | % SH_MIN_ALIGN_FOR_CALLEE_COPY == 0)); | |
8121 | } | |
8122 | ||
78a52f11 RH |
8123 | static int |
8124 | sh_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
8125 | tree type, bool named ATTRIBUTE_UNUSED) | |
8126 | { | |
8127 | int words = 0; | |
8128 | ||
8129 | if (!TARGET_SH5 | |
8130 | && PASS_IN_REG_P (*cum, mode, type) | |
8131 | && !(TARGET_SH4 || TARGET_SH2A_DOUBLE) | |
8132 | && (ROUND_REG (*cum, mode) | |
8133 | + (mode != BLKmode | |
8134 | ? ROUND_ADVANCE (GET_MODE_SIZE (mode)) | |
8135 | : ROUND_ADVANCE (int_size_in_bytes (type))) | |
8136 | > NPARM_REGS (mode))) | |
8137 | words = NPARM_REGS (mode) - ROUND_REG (*cum, mode); | |
8138 | ||
8139 | else if (!TARGET_SHCOMPACT | |
8140 | && SH5_WOULD_BE_PARTIAL_NREGS (*cum, mode, type, named)) | |
8141 | words = NPARM_REGS (SImode) - cum->arg_count[SH_ARG_INT]; | |
8142 | ||
8143 | return words * UNITS_PER_WORD; | |
8144 | } | |
8145 | ||
8146 | ||
61f71b34 DD |
8147 | /* Define where to put the arguments to a function. |
8148 | Value is zero to push the argument on the stack, | |
8149 | or a hard register in which to store the argument. | |
8150 | ||
8151 | MODE is the argument's machine mode. | |
8152 | TYPE is the data type of the argument (as a tree). | |
8153 | This is null for libcalls where that information may | |
8154 | not be available. | |
8155 | CUM is a variable of type CUMULATIVE_ARGS which gives info about | |
8156 | the preceding args and about the function being called. | |
8157 | NAMED is nonzero if this argument is a named parameter | |
8158 | (otherwise it is an extra parameter matching an ellipsis). | |
8159 | ||
8160 | On SH the first args are normally in registers | |
8161 | and the rest are pushed. Any arg that starts within the first | |
8162 | NPARM_REGS words is at least partially passed in a register unless | |
8163 | its data type forbids. */ | |
8164 | ||
8165 | ||
8166 | rtx | |
cf277499 SB |
8167 | sh_function_arg (CUMULATIVE_ARGS *ca, enum machine_mode mode, |
8168 | tree type, int named) | |
61f71b34 DD |
8169 | { |
8170 | if (! TARGET_SH5 && mode == VOIDmode) | |
8171 | return GEN_INT (ca->renesas_abi ? 1 : 0); | |
8172 | ||
8173 | if (! TARGET_SH5 | |
8174 | && PASS_IN_REG_P (*ca, mode, type) | |
8175 | && (named || ! (TARGET_HITACHI || ca->renesas_abi))) | |
8176 | { | |
8177 | int regno; | |
8178 | ||
8179 | if (mode == SCmode && TARGET_SH4 && TARGET_LITTLE_ENDIAN | |
8180 | && (! FUNCTION_ARG_SCmode_WART || (ROUND_REG (*ca, mode) & 1))) | |
8181 | { | |
8182 | rtx r1 = gen_rtx_EXPR_LIST (VOIDmode, | |
8183 | gen_rtx_REG (SFmode, | |
8184 | BASE_ARG_REG (mode) | |
8185 | + (ROUND_REG (*ca, mode) ^ 1)), | |
8186 | const0_rtx); | |
832a3292 KH |
8187 | rtx r2 = gen_rtx_EXPR_LIST (VOIDmode, |
8188 | gen_rtx_REG (SFmode, | |
8189 | BASE_ARG_REG (mode) | |
8190 | + ((ROUND_REG (*ca, mode) + 1) ^ 1)), | |
8191 | GEN_INT (4)); | |
61f71b34 DD |
8192 | return gen_rtx_PARALLEL(SCmode, gen_rtvec(2, r1, r2)); |
8193 | } | |
8194 | ||
8195 | /* If the alignment of a DF value causes an SF register to be | |
8196 | skipped, we will use that skipped register for the next SF | |
8197 | value. */ | |
8198 | if ((TARGET_HITACHI || ca->renesas_abi) | |
8199 | && ca->free_single_fp_reg | |
8200 | && mode == SFmode) | |
8201 | return gen_rtx_REG (mode, ca->free_single_fp_reg); | |
8202 | ||
8203 | regno = (BASE_ARG_REG (mode) + ROUND_REG (*ca, mode)) | |
8204 | ^ (mode == SFmode && TARGET_SH4 | |
8205 | && TARGET_LITTLE_ENDIAN != 0 | |
8206 | && ! TARGET_HITACHI && ! ca->renesas_abi); | |
8207 | return gen_rtx_REG (mode, regno); | |
8208 | ||
8209 | } | |
f676971a | 8210 | |
61f71b34 DD |
8211 | if (TARGET_SH5) |
8212 | { | |
8213 | if (mode == VOIDmode && TARGET_SHCOMPACT) | |
8214 | return GEN_INT (ca->call_cookie); | |
8215 | ||
8216 | /* The following test assumes unnamed arguments are promoted to | |
8217 | DFmode. */ | |
8218 | if (mode == SFmode && ca->free_single_fp_reg) | |
8219 | return SH5_PROTOTYPED_FLOAT_ARG (*ca, mode, ca->free_single_fp_reg); | |
8220 | ||
8221 | if ((GET_SH_ARG_CLASS (mode) == SH_ARG_FLOAT) | |
8222 | && (named || ! ca->prototype_p) | |
8223 | && ca->arg_count[(int) SH_ARG_FLOAT] < NPARM_REGS (SFmode)) | |
8224 | { | |
8225 | if (! ca->prototype_p && TARGET_SHMEDIA) | |
8226 | return SH5_PROTOTYPELESS_FLOAT_ARG (*ca, mode); | |
8227 | ||
8228 | return SH5_PROTOTYPED_FLOAT_ARG (*ca, mode, | |
8229 | FIRST_FP_PARM_REG | |
8230 | + ca->arg_count[(int) SH_ARG_FLOAT]); | |
8231 | } | |
8232 | ||
8233 | if (ca->arg_count[(int) SH_ARG_INT] < NPARM_REGS (SImode) | |
8234 | && (! TARGET_SHCOMPACT | |
8235 | || (! SHCOMPACT_FORCE_ON_STACK (mode, type) | |
8236 | && ! SH5_WOULD_BE_PARTIAL_NREGS (*ca, mode, | |
8237 | type, named)))) | |
8238 | { | |
8239 | return gen_rtx_REG (mode, (FIRST_PARM_REG | |
8240 | + ca->arg_count[(int) SH_ARG_INT])); | |
8241 | } | |
8242 | ||
8243 | return 0; | |
8244 | } | |
8245 | ||
8246 | return 0; | |
8247 | } | |
f676971a | 8248 | |
61f71b34 DD |
8249 | /* Update the data in CUM to advance over an argument |
8250 | of mode MODE and data type TYPE. | |
8251 | (TYPE is null for libcalls where that information may not be | |
8252 | available.) */ | |
8253 | ||
8254 | void | |
cf277499 SB |
8255 | sh_function_arg_advance (CUMULATIVE_ARGS *ca, enum machine_mode mode, |
8256 | tree type, int named) | |
61f71b34 | 8257 | { |
832a3292 KH |
8258 | if (ca->force_mem) |
8259 | ca->force_mem = 0; | |
8260 | else if (TARGET_SH5) | |
8261 | { | |
8262 | tree type2 = (ca->byref && type | |
8263 | ? TREE_TYPE (type) | |
8264 | : type); | |
8265 | enum machine_mode mode2 = (ca->byref && type | |
8266 | ? TYPE_MODE (type2) | |
8267 | : mode); | |
8268 | int dwords = ((ca->byref | |
8269 | ? ca->byref | |
8270 | : mode2 == BLKmode | |
8271 | ? int_size_in_bytes (type2) | |
8272 | : GET_MODE_SIZE (mode2)) + 7) / 8; | |
8273 | int numregs = MIN (dwords, NPARM_REGS (SImode) | |
8274 | - ca->arg_count[(int) SH_ARG_INT]); | |
8275 | ||
8276 | if (numregs) | |
8277 | { | |
8278 | ca->arg_count[(int) SH_ARG_INT] += numregs; | |
8279 | if (TARGET_SHCOMPACT | |
8280 | && SHCOMPACT_FORCE_ON_STACK (mode2, type2)) | |
8281 | { | |
8282 | ca->call_cookie | |
8283 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
8284 | - numregs, 1); | |
8285 | /* N.B. We want this also for outgoing. */ | |
8286 | ca->stack_regs += numregs; | |
8287 | } | |
8288 | else if (ca->byref) | |
8289 | { | |
8290 | if (! ca->outgoing) | |
8291 | ca->stack_regs += numregs; | |
8292 | ca->byref_regs += numregs; | |
8293 | ca->byref = 0; | |
8294 | do | |
8295 | ca->call_cookie | |
8296 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
8297 | - numregs, 2); | |
8298 | while (--numregs); | |
8299 | ca->call_cookie | |
8300 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
8301 | - 1, 1); | |
8302 | } | |
8303 | else if (dwords > numregs) | |
8304 | { | |
8305 | int pushregs = numregs; | |
8306 | ||
8307 | if (TARGET_SHCOMPACT) | |
8308 | ca->stack_regs += numregs; | |
8309 | while (pushregs < NPARM_REGS (SImode) - 1 | |
8310 | && (CALL_COOKIE_INT_REG_GET | |
8311 | (ca->call_cookie, | |
8312 | NPARM_REGS (SImode) - pushregs) | |
8313 | == 1)) | |
8314 | { | |
8315 | ca->call_cookie | |
8316 | &= ~ CALL_COOKIE_INT_REG (NPARM_REGS (SImode) | |
8317 | - pushregs, 1); | |
8318 | pushregs++; | |
8319 | } | |
8320 | if (numregs == NPARM_REGS (SImode)) | |
8321 | ca->call_cookie | |
8322 | |= CALL_COOKIE_INT_REG (0, 1) | |
8323 | | CALL_COOKIE_STACKSEQ (numregs - 1); | |
8324 | else | |
8325 | ca->call_cookie | |
8326 | |= CALL_COOKIE_STACKSEQ (numregs); | |
8327 | } | |
8328 | } | |
8329 | if (GET_SH_ARG_CLASS (mode2) == SH_ARG_FLOAT | |
8330 | && (named || ! ca->prototype_p)) | |
8331 | { | |
8332 | if (mode2 == SFmode && ca->free_single_fp_reg) | |
8333 | ca->free_single_fp_reg = 0; | |
8334 | else if (ca->arg_count[(int) SH_ARG_FLOAT] | |
8335 | < NPARM_REGS (SFmode)) | |
8336 | { | |
8337 | int numfpregs | |
8338 | = MIN ((GET_MODE_SIZE (mode2) + 7) / 8 * 2, | |
8339 | NPARM_REGS (SFmode) | |
8340 | - ca->arg_count[(int) SH_ARG_FLOAT]); | |
8341 | ||
8342 | ca->arg_count[(int) SH_ARG_FLOAT] += numfpregs; | |
8343 | ||
8344 | if (TARGET_SHCOMPACT && ! ca->prototype_p) | |
8345 | { | |
8346 | if (ca->outgoing && numregs > 0) | |
8347 | do | |
8348 | { | |
8349 | ca->call_cookie | |
8350 | |= (CALL_COOKIE_INT_REG | |
8351 | (ca->arg_count[(int) SH_ARG_INT] | |
8352 | - numregs + ((numfpregs - 2) / 2), | |
8353 | 4 + (ca->arg_count[(int) SH_ARG_FLOAT] | |
8354 | - numfpregs) / 2)); | |
8355 | } | |
8356 | while (numfpregs -= 2); | |
8357 | } | |
8358 | else if (mode2 == SFmode && (named) | |
8359 | && (ca->arg_count[(int) SH_ARG_FLOAT] | |
8360 | < NPARM_REGS (SFmode))) | |
8361 | ca->free_single_fp_reg | |
8362 | = FIRST_FP_PARM_REG - numfpregs | |
8363 | + ca->arg_count[(int) SH_ARG_FLOAT] + 1; | |
8364 | } | |
8365 | } | |
8366 | return; | |
8367 | } | |
8368 | ||
8369 | if ((TARGET_HITACHI || ca->renesas_abi) && TARGET_FPU_DOUBLE) | |
8370 | { | |
8371 | /* Note that we've used the skipped register. */ | |
8372 | if (mode == SFmode && ca->free_single_fp_reg) | |
8373 | { | |
8374 | ca->free_single_fp_reg = 0; | |
8375 | return; | |
8376 | } | |
8377 | /* When we have a DF after an SF, there's an SF register that get | |
8378 | skipped in order to align the DF value. We note this skipped | |
8379 | register, because the next SF value will use it, and not the | |
8380 | SF that follows the DF. */ | |
8381 | if (mode == DFmode | |
8382 | && ROUND_REG (*ca, DFmode) != ROUND_REG (*ca, SFmode)) | |
8383 | { | |
8384 | ca->free_single_fp_reg = (ROUND_REG (*ca, SFmode) | |
8385 | + BASE_ARG_REG (mode)); | |
8386 | } | |
8387 | } | |
8388 | ||
157371cf | 8389 | if (! ((TARGET_SH4 || TARGET_SH2A) || ca->renesas_abi) |
832a3292 KH |
8390 | || PASS_IN_REG_P (*ca, mode, type)) |
8391 | (ca->arg_count[(int) GET_SH_ARG_CLASS (mode)] | |
8392 | = (ROUND_REG (*ca, mode) | |
8393 | + (mode == BLKmode | |
8394 | ? ROUND_ADVANCE (int_size_in_bytes (type)) | |
8395 | : ROUND_ADVANCE (GET_MODE_SIZE (mode))))); | |
61f71b34 DD |
8396 | } |
8397 | ||
61f71b34 DD |
8398 | /* The Renesas calling convention doesn't quite fit into this scheme since |
8399 | the address is passed like an invisible argument, but one that is always | |
8400 | passed in memory. */ | |
8401 | static rtx | |
cf277499 | 8402 | sh_struct_value_rtx (tree fndecl, int incoming ATTRIBUTE_UNUSED) |
61f71b34 DD |
8403 | { |
8404 | if (TARGET_HITACHI || sh_attr_renesas_p (fndecl)) | |
8405 | return 0; | |
8406 | return gen_rtx_REG (Pmode, 2); | |
8407 | } | |
8408 | ||
d64728cf AS |
8409 | /* Worker function for TARGET_FUNCTION_VALUE. |
8410 | ||
8411 | For the SH, this is like LIBCALL_VALUE, except that we must change the | |
8412 | mode like PROMOTE_MODE does. | |
8413 | ??? PROMOTE_MODE is ignored for non-scalar types. The set of types | |
8414 | tested here has to be kept in sync with the one in explow.c:promote_mode. | |
8415 | */ | |
8416 | ||
8417 | static rtx | |
8418 | sh_function_value (const_tree valtype, | |
8419 | const_tree fn_decl_or_type, | |
8420 | bool outgoing ATTRIBUTE_UNUSED) | |
8421 | { | |
8422 | if (fn_decl_or_type | |
8423 | && !DECL_P (fn_decl_or_type)) | |
8424 | fn_decl_or_type = NULL; | |
8425 | ||
8426 | return gen_rtx_REG ( | |
8427 | ((GET_MODE_CLASS (TYPE_MODE (valtype)) == MODE_INT | |
8428 | && GET_MODE_SIZE (TYPE_MODE (valtype)) < 4 | |
8429 | && (TREE_CODE (valtype) == INTEGER_TYPE | |
8430 | || TREE_CODE (valtype) == ENUMERAL_TYPE | |
8431 | || TREE_CODE (valtype) == BOOLEAN_TYPE | |
8432 | || TREE_CODE (valtype) == REAL_TYPE | |
8433 | || TREE_CODE (valtype) == OFFSET_TYPE)) | |
8434 | && sh_promote_prototypes (fn_decl_or_type) | |
8435 | ? (TARGET_SHMEDIA64 ? DImode : SImode) : TYPE_MODE (valtype)), | |
8436 | BASE_RETURN_VALUE_REG (TYPE_MODE (valtype))); | |
8437 | } | |
8438 | ||
8439 | /* Worker function for TARGET_LIBCALL_VALUE. */ | |
8440 | ||
8441 | static rtx | |
8442 | sh_libcall_value (enum machine_mode mode, const_rtx fun ATTRIBUTE_UNUSED) | |
8443 | { | |
8444 | return gen_rtx_REG (mode, BASE_RETURN_VALUE_REG (mode)); | |
8445 | } | |
8446 | ||
8447 | /* Worker function for FUNCTION_VALUE_REGNO_P. */ | |
8448 | ||
8449 | bool | |
8450 | sh_function_value_regno_p (const unsigned int regno) | |
8451 | { | |
8452 | return ((regno) == FIRST_RET_REG | |
8453 | || (TARGET_SH2E && (regno) == FIRST_FP_RET_REG) | |
8454 | || (TARGET_SHMEDIA_FPU && (regno) == FIRST_FP_RET_REG)); | |
8455 | } | |
8456 | ||
bd5bd7ac KH |
8457 | /* Worker function for TARGET_RETURN_IN_MEMORY. */ |
8458 | ||
61f71b34 | 8459 | static bool |
586de218 | 8460 | sh_return_in_memory (const_tree type, const_tree fndecl) |
61f71b34 DD |
8461 | { |
8462 | if (TARGET_SH5) | |
8463 | { | |
8464 | if (TYPE_MODE (type) == BLKmode) | |
8465 | return ((unsigned HOST_WIDE_INT) int_size_in_bytes (type)) > 8; | |
8466 | else | |
8467 | return GET_MODE_SIZE (TYPE_MODE (type)) > 8; | |
8468 | } | |
8469 | else | |
8470 | { | |
8471 | return (TYPE_MODE (type) == BLKmode | |
8472 | || ((TARGET_HITACHI || sh_attr_renesas_p (fndecl)) | |
8473 | && TREE_CODE (type) == RECORD_TYPE)); | |
8474 | } | |
8475 | } | |
8476 | ||
8477 | /* We actually emit the code in sh_expand_prologue. We used to use | |
8478 | a static variable to flag that we need to emit this code, but that | |
8479 | doesn't when inlining, when functions are deferred and then emitted | |
8480 | later. Fortunately, we already have two flags that are part of struct | |
8481 | function that tell if a function uses varargs or stdarg. */ | |
8482 | static void | |
49616835 JR |
8483 | sh_setup_incoming_varargs (CUMULATIVE_ARGS *ca, |
8484 | enum machine_mode mode, | |
8485 | tree type, | |
8486 | int *pretend_arg_size, | |
cf277499 | 8487 | int second_time ATTRIBUTE_UNUSED) |
61f71b34 | 8488 | { |
e3b5732b | 8489 | gcc_assert (cfun->stdarg); |
49616835 JR |
8490 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl)) |
8491 | { | |
8492 | int named_parm_regs, anon_parm_regs; | |
8493 | ||
8494 | named_parm_regs = (ROUND_REG (*ca, mode) | |
8495 | + (mode == BLKmode | |
8496 | ? ROUND_ADVANCE (int_size_in_bytes (type)) | |
8497 | : ROUND_ADVANCE (GET_MODE_SIZE (mode)))); | |
8498 | anon_parm_regs = NPARM_REGS (SImode) - named_parm_regs; | |
8499 | if (anon_parm_regs > 0) | |
8500 | *pretend_arg_size = anon_parm_regs * 4; | |
8501 | } | |
61f71b34 DD |
8502 | } |
8503 | ||
8504 | static bool | |
cf277499 | 8505 | sh_strict_argument_naming (CUMULATIVE_ARGS *ca ATTRIBUTE_UNUSED) |
61f71b34 DD |
8506 | { |
8507 | return TARGET_SH5; | |
8508 | } | |
8509 | ||
8510 | static bool | |
cf277499 | 8511 | sh_pretend_outgoing_varargs_named (CUMULATIVE_ARGS *ca) |
61f71b34 DD |
8512 | { |
8513 | return ! (TARGET_HITACHI || ca->renesas_abi) && ! TARGET_SH5; | |
8514 | } | |
8515 | ||
8516 | ||
0d7e008e SC |
8517 | /* Define the offset between two registers, one to be eliminated, and |
8518 | the other its replacement, at the start of a routine. */ | |
8519 | ||
8520 | int | |
cf277499 | 8521 | initial_elimination_offset (int from, int to) |
0d7e008e SC |
8522 | { |
8523 | int regs_saved; | |
fa5322fa | 8524 | int regs_saved_rounding = 0; |
0d7e008e | 8525 | int total_saved_regs_space; |
e5f8d4c0 | 8526 | int total_auto_space; |
1245df60 | 8527 | int save_flags = target_flags; |
fa5322fa | 8528 | int copy_flags; |
490dbac7 | 8529 | HARD_REG_SET live_regs_mask; |
fe3ad572 SC |
8530 | |
8531 | shmedia_space_reserved_for_target_registers = false; | |
490dbac7 | 8532 | regs_saved = calc_live_regs (&live_regs_mask); |
fa5322fa | 8533 | regs_saved += SHMEDIA_REGS_STACK_ADJUST (); |
fe3ad572 SC |
8534 | |
8535 | if (shmedia_reserve_space_for_target_registers_p (regs_saved, &live_regs_mask)) | |
8536 | { | |
8537 | shmedia_space_reserved_for_target_registers = true; | |
8538 | regs_saved += shmedia_target_regs_stack_adjust (&live_regs_mask); | |
8539 | } | |
8540 | ||
fa5322fa AO |
8541 | if (TARGET_SH5 && regs_saved % (STACK_BOUNDARY / BITS_PER_UNIT)) |
8542 | regs_saved_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) | |
8543 | - regs_saved % (STACK_BOUNDARY / BITS_PER_UNIT)); | |
8e87e161 | 8544 | |
fa5322fa AO |
8545 | total_auto_space = rounded_frame_size (regs_saved) - regs_saved_rounding; |
8546 | copy_flags = target_flags; | |
1245df60 | 8547 | target_flags = save_flags; |
1a95a963 | 8548 | |
fa5322fa | 8549 | total_saved_regs_space = regs_saved + regs_saved_rounding; |
b9654711 | 8550 | |
96a2347e | 8551 | if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) |
fa5322fa | 8552 | return total_saved_regs_space + total_auto_space |
38173d38 | 8553 | + crtl->args.info.byref_regs * 8; |
8aa2a305 | 8554 | |
0d7e008e | 8555 | if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
fa5322fa | 8556 | return total_saved_regs_space + total_auto_space |
38173d38 | 8557 | + crtl->args.info.byref_regs * 8; |
8aa2a305 JW |
8558 | |
8559 | /* Initial gap between fp and sp is 0. */ | |
96a2347e | 8560 | if (from == HARD_FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
8aa2a305 JW |
8561 | return 0; |
8562 | ||
96a2347e KK |
8563 | if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
8564 | return rounded_frame_size (0); | |
8565 | ||
8566 | if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
8567 | return rounded_frame_size (0); | |
8568 | ||
f5b9e7c9 | 8569 | gcc_assert (from == RETURN_ADDRESS_POINTER_REGNUM |
96a2347e KK |
8570 | && (to == HARD_FRAME_POINTER_REGNUM |
8571 | || to == STACK_POINTER_REGNUM)); | |
f5b9e7c9 | 8572 | if (TARGET_SH5) |
d042370e | 8573 | { |
f5b9e7c9 NS |
8574 | int n = total_saved_regs_space; |
8575 | int pr_reg = TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG; | |
8576 | save_schedule schedule; | |
8577 | save_entry *entry; | |
8578 | ||
8579 | n += total_auto_space; | |
8580 | ||
8581 | /* If it wasn't saved, there's not much we can do. */ | |
8582 | if (! TEST_HARD_REG_BIT (live_regs_mask, pr_reg)) | |
8583 | return n; | |
8584 | ||
8585 | target_flags = copy_flags; | |
8586 | ||
8587 | sh5_schedule_saves (&live_regs_mask, &schedule, n); | |
8588 | for (entry = &schedule.entries[1]; entry->mode != VOIDmode; entry++) | |
8589 | if (entry->reg == pr_reg) | |
8590 | { | |
8591 | target_flags = save_flags; | |
8592 | return entry->offset; | |
8593 | } | |
8594 | gcc_unreachable (); | |
d042370e | 8595 | } |
f5b9e7c9 NS |
8596 | else |
8597 | return total_auto_space; | |
0d7e008e | 8598 | } |
3217af3e KK |
8599 | |
8600 | /* Parse the -mfixed-range= option string. */ | |
8601 | void | |
8602 | sh_fix_range (const char *const_str) | |
8603 | { | |
8604 | int i, first, last; | |
8605 | char *str, *dash, *comma; | |
8606 | ||
8607 | /* str must be of the form REG1'-'REG2{,REG1'-'REG} where REG1 and | |
8608 | REG2 are either register names or register numbers. The effect | |
8609 | of this option is to mark the registers in the range from REG1 to | |
8610 | REG2 as ``fixed'' so they won't be used by the compiler. */ | |
8611 | ||
8612 | i = strlen (const_str); | |
8613 | str = (char *) alloca (i + 1); | |
8614 | memcpy (str, const_str, i + 1); | |
8615 | ||
8616 | while (1) | |
8617 | { | |
8618 | dash = strchr (str, '-'); | |
8619 | if (!dash) | |
8620 | { | |
8621 | warning (0, "value of -mfixed-range must have form REG1-REG2"); | |
8622 | return; | |
8623 | } | |
8624 | *dash = '\0'; | |
8625 | comma = strchr (dash + 1, ','); | |
8626 | if (comma) | |
8627 | *comma = '\0'; | |
8628 | ||
8629 | first = decode_reg_name (str); | |
8630 | if (first < 0) | |
8631 | { | |
8632 | warning (0, "unknown register name: %s", str); | |
8633 | return; | |
8634 | } | |
8635 | ||
8636 | last = decode_reg_name (dash + 1); | |
8637 | if (last < 0) | |
8638 | { | |
8639 | warning (0, "unknown register name: %s", dash + 1); | |
8640 | return; | |
8641 | } | |
8642 | ||
8643 | *dash = '-'; | |
8644 | ||
8645 | if (first > last) | |
8646 | { | |
8647 | warning (0, "%s-%s is an empty range", str, dash + 1); | |
8648 | return; | |
8649 | } | |
8650 | ||
8651 | for (i = first; i <= last; ++i) | |
8652 | fixed_regs[i] = call_used_regs[i] = 1; | |
8653 | ||
8654 | if (!comma) | |
8655 | break; | |
8656 | ||
8657 | *comma = ','; | |
8658 | str = comma + 1; | |
8659 | } | |
8660 | } | |
8aa2a305 | 8661 | \f |
a6ab9fc0 | 8662 | /* Insert any deferred function attributes from earlier pragmas. */ |
12a68f1f | 8663 | static void |
cf277499 | 8664 | sh_insert_attributes (tree node, tree *attributes) |
157131d7 | 8665 | { |
a6ab9fc0 R |
8666 | tree attrs; |
8667 | ||
8668 | if (TREE_CODE (node) != FUNCTION_DECL) | |
157131d7 SC |
8669 | return; |
8670 | ||
8671 | /* We are only interested in fields. */ | |
6615c446 | 8672 | if (!DECL_P (node)) |
157131d7 SC |
8673 | return; |
8674 | ||
a6ab9fc0 R |
8675 | /* Append the attributes to the deferred attributes. */ |
8676 | *sh_deferred_function_attributes_tail = *attributes; | |
8677 | attrs = sh_deferred_function_attributes; | |
8678 | if (!attrs) | |
8679 | return; | |
8680 | ||
8681 | /* Some attributes imply or require the interrupt attribute. */ | |
8682 | if (!lookup_attribute ("interrupt_handler", attrs) | |
8683 | && !lookup_attribute ("interrupt_handler", DECL_ATTRIBUTES (node))) | |
8684 | { | |
8685 | /* If we have a trapa_handler, but no interrupt_handler attribute, | |
8686 | insert an interrupt_handler attribute. */ | |
8687 | if (lookup_attribute ("trapa_handler", attrs) != NULL_TREE) | |
8688 | /* We can't use sh_pr_interrupt here because that's not in the | |
8689 | java frontend. */ | |
8690 | attrs | |
8691 | = tree_cons (get_identifier("interrupt_handler"), NULL_TREE, attrs); | |
561642fa AP |
8692 | /* However, for sp_switch, trap_exit, nosave_low_regs and resbank, |
8693 | if the interrupt attribute is missing, we ignore the attribute | |
8694 | and warn. */ | |
a6ab9fc0 R |
8695 | else if (lookup_attribute ("sp_switch", attrs) |
8696 | || lookup_attribute ("trap_exit", attrs) | |
561642fa AP |
8697 | || lookup_attribute ("nosave_low_regs", attrs) |
8698 | || lookup_attribute ("resbank", attrs)) | |
a6ab9fc0 R |
8699 | { |
8700 | tree *tail; | |
8701 | ||
8702 | for (tail = attributes; attrs; attrs = TREE_CHAIN (attrs)) | |
8703 | { | |
8704 | if (is_attribute_p ("sp_switch", TREE_PURPOSE (attrs)) | |
8705 | || is_attribute_p ("trap_exit", TREE_PURPOSE (attrs)) | |
561642fa AP |
8706 | || is_attribute_p ("nosave_low_regs", TREE_PURPOSE (attrs)) |
8707 | || is_attribute_p ("resbank", TREE_PURPOSE (attrs))) | |
a6ab9fc0 | 8708 | warning (OPT_Wattributes, |
29d08eba JM |
8709 | "%qE attribute only applies to interrupt functions", |
8710 | TREE_PURPOSE (attrs)); | |
a6ab9fc0 R |
8711 | else |
8712 | { | |
8713 | *tail = tree_cons (TREE_PURPOSE (attrs), NULL_TREE, | |
8714 | NULL_TREE); | |
8715 | tail = &TREE_CHAIN (*tail); | |
8716 | } | |
8717 | } | |
8718 | attrs = *attributes; | |
8719 | } | |
8720 | } | |
8721 | ||
8722 | /* Install the processed list. */ | |
8723 | *attributes = attrs; | |
8724 | ||
8725 | /* Clear deferred attributes. */ | |
8726 | sh_deferred_function_attributes = NULL_TREE; | |
8727 | sh_deferred_function_attributes_tail = &sh_deferred_function_attributes; | |
157131d7 SC |
8728 | |
8729 | return; | |
8730 | } | |
8731 | ||
91d231cb | 8732 | /* Supported attributes: |
4408efce JL |
8733 | |
8734 | interrupt_handler -- specifies this function is an interrupt handler. | |
8735 | ||
a6ab9fc0 R |
8736 | trapa_handler - like above, but don't save all registers. |
8737 | ||
4408efce JL |
8738 | sp_switch -- specifies an alternate stack for an interrupt handler |
8739 | to run on. | |
8740 | ||
956d6950 | 8741 | trap_exit -- use a trapa to exit an interrupt function instead of |
61f71b34 DD |
8742 | an rte instruction. |
8743 | ||
a6ab9fc0 R |
8744 | nosave_low_regs - don't save r0..r7 in an interrupt handler. |
8745 | This is useful on the SH3 and upwards, | |
8746 | which has a separate set of low regs for User and Supervisor modes. | |
8747 | This should only be used for the lowest level of interrupts. Higher levels | |
8748 | of interrupts must save the registers in case they themselves are | |
8749 | interrupted. | |
8750 | ||
61f71b34 DD |
8751 | renesas -- use Renesas calling/layout conventions (functions and |
8752 | structures). | |
8753 | ||
561642fa AP |
8754 | resbank -- In case of an ISR, use a register bank to save registers |
8755 | R0-R14, MACH, MACL, GBR and PR. This is useful only on SH2A targets. | |
61f71b34 | 8756 | */ |
4408efce | 8757 | |
561642fa AP |
8758 | /* Handle a 'resbank' attribute. */ |
8759 | static tree | |
8760 | sh_handle_resbank_handler_attribute (tree * node, tree name, | |
8761 | tree args ATTRIBUTE_UNUSED, | |
8762 | int flags ATTRIBUTE_UNUSED, | |
8763 | bool * no_add_attrs) | |
8764 | { | |
8765 | if (!TARGET_SH2A) | |
8766 | { | |
29d08eba JM |
8767 | warning (OPT_Wattributes, "%qE attribute is supported only for SH2A", |
8768 | name); | |
561642fa AP |
8769 | *no_add_attrs = true; |
8770 | } | |
8771 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
8772 | { | |
29d08eba JM |
8773 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
8774 | name); | |
561642fa AP |
8775 | *no_add_attrs = true; |
8776 | } | |
8777 | ||
8778 | return NULL_TREE; | |
8779 | } | |
8780 | ||
91d231cb JM |
8781 | /* Handle an "interrupt_handler" attribute; arguments as in |
8782 | struct attribute_spec.handler. */ | |
8783 | static tree | |
cf277499 | 8784 | sh_handle_interrupt_handler_attribute (tree *node, tree name, |
561642fa AP |
8785 | tree args ATTRIBUTE_UNUSED, |
8786 | int flags ATTRIBUTE_UNUSED, | |
8787 | bool *no_add_attrs) | |
91d231cb JM |
8788 | { |
8789 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
4408efce | 8790 | { |
29d08eba JM |
8791 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
8792 | name); | |
91d231cb | 8793 | *no_add_attrs = true; |
4408efce | 8794 | } |
58ab7171 R |
8795 | else if (TARGET_SHCOMPACT) |
8796 | { | |
8797 | error ("attribute interrupt_handler is not compatible with -m5-compact"); | |
8798 | *no_add_attrs = true; | |
8799 | } | |
4408efce | 8800 | |
91d231cb JM |
8801 | return NULL_TREE; |
8802 | } | |
8803 | ||
561642fa AP |
8804 | /* Handle an 'function_vector' attribute; arguments as in |
8805 | struct attribute_spec.handler. */ | |
8806 | static tree | |
8807 | sh2a_handle_function_vector_handler_attribute (tree * node, tree name, | |
8808 | tree args ATTRIBUTE_UNUSED, | |
8809 | int flags ATTRIBUTE_UNUSED, | |
8810 | bool * no_add_attrs) | |
8811 | { | |
8812 | if (!TARGET_SH2A) | |
8813 | { | |
29d08eba JM |
8814 | warning (OPT_Wattributes, "%qE attribute only applies to SH2A", |
8815 | name); | |
561642fa AP |
8816 | *no_add_attrs = true; |
8817 | } | |
8818 | else if (TREE_CODE (*node) != FUNCTION_DECL) | |
8819 | { | |
29d08eba JM |
8820 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
8821 | name); | |
561642fa AP |
8822 | *no_add_attrs = true; |
8823 | } | |
8824 | else if (TREE_CODE (TREE_VALUE (args)) != INTEGER_CST) | |
8825 | { | |
8826 | /* The argument must be a constant integer. */ | |
8827 | warning (OPT_Wattributes, | |
29d08eba JM |
8828 | "%qE attribute argument not an integer constant", |
8829 | name); | |
561642fa AP |
8830 | *no_add_attrs = true; |
8831 | } | |
8832 | else if (TREE_INT_CST_LOW (TREE_VALUE (args)) > 255) | |
8833 | { | |
8834 | /* The argument value must be between 0 to 255. */ | |
8835 | warning (OPT_Wattributes, | |
29d08eba JM |
8836 | "%qE attribute argument should be between 0 to 255", |
8837 | name); | |
561642fa AP |
8838 | *no_add_attrs = true; |
8839 | } | |
8840 | return NULL_TREE; | |
8841 | } | |
8842 | ||
8843 | /* Returns 1 if current function has been assigned the attribute | |
8844 | 'function_vector'. */ | |
8845 | int | |
8846 | sh2a_is_function_vector_call (rtx x) | |
8847 | { | |
8848 | if (GET_CODE (x) == SYMBOL_REF | |
8849 | && (SYMBOL_REF_FLAGS (x) & SYMBOL_FLAG_FUNCVEC_FUNCTION)) | |
8850 | { | |
8851 | tree tr = SYMBOL_REF_DECL (x); | |
8852 | ||
8853 | if (sh2a_function_vector_p (tr)) | |
8854 | return 1; | |
8855 | } | |
8856 | ||
8857 | return 0; | |
8858 | } | |
8859 | ||
8860 | /* Returns the function vector number, if the the attribute | |
8861 | 'function_vector' is assigned, otherwise returns zero. */ | |
8862 | int | |
8863 | sh2a_get_function_vector_number (rtx x) | |
8864 | { | |
8865 | int num; | |
8866 | tree list, t; | |
8867 | ||
8868 | if ((GET_CODE (x) == SYMBOL_REF) | |
8869 | && (SYMBOL_REF_FLAGS (x) & SYMBOL_FLAG_FUNCVEC_FUNCTION)) | |
8870 | { | |
8871 | t = SYMBOL_REF_DECL (x); | |
8872 | ||
8873 | if (TREE_CODE (t) != FUNCTION_DECL) | |
8874 | return 0; | |
8875 | ||
8876 | list = SH_ATTRIBUTES (t); | |
8877 | while (list) | |
8878 | { | |
8879 | if (is_attribute_p ("function_vector", TREE_PURPOSE (list))) | |
8880 | { | |
8881 | num = TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (list))); | |
8882 | return num; | |
8883 | } | |
8884 | ||
8885 | list = TREE_CHAIN (list); | |
8886 | } | |
8887 | ||
8888 | return 0; | |
8889 | } | |
8890 | else | |
8891 | return 0; | |
8892 | } | |
8893 | ||
91d231cb JM |
8894 | /* Handle an "sp_switch" attribute; arguments as in |
8895 | struct attribute_spec.handler. */ | |
8896 | static tree | |
cf277499 SB |
8897 | sh_handle_sp_switch_attribute (tree *node, tree name, tree args, |
8898 | int flags ATTRIBUTE_UNUSED, bool *no_add_attrs) | |
91d231cb JM |
8899 | { |
8900 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
8901 | { | |
29d08eba JM |
8902 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
8903 | name); | |
91d231cb JM |
8904 | *no_add_attrs = true; |
8905 | } | |
91d231cb JM |
8906 | else if (TREE_CODE (TREE_VALUE (args)) != STRING_CST) |
8907 | { | |
4408efce | 8908 | /* The argument must be a constant string. */ |
29d08eba JM |
8909 | warning (OPT_Wattributes, "%qE attribute argument not a string constant", |
8910 | name); | |
91d231cb JM |
8911 | *no_add_attrs = true; |
8912 | } | |
4408efce | 8913 | |
91d231cb JM |
8914 | return NULL_TREE; |
8915 | } | |
8916 | ||
8917 | /* Handle an "trap_exit" attribute; arguments as in | |
8918 | struct attribute_spec.handler. */ | |
8919 | static tree | |
cf277499 SB |
8920 | sh_handle_trap_exit_attribute (tree *node, tree name, tree args, |
8921 | int flags ATTRIBUTE_UNUSED, bool *no_add_attrs) | |
91d231cb JM |
8922 | { |
8923 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
8924 | { | |
29d08eba JM |
8925 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
8926 | name); | |
91d231cb JM |
8927 | *no_add_attrs = true; |
8928 | } | |
a6ab9fc0 R |
8929 | /* The argument specifies a trap number to be used in a trapa instruction |
8930 | at function exit (instead of an rte instruction). */ | |
91d231cb JM |
8931 | else if (TREE_CODE (TREE_VALUE (args)) != INTEGER_CST) |
8932 | { | |
4408efce | 8933 | /* The argument must be a constant integer. */ |
29d08eba JM |
8934 | warning (OPT_Wattributes, "%qE attribute argument not an " |
8935 | "integer constant", name); | |
91d231cb JM |
8936 | *no_add_attrs = true; |
8937 | } | |
318881c0 | 8938 | |
91d231cb | 8939 | return NULL_TREE; |
4408efce JL |
8940 | } |
8941 | ||
61f71b34 | 8942 | static tree |
cf277499 SB |
8943 | sh_handle_renesas_attribute (tree *node ATTRIBUTE_UNUSED, |
8944 | tree name ATTRIBUTE_UNUSED, | |
8945 | tree args ATTRIBUTE_UNUSED, | |
8946 | int flags ATTRIBUTE_UNUSED, | |
8947 | bool *no_add_attrs ATTRIBUTE_UNUSED) | |
61f71b34 DD |
8948 | { |
8949 | return NULL_TREE; | |
8950 | } | |
8951 | ||
8952 | /* True if __attribute__((renesas)) or -mrenesas. */ | |
8953 | int | |
586de218 | 8954 | sh_attr_renesas_p (const_tree td) |
61f71b34 DD |
8955 | { |
8956 | if (TARGET_HITACHI) | |
8957 | return 1; | |
8958 | if (td == 0) | |
8959 | return 0; | |
8960 | if (DECL_P (td)) | |
8961 | td = TREE_TYPE (td); | |
a11c004b R |
8962 | if (td == error_mark_node) |
8963 | return 0; | |
61f71b34 DD |
8964 | return (lookup_attribute ("renesas", TYPE_ATTRIBUTES (td)) |
8965 | != NULL_TREE); | |
8966 | } | |
8967 | ||
8968 | /* True if __attribute__((renesas)) or -mrenesas, for the current | |
8969 | function. */ | |
8970 | int | |
cf277499 | 8971 | sh_cfun_attr_renesas_p (void) |
61f71b34 DD |
8972 | { |
8973 | return sh_attr_renesas_p (current_function_decl); | |
8974 | } | |
8975 | ||
e3ba8d11 | 8976 | int |
cf277499 | 8977 | sh_cfun_interrupt_handler_p (void) |
e3ba8d11 R |
8978 | { |
8979 | return (lookup_attribute ("interrupt_handler", | |
8980 | DECL_ATTRIBUTES (current_function_decl)) | |
8981 | != NULL_TREE); | |
8982 | } | |
bcc8cc82 | 8983 | |
561642fa AP |
8984 | /* Returns 1 if FUNC has been assigned the attribute |
8985 | "function_vector". */ | |
8986 | int | |
8987 | sh2a_function_vector_p (tree func) | |
8988 | { | |
8989 | tree list; | |
8990 | if (TREE_CODE (func) != FUNCTION_DECL) | |
8991 | return 0; | |
8992 | ||
8993 | list = SH_ATTRIBUTES (func); | |
8994 | while (list) | |
8995 | { | |
8996 | if (is_attribute_p ("function_vector", TREE_PURPOSE (list))) | |
8997 | return 1; | |
8998 | ||
8999 | list = TREE_CHAIN (list); | |
9000 | } | |
9001 | return 0; | |
9002 | } | |
9003 | ||
9004 | /* Returns TRUE if given tree has the "resbank" attribute. */ | |
9005 | ||
9006 | int | |
9007 | sh_cfun_resbank_handler_p (void) | |
9008 | { | |
9009 | return ((lookup_attribute ("resbank", | |
9010 | DECL_ATTRIBUTES (current_function_decl)) | |
9011 | != NULL_TREE) | |
9012 | && (lookup_attribute ("interrupt_handler", | |
9013 | DECL_ATTRIBUTES (current_function_decl)) | |
9014 | != NULL_TREE) && TARGET_SH2A); | |
9015 | } | |
9016 | ||
8d932be3 | 9017 | /* Implement TARGET_CHECK_PCH_TARGET_FLAGS. */ |
73a4d10b | 9018 | |
8d932be3 RS |
9019 | static const char * |
9020 | sh_check_pch_target_flags (int old_flags) | |
9021 | { | |
c0fb94d7 RS |
9022 | if ((old_flags ^ target_flags) & (MASK_SH1 | MASK_SH2 | MASK_SH3 |
9023 | | MASK_SH_E | MASK_HARD_SH4 | |
9024 | | MASK_FPU_SINGLE | MASK_SH4)) | |
9025 | return _("created and used with different architectures / ABIs"); | |
9026 | if ((old_flags ^ target_flags) & MASK_HITACHI) | |
9027 | return _("created and used with different ABIs"); | |
9028 | if ((old_flags ^ target_flags) & MASK_LITTLE_ENDIAN) | |
9029 | return _("created and used with different endianness"); | |
bcc8cc82 | 9030 | return NULL; |
bcc8cc82 | 9031 | } |
0d7e008e | 9032 | \f |
8aa2a305 | 9033 | /* Predicates used by the templates. */ |
0d7e008e | 9034 | |
8aa2a305 JW |
9035 | /* Returns 1 if OP is MACL, MACH or PR. The input must be a REG rtx. |
9036 | Used only in general_movsrc_operand. */ | |
0d7e008e | 9037 | |
8aa2a305 | 9038 | int |
cf277499 | 9039 | system_reg_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
0d7e008e | 9040 | { |
8aa2a305 | 9041 | switch (REGNO (op)) |
0d7e008e | 9042 | { |
8aa2a305 JW |
9043 | case PR_REG: |
9044 | case MACL_REG: | |
9045 | case MACH_REG: | |
9046 | return 1; | |
0d7e008e | 9047 | } |
8aa2a305 | 9048 | return 0; |
0d7e008e | 9049 | } |
0d7e008e | 9050 | |
1a95a963 JW |
9051 | /* Nonzero if OP is a floating point value with value 0.0. */ |
9052 | ||
9053 | int | |
cf277499 | 9054 | fp_zero_operand (rtx op) |
d3ae8277 | 9055 | { |
1a95a963 | 9056 | REAL_VALUE_TYPE r; |
d3ae8277 | 9057 | |
1a95a963 JW |
9058 | if (GET_MODE (op) != SFmode) |
9059 | return 0; | |
9060 | ||
9061 | REAL_VALUE_FROM_CONST_DOUBLE (r, op); | |
e4fa6b06 | 9062 | return REAL_VALUES_EQUAL (r, dconst0) && ! REAL_VALUE_MINUS_ZERO (r); |
d3ae8277 SC |
9063 | } |
9064 | ||
1a95a963 | 9065 | /* Nonzero if OP is a floating point value with value 1.0. */ |
d3ae8277 SC |
9066 | |
9067 | int | |
cf277499 | 9068 | fp_one_operand (rtx op) |
d3ae8277 | 9069 | { |
1a95a963 JW |
9070 | REAL_VALUE_TYPE r; |
9071 | ||
9072 | if (GET_MODE (op) != SFmode) | |
9073 | return 0; | |
9074 | ||
9075 | REAL_VALUE_FROM_CONST_DOUBLE (r, op); | |
9076 | return REAL_VALUES_EQUAL (r, dconst1); | |
d3ae8277 | 9077 | } |
1245df60 | 9078 | |
7a61cf6f | 9079 | /* In general mode switching is used. If we are |
9f09b1f2 R |
9080 | compiling without -mfmovd, movsf_ie isn't taken into account for |
9081 | mode switching. We could check in machine_dependent_reorg for | |
9082 | cases where we know we are in single precision mode, but there is | |
9083 | interface to find that out during reload, so we must avoid | |
9084 | choosing an fldi alternative during reload and thus failing to | |
9085 | allocate a scratch register for the constant loading. */ | |
9086 | int | |
cf277499 | 9087 | fldi_ok (void) |
9f09b1f2 | 9088 | { |
7a61cf6f | 9089 | return 1; |
9f09b1f2 R |
9090 | } |
9091 | ||
225e4f43 | 9092 | int |
cf277499 | 9093 | tertiary_reload_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
225e4f43 R |
9094 | { |
9095 | enum rtx_code code = GET_CODE (op); | |
9096 | return code == MEM || (TARGET_SH4 && code == CONST_DOUBLE); | |
9097 | } | |
9098 | ||
463f02cd | 9099 | /* Return the TLS type for TLS symbols, 0 for otherwise. */ |
81f40b79 | 9100 | enum tls_model |
cf277499 | 9101 | tls_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
463f02cd | 9102 | { |
463f02cd | 9103 | if (GET_CODE (op) != SYMBOL_REF) |
81f40b79 | 9104 | return TLS_MODEL_NONE; |
675ff4c7 | 9105 | return SYMBOL_REF_TLS_MODEL (op); |
463f02cd | 9106 | } |
1245df60 | 9107 | \f |
33f7f353 | 9108 | /* Return the destination address of a branch. */ |
f676971a | 9109 | |
318881c0 | 9110 | static int |
cf277499 | 9111 | branch_dest (rtx branch) |
1245df60 | 9112 | { |
33f7f353 JR |
9113 | rtx dest = SET_SRC (PATTERN (branch)); |
9114 | int dest_uid; | |
1245df60 R |
9115 | |
9116 | if (GET_CODE (dest) == IF_THEN_ELSE) | |
9117 | dest = XEXP (dest, 1); | |
9118 | dest = XEXP (dest, 0); | |
9119 | dest_uid = INSN_UID (dest); | |
9d98a694 | 9120 | return INSN_ADDRESSES (dest_uid); |
1245df60 | 9121 | } |
a55e9d2b | 9122 | \f |
5e7a8ee0 | 9123 | /* Return nonzero if REG is not used after INSN. |
a55e9d2b | 9124 | We assume REG is a reload reg, and therefore does |
8b760293 | 9125 | not live past labels. It may live past calls or jumps though. */ |
a55e9d2b | 9126 | int |
cf277499 | 9127 | reg_unused_after (rtx reg, rtx insn) |
a55e9d2b | 9128 | { |
8783b15e | 9129 | enum rtx_code code; |
a55e9d2b RK |
9130 | rtx set; |
9131 | ||
9132 | /* If the reg is set by this instruction, then it is safe for our | |
9133 | case. Disregard the case where this is a store to memory, since | |
9134 | we are checking a register used in the store address. */ | |
9135 | set = single_set (insn); | |
f3536097 | 9136 | if (set && !MEM_P (SET_DEST (set)) |
a55e9d2b RK |
9137 | && reg_overlap_mentioned_p (reg, SET_DEST (set))) |
9138 | return 1; | |
9139 | ||
318881c0 | 9140 | while ((insn = NEXT_INSN (insn))) |
a55e9d2b | 9141 | { |
ec8e098d PB |
9142 | rtx set; |
9143 | if (!INSN_P (insn)) | |
9144 | continue; | |
9145 | ||
a55e9d2b | 9146 | code = GET_CODE (insn); |
8783b15e | 9147 | |
c8f6f18d RK |
9148 | #if 0 |
9149 | /* If this is a label that existed before reload, then the register | |
9150 | if dead here. However, if this is a label added by reorg, then | |
9151 | the register may still be live here. We can't tell the difference, | |
9152 | so we just ignore labels completely. */ | |
8783b15e | 9153 | if (code == CODE_LABEL) |
a55e9d2b | 9154 | return 1; |
c8f6f18d RK |
9155 | /* else */ |
9156 | #endif | |
a55e9d2b | 9157 | |
8b760293 JW |
9158 | if (code == JUMP_INSN) |
9159 | return 0; | |
9160 | ||
8783b15e RK |
9161 | /* If this is a sequence, we must handle them all at once. |
9162 | We could have for instance a call that sets the target register, | |
e03f5d43 | 9163 | and an insn in a delay slot that uses the register. In this case, |
8783b15e | 9164 | we must return 0. */ |
8b760293 | 9165 | else if (code == INSN && GET_CODE (PATTERN (insn)) == SEQUENCE) |
a55e9d2b | 9166 | { |
8783b15e RK |
9167 | int i; |
9168 | int retval = 0; | |
9169 | ||
9170 | for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++) | |
9171 | { | |
9172 | rtx this_insn = XVECEXP (PATTERN (insn), 0, i); | |
9173 | rtx set = single_set (this_insn); | |
9174 | ||
f3536097 | 9175 | if (CALL_P (this_insn)) |
8783b15e | 9176 | code = CALL_INSN; |
f3536097 | 9177 | else if (JUMP_P (this_insn)) |
38f35781 JW |
9178 | { |
9179 | if (INSN_ANNULLED_BRANCH_P (this_insn)) | |
9180 | return 0; | |
9181 | code = JUMP_INSN; | |
9182 | } | |
a55e9d2b | 9183 | |
8783b15e RK |
9184 | if (set && reg_overlap_mentioned_p (reg, SET_SRC (set))) |
9185 | return 0; | |
9186 | if (set && reg_overlap_mentioned_p (reg, SET_DEST (set))) | |
9187 | { | |
f3536097 | 9188 | if (!MEM_P (SET_DEST (set))) |
8783b15e RK |
9189 | retval = 1; |
9190 | else | |
9191 | return 0; | |
9192 | } | |
9193 | if (set == 0 | |
9194 | && reg_overlap_mentioned_p (reg, PATTERN (this_insn))) | |
9195 | return 0; | |
9196 | } | |
9197 | if (retval == 1) | |
9198 | return 1; | |
38f35781 JW |
9199 | else if (code == JUMP_INSN) |
9200 | return 0; | |
8783b15e | 9201 | } |
a55e9d2b | 9202 | |
ec8e098d PB |
9203 | set = single_set (insn); |
9204 | if (set && reg_overlap_mentioned_p (reg, SET_SRC (set))) | |
9205 | return 0; | |
9206 | if (set && reg_overlap_mentioned_p (reg, SET_DEST (set))) | |
f3536097 | 9207 | return !MEM_P (SET_DEST (set)); |
ec8e098d PB |
9208 | if (set == 0 && reg_overlap_mentioned_p (reg, PATTERN (insn))) |
9209 | return 0; | |
8783b15e | 9210 | |
62164eb4 | 9211 | if (code == CALL_INSN && call_really_used_regs[REGNO (reg)]) |
8783b15e | 9212 | return 1; |
a55e9d2b RK |
9213 | } |
9214 | return 1; | |
9215 | } | |
225e4f43 | 9216 | \f |
c5b9ef02 | 9217 | #include "ggc.h" |
225e4f43 | 9218 | |
e2500fed | 9219 | static GTY(()) rtx fpscr_rtx; |
225e4f43 | 9220 | rtx |
cf277499 | 9221 | get_fpscr_rtx (void) |
225e4f43 | 9222 | { |
225e4f43 R |
9223 | if (! fpscr_rtx) |
9224 | { | |
c0d4e710 | 9225 | fpscr_rtx = gen_rtx_REG (PSImode, FPSCR_REG); |
225e4f43 | 9226 | REG_USERVAR_P (fpscr_rtx) = 1; |
225e4f43 R |
9227 | mark_user_reg (fpscr_rtx); |
9228 | } | |
9229 | if (! reload_completed || mdep_reorg_phase != SH_AFTER_MDEP_REORG) | |
9230 | mark_user_reg (fpscr_rtx); | |
9231 | return fpscr_rtx; | |
9232 | } | |
9233 | ||
24c2fde2 RH |
9234 | static GTY(()) tree fpscr_values; |
9235 | ||
9236 | static void | |
9237 | emit_fpu_switch (rtx scratch, int index) | |
9238 | { | |
9239 | rtx dst, src; | |
9240 | ||
9241 | if (fpscr_values == NULL) | |
9242 | { | |
9243 | tree t; | |
9244 | ||
9245 | t = build_index_type (integer_one_node); | |
9246 | t = build_array_type (integer_type_node, t); | |
4c4bde29 AH |
9247 | t = build_decl (BUILTINS_LOCATION, |
9248 | VAR_DECL, get_identifier ("__fpscr_values"), t); | |
24c2fde2 RH |
9249 | DECL_ARTIFICIAL (t) = 1; |
9250 | DECL_IGNORED_P (t) = 1; | |
9251 | DECL_EXTERNAL (t) = 1; | |
9252 | TREE_STATIC (t) = 1; | |
a5f5c6be | 9253 | TREE_PUBLIC (t) = 1; |
24c2fde2 RH |
9254 | TREE_USED (t) = 1; |
9255 | ||
9256 | fpscr_values = t; | |
9257 | } | |
9258 | ||
9259 | src = DECL_RTL (fpscr_values); | |
b3a13419 | 9260 | if (!can_create_pseudo_p ()) |
24c2fde2 RH |
9261 | { |
9262 | emit_move_insn (scratch, XEXP (src, 0)); | |
9263 | if (index != 0) | |
9264 | emit_insn (gen_addsi3 (scratch, scratch, GEN_INT (index * 4))); | |
9265 | src = adjust_automodify_address (src, PSImode, scratch, index * 4); | |
9266 | } | |
9267 | else | |
9268 | src = adjust_address (src, PSImode, index * 4); | |
9269 | ||
9270 | dst = get_fpscr_rtx (); | |
9271 | emit_move_insn (dst, src); | |
9272 | } | |
9273 | ||
225e4f43 | 9274 | void |
cf277499 | 9275 | emit_sf_insn (rtx pat) |
225e4f43 | 9276 | { |
225e4f43 | 9277 | emit_insn (pat); |
225e4f43 R |
9278 | } |
9279 | ||
9280 | void | |
cf277499 | 9281 | emit_df_insn (rtx pat) |
225e4f43 | 9282 | { |
225e4f43 | 9283 | emit_insn (pat); |
225e4f43 R |
9284 | } |
9285 | ||
9286 | void | |
cf277499 | 9287 | expand_sf_unop (rtx (*fun) (rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
9288 | { |
9289 | emit_sf_insn ((*fun) (operands[0], operands[1], get_fpscr_rtx ())); | |
9290 | } | |
9291 | ||
9292 | void | |
cf277499 | 9293 | expand_sf_binop (rtx (*fun) (rtx, rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
9294 | { |
9295 | emit_sf_insn ((*fun) (operands[0], operands[1], operands[2], | |
9296 | get_fpscr_rtx ())); | |
9297 | } | |
9298 | ||
9299 | void | |
cf277499 | 9300 | expand_df_unop (rtx (*fun) (rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
9301 | { |
9302 | emit_df_insn ((*fun) (operands[0], operands[1], get_fpscr_rtx ())); | |
9303 | } | |
9304 | ||
9305 | void | |
cf277499 | 9306 | expand_df_binop (rtx (*fun) (rtx, rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
9307 | { |
9308 | emit_df_insn ((*fun) (operands[0], operands[1], operands[2], | |
832a3292 | 9309 | get_fpscr_rtx ())); |
225e4f43 | 9310 | } |
225e4f43 | 9311 | \f |
cf277499 | 9312 | static rtx get_free_reg (HARD_REG_SET); |
9f09b1f2 R |
9313 | |
9314 | /* This function returns a register to use to load the address to load | |
9315 | the fpscr from. Currently it always returns r1 or r7, but when we are | |
9316 | able to use pseudo registers after combine, or have a better mechanism | |
9317 | for choosing a register, it should be done here. */ | |
9318 | /* REGS_LIVE is the liveness information for the point for which we | |
9319 | need this allocation. In some bare-bones exit blocks, r1 is live at the | |
9320 | start. We can even have all of r0..r3 being live: | |
9321 | __complex__ long long f (double d) { if (d == 0) return 2; else return 3; } | |
9322 | INSN before which new insns are placed with will clobber the register | |
9323 | we return. If a basic block consists only of setting the return value | |
9324 | register to a pseudo and using that register, the return value is not | |
9325 | live before or after this block, yet we we'll insert our insns right in | |
9326 | the middle. */ | |
225e4f43 | 9327 | |
9f09b1f2 | 9328 | static rtx |
cf277499 | 9329 | get_free_reg (HARD_REG_SET regs_live) |
225e4f43 | 9330 | { |
9f09b1f2 R |
9331 | if (! TEST_HARD_REG_BIT (regs_live, 1)) |
9332 | return gen_rtx_REG (Pmode, 1); | |
225e4f43 | 9333 | |
42db504c | 9334 | /* Hard reg 1 is live; since this is a small register classes target, |
9f09b1f2 | 9335 | there shouldn't be anything but a jump before the function end. */ |
f5b9e7c9 NS |
9336 | gcc_assert (!TEST_HARD_REG_BIT (regs_live, 7)); |
9337 | return gen_rtx_REG (Pmode, 7); | |
9f09b1f2 | 9338 | } |
225e4f43 | 9339 | |
f676971a | 9340 | /* This function will set the fpscr from memory. |
9f09b1f2 R |
9341 | MODE is the mode we are setting it to. */ |
9342 | void | |
cf277499 | 9343 | fpscr_set_from_mem (int mode, HARD_REG_SET regs_live) |
9f09b1f2 | 9344 | { |
81f40b79 | 9345 | enum attr_fp_mode fp_mode = (enum attr_fp_mode) mode; |
24c2fde2 | 9346 | enum attr_fp_mode norm_mode = ACTUAL_NORMAL_MODE (FP_MODE); |
5e77c6f9 | 9347 | rtx addr_reg; |
9f09b1f2 | 9348 | |
b3a13419 | 9349 | addr_reg = !can_create_pseudo_p () ? get_free_reg (regs_live) : NULL_RTX; |
24c2fde2 | 9350 | emit_fpu_switch (addr_reg, fp_mode == norm_mode); |
225e4f43 | 9351 | } |
9342c0c4 R |
9352 | |
9353 | /* Is the given character a logical line separator for the assembler? */ | |
9354 | #ifndef IS_ASM_LOGICAL_LINE_SEPARATOR | |
980d8882 | 9355 | #define IS_ASM_LOGICAL_LINE_SEPARATOR(C, STR) ((C) == ';') |
9342c0c4 R |
9356 | #endif |
9357 | ||
9358 | int | |
cf277499 | 9359 | sh_insn_length_adjustment (rtx insn) |
9342c0c4 R |
9360 | { |
9361 | /* Instructions with unfilled delay slots take up an extra two bytes for | |
9362 | the nop in the delay slot. */ | |
f3536097 | 9363 | if (((NONJUMP_INSN_P (insn) |
c1b92d09 R |
9364 | && GET_CODE (PATTERN (insn)) != USE |
9365 | && GET_CODE (PATTERN (insn)) != CLOBBER) | |
f3536097 | 9366 | || CALL_P (insn) |
69b18163 | 9367 | || (JUMP_P (insn) && !JUMP_TABLE_DATA_P (insn))) |
9342c0c4 R |
9368 | && GET_CODE (PATTERN (NEXT_INSN (PREV_INSN (insn)))) != SEQUENCE |
9369 | && get_attr_needs_delay_slot (insn) == NEEDS_DELAY_SLOT_YES) | |
9370 | return 2; | |
9371 | ||
3a8699c7 AO |
9372 | /* SH2e has a bug that prevents the use of annulled branches, so if |
9373 | the delay slot is not filled, we'll have to put a NOP in it. */ | |
b757d36f | 9374 | if (sh_cpu_attr == CPU_SH2E |
69b18163 | 9375 | && JUMP_P (insn) && !JUMP_TABLE_DATA_P (insn) |
3a8699c7 AO |
9376 | && get_attr_type (insn) == TYPE_CBRANCH |
9377 | && GET_CODE (PATTERN (NEXT_INSN (PREV_INSN (insn)))) != SEQUENCE) | |
9378 | return 2; | |
9379 | ||
9342c0c4 | 9380 | /* sh-dsp parallel processing insn take four bytes instead of two. */ |
f676971a | 9381 | |
f3536097 | 9382 | if (NONJUMP_INSN_P (insn)) |
9342c0c4 R |
9383 | { |
9384 | int sum = 0; | |
9385 | rtx body = PATTERN (insn); | |
0a2aaacc | 9386 | const char *templ; |
8ceaac3c | 9387 | char c; |
9342c0c4 R |
9388 | int maybe_label = 1; |
9389 | ||
9390 | if (GET_CODE (body) == ASM_INPUT) | |
0a2aaacc | 9391 | templ = XSTR (body, 0); |
9342c0c4 | 9392 | else if (asm_noperands (body) >= 0) |
0a2aaacc | 9393 | templ |
343d66ad | 9394 | = decode_asm_operands (body, NULL, NULL, NULL, NULL, NULL); |
9342c0c4 R |
9395 | else |
9396 | return 0; | |
9397 | do | |
9398 | { | |
9399 | int ppi_adjust = 0; | |
9400 | ||
9401 | do | |
0a2aaacc | 9402 | c = *templ++; |
9342c0c4 R |
9403 | while (c == ' ' || c == '\t'); |
9404 | /* all sh-dsp parallel-processing insns start with p. | |
9405 | The only non-ppi sh insn starting with p is pref. | |
9406 | The only ppi starting with pr is prnd. */ | |
0a2aaacc | 9407 | if ((c == 'p' || c == 'P') && strncasecmp ("re", templ, 2)) |
9342c0c4 R |
9408 | ppi_adjust = 2; |
9409 | /* The repeat pseudo-insn expands two three insns, a total of | |
9410 | six bytes in size. */ | |
9411 | else if ((c == 'r' || c == 'R') | |
0a2aaacc | 9412 | && ! strncasecmp ("epeat", templ, 5)) |
9342c0c4 | 9413 | ppi_adjust = 4; |
980d8882 | 9414 | while (c && c != '\n' |
0a2aaacc | 9415 | && ! IS_ASM_LOGICAL_LINE_SEPARATOR (c, templ)) |
9342c0c4 R |
9416 | { |
9417 | /* If this is a label, it is obviously not a ppi insn. */ | |
9418 | if (c == ':' && maybe_label) | |
9419 | { | |
9420 | ppi_adjust = 0; | |
9421 | break; | |
9422 | } | |
9423 | else if (c == '\'' || c == '"') | |
9424 | maybe_label = 0; | |
0a2aaacc | 9425 | c = *templ++; |
9342c0c4 R |
9426 | } |
9427 | sum += ppi_adjust; | |
9428 | maybe_label = c != ':'; | |
9429 | } | |
9430 | while (c); | |
9431 | return sum; | |
9432 | } | |
9433 | return 0; | |
9434 | } | |
1a66cd67 | 9435 | \f |
cdeed45a KK |
9436 | /* Return TRUE for a valid displacement for the REG+disp addressing |
9437 | with MODE. */ | |
9438 | ||
9439 | /* ??? The SH2e does not have the REG+disp addressing mode when loading values | |
9440 | into the FRx registers. We implement this by setting the maximum offset | |
9441 | to zero when the value is SFmode. This also restricts loading of SFmode | |
9442 | values into the integer registers, but that can't be helped. */ | |
9443 | ||
9444 | /* The SH allows a displacement in a QI or HI amode, but only when the | |
9445 | other operand is R0. GCC doesn't handle this very well, so we forgot | |
9446 | all of that. | |
9447 | ||
9448 | A legitimate index for a QI or HI is 0, SI can be any number 0..63, | |
9449 | DI can be any number 0..60. */ | |
9450 | ||
9451 | bool | |
9452 | sh_legitimate_index_p (enum machine_mode mode, rtx op) | |
9453 | { | |
f3536097 | 9454 | if (CONST_INT_P (op)) |
cdeed45a KK |
9455 | { |
9456 | if (TARGET_SHMEDIA) | |
9457 | { | |
9458 | int size; | |
9459 | ||
9460 | /* Check if this the address of an unaligned load / store. */ | |
9461 | if (mode == VOIDmode) | |
9462 | return CONST_OK_FOR_I06 (INTVAL (op)); | |
9463 | ||
9464 | size = GET_MODE_SIZE (mode); | |
9465 | return (!(INTVAL (op) & (size - 1)) | |
9466 | && INTVAL (op) >= -512 * size | |
9467 | && INTVAL (op) < 512 * size); | |
9468 | } | |
9469 | ||
9470 | if (TARGET_SH2A) | |
9471 | { | |
9472 | if (GET_MODE_SIZE (mode) == 1 | |
9473 | && (unsigned) INTVAL (op) < 4096) | |
9474 | return true; | |
9475 | } | |
9476 | ||
9477 | if ((GET_MODE_SIZE (mode) == 4 | |
9478 | && (unsigned) INTVAL (op) < 64 | |
9479 | && !(INTVAL (op) & 3) | |
9480 | && !(TARGET_SH2E && mode == SFmode)) | |
9481 | || (GET_MODE_SIZE (mode) == 4 | |
9482 | && (unsigned) INTVAL (op) < 16383 | |
9483 | && !(INTVAL (op) & 3) && TARGET_SH2A)) | |
9484 | return true; | |
9485 | ||
9486 | if ((GET_MODE_SIZE (mode) == 8 | |
9487 | && (unsigned) INTVAL (op) < 60 | |
9488 | && !(INTVAL (op) & 3) | |
9489 | && !((TARGET_SH4 || TARGET_SH2A) && mode == DFmode)) | |
9490 | || ((GET_MODE_SIZE (mode)==8) | |
9491 | && (unsigned) INTVAL (op) < 8192 | |
9492 | && !(INTVAL (op) & (TARGET_SH2A_DOUBLE ? 7 : 3)) | |
9493 | && (TARGET_SH2A && mode == DFmode))) | |
9494 | return true; | |
9495 | } | |
9496 | ||
9497 | return false; | |
9498 | } | |
9499 | ||
9500 | /* Recognize an RTL expression that is a valid memory address for | |
9501 | an instruction. | |
9502 | The MODE argument is the machine mode for the MEM expression | |
9503 | that wants to use this address. | |
9504 | Allow REG | |
9505 | REG+disp | |
9506 | REG+r0 | |
9507 | REG++ | |
9508 | --REG */ | |
9509 | ||
1bb99877 | 9510 | static bool |
cdeed45a KK |
9511 | sh_legitimate_address_p (enum machine_mode mode, rtx x, bool strict) |
9512 | { | |
9513 | if (MAYBE_BASE_REGISTER_RTX_P (x, strict)) | |
9514 | return true; | |
9515 | else if ((GET_CODE (x) == POST_INC || GET_CODE (x) == PRE_DEC) | |
9516 | && ! TARGET_SHMEDIA | |
9517 | && MAYBE_BASE_REGISTER_RTX_P (XEXP (x, 0), strict)) | |
9518 | return true; | |
9519 | else if (GET_CODE (x) == PLUS | |
9520 | && (mode != PSImode || reload_completed)) | |
9521 | { | |
9522 | rtx xop0 = XEXP (x, 0); | |
9523 | rtx xop1 = XEXP (x, 1); | |
9524 | ||
9525 | if (GET_MODE_SIZE (mode) <= 8 | |
9526 | && MAYBE_BASE_REGISTER_RTX_P (xop0, strict) | |
9527 | && sh_legitimate_index_p (mode, xop1)) | |
9528 | return true; | |
9529 | ||
9530 | if ((ALLOW_INDEXED_ADDRESS || GET_MODE (x) == DImode | |
9531 | || ((xop0 == stack_pointer_rtx | |
9532 | || xop0 == hard_frame_pointer_rtx) | |
9533 | && REG_P (xop1) && REGNO (xop1) == R0_REG) | |
9534 | || ((xop1 == stack_pointer_rtx | |
9535 | || xop1 == hard_frame_pointer_rtx) | |
9536 | && REG_P (xop0) && REGNO (xop0) == R0_REG)) | |
9537 | && ((!TARGET_SHMEDIA && GET_MODE_SIZE (mode) <= 4) | |
9538 | || (TARGET_SHMEDIA && GET_MODE_SIZE (mode) <= 8) | |
9539 | || ((TARGET_SH4 || TARGET_SH2A_DOUBLE) | |
9540 | && TARGET_FMOVD && mode == DFmode))) | |
9541 | { | |
9542 | if (MAYBE_BASE_REGISTER_RTX_P (xop1, strict) | |
9543 | && MAYBE_INDEX_REGISTER_RTX_P (xop0, strict)) | |
9544 | return true; | |
9545 | if (MAYBE_INDEX_REGISTER_RTX_P (xop1, strict) | |
9546 | && MAYBE_BASE_REGISTER_RTX_P (xop0, strict)) | |
9547 | return true; | |
9548 | } | |
9549 | } | |
9550 | ||
9551 | return false; | |
9552 | } | |
9553 | \f | |
55eec660 AO |
9554 | /* Return TRUE if X references a SYMBOL_REF or LABEL_REF whose symbol |
9555 | isn't protected by a PIC unspec. */ | |
1a66cd67 | 9556 | int |
cf277499 | 9557 | nonpic_symbol_mentioned_p (rtx x) |
1a66cd67 AO |
9558 | { |
9559 | register const char *fmt; | |
9560 | register int i; | |
9561 | ||
fa5322fa AO |
9562 | if (GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == LABEL_REF |
9563 | || GET_CODE (x) == PC) | |
1a66cd67 AO |
9564 | return 1; |
9565 | ||
fa5322fa AO |
9566 | /* We don't want to look into the possible MEM location of a |
9567 | CONST_DOUBLE, since we're not going to use it, in general. */ | |
9568 | if (GET_CODE (x) == CONST_DOUBLE) | |
9569 | return 0; | |
9570 | ||
1a66cd67 | 9571 | if (GET_CODE (x) == UNSPEC |
4773afa4 AO |
9572 | && (XINT (x, 1) == UNSPEC_PIC |
9573 | || XINT (x, 1) == UNSPEC_GOT | |
9574 | || XINT (x, 1) == UNSPEC_GOTOFF | |
fa5322fa | 9575 | || XINT (x, 1) == UNSPEC_GOTPLT |
463f02cd KK |
9576 | || XINT (x, 1) == UNSPEC_GOTTPOFF |
9577 | || XINT (x, 1) == UNSPEC_DTPOFF | |
58f451f0 | 9578 | || XINT (x, 1) == UNSPEC_TPOFF |
dc3ba671 RS |
9579 | || XINT (x, 1) == UNSPEC_PLT |
9580 | || XINT (x, 1) == UNSPEC_SYMOFF | |
9581 | || XINT (x, 1) == UNSPEC_PCREL_SYMOFF)) | |
832a3292 | 9582 | return 0; |
1a66cd67 AO |
9583 | |
9584 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
9585 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
9586 | { | |
9587 | if (fmt[i] == 'E') | |
9588 | { | |
9589 | register int j; | |
9590 | ||
9591 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
9592 | if (nonpic_symbol_mentioned_p (XVECEXP (x, i, j))) | |
9593 | return 1; | |
9594 | } | |
9595 | else if (fmt[i] == 'e' && nonpic_symbol_mentioned_p (XEXP (x, i))) | |
9596 | return 1; | |
9597 | } | |
9598 | ||
9599 | return 0; | |
9600 | } | |
9601 | ||
9602 | /* Convert a non-PIC address in `orig' to a PIC address using @GOT or | |
6f317ef3 | 9603 | @GOTOFF in `reg'. */ |
1a66cd67 | 9604 | rtx |
cf277499 SB |
9605 | legitimize_pic_address (rtx orig, enum machine_mode mode ATTRIBUTE_UNUSED, |
9606 | rtx reg) | |
1a66cd67 | 9607 | { |
81f40b79 | 9608 | if (tls_symbolic_operand (orig, Pmode) != TLS_MODEL_NONE) |
463f02cd KK |
9609 | return orig; |
9610 | ||
1a66cd67 | 9611 | if (GET_CODE (orig) == LABEL_REF |
675ff4c7 | 9612 | || (GET_CODE (orig) == SYMBOL_REF && SYMBOL_REF_LOCAL_P (orig))) |
1a66cd67 AO |
9613 | { |
9614 | if (reg == 0) | |
9615 | reg = gen_reg_rtx (Pmode); | |
9616 | ||
9617 | emit_insn (gen_symGOTOFF2reg (reg, orig)); | |
9618 | return reg; | |
9619 | } | |
9620 | else if (GET_CODE (orig) == SYMBOL_REF) | |
9621 | { | |
9622 | if (reg == 0) | |
9623 | reg = gen_reg_rtx (Pmode); | |
9624 | ||
9625 | emit_insn (gen_symGOT2reg (reg, orig)); | |
9626 | return reg; | |
9627 | } | |
9628 | return orig; | |
9629 | } | |
b91455de | 9630 | |
506d7b68 PB |
9631 | /* Try machine-dependent ways of modifying an illegitimate address |
9632 | to be legitimate. If we find one, return the new, valid address. | |
9633 | Otherwise, return X. | |
9634 | ||
9635 | For the SH, if X is almost suitable for indexing, but the offset is | |
9636 | out of range, convert it into a normal form so that CSE has a chance | |
9637 | of reducing the number of address registers used. */ | |
9638 | ||
9639 | static rtx | |
9640 | sh_legitimize_address (rtx x, rtx oldx, enum machine_mode mode) | |
9641 | { | |
9642 | if (flag_pic) | |
9643 | x = legitimize_pic_address (oldx, mode, NULL_RTX); | |
9644 | ||
9645 | if (GET_CODE (x) == PLUS | |
9646 | && (GET_MODE_SIZE (mode) == 4 | |
9647 | || GET_MODE_SIZE (mode) == 8) | |
f3536097 | 9648 | && CONST_INT_P (XEXP (x, 1)) |
506d7b68 PB |
9649 | && BASE_REGISTER_RTX_P (XEXP (x, 0)) |
9650 | && ! TARGET_SHMEDIA | |
9651 | && ! ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && mode == DFmode) | |
9652 | && ! (TARGET_SH2E && mode == SFmode)) | |
9653 | { | |
9654 | rtx index_rtx = XEXP (x, 1); | |
9655 | HOST_WIDE_INT offset = INTVAL (index_rtx), offset_base; | |
9656 | rtx sum; | |
9657 | ||
9658 | /* On rare occasions, we might get an unaligned pointer | |
9659 | that is indexed in a way to give an aligned address. | |
9660 | Therefore, keep the lower two bits in offset_base. */ | |
9661 | /* Instead of offset_base 128..131 use 124..127, so that | |
9662 | simple add suffices. */ | |
9663 | if (offset > 127) | |
9664 | offset_base = ((offset + 4) & ~60) - 4; | |
9665 | else | |
9666 | offset_base = offset & ~60; | |
9667 | ||
9668 | /* Sometimes the normal form does not suit DImode. We | |
9669 | could avoid that by using smaller ranges, but that | |
9670 | would give less optimized code when SImode is | |
9671 | prevalent. */ | |
9672 | if (GET_MODE_SIZE (mode) + offset - offset_base <= 64) | |
9673 | { | |
9674 | sum = expand_binop (Pmode, add_optab, XEXP (x, 0), | |
9675 | GEN_INT (offset_base), NULL_RTX, 0, | |
9676 | OPTAB_LIB_WIDEN); | |
9677 | ||
9678 | return gen_rtx_PLUS (Pmode, sum, GEN_INT (offset - offset_base)); | |
9679 | } | |
9680 | } | |
9681 | ||
9682 | return x; | |
9683 | } | |
9684 | ||
6f50eb9c KK |
9685 | /* Attempt to replace *P, which is an address that needs reloading, with |
9686 | a valid memory address for an operand of mode MODE. | |
9687 | Like for sh_legitimize_address, for the SH we try to get a normal form | |
9688 | of the address. That will allow inheritance of the address reloads. */ | |
9689 | ||
9690 | bool | |
9691 | sh_legitimize_reload_address (rtx *p, enum machine_mode mode, int opnum, | |
9692 | int itype) | |
9693 | { | |
9694 | enum reload_type type = (enum reload_type) itype; | |
9695 | ||
9696 | if (GET_CODE (*p) == PLUS | |
9697 | && (GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) | |
9698 | && CONST_INT_P (XEXP (*p, 1)) | |
7140f0f9 | 9699 | && MAYBE_BASE_REGISTER_RTX_P (XEXP (*p, 0), true) |
6f50eb9c KK |
9700 | && ! TARGET_SHMEDIA |
9701 | && ! (TARGET_SH4 && mode == DFmode) | |
9702 | && ! (mode == PSImode && type == RELOAD_FOR_INPUT_ADDRESS) | |
9703 | && (ALLOW_INDEXED_ADDRESS | |
9704 | || XEXP (*p, 0) == stack_pointer_rtx | |
9705 | || XEXP (*p, 0) == hard_frame_pointer_rtx)) | |
9706 | { | |
9707 | rtx index_rtx = XEXP (*p, 1); | |
9708 | HOST_WIDE_INT offset = INTVAL (index_rtx), offset_base; | |
9709 | rtx sum; | |
9710 | ||
9711 | if (TARGET_SH2A && mode == DFmode && (offset & 0x7)) | |
9712 | { | |
9713 | push_reload (*p, NULL_RTX, p, NULL, | |
9714 | BASE_REG_CLASS, Pmode, VOIDmode, 0, 0, opnum, type); | |
9715 | goto win; | |
9716 | } | |
9717 | if (TARGET_SH2E && mode == SFmode) | |
9718 | { | |
9719 | *p = copy_rtx (*p); | |
9720 | push_reload (*p, NULL_RTX, p, NULL, | |
9721 | BASE_REG_CLASS, Pmode, VOIDmode, 0, 0, opnum, type); | |
9722 | goto win; | |
9723 | } | |
9724 | /* Instead of offset_base 128..131 use 124..127, so that | |
9725 | simple add suffices. */ | |
9726 | if (offset > 127) | |
9727 | offset_base = ((offset + 4) & ~60) - 4; | |
9728 | else | |
9729 | offset_base = offset & ~60; | |
9730 | /* Sometimes the normal form does not suit DImode. We could avoid | |
9731 | that by using smaller ranges, but that would give less optimized | |
9732 | code when SImode is prevalent. */ | |
7140f0f9 | 9733 | if (GET_MODE_SIZE (mode) + offset - offset_base <= 64) |
6f50eb9c KK |
9734 | { |
9735 | sum = gen_rtx_PLUS (Pmode, XEXP (*p, 0), GEN_INT (offset_base)); | |
9736 | *p = gen_rtx_PLUS (Pmode, sum, GEN_INT (offset - offset_base)); | |
9737 | push_reload (sum, NULL_RTX, &XEXP (*p, 0), NULL, | |
9738 | BASE_REG_CLASS, Pmode, VOIDmode, 0, 0, opnum, type); | |
9739 | goto win; | |
9740 | } | |
9741 | } | |
9742 | /* We must re-recognize what we created before. */ | |
9743 | else if (GET_CODE (*p) == PLUS | |
9744 | && (GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) | |
9745 | && GET_CODE (XEXP (*p, 0)) == PLUS | |
9746 | && CONST_INT_P (XEXP (XEXP (*p, 0), 1)) | |
7140f0f9 | 9747 | && MAYBE_BASE_REGISTER_RTX_P (XEXP (XEXP (*p, 0), 0), true) |
6f50eb9c KK |
9748 | && CONST_INT_P (XEXP (*p, 1)) |
9749 | && ! TARGET_SHMEDIA | |
9750 | && ! (TARGET_SH2E && mode == SFmode)) | |
9751 | { | |
9752 | /* Because this address is so complex, we know it must have | |
9753 | been created by LEGITIMIZE_RELOAD_ADDRESS before; thus, | |
9754 | it is already unshared, and needs no further unsharing. */ | |
9755 | push_reload (XEXP (*p, 0), NULL_RTX, &XEXP (*p, 0), NULL, | |
9756 | BASE_REG_CLASS, Pmode, VOIDmode, 0, 0, opnum, type); | |
9757 | goto win; | |
9758 | } | |
9759 | ||
9760 | return false; | |
9761 | ||
9762 | win: | |
9763 | return true; | |
9764 | } | |
9765 | ||
b91455de KK |
9766 | /* Mark the use of a constant in the literal table. If the constant |
9767 | has multiple labels, make it unique. */ | |
c1b92d09 | 9768 | static rtx |
cf277499 | 9769 | mark_constant_pool_use (rtx x) |
b91455de KK |
9770 | { |
9771 | rtx insn, lab, pattern; | |
9772 | ||
9773 | if (x == NULL) | |
9774 | return x; | |
9775 | ||
9776 | switch (GET_CODE (x)) | |
9777 | { | |
9778 | case LABEL_REF: | |
9779 | x = XEXP (x, 0); | |
9780 | case CODE_LABEL: | |
9781 | break; | |
9782 | default: | |
9783 | return x; | |
9784 | } | |
9785 | ||
9786 | /* Get the first label in the list of labels for the same constant | |
9787 | and delete another labels in the list. */ | |
9788 | lab = x; | |
9789 | for (insn = PREV_INSN (x); insn; insn = PREV_INSN (insn)) | |
9790 | { | |
f3536097 | 9791 | if (!LABEL_P (insn) |
b91455de KK |
9792 | || LABEL_REFS (insn) != NEXT_INSN (insn)) |
9793 | break; | |
9794 | lab = insn; | |
9795 | } | |
9796 | ||
9797 | for (insn = LABEL_REFS (lab); insn; insn = LABEL_REFS (insn)) | |
9798 | INSN_DELETED_P (insn) = 1; | |
9799 | ||
9800 | /* Mark constants in a window. */ | |
9801 | for (insn = NEXT_INSN (x); insn; insn = NEXT_INSN (insn)) | |
9802 | { | |
f3536097 | 9803 | if (!NONJUMP_INSN_P (insn)) |
b91455de KK |
9804 | continue; |
9805 | ||
9806 | pattern = PATTERN (insn); | |
9807 | if (GET_CODE (pattern) != UNSPEC_VOLATILE) | |
9808 | continue; | |
9809 | ||
9810 | switch (XINT (pattern, 1)) | |
9811 | { | |
9812 | case UNSPECV_CONST2: | |
9813 | case UNSPECV_CONST4: | |
9814 | case UNSPECV_CONST8: | |
9815 | XVECEXP (pattern, 0, 1) = const1_rtx; | |
9816 | break; | |
9817 | case UNSPECV_WINDOW_END: | |
9818 | if (XVECEXP (pattern, 0, 0) == x) | |
9819 | return lab; | |
9820 | break; | |
9821 | case UNSPECV_CONST_END: | |
9822 | return lab; | |
9823 | default: | |
9824 | break; | |
9825 | } | |
9826 | } | |
9827 | ||
9828 | return lab; | |
9829 | } | |
6e7b03e1 AH |
9830 | \f |
9831 | /* Return true if it's possible to redirect BRANCH1 to the destination | |
9832 | of an unconditional jump BRANCH2. We only want to do this if the | |
9833 | resulting branch will have a short displacement. */ | |
f676971a | 9834 | int |
cf277499 | 9835 | sh_can_redirect_branch (rtx branch1, rtx branch2) |
6e7b03e1 AH |
9836 | { |
9837 | if (flag_expensive_optimizations && simplejump_p (branch2)) | |
9838 | { | |
9839 | rtx dest = XEXP (SET_SRC (single_set (branch2)), 0); | |
9840 | rtx insn; | |
9841 | int distance; | |
f676971a EC |
9842 | |
9843 | for (distance = 0, insn = NEXT_INSN (branch1); | |
9844 | insn && distance < 256; | |
6e7b03e1 AH |
9845 | insn = PREV_INSN (insn)) |
9846 | { | |
f676971a | 9847 | if (insn == dest) |
6e7b03e1 AH |
9848 | return 1; |
9849 | else | |
9850 | distance += get_attr_length (insn); | |
9851 | } | |
f676971a EC |
9852 | for (distance = 0, insn = NEXT_INSN (branch1); |
9853 | insn && distance < 256; | |
6e7b03e1 AH |
9854 | insn = NEXT_INSN (insn)) |
9855 | { | |
f676971a | 9856 | if (insn == dest) |
6e7b03e1 AH |
9857 | return 1; |
9858 | else | |
9859 | distance += get_attr_length (insn); | |
9860 | } | |
9861 | } | |
9862 | return 0; | |
9863 | } | |
9864 | ||
5e7a8ee0 | 9865 | /* Return nonzero if register old_reg can be renamed to register new_reg. */ |
03b8ec29 | 9866 | int |
cf277499 SB |
9867 | sh_hard_regno_rename_ok (unsigned int old_reg ATTRIBUTE_UNUSED, |
9868 | unsigned int new_reg) | |
03b8ec29 | 9869 | { |
832a3292 KH |
9870 | /* Interrupt functions can only use registers that have already been |
9871 | saved by the prologue, even if they would normally be | |
9872 | call-clobbered. */ | |
03b8ec29 | 9873 | |
6fb5fa3c | 9874 | if (sh_cfun_interrupt_handler_p () && !df_regs_ever_live_p (new_reg)) |
832a3292 | 9875 | return 0; |
03b8ec29 | 9876 | |
832a3292 | 9877 | return 1; |
03b8ec29 DD |
9878 | } |
9879 | ||
95c0af87 | 9880 | /* Function to update the integer COST |
c237e94a ZW |
9881 | based on the relationship between INSN that is dependent on |
9882 | DEP_INSN through the dependence LINK. The default is to make no | |
9883 | adjustment to COST. This can be used for example to specify to | |
9884 | the scheduler that an output- or anti-dependence does not incur | |
95c0af87 JR |
9885 | the same cost as a data-dependence. The return value should be |
9886 | the new value for COST. */ | |
c237e94a | 9887 | static int |
cf277499 | 9888 | sh_adjust_cost (rtx insn, rtx link ATTRIBUTE_UNUSED, rtx dep_insn, int cost) |
c237e94a | 9889 | { |
c49439f1 | 9890 | rtx reg, use_pat; |
c237e94a | 9891 | |
2ad65b0e SC |
9892 | if (TARGET_SHMEDIA) |
9893 | { | |
9894 | /* On SHmedia, if the dependence is an anti-dependence or | |
24746a42 | 9895 | output-dependence, there is no cost. */ |
2ad65b0e | 9896 | if (REG_NOTE_KIND (link) != 0) |
73a4d10b R |
9897 | { |
9898 | /* However, dependencies between target register loads and | |
9899 | uses of the register in a subsequent block that are separated | |
9900 | by a conditional branch are not modelled - we have to do with | |
9901 | the anti-dependency between the target register load and the | |
9902 | conditional branch that ends the current block. */ | |
9903 | if (REG_NOTE_KIND (link) == REG_DEP_ANTI | |
9904 | && GET_CODE (PATTERN (dep_insn)) == SET | |
9905 | && (get_attr_type (dep_insn) == TYPE_PT_MEDIA | |
9906 | || get_attr_type (dep_insn) == TYPE_PTABS_MEDIA) | |
9907 | && get_attr_type (insn) == TYPE_CBRANCH_MEDIA) | |
9908 | { | |
9909 | int orig_cost = cost; | |
9910 | rtx note = find_reg_note (insn, REG_BR_PROB, 0); | |
9911 | rtx target = ((! note | |
9912 | || INTVAL (XEXP (note, 0)) * 2 < REG_BR_PROB_BASE) | |
9913 | ? insn : JUMP_LABEL (insn)); | |
9914 | /* On the likely path, the branch costs 1, on the unlikely path, | |
9915 | it costs 3. */ | |
9916 | cost--; | |
9917 | do | |
9918 | target = next_active_insn (target); | |
9919 | while (target && ! flow_dependent_p (target, dep_insn) | |
9920 | && --cost > 0); | |
9921 | /* If two branches are executed in immediate succession, with the | |
9922 | first branch properly predicted, this causes a stall at the | |
9923 | second branch, hence we won't need the target for the | |
9924 | second branch for two cycles after the launch of the first | |
9925 | branch. */ | |
9926 | if (cost > orig_cost - 2) | |
9927 | cost = orig_cost - 2; | |
9928 | } | |
9929 | else | |
9930 | cost = 0; | |
9931 | } | |
2ad65b0e | 9932 | |
73a4d10b R |
9933 | else if (get_attr_is_mac_media (insn) |
9934 | && get_attr_is_mac_media (dep_insn)) | |
9935 | cost = 1; | |
9936 | ||
9937 | else if (! reload_completed | |
9938 | && GET_CODE (PATTERN (insn)) == SET | |
9939 | && GET_CODE (SET_SRC (PATTERN (insn))) == FLOAT | |
9940 | && GET_CODE (PATTERN (dep_insn)) == SET | |
9941 | && fp_arith_reg_operand (SET_SRC (PATTERN (dep_insn)), VOIDmode) | |
9942 | && cost < 4) | |
9943 | cost = 4; | |
9944 | /* Schedule the ptabs for a casesi_jump_media in preference to stuff | |
9945 | that is needed at the target. */ | |
9946 | else if (get_attr_type (insn) == TYPE_JUMP_MEDIA | |
9947 | && ! flow_dependent_p (insn, dep_insn)) | |
9948 | cost--; | |
2ad65b0e | 9949 | } |
c49439f1 | 9950 | else if (REG_NOTE_KIND (link) == 0) |
c237e94a | 9951 | { |
78d310c2 R |
9952 | enum attr_type type; |
9953 | rtx dep_set; | |
c49439f1 R |
9954 | |
9955 | if (recog_memoized (insn) < 0 | |
9956 | || recog_memoized (dep_insn) < 0) | |
95c0af87 | 9957 | return cost; |
c49439f1 | 9958 | |
78d310c2 | 9959 | dep_set = single_set (dep_insn); |
c49439f1 | 9960 | |
78d310c2 R |
9961 | /* The latency that we specify in the scheduling description refers |
9962 | to the actual output, not to an auto-increment register; for that, | |
9963 | the latency is one. */ | |
9964 | if (dep_set && MEM_P (SET_SRC (dep_set)) && cost > 1) | |
9965 | { | |
9966 | rtx set = single_set (insn); | |
9967 | ||
9968 | if (set | |
9969 | && !reg_mentioned_p (SET_DEST (dep_set), SET_SRC (set)) | |
9970 | && (!MEM_P (SET_DEST (set)) | |
9971 | || !reg_mentioned_p (SET_DEST (dep_set), | |
9972 | XEXP (SET_DEST (set), 0)))) | |
9973 | cost = 1; | |
9974 | } | |
c237e94a ZW |
9975 | /* The only input for a call that is timing-critical is the |
9976 | function's address. */ | |
f3536097 | 9977 | if (CALL_P (insn)) |
c49439f1 R |
9978 | { |
9979 | rtx call = PATTERN (insn); | |
9980 | ||
9981 | if (GET_CODE (call) == PARALLEL) | |
9982 | call = XVECEXP (call, 0 ,0); | |
9983 | if (GET_CODE (call) == SET) | |
9984 | call = SET_SRC (call); | |
f3536097 | 9985 | if (GET_CODE (call) == CALL && MEM_P (XEXP (call, 0)) |
73a4d10b R |
9986 | /* sibcalli_thunk uses a symbol_ref in an unspec. */ |
9987 | && (GET_CODE (XEXP (XEXP (call, 0), 0)) == UNSPEC | |
9988 | || ! reg_set_p (XEXP (XEXP (call, 0), 0), dep_insn))) | |
78d310c2 | 9989 | cost -= TARGET_SH4_300 ? 3 : 6; |
c49439f1 | 9990 | } |
c237e94a ZW |
9991 | /* Likewise, the most timing critical input for an sfuncs call |
9992 | is the function address. However, sfuncs typically start | |
9993 | using their arguments pretty quickly. | |
78d310c2 R |
9994 | Assume a four cycle delay for SH4 before they are needed. |
9995 | Cached ST40-300 calls are quicker, so assume only a one | |
9996 | cycle delay there. | |
9997 | ??? Maybe we should encode the delays till input registers | |
9998 | are needed by sfuncs into the sfunc call insn. */ | |
c49439f1 R |
9999 | /* All sfunc calls are parallels with at least four components. |
10000 | Exploit this to avoid unnecessary calls to sfunc_uses_reg. */ | |
10001 | else if (GET_CODE (PATTERN (insn)) == PARALLEL | |
10002 | && XVECLEN (PATTERN (insn), 0) >= 4 | |
10003 | && (reg = sfunc_uses_reg (insn))) | |
10004 | { | |
10005 | if (! reg_set_p (reg, dep_insn)) | |
78d310c2 R |
10006 | cost -= TARGET_SH4_300 ? 1 : 4; |
10007 | } | |
10008 | if (TARGET_HARD_SH4 && !TARGET_SH4_300) | |
10009 | { | |
10010 | enum attr_type dep_type = get_attr_type (dep_insn); | |
10011 | ||
10012 | if (dep_type == TYPE_FLOAD || dep_type == TYPE_PCFLOAD) | |
10013 | cost--; | |
10014 | else if ((dep_type == TYPE_LOAD_SI || dep_type == TYPE_PCLOAD_SI) | |
10015 | && (type = get_attr_type (insn)) != TYPE_CALL | |
10016 | && type != TYPE_SFUNC) | |
10017 | cost--; | |
10018 | /* When the preceding instruction loads the shift amount of | |
10019 | the following SHAD/SHLD, the latency of the load is increased | |
10020 | by 1 cycle. */ | |
10021 | if (get_attr_type (insn) == TYPE_DYN_SHIFT | |
10022 | && get_attr_any_int_load (dep_insn) == ANY_INT_LOAD_YES | |
c6ac6e82 | 10023 | && reg_overlap_mentioned_p (SET_DEST (dep_set), |
78d310c2 R |
10024 | XEXP (SET_SRC (single_set (insn)), |
10025 | 1))) | |
10026 | cost++; | |
10027 | /* When an LS group instruction with a latency of less than | |
10028 | 3 cycles is followed by a double-precision floating-point | |
10029 | instruction, FIPR, or FTRV, the latency of the first | |
10030 | instruction is increased to 3 cycles. */ | |
10031 | else if (cost < 3 | |
10032 | && get_attr_insn_class (dep_insn) == INSN_CLASS_LS_GROUP | |
10033 | && get_attr_dfp_comp (insn) == DFP_COMP_YES) | |
10034 | cost = 3; | |
10035 | /* The lsw register of a double-precision computation is ready one | |
10036 | cycle earlier. */ | |
10037 | else if (reload_completed | |
10038 | && get_attr_dfp_comp (dep_insn) == DFP_COMP_YES | |
10039 | && (use_pat = single_set (insn)) | |
10040 | && ! regno_use_in (REGNO (SET_DEST (single_set (dep_insn))), | |
10041 | SET_SRC (use_pat))) | |
10042 | cost -= 1; | |
10043 | ||
10044 | if (get_attr_any_fp_comp (dep_insn) == ANY_FP_COMP_YES | |
10045 | && get_attr_late_fp_use (insn) == LATE_FP_USE_YES) | |
10046 | cost -= 1; | |
10047 | } | |
10048 | else if (TARGET_SH4_300) | |
10049 | { | |
10050 | /* Stores need their input register two cycles later. */ | |
10051 | if (dep_set && cost >= 1 | |
10052 | && ((type = get_attr_type (insn)) == TYPE_STORE | |
10053 | || type == TYPE_PSTORE | |
10054 | || type == TYPE_FSTORE || type == TYPE_MAC_MEM)) | |
10055 | { | |
10056 | rtx set = single_set (insn); | |
10057 | ||
10058 | if (!reg_mentioned_p (SET_SRC (set), XEXP (SET_DEST (set), 0)) | |
10059 | && rtx_equal_p (SET_SRC (set), SET_DEST (dep_set))) | |
10060 | { | |
10061 | cost -= 2; | |
10062 | /* But don't reduce the cost below 1 if the address depends | |
10063 | on a side effect of dep_insn. */ | |
10064 | if (cost < 1 | |
10065 | && modified_in_p (XEXP (SET_DEST (set), 0), dep_insn)) | |
10066 | cost = 1; | |
10067 | } | |
10068 | } | |
10069 | } | |
c49439f1 R |
10070 | } |
10071 | /* An anti-dependence penalty of two applies if the first insn is a double | |
10072 | precision fadd / fsub / fmul. */ | |
78d310c2 R |
10073 | else if (!TARGET_SH4_300 |
10074 | && REG_NOTE_KIND (link) == REG_DEP_ANTI | |
c49439f1 | 10075 | && recog_memoized (dep_insn) >= 0 |
78d310c2 R |
10076 | && (get_attr_type (dep_insn) == TYPE_DFP_ARITH |
10077 | || get_attr_type (dep_insn) == TYPE_DFP_MUL) | |
c49439f1 R |
10078 | /* A lot of alleged anti-flow dependences are fake, |
10079 | so check this one is real. */ | |
10080 | && flow_dependent_p (dep_insn, insn)) | |
c237e94a | 10081 | cost = 2; |
c49439f1 | 10082 | |
c237e94a ZW |
10083 | return cost; |
10084 | } | |
3007d592 | 10085 | |
c49439f1 R |
10086 | /* Check if INSN is flow-dependent on DEP_INSN. Can also be used to check |
10087 | if DEP_INSN is anti-flow dependent on INSN. */ | |
10088 | static int | |
cf277499 | 10089 | flow_dependent_p (rtx insn, rtx dep_insn) |
c49439f1 R |
10090 | { |
10091 | rtx tmp = PATTERN (insn); | |
10092 | ||
10093 | note_stores (PATTERN (dep_insn), flow_dependent_p_1, &tmp); | |
10094 | return tmp == NULL_RTX; | |
10095 | } | |
10096 | ||
10097 | /* A helper function for flow_dependent_p called through note_stores. */ | |
10098 | static void | |
7bc980e1 | 10099 | flow_dependent_p_1 (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) |
c49439f1 R |
10100 | { |
10101 | rtx * pinsn = (rtx *) data; | |
10102 | ||
10103 | if (*pinsn && reg_referenced_p (x, *pinsn)) | |
10104 | *pinsn = NULL_RTX; | |
10105 | } | |
10106 | ||
b48f503c | 10107 | /* For use by sh_allocate_initial_value. Note that sh.md contains some |
3007d592 R |
10108 | 'special function' patterns (type sfunc) that clobber pr, but that |
10109 | do not look like function calls to leaf_function_p. Hence we must | |
10110 | do this extra check. */ | |
b48f503c | 10111 | static int |
cf277499 | 10112 | sh_pr_n_sets (void) |
3007d592 | 10113 | { |
6fb5fa3c | 10114 | return DF_REG_DEF_COUNT (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG); |
3007d592 | 10115 | } |
fa5322fa | 10116 | |
b48f503c KK |
10117 | /* Return where to allocate pseudo for a given hard register initial |
10118 | value. */ | |
10119 | static rtx | |
10120 | sh_allocate_initial_value (rtx hard_reg) | |
10121 | { | |
10122 | rtx x; | |
10123 | ||
10124 | if (REGNO (hard_reg) == (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG)) | |
10125 | { | |
10126 | if (current_function_is_leaf | |
10127 | && ! sh_pr_n_sets () | |
10128 | && ! (TARGET_SHCOMPACT | |
38173d38 | 10129 | && ((crtl->args.info.call_cookie |
b48f503c | 10130 | & ~ CALL_COOKIE_RET_TRAMP (1)) |
e3b5732b | 10131 | || crtl->saves_all_registers))) |
b48f503c KK |
10132 | x = hard_reg; |
10133 | else | |
57d38024 | 10134 | x = gen_frame_mem (Pmode, return_address_pointer_rtx); |
b48f503c KK |
10135 | } |
10136 | else | |
10137 | x = NULL_RTX; | |
10138 | ||
10139 | return x; | |
10140 | } | |
10141 | ||
c49439f1 R |
10142 | /* This function returns "2" to indicate dual issue for the SH4 |
10143 | processor. To be used by the DFA pipeline description. */ | |
fae15c93 | 10144 | static int |
832a3292 | 10145 | sh_issue_rate (void) |
fae15c93 | 10146 | { |
c49439f1 R |
10147 | if (TARGET_SUPERSCALAR) |
10148 | return 2; | |
10149 | else | |
10150 | return 1; | |
fae15c93 VM |
10151 | } |
10152 | ||
58565a33 SKG |
10153 | /* Functions for ready queue reordering for sched1. */ |
10154 | ||
10155 | /* Get weight for mode for a set x. */ | |
10156 | static short | |
508ea1c5 | 10157 | find_set_regmode_weight (rtx x, enum machine_mode mode) |
58565a33 SKG |
10158 | { |
10159 | if (GET_CODE (x) == CLOBBER && register_operand (SET_DEST (x), mode)) | |
10160 | return 1; | |
10161 | if (GET_CODE (x) == SET && register_operand (SET_DEST (x), mode)) | |
10162 | { | |
f3536097 | 10163 | if (REG_P (SET_DEST (x))) |
58565a33 SKG |
10164 | { |
10165 | if (!reg_mentioned_p (SET_DEST (x), SET_SRC (x))) | |
10166 | return 1; | |
10167 | else | |
10168 | return 0; | |
10169 | } | |
10170 | return 1; | |
10171 | } | |
10172 | return 0; | |
10173 | } | |
10174 | ||
10175 | /* Get regmode weight for insn. */ | |
10176 | static short | |
508ea1c5 | 10177 | find_insn_regmode_weight (rtx insn, enum machine_mode mode) |
58565a33 SKG |
10178 | { |
10179 | short reg_weight = 0; | |
10180 | rtx x; | |
10181 | ||
10182 | /* Increment weight for each register born here. */ | |
10183 | x = PATTERN (insn); | |
10184 | reg_weight += find_set_regmode_weight (x, mode); | |
10185 | if (GET_CODE (x) == PARALLEL) | |
10186 | { | |
10187 | int j; | |
10188 | for (j = XVECLEN (x, 0) - 1; j >= 0; j--) | |
10189 | { | |
10190 | x = XVECEXP (PATTERN (insn), 0, j); | |
10191 | reg_weight += find_set_regmode_weight (x, mode); | |
10192 | } | |
10193 | } | |
10194 | /* Decrement weight for each register that dies here. */ | |
10195 | for (x = REG_NOTES (insn); x; x = XEXP (x, 1)) | |
10196 | { | |
10197 | if (REG_NOTE_KIND (x) == REG_DEAD || REG_NOTE_KIND (x) == REG_UNUSED) | |
10198 | { | |
10199 | rtx note = XEXP (x, 0); | |
f3536097 | 10200 | if (REG_P (note) && GET_MODE (note) == mode) |
58565a33 SKG |
10201 | reg_weight--; |
10202 | } | |
10203 | } | |
10204 | return reg_weight; | |
10205 | } | |
10206 | ||
10207 | /* Calculate regmode weights for all insns of a basic block. */ | |
10208 | static void | |
c6d56443 | 10209 | find_regmode_weight (basic_block b, enum machine_mode mode) |
58565a33 SKG |
10210 | { |
10211 | rtx insn, next_tail, head, tail; | |
10212 | ||
c6d56443 | 10213 | get_ebb_head_tail (b, b, &head, &tail); |
58565a33 SKG |
10214 | next_tail = NEXT_INSN (tail); |
10215 | ||
10216 | for (insn = head; insn != next_tail; insn = NEXT_INSN (insn)) | |
10217 | { | |
10218 | /* Handle register life information. */ | |
10219 | if (!INSN_P (insn)) | |
10220 | continue; | |
10221 | ||
10222 | if (mode == SFmode) | |
10223 | INSN_REGMODE_WEIGHT (insn, mode) = | |
10224 | find_insn_regmode_weight (insn, mode) + 2 * find_insn_regmode_weight (insn, DFmode); | |
10225 | else if (mode == SImode) | |
10226 | INSN_REGMODE_WEIGHT (insn, mode) = | |
10227 | find_insn_regmode_weight (insn, mode) + 2 * find_insn_regmode_weight (insn, DImode); | |
10228 | } | |
10229 | } | |
10230 | ||
10231 | /* Comparison function for ready queue sorting. */ | |
10232 | static int | |
508ea1c5 | 10233 | rank_for_reorder (const void *x, const void *y) |
58565a33 SKG |
10234 | { |
10235 | rtx tmp = *(const rtx *) y; | |
10236 | rtx tmp2 = *(const rtx *) x; | |
10237 | ||
10238 | /* The insn in a schedule group should be issued the first. */ | |
10239 | if (SCHED_GROUP_P (tmp) != SCHED_GROUP_P (tmp2)) | |
10240 | return SCHED_GROUP_P (tmp2) ? 1 : -1; | |
10241 | ||
f676971a | 10242 | /* If insns are equally good, sort by INSN_LUID (original insn order), This |
58565a33 SKG |
10243 | minimizes instruction movement, thus minimizing sched's effect on |
10244 | register pressure. */ | |
10245 | return INSN_LUID (tmp) - INSN_LUID (tmp2); | |
10246 | } | |
10247 | ||
10248 | /* Resort the array A in which only element at index N may be out of order. */ | |
10249 | static void | |
508ea1c5 | 10250 | swap_reorder (rtx *a, int n) |
58565a33 SKG |
10251 | { |
10252 | rtx insn = a[n - 1]; | |
10253 | int i = n - 2; | |
10254 | ||
10255 | while (i >= 0 && rank_for_reorder (a + i, &insn) >= 0) | |
10256 | { | |
10257 | a[i + 1] = a[i]; | |
10258 | i -= 1; | |
10259 | } | |
10260 | a[i + 1] = insn; | |
10261 | } | |
10262 | ||
832a3292 KH |
10263 | #define SCHED_REORDER(READY, N_READY) \ |
10264 | do \ | |
10265 | { \ | |
10266 | if ((N_READY) == 2) \ | |
10267 | swap_reorder (READY, N_READY); \ | |
10268 | else if ((N_READY) > 2) \ | |
10269 | qsort (READY, N_READY, sizeof (rtx), rank_for_reorder); \ | |
10270 | } \ | |
10271 | while (0) | |
58565a33 SKG |
10272 | |
10273 | /* Sort the ready list READY by ascending priority, using the SCHED_REORDER | |
10274 | macro. */ | |
10275 | static void | |
508ea1c5 | 10276 | ready_reorder (rtx *ready, int nready) |
58565a33 SKG |
10277 | { |
10278 | SCHED_REORDER (ready, nready); | |
10279 | } | |
10280 | ||
6fb5fa3c DB |
10281 | /* Count life regions of r0 for a block. */ |
10282 | static int | |
10283 | find_r0_life_regions (basic_block b) | |
10284 | { | |
10285 | rtx end, insn; | |
10286 | rtx pset; | |
10287 | rtx r0_reg; | |
10288 | int live; | |
10289 | int set; | |
10290 | int death = 0; | |
10291 | ||
89a95777 | 10292 | if (REGNO_REG_SET_P (df_get_live_in (b), R0_REG)) |
6fb5fa3c DB |
10293 | { |
10294 | set = 1; | |
10295 | live = 1; | |
10296 | } | |
10297 | else | |
10298 | { | |
10299 | set = 0; | |
10300 | live = 0; | |
10301 | } | |
10302 | ||
10303 | insn = BB_HEAD (b); | |
10304 | end = BB_END (b); | |
10305 | r0_reg = gen_rtx_REG (SImode, R0_REG); | |
10306 | while (1) | |
10307 | { | |
10308 | if (INSN_P (insn)) | |
10309 | { | |
10310 | if (find_regno_note (insn, REG_DEAD, R0_REG)) | |
10311 | { | |
10312 | death++; | |
10313 | live = 0; | |
10314 | } | |
10315 | if (!live | |
10316 | && (pset = single_set (insn)) | |
10317 | && reg_overlap_mentioned_p (r0_reg, SET_DEST (pset)) | |
10318 | && !find_regno_note (insn, REG_UNUSED, R0_REG)) | |
10319 | { | |
10320 | set++; | |
10321 | live = 1; | |
10322 | } | |
10323 | } | |
10324 | if (insn == end) | |
10325 | break; | |
10326 | insn = NEXT_INSN (insn); | |
10327 | } | |
10328 | return set - death; | |
10329 | } | |
10330 | ||
58565a33 SKG |
10331 | /* Calculate regmode weights for all insns of all basic block. */ |
10332 | static void | |
508ea1c5 KH |
10333 | sh_md_init_global (FILE *dump ATTRIBUTE_UNUSED, |
10334 | int verbose ATTRIBUTE_UNUSED, | |
10335 | int old_max_uid) | |
58565a33 SKG |
10336 | { |
10337 | basic_block b; | |
10338 | ||
10339 | regmode_weight[0] = (short *) xcalloc (old_max_uid, sizeof (short)); | |
10340 | regmode_weight[1] = (short *) xcalloc (old_max_uid, sizeof (short)); | |
6fb5fa3c | 10341 | r0_life_regions = 0; |
58565a33 SKG |
10342 | |
10343 | FOR_EACH_BB_REVERSE (b) | |
10344 | { | |
c6d56443 KK |
10345 | find_regmode_weight (b, SImode); |
10346 | find_regmode_weight (b, SFmode); | |
6fb5fa3c DB |
10347 | if (!reload_completed) |
10348 | r0_life_regions += find_r0_life_regions (b); | |
58565a33 SKG |
10349 | } |
10350 | ||
10351 | CURR_REGMODE_PRESSURE (SImode) = 0; | |
10352 | CURR_REGMODE_PRESSURE (SFmode) = 0; | |
10353 | ||
10354 | } | |
10355 | ||
10356 | /* Cleanup. */ | |
10357 | static void | |
508ea1c5 KH |
10358 | sh_md_finish_global (FILE *dump ATTRIBUTE_UNUSED, |
10359 | int verbose ATTRIBUTE_UNUSED) | |
58565a33 SKG |
10360 | { |
10361 | if (regmode_weight[0]) | |
10362 | { | |
10363 | free (regmode_weight[0]); | |
10364 | regmode_weight[0] = NULL; | |
10365 | } | |
10366 | if (regmode_weight[1]) | |
10367 | { | |
10368 | free (regmode_weight[1]); | |
10369 | regmode_weight[1] = NULL; | |
10370 | } | |
10371 | } | |
10372 | ||
b0fbea13 KK |
10373 | /* The scalar modes supported differs from the default version in TImode |
10374 | for 32-bit SHMEDIA. */ | |
10375 | static bool | |
10376 | sh_scalar_mode_supported_p (enum machine_mode mode) | |
10377 | { | |
10378 | if (TARGET_SHMEDIA32 && mode == TImode) | |
10379 | return false; | |
10380 | ||
10381 | return default_scalar_mode_supported_p (mode); | |
10382 | } | |
10383 | ||
58565a33 SKG |
10384 | /* Cache the can_issue_more so that we can return it from reorder2. Also, |
10385 | keep count of register pressures on SImode and SFmode. */ | |
10386 | static int | |
508ea1c5 KH |
10387 | sh_variable_issue (FILE *dump ATTRIBUTE_UNUSED, |
10388 | int sched_verbose ATTRIBUTE_UNUSED, | |
10389 | rtx insn, | |
10390 | int can_issue_more) | |
58565a33 SKG |
10391 | { |
10392 | if (GET_CODE (PATTERN (insn)) != USE | |
10393 | && GET_CODE (PATTERN (insn)) != CLOBBER) | |
10394 | cached_can_issue_more = can_issue_more - 1; | |
10395 | else | |
10396 | cached_can_issue_more = can_issue_more; | |
10397 | ||
10398 | if (reload_completed) | |
10399 | return cached_can_issue_more; | |
10400 | ||
10401 | CURR_REGMODE_PRESSURE (SImode) += INSN_REGMODE_WEIGHT (insn, SImode); | |
10402 | CURR_REGMODE_PRESSURE (SFmode) += INSN_REGMODE_WEIGHT (insn, SFmode); | |
10403 | ||
10404 | return cached_can_issue_more; | |
10405 | } | |
10406 | ||
10407 | static void | |
508ea1c5 KH |
10408 | sh_md_init (FILE *dump ATTRIBUTE_UNUSED, |
10409 | int verbose ATTRIBUTE_UNUSED, | |
10410 | int veclen ATTRIBUTE_UNUSED) | |
58565a33 SKG |
10411 | { |
10412 | CURR_REGMODE_PRESSURE (SImode) = 0; | |
10413 | CURR_REGMODE_PRESSURE (SFmode) = 0; | |
10414 | } | |
10415 | ||
10416 | /* Some magic numbers. */ | |
10417 | /* Pressure on register r0 can lead to spill failures. so avoid sched1 for | |
10418 | functions that already have high pressure on r0. */ | |
10419 | #define R0_MAX_LIFE_REGIONS 2 | |
4ee31f1e | 10420 | /* Register Pressure thresholds for SImode and SFmode registers. */ |
58565a33 SKG |
10421 | #define SIMODE_MAX_WEIGHT 5 |
10422 | #define SFMODE_MAX_WEIGHT 10 | |
10423 | ||
10424 | /* Return true if the pressure is high for MODE. */ | |
10425 | static short | |
508ea1c5 | 10426 | high_pressure (enum machine_mode mode) |
58565a33 SKG |
10427 | { |
10428 | /* Pressure on register r0 can lead to spill failures. so avoid sched1 for | |
10429 | functions that already have high pressure on r0. */ | |
6fb5fa3c DB |
10430 | if (r0_life_regions >= R0_MAX_LIFE_REGIONS) |
10431 | return 1; | |
58565a33 SKG |
10432 | |
10433 | if (mode == SFmode) | |
10434 | return (CURR_REGMODE_PRESSURE (SFmode) > SFMODE_MAX_WEIGHT); | |
10435 | else | |
10436 | return (CURR_REGMODE_PRESSURE (SImode) > SIMODE_MAX_WEIGHT); | |
10437 | } | |
10438 | ||
10439 | /* Reorder ready queue if register pressure is high. */ | |
10440 | static int | |
508ea1c5 KH |
10441 | sh_reorder (FILE *dump ATTRIBUTE_UNUSED, |
10442 | int sched_verbose ATTRIBUTE_UNUSED, | |
10443 | rtx *ready, | |
10444 | int *n_readyp, | |
10445 | int clock_var ATTRIBUTE_UNUSED) | |
58565a33 SKG |
10446 | { |
10447 | if (reload_completed) | |
10448 | return sh_issue_rate (); | |
10449 | ||
10450 | if (high_pressure (SFmode) || high_pressure (SImode)) | |
10451 | { | |
10452 | ready_reorder (ready, *n_readyp); | |
10453 | } | |
10454 | ||
10455 | return sh_issue_rate (); | |
10456 | } | |
10457 | ||
10458 | /* Skip cycles if the current register pressure is high. */ | |
f676971a | 10459 | static int |
508ea1c5 KH |
10460 | sh_reorder2 (FILE *dump ATTRIBUTE_UNUSED, |
10461 | int sched_verbose ATTRIBUTE_UNUSED, | |
10462 | rtx *ready ATTRIBUTE_UNUSED, | |
10463 | int *n_readyp ATTRIBUTE_UNUSED, | |
10464 | int clock_var ATTRIBUTE_UNUSED) | |
58565a33 SKG |
10465 | { |
10466 | if (reload_completed) | |
10467 | return cached_can_issue_more; | |
10468 | ||
f676971a | 10469 | if (high_pressure(SFmode) || high_pressure (SImode)) |
58565a33 SKG |
10470 | skip_cycles = 1; |
10471 | ||
10472 | return cached_can_issue_more; | |
10473 | } | |
10474 | ||
10475 | /* Skip cycles without sorting the ready queue. This will move insn from | |
10476 | Q->R. If this is the last cycle we are skipping; allow sorting of ready | |
f676971a | 10477 | queue by sh_reorder. */ |
58565a33 | 10478 | |
f676971a EC |
10479 | /* Generally, skipping these many cycles are sufficient for all insns to move |
10480 | from Q -> R. */ | |
10481 | #define MAX_SKIPS 8 | |
58565a33 SKG |
10482 | |
10483 | static int | |
508ea1c5 KH |
10484 | sh_dfa_new_cycle (FILE *sched_dump ATTRIBUTE_UNUSED, |
10485 | int sched_verbose ATTRIBUTE_UNUSED, | |
10486 | rtx insn ATTRIBUTE_UNUSED, | |
10487 | int last_clock_var, | |
10488 | int clock_var, | |
10489 | int *sort_p) | |
58565a33 SKG |
10490 | { |
10491 | if (reload_completed) | |
832a3292 | 10492 | return 0; |
58565a33 | 10493 | |
f676971a | 10494 | if (skip_cycles) |
58565a33 | 10495 | { |
832a3292 KH |
10496 | if ((clock_var - last_clock_var) < MAX_SKIPS) |
10497 | { | |
10498 | *sort_p = 0; | |
10499 | return 1; | |
10500 | } | |
10501 | /* If this is the last cycle we are skipping, allow reordering of R. */ | |
10502 | if ((clock_var - last_clock_var) == MAX_SKIPS) | |
10503 | { | |
10504 | *sort_p = 1; | |
10505 | return 1; | |
10506 | } | |
58565a33 | 10507 | } |
58565a33 SKG |
10508 | |
10509 | skip_cycles = 0; | |
10510 | ||
10511 | return 0; | |
10512 | } | |
10513 | ||
fa5322fa AO |
10514 | /* SHmedia requires registers for branches, so we can't generate new |
10515 | branches past reload. */ | |
10516 | static bool | |
cf277499 | 10517 | sh_cannot_modify_jumps_p (void) |
fa5322fa AO |
10518 | { |
10519 | return (TARGET_SHMEDIA && (reload_in_progress || reload_completed)); | |
10520 | } | |
10521 | ||
a87cf97e | 10522 | static reg_class_t |
fe3ad572 SC |
10523 | sh_target_reg_class (void) |
10524 | { | |
10525 | return TARGET_SHMEDIA ? TARGET_REGS : NO_REGS; | |
10526 | } | |
10527 | ||
10528 | static bool | |
10529 | sh_optimize_target_register_callee_saved (bool after_prologue_epilogue_gen) | |
10530 | { | |
73a4d10b | 10531 | HARD_REG_SET dummy; |
cb3d3842 | 10532 | #if 0 |
73a4d10b | 10533 | rtx insn; |
cb3d3842 | 10534 | #endif |
73a4d10b R |
10535 | |
10536 | if (! shmedia_space_reserved_for_target_registers) | |
10537 | return 0; | |
10538 | if (after_prologue_epilogue_gen && ! TARGET_SAVE_ALL_TARGET_REGS) | |
10539 | return 0; | |
10540 | if (calc_live_regs (&dummy) >= 6 * 8) | |
10541 | return 1; | |
73a4d10b | 10542 | return 0; |
fe3ad572 SC |
10543 | } |
10544 | ||
fa5322fa | 10545 | static bool |
3101faab | 10546 | sh_ms_bitfield_layout_p (const_tree record_type ATTRIBUTE_UNUSED) |
fa5322fa | 10547 | { |
61f71b34 | 10548 | return (TARGET_SH5 || TARGET_HITACHI || sh_attr_renesas_p (record_type)); |
fa5322fa | 10549 | } |
ca903bba | 10550 | \f |
f676971a | 10551 | /* |
ca903bba R |
10552 | On the SH1..SH4, the trampoline looks like |
10553 | 2 0002 D202 mov.l l2,r2 | |
10554 | 1 0000 D301 mov.l l1,r3 | |
10555 | 3 0004 422B jmp @r2 | |
10556 | 4 0006 0009 nop | |
10557 | 5 0008 00000000 l1: .long area | |
10558 | 6 000c 00000000 l2: .long function | |
10559 | ||
10560 | SH5 (compact) uses r1 instead of r3 for the static chain. */ | |
10561 | ||
10562 | ||
10563 | /* Emit RTL insns to initialize the variable parts of a trampoline. | |
10564 | FNADDR is an RTX for the address of the function's pure code. | |
10565 | CXT is an RTX for the static chain value for the function. */ | |
10566 | ||
994322ef RH |
10567 | static void |
10568 | sh_trampoline_init (rtx tramp_mem, tree fndecl, rtx cxt) | |
ca903bba | 10569 | { |
994322ef RH |
10570 | rtx fnaddr = XEXP (DECL_RTL (fndecl), 0); |
10571 | rtx tramp = force_reg (Pmode, XEXP (tramp_mem, 0)); | |
57d38024 | 10572 | |
ca903bba R |
10573 | if (TARGET_SHMEDIA64) |
10574 | { | |
10575 | rtx tramp_templ; | |
10576 | int fixed_len; | |
10577 | ||
10578 | rtx movi1 = GEN_INT (0xcc000010); | |
10579 | rtx shori1 = GEN_INT (0xc8000010); | |
10580 | rtx src, dst; | |
10581 | ||
10582 | /* The following trampoline works within a +- 128 KB range for cxt: | |
10583 | ptb/u cxt,tr1; movi fnaddr >> 48,r0; shori fnaddr >> 32,r0; | |
10584 | shori fnaddr >> 16,r0; shori fnaddr,r0; ptabs/l r0,tr0 | |
10585 | gettr tr1,r1; blink tr0,r63 */ | |
10586 | /* Address rounding makes it hard to compute the exact bounds of the | |
10587 | offset for this trampoline, but we have a rather generous offset | |
10588 | range, so frame_offset should do fine as an upper bound. */ | |
10589 | if (cxt == virtual_stack_vars_rtx && frame_offset < 0x20000) | |
10590 | { | |
10591 | /* ??? could optimize this trampoline initialization | |
10592 | by writing DImode words with two insns each. */ | |
10593 | rtx mask = force_reg (DImode, GEN_INT (0x3fffc00)); | |
10594 | rtx insn = gen_rtx_MINUS (DImode, cxt, tramp); | |
10595 | insn = gen_rtx_ASHIFT (DImode, insn, GEN_INT (10-2)); | |
10596 | insn = gen_rtx_AND (DImode, insn, mask); | |
10597 | /* Or in ptb/u .,tr1 pattern */ | |
10598 | insn = gen_rtx_IOR (DImode, insn, gen_int_mode (0xec000010, SImode)); | |
10599 | insn = force_operand (insn, NULL_RTX); | |
10600 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 10601 | emit_move_insn (change_address (tramp_mem, SImode, NULL_RTX), insn); |
ca903bba R |
10602 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (38)); |
10603 | insn = gen_rtx_AND (DImode, insn, mask); | |
10604 | insn = force_operand (gen_rtx_IOR (DImode, movi1, insn), NULL_RTX); | |
10605 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 10606 | emit_move_insn (adjust_address (tramp_mem, SImode, 4), insn); |
ca903bba R |
10607 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (22)); |
10608 | insn = gen_rtx_AND (DImode, insn, mask); | |
10609 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
10610 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 10611 | emit_move_insn (adjust_address (tramp_mem, SImode, 8), insn); |
ca903bba R |
10612 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (6)); |
10613 | insn = gen_rtx_AND (DImode, insn, mask); | |
10614 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
10615 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 10616 | emit_move_insn (adjust_address (tramp_mem, SImode, 12), insn); |
ca903bba R |
10617 | insn = gen_rtx_ASHIFT (DImode, fnaddr, GEN_INT (10)); |
10618 | insn = gen_rtx_AND (DImode, insn, mask); | |
10619 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
10620 | insn = gen_lowpart (SImode, insn); | |
57d38024 R |
10621 | emit_move_insn (adjust_address (tramp_mem, SImode, 16), insn); |
10622 | emit_move_insn (adjust_address (tramp_mem, SImode, 20), | |
ca903bba | 10623 | GEN_INT (0x6bf10600)); |
57d38024 | 10624 | emit_move_insn (adjust_address (tramp_mem, SImode, 24), |
ca903bba | 10625 | GEN_INT (0x4415fc10)); |
57d38024 | 10626 | emit_move_insn (adjust_address (tramp_mem, SImode, 28), |
ca903bba R |
10627 | GEN_INT (0x4401fff0)); |
10628 | emit_insn (gen_ic_invalidate_line (tramp)); | |
10629 | return; | |
10630 | } | |
10631 | tramp_templ = gen_rtx_SYMBOL_REF (Pmode,"__GCC_nested_trampoline"); | |
10632 | fixed_len = TRAMPOLINE_SIZE - 2 * GET_MODE_SIZE (Pmode); | |
10633 | ||
10634 | tramp_templ = gen_datalabel_ref (tramp_templ); | |
57d38024 R |
10635 | dst = tramp_mem; |
10636 | src = gen_const_mem (BLKmode, tramp_templ); | |
ca903bba R |
10637 | set_mem_align (dst, 256); |
10638 | set_mem_align (src, 64); | |
44bb111a | 10639 | emit_block_move (dst, src, GEN_INT (fixed_len), BLOCK_OP_NORMAL); |
ca903bba | 10640 | |
57d38024 R |
10641 | emit_move_insn (adjust_address (tramp_mem, Pmode, fixed_len), fnaddr); |
10642 | emit_move_insn (adjust_address (tramp_mem, Pmode, | |
10643 | fixed_len + GET_MODE_SIZE (Pmode)), | |
ca903bba R |
10644 | cxt); |
10645 | emit_insn (gen_ic_invalidate_line (tramp)); | |
10646 | return; | |
10647 | } | |
10648 | else if (TARGET_SHMEDIA) | |
10649 | { | |
10650 | /* movi fnaddr >> 16,r1; shori fnaddr,r1; ptabs/l r1,tr0 | |
10651 | movi cxt >> 16,r1; shori cxt,r1; blink tr0,r63 */ | |
10652 | rtx quad0 = gen_reg_rtx (DImode), cxtload = gen_reg_rtx (DImode); | |
10653 | rtx quad1 = gen_reg_rtx (DImode), quad2 = gen_reg_rtx (DImode); | |
10654 | /* movi 0,r1: 0xcc000010 shori 0,r1: c8000010 concatenated, | |
825db093 | 10655 | rotated 10 right, and higher 16 bit of every 32 selected. */ |
ca903bba R |
10656 | rtx movishori |
10657 | = force_reg (V2HImode, (simplify_gen_subreg | |
10658 | (V2HImode, GEN_INT (0x4330432), SImode, 0))); | |
10659 | rtx ptabs = force_reg (DImode, GEN_INT (0x6bf10600)); | |
10660 | rtx blink = force_reg (DImode, GEN_INT (0x4401fff0)); | |
10661 | ||
ca903bba R |
10662 | fnaddr = force_reg (SImode, fnaddr); |
10663 | cxt = force_reg (SImode, cxt); | |
10664 | emit_insn (gen_mshflo_w_x (gen_rtx_SUBREG (V4HImode, quad0, 0), | |
10665 | gen_rtx_SUBREG (V2HImode, fnaddr, 0), | |
10666 | movishori)); | |
f1e2239f | 10667 | emit_insn (gen_rotrdi3_mextr (quad0, quad0, |
ca903bba | 10668 | GEN_INT (TARGET_LITTLE_ENDIAN ? 24 : 56))); |
a556fd39 | 10669 | emit_insn (gen_ashldi3_media (quad0, quad0, const2_rtx)); |
57d38024 | 10670 | emit_move_insn (change_address (tramp_mem, DImode, NULL_RTX), quad0); |
ca903bba R |
10671 | emit_insn (gen_mshflo_w_x (gen_rtx_SUBREG (V4HImode, cxtload, 0), |
10672 | gen_rtx_SUBREG (V2HImode, cxt, 0), | |
10673 | movishori)); | |
f1e2239f | 10674 | emit_insn (gen_rotrdi3_mextr (cxtload, cxtload, |
ca903bba | 10675 | GEN_INT (TARGET_LITTLE_ENDIAN ? 24 : 56))); |
a556fd39 | 10676 | emit_insn (gen_ashldi3_media (cxtload, cxtload, const2_rtx)); |
ca903bba R |
10677 | if (TARGET_LITTLE_ENDIAN) |
10678 | { | |
10679 | emit_insn (gen_mshflo_l_di (quad1, ptabs, cxtload)); | |
10680 | emit_insn (gen_mextr4 (quad2, cxtload, blink)); | |
10681 | } | |
10682 | else | |
10683 | { | |
10684 | emit_insn (gen_mextr4 (quad1, cxtload, ptabs)); | |
10685 | emit_insn (gen_mshflo_l_di (quad2, blink, cxtload)); | |
10686 | } | |
57d38024 R |
10687 | emit_move_insn (adjust_address (tramp_mem, DImode, 8), quad1); |
10688 | emit_move_insn (adjust_address (tramp_mem, DImode, 16), quad2); | |
ca903bba R |
10689 | emit_insn (gen_ic_invalidate_line (tramp)); |
10690 | return; | |
10691 | } | |
10692 | else if (TARGET_SHCOMPACT) | |
10693 | { | |
10694 | emit_insn (gen_initialize_trampoline (tramp, cxt, fnaddr)); | |
10695 | return; | |
10696 | } | |
57d38024 | 10697 | emit_move_insn (change_address (tramp_mem, SImode, NULL_RTX), |
ca903bba R |
10698 | gen_int_mode (TARGET_LITTLE_ENDIAN ? 0xd301d202 : 0xd202d301, |
10699 | SImode)); | |
57d38024 | 10700 | emit_move_insn (adjust_address (tramp_mem, SImode, 4), |
ca903bba R |
10701 | gen_int_mode (TARGET_LITTLE_ENDIAN ? 0x0009422b : 0x422b0009, |
10702 | SImode)); | |
57d38024 R |
10703 | emit_move_insn (adjust_address (tramp_mem, SImode, 8), cxt); |
10704 | emit_move_insn (adjust_address (tramp_mem, SImode, 12), fnaddr); | |
ca903bba R |
10705 | if (TARGET_HARVARD) |
10706 | { | |
dc557046 | 10707 | if (!TARGET_INLINE_IC_INVALIDATE |
6f8eec29 | 10708 | || (!(TARGET_SH4A_ARCH || TARGET_SH4_300) && TARGET_USERMODE)) |
73a4d10b R |
10709 | emit_library_call (function_symbol (NULL, "__ic_invalidate", |
10710 | FUNCTION_ORDINARY), | |
b757d36f | 10711 | LCT_NORMAL, VOIDmode, 1, tramp, SImode); |
ca903bba R |
10712 | else |
10713 | emit_insn (gen_ic_invalidate_line (tramp)); | |
10714 | } | |
10715 | } | |
10716 | ||
994322ef RH |
10717 | /* On SH5, trampolines are SHmedia code, so add 1 to the address. */ |
10718 | ||
10719 | static rtx | |
10720 | sh_trampoline_adjust_address (rtx tramp) | |
10721 | { | |
10722 | if (TARGET_SHMEDIA) | |
10723 | tramp = expand_simple_binop (Pmode, PLUS, tramp, const1_rtx, | |
10724 | gen_reg_rtx (Pmode), 0, OPTAB_LIB_WIDEN); | |
10725 | return tramp; | |
10726 | } | |
10727 | ||
4977bab6 ZW |
10728 | /* FIXME: This is overly conservative. A SHcompact function that |
10729 | receives arguments ``by reference'' will have them stored in its | |
10730 | own stack frame, so it must not pass pointers or references to | |
10731 | these arguments to other functions by means of sibling calls. */ | |
73a4d10b R |
10732 | /* If PIC, we cannot make sibling calls to global functions |
10733 | because the PLT requires r12 to be live. */ | |
4977bab6 | 10734 | static bool |
cf277499 | 10735 | sh_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED) |
4977bab6 | 10736 | { |
73a4d10b | 10737 | return (1 |
4977bab6 | 10738 | && (! TARGET_SHCOMPACT |
38173d38 | 10739 | || crtl->args.info.stack_regs == 0) |
73a4d10b R |
10740 | && ! sh_cfun_interrupt_handler_p () |
10741 | && (! flag_pic | |
10742 | || (decl && ! TREE_PUBLIC (decl)) | |
10743 | || (decl && DECL_VISIBILITY (decl) != VISIBILITY_DEFAULT))); | |
4977bab6 | 10744 | } |
c1b92d09 R |
10745 | \f |
10746 | /* Machine specific built-in functions. */ | |
10747 | ||
10748 | struct builtin_description | |
10749 | { | |
10750 | const enum insn_code icode; | |
10751 | const char *const name; | |
10752 | int signature; | |
1cc24576 | 10753 | tree fndecl; |
c1b92d09 R |
10754 | }; |
10755 | ||
10756 | /* describe number and signedness of arguments; arg[0] == result | |
10757 | (1: unsigned, 2: signed, 4: don't care, 8: pointer 0: no argument */ | |
3b962ea0 | 10758 | /* 9: 64-bit pointer, 10: 32-bit pointer */ |
c1b92d09 R |
10759 | static const char signature_args[][4] = |
10760 | { | |
10761 | #define SH_BLTIN_V2SI2 0 | |
10762 | { 4, 4 }, | |
10763 | #define SH_BLTIN_V4HI2 1 | |
10764 | { 4, 4 }, | |
10765 | #define SH_BLTIN_V2SI3 2 | |
10766 | { 4, 4, 4 }, | |
10767 | #define SH_BLTIN_V4HI3 3 | |
10768 | { 4, 4, 4 }, | |
10769 | #define SH_BLTIN_V8QI3 4 | |
10770 | { 4, 4, 4 }, | |
10771 | #define SH_BLTIN_MAC_HISI 5 | |
10772 | { 1, 4, 4, 1 }, | |
10773 | #define SH_BLTIN_SH_HI 6 | |
10774 | { 4, 4, 1 }, | |
10775 | #define SH_BLTIN_SH_SI 7 | |
10776 | { 4, 4, 1 }, | |
10777 | #define SH_BLTIN_V4HI2V2SI 8 | |
10778 | { 4, 4, 4 }, | |
10779 | #define SH_BLTIN_V4HI2V8QI 9 | |
10780 | { 4, 4, 4 }, | |
10781 | #define SH_BLTIN_SISF 10 | |
10782 | { 4, 2 }, | |
10783 | #define SH_BLTIN_LDUA_L 11 | |
73a4d10b | 10784 | { 2, 10 }, |
c1b92d09 | 10785 | #define SH_BLTIN_LDUA_Q 12 |
73a4d10b | 10786 | { 1, 10 }, |
c1b92d09 | 10787 | #define SH_BLTIN_STUA_L 13 |
73a4d10b | 10788 | { 0, 10, 2 }, |
c1b92d09 | 10789 | #define SH_BLTIN_STUA_Q 14 |
73a4d10b R |
10790 | { 0, 10, 1 }, |
10791 | #define SH_BLTIN_LDUA_L64 15 | |
10792 | { 2, 9 }, | |
10793 | #define SH_BLTIN_LDUA_Q64 16 | |
10794 | { 1, 9 }, | |
10795 | #define SH_BLTIN_STUA_L64 17 | |
10796 | { 0, 9, 2 }, | |
10797 | #define SH_BLTIN_STUA_Q64 18 | |
10798 | { 0, 9, 1 }, | |
10799 | #define SH_BLTIN_NUM_SHARED_SIGNATURES 19 | |
10800 | #define SH_BLTIN_2 19 | |
10801 | #define SH_BLTIN_SU 19 | |
c1b92d09 | 10802 | { 1, 2 }, |
73a4d10b R |
10803 | #define SH_BLTIN_3 20 |
10804 | #define SH_BLTIN_SUS 20 | |
c1b92d09 | 10805 | { 2, 2, 1 }, |
73a4d10b | 10806 | #define SH_BLTIN_PSSV 21 |
c1b92d09 | 10807 | { 0, 8, 2, 2 }, |
73a4d10b R |
10808 | #define SH_BLTIN_XXUU 22 |
10809 | #define SH_BLTIN_UUUU 22 | |
c1b92d09 | 10810 | { 1, 1, 1, 1 }, |
73a4d10b | 10811 | #define SH_BLTIN_PV 23 |
c1b92d09 R |
10812 | { 0, 8 }, |
10813 | }; | |
24746a42 | 10814 | /* mcmv: operands considered unsigned. */ |
c1b92d09 | 10815 | /* mmulsum_wq, msad_ubq: result considered unsigned long long. */ |
24746a42 KH |
10816 | /* mperm: control value considered unsigned int. */ |
10817 | /* mshalds, mshard, mshards, mshlld, mshlrd: shift count is unsigned int. */ | |
c1b92d09 R |
10818 | /* mshards_q: returns signed short. */ |
10819 | /* nsb: takes long long arg, returns unsigned char. */ | |
1cc24576 KK |
10820 | static struct builtin_description bdesc[] = |
10821 | { | |
10822 | { CODE_FOR_absv2si2, "__builtin_absv2si2", SH_BLTIN_V2SI2, 0 }, | |
10823 | { CODE_FOR_absv4hi2, "__builtin_absv4hi2", SH_BLTIN_V4HI2, 0 }, | |
10824 | { CODE_FOR_addv2si3, "__builtin_addv2si3", SH_BLTIN_V2SI3, 0 }, | |
10825 | { CODE_FOR_addv4hi3, "__builtin_addv4hi3", SH_BLTIN_V4HI3, 0 }, | |
10826 | { CODE_FOR_ssaddv2si3,"__builtin_ssaddv2si3", SH_BLTIN_V2SI3, 0 }, | |
10827 | { CODE_FOR_usaddv8qi3,"__builtin_usaddv8qi3", SH_BLTIN_V8QI3, 0 }, | |
10828 | { CODE_FOR_ssaddv4hi3,"__builtin_ssaddv4hi3", SH_BLTIN_V4HI3, 0 }, | |
10829 | { CODE_FOR_alloco_i, "__builtin_sh_media_ALLOCO", SH_BLTIN_PV, 0 }, | |
10830 | { CODE_FOR_negcmpeqv8qi,"__builtin_sh_media_MCMPEQ_B", SH_BLTIN_V8QI3, 0 }, | |
10831 | { CODE_FOR_negcmpeqv2si,"__builtin_sh_media_MCMPEQ_L", SH_BLTIN_V2SI3, 0 }, | |
10832 | { CODE_FOR_negcmpeqv4hi,"__builtin_sh_media_MCMPEQ_W", SH_BLTIN_V4HI3, 0 }, | |
10833 | { CODE_FOR_negcmpgtuv8qi,"__builtin_sh_media_MCMPGT_UB", SH_BLTIN_V8QI3, 0 }, | |
10834 | { CODE_FOR_negcmpgtv2si,"__builtin_sh_media_MCMPGT_L", SH_BLTIN_V2SI3, 0 }, | |
10835 | { CODE_FOR_negcmpgtv4hi,"__builtin_sh_media_MCMPGT_W", SH_BLTIN_V4HI3, 0 }, | |
10836 | { CODE_FOR_mcmv, "__builtin_sh_media_MCMV", SH_BLTIN_UUUU, 0 }, | |
10837 | { CODE_FOR_mcnvs_lw, "__builtin_sh_media_MCNVS_LW", SH_BLTIN_3, 0 }, | |
10838 | { CODE_FOR_mcnvs_wb, "__builtin_sh_media_MCNVS_WB", SH_BLTIN_V4HI2V8QI, 0 }, | |
10839 | { CODE_FOR_mcnvs_wub, "__builtin_sh_media_MCNVS_WUB", SH_BLTIN_V4HI2V8QI, 0 }, | |
10840 | { CODE_FOR_mextr1, "__builtin_sh_media_MEXTR1", SH_BLTIN_V8QI3, 0 }, | |
10841 | { CODE_FOR_mextr2, "__builtin_sh_media_MEXTR2", SH_BLTIN_V8QI3, 0 }, | |
10842 | { CODE_FOR_mextr3, "__builtin_sh_media_MEXTR3", SH_BLTIN_V8QI3, 0 }, | |
10843 | { CODE_FOR_mextr4, "__builtin_sh_media_MEXTR4", SH_BLTIN_V8QI3, 0 }, | |
10844 | { CODE_FOR_mextr5, "__builtin_sh_media_MEXTR5", SH_BLTIN_V8QI3, 0 }, | |
10845 | { CODE_FOR_mextr6, "__builtin_sh_media_MEXTR6", SH_BLTIN_V8QI3, 0 }, | |
10846 | { CODE_FOR_mextr7, "__builtin_sh_media_MEXTR7", SH_BLTIN_V8QI3, 0 }, | |
10847 | { CODE_FOR_mmacfx_wl, "__builtin_sh_media_MMACFX_WL", SH_BLTIN_MAC_HISI, 0 }, | |
10848 | { CODE_FOR_mmacnfx_wl,"__builtin_sh_media_MMACNFX_WL", SH_BLTIN_MAC_HISI, 0 }, | |
10849 | { CODE_FOR_mulv2si3, "__builtin_mulv2si3", SH_BLTIN_V2SI3, 0 }, | |
10850 | { CODE_FOR_mulv4hi3, "__builtin_mulv4hi3", SH_BLTIN_V4HI3, 0 }, | |
10851 | { CODE_FOR_mmulfx_l, "__builtin_sh_media_MMULFX_L", SH_BLTIN_V2SI3, 0 }, | |
10852 | { CODE_FOR_mmulfx_w, "__builtin_sh_media_MMULFX_W", SH_BLTIN_V4HI3, 0 }, | |
10853 | { CODE_FOR_mmulfxrp_w,"__builtin_sh_media_MMULFXRP_W", SH_BLTIN_V4HI3, 0 }, | |
10854 | { CODE_FOR_mmulhi_wl, "__builtin_sh_media_MMULHI_WL", SH_BLTIN_V4HI2V2SI, 0 }, | |
10855 | { CODE_FOR_mmullo_wl, "__builtin_sh_media_MMULLO_WL", SH_BLTIN_V4HI2V2SI, 0 }, | |
10856 | { CODE_FOR_mmulsum_wq,"__builtin_sh_media_MMULSUM_WQ", SH_BLTIN_XXUU, 0 }, | |
10857 | { CODE_FOR_mperm_w, "__builtin_sh_media_MPERM_W", SH_BLTIN_SH_HI, 0 }, | |
10858 | { CODE_FOR_msad_ubq, "__builtin_sh_media_MSAD_UBQ", SH_BLTIN_XXUU, 0 }, | |
10859 | { CODE_FOR_mshalds_l, "__builtin_sh_media_MSHALDS_L", SH_BLTIN_SH_SI, 0 }, | |
10860 | { CODE_FOR_mshalds_w, "__builtin_sh_media_MSHALDS_W", SH_BLTIN_SH_HI, 0 }, | |
10861 | { CODE_FOR_ashrv2si3, "__builtin_ashrv2si3", SH_BLTIN_SH_SI, 0 }, | |
10862 | { CODE_FOR_ashrv4hi3, "__builtin_ashrv4hi3", SH_BLTIN_SH_HI, 0 }, | |
10863 | { CODE_FOR_mshards_q, "__builtin_sh_media_MSHARDS_Q", SH_BLTIN_SUS, 0 }, | |
10864 | { CODE_FOR_mshfhi_b, "__builtin_sh_media_MSHFHI_B", SH_BLTIN_V8QI3, 0 }, | |
10865 | { CODE_FOR_mshfhi_l, "__builtin_sh_media_MSHFHI_L", SH_BLTIN_V2SI3, 0 }, | |
10866 | { CODE_FOR_mshfhi_w, "__builtin_sh_media_MSHFHI_W", SH_BLTIN_V4HI3, 0 }, | |
10867 | { CODE_FOR_mshflo_b, "__builtin_sh_media_MSHFLO_B", SH_BLTIN_V8QI3, 0 }, | |
10868 | { CODE_FOR_mshflo_l, "__builtin_sh_media_MSHFLO_L", SH_BLTIN_V2SI3, 0 }, | |
10869 | { CODE_FOR_mshflo_w, "__builtin_sh_media_MSHFLO_W", SH_BLTIN_V4HI3, 0 }, | |
10870 | { CODE_FOR_ashlv2si3, "__builtin_ashlv2si3", SH_BLTIN_SH_SI, 0 }, | |
10871 | { CODE_FOR_ashlv4hi3, "__builtin_ashlv4hi3", SH_BLTIN_SH_HI, 0 }, | |
10872 | { CODE_FOR_lshrv2si3, "__builtin_lshrv2si3", SH_BLTIN_SH_SI, 0 }, | |
10873 | { CODE_FOR_lshrv4hi3, "__builtin_lshrv4hi3", SH_BLTIN_SH_HI, 0 }, | |
10874 | { CODE_FOR_subv2si3, "__builtin_subv2si3", SH_BLTIN_V2SI3, 0 }, | |
10875 | { CODE_FOR_subv4hi3, "__builtin_subv4hi3", SH_BLTIN_V4HI3, 0 }, | |
10876 | { CODE_FOR_sssubv2si3,"__builtin_sssubv2si3", SH_BLTIN_V2SI3, 0 }, | |
10877 | { CODE_FOR_ussubv8qi3,"__builtin_ussubv8qi3", SH_BLTIN_V8QI3, 0 }, | |
10878 | { CODE_FOR_sssubv4hi3,"__builtin_sssubv4hi3", SH_BLTIN_V4HI3, 0 }, | |
10879 | { CODE_FOR_fcosa_s, "__builtin_sh_media_FCOSA_S", SH_BLTIN_SISF, 0 }, | |
10880 | { CODE_FOR_fsina_s, "__builtin_sh_media_FSINA_S", SH_BLTIN_SISF, 0 }, | |
10881 | { CODE_FOR_fipr, "__builtin_sh_media_FIPR_S", SH_BLTIN_3, 0 }, | |
10882 | { CODE_FOR_ftrv, "__builtin_sh_media_FTRV_S", SH_BLTIN_3, 0 }, | |
10883 | { CODE_FOR_mac_media, "__builtin_sh_media_FMAC_S", SH_BLTIN_3, 0 }, | |
10884 | { CODE_FOR_sqrtdf2, "__builtin_sh_media_FSQRT_D", SH_BLTIN_2, 0 }, | |
10885 | { CODE_FOR_sqrtsf2, "__builtin_sh_media_FSQRT_S", SH_BLTIN_2, 0 }, | |
10886 | { CODE_FOR_fsrra_s, "__builtin_sh_media_FSRRA_S", SH_BLTIN_2, 0 }, | |
10887 | { CODE_FOR_ldhi_l, "__builtin_sh_media_LDHI_L", SH_BLTIN_LDUA_L, 0 }, | |
10888 | { CODE_FOR_ldhi_q, "__builtin_sh_media_LDHI_Q", SH_BLTIN_LDUA_Q, 0 }, | |
10889 | { CODE_FOR_ldlo_l, "__builtin_sh_media_LDLO_L", SH_BLTIN_LDUA_L, 0 }, | |
10890 | { CODE_FOR_ldlo_q, "__builtin_sh_media_LDLO_Q", SH_BLTIN_LDUA_Q, 0 }, | |
10891 | { CODE_FOR_sthi_l, "__builtin_sh_media_STHI_L", SH_BLTIN_STUA_L, 0 }, | |
10892 | { CODE_FOR_sthi_q, "__builtin_sh_media_STHI_Q", SH_BLTIN_STUA_Q, 0 }, | |
10893 | { CODE_FOR_stlo_l, "__builtin_sh_media_STLO_L", SH_BLTIN_STUA_L, 0 }, | |
10894 | { CODE_FOR_stlo_q, "__builtin_sh_media_STLO_Q", SH_BLTIN_STUA_Q, 0 }, | |
10895 | { CODE_FOR_ldhi_l64, "__builtin_sh_media_LDHI_L", SH_BLTIN_LDUA_L64, 0 }, | |
10896 | { CODE_FOR_ldhi_q64, "__builtin_sh_media_LDHI_Q", SH_BLTIN_LDUA_Q64, 0 }, | |
10897 | { CODE_FOR_ldlo_l64, "__builtin_sh_media_LDLO_L", SH_BLTIN_LDUA_L64, 0 }, | |
10898 | { CODE_FOR_ldlo_q64, "__builtin_sh_media_LDLO_Q", SH_BLTIN_LDUA_Q64, 0 }, | |
10899 | { CODE_FOR_sthi_l64, "__builtin_sh_media_STHI_L", SH_BLTIN_STUA_L64, 0 }, | |
10900 | { CODE_FOR_sthi_q64, "__builtin_sh_media_STHI_Q", SH_BLTIN_STUA_Q64, 0 }, | |
10901 | { CODE_FOR_stlo_l64, "__builtin_sh_media_STLO_L", SH_BLTIN_STUA_L64, 0 }, | |
10902 | { CODE_FOR_stlo_q64, "__builtin_sh_media_STLO_Q", SH_BLTIN_STUA_Q64, 0 }, | |
10903 | { CODE_FOR_nsb, "__builtin_sh_media_NSB", SH_BLTIN_SU, 0 }, | |
10904 | { CODE_FOR_byterev, "__builtin_sh_media_BYTEREV", SH_BLTIN_2, 0 }, | |
10905 | { CODE_FOR_prefetch, "__builtin_sh_media_PREFO", SH_BLTIN_PSSV, 0 }, | |
c1b92d09 R |
10906 | }; |
10907 | ||
10908 | static void | |
cf277499 | 10909 | sh_media_init_builtins (void) |
c1b92d09 R |
10910 | { |
10911 | tree shared[SH_BLTIN_NUM_SHARED_SIGNATURES]; | |
1cc24576 | 10912 | struct builtin_description *d; |
c1b92d09 | 10913 | |
a3acdc0c | 10914 | memset (shared, 0, sizeof shared); |
e97a46ce | 10915 | for (d = bdesc; d - bdesc < (int) ARRAY_SIZE (bdesc); d++) |
c1b92d09 | 10916 | { |
73a4d10b | 10917 | tree type, arg_type = 0; |
c1b92d09 R |
10918 | int signature = d->signature; |
10919 | int i; | |
10920 | ||
10921 | if (signature < SH_BLTIN_NUM_SHARED_SIGNATURES && shared[signature]) | |
10922 | type = shared[signature]; | |
10923 | else | |
10924 | { | |
10925 | int has_result = signature_args[signature][0] != 0; | |
10926 | ||
73a4d10b R |
10927 | if ((signature_args[signature][1] & 8) |
10928 | && (((signature_args[signature][1] & 1) && TARGET_SHMEDIA32) | |
10929 | || ((signature_args[signature][1] & 2) && TARGET_SHMEDIA64))) | |
c1b92d09 R |
10930 | continue; |
10931 | if (! TARGET_FPU_ANY | |
10932 | && FLOAT_MODE_P (insn_data[d->icode].operand[0].mode)) | |
10933 | continue; | |
10934 | type = void_list_node; | |
10935 | for (i = 3; ; i--) | |
10936 | { | |
10937 | int arg = signature_args[signature][i]; | |
10938 | int opno = i - 1 + has_result; | |
10939 | ||
73a4d10b | 10940 | if (arg & 8) |
c1b92d09 R |
10941 | arg_type = ptr_type_node; |
10942 | else if (arg) | |
73a4d10b R |
10943 | arg_type = (*lang_hooks.types.type_for_mode) |
10944 | (insn_data[d->icode].operand[opno].mode, | |
10945 | (arg & 1)); | |
c1b92d09 R |
10946 | else if (i) |
10947 | continue; | |
10948 | else | |
10949 | arg_type = void_type_node; | |
10950 | if (i == 0) | |
10951 | break; | |
10952 | type = tree_cons (NULL_TREE, arg_type, type); | |
10953 | } | |
10954 | type = build_function_type (arg_type, type); | |
10955 | if (signature < SH_BLTIN_NUM_SHARED_SIGNATURES) | |
10956 | shared[signature] = type; | |
10957 | } | |
1cc24576 KK |
10958 | d->fndecl = |
10959 | add_builtin_function (d->name, type, d - bdesc, BUILT_IN_MD, | |
10960 | NULL, NULL_TREE); | |
c1b92d09 R |
10961 | } |
10962 | } | |
10963 | ||
1cc24576 KK |
10964 | /* Returns the shmedia builtin decl for CODE. */ |
10965 | ||
10966 | static tree | |
10967 | sh_media_builtin_decl (unsigned code, bool initialize_p ATTRIBUTE_UNUSED) | |
10968 | { | |
10969 | if (code >= ARRAY_SIZE (bdesc)) | |
10970 | return error_mark_node; | |
10971 | ||
10972 | return bdesc[code].fndecl; | |
10973 | } | |
10974 | ||
f676971a EC |
10975 | /* Implements target hook vector_mode_supported_p. */ |
10976 | bool | |
10977 | sh_vector_mode_supported_p (enum machine_mode mode) | |
10978 | { | |
10979 | if (TARGET_FPU_ANY | |
10980 | && ((mode == V2SFmode) | |
10981 | || (mode == V4SFmode) | |
10982 | || (mode == V16SFmode))) | |
10983 | return true; | |
10984 | ||
10985 | else if (TARGET_SHMEDIA | |
10986 | && ((mode == V8QImode) | |
10987 | || (mode == V2HImode) | |
10988 | || (mode == V4HImode) | |
10989 | || (mode == V2SImode))) | |
10990 | return true; | |
10991 | ||
10992 | return false; | |
10993 | } | |
10994 | ||
7a296495 CB |
10995 | bool |
10996 | sh_frame_pointer_required (void) | |
10997 | { | |
10998 | /* If needed override this in other tm.h files to cope with various OS | |
10999 | lossage requiring a frame pointer. */ | |
11000 | if (SUBTARGET_FRAME_POINTER_REQUIRED) | |
11001 | return true; | |
11002 | ||
11003 | if (crtl->profile) | |
11004 | return true; | |
11005 | ||
11006 | return false; | |
11007 | } | |
11008 | ||
a1c496cb EC |
11009 | /* Implements target hook dwarf_calling_convention. Return an enum |
11010 | of dwarf_calling_convention. */ | |
11011 | int | |
3101faab | 11012 | sh_dwarf_calling_convention (const_tree func) |
a1c496cb EC |
11013 | { |
11014 | if (sh_attr_renesas_p (func)) | |
29832164 | 11015 | return DW_CC_GNU_renesas_sh; |
a1c496cb EC |
11016 | |
11017 | return DW_CC_normal; | |
11018 | } | |
11019 | ||
c1b92d09 | 11020 | static void |
cf277499 | 11021 | sh_init_builtins (void) |
c1b92d09 R |
11022 | { |
11023 | if (TARGET_SHMEDIA) | |
11024 | sh_media_init_builtins (); | |
11025 | } | |
11026 | ||
1cc24576 KK |
11027 | /* Returns the sh builtin decl for CODE. */ |
11028 | ||
11029 | static tree | |
11030 | sh_builtin_decl (unsigned code, bool initialize_p ATTRIBUTE_UNUSED) | |
11031 | { | |
11032 | if (TARGET_SHMEDIA) | |
11033 | return sh_media_builtin_decl (code, initialize_p); | |
11034 | ||
11035 | return error_mark_node; | |
11036 | } | |
11037 | ||
c1b92d09 R |
11038 | /* Expand an expression EXP that calls a built-in function, |
11039 | with result going to TARGET if that's convenient | |
11040 | (and in mode MODE if that's convenient). | |
11041 | SUBTARGET may be used as the target for computing one of EXP's operands. | |
11042 | IGNORE is nonzero if the value is to be ignored. */ | |
11043 | ||
11044 | static rtx | |
cf277499 SB |
11045 | sh_expand_builtin (tree exp, rtx target, rtx subtarget ATTRIBUTE_UNUSED, |
11046 | enum machine_mode mode ATTRIBUTE_UNUSED, int ignore) | |
c1b92d09 | 11047 | { |
5039610b | 11048 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); |
c1b92d09 R |
11049 | unsigned int fcode = DECL_FUNCTION_CODE (fndecl); |
11050 | const struct builtin_description *d = &bdesc[fcode]; | |
11051 | enum insn_code icode = d->icode; | |
11052 | int signature = d->signature; | |
11053 | enum machine_mode tmode = VOIDmode; | |
11054 | int nop = 0, i; | |
11055 | rtx op[4]; | |
73a4d10b | 11056 | rtx pat = 0; |
c1b92d09 R |
11057 | |
11058 | if (signature_args[signature][0]) | |
11059 | { | |
11060 | if (ignore) | |
11061 | return 0; | |
11062 | ||
11063 | tmode = insn_data[icode].operand[0].mode; | |
11064 | if (! target | |
11065 | || GET_MODE (target) != tmode | |
11066 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
11067 | target = gen_reg_rtx (tmode); | |
11068 | op[nop++] = target; | |
11069 | } | |
11070 | else | |
11071 | target = 0; | |
11072 | ||
11073 | for (i = 1; i <= 3; i++, nop++) | |
11074 | { | |
11075 | tree arg; | |
11076 | enum machine_mode opmode, argmode; | |
73a4d10b | 11077 | tree optype; |
c1b92d09 R |
11078 | |
11079 | if (! signature_args[signature][i]) | |
11080 | break; | |
5039610b | 11081 | arg = CALL_EXPR_ARG (exp, i - 1); |
3578cf63 R |
11082 | if (arg == error_mark_node) |
11083 | return const0_rtx; | |
73a4d10b R |
11084 | if (signature_args[signature][i] & 8) |
11085 | { | |
11086 | opmode = ptr_mode; | |
11087 | optype = ptr_type_node; | |
11088 | } | |
11089 | else | |
11090 | { | |
11091 | opmode = insn_data[icode].operand[nop].mode; | |
11092 | optype = (*lang_hooks.types.type_for_mode) (opmode, 0); | |
11093 | } | |
c1b92d09 R |
11094 | argmode = TYPE_MODE (TREE_TYPE (arg)); |
11095 | if (argmode != opmode) | |
73a4d10b | 11096 | arg = build1 (NOP_EXPR, optype, arg); |
b757d36f | 11097 | op[nop] = expand_expr (arg, NULL_RTX, opmode, EXPAND_NORMAL); |
c1b92d09 R |
11098 | if (! (*insn_data[icode].operand[nop].predicate) (op[nop], opmode)) |
11099 | op[nop] = copy_to_mode_reg (opmode, op[nop]); | |
11100 | } | |
11101 | ||
11102 | switch (nop) | |
11103 | { | |
11104 | case 1: | |
11105 | pat = (*insn_data[d->icode].genfun) (op[0]); | |
11106 | break; | |
11107 | case 2: | |
11108 | pat = (*insn_data[d->icode].genfun) (op[0], op[1]); | |
11109 | break; | |
11110 | case 3: | |
11111 | pat = (*insn_data[d->icode].genfun) (op[0], op[1], op[2]); | |
11112 | break; | |
11113 | case 4: | |
11114 | pat = (*insn_data[d->icode].genfun) (op[0], op[1], op[2], op[3]); | |
11115 | break; | |
95c0af87 | 11116 | default: |
f5b9e7c9 | 11117 | gcc_unreachable (); |
c1b92d09 R |
11118 | } |
11119 | if (! pat) | |
11120 | return 0; | |
11121 | emit_insn (pat); | |
11122 | return target; | |
11123 | } | |
0ac78517 R |
11124 | |
11125 | void | |
cf277499 | 11126 | sh_expand_unop_v2sf (enum rtx_code code, rtx op0, rtx op1) |
0ac78517 R |
11127 | { |
11128 | rtx sel0 = const0_rtx; | |
11129 | rtx sel1 = const1_rtx; | |
cf277499 | 11130 | rtx (*fn) (rtx, rtx, rtx, rtx, rtx) = gen_unary_sf_op; |
0ac78517 R |
11131 | rtx op = gen_rtx_fmt_e (code, SFmode, op1); |
11132 | ||
11133 | emit_insn ((*fn) (op0, op1, op, sel0, sel0)); | |
11134 | emit_insn ((*fn) (op0, op1, op, sel1, sel1)); | |
11135 | } | |
11136 | ||
11137 | void | |
cf277499 | 11138 | sh_expand_binop_v2sf (enum rtx_code code, rtx op0, rtx op1, rtx op2) |
0ac78517 | 11139 | { |
0ac78517 R |
11140 | rtx op = gen_rtx_fmt_ee (code, SFmode, op1, op2); |
11141 | ||
f8be3f56 KK |
11142 | emit_insn (gen_binary_sf_op0 (op0, op1, op2, op)); |
11143 | emit_insn (gen_binary_sf_op1 (op0, op1, op2, op)); | |
0ac78517 R |
11144 | } |
11145 | ||
347146a4 KK |
11146 | /* Return true if hard register REGNO can hold a value of machine-mode MODE. |
11147 | We can allow any mode in any general register. The special registers | |
11148 | only allow SImode. Don't allow any mode in the PR. | |
11149 | ||
11150 | We cannot hold DCmode values in the XD registers because alter_reg | |
11151 | handles subregs of them incorrectly. We could work around this by | |
11152 | spacing the XD registers like the DR registers, but this would require | |
11153 | additional memory in every compilation to hold larger register vectors. | |
11154 | We could hold SFmode / SCmode values in XD registers, but that | |
11155 | would require a tertiary reload when reloading from / to memory, | |
11156 | and a secondary reload to reload from / to general regs; that | |
11157 | seems to be a loosing proposition. | |
11158 | ||
11159 | We want to allow TImode FP regs so that when V4SFmode is loaded as TImode, | |
11160 | it won't be ferried through GP registers first. */ | |
11161 | ||
11162 | bool | |
11163 | sh_hard_regno_mode_ok (unsigned int regno, enum machine_mode mode) | |
11164 | { | |
11165 | if (SPECIAL_REGISTER_P (regno)) | |
11166 | return mode == SImode; | |
11167 | ||
11168 | if (regno == FPUL_REG) | |
11169 | return (mode == SImode || mode == SFmode); | |
11170 | ||
11171 | if (FP_REGISTER_P (regno) && mode == SFmode) | |
11172 | return true; | |
11173 | ||
11174 | if (mode == V2SFmode) | |
11175 | { | |
11176 | if (((FP_REGISTER_P (regno) && (regno - FIRST_FP_REG) % 2 == 0) | |
11177 | || GENERAL_REGISTER_P (regno))) | |
11178 | return true; | |
11179 | else | |
11180 | return false; | |
11181 | } | |
11182 | ||
11183 | if (mode == V4SFmode) | |
11184 | { | |
11185 | if ((FP_REGISTER_P (regno) && (regno - FIRST_FP_REG) % 4 == 0) | |
11186 | || GENERAL_REGISTER_P (regno)) | |
11187 | return true; | |
11188 | else | |
11189 | return false; | |
11190 | } | |
11191 | ||
11192 | if (mode == V16SFmode) | |
11193 | { | |
11194 | if (TARGET_SHMEDIA) | |
11195 | { | |
11196 | if (FP_REGISTER_P (regno) && (regno - FIRST_FP_REG) % 16 == 0) | |
11197 | return true; | |
11198 | else | |
11199 | return false; | |
11200 | } | |
11201 | else | |
11202 | return regno == FIRST_XD_REG; | |
11203 | } | |
11204 | ||
11205 | if (FP_REGISTER_P (regno)) | |
11206 | { | |
11207 | if (mode == SFmode | |
11208 | || mode == SImode | |
11209 | || ((TARGET_SH2E || TARGET_SHMEDIA) && mode == SCmode) | |
11210 | || ((((TARGET_SH4 || TARGET_SH2A_DOUBLE) && mode == DFmode) | |
11211 | || mode == DCmode | |
11212 | || (TARGET_SHMEDIA | |
11213 | && (mode == DFmode || mode == DImode | |
11214 | || mode == V2SFmode || mode == TImode))) | |
11215 | && ((regno - FIRST_FP_REG) & 1) == 0) | |
11216 | || ((TARGET_SH4 || TARGET_SHMEDIA) && mode == TImode | |
11217 | && ((regno - FIRST_FP_REG) & 3) == 0)) | |
11218 | return true; | |
11219 | else | |
11220 | return false; | |
11221 | } | |
11222 | ||
11223 | if (XD_REGISTER_P (regno)) | |
11224 | return mode == DFmode; | |
11225 | ||
11226 | if (TARGET_REGISTER_P (regno)) | |
11227 | return (mode == DImode || mode == SImode || mode == PDImode); | |
11228 | ||
11229 | if (regno == PR_REG) | |
11230 | return mode == SImode; | |
11231 | ||
11232 | if (regno == FPSCR_REG) | |
11233 | return mode == PSImode; | |
11234 | ||
11235 | /* FIXME. This works around PR target/37633 for -O0. */ | |
11236 | if (!optimize && TARGET_SHMEDIA32 && GET_MODE_SIZE (mode) > 4) | |
11237 | { | |
11238 | unsigned int n = GET_MODE_SIZE (mode) / 8; | |
11239 | ||
11240 | if (regno >= FIRST_GENERAL_REG + 10 - n + 1 | |
11241 | && regno <= FIRST_GENERAL_REG + 14) | |
11242 | return false; | |
11243 | } | |
11244 | ||
11245 | return true; | |
11246 | } | |
11247 | ||
cff9f8d5 AH |
11248 | /* Return the class of registers for which a mode change from FROM to TO |
11249 | is invalid. */ | |
b0c42aed | 11250 | bool |
cf277499 | 11251 | sh_cannot_change_mode_class (enum machine_mode from, enum machine_mode to, |
0a2aaacc | 11252 | enum reg_class rclass) |
cff9f8d5 | 11253 | { |
312209c6 AO |
11254 | /* We want to enable the use of SUBREGs as a means to |
11255 | VEC_SELECT a single element of a vector. */ | |
11256 | if (to == SFmode && VECTOR_MODE_P (from) && GET_MODE_INNER (from) == SFmode) | |
0a2aaacc | 11257 | return (reg_classes_intersect_p (GENERAL_REGS, rclass)); |
312209c6 | 11258 | |
cff9f8d5 AH |
11259 | if (GET_MODE_SIZE (from) != GET_MODE_SIZE (to)) |
11260 | { | |
832a3292 KH |
11261 | if (TARGET_LITTLE_ENDIAN) |
11262 | { | |
11263 | if (GET_MODE_SIZE (to) < 8 || GET_MODE_SIZE (from) < 8) | |
0a2aaacc | 11264 | return reg_classes_intersect_p (DF_REGS, rclass); |
832a3292 KH |
11265 | } |
11266 | else | |
11267 | { | |
11268 | if (GET_MODE_SIZE (from) < 8) | |
0a2aaacc | 11269 | return reg_classes_intersect_p (DF_HI_REGS, rclass); |
832a3292 | 11270 | } |
cff9f8d5 | 11271 | } |
b0c42aed | 11272 | return 0; |
cff9f8d5 AH |
11273 | } |
11274 | ||
42db504c SB |
11275 | /* Return true if registers in machine mode MODE will likely be |
11276 | allocated to registers in small register classes. */ | |
11277 | ||
9c265c4d | 11278 | bool |
42db504c SB |
11279 | sh_small_register_classes_for_mode_p (enum machine_mode mode ATTRIBUTE_UNUSED) |
11280 | { | |
11281 | return (! TARGET_SHMEDIA); | |
11282 | } | |
ea4210ef RS |
11283 | |
11284 | /* If ADDRESS refers to a CODE_LABEL, add NUSES to the number of times | |
11285 | that label is used. */ | |
11286 | ||
11287 | void | |
cf277499 | 11288 | sh_mark_label (rtx address, int nuses) |
ea4210ef RS |
11289 | { |
11290 | if (GOTOFF_P (address)) | |
11291 | { | |
11292 | /* Extract the label or symbol. */ | |
11293 | address = XEXP (address, 0); | |
11294 | if (GET_CODE (address) == PLUS) | |
11295 | address = XEXP (address, 0); | |
11296 | address = XVECEXP (address, 0, 0); | |
11297 | } | |
11298 | if (GET_CODE (address) == LABEL_REF | |
f3536097 | 11299 | && LABEL_P (XEXP (address, 0))) |
ea4210ef RS |
11300 | LABEL_NUSES (XEXP (address, 0)) += nuses; |
11301 | } | |
11302 | ||
95c0af87 JR |
11303 | /* Compute extra cost of moving data between one register class |
11304 | and another. */ | |
11305 | ||
ec555f32 R |
11306 | /* If SECONDARY*_RELOAD_CLASS says something about the src/dst pair, regclass |
11307 | uses this information. Hence, the general register <-> floating point | |
95c0af87 JR |
11308 | register information here is not used for SFmode. */ |
11309 | ||
11310 | int | |
cf277499 SB |
11311 | sh_register_move_cost (enum machine_mode mode, |
11312 | enum reg_class srcclass, enum reg_class dstclass) | |
95c0af87 JR |
11313 | { |
11314 | if (dstclass == T_REGS || dstclass == PR_REGS) | |
11315 | return 10; | |
11316 | ||
b1dcf523 DD |
11317 | if (dstclass == MAC_REGS && srcclass == MAC_REGS) |
11318 | return 4; | |
11319 | ||
ec555f32 R |
11320 | if (mode == SImode && ! TARGET_SHMEDIA && TARGET_FMOVD |
11321 | && REGCLASS_HAS_FP_REG (srcclass) | |
11322 | && REGCLASS_HAS_FP_REG (dstclass)) | |
11323 | return 4; | |
11324 | ||
9c24200c KK |
11325 | if (REGCLASS_HAS_FP_REG (dstclass) && srcclass == T_REGS) |
11326 | return ((TARGET_HARD_SH4 && !optimize_size) ? 10 : 7); | |
11327 | ||
a60a9553 | 11328 | if ((REGCLASS_HAS_FP_REG (dstclass) && srcclass == MAC_REGS) |
832a3292 | 11329 | || (dstclass == MAC_REGS && REGCLASS_HAS_FP_REG (srcclass))) |
a60a9553 R |
11330 | return 9; |
11331 | ||
95c0af87 JR |
11332 | if ((REGCLASS_HAS_FP_REG (dstclass) |
11333 | && REGCLASS_HAS_GENERAL_REG (srcclass)) | |
11334 | || (REGCLASS_HAS_GENERAL_REG (dstclass) | |
11335 | && REGCLASS_HAS_FP_REG (srcclass))) | |
832a3292 KH |
11336 | return ((TARGET_SHMEDIA ? 4 : TARGET_FMOVD ? 8 : 12) |
11337 | * ((GET_MODE_SIZE (mode) + 7) / 8U)); | |
95c0af87 JR |
11338 | |
11339 | if ((dstclass == FPUL_REGS | |
11340 | && REGCLASS_HAS_GENERAL_REG (srcclass)) | |
11341 | || (srcclass == FPUL_REGS | |
11342 | && REGCLASS_HAS_GENERAL_REG (dstclass))) | |
11343 | return 5; | |
11344 | ||
11345 | if ((dstclass == FPUL_REGS | |
11346 | && (srcclass == PR_REGS || srcclass == MAC_REGS || srcclass == T_REGS)) | |
f676971a | 11347 | || (srcclass == FPUL_REGS |
95c0af87 JR |
11348 | && (dstclass == PR_REGS || dstclass == MAC_REGS))) |
11349 | return 7; | |
11350 | ||
11351 | if ((srcclass == TARGET_REGS && ! REGCLASS_HAS_GENERAL_REG (dstclass)) | |
11352 | || ((dstclass) == TARGET_REGS && ! REGCLASS_HAS_GENERAL_REG (srcclass))) | |
11353 | return 20; | |
11354 | ||
73a4d10b R |
11355 | /* ??? ptabs faults on (value & 0x3) == 0x3 */ |
11356 | if (TARGET_SHMEDIA | |
11357 | && ((srcclass) == TARGET_REGS || (srcclass) == SIBCALL_REGS)) | |
11358 | { | |
7dd2f19b RS |
11359 | if (sh_gettrcost >= 0) |
11360 | return sh_gettrcost; | |
73a4d10b R |
11361 | else if (!TARGET_PT_FIXED) |
11362 | return 100; | |
11363 | } | |
11364 | ||
95c0af87 JR |
11365 | if ((srcclass == FPSCR_REGS && ! REGCLASS_HAS_GENERAL_REG (dstclass)) |
11366 | || (dstclass == FPSCR_REGS && ! REGCLASS_HAS_GENERAL_REG (srcclass))) | |
11367 | return 4; | |
11368 | ||
ec555f32 R |
11369 | if (TARGET_SHMEDIA |
11370 | || (TARGET_FMOVD | |
11371 | && ! REGCLASS_HAS_GENERAL_REG (srcclass) | |
11372 | && ! REGCLASS_HAS_GENERAL_REG (dstclass))) | |
11373 | return 2 * ((GET_MODE_SIZE (mode) + 7) / 8U); | |
11374 | ||
11375 | return 2 * ((GET_MODE_SIZE (mode) + 3) / 4U); | |
95c0af87 JR |
11376 | } |
11377 | ||
cf277499 | 11378 | static rtx emit_load_ptr (rtx, rtx); |
1aa03f38 R |
11379 | |
11380 | static rtx | |
cf277499 | 11381 | emit_load_ptr (rtx reg, rtx addr) |
1aa03f38 | 11382 | { |
57d38024 | 11383 | rtx mem = gen_const_mem (ptr_mode, addr); |
1aa03f38 R |
11384 | |
11385 | if (Pmode != ptr_mode) | |
11386 | mem = gen_rtx_SIGN_EXTEND (Pmode, mem); | |
11387 | return emit_move_insn (reg, mem); | |
11388 | } | |
11389 | ||
73a4d10b | 11390 | static void |
cf277499 SB |
11391 | sh_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED, |
11392 | HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset, | |
11393 | tree function) | |
1aa03f38 R |
11394 | { |
11395 | CUMULATIVE_ARGS cum; | |
11396 | int structure_value_byref = 0; | |
0a2aaacc | 11397 | rtx this_rtx, this_value, sibcall, insns, funexp; |
1aa03f38 | 11398 | tree funtype = TREE_TYPE (function); |
735cb76e | 11399 | int simple_add = CONST_OK_FOR_ADD (delta); |
1aa03f38 R |
11400 | int did_load = 0; |
11401 | rtx scratch0, scratch1, scratch2; | |
73a4d10b | 11402 | unsigned i; |
1aa03f38 R |
11403 | |
11404 | reload_completed = 1; | |
fe3ad572 | 11405 | epilogue_completed = 1; |
1aa03f38 R |
11406 | current_function_uses_only_leaf_regs = 1; |
11407 | ||
2e040219 | 11408 | emit_note (NOTE_INSN_PROLOGUE_END); |
1aa03f38 R |
11409 | |
11410 | /* Find the "this" pointer. We have such a wide range of ABIs for the | |
11411 | SH that it's best to do this completely machine independently. | |
f676971a | 11412 | "this" is passed as first argument, unless a structure return pointer |
1aa03f38 | 11413 | comes first, in which case "this" comes second. */ |
0f6937fe | 11414 | INIT_CUMULATIVE_ARGS (cum, funtype, NULL_RTX, 0, 1); |
1aa03f38 | 11415 | #ifndef PCC_STATIC_STRUCT_RETURN |
61f71b34 | 11416 | if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function)) |
1aa03f38 R |
11417 | structure_value_byref = 1; |
11418 | #endif /* not PCC_STATIC_STRUCT_RETURN */ | |
61f71b34 | 11419 | if (structure_value_byref && sh_struct_value_rtx (function, 0) == 0) |
f676971a | 11420 | { |
1aa03f38 R |
11421 | tree ptype = build_pointer_type (TREE_TYPE (funtype)); |
11422 | ||
11423 | FUNCTION_ARG_ADVANCE (cum, Pmode, ptype, 1); | |
11424 | } | |
0a2aaacc | 11425 | this_rtx = FUNCTION_ARG (cum, Pmode, ptr_type_node, 1); |
1aa03f38 R |
11426 | |
11427 | /* For SHcompact, we only have r0 for a scratch register: r1 is the | |
11428 | static chain pointer (even if you can't have nested virtual functions | |
11429 | right now, someone might implement them sometime), and the rest of the | |
11430 | registers are used for argument passing, are callee-saved, or reserved. */ | |
73a4d10b R |
11431 | /* We need to check call_used_regs / fixed_regs in case -fcall_saved-reg / |
11432 | -ffixed-reg has been used. */ | |
11433 | if (! call_used_regs[0] || fixed_regs[0]) | |
11434 | error ("r0 needs to be available as a call-clobbered register"); | |
1aa03f38 R |
11435 | scratch0 = scratch1 = scratch2 = gen_rtx_REG (Pmode, 0); |
11436 | if (! TARGET_SH5) | |
11437 | { | |
73a4d10b R |
11438 | if (call_used_regs[1] && ! fixed_regs[1]) |
11439 | scratch1 = gen_rtx_REG (ptr_mode, 1); | |
1aa03f38 R |
11440 | /* N.B., if not TARGET_HITACHI, register 2 is used to pass the pointer |
11441 | pointing where to return struct values. */ | |
73a4d10b R |
11442 | if (call_used_regs[3] && ! fixed_regs[3]) |
11443 | scratch2 = gen_rtx_REG (Pmode, 3); | |
1aa03f38 R |
11444 | } |
11445 | else if (TARGET_SHMEDIA) | |
11446 | { | |
73a4d10b R |
11447 | for (i = FIRST_GENERAL_REG; i <= LAST_GENERAL_REG; i++) |
11448 | if (i != REGNO (scratch0) && | |
11449 | call_used_regs[i] && ! fixed_regs[i] && ! FUNCTION_ARG_REGNO_P (i)) | |
11450 | { | |
11451 | scratch1 = gen_rtx_REG (ptr_mode, i); | |
11452 | break; | |
11453 | } | |
11454 | if (scratch1 == scratch0) | |
11455 | error ("Need a second call-clobbered general purpose register"); | |
11456 | for (i = FIRST_TARGET_REG; i <= LAST_TARGET_REG; i++) | |
11457 | if (call_used_regs[i] && ! fixed_regs[i]) | |
11458 | { | |
11459 | scratch2 = gen_rtx_REG (Pmode, i); | |
11460 | break; | |
11461 | } | |
11462 | if (scratch2 == scratch0) | |
11463 | error ("Need a call-clobbered target register"); | |
1aa03f38 R |
11464 | } |
11465 | ||
0a2aaacc | 11466 | this_value = plus_constant (this_rtx, delta); |
1aa03f38 R |
11467 | if (vcall_offset |
11468 | && (simple_add || scratch0 != scratch1) | |
11469 | && strict_memory_address_p (ptr_mode, this_value)) | |
11470 | { | |
11471 | emit_load_ptr (scratch0, this_value); | |
11472 | did_load = 1; | |
11473 | } | |
11474 | ||
11475 | if (!delta) | |
11476 | ; /* Do nothing. */ | |
11477 | else if (simple_add) | |
0a2aaacc | 11478 | emit_move_insn (this_rtx, this_value); |
1aa03f38 R |
11479 | else |
11480 | { | |
11481 | emit_move_insn (scratch1, GEN_INT (delta)); | |
0a2aaacc | 11482 | emit_insn (gen_add2_insn (this_rtx, scratch1)); |
1aa03f38 R |
11483 | } |
11484 | ||
11485 | if (vcall_offset) | |
11486 | { | |
11487 | rtx offset_addr; | |
11488 | ||
11489 | if (!did_load) | |
0a2aaacc | 11490 | emit_load_ptr (scratch0, this_rtx); |
1aa03f38 R |
11491 | |
11492 | offset_addr = plus_constant (scratch0, vcall_offset); | |
11493 | if (strict_memory_address_p (ptr_mode, offset_addr)) | |
11494 | ; /* Do nothing. */ | |
73a4d10b | 11495 | else if (! TARGET_SH5 && scratch0 != scratch1) |
1aa03f38 R |
11496 | { |
11497 | /* scratch0 != scratch1, and we have indexed loads. Get better | |
11498 | schedule by loading the offset into r1 and using an indexed | |
11499 | load - then the load of r1 can issue before the load from | |
0a2aaacc | 11500 | (this_rtx + delta) finishes. */ |
1aa03f38 R |
11501 | emit_move_insn (scratch1, GEN_INT (vcall_offset)); |
11502 | offset_addr = gen_rtx_PLUS (Pmode, scratch0, scratch1); | |
11503 | } | |
735cb76e | 11504 | else if (CONST_OK_FOR_ADD (vcall_offset)) |
1aa03f38 R |
11505 | { |
11506 | emit_insn (gen_add2_insn (scratch0, GEN_INT (vcall_offset))); | |
11507 | offset_addr = scratch0; | |
11508 | } | |
11509 | else if (scratch0 != scratch1) | |
11510 | { | |
11511 | emit_move_insn (scratch1, GEN_INT (vcall_offset)); | |
11512 | emit_insn (gen_add2_insn (scratch0, scratch1)); | |
11513 | offset_addr = scratch0; | |
11514 | } | |
11515 | else | |
f5b9e7c9 | 11516 | gcc_unreachable (); /* FIXME */ |
1aa03f38 R |
11517 | emit_load_ptr (scratch0, offset_addr); |
11518 | ||
832a3292 | 11519 | if (Pmode != ptr_mode) |
1aa03f38 | 11520 | scratch0 = gen_rtx_TRUNCATE (ptr_mode, scratch0); |
0a2aaacc | 11521 | emit_insn (gen_add2_insn (this_rtx, scratch0)); |
1aa03f38 R |
11522 | } |
11523 | ||
11524 | /* Generate a tail call to the target function. */ | |
11525 | if (! TREE_USED (function)) | |
11526 | { | |
11527 | assemble_external (function); | |
11528 | TREE_USED (function) = 1; | |
11529 | } | |
11530 | funexp = XEXP (DECL_RTL (function), 0); | |
73a4d10b R |
11531 | /* If the function is overridden, so is the thunk, hence we don't |
11532 | need GOT addressing even if this is a public symbol. */ | |
11533 | #if 0 | |
11534 | if (TARGET_SH1 && ! flag_weak) | |
11535 | sibcall = gen_sibcalli_thunk (funexp, const0_rtx); | |
11536 | else | |
11537 | #endif | |
11538 | if (TARGET_SH2 && flag_pic) | |
11539 | { | |
11540 | sibcall = gen_sibcall_pcrel (funexp, const0_rtx); | |
11541 | XEXP (XVECEXP (sibcall, 0, 2), 0) = scratch2; | |
11542 | } | |
11543 | else | |
11544 | { | |
11545 | if (TARGET_SHMEDIA && flag_pic) | |
11546 | { | |
11547 | funexp = gen_sym2PIC (funexp); | |
11548 | PUT_MODE (funexp, Pmode); | |
11549 | } | |
11550 | emit_move_insn (scratch2, funexp); | |
11551 | funexp = gen_rtx_MEM (FUNCTION_MODE, scratch2); | |
11552 | sibcall = gen_sibcall (funexp, const0_rtx, NULL_RTX); | |
11553 | } | |
11554 | sibcall = emit_call_insn (sibcall); | |
1aa03f38 | 11555 | SIBLING_CALL_P (sibcall) = 1; |
0a2aaacc | 11556 | use_reg (&CALL_INSN_FUNCTION_USAGE (sibcall), this_rtx); |
1aa03f38 R |
11557 | emit_barrier (); |
11558 | ||
a2855205 | 11559 | /* Run just enough of rest_of_compilation to do scheduling and get |
1aa03f38 R |
11560 | the insns emitted. Note that use_thunk calls |
11561 | assemble_start_function and assemble_end_function. */ | |
a2855205 | 11562 | |
55e092c4 | 11563 | insn_locators_alloc (); |
1aa03f38 R |
11564 | insns = get_insns (); |
11565 | ||
6fb5fa3c DB |
11566 | if (optimize > 0) |
11567 | { | |
11568 | if (! cfun->cfg) | |
eb6cbf91 | 11569 | init_flow (cfun); |
6fb5fa3c DB |
11570 | split_all_insns_noflow (); |
11571 | } | |
1aa03f38 | 11572 | |
18dbd950 | 11573 | sh_reorg (); |
1aa03f38 R |
11574 | |
11575 | if (optimize > 0 && flag_delayed_branch) | |
10d22567 | 11576 | dbr_schedule (insns); |
384a8760 | 11577 | |
1aa03f38 R |
11578 | shorten_branches (insns); |
11579 | final_start_function (insns, file, 1); | |
c9d691e9 | 11580 | final (insns, file, 1); |
1aa03f38 R |
11581 | final_end_function (); |
11582 | ||
1aa03f38 | 11583 | reload_completed = 0; |
fe3ad572 | 11584 | epilogue_completed = 0; |
1aa03f38 R |
11585 | } |
11586 | ||
90534361 | 11587 | rtx |
73a4d10b | 11588 | function_symbol (rtx target, const char *name, enum sh_function_kind kind) |
90534361 | 11589 | { |
73a4d10b R |
11590 | rtx sym; |
11591 | ||
11592 | /* If this is not an ordinary function, the name usually comes from a | |
11593 | string literal or an sprintf buffer. Make sure we use the same | |
11594 | string consistently, so that cse will be able to unify address loads. */ | |
11595 | if (kind != FUNCTION_ORDINARY) | |
11596 | name = IDENTIFIER_POINTER (get_identifier (name)); | |
11597 | sym = gen_rtx_SYMBOL_REF (Pmode, name); | |
90534361 | 11598 | SYMBOL_REF_FLAGS (sym) = SYMBOL_FLAG_FUNCTION; |
73a4d10b R |
11599 | if (flag_pic) |
11600 | switch (kind) | |
11601 | { | |
11602 | case FUNCTION_ORDINARY: | |
11603 | break; | |
11604 | case SFUNC_GOT: | |
11605 | { | |
11606 | rtx reg = target ? target : gen_reg_rtx (Pmode); | |
11607 | ||
11608 | emit_insn (gen_symGOT2reg (reg, sym)); | |
11609 | sym = reg; | |
11610 | break; | |
11611 | } | |
11612 | case SFUNC_STATIC: | |
11613 | { | |
11614 | /* ??? To allow cse to work, we use GOTOFF relocations. | |
11615 | we could add combiner patterns to transform this into | |
11616 | straight pc-relative calls with sym2PIC / bsrf when | |
11617 | label load and function call are still 1:1 and in the | |
11618 | same basic block during combine. */ | |
11619 | rtx reg = target ? target : gen_reg_rtx (Pmode); | |
11620 | ||
11621 | emit_insn (gen_symGOTOFF2reg (reg, sym)); | |
11622 | sym = reg; | |
11623 | break; | |
11624 | } | |
11625 | } | |
11626 | if (target && sym != target) | |
11627 | { | |
11628 | emit_move_insn (target, sym); | |
11629 | return target; | |
11630 | } | |
90534361 R |
11631 | return sym; |
11632 | } | |
11633 | ||
fada1961 R |
11634 | /* Find the number of a general purpose register in S. */ |
11635 | static int | |
11636 | scavenge_reg (HARD_REG_SET *s) | |
11637 | { | |
11638 | int r; | |
11639 | for (r = FIRST_GENERAL_REG; r <= LAST_GENERAL_REG; r++) | |
11640 | if (TEST_HARD_REG_BIT (*s, r)) | |
11641 | return r; | |
11642 | return -1; | |
11643 | } | |
11644 | ||
11645 | rtx | |
11646 | sh_get_pr_initial_val (void) | |
11647 | { | |
7d73a2ba R |
11648 | rtx val; |
11649 | ||
b869f904 R |
11650 | /* ??? Unfortunately, get_hard_reg_initial_val doesn't always work for the |
11651 | PR register on SHcompact, because it might be clobbered by the prologue. | |
7d73a2ba | 11652 | We check first if that is known to be the case. */ |
b869f904 | 11653 | if (TARGET_SHCOMPACT |
38173d38 | 11654 | && ((crtl->args.info.call_cookie |
7d73a2ba | 11655 | & ~ CALL_COOKIE_RET_TRAMP (1)) |
e3b5732b | 11656 | || crtl->saves_all_registers)) |
57d38024 | 11657 | return gen_frame_mem (SImode, return_address_pointer_rtx); |
7d73a2ba R |
11658 | |
11659 | /* If we haven't finished rtl generation, there might be a nonlocal label | |
11660 | that we haven't seen yet. | |
b3a13419 ILT |
11661 | ??? get_hard_reg_initial_val fails if it is called after register |
11662 | allocation has started, unless it has been called before for the | |
11663 | same register. And even then, we end in trouble if we didn't use | |
11664 | the register in the same basic block before. So call | |
11665 | get_hard_reg_initial_val now and wrap it in an unspec if we might | |
11666 | need to replace it. */ | |
dce20bbc R |
11667 | /* ??? We also must do this for TARGET_SH1 in general, because otherwise |
11668 | combine can put the pseudo returned by get_hard_reg_initial_val into | |
11669 | instructions that need a general purpose registers, which will fail to | |
11670 | be recognized when the pseudo becomes allocated to PR. */ | |
7d73a2ba R |
11671 | val |
11672 | = get_hard_reg_initial_val (Pmode, TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG); | |
dce20bbc | 11673 | if (TARGET_SH1) |
7d73a2ba R |
11674 | return gen_rtx_UNSPEC (SImode, gen_rtvec (1, val), UNSPEC_RA); |
11675 | return val; | |
fada1961 R |
11676 | } |
11677 | ||
3db1b434 | 11678 | int |
f90b7a5a | 11679 | sh_expand_t_scc (rtx operands[]) |
3db1b434 | 11680 | { |
f90b7a5a PB |
11681 | enum rtx_code code = GET_CODE (operands[1]); |
11682 | rtx target = operands[0]; | |
11683 | rtx op0 = operands[2]; | |
11684 | rtx op1 = operands[3]; | |
3db1b434 R |
11685 | rtx result = target; |
11686 | HOST_WIDE_INT val; | |
11687 | ||
f3536097 SZ |
11688 | if (!REG_P (op0) || REGNO (op0) != T_REG |
11689 | || !CONST_INT_P (op1)) | |
3db1b434 | 11690 | return 0; |
f3536097 | 11691 | if (!REG_P (result)) |
3db1b434 | 11692 | result = gen_reg_rtx (SImode); |
f90b7a5a | 11693 | val = INTVAL (op1); |
3db1b434 R |
11694 | if ((code == EQ && val == 1) || (code == NE && val == 0)) |
11695 | emit_insn (gen_movt (result)); | |
f326a6cb N |
11696 | else if (TARGET_SH2A && ((code == EQ && val == 0) |
11697 | || (code == NE && val == 1))) | |
c185de40 | 11698 | emit_insn (gen_xorsi3_movrt (result)); |
3db1b434 R |
11699 | else if ((code == EQ && val == 0) || (code == NE && val == 1)) |
11700 | { | |
c41c1387 | 11701 | emit_clobber (result); |
3db1b434 | 11702 | emit_insn (gen_subc (result, result, result)); |
a556fd39 | 11703 | emit_insn (gen_addsi3 (result, result, const1_rtx)); |
3db1b434 R |
11704 | } |
11705 | else if (code == EQ || code == NE) | |
11706 | emit_insn (gen_move_insn (result, GEN_INT (code == NE))); | |
11707 | else | |
11708 | return 0; | |
11709 | if (result != target) | |
11710 | emit_move_insn (target, result); | |
11711 | return 1; | |
11712 | } | |
11713 | ||
07d7d2f4 R |
11714 | /* INSN is an sfunc; return the rtx that describes the address used. */ |
11715 | static rtx | |
11716 | extract_sfunc_addr (rtx insn) | |
11717 | { | |
11718 | rtx pattern, part = NULL_RTX; | |
11719 | int len, i; | |
11720 | ||
11721 | pattern = PATTERN (insn); | |
11722 | len = XVECLEN (pattern, 0); | |
11723 | for (i = 0; i < len; i++) | |
11724 | { | |
11725 | part = XVECEXP (pattern, 0, i); | |
11726 | if (GET_CODE (part) == USE && GET_MODE (XEXP (part, 0)) == Pmode | |
11727 | && GENERAL_REGISTER_P (true_regnum (XEXP (part, 0)))) | |
11728 | return XEXP (part, 0); | |
11729 | } | |
f5b9e7c9 NS |
11730 | gcc_assert (GET_CODE (XVECEXP (pattern, 0, 0)) == UNSPEC_VOLATILE); |
11731 | return XVECEXP (XVECEXP (pattern, 0, 0), 0, 1); | |
07d7d2f4 R |
11732 | } |
11733 | ||
11734 | /* Verify that the register in use_sfunc_addr still agrees with the address | |
11735 | used in the sfunc. This prevents fill_slots_from_thread from changing | |
11736 | use_sfunc_addr. | |
11737 | INSN is the use_sfunc_addr instruction, and REG is the register it | |
11738 | guards. */ | |
11739 | int | |
11740 | check_use_sfunc_addr (rtx insn, rtx reg) | |
11741 | { | |
11742 | /* Search for the sfunc. It should really come right after INSN. */ | |
11743 | while ((insn = NEXT_INSN (insn))) | |
11744 | { | |
f3536097 | 11745 | if (LABEL_P (insn) || JUMP_P (insn)) |
07d7d2f4 R |
11746 | break; |
11747 | if (! INSN_P (insn)) | |
11748 | continue; | |
f676971a | 11749 | |
07d7d2f4 R |
11750 | if (GET_CODE (PATTERN (insn)) == SEQUENCE) |
11751 | insn = XVECEXP (PATTERN (insn), 0, 0); | |
11752 | if (GET_CODE (PATTERN (insn)) != PARALLEL | |
11753 | || get_attr_type (insn) != TYPE_SFUNC) | |
11754 | continue; | |
11755 | return rtx_equal_p (extract_sfunc_addr (insn), reg); | |
11756 | } | |
f5b9e7c9 | 11757 | gcc_unreachable (); |
07d7d2f4 R |
11758 | } |
11759 | ||
312209c6 AO |
11760 | /* This function returns a constant rtx that represents pi / 2**15 in |
11761 | SFmode. it's used to scale SFmode angles, in radians, to a | |
11762 | fixed-point signed 16.16-bit fraction of a full circle, i.e., 2*pi | |
11763 | maps to 0x10000). */ | |
11764 | ||
11765 | static GTY(()) rtx sh_fsca_sf2int_rtx; | |
11766 | ||
11767 | rtx | |
11768 | sh_fsca_sf2int (void) | |
11769 | { | |
11770 | if (! sh_fsca_sf2int_rtx) | |
11771 | { | |
11772 | REAL_VALUE_TYPE rv; | |
11773 | ||
11774 | real_from_string (&rv, "10430.378350470453"); | |
11775 | sh_fsca_sf2int_rtx = const_double_from_real_value (rv, SFmode); | |
11776 | } | |
11777 | ||
11778 | return sh_fsca_sf2int_rtx; | |
11779 | } | |
f676971a | 11780 | |
312209c6 AO |
11781 | /* This function returns a constant rtx that represents pi / 2**15 in |
11782 | DFmode. it's used to scale DFmode angles, in radians, to a | |
11783 | fixed-point signed 16.16-bit fraction of a full circle, i.e., 2*pi | |
11784 | maps to 0x10000). */ | |
11785 | ||
11786 | static GTY(()) rtx sh_fsca_df2int_rtx; | |
11787 | ||
11788 | rtx | |
11789 | sh_fsca_df2int (void) | |
11790 | { | |
11791 | if (! sh_fsca_df2int_rtx) | |
11792 | { | |
11793 | REAL_VALUE_TYPE rv; | |
11794 | ||
11795 | real_from_string (&rv, "10430.378350470453"); | |
11796 | sh_fsca_df2int_rtx = const_double_from_real_value (rv, DFmode); | |
11797 | } | |
11798 | ||
11799 | return sh_fsca_df2int_rtx; | |
11800 | } | |
f676971a | 11801 | |
312209c6 AO |
11802 | /* This function returns a constant rtx that represents 2**15 / pi in |
11803 | SFmode. it's used to scale a fixed-point signed 16.16-bit fraction | |
11804 | of a full circle back to a SFmode value, i.e., 0x10000 maps to | |
11805 | 2*pi). */ | |
11806 | ||
11807 | static GTY(()) rtx sh_fsca_int2sf_rtx; | |
11808 | ||
11809 | rtx | |
11810 | sh_fsca_int2sf (void) | |
11811 | { | |
11812 | if (! sh_fsca_int2sf_rtx) | |
11813 | { | |
11814 | REAL_VALUE_TYPE rv; | |
11815 | ||
11816 | real_from_string (&rv, "9.587379924285257e-5"); | |
11817 | sh_fsca_int2sf_rtx = const_double_from_real_value (rv, SFmode); | |
11818 | } | |
11819 | ||
11820 | return sh_fsca_int2sf_rtx; | |
11821 | } | |
cc15e98f | 11822 | |
8b0add27 | 11823 | /* Initialize the CUMULATIVE_ARGS structure. */ |
cc15e98f NC |
11824 | |
11825 | void | |
11826 | sh_init_cumulative_args (CUMULATIVE_ARGS * pcum, | |
11827 | tree fntype, | |
11828 | rtx libname ATTRIBUTE_UNUSED, | |
11829 | tree fndecl, | |
11830 | signed int n_named_args, | |
11831 | enum machine_mode mode) | |
11832 | { | |
11833 | pcum->arg_count [(int) SH_ARG_FLOAT] = 0; | |
11834 | pcum->free_single_fp_reg = 0; | |
11835 | pcum->stack_regs = 0; | |
11836 | pcum->byref_regs = 0; | |
11837 | pcum->byref = 0; | |
11838 | pcum->outgoing = (n_named_args == -1) ? 0 : 1; | |
11839 | ||
11840 | /* XXX - Should we check TARGET_HITACHI here ??? */ | |
11841 | pcum->renesas_abi = sh_attr_renesas_p (fntype) ? 1 : 0; | |
11842 | ||
11843 | if (fntype) | |
11844 | { | |
11845 | pcum->force_mem = ((TARGET_HITACHI || pcum->renesas_abi) | |
11846 | && aggregate_value_p (TREE_TYPE (fntype), fndecl)); | |
11847 | pcum->prototype_p = TYPE_ARG_TYPES (fntype) ? TRUE : FALSE; | |
11848 | pcum->arg_count [(int) SH_ARG_INT] | |
11849 | = TARGET_SH5 && aggregate_value_p (TREE_TYPE (fntype), fndecl); | |
11850 | ||
11851 | pcum->call_cookie | |
11852 | = CALL_COOKIE_RET_TRAMP (TARGET_SHCOMPACT | |
11853 | && pcum->arg_count [(int) SH_ARG_INT] == 0 | |
11854 | && (TYPE_MODE (TREE_TYPE (fntype)) == BLKmode | |
11855 | ? int_size_in_bytes (TREE_TYPE (fntype)) | |
11856 | : GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (fntype)))) > 4 | |
11857 | && (BASE_RETURN_VALUE_REG (TYPE_MODE (TREE_TYPE (fntype))) | |
11858 | == FIRST_RET_REG)); | |
11859 | } | |
11860 | else | |
11861 | { | |
11862 | pcum->arg_count [(int) SH_ARG_INT] = 0; | |
11863 | pcum->prototype_p = FALSE; | |
11864 | if (mode != VOIDmode) | |
11865 | { | |
29832164 | 11866 | pcum->call_cookie = |
cc15e98f NC |
11867 | CALL_COOKIE_RET_TRAMP (TARGET_SHCOMPACT |
11868 | && GET_MODE_SIZE (mode) > 4 | |
11869 | && BASE_RETURN_VALUE_REG (mode) == FIRST_RET_REG); | |
11870 | ||
11871 | /* If the default ABI is the Renesas ABI then all library | |
11872 | calls must assume that the library will be using the | |
11873 | Renesas ABI. So if the function would return its result | |
11874 | in memory then we must force the address of this memory | |
11875 | block onto the stack. Ideally we would like to call | |
11876 | targetm.calls.return_in_memory() here but we do not have | |
8b0add27 | 11877 | the TYPE or the FNDECL available so we synthesize the |
cc15e98f NC |
11878 | contents of that function as best we can. */ |
11879 | pcum->force_mem = | |
c0fb94d7 | 11880 | (TARGET_DEFAULT & MASK_HITACHI) |
cc15e98f NC |
11881 | && (mode == BLKmode |
11882 | || (GET_MODE_SIZE (mode) > 4 | |
11883 | && !(mode == DFmode | |
11884 | && TARGET_FPU_DOUBLE))); | |
11885 | } | |
11886 | else | |
11887 | { | |
11888 | pcum->call_cookie = 0; | |
11889 | pcum->force_mem = FALSE; | |
11890 | } | |
11891 | } | |
11892 | } | |
11893 | ||
73a4d10b R |
11894 | /* Replace any occurrence of FROM(n) in X with TO(n). The function does |
11895 | not enter into CONST_DOUBLE for the replace. | |
11896 | ||
11897 | Note that copying is not done so X must not be shared unless all copies | |
11898 | are to be modified. | |
11899 | ||
11900 | This is like replace_rtx, except that we operate on N_REPLACEMENTS | |
30dc60c7 | 11901 | replacements simultaneously - FROM(n) is replacements[n*2] and to(n) is |
73a4d10b R |
11902 | replacements[n*2+1] - and that we take mode changes into account. |
11903 | ||
30dc60c7 | 11904 | If a replacement is ambiguous, return NULL_RTX. |
73a4d10b R |
11905 | |
11906 | If MODIFY is zero, don't modify any rtl in place, | |
11907 | just return zero or nonzero for failure / success. */ | |
11908 | ||
11909 | rtx | |
11910 | replace_n_hard_rtx (rtx x, rtx *replacements, int n_replacements, int modify) | |
11911 | { | |
11912 | int i, j; | |
11913 | const char *fmt; | |
11914 | ||
11915 | /* The following prevents loops occurrence when we change MEM in | |
11916 | CONST_DOUBLE onto the same CONST_DOUBLE. */ | |
11917 | if (x != 0 && GET_CODE (x) == CONST_DOUBLE) | |
11918 | return x; | |
11919 | ||
11920 | for (i = n_replacements - 1; i >= 0 ; i--) | |
11921 | if (x == replacements[i*2] && GET_MODE (x) == GET_MODE (replacements[i*2+1])) | |
11922 | return replacements[i*2+1]; | |
11923 | ||
11924 | /* Allow this function to make replacements in EXPR_LISTs. */ | |
11925 | if (x == 0) | |
11926 | return 0; | |
11927 | ||
11928 | if (GET_CODE (x) == SUBREG) | |
11929 | { | |
0a2aaacc | 11930 | rtx new_rtx = replace_n_hard_rtx (SUBREG_REG (x), replacements, |
73a4d10b R |
11931 | n_replacements, modify); |
11932 | ||
f3536097 | 11933 | if (CONST_INT_P (new_rtx)) |
73a4d10b | 11934 | { |
0a2aaacc | 11935 | x = simplify_subreg (GET_MODE (x), new_rtx, |
73a4d10b R |
11936 | GET_MODE (SUBREG_REG (x)), |
11937 | SUBREG_BYTE (x)); | |
11938 | if (! x) | |
11939 | abort (); | |
11940 | } | |
11941 | else if (modify) | |
0a2aaacc | 11942 | SUBREG_REG (x) = new_rtx; |
73a4d10b R |
11943 | |
11944 | return x; | |
11945 | } | |
f3536097 | 11946 | else if (REG_P (x)) |
73a4d10b R |
11947 | { |
11948 | unsigned regno = REGNO (x); | |
11949 | unsigned nregs = (regno < FIRST_PSEUDO_REGISTER | |
11950 | ? HARD_REGNO_NREGS (regno, GET_MODE (x)) : 1); | |
11951 | rtx result = NULL_RTX; | |
11952 | ||
11953 | for (i = n_replacements - 1; i >= 0; i--) | |
11954 | { | |
11955 | rtx from = replacements[i*2]; | |
11956 | rtx to = replacements[i*2+1]; | |
11957 | unsigned from_regno, from_nregs, to_regno, new_regno; | |
11958 | ||
f3536097 | 11959 | if (!REG_P (from)) |
73a4d10b R |
11960 | continue; |
11961 | from_regno = REGNO (from); | |
11962 | from_nregs = (from_regno < FIRST_PSEUDO_REGISTER | |
11963 | ? HARD_REGNO_NREGS (from_regno, GET_MODE (from)) : 1); | |
11964 | if (regno < from_regno + from_nregs && regno + nregs > from_regno) | |
11965 | { | |
11966 | if (regno < from_regno | |
11967 | || regno + nregs > from_regno + nregs | |
f3536097 | 11968 | || !REG_P (to) |
73a4d10b R |
11969 | || result) |
11970 | return NULL_RTX; | |
11971 | to_regno = REGNO (to); | |
11972 | if (to_regno < FIRST_PSEUDO_REGISTER) | |
11973 | { | |
11974 | new_regno = regno + to_regno - from_regno; | |
11975 | if ((unsigned) HARD_REGNO_NREGS (new_regno, GET_MODE (x)) | |
11976 | != nregs) | |
11977 | return NULL_RTX; | |
11978 | result = gen_rtx_REG (GET_MODE (x), new_regno); | |
11979 | } | |
11980 | else if (GET_MODE (x) <= GET_MODE (to)) | |
11981 | result = gen_lowpart_common (GET_MODE (x), to); | |
11982 | else | |
11983 | result = gen_lowpart_SUBREG (GET_MODE (x), to); | |
11984 | } | |
11985 | } | |
11986 | return result ? result : x; | |
11987 | } | |
11988 | else if (GET_CODE (x) == ZERO_EXTEND) | |
11989 | { | |
0a2aaacc | 11990 | rtx new_rtx = replace_n_hard_rtx (XEXP (x, 0), replacements, |
73a4d10b R |
11991 | n_replacements, modify); |
11992 | ||
f3536097 | 11993 | if (CONST_INT_P (new_rtx)) |
73a4d10b R |
11994 | { |
11995 | x = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x), | |
0a2aaacc | 11996 | new_rtx, GET_MODE (XEXP (x, 0))); |
73a4d10b R |
11997 | if (! x) |
11998 | abort (); | |
11999 | } | |
12000 | else if (modify) | |
0a2aaacc | 12001 | XEXP (x, 0) = new_rtx; |
73a4d10b R |
12002 | |
12003 | return x; | |
12004 | } | |
12005 | ||
12006 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
12007 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
12008 | { | |
0a2aaacc | 12009 | rtx new_rtx; |
73a4d10b R |
12010 | |
12011 | if (fmt[i] == 'e') | |
12012 | { | |
0a2aaacc | 12013 | new_rtx = replace_n_hard_rtx (XEXP (x, i), replacements, |
73a4d10b | 12014 | n_replacements, modify); |
0a2aaacc | 12015 | if (!new_rtx) |
73a4d10b R |
12016 | return NULL_RTX; |
12017 | if (modify) | |
0a2aaacc | 12018 | XEXP (x, i) = new_rtx; |
73a4d10b R |
12019 | } |
12020 | else if (fmt[i] == 'E') | |
12021 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
12022 | { | |
0a2aaacc | 12023 | new_rtx = replace_n_hard_rtx (XVECEXP (x, i, j), replacements, |
73a4d10b | 12024 | n_replacements, modify); |
0a2aaacc | 12025 | if (!new_rtx) |
73a4d10b R |
12026 | return NULL_RTX; |
12027 | if (modify) | |
0a2aaacc | 12028 | XVECEXP (x, i, j) = new_rtx; |
73a4d10b R |
12029 | } |
12030 | } | |
12031 | ||
12032 | return x; | |
12033 | } | |
12034 | ||
12035 | rtx | |
12036 | sh_gen_truncate (enum machine_mode mode, rtx x, int need_sign_ext) | |
12037 | { | |
12038 | enum rtx_code code = TRUNCATE; | |
12039 | ||
12040 | if (GET_CODE (x) == ZERO_EXTEND || GET_CODE (x) == SIGN_EXTEND) | |
12041 | { | |
12042 | rtx inner = XEXP (x, 0); | |
12043 | enum machine_mode inner_mode = GET_MODE (inner); | |
12044 | ||
12045 | if (inner_mode == mode) | |
12046 | return inner; | |
12047 | else if (GET_MODE_SIZE (inner_mode) >= GET_MODE_SIZE (mode)) | |
12048 | x = inner; | |
12049 | else if (GET_MODE_SIZE (inner_mode) < GET_MODE_SIZE (mode) | |
12050 | && (! need_sign_ext || GET_CODE (x) == SIGN_EXTEND)) | |
12051 | { | |
12052 | code = GET_CODE (x); | |
12053 | x = inner; | |
12054 | } | |
12055 | } | |
12056 | return gen_rtx_fmt_e (code, mode, x); | |
12057 | } | |
12058 | ||
12059 | /* called via for_each_rtx after reload, to clean up truncates of | |
12060 | registers that span multiple actual hard registers. */ | |
12061 | int | |
12062 | shmedia_cleanup_truncate (rtx *p, void *n_changes) | |
12063 | { | |
12064 | rtx x = *p, reg; | |
12065 | ||
12066 | if (GET_CODE (x) != TRUNCATE) | |
12067 | return 0; | |
12068 | reg = XEXP (x, 0); | |
f3536097 | 12069 | if (GET_MODE_SIZE (GET_MODE (reg)) > 8 && REG_P (reg)) |
73a4d10b R |
12070 | { |
12071 | enum machine_mode reg_mode = GET_MODE (reg); | |
12072 | XEXP (x, 0) = simplify_subreg (DImode, reg, reg_mode, | |
12073 | subreg_lowpart_offset (DImode, reg_mode)); | |
12074 | *(int*) n_changes += 1; | |
12075 | return -1; | |
12076 | } | |
12077 | return 0; | |
12078 | } | |
12079 | ||
12080 | /* Load and store depend on the highpart of the address. However, | |
12081 | set_attr_alternative does not give well-defined results before reload, | |
12082 | so we must look at the rtl ourselves to see if any of the feeding | |
12083 | registers is used in a memref. */ | |
12084 | ||
12085 | /* Called by sh_contains_memref_p via for_each_rtx. */ | |
12086 | static int | |
12087 | sh_contains_memref_p_1 (rtx *loc, void *data ATTRIBUTE_UNUSED) | |
12088 | { | |
f3536097 | 12089 | return (MEM_P (*loc)); |
73a4d10b R |
12090 | } |
12091 | ||
a4d05547 | 12092 | /* Return nonzero iff INSN contains a MEM. */ |
73a4d10b R |
12093 | int |
12094 | sh_contains_memref_p (rtx insn) | |
12095 | { | |
12096 | return for_each_rtx (&PATTERN (insn), &sh_contains_memref_p_1, NULL); | |
12097 | } | |
12098 | ||
5c7cafa8 CB |
12099 | /* Return nonzero iff INSN loads a banked register. */ |
12100 | int | |
12101 | sh_loads_bankedreg_p (rtx insn) | |
12102 | { | |
12103 | if (GET_CODE (PATTERN (insn)) == SET) | |
12104 | { | |
12105 | rtx op = SET_DEST (PATTERN(insn)); | |
12106 | if (REG_P (op) && BANKED_REGISTER_P (REGNO (op))) | |
12107 | return 1; | |
12108 | } | |
12109 | ||
12110 | return 0; | |
12111 | } | |
12112 | ||
73a4d10b R |
12113 | /* FNADDR is the MEM expression from a call expander. Return an address |
12114 | to use in an SHmedia insn pattern. */ | |
12115 | rtx | |
12116 | shmedia_prepare_call_address (rtx fnaddr, int is_sibcall) | |
12117 | { | |
12118 | int is_sym; | |
12119 | ||
12120 | fnaddr = XEXP (fnaddr, 0); | |
12121 | is_sym = GET_CODE (fnaddr) == SYMBOL_REF; | |
12122 | if (flag_pic && is_sym) | |
12123 | { | |
12124 | if (! SYMBOL_REF_LOCAL_P (fnaddr)) | |
12125 | { | |
12126 | rtx reg = gen_reg_rtx (Pmode); | |
12127 | ||
12128 | /* We must not use GOTPLT for sibcalls, because PIC_REG | |
12129 | must be restored before the PLT code gets to run. */ | |
12130 | if (is_sibcall) | |
12131 | emit_insn (gen_symGOT2reg (reg, fnaddr)); | |
12132 | else | |
12133 | emit_insn (gen_symGOTPLT2reg (reg, fnaddr)); | |
12134 | fnaddr = reg; | |
12135 | } | |
12136 | else | |
12137 | { | |
12138 | fnaddr = gen_sym2PIC (fnaddr); | |
12139 | PUT_MODE (fnaddr, Pmode); | |
12140 | } | |
12141 | } | |
12142 | /* If ptabs might trap, make this visible to the rest of the compiler. | |
12143 | We generally assume that symbols pertain to valid locations, but | |
12144 | it is possible to generate invalid symbols with asm or linker tricks. | |
30dc60c7 | 12145 | In a list of functions where each returns its successor, an invalid |
73a4d10b R |
12146 | symbol might denote an empty list. */ |
12147 | if (!TARGET_PT_FIXED | |
12148 | && (!is_sym || TARGET_INVALID_SYMBOLS) | |
12149 | && (!REG_P (fnaddr) || ! TARGET_REGISTER_P (REGNO (fnaddr)))) | |
12150 | { | |
12151 | rtx tr = gen_reg_rtx (PDImode); | |
12152 | ||
12153 | emit_insn (gen_ptabs (tr, fnaddr)); | |
12154 | fnaddr = tr; | |
12155 | } | |
12156 | else if (! target_reg_operand (fnaddr, Pmode)) | |
12157 | fnaddr = copy_to_mode_reg (Pmode, fnaddr); | |
12158 | return fnaddr; | |
12159 | } | |
12160 | ||
a87cf97e JR |
12161 | reg_class_t |
12162 | sh_secondary_reload (bool in_p, rtx x, reg_class_t rclass_i, | |
8a99f6f9 R |
12163 | enum machine_mode mode, secondary_reload_info *sri) |
12164 | { | |
a87cf97e JR |
12165 | enum reg_class rclass = (enum reg_class) rclass_i; |
12166 | ||
8a99f6f9 R |
12167 | if (in_p) |
12168 | { | |
0a2aaacc | 12169 | if (REGCLASS_HAS_FP_REG (rclass) |
8a99f6f9 R |
12170 | && ! TARGET_SHMEDIA |
12171 | && immediate_operand ((x), mode) | |
12172 | && ! ((fp_zero_operand (x) || fp_one_operand (x)) | |
12173 | && mode == SFmode && fldi_ok ())) | |
12174 | switch (mode) | |
12175 | { | |
12176 | case SFmode: | |
12177 | sri->icode = CODE_FOR_reload_insf__frn; | |
12178 | return NO_REGS; | |
12179 | case DFmode: | |
12180 | sri->icode = CODE_FOR_reload_indf__frn; | |
12181 | return NO_REGS; | |
12182 | case SImode: | |
12183 | /* ??? If we knew that we are in the appropriate mode - | |
12184 | single precision - we could use a reload pattern directly. */ | |
12185 | return FPUL_REGS; | |
12186 | default: | |
12187 | abort (); | |
12188 | } | |
0a2aaacc | 12189 | if (rclass == FPUL_REGS |
f3536097 | 12190 | && ((REG_P (x) |
8a99f6f9 R |
12191 | && (REGNO (x) == MACL_REG || REGNO (x) == MACH_REG |
12192 | || REGNO (x) == T_REG)) | |
12193 | || GET_CODE (x) == PLUS)) | |
12194 | return GENERAL_REGS; | |
0a2aaacc | 12195 | if (rclass == FPUL_REGS && immediate_operand (x, mode)) |
8a99f6f9 | 12196 | { |
d973ac5a | 12197 | if (satisfies_constraint_I08 (x) || fp_zero_operand (x)) |
8a99f6f9 | 12198 | return GENERAL_REGS; |
d973ac5a KK |
12199 | else if (mode == SFmode) |
12200 | return FP_REGS; | |
8a99f6f9 R |
12201 | sri->icode = CODE_FOR_reload_insi__i_fpul; |
12202 | return NO_REGS; | |
12203 | } | |
0a2aaacc | 12204 | if (rclass == FPSCR_REGS |
f3536097 SZ |
12205 | && ((REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER) |
12206 | || (MEM_P (x) && GET_CODE (XEXP (x, 0)) == PLUS))) | |
8a99f6f9 | 12207 | return GENERAL_REGS; |
0a2aaacc | 12208 | if (REGCLASS_HAS_FP_REG (rclass) |
8a99f6f9 R |
12209 | && TARGET_SHMEDIA |
12210 | && immediate_operand (x, mode) | |
12211 | && x != CONST0_RTX (GET_MODE (x)) | |
12212 | && GET_MODE (x) != V4SFmode) | |
12213 | return GENERAL_REGS; | |
12214 | if ((mode == QImode || mode == HImode) | |
12215 | && TARGET_SHMEDIA && inqhi_operand (x, mode)) | |
12216 | { | |
12217 | sri->icode = ((mode == QImode) | |
12218 | ? CODE_FOR_reload_inqi : CODE_FOR_reload_inhi); | |
12219 | return NO_REGS; | |
12220 | } | |
0a2aaacc | 12221 | if (TARGET_SHMEDIA && rclass == GENERAL_REGS |
dc3ba671 | 12222 | && (GET_CODE (x) == LABEL_REF || PIC_ADDR_P (x))) |
8a99f6f9 R |
12223 | return TARGET_REGS; |
12224 | } /* end of input-only processing. */ | |
12225 | ||
0a2aaacc | 12226 | if (((REGCLASS_HAS_FP_REG (rclass) |
f3536097 | 12227 | && (REG_P (x) |
8a99f6f9 R |
12228 | && (GENERAL_OR_AP_REGISTER_P (REGNO (x)) |
12229 | || (FP_REGISTER_P (REGNO (x)) && mode == SImode | |
12230 | && TARGET_FMOVD)))) | |
0a2aaacc | 12231 | || (REGCLASS_HAS_GENERAL_REG (rclass) |
f3536097 | 12232 | && REG_P (x) |
8a99f6f9 R |
12233 | && FP_REGISTER_P (REGNO (x)))) |
12234 | && ! TARGET_SHMEDIA | |
12235 | && (mode == SFmode || mode == SImode)) | |
12236 | return FPUL_REGS; | |
0a2aaacc KG |
12237 | if ((rclass == FPUL_REGS |
12238 | || (REGCLASS_HAS_FP_REG (rclass) | |
8a99f6f9 | 12239 | && ! TARGET_SHMEDIA && mode == SImode)) |
f3536097 SZ |
12240 | && (MEM_P (x) |
12241 | || (REG_P (x) | |
8a99f6f9 R |
12242 | && (REGNO (x) >= FIRST_PSEUDO_REGISTER |
12243 | || REGNO (x) == T_REG | |
12244 | || system_reg_operand (x, VOIDmode))))) | |
12245 | { | |
0a2aaacc | 12246 | if (rclass == FPUL_REGS) |
8a99f6f9 R |
12247 | return GENERAL_REGS; |
12248 | return FPUL_REGS; | |
12249 | } | |
0a2aaacc KG |
12250 | if ((rclass == TARGET_REGS |
12251 | || (TARGET_SHMEDIA && rclass == SIBCALL_REGS)) | |
32a7ab3d | 12252 | && !satisfies_constraint_Csy (x) |
f3536097 | 12253 | && (!REG_P (x) || ! GENERAL_REGISTER_P (REGNO (x)))) |
8a99f6f9 | 12254 | return GENERAL_REGS; |
0a2aaacc | 12255 | if ((rclass == MAC_REGS || rclass == PR_REGS) |
f3536097 | 12256 | && REG_P (x) && ! GENERAL_REGISTER_P (REGNO (x)) |
0a2aaacc | 12257 | && rclass != REGNO_REG_CLASS (REGNO (x))) |
8a99f6f9 | 12258 | return GENERAL_REGS; |
f3536097 | 12259 | if (rclass != GENERAL_REGS && REG_P (x) |
8a99f6f9 R |
12260 | && TARGET_REGISTER_P (REGNO (x))) |
12261 | return GENERAL_REGS; | |
12262 | return NO_REGS; | |
12263 | } | |
12264 | ||
73a4d10b R |
12265 | enum sh_divide_strategy_e sh_div_strategy = SH_DIV_STRATEGY_DEFAULT; |
12266 | ||
e2500fed | 12267 | #include "gt-sh.h" |