]>
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, |
01e935c1 | 3 | 2003, 2004, 2005, 2006, 2007, 2008, 2009 Free Software Foundation, Inc. |
c5c76735 | 4 | Contributed by Steve Chamberlain (sac@cygnus.com). |
f676971a | 5 | Improved by Jim Wilson (wilson@cygnus.com). |
bc45ade3 | 6 | |
7ec022b2 | 7 | This file is part of GCC. |
c15c9075 | 8 | |
7ec022b2 | 9 | GCC is free software; you can redistribute it and/or modify |
c15c9075 | 10 | it under the terms of the GNU General Public License as published by |
2f83c7d6 | 11 | the Free Software Foundation; either version 3, or (at your option) |
c15c9075 RK |
12 | any later version. |
13 | ||
7ec022b2 | 14 | GCC is distributed in the hope that it will be useful, |
c15c9075 RK |
15 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
16 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
17 | GNU General Public License for more details. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
2f83c7d6 NC |
20 | along with GCC; see the file COPYING3. If not see |
21 | <http://www.gnu.org/licenses/>. */ | |
bc45ade3 | 22 | |
bc45ade3 | 23 | #include "config.h" |
c5c76735 | 24 | #include "system.h" |
4977bab6 ZW |
25 | #include "coretypes.h" |
26 | #include "tm.h" | |
318881c0 | 27 | #include "insn-config.h" |
bc45ade3 | 28 | #include "rtl.h" |
bc45ade3 | 29 | #include "tree.h" |
bc45ade3 | 30 | #include "flags.h" |
bc45ade3 | 31 | #include "expr.h" |
e78d8e51 | 32 | #include "optabs.h" |
49ad7cfa | 33 | #include "function.h" |
8aa2a305 JW |
34 | #include "regs.h" |
35 | #include "hard-reg-set.h" | |
36 | #include "output.h" | |
956a5206 | 37 | #include "insn-attr.h" |
318881c0 KG |
38 | #include "toplev.h" |
39 | #include "recog.h" | |
950a3816 | 40 | #include "integrate.h" |
bbf531f0 | 41 | #include "elf/dwarf2.h" |
318881c0 | 42 | #include "tm_p.h" |
672a6f42 NB |
43 | #include "target.h" |
44 | #include "target-def.h" | |
62526ec3 | 45 | #include "real.h" |
c1b92d09 | 46 | #include "langhooks.h" |
1aa03f38 | 47 | #include "basic-block.h" |
6fb5fa3c | 48 | #include "df.h" |
117dca74 | 49 | #include "cfglayout.h" |
bcc8cc82 | 50 | #include "intl.h" |
58565a33 | 51 | #include "sched-int.h" |
bd9a3465 | 52 | #include "params.h" |
8582b18a | 53 | #include "ggc.h" |
726a989a | 54 | #include "gimple.h" |
73a4d10b | 55 | #include "cfgloop.h" |
6dd3c0a5 | 56 | #include "alloc-pool.h" |
32a7ab3d | 57 | #include "tm-constrs.h" |
5fac2ed2 | 58 | |
bc45ade3 | 59 | |
1245df60 R |
60 | int code_for_indirect_jump_scratch = CODE_FOR_indirect_jump_scratch; |
61 | ||
00f8ff66 SC |
62 | #define MSW (TARGET_LITTLE_ENDIAN ? 1 : 0) |
63 | #define LSW (TARGET_LITTLE_ENDIAN ? 0 : 1) | |
64 | ||
fa5322fa AO |
65 | /* These are some macros to abstract register modes. */ |
66 | #define CONST_OK_FOR_ADD(size) \ | |
735cb76e | 67 | (TARGET_SHMEDIA ? CONST_OK_FOR_I10 (size) : CONST_OK_FOR_I08 (size)) |
fa5322fa AO |
68 | #define GEN_MOV (*(TARGET_SHMEDIA64 ? gen_movdi : gen_movsi)) |
69 | #define GEN_ADD3 (*(TARGET_SHMEDIA64 ? gen_adddi3 : gen_addsi3)) | |
70 | #define GEN_SUB3 (*(TARGET_SHMEDIA64 ? gen_subdi3 : gen_subsi3)) | |
71 | ||
561642fa AP |
72 | /* Used to simplify the logic below. Find the attributes wherever |
73 | they may be. */ | |
74 | #define SH_ATTRIBUTES(decl) \ | |
75 | (TYPE_P (decl)) ? TYPE_ATTRIBUTES (decl) \ | |
76 | : DECL_ATTRIBUTES (decl) \ | |
77 | ? (DECL_ATTRIBUTES (decl)) \ | |
78 | : TYPE_ATTRIBUTES (TREE_TYPE (decl)) | |
79 | ||
552ecbd9 AH |
80 | /* Set to 1 by expand_prologue() when the function is an interrupt handler. */ |
81 | int current_function_interrupt; | |
82 | ||
a6ab9fc0 R |
83 | tree sh_deferred_function_attributes; |
84 | tree *sh_deferred_function_attributes_tail = &sh_deferred_function_attributes; | |
16bea517 | 85 | |
6f317ef3 | 86 | /* Global variables for machine-dependent things. */ |
bc45ade3 | 87 | |
16bea517 JW |
88 | /* Which cpu are we scheduling for. */ |
89 | enum processor_type sh_cpu; | |
90 | ||
58565a33 SKG |
91 | /* Definitions used in ready queue reordering for first scheduling pass. */ |
92 | ||
93 | /* Reg weights arrays for modes SFmode and SImode, indexed by insn LUID. */ | |
94 | static short *regmode_weight[2]; | |
95 | ||
96 | /* Total SFmode and SImode weights of scheduled insns. */ | |
97 | static int curr_regmode_pressure[2]; | |
98 | ||
6fb5fa3c DB |
99 | /* Number of r0 life regions. */ |
100 | static int r0_life_regions; | |
101 | ||
58565a33 SKG |
102 | /* If true, skip cycles for Q -> R movement. */ |
103 | static int skip_cycles = 0; | |
104 | ||
105 | /* Cached value of can_issue_more. This is cached in sh_variable_issue hook | |
106 | and returned from sh_reorder2. */ | |
107 | static short cached_can_issue_more; | |
108 | ||
bc45ade3 | 109 | /* Provides the class number of the smallest class containing |
16bea517 | 110 | reg number. */ |
bc45ade3 | 111 | |
fada1961 | 112 | enum reg_class regno_reg_class[FIRST_PSEUDO_REGISTER] = |
bc45ade3 | 113 | { |
8e87e161 | 114 | R0_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, |
bc45ade3 SC |
115 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, |
116 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
117 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
fa5322fa AO |
118 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, |
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, | |
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, | |
1a95a963 JW |
130 | FP0_REGS,FP_REGS, FP_REGS, FP_REGS, |
131 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
132 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
133 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
fa5322fa AO |
134 | FP_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, | |
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 | TARGET_REGS, TARGET_REGS, TARGET_REGS, TARGET_REGS, | |
147 | TARGET_REGS, TARGET_REGS, TARGET_REGS, TARGET_REGS, | |
225e4f43 R |
148 | DF_REGS, DF_REGS, DF_REGS, DF_REGS, |
149 | DF_REGS, DF_REGS, DF_REGS, DF_REGS, | |
fa5322fa AO |
150 | NO_REGS, GENERAL_REGS, PR_REGS, T_REGS, |
151 | MAC_REGS, MAC_REGS, FPUL_REGS, FPSCR_REGS, | |
96a2347e | 152 | GENERAL_REGS, GENERAL_REGS, |
225e4f43 R |
153 | }; |
154 | ||
fa5322fa AO |
155 | char sh_register_names[FIRST_PSEUDO_REGISTER] \ |
156 | [MAX_REGISTER_NAME_LENGTH + 1] = SH_REGISTER_NAMES_INITIALIZER; | |
157 | ||
158 | char sh_additional_register_names[ADDREGNAMES_SIZE] \ | |
159 | [MAX_ADDITIONAL_REGISTER_NAME_LENGTH + 1] | |
160 | = SH_ADDITIONAL_REGISTER_NAMES_INITIALIZER; | |
bc45ade3 | 161 | |
225e4f43 R |
162 | int assembler_dialect; |
163 | ||
fe3ad572 SC |
164 | static bool shmedia_space_reserved_for_target_registers; |
165 | ||
c0fb94d7 | 166 | static bool sh_handle_option (size_t, const char *, int); |
cf277499 SB |
167 | static void split_branches (rtx); |
168 | static int branch_dest (rtx); | |
169 | static void force_into (rtx, rtx); | |
170 | static void print_slot (rtx); | |
171 | static rtx add_constant (rtx, enum machine_mode, rtx); | |
078c8b08 | 172 | static void dump_table (rtx, rtx); |
cf277499 SB |
173 | static int hi_const (rtx); |
174 | static int broken_move (rtx); | |
175 | static int mova_p (rtx); | |
176 | static rtx find_barrier (int, rtx, rtx); | |
177 | static int noncall_uses_reg (rtx, rtx, rtx *); | |
178 | static rtx gen_block_redirect (rtx, int, int); | |
179 | static void sh_reorg (void); | |
fada1961 | 180 | static void output_stack_adjust (int, rtx, int, HARD_REG_SET *); |
cf277499 SB |
181 | static rtx frame_insn (rtx); |
182 | static rtx push (int); | |
183 | static void pop (int); | |
184 | static void push_regs (HARD_REG_SET *, int); | |
185 | static int calc_live_regs (HARD_REG_SET *); | |
cf277499 SB |
186 | static HOST_WIDE_INT rounded_frame_size (int); |
187 | static rtx mark_constant_pool_use (rtx); | |
cf277499 | 188 | static tree sh_handle_interrupt_handler_attribute (tree *, tree, tree, int, bool *); |
561642fa AP |
189 | static tree sh_handle_resbank_handler_attribute (tree *, tree, |
190 | tree, int, bool *); | |
191 | static tree sh2a_handle_function_vector_handler_attribute (tree *, tree, | |
192 | tree, int, bool *); | |
cf277499 SB |
193 | static tree sh_handle_sp_switch_attribute (tree *, tree, tree, int, bool *); |
194 | static tree sh_handle_trap_exit_attribute (tree *, tree, tree, int, bool *); | |
195 | static tree sh_handle_renesas_attribute (tree *, tree, tree, int, bool *); | |
196 | static void sh_output_function_epilogue (FILE *, HOST_WIDE_INT); | |
197 | static void sh_insert_attributes (tree, tree *); | |
8d932be3 | 198 | static const char *sh_check_pch_target_flags (int); |
cf277499 | 199 | static int sh_adjust_cost (rtx, rtx, rtx, int); |
cf277499 | 200 | static int sh_issue_rate (void); |
58565a33 SKG |
201 | static int sh_dfa_new_cycle (FILE *, int, rtx, int, int, int *sort_p); |
202 | static short find_set_regmode_weight (rtx, enum machine_mode); | |
203 | static short find_insn_regmode_weight (rtx, enum machine_mode); | |
c6d56443 | 204 | static void find_regmode_weight (basic_block, enum machine_mode); |
6fb5fa3c | 205 | static int find_r0_life_regions (basic_block); |
58565a33 SKG |
206 | static void sh_md_init_global (FILE *, int, int); |
207 | static void sh_md_finish_global (FILE *, int); | |
208 | static int rank_for_reorder (const void *, const void *); | |
209 | static void swap_reorder (rtx *, int); | |
210 | static void ready_reorder (rtx *, int); | |
211 | static short high_pressure (enum machine_mode); | |
212 | static int sh_reorder (FILE *, int, rtx *, int *, int); | |
213 | static int sh_reorder2 (FILE *, int, rtx *, int *, int); | |
214 | static void sh_md_init (FILE *, int, int); | |
215 | static int sh_variable_issue (FILE *, int, rtx, int); | |
f676971a | 216 | |
cf277499 SB |
217 | static bool sh_function_ok_for_sibcall (tree, tree); |
218 | ||
219 | static bool sh_cannot_modify_jumps_p (void); | |
81f40b79 | 220 | static enum reg_class sh_target_reg_class (void); |
fe3ad572 | 221 | static bool sh_optimize_target_register_callee_saved (bool); |
3101faab | 222 | static bool sh_ms_bitfield_layout_p (const_tree); |
cf277499 SB |
223 | |
224 | static void sh_init_builtins (void); | |
225 | static void sh_media_init_builtins (void); | |
226 | static rtx sh_expand_builtin (tree, rtx, rtx, enum machine_mode, int); | |
227 | static void sh_output_mi_thunk (FILE *, tree, HOST_WIDE_INT, HOST_WIDE_INT, tree); | |
228 | static void sh_file_start (void); | |
229 | static int flow_dependent_p (rtx, rtx); | |
7bc980e1 | 230 | static void flow_dependent_p_1 (rtx, const_rtx, void *); |
cf277499 SB |
231 | static int shiftcosts (rtx); |
232 | static int andcosts (rtx); | |
233 | static int addsubcosts (rtx); | |
234 | static int multcosts (rtx); | |
235 | static bool unspec_caller_rtx_p (rtx); | |
236 | static bool sh_cannot_copy_insn_p (rtx); | |
f40751dd JH |
237 | static bool sh_rtx_costs (rtx, int, int, int *, bool); |
238 | static int sh_address_cost (rtx, bool); | |
b48f503c KK |
239 | static int sh_pr_n_sets (void); |
240 | static rtx sh_allocate_initial_value (rtx); | |
1bb99877 | 241 | static bool sh_legitimate_address_p (enum machine_mode, rtx, bool); |
506d7b68 | 242 | static rtx sh_legitimize_address (rtx, rtx, enum machine_mode); |
fe3ad572 SC |
243 | static int shmedia_target_regs_stack_space (HARD_REG_SET *); |
244 | static int shmedia_reserve_space_for_target_registers_p (int, HARD_REG_SET *); | |
245 | static int shmedia_target_regs_stack_adjust (HARD_REG_SET *); | |
fada1961 R |
246 | static int scavenge_reg (HARD_REG_SET *s); |
247 | struct save_schedule_s; | |
248 | static struct save_entry_s *sh5_schedule_saves (HARD_REG_SET *, | |
249 | struct save_schedule_s *, int); | |
61f71b34 | 250 | |
cf277499 | 251 | static rtx sh_struct_value_rtx (tree, int); |
586de218 | 252 | static bool sh_return_in_memory (const_tree, const_tree); |
cf277499 SB |
253 | static rtx sh_builtin_saveregs (void); |
254 | static void sh_setup_incoming_varargs (CUMULATIVE_ARGS *, enum machine_mode, tree, int *, int); | |
255 | static bool sh_strict_argument_naming (CUMULATIVE_ARGS *); | |
256 | static bool sh_pretend_outgoing_varargs_named (CUMULATIVE_ARGS *); | |
c35d187f | 257 | static tree sh_build_builtin_va_list (void); |
d7bd8aeb | 258 | static void sh_va_start (tree, rtx); |
726a989a | 259 | static tree sh_gimplify_va_arg_expr (tree, tree, gimple_seq *, gimple_seq *); |
8cd5a4e0 | 260 | static bool sh_pass_by_reference (CUMULATIVE_ARGS *, enum machine_mode, |
586de218 | 261 | const_tree, bool); |
6cdd5672 | 262 | static bool sh_callee_copies (CUMULATIVE_ARGS *, enum machine_mode, |
586de218 | 263 | const_tree, bool); |
78a52f11 RH |
264 | static int sh_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode, |
265 | tree, bool); | |
b0fbea13 | 266 | static bool sh_scalar_mode_supported_p (enum machine_mode); |
3101faab | 267 | static int sh_dwarf_calling_convention (const_tree); |
561642fa AP |
268 | static void sh_encode_section_info (tree, rtx, int); |
269 | static int sh2a_function_vector_p (tree); | |
5a82ecd9 ILT |
270 | \f |
271 | static const struct attribute_spec sh_attribute_table[] = | |
272 | { | |
273 | /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */ | |
274 | { "interrupt_handler", 0, 0, true, false, false, sh_handle_interrupt_handler_attribute }, | |
275 | { "sp_switch", 1, 1, true, false, false, sh_handle_sp_switch_attribute }, | |
276 | { "trap_exit", 1, 1, true, false, false, sh_handle_trap_exit_attribute }, | |
277 | { "renesas", 0, 0, false, true, false, sh_handle_renesas_attribute }, | |
278 | { "trapa_handler", 0, 0, true, false, false, sh_handle_interrupt_handler_attribute }, | |
279 | { "nosave_low_regs", 0, 0, true, false, false, sh_handle_interrupt_handler_attribute }, | |
280 | { "resbank", 0, 0, true, false, false, sh_handle_resbank_handler_attribute }, | |
281 | { "function_vector", 1, 1, true, false, false, sh2a_handle_function_vector_handler_attribute }, | |
282 | #ifdef SYMBIAN | |
283 | /* Symbian support adds three new attributes: | |
284 | dllexport - for exporting a function/variable that will live in a dll | |
285 | dllimport - for importing a function/variable from a dll | |
61f71b34 | 286 | |
5a82ecd9 ILT |
287 | Microsoft allows multiple declspecs in one __declspec, separating |
288 | them with spaces. We do NOT support this. Instead, use __declspec | |
289 | multiple times. */ | |
290 | { "dllimport", 0, 0, true, false, false, sh_symbian_handle_dll_attribute }, | |
291 | { "dllexport", 0, 0, true, false, false, sh_symbian_handle_dll_attribute }, | |
292 | #endif | |
293 | { NULL, 0, 0, false, false, false, NULL } | |
294 | }; | |
672a6f42 NB |
295 | \f |
296 | /* Initialize the GCC target structure. */ | |
91d231cb JM |
297 | #undef TARGET_ATTRIBUTE_TABLE |
298 | #define TARGET_ATTRIBUTE_TABLE sh_attribute_table | |
672a6f42 | 299 | |
301d03af RS |
300 | /* The next two are used for debug info when compiling with -gdwarf. */ |
301 | #undef TARGET_ASM_UNALIGNED_HI_OP | |
302 | #define TARGET_ASM_UNALIGNED_HI_OP "\t.uaword\t" | |
303 | #undef TARGET_ASM_UNALIGNED_SI_OP | |
304 | #define TARGET_ASM_UNALIGNED_SI_OP "\t.ualong\t" | |
305 | ||
fa5322fa AO |
306 | /* These are NULLed out on non-SH5 in OVERRIDE_OPTIONS. */ |
307 | #undef TARGET_ASM_UNALIGNED_DI_OP | |
308 | #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaquad\t" | |
309 | #undef TARGET_ASM_ALIGNED_DI_OP | |
310 | #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t" | |
311 | ||
08c148a8 NB |
312 | #undef TARGET_ASM_FUNCTION_EPILOGUE |
313 | #define TARGET_ASM_FUNCTION_EPILOGUE sh_output_function_epilogue | |
314 | ||
1aa03f38 R |
315 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
316 | #define TARGET_ASM_OUTPUT_MI_THUNK sh_output_mi_thunk | |
317 | ||
318 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
3101faab | 319 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true |
1aa03f38 | 320 | |
1bc7c5b6 ZW |
321 | #undef TARGET_ASM_FILE_START |
322 | #define TARGET_ASM_FILE_START sh_file_start | |
323 | #undef TARGET_ASM_FILE_START_FILE_DIRECTIVE | |
324 | #define TARGET_ASM_FILE_START_FILE_DIRECTIVE true | |
325 | ||
c0fb94d7 RS |
326 | #undef TARGET_DEFAULT_TARGET_FLAGS |
327 | #define TARGET_DEFAULT_TARGET_FLAGS TARGET_DEFAULT | |
328 | #undef TARGET_HANDLE_OPTION | |
329 | #define TARGET_HANDLE_OPTION sh_handle_option | |
330 | ||
12a68f1f JM |
331 | #undef TARGET_INSERT_ATTRIBUTES |
332 | #define TARGET_INSERT_ATTRIBUTES sh_insert_attributes | |
333 | ||
c237e94a ZW |
334 | #undef TARGET_SCHED_ADJUST_COST |
335 | #define TARGET_SCHED_ADJUST_COST sh_adjust_cost | |
336 | ||
fae15c93 VM |
337 | #undef TARGET_SCHED_ISSUE_RATE |
338 | #define TARGET_SCHED_ISSUE_RATE sh_issue_rate | |
339 | ||
58565a33 SKG |
340 | /* The next 5 hooks have been implemented for reenabling sched1. With the |
341 | help of these macros we are limiting the movement of insns in sched1 to | |
f676971a | 342 | reduce the register pressure. The overall idea is to keep count of SImode |
58565a33 SKG |
343 | and SFmode regs required by already scheduled insns. When these counts |
344 | cross some threshold values; give priority to insns that free registers. | |
345 | The insn that frees registers is most likely to be the insn with lowest | |
f676971a | 346 | LUID (original insn order); but such an insn might be there in the stalled |
58565a33 SKG |
347 | queue (Q) instead of the ready queue (R). To solve this, we skip cycles |
348 | upto a max of 8 cycles so that such insns may move from Q -> R. | |
349 | ||
350 | The description of the hooks are as below: | |
351 | ||
352 | TARGET_SCHED_INIT_GLOBAL: Added a new target hook in the generic | |
353 | scheduler; it is called inside the sched_init function just after | |
354 | find_insn_reg_weights function call. It is used to calculate the SImode | |
1ae58c30 | 355 | and SFmode weights of insns of basic blocks; much similar to what |
f676971a | 356 | find_insn_reg_weights does. |
58565a33 SKG |
357 | TARGET_SCHED_FINISH_GLOBAL: Corresponding cleanup hook. |
358 | ||
359 | TARGET_SCHED_DFA_NEW_CYCLE: Skip cycles if high register pressure is | |
360 | indicated by TARGET_SCHED_REORDER2; doing this may move insns from | |
361 | (Q)->(R). | |
362 | ||
363 | TARGET_SCHED_REORDER: If the register pressure for SImode or SFmode is | |
364 | high; reorder the ready queue so that the insn with lowest LUID will be | |
365 | issued next. | |
366 | ||
367 | TARGET_SCHED_REORDER2: If the register pressure is high, indicate to | |
368 | TARGET_SCHED_DFA_NEW_CYCLE to skip cycles. | |
369 | ||
370 | TARGET_SCHED_VARIABLE_ISSUE: Cache the value of can_issue_more so that it | |
371 | can be returned from TARGET_SCHED_REORDER2. | |
372 | ||
373 | TARGET_SCHED_INIT: Reset the register pressure counting variables. */ | |
374 | ||
375 | #undef TARGET_SCHED_DFA_NEW_CYCLE | |
376 | #define TARGET_SCHED_DFA_NEW_CYCLE sh_dfa_new_cycle | |
377 | ||
378 | #undef TARGET_SCHED_INIT_GLOBAL | |
379 | #define TARGET_SCHED_INIT_GLOBAL sh_md_init_global | |
380 | ||
381 | #undef TARGET_SCHED_FINISH_GLOBAL | |
382 | #define TARGET_SCHED_FINISH_GLOBAL sh_md_finish_global | |
383 | ||
384 | #undef TARGET_SCHED_VARIABLE_ISSUE | |
385 | #define TARGET_SCHED_VARIABLE_ISSUE sh_variable_issue | |
386 | ||
387 | #undef TARGET_SCHED_REORDER | |
388 | #define TARGET_SCHED_REORDER sh_reorder | |
389 | ||
390 | #undef TARGET_SCHED_REORDER2 | |
391 | #define TARGET_SCHED_REORDER2 sh_reorder2 | |
392 | ||
393 | #undef TARGET_SCHED_INIT | |
394 | #define TARGET_SCHED_INIT sh_md_init | |
395 | ||
506d7b68 PB |
396 | #undef TARGET_LEGITIMIZE_ADDRESS |
397 | #define TARGET_LEGITIMIZE_ADDRESS sh_legitimize_address | |
398 | ||
fa5322fa AO |
399 | #undef TARGET_CANNOT_MODIFY_JUMPS_P |
400 | #define TARGET_CANNOT_MODIFY_JUMPS_P sh_cannot_modify_jumps_p | |
fe3ad572 SC |
401 | #undef TARGET_BRANCH_TARGET_REGISTER_CLASS |
402 | #define TARGET_BRANCH_TARGET_REGISTER_CLASS sh_target_reg_class | |
403 | #undef TARGET_BRANCH_TARGET_REGISTER_CALLEE_SAVED | |
404 | #define TARGET_BRANCH_TARGET_REGISTER_CALLEE_SAVED \ | |
405 | sh_optimize_target_register_callee_saved | |
fa5322fa AO |
406 | |
407 | #undef TARGET_MS_BITFIELD_LAYOUT_P | |
408 | #define TARGET_MS_BITFIELD_LAYOUT_P sh_ms_bitfield_layout_p | |
409 | ||
c1b92d09 R |
410 | #undef TARGET_INIT_BUILTINS |
411 | #define TARGET_INIT_BUILTINS sh_init_builtins | |
412 | #undef TARGET_EXPAND_BUILTIN | |
413 | #define TARGET_EXPAND_BUILTIN sh_expand_builtin | |
414 | ||
4977bab6 ZW |
415 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL |
416 | #define TARGET_FUNCTION_OK_FOR_SIBCALL sh_function_ok_for_sibcall | |
417 | ||
9f3a9a08 KK |
418 | #undef TARGET_CANNOT_COPY_INSN_P |
419 | #define TARGET_CANNOT_COPY_INSN_P sh_cannot_copy_insn_p | |
3c50106f RH |
420 | #undef TARGET_RTX_COSTS |
421 | #define TARGET_RTX_COSTS sh_rtx_costs | |
dcefdf67 RH |
422 | #undef TARGET_ADDRESS_COST |
423 | #define TARGET_ADDRESS_COST sh_address_cost | |
b48f503c KK |
424 | #undef TARGET_ALLOCATE_INITIAL_VALUE |
425 | #define TARGET_ALLOCATE_INITIAL_VALUE sh_allocate_initial_value | |
3c50106f | 426 | |
18dbd950 RS |
427 | #undef TARGET_MACHINE_DEPENDENT_REORG |
428 | #define TARGET_MACHINE_DEPENDENT_REORG sh_reorg | |
429 | ||
63a9de54 CB |
430 | #undef TARGET_DWARF_REGISTER_SPAN |
431 | #define TARGET_DWARF_REGISTER_SPAN sh_dwarf_register_span | |
432 | ||
463f02cd KK |
433 | #ifdef HAVE_AS_TLS |
434 | #undef TARGET_HAVE_TLS | |
435 | #define TARGET_HAVE_TLS true | |
436 | #endif | |
437 | ||
61f71b34 DD |
438 | #undef TARGET_PROMOTE_PROTOTYPES |
439 | #define TARGET_PROMOTE_PROTOTYPES sh_promote_prototypes | |
440 | #undef TARGET_PROMOTE_FUNCTION_ARGS | |
441 | #define TARGET_PROMOTE_FUNCTION_ARGS sh_promote_prototypes | |
442 | #undef TARGET_PROMOTE_FUNCTION_RETURN | |
443 | #define TARGET_PROMOTE_FUNCTION_RETURN sh_promote_prototypes | |
444 | ||
445 | #undef TARGET_STRUCT_VALUE_RTX | |
446 | #define TARGET_STRUCT_VALUE_RTX sh_struct_value_rtx | |
447 | #undef TARGET_RETURN_IN_MEMORY | |
448 | #define TARGET_RETURN_IN_MEMORY sh_return_in_memory | |
449 | ||
450 | #undef TARGET_EXPAND_BUILTIN_SAVEREGS | |
451 | #define TARGET_EXPAND_BUILTIN_SAVEREGS sh_builtin_saveregs | |
452 | #undef TARGET_SETUP_INCOMING_VARARGS | |
453 | #define TARGET_SETUP_INCOMING_VARARGS sh_setup_incoming_varargs | |
454 | #undef TARGET_STRICT_ARGUMENT_NAMING | |
455 | #define TARGET_STRICT_ARGUMENT_NAMING sh_strict_argument_naming | |
456 | #undef TARGET_PRETEND_OUTGOING_VARARGS_NAMED | |
457 | #define TARGET_PRETEND_OUTGOING_VARARGS_NAMED sh_pretend_outgoing_varargs_named | |
fe984136 RH |
458 | #undef TARGET_MUST_PASS_IN_STACK |
459 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
8cd5a4e0 RH |
460 | #undef TARGET_PASS_BY_REFERENCE |
461 | #define TARGET_PASS_BY_REFERENCE sh_pass_by_reference | |
6cdd5672 RH |
462 | #undef TARGET_CALLEE_COPIES |
463 | #define TARGET_CALLEE_COPIES sh_callee_copies | |
78a52f11 RH |
464 | #undef TARGET_ARG_PARTIAL_BYTES |
465 | #define TARGET_ARG_PARTIAL_BYTES sh_arg_partial_bytes | |
61f71b34 | 466 | |
c35d187f RH |
467 | #undef TARGET_BUILD_BUILTIN_VA_LIST |
468 | #define TARGET_BUILD_BUILTIN_VA_LIST sh_build_builtin_va_list | |
d7bd8aeb JJ |
469 | #undef TARGET_EXPAND_BUILTIN_VA_START |
470 | #define TARGET_EXPAND_BUILTIN_VA_START sh_va_start | |
5fac2ed2 RH |
471 | #undef TARGET_GIMPLIFY_VA_ARG_EXPR |
472 | #define TARGET_GIMPLIFY_VA_ARG_EXPR sh_gimplify_va_arg_expr | |
c35d187f | 473 | |
b0fbea13 KK |
474 | #undef TARGET_SCALAR_MODE_SUPPORTED_P |
475 | #define TARGET_SCALAR_MODE_SUPPORTED_P sh_scalar_mode_supported_p | |
f676971a EC |
476 | #undef TARGET_VECTOR_MODE_SUPPORTED_P |
477 | #define TARGET_VECTOR_MODE_SUPPORTED_P sh_vector_mode_supported_p | |
478 | ||
8d932be3 RS |
479 | #undef TARGET_CHECK_PCH_TARGET_FLAGS |
480 | #define TARGET_CHECK_PCH_TARGET_FLAGS sh_check_pch_target_flags | |
bcc8cc82 | 481 | |
a1c496cb EC |
482 | #undef TARGET_DWARF_CALLING_CONVENTION |
483 | #define TARGET_DWARF_CALLING_CONVENTION sh_dwarf_calling_convention | |
484 | ||
58565a33 SKG |
485 | /* Return regmode weight for insn. */ |
486 | #define INSN_REGMODE_WEIGHT(INSN, MODE) regmode_weight[((MODE) == SImode) ? 0 : 1][INSN_UID (INSN)] | |
487 | ||
488 | /* Return current register pressure for regmode. */ | |
489 | #define CURR_REGMODE_PRESSURE(MODE) curr_regmode_pressure[((MODE) == SImode) ? 0 : 1] | |
490 | ||
561642fa AP |
491 | #undef TARGET_ENCODE_SECTION_INFO |
492 | #define TARGET_ENCODE_SECTION_INFO sh_encode_section_info | |
493 | ||
b55e3aad NC |
494 | #ifdef SYMBIAN |
495 | ||
496 | #undef TARGET_ENCODE_SECTION_INFO | |
497 | #define TARGET_ENCODE_SECTION_INFO sh_symbian_encode_section_info | |
498 | #undef TARGET_STRIP_NAME_ENCODING | |
499 | #define TARGET_STRIP_NAME_ENCODING sh_symbian_strip_name_encoding | |
500 | #undef TARGET_CXX_IMPORT_EXPORT_CLASS | |
501 | #define TARGET_CXX_IMPORT_EXPORT_CLASS symbian_import_export_class | |
502 | ||
503 | #endif /* SYMBIAN */ | |
504 | ||
8a99f6f9 R |
505 | #undef TARGET_SECONDARY_RELOAD |
506 | #define TARGET_SECONDARY_RELOAD sh_secondary_reload | |
507 | ||
1bb99877 KK |
508 | #undef TARGET_LEGITIMATE_ADDRESS_P |
509 | #define TARGET_LEGITIMATE_ADDRESS_P sh_legitimate_address_p | |
510 | ||
561642fa AP |
511 | /* Machine-specific symbol_ref flags. */ |
512 | #define SYMBOL_FLAG_FUNCVEC_FUNCTION (SYMBOL_FLAG_MACH_DEP << 0) | |
513 | ||
f6897b10 | 514 | struct gcc_target targetm = TARGET_INITIALIZER; |
b9654711 | 515 | \f |
c0fb94d7 RS |
516 | /* Implement TARGET_HANDLE_OPTION. */ |
517 | ||
518 | static bool | |
519 | sh_handle_option (size_t code, const char *arg ATTRIBUTE_UNUSED, | |
520 | int value ATTRIBUTE_UNUSED) | |
521 | { | |
522 | switch (code) | |
523 | { | |
524 | case OPT_m1: | |
525 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH1; | |
526 | return true; | |
527 | ||
528 | case OPT_m2: | |
529 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2; | |
530 | return true; | |
531 | ||
532 | case OPT_m2a: | |
533 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A; | |
534 | return true; | |
535 | ||
536 | case OPT_m2a_nofpu: | |
537 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_NOFPU; | |
538 | return true; | |
539 | ||
540 | case OPT_m2a_single: | |
541 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_SINGLE; | |
542 | return true; | |
543 | ||
544 | case OPT_m2a_single_only: | |
545 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_SINGLE_ONLY; | |
546 | return true; | |
547 | ||
548 | case OPT_m2e: | |
549 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2E; | |
550 | return true; | |
551 | ||
552 | case OPT_m3: | |
553 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH3; | |
554 | return true; | |
555 | ||
556 | case OPT_m3e: | |
557 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH3E; | |
558 | return true; | |
559 | ||
560 | case OPT_m4: | |
25e651ca AS |
561 | case OPT_m4_100: |
562 | case OPT_m4_200: | |
78d310c2 | 563 | case OPT_m4_300: |
c0fb94d7 RS |
564 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4; |
565 | return true; | |
566 | ||
567 | case OPT_m4_nofpu: | |
78d310c2 R |
568 | case OPT_m4_100_nofpu: |
569 | case OPT_m4_200_nofpu: | |
570 | case OPT_m4_300_nofpu: | |
571 | case OPT_m4_340: | |
25e651ca AS |
572 | case OPT_m4_400: |
573 | case OPT_m4_500: | |
c0fb94d7 RS |
574 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_NOFPU; |
575 | return true; | |
576 | ||
577 | case OPT_m4_single: | |
25e651ca AS |
578 | case OPT_m4_100_single: |
579 | case OPT_m4_200_single: | |
78d310c2 | 580 | case OPT_m4_300_single: |
c0fb94d7 RS |
581 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_SINGLE; |
582 | return true; | |
583 | ||
584 | case OPT_m4_single_only: | |
25e651ca AS |
585 | case OPT_m4_100_single_only: |
586 | case OPT_m4_200_single_only: | |
78d310c2 | 587 | case OPT_m4_300_single_only: |
c0fb94d7 RS |
588 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_SINGLE_ONLY; |
589 | return true; | |
590 | ||
591 | case OPT_m4a: | |
592 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A; | |
593 | return true; | |
594 | ||
595 | case OPT_m4a_nofpu: | |
596 | case OPT_m4al: | |
597 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_NOFPU; | |
598 | return true; | |
599 | ||
600 | case OPT_m4a_single: | |
601 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_SINGLE; | |
602 | return true; | |
603 | ||
604 | case OPT_m4a_single_only: | |
605 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_SINGLE_ONLY; | |
606 | return true; | |
607 | ||
608 | case OPT_m5_32media: | |
609 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_32MEDIA; | |
610 | return true; | |
611 | ||
612 | case OPT_m5_32media_nofpu: | |
613 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_32MEDIA_NOFPU; | |
614 | return true; | |
615 | ||
616 | case OPT_m5_64media: | |
617 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_64MEDIA; | |
618 | return true; | |
619 | ||
620 | case OPT_m5_64media_nofpu: | |
621 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_64MEDIA_NOFPU; | |
622 | return true; | |
623 | ||
624 | case OPT_m5_compact: | |
625 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_COMPACT; | |
626 | return true; | |
627 | ||
628 | case OPT_m5_compact_nofpu: | |
629 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_COMPACT_NOFPU; | |
630 | return true; | |
631 | ||
632 | default: | |
633 | return true; | |
634 | } | |
635 | } | |
636 | \f | |
bd9a3465 KK |
637 | /* Set default optimization options. */ |
638 | void | |
639 | sh_optimization_options (int level ATTRIBUTE_UNUSED, int size ATTRIBUTE_UNUSED) | |
640 | { | |
641 | if (level) | |
642 | { | |
643 | flag_omit_frame_pointer = 2; | |
644 | if (!size) | |
645 | sh_div_str = "inv:minlat"; | |
646 | } | |
647 | if (size) | |
648 | { | |
649 | target_flags |= MASK_SMALLCODE; | |
650 | sh_div_str = SH_DIV_STR_FOR_SIZE ; | |
651 | } | |
652 | else | |
653 | TARGET_CBRANCHDI4 = 1; | |
654 | /* We can't meaningfully test TARGET_SHMEDIA here, because -m options | |
655 | haven't been parsed yet, hence we'd read only the default. | |
656 | sh_target_reg_class will return NO_REGS if this is not SHMEDIA, so | |
657 | it's OK to always set flag_branch_target_load_optimize. */ | |
658 | if (level > 1) | |
659 | { | |
660 | flag_branch_target_load_optimize = 1; | |
661 | if (!size) | |
662 | target_flags |= MASK_SAVE_ALL_TARGET_REGS; | |
663 | } | |
664 | /* Likewise, we can't meaningfully test TARGET_SH2E / TARGET_IEEE | |
665 | here, so leave it to OVERRIDE_OPTIONS to set | |
666 | flag_finite_math_only. We set it to 2 here so we know if the user | |
667 | explicitly requested this to be on or off. */ | |
668 | flag_finite_math_only = 2; | |
669 | /* If flag_schedule_insns is 1, we set it to 2 here so we know if | |
670 | the user explicitly requested this to be on or off. */ | |
671 | if (flag_schedule_insns > 0) | |
672 | flag_schedule_insns = 2; | |
673 | ||
674 | set_param_value ("simultaneous-prefetches", 2); | |
675 | } | |
676 | ||
677 | /* Implement OVERRIDE_OPTIONS macro. Validate and override various | |
678 | options, and do some machine dependent initialization. */ | |
679 | void | |
680 | sh_override_options (void) | |
681 | { | |
682 | int regno; | |
683 | ||
684 | SUBTARGET_OVERRIDE_OPTIONS; | |
685 | if (flag_finite_math_only == 2) | |
686 | flag_finite_math_only | |
687 | = !flag_signaling_nans && TARGET_SH2E && ! TARGET_IEEE; | |
688 | if (TARGET_SH2E && !flag_finite_math_only) | |
689 | target_flags |= MASK_IEEE; | |
690 | sh_cpu = PROCESSOR_SH1; | |
691 | assembler_dialect = 0; | |
692 | if (TARGET_SH2) | |
693 | sh_cpu = PROCESSOR_SH2; | |
694 | if (TARGET_SH2E) | |
695 | sh_cpu = PROCESSOR_SH2E; | |
696 | if (TARGET_SH2A) | |
697 | { | |
698 | sh_cpu = PROCESSOR_SH2A; | |
699 | if (TARGET_SH2A_DOUBLE) | |
700 | target_flags |= MASK_FMOVD; | |
701 | } | |
702 | if (TARGET_SH3) | |
703 | sh_cpu = PROCESSOR_SH3; | |
704 | if (TARGET_SH3E) | |
705 | sh_cpu = PROCESSOR_SH3E; | |
706 | if (TARGET_SH4) | |
707 | { | |
708 | assembler_dialect = 1; | |
709 | sh_cpu = PROCESSOR_SH4; | |
710 | } | |
711 | if (TARGET_SH4A_ARCH) | |
712 | { | |
713 | assembler_dialect = 1; | |
714 | sh_cpu = PROCESSOR_SH4A; | |
715 | } | |
716 | if (TARGET_SH5) | |
717 | { | |
718 | sh_cpu = PROCESSOR_SH5; | |
719 | target_flags |= MASK_ALIGN_DOUBLE; | |
720 | if (TARGET_SHMEDIA_FPU) | |
721 | target_flags |= MASK_FMOVD; | |
722 | if (TARGET_SHMEDIA) | |
723 | { | |
724 | /* There are no delay slots on SHmedia. */ | |
725 | flag_delayed_branch = 0; | |
726 | /* Relaxation isn't yet supported for SHmedia */ | |
727 | target_flags &= ~MASK_RELAX; | |
728 | /* After reload, if conversion does little good but can cause | |
729 | ICEs: | |
730 | - find_if_block doesn't do anything for SH because we don't | |
731 | have conditional execution patterns. (We use conditional | |
732 | move patterns, which are handled differently, and only | |
733 | before reload). | |
734 | - find_cond_trap doesn't do anything for the SH because we | |
735 | don't have conditional traps. | |
736 | - find_if_case_1 uses redirect_edge_and_branch_force in | |
737 | the only path that does an optimization, and this causes | |
738 | an ICE when branch targets are in registers. | |
739 | - find_if_case_2 doesn't do anything for the SHmedia after | |
740 | reload except when it can redirect a tablejump - and | |
741 | that's rather rare. */ | |
742 | flag_if_conversion2 = 0; | |
743 | if (! strcmp (sh_div_str, "call")) | |
744 | sh_div_strategy = SH_DIV_CALL; | |
745 | else if (! strcmp (sh_div_str, "call2")) | |
746 | sh_div_strategy = SH_DIV_CALL2; | |
747 | if (! strcmp (sh_div_str, "fp") && TARGET_FPU_ANY) | |
748 | sh_div_strategy = SH_DIV_FP; | |
749 | else if (! strcmp (sh_div_str, "inv")) | |
750 | sh_div_strategy = SH_DIV_INV; | |
751 | else if (! strcmp (sh_div_str, "inv:minlat")) | |
752 | sh_div_strategy = SH_DIV_INV_MINLAT; | |
753 | else if (! strcmp (sh_div_str, "inv20u")) | |
754 | sh_div_strategy = SH_DIV_INV20U; | |
755 | else if (! strcmp (sh_div_str, "inv20l")) | |
756 | sh_div_strategy = SH_DIV_INV20L; | |
757 | else if (! strcmp (sh_div_str, "inv:call2")) | |
758 | sh_div_strategy = SH_DIV_INV_CALL2; | |
759 | else if (! strcmp (sh_div_str, "inv:call")) | |
760 | sh_div_strategy = SH_DIV_INV_CALL; | |
761 | else if (! strcmp (sh_div_str, "inv:fp")) | |
762 | { | |
763 | if (TARGET_FPU_ANY) | |
764 | sh_div_strategy = SH_DIV_INV_FP; | |
765 | else | |
766 | sh_div_strategy = SH_DIV_INV; | |
767 | } | |
768 | TARGET_CBRANCHDI4 = 0; | |
769 | /* Assembler CFI isn't yet fully supported for SHmedia. */ | |
770 | flag_dwarf2_cfi_asm = 0; | |
771 | } | |
772 | } | |
773 | else | |
774 | { | |
775 | /* Only the sh64-elf assembler fully supports .quad properly. */ | |
776 | targetm.asm_out.aligned_op.di = NULL; | |
777 | targetm.asm_out.unaligned_op.di = NULL; | |
778 | } | |
779 | if (TARGET_SH1) | |
780 | { | |
781 | if (! strcmp (sh_div_str, "call-div1")) | |
782 | sh_div_strategy = SH_DIV_CALL_DIV1; | |
783 | else if (! strcmp (sh_div_str, "call-fp") | |
784 | && (TARGET_FPU_DOUBLE | |
785 | || (TARGET_HARD_SH4 && TARGET_SH2E) | |
786 | || (TARGET_SHCOMPACT && TARGET_FPU_ANY))) | |
787 | sh_div_strategy = SH_DIV_CALL_FP; | |
788 | else if (! strcmp (sh_div_str, "call-table") && TARGET_SH2) | |
789 | sh_div_strategy = SH_DIV_CALL_TABLE; | |
790 | else | |
791 | /* Pick one that makes most sense for the target in general. | |
792 | It is not much good to use different functions depending | |
793 | on -Os, since then we'll end up with two different functions | |
794 | when some of the code is compiled for size, and some for | |
795 | speed. */ | |
796 | ||
797 | /* SH4 tends to emphasize speed. */ | |
798 | if (TARGET_HARD_SH4) | |
799 | sh_div_strategy = SH_DIV_CALL_TABLE; | |
800 | /* These have their own way of doing things. */ | |
801 | else if (TARGET_SH2A) | |
802 | sh_div_strategy = SH_DIV_INTRINSIC; | |
803 | /* ??? Should we use the integer SHmedia function instead? */ | |
804 | else if (TARGET_SHCOMPACT && TARGET_FPU_ANY) | |
805 | sh_div_strategy = SH_DIV_CALL_FP; | |
806 | /* SH1 .. SH3 cores often go into small-footprint systems, so | |
807 | default to the smallest implementation available. */ | |
808 | else if (TARGET_SH2) /* ??? EXPERIMENTAL */ | |
809 | sh_div_strategy = SH_DIV_CALL_TABLE; | |
810 | else | |
811 | sh_div_strategy = SH_DIV_CALL_DIV1; | |
812 | } | |
813 | if (!TARGET_SH1) | |
814 | TARGET_PRETEND_CMOVE = 0; | |
815 | if (sh_divsi3_libfunc[0]) | |
816 | ; /* User supplied - leave it alone. */ | |
817 | else if (TARGET_DIVIDE_CALL_FP) | |
818 | sh_divsi3_libfunc = "__sdivsi3_i4"; | |
819 | else if (TARGET_DIVIDE_CALL_TABLE) | |
820 | sh_divsi3_libfunc = "__sdivsi3_i4i"; | |
821 | else if (TARGET_SH5) | |
822 | sh_divsi3_libfunc = "__sdivsi3_1"; | |
823 | else | |
824 | sh_divsi3_libfunc = "__sdivsi3"; | |
825 | if (sh_branch_cost == -1) | |
826 | sh_branch_cost | |
827 | = TARGET_SH5 ? 1 : ! TARGET_SH2 || TARGET_HARD_SH4 ? 2 : 1; | |
828 | ||
829 | for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++) | |
830 | if (! VALID_REGISTER_P (regno)) | |
831 | sh_register_names[regno][0] = '\0'; | |
832 | ||
833 | for (regno = 0; regno < ADDREGNAMES_SIZE; regno++) | |
834 | if (! VALID_REGISTER_P (ADDREGNAMES_REGNO (regno))) | |
835 | sh_additional_register_names[regno][0] = '\0'; | |
836 | ||
837 | if (flag_omit_frame_pointer == 2) | |
838 | { | |
839 | /* The debugging information is sufficient, | |
840 | but gdb doesn't implement this yet */ | |
841 | if (0) | |
842 | flag_omit_frame_pointer | |
843 | = (PREFERRED_DEBUGGING_TYPE == DWARF2_DEBUG); | |
844 | else | |
845 | flag_omit_frame_pointer = 0; | |
846 | } | |
847 | ||
848 | if ((flag_pic && ! TARGET_PREFERGOT) | |
849 | || (TARGET_SHMEDIA && !TARGET_PT_FIXED)) | |
850 | flag_no_function_cse = 1; | |
851 | ||
852 | if (SMALL_REGISTER_CLASSES) | |
853 | { | |
854 | /* Never run scheduling before reload, since that can | |
855 | break global alloc, and generates slower code anyway due | |
856 | to the pressure on R0. */ | |
857 | /* Enable sched1 for SH4 if the user explicitly requests. | |
858 | When sched1 is enabled, the ready queue will be reordered by | |
859 | the target hooks if pressure is high. We can not do this for | |
860 | PIC, SH3 and lower as they give spill failures for R0. */ | |
861 | if (!TARGET_HARD_SH4 || flag_pic) | |
862 | flag_schedule_insns = 0; | |
863 | /* ??? Current exception handling places basic block boundaries | |
864 | after call_insns. It causes the high pressure on R0 and gives | |
865 | spill failures for R0 in reload. See PR 22553 and the thread | |
866 | on gcc-patches | |
867 | <http://gcc.gnu.org/ml/gcc-patches/2005-10/msg00816.html>. */ | |
868 | else if (flag_exceptions) | |
869 | { | |
870 | if (flag_schedule_insns == 1) | |
871 | warning (0, "ignoring -fschedule-insns because of exception handling bug"); | |
872 | flag_schedule_insns = 0; | |
873 | } | |
874 | else if (flag_schedule_insns == 2) | |
875 | flag_schedule_insns = 0; | |
876 | } | |
877 | ||
878 | if (align_loops == 0) | |
879 | align_loops = 1 << (TARGET_SH5 ? 3 : 2); | |
880 | if (align_jumps == 0) | |
881 | align_jumps = 1 << CACHE_LOG; | |
882 | else if (align_jumps < (TARGET_SHMEDIA ? 4 : 2)) | |
883 | align_jumps = TARGET_SHMEDIA ? 4 : 2; | |
884 | ||
885 | /* Allocation boundary (in *bytes*) for the code of a function. | |
886 | SH1: 32 bit alignment is faster, because instructions are always | |
887 | fetched as a pair from a longword boundary. | |
888 | SH2 .. SH5 : align to cache line start. */ | |
889 | if (align_functions == 0) | |
890 | align_functions | |
891 | = TARGET_SMALLCODE ? FUNCTION_BOUNDARY/8 : (1 << CACHE_LOG); | |
892 | /* The linker relaxation code breaks when a function contains | |
893 | alignments that are larger than that at the start of a | |
894 | compilation unit. */ | |
895 | if (TARGET_RELAX) | |
896 | { | |
897 | int min_align | |
898 | = align_loops > align_jumps ? align_loops : align_jumps; | |
899 | ||
900 | /* Also take possible .long constants / mova tables int account. */ | |
901 | if (min_align < 4) | |
902 | min_align = 4; | |
903 | if (align_functions < min_align) | |
904 | align_functions = min_align; | |
905 | } | |
906 | ||
907 | if (sh_fixed_range_str) | |
908 | sh_fix_range (sh_fixed_range_str); | |
909 | } | |
910 | \f | |
16bea517 | 911 | /* Print the operand address in x to the stream. */ |
bc45ade3 SC |
912 | |
913 | void | |
cf277499 | 914 | print_operand_address (FILE *stream, rtx x) |
bc45ade3 SC |
915 | { |
916 | switch (GET_CODE (x)) | |
917 | { | |
918 | case REG: | |
225e4f43 R |
919 | case SUBREG: |
920 | fprintf (stream, "@%s", reg_names[true_regnum (x)]); | |
bc45ade3 | 921 | break; |
8aa2a305 | 922 | |
bc45ade3 SC |
923 | case PLUS: |
924 | { | |
925 | rtx base = XEXP (x, 0); | |
926 | rtx index = XEXP (x, 1); | |
927 | ||
bc45ade3 SC |
928 | switch (GET_CODE (index)) |
929 | { | |
930 | case CONST_INT: | |
c5b9ef02 | 931 | fprintf (stream, "@(%d,%s)", (int) INTVAL (index), |
225e4f43 | 932 | reg_names[true_regnum (base)]); |
bc45ade3 SC |
933 | break; |
934 | ||
935 | case REG: | |
225e4f43 R |
936 | case SUBREG: |
937 | { | |
938 | int base_num = true_regnum (base); | |
939 | int index_num = true_regnum (index); | |
940 | ||
941 | fprintf (stream, "@(r0,%s)", | |
942 | reg_names[MAX (base_num, index_num)]); | |
943 | break; | |
944 | } | |
bc45ade3 SC |
945 | |
946 | default: | |
f5b9e7c9 | 947 | gcc_unreachable (); |
bc45ade3 SC |
948 | } |
949 | } | |
bc45ade3 | 950 | break; |
8aa2a305 | 951 | |
bc45ade3 | 952 | case PRE_DEC: |
225e4f43 | 953 | fprintf (stream, "@-%s", reg_names[true_regnum (XEXP (x, 0))]); |
bc45ade3 SC |
954 | break; |
955 | ||
956 | case POST_INC: | |
225e4f43 | 957 | fprintf (stream, "@%s+", reg_names[true_regnum (XEXP (x, 0))]); |
bc45ade3 SC |
958 | break; |
959 | ||
960 | default: | |
b91455de | 961 | x = mark_constant_pool_use (x); |
422be3c3 | 962 | output_addr_const (stream, x); |
bc45ade3 SC |
963 | break; |
964 | } | |
965 | } | |
966 | ||
967 | /* Print operand x (an rtx) in assembler syntax to file stream | |
968 | according to modifier code. | |
969 | ||
b9654711 | 970 | '.' print a .s if insn needs delay slot |
1245df60 | 971 | ',' print LOCAL_LABEL_PREFIX |
4408efce | 972 | '@' print trap, rte or rts depending upon pragma interruptness |
8aa2a305 | 973 | '#' output a nop if there is nothing to put in the delay slot |
fae778eb | 974 | ''' print likelihood suffix (/u for unlikely). |
73a4d10b | 975 | '>' print branch target if -fverbose-asm |
8aa2a305 | 976 | 'O' print a constant without the # |
00f8ff66 | 977 | 'R' print the LSW of a dp value - changes if in little endian |
00f8ff66 | 978 | 'S' print the MSW of a dp value - changes if in little endian |
225e4f43 | 979 | 'T' print the next word of a dp value - same as 'R' in big endian mode. |
31b6f0ae R |
980 | 'M' SHMEDIA: print an `x' if `m' will print `base,index'. |
981 | otherwise: print .b / .w / .l / .s / .d suffix if operand is a MEM. | |
c1b92d09 | 982 | 'N' print 'r63' if the operand is (const_int 0). |
312209c6 | 983 | 'd' print a V2SF reg as dN instead of fpN. |
fa5322fa | 984 | 'm' print a pair `base,offset' or `base,index', for LD and ST. |
73a4d10b | 985 | 'U' Likewise for {LD,ST}{HI,LO}. |
de6adfa2 N |
986 | 'V' print the position of a single bit set. |
987 | 'W' print the position of a single bit cleared. | |
9eb3a0dd | 988 | 't' print a memory address which is a register. |
fa5322fa | 989 | 'u' prints the lowest 16 bits of CONST_INT, as an unsigned value. |
225e4f43 | 990 | 'o' output an operator. */ |
bc45ade3 SC |
991 | |
992 | void | |
cf277499 | 993 | print_operand (FILE *stream, rtx x, int code) |
bc45ade3 | 994 | { |
73a4d10b R |
995 | int regno; |
996 | enum machine_mode mode; | |
997 | ||
bc45ade3 SC |
998 | switch (code) |
999 | { | |
a6ab9fc0 R |
1000 | tree trapa_attr; |
1001 | ||
b9654711 | 1002 | case '.': |
79b2746a | 1003 | if (final_sequence |
c608a684 R |
1004 | && ! INSN_ANNULLED_BRANCH_P (XVECEXP (final_sequence, 0, 0)) |
1005 | && get_attr_length (XVECEXP (final_sequence, 0, 1))) | |
1245df60 R |
1006 | fprintf (stream, ASSEMBLER_DIALECT ? "/s" : ".s"); |
1007 | break; | |
1008 | case ',': | |
1009 | fprintf (stream, "%s", LOCAL_LABEL_PREFIX); | |
b9654711 | 1010 | break; |
d3ae8277 | 1011 | case '@': |
a6ab9fc0 R |
1012 | trapa_attr = lookup_attribute ("trap_exit", |
1013 | DECL_ATTRIBUTES (current_function_decl)); | |
1014 | if (trapa_attr) | |
1015 | fprintf (stream, "trapa #%ld", | |
1016 | (long) TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (trapa_attr)))); | |
e3ba8d11 | 1017 | else if (sh_cfun_interrupt_handler_p ()) |
561642fa AP |
1018 | { |
1019 | if (sh_cfun_resbank_handler_p ()) | |
1020 | fprintf (stream, "resbank\n"); | |
1021 | fprintf (stream, "rte"); | |
1022 | } | |
d3ae8277 | 1023 | else |
8e87e161 | 1024 | fprintf (stream, "rts"); |
d3ae8277 | 1025 | break; |
bc45ade3 | 1026 | case '#': |
16bea517 | 1027 | /* Output a nop if there's nothing in the delay slot. */ |
bc45ade3 | 1028 | if (dbr_sequence_length () == 0) |
8aa2a305 | 1029 | fprintf (stream, "\n\tnop"); |
bc45ade3 | 1030 | break; |
c8cc4417 R |
1031 | case '\'': |
1032 | { | |
1033 | rtx note = find_reg_note (current_output_insn, REG_BR_PROB, 0); | |
1034 | ||
1035 | if (note && INTVAL (XEXP (note, 0)) * 2 < REG_BR_PROB_BASE) | |
1036 | fputs ("/u", stream); | |
1037 | break; | |
1038 | } | |
73a4d10b R |
1039 | case '>': |
1040 | if (flag_verbose_asm && JUMP_LABEL (current_output_insn)) | |
1041 | { | |
1042 | fputs ("\t! target: ", stream); | |
1043 | output_addr_const (stream, JUMP_LABEL (current_output_insn)); | |
1044 | } | |
1045 | break; | |
b9654711 | 1046 | case 'O': |
b91455de | 1047 | x = mark_constant_pool_use (x); |
0d7e008e | 1048 | output_addr_const (stream, x); |
bc45ade3 | 1049 | break; |
4802062c R |
1050 | /* N.B.: %R / %S / %T adjust memory addresses by four. |
1051 | For SHMEDIA, that means they can be used to access the first and | |
1052 | second 32 bit part of a 64 bit (or larger) value that | |
1053 | might be held in floating point registers or memory. | |
1054 | While they can be used to access 64 bit parts of a larger value | |
1055 | held in general purpose registers, that won't work with memory - | |
1056 | neither for fp registers, since the frxx names are used. */ | |
bc45ade3 | 1057 | case 'R': |
4802062c R |
1058 | if (REG_P (x) || GET_CODE (x) == SUBREG) |
1059 | { | |
1060 | regno = true_regnum (x); | |
1061 | regno += FP_REGISTER_P (regno) ? 1 : LSW; | |
1062 | fputs (reg_names[regno], (stream)); | |
1063 | } | |
1064 | else if (MEM_P (x)) | |
1065 | { | |
1066 | x = adjust_address (x, SImode, 4 * LSW); | |
1067 | print_operand_address (stream, XEXP (x, 0)); | |
1068 | } | |
1069 | else | |
1070 | { | |
1071 | rtx sub = NULL_RTX; | |
1072 | ||
1073 | mode = GET_MODE (x); | |
1074 | if (mode == VOIDmode) | |
1075 | mode = DImode; | |
1076 | if (GET_MODE_SIZE (mode) >= 8) | |
1077 | sub = simplify_subreg (SImode, x, mode, 4 * LSW); | |
1078 | if (sub) | |
1079 | print_operand (stream, sub, 0); | |
1080 | else | |
1081 | output_operand_lossage ("invalid operand to %%R"); | |
1082 | } | |
8aa2a305 JW |
1083 | break; |
1084 | case 'S': | |
4802062c R |
1085 | if (REG_P (x) || GET_CODE (x) == SUBREG) |
1086 | { | |
1087 | regno = true_regnum (x); | |
1088 | regno += FP_REGISTER_P (regno) ? 0 : MSW; | |
1089 | fputs (reg_names[regno], (stream)); | |
1090 | } | |
1091 | else if (MEM_P (x)) | |
1092 | { | |
1093 | x = adjust_address (x, SImode, 4 * MSW); | |
1094 | print_operand_address (stream, XEXP (x, 0)); | |
1095 | } | |
1096 | else | |
1097 | { | |
1098 | rtx sub = NULL_RTX; | |
1099 | ||
1100 | mode = GET_MODE (x); | |
1101 | if (mode == VOIDmode) | |
1102 | mode = DImode; | |
1103 | if (GET_MODE_SIZE (mode) >= 8) | |
1104 | sub = simplify_subreg (SImode, x, mode, 4 * MSW); | |
1105 | if (sub) | |
1106 | print_operand (stream, sub, 0); | |
1107 | else | |
1108 | output_operand_lossage ("invalid operand to %%S"); | |
1109 | } | |
00f8ff66 SC |
1110 | break; |
1111 | case 'T': | |
16bea517 | 1112 | /* Next word of a double. */ |
bc45ade3 SC |
1113 | switch (GET_CODE (x)) |
1114 | { | |
1115 | case REG: | |
1116 | fputs (reg_names[REGNO (x) + 1], (stream)); | |
1117 | break; | |
1118 | case MEM: | |
225e4f43 R |
1119 | if (GET_CODE (XEXP (x, 0)) != PRE_DEC |
1120 | && GET_CODE (XEXP (x, 0)) != POST_INC) | |
b72f00af | 1121 | x = adjust_address (x, SImode, 4); |
225e4f43 | 1122 | print_operand_address (stream, XEXP (x, 0)); |
bc45ade3 | 1123 | break; |
318881c0 KG |
1124 | default: |
1125 | break; | |
bc45ade3 SC |
1126 | } |
1127 | break; | |
9eb3a0dd N |
1128 | |
1129 | case 't': | |
f3536097 | 1130 | gcc_assert (MEM_P (x)); |
9eb3a0dd N |
1131 | x = XEXP (x, 0); |
1132 | switch (GET_CODE (x)) | |
1133 | { | |
1134 | case REG: | |
1135 | case SUBREG: | |
1136 | print_operand (stream, x, 0); | |
1137 | break; | |
1138 | default: | |
1139 | break; | |
1140 | } | |
1141 | break; | |
1142 | ||
225e4f43 R |
1143 | case 'o': |
1144 | switch (GET_CODE (x)) | |
1145 | { | |
1146 | case PLUS: fputs ("add", stream); break; | |
1147 | case MINUS: fputs ("sub", stream); break; | |
1148 | case MULT: fputs ("mul", stream); break; | |
1149 | case DIV: fputs ("div", stream); break; | |
c8cc4417 R |
1150 | case EQ: fputs ("eq", stream); break; |
1151 | case NE: fputs ("ne", stream); break; | |
1152 | case GT: case LT: fputs ("gt", stream); break; | |
1153 | case GE: case LE: fputs ("ge", stream); break; | |
1154 | case GTU: case LTU: fputs ("gtu", stream); break; | |
1155 | case GEU: case LEU: fputs ("geu", stream); break; | |
318881c0 KG |
1156 | default: |
1157 | break; | |
225e4f43 R |
1158 | } |
1159 | break; | |
fa5322fa | 1160 | case 'M': |
31b6f0ae R |
1161 | if (TARGET_SHMEDIA) |
1162 | { | |
f3536097 | 1163 | if (MEM_P (x) |
31b6f0ae | 1164 | && GET_CODE (XEXP (x, 0)) == PLUS |
f3536097 | 1165 | && (REG_P (XEXP (XEXP (x, 0), 1)) |
31b6f0ae R |
1166 | || GET_CODE (XEXP (XEXP (x, 0), 1)) == SUBREG)) |
1167 | fputc ('x', stream); | |
1168 | } | |
1169 | else | |
1170 | { | |
f3536097 | 1171 | if (MEM_P (x)) |
31b6f0ae R |
1172 | { |
1173 | switch (GET_MODE (x)) | |
1174 | { | |
1175 | case QImode: fputs (".b", stream); break; | |
1176 | case HImode: fputs (".w", stream); break; | |
1177 | case SImode: fputs (".l", stream); break; | |
1178 | case SFmode: fputs (".s", stream); break; | |
1179 | case DFmode: fputs (".d", stream); break; | |
1180 | default: gcc_unreachable (); | |
1181 | } | |
1182 | } | |
1183 | } | |
fa5322fa AO |
1184 | break; |
1185 | ||
1186 | case 'm': | |
f3536097 | 1187 | gcc_assert (MEM_P (x)); |
fa5322fa | 1188 | x = XEXP (x, 0); |
73a4d10b R |
1189 | /* Fall through. */ |
1190 | case 'U': | |
fa5322fa AO |
1191 | switch (GET_CODE (x)) |
1192 | { | |
1193 | case REG: | |
1194 | case SUBREG: | |
1195 | print_operand (stream, x, 0); | |
1196 | fputs (", 0", stream); | |
1197 | break; | |
1198 | ||
1199 | case PLUS: | |
1200 | print_operand (stream, XEXP (x, 0), 0); | |
1201 | fputs (", ", stream); | |
1202 | print_operand (stream, XEXP (x, 1), 0); | |
1203 | break; | |
1204 | ||
1205 | default: | |
f5b9e7c9 | 1206 | gcc_unreachable (); |
fa5322fa AO |
1207 | } |
1208 | break; | |
1209 | ||
de6adfa2 N |
1210 | case 'V': |
1211 | { | |
1212 | int num = exact_log2 (INTVAL (x)); | |
1213 | gcc_assert (num >= 0); | |
1214 | fprintf (stream, "#%d", num); | |
1215 | } | |
1216 | break; | |
1217 | ||
1218 | case 'W': | |
1219 | { | |
1220 | int num = exact_log2 (~INTVAL (x)); | |
1221 | gcc_assert (num >= 0); | |
1222 | fprintf (stream, "#%d", num); | |
1223 | } | |
1224 | break; | |
1225 | ||
312209c6 | 1226 | case 'd': |
f3536097 | 1227 | gcc_assert (REG_P (x) && GET_MODE (x) == V2SFmode); |
312209c6 AO |
1228 | |
1229 | fprintf ((stream), "d%s", reg_names[REGNO (x)] + 1); | |
1230 | break; | |
f676971a | 1231 | |
c1b92d09 | 1232 | case 'N': |
0ac78517 | 1233 | if (x == CONST0_RTX (GET_MODE (x))) |
c1b92d09 R |
1234 | { |
1235 | fprintf ((stream), "r63"); | |
1236 | break; | |
1237 | } | |
1238 | goto default_output; | |
fa5322fa | 1239 | case 'u': |
f3536097 | 1240 | if (CONST_INT_P (x)) |
c8cc4417 | 1241 | { |
fa5322fa AO |
1242 | fprintf ((stream), "%u", (unsigned) INTVAL (x) & (0x10000 - 1)); |
1243 | break; | |
1244 | } | |
1245 | /* Fall through. */ | |
1246 | ||
c1b92d09 | 1247 | default_output: |
bc45ade3 | 1248 | default: |
73a4d10b R |
1249 | regno = 0; |
1250 | mode = GET_MODE (x); | |
1251 | ||
bc45ade3 SC |
1252 | switch (GET_CODE (x)) |
1253 | { | |
73a4d10b R |
1254 | case TRUNCATE: |
1255 | { | |
1256 | rtx inner = XEXP (x, 0); | |
1257 | int offset = 0; | |
1258 | enum machine_mode inner_mode; | |
1259 | ||
1260 | /* We might see SUBREGs with vector mode registers inside. */ | |
1261 | if (GET_CODE (inner) == SUBREG | |
1262 | && (GET_MODE_SIZE (GET_MODE (inner)) | |
1263 | == GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner)))) | |
1264 | && subreg_lowpart_p (inner)) | |
1265 | inner = SUBREG_REG (inner); | |
f3536097 | 1266 | if (CONST_INT_P (inner)) |
73a4d10b R |
1267 | { |
1268 | x = GEN_INT (trunc_int_for_mode (INTVAL (inner), GET_MODE (x))); | |
1269 | goto default_output; | |
1270 | } | |
1271 | inner_mode = GET_MODE (inner); | |
1272 | if (GET_CODE (inner) == SUBREG | |
1273 | && (GET_MODE_SIZE (GET_MODE (inner)) | |
1274 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner)))) | |
f3536097 | 1275 | && REG_P (SUBREG_REG (inner))) |
73a4d10b R |
1276 | { |
1277 | offset = subreg_regno_offset (REGNO (SUBREG_REG (inner)), | |
1278 | GET_MODE (SUBREG_REG (inner)), | |
1279 | SUBREG_BYTE (inner), | |
1280 | GET_MODE (inner)); | |
1281 | inner = SUBREG_REG (inner); | |
1282 | } | |
f3536097 | 1283 | if (!REG_P (inner) || GET_MODE_SIZE (inner_mode) > 8) |
73a4d10b R |
1284 | abort (); |
1285 | /* Floating point register pairs are always big endian; | |
30dc60c7 | 1286 | general purpose registers are 64 bit wide. */ |
73a4d10b R |
1287 | regno = REGNO (inner); |
1288 | regno = (HARD_REGNO_NREGS (regno, inner_mode) | |
1289 | - HARD_REGNO_NREGS (regno, mode)) | |
1290 | + offset; | |
1291 | x = inner; | |
1292 | goto reg; | |
1293 | } | |
1294 | case SIGN_EXTEND: | |
1295 | x = XEXP (x, 0); | |
1296 | goto reg; | |
fa5322fa AO |
1297 | /* FIXME: We need this on SHmedia32 because reload generates |
1298 | some sign-extended HI or QI loads into DImode registers | |
1299 | but, because Pmode is SImode, the address ends up with a | |
1300 | subreg:SI of the DImode register. Maybe reload should be | |
1301 | fixed so as to apply alter_subreg to such loads? */ | |
73a4d10b R |
1302 | case IF_THEN_ELSE: |
1303 | gcc_assert (trapping_target_operand (x, VOIDmode)); | |
1304 | x = XEXP (XEXP (x, 2), 0); | |
1305 | goto default_output; | |
fa5322fa | 1306 | case SUBREG: |
f5b9e7c9 | 1307 | gcc_assert (SUBREG_BYTE (x) == 0 |
f3536097 | 1308 | && REG_P (SUBREG_REG (x))); |
fa5322fa AO |
1309 | |
1310 | x = SUBREG_REG (x); | |
1311 | /* Fall through. */ | |
1312 | ||
73a4d10b | 1313 | reg: |
bc45ade3 | 1314 | case REG: |
73a4d10b R |
1315 | regno += REGNO (x); |
1316 | if (FP_REGISTER_P (regno) | |
1317 | && mode == V16SFmode) | |
1318 | fprintf ((stream), "mtrx%s", reg_names[regno] + 2); | |
fa5322fa | 1319 | else if (FP_REGISTER_P (REGNO (x)) |
73a4d10b R |
1320 | && mode == V4SFmode) |
1321 | fprintf ((stream), "fv%s", reg_names[regno] + 2); | |
f3536097 | 1322 | else if (REG_P (x) |
73a4d10b R |
1323 | && mode == V2SFmode) |
1324 | fprintf ((stream), "fp%s", reg_names[regno] + 2); | |
fa5322fa | 1325 | else if (FP_REGISTER_P (REGNO (x)) |
73a4d10b R |
1326 | && GET_MODE_SIZE (mode) > 4) |
1327 | fprintf ((stream), "d%s", reg_names[regno] + 1); | |
225e4f43 | 1328 | else |
73a4d10b | 1329 | fputs (reg_names[regno], (stream)); |
bc45ade3 | 1330 | break; |
fa5322fa | 1331 | |
bc45ade3 SC |
1332 | case MEM: |
1333 | output_address (XEXP (x, 0)); | |
1334 | break; | |
f676971a | 1335 | |
bc45ade3 | 1336 | default: |
fa5322fa AO |
1337 | if (TARGET_SH1) |
1338 | fputc ('#', stream); | |
bc45ade3 SC |
1339 | output_addr_const (stream, x); |
1340 | break; | |
bc45ade3 SC |
1341 | } |
1342 | break; | |
1343 | } | |
1344 | } | |
bc45ade3 | 1345 | \f |
561642fa AP |
1346 | |
1347 | /* Encode symbol attributes of a SYMBOL_REF into its | |
1348 | SYMBOL_REF_FLAGS. */ | |
1349 | static void | |
1350 | sh_encode_section_info (tree decl, rtx rtl, int first) | |
1351 | { | |
1352 | default_encode_section_info (decl, rtl, first); | |
1353 | ||
1354 | if (TREE_CODE (decl) == FUNCTION_DECL | |
1355 | && sh2a_function_vector_p (decl) && TARGET_SH2A) | |
1356 | SYMBOL_REF_FLAGS (XEXP (rtl, 0)) |= SYMBOL_FLAG_FUNCVEC_FUNCTION; | |
1357 | } | |
1358 | ||
0aa692fc R |
1359 | /* Like force_operand, but guarantees that VALUE ends up in TARGET. */ |
1360 | static void | |
cf277499 | 1361 | force_into (rtx value, rtx target) |
0aa692fc R |
1362 | { |
1363 | value = force_operand (value, target); | |
1364 | if (! rtx_equal_p (value, target)) | |
1365 | emit_insn (gen_move_insn (target, value)); | |
1366 | } | |
1367 | ||
0d7e008e SC |
1368 | /* Emit code to perform a block move. Choose the best method. |
1369 | ||
1370 | OPERANDS[0] is the destination. | |
1371 | OPERANDS[1] is the source. | |
1372 | OPERANDS[2] is the size. | |
1373 | OPERANDS[3] is the alignment safe to use. */ | |
1374 | ||
0d7e008e | 1375 | int |
cf277499 | 1376 | expand_block_move (rtx *operands) |
0d7e008e SC |
1377 | { |
1378 | int align = INTVAL (operands[3]); | |
f3536097 | 1379 | int constp = (CONST_INT_P (operands[2])); |
0d7e008e | 1380 | int bytes = (constp ? INTVAL (operands[2]) : 0); |
0d7e008e | 1381 | |
312209c6 AO |
1382 | if (! constp) |
1383 | return 0; | |
1384 | ||
1385 | /* If we could use mov.l to move words and dest is word-aligned, we | |
1386 | can use movua.l for loads and still generate a relatively short | |
1387 | and efficient sequence. */ | |
1388 | if (TARGET_SH4A_ARCH && align < 4 | |
1389 | && MEM_ALIGN (operands[0]) >= 32 | |
1390 | && can_move_by_pieces (bytes, 32)) | |
1391 | { | |
1392 | rtx dest = copy_rtx (operands[0]); | |
1393 | rtx src = copy_rtx (operands[1]); | |
1394 | /* We could use different pseudos for each copied word, but | |
1395 | since movua can only load into r0, it's kind of | |
1396 | pointless. */ | |
1397 | rtx temp = gen_reg_rtx (SImode); | |
1398 | rtx src_addr = copy_addr_to_reg (XEXP (src, 0)); | |
1399 | int copied = 0; | |
1400 | ||
1401 | while (copied + 4 <= bytes) | |
1402 | { | |
1403 | rtx to = adjust_address (dest, SImode, copied); | |
85144c5c RS |
1404 | rtx from = adjust_automodify_address (src, BLKmode, |
1405 | src_addr, copied); | |
312209c6 | 1406 | |
85144c5c | 1407 | set_mem_size (from, GEN_INT (4)); |
312209c6 AO |
1408 | emit_insn (gen_movua (temp, from)); |
1409 | emit_move_insn (src_addr, plus_constant (src_addr, 4)); | |
1410 | emit_move_insn (to, temp); | |
1411 | copied += 4; | |
1412 | } | |
1413 | ||
1414 | if (copied < bytes) | |
1415 | move_by_pieces (adjust_address (dest, BLKmode, copied), | |
1416 | adjust_automodify_address (src, BLKmode, | |
1417 | src_addr, copied), | |
1418 | bytes - copied, align, 0); | |
1419 | ||
1420 | return 1; | |
1421 | } | |
1422 | ||
8aa2a305 JW |
1423 | /* If it isn't a constant number of bytes, or if it doesn't have 4 byte |
1424 | alignment, or if it isn't a multiple of 4 bytes, then fail. */ | |
312209c6 | 1425 | if (align < 4 || (bytes % 4 != 0)) |
d3ae8277 SC |
1426 | return 0; |
1427 | ||
225e4f43 R |
1428 | if (TARGET_HARD_SH4) |
1429 | { | |
1430 | if (bytes < 12) | |
1431 | return 0; | |
1432 | else if (bytes == 12) | |
1433 | { | |
73a4d10b | 1434 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
c0d4e710 KH |
1435 | rtx r4 = gen_rtx_REG (SImode, 4); |
1436 | rtx r5 = gen_rtx_REG (SImode, 5); | |
225e4f43 | 1437 | |
73a4d10b | 1438 | function_symbol (func_addr_rtx, "__movmemSI12_i4", SFUNC_STATIC); |
0aa692fc R |
1439 | force_into (XEXP (operands[0], 0), r4); |
1440 | force_into (XEXP (operands[1], 0), r5); | |
225e4f43 R |
1441 | emit_insn (gen_block_move_real_i4 (func_addr_rtx)); |
1442 | return 1; | |
1443 | } | |
1444 | else if (! TARGET_SMALLCODE) | |
1445 | { | |
73a4d10b R |
1446 | const char *entry_name; |
1447 | rtx func_addr_rtx = gen_reg_rtx (Pmode); | |
225e4f43 | 1448 | int dwords; |
c0d4e710 KH |
1449 | rtx r4 = gen_rtx_REG (SImode, 4); |
1450 | rtx r5 = gen_rtx_REG (SImode, 5); | |
1451 | rtx r6 = gen_rtx_REG (SImode, 6); | |
225e4f43 | 1452 | |
73a4d10b R |
1453 | entry_name = (bytes & 4 ? "__movmem_i4_odd" : "__movmem_i4_even"); |
1454 | function_symbol (func_addr_rtx, entry_name, SFUNC_STATIC); | |
0aa692fc R |
1455 | force_into (XEXP (operands[0], 0), r4); |
1456 | force_into (XEXP (operands[1], 0), r5); | |
225e4f43 R |
1457 | |
1458 | dwords = bytes >> 3; | |
1459 | emit_insn (gen_move_insn (r6, GEN_INT (dwords - 1))); | |
1460 | emit_insn (gen_block_lump_real_i4 (func_addr_rtx)); | |
1461 | return 1; | |
1462 | } | |
1463 | else | |
1464 | return 0; | |
1465 | } | |
8aa2a305 | 1466 | if (bytes < 64) |
0d7e008e SC |
1467 | { |
1468 | char entry[30]; | |
73a4d10b | 1469 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
c5c76735 JL |
1470 | rtx r4 = gen_rtx_REG (SImode, 4); |
1471 | rtx r5 = gen_rtx_REG (SImode, 5); | |
8aa2a305 | 1472 | |
70128ad9 | 1473 | sprintf (entry, "__movmemSI%d", bytes); |
73a4d10b | 1474 | function_symbol (func_addr_rtx, entry, SFUNC_STATIC); |
0aa692fc R |
1475 | force_into (XEXP (operands[0], 0), r4); |
1476 | force_into (XEXP (operands[1], 0), r5); | |
0d7e008e SC |
1477 | emit_insn (gen_block_move_real (func_addr_rtx)); |
1478 | return 1; | |
1479 | } | |
8aa2a305 JW |
1480 | |
1481 | /* This is the same number of bytes as a memcpy call, but to a different | |
1482 | less common function name, so this will occasionally use more space. */ | |
1483 | if (! TARGET_SMALLCODE) | |
0d7e008e | 1484 | { |
73a4d10b | 1485 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
8aa2a305 | 1486 | int final_switch, while_loop; |
c5c76735 JL |
1487 | rtx r4 = gen_rtx_REG (SImode, 4); |
1488 | rtx r5 = gen_rtx_REG (SImode, 5); | |
1489 | rtx r6 = gen_rtx_REG (SImode, 6); | |
0d7e008e | 1490 | |
73a4d10b | 1491 | function_symbol (func_addr_rtx, "__movmem", SFUNC_STATIC); |
0aa692fc R |
1492 | force_into (XEXP (operands[0], 0), r4); |
1493 | force_into (XEXP (operands[1], 0), r5); | |
0d7e008e | 1494 | |
8aa2a305 JW |
1495 | /* r6 controls the size of the move. 16 is decremented from it |
1496 | for each 64 bytes moved. Then the negative bit left over is used | |
1497 | as an index into a list of move instructions. e.g., a 72 byte move | |
1498 | would be set up with size(r6) = 14, for one iteration through the | |
1499 | big while loop, and a switch of -2 for the last part. */ | |
0d7e008e | 1500 | |
8aa2a305 JW |
1501 | final_switch = 16 - ((bytes / 4) % 16); |
1502 | while_loop = ((bytes / 4) / 16 - 1) * 16; | |
1503 | emit_insn (gen_move_insn (r6, GEN_INT (while_loop + final_switch))); | |
1504 | emit_insn (gen_block_lump_real (func_addr_rtx)); | |
1505 | return 1; | |
0d7e008e | 1506 | } |
0d7e008e | 1507 | |
d3ae8277 | 1508 | return 0; |
0d7e008e SC |
1509 | } |
1510 | ||
bc45ade3 | 1511 | /* Prepare operands for a move define_expand; specifically, one of the |
8aa2a305 | 1512 | operands must be in a register. */ |
bc45ade3 | 1513 | |
b9654711 | 1514 | int |
cf277499 | 1515 | prepare_move_operands (rtx operands[], enum machine_mode mode) |
bc45ade3 | 1516 | { |
463f02cd KK |
1517 | if ((mode == SImode || mode == DImode) |
1518 | && flag_pic | |
1519 | && ! ((mode == Pmode || mode == ptr_mode) | |
81f40b79 | 1520 | && tls_symbolic_operand (operands[1], Pmode) != TLS_MODEL_NONE)) |
1a66cd67 AO |
1521 | { |
1522 | rtx temp; | |
1523 | if (SYMBOLIC_CONST_P (operands[1])) | |
1524 | { | |
f3536097 | 1525 | if (MEM_P (operands[0])) |
1a66cd67 | 1526 | operands[1] = force_reg (Pmode, operands[1]); |
5615d8fd SC |
1527 | else if (TARGET_SHMEDIA |
1528 | && GET_CODE (operands[1]) == LABEL_REF | |
fa5322fa AO |
1529 | && target_reg_operand (operands[0], mode)) |
1530 | /* It's ok. */; | |
1a66cd67 AO |
1531 | else |
1532 | { | |
b3a13419 ILT |
1533 | temp = (!can_create_pseudo_p () |
1534 | ? operands[0] | |
1535 | : gen_reg_rtx (Pmode)); | |
fa5322fa | 1536 | operands[1] = legitimize_pic_address (operands[1], mode, temp); |
1a66cd67 AO |
1537 | } |
1538 | } | |
1539 | else if (GET_CODE (operands[1]) == CONST | |
1540 | && GET_CODE (XEXP (operands[1], 0)) == PLUS | |
1541 | && SYMBOLIC_CONST_P (XEXP (XEXP (operands[1], 0), 0))) | |
1542 | { | |
b3a13419 | 1543 | temp = !can_create_pseudo_p () ? operands[0] : gen_reg_rtx (Pmode); |
1a66cd67 | 1544 | temp = legitimize_pic_address (XEXP (XEXP (operands[1], 0), 0), |
fa5322fa AO |
1545 | mode, temp); |
1546 | operands[1] = expand_binop (mode, add_optab, temp, | |
1a66cd67 | 1547 | XEXP (XEXP (operands[1], 0), 1), |
b3a13419 ILT |
1548 | (!can_create_pseudo_p () |
1549 | ? temp | |
1550 | : gen_reg_rtx (Pmode)), | |
68c37dc9 | 1551 | 0, OPTAB_LIB_WIDEN); |
1a66cd67 AO |
1552 | } |
1553 | } | |
1554 | ||
e22856e8 JW |
1555 | if (! reload_in_progress && ! reload_completed) |
1556 | { | |
1557 | /* Copy the source to a register if both operands aren't registers. */ | |
1558 | if (! register_operand (operands[0], mode) | |
d9da94a1 | 1559 | && ! sh_register_operand (operands[1], mode)) |
e22856e8 JW |
1560 | operands[1] = copy_to_mode_reg (mode, operands[1]); |
1561 | ||
f3536097 | 1562 | if (MEM_P (operands[0]) && ! memory_operand (operands[0], mode)) |
266a2732 R |
1563 | { |
1564 | /* This is like change_address_1 (operands[0], mode, 0, 1) , | |
1565 | except that we can't use that function because it is static. */ | |
0a2aaacc KG |
1566 | rtx new_rtx = change_address (operands[0], mode, 0); |
1567 | MEM_COPY_ATTRIBUTES (new_rtx, operands[0]); | |
1568 | operands[0] = new_rtx; | |
266a2732 R |
1569 | } |
1570 | ||
e22856e8 JW |
1571 | /* This case can happen while generating code to move the result |
1572 | of a library call to the target. Reject `st r0,@(rX,rY)' because | |
1573 | reload will fail to find a spill register for rX, since r0 is already | |
1574 | being used for the source. */ | |
73a4d10b R |
1575 | else if (TARGET_SH1 |
1576 | && refers_to_regno_p (R0_REG, R0_REG + 1, operands[1], (rtx *)0) | |
f3536097 | 1577 | && MEM_P (operands[0]) |
e22856e8 | 1578 | && GET_CODE (XEXP (operands[0], 0)) == PLUS |
f3536097 | 1579 | && REG_P (XEXP (XEXP (operands[0], 0), 1))) |
e22856e8 JW |
1580 | operands[1] = copy_to_mode_reg (mode, operands[1]); |
1581 | } | |
8aa2a305 | 1582 | |
463f02cd KK |
1583 | if (mode == Pmode || mode == ptr_mode) |
1584 | { | |
6058b2fa | 1585 | rtx op0, op1, opc; |
463f02cd KK |
1586 | enum tls_model tls_kind; |
1587 | ||
1588 | op0 = operands[0]; | |
1589 | op1 = operands[1]; | |
6058b2fa KK |
1590 | if (GET_CODE (op1) == CONST |
1591 | && GET_CODE (XEXP (op1, 0)) == PLUS | |
81f40b79 ILT |
1592 | && (tls_symbolic_operand (XEXP (XEXP (op1, 0), 0), Pmode) |
1593 | != TLS_MODEL_NONE)) | |
6058b2fa KK |
1594 | { |
1595 | opc = XEXP (XEXP (op1, 0), 1); | |
1596 | op1 = XEXP (XEXP (op1, 0), 0); | |
1597 | } | |
1598 | else | |
1599 | opc = NULL_RTX; | |
1600 | ||
81f40b79 | 1601 | if ((tls_kind = tls_symbolic_operand (op1, Pmode)) != TLS_MODEL_NONE) |
463f02cd KK |
1602 | { |
1603 | rtx tga_op1, tga_ret, tmp, tmp2; | |
1604 | ||
463f02cd KK |
1605 | switch (tls_kind) |
1606 | { | |
1607 | case TLS_MODEL_GLOBAL_DYNAMIC: | |
1608 | tga_ret = gen_rtx_REG (Pmode, R0_REG); | |
13ecc9e0 | 1609 | emit_call_insn (gen_tls_global_dynamic (tga_ret, op1)); |
463f02cd KK |
1610 | op1 = tga_ret; |
1611 | break; | |
1612 | ||
1613 | case TLS_MODEL_LOCAL_DYNAMIC: | |
1614 | tga_ret = gen_rtx_REG (Pmode, R0_REG); | |
13ecc9e0 | 1615 | emit_call_insn (gen_tls_local_dynamic (tga_ret, op1)); |
463f02cd KK |
1616 | |
1617 | tmp = gen_reg_rtx (Pmode); | |
1618 | emit_move_insn (tmp, tga_ret); | |
1619 | ||
1620 | if (register_operand (op0, Pmode)) | |
1621 | tmp2 = op0; | |
1622 | else | |
1623 | tmp2 = gen_reg_rtx (Pmode); | |
1624 | ||
1625 | emit_insn (gen_symDTPOFF2reg (tmp2, op1, tmp)); | |
1626 | op1 = tmp2; | |
1627 | break; | |
1628 | ||
1629 | case TLS_MODEL_INITIAL_EXEC: | |
1630 | if (! flag_pic) | |
b741fa00 | 1631 | { |
b92c3264 KK |
1632 | /* Don't schedule insns for getting GOT address when |
1633 | the first scheduling is enabled, to avoid spill | |
1634 | failures for R0. */ | |
1635 | if (flag_schedule_insns) | |
1636 | emit_insn (gen_blockage ()); | |
b741fa00 | 1637 | emit_insn (gen_GOTaddr2picreg ()); |
c41c1387 | 1638 | emit_use (gen_rtx_REG (SImode, PIC_REG)); |
b92c3264 KK |
1639 | if (flag_schedule_insns) |
1640 | emit_insn (gen_blockage ()); | |
b741fa00 | 1641 | } |
b3a13419 | 1642 | tga_op1 = !can_create_pseudo_p () ? op0 : gen_reg_rtx (Pmode); |
463f02cd KK |
1643 | tmp = gen_sym2GOTTPOFF (op1); |
1644 | emit_insn (gen_tls_initial_exec (tga_op1, tmp)); | |
1645 | op1 = tga_op1; | |
1646 | break; | |
1647 | ||
1648 | case TLS_MODEL_LOCAL_EXEC: | |
1649 | tmp2 = gen_reg_rtx (Pmode); | |
1650 | emit_insn (gen_load_gbr (tmp2)); | |
1651 | tmp = gen_reg_rtx (Pmode); | |
1652 | emit_insn (gen_symTPOFF2reg (tmp, op1)); | |
463f02cd KK |
1653 | |
1654 | if (register_operand (op0, Pmode)) | |
1655 | op1 = op0; | |
1656 | else | |
1657 | op1 = gen_reg_rtx (Pmode); | |
1658 | ||
1659 | emit_insn (gen_addsi3 (op1, tmp, tmp2)); | |
1660 | break; | |
1661 | ||
1662 | default: | |
f5b9e7c9 | 1663 | gcc_unreachable (); |
463f02cd | 1664 | } |
6058b2fa KK |
1665 | if (opc) |
1666 | emit_insn (gen_addsi3 (op1, op1, force_reg (SImode, opc))); | |
463f02cd KK |
1667 | operands[1] = op1; |
1668 | } | |
1669 | } | |
1670 | ||
8aa2a305 JW |
1671 | return 0; |
1672 | } | |
1673 | ||
78d310c2 R |
1674 | enum rtx_code |
1675 | prepare_cbranch_operands (rtx *operands, enum machine_mode mode, | |
1676 | enum rtx_code comparison) | |
1677 | { | |
1678 | rtx op1; | |
1679 | rtx scratch = NULL_RTX; | |
1680 | ||
b757d36f | 1681 | if (comparison == LAST_AND_UNUSED_RTX_CODE) |
78d310c2 R |
1682 | comparison = GET_CODE (operands[0]); |
1683 | else | |
1684 | scratch = operands[4]; | |
f3536097 SZ |
1685 | if (CONST_INT_P (operands[1]) |
1686 | && !CONST_INT_P (operands[2])) | |
78d310c2 R |
1687 | { |
1688 | rtx tmp = operands[1]; | |
1689 | ||
1690 | operands[1] = operands[2]; | |
1691 | operands[2] = tmp; | |
1692 | comparison = swap_condition (comparison); | |
1693 | } | |
f3536097 | 1694 | if (CONST_INT_P (operands[2])) |
78d310c2 R |
1695 | { |
1696 | HOST_WIDE_INT val = INTVAL (operands[2]); | |
1697 | if ((val == -1 || val == -0x81) | |
1698 | && (comparison == GT || comparison == LE)) | |
1699 | { | |
1700 | comparison = (comparison == GT) ? GE : LT; | |
1701 | operands[2] = gen_int_mode (val + 1, mode); | |
1702 | } | |
1703 | else if ((val == 1 || val == 0x80) | |
1704 | && (comparison == GE || comparison == LT)) | |
1705 | { | |
1706 | comparison = (comparison == GE) ? GT : LE; | |
1707 | operands[2] = gen_int_mode (val - 1, mode); | |
1708 | } | |
1709 | else if (val == 1 && (comparison == GEU || comparison == LTU)) | |
1710 | { | |
1711 | comparison = (comparison == GEU) ? NE : EQ; | |
1712 | operands[2] = CONST0_RTX (mode); | |
1713 | } | |
1714 | else if (val == 0x80 && (comparison == GEU || comparison == LTU)) | |
1715 | { | |
1716 | comparison = (comparison == GEU) ? GTU : LEU; | |
1717 | operands[2] = gen_int_mode (val - 1, mode); | |
1718 | } | |
1719 | else if (val == 0 && (comparison == GTU || comparison == LEU)) | |
1720 | comparison = (comparison == GTU) ? NE : EQ; | |
1721 | else if (mode == SImode | |
1722 | && ((val == 0x7fffffff | |
1723 | && (comparison == GTU || comparison == LEU)) | |
1724 | || ((unsigned HOST_WIDE_INT) val | |
1725 | == (unsigned HOST_WIDE_INT) 0x7fffffff + 1 | |
1726 | && (comparison == GEU || comparison == LTU)))) | |
1727 | { | |
1728 | comparison = (comparison == GTU || comparison == GEU) ? LT : GE; | |
1729 | operands[2] = CONST0_RTX (mode); | |
1730 | } | |
1731 | } | |
1732 | op1 = operands[1]; | |
b3a13419 | 1733 | if (can_create_pseudo_p ()) |
78d310c2 R |
1734 | operands[1] = force_reg (mode, op1); |
1735 | /* When we are handling DImode comparisons, we want to keep constants so | |
1736 | that we can optimize the component comparisons; however, memory loads | |
1737 | are better issued as a whole so that they can be scheduled well. | |
1738 | SImode equality comparisons allow I08 constants, but only when they | |
1739 | compare r0. Hence, if operands[1] has to be loaded from somewhere else | |
1740 | into a register, that register might as well be r0, and we allow the | |
1741 | constant. If it is already in a register, this is likely to be | |
2f8e468b | 1742 | allocated to a different hard register, thus we load the constant into |
78d310c2 R |
1743 | a register unless it is zero. */ |
1744 | if (!REG_P (operands[2]) | |
f3536097 | 1745 | && (!CONST_INT_P (operands[2]) |
78d310c2 R |
1746 | || (mode == SImode && operands[2] != CONST0_RTX (SImode) |
1747 | && ((comparison != EQ && comparison != NE) | |
1748 | || (REG_P (op1) && REGNO (op1) != R0_REG) | |
32a7ab3d | 1749 | || !satisfies_constraint_I08 (operands[2]))))) |
78d310c2 R |
1750 | { |
1751 | if (scratch && GET_MODE (scratch) == mode) | |
1752 | { | |
1753 | emit_move_insn (scratch, operands[2]); | |
1754 | operands[2] = scratch; | |
1755 | } | |
b3a13419 | 1756 | else if (can_create_pseudo_p ()) |
78d310c2 R |
1757 | operands[2] = force_reg (mode, operands[2]); |
1758 | } | |
1759 | return comparison; | |
1760 | } | |
1761 | ||
1762 | void | |
1763 | expand_cbranchsi4 (rtx *operands, enum rtx_code comparison, int probability) | |
1764 | { | |
1765 | rtx (*branch_expander) (rtx) = gen_branch_true; | |
1766 | rtx jump; | |
1767 | ||
1768 | comparison = prepare_cbranch_operands (operands, SImode, comparison); | |
1769 | switch (comparison) | |
1770 | { | |
1771 | case NE: case LT: case LE: case LTU: case LEU: | |
1772 | comparison = reverse_condition (comparison); | |
1773 | branch_expander = gen_branch_false; | |
1774 | default: ; | |
1775 | } | |
1776 | emit_insn (gen_rtx_SET (VOIDmode, gen_rtx_REG (SImode, T_REG), | |
1777 | gen_rtx_fmt_ee (comparison, SImode, | |
1778 | operands[1], operands[2]))); | |
1779 | jump = emit_jump_insn (branch_expander (operands[3])); | |
1780 | if (probability >= 0) | |
b757d36f | 1781 | add_reg_note (jump, REG_BR_PROB, GEN_INT (probability)); |
78d310c2 R |
1782 | |
1783 | } | |
1784 | ||
1785 | /* ??? How should we distribute probabilities when more than one branch | |
1786 | is generated. So far we only have soem ad-hoc observations: | |
1787 | - If the operands are random, they are likely to differ in both parts. | |
1788 | - If comparing items in a hash chain, the operands are random or equal; | |
1789 | operation should be EQ or NE. | |
1790 | - If items are searched in an ordered tree from the root, we can expect | |
1791 | the highpart to be unequal about half of the time; operation should be | |
2f8e468b | 1792 | an inequality comparison, operands non-constant, and overall probability |
78d310c2 R |
1793 | about 50%. Likewise for quicksort. |
1794 | - Range checks will be often made against constants. Even if we assume for | |
1795 | simplicity an even distribution of the non-constant operand over a | |
1796 | sub-range here, the same probability could be generated with differently | |
1797 | wide sub-ranges - as long as the ratio of the part of the subrange that | |
1798 | is before the threshold to the part that comes after the threshold stays | |
1799 | the same. Thus, we can't really tell anything here; | |
1800 | assuming random distribution is at least simple. | |
1801 | */ | |
1802 | ||
1803 | bool | |
1804 | expand_cbranchdi4 (rtx *operands, enum rtx_code comparison) | |
1805 | { | |
1806 | enum rtx_code msw_taken, msw_skip, lsw_taken; | |
cb3d3842 | 1807 | rtx skip_label = NULL_RTX; |
78d310c2 R |
1808 | rtx op1h, op1l, op2h, op2l; |
1809 | int num_branches; | |
1810 | int prob, rev_prob; | |
1811 | int msw_taken_prob = -1, msw_skip_prob = -1, lsw_taken_prob = -1; | |
6f8eec29 | 1812 | rtx scratch = operands[4]; |
78d310c2 R |
1813 | |
1814 | comparison = prepare_cbranch_operands (operands, DImode, comparison); | |
1815 | op1h = gen_highpart_mode (SImode, DImode, operands[1]); | |
1816 | op2h = gen_highpart_mode (SImode, DImode, operands[2]); | |
1817 | op1l = gen_lowpart (SImode, operands[1]); | |
1818 | op2l = gen_lowpart (SImode, operands[2]); | |
b757d36f | 1819 | msw_taken = msw_skip = lsw_taken = LAST_AND_UNUSED_RTX_CODE; |
78d310c2 R |
1820 | prob = split_branch_probability; |
1821 | rev_prob = REG_BR_PROB_BASE - prob; | |
1822 | switch (comparison) | |
1823 | { | |
1824 | /* ??? Should we use the cmpeqdi_t pattern for equality comparisons? | |
1825 | That costs 1 cycle more when the first branch can be predicted taken, | |
1826 | but saves us mispredicts because only one branch needs prediction. | |
1827 | It also enables generating the cmpeqdi_t-1 pattern. */ | |
1828 | case EQ: | |
1829 | if (TARGET_CMPEQDI_T) | |
1830 | { | |
1831 | emit_insn (gen_cmpeqdi_t (operands[1], operands[2])); | |
1832 | emit_jump_insn (gen_branch_true (operands[3])); | |
1833 | return true; | |
1834 | } | |
1835 | msw_skip = NE; | |
1836 | lsw_taken = EQ; | |
1837 | if (prob >= 0) | |
1838 | { | |
1839 | /* If we had more precision, we'd use rev_prob - (rev_prob >> 32) . | |
1840 | */ | |
1841 | msw_skip_prob = rev_prob; | |
1842 | if (REG_BR_PROB_BASE <= 65535) | |
1843 | lsw_taken_prob = prob ? REG_BR_PROB_BASE : 0; | |
1844 | else | |
1845 | { | |
1846 | gcc_assert (HOST_BITS_PER_WIDEST_INT >= 64); | |
1847 | lsw_taken_prob | |
1848 | = (prob | |
1849 | ? (REG_BR_PROB_BASE | |
1850 | - ((HOST_WIDEST_INT) REG_BR_PROB_BASE * rev_prob | |
1851 | / ((HOST_WIDEST_INT) prob << 32))) | |
1852 | : 0); | |
1853 | } | |
1854 | } | |
1855 | break; | |
1856 | case NE: | |
1857 | if (TARGET_CMPEQDI_T) | |
1858 | { | |
1859 | emit_insn (gen_cmpeqdi_t (operands[1], operands[2])); | |
1860 | emit_jump_insn (gen_branch_false (operands[3])); | |
1861 | return true; | |
1862 | } | |
1863 | msw_taken = NE; | |
6f8eec29 | 1864 | msw_taken_prob = prob; |
78d310c2 R |
1865 | lsw_taken = NE; |
1866 | lsw_taken_prob = 0; | |
1867 | break; | |
1868 | case GTU: case GT: | |
1869 | msw_taken = comparison; | |
f3536097 | 1870 | if (CONST_INT_P (op2l) && INTVAL (op2l) == -1) |
78d310c2 R |
1871 | break; |
1872 | if (comparison != GTU || op2h != CONST0_RTX (SImode)) | |
1873 | msw_skip = swap_condition (msw_taken); | |
1874 | lsw_taken = GTU; | |
1875 | break; | |
1876 | case GEU: case GE: | |
1877 | if (op2l == CONST0_RTX (SImode)) | |
1878 | msw_taken = comparison; | |
1879 | else | |
1880 | { | |
1881 | msw_taken = comparison == GE ? GT : GTU; | |
1882 | msw_skip = swap_condition (msw_taken); | |
1883 | lsw_taken = GEU; | |
1884 | } | |
1885 | break; | |
1886 | case LTU: case LT: | |
1887 | msw_taken = comparison; | |
1888 | if (op2l == CONST0_RTX (SImode)) | |
1889 | break; | |
1890 | msw_skip = swap_condition (msw_taken); | |
1891 | lsw_taken = LTU; | |
1892 | break; | |
1893 | case LEU: case LE: | |
f3536097 | 1894 | if (CONST_INT_P (op2l) && INTVAL (op2l) == -1) |
78d310c2 R |
1895 | msw_taken = comparison; |
1896 | else | |
1897 | { | |
1898 | lsw_taken = LEU; | |
1899 | if (comparison == LE) | |
1900 | msw_taken = LT; | |
1901 | else if (op2h != CONST0_RTX (SImode)) | |
1902 | msw_taken = LTU; | |
1903 | else | |
1904 | break; | |
1905 | msw_skip = swap_condition (msw_taken); | |
1906 | } | |
1907 | break; | |
1908 | default: return false; | |
1909 | } | |
b757d36f KK |
1910 | num_branches = ((msw_taken != LAST_AND_UNUSED_RTX_CODE) |
1911 | + (msw_skip != LAST_AND_UNUSED_RTX_CODE) | |
1912 | + (lsw_taken != LAST_AND_UNUSED_RTX_CODE)); | |
78d310c2 R |
1913 | if (comparison != EQ && comparison != NE && num_branches > 1) |
1914 | { | |
1915 | if (!CONSTANT_P (operands[2]) | |
1916 | && prob >= (int) (REG_BR_PROB_BASE * 3 / 8U) | |
1917 | && prob <= (int) (REG_BR_PROB_BASE * 5 / 8U)) | |
1918 | { | |
1919 | msw_taken_prob = prob / 2U; | |
1920 | msw_skip_prob | |
1921 | = REG_BR_PROB_BASE * rev_prob / (REG_BR_PROB_BASE + rev_prob); | |
1922 | lsw_taken_prob = prob; | |
1923 | } | |
1924 | else | |
1925 | { | |
1926 | msw_taken_prob = prob; | |
1927 | msw_skip_prob = REG_BR_PROB_BASE; | |
1928 | /* ??? If we have a constant op2h, should we use that when | |
1929 | calculating lsw_taken_prob? */ | |
1930 | lsw_taken_prob = prob; | |
1931 | } | |
1932 | } | |
1933 | operands[1] = op1h; | |
1934 | operands[2] = op2h; | |
1935 | operands[4] = NULL_RTX; | |
6f8eec29 | 1936 | if (reload_completed |
99052692 KK |
1937 | && ! arith_reg_or_0_operand (op2h, SImode) |
1938 | && (true_regnum (op1h) || (comparison != EQ && comparison != NE)) | |
b757d36f KK |
1939 | && (msw_taken != LAST_AND_UNUSED_RTX_CODE |
1940 | || msw_skip != LAST_AND_UNUSED_RTX_CODE)) | |
6f8eec29 R |
1941 | { |
1942 | emit_move_insn (scratch, operands[2]); | |
1943 | operands[2] = scratch; | |
1944 | } | |
b757d36f | 1945 | if (msw_taken != LAST_AND_UNUSED_RTX_CODE) |
78d310c2 | 1946 | expand_cbranchsi4 (operands, msw_taken, msw_taken_prob); |
b757d36f | 1947 | if (msw_skip != LAST_AND_UNUSED_RTX_CODE) |
78d310c2 R |
1948 | { |
1949 | rtx taken_label = operands[3]; | |
1950 | ||
d60819f8 CB |
1951 | /* Operands were possibly modified, but msw_skip doesn't expect this. |
1952 | Always use the original ones. */ | |
b757d36f | 1953 | if (msw_taken != LAST_AND_UNUSED_RTX_CODE) |
d60819f8 CB |
1954 | { |
1955 | operands[1] = op1h; | |
1956 | operands[2] = op2h; | |
1957 | } | |
1958 | ||
78d310c2 R |
1959 | operands[3] = skip_label = gen_label_rtx (); |
1960 | expand_cbranchsi4 (operands, msw_skip, msw_skip_prob); | |
1961 | operands[3] = taken_label; | |
1962 | } | |
1963 | operands[1] = op1l; | |
1964 | operands[2] = op2l; | |
b757d36f | 1965 | if (lsw_taken != LAST_AND_UNUSED_RTX_CODE) |
6f8eec29 R |
1966 | { |
1967 | if (reload_completed | |
99052692 KK |
1968 | && ! arith_reg_or_0_operand (op2l, SImode) |
1969 | && (true_regnum (op1l) || (lsw_taken != EQ && lsw_taken != NE))) | |
1970 | { | |
1971 | emit_move_insn (scratch, operands[2]); | |
1972 | operands[2] = scratch; | |
1973 | } | |
6f8eec29 R |
1974 | expand_cbranchsi4 (operands, lsw_taken, lsw_taken_prob); |
1975 | } | |
b757d36f | 1976 | if (msw_skip != LAST_AND_UNUSED_RTX_CODE) |
78d310c2 R |
1977 | emit_label (skip_label); |
1978 | return true; | |
1979 | } | |
1980 | ||
f90b7a5a PB |
1981 | /* Emit INSN, possibly in a PARALLEL with an USE of fpscr for SH4. */ |
1982 | ||
1983 | static void | |
1984 | sh_emit_set_t_insn (rtx insn, enum machine_mode mode) | |
1985 | { | |
1986 | if ((TARGET_SH4 || TARGET_SH2A) && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
1987 | { | |
1988 | insn = gen_rtx_PARALLEL (VOIDmode, | |
1989 | gen_rtvec (2, insn, | |
1990 | gen_rtx_USE (VOIDmode, get_fpscr_rtx ()))); | |
1991 | (mode == SFmode ? emit_sf_insn : emit_df_insn) (insn); | |
1992 | } | |
1993 | else | |
1994 | emit_insn (insn); | |
1995 | } | |
1996 | ||
8aa2a305 | 1997 | /* Prepare the operands for an scc instruction; make sure that the |
f90b7a5a PB |
1998 | compare has been done and the result is in T_REG. */ |
1999 | void | |
2000 | sh_emit_scc_to_t (enum rtx_code code, rtx op0, rtx op1) | |
8aa2a305 | 2001 | { |
c5c76735 | 2002 | rtx t_reg = gen_rtx_REG (SImode, T_REG); |
8aa2a305 | 2003 | enum rtx_code oldcode = code; |
9374bd85 | 2004 | enum machine_mode mode; |
8aa2a305 JW |
2005 | |
2006 | /* First need a compare insn. */ | |
2007 | switch (code) | |
bc45ade3 | 2008 | { |
8aa2a305 JW |
2009 | case NE: |
2010 | /* It isn't possible to handle this case. */ | |
f5b9e7c9 | 2011 | gcc_unreachable (); |
8aa2a305 JW |
2012 | case LT: |
2013 | code = GT; | |
2014 | break; | |
2015 | case LE: | |
2016 | code = GE; | |
2017 | break; | |
2018 | case LTU: | |
2019 | code = GTU; | |
2020 | break; | |
2021 | case LEU: | |
2022 | code = GEU; | |
2023 | break; | |
10db12e0 R |
2024 | default: |
2025 | break; | |
bc45ade3 | 2026 | } |
8aa2a305 | 2027 | if (code != oldcode) |
b9654711 | 2028 | { |
f90b7a5a PB |
2029 | rtx tmp = op0; |
2030 | op0 = op1; | |
2031 | op1 = tmp; | |
0d7e008e | 2032 | } |
b9654711 | 2033 | |
f90b7a5a | 2034 | mode = GET_MODE (op0); |
9374bd85 | 2035 | if (mode == VOIDmode) |
f90b7a5a | 2036 | mode = GET_MODE (op1); |
9374bd85 | 2037 | |
f90b7a5a | 2038 | op0 = force_reg (mode, op0); |
1245df60 | 2039 | if ((code != EQ && code != NE |
f90b7a5a | 2040 | && (op1 != const0_rtx |
1245df60 | 2041 | || code == GTU || code == GEU || code == LTU || code == LEU)) |
f90b7a5a | 2042 | || (mode == DImode && op1 != const0_rtx) |
3a8699c7 | 2043 | || (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT)) |
f90b7a5a | 2044 | op1 = force_reg (mode, op1); |
0d7e008e | 2045 | |
f90b7a5a PB |
2046 | sh_emit_set_t_insn (gen_rtx_SET (VOIDmode, t_reg, |
2047 | gen_rtx_fmt_ee (code, SImode, op0, op1)), | |
2048 | mode); | |
2049 | } | |
0d7e008e | 2050 | |
f90b7a5a PB |
2051 | rtx |
2052 | sh_emit_cheap_store_flag (enum machine_mode mode, enum rtx_code code, | |
2053 | rtx op0, rtx op1) | |
2054 | { | |
2055 | rtx target = gen_reg_rtx (SImode); | |
2056 | rtx tmp; | |
2057 | ||
2058 | gcc_assert (TARGET_SHMEDIA); | |
2059 | switch (code) | |
2060 | { | |
2061 | case EQ: | |
2062 | case GT: | |
2063 | case LT: | |
2064 | case UNORDERED: | |
2065 | case GTU: | |
2066 | case LTU: | |
2067 | tmp = gen_rtx_fmt_ee (code, SImode, op0, op1); | |
2068 | emit_insn (gen_cstore4_media (target, tmp, op0, op1)); | |
2069 | code = NE; | |
2070 | break; | |
2071 | ||
2072 | case NE: | |
2073 | case GE: | |
2074 | case LE: | |
2075 | case ORDERED: | |
2076 | case GEU: | |
2077 | case LEU: | |
2078 | tmp = gen_rtx_fmt_ee (reverse_condition (code), mode, op0, op1); | |
2079 | emit_insn (gen_cstore4_media (target, tmp, op0, op1)); | |
2080 | code = EQ; | |
2081 | break; | |
2082 | ||
2083 | case UNEQ: | |
2084 | case UNGE: | |
2085 | case UNGT: | |
2086 | case UNLE: | |
2087 | case UNLT: | |
2088 | case LTGT: | |
2089 | return NULL_RTX; | |
2090 | ||
2091 | default: | |
2092 | gcc_unreachable (); | |
2093 | } | |
2094 | ||
2095 | if (mode == DImode) | |
2096 | { | |
2097 | rtx t2 = gen_reg_rtx (DImode); | |
2098 | emit_insn (gen_extendsidi2 (t2, target)); | |
2099 | target = t2; | |
2100 | } | |
2101 | ||
2102 | return gen_rtx_fmt_ee (code, VOIDmode, target, const0_rtx); | |
bc45ade3 SC |
2103 | } |
2104 | ||
8aa2a305 JW |
2105 | /* Called from the md file, set up the operands of a compare instruction. */ |
2106 | ||
2107 | void | |
f90b7a5a | 2108 | sh_emit_compare_and_branch (rtx *operands, enum machine_mode mode) |
bc45ade3 | 2109 | { |
f90b7a5a PB |
2110 | enum rtx_code code = GET_CODE (operands[0]); |
2111 | enum rtx_code branch_code; | |
2112 | rtx op0 = operands[1]; | |
2113 | rtx op1 = operands[2]; | |
2114 | rtx insn, tem; | |
2115 | bool need_ccmpeq = false; | |
2116 | ||
2117 | if (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
bc45ade3 | 2118 | { |
f90b7a5a PB |
2119 | op0 = force_reg (mode, op0); |
2120 | op1 = force_reg (mode, op1); | |
bc45ade3 | 2121 | } |
f90b7a5a | 2122 | else |
1245df60 | 2123 | { |
f90b7a5a PB |
2124 | if (code != EQ || mode == DImode) |
2125 | { | |
2126 | /* Force args into regs, since we can't use constants here. */ | |
2127 | op0 = force_reg (mode, op0); | |
2128 | if (op1 != const0_rtx || code == GTU || code == GEU) | |
2129 | op1 = force_reg (mode, op1); | |
2130 | } | |
1245df60 | 2131 | } |
f90b7a5a PB |
2132 | |
2133 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
2134 | { | |
2135 | if (code == LT | |
2136 | || (code == LE && TARGET_IEEE && TARGET_SH2E) | |
2137 | || (code == GE && !(TARGET_IEEE && TARGET_SH2E))) | |
2138 | { | |
2139 | tem = op0, op0 = op1, op1 = tem; | |
2140 | code = swap_condition (code); | |
2141 | } | |
2142 | ||
2143 | /* GE becomes fcmp/gt+fcmp/eq, for SH2E and TARGET_IEEE only. */ | |
2144 | if (code == GE) | |
2145 | { | |
2146 | gcc_assert (TARGET_IEEE && TARGET_SH2E); | |
2147 | need_ccmpeq = true; | |
2148 | code = GT; | |
2149 | } | |
2150 | ||
2151 | /* Now we can have EQ, NE, GT, LE. NE and LE are then transformed | |
2152 | to EQ/GT respectively. */ | |
2153 | gcc_assert (code == EQ || code == GT || code == NE || code == LE); | |
2154 | } | |
2155 | ||
2156 | switch (code) | |
2157 | { | |
2158 | case EQ: | |
2159 | case GT: | |
2160 | case GE: | |
2161 | case GTU: | |
2162 | case GEU: | |
2163 | branch_code = code; | |
2164 | break; | |
2165 | case NE: | |
2166 | case LT: | |
2167 | case LE: | |
2168 | case LTU: | |
2169 | case LEU: | |
2170 | branch_code = reverse_condition (code); | |
2171 | break; | |
2172 | default: | |
2173 | gcc_unreachable (); | |
2174 | } | |
2175 | ||
2176 | insn = gen_rtx_SET (VOIDmode, | |
2177 | gen_rtx_REG (SImode, T_REG), | |
2178 | gen_rtx_fmt_ee (branch_code, SImode, op0, op1)); | |
2179 | ||
2180 | sh_emit_set_t_insn (insn, mode); | |
2181 | if (need_ccmpeq) | |
2182 | sh_emit_set_t_insn (gen_ieee_ccmpeqsf_t (op0, op1), mode); | |
2183 | ||
2184 | if (branch_code == code) | |
2185 | emit_jump_insn (gen_branch_true (operands[3])); | |
1245df60 | 2186 | else |
f90b7a5a PB |
2187 | emit_jump_insn (gen_branch_false (operands[3])); |
2188 | } | |
2189 | ||
2190 | void | |
2191 | sh_emit_compare_and_set (rtx *operands, enum machine_mode mode) | |
2192 | { | |
2193 | enum rtx_code code = GET_CODE (operands[1]); | |
2194 | rtx op0 = operands[2]; | |
2195 | rtx op1 = operands[3]; | |
2196 | rtx lab = NULL_RTX; | |
2197 | bool invert = false; | |
2198 | rtx tem; | |
2199 | ||
2200 | op0 = force_reg (mode, op0); | |
2201 | if ((code != EQ && code != NE | |
2202 | && (op1 != const0_rtx | |
2203 | || code == GTU || code == GEU || code == LTU || code == LEU)) | |
2204 | || (mode == DImode && op1 != const0_rtx) | |
2205 | || (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT)) | |
2206 | op1 = force_reg (mode, op1); | |
2207 | ||
2208 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
225e4f43 | 2209 | { |
f90b7a5a PB |
2210 | if (code == LT || code == LE) |
2211 | { | |
2212 | code = swap_condition (code); | |
2213 | tem = op0, op0 = op1, op1 = tem; | |
2214 | } | |
2215 | if (code == GE) | |
2216 | { | |
2217 | if (TARGET_IEEE) | |
2218 | { | |
2219 | lab = gen_label_rtx (); | |
2220 | sh_emit_scc_to_t (EQ, op0, op1); | |
2221 | emit_jump_insn (gen_branch_true (lab)); | |
2222 | code = GT; | |
2223 | } | |
2224 | else | |
2225 | { | |
2226 | code = LT; | |
2227 | invert = true; | |
2228 | } | |
2229 | } | |
225e4f43 | 2230 | } |
f90b7a5a PB |
2231 | |
2232 | if (code == NE) | |
2233 | { | |
2234 | code = EQ; | |
2235 | invert = true; | |
2236 | } | |
2237 | ||
2238 | sh_emit_scc_to_t (code, op0, op1); | |
2239 | if (lab) | |
2240 | emit_label (lab); | |
2241 | if (invert) | |
2242 | emit_insn (gen_movnegt (operands[0])); | |
225e4f43 | 2243 | else |
f90b7a5a | 2244 | emit_move_insn (operands[0], gen_rtx_REG (SImode, T_REG)); |
bc45ade3 SC |
2245 | } |
2246 | \f | |
16bea517 | 2247 | /* Functions to output assembly code. */ |
bc45ade3 | 2248 | |
b9654711 | 2249 | /* Return a sequence of instructions to perform DI or DF move. |
bc45ade3 | 2250 | |
b9654711 | 2251 | Since the SH cannot move a DI or DF in one instruction, we have |
16bea517 | 2252 | to take care when we see overlapping source and dest registers. */ |
0d7e008e | 2253 | |
318881c0 | 2254 | const char * |
cf277499 SB |
2255 | output_movedouble (rtx insn ATTRIBUTE_UNUSED, rtx operands[], |
2256 | enum machine_mode mode) | |
bc45ade3 | 2257 | { |
b9654711 SC |
2258 | rtx dst = operands[0]; |
2259 | rtx src = operands[1]; | |
b9654711 | 2260 | |
f3536097 | 2261 | if (MEM_P (dst) |
16bea517 | 2262 | && GET_CODE (XEXP (dst, 0)) == PRE_DEC) |
8aa2a305 JW |
2263 | return "mov.l %T1,%0\n\tmov.l %1,%0"; |
2264 | ||
b9654711 SC |
2265 | if (register_operand (dst, mode) |
2266 | && register_operand (src, mode)) | |
bc45ade3 | 2267 | { |
b9654711 | 2268 | if (REGNO (src) == MACH_REG) |
00f8ff66 | 2269 | return "sts mach,%S0\n\tsts macl,%R0"; |
bc45ade3 | 2270 | |
8aa2a305 JW |
2271 | /* When mov.d r1,r2 do r2->r3 then r1->r2; |
2272 | when mov.d r1,r0 do r1->r0 then r2->r1. */ | |
b9654711 SC |
2273 | |
2274 | if (REGNO (src) + 1 == REGNO (dst)) | |
5325c0fa | 2275 | return "mov %T1,%T0\n\tmov %1,%0"; |
b9654711 | 2276 | else |
5325c0fa | 2277 | return "mov %1,%0\n\tmov %T1,%T0"; |
b9654711 | 2278 | } |
f3536097 | 2279 | else if (CONST_INT_P (src)) |
bc45ade3 | 2280 | { |
8aa2a305 JW |
2281 | if (INTVAL (src) < 0) |
2282 | output_asm_insn ("mov #-1,%S0", operands); | |
bc45ade3 | 2283 | else |
8aa2a305 | 2284 | output_asm_insn ("mov #0,%S0", operands); |
bc45ade3 | 2285 | |
8aa2a305 | 2286 | return "mov %1,%R0"; |
0d7e008e | 2287 | } |
f3536097 | 2288 | else if (MEM_P (src)) |
bc45ade3 | 2289 | { |
8aa2a305 | 2290 | int ptrreg = -1; |
b9654711 SC |
2291 | int dreg = REGNO (dst); |
2292 | rtx inside = XEXP (src, 0); | |
bc45ade3 | 2293 | |
f5b9e7c9 | 2294 | switch (GET_CODE (inside)) |
bc45ade3 | 2295 | { |
f5b9e7c9 NS |
2296 | case REG: |
2297 | ptrreg = REGNO (inside); | |
2298 | break; | |
2299 | ||
2300 | case SUBREG: | |
2301 | ptrreg = subreg_regno (inside); | |
2302 | break; | |
2303 | ||
2304 | case PLUS: | |
8aa2a305 JW |
2305 | ptrreg = REGNO (XEXP (inside, 0)); |
2306 | /* ??? A r0+REG address shouldn't be possible here, because it isn't | |
2307 | an offsettable address. Unfortunately, offsettable addresses use | |
2308 | QImode to check the offset, and a QImode offsettable address | |
2309 | requires r0 for the other operand, which is not currently | |
2310 | supported, so we can't use the 'o' constraint. | |
2311 | Thus we must check for and handle r0+REG addresses here. | |
2312 | We punt for now, since this is likely very rare. */ | |
f3536097 | 2313 | gcc_assert (!REG_P (XEXP (inside, 1))); |
f5b9e7c9 NS |
2314 | break; |
2315 | ||
2316 | case LABEL_REF: | |
2317 | return "mov.l %1,%0\n\tmov.l %1+4,%T0"; | |
2318 | case POST_INC: | |
2319 | return "mov.l %1,%0\n\tmov.l %1,%T0"; | |
2320 | default: | |
2321 | gcc_unreachable (); | |
bc45ade3 | 2322 | } |
bc45ade3 | 2323 | |
8aa2a305 JW |
2324 | /* Work out the safe way to copy. Copy into the second half first. */ |
2325 | if (dreg == ptrreg) | |
2326 | return "mov.l %T1,%T0\n\tmov.l %1,%0"; | |
bc45ade3 SC |
2327 | } |
2328 | ||
00f8ff66 | 2329 | return "mov.l %1,%0\n\tmov.l %T1,%T0"; |
bc45ade3 SC |
2330 | } |
2331 | ||
8aa2a305 JW |
2332 | /* Print an instruction which would have gone into a delay slot after |
2333 | another instruction, but couldn't because the other instruction expanded | |
2334 | into a sequence where putting the slot insn at the end wouldn't work. */ | |
0d7e008e | 2335 | |
8aa2a305 | 2336 | static void |
cf277499 | 2337 | print_slot (rtx insn) |
8aa2a305 | 2338 | { |
c9d691e9 | 2339 | final_scan_insn (XVECEXP (insn, 0, 1), asm_out_file, optimize, 1, NULL); |
b9654711 | 2340 | |
8aa2a305 JW |
2341 | INSN_DELETED_P (XVECEXP (insn, 0, 1)) = 1; |
2342 | } | |
0d7e008e | 2343 | |
318881c0 | 2344 | const char * |
cf277499 | 2345 | output_far_jump (rtx insn, rtx op) |
0d7e008e | 2346 | { |
0a2aaacc | 2347 | struct { rtx lab, reg, op; } this_jmp; |
950a3816 | 2348 | rtx braf_base_lab = NULL_RTX; |
318881c0 | 2349 | const char *jump; |
1245df60 | 2350 | int far; |
9d98a694 | 2351 | int offset = branch_dest (insn) - INSN_ADDRESSES (INSN_UID (insn)); |
10f4f635 | 2352 | rtx prev; |
0d7e008e | 2353 | |
0a2aaacc | 2354 | this_jmp.lab = gen_label_rtx (); |
b9654711 | 2355 | |
1a953d0f R |
2356 | if (TARGET_SH2 |
2357 | && offset >= -32764 | |
2358 | && offset - get_attr_length (insn) <= 32766) | |
1245df60 R |
2359 | { |
2360 | far = 0; | |
e6dfd05f | 2361 | jump = "mov.w %O0,%1; braf %1"; |
1245df60 R |
2362 | } |
2363 | else | |
2364 | { | |
2365 | far = 1; | |
1a66cd67 | 2366 | if (flag_pic) |
e6dfd05f AO |
2367 | { |
2368 | if (TARGET_SH2) | |
2369 | jump = "mov.l %O0,%1; braf %1"; | |
2370 | else | |
2371 | jump = "mov.l r0,@-r15; mova %O0,r0; mov.l @r0,%1; add r0,%1; mov.l @r15+,r0; jmp @%1"; | |
2372 | } | |
1a66cd67 | 2373 | else |
e6dfd05f | 2374 | jump = "mov.l %O0,%1; jmp @%1"; |
1245df60 R |
2375 | } |
2376 | /* If we have a scratch register available, use it. */ | |
f3536097 | 2377 | if (NONJUMP_INSN_P ((prev = prev_nonnote_insn (insn))) |
10f4f635 | 2378 | && INSN_CODE (prev) == CODE_FOR_indirect_jump_scratch) |
1245df60 | 2379 | { |
0a2aaacc KG |
2380 | this_jmp.reg = SET_DEST (XVECEXP (PATTERN (prev), 0, 0)); |
2381 | if (REGNO (this_jmp.reg) == R0_REG && flag_pic && ! TARGET_SH2) | |
e6dfd05f | 2382 | jump = "mov.l r1,@-r15; mova %O0,r0; mov.l @r0,r1; add r1,r0; mov.l @r15+,r1; jmp @%1"; |
0a2aaacc | 2383 | output_asm_insn (jump, &this_jmp.lab); |
1245df60 R |
2384 | if (dbr_sequence_length ()) |
2385 | print_slot (final_sequence); | |
2386 | else | |
2387 | output_asm_insn ("nop", 0); | |
2388 | } | |
2389 | else | |
2390 | { | |
2391 | /* Output the delay slot insn first if any. */ | |
2392 | if (dbr_sequence_length ()) | |
2393 | print_slot (final_sequence); | |
2394 | ||
0a2aaacc | 2395 | this_jmp.reg = gen_rtx_REG (SImode, 13); |
fa5322fa AO |
2396 | /* We must keep the stack aligned to 8-byte boundaries on SH5. |
2397 | Fortunately, MACL is fixed and call-clobbered, and we never | |
2398 | need its value across jumps, so save r13 in it instead of in | |
2399 | the stack. */ | |
2400 | if (TARGET_SH5) | |
2401 | output_asm_insn ("lds r13, macl", 0); | |
2402 | else | |
2403 | output_asm_insn ("mov.l r13,@-r15", 0); | |
0a2aaacc | 2404 | output_asm_insn (jump, &this_jmp.lab); |
fa5322fa AO |
2405 | if (TARGET_SH5) |
2406 | output_asm_insn ("sts macl, r13", 0); | |
2407 | else | |
2408 | output_asm_insn ("mov.l @r15+,r13", 0); | |
1245df60 | 2409 | } |
e6dfd05f AO |
2410 | if (far && flag_pic && TARGET_SH2) |
2411 | { | |
2412 | braf_base_lab = gen_label_rtx (); | |
4977bab6 | 2413 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
e6dfd05f AO |
2414 | CODE_LABEL_NUMBER (braf_base_lab)); |
2415 | } | |
1245df60 R |
2416 | if (far) |
2417 | output_asm_insn (".align 2", 0); | |
0a2aaacc KG |
2418 | (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (this_jmp.lab)); |
2419 | this_jmp.op = op; | |
1a66cd67 | 2420 | if (far && flag_pic) |
e6dfd05f AO |
2421 | { |
2422 | if (TARGET_SH2) | |
0a2aaacc KG |
2423 | this_jmp.lab = braf_base_lab; |
2424 | output_asm_insn (".long %O2-%O0", &this_jmp.lab); | |
e6dfd05f | 2425 | } |
1a66cd67 | 2426 | else |
0a2aaacc | 2427 | output_asm_insn (far ? ".long %O2" : ".word %O2-%O0", &this_jmp.lab); |
0d7e008e SC |
2428 | return ""; |
2429 | } | |
bc45ade3 | 2430 | |
8aa2a305 JW |
2431 | /* Local label counter, used for constants in the pool and inside |
2432 | pattern branches. */ | |
2433 | ||
2434 | static int lf = 100; | |
2435 | ||
2436 | /* Output code for ordinary branches. */ | |
2437 | ||
318881c0 | 2438 | const char * |
cf277499 | 2439 | output_branch (int logic, rtx insn, rtx *operands) |
bc45ade3 | 2440 | { |
afbc5d86 | 2441 | switch (get_attr_length (insn)) |
bc45ade3 | 2442 | { |
33f7f353 JR |
2443 | case 6: |
2444 | /* This can happen if filling the delay slot has caused a forward | |
2445 | branch to exceed its range (we could reverse it, but only | |
2446 | when we know we won't overextend other branches; this should | |
2447 | best be handled by relaxation). | |
2448 | It can also happen when other condbranches hoist delay slot insn | |
1245df60 R |
2449 | from their destination, thus leading to code size increase. |
2450 | But the branch will still be in the range -4092..+4098 bytes. */ | |
bc45ade3 | 2451 | |
33f7f353 | 2452 | if (! TARGET_RELAX) |
1245df60 | 2453 | { |
33f7f353 JR |
2454 | int label = lf++; |
2455 | /* The call to print_slot will clobber the operands. */ | |
2456 | rtx op0 = operands[0]; | |
f676971a | 2457 | |
33f7f353 JR |
2458 | /* If the instruction in the delay slot is annulled (true), then |
2459 | there is no delay slot where we can put it now. The only safe | |
2460 | place for it is after the label. final will do that by default. */ | |
f676971a | 2461 | |
33f7f353 | 2462 | if (final_sequence |
bee9efbc KK |
2463 | && ! INSN_ANNULLED_BRANCH_P (XVECEXP (final_sequence, 0, 0)) |
2464 | && get_attr_length (XVECEXP (final_sequence, 0, 1))) | |
33f7f353 | 2465 | { |
afbc5d86 | 2466 | asm_fprintf (asm_out_file, "\tb%s%ss\t%LLF%d\n", logic ? "f" : "t", |
33f7f353 JR |
2467 | ASSEMBLER_DIALECT ? "/" : ".", label); |
2468 | print_slot (final_sequence); | |
2469 | } | |
2470 | else | |
afbc5d86 | 2471 | asm_fprintf (asm_out_file, "\tb%s\t%LLF%d\n", logic ? "f" : "t", label); |
f676971a | 2472 | |
afbc5d86 R |
2473 | output_asm_insn ("bra\t%l0", &op0); |
2474 | fprintf (asm_out_file, "\tnop\n"); | |
832a3292 | 2475 | (*targetm.asm_out.internal_label) (asm_out_file, "LF", label); |
f676971a | 2476 | |
33f7f353 | 2477 | return ""; |
1245df60 | 2478 | } |
33f7f353 JR |
2479 | /* When relaxing, handle this like a short branch. The linker |
2480 | will fix it up if it still doesn't fit after relaxation. */ | |
2481 | case 2: | |
2482 | return logic ? "bt%.\t%l0" : "bf%.\t%l0"; | |
3a8699c7 AO |
2483 | |
2484 | /* These are for SH2e, in which we have to account for the | |
2485 | extra nop because of the hardware bug in annulled branches. */ | |
2486 | case 8: | |
2487 | if (! TARGET_RELAX) | |
2488 | { | |
2489 | int label = lf++; | |
2490 | ||
f5b9e7c9 NS |
2491 | gcc_assert (!final_sequence |
2492 | || !(INSN_ANNULLED_BRANCH_P | |
2493 | (XVECEXP (final_sequence, 0, 0)))); | |
3a8699c7 AO |
2494 | asm_fprintf (asm_out_file, "b%s%ss\t%LLF%d\n", |
2495 | logic ? "f" : "t", | |
2496 | ASSEMBLER_DIALECT ? "/" : ".", label); | |
2497 | fprintf (asm_out_file, "\tnop\n"); | |
2498 | output_asm_insn ("bra\t%l0", operands); | |
2499 | fprintf (asm_out_file, "\tnop\n"); | |
2500 | (*targetm.asm_out.internal_label) (asm_out_file, "LF", label); | |
2501 | ||
2502 | return ""; | |
2503 | } | |
2504 | /* When relaxing, fall through. */ | |
2505 | case 4: | |
2506 | { | |
2507 | char buffer[10]; | |
f676971a | 2508 | |
3a8699c7 AO |
2509 | sprintf (buffer, "b%s%ss\t%%l0", |
2510 | logic ? "t" : "f", | |
2511 | ASSEMBLER_DIALECT ? "/" : "."); | |
2512 | output_asm_insn (buffer, &operands[0]); | |
2513 | return "nop"; | |
2514 | } | |
2515 | ||
33f7f353 | 2516 | default: |
afbc5d86 R |
2517 | /* There should be no longer branches now - that would |
2518 | indicate that something has destroyed the branches set | |
2519 | up in machine_dependent_reorg. */ | |
f5b9e7c9 | 2520 | gcc_unreachable (); |
1245df60 | 2521 | } |
1245df60 | 2522 | } |
bc45ade3 | 2523 | |
0a2aaacc | 2524 | /* Output a code sequence for INSN using TEMPL with OPERANDS; but before, |
78d310c2 R |
2525 | fill in operands 9 as a label to the successor insn. |
2526 | We try to use jump threading where possible. | |
2527 | IF CODE matches the comparison in the IF_THEN_ELSE of a following jump, | |
2528 | we assume the jump is taken. I.e. EQ means follow jmp and bf, NE means | |
2529 | follow jmp and bt, if the address is in range. */ | |
318881c0 | 2530 | const char * |
0a2aaacc | 2531 | output_branchy_insn (enum rtx_code code, const char *templ, |
cf277499 | 2532 | rtx insn, rtx *operands) |
1245df60 R |
2533 | { |
2534 | rtx next_insn = NEXT_INSN (insn); | |
79b2746a | 2535 | |
f3536097 | 2536 | if (next_insn && JUMP_P (next_insn) && condjump_p (next_insn)) |
1245df60 R |
2537 | { |
2538 | rtx src = SET_SRC (PATTERN (next_insn)); | |
2539 | if (GET_CODE (src) == IF_THEN_ELSE && GET_CODE (XEXP (src, 0)) != code) | |
2540 | { | |
2541 | /* Following branch not taken */ | |
2542 | operands[9] = gen_label_rtx (); | |
2543 | emit_label_after (operands[9], next_insn); | |
9d98a694 AO |
2544 | INSN_ADDRESSES_NEW (operands[9], |
2545 | INSN_ADDRESSES (INSN_UID (next_insn)) | |
2546 | + get_attr_length (next_insn)); | |
0a2aaacc | 2547 | return templ; |
1245df60 R |
2548 | } |
2549 | else | |
2550 | { | |
33f7f353 | 2551 | int offset = (branch_dest (next_insn) |
9d98a694 | 2552 | - INSN_ADDRESSES (INSN_UID (next_insn)) + 4); |
33f7f353 | 2553 | if (offset >= -252 && offset <= 258) |
1245df60 R |
2554 | { |
2555 | if (GET_CODE (src) == IF_THEN_ELSE) | |
2556 | /* branch_true */ | |
2557 | src = XEXP (src, 1); | |
2558 | operands[9] = src; | |
0a2aaacc | 2559 | return templ; |
1245df60 R |
2560 | } |
2561 | } | |
bc45ade3 | 2562 | } |
1245df60 R |
2563 | operands[9] = gen_label_rtx (); |
2564 | emit_label_after (operands[9], insn); | |
9d98a694 AO |
2565 | INSN_ADDRESSES_NEW (operands[9], |
2566 | INSN_ADDRESSES (INSN_UID (insn)) | |
2567 | + get_attr_length (insn)); | |
0a2aaacc | 2568 | return templ; |
1245df60 | 2569 | } |
e4fa6b06 | 2570 | |
318881c0 | 2571 | const char * |
cf277499 | 2572 | output_ieee_ccmpeq (rtx insn, rtx *operands) |
1245df60 | 2573 | { |
9fd0c419 R |
2574 | return output_branchy_insn (NE, "bt\t%l9\n\tfcmp/eq\t%1,%0", |
2575 | insn, operands); | |
bc45ade3 | 2576 | } |
8e87e161 | 2577 | \f |
1bc7c5b6 | 2578 | /* Output the start of the assembler file. */ |
8aa2a305 | 2579 | |
1bc7c5b6 | 2580 | static void |
cf277499 | 2581 | sh_file_start (void) |
bc45ade3 | 2582 | { |
1bc7c5b6 ZW |
2583 | default_file_start (); |
2584 | ||
b55e3aad NC |
2585 | #ifdef SYMBIAN |
2586 | /* Declare the .directive section before it is used. */ | |
2587 | fputs ("\t.section .directive, \"SM\", @progbits, 1\n", asm_out_file); | |
2588 | fputs ("\t.asciz \"#<SYMEDIT>#\\n\"\n", asm_out_file); | |
2589 | #endif | |
f676971a | 2590 | |
1bc7c5b6 ZW |
2591 | if (TARGET_ELF) |
2592 | /* We need to show the text section with the proper | |
2593 | attributes as in TEXT_SECTION_ASM_OP, before dwarf2out | |
0fe7abcc | 2594 | emits it without attributes in TEXT_SECTION_ASM_OP, else GAS |
1bc7c5b6 ZW |
2595 | will complain. We can teach GAS specifically about the |
2596 | default attributes for our choice of text section, but | |
2597 | then we would have to change GAS again if/when we change | |
2598 | the text section name. */ | |
2599 | fprintf (asm_out_file, "%s\n", TEXT_SECTION_ASM_OP); | |
2600 | else | |
2601 | /* Switch to the data section so that the coffsem symbol | |
2602 | isn't in the text section. */ | |
d6b5193b | 2603 | switch_to_section (data_section); |
b9654711 | 2604 | |
8aa2a305 | 2605 | if (TARGET_LITTLE_ENDIAN) |
1bc7c5b6 | 2606 | fputs ("\t.little\n", asm_out_file); |
fa5322fa | 2607 | |
1bc7c5b6 ZW |
2608 | if (!TARGET_ELF) |
2609 | { | |
2610 | if (TARGET_SHCOMPACT) | |
2611 | fputs ("\t.mode\tSHcompact\n", asm_out_file); | |
2612 | else if (TARGET_SHMEDIA) | |
2613 | fprintf (asm_out_file, "\t.mode\tSHmedia\n\t.abi\t%i\n", | |
2614 | TARGET_SHMEDIA64 ? 64 : 32); | |
2615 | } | |
bc45ade3 | 2616 | } |
0d7e008e | 2617 | \f |
9f3a9a08 KK |
2618 | /* Check if PAT includes UNSPEC_CALLER unspec pattern. */ |
2619 | ||
2620 | static bool | |
cf277499 | 2621 | unspec_caller_rtx_p (rtx pat) |
9f3a9a08 | 2622 | { |
dc3ba671 RS |
2623 | rtx base, offset; |
2624 | int i; | |
2625 | ||
2626 | split_const (pat, &base, &offset); | |
2627 | if (GET_CODE (base) == UNSPEC) | |
9f3a9a08 | 2628 | { |
dc3ba671 | 2629 | if (XINT (base, 1) == UNSPEC_CALLER) |
9f3a9a08 | 2630 | return true; |
dc3ba671 RS |
2631 | for (i = 0; i < XVECLEN (base, 0); i++) |
2632 | if (unspec_caller_rtx_p (XVECEXP (base, 0, i))) | |
2633 | return true; | |
9f3a9a08 | 2634 | } |
9f3a9a08 KK |
2635 | return false; |
2636 | } | |
2637 | ||
2638 | /* Indicate that INSN cannot be duplicated. This is true for insn | |
569b7f6a | 2639 | that generates a unique label. */ |
9f3a9a08 KK |
2640 | |
2641 | static bool | |
cf277499 | 2642 | sh_cannot_copy_insn_p (rtx insn) |
9f3a9a08 KK |
2643 | { |
2644 | rtx pat; | |
2645 | ||
2646 | if (!reload_completed || !flag_pic) | |
2647 | return false; | |
2648 | ||
f3536097 | 2649 | if (!NONJUMP_INSN_P (insn)) |
9f3a9a08 KK |
2650 | return false; |
2651 | if (asm_noperands (insn) >= 0) | |
2652 | return false; | |
2653 | ||
2654 | pat = PATTERN (insn); | |
2655 | if (GET_CODE (pat) != SET) | |
2656 | return false; | |
2657 | pat = SET_SRC (pat); | |
2658 | ||
2659 | if (unspec_caller_rtx_p (pat)) | |
2660 | return true; | |
2661 | ||
2662 | return false; | |
2663 | } | |
2664 | \f | |
16bea517 | 2665 | /* Actual number of instructions used to make a shift by N. */ |
0b5826ac | 2666 | static const char ashiftrt_insns[] = |
16bea517 JW |
2667 | { 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}; |
2668 | ||
2669 | /* Left shift and logical right shift are the same. */ | |
0b5826ac | 2670 | static const char shift_insns[] = |
16bea517 | 2671 | { 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 | 2672 | |
16bea517 JW |
2673 | /* Individual shift amounts needed to get the above length sequences. |
2674 | One bit right shifts clobber the T bit, so when possible, put one bit | |
2675 | shifts in the middle of the sequence, so the ends are eligible for | |
2676 | branch delay slots. */ | |
0139adca | 2677 | static const short shift_amounts[32][5] = { |
16bea517 JW |
2678 | {0}, {1}, {2}, {2, 1}, |
2679 | {2, 2}, {2, 1, 2}, {2, 2, 2}, {2, 2, 1, 2}, | |
2680 | {8}, {8, 1}, {8, 2}, {8, 1, 2}, | |
2681 | {8, 2, 2}, {8, 2, 1, 2}, {8, -2, 8}, {8, -1, 8}, | |
2682 | {16}, {16, 1}, {16, 2}, {16, 1, 2}, | |
2683 | {16, 2, 2}, {16, 2, 1, 2}, {16, -2, 8}, {16, -1, 8}, | |
2684 | {16, 8}, {16, 1, 8}, {16, 8, 2}, {16, 8, 1, 2}, | |
1d3534c0 | 2685 | {16, 8, 2, 2}, {16, -1, -2, 16}, {16, -2, 16}, {16, -1, 16}}; |
16bea517 | 2686 | |
8d481241 JW |
2687 | /* Likewise, but for shift amounts < 16, up to three highmost bits |
2688 | might be clobbered. This is typically used when combined with some | |
2689 | kind of sign or zero extension. */ | |
f676971a | 2690 | |
0b5826ac | 2691 | static const char ext_shift_insns[] = |
8d481241 JW |
2692 | { 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}; |
2693 | ||
0b5826ac | 2694 | static const short ext_shift_amounts[32][4] = { |
8d481241 JW |
2695 | {0}, {1}, {2}, {2, 1}, |
2696 | {2, 2}, {2, 1, 2}, {8, -2}, {8, -1}, | |
2697 | {8}, {8, 1}, {8, 2}, {8, 1, 2}, | |
2698 | {8, 2, 2}, {16, -2, -1}, {16, -2}, {16, -1}, | |
2699 | {16}, {16, 1}, {16, 2}, {16, 1, 2}, | |
2700 | {16, 2, 2}, {16, 2, 1, 2}, {16, -2, 8}, {16, -1, 8}, | |
2701 | {16, 8}, {16, 1, 8}, {16, 8, 2}, {16, 8, 1, 2}, | |
2702 | {16, 8, 2, 2}, {16, -1, -2, 16}, {16, -2, 16}, {16, -1, 16}}; | |
2703 | ||
ae9d19c0 JR |
2704 | /* Assuming we have a value that has been sign-extended by at least one bit, |
2705 | can we use the ext_shift_amounts with the last shift turned to an arithmetic shift | |
2706 | to shift it by N without data loss, and quicker than by other means? */ | |
2707 | #define EXT_SHIFT_SIGNED(n) (((n) | 8) == 15) | |
2708 | ||
16bea517 JW |
2709 | /* This is used in length attributes in sh.md to help compute the length |
2710 | of arbitrary constant shift instructions. */ | |
bc45ade3 | 2711 | |
16bea517 | 2712 | int |
cf277499 | 2713 | shift_insns_rtx (rtx insn) |
16bea517 JW |
2714 | { |
2715 | rtx set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
6c34a092 | 2716 | int shift_count = INTVAL (XEXP (set_src, 1)) & 31; |
16bea517 | 2717 | enum rtx_code shift_code = GET_CODE (set_src); |
00f8ff66 | 2718 | |
16bea517 JW |
2719 | switch (shift_code) |
2720 | { | |
2721 | case ASHIFTRT: | |
2722 | return ashiftrt_insns[shift_count]; | |
2723 | case LSHIFTRT: | |
2724 | case ASHIFT: | |
2725 | return shift_insns[shift_count]; | |
2726 | default: | |
f5b9e7c9 | 2727 | gcc_unreachable (); |
16bea517 JW |
2728 | } |
2729 | } | |
2730 | ||
16bea517 | 2731 | /* Return the cost of a shift. */ |
bc45ade3 | 2732 | |
3c50106f | 2733 | static inline int |
cf277499 | 2734 | shiftcosts (rtx x) |
bc45ade3 | 2735 | { |
3d422cb1 | 2736 | int value; |
8aa2a305 | 2737 | |
fa5322fa AO |
2738 | if (TARGET_SHMEDIA) |
2739 | return 1; | |
2740 | ||
c762ab6e R |
2741 | if (GET_MODE_SIZE (GET_MODE (x)) > UNITS_PER_WORD) |
2742 | { | |
2743 | if (GET_MODE (x) == DImode | |
f3536097 | 2744 | && CONST_INT_P (XEXP (x, 1)) |
c762ab6e R |
2745 | && INTVAL (XEXP (x, 1)) == 1) |
2746 | return 2; | |
2747 | ||
2748 | /* Everything else is invalid, because there is no pattern for it. */ | |
cb2eb96f | 2749 | return MAX_COST; |
c762ab6e | 2750 | } |
16bea517 | 2751 | /* If shift by a non constant, then this will be expensive. */ |
f3536097 | 2752 | if (!CONST_INT_P (XEXP (x, 1))) |
1245df60 | 2753 | return SH_DYNAMIC_SHIFT_COST; |
bc45ade3 | 2754 | |
6c34a092 PB |
2755 | /* Otherwise, return the true cost in instructions. Cope with out of range |
2756 | shift counts more or less arbitrarily. */ | |
2757 | value = INTVAL (XEXP (x, 1)) & 31; | |
3d422cb1 | 2758 | |
8aa2a305 | 2759 | if (GET_CODE (x) == ASHIFTRT) |
49b6d06b JW |
2760 | { |
2761 | int cost = ashiftrt_insns[value]; | |
2762 | /* If SH3, then we put the constant in a reg and use shad. */ | |
1245df60 R |
2763 | if (cost > 1 + SH_DYNAMIC_SHIFT_COST) |
2764 | cost = 1 + SH_DYNAMIC_SHIFT_COST; | |
49b6d06b JW |
2765 | return cost; |
2766 | } | |
8aa2a305 JW |
2767 | else |
2768 | return shift_insns[value]; | |
0d7e008e | 2769 | } |
b9654711 | 2770 | |
8aa2a305 JW |
2771 | /* Return the cost of an AND operation. */ |
2772 | ||
3c50106f | 2773 | static inline int |
cf277499 | 2774 | andcosts (rtx x) |
0d7e008e SC |
2775 | { |
2776 | int i; | |
8aa2a305 | 2777 | |
a6f71af5 | 2778 | /* Anding with a register is a single cycle and instruction. */ |
f3536097 | 2779 | if (!CONST_INT_P (XEXP (x, 1))) |
a6f71af5 | 2780 | return 1; |
8aa2a305 JW |
2781 | |
2782 | i = INTVAL (XEXP (x, 1)); | |
fa5322fa AO |
2783 | |
2784 | if (TARGET_SHMEDIA) | |
2785 | { | |
32a7ab3d KK |
2786 | if (satisfies_constraint_I10 (XEXP (x, 1)) |
2787 | || satisfies_constraint_J16 (XEXP (x, 1))) | |
fa5322fa AO |
2788 | return 1; |
2789 | else | |
f40751dd | 2790 | return 1 + rtx_cost (XEXP (x, 1), AND, !optimize_size); |
fa5322fa AO |
2791 | } |
2792 | ||
a6f71af5 | 2793 | /* These constants are single cycle extu.[bw] instructions. */ |
0d7e008e | 2794 | if (i == 0xff || i == 0xffff) |
a6f71af5 | 2795 | return 1; |
735cb76e | 2796 | /* Constants that can be used in an and immediate instruction in a single |
a6f71af5 | 2797 | cycle, but this requires r0, so make it a little more expensive. */ |
735cb76e | 2798 | if (CONST_OK_FOR_K08 (i)) |
0d7e008e | 2799 | return 2; |
a6f71af5 JW |
2800 | /* Constants that can be loaded with a mov immediate and an and. |
2801 | This case is probably unnecessary. */ | |
735cb76e | 2802 | if (CONST_OK_FOR_I08 (i)) |
a6f71af5 JW |
2803 | return 2; |
2804 | /* Any other constants requires a 2 cycle pc-relative load plus an and. | |
2805 | This case is probably unnecessary. */ | |
2806 | return 3; | |
0d7e008e | 2807 | } |
d3ae8277 | 2808 | |
630c79be BS |
2809 | /* Return the cost of an addition or a subtraction. */ |
2810 | ||
3c50106f | 2811 | static inline int |
cf277499 | 2812 | addsubcosts (rtx x) |
630c79be BS |
2813 | { |
2814 | /* Adding a register is a single cycle insn. */ | |
f3536097 | 2815 | if (REG_P (XEXP (x, 1)) |
50ceefc2 | 2816 | || GET_CODE (XEXP (x, 1)) == SUBREG) |
630c79be BS |
2817 | return 1; |
2818 | ||
2819 | /* Likewise for small constants. */ | |
f3536097 | 2820 | if (CONST_INT_P (XEXP (x, 1)) |
fa5322fa | 2821 | && CONST_OK_FOR_ADD (INTVAL (XEXP (x, 1)))) |
630c79be BS |
2822 | return 1; |
2823 | ||
fa5322fa AO |
2824 | if (TARGET_SHMEDIA) |
2825 | switch (GET_CODE (XEXP (x, 1))) | |
2826 | { | |
2827 | case CONST: | |
2828 | case LABEL_REF: | |
2829 | case SYMBOL_REF: | |
2830 | return TARGET_SHMEDIA64 ? 5 : 3; | |
2831 | ||
2832 | case CONST_INT: | |
735cb76e | 2833 | if (CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1)))) |
fa5322fa | 2834 | return 2; |
735cb76e | 2835 | else if (CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1)) >> 16)) |
fa5322fa | 2836 | return 3; |
735cb76e | 2837 | else if (CONST_OK_FOR_I16 ((INTVAL (XEXP (x, 1)) >> 16) >> 16)) |
fa5322fa AO |
2838 | return 4; |
2839 | ||
2840 | /* Fall through. */ | |
2841 | default: | |
832a3292 | 2842 | return 5; |
fa5322fa AO |
2843 | } |
2844 | ||
630c79be BS |
2845 | /* Any other constant requires a 2 cycle pc-relative load plus an |
2846 | addition. */ | |
2847 | return 3; | |
2848 | } | |
2849 | ||
16bea517 | 2850 | /* Return the cost of a multiply. */ |
3c50106f | 2851 | static inline int |
cf277499 | 2852 | multcosts (rtx x ATTRIBUTE_UNUSED) |
0d7e008e | 2853 | { |
7dd2f19b RS |
2854 | if (sh_multcost >= 0) |
2855 | return sh_multcost; | |
fa5322fa | 2856 | if (TARGET_SHMEDIA) |
73a4d10b R |
2857 | /* ??? We have a mul insn, but it has a latency of three, and doesn't |
2858 | accept constants. Ideally, we would use a cost of one or two and | |
2859 | add the cost of the operand, but disregard the latter when inside loops | |
2860 | and loop invariant code motion is still to follow. | |
2861 | Using a multiply first and splitting it later if it's a loss | |
2862 | doesn't work because of different sign / zero extension semantics | |
2863 | of multiplies vs. shifts. */ | |
2864 | return TARGET_SMALLCODE ? 2 : 3; | |
fa5322fa | 2865 | |
0d7e008e | 2866 | if (TARGET_SH2) |
d3ae8277 SC |
2867 | { |
2868 | /* We have a mul insn, so we can never take more than the mul and the | |
a7771f78 | 2869 | read of the mac reg, but count more because of the latency and extra |
16bea517 | 2870 | reg usage. */ |
d3ae8277 | 2871 | if (TARGET_SMALLCODE) |
8e87e161 | 2872 | return 2; |
a7771f78 | 2873 | return 3; |
d3ae8277 SC |
2874 | } |
2875 | ||
a7771f78 | 2876 | /* If we're aiming at small code, then just count the number of |
16bea517 | 2877 | insns in a multiply call sequence. */ |
8e87e161 | 2878 | if (TARGET_SMALLCODE) |
8aa2a305 | 2879 | return 5; |
d3ae8277 | 2880 | |
16bea517 | 2881 | /* Otherwise count all the insns in the routine we'd be calling too. */ |
d3ae8277 | 2882 | return 20; |
0d7e008e | 2883 | } |
b9654711 | 2884 | |
3c50106f RH |
2885 | /* Compute a (partial) cost for rtx X. Return true if the complete |
2886 | cost has been computed, and false if subexpressions should be | |
2887 | scanned. In either case, *TOTAL contains the cost result. */ | |
2888 | ||
2889 | static bool | |
f40751dd JH |
2890 | sh_rtx_costs (rtx x, int code, int outer_code, int *total, |
2891 | bool speed ATTRIBUTE_UNUSED) | |
3c50106f RH |
2892 | { |
2893 | switch (code) | |
2894 | { | |
2895 | case CONST_INT: | |
2896 | if (TARGET_SHMEDIA) | |
2897 | { | |
2898 | if (INTVAL (x) == 0) | |
2899 | *total = 0; | |
2900 | else if (outer_code == AND && and_operand ((x), DImode)) | |
2901 | *total = 0; | |
2902 | else if ((outer_code == IOR || outer_code == XOR | |
2903 | || outer_code == PLUS) | |
735cb76e | 2904 | && CONST_OK_FOR_I10 (INTVAL (x))) |
3c50106f | 2905 | *total = 0; |
735cb76e | 2906 | else if (CONST_OK_FOR_I16 (INTVAL (x))) |
3c50106f | 2907 | *total = COSTS_N_INSNS (outer_code != SET); |
735cb76e | 2908 | else if (CONST_OK_FOR_I16 (INTVAL (x) >> 16)) |
73a4d10b | 2909 | *total = COSTS_N_INSNS ((outer_code != SET) + 1); |
735cb76e | 2910 | else if (CONST_OK_FOR_I16 ((INTVAL (x) >> 16) >> 16)) |
42201282 | 2911 | *total = COSTS_N_INSNS ((outer_code != SET) + 2); |
3c50106f | 2912 | else |
42201282 | 2913 | *total = COSTS_N_INSNS ((outer_code != SET) + 3); |
3c50106f RH |
2914 | return true; |
2915 | } | |
735cb76e | 2916 | if (CONST_OK_FOR_I08 (INTVAL (x))) |
3c50106f RH |
2917 | *total = 0; |
2918 | else if ((outer_code == AND || outer_code == IOR || outer_code == XOR) | |
735cb76e | 2919 | && CONST_OK_FOR_K08 (INTVAL (x))) |
3c50106f | 2920 | *total = 1; |
78d310c2 | 2921 | /* prepare_cmp_insn will force costly constants int registers before |
35912544 | 2922 | the cbranch[sd]i4 patterns can see them, so preserve potentially |
78d310c2 R |
2923 | interesting ones not covered by I08 above. */ |
2924 | else if (outer_code == COMPARE | |
2925 | && ((unsigned HOST_WIDE_INT) INTVAL (x) | |
2926 | == (unsigned HOST_WIDE_INT) 0x7fffffff + 1 | |
2927 | || INTVAL (x) == 0x7fffffff | |
2928 | || INTVAL (x) == 0x80 || INTVAL (x) == -0x81)) | |
2929 | *total = 1; | |
3c50106f RH |
2930 | else |
2931 | *total = 8; | |
2932 | return true; | |
2933 | ||
2934 | case CONST: | |
2935 | case LABEL_REF: | |
2936 | case SYMBOL_REF: | |
2937 | if (TARGET_SHMEDIA64) | |
2938 | *total = COSTS_N_INSNS (4); | |
2939 | else if (TARGET_SHMEDIA32) | |
2940 | *total = COSTS_N_INSNS (2); | |
2941 | else | |
2942 | *total = 5; | |
2943 | return true; | |
2944 | ||
2945 | case CONST_DOUBLE: | |
2946 | if (TARGET_SHMEDIA) | |
2947 | *total = COSTS_N_INSNS (4); | |
78d310c2 | 2948 | /* prepare_cmp_insn will force costly constants int registers before |
35912544 | 2949 | the cbranchdi4 pattern can see them, so preserve potentially |
78d310c2 R |
2950 | interesting ones. */ |
2951 | else if (outer_code == COMPARE && GET_MODE (x) == DImode) | |
2952 | *total = 1; | |
3c50106f RH |
2953 | else |
2954 | *total = 10; | |
2955 | return true; | |
73a4d10b R |
2956 | case CONST_VECTOR: |
2957 | if (x == CONST0_RTX (GET_MODE (x))) | |
2958 | *total = 0; | |
2959 | else if (sh_1el_vec (x, VOIDmode)) | |
2960 | *total = outer_code != SET; | |
2961 | if (sh_rep_vec (x, VOIDmode)) | |
2962 | *total = ((GET_MODE_UNIT_SIZE (GET_MODE (x)) + 3) / 4 | |
2963 | + (outer_code != SET)); | |
2964 | *total = COSTS_N_INSNS (3) + (outer_code != SET); | |
2965 | return true; | |
3c50106f RH |
2966 | |
2967 | case PLUS: | |
73a4d10b | 2968 | case MINUS: |
3c50106f RH |
2969 | *total = COSTS_N_INSNS (addsubcosts (x)); |
2970 | return true; | |
2971 | ||
2972 | case AND: | |
2973 | *total = COSTS_N_INSNS (andcosts (x)); | |
2974 | return true; | |
2975 | ||
2976 | case MULT: | |
2977 | *total = COSTS_N_INSNS (multcosts (x)); | |
2978 | return true; | |
2979 | ||
2980 | case ASHIFT: | |
2981 | case ASHIFTRT: | |
2982 | case LSHIFTRT: | |
2983 | *total = COSTS_N_INSNS (shiftcosts (x)); | |
2984 | return true; | |
2985 | ||
2986 | case DIV: | |
2987 | case UDIV: | |
2988 | case MOD: | |
2989 | case UMOD: | |
2990 | *total = COSTS_N_INSNS (20); | |
2991 | return true; | |
2992 | ||
73a4d10b R |
2993 | case PARALLEL: |
2994 | if (sh_1el_vec (x, VOIDmode)) | |
2995 | *total = outer_code != SET; | |
2996 | if (sh_rep_vec (x, VOIDmode)) | |
2997 | *total = ((GET_MODE_UNIT_SIZE (GET_MODE (x)) + 3) / 4 | |
2998 | + (outer_code != SET)); | |
2999 | *total = COSTS_N_INSNS (3) + (outer_code != SET); | |
3000 | return true; | |
3001 | ||
3c50106f RH |
3002 | case FLOAT: |
3003 | case FIX: | |
3004 | *total = 100; | |
3005 | return true; | |
3006 | ||
3007 | default: | |
3008 | return false; | |
3009 | } | |
3010 | } | |
3011 | ||
dcefdf67 RH |
3012 | /* Compute the cost of an address. For the SH, all valid addresses are |
3013 | the same cost. Use a slightly higher cost for reg + reg addressing, | |
3014 | since it increases pressure on r0. */ | |
3015 | ||
3016 | static int | |
f40751dd JH |
3017 | sh_address_cost (rtx X, |
3018 | bool speed ATTRIBUTE_UNUSED) | |
dcefdf67 RH |
3019 | { |
3020 | return (GET_CODE (X) == PLUS | |
3021 | && ! CONSTANT_P (XEXP (X, 1)) | |
3022 | && ! TARGET_SHMEDIA ? 1 : 0); | |
3023 | } | |
3c50106f | 3024 | |
16bea517 | 3025 | /* Code to expand a shift. */ |
b9654711 | 3026 | |
0d7e008e | 3027 | void |
cf277499 | 3028 | gen_ashift (int type, int n, rtx reg) |
0d7e008e | 3029 | { |
16bea517 JW |
3030 | /* Negative values here come from the shift_amounts array. */ |
3031 | if (n < 0) | |
3032 | { | |
3033 | if (type == ASHIFT) | |
3034 | type = LSHIFTRT; | |
3035 | else | |
3036 | type = ASHIFT; | |
3037 | n = -n; | |
3038 | } | |
3039 | ||
0d7e008e | 3040 | switch (type) |
bc45ade3 | 3041 | { |
0d7e008e SC |
3042 | case ASHIFTRT: |
3043 | emit_insn (gen_ashrsi3_k (reg, reg, GEN_INT (n))); | |
3044 | break; | |
3045 | case LSHIFTRT: | |
16bea517 JW |
3046 | if (n == 1) |
3047 | emit_insn (gen_lshrsi3_m (reg, reg, GEN_INT (n))); | |
3048 | else | |
3049 | emit_insn (gen_lshrsi3_k (reg, reg, GEN_INT (n))); | |
0d7e008e SC |
3050 | break; |
3051 | case ASHIFT: | |
7e2fda6e | 3052 | emit_insn (gen_ashlsi3_std (reg, reg, GEN_INT (n))); |
0d7e008e | 3053 | break; |
bc45ade3 | 3054 | } |
bc45ade3 | 3055 | } |
bc45ade3 | 3056 | |
8d481241 JW |
3057 | /* Same for HImode */ |
3058 | ||
3059 | void | |
cf277499 | 3060 | gen_ashift_hi (int type, int n, rtx reg) |
8d481241 JW |
3061 | { |
3062 | /* Negative values here come from the shift_amounts array. */ | |
3063 | if (n < 0) | |
3064 | { | |
3065 | if (type == ASHIFT) | |
3066 | type = LSHIFTRT; | |
3067 | else | |
3068 | type = ASHIFT; | |
3069 | n = -n; | |
3070 | } | |
3071 | ||
3072 | switch (type) | |
3073 | { | |
3074 | case ASHIFTRT: | |
8d481241 | 3075 | case LSHIFTRT: |
d0c42859 R |
3076 | /* We don't have HImode right shift operations because using the |
3077 | ordinary 32 bit shift instructions for that doesn't generate proper | |
3078 | zero/sign extension. | |
3079 | gen_ashift_hi is only called in contexts where we know that the | |
3080 | sign extension works out correctly. */ | |
97d6fd65 | 3081 | { |
ddef6bc7 | 3082 | int offset = 0; |
97d6fd65 R |
3083 | if (GET_CODE (reg) == SUBREG) |
3084 | { | |
ddef6bc7 | 3085 | offset = SUBREG_BYTE (reg); |
97d6fd65 R |
3086 | reg = SUBREG_REG (reg); |
3087 | } | |
ddef6bc7 | 3088 | gen_ashift (type, n, gen_rtx_SUBREG (SImode, reg, offset)); |
97d6fd65 R |
3089 | break; |
3090 | } | |
8d481241 JW |
3091 | case ASHIFT: |
3092 | emit_insn (gen_ashlhi3_k (reg, reg, GEN_INT (n))); | |
3093 | break; | |
3094 | } | |
3095 | } | |
3096 | ||
8aa2a305 JW |
3097 | /* Output RTL to split a constant shift into its component SH constant |
3098 | shift instructions. */ | |
f676971a | 3099 | |
318881c0 | 3100 | void |
cf277499 | 3101 | gen_shifty_op (int code, rtx *operands) |
bc45ade3 | 3102 | { |
16bea517 | 3103 | int value = INTVAL (operands[2]); |
8aa2a305 | 3104 | int max, i; |
00f8ff66 | 3105 | |
cff3d762 | 3106 | /* Truncate the shift count in case it is out of bounds. */ |
6c34a092 | 3107 | value = value & 31; |
f676971a | 3108 | |
8aa2a305 | 3109 | if (value == 31) |
16bea517 | 3110 | { |
8aa2a305 | 3111 | if (code == LSHIFTRT) |
0d7e008e | 3112 | { |
8aa2a305 JW |
3113 | emit_insn (gen_rotlsi3_1 (operands[0], operands[0])); |
3114 | emit_insn (gen_movt (operands[0])); | |
3115 | return; | |
16bea517 | 3116 | } |
8aa2a305 | 3117 | else if (code == ASHIFT) |
16bea517 | 3118 | { |
8aa2a305 JW |
3119 | /* There is a two instruction sequence for 31 bit left shifts, |
3120 | but it requires r0. */ | |
f3536097 | 3121 | if (REG_P (operands[0]) && REGNO (operands[0]) == 0) |
0d7e008e | 3122 | { |
8aa2a305 JW |
3123 | emit_insn (gen_andsi3 (operands[0], operands[0], const1_rtx)); |
3124 | emit_insn (gen_rotlsi3_31 (operands[0], operands[0])); | |
3125 | return; | |
0d7e008e | 3126 | } |
16bea517 | 3127 | } |
8aa2a305 | 3128 | } |
c17f53a8 JW |
3129 | else if (value == 0) |
3130 | { | |
73a4d10b R |
3131 | /* This can happen even when optimizing, if there were subregs before |
3132 | reload. Don't output a nop here, as this is never optimized away; | |
3133 | use a no-op move instead. */ | |
3134 | emit_insn (gen_rtx_SET (VOIDmode, operands[0], operands[0])); | |
c17f53a8 JW |
3135 | return; |
3136 | } | |
a9f71ad8 | 3137 | |
8aa2a305 JW |
3138 | max = shift_insns[value]; |
3139 | for (i = 0; i < max; i++) | |
3140 | gen_ashift (code, shift_amounts[value][i], operands[0]); | |
3141 | } | |
f676971a | 3142 | |
8d481241 JW |
3143 | /* Same as above, but optimized for values where the topmost bits don't |
3144 | matter. */ | |
3145 | ||
318881c0 | 3146 | void |
cf277499 | 3147 | gen_shifty_hi_op (int code, rtx *operands) |
8d481241 JW |
3148 | { |
3149 | int value = INTVAL (operands[2]); | |
3150 | int max, i; | |
cf277499 | 3151 | void (*gen_fun) (int, int, rtx); |
8d481241 JW |
3152 | |
3153 | /* This operation is used by and_shl for SImode values with a few | |
3154 | high bits known to be cleared. */ | |
3155 | value &= 31; | |
3156 | if (value == 0) | |
3157 | { | |
3158 | emit_insn (gen_nop ()); | |
3159 | return; | |
3160 | } | |
3161 | ||
3162 | gen_fun = GET_MODE (operands[0]) == HImode ? gen_ashift_hi : gen_ashift; | |
3163 | if (code == ASHIFT) | |
3164 | { | |
3165 | max = ext_shift_insns[value]; | |
3166 | for (i = 0; i < max; i++) | |
3167 | gen_fun (code, ext_shift_amounts[value][i], operands[0]); | |
3168 | } | |
3169 | else | |
3170 | /* When shifting right, emit the shifts in reverse order, so that | |
3171 | solitary negative values come first. */ | |
3172 | for (i = ext_shift_insns[value] - 1; i >= 0; i--) | |
3173 | gen_fun (code, ext_shift_amounts[value][i], operands[0]); | |
3174 | } | |
8aa2a305 JW |
3175 | |
3176 | /* Output RTL for an arithmetic right shift. */ | |
3177 | ||
3178 | /* ??? Rewrite to use super-optimizer sequences. */ | |
3179 | ||
3180 | int | |
cf277499 | 3181 | expand_ashiftrt (rtx *operands) |
8aa2a305 JW |
3182 | { |
3183 | rtx wrk; | |
3184 | char func[18]; | |
8aa2a305 JW |
3185 | int value; |
3186 | ||
49b6d06b | 3187 | if (TARGET_SH3) |
20b04867 | 3188 | { |
f3536097 | 3189 | if (!CONST_INT_P (operands[2])) |
49b6d06b JW |
3190 | { |
3191 | rtx count = copy_to_mode_reg (SImode, operands[2]); | |
3192 | emit_insn (gen_negsi2 (count, count)); | |
3193 | emit_insn (gen_ashrsi3_d (operands[0], operands[1], count)); | |
3194 | return 1; | |
3195 | } | |
1245df60 R |
3196 | else if (ashiftrt_insns[INTVAL (operands[2]) & 31] |
3197 | > 1 + SH_DYNAMIC_SHIFT_COST) | |
49b6d06b | 3198 | { |
1245df60 R |
3199 | rtx count |
3200 | = force_reg (SImode, GEN_INT (- (INTVAL (operands[2]) & 31))); | |
49b6d06b JW |
3201 | emit_insn (gen_ashrsi3_d (operands[0], operands[1], count)); |
3202 | return 1; | |
3203 | } | |
20b04867 | 3204 | } |
f3536097 | 3205 | if (!CONST_INT_P (operands[2])) |
8aa2a305 JW |
3206 | return 0; |
3207 | ||
1245df60 | 3208 | value = INTVAL (operands[2]) & 31; |
8aa2a305 JW |
3209 | |
3210 | if (value == 31) | |
3211 | { | |
73a4d10b R |
3212 | /* If we are called from abs expansion, arrange things so that we |
3213 | we can use a single MT instruction that doesn't clobber the source, | |
3214 | if LICM can hoist out the load of the constant zero. */ | |
3215 | if (currently_expanding_to_rtl) | |
3216 | { | |
3217 | emit_insn (gen_cmpgtsi_t (force_reg (SImode, CONST0_RTX (SImode)), | |
3218 | operands[1])); | |
3219 | emit_insn (gen_mov_neg_si_t (operands[0])); | |
3220 | return 1; | |
3221 | } | |
8aa2a305 JW |
3222 | emit_insn (gen_ashrsi2_31 (operands[0], operands[1])); |
3223 | return 1; | |
3224 | } | |
3225 | else if (value >= 16 && value <= 19) | |
3226 | { | |
3227 | wrk = gen_reg_rtx (SImode); | |
3228 | emit_insn (gen_ashrsi2_16 (wrk, operands[1])); | |
3229 | value -= 16; | |
3230 | while (value--) | |
3231 | gen_ashift (ASHIFTRT, 1, wrk); | |
3232 | emit_move_insn (operands[0], wrk); | |
3233 | return 1; | |
a9f71ad8 | 3234 | } |
8aa2a305 JW |
3235 | /* Expand a short sequence inline, longer call a magic routine. */ |
3236 | else if (value <= 5) | |
3237 | { | |
3238 | wrk = gen_reg_rtx (SImode); | |
3239 | emit_move_insn (wrk, operands[1]); | |
3240 | while (value--) | |
3241 | gen_ashift (ASHIFTRT, 1, wrk); | |
3242 | emit_move_insn (operands[0], wrk); | |
3243 | return 1; | |
3244 | } | |
3245 | ||
3246 | wrk = gen_reg_rtx (Pmode); | |
3247 | ||
3248 | /* Load the value into an arg reg and call a helper. */ | |
c5c76735 | 3249 | emit_move_insn (gen_rtx_REG (SImode, 4), operands[1]); |
8aa2a305 | 3250 | sprintf (func, "__ashiftrt_r4_%d", value); |
73a4d10b | 3251 | function_symbol (wrk, func, SFUNC_STATIC); |
8aa2a305 | 3252 | emit_insn (gen_ashrsi3_n (GEN_INT (value), wrk)); |
c5c76735 | 3253 | emit_move_insn (operands[0], gen_rtx_REG (SImode, 4)); |
8aa2a305 | 3254 | return 1; |
bc45ade3 | 3255 | } |
8d481241 | 3256 | |
318881c0 | 3257 | int |
cf277499 | 3258 | sh_dynamicalize_shift_p (rtx count) |
1245df60 | 3259 | { |
6c34a092 | 3260 | return shift_insns[INTVAL (count) & 31] > 1 + SH_DYNAMIC_SHIFT_COST; |
1245df60 R |
3261 | } |
3262 | ||
8d481241 JW |
3263 | /* Try to find a good way to implement the combiner pattern |
3264 | [(set (match_operand:SI 0 "register_operand" "r") | |
3265 | (and:SI (ashift:SI (match_operand:SI 1 "register_operand" "r") | |
3266 | (match_operand:SI 2 "const_int_operand" "n")) | |
3267 | (match_operand:SI 3 "const_int_operand" "n"))) . | |
3268 | LEFT_RTX is operand 2 in the above pattern, and MASK_RTX is operand 3. | |
3269 | return 0 for simple right / left or left/right shift combination. | |
3270 | return 1 for a combination of shifts with zero_extend. | |
3271 | return 2 for a combination of shifts with an AND that needs r0. | |
3272 | return 3 for a combination of shifts with an AND that needs an extra | |
3273 | scratch register, when the three highmost bits of the AND mask are clear. | |
3274 | return 4 for a combination of shifts with an AND that needs an extra | |
3275 | scratch register, when any of the three highmost bits of the AND mask | |
3276 | is set. | |
3277 | If ATTRP is set, store an initial right shift width in ATTRP[0], | |
3278 | and the instruction length in ATTRP[1] . These values are not valid | |
3279 | when returning 0. | |
3280 | When ATTRP is set and returning 1, ATTRP[2] gets set to the index into | |
3281 | shift_amounts for the last shift value that is to be used before the | |
3282 | sign extend. */ | |
3283 | int | |
cf277499 | 3284 | shl_and_kind (rtx left_rtx, rtx mask_rtx, int *attrp) |
8d481241 JW |
3285 | { |
3286 | unsigned HOST_WIDE_INT mask, lsb, mask2, lsb2; | |
3287 | int left = INTVAL (left_rtx), right; | |
3288 | int best = 0; | |
3289 | int cost, best_cost = 10000; | |
3290 | int best_right = 0, best_len = 0; | |
3291 | int i; | |
3292 | int can_ext; | |
3293 | ||
3294 | if (left < 0 || left > 31) | |
3295 | return 0; | |
f3536097 | 3296 | if (CONST_INT_P (mask_rtx)) |
8d481241 JW |
3297 | mask = (unsigned HOST_WIDE_INT) INTVAL (mask_rtx) >> left; |
3298 | else | |
3299 | mask = (unsigned HOST_WIDE_INT) GET_MODE_MASK (SImode) >> left; | |
832a3292 | 3300 | /* Can this be expressed as a right shift / left shift pair? */ |
8d481241 JW |
3301 | lsb = ((mask ^ (mask - 1)) >> 1) + 1; |
3302 | right = exact_log2 (lsb); | |
3303 | mask2 = ~(mask + lsb - 1); | |
3304 | lsb2 = ((mask2 ^ (mask2 - 1)) >> 1) + 1; | |
3305 | /* mask has no zeroes but trailing zeroes <==> ! mask2 */ | |
3306 | if (! mask2) | |
3307 | best_cost = shift_insns[right] + shift_insns[right + left]; | |
3308 | /* mask has no trailing zeroes <==> ! right */ | |
3309 | else if (! right && mask2 == ~(lsb2 - 1)) | |
3310 | { | |
3311 | int late_right = exact_log2 (lsb2); | |
3312 | best_cost = shift_insns[left + late_right] + shift_insns[late_right]; | |
3313 | } | |
832a3292 | 3314 | /* Try to use zero extend. */ |
8d481241 JW |
3315 | if (mask2 == ~(lsb2 - 1)) |
3316 | { | |
3317 | int width, first; | |
3318 | ||
3319 | for (width = 8; width <= 16; width += 8) | |
3320 | { | |
832a3292 KH |
3321 | /* Can we zero-extend right away? */ |
3322 | if (lsb2 == (unsigned HOST_WIDE_INT) 1 << width) | |
8d481241 JW |
3323 | { |
3324 | cost | |
3325 | = 1 + ext_shift_insns[right] + ext_shift_insns[left + right]; | |
3326 | if (cost < best_cost) | |
3327 | { | |
3328 | best = 1; | |
3329 | best_cost = cost; | |
3330 | best_right = right; | |
3331 | best_len = cost; | |
3332 | if (attrp) | |
3333 | attrp[2] = -1; | |
3334 | } | |
3335 | continue; | |
3336 | } | |
3337 | /* ??? Could try to put zero extend into initial right shift, | |
6f317ef3 | 3338 | or even shift a bit left before the right shift. */ |
8d481241 JW |
3339 | /* Determine value of first part of left shift, to get to the |
3340 | zero extend cut-off point. */ | |
3341 | first = width - exact_log2 (lsb2) + right; | |
3342 | if (first >= 0 && right + left - first >= 0) | |
3343 | { | |
3344 | cost = ext_shift_insns[right] + ext_shift_insns[first] + 1 | |
3345 | + ext_shift_insns[right + left - first]; | |
3346 | if (cost < best_cost) | |
3347 | { | |
3348 | best = 1; | |
3349 | best_cost = cost; | |
3350 | best_right = right; | |
3351 | best_len = cost; | |
3352 | if (attrp) | |
3353 | attrp[2] = first; | |
832a3292 | 3354 | } |
8d481241 JW |
3355 | } |
3356 | } | |
3357 | } | |
3358 | /* Try to use r0 AND pattern */ | |
3359 | for (i = 0; i <= 2; i++) | |
3360 | { | |
3361 | if (i > right) | |
3362 | break; | |
735cb76e | 3363 | if (! CONST_OK_FOR_K08 (mask >> i)) |
8d481241 JW |
3364 | continue; |
3365 | cost = (i != 0) + 2 + ext_shift_insns[left + i]; | |
3366 | if (cost < best_cost) | |
3367 | { | |
3368 | best = 2; | |
3369 | best_cost = cost; | |
3370 | best_right = i; | |
3371 | best_len = cost - 1; | |
3372 | } | |
3373 | } | |
3374 | /* Try to use a scratch register to hold the AND operand. */ | |
832a3292 | 3375 | can_ext = ((mask << left) & ((unsigned HOST_WIDE_INT) 3 << 30)) == 0; |
8d481241 JW |
3376 | for (i = 0; i <= 2; i++) |
3377 | { | |
3378 | if (i > right) | |
3379 | break; | |
735cb76e | 3380 | cost = (i != 0) + (CONST_OK_FOR_I08 (mask >> i) ? 2 : 3) |
6ab911bb | 3381 | + (can_ext ? ext_shift_insns : shift_insns)[left + i]; |
8d481241 JW |
3382 | if (cost < best_cost) |
3383 | { | |
3384 | best = 4 - can_ext; | |
3385 | best_cost = cost; | |
3386 | best_right = i; | |
735cb76e | 3387 | best_len = cost - 1 - ! CONST_OK_FOR_I08 (mask >> i); |
8d481241 JW |
3388 | } |
3389 | } | |
3390 | ||
3391 | if (attrp) | |
3392 | { | |
3393 | attrp[0] = best_right; | |
3394 | attrp[1] = best_len; | |
3395 | } | |
3396 | return best; | |
3397 | } | |
3398 | ||
3399 | /* This is used in length attributes of the unnamed instructions | |
3400 | corresponding to shl_and_kind return values of 1 and 2. */ | |
3401 | int | |
cf277499 | 3402 | shl_and_length (rtx insn) |
8d481241 JW |
3403 | { |
3404 | rtx set_src, left_rtx, mask_rtx; | |
3405 | int attributes[3]; | |
3406 | ||
3407 | set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
3408 | left_rtx = XEXP (XEXP (set_src, 0), 1); | |
3409 | mask_rtx = XEXP (set_src, 1); | |
3410 | shl_and_kind (left_rtx, mask_rtx, attributes); | |
3411 | return attributes[1]; | |
3412 | } | |
3413 | ||
3414 | /* This is used in length attribute of the and_shl_scratch instruction. */ | |
3415 | ||
3416 | int | |
cf277499 | 3417 | shl_and_scr_length (rtx insn) |
8d481241 JW |
3418 | { |
3419 | rtx set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
6c34a092 | 3420 | int len = shift_insns[INTVAL (XEXP (set_src, 1)) & 31]; |
8d481241 | 3421 | rtx op = XEXP (set_src, 0); |
6c34a092 | 3422 | len += shift_insns[INTVAL (XEXP (op, 1)) & 31] + 1; |
8d481241 | 3423 | op = XEXP (XEXP (op, 0), 0); |
6c34a092 | 3424 | return len + shift_insns[INTVAL (XEXP (op, 1)) & 31]; |
8d481241 JW |
3425 | } |
3426 | ||
8d481241 JW |
3427 | /* Generate rtl for instructions for which shl_and_kind advised a particular |
3428 | method of generating them, i.e. returned zero. */ | |
3429 | ||
3430 | int | |
cf277499 | 3431 | gen_shl_and (rtx dest, rtx left_rtx, rtx mask_rtx, rtx source) |
8d481241 JW |
3432 | { |
3433 | int attributes[3]; | |
3434 | unsigned HOST_WIDE_INT mask; | |
3435 | int kind = shl_and_kind (left_rtx, mask_rtx, attributes); | |
3436 | int right, total_shift; | |
832a3292 | 3437 | void (*shift_gen_fun) (int, rtx *) = gen_shifty_hi_op; |
8d481241 JW |
3438 | |
3439 | right = attributes[0]; | |
3440 | total_shift = INTVAL (left_rtx) + right; | |
3441 | mask = (unsigned HOST_WIDE_INT) INTVAL (mask_rtx) >> total_shift; | |
3442 | switch (kind) | |
3443 | { | |
3444 | default: | |
3445 | return -1; | |
3446 | case 1: | |
3447 | { | |
3448 | int first = attributes[2]; | |
3449 | rtx operands[3]; | |
3450 | ||
3451 | if (first < 0) | |
3452 | { | |
7174c937 | 3453 | emit_insn ((mask << right) <= 0xff |
832a3292 KH |
3454 | ? gen_zero_extendqisi2 (dest, |
3455 | gen_lowpart (QImode, source)) | |
3456 | : gen_zero_extendhisi2 (dest, | |
3457 | gen_lowpart (HImode, source))); | |
8d481241 JW |
3458 | source = dest; |
3459 | } | |
3460 | if (source != dest) | |
3461 | emit_insn (gen_movsi (dest, source)); | |
3462 | operands[0] = dest; | |
3463 | if (right) | |
3464 | { | |
3465 | operands[2] = GEN_INT (right); | |
3466 | gen_shifty_hi_op (LSHIFTRT, operands); | |
3467 | } | |
3468 | if (first > 0) | |
3469 | { | |
3470 | operands[2] = GEN_INT (first); | |
3471 | gen_shifty_hi_op (ASHIFT, operands); | |
3472 | total_shift -= first; | |
3473 | mask <<= first; | |
3474 | } | |
3475 | if (first >= 0) | |
85af47b9 | 3476 | emit_insn (mask <= 0xff |
832a3292 KH |
3477 | ? gen_zero_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
3478 | : gen_zero_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
8d481241 JW |
3479 | if (total_shift > 0) |
3480 | { | |
3481 | operands[2] = GEN_INT (total_shift); | |
3482 | gen_shifty_hi_op (ASHIFT, operands); | |
3483 | } | |
3484 | break; | |
3485 | } | |
3486 | case 4: | |
3487 | shift_gen_fun = gen_shifty_op; | |
8d481241 | 3488 | case 3: |
24c50999 JR |
3489 | /* If the topmost bit that matters is set, set the topmost bits |
3490 | that don't matter. This way, we might be able to get a shorter | |
3491 | signed constant. */ | |
832a3292 KH |
3492 | if (mask & ((HOST_WIDE_INT) 1 << (31 - total_shift))) |
3493 | mask |= (HOST_WIDE_INT) ~0 << (31 - total_shift); | |
61fb6bac | 3494 | case 2: |
8d481241 JW |
3495 | /* Don't expand fine-grained when combining, because that will |
3496 | make the pattern fail. */ | |
4586b4ca | 3497 | if (currently_expanding_to_rtl |
8d481241 JW |
3498 | || reload_in_progress || reload_completed) |
3499 | { | |
3500 | rtx operands[3]; | |
f676971a | 3501 | |
61fb6bac R |
3502 | /* Cases 3 and 4 should be handled by this split |
3503 | only while combining */ | |
f5b9e7c9 | 3504 | gcc_assert (kind <= 2); |
8d481241 JW |
3505 | if (right) |
3506 | { | |
3507 | emit_insn (gen_lshrsi3 (dest, source, GEN_INT (right))); | |
3508 | source = dest; | |
3509 | } | |
3510 | emit_insn (gen_andsi3 (dest, source, GEN_INT (mask))); | |
afad3d2c JW |
3511 | if (total_shift) |
3512 | { | |
3513 | operands[0] = dest; | |
3514 | operands[1] = dest; | |
3515 | operands[2] = GEN_INT (total_shift); | |
3516 | shift_gen_fun (ASHIFT, operands); | |
3517 | } | |
8d481241 JW |
3518 | break; |
3519 | } | |
3520 | else | |
3521 | { | |
3522 | int neg = 0; | |
3523 | if (kind != 4 && total_shift < 16) | |
3524 | { | |
3525 | neg = -ext_shift_amounts[total_shift][1]; | |
3526 | if (neg > 0) | |
3527 | neg -= ext_shift_amounts[total_shift][2]; | |
3528 | else | |
3529 | neg = 0; | |
3530 | } | |
3531 | emit_insn (gen_and_shl_scratch (dest, source, | |
3532 | GEN_INT (right), | |
3533 | GEN_INT (mask), | |
3534 | GEN_INT (total_shift + neg), | |
3535 | GEN_INT (neg))); | |
3536 | emit_insn (gen_movsi (dest, dest)); | |
3537 | break; | |
3538 | } | |
3539 | } | |
3540 | return 0; | |
3541 | } | |
3542 | ||
3543 | /* Try to find a good way to implement the combiner pattern | |
3544 | [(set (match_operand:SI 0 "register_operand" "=r") | |
3545 | (sign_extract:SI (ashift:SI (match_operand:SI 1 "register_operand" "r") | |
3546 | (match_operand:SI 2 "const_int_operand" "n") | |
3547 | (match_operand:SI 3 "const_int_operand" "n") | |
3548 | (const_int 0))) | |
4773afa4 | 3549 | (clobber (reg:SI T_REG))] |
8d481241 JW |
3550 | LEFT_RTX is operand 2 in the above pattern, and SIZE_RTX is operand 3. |
3551 | return 0 for simple left / right shift combination. | |
3552 | return 1 for left shift / 8 bit sign extend / left shift. | |
3553 | return 2 for left shift / 16 bit sign extend / left shift. | |
3554 | return 3 for left shift / 8 bit sign extend / shift / sign extend. | |
3555 | return 4 for left shift / 16 bit sign extend / shift / sign extend. | |
3556 | return 5 for left shift / 16 bit sign extend / right shift | |
3557 | return 6 for < 8 bit sign extend / left shift. | |
3558 | return 7 for < 8 bit sign extend / left shift / single right shift. | |
3559 | If COSTP is nonzero, assign the calculated cost to *COSTP. */ | |
3560 | ||
3561 | int | |
cf277499 | 3562 | shl_sext_kind (rtx left_rtx, rtx size_rtx, int *costp) |
8d481241 JW |
3563 | { |
3564 | int left, size, insize, ext; | |
3a8699c7 | 3565 | int cost = 0, best_cost; |
8d481241 JW |
3566 | int kind; |
3567 | ||
3568 | left = INTVAL (left_rtx); | |
3569 | size = INTVAL (size_rtx); | |
3570 | insize = size - left; | |
f5b9e7c9 | 3571 | gcc_assert (insize > 0); |
8d481241 JW |
3572 | /* Default to left / right shift. */ |
3573 | kind = 0; | |
3574 | best_cost = shift_insns[32 - insize] + ashiftrt_insns[32 - size]; | |
3575 | if (size <= 16) | |
3576 | { | |
3577 | /* 16 bit shift / sign extend / 16 bit shift */ | |
3578 | cost = shift_insns[16 - insize] + 1 + ashiftrt_insns[16 - size]; | |
afad3d2c JW |
3579 | /* If ashiftrt_insns[16 - size] is 8, this choice will be overridden |
3580 | below, by alternative 3 or something even better. */ | |
8d481241 JW |
3581 | if (cost < best_cost) |
3582 | { | |
3583 | kind = 5; | |
3584 | best_cost = cost; | |
3585 | } | |
3586 | } | |
3587 | /* Try a plain sign extend between two shifts. */ | |
3588 | for (ext = 16; ext >= insize; ext -= 8) | |
3589 | { | |
3590 | if (ext <= size) | |
3591 | { | |
3592 | cost = ext_shift_insns[ext - insize] + 1 + shift_insns[size - ext]; | |
3593 | if (cost < best_cost) | |
3594 | { | |
11f9ed1a | 3595 | kind = ext / (unsigned) 8; |
8d481241 JW |
3596 | best_cost = cost; |
3597 | } | |
3598 | } | |
ae9d19c0 JR |
3599 | /* Check if we can do a sloppy shift with a final signed shift |
3600 | restoring the sign. */ | |
3601 | if (EXT_SHIFT_SIGNED (size - ext)) | |
3602 | cost = ext_shift_insns[ext - insize] + ext_shift_insns[size - ext] + 1; | |
3603 | /* If not, maybe it's still cheaper to do the second shift sloppy, | |
3604 | and do a final sign extend? */ | |
3605 | else if (size <= 16) | |
3606 | cost = ext_shift_insns[ext - insize] + 1 | |
3607 | + ext_shift_insns[size > ext ? size - ext : ext - size] + 1; | |
3608 | else | |
3609 | continue; | |
3610 | if (cost < best_cost) | |
8d481241 | 3611 | { |
11f9ed1a | 3612 | kind = ext / (unsigned) 8 + 2; |
ae9d19c0 | 3613 | best_cost = cost; |
8d481241 JW |
3614 | } |
3615 | } | |
3616 | /* Check if we can sign extend in r0 */ | |
3617 | if (insize < 8) | |
3618 | { | |
3619 | cost = 3 + shift_insns[left]; | |
3620 | if (cost < best_cost) | |
3621 | { | |
3622 | kind = 6; | |
3623 | best_cost = cost; | |
3624 | } | |
3625 | /* Try the same with a final signed shift. */ | |
3626 | if (left < 31) | |
3627 | { | |
3628 | cost = 3 + ext_shift_insns[left + 1] + 1; | |
3629 | if (cost < best_cost) | |
3630 | { | |
3631 | kind = 7; | |
3632 | best_cost = cost; | |
3633 | } | |
3634 | } | |
3635 | } | |
3636 | if (TARGET_SH3) | |
3637 | { | |
3638 | /* Try to use a dynamic shift. */ | |
1245df60 | 3639 | cost = shift_insns[32 - insize] + 1 + SH_DYNAMIC_SHIFT_COST; |
8d481241 JW |
3640 | if (cost < best_cost) |
3641 | { | |
3642 | kind = 0; | |
3643 | best_cost = cost; | |
3644 | } | |
3645 | } | |
3646 | if (costp) | |
3647 | *costp = cost; | |
3648 | return kind; | |
3649 | } | |
3650 | ||
3651 | /* Function to be used in the length attribute of the instructions | |
3652 | implementing this pattern. */ | |
3653 | ||
3654 | int | |
cf277499 | 3655 | shl_sext_length (rtx insn) |
8d481241 JW |
3656 | { |
3657 | rtx set_src, left_rtx, size_rtx; | |
3658 | int cost; | |
3659 | ||
3660 | set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
3661 | left_rtx = XEXP (XEXP (set_src, 0), 1); | |
3662 | size_rtx = XEXP (set_src, 1); | |
3663 | shl_sext_kind (left_rtx, size_rtx, &cost); | |
3664 | return cost; | |
3665 | } | |
3666 | ||
3667 | /* Generate rtl for this pattern */ | |
3668 | ||
3669 | int | |
cf277499 | 3670 | gen_shl_sext (rtx dest, rtx left_rtx, rtx size_rtx, rtx source) |
8d481241 JW |
3671 | { |
3672 | int kind; | |
d00d338c | 3673 | int left, size, insize, cost; |
8d481241 JW |
3674 | rtx operands[3]; |
3675 | ||
d00d338c | 3676 | kind = shl_sext_kind (left_rtx, size_rtx, &cost); |
8d481241 JW |
3677 | left = INTVAL (left_rtx); |
3678 | size = INTVAL (size_rtx); | |
3679 | insize = size - left; | |
3680 | switch (kind) | |
3681 | { | |
3682 | case 1: | |
3683 | case 2: | |
3684 | case 3: | |
3685 | case 4: | |
3686 | { | |
3687 | int ext = kind & 1 ? 8 : 16; | |
3688 | int shift2 = size - ext; | |
3689 | ||
3690 | /* Don't expand fine-grained when combining, because that will | |
3691 | make the pattern fail. */ | |
4586b4ca | 3692 | if (! currently_expanding_to_rtl |
8d481241 JW |
3693 | && ! reload_in_progress && ! reload_completed) |
3694 | { | |
3695 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
3696 | emit_insn (gen_movsi (dest, source)); | |
3697 | break; | |
3698 | } | |
3699 | if (dest != source) | |
3700 | emit_insn (gen_movsi (dest, source)); | |
3701 | operands[0] = dest; | |
afad3d2c JW |
3702 | if (ext - insize) |
3703 | { | |
3704 | operands[2] = GEN_INT (ext - insize); | |
3705 | gen_shifty_hi_op (ASHIFT, operands); | |
3706 | } | |
8d481241 | 3707 | emit_insn (kind & 1 |
832a3292 KH |
3708 | ? gen_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
3709 | : gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
8d481241 JW |
3710 | if (kind <= 2) |
3711 | { | |
afad3d2c JW |
3712 | if (shift2) |
3713 | { | |
3714 | operands[2] = GEN_INT (shift2); | |
3715 | gen_shifty_op (ASHIFT, operands); | |
3716 | } | |
8d481241 JW |
3717 | } |
3718 | else | |
3719 | { | |
afad3d2c | 3720 | if (shift2 > 0) |
8d481241 | 3721 | { |
ae9d19c0 JR |
3722 | if (EXT_SHIFT_SIGNED (shift2)) |
3723 | { | |
3724 | operands[2] = GEN_INT (shift2 + 1); | |
3725 | gen_shifty_op (ASHIFT, operands); | |
a556fd39 | 3726 | operands[2] = const1_rtx; |
ae9d19c0 JR |
3727 | gen_shifty_op (ASHIFTRT, operands); |
3728 | break; | |
3729 | } | |
8d481241 JW |
3730 | operands[2] = GEN_INT (shift2); |
3731 | gen_shifty_hi_op (ASHIFT, operands); | |
3732 | } | |
afad3d2c | 3733 | else if (shift2) |
8d481241 JW |
3734 | { |
3735 | operands[2] = GEN_INT (-shift2); | |
3736 | gen_shifty_hi_op (LSHIFTRT, operands); | |
3737 | } | |
3738 | emit_insn (size <= 8 | |
afad3d2c JW |
3739 | ? gen_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
3740 | : gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
8d481241 JW |
3741 | } |
3742 | break; | |
3743 | } | |
3744 | case 5: | |
afad3d2c JW |
3745 | { |
3746 | int i = 16 - size; | |
4586b4ca | 3747 | if (! currently_expanding_to_rtl |
913d8e13 R |
3748 | && ! reload_in_progress && ! reload_completed) |
3749 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
3750 | else | |
3751 | { | |
3752 | operands[0] = dest; | |
3753 | operands[2] = GEN_INT (16 - insize); | |
3754 | gen_shifty_hi_op (ASHIFT, operands); | |
3755 | emit_insn (gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
3756 | } | |
afad3d2c JW |
3757 | /* Don't use gen_ashrsi3 because it generates new pseudos. */ |
3758 | while (--i >= 0) | |
3759 | gen_ashift (ASHIFTRT, 1, dest); | |
3760 | break; | |
3761 | } | |
8d481241 JW |
3762 | case 6: |
3763 | case 7: | |
3764 | /* Don't expand fine-grained when combining, because that will | |
3765 | make the pattern fail. */ | |
4586b4ca | 3766 | if (! currently_expanding_to_rtl |
8d481241 JW |
3767 | && ! reload_in_progress && ! reload_completed) |
3768 | { | |
3769 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
3770 | emit_insn (gen_movsi (dest, source)); | |
3771 | break; | |
3772 | } | |
3773 | emit_insn (gen_andsi3 (dest, source, GEN_INT ((1 << insize) - 1))); | |
3774 | emit_insn (gen_xorsi3 (dest, dest, GEN_INT (1 << (insize - 1)))); | |
3775 | emit_insn (gen_addsi3 (dest, dest, GEN_INT (-1 << (insize - 1)))); | |
3776 | operands[0] = dest; | |
3777 | operands[2] = kind == 7 ? GEN_INT (left + 1) : left_rtx; | |
3778 | gen_shifty_op (ASHIFT, operands); | |
3779 | if (kind == 7) | |
a556fd39 | 3780 | emit_insn (gen_ashrsi3_k (dest, dest, const1_rtx)); |
8d481241 JW |
3781 | break; |
3782 | default: | |
3783 | return -1; | |
3784 | } | |
3785 | return 0; | |
3786 | } | |
fa5322fa AO |
3787 | |
3788 | /* Prefix a symbol_ref name with "datalabel". */ | |
3789 | ||
3790 | rtx | |
cf277499 | 3791 | gen_datalabel_ref (rtx sym) |
fa5322fa | 3792 | { |
73a4d10b R |
3793 | const char *str; |
3794 | ||
fa5322fa AO |
3795 | if (GET_CODE (sym) == LABEL_REF) |
3796 | return gen_rtx_CONST (GET_MODE (sym), | |
3797 | gen_rtx_UNSPEC (GET_MODE (sym), | |
3798 | gen_rtvec (1, sym), | |
3799 | UNSPEC_DATALABEL)); | |
f676971a | 3800 | |
f5b9e7c9 | 3801 | gcc_assert (GET_CODE (sym) == SYMBOL_REF); |
fa5322fa | 3802 | |
73a4d10b R |
3803 | str = XSTR (sym, 0); |
3804 | /* Share all SYMBOL_REF strings with the same value - that is important | |
3805 | for cse. */ | |
3806 | str = IDENTIFIER_POINTER (get_identifier (str)); | |
3807 | XSTR (sym, 0) = str; | |
3808 | ||
fa5322fa AO |
3809 | return sym; |
3810 | } | |
3811 | ||
8aa2a305 | 3812 | \f |
6dd3c0a5 SB |
3813 | static alloc_pool label_ref_list_pool; |
3814 | ||
3815 | typedef struct label_ref_list_d | |
3816 | { | |
3817 | rtx label; | |
3818 | struct label_ref_list_d *next; | |
3819 | } *label_ref_list_t; | |
3820 | ||
8aa2a305 JW |
3821 | /* The SH cannot load a large constant into a register, constants have to |
3822 | come from a pc relative load. The reference of a pc relative load | |
0fa2e4df | 3823 | instruction must be less than 1k in front of the instruction. This |
8aa2a305 JW |
3824 | means that we often have to dump a constant inside a function, and |
3825 | generate code to branch around it. | |
bc45ade3 | 3826 | |
8aa2a305 JW |
3827 | It is important to minimize this, since the branches will slow things |
3828 | down and make things bigger. | |
3829 | ||
3830 | Worst case code looks like: | |
3831 | ||
3832 | mov.l L1,rn | |
3833 | bra L2 | |
3834 | nop | |
3835 | align | |
3836 | L1: .long value | |
3837 | L2: | |
3838 | .. | |
3839 | ||
3840 | mov.l L3,rn | |
3841 | bra L4 | |
3842 | nop | |
3843 | align | |
3844 | L3: .long value | |
3845 | L4: | |
3846 | .. | |
3847 | ||
3848 | We fix this by performing a scan before scheduling, which notices which | |
3849 | instructions need to have their operands fetched from the constant table | |
3850 | and builds the table. | |
3851 | ||
3852 | The algorithm is: | |
3853 | ||
3854 | scan, find an instruction which needs a pcrel move. Look forward, find the | |
3855 | last barrier which is within MAX_COUNT bytes of the requirement. | |
3856 | If there isn't one, make one. Process all the instructions between | |
3857 | the find and the barrier. | |
3858 | ||
3859 | In the above example, we can tell that L3 is within 1k of L1, so | |
3860 | the first move can be shrunk from the 3 insn+constant sequence into | |
3861 | just 1 insn, and the constant moved to L3 to make: | |
3862 | ||
3863 | mov.l L1,rn | |
3864 | .. | |
3865 | mov.l L3,rn | |
3866 | bra L4 | |
3867 | nop | |
3868 | align | |
3869 | L3:.long value | |
3870 | L4:.long value | |
3871 | ||
3872 | Then the second move becomes the target for the shortening process. */ | |
3873 | ||
3874 | typedef struct | |
3875 | { | |
3876 | rtx value; /* Value in table. */ | |
3877 | rtx label; /* Label of value. */ | |
6dd3c0a5 | 3878 | label_ref_list_t wend; /* End of window. */ |
8aa2a305 | 3879 | enum machine_mode mode; /* Mode of value. */ |
3503150c RS |
3880 | |
3881 | /* True if this constant is accessed as part of a post-increment | |
3882 | sequence. Note that HImode constants are never accessed in this way. */ | |
3883 | bool part_of_sequence_p; | |
8aa2a305 JW |
3884 | } pool_node; |
3885 | ||
3886 | /* The maximum number of constants that can fit into one pool, since | |
73a4d10b R |
3887 | constants in the range 0..510 are at least 2 bytes long, and in the |
3888 | range from there to 1018 at least 4 bytes. */ | |
8aa2a305 | 3889 | |
73a4d10b | 3890 | #define MAX_POOL_SIZE 372 |
8aa2a305 JW |
3891 | static pool_node pool_vector[MAX_POOL_SIZE]; |
3892 | static int pool_size; | |
b91455de KK |
3893 | static rtx pool_window_label; |
3894 | static int pool_window_last; | |
8aa2a305 | 3895 | |
4241ecb0 R |
3896 | static int max_labelno_before_reorg; |
3897 | ||
8aa2a305 JW |
3898 | /* ??? If we need a constant in HImode which is the truncated value of a |
3899 | constant we need in SImode, we could combine the two entries thus saving | |
3900 | two bytes. Is this common enough to be worth the effort of implementing | |
3901 | it? */ | |
3902 | ||
3903 | /* ??? This stuff should be done at the same time that we shorten branches. | |
3904 | As it is now, we must assume that all branches are the maximum size, and | |
3905 | this causes us to almost always output constant pools sooner than | |
3906 | necessary. */ | |
3907 | ||
3908 | /* Add a constant to the pool and return its label. */ | |
3909 | ||
3910 | static rtx | |
cf277499 | 3911 | add_constant (rtx x, enum machine_mode mode, rtx last_value) |
0d7e008e SC |
3912 | { |
3913 | int i; | |
0a2aaacc | 3914 | rtx lab, new_rtx; |
6dd3c0a5 | 3915 | label_ref_list_t ref, newref; |
8aa2a305 JW |
3916 | |
3917 | /* First see if we've already got it. */ | |
3918 | for (i = 0; i < pool_size; i++) | |
0d7e008e | 3919 | { |
8aa2a305 JW |
3920 | if (x->code == pool_vector[i].value->code |
3921 | && mode == pool_vector[i].mode) | |
0d7e008e | 3922 | { |
8aa2a305 JW |
3923 | if (x->code == CODE_LABEL) |
3924 | { | |
3925 | if (XINT (x, 3) != XINT (pool_vector[i].value, 3)) | |
3926 | continue; | |
3927 | } | |
3928 | if (rtx_equal_p (x, pool_vector[i].value)) | |
225e4f43 | 3929 | { |
0a2aaacc | 3930 | lab = new_rtx = 0; |
225e4f43 R |
3931 | if (! last_value |
3932 | || ! i | |
3933 | || ! rtx_equal_p (last_value, pool_vector[i-1].value)) | |
3934 | { | |
0a2aaacc KG |
3935 | new_rtx = gen_label_rtx (); |
3936 | LABEL_REFS (new_rtx) = pool_vector[i].label; | |
3937 | pool_vector[i].label = lab = new_rtx; | |
225e4f43 | 3938 | } |
b91455de KK |
3939 | if (lab && pool_window_label) |
3940 | { | |
6dd3c0a5 SB |
3941 | newref = (label_ref_list_t) pool_alloc (label_ref_list_pool); |
3942 | newref->label = pool_window_label; | |
b91455de | 3943 | ref = pool_vector[pool_window_last].wend; |
6dd3c0a5 | 3944 | newref->next = ref; |
b91455de KK |
3945 | pool_vector[pool_window_last].wend = newref; |
3946 | } | |
0a2aaacc KG |
3947 | if (new_rtx) |
3948 | pool_window_label = new_rtx; | |
b91455de | 3949 | pool_window_last = i; |
225e4f43 R |
3950 | return lab; |
3951 | } | |
0d7e008e | 3952 | } |
0d7e008e | 3953 | } |
b9654711 | 3954 | |
8aa2a305 JW |
3955 | /* Need a new one. */ |
3956 | pool_vector[pool_size].value = x; | |
225e4f43 | 3957 | if (last_value && rtx_equal_p (last_value, pool_vector[pool_size - 1].value)) |
3503150c RS |
3958 | { |
3959 | lab = 0; | |
3960 | pool_vector[pool_size - 1].part_of_sequence_p = true; | |
3961 | } | |
225e4f43 R |
3962 | else |
3963 | lab = gen_label_rtx (); | |
8aa2a305 JW |
3964 | pool_vector[pool_size].mode = mode; |
3965 | pool_vector[pool_size].label = lab; | |
6dd3c0a5 | 3966 | pool_vector[pool_size].wend = NULL; |
3503150c | 3967 | pool_vector[pool_size].part_of_sequence_p = (lab == 0); |
b91455de KK |
3968 | if (lab && pool_window_label) |
3969 | { | |
6dd3c0a5 SB |
3970 | newref = (label_ref_list_t) pool_alloc (label_ref_list_pool); |
3971 | newref->label = pool_window_label; | |
b91455de | 3972 | ref = pool_vector[pool_window_last].wend; |
6dd3c0a5 | 3973 | newref->next = ref; |
b91455de KK |
3974 | pool_vector[pool_window_last].wend = newref; |
3975 | } | |
3976 | if (lab) | |
3977 | pool_window_label = lab; | |
3978 | pool_window_last = pool_size; | |
8aa2a305 JW |
3979 | pool_size++; |
3980 | return lab; | |
0d7e008e | 3981 | } |
16bea517 | 3982 | |
078c8b08 R |
3983 | /* Output the literal table. START, if nonzero, is the first instruction |
3984 | this table is needed for, and also indicates that there is at least one | |
3985 | casesi_worker_2 instruction; We have to emit the operand3 labels from | |
3986 | these insns at a 4-byte aligned position. BARRIER is the barrier | |
3987 | after which we are to place the table. */ | |
b9654711 | 3988 | |
b9654711 | 3989 | static void |
078c8b08 | 3990 | dump_table (rtx start, rtx barrier) |
b9654711 | 3991 | { |
078c8b08 | 3992 | rtx scan = barrier; |
0d7e008e | 3993 | int i; |
0d7e008e | 3994 | int need_align = 1; |
6dd3c0a5 SB |
3995 | rtx lab; |
3996 | label_ref_list_t ref; | |
aa06e8f5 | 3997 | int have_df = 0; |
b9654711 | 3998 | |
16bea517 | 3999 | /* Do two passes, first time dump out the HI sized constants. */ |
b9654711 | 4000 | |
0d7e008e | 4001 | for (i = 0; i < pool_size; i++) |
b9654711 | 4002 | { |
8aa2a305 JW |
4003 | pool_node *p = &pool_vector[i]; |
4004 | ||
0d7e008e SC |
4005 | if (p->mode == HImode) |
4006 | { | |
4007 | if (need_align) | |
4008 | { | |
4009 | scan = emit_insn_after (gen_align_2 (), scan); | |
4010 | need_align = 0; | |
4011 | } | |
b91455de KK |
4012 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
4013 | scan = emit_label_after (lab, scan); | |
4014 | scan = emit_insn_after (gen_consttable_2 (p->value, const0_rtx), | |
4015 | scan); | |
6dd3c0a5 | 4016 | for (ref = p->wend; ref; ref = ref->next) |
b91455de | 4017 | { |
6dd3c0a5 | 4018 | lab = ref->label; |
b91455de KK |
4019 | scan = emit_insn_after (gen_consttable_window_end (lab), scan); |
4020 | } | |
0d7e008e | 4021 | } |
aa06e8f5 R |
4022 | else if (p->mode == DFmode) |
4023 | have_df = 1; | |
b9654711 | 4024 | } |
8aa2a305 | 4025 | |
0d7e008e | 4026 | need_align = 1; |
b9654711 | 4027 | |
078c8b08 R |
4028 | if (start) |
4029 | { | |
4030 | scan = emit_insn_after (gen_align_4 (), scan); | |
4031 | need_align = 0; | |
4032 | for (; start != barrier; start = NEXT_INSN (start)) | |
f3536097 | 4033 | if (NONJUMP_INSN_P (start) |
078c8b08 R |
4034 | && recog_memoized (start) == CODE_FOR_casesi_worker_2) |
4035 | { | |
4036 | rtx src = SET_SRC (XVECEXP (PATTERN (start), 0, 0)); | |
4037 | rtx lab = XEXP (XVECEXP (src, 0, 3), 0); | |
4038 | ||
4039 | scan = emit_label_after (lab, scan); | |
4040 | } | |
4041 | } | |
aa06e8f5 | 4042 | if (TARGET_FMOVD && TARGET_ALIGN_DOUBLE && have_df) |
fa5322fa AO |
4043 | { |
4044 | rtx align_insn = NULL_RTX; | |
4045 | ||
4046 | scan = emit_label_after (gen_label_rtx (), scan); | |
4047 | scan = emit_insn_after (gen_align_log (GEN_INT (3)), scan); | |
4048 | need_align = 0; | |
4049 | ||
4050 | for (i = 0; i < pool_size; i++) | |
4051 | { | |
4052 | pool_node *p = &pool_vector[i]; | |
4053 | ||
4054 | switch (p->mode) | |
4055 | { | |
4056 | case HImode: | |
4057 | break; | |
4058 | case SImode: | |
4059 | case SFmode: | |
3503150c | 4060 | if (align_insn && !p->part_of_sequence_p) |
fa5322fa AO |
4061 | { |
4062 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) | |
4063 | emit_label_before (lab, align_insn); | |
4064 | emit_insn_before (gen_consttable_4 (p->value, const0_rtx), | |
4065 | align_insn); | |
6dd3c0a5 | 4066 | for (ref = p->wend; ref; ref = ref->next) |
fa5322fa | 4067 | { |
6dd3c0a5 | 4068 | lab = ref->label; |
fa5322fa | 4069 | emit_insn_before (gen_consttable_window_end (lab), |
832a3292 | 4070 | align_insn); |
fa5322fa AO |
4071 | } |
4072 | delete_insn (align_insn); | |
4073 | align_insn = NULL_RTX; | |
4074 | continue; | |
4075 | } | |
4076 | else | |
4077 | { | |
4078 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) | |
4079 | scan = emit_label_after (lab, scan); | |
4080 | scan = emit_insn_after (gen_consttable_4 (p->value, | |
4081 | const0_rtx), scan); | |
4082 | need_align = ! need_align; | |
4083 | } | |
4084 | break; | |
4085 | case DFmode: | |
fa5322fa AO |
4086 | if (need_align) |
4087 | { | |
4088 | scan = emit_insn_after (gen_align_log (GEN_INT (3)), scan); | |
4089 | align_insn = scan; | |
4090 | need_align = 0; | |
4091 | } | |
aa06e8f5 | 4092 | case DImode: |
fa5322fa AO |
4093 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
4094 | scan = emit_label_after (lab, scan); | |
4095 | scan = emit_insn_after (gen_consttable_8 (p->value, const0_rtx), | |
4096 | scan); | |
4097 | break; | |
4098 | default: | |
f5b9e7c9 | 4099 | gcc_unreachable (); |
fa5322fa AO |
4100 | } |
4101 | ||
4102 | if (p->mode != HImode) | |
4103 | { | |
6dd3c0a5 | 4104 | for (ref = p->wend; ref; ref = ref->next) |
fa5322fa | 4105 | { |
6dd3c0a5 | 4106 | lab = ref->label; |
fa5322fa AO |
4107 | scan = emit_insn_after (gen_consttable_window_end (lab), |
4108 | scan); | |
4109 | } | |
4110 | } | |
4111 | } | |
4112 | ||
4113 | pool_size = 0; | |
4114 | } | |
f676971a | 4115 | |
0d7e008e | 4116 | for (i = 0; i < pool_size; i++) |
b9654711 | 4117 | { |
8aa2a305 | 4118 | pool_node *p = &pool_vector[i]; |
b9654711 | 4119 | |
0d7e008e | 4120 | switch (p->mode) |
b9654711 | 4121 | { |
0d7e008e SC |
4122 | case HImode: |
4123 | break; | |
4124 | case SImode: | |
e577c183 | 4125 | case SFmode: |
0d7e008e | 4126 | if (need_align) |
b9654711 | 4127 | { |
0d7e008e | 4128 | need_align = 0; |
d3ae8277 | 4129 | scan = emit_label_after (gen_label_rtx (), scan); |
0d7e008e | 4130 | scan = emit_insn_after (gen_align_4 (), scan); |
b9654711 | 4131 | } |
b91455de KK |
4132 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
4133 | scan = emit_label_after (lab, scan); | |
4134 | scan = emit_insn_after (gen_consttable_4 (p->value, const0_rtx), | |
4135 | scan); | |
0d7e008e | 4136 | break; |
e577c183 | 4137 | case DFmode: |
0d7e008e SC |
4138 | case DImode: |
4139 | if (need_align) | |
4140 | { | |
4141 | need_align = 0; | |
d3ae8277 | 4142 | scan = emit_label_after (gen_label_rtx (), scan); |
0d7e008e SC |
4143 | scan = emit_insn_after (gen_align_4 (), scan); |
4144 | } | |
b91455de KK |
4145 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
4146 | scan = emit_label_after (lab, scan); | |
4147 | scan = emit_insn_after (gen_consttable_8 (p->value, const0_rtx), | |
4148 | scan); | |
0d7e008e SC |
4149 | break; |
4150 | default: | |
f5b9e7c9 | 4151 | gcc_unreachable (); |
b9654711 | 4152 | } |
b91455de KK |
4153 | |
4154 | if (p->mode != HImode) | |
4155 | { | |
6dd3c0a5 | 4156 | for (ref = p->wend; ref; ref = ref->next) |
b91455de | 4157 | { |
6dd3c0a5 | 4158 | lab = ref->label; |
b91455de KK |
4159 | scan = emit_insn_after (gen_consttable_window_end (lab), scan); |
4160 | } | |
4161 | } | |
b9654711 | 4162 | } |
b9654711 | 4163 | |
0d7e008e SC |
4164 | scan = emit_insn_after (gen_consttable_end (), scan); |
4165 | scan = emit_barrier_after (scan); | |
4166 | pool_size = 0; | |
b91455de KK |
4167 | pool_window_label = NULL_RTX; |
4168 | pool_window_last = 0; | |
0d7e008e | 4169 | } |
b9654711 | 4170 | |
5e7a8ee0 | 4171 | /* Return nonzero if constant would be an ok source for a |
8aa2a305 | 4172 | mov.w instead of a mov.l. */ |
b9654711 | 4173 | |
16bea517 | 4174 | static int |
cf277499 | 4175 | hi_const (rtx src) |
0d7e008e | 4176 | { |
f3536097 | 4177 | return (CONST_INT_P (src) |
8aa2a305 JW |
4178 | && INTVAL (src) >= -32768 |
4179 | && INTVAL (src) <= 32767); | |
b9654711 SC |
4180 | } |
4181 | ||
4241ecb0 R |
4182 | #define MOVA_LABELREF(mova) XVECEXP (SET_SRC (PATTERN (mova)), 0, 0) |
4183 | ||
5e7a8ee0 | 4184 | /* Nonzero if the insn is a move instruction which needs to be fixed. */ |
8aa2a305 JW |
4185 | |
4186 | /* ??? For a DImode/DFmode moves, we don't need to fix it if each half of the | |
735cb76e R |
4187 | CONST_DOUBLE input value is CONST_OK_FOR_I08. For a SFmode move, we don't |
4188 | need to fix it if the input value is CONST_OK_FOR_I08. */ | |
8aa2a305 JW |
4189 | |
4190 | static int | |
cf277499 | 4191 | broken_move (rtx insn) |
b9654711 | 4192 | { |
f3536097 | 4193 | if (NONJUMP_INSN_P (insn)) |
e577c183 JR |
4194 | { |
4195 | rtx pat = PATTERN (insn); | |
4196 | if (GET_CODE (pat) == PARALLEL) | |
4197 | pat = XVECEXP (pat, 0, 0); | |
4198 | if (GET_CODE (pat) == SET | |
3b962ea0 | 4199 | /* We can load any 8-bit value if we don't care what the high |
e577c183 JR |
4200 | order bits end up as. */ |
4201 | && GET_MODE (SET_DEST (pat)) != QImode | |
43c05634 AO |
4202 | && (CONSTANT_P (SET_SRC (pat)) |
4203 | /* Match mova_const. */ | |
4204 | || (GET_CODE (SET_SRC (pat)) == UNSPEC | |
4205 | && XINT (SET_SRC (pat), 1) == UNSPEC_MOVA | |
4206 | && GET_CODE (XVECEXP (SET_SRC (pat), 0, 0)) == CONST)) | |
3a8699c7 | 4207 | && ! (TARGET_SH2E |
1245df60 | 4208 | && GET_CODE (SET_SRC (pat)) == CONST_DOUBLE |
e577c183 JR |
4209 | && (fp_zero_operand (SET_SRC (pat)) |
4210 | || fp_one_operand (SET_SRC (pat))) | |
18778292 R |
4211 | /* ??? If this is a -m4 or -m4-single compilation, in general |
4212 | we don't know the current setting of fpscr, so disable fldi. | |
4213 | There is an exception if this was a register-register move | |
4214 | before reload - and hence it was ascertained that we have | |
4215 | single precision setting - and in a post-reload optimization | |
4216 | we changed this to do a constant load. In that case | |
4217 | we don't have an r0 clobber, hence we must use fldi. */ | |
4218 | && (! TARGET_SH4 || TARGET_FMOVD | |
4219 | || (GET_CODE (XEXP (XVECEXP (PATTERN (insn), 0, 2), 0)) | |
4220 | == SCRATCH)) | |
f3536097 | 4221 | && REG_P (SET_DEST (pat)) |
104ee20b | 4222 | && FP_REGISTER_P (REGNO (SET_DEST (pat)))) |
157371cf AO |
4223 | && ! (TARGET_SH2A |
4224 | && GET_MODE (SET_DEST (pat)) == SImode | |
e990551a JS |
4225 | && (satisfies_constraint_I20 (SET_SRC (pat)) |
4226 | || satisfies_constraint_I28 (SET_SRC (pat)))) | |
32a7ab3d | 4227 | && ! satisfies_constraint_I08 (SET_SRC (pat))) |
e577c183 JR |
4228 | return 1; |
4229 | } | |
d3ae8277 | 4230 | |
8aa2a305 | 4231 | return 0; |
b9654711 | 4232 | } |
b9654711 | 4233 | |
1245df60 | 4234 | static int |
cf277499 | 4235 | mova_p (rtx insn) |
1245df60 | 4236 | { |
f3536097 | 4237 | return (NONJUMP_INSN_P (insn) |
1245df60 R |
4238 | && GET_CODE (PATTERN (insn)) == SET |
4239 | && GET_CODE (SET_SRC (PATTERN (insn))) == UNSPEC | |
43c05634 AO |
4240 | && XINT (SET_SRC (PATTERN (insn)), 1) == UNSPEC_MOVA |
4241 | /* Don't match mova_const. */ | |
4241ecb0 | 4242 | && GET_CODE (MOVA_LABELREF (insn)) == LABEL_REF); |
1245df60 R |
4243 | } |
4244 | ||
078c8b08 R |
4245 | /* Fix up a mova from a switch that went out of range. */ |
4246 | static void | |
4247 | fixup_mova (rtx mova) | |
4248 | { | |
4241ecb0 | 4249 | PUT_MODE (XEXP (MOVA_LABELREF (mova), 0), QImode); |
078c8b08 R |
4250 | if (! flag_pic) |
4251 | { | |
4241ecb0 | 4252 | SET_SRC (PATTERN (mova)) = MOVA_LABELREF (mova); |
078c8b08 R |
4253 | INSN_CODE (mova) = -1; |
4254 | } | |
4255 | else | |
4256 | { | |
4257 | rtx worker = mova; | |
4258 | rtx lab = gen_label_rtx (); | |
dc3ba671 | 4259 | rtx wpat, wpat0, wpat1, wsrc, target, base, diff; |
078c8b08 R |
4260 | |
4261 | do | |
4262 | { | |
4263 | worker = NEXT_INSN (worker); | |
f5b9e7c9 | 4264 | gcc_assert (worker |
f3536097 SZ |
4265 | && !LABEL_P (worker) |
4266 | && !JUMP_P (worker)); | |
4267 | } while (NOTE_P (worker) | |
0bf72672 | 4268 | || recog_memoized (worker) != CODE_FOR_casesi_worker_1); |
078c8b08 R |
4269 | wpat = PATTERN (worker); |
4270 | wpat0 = XVECEXP (wpat, 0, 0); | |
4271 | wpat1 = XVECEXP (wpat, 0, 1); | |
4272 | wsrc = SET_SRC (wpat0); | |
4273 | PATTERN (worker) = (gen_casesi_worker_2 | |
4274 | (SET_DEST (wpat0), XVECEXP (wsrc, 0, 1), | |
4275 | XEXP (XVECEXP (wsrc, 0, 2), 0), lab, | |
4276 | XEXP (wpat1, 0))); | |
4277 | INSN_CODE (worker) = -1; | |
dc3ba671 RS |
4278 | target = XVECEXP (SET_SRC (PATTERN (mova)), 0, 0); |
4279 | base = gen_rtx_LABEL_REF (Pmode, lab); | |
4280 | diff = gen_rtx_UNSPEC (Pmode, gen_rtvec (2, target, base), UNSPEC_SYMOFF); | |
078c8b08 R |
4281 | SET_SRC (PATTERN (mova)) = gen_rtx_CONST (Pmode, diff); |
4282 | INSN_CODE (mova) = -1; | |
4283 | } | |
4284 | } | |
4285 | ||
4241ecb0 R |
4286 | /* NEW_MOVA is a mova we've just encountered while scanning forward. Update |
4287 | *num_mova, and check if the new mova is not nested within the first one. | |
4288 | return 0 if *first_mova was replaced, 1 if new_mova was replaced, | |
4289 | 2 if new_mova has been assigned to *first_mova, -1 otherwise.. */ | |
4290 | static int | |
4291 | untangle_mova (int *num_mova, rtx *first_mova, rtx new_mova) | |
4292 | { | |
c6d56443 KK |
4293 | int n_addr = 0; /* Initialization to shut up spurious warning. */ |
4294 | int f_target, n_target = 0; /* Likewise. */ | |
4241ecb0 R |
4295 | |
4296 | if (optimize) | |
4297 | { | |
77253e56 KK |
4298 | /* If NEW_MOVA has no address yet, it will be handled later. */ |
4299 | if (INSN_ADDRESSES_SIZE() <= (unsigned) INSN_UID (new_mova)) | |
4300 | return -1; | |
4301 | ||
4241ecb0 R |
4302 | n_addr = INSN_ADDRESSES (INSN_UID (new_mova)); |
4303 | n_target = INSN_ADDRESSES (INSN_UID (XEXP (MOVA_LABELREF (new_mova), 0))); | |
4304 | if (n_addr > n_target || n_addr + 1022 < n_target) | |
4305 | { | |
4306 | /* Change the mova into a load. | |
4307 | broken_move will then return true for it. */ | |
4308 | fixup_mova (new_mova); | |
4309 | return 1; | |
4310 | } | |
4311 | } | |
4312 | if (!(*num_mova)++) | |
4313 | { | |
4314 | *first_mova = new_mova; | |
4315 | return 2; | |
4316 | } | |
4317 | if (!optimize | |
4318 | || ((f_target | |
4319 | = INSN_ADDRESSES (INSN_UID (XEXP (MOVA_LABELREF (*first_mova), 0)))) | |
4320 | >= n_target)) | |
4321 | return -1; | |
4322 | ||
4323 | (*num_mova)--; | |
4324 | if (f_target - INSN_ADDRESSES (INSN_UID (*first_mova)) | |
4325 | > n_target - n_addr) | |
4326 | { | |
4327 | fixup_mova (*first_mova); | |
4328 | return 0; | |
4329 | } | |
4330 | else | |
4331 | { | |
4332 | fixup_mova (new_mova); | |
4333 | return 1; | |
4334 | } | |
4335 | } | |
4336 | ||
8aa2a305 JW |
4337 | /* Find the last barrier from insn FROM which is close enough to hold the |
4338 | constant pool. If we can't find one, then create one near the end of | |
4339 | the range. */ | |
b9654711 | 4340 | |
8aa2a305 | 4341 | static rtx |
cf277499 | 4342 | find_barrier (int num_mova, rtx mova, rtx from) |
b9654711 | 4343 | { |
0d7e008e SC |
4344 | int count_si = 0; |
4345 | int count_hi = 0; | |
4346 | int found_hi = 0; | |
4347 | int found_si = 0; | |
fa5322fa | 4348 | int found_di = 0; |
33f7f353 JR |
4349 | int hi_align = 2; |
4350 | int si_align = 2; | |
1245df60 | 4351 | int leading_mova = num_mova; |
3a8699c7 | 4352 | rtx barrier_before_mova = 0, found_barrier = 0, good_barrier = 0; |
c17f53a8 JW |
4353 | int si_limit; |
4354 | int hi_limit; | |
7a3cb3ab | 4355 | rtx orig = from; |
8aa2a305 JW |
4356 | |
4357 | /* For HImode: range is 510, add 4 because pc counts from address of | |
4358 | second instruction after this one, subtract 2 for the jump instruction | |
3adbde60 JW |
4359 | that we may need to emit before the table, subtract 2 for the instruction |
4360 | that fills the jump delay slot (in very rare cases, reorg will take an | |
4361 | instruction from after the constant pool or will leave the delay slot | |
4362 | empty). This gives 510. | |
8aa2a305 JW |
4363 | For SImode: range is 1020, add 4 because pc counts from address of |
4364 | second instruction after this one, subtract 2 in case pc is 2 byte | |
4365 | aligned, subtract 2 for the jump instruction that we may need to emit | |
3adbde60 JW |
4366 | before the table, subtract 2 for the instruction that fills the jump |
4367 | delay slot. This gives 1018. */ | |
c17f53a8 | 4368 | |
1245df60 | 4369 | /* The branch will always be shortened now that the reference address for |
956d6950 | 4370 | forward branches is the successor address, thus we need no longer make |
1245df60 | 4371 | adjustments to the [sh]i_limit for -O0. */ |
c17f53a8 | 4372 | |
1245df60 R |
4373 | si_limit = 1018; |
4374 | hi_limit = 510; | |
e4fa6b06 | 4375 | |
c17f53a8 | 4376 | while (from && count_si < si_limit && count_hi < hi_limit) |
0d7e008e | 4377 | { |
33f7f353 JR |
4378 | int inc = get_attr_length (from); |
4379 | int new_align = 1; | |
1245df60 | 4380 | |
4241ecb0 R |
4381 | /* If this is a label that existed at the time of the compute_alignments |
4382 | call, determine the alignment. N.B. When find_barrier recurses for | |
4383 | an out-of-reach mova, we might see labels at the start of previously | |
4384 | inserted constant tables. */ | |
f3536097 | 4385 | if (LABEL_P (from) |
4241ecb0 | 4386 | && CODE_LABEL_NUMBER (from) <= max_labelno_before_reorg) |
77008a44 R |
4387 | { |
4388 | if (optimize) | |
4389 | new_align = 1 << label_to_alignment (from); | |
f3536097 | 4390 | else if (BARRIER_P (prev_nonnote_insn (from))) |
77008a44 R |
4391 | new_align = 1 << barrier_align (from); |
4392 | else | |
4393 | new_align = 1; | |
4394 | inc = 0; | |
4395 | } | |
4241ecb0 R |
4396 | /* In case we are scanning a constant table because of recursion, check |
4397 | for explicit alignments. If the table is long, we might be forced | |
4398 | to emit the new table in front of it; the length of the alignment | |
4399 | might be the last straw. */ | |
f3536097 | 4400 | else if (NONJUMP_INSN_P (from) |
4241ecb0 R |
4401 | && GET_CODE (PATTERN (from)) == UNSPEC_VOLATILE |
4402 | && XINT (PATTERN (from), 1) == UNSPECV_ALIGN) | |
4403 | new_align = INTVAL (XVECEXP (PATTERN (from), 0, 0)); | |
4404 | /* When we find the end of a constant table, paste the new constant | |
4405 | at the end. That is better than putting it in front because | |
4406 | this way, we don't need extra alignment for adding a 4-byte-aligned | |
4407 | mov(a) label to a 2/4 or 8/4 byte aligned table. */ | |
f3536097 | 4408 | else if (NONJUMP_INSN_P (from) |
4241ecb0 R |
4409 | && GET_CODE (PATTERN (from)) == UNSPEC_VOLATILE |
4410 | && XINT (PATTERN (from), 1) == UNSPECV_CONST_END) | |
4411 | return from; | |
8aa2a305 | 4412 | |
f3536097 | 4413 | if (BARRIER_P (from)) |
1245df60 | 4414 | { |
6d4d15b8 | 4415 | rtx next; |
33f7f353 | 4416 | |
1245df60 | 4417 | found_barrier = from; |
33f7f353 | 4418 | |
956d6950 | 4419 | /* If we are at the end of the function, or in front of an alignment |
1245df60 R |
4420 | instruction, we need not insert an extra alignment. We prefer |
4421 | this kind of barrier. */ | |
33f7f353 | 4422 | if (barrier_align (from) > 2) |
1245df60 | 4423 | good_barrier = from; |
6d4d15b8 KK |
4424 | |
4425 | /* If we are at the end of a hot/cold block, dump the constants | |
4426 | here. */ | |
4427 | next = NEXT_INSN (from); | |
4428 | if (next | |
4429 | && NOTE_P (next) | |
4430 | && NOTE_KIND (next) == NOTE_INSN_SWITCH_TEXT_SECTIONS) | |
4431 | break; | |
1245df60 | 4432 | } |
8aa2a305 | 4433 | |
8aa2a305 | 4434 | if (broken_move (from)) |
0d7e008e | 4435 | { |
1245df60 R |
4436 | rtx pat, src, dst; |
4437 | enum machine_mode mode; | |
4438 | ||
4439 | pat = PATTERN (from); | |
4440 | if (GET_CODE (pat) == PARALLEL) | |
4441 | pat = XVECEXP (pat, 0, 0); | |
4442 | src = SET_SRC (pat); | |
4443 | dst = SET_DEST (pat); | |
4444 | mode = GET_MODE (dst); | |
c17f53a8 JW |
4445 | |
4446 | /* We must explicitly check the mode, because sometimes the | |
4447 | front end will generate code to load unsigned constants into | |
4448 | HImode targets without properly sign extending them. */ | |
225e4f43 R |
4449 | if (mode == HImode |
4450 | || (mode == SImode && hi_const (src) && REGNO (dst) != FPUL_REG)) | |
00e94d65 | 4451 | { |
1245df60 | 4452 | found_hi += 2; |
00e94d65 RK |
4453 | /* We put the short constants before the long constants, so |
4454 | we must count the length of short constants in the range | |
4455 | for the long constants. */ | |
4456 | /* ??? This isn't optimal, but is easy to do. */ | |
1245df60 | 4457 | si_limit -= 2; |
00e94d65 | 4458 | } |
0d7e008e | 4459 | else |
1245df60 | 4460 | { |
fa5322fa AO |
4461 | /* We dump DF/DI constants before SF/SI ones, because |
4462 | the limit is the same, but the alignment requirements | |
4463 | are higher. We may waste up to 4 additional bytes | |
4464 | for alignment, and the DF/DI constant may have | |
24746a42 | 4465 | another SF/SI constant placed before it. */ |
fa5322fa AO |
4466 | if (TARGET_SHCOMPACT |
4467 | && ! found_di | |
4468 | && (mode == DFmode || mode == DImode)) | |
4469 | { | |
4470 | found_di = 1; | |
4471 | si_limit -= 8; | |
4472 | } | |
33f7f353 JR |
4473 | while (si_align > 2 && found_si + si_align - 2 > count_si) |
4474 | si_align >>= 1; | |
1245df60 R |
4475 | if (found_si > count_si) |
4476 | count_si = found_si; | |
4477 | found_si += GET_MODE_SIZE (mode); | |
4478 | if (num_mova) | |
4479 | si_limit -= GET_MODE_SIZE (mode); | |
4480 | } | |
0d7e008e | 4481 | } |
5325c0fa | 4482 | |
33f7f353 | 4483 | if (mova_p (from)) |
1245df60 | 4484 | { |
4241ecb0 | 4485 | switch (untangle_mova (&num_mova, &mova, from)) |
1245df60 | 4486 | { |
4241ecb0 R |
4487 | case 0: return find_barrier (0, 0, mova); |
4488 | case 2: | |
4489 | { | |
4490 | leading_mova = 0; | |
4491 | barrier_before_mova | |
4492 | = good_barrier ? good_barrier : found_barrier; | |
4493 | } | |
4494 | default: break; | |
1245df60 R |
4495 | } |
4496 | if (found_si > count_si) | |
4497 | count_si = found_si; | |
4498 | } | |
f3536097 | 4499 | else if (JUMP_TABLE_DATA_P (from)) |
1245df60 | 4500 | { |
4241ecb0 R |
4501 | if ((num_mova > 1 && GET_MODE (prev_nonnote_insn (from)) == VOIDmode) |
4502 | || (num_mova | |
4503 | && (prev_nonnote_insn (from) | |
4504 | == XEXP (MOVA_LABELREF (mova), 0)))) | |
1245df60 | 4505 | num_mova--; |
10f4f635 | 4506 | if (barrier_align (next_real_insn (from)) == align_jumps_log) |
1245df60 | 4507 | { |
38e01259 | 4508 | /* We have just passed the barrier in front of the |
a0798779 R |
4509 | ADDR_DIFF_VEC, which is stored in found_barrier. Since |
4510 | the ADDR_DIFF_VEC is accessed as data, just like our pool | |
4511 | constants, this is a good opportunity to accommodate what | |
4512 | we have gathered so far. | |
1245df60 R |
4513 | If we waited any longer, we could end up at a barrier in |
4514 | front of code, which gives worse cache usage for separated | |
4515 | instruction / data caches. */ | |
a0798779 | 4516 | good_barrier = found_barrier; |
1245df60 R |
4517 | break; |
4518 | } | |
a0798779 R |
4519 | else |
4520 | { | |
4521 | rtx body = PATTERN (from); | |
4522 | inc = XVECLEN (body, 1) * GET_MODE_SIZE (GET_MODE (body)); | |
4523 | } | |
1245df60 | 4524 | } |
50ceefc2 | 4525 | /* For the SH1, we generate alignments even after jumps-around-jumps. */ |
f3536097 | 4526 | else if (JUMP_P (from) |
50ceefc2 R |
4527 | && ! TARGET_SH2 |
4528 | && ! TARGET_SMALLCODE) | |
4529 | new_align = 4; | |
5325c0fa | 4530 | |
0d7e008e | 4531 | if (found_si) |
33f7f353 | 4532 | { |
50ceefc2 | 4533 | count_si += inc; |
33f7f353 JR |
4534 | if (new_align > si_align) |
4535 | { | |
318881c0 | 4536 | si_limit -= (count_si - 1) & (new_align - si_align); |
33f7f353 JR |
4537 | si_align = new_align; |
4538 | } | |
318881c0 | 4539 | count_si = (count_si + new_align - 1) & -new_align; |
33f7f353 | 4540 | } |
0d7e008e | 4541 | if (found_hi) |
33f7f353 | 4542 | { |
50ceefc2 | 4543 | count_hi += inc; |
33f7f353 JR |
4544 | if (new_align > hi_align) |
4545 | { | |
318881c0 | 4546 | hi_limit -= (count_hi - 1) & (new_align - hi_align); |
33f7f353 JR |
4547 | hi_align = new_align; |
4548 | } | |
318881c0 | 4549 | count_hi = (count_hi + new_align - 1) & -new_align; |
33f7f353 | 4550 | } |
0d7e008e SC |
4551 | from = NEXT_INSN (from); |
4552 | } | |
4553 | ||
1245df60 | 4554 | if (num_mova) |
eda44c78 R |
4555 | { |
4556 | if (leading_mova) | |
4557 | { | |
4558 | /* Try as we might, the leading mova is out of range. Change | |
4559 | it into a load (which will become a pcload) and retry. */ | |
078c8b08 | 4560 | fixup_mova (mova); |
eda44c78 R |
4561 | return find_barrier (0, 0, mova); |
4562 | } | |
4563 | else | |
4564 | { | |
4565 | /* Insert the constant pool table before the mova instruction, | |
4566 | to prevent the mova label reference from going out of range. */ | |
4567 | from = mova; | |
4568 | good_barrier = found_barrier = barrier_before_mova; | |
4569 | } | |
4570 | } | |
5325c0fa | 4571 | |
1245df60 R |
4572 | if (found_barrier) |
4573 | { | |
33f7f353 | 4574 | if (good_barrier && next_real_insn (found_barrier)) |
1245df60 | 4575 | found_barrier = good_barrier; |
1245df60 R |
4576 | } |
4577 | else | |
b9654711 | 4578 | { |
8aa2a305 JW |
4579 | /* We didn't find a barrier in time to dump our stuff, |
4580 | so we'll make one. */ | |
0d7e008e | 4581 | rtx label = gen_label_rtx (); |
8aa2a305 | 4582 | |
5dbcc9c0 RK |
4583 | /* If we exceeded the range, then we must back up over the last |
4584 | instruction we looked at. Otherwise, we just need to undo the | |
4585 | NEXT_INSN at the end of the loop. */ | |
7a3cb3ab KK |
4586 | if (PREV_INSN (from) != orig |
4587 | && (count_hi > hi_limit || count_si > si_limit)) | |
5dbcc9c0 RK |
4588 | from = PREV_INSN (PREV_INSN (from)); |
4589 | else | |
4590 | from = PREV_INSN (from); | |
4591 | ||
8aa2a305 JW |
4592 | /* Walk back to be just before any jump or label. |
4593 | Putting it before a label reduces the number of times the branch | |
4594 | around the constant pool table will be hit. Putting it before | |
4595 | a jump makes it more likely that the bra delay slot will be | |
4596 | filled. */ | |
f3536097 SZ |
4597 | while (NOTE_P (from) || JUMP_P (from) |
4598 | || LABEL_P (from)) | |
8aa2a305 JW |
4599 | from = PREV_INSN (from); |
4600 | ||
0d7e008e SC |
4601 | from = emit_jump_insn_after (gen_jump (label), from); |
4602 | JUMP_LABEL (from) = label; | |
e081ed26 | 4603 | LABEL_NUSES (label) = 1; |
0d7e008e SC |
4604 | found_barrier = emit_barrier_after (from); |
4605 | emit_label_after (label, found_barrier); | |
b9654711 | 4606 | } |
b9654711 | 4607 | |
8aa2a305 | 4608 | return found_barrier; |
0d7e008e | 4609 | } |
b9654711 | 4610 | |
4787bce0 JW |
4611 | /* If the instruction INSN is implemented by a special function, and we can |
4612 | positively find the register that is used to call the sfunc, and this | |
4613 | register is not used anywhere else in this instruction - except as the | |
4614 | destination of a set, return this register; else, return 0. */ | |
1245df60 | 4615 | rtx |
cf277499 | 4616 | sfunc_uses_reg (rtx insn) |
4787bce0 JW |
4617 | { |
4618 | int i; | |
4619 | rtx pattern, part, reg_part, reg; | |
4620 | ||
f3536097 | 4621 | if (!NONJUMP_INSN_P (insn)) |
4787bce0 JW |
4622 | return 0; |
4623 | pattern = PATTERN (insn); | |
4624 | if (GET_CODE (pattern) != PARALLEL || get_attr_type (insn) != TYPE_SFUNC) | |
4625 | return 0; | |
4626 | ||
4627 | for (reg_part = 0, i = XVECLEN (pattern, 0) - 1; i >= 1; i--) | |
4628 | { | |
4629 | part = XVECEXP (pattern, 0, i); | |
1245df60 | 4630 | if (GET_CODE (part) == USE && GET_MODE (XEXP (part, 0)) == SImode) |
4787bce0 JW |
4631 | reg_part = part; |
4632 | } | |
4633 | if (! reg_part) | |
4634 | return 0; | |
4635 | reg = XEXP (reg_part, 0); | |
4636 | for (i = XVECLEN (pattern, 0) - 1; i >= 0; i--) | |
4637 | { | |
4638 | part = XVECEXP (pattern, 0, i); | |
225e4f43 | 4639 | if (part == reg_part || GET_CODE (part) == CLOBBER) |
4787bce0 JW |
4640 | continue; |
4641 | if (reg_mentioned_p (reg, ((GET_CODE (part) == SET | |
f3536097 | 4642 | && REG_P (SET_DEST (part))) |
4787bce0 JW |
4643 | ? SET_SRC (part) : part))) |
4644 | return 0; | |
4645 | } | |
4646 | return reg; | |
4647 | } | |
4648 | ||
933c3ba3 JW |
4649 | /* See if the only way in which INSN uses REG is by calling it, or by |
4650 | setting it while calling it. Set *SET to a SET rtx if the register | |
4651 | is set by INSN. */ | |
4652 | ||
4653 | static int | |
cf277499 | 4654 | noncall_uses_reg (rtx reg, rtx insn, rtx *set) |
933c3ba3 | 4655 | { |
4787bce0 | 4656 | rtx pattern, reg2; |
933c3ba3 JW |
4657 | |
4658 | *set = NULL_RTX; | |
4659 | ||
4787bce0 JW |
4660 | reg2 = sfunc_uses_reg (insn); |
4661 | if (reg2 && REGNO (reg2) == REGNO (reg)) | |
4662 | { | |
4663 | pattern = single_set (insn); | |
4664 | if (pattern | |
f3536097 | 4665 | && REG_P (SET_DEST (pattern)) |
4787bce0 JW |
4666 | && REGNO (reg) == REGNO (SET_DEST (pattern))) |
4667 | *set = pattern; | |
4668 | return 0; | |
4669 | } | |
f3536097 | 4670 | if (!CALL_P (insn)) |
933c3ba3 JW |
4671 | { |
4672 | /* We don't use rtx_equal_p because we don't care if the mode is | |
4673 | different. */ | |
4674 | pattern = single_set (insn); | |
4675 | if (pattern | |
f3536097 | 4676 | && REG_P (SET_DEST (pattern)) |
933c3ba3 JW |
4677 | && REGNO (reg) == REGNO (SET_DEST (pattern))) |
4678 | { | |
4787bce0 JW |
4679 | rtx par, part; |
4680 | int i; | |
4681 | ||
933c3ba3 | 4682 | *set = pattern; |
4787bce0 JW |
4683 | par = PATTERN (insn); |
4684 | if (GET_CODE (par) == PARALLEL) | |
4685 | for (i = XVECLEN (par, 0) - 1; i >= 0; i--) | |
4686 | { | |
4687 | part = XVECEXP (par, 0, i); | |
4688 | if (GET_CODE (part) != SET && reg_mentioned_p (reg, part)) | |
4689 | return 1; | |
4690 | } | |
4691 | return reg_mentioned_p (reg, SET_SRC (pattern)); | |
933c3ba3 JW |
4692 | } |
4693 | ||
4694 | return 1; | |
4695 | } | |
4696 | ||
4697 | pattern = PATTERN (insn); | |
4698 | ||
4699 | if (GET_CODE (pattern) == PARALLEL) | |
4700 | { | |
4701 | int i; | |
4702 | ||
4703 | for (i = XVECLEN (pattern, 0) - 1; i >= 1; i--) | |
4704 | if (reg_mentioned_p (reg, XVECEXP (pattern, 0, i))) | |
4705 | return 1; | |
4706 | pattern = XVECEXP (pattern, 0, 0); | |
4707 | } | |
4708 | ||
4709 | if (GET_CODE (pattern) == SET) | |
4710 | { | |
4711 | if (reg_mentioned_p (reg, SET_DEST (pattern))) | |
4712 | { | |
4713 | /* We don't use rtx_equal_p, because we don't care if the | |
4714 | mode is different. */ | |
f3536097 | 4715 | if (!REG_P (SET_DEST (pattern)) |
933c3ba3 JW |
4716 | || REGNO (reg) != REGNO (SET_DEST (pattern))) |
4717 | return 1; | |
4718 | ||
4719 | *set = pattern; | |
4720 | } | |
4721 | ||
4722 | pattern = SET_SRC (pattern); | |
4723 | } | |
4724 | ||
4725 | if (GET_CODE (pattern) != CALL | |
f3536097 | 4726 | || !MEM_P (XEXP (pattern, 0)) |
933c3ba3 JW |
4727 | || ! rtx_equal_p (reg, XEXP (XEXP (pattern, 0), 0))) |
4728 | return 1; | |
4729 | ||
4730 | return 0; | |
4731 | } | |
4732 | ||
1245df60 R |
4733 | /* Given a X, a pattern of an insn or a part of it, return a mask of used |
4734 | general registers. Bits 0..15 mean that the respective registers | |
4735 | are used as inputs in the instruction. Bits 16..31 mean that the | |
4736 | registers 0..15, respectively, are used as outputs, or are clobbered. | |
4737 | IS_DEST should be set to 16 if X is the destination of a SET, else to 0. */ | |
4738 | int | |
cf277499 | 4739 | regs_used (rtx x, int is_dest) |
1245df60 R |
4740 | { |
4741 | enum rtx_code code; | |
6f7d635c | 4742 | const char *fmt; |
1245df60 R |
4743 | int i, used = 0; |
4744 | ||
4745 | if (! x) | |
4746 | return used; | |
4747 | code = GET_CODE (x); | |
4748 | switch (code) | |
4749 | { | |
4750 | case REG: | |
4751 | if (REGNO (x) < 16) | |
4752 | return (((1 << HARD_REGNO_NREGS (0, GET_MODE (x))) - 1) | |
4753 | << (REGNO (x) + is_dest)); | |
4754 | return 0; | |
4755 | case SUBREG: | |
4756 | { | |
4757 | rtx y = SUBREG_REG (x); | |
f676971a | 4758 | |
f3536097 | 4759 | if (!REG_P (y)) |
1245df60 R |
4760 | break; |
4761 | if (REGNO (y) < 16) | |
4762 | return (((1 << HARD_REGNO_NREGS (0, GET_MODE (x))) - 1) | |
ddef6bc7 JJ |
4763 | << (REGNO (y) + |
4764 | subreg_regno_offset (REGNO (y), | |
4765 | GET_MODE (y), | |
4766 | SUBREG_BYTE (x), | |
4767 | GET_MODE (x)) + is_dest)); | |
1245df60 R |
4768 | return 0; |
4769 | } | |
4770 | case SET: | |
4771 | return regs_used (SET_SRC (x), 0) | regs_used (SET_DEST (x), 16); | |
4772 | case RETURN: | |
4773 | /* If there was a return value, it must have been indicated with USE. */ | |
4774 | return 0x00ffff00; | |
4775 | case CLOBBER: | |
4776 | is_dest = 1; | |
4777 | break; | |
4778 | case MEM: | |
4779 | is_dest = 0; | |
4780 | break; | |
4781 | case CALL: | |
4782 | used |= 0x00ff00f0; | |
4783 | break; | |
318881c0 KG |
4784 | default: |
4785 | break; | |
1245df60 R |
4786 | } |
4787 | ||
4788 | fmt = GET_RTX_FORMAT (code); | |
4789 | ||
4790 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
4791 | { | |
4792 | if (fmt[i] == 'E') | |
4793 | { | |
4794 | register int j; | |
4795 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
4796 | used |= regs_used (XVECEXP (x, i, j), is_dest); | |
4797 | } | |
4798 | else if (fmt[i] == 'e') | |
4799 | used |= regs_used (XEXP (x, i), is_dest); | |
4800 | } | |
4801 | return used; | |
4802 | } | |
4803 | ||
4804 | /* Create an instruction that prevents redirection of a conditional branch | |
956d6950 | 4805 | to the destination of the JUMP with address ADDR. |
1245df60 R |
4806 | If the branch needs to be implemented as an indirect jump, try to find |
4807 | a scratch register for it. | |
4808 | If NEED_BLOCK is 0, don't do anything unless we need a scratch register. | |
4809 | If any preceding insn that doesn't fit into a delay slot is good enough, | |
4810 | pass 1. Pass 2 if a definite blocking insn is needed. | |
4811 | -1 is used internally to avoid deep recursion. | |
4812 | If a blocking instruction is made or recognized, return it. */ | |
f676971a | 4813 | |
1245df60 | 4814 | static rtx |
cf277499 | 4815 | gen_block_redirect (rtx jump, int addr, int need_block) |
1245df60 R |
4816 | { |
4817 | int dead = 0; | |
4818 | rtx prev = prev_nonnote_insn (jump); | |
4819 | rtx dest; | |
4820 | ||
4821 | /* First, check if we already have an instruction that satisfies our need. */ | |
f3536097 | 4822 | if (prev && NONJUMP_INSN_P (prev) && ! INSN_DELETED_P (prev)) |
1245df60 R |
4823 | { |
4824 | if (INSN_CODE (prev) == CODE_FOR_indirect_jump_scratch) | |
4825 | return prev; | |
4826 | if (GET_CODE (PATTERN (prev)) == USE | |
4827 | || GET_CODE (PATTERN (prev)) == CLOBBER | |
4828 | || get_attr_in_delay_slot (prev) == IN_DELAY_SLOT_YES) | |
4829 | prev = jump; | |
4830 | else if ((need_block &= ~1) < 0) | |
4831 | return prev; | |
4832 | else if (recog_memoized (prev) == CODE_FOR_block_branch_redirect) | |
4833 | need_block = 0; | |
4834 | } | |
f9654065 R |
4835 | if (GET_CODE (PATTERN (jump)) == RETURN) |
4836 | { | |
4837 | if (! need_block) | |
4838 | return prev; | |
4839 | /* Reorg even does nasty things with return insns that cause branches | |
4840 | to go out of range - see find_end_label and callers. */ | |
a556fd39 | 4841 | return emit_insn_before (gen_block_branch_redirect (const0_rtx) , jump); |
f9654065 | 4842 | } |
1245df60 R |
4843 | /* We can't use JUMP_LABEL here because it might be undefined |
4844 | when not optimizing. */ | |
4845 | dest = XEXP (SET_SRC (PATTERN (jump)), 0); | |
4846 | /* If the branch is out of range, try to find a scratch register for it. */ | |
4847 | if (optimize | |
11f9ed1a KG |
4848 | && (INSN_ADDRESSES (INSN_UID (dest)) - addr + (unsigned) 4092 |
4849 | > 4092 + 4098)) | |
1245df60 R |
4850 | { |
4851 | rtx scan; | |
4852 | /* Don't look for the stack pointer as a scratch register, | |
956d6950 | 4853 | it would cause trouble if an interrupt occurred. */ |
0a2aaacc | 4854 | unsigned attempt = 0x7fff, used; |
1245df60 | 4855 | int jump_left = flag_expensive_optimizations + 1; |
f676971a | 4856 | |
1245df60 R |
4857 | /* It is likely that the most recent eligible instruction is wanted for |
4858 | the delay slot. Therefore, find out which registers it uses, and | |
4859 | try to avoid using them. */ | |
f676971a | 4860 | |
318881c0 | 4861 | for (scan = jump; (scan = PREV_INSN (scan)); ) |
1245df60 R |
4862 | { |
4863 | enum rtx_code code; | |
4864 | ||
4865 | if (INSN_DELETED_P (scan)) | |
4866 | continue; | |
4867 | code = GET_CODE (scan); | |
4868 | if (code == CODE_LABEL || code == JUMP_INSN) | |
4869 | break; | |
4870 | if (code == INSN | |
4871 | && GET_CODE (PATTERN (scan)) != USE | |
4872 | && GET_CODE (PATTERN (scan)) != CLOBBER | |
4873 | && get_attr_in_delay_slot (scan) == IN_DELAY_SLOT_YES) | |
4874 | { | |
0a2aaacc | 4875 | attempt &= ~regs_used (PATTERN (scan), 0); |
1245df60 R |
4876 | break; |
4877 | } | |
4878 | } | |
eda44c78 R |
4879 | for (used = dead = 0, scan = JUMP_LABEL (jump); |
4880 | (scan = NEXT_INSN (scan)); ) | |
1245df60 R |
4881 | { |
4882 | enum rtx_code code; | |
4883 | ||
4884 | if (INSN_DELETED_P (scan)) | |
4885 | continue; | |
4886 | code = GET_CODE (scan); | |
ec8e098d | 4887 | if (INSN_P (scan)) |
1245df60 R |
4888 | { |
4889 | used |= regs_used (PATTERN (scan), 0); | |
4890 | if (code == CALL_INSN) | |
4891 | used |= regs_used (CALL_INSN_FUNCTION_USAGE (scan), 0); | |
4892 | dead |= (used >> 16) & ~used; | |
0a2aaacc | 4893 | if (dead & attempt) |
1245df60 | 4894 | { |
0a2aaacc | 4895 | dead &= attempt; |
1245df60 R |
4896 | break; |
4897 | } | |
4898 | if (code == JUMP_INSN) | |
eda44c78 R |
4899 | { |
4900 | if (jump_left-- && simplejump_p (scan)) | |
4901 | scan = JUMP_LABEL (scan); | |
4902 | else | |
4903 | break; | |
4904 | } | |
1245df60 R |
4905 | } |
4906 | } | |
4907 | /* Mask out the stack pointer again, in case it was | |
4908 | the only 'free' register we have found. */ | |
4909 | dead &= 0x7fff; | |
4910 | } | |
4911 | /* If the immediate destination is still in range, check for possible | |
4912 | threading with a jump beyond the delay slot insn. | |
4913 | Don't check if we are called recursively; the jump has been or will be | |
956d6950 | 4914 | checked in a different invocation then. */ |
f676971a | 4915 | |
1245df60 R |
4916 | else if (optimize && need_block >= 0) |
4917 | { | |
4918 | rtx next = next_active_insn (next_active_insn (dest)); | |
f3536097 | 4919 | if (next && JUMP_P (next) |
1245df60 | 4920 | && GET_CODE (PATTERN (next)) == SET |
ddf8ec7e | 4921 | && recog_memoized (next) == CODE_FOR_jump_compact) |
1245df60 R |
4922 | { |
4923 | dest = JUMP_LABEL (next); | |
4924 | if (dest | |
11f9ed1a | 4925 | && (INSN_ADDRESSES (INSN_UID (dest)) - addr + (unsigned) 4092 |
9d98a694 AO |
4926 | > 4092 + 4098)) |
4927 | gen_block_redirect (next, INSN_ADDRESSES (INSN_UID (next)), -1); | |
1245df60 R |
4928 | } |
4929 | } | |
4930 | ||
4931 | if (dead) | |
4932 | { | |
c5c76735 | 4933 | rtx reg = gen_rtx_REG (SImode, exact_log2 (dead & -dead)); |
1245df60 R |
4934 | |
4935 | /* It would be nice if we could convert the jump into an indirect | |
956d6950 | 4936 | jump / far branch right now, and thus exposing all constituent |
1245df60 R |
4937 | instructions to further optimization. However, reorg uses |
4938 | simplejump_p to determine if there is an unconditional jump where | |
4939 | it should try to schedule instructions from the target of the | |
4940 | branch; simplejump_p fails for indirect jumps even if they have | |
4941 | a JUMP_LABEL. */ | |
4942 | rtx insn = emit_insn_before (gen_indirect_jump_scratch | |
4943 | (reg, GEN_INT (INSN_UID (JUMP_LABEL (jump)))) | |
4944 | , jump); | |
10f4f635 R |
4945 | /* ??? We would like this to have the scope of the jump, but that |
4946 | scope will change when a delay slot insn of an inner scope is added. | |
4947 | Hence, after delay slot scheduling, we'll have to expect | |
4948 | NOTE_INSN_BLOCK_END notes between the indirect_jump_scratch and | |
4949 | the jump. */ | |
f676971a | 4950 | |
86143cb4 | 4951 | INSN_LOCATOR (insn) = INSN_LOCATOR (jump); |
1245df60 R |
4952 | INSN_CODE (insn) = CODE_FOR_indirect_jump_scratch; |
4953 | return insn; | |
4954 | } | |
4955 | else if (need_block) | |
4956 | /* We can't use JUMP_LABEL here because it might be undefined | |
4957 | when not optimizing. */ | |
4958 | return emit_insn_before (gen_block_branch_redirect | |
4959 | (GEN_INT (INSN_UID (XEXP (SET_SRC (PATTERN (jump)), 0)))) | |
4960 | , jump); | |
4961 | return prev; | |
4962 | } | |
4963 | ||
4964 | #define CONDJUMP_MIN -252 | |
4965 | #define CONDJUMP_MAX 262 | |
4966 | struct far_branch | |
4967 | { | |
4968 | /* A label (to be placed) in front of the jump | |
4969 | that jumps to our ultimate destination. */ | |
4970 | rtx near_label; | |
4971 | /* Where we are going to insert it if we cannot move the jump any farther, | |
4972 | or the jump itself if we have picked up an existing jump. */ | |
4973 | rtx insert_place; | |
4974 | /* The ultimate destination. */ | |
4975 | rtx far_label; | |
4976 | struct far_branch *prev; | |
4977 | /* If the branch has already been created, its address; | |
4978 | else the address of its first prospective user. */ | |
4979 | int address; | |
4980 | }; | |
4981 | ||
cf277499 | 4982 | static void gen_far_branch (struct far_branch *); |
1245df60 | 4983 | enum mdep_reorg_phase_e mdep_reorg_phase; |
c5b9ef02 | 4984 | static void |
cf277499 | 4985 | gen_far_branch (struct far_branch *bp) |
1245df60 R |
4986 | { |
4987 | rtx insn = bp->insert_place; | |
4988 | rtx jump; | |
4989 | rtx label = gen_label_rtx (); | |
f5b9e7c9 | 4990 | int ok; |
1245df60 R |
4991 | |
4992 | emit_label_after (label, insn); | |
4993 | if (bp->far_label) | |
4994 | { | |
4995 | jump = emit_jump_insn_after (gen_jump (bp->far_label), insn); | |
4996 | LABEL_NUSES (bp->far_label)++; | |
4997 | } | |
4998 | else | |
4999 | jump = emit_jump_insn_after (gen_return (), insn); | |
225e4f43 R |
5000 | /* Emit a barrier so that reorg knows that any following instructions |
5001 | are not reachable via a fall-through path. | |
fae778eb | 5002 | But don't do this when not optimizing, since we wouldn't suppress the |
225e4f43 R |
5003 | alignment for the barrier then, and could end up with out-of-range |
5004 | pc-relative loads. */ | |
5005 | if (optimize) | |
5006 | emit_barrier_after (jump); | |
1245df60 R |
5007 | emit_label_after (bp->near_label, insn); |
5008 | JUMP_LABEL (jump) = bp->far_label; | |
f5b9e7c9 NS |
5009 | ok = invert_jump (insn, label, 1); |
5010 | gcc_assert (ok); | |
5011 | ||
f9654065 R |
5012 | /* If we are branching around a jump (rather than a return), prevent |
5013 | reorg from using an insn from the jump target as the delay slot insn - | |
5014 | when reorg did this, it pessimized code (we rather hide the delay slot) | |
5015 | and it could cause branches to go out of range. */ | |
5016 | if (bp->far_label) | |
5017 | (emit_insn_after | |
5018 | (gen_stuff_delay_slot | |
5019 | (GEN_INT (INSN_UID (XEXP (SET_SRC (PATTERN (jump)), 0))), | |
5020 | GEN_INT (recog_memoized (insn) == CODE_FOR_branch_false)), | |
5021 | insn)); | |
1245df60 R |
5022 | /* Prevent reorg from undoing our splits. */ |
5023 | gen_block_redirect (jump, bp->address += 2, 2); | |
5024 | } | |
5025 | ||
1245df60 R |
5026 | /* Fix up ADDR_DIFF_VECs. */ |
5027 | void | |
cf277499 | 5028 | fixup_addr_diff_vecs (rtx first) |
1245df60 R |
5029 | { |
5030 | rtx insn; | |
33f7f353 | 5031 | |
1245df60 R |
5032 | for (insn = first; insn; insn = NEXT_INSN (insn)) |
5033 | { | |
eb3881bf | 5034 | rtx vec_lab, pat, prev, prevpat, x, braf_label; |
1245df60 | 5035 | |
f3536097 | 5036 | if (!JUMP_P (insn) |
1245df60 R |
5037 | || GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC) |
5038 | continue; | |
5039 | pat = PATTERN (insn); | |
33f7f353 | 5040 | vec_lab = XEXP (XEXP (pat, 0), 0); |
1245df60 | 5041 | |
33f7f353 JR |
5042 | /* Search the matching casesi_jump_2. */ |
5043 | for (prev = vec_lab; ; prev = PREV_INSN (prev)) | |
1245df60 | 5044 | { |
f3536097 | 5045 | if (!JUMP_P (prev)) |
33f7f353 JR |
5046 | continue; |
5047 | prevpat = PATTERN (prev); | |
5048 | if (GET_CODE (prevpat) != PARALLEL || XVECLEN (prevpat, 0) != 2) | |
5049 | continue; | |
5050 | x = XVECEXP (prevpat, 0, 1); | |
5051 | if (GET_CODE (x) != USE) | |
5052 | continue; | |
5053 | x = XEXP (x, 0); | |
5054 | if (GET_CODE (x) == LABEL_REF && XEXP (x, 0) == vec_lab) | |
5055 | break; | |
1245df60 | 5056 | } |
157371cf AO |
5057 | /* FIXME: This is a bug in the optimizer, but it seems harmless |
5058 | to just avoid panicing. */ | |
5059 | if (!prev) | |
5060 | continue; | |
eb3881bf R |
5061 | |
5062 | /* Emit the reference label of the braf where it belongs, right after | |
5063 | the casesi_jump_2 (i.e. braf). */ | |
5064 | braf_label = XEXP (XEXP (SET_SRC (XVECEXP (prevpat, 0, 0)), 1), 0); | |
5065 | emit_label_after (braf_label, prev); | |
5066 | ||
33f7f353 JR |
5067 | /* Fix up the ADDR_DIF_VEC to be relative |
5068 | to the reference address of the braf. */ | |
eb3881bf | 5069 | XEXP (XEXP (pat, 0), 0) = braf_label; |
1245df60 | 5070 | } |
1245df60 R |
5071 | } |
5072 | ||
33f7f353 JR |
5073 | /* BARRIER_OR_LABEL is either a BARRIER or a CODE_LABEL immediately following |
5074 | a barrier. Return the base 2 logarithm of the desired alignment. */ | |
1245df60 | 5075 | int |
cf277499 | 5076 | barrier_align (rtx barrier_or_label) |
1245df60 | 5077 | { |
33f7f353 | 5078 | rtx next = next_real_insn (barrier_or_label), pat, prev; |
3a8699c7 | 5079 | int slot, credit, jump_to_next = 0; |
f676971a | 5080 | |
33f7f353 | 5081 | if (! next) |
1245df60 | 5082 | return 0; |
1245df60 | 5083 | |
33f7f353 | 5084 | pat = PATTERN (next); |
1245df60 | 5085 | |
33f7f353 JR |
5086 | if (GET_CODE (pat) == ADDR_DIFF_VEC) |
5087 | return 2; | |
1245df60 | 5088 | |
4773afa4 | 5089 | if (GET_CODE (pat) == UNSPEC_VOLATILE && XINT (pat, 1) == UNSPECV_ALIGN) |
33f7f353 JR |
5090 | /* This is a barrier in front of a constant table. */ |
5091 | return 0; | |
5092 | ||
5093 | prev = prev_real_insn (barrier_or_label); | |
5094 | if (GET_CODE (PATTERN (prev)) == ADDR_DIFF_VEC) | |
1245df60 | 5095 | { |
33f7f353 JR |
5096 | pat = PATTERN (prev); |
5097 | /* If this is a very small table, we want to keep the alignment after | |
5098 | the table to the minimum for proper code alignment. */ | |
5099 | return ((TARGET_SMALLCODE | |
d042370e | 5100 | || ((unsigned) XVECLEN (pat, 1) * GET_MODE_SIZE (GET_MODE (pat)) |
832a3292 | 5101 | <= (unsigned) 1 << (CACHE_LOG - 2))) |
10f4f635 | 5102 | ? 1 << TARGET_SHMEDIA : align_jumps_log); |
1245df60 | 5103 | } |
33f7f353 JR |
5104 | |
5105 | if (TARGET_SMALLCODE) | |
5106 | return 0; | |
5107 | ||
4d070fd3 | 5108 | if (! TARGET_SH2 || ! optimize) |
10f4f635 | 5109 | return align_jumps_log; |
33f7f353 | 5110 | |
225e4f43 R |
5111 | /* When fixing up pcloads, a constant table might be inserted just before |
5112 | the basic block that ends with the barrier. Thus, we can't trust the | |
5113 | instruction lengths before that. */ | |
5114 | if (mdep_reorg_phase > SH_FIXUP_PCLOAD) | |
1245df60 | 5115 | { |
225e4f43 R |
5116 | /* Check if there is an immediately preceding branch to the insn beyond |
5117 | the barrier. We must weight the cost of discarding useful information | |
5118 | from the current cache line when executing this branch and there is | |
5119 | an alignment, against that of fetching unneeded insn in front of the | |
5120 | branch target when there is no alignment. */ | |
5121 | ||
f676971a EC |
5122 | /* There are two delay_slot cases to consider. One is the simple case |
5123 | where the preceding branch is to the insn beyond the barrier (simple | |
5124 | delay slot filling), and the other is where the preceding branch has | |
5125 | a delay slot that is a duplicate of the insn after the barrier | |
5126 | (fill_eager_delay_slots) and the branch is to the insn after the insn | |
b1b41fe4 CP |
5127 | after the barrier. */ |
5128 | ||
225e4f43 R |
5129 | /* PREV is presumed to be the JUMP_INSN for the barrier under |
5130 | investigation. Skip to the insn before it. */ | |
5131 | prev = prev_real_insn (prev); | |
5132 | ||
c5b9ef02 | 5133 | for (slot = 2, credit = (1 << (CACHE_LOG - 2)) + 2; |
f3536097 | 5134 | credit >= 0 && prev && NONJUMP_INSN_P (prev); |
225e4f43 R |
5135 | prev = prev_real_insn (prev)) |
5136 | { | |
b1b41fe4 | 5137 | jump_to_next = 0; |
225e4f43 R |
5138 | if (GET_CODE (PATTERN (prev)) == USE |
5139 | || GET_CODE (PATTERN (prev)) == CLOBBER) | |
5140 | continue; | |
5141 | if (GET_CODE (PATTERN (prev)) == SEQUENCE) | |
b1b41fe4 CP |
5142 | { |
5143 | prev = XVECEXP (PATTERN (prev), 0, 1); | |
f676971a | 5144 | if (INSN_UID (prev) == INSN_UID (next)) |
b1b41fe4 CP |
5145 | { |
5146 | /* Delay slot was filled with insn at jump target. */ | |
5147 | jump_to_next = 1; | |
5148 | continue; | |
5149 | } | |
5150 | } | |
5151 | ||
225e4f43 R |
5152 | if (slot && |
5153 | get_attr_in_delay_slot (prev) == IN_DELAY_SLOT_YES) | |
5154 | slot = 0; | |
5155 | credit -= get_attr_length (prev); | |
5156 | } | |
5157 | if (prev | |
f3536097 | 5158 | && JUMP_P (prev) |
2b9d17c7 AH |
5159 | && JUMP_LABEL (prev)) |
5160 | { | |
5161 | rtx x; | |
300c5453 | 5162 | if (jump_to_next |
2b9d17c7 | 5163 | || next_real_insn (JUMP_LABEL (prev)) == next |
1594c6bc AO |
5164 | /* If relax_delay_slots() decides NEXT was redundant |
5165 | with some previous instruction, it will have | |
5166 | redirected PREV's jump to the following insn. */ | |
5d291213 | 5167 | || JUMP_LABEL (prev) == next_nonnote_insn (next) |
300c5453 AO |
5168 | /* There is no upper bound on redundant instructions |
5169 | that might have been skipped, but we must not put an | |
2b9d17c7 | 5170 | alignment where none had been before. */ |
f676971a EC |
5171 | || (x = (NEXT_INSN (NEXT_INSN (PREV_INSN (prev)))), |
5172 | (INSN_P (x) | |
2b9d17c7 | 5173 | && (INSN_CODE (x) == CODE_FOR_block_branch_redirect |
9959c8c7 R |
5174 | || INSN_CODE (x) == CODE_FOR_indirect_jump_scratch |
5175 | || INSN_CODE (x) == CODE_FOR_stuff_delay_slot)))) | |
2b9d17c7 AH |
5176 | { |
5177 | rtx pat = PATTERN (prev); | |
5178 | if (GET_CODE (pat) == PARALLEL) | |
300c5453 | 5179 | pat = XVECEXP (pat, 0, 0); |
2b9d17c7 AH |
5180 | if (credit - slot >= (GET_CODE (SET_SRC (pat)) == PC ? 2 : 0)) |
5181 | return 0; | |
5182 | } | |
300c5453 | 5183 | } |
33f7f353 | 5184 | } |
f676971a | 5185 | |
10f4f635 | 5186 | return align_jumps_log; |
1245df60 R |
5187 | } |
5188 | ||
2148624a R |
5189 | /* If we are inside a phony loop, almost any kind of label can turn up as the |
5190 | first one in the loop. Aligning a braf label causes incorrect switch | |
5191 | destination addresses; we can detect braf labels because they are | |
5192 | followed by a BARRIER. | |
5193 | Applying loop alignment to small constant or switch tables is a waste | |
5194 | of space, so we suppress this too. */ | |
5195 | int | |
cf277499 | 5196 | sh_loop_align (rtx label) |
2148624a R |
5197 | { |
5198 | rtx next = label; | |
5199 | ||
5200 | do | |
5201 | next = next_nonnote_insn (next); | |
f3536097 | 5202 | while (next && LABEL_P (next)); |
2148624a R |
5203 | |
5204 | if (! next | |
2c3c49de | 5205 | || ! INSN_P (next) |
2148624a R |
5206 | || GET_CODE (PATTERN (next)) == ADDR_DIFF_VEC |
5207 | || recog_memoized (next) == CODE_FOR_consttable_2) | |
5208 | return 0; | |
fa5322fa | 5209 | |
10f4f635 | 5210 | return align_loops_log; |
2148624a R |
5211 | } |
5212 | ||
18dbd950 | 5213 | /* Do a final pass over the function, just before delayed branch |
933c3ba3 | 5214 | scheduling. */ |
b9654711 | 5215 | |
18dbd950 | 5216 | static void |
cf277499 | 5217 | sh_reorg (void) |
0d7e008e | 5218 | { |
18dbd950 | 5219 | rtx first, insn, mova = NULL_RTX; |
1245df60 | 5220 | int num_mova; |
c5c76735 JL |
5221 | rtx r0_rtx = gen_rtx_REG (Pmode, 0); |
5222 | rtx r0_inc_rtx = gen_rtx_POST_INC (Pmode, r0_rtx); | |
8aa2a305 | 5223 | |
18dbd950 | 5224 | first = get_insns (); |
4241ecb0 | 5225 | max_labelno_before_reorg = max_label_num (); |
18dbd950 | 5226 | |
cb51ecd2 AO |
5227 | /* We must split call insns before introducing `mova's. If we're |
5228 | optimizing, they'll have already been split. Otherwise, make | |
5229 | sure we don't split them too late. */ | |
5230 | if (! optimize) | |
6f862f2f | 5231 | split_all_insns_noflow (); |
cb51ecd2 | 5232 | |
fa5322fa AO |
5233 | if (TARGET_SHMEDIA) |
5234 | return; | |
5235 | ||
933c3ba3 JW |
5236 | /* If relaxing, generate pseudo-ops to associate function calls with |
5237 | the symbols they call. It does no harm to not generate these | |
5238 | pseudo-ops. However, when we can generate them, it enables to | |
5239 | linker to potentially relax the jsr to a bsr, and eliminate the | |
5240 | register load and, possibly, the constant pool entry. */ | |
5241 | ||
1245df60 | 5242 | mdep_reorg_phase = SH_INSERT_USES_LABELS; |
933c3ba3 JW |
5243 | if (TARGET_RELAX) |
5244 | { | |
cf7c4aa6 HPN |
5245 | /* Remove all REG_LABEL_OPERAND notes. We want to use them for our |
5246 | own purposes. This works because none of the remaining passes | |
933c3ba3 JW |
5247 | need to look at them. |
5248 | ||
5249 | ??? But it may break in the future. We should use a machine | |
5250 | dependent REG_NOTE, or some other approach entirely. */ | |
5251 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
5252 | { | |
2c3c49de | 5253 | if (INSN_P (insn)) |
933c3ba3 JW |
5254 | { |
5255 | rtx note; | |
5256 | ||
cf7c4aa6 HPN |
5257 | while ((note = find_reg_note (insn, REG_LABEL_OPERAND, |
5258 | NULL_RTX)) != 0) | |
933c3ba3 JW |
5259 | remove_note (insn, note); |
5260 | } | |
5261 | } | |
5262 | ||
5263 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
5264 | { | |
5265 | rtx pattern, reg, link, set, scan, dies, label; | |
5266 | int rescan = 0, foundinsn = 0; | |
5267 | ||
f3536097 | 5268 | if (CALL_P (insn)) |
4787bce0 JW |
5269 | { |
5270 | pattern = PATTERN (insn); | |
933c3ba3 | 5271 | |
4787bce0 JW |
5272 | if (GET_CODE (pattern) == PARALLEL) |
5273 | pattern = XVECEXP (pattern, 0, 0); | |
5274 | if (GET_CODE (pattern) == SET) | |
5275 | pattern = SET_SRC (pattern); | |
933c3ba3 | 5276 | |
4787bce0 | 5277 | if (GET_CODE (pattern) != CALL |
f3536097 | 5278 | || !MEM_P (XEXP (pattern, 0))) |
4787bce0 | 5279 | continue; |
933c3ba3 | 5280 | |
4787bce0 JW |
5281 | reg = XEXP (XEXP (pattern, 0), 0); |
5282 | } | |
5283 | else | |
5284 | { | |
5285 | reg = sfunc_uses_reg (insn); | |
5286 | if (! reg) | |
5287 | continue; | |
5288 | } | |
933c3ba3 | 5289 | |
f3536097 | 5290 | if (!REG_P (reg)) |
933c3ba3 JW |
5291 | continue; |
5292 | ||
6fb5fa3c DB |
5293 | /* Try scanning backward to find where the register is set. */ |
5294 | link = NULL; | |
5295 | for (scan = PREV_INSN (insn); | |
f3536097 | 5296 | scan && !LABEL_P (scan); |
6fb5fa3c | 5297 | scan = PREV_INSN (scan)) |
933c3ba3 | 5298 | { |
6fb5fa3c DB |
5299 | if (! INSN_P (scan)) |
5300 | continue; | |
933c3ba3 | 5301 | |
6fb5fa3c DB |
5302 | if (! reg_mentioned_p (reg, scan)) |
5303 | continue; | |
933c3ba3 | 5304 | |
6fb5fa3c DB |
5305 | if (noncall_uses_reg (reg, scan, &set)) |
5306 | break; | |
933c3ba3 | 5307 | |
6fb5fa3c DB |
5308 | if (set) |
5309 | { | |
5310 | link = scan; | |
5311 | break; | |
933c3ba3 JW |
5312 | } |
5313 | } | |
5314 | ||
5315 | if (! link) | |
5316 | continue; | |
5317 | ||
5318 | /* The register is set at LINK. */ | |
5319 | ||
5320 | /* We can only optimize the function call if the register is | |
5321 | being set to a symbol. In theory, we could sometimes | |
5322 | optimize calls to a constant location, but the assembler | |
5323 | and linker do not support that at present. */ | |
5324 | if (GET_CODE (SET_SRC (set)) != SYMBOL_REF | |
5325 | && GET_CODE (SET_SRC (set)) != LABEL_REF) | |
5326 | continue; | |
5327 | ||
5328 | /* Scan forward from LINK to the place where REG dies, and | |
5329 | make sure that the only insns which use REG are | |
5330 | themselves function calls. */ | |
5331 | ||
c1e8e6bd JW |
5332 | /* ??? This doesn't work for call targets that were allocated |
5333 | by reload, since there may not be a REG_DEAD note for the | |
5334 | register. */ | |
5335 | ||
933c3ba3 JW |
5336 | dies = NULL_RTX; |
5337 | for (scan = NEXT_INSN (link); scan; scan = NEXT_INSN (scan)) | |
5338 | { | |
5339 | rtx scanset; | |
5340 | ||
c1e8e6bd JW |
5341 | /* Don't try to trace forward past a CODE_LABEL if we haven't |
5342 | seen INSN yet. Ordinarily, we will only find the setting insn | |
6fb5fa3c | 5343 | if it is in the same basic block. However, |
c1e8e6bd JW |
5344 | cross-jumping can insert code labels in between the load and |
5345 | the call, and can result in situations where a single call | |
5346 | insn may have two targets depending on where we came from. */ | |
5347 | ||
f3536097 | 5348 | if (LABEL_P (scan) && ! foundinsn) |
c1e8e6bd JW |
5349 | break; |
5350 | ||
2c3c49de | 5351 | if (! INSN_P (scan)) |
933c3ba3 JW |
5352 | continue; |
5353 | ||
5354 | /* Don't try to trace forward past a JUMP. To optimize | |
5355 | safely, we would have to check that all the | |
c1e8e6bd | 5356 | instructions at the jump destination did not use REG. */ |
933c3ba3 | 5357 | |
f3536097 | 5358 | if (JUMP_P (scan)) |
933c3ba3 JW |
5359 | break; |
5360 | ||
5361 | if (! reg_mentioned_p (reg, scan)) | |
5362 | continue; | |
5363 | ||
5364 | if (noncall_uses_reg (reg, scan, &scanset)) | |
5365 | break; | |
5366 | ||
5367 | if (scan == insn) | |
5368 | foundinsn = 1; | |
5369 | ||
4787bce0 | 5370 | if (scan != insn |
f3536097 | 5371 | && (CALL_P (scan) || sfunc_uses_reg (scan))) |
933c3ba3 JW |
5372 | { |
5373 | /* There is a function call to this register other | |
5374 | than the one we are checking. If we optimize | |
5375 | this call, we need to rescan again below. */ | |
5376 | rescan = 1; | |
5377 | } | |
5378 | ||
5379 | /* ??? We shouldn't have to worry about SCANSET here. | |
5380 | We should just be able to check for a REG_DEAD note | |
5381 | on a function call. However, the REG_DEAD notes are | |
5382 | apparently not dependable around libcalls; c-torture | |
5383 | execute/920501-2 is a test case. If SCANSET is set, | |
5384 | then this insn sets the register, so it must have | |
5385 | died earlier. Unfortunately, this will only handle | |
5386 | the cases in which the register is, in fact, set in a | |
5387 | later insn. */ | |
5388 | ||
5389 | /* ??? We shouldn't have to use FOUNDINSN here. | |
6fb5fa3c DB |
5390 | This dates back to when we used LOG_LINKS to find |
5391 | the most recent insn which sets the register. */ | |
933c3ba3 JW |
5392 | |
5393 | if (foundinsn | |
5394 | && (scanset | |
5395 | || find_reg_note (scan, REG_DEAD, reg))) | |
5396 | { | |
5397 | dies = scan; | |
5398 | break; | |
5399 | } | |
5400 | } | |
5401 | ||
5402 | if (! dies) | |
5403 | { | |
5404 | /* Either there was a branch, or some insn used REG | |
5405 | other than as a function call address. */ | |
5406 | continue; | |
5407 | } | |
5408 | ||
cf7c4aa6 HPN |
5409 | /* Create a code label, and put it in a REG_LABEL_OPERAND note |
5410 | on the insn which sets the register, and on each call insn | |
5411 | which uses the register. In final_prescan_insn we look for | |
5412 | the REG_LABEL_OPERAND notes, and output the appropriate label | |
933c3ba3 JW |
5413 | or pseudo-op. */ |
5414 | ||
5415 | label = gen_label_rtx (); | |
65c5f2a6 ILT |
5416 | add_reg_note (link, REG_LABEL_OPERAND, label); |
5417 | add_reg_note (insn, REG_LABEL_OPERAND, label); | |
933c3ba3 JW |
5418 | if (rescan) |
5419 | { | |
5420 | scan = link; | |
5421 | do | |
5422 | { | |
4787bce0 JW |
5423 | rtx reg2; |
5424 | ||
933c3ba3 JW |
5425 | scan = NEXT_INSN (scan); |
5426 | if (scan != insn | |
f3536097 | 5427 | && ((CALL_P (scan) |
4787bce0 JW |
5428 | && reg_mentioned_p (reg, scan)) |
5429 | || ((reg2 = sfunc_uses_reg (scan)) | |
5430 | && REGNO (reg2) == REGNO (reg)))) | |
65c5f2a6 | 5431 | add_reg_note (scan, REG_LABEL_OPERAND, label); |
933c3ba3 JW |
5432 | } |
5433 | while (scan != dies); | |
5434 | } | |
5435 | } | |
5436 | } | |
5437 | ||
33f7f353 JR |
5438 | if (TARGET_SH2) |
5439 | fixup_addr_diff_vecs (first); | |
1245df60 R |
5440 | |
5441 | if (optimize) | |
5442 | { | |
1245df60 R |
5443 | mdep_reorg_phase = SH_SHORTEN_BRANCHES0; |
5444 | shorten_branches (first); | |
5445 | } | |
6dd3c0a5 | 5446 | |
933c3ba3 JW |
5447 | /* Scan the function looking for move instructions which have to be |
5448 | changed to pc-relative loads and insert the literal tables. */ | |
6dd3c0a5 SB |
5449 | label_ref_list_pool = create_alloc_pool ("label references list", |
5450 | sizeof (struct label_ref_list_d), | |
5451 | 30); | |
1245df60 R |
5452 | mdep_reorg_phase = SH_FIXUP_PCLOAD; |
5453 | for (insn = first, num_mova = 0; insn; insn = NEXT_INSN (insn)) | |
0d7e008e | 5454 | { |
1245df60 R |
5455 | if (mova_p (insn)) |
5456 | { | |
078c8b08 R |
5457 | /* ??? basic block reordering can move a switch table dispatch |
5458 | below the switch table. Check if that has happened. | |
5459 | We only have the addresses available when optimizing; but then, | |
5460 | this check shouldn't be needed when not optimizing. */ | |
4241ecb0 | 5461 | if (!untangle_mova (&num_mova, &mova, insn)) |
078c8b08 | 5462 | { |
4241ecb0 R |
5463 | insn = mova; |
5464 | num_mova = 0; | |
078c8b08 | 5465 | } |
1245df60 | 5466 | } |
f3536097 | 5467 | else if (JUMP_P (insn) |
1245df60 | 5468 | && GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC |
4241ecb0 R |
5469 | && num_mova |
5470 | /* ??? loop invariant motion can also move a mova out of a | |
5471 | loop. Since loop does this code motion anyway, maybe we | |
5472 | should wrap UNSPEC_MOVA into a CONST, so that reload can | |
5473 | move it back. */ | |
5474 | && ((num_mova > 1 | |
5475 | && GET_MODE (prev_nonnote_insn (insn)) == VOIDmode) | |
5476 | || (prev_nonnote_insn (insn) | |
5477 | == XEXP (MOVA_LABELREF (mova), 0)))) | |
1245df60 R |
5478 | { |
5479 | rtx scan; | |
5480 | int total; | |
5481 | ||
5482 | num_mova--; | |
5483 | ||
5484 | /* Some code might have been inserted between the mova and | |
5485 | its ADDR_DIFF_VEC. Check if the mova is still in range. */ | |
5486 | for (scan = mova, total = 0; scan != insn; scan = NEXT_INSN (scan)) | |
33f7f353 | 5487 | total += get_attr_length (scan); |
1245df60 R |
5488 | |
5489 | /* range of mova is 1020, add 4 because pc counts from address of | |
5490 | second instruction after this one, subtract 2 in case pc is 2 | |
5491 | byte aligned. Possible alignment needed for the ADDR_DIFF_VEC | |
956d6950 | 5492 | cancels out with alignment effects of the mova itself. */ |
1245df60 R |
5493 | if (total > 1022) |
5494 | { | |
5495 | /* Change the mova into a load, and restart scanning | |
5496 | there. broken_move will then return true for mova. */ | |
078c8b08 | 5497 | fixup_mova (mova); |
1245df60 R |
5498 | insn = mova; |
5499 | } | |
5500 | } | |
078c8b08 | 5501 | if (broken_move (insn) |
f3536097 | 5502 | || (NONJUMP_INSN_P (insn) |
078c8b08 | 5503 | && recog_memoized (insn) == CODE_FOR_casesi_worker_2)) |
0d7e008e | 5504 | { |
0d7e008e | 5505 | rtx scan; |
8aa2a305 JW |
5506 | /* Scan ahead looking for a barrier to stick the constant table |
5507 | behind. */ | |
1245df60 | 5508 | rtx barrier = find_barrier (num_mova, mova, insn); |
3a8699c7 | 5509 | rtx last_float_move = NULL_RTX, last_float = 0, *last_float_addr = NULL; |
078c8b08 | 5510 | int need_aligned_label = 0; |
b9654711 | 5511 | |
1245df60 R |
5512 | if (num_mova && ! mova_p (mova)) |
5513 | { | |
5514 | /* find_barrier had to change the first mova into a | |
5515 | pcload; thus, we have to start with this new pcload. */ | |
5516 | insn = mova; | |
5517 | num_mova = 0; | |
5518 | } | |
16bea517 | 5519 | /* Now find all the moves between the points and modify them. */ |
0d7e008e SC |
5520 | for (scan = insn; scan != barrier; scan = NEXT_INSN (scan)) |
5521 | { | |
f3536097 | 5522 | if (LABEL_P (scan)) |
1245df60 | 5523 | last_float = 0; |
f3536097 | 5524 | if (NONJUMP_INSN_P (scan) |
078c8b08 R |
5525 | && recog_memoized (scan) == CODE_FOR_casesi_worker_2) |
5526 | need_aligned_label = 1; | |
0d7e008e SC |
5527 | if (broken_move (scan)) |
5528 | { | |
e577c183 JR |
5529 | rtx *patp = &PATTERN (scan), pat = *patp; |
5530 | rtx src, dst; | |
0d7e008e | 5531 | rtx lab; |
0d7e008e | 5532 | rtx newsrc; |
e577c183 JR |
5533 | enum machine_mode mode; |
5534 | ||
5535 | if (GET_CODE (pat) == PARALLEL) | |
5536 | patp = &XVECEXP (pat, 0, 0), pat = *patp; | |
5537 | src = SET_SRC (pat); | |
5538 | dst = SET_DEST (pat); | |
5539 | mode = GET_MODE (dst); | |
0d7e008e | 5540 | |
225e4f43 R |
5541 | if (mode == SImode && hi_const (src) |
5542 | && REGNO (dst) != FPUL_REG) | |
0d7e008e | 5543 | { |
23ed92ca | 5544 | int offset = 0; |
8aa2a305 | 5545 | |
0d7e008e | 5546 | mode = HImode; |
d3ae8277 | 5547 | while (GET_CODE (dst) == SUBREG) |
23ed92ca | 5548 | { |
ddef6bc7 JJ |
5549 | offset += subreg_regno_offset (REGNO (SUBREG_REG (dst)), |
5550 | GET_MODE (SUBREG_REG (dst)), | |
5551 | SUBREG_BYTE (dst), | |
5552 | GET_MODE (dst)); | |
23ed92ca JW |
5553 | dst = SUBREG_REG (dst); |
5554 | } | |
c5c76735 | 5555 | dst = gen_rtx_REG (HImode, REGNO (dst) + offset); |
0d7e008e | 5556 | } |
f3536097 | 5557 | if (REG_P (dst) && FP_ANY_REGISTER_P (REGNO (dst))) |
1245df60 | 5558 | { |
b8794099 | 5559 | /* This must be an insn that clobbers r0. */ |
0f2b41b2 AO |
5560 | rtx *clobberp = &XVECEXP (PATTERN (scan), 0, |
5561 | XVECLEN (PATTERN (scan), 0) | |
5562 | - 1); | |
5563 | rtx clobber = *clobberp; | |
b8794099 | 5564 | |
f5b9e7c9 NS |
5565 | gcc_assert (GET_CODE (clobber) == CLOBBER |
5566 | && rtx_equal_p (XEXP (clobber, 0), r0_rtx)); | |
b8794099 | 5567 | |
1245df60 R |
5568 | if (last_float |
5569 | && reg_set_between_p (r0_rtx, last_float_move, scan)) | |
5570 | last_float = 0; | |
3503150c RS |
5571 | if (last_float |
5572 | && TARGET_SHCOMPACT | |
5573 | && GET_MODE_SIZE (mode) != 4 | |
5574 | && GET_MODE_SIZE (GET_MODE (last_float)) == 4) | |
5575 | last_float = 0; | |
1245df60 R |
5576 | lab = add_constant (src, mode, last_float); |
5577 | if (lab) | |
5578 | emit_insn_before (gen_mova (lab), scan); | |
5579 | else | |
b8794099 R |
5580 | { |
5581 | /* There will be a REG_UNUSED note for r0 on | |
5582 | LAST_FLOAT_MOVE; we have to change it to REG_INC, | |
5583 | lest reorg:mark_target_live_regs will not | |
5584 | consider r0 to be used, and we end up with delay | |
475f26dd | 5585 | slot insn in front of SCAN that clobbers r0. */ |
b8794099 R |
5586 | rtx note |
5587 | = find_regno_note (last_float_move, REG_UNUSED, 0); | |
5588 | ||
ab1672a3 JW |
5589 | /* If we are not optimizing, then there may not be |
5590 | a note. */ | |
5591 | if (note) | |
81f40b79 | 5592 | PUT_REG_NOTE_KIND (note, REG_INC); |
b8794099 R |
5593 | |
5594 | *last_float_addr = r0_inc_rtx; | |
5595 | } | |
1245df60 R |
5596 | last_float_move = scan; |
5597 | last_float = src; | |
57d38024 | 5598 | newsrc = gen_const_mem (mode, |
318881c0 | 5599 | (((TARGET_SH4 && ! TARGET_FMOVD) |
225e4f43 | 5600 | || REGNO (dst) == FPUL_REG) |
1245df60 R |
5601 | ? r0_inc_rtx |
5602 | : r0_rtx)); | |
5603 | last_float_addr = &XEXP (newsrc, 0); | |
b8794099 R |
5604 | |
5605 | /* Remove the clobber of r0. */ | |
0f2b41b2 AO |
5606 | *clobberp = gen_rtx_CLOBBER (GET_MODE (clobber), |
5607 | gen_rtx_SCRATCH (Pmode)); | |
1245df60 | 5608 | } |
1a66cd67 | 5609 | /* This is a mova needing a label. Create it. */ |
43c05634 AO |
5610 | else if (GET_CODE (src) == UNSPEC |
5611 | && XINT (src, 1) == UNSPEC_MOVA | |
5612 | && GET_CODE (XVECEXP (src, 0, 0)) == CONST) | |
1a66cd67 | 5613 | { |
43c05634 | 5614 | lab = add_constant (XVECEXP (src, 0, 0), mode, 0); |
1a66cd67 | 5615 | newsrc = gen_rtx_LABEL_REF (VOIDmode, lab); |
e69d1422 | 5616 | newsrc = gen_rtx_UNSPEC (SImode, |
4773afa4 AO |
5617 | gen_rtvec (1, newsrc), |
5618 | UNSPEC_MOVA); | |
1a66cd67 | 5619 | } |
1245df60 R |
5620 | else |
5621 | { | |
5622 | lab = add_constant (src, mode, 0); | |
542a8afa RH |
5623 | newsrc = gen_rtx_LABEL_REF (VOIDmode, lab); |
5624 | newsrc = gen_const_mem (mode, newsrc); | |
1245df60 | 5625 | } |
c5c76735 | 5626 | *patp = gen_rtx_SET (VOIDmode, dst, newsrc); |
e577c183 | 5627 | INSN_CODE (scan) = -1; |
0d7e008e SC |
5628 | } |
5629 | } | |
078c8b08 | 5630 | dump_table (need_aligned_label ? insn : 0, barrier); |
1245df60 | 5631 | insn = barrier; |
0d7e008e SC |
5632 | } |
5633 | } | |
6dd3c0a5 | 5634 | free_alloc_pool (label_ref_list_pool); |
4241ecb0 R |
5635 | for (insn = first; insn; insn = NEXT_INSN (insn)) |
5636 | PUT_MODE (insn, VOIDmode); | |
5637 | ||
1245df60 | 5638 | mdep_reorg_phase = SH_SHORTEN_BRANCHES1; |
9d98a694 | 5639 | INSN_ADDRESSES_FREE (); |
1245df60 R |
5640 | split_branches (first); |
5641 | ||
5642 | /* The INSN_REFERENCES_ARE_DELAYED in sh.h is problematic because it | |
a920aefe | 5643 | also has an effect on the register that holds the address of the sfunc. |
1245df60 R |
5644 | Insert an extra dummy insn in front of each sfunc that pretends to |
5645 | use this register. */ | |
5646 | if (flag_delayed_branch) | |
5647 | { | |
5648 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
5649 | { | |
5650 | rtx reg = sfunc_uses_reg (insn); | |
5651 | ||
5652 | if (! reg) | |
5653 | continue; | |
5654 | emit_insn_before (gen_use_sfunc_addr (reg), insn); | |
5655 | } | |
5656 | } | |
225e4f43 R |
5657 | #if 0 |
5658 | /* fpscr is not actually a user variable, but we pretend it is for the | |
5659 | sake of the previous optimization passes, since we want it handled like | |
9f09b1f2 | 5660 | one. However, we don't have any debugging information for it, so turn |
225e4f43 R |
5661 | it into a non-user variable now. */ |
5662 | if (TARGET_SH4) | |
5663 | REG_USERVAR_P (get_fpscr_rtx ()) = 0; | |
5664 | #endif | |
1245df60 R |
5665 | mdep_reorg_phase = SH_AFTER_MDEP_REORG; |
5666 | } | |
5667 | ||
5668 | int | |
cf277499 | 5669 | get_dest_uid (rtx label, int max_uid) |
1245df60 R |
5670 | { |
5671 | rtx dest = next_real_insn (label); | |
5672 | int dest_uid; | |
5673 | if (! dest) | |
5674 | /* This can happen for an undefined label. */ | |
5675 | return 0; | |
5676 | dest_uid = INSN_UID (dest); | |
5677 | /* If this is a newly created branch redirection blocking instruction, | |
5678 | we cannot index the branch_uid or insn_addresses arrays with its | |
5679 | uid. But then, we won't need to, because the actual destination is | |
5680 | the following branch. */ | |
5681 | while (dest_uid >= max_uid) | |
5682 | { | |
5683 | dest = NEXT_INSN (dest); | |
5684 | dest_uid = INSN_UID (dest); | |
5685 | } | |
f3536097 | 5686 | if (JUMP_P (dest) && GET_CODE (PATTERN (dest)) == RETURN) |
1245df60 R |
5687 | return 0; |
5688 | return dest_uid; | |
5689 | } | |
5690 | ||
5691 | /* Split condbranches that are out of range. Also add clobbers for | |
5692 | scratch registers that are needed in far jumps. | |
5693 | We do this before delay slot scheduling, so that it can take our | |
5694 | newly created instructions into account. It also allows us to | |
5695 | find branches with common targets more easily. */ | |
5696 | ||
5697 | static void | |
cf277499 | 5698 | split_branches (rtx first) |
1245df60 R |
5699 | { |
5700 | rtx insn; | |
5701 | struct far_branch **uid_branch, *far_branch_list = 0; | |
5702 | int max_uid = get_max_uid (); | |
f5b9e7c9 | 5703 | int ok; |
1245df60 R |
5704 | |
5705 | /* Find out which branches are out of range. */ | |
1245df60 R |
5706 | shorten_branches (first); |
5707 | ||
5708 | uid_branch = (struct far_branch **) alloca (max_uid * sizeof *uid_branch); | |
2e09e75a | 5709 | memset ((char *) uid_branch, 0, max_uid * sizeof *uid_branch); |
1245df60 R |
5710 | |
5711 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
2c3c49de | 5712 | if (! INSN_P (insn)) |
1245df60 R |
5713 | continue; |
5714 | else if (INSN_DELETED_P (insn)) | |
5715 | { | |
5716 | /* Shorten_branches would split this instruction again, | |
5717 | so transform it into a note. */ | |
a38e7aa5 | 5718 | SET_INSN_DELETED (insn); |
1245df60 | 5719 | } |
f3536097 | 5720 | else if (JUMP_P (insn) |
1245df60 R |
5721 | /* Don't mess with ADDR_DIFF_VEC */ |
5722 | && (GET_CODE (PATTERN (insn)) == SET | |
5723 | || GET_CODE (PATTERN (insn)) == RETURN)) | |
5724 | { | |
5725 | enum attr_type type = get_attr_type (insn); | |
5726 | if (type == TYPE_CBRANCH) | |
5727 | { | |
5728 | rtx next, beyond; | |
f676971a | 5729 | |
1245df60 R |
5730 | if (get_attr_length (insn) > 4) |
5731 | { | |
5732 | rtx src = SET_SRC (PATTERN (insn)); | |
1245df60 | 5733 | rtx olabel = XEXP (XEXP (src, 1), 0); |
9d98a694 | 5734 | int addr = INSN_ADDRESSES (INSN_UID (insn)); |
1245df60 R |
5735 | rtx label = 0; |
5736 | int dest_uid = get_dest_uid (olabel, max_uid); | |
5737 | struct far_branch *bp = uid_branch[dest_uid]; | |
f676971a | 5738 | |
1245df60 | 5739 | /* redirect_jump needs a valid JUMP_LABEL, and it might delete |
25938114 | 5740 | the label if the LABEL_NUSES count drops to zero. There is |
1245df60 R |
5741 | always a jump_optimize pass that sets these values, but it |
5742 | proceeds to delete unreferenced code, and then if not | |
956d6950 | 5743 | optimizing, to un-delete the deleted instructions, thus |
1245df60 R |
5744 | leaving labels with too low uses counts. */ |
5745 | if (! optimize) | |
5746 | { | |
5747 | JUMP_LABEL (insn) = olabel; | |
5748 | LABEL_NUSES (olabel)++; | |
5749 | } | |
5750 | if (! bp) | |
5751 | { | |
5752 | bp = (struct far_branch *) alloca (sizeof *bp); | |
5753 | uid_branch[dest_uid] = bp; | |
5754 | bp->prev = far_branch_list; | |
5755 | far_branch_list = bp; | |
5756 | bp->far_label | |
5757 | = XEXP (XEXP (SET_SRC (PATTERN (insn)), 1), 0); | |
5758 | LABEL_NUSES (bp->far_label)++; | |
5759 | } | |
5760 | else | |
5761 | { | |
5762 | label = bp->near_label; | |
5763 | if (! label && bp->address - addr >= CONDJUMP_MIN) | |
5764 | { | |
5765 | rtx block = bp->insert_place; | |
5766 | ||
5767 | if (GET_CODE (PATTERN (block)) == RETURN) | |
5768 | block = PREV_INSN (block); | |
5769 | else | |
5770 | block = gen_block_redirect (block, | |
5771 | bp->address, 2); | |
5772 | label = emit_label_after (gen_label_rtx (), | |
5773 | PREV_INSN (block)); | |
5774 | bp->near_label = label; | |
5775 | } | |
5776 | else if (label && ! NEXT_INSN (label)) | |
eda44c78 R |
5777 | { |
5778 | if (addr + 2 - bp->address <= CONDJUMP_MAX) | |
5779 | bp->insert_place = insn; | |
5780 | else | |
5781 | gen_far_branch (bp); | |
5782 | } | |
1245df60 R |
5783 | } |
5784 | if (! label | |
318881c0 | 5785 | || (NEXT_INSN (label) && bp->address - addr < CONDJUMP_MIN)) |
1245df60 R |
5786 | { |
5787 | bp->near_label = label = gen_label_rtx (); | |
5788 | bp->insert_place = insn; | |
5789 | bp->address = addr; | |
5790 | } | |
2acdc1fb | 5791 | ok = redirect_jump (insn, label, 0); |
f5b9e7c9 | 5792 | gcc_assert (ok); |
1245df60 R |
5793 | } |
5794 | else | |
5795 | { | |
5796 | /* get_attr_length (insn) == 2 */ | |
5797 | /* Check if we have a pattern where reorg wants to redirect | |
5798 | the branch to a label from an unconditional branch that | |
5799 | is too far away. */ | |
5800 | /* We can't use JUMP_LABEL here because it might be undefined | |
5801 | when not optimizing. */ | |
33f7f353 | 5802 | /* A syntax error might cause beyond to be NULL_RTX. */ |
1245df60 R |
5803 | beyond |
5804 | = next_active_insn (XEXP (XEXP (SET_SRC (PATTERN (insn)), 1), | |
5805 | 0)); | |
f676971a | 5806 | |
33f7f353 | 5807 | if (beyond |
f3536097 | 5808 | && (JUMP_P (beyond) |
b94544e4 | 5809 | || ((beyond = next_active_insn (beyond)) |
f3536097 | 5810 | && JUMP_P (beyond))) |
1245df60 | 5811 | && GET_CODE (PATTERN (beyond)) == SET |
ddf8ec7e | 5812 | && recog_memoized (beyond) == CODE_FOR_jump_compact |
9d98a694 AO |
5813 | && ((INSN_ADDRESSES |
5814 | (INSN_UID (XEXP (SET_SRC (PATTERN (beyond)), 0))) | |
11f9ed1a | 5815 | - INSN_ADDRESSES (INSN_UID (insn)) + (unsigned) 252) |
1245df60 R |
5816 | > 252 + 258 + 2)) |
5817 | gen_block_redirect (beyond, | |
9d98a694 | 5818 | INSN_ADDRESSES (INSN_UID (beyond)), 1); |
1245df60 | 5819 | } |
f676971a | 5820 | |
1245df60 R |
5821 | next = next_active_insn (insn); |
5822 | ||
f3536097 | 5823 | if ((JUMP_P (next) |
03904fb3 | 5824 | || ((next = next_active_insn (next)) |
f3536097 | 5825 | && JUMP_P (next))) |
1245df60 | 5826 | && GET_CODE (PATTERN (next)) == SET |
ddf8ec7e | 5827 | && recog_memoized (next) == CODE_FOR_jump_compact |
9d98a694 AO |
5828 | && ((INSN_ADDRESSES |
5829 | (INSN_UID (XEXP (SET_SRC (PATTERN (next)), 0))) | |
11f9ed1a | 5830 | - INSN_ADDRESSES (INSN_UID (insn)) + (unsigned) 252) |
1245df60 | 5831 | > 252 + 258 + 2)) |
9d98a694 | 5832 | gen_block_redirect (next, INSN_ADDRESSES (INSN_UID (next)), 1); |
1245df60 R |
5833 | } |
5834 | else if (type == TYPE_JUMP || type == TYPE_RETURN) | |
5835 | { | |
9d98a694 | 5836 | int addr = INSN_ADDRESSES (INSN_UID (insn)); |
1245df60 R |
5837 | rtx far_label = 0; |
5838 | int dest_uid = 0; | |
5839 | struct far_branch *bp; | |
5840 | ||
5841 | if (type == TYPE_JUMP) | |
5842 | { | |
5843 | far_label = XEXP (SET_SRC (PATTERN (insn)), 0); | |
5844 | dest_uid = get_dest_uid (far_label, max_uid); | |
5845 | if (! dest_uid) | |
5846 | { | |
5847 | /* Parse errors can lead to labels outside | |
5848 | the insn stream. */ | |
5849 | if (! NEXT_INSN (far_label)) | |
5850 | continue; | |
5851 | ||
5852 | if (! optimize) | |
5853 | { | |
5854 | JUMP_LABEL (insn) = far_label; | |
5855 | LABEL_NUSES (far_label)++; | |
5856 | } | |
8ceaac3c | 5857 | redirect_jump (insn, NULL_RTX, 1); |
1245df60 R |
5858 | far_label = 0; |
5859 | } | |
5860 | } | |
5861 | bp = uid_branch[dest_uid]; | |
5862 | if (! bp) | |
5863 | { | |
5864 | bp = (struct far_branch *) alloca (sizeof *bp); | |
5865 | uid_branch[dest_uid] = bp; | |
5866 | bp->prev = far_branch_list; | |
5867 | far_branch_list = bp; | |
5868 | bp->near_label = 0; | |
5869 | bp->far_label = far_label; | |
5870 | if (far_label) | |
5871 | LABEL_NUSES (far_label)++; | |
5872 | } | |
5873 | else if (bp->near_label && ! NEXT_INSN (bp->near_label)) | |
5874 | if (addr - bp->address <= CONDJUMP_MAX) | |
5875 | emit_label_after (bp->near_label, PREV_INSN (insn)); | |
5876 | else | |
5877 | { | |
5878 | gen_far_branch (bp); | |
5879 | bp->near_label = 0; | |
5880 | } | |
5881 | else | |
5882 | bp->near_label = 0; | |
5883 | bp->address = addr; | |
5884 | bp->insert_place = insn; | |
5885 | if (! far_label) | |
5886 | emit_insn_before (gen_block_branch_redirect (const0_rtx), insn); | |
5887 | else | |
5888 | gen_block_redirect (insn, addr, bp->near_label ? 2 : 0); | |
5889 | } | |
5890 | } | |
5891 | /* Generate all pending far branches, | |
5892 | and free our references to the far labels. */ | |
5893 | while (far_branch_list) | |
5894 | { | |
5895 | if (far_branch_list->near_label | |
5896 | && ! NEXT_INSN (far_branch_list->near_label)) | |
5897 | gen_far_branch (far_branch_list); | |
5898 | if (optimize | |
5899 | && far_branch_list->far_label | |
5900 | && ! --LABEL_NUSES (far_branch_list->far_label)) | |
5901 | delete_insn (far_branch_list->far_label); | |
5902 | far_branch_list = far_branch_list->prev; | |
5903 | } | |
a0798779 R |
5904 | |
5905 | /* Instruction length information is no longer valid due to the new | |
5906 | instructions that have been generated. */ | |
5907 | init_insn_lengths (); | |
b9654711 SC |
5908 | } |
5909 | ||
8aa2a305 | 5910 | /* Dump out instruction addresses, which is useful for debugging the |
933c3ba3 JW |
5911 | constant pool table stuff. |
5912 | ||
5913 | If relaxing, output the label and pseudo-ops used to link together | |
5914 | calls and the instruction which set the registers. */ | |
8aa2a305 | 5915 | |
8aa2a305 JW |
5916 | /* ??? The addresses printed by this routine for insns are nonsense for |
5917 | insns which are inside of a sequence where none of the inner insns have | |
5918 | variable length. This is because the second pass of shorten_branches | |
5919 | does not bother to update them. */ | |
0d7e008e | 5920 | |
8e87e161 | 5921 | void |
cf277499 SB |
5922 | final_prescan_insn (rtx insn, rtx *opvec ATTRIBUTE_UNUSED, |
5923 | int noperands ATTRIBUTE_UNUSED) | |
b9654711 | 5924 | { |
8aa2a305 | 5925 | if (TARGET_DUMPISIZE) |
9d98a694 | 5926 | fprintf (asm_out_file, "\n! at %04x\n", INSN_ADDRESSES (INSN_UID (insn))); |
933c3ba3 JW |
5927 | |
5928 | if (TARGET_RELAX) | |
5929 | { | |
5930 | rtx note; | |
5931 | ||
cf7c4aa6 | 5932 | note = find_reg_note (insn, REG_LABEL_OPERAND, NULL_RTX); |
933c3ba3 JW |
5933 | if (note) |
5934 | { | |
5935 | rtx pattern; | |
5936 | ||
5937 | pattern = PATTERN (insn); | |
4b7d2f07 R |
5938 | if (GET_CODE (pattern) == PARALLEL) |
5939 | pattern = XVECEXP (pattern, 0, 0); | |
f5b9e7c9 NS |
5940 | switch (GET_CODE (pattern)) |
5941 | { | |
f5b9e7c9 NS |
5942 | case SET: |
5943 | if (GET_CODE (SET_SRC (pattern)) != CALL | |
5944 | && get_attr_type (insn) != TYPE_SFUNC) | |
5945 | { | |
5946 | targetm.asm_out.internal_label | |
5947 | (asm_out_file, "L", CODE_LABEL_NUMBER (XEXP (note, 0))); | |
5948 | break; | |
5949 | } | |
5950 | /* else FALLTHROUGH */ | |
5951 | case CALL: | |
5952 | asm_fprintf (asm_out_file, "\t.uses %LL%d\n", | |
5953 | CODE_LABEL_NUMBER (XEXP (note, 0))); | |
5954 | break; | |
5955 | ||
5956 | default: | |
5957 | gcc_unreachable (); | |
5958 | } | |
933c3ba3 JW |
5959 | } |
5960 | } | |
0d7e008e | 5961 | } |
b9654711 | 5962 | |
8aa2a305 | 5963 | /* Dump out any constants accumulated in the final pass. These will |
38e01259 | 5964 | only be labels. */ |
b9654711 | 5965 | |
318881c0 | 5966 | const char * |
cf277499 | 5967 | output_jump_label_table (void) |
0d7e008e | 5968 | { |
8aa2a305 JW |
5969 | int i; |
5970 | ||
5971 | if (pool_size) | |
5972 | { | |
5973 | fprintf (asm_out_file, "\t.align 2\n"); | |
5974 | for (i = 0; i < pool_size; i++) | |
5975 | { | |
5976 | pool_node *p = &pool_vector[i]; | |
5977 | ||
4977bab6 | 5978 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
8aa2a305 JW |
5979 | CODE_LABEL_NUMBER (p->label)); |
5980 | output_asm_insn (".long %O0", &p->value); | |
5981 | } | |
5982 | pool_size = 0; | |
5983 | } | |
b9654711 | 5984 | |
8aa2a305 JW |
5985 | return ""; |
5986 | } | |
5987 | \f | |
5988 | /* A full frame looks like: | |
16bea517 JW |
5989 | |
5990 | arg-5 | |
5991 | arg-4 | |
5992 | [ if current_function_anonymous_args | |
5993 | arg-3 | |
5994 | arg-2 | |
5995 | arg-1 | |
5996 | arg-0 ] | |
5997 | saved-fp | |
5998 | saved-r10 | |
5999 | saved-r11 | |
6000 | saved-r12 | |
6001 | saved-pr | |
6002 | local-n | |
6003 | .. | |
6004 | local-1 | |
6005 | local-0 <- fp points here. */ | |
8e87e161 | 6006 | |
8aa2a305 JW |
6007 | /* Number of bytes pushed for anonymous args, used to pass information |
6008 | between expand_prologue and expand_epilogue. */ | |
6009 | ||
fada1961 R |
6010 | /* Adjust the stack by SIZE bytes. REG holds the rtl of the register to be |
6011 | adjusted. If epilogue_p is zero, this is for a prologue; otherwise, it's | |
726d4cb7 KK |
6012 | for an epilogue and a negative value means that it's for a sibcall |
6013 | epilogue. If LIVE_REGS_MASK is nonzero, it points to a HARD_REG_SET of | |
6014 | all the registers that are about to be restored, and hence dead. */ | |
8aa2a305 JW |
6015 | |
6016 | static void | |
fada1961 R |
6017 | output_stack_adjust (int size, rtx reg, int epilogue_p, |
6018 | HARD_REG_SET *live_regs_mask) | |
8aa2a305 | 6019 | { |
fada1961 | 6020 | rtx (*emit_fn) (rtx) = epilogue_p ? &emit_insn : &frame_insn; |
8aa2a305 JW |
6021 | if (size) |
6022 | { | |
fa5322fa AO |
6023 | HOST_WIDE_INT align = STACK_BOUNDARY / BITS_PER_UNIT; |
6024 | ||
157371cf AO |
6025 | /* This test is bogus, as output_stack_adjust is used to re-align the |
6026 | stack. */ | |
6027 | #if 0 | |
f5b9e7c9 | 6028 | gcc_assert (!(size % align)); |
157371cf | 6029 | #endif |
fa5322fa AO |
6030 | |
6031 | if (CONST_OK_FOR_ADD (size)) | |
2754d3c5 | 6032 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size))); |
f3cd5375 JW |
6033 | /* Try to do it with two partial adjustments; however, we must make |
6034 | sure that the stack is properly aligned at all times, in case | |
6f317ef3 | 6035 | an interrupt occurs between the two partial adjustments. */ |
fa5322fa AO |
6036 | else if (CONST_OK_FOR_ADD (size / 2 & -align) |
6037 | && CONST_OK_FOR_ADD (size - (size / 2 & -align))) | |
f3cd5375 | 6038 | { |
2754d3c5 R |
6039 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size / 2 & -align))); |
6040 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size - (size / 2 & -align)))); | |
f3cd5375 JW |
6041 | } |
6042 | else | |
8aa2a305 | 6043 | { |
f3cd5375 | 6044 | rtx const_reg; |
2754d3c5 | 6045 | rtx insn; |
fada1961 R |
6046 | int temp = epilogue_p ? 7 : (TARGET_SH5 ? 0 : 1); |
6047 | int i; | |
f3cd5375 JW |
6048 | |
6049 | /* If TEMP is invalid, we could temporarily save a general | |
6050 | register to MACL. However, there is currently no need | |
f5b9e7c9 | 6051 | to handle this case, so just die when we see it. */ |
726d4cb7 KK |
6052 | if (epilogue_p < 0 |
6053 | || current_function_interrupt | |
62164eb4 | 6054 | || ! call_really_used_regs[temp] || fixed_regs[temp]) |
fada1961 | 6055 | temp = -1; |
726d4cb7 KK |
6056 | if (temp < 0 && ! current_function_interrupt |
6057 | && (TARGET_SHMEDIA || epilogue_p >= 0)) | |
fada1961 R |
6058 | { |
6059 | HARD_REG_SET temps; | |
6060 | COPY_HARD_REG_SET (temps, call_used_reg_set); | |
6061 | AND_COMPL_HARD_REG_SET (temps, call_fixed_reg_set); | |
726d4cb7 | 6062 | if (epilogue_p > 0) |
fada1961 | 6063 | { |
726d4cb7 | 6064 | int nreg = 0; |
38173d38 | 6065 | if (crtl->return_rtx) |
726d4cb7 KK |
6066 | { |
6067 | enum machine_mode mode; | |
38173d38 | 6068 | mode = GET_MODE (crtl->return_rtx); |
726d4cb7 KK |
6069 | if (BASE_RETURN_VALUE_REG (mode) == FIRST_RET_REG) |
6070 | nreg = HARD_REGNO_NREGS (FIRST_RET_REG, mode); | |
6071 | } | |
6072 | for (i = 0; i < nreg; i++) | |
fada1961 | 6073 | CLEAR_HARD_REG_BIT (temps, FIRST_RET_REG + i); |
e3b5732b | 6074 | if (crtl->calls_eh_return) |
fada1961 R |
6075 | { |
6076 | CLEAR_HARD_REG_BIT (temps, EH_RETURN_STACKADJ_REGNO); | |
6077 | for (i = 0; i <= 3; i++) | |
6078 | CLEAR_HARD_REG_BIT (temps, EH_RETURN_DATA_REGNO (i)); | |
6079 | } | |
6080 | } | |
726d4cb7 KK |
6081 | if (TARGET_SHMEDIA && epilogue_p < 0) |
6082 | for (i = FIRST_TARGET_REG; i <= LAST_TARGET_REG; i++) | |
6083 | CLEAR_HARD_REG_BIT (temps, i); | |
6084 | if (epilogue_p <= 0) | |
fada1961 R |
6085 | { |
6086 | for (i = FIRST_PARM_REG; | |
6087 | i < FIRST_PARM_REG + NPARM_REGS (SImode); i++) | |
6088 | CLEAR_HARD_REG_BIT (temps, i); | |
6de9cd9a | 6089 | if (cfun->static_chain_decl != NULL) |
fada1961 R |
6090 | CLEAR_HARD_REG_BIT (temps, STATIC_CHAIN_REGNUM); |
6091 | } | |
6092 | temp = scavenge_reg (&temps); | |
6093 | } | |
6094 | if (temp < 0 && live_regs_mask) | |
eb996a4a KK |
6095 | { |
6096 | HARD_REG_SET temps; | |
6097 | ||
6098 | COPY_HARD_REG_SET (temps, *live_regs_mask); | |
6099 | CLEAR_HARD_REG_BIT (temps, REGNO (reg)); | |
6100 | temp = scavenge_reg (&temps); | |
6101 | } | |
f3cd5375 | 6102 | if (temp < 0) |
726d4cb7 | 6103 | { |
f5b9e7c9 NS |
6104 | rtx adj_reg, tmp_reg, mem; |
6105 | ||
726d4cb7 KK |
6106 | /* If we reached here, the most likely case is the (sibcall) |
6107 | epilogue for non SHmedia. Put a special push/pop sequence | |
6108 | for such case as the last resort. This looks lengthy but | |
f5b9e7c9 NS |
6109 | would not be problem because it seems to be very |
6110 | rare. */ | |
6111 | ||
6112 | gcc_assert (!TARGET_SHMEDIA && epilogue_p); | |
6113 | ||
6114 | ||
6115 | /* ??? There is still the slight possibility that r4 or | |
6116 | r5 have been reserved as fixed registers or assigned | |
6117 | as global registers, and they change during an | |
6118 | interrupt. There are possible ways to handle this: | |
6119 | ||
6120 | - If we are adjusting the frame pointer (r14), we can do | |
6121 | with a single temp register and an ordinary push / pop | |
6122 | on the stack. | |
6123 | - Grab any call-used or call-saved registers (i.e. not | |
6124 | fixed or globals) for the temps we need. We might | |
6125 | also grab r14 if we are adjusting the stack pointer. | |
6126 | If we can't find enough available registers, issue | |
6127 | a diagnostic and die - the user must have reserved | |
6128 | way too many registers. | |
6129 | But since all this is rather unlikely to happen and | |
6130 | would require extra testing, we just die if r4 / r5 | |
6131 | are not available. */ | |
6132 | gcc_assert (!fixed_regs[4] && !fixed_regs[5] | |
6133 | && !global_regs[4] && !global_regs[5]); | |
6134 | ||
6135 | adj_reg = gen_rtx_REG (GET_MODE (reg), 4); | |
6136 | tmp_reg = gen_rtx_REG (GET_MODE (reg), 5); | |
57d38024 | 6137 | emit_move_insn (gen_tmp_stack_mem (Pmode, reg), adj_reg); |
f5b9e7c9 NS |
6138 | emit_insn (GEN_MOV (adj_reg, GEN_INT (size))); |
6139 | emit_insn (GEN_ADD3 (adj_reg, adj_reg, reg)); | |
57d38024 | 6140 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_PRE_DEC (Pmode, adj_reg)); |
f5b9e7c9 | 6141 | emit_move_insn (mem, tmp_reg); |
57d38024 R |
6142 | emit_move_insn (tmp_reg, gen_tmp_stack_mem (Pmode, reg)); |
6143 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_PRE_DEC (Pmode, adj_reg)); | |
6144 | emit_move_insn (mem, tmp_reg); | |
6145 | emit_move_insn (reg, adj_reg); | |
6146 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_POST_INC (Pmode, reg)); | |
6147 | emit_move_insn (adj_reg, mem); | |
6148 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_POST_INC (Pmode, reg)); | |
6149 | emit_move_insn (tmp_reg, mem); | |
eb996a4a | 6150 | /* Tell flow the insns that pop r4/r5 aren't dead. */ |
c41c1387 RS |
6151 | emit_use (tmp_reg); |
6152 | emit_use (adj_reg); | |
57d38024 | 6153 | return; |
726d4cb7 | 6154 | } |
fa5322fa | 6155 | const_reg = gen_rtx_REG (GET_MODE (reg), temp); |
f3cd5375 JW |
6156 | |
6157 | /* If SIZE is negative, subtract the positive value. | |
6158 | This sometimes allows a constant pool entry to be shared | |
6159 | between prologue and epilogue code. */ | |
6160 | if (size < 0) | |
885c9a39 | 6161 | { |
fa5322fa | 6162 | emit_insn (GEN_MOV (const_reg, GEN_INT (-size))); |
2754d3c5 | 6163 | insn = emit_fn (GEN_SUB3 (reg, reg, const_reg)); |
885c9a39 JR |
6164 | } |
6165 | else | |
6166 | { | |
fa5322fa | 6167 | emit_insn (GEN_MOV (const_reg, GEN_INT (size))); |
2754d3c5 | 6168 | insn = emit_fn (GEN_ADD3 (reg, reg, const_reg)); |
885c9a39 | 6169 | } |
fada1961 | 6170 | if (! epilogue_p) |
b757d36f KK |
6171 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, |
6172 | gen_rtx_SET (VOIDmode, reg, | |
6173 | gen_rtx_PLUS (SImode, reg, | |
6174 | GEN_INT (size)))); | |
8aa2a305 | 6175 | } |
8aa2a305 JW |
6176 | } |
6177 | } | |
6178 | ||
2754d3c5 | 6179 | static rtx |
cf277499 | 6180 | frame_insn (rtx x) |
2754d3c5 R |
6181 | { |
6182 | x = emit_insn (x); | |
6183 | RTX_FRAME_RELATED_P (x) = 1; | |
6184 | return x; | |
6185 | } | |
6186 | ||
8aa2a305 JW |
6187 | /* Output RTL to push register RN onto the stack. */ |
6188 | ||
2754d3c5 | 6189 | static rtx |
cf277499 | 6190 | push (int rn) |
8aa2a305 JW |
6191 | { |
6192 | rtx x; | |
225e4f43 R |
6193 | if (rn == FPUL_REG) |
6194 | x = gen_push_fpul (); | |
7144b2d8 DD |
6195 | else if (rn == FPSCR_REG) |
6196 | x = gen_push_fpscr (); | |
157371cf | 6197 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && ! TARGET_FPU_SINGLE |
104ee20b | 6198 | && FP_OR_XD_REGISTER_P (rn)) |
225e4f43 | 6199 | { |
104ee20b | 6200 | if (FP_REGISTER_P (rn) && (rn - FIRST_FP_REG) & 1) |
d042370e | 6201 | return NULL_RTX; |
c5c76735 | 6202 | x = gen_push_4 (gen_rtx_REG (DFmode, rn)); |
225e4f43 | 6203 | } |
3a8699c7 | 6204 | else if (TARGET_SH2E && FP_REGISTER_P (rn)) |
c5c76735 | 6205 | x = gen_push_e (gen_rtx_REG (SFmode, rn)); |
1a95a963 | 6206 | else |
c5c76735 | 6207 | x = gen_push (gen_rtx_REG (SImode, rn)); |
1a95a963 | 6208 | |
2754d3c5 | 6209 | x = frame_insn (x); |
b757d36f | 6210 | add_reg_note (x, REG_INC, gen_rtx_REG (SImode, STACK_POINTER_REGNUM)); |
2754d3c5 | 6211 | return x; |
8aa2a305 JW |
6212 | } |
6213 | ||
6214 | /* Output RTL to pop register RN from the stack. */ | |
6215 | ||
6216 | static void | |
cf277499 | 6217 | pop (int rn) |
8aa2a305 JW |
6218 | { |
6219 | rtx x; | |
225e4f43 R |
6220 | if (rn == FPUL_REG) |
6221 | x = gen_pop_fpul (); | |
7144b2d8 DD |
6222 | else if (rn == FPSCR_REG) |
6223 | x = gen_pop_fpscr (); | |
157371cf | 6224 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && ! TARGET_FPU_SINGLE |
104ee20b | 6225 | && FP_OR_XD_REGISTER_P (rn)) |
225e4f43 | 6226 | { |
104ee20b | 6227 | if (FP_REGISTER_P (rn) && (rn - FIRST_FP_REG) & 1) |
225e4f43 | 6228 | return; |
c5c76735 | 6229 | x = gen_pop_4 (gen_rtx_REG (DFmode, rn)); |
225e4f43 | 6230 | } |
3a8699c7 | 6231 | else if (TARGET_SH2E && FP_REGISTER_P (rn)) |
c5c76735 | 6232 | x = gen_pop_e (gen_rtx_REG (SFmode, rn)); |
1a95a963 | 6233 | else |
c5c76735 | 6234 | x = gen_pop (gen_rtx_REG (SImode, rn)); |
f676971a | 6235 | |
1245df60 | 6236 | x = emit_insn (x); |
b757d36f | 6237 | add_reg_note (x, REG_INC, gen_rtx_REG (SImode, STACK_POINTER_REGNUM)); |
8aa2a305 JW |
6238 | } |
6239 | ||
1245df60 | 6240 | /* Generate code to push the regs specified in the mask. */ |
8aa2a305 JW |
6241 | |
6242 | static void | |
cf277499 | 6243 | push_regs (HARD_REG_SET *mask, int interrupt_handler) |
8aa2a305 | 6244 | { |
5c7cafa8 | 6245 | int i = interrupt_handler ? LAST_BANKED_REG + 1 : 0; |
ac55736a | 6246 | int skip_fpscr = 0; |
8aa2a305 | 6247 | |
1245df60 R |
6248 | /* Push PR last; this gives better latencies after the prologue, and |
6249 | candidates for the return delay slot when there are no general | |
6250 | registers pushed. */ | |
5c7cafa8 | 6251 | for (; i < FIRST_PSEUDO_REGISTER; i++) |
ac55736a R |
6252 | { |
6253 | /* If this is an interrupt handler, and the SZ bit varies, | |
6254 | and we have to push any floating point register, we need | |
6255 | to switch to the correct precision first. */ | |
6256 | if (i == FIRST_FP_REG && interrupt_handler && TARGET_FMOVD | |
56b138ae | 6257 | && hard_reg_set_intersect_p (*mask, reg_class_contents[DF_REGS])) |
ac55736a R |
6258 | { |
6259 | HARD_REG_SET unsaved; | |
6260 | ||
6261 | push (FPSCR_REG); | |
832a3292 | 6262 | COMPL_HARD_REG_SET (unsaved, *mask); |
ac55736a R |
6263 | fpscr_set_from_mem (NORMAL_MODE (FP_MODE), unsaved); |
6264 | skip_fpscr = 1; | |
6265 | } | |
6266 | if (i != PR_REG | |
6267 | && (i != FPSCR_REG || ! skip_fpscr) | |
6268 | && TEST_HARD_REG_BIT (*mask, i)) | |
561642fa AP |
6269 | { |
6270 | /* If the ISR has RESBANK attribute assigned, don't push any of | |
6271 | the following registers - R0-R14, MACH, MACL and GBR. */ | |
6272 | if (! (sh_cfun_resbank_handler_p () | |
6273 | && ((i >= FIRST_GENERAL_REG && i < LAST_GENERAL_REG) | |
6274 | || i == MACH_REG | |
6275 | || i == MACL_REG | |
6276 | || i == GBR_REG))) | |
6277 | push (i); | |
6278 | } | |
ac55736a | 6279 | } |
5c7cafa8 CB |
6280 | |
6281 | /* Push banked registers last to improve delay slot opportunities. */ | |
6282 | if (interrupt_handler) | |
6283 | for (i = FIRST_BANKED_REG; i <= LAST_BANKED_REG; i++) | |
6284 | if (TEST_HARD_REG_BIT (*mask, i)) | |
6285 | push (i); | |
6286 | ||
561642fa AP |
6287 | /* Don't push PR register for an ISR with RESBANK attribute assigned. */ |
6288 | if (TEST_HARD_REG_BIT (*mask, PR_REG) && !sh_cfun_resbank_handler_p ()) | |
1245df60 | 6289 | push (PR_REG); |
8aa2a305 JW |
6290 | } |
6291 | ||
fe3ad572 SC |
6292 | /* Calculate how much extra space is needed to save all callee-saved |
6293 | target registers. | |
6294 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
6295 | ||
6296 | static int | |
6297 | shmedia_target_regs_stack_space (HARD_REG_SET *live_regs_mask) | |
6298 | { | |
6299 | int reg; | |
6300 | int stack_space = 0; | |
6301 | int interrupt_handler = sh_cfun_interrupt_handler_p (); | |
6302 | ||
6303 | for (reg = LAST_TARGET_REG; reg >= FIRST_TARGET_REG; reg--) | |
62164eb4 | 6304 | if ((! call_really_used_regs[reg] || interrupt_handler) |
fe3ad572 SC |
6305 | && ! TEST_HARD_REG_BIT (*live_regs_mask, reg)) |
6306 | /* Leave space to save this target register on the stack, | |
24746a42 | 6307 | in case target register allocation wants to use it. */ |
fe3ad572 SC |
6308 | stack_space += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); |
6309 | return stack_space; | |
6310 | } | |
f676971a | 6311 | |
fe3ad572 SC |
6312 | /* Decide whether we should reserve space for callee-save target registers, |
6313 | in case target register allocation wants to use them. REGS_SAVED is | |
6314 | the space, in bytes, that is already required for register saves. | |
6315 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
6316 | ||
6317 | static int | |
6318 | shmedia_reserve_space_for_target_registers_p (int regs_saved, | |
6319 | HARD_REG_SET *live_regs_mask) | |
6320 | { | |
6321 | if (optimize_size) | |
6322 | return 0; | |
6323 | return shmedia_target_regs_stack_space (live_regs_mask) <= regs_saved; | |
6324 | } | |
6325 | ||
6326 | /* Decide how much space to reserve for callee-save target registers | |
6327 | in case target register allocation wants to use them. | |
6328 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
6329 | ||
6330 | static int | |
6331 | shmedia_target_regs_stack_adjust (HARD_REG_SET *live_regs_mask) | |
6332 | { | |
6333 | if (shmedia_space_reserved_for_target_registers) | |
6334 | return shmedia_target_regs_stack_space (live_regs_mask); | |
6335 | else | |
6336 | return 0; | |
6337 | } | |
6338 | ||
8aa2a305 | 6339 | /* Work out the registers which need to be saved, both as a mask and a |
490dbac7 | 6340 | count of saved words. Return the count. |
8aa2a305 JW |
6341 | |
6342 | If doing a pragma interrupt function, then push all regs used by the | |
6343 | function, and if we call another function (we can tell by looking at PR), | |
6344 | make sure that all the regs it clobbers are safe too. */ | |
6345 | ||
490dbac7 | 6346 | static int |
cf277499 | 6347 | calc_live_regs (HARD_REG_SET *live_regs_mask) |
8aa2a305 | 6348 | { |
0ef2c525 | 6349 | unsigned int reg; |
1245df60 | 6350 | int count; |
a6ab9fc0 R |
6351 | tree attrs; |
6352 | bool interrupt_or_trapa_handler, trapa_handler, interrupt_handler; | |
6353 | bool nosave_low_regs; | |
fada1961 | 6354 | int pr_live, has_call; |
157131d7 | 6355 | |
a6ab9fc0 R |
6356 | attrs = DECL_ATTRIBUTES (current_function_decl); |
6357 | interrupt_or_trapa_handler = sh_cfun_interrupt_handler_p (); | |
6358 | trapa_handler = lookup_attribute ("trapa_handler", attrs) != NULL_TREE; | |
6359 | interrupt_handler = interrupt_or_trapa_handler && ! trapa_handler; | |
6360 | nosave_low_regs = lookup_attribute ("nosave_low_regs", attrs) != NULL_TREE; | |
8aa2a305 | 6361 | |
fada1961 | 6362 | CLEAR_HARD_REG_SET (*live_regs_mask); |
157371cf | 6363 | if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && interrupt_handler |
6fb5fa3c | 6364 | && df_regs_ever_live_p (FPSCR_REG)) |
c0fb94d7 | 6365 | target_flags &= ~MASK_FPU_SINGLE; |
225e4f43 | 6366 | /* If we can save a lot of saves by switching to double mode, do that. */ |
157371cf | 6367 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && TARGET_FPU_SINGLE) |
225e4f43 | 6368 | for (count = 0, reg = FIRST_FP_REG; reg <= LAST_FP_REG; reg += 2) |
6fb5fa3c | 6369 | if (df_regs_ever_live_p (reg) && df_regs_ever_live_p (reg+1) |
62164eb4 | 6370 | && (! call_really_used_regs[reg] |
a6ab9fc0 | 6371 | || interrupt_handler) |
225e4f43 R |
6372 | && ++count > 2) |
6373 | { | |
c0fb94d7 | 6374 | target_flags &= ~MASK_FPU_SINGLE; |
225e4f43 R |
6375 | break; |
6376 | } | |
9d7ed806 R |
6377 | /* PR_MEDIA_REG is a general purpose register, thus global_alloc already |
6378 | knows how to use it. That means the pseudo originally allocated for | |
6379 | the initial value can become the PR_MEDIA_REG hard register, as seen for | |
6380 | execute/20010122-1.c:test9. */ | |
6381 | if (TARGET_SHMEDIA) | |
7d73a2ba R |
6382 | /* ??? this function is called from initial_elimination_offset, hence we |
6383 | can't use the result of sh_media_register_for_return here. */ | |
97b3d0f0 | 6384 | pr_live = sh_pr_n_sets (); |
9d7ed806 R |
6385 | else |
6386 | { | |
6387 | rtx pr_initial = has_hard_reg_initial_val (Pmode, PR_REG); | |
6388 | pr_live = (pr_initial | |
f3536097 | 6389 | ? (!REG_P (pr_initial) |
29db2a1e | 6390 | || REGNO (pr_initial) != (PR_REG)) |
6fb5fa3c | 6391 | : df_regs_ever_live_p (PR_REG)); |
7d73a2ba R |
6392 | /* For Shcompact, if not optimizing, we end up with a memory reference |
6393 | using the return address pointer for __builtin_return_address even | |
6394 | though there is no actual need to put the PR register on the stack. */ | |
6fb5fa3c | 6395 | pr_live |= df_regs_ever_live_p (RETURN_ADDRESS_POINTER_REGNUM); |
9d7ed806 | 6396 | } |
fa5322fa AO |
6397 | /* Force PR to be live if the prologue has to call the SHmedia |
6398 | argument decoder or register saver. */ | |
6399 | if (TARGET_SHCOMPACT | |
38173d38 | 6400 | && ((crtl->args.info.call_cookie |
fa5322fa | 6401 | & ~ CALL_COOKIE_RET_TRAMP (1)) |
e3b5732b | 6402 | || crtl->saves_all_registers)) |
fa5322fa | 6403 | pr_live = 1; |
fada1961 | 6404 | has_call = TARGET_SHMEDIA ? ! leaf_function_p () : pr_live; |
0ef2c525 | 6405 | for (count = 0, reg = FIRST_PSEUDO_REGISTER; reg-- != 0; ) |
8aa2a305 | 6406 | { |
97b3d0f0 | 6407 | if (reg == (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG) |
3007d592 | 6408 | ? pr_live |
a6ab9fc0 | 6409 | : interrupt_handler |
1245df60 | 6410 | ? (/* Need to save all the regs ever live. */ |
6fb5fa3c | 6411 | (df_regs_ever_live_p (reg) |
62164eb4 R |
6412 | || (call_really_used_regs[reg] |
6413 | && (! fixed_regs[reg] || reg == MACH_REG || reg == MACL_REG | |
6414 | || reg == PIC_OFFSET_TABLE_REGNUM) | |
fada1961 | 6415 | && has_call) |
a6ab9fc0 R |
6416 | || (TARGET_SHMEDIA && has_call |
6417 | && REGISTER_NATURAL_MODE (reg) == SImode | |
fada1961 | 6418 | && (GENERAL_REGISTER_P (reg) || TARGET_REGISTER_P (reg)))) |
1245df60 R |
6419 | && reg != STACK_POINTER_REGNUM && reg != ARG_POINTER_REGNUM |
6420 | && reg != RETURN_ADDRESS_POINTER_REGNUM | |
7144b2d8 DD |
6421 | && reg != T_REG && reg != GBR_REG |
6422 | /* Push fpscr only on targets which have FPU */ | |
6423 | && (reg != FPSCR_REG || TARGET_FPU_ANY)) | |
1245df60 | 6424 | : (/* Only push those regs which are used and need to be saved. */ |
180bde4f RS |
6425 | (TARGET_SHCOMPACT |
6426 | && flag_pic | |
38173d38 | 6427 | && crtl->args.info.call_cookie |
0ef2c525 | 6428 | && reg == PIC_OFFSET_TABLE_REGNUM) |
6fb5fa3c | 6429 | || (df_regs_ever_live_p (reg) |
9fdcdc74 KK |
6430 | && ((!call_really_used_regs[reg] |
6431 | && !(reg != PIC_OFFSET_TABLE_REGNUM | |
6432 | && fixed_regs[reg] && call_used_regs[reg])) | |
a6ab9fc0 | 6433 | || (trapa_handler && reg == FPSCR_REG && TARGET_FPU_ANY))) |
e3b5732b | 6434 | || (crtl->calls_eh_return |
0ef2c525 KK |
6435 | && (reg == EH_RETURN_DATA_REGNO (0) |
6436 | || reg == EH_RETURN_DATA_REGNO (1) | |
6437 | || reg == EH_RETURN_DATA_REGNO (2) | |
6438 | || reg == EH_RETURN_DATA_REGNO (3))) | |
61f71b34 | 6439 | || ((reg == MACL_REG || reg == MACH_REG) |
6fb5fa3c | 6440 | && df_regs_ever_live_p (reg) |
61f71b34 DD |
6441 | && sh_cfun_attr_renesas_p ()) |
6442 | )) | |
8aa2a305 | 6443 | { |
490dbac7 | 6444 | SET_HARD_REG_BIT (*live_regs_mask, reg); |
fa5322fa AO |
6445 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); |
6446 | ||
157371cf | 6447 | if ((TARGET_SH4 || TARGET_SH2A_DOUBLE || TARGET_SH5) && TARGET_FMOVD |
fa5322fa | 6448 | && GET_MODE_CLASS (REGISTER_NATURAL_MODE (reg)) == MODE_FLOAT) |
eda44c78 | 6449 | { |
104ee20b | 6450 | if (FP_REGISTER_P (reg)) |
eda44c78 | 6451 | { |
6fb5fa3c | 6452 | if (! TARGET_FPU_SINGLE && ! df_regs_ever_live_p (reg ^ 1)) |
eda44c78 | 6453 | { |
490dbac7 | 6454 | SET_HARD_REG_BIT (*live_regs_mask, (reg ^ 1)); |
fa5322fa | 6455 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg ^ 1)); |
eda44c78 R |
6456 | } |
6457 | } | |
fa5322fa | 6458 | else if (XD_REGISTER_P (reg)) |
eda44c78 R |
6459 | { |
6460 | /* Must switch to double mode to access these registers. */ | |
c0fb94d7 | 6461 | target_flags &= ~MASK_FPU_SINGLE; |
eda44c78 R |
6462 | } |
6463 | } | |
8aa2a305 | 6464 | } |
a6ab9fc0 R |
6465 | if (nosave_low_regs && reg == R8_REG) |
6466 | break; | |
8aa2a305 | 6467 | } |
fe3ad572 SC |
6468 | /* If we have a target register optimization pass after prologue / epilogue |
6469 | threading, we need to assume all target registers will be live even if | |
6470 | they aren't now. */ | |
6471 | if (flag_branch_target_load_optimize2 | |
6472 | && TARGET_SAVE_ALL_TARGET_REGS | |
6473 | && shmedia_space_reserved_for_target_registers) | |
6474 | for (reg = LAST_TARGET_REG; reg >= FIRST_TARGET_REG; reg--) | |
62164eb4 | 6475 | if ((! call_really_used_regs[reg] || interrupt_handler) |
fe3ad572 SC |
6476 | && ! TEST_HARD_REG_BIT (*live_regs_mask, reg)) |
6477 | { | |
6478 | SET_HARD_REG_BIT (*live_regs_mask, reg); | |
6479 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); | |
6480 | } | |
fada1961 R |
6481 | /* If this is an interrupt handler, we don't have any call-clobbered |
6482 | registers we can conveniently use for target register save/restore. | |
6483 | Make sure we save at least one general purpose register when we need | |
6484 | to save target registers. */ | |
6485 | if (interrupt_handler | |
56b138ae RS |
6486 | && hard_reg_set_intersect_p (*live_regs_mask, |
6487 | reg_class_contents[TARGET_REGS]) | |
6488 | && ! hard_reg_set_intersect_p (*live_regs_mask, | |
6489 | reg_class_contents[GENERAL_REGS])) | |
fada1961 R |
6490 | { |
6491 | SET_HARD_REG_BIT (*live_regs_mask, R0_REG); | |
6492 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (R0_REG)); | |
6493 | } | |
8aa2a305 | 6494 | |
490dbac7 | 6495 | return count; |
8aa2a305 JW |
6496 | } |
6497 | ||
6498 | /* Code to generate prologue and epilogue sequences */ | |
b9654711 | 6499 | |
a93d1ba2 | 6500 | /* PUSHED is the number of bytes that are being pushed on the |
0ad913af R |
6501 | stack for register saves. Return the frame size, padded |
6502 | appropriately so that the stack stays properly aligned. */ | |
e5f8d4c0 | 6503 | static HOST_WIDE_INT |
cf277499 | 6504 | rounded_frame_size (int pushed) |
e5f8d4c0 R |
6505 | { |
6506 | HOST_WIDE_INT size = get_frame_size (); | |
6507 | HOST_WIDE_INT align = STACK_BOUNDARY / BITS_PER_UNIT; | |
6508 | ||
c5b9ef02 | 6509 | return ((size + pushed + align - 1) & -align) - pushed; |
e5f8d4c0 R |
6510 | } |
6511 | ||
fa5322fa AO |
6512 | /* Choose a call-clobbered target-branch register that remains |
6513 | unchanged along the whole function. We set it up as the return | |
6514 | value in the prologue. */ | |
6515 | int | |
cf277499 | 6516 | sh_media_register_for_return (void) |
fa5322fa AO |
6517 | { |
6518 | int regno; | |
6519 | int tr0_used; | |
6520 | ||
6521 | if (! current_function_is_leaf) | |
6522 | return -1; | |
fada1961 R |
6523 | if (lookup_attribute ("interrupt_handler", |
6524 | DECL_ATTRIBUTES (current_function_decl))) | |
6525 | return -1; | |
73a4d10b R |
6526 | if (sh_cfun_interrupt_handler_p ()) |
6527 | return -1; | |
fa5322fa | 6528 | |
6fb5fa3c | 6529 | tr0_used = flag_pic && df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM); |
fa5322fa AO |
6530 | |
6531 | for (regno = FIRST_TARGET_REG + tr0_used; regno <= LAST_TARGET_REG; regno++) | |
6fb5fa3c | 6532 | if (call_really_used_regs[regno] && ! df_regs_ever_live_p (regno)) |
fa5322fa AO |
6533 | return regno; |
6534 | ||
6535 | return -1; | |
6536 | } | |
6537 | ||
fada1961 R |
6538 | /* The maximum registers we need to save are: |
6539 | - 62 general purpose registers (r15 is stack pointer, r63 is zero) | |
6540 | - 32 floating point registers (for each pair, we save none, | |
6541 | one single precision value, or a double precision value). | |
6542 | - 8 target registers | |
6543 | - add 1 entry for a delimiter. */ | |
6544 | #define MAX_SAVED_REGS (62+32+8) | |
6545 | ||
6546 | typedef struct save_entry_s | |
6547 | { | |
6548 | unsigned char reg; | |
6549 | unsigned char mode; | |
6550 | short offset; | |
6551 | } save_entry; | |
6552 | ||
6553 | #define MAX_TEMPS 4 | |
6554 | ||
6555 | /* There will be a delimiter entry with VOIDmode both at the start and the | |
6556 | end of a filled in schedule. The end delimiter has the offset of the | |
6557 | save with the smallest (i.e. most negative) offset. */ | |
6558 | typedef struct save_schedule_s | |
6559 | { | |
6560 | save_entry entries[MAX_SAVED_REGS + 2]; | |
6561 | int temps[MAX_TEMPS+1]; | |
6562 | } save_schedule; | |
6563 | ||
6564 | /* Fill in SCHEDULE according to LIVE_REGS_MASK. If RESTORE is nonzero, | |
6565 | use reverse order. Returns the last entry written to (not counting | |
6566 | the delimiter). OFFSET_BASE is a number to be added to all offset | |
6567 | entries. */ | |
f676971a | 6568 | |
fada1961 R |
6569 | static save_entry * |
6570 | sh5_schedule_saves (HARD_REG_SET *live_regs_mask, save_schedule *schedule, | |
6571 | int offset_base) | |
6572 | { | |
6573 | int align, i; | |
6574 | save_entry *entry = schedule->entries; | |
6575 | int tmpx = 0; | |
6576 | int offset; | |
6577 | ||
6578 | if (! current_function_interrupt) | |
6579 | for (i = FIRST_GENERAL_REG; tmpx < MAX_TEMPS && i <= LAST_GENERAL_REG; i++) | |
62164eb4 | 6580 | if (call_really_used_regs[i] && ! fixed_regs[i] && i != PR_MEDIA_REG |
fada1961 R |
6581 | && ! FUNCTION_ARG_REGNO_P (i) |
6582 | && i != FIRST_RET_REG | |
6de9cd9a | 6583 | && ! (cfun->static_chain_decl != NULL && i == STATIC_CHAIN_REGNUM) |
e3b5732b | 6584 | && ! (crtl->calls_eh_return |
fada1961 | 6585 | && (i == EH_RETURN_STACKADJ_REGNO |
f94cef0a KK |
6586 | || ((unsigned) i >= EH_RETURN_DATA_REGNO (0) |
6587 | && (unsigned) i <= EH_RETURN_DATA_REGNO (3))))) | |
fada1961 R |
6588 | schedule->temps[tmpx++] = i; |
6589 | entry->reg = -1; | |
6590 | entry->mode = VOIDmode; | |
6591 | entry->offset = offset_base; | |
6592 | entry++; | |
6593 | /* We loop twice: first, we save 8-byte aligned registers in the | |
6594 | higher addresses, that are known to be aligned. Then, we | |
6595 | proceed to saving 32-bit registers that don't need 8-byte | |
6596 | alignment. | |
6597 | If this is an interrupt function, all registers that need saving | |
6598 | need to be saved in full. moreover, we need to postpone saving | |
6599 | target registers till we have saved some general purpose registers | |
6600 | we can then use as scratch registers. */ | |
6601 | offset = offset_base; | |
6602 | for (align = 1; align >= 0; align--) | |
6603 | { | |
6604 | for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; i--) | |
6605 | if (TEST_HARD_REG_BIT (*live_regs_mask, i)) | |
6606 | { | |
6607 | enum machine_mode mode = REGISTER_NATURAL_MODE (i); | |
6608 | int reg = i; | |
6609 | ||
6610 | if (current_function_interrupt) | |
6611 | { | |
6612 | if (TARGET_REGISTER_P (i)) | |
6613 | continue; | |
6614 | if (GENERAL_REGISTER_P (i)) | |
6615 | mode = DImode; | |
6616 | } | |
6617 | if (mode == SFmode && (i % 2) == 1 | |
6618 | && ! TARGET_FPU_SINGLE && FP_REGISTER_P (i) | |
6619 | && (TEST_HARD_REG_BIT (*live_regs_mask, (i ^ 1)))) | |
6620 | { | |
6621 | mode = DFmode; | |
6622 | i--; | |
6623 | reg--; | |
6624 | } | |
6625 | ||
6626 | /* If we're doing the aligned pass and this is not aligned, | |
6627 | or we're doing the unaligned pass and this is aligned, | |
6628 | skip it. */ | |
6629 | if ((GET_MODE_SIZE (mode) % (STACK_BOUNDARY / BITS_PER_UNIT) == 0) | |
6630 | != align) | |
6631 | continue; | |
6632 | ||
6633 | if (current_function_interrupt | |
6634 | && GENERAL_REGISTER_P (i) | |
6635 | && tmpx < MAX_TEMPS) | |
6636 | schedule->temps[tmpx++] = i; | |
6637 | ||
6638 | offset -= GET_MODE_SIZE (mode); | |
6639 | entry->reg = i; | |
6640 | entry->mode = mode; | |
6641 | entry->offset = offset; | |
6642 | entry++; | |
6643 | } | |
6644 | if (align && current_function_interrupt) | |
6645 | for (i = LAST_TARGET_REG; i >= FIRST_TARGET_REG; i--) | |
6646 | if (TEST_HARD_REG_BIT (*live_regs_mask, i)) | |
6647 | { | |
6648 | offset -= GET_MODE_SIZE (DImode); | |
6649 | entry->reg = i; | |
6650 | entry->mode = DImode; | |
6651 | entry->offset = offset; | |
6652 | entry++; | |
6653 | } | |
6654 | } | |
6655 | entry->reg = -1; | |
6656 | entry->mode = VOIDmode; | |
6657 | entry->offset = offset; | |
6658 | schedule->temps[tmpx] = -1; | |
6659 | return entry - 1; | |
6660 | } | |
6661 | ||
b9654711 | 6662 | void |
cf277499 | 6663 | sh_expand_prologue (void) |
b9654711 | 6664 | { |
490dbac7 | 6665 | HARD_REG_SET live_regs_mask; |
40d2032b | 6666 | int d, i; |
fa5322fa | 6667 | int d_rounding = 0; |
225e4f43 | 6668 | int save_flags = target_flags; |
49616835 | 6669 | int pretend_args; |
a6ab9fc0 R |
6670 | tree sp_switch_attr |
6671 | = lookup_attribute ("sp_switch", DECL_ATTRIBUTES (current_function_decl)); | |
b9654711 | 6672 | |
e3ba8d11 | 6673 | current_function_interrupt = sh_cfun_interrupt_handler_p (); |
552ecbd9 | 6674 | |
0d7e008e | 6675 | /* We have pretend args if we had an object sent partially in registers |
8aa2a305 | 6676 | and partially on the stack, e.g. a large structure. */ |
38173d38 | 6677 | pretend_args = crtl->args.pretend_args_size; |
49616835 JR |
6678 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl) |
6679 | && (NPARM_REGS(SImode) | |
38173d38 | 6680 | > crtl->args.info.arg_count[(int) SH_ARG_INT])) |
49616835 JR |
6681 | pretend_args = 0; |
6682 | output_stack_adjust (-pretend_args | |
38173d38 | 6683 | - crtl->args.info.stack_regs * 8, |
fada1961 | 6684 | stack_pointer_rtx, 0, NULL); |
b9654711 | 6685 | |
38173d38 | 6686 | if (TARGET_SHCOMPACT && flag_pic && crtl->args.info.call_cookie) |
2b1fd83f AO |
6687 | /* We're going to use the PIC register to load the address of the |
6688 | incoming-argument decoder and/or of the return trampoline from | |
6689 | the GOT, so make sure the PIC register is preserved and | |
6690 | initialized. */ | |
6fb5fa3c | 6691 | df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true); |
2b1fd83f | 6692 | |
fa5322fa | 6693 | if (TARGET_SHCOMPACT |
38173d38 | 6694 | && (crtl->args.info.call_cookie & ~ CALL_COOKIE_RET_TRAMP(1))) |
fa5322fa AO |
6695 | { |
6696 | int reg; | |
6697 | ||
6698 | /* First, make all registers with incoming arguments that will | |
6699 | be pushed onto the stack live, so that register renaming | |
6700 | doesn't overwrite them. */ | |
6701 | for (reg = 0; reg < NPARM_REGS (SImode); reg++) | |
38173d38 | 6702 | if (CALL_COOKIE_STACKSEQ_GET (crtl->args.info.call_cookie) |
fa5322fa AO |
6703 | >= NPARM_REGS (SImode) - reg) |
6704 | for (; reg < NPARM_REGS (SImode); reg++) | |
6705 | emit_insn (gen_shcompact_preserve_incoming_args | |
6706 | (gen_rtx_REG (SImode, FIRST_PARM_REG + reg))); | |
6707 | else if (CALL_COOKIE_INT_REG_GET | |
38173d38 | 6708 | (crtl->args.info.call_cookie, reg) == 1) |
fa5322fa AO |
6709 | emit_insn (gen_shcompact_preserve_incoming_args |
6710 | (gen_rtx_REG (SImode, FIRST_PARM_REG + reg))); | |
6711 | ||
6712 | emit_move_insn (gen_rtx_REG (Pmode, MACL_REG), | |
6713 | stack_pointer_rtx); | |
6714 | emit_move_insn (gen_rtx_REG (SImode, R0_REG), | |
38173d38 | 6715 | GEN_INT (crtl->args.info.call_cookie)); |
fa5322fa AO |
6716 | emit_move_insn (gen_rtx_REG (SImode, MACH_REG), |
6717 | gen_rtx_REG (SImode, R0_REG)); | |
6718 | } | |
6719 | else if (TARGET_SHMEDIA) | |
6720 | { | |
6721 | int tr = sh_media_register_for_return (); | |
6722 | ||
6723 | if (tr >= 0) | |
6fb5fa3c DB |
6724 | emit_move_insn (gen_rtx_REG (DImode, tr), |
6725 | gen_rtx_REG (DImode, PR_MEDIA_REG)); | |
fa5322fa AO |
6726 | } |
6727 | ||
7ab56274 | 6728 | /* Emit the code for SETUP_VARARGS. */ |
e3b5732b | 6729 | if (cfun->stdarg) |
b9654711 | 6730 | { |
49616835 | 6731 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl)) |
1245df60 | 6732 | { |
1a95a963 JW |
6733 | /* Push arg regs as if they'd been provided by caller in stack. */ |
6734 | for (i = 0; i < NPARM_REGS(SImode); i++) | |
6735 | { | |
6736 | int rn = NPARM_REGS(SImode) + FIRST_PARM_REG - i - 1; | |
2754d3c5 R |
6737 | rtx insn; |
6738 | ||
f676971a | 6739 | if (i >= (NPARM_REGS(SImode) |
38173d38 | 6740 | - crtl->args.info.arg_count[(int) SH_ARG_INT] |
aec373f1 | 6741 | )) |
1a95a963 | 6742 | break; |
2754d3c5 | 6743 | insn = push (rn); |
1a95a963 | 6744 | } |
1245df60 | 6745 | } |
b9654711 | 6746 | } |
1a95a963 | 6747 | |
4408efce | 6748 | /* If we're supposed to switch stacks at function entry, do so now. */ |
a6ab9fc0 R |
6749 | if (sp_switch_attr) |
6750 | { | |
6751 | /* The argument specifies a variable holding the address of the | |
6752 | stack the interrupt function should switch to/from at entry/exit. */ | |
6753 | const char *s | |
6754 | = ggc_strdup (TREE_STRING_POINTER (TREE_VALUE (sp_switch_attr))); | |
6755 | rtx sp_switch = gen_rtx_SYMBOL_REF (Pmode, s); | |
6756 | ||
6757 | emit_insn (gen_sp_switch_1 (sp_switch)); | |
6758 | } | |
4408efce | 6759 | |
490dbac7 | 6760 | d = calc_live_regs (&live_regs_mask); |
225e4f43 R |
6761 | /* ??? Maybe we could save some switching if we can move a mode switch |
6762 | that already happens to be at the function start into the prologue. */ | |
ac55736a | 6763 | if (target_flags != save_flags && ! current_function_interrupt) |
225e4f43 | 6764 | emit_insn (gen_toggle_sz ()); |
f676971a | 6765 | |
fa5322fa AO |
6766 | if (TARGET_SH5) |
6767 | { | |
fada1961 R |
6768 | int offset_base, offset; |
6769 | rtx r0 = NULL_RTX; | |
fa5322fa AO |
6770 | int offset_in_r0 = -1; |
6771 | int sp_in_r0 = 0; | |
fe3ad572 SC |
6772 | int tregs_space = shmedia_target_regs_stack_adjust (&live_regs_mask); |
6773 | int total_size, save_size; | |
fada1961 R |
6774 | save_schedule schedule; |
6775 | save_entry *entry; | |
6776 | int *tmp_pnt; | |
6777 | ||
62164eb4 | 6778 | if (call_really_used_regs[R0_REG] && ! fixed_regs[R0_REG] |
fada1961 R |
6779 | && ! current_function_interrupt) |
6780 | r0 = gen_rtx_REG (Pmode, R0_REG); | |
fe3ad572 SC |
6781 | |
6782 | /* D is the actual number of bytes that we need for saving registers, | |
6783 | however, in initial_elimination_offset we have committed to using | |
6784 | an additional TREGS_SPACE amount of bytes - in order to keep both | |
6785 | addresses to arguments supplied by the caller and local variables | |
6786 | valid, we must keep this gap. Place it between the incoming | |
6787 | arguments and the actually saved registers in a bid to optimize | |
6788 | locality of reference. */ | |
6789 | total_size = d + tregs_space; | |
6790 | total_size += rounded_frame_size (total_size); | |
6791 | save_size = total_size - rounded_frame_size (d); | |
6792 | if (save_size % (STACK_BOUNDARY / BITS_PER_UNIT)) | |
fa5322fa | 6793 | d_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) |
fe3ad572 SC |
6794 | - save_size % (STACK_BOUNDARY / BITS_PER_UNIT)); |
6795 | ||
6796 | /* If adjusting the stack in a single step costs nothing extra, do so. | |
6797 | I.e. either if a single addi is enough, or we need a movi anyway, | |
6798 | and we don't exceed the maximum offset range (the test for the | |
6799 | latter is conservative for simplicity). */ | |
6800 | if (TARGET_SHMEDIA | |
6801 | && (CONST_OK_FOR_I10 (-total_size) | |
6802 | || (! CONST_OK_FOR_I10 (-(save_size + d_rounding)) | |
6803 | && total_size <= 2044))) | |
6804 | d_rounding = total_size - save_size; | |
fa5322fa | 6805 | |
fada1961 | 6806 | offset_base = d + d_rounding; |
fe3ad572 SC |
6807 | |
6808 | output_stack_adjust (-(save_size + d_rounding), stack_pointer_rtx, | |
fada1961 | 6809 | 0, NULL); |
fa5322fa | 6810 | |
fada1961 R |
6811 | sh5_schedule_saves (&live_regs_mask, &schedule, offset_base); |
6812 | tmp_pnt = schedule.temps; | |
6813 | for (entry = &schedule.entries[1]; entry->mode != VOIDmode; entry++) | |
6814 | { | |
81f40b79 | 6815 | enum machine_mode mode = (enum machine_mode) entry->mode; |
0ef2c525 | 6816 | unsigned int reg = entry->reg; |
fada1961 | 6817 | rtx reg_rtx, mem_rtx, pre_dec = NULL_RTX; |
455786b6 | 6818 | rtx orig_reg_rtx; |
fa5322fa | 6819 | |
fada1961 | 6820 | offset = entry->offset; |
fa5322fa | 6821 | |
fada1961 | 6822 | reg_rtx = gen_rtx_REG (mode, reg); |
fa5322fa | 6823 | |
57d38024 R |
6824 | mem_rtx = gen_frame_mem (mode, |
6825 | gen_rtx_PLUS (Pmode, | |
6826 | stack_pointer_rtx, | |
6827 | GEN_INT (offset))); | |
fa5322fa | 6828 | |
6de7a512 PB |
6829 | if (!memory_address_p (mode, XEXP (mem_rtx, 0))) |
6830 | { | |
6831 | gcc_assert (r0); | |
6832 | mem_rtx = NULL_RTX; | |
6833 | } | |
fa5322fa | 6834 | |
6de7a512 PB |
6835 | if (HAVE_PRE_DECREMENT |
6836 | && (offset_in_r0 - offset == GET_MODE_SIZE (mode) | |
6837 | || mem_rtx == NULL_RTX | |
6838 | || reg == PR_REG || SPECIAL_REGISTER_P (reg))) | |
6839 | { | |
6840 | pre_dec = gen_frame_mem (mode, gen_rtx_PRE_DEC (Pmode, r0)); | |
fa5322fa | 6841 | |
6de7a512 | 6842 | if (!memory_address_p (mode, XEXP (pre_dec, 0))) |
fada1961 | 6843 | pre_dec = NULL_RTX; |
6de7a512 PB |
6844 | else |
6845 | { | |
6846 | mem_rtx = NULL_RTX; | |
6847 | offset += GET_MODE_SIZE (mode); | |
6848 | } | |
6849 | } | |
fa5322fa | 6850 | |
fada1961 R |
6851 | if (mem_rtx != NULL_RTX) |
6852 | goto addr_ok; | |
fa5322fa | 6853 | |
fada1961 R |
6854 | if (offset_in_r0 == -1) |
6855 | { | |
6856 | emit_move_insn (r0, GEN_INT (offset)); | |
6857 | offset_in_r0 = offset; | |
6858 | } | |
6859 | else if (offset != offset_in_r0) | |
6860 | { | |
6861 | emit_move_insn (r0, | |
6862 | gen_rtx_PLUS | |
6863 | (Pmode, r0, | |
6864 | GEN_INT (offset - offset_in_r0))); | |
6865 | offset_in_r0 += offset - offset_in_r0; | |
6866 | } | |
f676971a | 6867 | |
fada1961 R |
6868 | if (pre_dec != NULL_RTX) |
6869 | { | |
6870 | if (! sp_in_r0) | |
fa5322fa AO |
6871 | { |
6872 | emit_move_insn (r0, | |
6873 | gen_rtx_PLUS | |
fada1961 R |
6874 | (Pmode, r0, stack_pointer_rtx)); |
6875 | sp_in_r0 = 1; | |
fa5322fa | 6876 | } |
fa5322fa | 6877 | |
fada1961 R |
6878 | offset -= GET_MODE_SIZE (mode); |
6879 | offset_in_r0 -= GET_MODE_SIZE (mode); | |
fa5322fa | 6880 | |
fada1961 R |
6881 | mem_rtx = pre_dec; |
6882 | } | |
6883 | else if (sp_in_r0) | |
57d38024 | 6884 | mem_rtx = gen_frame_mem (mode, r0); |
fada1961 | 6885 | else |
57d38024 R |
6886 | mem_rtx = gen_frame_mem (mode, |
6887 | gen_rtx_PLUS (Pmode, | |
6888 | stack_pointer_rtx, | |
6889 | r0)); | |
fada1961 R |
6890 | |
6891 | /* We must not use an r0-based address for target-branch | |
6892 | registers or for special registers without pre-dec | |
6893 | memory addresses, since we store their values in r0 | |
6894 | first. */ | |
f5b9e7c9 NS |
6895 | gcc_assert (!TARGET_REGISTER_P (reg) |
6896 | && ((reg != PR_REG && !SPECIAL_REGISTER_P (reg)) | |
6897 | || mem_rtx == pre_dec)); | |
6898 | ||
fada1961 | 6899 | addr_ok: |
455786b6 | 6900 | orig_reg_rtx = reg_rtx; |
fada1961 R |
6901 | if (TARGET_REGISTER_P (reg) |
6902 | || ((reg == PR_REG || SPECIAL_REGISTER_P (reg)) | |
6903 | && mem_rtx != pre_dec)) | |
6904 | { | |
6905 | rtx tmp_reg = gen_rtx_REG (GET_MODE (reg_rtx), *tmp_pnt); | |
fa5322fa | 6906 | |
fada1961 | 6907 | emit_move_insn (tmp_reg, reg_rtx); |
fa5322fa | 6908 | |
fada1961 R |
6909 | if (REGNO (tmp_reg) == R0_REG) |
6910 | { | |
fa5322fa AO |
6911 | offset_in_r0 = -1; |
6912 | sp_in_r0 = 0; | |
f5b9e7c9 NS |
6913 | gcc_assert (!refers_to_regno_p |
6914 | (R0_REG, R0_REG+1, mem_rtx, (rtx *) 0)); | |
fa5322fa AO |
6915 | } |
6916 | ||
fada1961 R |
6917 | if (*++tmp_pnt <= 0) |
6918 | tmp_pnt = schedule.temps; | |
6919 | ||
6920 | reg_rtx = tmp_reg; | |
fa5322fa | 6921 | } |
fada1961 R |
6922 | { |
6923 | rtx insn; | |
6924 | ||
6925 | /* Mark as interesting for dwarf cfi generator */ | |
6926 | insn = emit_move_insn (mem_rtx, reg_rtx); | |
6927 | RTX_FRAME_RELATED_P (insn) = 1; | |
455786b6 KK |
6928 | /* If we use an intermediate register for the save, we can't |
6929 | describe this exactly in cfi as a copy of the to-be-saved | |
6930 | register into the temporary register and then the temporary | |
6931 | register on the stack, because the temporary register can | |
6932 | have a different natural size than the to-be-saved register. | |
6933 | Thus, we gloss over the intermediate copy and pretend we do | |
6934 | a direct save from the to-be-saved register. */ | |
6935 | if (REGNO (reg_rtx) != reg) | |
6936 | { | |
b757d36f | 6937 | rtx set; |
455786b6 KK |
6938 | |
6939 | set = gen_rtx_SET (VOIDmode, mem_rtx, orig_reg_rtx); | |
b757d36f | 6940 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, set); |
455786b6 | 6941 | } |
fada1961 | 6942 | |
f676971a | 6943 | if (TARGET_SHCOMPACT && (offset_in_r0 != -1)) |
fada1961 R |
6944 | { |
6945 | rtx reg_rtx = gen_rtx_REG (mode, reg); | |
b757d36f | 6946 | rtx set; |
57d38024 R |
6947 | rtx mem_rtx = gen_frame_mem (mode, |
6948 | gen_rtx_PLUS (Pmode, | |
6949 | stack_pointer_rtx, | |
6950 | GEN_INT (offset))); | |
fada1961 R |
6951 | |
6952 | set = gen_rtx_SET (VOIDmode, mem_rtx, reg_rtx); | |
b757d36f | 6953 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, set); |
fada1961 R |
6954 | } |
6955 | } | |
6956 | } | |
fa5322fa | 6957 | |
f5b9e7c9 | 6958 | gcc_assert (entry->offset == d_rounding); |
fa5322fa AO |
6959 | } |
6960 | else | |
ac55736a | 6961 | push_regs (&live_regs_mask, current_function_interrupt); |
1a66cd67 | 6962 | |
6fb5fa3c DB |
6963 | if (flag_pic && df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM)) |
6964 | emit_insn (gen_GOTaddr2picreg ()); | |
1a66cd67 | 6965 | |
fa5322fa AO |
6966 | if (SHMEDIA_REGS_STACK_ADJUST ()) |
6967 | { | |
fa5322fa AO |
6968 | /* This must NOT go through the PLT, otherwise mach and macl |
6969 | may be clobbered. */ | |
73a4d10b R |
6970 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
6971 | (TARGET_FPU_ANY | |
6972 | ? "__GCC_push_shmedia_regs" | |
6973 | : "__GCC_push_shmedia_regs_nofpu"), SFUNC_GOT); | |
fa5322fa AO |
6974 | emit_insn (gen_shmedia_save_restore_regs_compact |
6975 | (GEN_INT (-SHMEDIA_REGS_STACK_ADJUST ()))); | |
6976 | } | |
6977 | ||
ac55736a | 6978 | if (target_flags != save_flags && ! current_function_interrupt) |
6fb5fa3c | 6979 | emit_insn (gen_toggle_sz ()); |
1a95a963 | 6980 | |
225e4f43 R |
6981 | target_flags = save_flags; |
6982 | ||
fa5322fa | 6983 | output_stack_adjust (-rounded_frame_size (d) + d_rounding, |
fada1961 | 6984 | stack_pointer_rtx, 0, NULL); |
b9654711 SC |
6985 | |
6986 | if (frame_pointer_needed) | |
96a2347e | 6987 | frame_insn (GEN_MOV (hard_frame_pointer_rtx, stack_pointer_rtx)); |
fa5322fa | 6988 | |
fa5322fa | 6989 | if (TARGET_SHCOMPACT |
38173d38 | 6990 | && (crtl->args.info.call_cookie & ~ CALL_COOKIE_RET_TRAMP(1))) |
fa5322fa AO |
6991 | { |
6992 | /* This must NOT go through the PLT, otherwise mach and macl | |
6993 | may be clobbered. */ | |
73a4d10b R |
6994 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
6995 | "__GCC_shcompact_incoming_args", SFUNC_GOT); | |
fa5322fa AO |
6996 | emit_insn (gen_shcompact_incoming_args ()); |
6997 | } | |
b9654711 SC |
6998 | } |
6999 | ||
7000 | void | |
726d4cb7 | 7001 | sh_expand_epilogue (bool sibcall_p) |
b9654711 | 7002 | { |
490dbac7 | 7003 | HARD_REG_SET live_regs_mask; |
40d2032b | 7004 | int d, i; |
fa5322fa | 7005 | int d_rounding = 0; |
b9654711 | 7006 | |
225e4f43 | 7007 | int save_flags = target_flags; |
fe3ad572 | 7008 | int frame_size, save_size; |
ac55736a | 7009 | int fpscr_deferred = 0; |
726d4cb7 | 7010 | int e = sibcall_p ? -1 : 1; |
3d5a0820 | 7011 | |
490dbac7 | 7012 | d = calc_live_regs (&live_regs_mask); |
3d5a0820 | 7013 | |
fe3ad572 SC |
7014 | save_size = d; |
7015 | frame_size = rounded_frame_size (d); | |
7016 | ||
7017 | if (TARGET_SH5) | |
7018 | { | |
7019 | int tregs_space = shmedia_target_regs_stack_adjust (&live_regs_mask); | |
7020 | int total_size; | |
7021 | if (d % (STACK_BOUNDARY / BITS_PER_UNIT)) | |
7022 | d_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) | |
7023 | - d % (STACK_BOUNDARY / BITS_PER_UNIT)); | |
7024 | ||
7025 | total_size = d + tregs_space; | |
7026 | total_size += rounded_frame_size (total_size); | |
7027 | save_size = total_size - frame_size; | |
7028 | ||
7029 | /* If adjusting the stack in a single step costs nothing extra, do so. | |
7030 | I.e. either if a single addi is enough, or we need a movi anyway, | |
7031 | and we don't exceed the maximum offset range (the test for the | |
7032 | latter is conservative for simplicity). */ | |
7033 | if (TARGET_SHMEDIA | |
7034 | && ! frame_pointer_needed | |
7035 | && (CONST_OK_FOR_I10 (total_size) | |
7036 | || (! CONST_OK_FOR_I10 (save_size + d_rounding) | |
7037 | && total_size <= 2044))) | |
7038 | d_rounding = frame_size; | |
fa5322fa | 7039 | |
fe3ad572 SC |
7040 | frame_size -= d_rounding; |
7041 | } | |
16bea517 | 7042 | |
b9654711 | 7043 | if (frame_pointer_needed) |
46d81ffa | 7044 | { |
1964788a | 7045 | /* We must avoid scheduling the epilogue with previous basic blocks. |
47ea8d4e | 7046 | See PR/18032 and PR/40313. */ |
1964788a | 7047 | emit_insn (gen_blockage ()); |
96a2347e KK |
7048 | output_stack_adjust (frame_size, hard_frame_pointer_rtx, e, |
7049 | &live_regs_mask); | |
07f5b9aa JL |
7050 | |
7051 | /* We must avoid moving the stack pointer adjustment past code | |
7052 | which reads from the local frame, else an interrupt could | |
7053 | occur after the SP adjustment and clobber data in the local | |
7054 | frame. */ | |
7055 | emit_insn (gen_blockage ()); | |
96a2347e | 7056 | emit_insn (GEN_MOV (stack_pointer_rtx, hard_frame_pointer_rtx)); |
46d81ffa | 7057 | } |
3d5a0820 | 7058 | else if (frame_size) |
07f5b9aa JL |
7059 | { |
7060 | /* We must avoid moving the stack pointer adjustment past code | |
7061 | which reads from the local frame, else an interrupt could | |
7062 | occur after the SP adjustment and clobber data in the local | |
7063 | frame. */ | |
7064 | emit_insn (gen_blockage ()); | |
726d4cb7 | 7065 | output_stack_adjust (frame_size, stack_pointer_rtx, e, &live_regs_mask); |
07f5b9aa | 7066 | } |
b9654711 | 7067 | |
fa5322fa AO |
7068 | if (SHMEDIA_REGS_STACK_ADJUST ()) |
7069 | { | |
73a4d10b R |
7070 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
7071 | (TARGET_FPU_ANY | |
7072 | ? "__GCC_pop_shmedia_regs" | |
7073 | : "__GCC_pop_shmedia_regs_nofpu"), SFUNC_GOT); | |
fa5322fa AO |
7074 | /* This must NOT go through the PLT, otherwise mach and macl |
7075 | may be clobbered. */ | |
7076 | emit_insn (gen_shmedia_save_restore_regs_compact | |
7077 | (GEN_INT (SHMEDIA_REGS_STACK_ADJUST ()))); | |
7078 | } | |
7079 | ||
16bea517 | 7080 | /* Pop all the registers. */ |
0d7e008e | 7081 | |
ac55736a | 7082 | if (target_flags != save_flags && ! current_function_interrupt) |
225e4f43 | 7083 | emit_insn (gen_toggle_sz ()); |
fa5322fa AO |
7084 | if (TARGET_SH5) |
7085 | { | |
fada1961 | 7086 | int offset_base, offset; |
fa5322fa AO |
7087 | int offset_in_r0 = -1; |
7088 | int sp_in_r0 = 0; | |
fa5322fa | 7089 | rtx r0 = gen_rtx_REG (Pmode, R0_REG); |
fada1961 R |
7090 | save_schedule schedule; |
7091 | save_entry *entry; | |
7092 | int *tmp_pnt; | |
7d73a2ba | 7093 | |
fada1961 R |
7094 | entry = sh5_schedule_saves (&live_regs_mask, &schedule, d_rounding); |
7095 | offset_base = -entry[1].offset + d_rounding; | |
7096 | tmp_pnt = schedule.temps; | |
7097 | for (; entry->mode != VOIDmode; entry--) | |
7098 | { | |
81f40b79 | 7099 | enum machine_mode mode = (enum machine_mode) entry->mode; |
fada1961 R |
7100 | int reg = entry->reg; |
7101 | rtx reg_rtx, mem_rtx, post_inc = NULL_RTX, insn; | |
fa5322fa | 7102 | |
fada1961 R |
7103 | offset = offset_base + entry->offset; |
7104 | reg_rtx = gen_rtx_REG (mode, reg); | |
fa5322fa | 7105 | |
57d38024 R |
7106 | mem_rtx = gen_frame_mem (mode, |
7107 | gen_rtx_PLUS (Pmode, | |
7108 | stack_pointer_rtx, | |
7109 | GEN_INT (offset))); | |
fa5322fa | 7110 | |
6de7a512 PB |
7111 | if (!memory_address_p (mode, XEXP (mem_rtx, 0))) |
7112 | mem_rtx = NULL_RTX; | |
fa5322fa | 7113 | |
6de7a512 PB |
7114 | if (HAVE_POST_INCREMENT |
7115 | && (offset == offset_in_r0 | |
7116 | || (offset + GET_MODE_SIZE (mode) != d + d_rounding | |
7117 | && mem_rtx == NULL_RTX) | |
7118 | || reg == PR_REG || SPECIAL_REGISTER_P (reg))) | |
7119 | { | |
7120 | post_inc = gen_frame_mem (mode, gen_rtx_POST_INC (Pmode, r0)); | |
fa5322fa | 7121 | |
6de7a512 | 7122 | if (!memory_address_p (mode, XEXP (post_inc, 0))) |
fada1961 | 7123 | post_inc = NULL_RTX; |
6de7a512 | 7124 | else |
fada1961 | 7125 | mem_rtx = NULL_RTX; |
6de7a512 | 7126 | } |
f676971a | 7127 | |
fada1961 R |
7128 | if (mem_rtx != NULL_RTX) |
7129 | goto addr_ok; | |
fa5322fa | 7130 | |
fada1961 R |
7131 | if (offset_in_r0 == -1) |
7132 | { | |
7133 | emit_move_insn (r0, GEN_INT (offset)); | |
7134 | offset_in_r0 = offset; | |
7135 | } | |
7136 | else if (offset != offset_in_r0) | |
7137 | { | |
7138 | emit_move_insn (r0, | |
7139 | gen_rtx_PLUS | |
7140 | (Pmode, r0, | |
7141 | GEN_INT (offset - offset_in_r0))); | |
7142 | offset_in_r0 += offset - offset_in_r0; | |
7143 | } | |
f676971a | 7144 | |
fada1961 R |
7145 | if (post_inc != NULL_RTX) |
7146 | { | |
7147 | if (! sp_in_r0) | |
fa5322fa AO |
7148 | { |
7149 | emit_move_insn (r0, | |
7150 | gen_rtx_PLUS | |
fada1961 R |
7151 | (Pmode, r0, stack_pointer_rtx)); |
7152 | sp_in_r0 = 1; | |
fa5322fa | 7153 | } |
f676971a | 7154 | |
fada1961 | 7155 | mem_rtx = post_inc; |
fa5322fa | 7156 | |
fada1961 R |
7157 | offset_in_r0 += GET_MODE_SIZE (mode); |
7158 | } | |
7159 | else if (sp_in_r0) | |
57d38024 | 7160 | mem_rtx = gen_frame_mem (mode, r0); |
fada1961 | 7161 | else |
57d38024 R |
7162 | mem_rtx = gen_frame_mem (mode, |
7163 | gen_rtx_PLUS (Pmode, | |
7164 | stack_pointer_rtx, | |
7165 | r0)); | |
fa5322fa | 7166 | |
f5b9e7c9 NS |
7167 | gcc_assert ((reg != PR_REG && !SPECIAL_REGISTER_P (reg)) |
7168 | || mem_rtx == post_inc); | |
fa5322fa | 7169 | |
fada1961 R |
7170 | addr_ok: |
7171 | if ((reg == PR_REG || SPECIAL_REGISTER_P (reg)) | |
7172 | && mem_rtx != post_inc) | |
7173 | { | |
7174 | insn = emit_move_insn (r0, mem_rtx); | |
7175 | mem_rtx = r0; | |
7176 | } | |
7177 | else if (TARGET_REGISTER_P (reg)) | |
7178 | { | |
7179 | rtx tmp_reg = gen_rtx_REG (mode, *tmp_pnt); | |
7180 | ||
7181 | /* Give the scheduler a bit of freedom by using up to | |
7182 | MAX_TEMPS registers in a round-robin fashion. */ | |
7183 | insn = emit_move_insn (tmp_reg, mem_rtx); | |
7184 | mem_rtx = tmp_reg; | |
7185 | if (*++tmp_pnt < 0) | |
7186 | tmp_pnt = schedule.temps; | |
fa5322fa AO |
7187 | } |
7188 | ||
fada1961 | 7189 | insn = emit_move_insn (reg_rtx, mem_rtx); |
fada1961 R |
7190 | } |
7191 | ||
f5b9e7c9 | 7192 | gcc_assert (entry->offset + offset_base == d + d_rounding); |
fa5322fa | 7193 | } |
fe3ad572 | 7194 | else /* ! TARGET_SH5 */ |
b9654711 | 7195 | { |
5c7cafa8 CB |
7196 | int last_reg; |
7197 | ||
fe3ad572 | 7198 | save_size = 0; |
561642fa AP |
7199 | /* For an ISR with RESBANK attribute assigned, don't pop PR |
7200 | register. */ | |
7201 | if (TEST_HARD_REG_BIT (live_regs_mask, PR_REG) | |
7202 | && !sh_cfun_resbank_handler_p ()) | |
6fb5fa3c DB |
7203 | { |
7204 | if (!frame_pointer_needed) | |
7205 | emit_insn (gen_blockage ()); | |
7206 | pop (PR_REG); | |
7207 | } | |
5c7cafa8 CB |
7208 | |
7209 | /* Banked registers are poped first to avoid being scheduled in the | |
7210 | delay slot. RTE switches banks before the ds instruction. */ | |
7211 | if (current_function_interrupt) | |
7212 | { | |
7213 | for (i = FIRST_BANKED_REG; i <= LAST_BANKED_REG; i++) | |
7214 | if (TEST_HARD_REG_BIT (live_regs_mask, i)) | |
7215 | pop (LAST_BANKED_REG - i); | |
7216 | ||
7217 | last_reg = FIRST_PSEUDO_REGISTER - LAST_BANKED_REG - 1; | |
7218 | } | |
7219 | else | |
7220 | last_reg = FIRST_PSEUDO_REGISTER; | |
7221 | ||
7222 | for (i = 0; i < last_reg; i++) | |
fe3ad572 SC |
7223 | { |
7224 | int j = (FIRST_PSEUDO_REGISTER - 1) - i; | |
f676971a | 7225 | |
fe3ad572 | 7226 | if (j == FPSCR_REG && current_function_interrupt && TARGET_FMOVD |
56b138ae RS |
7227 | && hard_reg_set_intersect_p (live_regs_mask, |
7228 | reg_class_contents[DF_REGS])) | |
fe3ad572 | 7229 | fpscr_deferred = 1; |
561642fa AP |
7230 | /* For an ISR with RESBANK attribute assigned, don't pop |
7231 | following registers, R0-R14, MACH, MACL and GBR. */ | |
7232 | else if (j != PR_REG && TEST_HARD_REG_BIT (live_regs_mask, j) | |
7233 | && ! (sh_cfun_resbank_handler_p () | |
7234 | && ((j >= FIRST_GENERAL_REG | |
7235 | && j < LAST_GENERAL_REG) | |
7236 | || j == MACH_REG | |
7237 | || j == MACL_REG | |
7238 | || j == GBR_REG))) | |
fe3ad572 | 7239 | pop (j); |
5c7cafa8 | 7240 | |
fe3ad572 SC |
7241 | if (j == FIRST_FP_REG && fpscr_deferred) |
7242 | pop (FPSCR_REG); | |
fe3ad572 | 7243 | } |
b9654711 | 7244 | } |
ac55736a | 7245 | if (target_flags != save_flags && ! current_function_interrupt) |
225e4f43 R |
7246 | emit_insn (gen_toggle_sz ()); |
7247 | target_flags = save_flags; | |
b9654711 | 7248 | |
38173d38 | 7249 | output_stack_adjust (crtl->args.pretend_args_size |
fe3ad572 | 7250 | + save_size + d_rounding |
38173d38 | 7251 | + crtl->args.info.stack_regs * 8, |
726d4cb7 | 7252 | stack_pointer_rtx, e, NULL); |
4408efce | 7253 | |
e3b5732b | 7254 | if (crtl->calls_eh_return) |
4977bab6 ZW |
7255 | emit_insn (GEN_ADD3 (stack_pointer_rtx, stack_pointer_rtx, |
7256 | EH_RETURN_STACKADJ_RTX)); | |
7257 | ||
4408efce | 7258 | /* Switch back to the normal stack if necessary. */ |
a6ab9fc0 | 7259 | if (lookup_attribute ("sp_switch", DECL_ATTRIBUTES (current_function_decl))) |
4408efce | 7260 | emit_insn (gen_sp_switch_2 ()); |
89d4bc35 AO |
7261 | |
7262 | /* Tell flow the insn that pops PR isn't dead. */ | |
fa5322fa AO |
7263 | /* PR_REG will never be live in SHmedia mode, and we don't need to |
7264 | USE PR_MEDIA_REG, since it will be explicitly copied to TR0_REG | |
7265 | by the return pattern. */ | |
490dbac7 | 7266 | if (TEST_HARD_REG_BIT (live_regs_mask, PR_REG)) |
c41c1387 | 7267 | emit_use (gen_rtx_REG (SImode, PR_REG)); |
b9654711 SC |
7268 | } |
7269 | ||
afbc2905 R |
7270 | static int sh_need_epilogue_known = 0; |
7271 | ||
7272 | int | |
cf277499 | 7273 | sh_need_epilogue (void) |
afbc2905 R |
7274 | { |
7275 | if (! sh_need_epilogue_known) | |
7276 | { | |
7277 | rtx epilogue; | |
7278 | ||
7279 | start_sequence (); | |
726d4cb7 | 7280 | sh_expand_epilogue (0); |
2f937369 | 7281 | epilogue = get_insns (); |
afbc2905 | 7282 | end_sequence (); |
2f937369 | 7283 | sh_need_epilogue_known = (epilogue == NULL ? -1 : 1); |
afbc2905 R |
7284 | } |
7285 | return sh_need_epilogue_known > 0; | |
7286 | } | |
7287 | ||
4977bab6 ZW |
7288 | /* Emit code to change the current function's return address to RA. |
7289 | TEMP is available as a scratch register, if needed. */ | |
7290 | ||
7291 | void | |
cf277499 | 7292 | sh_set_return_address (rtx ra, rtx tmp) |
4977bab6 | 7293 | { |
490dbac7 | 7294 | HARD_REG_SET live_regs_mask; |
4977bab6 | 7295 | int d; |
4977bab6 ZW |
7296 | int pr_reg = TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG; |
7297 | int pr_offset; | |
7298 | ||
490dbac7 | 7299 | d = calc_live_regs (&live_regs_mask); |
4977bab6 ZW |
7300 | |
7301 | /* If pr_reg isn't life, we can set it (or the register given in | |
7302 | sh_media_register_for_return) directly. */ | |
490dbac7 | 7303 | if (! TEST_HARD_REG_BIT (live_regs_mask, pr_reg)) |
4977bab6 ZW |
7304 | { |
7305 | rtx rr; | |
7306 | ||
7307 | if (TARGET_SHMEDIA) | |
7308 | { | |
7309 | int rr_regno = sh_media_register_for_return (); | |
7310 | ||
7311 | if (rr_regno < 0) | |
7312 | rr_regno = pr_reg; | |
7313 | ||
7314 | rr = gen_rtx_REG (DImode, rr_regno); | |
7315 | } | |
7316 | else | |
7317 | rr = gen_rtx_REG (SImode, pr_reg); | |
7318 | ||
7319 | emit_insn (GEN_MOV (rr, ra)); | |
7320 | /* Tell flow the register for return isn't dead. */ | |
c41c1387 | 7321 | emit_use (rr); |
4977bab6 ZW |
7322 | return; |
7323 | } | |
7324 | ||
7325 | if (TARGET_SH5) | |
7326 | { | |
4977bab6 | 7327 | int offset; |
fada1961 R |
7328 | save_schedule schedule; |
7329 | save_entry *entry; | |
f676971a | 7330 | |
fada1961 R |
7331 | entry = sh5_schedule_saves (&live_regs_mask, &schedule, 0); |
7332 | offset = entry[1].offset; | |
7333 | for (; entry->mode != VOIDmode; entry--) | |
7334 | if (entry->reg == pr_reg) | |
7335 | goto found; | |
4977bab6 ZW |
7336 | |
7337 | /* We can't find pr register. */ | |
f5b9e7c9 | 7338 | gcc_unreachable (); |
4977bab6 ZW |
7339 | |
7340 | found: | |
fada1961 R |
7341 | offset = entry->offset - offset; |
7342 | pr_offset = (rounded_frame_size (d) + offset | |
4977bab6 ZW |
7343 | + SHMEDIA_REGS_STACK_ADJUST ()); |
7344 | } | |
7345 | else | |
fada1961 | 7346 | pr_offset = rounded_frame_size (d); |
4977bab6 ZW |
7347 | |
7348 | emit_insn (GEN_MOV (tmp, GEN_INT (pr_offset))); | |
96a2347e | 7349 | emit_insn (GEN_ADD3 (tmp, tmp, hard_frame_pointer_rtx)); |
4977bab6 | 7350 | |
57d38024 | 7351 | tmp = gen_frame_mem (Pmode, tmp); |
4977bab6 | 7352 | emit_insn (GEN_MOV (tmp, ra)); |
7f9a14e3 KK |
7353 | /* Tell this store isn't dead. */ |
7354 | emit_use (tmp); | |
4977bab6 ZW |
7355 | } |
7356 | ||
8aa2a305 JW |
7357 | /* Clear variables at function end. */ |
7358 | ||
08c148a8 | 7359 | static void |
cf277499 SB |
7360 | sh_output_function_epilogue (FILE *file ATTRIBUTE_UNUSED, |
7361 | HOST_WIDE_INT size ATTRIBUTE_UNUSED) | |
8aa2a305 | 7362 | { |
afbc2905 | 7363 | sh_need_epilogue_known = 0; |
8aa2a305 JW |
7364 | } |
7365 | ||
61f71b34 | 7366 | static rtx |
cf277499 | 7367 | sh_builtin_saveregs (void) |
1a95a963 | 7368 | { |
1a95a963 | 7369 | /* First unnamed integer register. */ |
38173d38 | 7370 | int first_intreg = crtl->args.info.arg_count[(int) SH_ARG_INT]; |
1a95a963 JW |
7371 | /* Number of integer registers we need to save. */ |
7372 | int n_intregs = MAX (0, NPARM_REGS (SImode) - first_intreg); | |
7373 | /* First unnamed SFmode float reg */ | |
38173d38 | 7374 | int first_floatreg = crtl->args.info.arg_count[(int) SH_ARG_FLOAT]; |
1a95a963 JW |
7375 | /* Number of SFmode float regs to save. */ |
7376 | int n_floatregs = MAX (0, NPARM_REGS (SFmode) - first_floatreg); | |
318881c0 | 7377 | rtx regbuf, fpregs; |
ba4828e0 | 7378 | int bufsize, regno; |
4862826d | 7379 | alias_set_type alias_set; |
1a95a963 | 7380 | |
fa5322fa AO |
7381 | if (TARGET_SH5) |
7382 | { | |
7383 | if (n_intregs) | |
7384 | { | |
7385 | int pushregs = n_intregs; | |
7386 | ||
7387 | while (pushregs < NPARM_REGS (SImode) - 1 | |
7388 | && (CALL_COOKIE_INT_REG_GET | |
38173d38 | 7389 | (crtl->args.info.call_cookie, |
fa5322fa AO |
7390 | NPARM_REGS (SImode) - pushregs) |
7391 | == 1)) | |
7392 | { | |
38173d38 | 7393 | crtl->args.info.call_cookie |
fa5322fa AO |
7394 | &= ~ CALL_COOKIE_INT_REG (NPARM_REGS (SImode) |
7395 | - pushregs, 1); | |
7396 | pushregs++; | |
7397 | } | |
7398 | ||
7399 | if (pushregs == NPARM_REGS (SImode)) | |
38173d38 | 7400 | crtl->args.info.call_cookie |
fa5322fa AO |
7401 | |= (CALL_COOKIE_INT_REG (0, 1) |
7402 | | CALL_COOKIE_STACKSEQ (pushregs - 1)); | |
7403 | else | |
38173d38 | 7404 | crtl->args.info.call_cookie |
fa5322fa AO |
7405 | |= CALL_COOKIE_STACKSEQ (pushregs); |
7406 | ||
38173d38 | 7407 | crtl->args.pretend_args_size += 8 * n_intregs; |
fa5322fa AO |
7408 | } |
7409 | if (TARGET_SHCOMPACT) | |
7410 | return const0_rtx; | |
7411 | } | |
f676971a | 7412 | |
3a8699c7 | 7413 | if (! TARGET_SH2E && ! TARGET_SH4 && ! TARGET_SH5) |
fa5322fa AO |
7414 | { |
7415 | error ("__builtin_saveregs not supported by this subtarget"); | |
7416 | return const0_rtx; | |
7417 | } | |
7418 | ||
7419 | if (TARGET_SHMEDIA) | |
7420 | n_floatregs = 0; | |
7421 | ||
6f317ef3 | 7422 | /* Allocate block of memory for the regs. */ |
1a95a963 JW |
7423 | /* ??? If n_intregs + n_floatregs == 0, should we allocate at least 1 byte? |
7424 | Or can assign_stack_local accept a 0 SIZE argument? */ | |
7425 | bufsize = (n_intregs * UNITS_PER_WORD) + (n_floatregs * UNITS_PER_WORD); | |
7426 | ||
fa5322fa | 7427 | if (TARGET_SHMEDIA) |
57d38024 | 7428 | regbuf = gen_frame_mem (BLKmode, gen_rtx_REG (Pmode, ARG_POINTER_REGNUM)); |
f78ec36a R |
7429 | else if (n_floatregs & 1) |
7430 | { | |
7431 | rtx addr; | |
7432 | ||
7433 | regbuf = assign_stack_local (BLKmode, bufsize + UNITS_PER_WORD, 0); | |
7434 | addr = copy_to_mode_reg (Pmode, XEXP (regbuf, 0)); | |
7435 | emit_insn (gen_iorsi3 (addr, addr, GEN_INT (UNITS_PER_WORD))); | |
7436 | regbuf = change_address (regbuf, BLKmode, addr); | |
7437 | } | |
a11c004b R |
7438 | else if (STACK_BOUNDARY < 64 && TARGET_FPU_DOUBLE && n_floatregs) |
7439 | { | |
7440 | rtx addr, mask; | |
7441 | ||
7442 | regbuf = assign_stack_local (BLKmode, bufsize + UNITS_PER_WORD, 0); | |
7443 | addr = copy_to_mode_reg (Pmode, plus_constant (XEXP (regbuf, 0), 4)); | |
7444 | mask = copy_to_mode_reg (Pmode, GEN_INT (-8)); | |
7445 | emit_insn (gen_andsi3 (addr, addr, mask)); | |
7446 | regbuf = change_address (regbuf, BLKmode, addr); | |
7447 | } | |
fa5322fa | 7448 | else |
a11c004b | 7449 | regbuf = assign_stack_local (BLKmode, bufsize, TARGET_FPU_DOUBLE ? 64 : 0); |
514066a1 | 7450 | alias_set = get_varargs_alias_set (); |
ba4828e0 | 7451 | set_mem_alias_set (regbuf, alias_set); |
1a95a963 JW |
7452 | |
7453 | /* Save int args. | |
7454 | This is optimized to only save the regs that are necessary. Explicitly | |
7455 | named args need not be saved. */ | |
7456 | if (n_intregs > 0) | |
7457 | move_block_from_reg (BASE_ARG_REG (SImode) + first_intreg, | |
f4ef873c RK |
7458 | adjust_address (regbuf, BLKmode, |
7459 | n_floatregs * UNITS_PER_WORD), | |
c6b97fac | 7460 | n_intregs); |
1a95a963 | 7461 | |
fa5322fa AO |
7462 | if (TARGET_SHMEDIA) |
7463 | /* Return the address of the regbuf. */ | |
7464 | return XEXP (regbuf, 0); | |
7465 | ||
1a95a963 JW |
7466 | /* Save float args. |
7467 | This is optimized to only save the regs that are necessary. Explicitly | |
7468 | named args need not be saved. | |
7469 | We explicitly build a pointer to the buffer because it halves the insn | |
7470 | count when not optimizing (otherwise the pointer is built for each reg | |
1245df60 R |
7471 | saved). |
7472 | We emit the moves in reverse order so that we can use predecrement. */ | |
1a95a963 | 7473 | |
7d9c5aa1 R |
7474 | fpregs = copy_to_mode_reg (Pmode, |
7475 | plus_constant (XEXP (regbuf, 0), | |
7476 | n_floatregs * UNITS_PER_WORD)); | |
157371cf | 7477 | if (TARGET_SH4 || TARGET_SH2A_DOUBLE) |
225e4f43 | 7478 | { |
514066a1 | 7479 | rtx mem; |
225e4f43 R |
7480 | for (regno = NPARM_REGS (DFmode) - 2; regno >= first_floatreg; regno -= 2) |
7481 | { | |
7482 | emit_insn (gen_addsi3 (fpregs, fpregs, | |
7483 | GEN_INT (-2 * UNITS_PER_WORD))); | |
57d38024 | 7484 | mem = change_address (regbuf, DFmode, fpregs); |
f676971a | 7485 | emit_move_insn (mem, |
c0d4e710 | 7486 | gen_rtx_REG (DFmode, BASE_ARG_REG (DFmode) + regno)); |
225e4f43 R |
7487 | } |
7488 | regno = first_floatreg; | |
7489 | if (regno & 1) | |
7490 | { | |
832a3292 | 7491 | emit_insn (gen_addsi3 (fpregs, fpregs, GEN_INT (-UNITS_PER_WORD))); |
57d38024 | 7492 | mem = change_address (regbuf, SFmode, fpregs); |
514066a1 | 7493 | emit_move_insn (mem, |
c0d4e710 | 7494 | gen_rtx_REG (SFmode, BASE_ARG_REG (SFmode) + regno |
225e4f43 R |
7495 | - (TARGET_LITTLE_ENDIAN != 0))); |
7496 | } | |
7497 | } | |
7498 | else | |
1245df60 R |
7499 | for (regno = NPARM_REGS (SFmode) - 1; regno >= first_floatreg; regno--) |
7500 | { | |
514066a1 | 7501 | rtx mem; |
ba4828e0 | 7502 | |
832a3292 | 7503 | emit_insn (gen_addsi3 (fpregs, fpregs, GEN_INT (-UNITS_PER_WORD))); |
57d38024 | 7504 | mem = change_address (regbuf, SFmode, fpregs); |
514066a1 | 7505 | emit_move_insn (mem, |
c5c76735 | 7506 | gen_rtx_REG (SFmode, BASE_ARG_REG (SFmode) + regno)); |
1245df60 | 7507 | } |
1a95a963 JW |
7508 | |
7509 | /* Return the address of the regbuf. */ | |
7510 | return XEXP (regbuf, 0); | |
7511 | } | |
7512 | ||
514066a1 RH |
7513 | /* Define the `__builtin_va_list' type for the ABI. */ |
7514 | ||
c35d187f RH |
7515 | static tree |
7516 | sh_build_builtin_va_list (void) | |
514066a1 RH |
7517 | { |
7518 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
7519 | tree record; | |
7520 | ||
61f71b34 DD |
7521 | if (TARGET_SH5 || (! TARGET_SH2E && ! TARGET_SH4) |
7522 | || TARGET_HITACHI || sh_cfun_attr_renesas_p ()) | |
514066a1 RH |
7523 | return ptr_type_node; |
7524 | ||
5d27ef94 | 7525 | record = (*lang_hooks.types.make_type) (RECORD_TYPE); |
514066a1 | 7526 | |
4c4bde29 AH |
7527 | f_next_o = build_decl (BUILTINS_LOCATION, |
7528 | FIELD_DECL, get_identifier ("__va_next_o"), | |
514066a1 | 7529 | ptr_type_node); |
4c4bde29 AH |
7530 | f_next_o_limit = build_decl (BUILTINS_LOCATION, |
7531 | FIELD_DECL, | |
514066a1 RH |
7532 | get_identifier ("__va_next_o_limit"), |
7533 | ptr_type_node); | |
4c4bde29 AH |
7534 | f_next_fp = build_decl (BUILTINS_LOCATION, |
7535 | FIELD_DECL, get_identifier ("__va_next_fp"), | |
514066a1 | 7536 | ptr_type_node); |
4c4bde29 AH |
7537 | f_next_fp_limit = build_decl (BUILTINS_LOCATION, |
7538 | FIELD_DECL, | |
514066a1 RH |
7539 | get_identifier ("__va_next_fp_limit"), |
7540 | ptr_type_node); | |
4c4bde29 AH |
7541 | f_next_stack = build_decl (BUILTINS_LOCATION, |
7542 | FIELD_DECL, get_identifier ("__va_next_stack"), | |
514066a1 RH |
7543 | ptr_type_node); |
7544 | ||
7545 | DECL_FIELD_CONTEXT (f_next_o) = record; | |
7546 | DECL_FIELD_CONTEXT (f_next_o_limit) = record; | |
7547 | DECL_FIELD_CONTEXT (f_next_fp) = record; | |
7548 | DECL_FIELD_CONTEXT (f_next_fp_limit) = record; | |
7549 | DECL_FIELD_CONTEXT (f_next_stack) = record; | |
7550 | ||
7551 | TYPE_FIELDS (record) = f_next_o; | |
7552 | TREE_CHAIN (f_next_o) = f_next_o_limit; | |
7553 | TREE_CHAIN (f_next_o_limit) = f_next_fp; | |
7554 | TREE_CHAIN (f_next_fp) = f_next_fp_limit; | |
7555 | TREE_CHAIN (f_next_fp_limit) = f_next_stack; | |
7556 | ||
7557 | layout_type (record); | |
7558 | ||
7559 | return record; | |
7560 | } | |
7561 | ||
7562 | /* Implement `va_start' for varargs and stdarg. */ | |
7563 | ||
d7bd8aeb | 7564 | static void |
cf277499 | 7565 | sh_va_start (tree valist, rtx nextarg) |
514066a1 RH |
7566 | { |
7567 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
7568 | tree next_o, next_o_limit, next_fp, next_fp_limit, next_stack; | |
7569 | tree t, u; | |
7570 | int nfp, nint; | |
7571 | ||
fa5322fa AO |
7572 | if (TARGET_SH5) |
7573 | { | |
7574 | expand_builtin_saveregs (); | |
e5faf155 | 7575 | std_expand_builtin_va_start (valist, nextarg); |
fa5322fa AO |
7576 | return; |
7577 | } | |
7578 | ||
61f71b34 DD |
7579 | if ((! TARGET_SH2E && ! TARGET_SH4) |
7580 | || TARGET_HITACHI || sh_cfun_attr_renesas_p ()) | |
514066a1 | 7581 | { |
e5faf155 | 7582 | std_expand_builtin_va_start (valist, nextarg); |
514066a1 RH |
7583 | return; |
7584 | } | |
7585 | ||
7586 | f_next_o = TYPE_FIELDS (va_list_type_node); | |
7587 | f_next_o_limit = TREE_CHAIN (f_next_o); | |
7588 | f_next_fp = TREE_CHAIN (f_next_o_limit); | |
7589 | f_next_fp_limit = TREE_CHAIN (f_next_fp); | |
7590 | f_next_stack = TREE_CHAIN (f_next_fp_limit); | |
7591 | ||
47a25a46 | 7592 | next_o = build3 (COMPONENT_REF, TREE_TYPE (f_next_o), valist, f_next_o, |
44de5aeb | 7593 | NULL_TREE); |
47a25a46 RG |
7594 | next_o_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_o_limit), |
7595 | valist, f_next_o_limit, NULL_TREE); | |
7596 | next_fp = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp), valist, f_next_fp, | |
7597 | NULL_TREE); | |
7598 | next_fp_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp_limit), | |
7599 | valist, f_next_fp_limit, NULL_TREE); | |
7600 | next_stack = build3 (COMPONENT_REF, TREE_TYPE (f_next_stack), | |
7601 | valist, f_next_stack, NULL_TREE); | |
514066a1 RH |
7602 | |
7603 | /* Call __builtin_saveregs. */ | |
5be014d5 AP |
7604 | u = make_tree (sizetype, expand_builtin_saveregs ()); |
7605 | u = fold_convert (ptr_type_node, u); | |
726a989a | 7606 | t = build2 (MODIFY_EXPR, ptr_type_node, next_fp, u); |
514066a1 RH |
7607 | TREE_SIDE_EFFECTS (t) = 1; |
7608 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7609 | ||
38173d38 | 7610 | nfp = crtl->args.info.arg_count[SH_ARG_FLOAT]; |
514066a1 RH |
7611 | if (nfp < 8) |
7612 | nfp = 8 - nfp; | |
7613 | else | |
7614 | nfp = 0; | |
5be014d5 AP |
7615 | u = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, u, |
7616 | size_int (UNITS_PER_WORD * nfp)); | |
726a989a | 7617 | t = build2 (MODIFY_EXPR, ptr_type_node, next_fp_limit, u); |
514066a1 RH |
7618 | TREE_SIDE_EFFECTS (t) = 1; |
7619 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7620 | ||
726a989a | 7621 | t = build2 (MODIFY_EXPR, ptr_type_node, next_o, u); |
514066a1 RH |
7622 | TREE_SIDE_EFFECTS (t) = 1; |
7623 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7624 | ||
38173d38 | 7625 | nint = crtl->args.info.arg_count[SH_ARG_INT]; |
514066a1 RH |
7626 | if (nint < 4) |
7627 | nint = 4 - nint; | |
7628 | else | |
7629 | nint = 0; | |
5be014d5 AP |
7630 | u = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, u, |
7631 | size_int (UNITS_PER_WORD * nint)); | |
726a989a | 7632 | t = build2 (MODIFY_EXPR, ptr_type_node, next_o_limit, u); |
514066a1 RH |
7633 | TREE_SIDE_EFFECTS (t) = 1; |
7634 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7635 | ||
7636 | u = make_tree (ptr_type_node, nextarg); | |
726a989a | 7637 | t = build2 (MODIFY_EXPR, ptr_type_node, next_stack, u); |
514066a1 RH |
7638 | TREE_SIDE_EFFECTS (t) = 1; |
7639 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7640 | } | |
7641 | ||
917f1b7e | 7642 | /* TYPE is a RECORD_TYPE. If there is only a single nonzero-sized |
21bdb106 JR |
7643 | member, return it. */ |
7644 | static tree | |
7645 | find_sole_member (tree type) | |
7646 | { | |
7647 | tree field, member = NULL_TREE; | |
7648 | ||
7649 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
7650 | { | |
7651 | if (TREE_CODE (field) != FIELD_DECL) | |
7652 | continue; | |
7653 | if (!DECL_SIZE (field)) | |
7654 | return NULL_TREE; | |
7655 | if (integer_zerop (DECL_SIZE (field))) | |
7656 | continue; | |
7657 | if (member) | |
7658 | return NULL_TREE; | |
7659 | member = field; | |
7660 | } | |
7661 | return member; | |
7662 | } | |
514066a1 RH |
7663 | /* Implement `va_arg'. */ |
7664 | ||
5fac2ed2 | 7665 | static tree |
726a989a RB |
7666 | sh_gimplify_va_arg_expr (tree valist, tree type, gimple_seq *pre_p, |
7667 | gimple_seq *post_p ATTRIBUTE_UNUSED) | |
514066a1 RH |
7668 | { |
7669 | HOST_WIDE_INT size, rsize; | |
318881c0 | 7670 | tree tmp, pptr_type_node; |
0ef2c525 | 7671 | tree addr, lab_over = NULL, result = NULL; |
03603192 | 7672 | int pass_by_ref = targetm.calls.must_pass_in_stack (TYPE_MODE (type), type); |
11160b58 | 7673 | tree eff_type; |
5fac2ed2 RH |
7674 | |
7675 | if (pass_by_ref) | |
7676 | type = build_pointer_type (type); | |
514066a1 RH |
7677 | |
7678 | size = int_size_in_bytes (type); | |
7679 | rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD; | |
7680 | pptr_type_node = build_pointer_type (ptr_type_node); | |
7681 | ||
61f71b34 DD |
7682 | if (! TARGET_SH5 && (TARGET_SH2E || TARGET_SH4) |
7683 | && ! (TARGET_HITACHI || sh_cfun_attr_renesas_p ())) | |
514066a1 RH |
7684 | { |
7685 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
7686 | tree next_o, next_o_limit, next_fp, next_fp_limit, next_stack; | |
7687 | int pass_as_float; | |
5fac2ed2 | 7688 | tree lab_false; |
21bdb106 | 7689 | tree member; |
514066a1 RH |
7690 | |
7691 | f_next_o = TYPE_FIELDS (va_list_type_node); | |
7692 | f_next_o_limit = TREE_CHAIN (f_next_o); | |
7693 | f_next_fp = TREE_CHAIN (f_next_o_limit); | |
7694 | f_next_fp_limit = TREE_CHAIN (f_next_fp); | |
7695 | f_next_stack = TREE_CHAIN (f_next_fp_limit); | |
7696 | ||
47a25a46 RG |
7697 | next_o = build3 (COMPONENT_REF, TREE_TYPE (f_next_o), valist, f_next_o, |
7698 | NULL_TREE); | |
7699 | next_o_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_o_limit), | |
7700 | valist, f_next_o_limit, NULL_TREE); | |
7701 | next_fp = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp), | |
7702 | valist, f_next_fp, NULL_TREE); | |
7703 | next_fp_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp_limit), | |
7704 | valist, f_next_fp_limit, NULL_TREE); | |
7705 | next_stack = build3 (COMPONENT_REF, TREE_TYPE (f_next_stack), | |
7706 | valist, f_next_stack, NULL_TREE); | |
514066a1 | 7707 | |
6d1cb95f R |
7708 | /* Structures with a single member with a distinct mode are passed |
7709 | like their member. This is relevant if the latter has a REAL_TYPE | |
7710 | or COMPLEX_TYPE type. */ | |
11160b58 R |
7711 | eff_type = type; |
7712 | while (TREE_CODE (eff_type) == RECORD_TYPE | |
7713 | && (member = find_sole_member (eff_type)) | |
21bdb106 JR |
7714 | && (TREE_CODE (TREE_TYPE (member)) == REAL_TYPE |
7715 | || TREE_CODE (TREE_TYPE (member)) == COMPLEX_TYPE | |
7716 | || TREE_CODE (TREE_TYPE (member)) == RECORD_TYPE)) | |
c26104e4 | 7717 | { |
21bdb106 | 7718 | tree field_type = TREE_TYPE (member); |
c26104e4 | 7719 | |
11160b58 R |
7720 | if (TYPE_MODE (eff_type) == TYPE_MODE (field_type)) |
7721 | eff_type = field_type; | |
c26104e4 | 7722 | else |
8a2209bf | 7723 | { |
11160b58 | 7724 | gcc_assert ((TYPE_ALIGN (eff_type) |
21bdb106 | 7725 | < GET_MODE_ALIGNMENT (TYPE_MODE (field_type))) |
11160b58 | 7726 | || (TYPE_ALIGN (eff_type) |
21bdb106 JR |
7727 | > GET_MODE_BITSIZE (TYPE_MODE (field_type)))); |
7728 | break; | |
8a2209bf | 7729 | } |
c26104e4 | 7730 | } |
5fac2ed2 | 7731 | |
bdf57c40 | 7732 | if (TARGET_SH4 || TARGET_SH2A_DOUBLE) |
514066a1 | 7733 | { |
11160b58 R |
7734 | pass_as_float = ((TREE_CODE (eff_type) == REAL_TYPE && size <= 8) |
7735 | || (TREE_CODE (eff_type) == COMPLEX_TYPE | |
7736 | && TREE_CODE (TREE_TYPE (eff_type)) == REAL_TYPE | |
514066a1 RH |
7737 | && size <= 16)); |
7738 | } | |
7739 | else | |
7740 | { | |
11160b58 | 7741 | pass_as_float = (TREE_CODE (eff_type) == REAL_TYPE && size == 4); |
514066a1 RH |
7742 | } |
7743 | ||
5fac2ed2 | 7744 | addr = create_tmp_var (pptr_type_node, NULL); |
4c4bde29 AH |
7745 | lab_false = create_artificial_label (UNKNOWN_LOCATION); |
7746 | lab_over = create_artificial_label (UNKNOWN_LOCATION); | |
514066a1 | 7747 | |
5fac2ed2 | 7748 | valist = build1 (INDIRECT_REF, ptr_type_node, addr); |
6d1cb95f | 7749 | |
514066a1 RH |
7750 | if (pass_as_float) |
7751 | { | |
a11c004b R |
7752 | tree next_fp_tmp = create_tmp_var (TREE_TYPE (f_next_fp), NULL); |
7753 | tree cmp; | |
11160b58 | 7754 | bool is_double = size == 8 && TREE_CODE (eff_type) == REAL_TYPE; |
f78ec36a | 7755 | |
8c54989a KK |
7756 | tmp = build1 (ADDR_EXPR, pptr_type_node, unshare_expr (next_fp)); |
7757 | gimplify_assign (unshare_expr (addr), tmp, pre_p); | |
a11c004b | 7758 | |
8c54989a | 7759 | gimplify_assign (unshare_expr (next_fp_tmp), valist, pre_p); |
a11c004b R |
7760 | tmp = next_fp_limit; |
7761 | if (size > 4 && !is_double) | |
8c54989a KK |
7762 | tmp = build2 (POINTER_PLUS_EXPR, TREE_TYPE (tmp), |
7763 | unshare_expr (tmp), size_int (4 - size)); | |
7764 | tmp = build2 (GE_EXPR, boolean_type_node, | |
7765 | unshare_expr (next_fp_tmp), unshare_expr (tmp)); | |
47a25a46 | 7766 | cmp = build3 (COND_EXPR, void_type_node, tmp, |
8c54989a KK |
7767 | build1 (GOTO_EXPR, void_type_node, |
7768 | unshare_expr (lab_false)), NULL_TREE); | |
a11c004b R |
7769 | if (!is_double) |
7770 | gimplify_and_add (cmp, pre_p); | |
514066a1 | 7771 | |
11160b58 R |
7772 | if (TYPE_ALIGN (eff_type) > BITS_PER_WORD |
7773 | || (is_double || size == 16)) | |
514066a1 | 7774 | { |
5be014d5 AP |
7775 | tmp = fold_convert (sizetype, next_fp_tmp); |
7776 | tmp = build2 (BIT_AND_EXPR, sizetype, tmp, | |
7777 | size_int (UNITS_PER_WORD)); | |
7778 | tmp = build2 (POINTER_PLUS_EXPR, ptr_type_node, | |
8c54989a KK |
7779 | unshare_expr (next_fp_tmp), tmp); |
7780 | gimplify_assign (unshare_expr (next_fp_tmp), tmp, pre_p); | |
514066a1 | 7781 | } |
a11c004b R |
7782 | if (is_double) |
7783 | gimplify_and_add (cmp, pre_p); | |
514066a1 | 7784 | |
6d1cb95f | 7785 | #ifdef FUNCTION_ARG_SCmode_WART |
11160b58 R |
7786 | if (TYPE_MODE (eff_type) == SCmode |
7787 | && TARGET_SH4 && TARGET_LITTLE_ENDIAN) | |
6d1cb95f | 7788 | { |
11160b58 | 7789 | tree subtype = TREE_TYPE (eff_type); |
5fac2ed2 | 7790 | tree real, imag; |
6d1cb95f | 7791 | |
a11c004b R |
7792 | imag |
7793 | = std_gimplify_va_arg_expr (next_fp_tmp, subtype, pre_p, NULL); | |
5fac2ed2 | 7794 | imag = get_initialized_tmp_var (imag, pre_p, NULL); |
6d1cb95f | 7795 | |
a11c004b R |
7796 | real |
7797 | = std_gimplify_va_arg_expr (next_fp_tmp, subtype, pre_p, NULL); | |
5fac2ed2 RH |
7798 | real = get_initialized_tmp_var (real, pre_p, NULL); |
7799 | ||
f2a945b1 KK |
7800 | result = build2 (COMPLEX_EXPR, eff_type, real, imag); |
7801 | if (type != eff_type) | |
7802 | result = build1 (VIEW_CONVERT_EXPR, type, result); | |
5fac2ed2 | 7803 | result = get_initialized_tmp_var (result, pre_p, NULL); |
6d1cb95f R |
7804 | } |
7805 | #endif /* FUNCTION_ARG_SCmode_WART */ | |
7806 | ||
8c54989a | 7807 | tmp = build1 (GOTO_EXPR, void_type_node, unshare_expr (lab_over)); |
5fac2ed2 RH |
7808 | gimplify_and_add (tmp, pre_p); |
7809 | ||
8c54989a | 7810 | tmp = build1 (LABEL_EXPR, void_type_node, unshare_expr (lab_false)); |
5fac2ed2 | 7811 | gimplify_and_add (tmp, pre_p); |
514066a1 | 7812 | |
8c54989a KK |
7813 | tmp = build1 (ADDR_EXPR, pptr_type_node, unshare_expr (next_stack)); |
7814 | gimplify_assign (unshare_expr (addr), tmp, pre_p); | |
7815 | gimplify_assign (unshare_expr (next_fp_tmp), | |
7816 | unshare_expr (valist), pre_p); | |
a11c004b | 7817 | |
8c54989a KK |
7818 | gimplify_assign (unshare_expr (valist), |
7819 | unshare_expr (next_fp_tmp), post_p); | |
a11c004b | 7820 | valist = next_fp_tmp; |
514066a1 RH |
7821 | } |
7822 | else | |
7823 | { | |
8c54989a KK |
7824 | tmp = build2 (POINTER_PLUS_EXPR, ptr_type_node, |
7825 | unshare_expr (next_o), size_int (rsize)); | |
7826 | tmp = build2 (GT_EXPR, boolean_type_node, tmp, | |
7827 | unshare_expr (next_o_limit)); | |
47a25a46 | 7828 | tmp = build3 (COND_EXPR, void_type_node, tmp, |
8c54989a KK |
7829 | build1 (GOTO_EXPR, void_type_node, |
7830 | unshare_expr (lab_false)), | |
7831 | NULL_TREE); | |
5fac2ed2 | 7832 | gimplify_and_add (tmp, pre_p); |
514066a1 | 7833 | |
8c54989a KK |
7834 | tmp = build1 (ADDR_EXPR, pptr_type_node, unshare_expr (next_o)); |
7835 | gimplify_assign (unshare_expr (addr), tmp, pre_p); | |
514066a1 | 7836 | |
8c54989a | 7837 | tmp = build1 (GOTO_EXPR, void_type_node, unshare_expr (lab_over)); |
5fac2ed2 RH |
7838 | gimplify_and_add (tmp, pre_p); |
7839 | ||
8c54989a | 7840 | tmp = build1 (LABEL_EXPR, void_type_node, unshare_expr (lab_false)); |
5fac2ed2 | 7841 | gimplify_and_add (tmp, pre_p); |
514066a1 | 7842 | |
bdf57c40 | 7843 | if (size > 4 && ! (TARGET_SH4 || TARGET_SH2A)) |
8c54989a KK |
7844 | gimplify_assign (unshare_expr (next_o), |
7845 | unshare_expr (next_o_limit), pre_p); | |
514066a1 | 7846 | |
8c54989a KK |
7847 | tmp = build1 (ADDR_EXPR, pptr_type_node, unshare_expr (next_stack)); |
7848 | gimplify_assign (unshare_expr (addr), tmp, pre_p); | |
514066a1 RH |
7849 | } |
7850 | ||
5fac2ed2 RH |
7851 | if (!result) |
7852 | { | |
8c54989a | 7853 | tmp = build1 (LABEL_EXPR, void_type_node, unshare_expr (lab_over)); |
5fac2ed2 RH |
7854 | gimplify_and_add (tmp, pre_p); |
7855 | } | |
514066a1 RH |
7856 | } |
7857 | ||
7858 | /* ??? In va-sh.h, there had been code to make values larger than | |
7859 | size 8 indirect. This does not match the FUNCTION_ARG macros. */ | |
7860 | ||
5fac2ed2 | 7861 | tmp = std_gimplify_va_arg_expr (valist, type, pre_p, NULL); |
6d1cb95f R |
7862 | if (result) |
7863 | { | |
726a989a | 7864 | gimplify_assign (result, tmp, pre_p); |
5fac2ed2 | 7865 | |
8c54989a | 7866 | tmp = build1 (LABEL_EXPR, void_type_node, unshare_expr (lab_over)); |
5fac2ed2 | 7867 | gimplify_and_add (tmp, pre_p); |
6d1cb95f R |
7868 | } |
7869 | else | |
5fac2ed2 | 7870 | result = tmp; |
6d1cb95f | 7871 | |
5e039e6b | 7872 | if (pass_by_ref) |
d6e9821f | 7873 | result = build_va_arg_indirect_ref (result); |
5fac2ed2 | 7874 | |
5e039e6b | 7875 | return result; |
514066a1 RH |
7876 | } |
7877 | ||
63a9de54 CB |
7878 | /* 64 bit floating points memory transfers are paired single precision loads |
7879 | or store. So DWARF information needs fixing in little endian (unless | |
7880 | PR=SZ=1 in FPSCR). */ | |
7881 | rtx | |
7882 | sh_dwarf_register_span (rtx reg) | |
7883 | { | |
7884 | unsigned regno = REGNO (reg); | |
7885 | ||
7886 | if (WORDS_BIG_ENDIAN || GET_MODE (reg) != DFmode) | |
7887 | return NULL_RTX; | |
7888 | ||
7889 | return | |
7890 | gen_rtx_PARALLEL (VOIDmode, | |
7891 | gen_rtvec (2, | |
7892 | gen_rtx_REG (SFmode, | |
7893 | DBX_REGISTER_NUMBER (regno+1)), | |
7894 | gen_rtx_REG (SFmode, | |
7895 | DBX_REGISTER_NUMBER (regno)))); | |
7896 | } | |
7897 | ||
150e407a | 7898 | bool |
586de218 | 7899 | sh_promote_prototypes (const_tree type) |
61f71b34 DD |
7900 | { |
7901 | if (TARGET_HITACHI) | |
7902 | return 0; | |
7903 | if (! type) | |
7904 | return 1; | |
7905 | return ! sh_attr_renesas_p (type); | |
7906 | } | |
7907 | ||
8cd5a4e0 RH |
7908 | /* Whether an argument must be passed by reference. On SHcompact, we |
7909 | pretend arguments wider than 32-bits that would have been passed in | |
7910 | registers are passed by reference, so that an SHmedia trampoline | |
7911 | loads them into the full 64-bits registers. */ | |
7912 | ||
7913 | static int | |
586de218 KG |
7914 | shcompact_byref (const CUMULATIVE_ARGS *cum, enum machine_mode mode, |
7915 | const_tree type, bool named) | |
8cd5a4e0 RH |
7916 | { |
7917 | unsigned HOST_WIDE_INT size; | |
7918 | ||
7919 | if (type) | |
7920 | size = int_size_in_bytes (type); | |
7921 | else | |
7922 | size = GET_MODE_SIZE (mode); | |
7923 | ||
7924 | if (cum->arg_count[SH_ARG_INT] < NPARM_REGS (SImode) | |
7925 | && (!named | |
7926 | || GET_SH_ARG_CLASS (mode) == SH_ARG_INT | |
7927 | || (GET_SH_ARG_CLASS (mode) == SH_ARG_FLOAT | |
7928 | && cum->arg_count[SH_ARG_FLOAT] >= NPARM_REGS (SFmode))) | |
7929 | && size > 4 | |
7930 | && !SHCOMPACT_FORCE_ON_STACK (mode, type) | |
7931 | && !SH5_WOULD_BE_PARTIAL_NREGS (*cum, mode, type, named)) | |
7932 | return size; | |
7933 | else | |
7934 | return 0; | |
7935 | } | |
7936 | ||
7937 | static bool | |
7938 | sh_pass_by_reference (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
586de218 | 7939 | const_tree type, bool named) |
8cd5a4e0 RH |
7940 | { |
7941 | if (targetm.calls.must_pass_in_stack (mode, type)) | |
7942 | return true; | |
7943 | ||
03603192 JR |
7944 | /* ??? std_gimplify_va_arg_expr passes NULL for cum. That function |
7945 | wants to know about pass-by-reference semantics for incoming | |
7946 | arguments. */ | |
7947 | if (! cum) | |
7948 | return false; | |
7949 | ||
8cd5a4e0 RH |
7950 | if (TARGET_SHCOMPACT) |
7951 | { | |
7952 | cum->byref = shcompact_byref (cum, mode, type, named); | |
7953 | return cum->byref != 0; | |
7954 | } | |
7955 | ||
7956 | return false; | |
7957 | } | |
7958 | ||
6cdd5672 RH |
7959 | static bool |
7960 | sh_callee_copies (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
586de218 | 7961 | const_tree type, bool named ATTRIBUTE_UNUSED) |
6cdd5672 RH |
7962 | { |
7963 | /* ??? How can it possibly be correct to return true only on the | |
7964 | caller side of the equation? Is there someplace else in the | |
7965 | sh backend that's magically producing the copies? */ | |
7966 | return (cum->outgoing | |
7967 | && ((mode == BLKmode ? TYPE_ALIGN (type) : GET_MODE_ALIGNMENT (mode)) | |
7968 | % SH_MIN_ALIGN_FOR_CALLEE_COPY == 0)); | |
7969 | } | |
7970 | ||
78a52f11 RH |
7971 | static int |
7972 | sh_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
7973 | tree type, bool named ATTRIBUTE_UNUSED) | |
7974 | { | |
7975 | int words = 0; | |
7976 | ||
7977 | if (!TARGET_SH5 | |
7978 | && PASS_IN_REG_P (*cum, mode, type) | |
7979 | && !(TARGET_SH4 || TARGET_SH2A_DOUBLE) | |
7980 | && (ROUND_REG (*cum, mode) | |
7981 | + (mode != BLKmode | |
7982 | ? ROUND_ADVANCE (GET_MODE_SIZE (mode)) | |
7983 | : ROUND_ADVANCE (int_size_in_bytes (type))) | |
7984 | > NPARM_REGS (mode))) | |
7985 | words = NPARM_REGS (mode) - ROUND_REG (*cum, mode); | |
7986 | ||
7987 | else if (!TARGET_SHCOMPACT | |
7988 | && SH5_WOULD_BE_PARTIAL_NREGS (*cum, mode, type, named)) | |
7989 | words = NPARM_REGS (SImode) - cum->arg_count[SH_ARG_INT]; | |
7990 | ||
7991 | return words * UNITS_PER_WORD; | |
7992 | } | |
7993 | ||
7994 | ||
61f71b34 DD |
7995 | /* Define where to put the arguments to a function. |
7996 | Value is zero to push the argument on the stack, | |
7997 | or a hard register in which to store the argument. | |
7998 | ||
7999 | MODE is the argument's machine mode. | |
8000 | TYPE is the data type of the argument (as a tree). | |
8001 | This is null for libcalls where that information may | |
8002 | not be available. | |
8003 | CUM is a variable of type CUMULATIVE_ARGS which gives info about | |
8004 | the preceding args and about the function being called. | |
8005 | NAMED is nonzero if this argument is a named parameter | |
8006 | (otherwise it is an extra parameter matching an ellipsis). | |
8007 | ||
8008 | On SH the first args are normally in registers | |
8009 | and the rest are pushed. Any arg that starts within the first | |
8010 | NPARM_REGS words is at least partially passed in a register unless | |
8011 | its data type forbids. */ | |
8012 | ||
8013 | ||
8014 | rtx | |
cf277499 SB |
8015 | sh_function_arg (CUMULATIVE_ARGS *ca, enum machine_mode mode, |
8016 | tree type, int named) | |
61f71b34 DD |
8017 | { |
8018 | if (! TARGET_SH5 && mode == VOIDmode) | |
8019 | return GEN_INT (ca->renesas_abi ? 1 : 0); | |
8020 | ||
8021 | if (! TARGET_SH5 | |
8022 | && PASS_IN_REG_P (*ca, mode, type) | |
8023 | && (named || ! (TARGET_HITACHI || ca->renesas_abi))) | |
8024 | { | |
8025 | int regno; | |
8026 | ||
8027 | if (mode == SCmode && TARGET_SH4 && TARGET_LITTLE_ENDIAN | |
8028 | && (! FUNCTION_ARG_SCmode_WART || (ROUND_REG (*ca, mode) & 1))) | |
8029 | { | |
8030 | rtx r1 = gen_rtx_EXPR_LIST (VOIDmode, | |
8031 | gen_rtx_REG (SFmode, | |
8032 | BASE_ARG_REG (mode) | |
8033 | + (ROUND_REG (*ca, mode) ^ 1)), | |
8034 | const0_rtx); | |
832a3292 KH |
8035 | rtx r2 = gen_rtx_EXPR_LIST (VOIDmode, |
8036 | gen_rtx_REG (SFmode, | |
8037 | BASE_ARG_REG (mode) | |
8038 | + ((ROUND_REG (*ca, mode) + 1) ^ 1)), | |
8039 | GEN_INT (4)); | |
61f71b34 DD |
8040 | return gen_rtx_PARALLEL(SCmode, gen_rtvec(2, r1, r2)); |
8041 | } | |
8042 | ||
8043 | /* If the alignment of a DF value causes an SF register to be | |
8044 | skipped, we will use that skipped register for the next SF | |
8045 | value. */ | |
8046 | if ((TARGET_HITACHI || ca->renesas_abi) | |
8047 | && ca->free_single_fp_reg | |
8048 | && mode == SFmode) | |
8049 | return gen_rtx_REG (mode, ca->free_single_fp_reg); | |
8050 | ||
8051 | regno = (BASE_ARG_REG (mode) + ROUND_REG (*ca, mode)) | |
8052 | ^ (mode == SFmode && TARGET_SH4 | |
8053 | && TARGET_LITTLE_ENDIAN != 0 | |
8054 | && ! TARGET_HITACHI && ! ca->renesas_abi); | |
8055 | return gen_rtx_REG (mode, regno); | |
8056 | ||
8057 | } | |
f676971a | 8058 | |
61f71b34 DD |
8059 | if (TARGET_SH5) |
8060 | { | |
8061 | if (mode == VOIDmode && TARGET_SHCOMPACT) | |
8062 | return GEN_INT (ca->call_cookie); | |
8063 | ||
8064 | /* The following test assumes unnamed arguments are promoted to | |
8065 | DFmode. */ | |
8066 | if (mode == SFmode && ca->free_single_fp_reg) | |
8067 | return SH5_PROTOTYPED_FLOAT_ARG (*ca, mode, ca->free_single_fp_reg); | |
8068 | ||
8069 | if ((GET_SH_ARG_CLASS (mode) == SH_ARG_FLOAT) | |
8070 | && (named || ! ca->prototype_p) | |
8071 | && ca->arg_count[(int) SH_ARG_FLOAT] < NPARM_REGS (SFmode)) | |
8072 | { | |
8073 | if (! ca->prototype_p && TARGET_SHMEDIA) | |
8074 | return SH5_PROTOTYPELESS_FLOAT_ARG (*ca, mode); | |
8075 | ||
8076 | return SH5_PROTOTYPED_FLOAT_ARG (*ca, mode, | |
8077 | FIRST_FP_PARM_REG | |
8078 | + ca->arg_count[(int) SH_ARG_FLOAT]); | |
8079 | } | |
8080 | ||
8081 | if (ca->arg_count[(int) SH_ARG_INT] < NPARM_REGS (SImode) | |
8082 | && (! TARGET_SHCOMPACT | |
8083 | || (! SHCOMPACT_FORCE_ON_STACK (mode, type) | |
8084 | && ! SH5_WOULD_BE_PARTIAL_NREGS (*ca, mode, | |
8085 | type, named)))) | |
8086 | { | |
8087 | return gen_rtx_REG (mode, (FIRST_PARM_REG | |
8088 | + ca->arg_count[(int) SH_ARG_INT])); | |
8089 | } | |
8090 | ||
8091 | return 0; | |
8092 | } | |
8093 | ||
8094 | return 0; | |
8095 | } | |
f676971a | 8096 | |
61f71b34 DD |
8097 | /* Update the data in CUM to advance over an argument |
8098 | of mode MODE and data type TYPE. | |
8099 | (TYPE is null for libcalls where that information may not be | |
8100 | available.) */ | |
8101 | ||
8102 | void | |
cf277499 SB |
8103 | sh_function_arg_advance (CUMULATIVE_ARGS *ca, enum machine_mode mode, |
8104 | tree type, int named) | |
61f71b34 | 8105 | { |
832a3292 KH |
8106 | if (ca->force_mem) |
8107 | ca->force_mem = 0; | |
8108 | else if (TARGET_SH5) | |
8109 | { | |
8110 | tree type2 = (ca->byref && type | |
8111 | ? TREE_TYPE (type) | |
8112 | : type); | |
8113 | enum machine_mode mode2 = (ca->byref && type | |
8114 | ? TYPE_MODE (type2) | |
8115 | : mode); | |
8116 | int dwords = ((ca->byref | |
8117 | ? ca->byref | |
8118 | : mode2 == BLKmode | |
8119 | ? int_size_in_bytes (type2) | |
8120 | : GET_MODE_SIZE (mode2)) + 7) / 8; | |
8121 | int numregs = MIN (dwords, NPARM_REGS (SImode) | |
8122 | - ca->arg_count[(int) SH_ARG_INT]); | |
8123 | ||
8124 | if (numregs) | |
8125 | { | |
8126 | ca->arg_count[(int) SH_ARG_INT] += numregs; | |
8127 | if (TARGET_SHCOMPACT | |
8128 | && SHCOMPACT_FORCE_ON_STACK (mode2, type2)) | |
8129 | { | |
8130 | ca->call_cookie | |
8131 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
8132 | - numregs, 1); | |
8133 | /* N.B. We want this also for outgoing. */ | |
8134 | ca->stack_regs += numregs; | |
8135 | } | |
8136 | else if (ca->byref) | |
8137 | { | |
8138 | if (! ca->outgoing) | |
8139 | ca->stack_regs += numregs; | |
8140 | ca->byref_regs += numregs; | |
8141 | ca->byref = 0; | |
8142 | do | |
8143 | ca->call_cookie | |
8144 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
8145 | - numregs, 2); | |
8146 | while (--numregs); | |
8147 | ca->call_cookie | |
8148 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
8149 | - 1, 1); | |
8150 | } | |
8151 | else if (dwords > numregs) | |
8152 | { | |
8153 | int pushregs = numregs; | |
8154 | ||
8155 | if (TARGET_SHCOMPACT) | |
8156 | ca->stack_regs += numregs; | |
8157 | while (pushregs < NPARM_REGS (SImode) - 1 | |
8158 | && (CALL_COOKIE_INT_REG_GET | |
8159 | (ca->call_cookie, | |
8160 | NPARM_REGS (SImode) - pushregs) | |
8161 | == 1)) | |
8162 | { | |
8163 | ca->call_cookie | |
8164 | &= ~ CALL_COOKIE_INT_REG (NPARM_REGS (SImode) | |
8165 | - pushregs, 1); | |
8166 | pushregs++; | |
8167 | } | |
8168 | if (numregs == NPARM_REGS (SImode)) | |
8169 | ca->call_cookie | |
8170 | |= CALL_COOKIE_INT_REG (0, 1) | |
8171 | | CALL_COOKIE_STACKSEQ (numregs - 1); | |
8172 | else | |
8173 | ca->call_cookie | |
8174 | |= CALL_COOKIE_STACKSEQ (numregs); | |
8175 | } | |
8176 | } | |
8177 | if (GET_SH_ARG_CLASS (mode2) == SH_ARG_FLOAT | |
8178 | && (named || ! ca->prototype_p)) | |
8179 | { | |
8180 | if (mode2 == SFmode && ca->free_single_fp_reg) | |
8181 | ca->free_single_fp_reg = 0; | |
8182 | else if (ca->arg_count[(int) SH_ARG_FLOAT] | |
8183 | < NPARM_REGS (SFmode)) | |
8184 | { | |
8185 | int numfpregs | |
8186 | = MIN ((GET_MODE_SIZE (mode2) + 7) / 8 * 2, | |
8187 | NPARM_REGS (SFmode) | |
8188 | - ca->arg_count[(int) SH_ARG_FLOAT]); | |
8189 | ||
8190 | ca->arg_count[(int) SH_ARG_FLOAT] += numfpregs; | |
8191 | ||
8192 | if (TARGET_SHCOMPACT && ! ca->prototype_p) | |
8193 | { | |
8194 | if (ca->outgoing && numregs > 0) | |
8195 | do | |
8196 | { | |
8197 | ca->call_cookie | |
8198 | |= (CALL_COOKIE_INT_REG | |
8199 | (ca->arg_count[(int) SH_ARG_INT] | |
8200 | - numregs + ((numfpregs - 2) / 2), | |
8201 | 4 + (ca->arg_count[(int) SH_ARG_FLOAT] | |
8202 | - numfpregs) / 2)); | |
8203 | } | |
8204 | while (numfpregs -= 2); | |
8205 | } | |
8206 | else if (mode2 == SFmode && (named) | |
8207 | && (ca->arg_count[(int) SH_ARG_FLOAT] | |
8208 | < NPARM_REGS (SFmode))) | |
8209 | ca->free_single_fp_reg | |
8210 | = FIRST_FP_PARM_REG - numfpregs | |
8211 | + ca->arg_count[(int) SH_ARG_FLOAT] + 1; | |
8212 | } | |
8213 | } | |
8214 | return; | |
8215 | } | |
8216 | ||
8217 | if ((TARGET_HITACHI || ca->renesas_abi) && TARGET_FPU_DOUBLE) | |
8218 | { | |
8219 | /* Note that we've used the skipped register. */ | |
8220 | if (mode == SFmode && ca->free_single_fp_reg) | |
8221 | { | |
8222 | ca->free_single_fp_reg = 0; | |
8223 | return; | |
8224 | } | |
8225 | /* When we have a DF after an SF, there's an SF register that get | |
8226 | skipped in order to align the DF value. We note this skipped | |
8227 | register, because the next SF value will use it, and not the | |
8228 | SF that follows the DF. */ | |
8229 | if (mode == DFmode | |
8230 | && ROUND_REG (*ca, DFmode) != ROUND_REG (*ca, SFmode)) | |
8231 | { | |
8232 | ca->free_single_fp_reg = (ROUND_REG (*ca, SFmode) | |
8233 | + BASE_ARG_REG (mode)); | |
8234 | } | |
8235 | } | |
8236 | ||
157371cf | 8237 | if (! ((TARGET_SH4 || TARGET_SH2A) || ca->renesas_abi) |
832a3292 KH |
8238 | || PASS_IN_REG_P (*ca, mode, type)) |
8239 | (ca->arg_count[(int) GET_SH_ARG_CLASS (mode)] | |
8240 | = (ROUND_REG (*ca, mode) | |
8241 | + (mode == BLKmode | |
8242 | ? ROUND_ADVANCE (int_size_in_bytes (type)) | |
8243 | : ROUND_ADVANCE (GET_MODE_SIZE (mode))))); | |
61f71b34 DD |
8244 | } |
8245 | ||
61f71b34 DD |
8246 | /* The Renesas calling convention doesn't quite fit into this scheme since |
8247 | the address is passed like an invisible argument, but one that is always | |
8248 | passed in memory. */ | |
8249 | static rtx | |
cf277499 | 8250 | sh_struct_value_rtx (tree fndecl, int incoming ATTRIBUTE_UNUSED) |
61f71b34 DD |
8251 | { |
8252 | if (TARGET_HITACHI || sh_attr_renesas_p (fndecl)) | |
8253 | return 0; | |
8254 | return gen_rtx_REG (Pmode, 2); | |
8255 | } | |
8256 | ||
bd5bd7ac KH |
8257 | /* Worker function for TARGET_RETURN_IN_MEMORY. */ |
8258 | ||
61f71b34 | 8259 | static bool |
586de218 | 8260 | sh_return_in_memory (const_tree type, const_tree fndecl) |
61f71b34 DD |
8261 | { |
8262 | if (TARGET_SH5) | |
8263 | { | |
8264 | if (TYPE_MODE (type) == BLKmode) | |
8265 | return ((unsigned HOST_WIDE_INT) int_size_in_bytes (type)) > 8; | |
8266 | else | |
8267 | return GET_MODE_SIZE (TYPE_MODE (type)) > 8; | |
8268 | } | |
8269 | else | |
8270 | { | |
8271 | return (TYPE_MODE (type) == BLKmode | |
8272 | || ((TARGET_HITACHI || sh_attr_renesas_p (fndecl)) | |
8273 | && TREE_CODE (type) == RECORD_TYPE)); | |
8274 | } | |
8275 | } | |
8276 | ||
8277 | /* We actually emit the code in sh_expand_prologue. We used to use | |
8278 | a static variable to flag that we need to emit this code, but that | |
8279 | doesn't when inlining, when functions are deferred and then emitted | |
8280 | later. Fortunately, we already have two flags that are part of struct | |
8281 | function that tell if a function uses varargs or stdarg. */ | |
8282 | static void | |
49616835 JR |
8283 | sh_setup_incoming_varargs (CUMULATIVE_ARGS *ca, |
8284 | enum machine_mode mode, | |
8285 | tree type, | |
8286 | int *pretend_arg_size, | |
cf277499 | 8287 | int second_time ATTRIBUTE_UNUSED) |
61f71b34 | 8288 | { |
e3b5732b | 8289 | gcc_assert (cfun->stdarg); |
49616835 JR |
8290 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl)) |
8291 | { | |
8292 | int named_parm_regs, anon_parm_regs; | |
8293 | ||
8294 | named_parm_regs = (ROUND_REG (*ca, mode) | |
8295 | + (mode == BLKmode | |
8296 | ? ROUND_ADVANCE (int_size_in_bytes (type)) | |
8297 | : ROUND_ADVANCE (GET_MODE_SIZE (mode)))); | |
8298 | anon_parm_regs = NPARM_REGS (SImode) - named_parm_regs; | |
8299 | if (anon_parm_regs > 0) | |
8300 | *pretend_arg_size = anon_parm_regs * 4; | |
8301 | } | |
61f71b34 DD |
8302 | } |
8303 | ||
8304 | static bool | |
cf277499 | 8305 | sh_strict_argument_naming (CUMULATIVE_ARGS *ca ATTRIBUTE_UNUSED) |
61f71b34 DD |
8306 | { |
8307 | return TARGET_SH5; | |
8308 | } | |
8309 | ||
8310 | static bool | |
cf277499 | 8311 | sh_pretend_outgoing_varargs_named (CUMULATIVE_ARGS *ca) |
61f71b34 DD |
8312 | { |
8313 | return ! (TARGET_HITACHI || ca->renesas_abi) && ! TARGET_SH5; | |
8314 | } | |
8315 | ||
8316 | ||
0d7e008e SC |
8317 | /* Define the offset between two registers, one to be eliminated, and |
8318 | the other its replacement, at the start of a routine. */ | |
8319 | ||
8320 | int | |
cf277499 | 8321 | initial_elimination_offset (int from, int to) |
0d7e008e SC |
8322 | { |
8323 | int regs_saved; | |
fa5322fa | 8324 | int regs_saved_rounding = 0; |
0d7e008e | 8325 | int total_saved_regs_space; |
e5f8d4c0 | 8326 | int total_auto_space; |
1245df60 | 8327 | int save_flags = target_flags; |
fa5322fa | 8328 | int copy_flags; |
490dbac7 | 8329 | HARD_REG_SET live_regs_mask; |
fe3ad572 SC |
8330 | |
8331 | shmedia_space_reserved_for_target_registers = false; | |
490dbac7 | 8332 | regs_saved = calc_live_regs (&live_regs_mask); |
fa5322fa | 8333 | regs_saved += SHMEDIA_REGS_STACK_ADJUST (); |
fe3ad572 SC |
8334 | |
8335 | if (shmedia_reserve_space_for_target_registers_p (regs_saved, &live_regs_mask)) | |
8336 | { | |
8337 | shmedia_space_reserved_for_target_registers = true; | |
8338 | regs_saved += shmedia_target_regs_stack_adjust (&live_regs_mask); | |
8339 | } | |
8340 | ||
fa5322fa AO |
8341 | if (TARGET_SH5 && regs_saved % (STACK_BOUNDARY / BITS_PER_UNIT)) |
8342 | regs_saved_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) | |
8343 | - regs_saved % (STACK_BOUNDARY / BITS_PER_UNIT)); | |
8e87e161 | 8344 | |
fa5322fa AO |
8345 | total_auto_space = rounded_frame_size (regs_saved) - regs_saved_rounding; |
8346 | copy_flags = target_flags; | |
1245df60 | 8347 | target_flags = save_flags; |
1a95a963 | 8348 | |
fa5322fa | 8349 | total_saved_regs_space = regs_saved + regs_saved_rounding; |
b9654711 | 8350 | |
96a2347e | 8351 | if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) |
fa5322fa | 8352 | return total_saved_regs_space + total_auto_space |
38173d38 | 8353 | + crtl->args.info.byref_regs * 8; |
8aa2a305 | 8354 | |
0d7e008e | 8355 | if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
fa5322fa | 8356 | return total_saved_regs_space + total_auto_space |
38173d38 | 8357 | + crtl->args.info.byref_regs * 8; |
8aa2a305 JW |
8358 | |
8359 | /* Initial gap between fp and sp is 0. */ | |
96a2347e | 8360 | if (from == HARD_FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
8aa2a305 JW |
8361 | return 0; |
8362 | ||
96a2347e KK |
8363 | if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
8364 | return rounded_frame_size (0); | |
8365 | ||
8366 | if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
8367 | return rounded_frame_size (0); | |
8368 | ||
f5b9e7c9 | 8369 | gcc_assert (from == RETURN_ADDRESS_POINTER_REGNUM |
96a2347e KK |
8370 | && (to == HARD_FRAME_POINTER_REGNUM |
8371 | || to == STACK_POINTER_REGNUM)); | |
f5b9e7c9 | 8372 | if (TARGET_SH5) |
d042370e | 8373 | { |
f5b9e7c9 NS |
8374 | int n = total_saved_regs_space; |
8375 | int pr_reg = TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG; | |
8376 | save_schedule schedule; | |
8377 | save_entry *entry; | |
8378 | ||
8379 | n += total_auto_space; | |
8380 | ||
8381 | /* If it wasn't saved, there's not much we can do. */ | |
8382 | if (! TEST_HARD_REG_BIT (live_regs_mask, pr_reg)) | |
8383 | return n; | |
8384 | ||
8385 | target_flags = copy_flags; | |
8386 | ||
8387 | sh5_schedule_saves (&live_regs_mask, &schedule, n); | |
8388 | for (entry = &schedule.entries[1]; entry->mode != VOIDmode; entry++) | |
8389 | if (entry->reg == pr_reg) | |
8390 | { | |
8391 | target_flags = save_flags; | |
8392 | return entry->offset; | |
8393 | } | |
8394 | gcc_unreachable (); | |
d042370e | 8395 | } |
f5b9e7c9 NS |
8396 | else |
8397 | return total_auto_space; | |
0d7e008e | 8398 | } |
3217af3e KK |
8399 | |
8400 | /* Parse the -mfixed-range= option string. */ | |
8401 | void | |
8402 | sh_fix_range (const char *const_str) | |
8403 | { | |
8404 | int i, first, last; | |
8405 | char *str, *dash, *comma; | |
8406 | ||
8407 | /* str must be of the form REG1'-'REG2{,REG1'-'REG} where REG1 and | |
8408 | REG2 are either register names or register numbers. The effect | |
8409 | of this option is to mark the registers in the range from REG1 to | |
8410 | REG2 as ``fixed'' so they won't be used by the compiler. */ | |
8411 | ||
8412 | i = strlen (const_str); | |
8413 | str = (char *) alloca (i + 1); | |
8414 | memcpy (str, const_str, i + 1); | |
8415 | ||
8416 | while (1) | |
8417 | { | |
8418 | dash = strchr (str, '-'); | |
8419 | if (!dash) | |
8420 | { | |
8421 | warning (0, "value of -mfixed-range must have form REG1-REG2"); | |
8422 | return; | |
8423 | } | |
8424 | *dash = '\0'; | |
8425 | comma = strchr (dash + 1, ','); | |
8426 | if (comma) | |
8427 | *comma = '\0'; | |
8428 | ||
8429 | first = decode_reg_name (str); | |
8430 | if (first < 0) | |
8431 | { | |
8432 | warning (0, "unknown register name: %s", str); | |
8433 | return; | |
8434 | } | |
8435 | ||
8436 | last = decode_reg_name (dash + 1); | |
8437 | if (last < 0) | |
8438 | { | |
8439 | warning (0, "unknown register name: %s", dash + 1); | |
8440 | return; | |
8441 | } | |
8442 | ||
8443 | *dash = '-'; | |
8444 | ||
8445 | if (first > last) | |
8446 | { | |
8447 | warning (0, "%s-%s is an empty range", str, dash + 1); | |
8448 | return; | |
8449 | } | |
8450 | ||
8451 | for (i = first; i <= last; ++i) | |
8452 | fixed_regs[i] = call_used_regs[i] = 1; | |
8453 | ||
8454 | if (!comma) | |
8455 | break; | |
8456 | ||
8457 | *comma = ','; | |
8458 | str = comma + 1; | |
8459 | } | |
8460 | } | |
8aa2a305 | 8461 | \f |
a6ab9fc0 | 8462 | /* Insert any deferred function attributes from earlier pragmas. */ |
12a68f1f | 8463 | static void |
cf277499 | 8464 | sh_insert_attributes (tree node, tree *attributes) |
157131d7 | 8465 | { |
a6ab9fc0 R |
8466 | tree attrs; |
8467 | ||
8468 | if (TREE_CODE (node) != FUNCTION_DECL) | |
157131d7 SC |
8469 | return; |
8470 | ||
8471 | /* We are only interested in fields. */ | |
6615c446 | 8472 | if (!DECL_P (node)) |
157131d7 SC |
8473 | return; |
8474 | ||
a6ab9fc0 R |
8475 | /* Append the attributes to the deferred attributes. */ |
8476 | *sh_deferred_function_attributes_tail = *attributes; | |
8477 | attrs = sh_deferred_function_attributes; | |
8478 | if (!attrs) | |
8479 | return; | |
8480 | ||
8481 | /* Some attributes imply or require the interrupt attribute. */ | |
8482 | if (!lookup_attribute ("interrupt_handler", attrs) | |
8483 | && !lookup_attribute ("interrupt_handler", DECL_ATTRIBUTES (node))) | |
8484 | { | |
8485 | /* If we have a trapa_handler, but no interrupt_handler attribute, | |
8486 | insert an interrupt_handler attribute. */ | |
8487 | if (lookup_attribute ("trapa_handler", attrs) != NULL_TREE) | |
8488 | /* We can't use sh_pr_interrupt here because that's not in the | |
8489 | java frontend. */ | |
8490 | attrs | |
8491 | = tree_cons (get_identifier("interrupt_handler"), NULL_TREE, attrs); | |
561642fa AP |
8492 | /* However, for sp_switch, trap_exit, nosave_low_regs and resbank, |
8493 | if the interrupt attribute is missing, we ignore the attribute | |
8494 | and warn. */ | |
a6ab9fc0 R |
8495 | else if (lookup_attribute ("sp_switch", attrs) |
8496 | || lookup_attribute ("trap_exit", attrs) | |
561642fa AP |
8497 | || lookup_attribute ("nosave_low_regs", attrs) |
8498 | || lookup_attribute ("resbank", attrs)) | |
a6ab9fc0 R |
8499 | { |
8500 | tree *tail; | |
8501 | ||
8502 | for (tail = attributes; attrs; attrs = TREE_CHAIN (attrs)) | |
8503 | { | |
8504 | if (is_attribute_p ("sp_switch", TREE_PURPOSE (attrs)) | |
8505 | || is_attribute_p ("trap_exit", TREE_PURPOSE (attrs)) | |
561642fa AP |
8506 | || is_attribute_p ("nosave_low_regs", TREE_PURPOSE (attrs)) |
8507 | || is_attribute_p ("resbank", TREE_PURPOSE (attrs))) | |
a6ab9fc0 | 8508 | warning (OPT_Wattributes, |
29d08eba JM |
8509 | "%qE attribute only applies to interrupt functions", |
8510 | TREE_PURPOSE (attrs)); | |
a6ab9fc0 R |
8511 | else |
8512 | { | |
8513 | *tail = tree_cons (TREE_PURPOSE (attrs), NULL_TREE, | |
8514 | NULL_TREE); | |
8515 | tail = &TREE_CHAIN (*tail); | |
8516 | } | |
8517 | } | |
8518 | attrs = *attributes; | |
8519 | } | |
8520 | } | |
8521 | ||
8522 | /* Install the processed list. */ | |
8523 | *attributes = attrs; | |
8524 | ||
8525 | /* Clear deferred attributes. */ | |
8526 | sh_deferred_function_attributes = NULL_TREE; | |
8527 | sh_deferred_function_attributes_tail = &sh_deferred_function_attributes; | |
157131d7 SC |
8528 | |
8529 | return; | |
8530 | } | |
8531 | ||
91d231cb | 8532 | /* Supported attributes: |
4408efce JL |
8533 | |
8534 | interrupt_handler -- specifies this function is an interrupt handler. | |
8535 | ||
a6ab9fc0 R |
8536 | trapa_handler - like above, but don't save all registers. |
8537 | ||
4408efce JL |
8538 | sp_switch -- specifies an alternate stack for an interrupt handler |
8539 | to run on. | |
8540 | ||
956d6950 | 8541 | trap_exit -- use a trapa to exit an interrupt function instead of |
61f71b34 DD |
8542 | an rte instruction. |
8543 | ||
a6ab9fc0 R |
8544 | nosave_low_regs - don't save r0..r7 in an interrupt handler. |
8545 | This is useful on the SH3 and upwards, | |
8546 | which has a separate set of low regs for User and Supervisor modes. | |
8547 | This should only be used for the lowest level of interrupts. Higher levels | |
8548 | of interrupts must save the registers in case they themselves are | |
8549 | interrupted. | |
8550 | ||
61f71b34 DD |
8551 | renesas -- use Renesas calling/layout conventions (functions and |
8552 | structures). | |
8553 | ||
561642fa AP |
8554 | resbank -- In case of an ISR, use a register bank to save registers |
8555 | R0-R14, MACH, MACL, GBR and PR. This is useful only on SH2A targets. | |
61f71b34 | 8556 | */ |
4408efce | 8557 | |
561642fa AP |
8558 | /* Handle a 'resbank' attribute. */ |
8559 | static tree | |
8560 | sh_handle_resbank_handler_attribute (tree * node, tree name, | |
8561 | tree args ATTRIBUTE_UNUSED, | |
8562 | int flags ATTRIBUTE_UNUSED, | |
8563 | bool * no_add_attrs) | |
8564 | { | |
8565 | if (!TARGET_SH2A) | |
8566 | { | |
29d08eba JM |
8567 | warning (OPT_Wattributes, "%qE attribute is supported only for SH2A", |
8568 | name); | |
561642fa AP |
8569 | *no_add_attrs = true; |
8570 | } | |
8571 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
8572 | { | |
29d08eba JM |
8573 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
8574 | name); | |
561642fa AP |
8575 | *no_add_attrs = true; |
8576 | } | |
8577 | ||
8578 | return NULL_TREE; | |
8579 | } | |
8580 | ||
91d231cb JM |
8581 | /* Handle an "interrupt_handler" attribute; arguments as in |
8582 | struct attribute_spec.handler. */ | |
8583 | static tree | |
cf277499 | 8584 | sh_handle_interrupt_handler_attribute (tree *node, tree name, |
561642fa AP |
8585 | tree args ATTRIBUTE_UNUSED, |
8586 | int flags ATTRIBUTE_UNUSED, | |
8587 | bool *no_add_attrs) | |
91d231cb JM |
8588 | { |
8589 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
4408efce | 8590 | { |
29d08eba JM |
8591 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
8592 | name); | |
91d231cb | 8593 | *no_add_attrs = true; |
4408efce | 8594 | } |
58ab7171 R |
8595 | else if (TARGET_SHCOMPACT) |
8596 | { | |
8597 | error ("attribute interrupt_handler is not compatible with -m5-compact"); | |
8598 | *no_add_attrs = true; | |
8599 | } | |
4408efce | 8600 | |
91d231cb JM |
8601 | return NULL_TREE; |
8602 | } | |
8603 | ||
561642fa AP |
8604 | /* Handle an 'function_vector' attribute; arguments as in |
8605 | struct attribute_spec.handler. */ | |
8606 | static tree | |
8607 | sh2a_handle_function_vector_handler_attribute (tree * node, tree name, | |
8608 | tree args ATTRIBUTE_UNUSED, | |
8609 | int flags ATTRIBUTE_UNUSED, | |
8610 | bool * no_add_attrs) | |
8611 | { | |
8612 | if (!TARGET_SH2A) | |
8613 | { | |
29d08eba JM |
8614 | warning (OPT_Wattributes, "%qE attribute only applies to SH2A", |
8615 | name); | |
561642fa AP |
8616 | *no_add_attrs = true; |
8617 | } | |
8618 | else if (TREE_CODE (*node) != FUNCTION_DECL) | |
8619 | { | |
29d08eba JM |
8620 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
8621 | name); | |
561642fa AP |
8622 | *no_add_attrs = true; |
8623 | } | |
8624 | else if (TREE_CODE (TREE_VALUE (args)) != INTEGER_CST) | |
8625 | { | |
8626 | /* The argument must be a constant integer. */ | |
8627 | warning (OPT_Wattributes, | |
29d08eba JM |
8628 | "%qE attribute argument not an integer constant", |
8629 | name); | |
561642fa AP |
8630 | *no_add_attrs = true; |
8631 | } | |
8632 | else if (TREE_INT_CST_LOW (TREE_VALUE (args)) > 255) | |
8633 | { | |
8634 | /* The argument value must be between 0 to 255. */ | |
8635 | warning (OPT_Wattributes, | |
29d08eba JM |
8636 | "%qE attribute argument should be between 0 to 255", |
8637 | name); | |
561642fa AP |
8638 | *no_add_attrs = true; |
8639 | } | |
8640 | return NULL_TREE; | |
8641 | } | |
8642 | ||
8643 | /* Returns 1 if current function has been assigned the attribute | |
8644 | 'function_vector'. */ | |
8645 | int | |
8646 | sh2a_is_function_vector_call (rtx x) | |
8647 | { | |
8648 | if (GET_CODE (x) == SYMBOL_REF | |
8649 | && (SYMBOL_REF_FLAGS (x) & SYMBOL_FLAG_FUNCVEC_FUNCTION)) | |
8650 | { | |
8651 | tree tr = SYMBOL_REF_DECL (x); | |
8652 | ||
8653 | if (sh2a_function_vector_p (tr)) | |
8654 | return 1; | |
8655 | } | |
8656 | ||
8657 | return 0; | |
8658 | } | |
8659 | ||
8660 | /* Returns the function vector number, if the the attribute | |
8661 | 'function_vector' is assigned, otherwise returns zero. */ | |
8662 | int | |
8663 | sh2a_get_function_vector_number (rtx x) | |
8664 | { | |
8665 | int num; | |
8666 | tree list, t; | |
8667 | ||
8668 | if ((GET_CODE (x) == SYMBOL_REF) | |
8669 | && (SYMBOL_REF_FLAGS (x) & SYMBOL_FLAG_FUNCVEC_FUNCTION)) | |
8670 | { | |
8671 | t = SYMBOL_REF_DECL (x); | |
8672 | ||
8673 | if (TREE_CODE (t) != FUNCTION_DECL) | |
8674 | return 0; | |
8675 | ||
8676 | list = SH_ATTRIBUTES (t); | |
8677 | while (list) | |
8678 | { | |
8679 | if (is_attribute_p ("function_vector", TREE_PURPOSE (list))) | |
8680 | { | |
8681 | num = TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (list))); | |
8682 | return num; | |
8683 | } | |
8684 | ||
8685 | list = TREE_CHAIN (list); | |
8686 | } | |
8687 | ||
8688 | return 0; | |
8689 | } | |
8690 | else | |
8691 | return 0; | |
8692 | } | |
8693 | ||
91d231cb JM |
8694 | /* Handle an "sp_switch" attribute; arguments as in |
8695 | struct attribute_spec.handler. */ | |
8696 | static tree | |
cf277499 SB |
8697 | sh_handle_sp_switch_attribute (tree *node, tree name, tree args, |
8698 | int flags ATTRIBUTE_UNUSED, bool *no_add_attrs) | |
91d231cb JM |
8699 | { |
8700 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
8701 | { | |
29d08eba JM |
8702 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
8703 | name); | |
91d231cb JM |
8704 | *no_add_attrs = true; |
8705 | } | |
91d231cb JM |
8706 | else if (TREE_CODE (TREE_VALUE (args)) != STRING_CST) |
8707 | { | |
4408efce | 8708 | /* The argument must be a constant string. */ |
29d08eba JM |
8709 | warning (OPT_Wattributes, "%qE attribute argument not a string constant", |
8710 | name); | |
91d231cb JM |
8711 | *no_add_attrs = true; |
8712 | } | |
4408efce | 8713 | |
91d231cb JM |
8714 | return NULL_TREE; |
8715 | } | |
8716 | ||
8717 | /* Handle an "trap_exit" attribute; arguments as in | |
8718 | struct attribute_spec.handler. */ | |
8719 | static tree | |
cf277499 SB |
8720 | sh_handle_trap_exit_attribute (tree *node, tree name, tree args, |
8721 | int flags ATTRIBUTE_UNUSED, bool *no_add_attrs) | |
91d231cb JM |
8722 | { |
8723 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
8724 | { | |
29d08eba JM |
8725 | warning (OPT_Wattributes, "%qE attribute only applies to functions", |
8726 | name); | |
91d231cb JM |
8727 | *no_add_attrs = true; |
8728 | } | |
a6ab9fc0 R |
8729 | /* The argument specifies a trap number to be used in a trapa instruction |
8730 | at function exit (instead of an rte instruction). */ | |
91d231cb JM |
8731 | else if (TREE_CODE (TREE_VALUE (args)) != INTEGER_CST) |
8732 | { | |
4408efce | 8733 | /* The argument must be a constant integer. */ |
29d08eba JM |
8734 | warning (OPT_Wattributes, "%qE attribute argument not an " |
8735 | "integer constant", name); | |
91d231cb JM |
8736 | *no_add_attrs = true; |
8737 | } | |
318881c0 | 8738 | |
91d231cb | 8739 | return NULL_TREE; |
4408efce JL |
8740 | } |
8741 | ||
61f71b34 | 8742 | static tree |
cf277499 SB |
8743 | sh_handle_renesas_attribute (tree *node ATTRIBUTE_UNUSED, |
8744 | tree name ATTRIBUTE_UNUSED, | |
8745 | tree args ATTRIBUTE_UNUSED, | |
8746 | int flags ATTRIBUTE_UNUSED, | |
8747 | bool *no_add_attrs ATTRIBUTE_UNUSED) | |
61f71b34 DD |
8748 | { |
8749 | return NULL_TREE; | |
8750 | } | |
8751 | ||
8752 | /* True if __attribute__((renesas)) or -mrenesas. */ | |
8753 | int | |
586de218 | 8754 | sh_attr_renesas_p (const_tree td) |
61f71b34 DD |
8755 | { |
8756 | if (TARGET_HITACHI) | |
8757 | return 1; | |
8758 | if (td == 0) | |
8759 | return 0; | |
8760 | if (DECL_P (td)) | |
8761 | td = TREE_TYPE (td); | |
a11c004b R |
8762 | if (td == error_mark_node) |
8763 | return 0; | |
61f71b34 DD |
8764 | return (lookup_attribute ("renesas", TYPE_ATTRIBUTES (td)) |
8765 | != NULL_TREE); | |
8766 | } | |
8767 | ||
8768 | /* True if __attribute__((renesas)) or -mrenesas, for the current | |
8769 | function. */ | |
8770 | int | |
cf277499 | 8771 | sh_cfun_attr_renesas_p (void) |
61f71b34 DD |
8772 | { |
8773 | return sh_attr_renesas_p (current_function_decl); | |
8774 | } | |
8775 | ||
e3ba8d11 | 8776 | int |
cf277499 | 8777 | sh_cfun_interrupt_handler_p (void) |
e3ba8d11 R |
8778 | { |
8779 | return (lookup_attribute ("interrupt_handler", | |
8780 | DECL_ATTRIBUTES (current_function_decl)) | |
8781 | != NULL_TREE); | |
8782 | } | |
bcc8cc82 | 8783 | |
561642fa AP |
8784 | /* Returns 1 if FUNC has been assigned the attribute |
8785 | "function_vector". */ | |
8786 | int | |
8787 | sh2a_function_vector_p (tree func) | |
8788 | { | |
8789 | tree list; | |
8790 | if (TREE_CODE (func) != FUNCTION_DECL) | |
8791 | return 0; | |
8792 | ||
8793 | list = SH_ATTRIBUTES (func); | |
8794 | while (list) | |
8795 | { | |
8796 | if (is_attribute_p ("function_vector", TREE_PURPOSE (list))) | |
8797 | return 1; | |
8798 | ||
8799 | list = TREE_CHAIN (list); | |
8800 | } | |
8801 | return 0; | |
8802 | } | |
8803 | ||
8804 | /* Returns TRUE if given tree has the "resbank" attribute. */ | |
8805 | ||
8806 | int | |
8807 | sh_cfun_resbank_handler_p (void) | |
8808 | { | |
8809 | return ((lookup_attribute ("resbank", | |
8810 | DECL_ATTRIBUTES (current_function_decl)) | |
8811 | != NULL_TREE) | |
8812 | && (lookup_attribute ("interrupt_handler", | |
8813 | DECL_ATTRIBUTES (current_function_decl)) | |
8814 | != NULL_TREE) && TARGET_SH2A); | |
8815 | } | |
8816 | ||
8d932be3 | 8817 | /* Implement TARGET_CHECK_PCH_TARGET_FLAGS. */ |
73a4d10b | 8818 | |
8d932be3 RS |
8819 | static const char * |
8820 | sh_check_pch_target_flags (int old_flags) | |
8821 | { | |
c0fb94d7 RS |
8822 | if ((old_flags ^ target_flags) & (MASK_SH1 | MASK_SH2 | MASK_SH3 |
8823 | | MASK_SH_E | MASK_HARD_SH4 | |
8824 | | MASK_FPU_SINGLE | MASK_SH4)) | |
8825 | return _("created and used with different architectures / ABIs"); | |
8826 | if ((old_flags ^ target_flags) & MASK_HITACHI) | |
8827 | return _("created and used with different ABIs"); | |
8828 | if ((old_flags ^ target_flags) & MASK_LITTLE_ENDIAN) | |
8829 | return _("created and used with different endianness"); | |
bcc8cc82 | 8830 | return NULL; |
bcc8cc82 | 8831 | } |
0d7e008e | 8832 | \f |
8aa2a305 | 8833 | /* Predicates used by the templates. */ |
0d7e008e | 8834 | |
8aa2a305 JW |
8835 | /* Returns 1 if OP is MACL, MACH or PR. The input must be a REG rtx. |
8836 | Used only in general_movsrc_operand. */ | |
0d7e008e | 8837 | |
8aa2a305 | 8838 | int |
cf277499 | 8839 | system_reg_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
0d7e008e | 8840 | { |
8aa2a305 | 8841 | switch (REGNO (op)) |
0d7e008e | 8842 | { |
8aa2a305 JW |
8843 | case PR_REG: |
8844 | case MACL_REG: | |
8845 | case MACH_REG: | |
8846 | return 1; | |
0d7e008e | 8847 | } |
8aa2a305 | 8848 | return 0; |
0d7e008e | 8849 | } |
0d7e008e | 8850 | |
1a95a963 JW |
8851 | /* Nonzero if OP is a floating point value with value 0.0. */ |
8852 | ||
8853 | int | |
cf277499 | 8854 | fp_zero_operand (rtx op) |
d3ae8277 | 8855 | { |
1a95a963 | 8856 | REAL_VALUE_TYPE r; |
d3ae8277 | 8857 | |
1a95a963 JW |
8858 | if (GET_MODE (op) != SFmode) |
8859 | return 0; | |
8860 | ||
8861 | REAL_VALUE_FROM_CONST_DOUBLE (r, op); | |
e4fa6b06 | 8862 | return REAL_VALUES_EQUAL (r, dconst0) && ! REAL_VALUE_MINUS_ZERO (r); |
d3ae8277 SC |
8863 | } |
8864 | ||
1a95a963 | 8865 | /* Nonzero if OP is a floating point value with value 1.0. */ |
d3ae8277 SC |
8866 | |
8867 | int | |
cf277499 | 8868 | fp_one_operand (rtx op) |
d3ae8277 | 8869 | { |
1a95a963 JW |
8870 | REAL_VALUE_TYPE r; |
8871 | ||
8872 | if (GET_MODE (op) != SFmode) | |
8873 | return 0; | |
8874 | ||
8875 | REAL_VALUE_FROM_CONST_DOUBLE (r, op); | |
8876 | return REAL_VALUES_EQUAL (r, dconst1); | |
d3ae8277 | 8877 | } |
1245df60 | 8878 | |
9f09b1f2 R |
8879 | /* For -m4 and -m4-single-only, mode switching is used. If we are |
8880 | compiling without -mfmovd, movsf_ie isn't taken into account for | |
8881 | mode switching. We could check in machine_dependent_reorg for | |
8882 | cases where we know we are in single precision mode, but there is | |
8883 | interface to find that out during reload, so we must avoid | |
8884 | choosing an fldi alternative during reload and thus failing to | |
8885 | allocate a scratch register for the constant loading. */ | |
8886 | int | |
cf277499 | 8887 | fldi_ok (void) |
9f09b1f2 R |
8888 | { |
8889 | return ! TARGET_SH4 || TARGET_FMOVD || reload_completed; | |
8890 | } | |
8891 | ||
225e4f43 | 8892 | int |
cf277499 | 8893 | tertiary_reload_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
225e4f43 R |
8894 | { |
8895 | enum rtx_code code = GET_CODE (op); | |
8896 | return code == MEM || (TARGET_SH4 && code == CONST_DOUBLE); | |
8897 | } | |
8898 | ||
463f02cd | 8899 | /* Return the TLS type for TLS symbols, 0 for otherwise. */ |
81f40b79 | 8900 | enum tls_model |
cf277499 | 8901 | tls_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
463f02cd | 8902 | { |
463f02cd | 8903 | if (GET_CODE (op) != SYMBOL_REF) |
81f40b79 | 8904 | return TLS_MODEL_NONE; |
675ff4c7 | 8905 | return SYMBOL_REF_TLS_MODEL (op); |
463f02cd | 8906 | } |
1245df60 | 8907 | \f |
33f7f353 | 8908 | /* Return the destination address of a branch. */ |
f676971a | 8909 | |
318881c0 | 8910 | static int |
cf277499 | 8911 | branch_dest (rtx branch) |
1245df60 | 8912 | { |
33f7f353 JR |
8913 | rtx dest = SET_SRC (PATTERN (branch)); |
8914 | int dest_uid; | |
1245df60 R |
8915 | |
8916 | if (GET_CODE (dest) == IF_THEN_ELSE) | |
8917 | dest = XEXP (dest, 1); | |
8918 | dest = XEXP (dest, 0); | |
8919 | dest_uid = INSN_UID (dest); | |
9d98a694 | 8920 | return INSN_ADDRESSES (dest_uid); |
1245df60 | 8921 | } |
a55e9d2b | 8922 | \f |
5e7a8ee0 | 8923 | /* Return nonzero if REG is not used after INSN. |
a55e9d2b | 8924 | We assume REG is a reload reg, and therefore does |
8b760293 | 8925 | not live past labels. It may live past calls or jumps though. */ |
a55e9d2b | 8926 | int |
cf277499 | 8927 | reg_unused_after (rtx reg, rtx insn) |
a55e9d2b | 8928 | { |
8783b15e | 8929 | enum rtx_code code; |
a55e9d2b RK |
8930 | rtx set; |
8931 | ||
8932 | /* If the reg is set by this instruction, then it is safe for our | |
8933 | case. Disregard the case where this is a store to memory, since | |
8934 | we are checking a register used in the store address. */ | |
8935 | set = single_set (insn); | |
f3536097 | 8936 | if (set && !MEM_P (SET_DEST (set)) |
a55e9d2b RK |
8937 | && reg_overlap_mentioned_p (reg, SET_DEST (set))) |
8938 | return 1; | |
8939 | ||
318881c0 | 8940 | while ((insn = NEXT_INSN (insn))) |
a55e9d2b | 8941 | { |
ec8e098d PB |
8942 | rtx set; |
8943 | if (!INSN_P (insn)) | |
8944 | continue; | |
8945 | ||
a55e9d2b | 8946 | code = GET_CODE (insn); |
8783b15e | 8947 | |
c8f6f18d RK |
8948 | #if 0 |
8949 | /* If this is a label that existed before reload, then the register | |
8950 | if dead here. However, if this is a label added by reorg, then | |
8951 | the register may still be live here. We can't tell the difference, | |
8952 | so we just ignore labels completely. */ | |
8783b15e | 8953 | if (code == CODE_LABEL) |
a55e9d2b | 8954 | return 1; |
c8f6f18d RK |
8955 | /* else */ |
8956 | #endif | |
a55e9d2b | 8957 | |
8b760293 JW |
8958 | if (code == JUMP_INSN) |
8959 | return 0; | |
8960 | ||
8783b15e RK |
8961 | /* If this is a sequence, we must handle them all at once. |
8962 | We could have for instance a call that sets the target register, | |
e03f5d43 | 8963 | and an insn in a delay slot that uses the register. In this case, |
8783b15e | 8964 | we must return 0. */ |
8b760293 | 8965 | else if (code == INSN && GET_CODE (PATTERN (insn)) == SEQUENCE) |
a55e9d2b | 8966 | { |
8783b15e RK |
8967 | int i; |
8968 | int retval = 0; | |
8969 | ||
8970 | for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++) | |
8971 | { | |
8972 | rtx this_insn = XVECEXP (PATTERN (insn), 0, i); | |
8973 | rtx set = single_set (this_insn); | |
8974 | ||
f3536097 | 8975 | if (CALL_P (this_insn)) |
8783b15e | 8976 | code = CALL_INSN; |
f3536097 | 8977 | else if (JUMP_P (this_insn)) |
38f35781 JW |
8978 | { |
8979 | if (INSN_ANNULLED_BRANCH_P (this_insn)) | |
8980 | return 0; | |
8981 | code = JUMP_INSN; | |
8982 | } | |
a55e9d2b | 8983 | |
8783b15e RK |
8984 | if (set && reg_overlap_mentioned_p (reg, SET_SRC (set))) |
8985 | return 0; | |
8986 | if (set && reg_overlap_mentioned_p (reg, SET_DEST (set))) | |
8987 | { | |
f3536097 | 8988 | if (!MEM_P (SET_DEST (set))) |
8783b15e RK |
8989 | retval = 1; |
8990 | else | |
8991 | return 0; | |
8992 | } | |
8993 | if (set == 0 | |
8994 | && reg_overlap_mentioned_p (reg, PATTERN (this_insn))) | |
8995 | return 0; | |
8996 | } | |
8997 | if (retval == 1) | |
8998 | return 1; | |
38f35781 JW |
8999 | else if (code == JUMP_INSN) |
9000 | return 0; | |
8783b15e | 9001 | } |
a55e9d2b | 9002 | |
ec8e098d PB |
9003 | set = single_set (insn); |
9004 | if (set && reg_overlap_mentioned_p (reg, SET_SRC (set))) | |
9005 | return 0; | |
9006 | if (set && reg_overlap_mentioned_p (reg, SET_DEST (set))) | |
f3536097 | 9007 | return !MEM_P (SET_DEST (set)); |
ec8e098d PB |
9008 | if (set == 0 && reg_overlap_mentioned_p (reg, PATTERN (insn))) |
9009 | return 0; | |
8783b15e | 9010 | |
62164eb4 | 9011 | if (code == CALL_INSN && call_really_used_regs[REGNO (reg)]) |
8783b15e | 9012 | return 1; |
a55e9d2b RK |
9013 | } |
9014 | return 1; | |
9015 | } | |
225e4f43 | 9016 | \f |
c5b9ef02 | 9017 | #include "ggc.h" |
225e4f43 | 9018 | |
e2500fed | 9019 | static GTY(()) rtx fpscr_rtx; |
225e4f43 | 9020 | rtx |
cf277499 | 9021 | get_fpscr_rtx (void) |
225e4f43 | 9022 | { |
225e4f43 R |
9023 | if (! fpscr_rtx) |
9024 | { | |
c0d4e710 | 9025 | fpscr_rtx = gen_rtx_REG (PSImode, FPSCR_REG); |
225e4f43 | 9026 | REG_USERVAR_P (fpscr_rtx) = 1; |
225e4f43 R |
9027 | mark_user_reg (fpscr_rtx); |
9028 | } | |
9029 | if (! reload_completed || mdep_reorg_phase != SH_AFTER_MDEP_REORG) | |
9030 | mark_user_reg (fpscr_rtx); | |
9031 | return fpscr_rtx; | |
9032 | } | |
9033 | ||
24c2fde2 RH |
9034 | static GTY(()) tree fpscr_values; |
9035 | ||
9036 | static void | |
9037 | emit_fpu_switch (rtx scratch, int index) | |
9038 | { | |
9039 | rtx dst, src; | |
9040 | ||
9041 | if (fpscr_values == NULL) | |
9042 | { | |
9043 | tree t; | |
9044 | ||
9045 | t = build_index_type (integer_one_node); | |
9046 | t = build_array_type (integer_type_node, t); | |
4c4bde29 AH |
9047 | t = build_decl (BUILTINS_LOCATION, |
9048 | VAR_DECL, get_identifier ("__fpscr_values"), t); | |
24c2fde2 RH |
9049 | DECL_ARTIFICIAL (t) = 1; |
9050 | DECL_IGNORED_P (t) = 1; | |
9051 | DECL_EXTERNAL (t) = 1; | |
9052 | TREE_STATIC (t) = 1; | |
a5f5c6be | 9053 | TREE_PUBLIC (t) = 1; |
24c2fde2 RH |
9054 | TREE_USED (t) = 1; |
9055 | ||
9056 | fpscr_values = t; | |
9057 | } | |
9058 | ||
9059 | src = DECL_RTL (fpscr_values); | |
b3a13419 | 9060 | if (!can_create_pseudo_p ()) |
24c2fde2 RH |
9061 | { |
9062 | emit_move_insn (scratch, XEXP (src, 0)); | |
9063 | if (index != 0) | |
9064 | emit_insn (gen_addsi3 (scratch, scratch, GEN_INT (index * 4))); | |
9065 | src = adjust_automodify_address (src, PSImode, scratch, index * 4); | |
9066 | } | |
9067 | else | |
9068 | src = adjust_address (src, PSImode, index * 4); | |
9069 | ||
9070 | dst = get_fpscr_rtx (); | |
9071 | emit_move_insn (dst, src); | |
9072 | } | |
9073 | ||
225e4f43 | 9074 | void |
cf277499 | 9075 | emit_sf_insn (rtx pat) |
225e4f43 | 9076 | { |
225e4f43 | 9077 | emit_insn (pat); |
225e4f43 R |
9078 | } |
9079 | ||
9080 | void | |
cf277499 | 9081 | emit_df_insn (rtx pat) |
225e4f43 | 9082 | { |
225e4f43 | 9083 | emit_insn (pat); |
225e4f43 R |
9084 | } |
9085 | ||
9086 | void | |
cf277499 | 9087 | expand_sf_unop (rtx (*fun) (rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
9088 | { |
9089 | emit_sf_insn ((*fun) (operands[0], operands[1], get_fpscr_rtx ())); | |
9090 | } | |
9091 | ||
9092 | void | |
cf277499 | 9093 | expand_sf_binop (rtx (*fun) (rtx, rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
9094 | { |
9095 | emit_sf_insn ((*fun) (operands[0], operands[1], operands[2], | |
9096 | get_fpscr_rtx ())); | |
9097 | } | |
9098 | ||
9099 | void | |
cf277499 | 9100 | expand_df_unop (rtx (*fun) (rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
9101 | { |
9102 | emit_df_insn ((*fun) (operands[0], operands[1], get_fpscr_rtx ())); | |
9103 | } | |
9104 | ||
9105 | void | |
cf277499 | 9106 | expand_df_binop (rtx (*fun) (rtx, rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
9107 | { |
9108 | emit_df_insn ((*fun) (operands[0], operands[1], operands[2], | |
832a3292 | 9109 | get_fpscr_rtx ())); |
225e4f43 | 9110 | } |
225e4f43 | 9111 | \f |
cf277499 | 9112 | static rtx get_free_reg (HARD_REG_SET); |
9f09b1f2 R |
9113 | |
9114 | /* This function returns a register to use to load the address to load | |
9115 | the fpscr from. Currently it always returns r1 or r7, but when we are | |
9116 | able to use pseudo registers after combine, or have a better mechanism | |
9117 | for choosing a register, it should be done here. */ | |
9118 | /* REGS_LIVE is the liveness information for the point for which we | |
9119 | need this allocation. In some bare-bones exit blocks, r1 is live at the | |
9120 | start. We can even have all of r0..r3 being live: | |
9121 | __complex__ long long f (double d) { if (d == 0) return 2; else return 3; } | |
9122 | INSN before which new insns are placed with will clobber the register | |
9123 | we return. If a basic block consists only of setting the return value | |
9124 | register to a pseudo and using that register, the return value is not | |
9125 | live before or after this block, yet we we'll insert our insns right in | |
9126 | the middle. */ | |
225e4f43 | 9127 | |
9f09b1f2 | 9128 | static rtx |
cf277499 | 9129 | get_free_reg (HARD_REG_SET regs_live) |
225e4f43 | 9130 | { |
9f09b1f2 R |
9131 | if (! TEST_HARD_REG_BIT (regs_live, 1)) |
9132 | return gen_rtx_REG (Pmode, 1); | |
225e4f43 | 9133 | |
9f09b1f2 R |
9134 | /* Hard reg 1 is live; since this is a SMALL_REGISTER_CLASSES target, |
9135 | there shouldn't be anything but a jump before the function end. */ | |
f5b9e7c9 NS |
9136 | gcc_assert (!TEST_HARD_REG_BIT (regs_live, 7)); |
9137 | return gen_rtx_REG (Pmode, 7); | |
9f09b1f2 | 9138 | } |
225e4f43 | 9139 | |
f676971a | 9140 | /* This function will set the fpscr from memory. |
9f09b1f2 R |
9141 | MODE is the mode we are setting it to. */ |
9142 | void | |
cf277499 | 9143 | fpscr_set_from_mem (int mode, HARD_REG_SET regs_live) |
9f09b1f2 | 9144 | { |
81f40b79 | 9145 | enum attr_fp_mode fp_mode = (enum attr_fp_mode) mode; |
24c2fde2 | 9146 | enum attr_fp_mode norm_mode = ACTUAL_NORMAL_MODE (FP_MODE); |
5e77c6f9 | 9147 | rtx addr_reg; |
9f09b1f2 | 9148 | |
b3a13419 | 9149 | addr_reg = !can_create_pseudo_p () ? get_free_reg (regs_live) : NULL_RTX; |
24c2fde2 | 9150 | emit_fpu_switch (addr_reg, fp_mode == norm_mode); |
225e4f43 | 9151 | } |
9342c0c4 R |
9152 | |
9153 | /* Is the given character a logical line separator for the assembler? */ | |
9154 | #ifndef IS_ASM_LOGICAL_LINE_SEPARATOR | |
980d8882 | 9155 | #define IS_ASM_LOGICAL_LINE_SEPARATOR(C, STR) ((C) == ';') |
9342c0c4 R |
9156 | #endif |
9157 | ||
9158 | int | |
cf277499 | 9159 | sh_insn_length_adjustment (rtx insn) |
9342c0c4 R |
9160 | { |
9161 | /* Instructions with unfilled delay slots take up an extra two bytes for | |
9162 | the nop in the delay slot. */ | |
f3536097 | 9163 | if (((NONJUMP_INSN_P (insn) |
c1b92d09 R |
9164 | && GET_CODE (PATTERN (insn)) != USE |
9165 | && GET_CODE (PATTERN (insn)) != CLOBBER) | |
f3536097 SZ |
9166 | || CALL_P (insn) |
9167 | || (JUMP_P (insn) | |
9342c0c4 R |
9168 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC |
9169 | && GET_CODE (PATTERN (insn)) != ADDR_VEC)) | |
9170 | && GET_CODE (PATTERN (NEXT_INSN (PREV_INSN (insn)))) != SEQUENCE | |
9171 | && get_attr_needs_delay_slot (insn) == NEEDS_DELAY_SLOT_YES) | |
9172 | return 2; | |
9173 | ||
3a8699c7 AO |
9174 | /* SH2e has a bug that prevents the use of annulled branches, so if |
9175 | the delay slot is not filled, we'll have to put a NOP in it. */ | |
b757d36f | 9176 | if (sh_cpu_attr == CPU_SH2E |
f3536097 | 9177 | && JUMP_P (insn) |
3a8699c7 AO |
9178 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC |
9179 | && GET_CODE (PATTERN (insn)) != ADDR_VEC | |
9180 | && get_attr_type (insn) == TYPE_CBRANCH | |
9181 | && GET_CODE (PATTERN (NEXT_INSN (PREV_INSN (insn)))) != SEQUENCE) | |
9182 | return 2; | |
9183 | ||
9342c0c4 | 9184 | /* sh-dsp parallel processing insn take four bytes instead of two. */ |
f676971a | 9185 | |
f3536097 | 9186 | if (NONJUMP_INSN_P (insn)) |
9342c0c4 R |
9187 | { |
9188 | int sum = 0; | |
9189 | rtx body = PATTERN (insn); | |
0a2aaacc | 9190 | const char *templ; |
8ceaac3c | 9191 | char c; |
9342c0c4 R |
9192 | int maybe_label = 1; |
9193 | ||
9194 | if (GET_CODE (body) == ASM_INPUT) | |
0a2aaacc | 9195 | templ = XSTR (body, 0); |
9342c0c4 | 9196 | else if (asm_noperands (body) >= 0) |
0a2aaacc | 9197 | templ |
343d66ad | 9198 | = decode_asm_operands (body, NULL, NULL, NULL, NULL, NULL); |
9342c0c4 R |
9199 | else |
9200 | return 0; | |
9201 | do | |
9202 | { | |
9203 | int ppi_adjust = 0; | |
9204 | ||
9205 | do | |
0a2aaacc | 9206 | c = *templ++; |
9342c0c4 R |
9207 | while (c == ' ' || c == '\t'); |
9208 | /* all sh-dsp parallel-processing insns start with p. | |
9209 | The only non-ppi sh insn starting with p is pref. | |
9210 | The only ppi starting with pr is prnd. */ | |
0a2aaacc | 9211 | if ((c == 'p' || c == 'P') && strncasecmp ("re", templ, 2)) |
9342c0c4 R |
9212 | ppi_adjust = 2; |
9213 | /* The repeat pseudo-insn expands two three insns, a total of | |
9214 | six bytes in size. */ | |
9215 | else if ((c == 'r' || c == 'R') | |
0a2aaacc | 9216 | && ! strncasecmp ("epeat", templ, 5)) |
9342c0c4 | 9217 | ppi_adjust = 4; |
980d8882 | 9218 | while (c && c != '\n' |
0a2aaacc | 9219 | && ! IS_ASM_LOGICAL_LINE_SEPARATOR (c, templ)) |
9342c0c4 R |
9220 | { |
9221 | /* If this is a label, it is obviously not a ppi insn. */ | |
9222 | if (c == ':' && maybe_label) | |
9223 | { | |
9224 | ppi_adjust = 0; | |
9225 | break; | |
9226 | } | |
9227 | else if (c == '\'' || c == '"') | |
9228 | maybe_label = 0; | |
0a2aaacc | 9229 | c = *templ++; |
9342c0c4 R |
9230 | } |
9231 | sum += ppi_adjust; | |
9232 | maybe_label = c != ':'; | |
9233 | } | |
9234 | while (c); | |
9235 | return sum; | |
9236 | } | |
9237 | return 0; | |
9238 | } | |
1a66cd67 | 9239 | \f |
cdeed45a KK |
9240 | /* Return TRUE for a valid displacement for the REG+disp addressing |
9241 | with MODE. */ | |
9242 | ||
9243 | /* ??? The SH2e does not have the REG+disp addressing mode when loading values | |
9244 | into the FRx registers. We implement this by setting the maximum offset | |
9245 | to zero when the value is SFmode. This also restricts loading of SFmode | |
9246 | values into the integer registers, but that can't be helped. */ | |
9247 | ||
9248 | /* The SH allows a displacement in a QI or HI amode, but only when the | |
9249 | other operand is R0. GCC doesn't handle this very well, so we forgot | |
9250 | all of that. | |
9251 | ||
9252 | A legitimate index for a QI or HI is 0, SI can be any number 0..63, | |
9253 | DI can be any number 0..60. */ | |
9254 | ||
9255 | bool | |
9256 | sh_legitimate_index_p (enum machine_mode mode, rtx op) | |
9257 | { | |
f3536097 | 9258 | if (CONST_INT_P (op)) |
cdeed45a KK |
9259 | { |
9260 | if (TARGET_SHMEDIA) | |
9261 | { | |
9262 | int size; | |
9263 | ||
9264 | /* Check if this the address of an unaligned load / store. */ | |
9265 | if (mode == VOIDmode) | |
9266 | return CONST_OK_FOR_I06 (INTVAL (op)); | |
9267 | ||
9268 | size = GET_MODE_SIZE (mode); | |
9269 | return (!(INTVAL (op) & (size - 1)) | |
9270 | && INTVAL (op) >= -512 * size | |
9271 | && INTVAL (op) < 512 * size); | |
9272 | } | |
9273 | ||
9274 | if (TARGET_SH2A) | |
9275 | { | |
9276 | if (GET_MODE_SIZE (mode) == 1 | |
9277 | && (unsigned) INTVAL (op) < 4096) | |
9278 | return true; | |
9279 | } | |
9280 | ||
9281 | if ((GET_MODE_SIZE (mode) == 4 | |
9282 | && (unsigned) INTVAL (op) < 64 | |
9283 | && !(INTVAL (op) & 3) | |
9284 | && !(TARGET_SH2E && mode == SFmode)) | |
9285 | || (GET_MODE_SIZE (mode) == 4 | |
9286 | && (unsigned) INTVAL (op) < 16383 | |
9287 | && !(INTVAL (op) & 3) && TARGET_SH2A)) | |
9288 | return true; | |
9289 | ||
9290 | if ((GET_MODE_SIZE (mode) == 8 | |
9291 | && (unsigned) INTVAL (op) < 60 | |
9292 | && !(INTVAL (op) & 3) | |
9293 | && !((TARGET_SH4 || TARGET_SH2A) && mode == DFmode)) | |
9294 | || ((GET_MODE_SIZE (mode)==8) | |
9295 | && (unsigned) INTVAL (op) < 8192 | |
9296 | && !(INTVAL (op) & (TARGET_SH2A_DOUBLE ? 7 : 3)) | |
9297 | && (TARGET_SH2A && mode == DFmode))) | |
9298 | return true; | |
9299 | } | |
9300 | ||
9301 | return false; | |
9302 | } | |
9303 | ||
9304 | /* Recognize an RTL expression that is a valid memory address for | |
9305 | an instruction. | |
9306 | The MODE argument is the machine mode for the MEM expression | |
9307 | that wants to use this address. | |
9308 | Allow REG | |
9309 | REG+disp | |
9310 | REG+r0 | |
9311 | REG++ | |
9312 | --REG */ | |
9313 | ||
1bb99877 | 9314 | static bool |
cdeed45a KK |
9315 | sh_legitimate_address_p (enum machine_mode mode, rtx x, bool strict) |
9316 | { | |
9317 | if (MAYBE_BASE_REGISTER_RTX_P (x, strict)) | |
9318 | return true; | |
9319 | else if ((GET_CODE (x) == POST_INC || GET_CODE (x) == PRE_DEC) | |
9320 | && ! TARGET_SHMEDIA | |
9321 | && MAYBE_BASE_REGISTER_RTX_P (XEXP (x, 0), strict)) | |
9322 | return true; | |
9323 | else if (GET_CODE (x) == PLUS | |
9324 | && (mode != PSImode || reload_completed)) | |
9325 | { | |
9326 | rtx xop0 = XEXP (x, 0); | |
9327 | rtx xop1 = XEXP (x, 1); | |
9328 | ||
9329 | if (GET_MODE_SIZE (mode) <= 8 | |
9330 | && MAYBE_BASE_REGISTER_RTX_P (xop0, strict) | |
9331 | && sh_legitimate_index_p (mode, xop1)) | |
9332 | return true; | |
9333 | ||
9334 | if ((ALLOW_INDEXED_ADDRESS || GET_MODE (x) == DImode | |
9335 | || ((xop0 == stack_pointer_rtx | |
9336 | || xop0 == hard_frame_pointer_rtx) | |
9337 | && REG_P (xop1) && REGNO (xop1) == R0_REG) | |
9338 | || ((xop1 == stack_pointer_rtx | |
9339 | || xop1 == hard_frame_pointer_rtx) | |
9340 | && REG_P (xop0) && REGNO (xop0) == R0_REG)) | |
9341 | && ((!TARGET_SHMEDIA && GET_MODE_SIZE (mode) <= 4) | |
9342 | || (TARGET_SHMEDIA && GET_MODE_SIZE (mode) <= 8) | |
9343 | || ((TARGET_SH4 || TARGET_SH2A_DOUBLE) | |
9344 | && TARGET_FMOVD && mode == DFmode))) | |
9345 | { | |
9346 | if (MAYBE_BASE_REGISTER_RTX_P (xop1, strict) | |
9347 | && MAYBE_INDEX_REGISTER_RTX_P (xop0, strict)) | |
9348 | return true; | |
9349 | if (MAYBE_INDEX_REGISTER_RTX_P (xop1, strict) | |
9350 | && MAYBE_BASE_REGISTER_RTX_P (xop0, strict)) | |
9351 | return true; | |
9352 | } | |
9353 | } | |
9354 | ||
9355 | return false; | |
9356 | } | |
9357 | \f | |
55eec660 AO |
9358 | /* Return TRUE if X references a SYMBOL_REF or LABEL_REF whose symbol |
9359 | isn't protected by a PIC unspec. */ | |
1a66cd67 | 9360 | int |
cf277499 | 9361 | nonpic_symbol_mentioned_p (rtx x) |
1a66cd67 AO |
9362 | { |
9363 | register const char *fmt; | |
9364 | register int i; | |
9365 | ||
fa5322fa AO |
9366 | if (GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == LABEL_REF |
9367 | || GET_CODE (x) == PC) | |
1a66cd67 AO |
9368 | return 1; |
9369 | ||
fa5322fa AO |
9370 | /* We don't want to look into the possible MEM location of a |
9371 | CONST_DOUBLE, since we're not going to use it, in general. */ | |
9372 | if (GET_CODE (x) == CONST_DOUBLE) | |
9373 | return 0; | |
9374 | ||
1a66cd67 | 9375 | if (GET_CODE (x) == UNSPEC |
4773afa4 AO |
9376 | && (XINT (x, 1) == UNSPEC_PIC |
9377 | || XINT (x, 1) == UNSPEC_GOT | |
9378 | || XINT (x, 1) == UNSPEC_GOTOFF | |
fa5322fa | 9379 | || XINT (x, 1) == UNSPEC_GOTPLT |
463f02cd KK |
9380 | || XINT (x, 1) == UNSPEC_GOTTPOFF |
9381 | || XINT (x, 1) == UNSPEC_DTPOFF | |
dc3ba671 RS |
9382 | || XINT (x, 1) == UNSPEC_PLT |
9383 | || XINT (x, 1) == UNSPEC_SYMOFF | |
9384 | || XINT (x, 1) == UNSPEC_PCREL_SYMOFF)) | |
832a3292 | 9385 | return 0; |
1a66cd67 AO |
9386 | |
9387 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
9388 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
9389 | { | |
9390 | if (fmt[i] == 'E') | |
9391 | { | |
9392 | register int j; | |
9393 | ||
9394 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
9395 | if (nonpic_symbol_mentioned_p (XVECEXP (x, i, j))) | |
9396 | return 1; | |
9397 | } | |
9398 | else if (fmt[i] == 'e' && nonpic_symbol_mentioned_p (XEXP (x, i))) | |
9399 | return 1; | |
9400 | } | |
9401 | ||
9402 | return 0; | |
9403 | } | |
9404 | ||
9405 | /* Convert a non-PIC address in `orig' to a PIC address using @GOT or | |
6f317ef3 | 9406 | @GOTOFF in `reg'. */ |
1a66cd67 | 9407 | rtx |
cf277499 SB |
9408 | legitimize_pic_address (rtx orig, enum machine_mode mode ATTRIBUTE_UNUSED, |
9409 | rtx reg) | |
1a66cd67 | 9410 | { |
81f40b79 | 9411 | if (tls_symbolic_operand (orig, Pmode) != TLS_MODEL_NONE) |
463f02cd KK |
9412 | return orig; |
9413 | ||
1a66cd67 | 9414 | if (GET_CODE (orig) == LABEL_REF |
675ff4c7 | 9415 | || (GET_CODE (orig) == SYMBOL_REF && SYMBOL_REF_LOCAL_P (orig))) |
1a66cd67 AO |
9416 | { |
9417 | if (reg == 0) | |
9418 | reg = gen_reg_rtx (Pmode); | |
9419 | ||
9420 | emit_insn (gen_symGOTOFF2reg (reg, orig)); | |
9421 | return reg; | |
9422 | } | |
9423 | else if (GET_CODE (orig) == SYMBOL_REF) | |
9424 | { | |
9425 | if (reg == 0) | |
9426 | reg = gen_reg_rtx (Pmode); | |
9427 | ||
9428 | emit_insn (gen_symGOT2reg (reg, orig)); | |
9429 | return reg; | |
9430 | } | |
9431 | return orig; | |
9432 | } | |
b91455de | 9433 | |
506d7b68 PB |
9434 | /* Try machine-dependent ways of modifying an illegitimate address |
9435 | to be legitimate. If we find one, return the new, valid address. | |
9436 | Otherwise, return X. | |
9437 | ||
9438 | For the SH, if X is almost suitable for indexing, but the offset is | |
9439 | out of range, convert it into a normal form so that CSE has a chance | |
9440 | of reducing the number of address registers used. */ | |
9441 | ||
9442 | static rtx | |
9443 | sh_legitimize_address (rtx x, rtx oldx, enum machine_mode mode) | |
9444 | { | |
9445 | if (flag_pic) | |
9446 | x = legitimize_pic_address (oldx, mode, NULL_RTX); | |
9447 | ||
9448 | if (GET_CODE (x) == PLUS | |
9449 | && (GET_MODE_SIZE (mode) == 4 | |
9450 | || GET_MODE_SIZE (mode) == 8) | |
f3536097 | 9451 | && CONST_INT_P (XEXP (x, 1)) |
506d7b68 PB |
9452 | && BASE_REGISTER_RTX_P (XEXP (x, 0)) |
9453 | && ! TARGET_SHMEDIA | |
9454 | && ! ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && mode == DFmode) | |
9455 | && ! (TARGET_SH2E && mode == SFmode)) | |
9456 | { | |
9457 | rtx index_rtx = XEXP (x, 1); | |
9458 | HOST_WIDE_INT offset = INTVAL (index_rtx), offset_base; | |
9459 | rtx sum; | |
9460 | ||
9461 | /* On rare occasions, we might get an unaligned pointer | |
9462 | that is indexed in a way to give an aligned address. | |
9463 | Therefore, keep the lower two bits in offset_base. */ | |
9464 | /* Instead of offset_base 128..131 use 124..127, so that | |
9465 | simple add suffices. */ | |
9466 | if (offset > 127) | |
9467 | offset_base = ((offset + 4) & ~60) - 4; | |
9468 | else | |
9469 | offset_base = offset & ~60; | |
9470 | ||
9471 | /* Sometimes the normal form does not suit DImode. We | |
9472 | could avoid that by using smaller ranges, but that | |
9473 | would give less optimized code when SImode is | |
9474 | prevalent. */ | |
9475 | if (GET_MODE_SIZE (mode) + offset - offset_base <= 64) | |
9476 | { | |
9477 | sum = expand_binop (Pmode, add_optab, XEXP (x, 0), | |
9478 | GEN_INT (offset_base), NULL_RTX, 0, | |
9479 | OPTAB_LIB_WIDEN); | |
9480 | ||
9481 | return gen_rtx_PLUS (Pmode, sum, GEN_INT (offset - offset_base)); | |
9482 | } | |
9483 | } | |
9484 | ||
9485 | return x; | |
9486 | } | |
9487 | ||
b91455de KK |
9488 | /* Mark the use of a constant in the literal table. If the constant |
9489 | has multiple labels, make it unique. */ | |
c1b92d09 | 9490 | static rtx |
cf277499 | 9491 | mark_constant_pool_use (rtx x) |
b91455de KK |
9492 | { |
9493 | rtx insn, lab, pattern; | |
9494 | ||
9495 | if (x == NULL) | |
9496 | return x; | |
9497 | ||
9498 | switch (GET_CODE (x)) | |
9499 | { | |
9500 | case LABEL_REF: | |
9501 | x = XEXP (x, 0); | |
9502 | case CODE_LABEL: | |
9503 | break; | |
9504 | default: | |
9505 | return x; | |
9506 | } | |
9507 | ||
9508 | /* Get the first label in the list of labels for the same constant | |
9509 | and delete another labels in the list. */ | |
9510 | lab = x; | |
9511 | for (insn = PREV_INSN (x); insn; insn = PREV_INSN (insn)) | |
9512 | { | |
f3536097 | 9513 | if (!LABEL_P (insn) |
b91455de KK |
9514 | || LABEL_REFS (insn) != NEXT_INSN (insn)) |
9515 | break; | |
9516 | lab = insn; | |
9517 | } | |
9518 | ||
9519 | for (insn = LABEL_REFS (lab); insn; insn = LABEL_REFS (insn)) | |
9520 | INSN_DELETED_P (insn) = 1; | |
9521 | ||
9522 | /* Mark constants in a window. */ | |
9523 | for (insn = NEXT_INSN (x); insn; insn = NEXT_INSN (insn)) | |
9524 | { | |
f3536097 | 9525 | if (!NONJUMP_INSN_P (insn)) |
b91455de KK |
9526 | continue; |
9527 | ||
9528 | pattern = PATTERN (insn); | |
9529 | if (GET_CODE (pattern) != UNSPEC_VOLATILE) | |
9530 | continue; | |
9531 | ||
9532 | switch (XINT (pattern, 1)) | |
9533 | { | |
9534 | case UNSPECV_CONST2: | |
9535 | case UNSPECV_CONST4: | |
9536 | case UNSPECV_CONST8: | |
9537 | XVECEXP (pattern, 0, 1) = const1_rtx; | |
9538 | break; | |
9539 | case UNSPECV_WINDOW_END: | |
9540 | if (XVECEXP (pattern, 0, 0) == x) | |
9541 | return lab; | |
9542 | break; | |
9543 | case UNSPECV_CONST_END: | |
9544 | return lab; | |
9545 | default: | |
9546 | break; | |
9547 | } | |
9548 | } | |
9549 | ||
9550 | return lab; | |
9551 | } | |
6e7b03e1 AH |
9552 | \f |
9553 | /* Return true if it's possible to redirect BRANCH1 to the destination | |
9554 | of an unconditional jump BRANCH2. We only want to do this if the | |
9555 | resulting branch will have a short displacement. */ | |
f676971a | 9556 | int |
cf277499 | 9557 | sh_can_redirect_branch (rtx branch1, rtx branch2) |
6e7b03e1 AH |
9558 | { |
9559 | if (flag_expensive_optimizations && simplejump_p (branch2)) | |
9560 | { | |
9561 | rtx dest = XEXP (SET_SRC (single_set (branch2)), 0); | |
9562 | rtx insn; | |
9563 | int distance; | |
f676971a EC |
9564 | |
9565 | for (distance = 0, insn = NEXT_INSN (branch1); | |
9566 | insn && distance < 256; | |
6e7b03e1 AH |
9567 | insn = PREV_INSN (insn)) |
9568 | { | |
f676971a | 9569 | if (insn == dest) |
6e7b03e1 AH |
9570 | return 1; |
9571 | else | |
9572 | distance += get_attr_length (insn); | |
9573 | } | |
f676971a EC |
9574 | for (distance = 0, insn = NEXT_INSN (branch1); |
9575 | insn && distance < 256; | |
6e7b03e1 AH |
9576 | insn = NEXT_INSN (insn)) |
9577 | { | |
f676971a | 9578 | if (insn == dest) |
6e7b03e1 AH |
9579 | return 1; |
9580 | else | |
9581 | distance += get_attr_length (insn); | |
9582 | } | |
9583 | } | |
9584 | return 0; | |
9585 | } | |
9586 | ||
5e7a8ee0 | 9587 | /* Return nonzero if register old_reg can be renamed to register new_reg. */ |
03b8ec29 | 9588 | int |
cf277499 SB |
9589 | sh_hard_regno_rename_ok (unsigned int old_reg ATTRIBUTE_UNUSED, |
9590 | unsigned int new_reg) | |
03b8ec29 | 9591 | { |
832a3292 KH |
9592 | /* Interrupt functions can only use registers that have already been |
9593 | saved by the prologue, even if they would normally be | |
9594 | call-clobbered. */ | |
03b8ec29 | 9595 | |
6fb5fa3c | 9596 | if (sh_cfun_interrupt_handler_p () && !df_regs_ever_live_p (new_reg)) |
832a3292 | 9597 | return 0; |
03b8ec29 | 9598 | |
832a3292 | 9599 | return 1; |
03b8ec29 DD |
9600 | } |
9601 | ||
95c0af87 | 9602 | /* Function to update the integer COST |
c237e94a ZW |
9603 | based on the relationship between INSN that is dependent on |
9604 | DEP_INSN through the dependence LINK. The default is to make no | |
9605 | adjustment to COST. This can be used for example to specify to | |
9606 | the scheduler that an output- or anti-dependence does not incur | |
95c0af87 JR |
9607 | the same cost as a data-dependence. The return value should be |
9608 | the new value for COST. */ | |
c237e94a | 9609 | static int |
cf277499 | 9610 | sh_adjust_cost (rtx insn, rtx link ATTRIBUTE_UNUSED, rtx dep_insn, int cost) |
c237e94a | 9611 | { |
c49439f1 | 9612 | rtx reg, use_pat; |
c237e94a | 9613 | |
2ad65b0e SC |
9614 | if (TARGET_SHMEDIA) |
9615 | { | |
9616 | /* On SHmedia, if the dependence is an anti-dependence or | |
24746a42 | 9617 | output-dependence, there is no cost. */ |
2ad65b0e | 9618 | if (REG_NOTE_KIND (link) != 0) |
73a4d10b R |
9619 | { |
9620 | /* However, dependencies between target register loads and | |
9621 | uses of the register in a subsequent block that are separated | |
9622 | by a conditional branch are not modelled - we have to do with | |
9623 | the anti-dependency between the target register load and the | |
9624 | conditional branch that ends the current block. */ | |
9625 | if (REG_NOTE_KIND (link) == REG_DEP_ANTI | |
9626 | && GET_CODE (PATTERN (dep_insn)) == SET | |
9627 | && (get_attr_type (dep_insn) == TYPE_PT_MEDIA | |
9628 | || get_attr_type (dep_insn) == TYPE_PTABS_MEDIA) | |
9629 | && get_attr_type (insn) == TYPE_CBRANCH_MEDIA) | |
9630 | { | |
9631 | int orig_cost = cost; | |
9632 | rtx note = find_reg_note (insn, REG_BR_PROB, 0); | |
9633 | rtx target = ((! note | |
9634 | || INTVAL (XEXP (note, 0)) * 2 < REG_BR_PROB_BASE) | |
9635 | ? insn : JUMP_LABEL (insn)); | |
9636 | /* On the likely path, the branch costs 1, on the unlikely path, | |
9637 | it costs 3. */ | |
9638 | cost--; | |
9639 | do | |
9640 | target = next_active_insn (target); | |
9641 | while (target && ! flow_dependent_p (target, dep_insn) | |
9642 | && --cost > 0); | |
9643 | /* If two branches are executed in immediate succession, with the | |
9644 | first branch properly predicted, this causes a stall at the | |
9645 | second branch, hence we won't need the target for the | |
9646 | second branch for two cycles after the launch of the first | |
9647 | branch. */ | |
9648 | if (cost > orig_cost - 2) | |
9649 | cost = orig_cost - 2; | |
9650 | } | |
9651 | else | |
9652 | cost = 0; | |
9653 | } | |
2ad65b0e | 9654 | |
73a4d10b R |
9655 | else if (get_attr_is_mac_media (insn) |
9656 | && get_attr_is_mac_media (dep_insn)) | |
9657 | cost = 1; | |
9658 | ||
9659 | else if (! reload_completed | |
9660 | && GET_CODE (PATTERN (insn)) == SET | |
9661 | && GET_CODE (SET_SRC (PATTERN (insn))) == FLOAT | |
9662 | && GET_CODE (PATTERN (dep_insn)) == SET | |
9663 | && fp_arith_reg_operand (SET_SRC (PATTERN (dep_insn)), VOIDmode) | |
9664 | && cost < 4) | |
9665 | cost = 4; | |
9666 | /* Schedule the ptabs for a casesi_jump_media in preference to stuff | |
9667 | that is needed at the target. */ | |
9668 | else if (get_attr_type (insn) == TYPE_JUMP_MEDIA | |
9669 | && ! flow_dependent_p (insn, dep_insn)) | |
9670 | cost--; | |
2ad65b0e | 9671 | } |
c49439f1 | 9672 | else if (REG_NOTE_KIND (link) == 0) |
c237e94a | 9673 | { |
78d310c2 R |
9674 | enum attr_type type; |
9675 | rtx dep_set; | |
c49439f1 R |
9676 | |
9677 | if (recog_memoized (insn) < 0 | |
9678 | || recog_memoized (dep_insn) < 0) | |
95c0af87 | 9679 | return cost; |
c49439f1 | 9680 | |
78d310c2 | 9681 | dep_set = single_set (dep_insn); |
c49439f1 | 9682 | |
78d310c2 R |
9683 | /* The latency that we specify in the scheduling description refers |
9684 | to the actual output, not to an auto-increment register; for that, | |
9685 | the latency is one. */ | |
9686 | if (dep_set && MEM_P (SET_SRC (dep_set)) && cost > 1) | |
9687 | { | |
9688 | rtx set = single_set (insn); | |
9689 | ||
9690 | if (set | |
9691 | && !reg_mentioned_p (SET_DEST (dep_set), SET_SRC (set)) | |
9692 | && (!MEM_P (SET_DEST (set)) | |
9693 | || !reg_mentioned_p (SET_DEST (dep_set), | |
9694 | XEXP (SET_DEST (set), 0)))) | |
9695 | cost = 1; | |
9696 | } | |
c237e94a ZW |
9697 | /* The only input for a call that is timing-critical is the |
9698 | function's address. */ | |
f3536097 | 9699 | if (CALL_P (insn)) |
c49439f1 R |
9700 | { |
9701 | rtx call = PATTERN (insn); | |
9702 | ||
9703 | if (GET_CODE (call) == PARALLEL) | |
9704 | call = XVECEXP (call, 0 ,0); | |
9705 | if (GET_CODE (call) == SET) | |
9706 | call = SET_SRC (call); | |
f3536097 | 9707 | if (GET_CODE (call) == CALL && MEM_P (XEXP (call, 0)) |
73a4d10b R |
9708 | /* sibcalli_thunk uses a symbol_ref in an unspec. */ |
9709 | && (GET_CODE (XEXP (XEXP (call, 0), 0)) == UNSPEC | |
9710 | || ! reg_set_p (XEXP (XEXP (call, 0), 0), dep_insn))) | |
78d310c2 | 9711 | cost -= TARGET_SH4_300 ? 3 : 6; |
c49439f1 | 9712 | } |
c237e94a ZW |
9713 | /* Likewise, the most timing critical input for an sfuncs call |
9714 | is the function address. However, sfuncs typically start | |
9715 | using their arguments pretty quickly. | |
78d310c2 R |
9716 | Assume a four cycle delay for SH4 before they are needed. |
9717 | Cached ST40-300 calls are quicker, so assume only a one | |
9718 | cycle delay there. | |
9719 | ??? Maybe we should encode the delays till input registers | |
9720 | are needed by sfuncs into the sfunc call insn. */ | |
c49439f1 R |
9721 | /* All sfunc calls are parallels with at least four components. |
9722 | Exploit this to avoid unnecessary calls to sfunc_uses_reg. */ | |
9723 | else if (GET_CODE (PATTERN (insn)) == PARALLEL | |
9724 | && XVECLEN (PATTERN (insn), 0) >= 4 | |
9725 | && (reg = sfunc_uses_reg (insn))) | |
9726 | { | |
9727 | if (! reg_set_p (reg, dep_insn)) | |
78d310c2 R |
9728 | cost -= TARGET_SH4_300 ? 1 : 4; |
9729 | } | |
9730 | if (TARGET_HARD_SH4 && !TARGET_SH4_300) | |
9731 | { | |
9732 | enum attr_type dep_type = get_attr_type (dep_insn); | |
9733 | ||
9734 | if (dep_type == TYPE_FLOAD || dep_type == TYPE_PCFLOAD) | |
9735 | cost--; | |
9736 | else if ((dep_type == TYPE_LOAD_SI || dep_type == TYPE_PCLOAD_SI) | |
9737 | && (type = get_attr_type (insn)) != TYPE_CALL | |
9738 | && type != TYPE_SFUNC) | |
9739 | cost--; | |
9740 | /* When the preceding instruction loads the shift amount of | |
9741 | the following SHAD/SHLD, the latency of the load is increased | |
9742 | by 1 cycle. */ | |
9743 | if (get_attr_type (insn) == TYPE_DYN_SHIFT | |
9744 | && get_attr_any_int_load (dep_insn) == ANY_INT_LOAD_YES | |
c6ac6e82 | 9745 | && reg_overlap_mentioned_p (SET_DEST (dep_set), |
78d310c2 R |
9746 | XEXP (SET_SRC (single_set (insn)), |
9747 | 1))) | |
9748 | cost++; | |
9749 | /* When an LS group instruction with a latency of less than | |
9750 | 3 cycles is followed by a double-precision floating-point | |
9751 | instruction, FIPR, or FTRV, the latency of the first | |
9752 | instruction is increased to 3 cycles. */ | |
9753 | else if (cost < 3 | |
9754 | && get_attr_insn_class (dep_insn) == INSN_CLASS_LS_GROUP | |
9755 | && get_attr_dfp_comp (insn) == DFP_COMP_YES) | |
9756 | cost = 3; | |
9757 | /* The lsw register of a double-precision computation is ready one | |
9758 | cycle earlier. */ | |
9759 | else if (reload_completed | |
9760 | && get_attr_dfp_comp (dep_insn) == DFP_COMP_YES | |
9761 | && (use_pat = single_set (insn)) | |
9762 | && ! regno_use_in (REGNO (SET_DEST (single_set (dep_insn))), | |
9763 | SET_SRC (use_pat))) | |
9764 | cost -= 1; | |
9765 | ||
9766 | if (get_attr_any_fp_comp (dep_insn) == ANY_FP_COMP_YES | |
9767 | && get_attr_late_fp_use (insn) == LATE_FP_USE_YES) | |
9768 | cost -= 1; | |
9769 | } | |
9770 | else if (TARGET_SH4_300) | |
9771 | { | |
9772 | /* Stores need their input register two cycles later. */ | |
9773 | if (dep_set && cost >= 1 | |
9774 | && ((type = get_attr_type (insn)) == TYPE_STORE | |
9775 | || type == TYPE_PSTORE | |
9776 | || type == TYPE_FSTORE || type == TYPE_MAC_MEM)) | |
9777 | { | |
9778 | rtx set = single_set (insn); | |
9779 | ||
9780 | if (!reg_mentioned_p (SET_SRC (set), XEXP (SET_DEST (set), 0)) | |
9781 | && rtx_equal_p (SET_SRC (set), SET_DEST (dep_set))) | |
9782 | { | |
9783 | cost -= 2; | |
9784 | /* But don't reduce the cost below 1 if the address depends | |
9785 | on a side effect of dep_insn. */ | |
9786 | if (cost < 1 | |
9787 | && modified_in_p (XEXP (SET_DEST (set), 0), dep_insn)) | |
9788 | cost = 1; | |
9789 | } | |
9790 | } | |
9791 | } | |
c49439f1 R |
9792 | } |
9793 | /* An anti-dependence penalty of two applies if the first insn is a double | |
9794 | precision fadd / fsub / fmul. */ | |
78d310c2 R |
9795 | else if (!TARGET_SH4_300 |
9796 | && REG_NOTE_KIND (link) == REG_DEP_ANTI | |
c49439f1 | 9797 | && recog_memoized (dep_insn) >= 0 |
78d310c2 R |
9798 | && (get_attr_type (dep_insn) == TYPE_DFP_ARITH |
9799 | || get_attr_type (dep_insn) == TYPE_DFP_MUL) | |
c49439f1 R |
9800 | /* A lot of alleged anti-flow dependences are fake, |
9801 | so check this one is real. */ | |
9802 | && flow_dependent_p (dep_insn, insn)) | |
c237e94a | 9803 | cost = 2; |
c49439f1 | 9804 | |
c237e94a ZW |
9805 | return cost; |
9806 | } | |
3007d592 | 9807 | |
c49439f1 R |
9808 | /* Check if INSN is flow-dependent on DEP_INSN. Can also be used to check |
9809 | if DEP_INSN is anti-flow dependent on INSN. */ | |
9810 | static int | |
cf277499 | 9811 | flow_dependent_p (rtx insn, rtx dep_insn) |
c49439f1 R |
9812 | { |
9813 | rtx tmp = PATTERN (insn); | |
9814 | ||
9815 | note_stores (PATTERN (dep_insn), flow_dependent_p_1, &tmp); | |
9816 | return tmp == NULL_RTX; | |
9817 | } | |
9818 | ||
9819 | /* A helper function for flow_dependent_p called through note_stores. */ | |
9820 | static void | |
7bc980e1 | 9821 | flow_dependent_p_1 (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) |
c49439f1 R |
9822 | { |
9823 | rtx * pinsn = (rtx *) data; | |
9824 | ||
9825 | if (*pinsn && reg_referenced_p (x, *pinsn)) | |
9826 | *pinsn = NULL_RTX; | |
9827 | } | |
9828 | ||
b48f503c | 9829 | /* For use by sh_allocate_initial_value. Note that sh.md contains some |
3007d592 R |
9830 | 'special function' patterns (type sfunc) that clobber pr, but that |
9831 | do not look like function calls to leaf_function_p. Hence we must | |
9832 | do this extra check. */ | |
b48f503c | 9833 | static int |
cf277499 | 9834 | sh_pr_n_sets (void) |
3007d592 | 9835 | { |
6fb5fa3c | 9836 | return DF_REG_DEF_COUNT (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG); |
3007d592 | 9837 | } |
fa5322fa | 9838 | |
b48f503c KK |
9839 | /* Return where to allocate pseudo for a given hard register initial |
9840 | value. */ | |
9841 | static rtx | |
9842 | sh_allocate_initial_value (rtx hard_reg) | |
9843 | { | |
9844 | rtx x; | |
9845 | ||
9846 | if (REGNO (hard_reg) == (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG)) | |
9847 | { | |
9848 | if (current_function_is_leaf | |
9849 | && ! sh_pr_n_sets () | |
9850 | && ! (TARGET_SHCOMPACT | |
38173d38 | 9851 | && ((crtl->args.info.call_cookie |
b48f503c | 9852 | & ~ CALL_COOKIE_RET_TRAMP (1)) |
e3b5732b | 9853 | || crtl->saves_all_registers))) |
b48f503c KK |
9854 | x = hard_reg; |
9855 | else | |
57d38024 | 9856 | x = gen_frame_mem (Pmode, return_address_pointer_rtx); |
b48f503c KK |
9857 | } |
9858 | else | |
9859 | x = NULL_RTX; | |
9860 | ||
9861 | return x; | |
9862 | } | |
9863 | ||
c49439f1 R |
9864 | /* This function returns "2" to indicate dual issue for the SH4 |
9865 | processor. To be used by the DFA pipeline description. */ | |
fae15c93 | 9866 | static int |
832a3292 | 9867 | sh_issue_rate (void) |
fae15c93 | 9868 | { |
c49439f1 R |
9869 | if (TARGET_SUPERSCALAR) |
9870 | return 2; | |
9871 | else | |
9872 | return 1; | |
fae15c93 VM |
9873 | } |
9874 | ||
58565a33 SKG |
9875 | /* Functions for ready queue reordering for sched1. */ |
9876 | ||
9877 | /* Get weight for mode for a set x. */ | |
9878 | static short | |
508ea1c5 | 9879 | find_set_regmode_weight (rtx x, enum machine_mode mode) |
58565a33 SKG |
9880 | { |
9881 | if (GET_CODE (x) == CLOBBER && register_operand (SET_DEST (x), mode)) | |
9882 | return 1; | |
9883 | if (GET_CODE (x) == SET && register_operand (SET_DEST (x), mode)) | |
9884 | { | |
f3536097 | 9885 | if (REG_P (SET_DEST (x))) |
58565a33 SKG |
9886 | { |
9887 | if (!reg_mentioned_p (SET_DEST (x), SET_SRC (x))) | |
9888 | return 1; | |
9889 | else | |
9890 | return 0; | |
9891 | } | |
9892 | return 1; | |
9893 | } | |
9894 | return 0; | |
9895 | } | |
9896 | ||
9897 | /* Get regmode weight for insn. */ | |
9898 | static short | |
508ea1c5 | 9899 | find_insn_regmode_weight (rtx insn, enum machine_mode mode) |
58565a33 SKG |
9900 | { |
9901 | short reg_weight = 0; | |
9902 | rtx x; | |
9903 | ||
9904 | /* Increment weight for each register born here. */ | |
9905 | x = PATTERN (insn); | |
9906 | reg_weight += find_set_regmode_weight (x, mode); | |
9907 | if (GET_CODE (x) == PARALLEL) | |
9908 | { | |
9909 | int j; | |
9910 | for (j = XVECLEN (x, 0) - 1; j >= 0; j--) | |
9911 | { | |
9912 | x = XVECEXP (PATTERN (insn), 0, j); | |
9913 | reg_weight += find_set_regmode_weight (x, mode); | |
9914 | } | |
9915 | } | |
9916 | /* Decrement weight for each register that dies here. */ | |
9917 | for (x = REG_NOTES (insn); x; x = XEXP (x, 1)) | |
9918 | { | |
9919 | if (REG_NOTE_KIND (x) == REG_DEAD || REG_NOTE_KIND (x) == REG_UNUSED) | |
9920 | { | |
9921 | rtx note = XEXP (x, 0); | |
f3536097 | 9922 | if (REG_P (note) && GET_MODE (note) == mode) |
58565a33 SKG |
9923 | reg_weight--; |
9924 | } | |
9925 | } | |
9926 | return reg_weight; | |
9927 | } | |
9928 | ||
9929 | /* Calculate regmode weights for all insns of a basic block. */ | |
9930 | static void | |
c6d56443 | 9931 | find_regmode_weight (basic_block b, enum machine_mode mode) |
58565a33 SKG |
9932 | { |
9933 | rtx insn, next_tail, head, tail; | |
9934 | ||
c6d56443 | 9935 | get_ebb_head_tail (b, b, &head, &tail); |
58565a33 SKG |
9936 | next_tail = NEXT_INSN (tail); |
9937 | ||
9938 | for (insn = head; insn != next_tail; insn = NEXT_INSN (insn)) | |
9939 | { | |
9940 | /* Handle register life information. */ | |
9941 | if (!INSN_P (insn)) | |
9942 | continue; | |
9943 | ||
9944 | if (mode == SFmode) | |
9945 | INSN_REGMODE_WEIGHT (insn, mode) = | |
9946 | find_insn_regmode_weight (insn, mode) + 2 * find_insn_regmode_weight (insn, DFmode); | |
9947 | else if (mode == SImode) | |
9948 | INSN_REGMODE_WEIGHT (insn, mode) = | |
9949 | find_insn_regmode_weight (insn, mode) + 2 * find_insn_regmode_weight (insn, DImode); | |
9950 | } | |
9951 | } | |
9952 | ||
9953 | /* Comparison function for ready queue sorting. */ | |
9954 | static int | |
508ea1c5 | 9955 | rank_for_reorder (const void *x, const void *y) |
58565a33 SKG |
9956 | { |
9957 | rtx tmp = *(const rtx *) y; | |
9958 | rtx tmp2 = *(const rtx *) x; | |
9959 | ||
9960 | /* The insn in a schedule group should be issued the first. */ | |
9961 | if (SCHED_GROUP_P (tmp) != SCHED_GROUP_P (tmp2)) | |
9962 | return SCHED_GROUP_P (tmp2) ? 1 : -1; | |
9963 | ||
f676971a | 9964 | /* If insns are equally good, sort by INSN_LUID (original insn order), This |
58565a33 SKG |
9965 | minimizes instruction movement, thus minimizing sched's effect on |
9966 | register pressure. */ | |
9967 | return INSN_LUID (tmp) - INSN_LUID (tmp2); | |
9968 | } | |
9969 | ||
9970 | /* Resort the array A in which only element at index N may be out of order. */ | |
9971 | static void | |
508ea1c5 | 9972 | swap_reorder (rtx *a, int n) |
58565a33 SKG |
9973 | { |
9974 | rtx insn = a[n - 1]; | |
9975 | int i = n - 2; | |
9976 | ||
9977 | while (i >= 0 && rank_for_reorder (a + i, &insn) >= 0) | |
9978 | { | |
9979 | a[i + 1] = a[i]; | |
9980 | i -= 1; | |
9981 | } | |
9982 | a[i + 1] = insn; | |
9983 | } | |
9984 | ||
832a3292 KH |
9985 | #define SCHED_REORDER(READY, N_READY) \ |
9986 | do \ | |
9987 | { \ | |
9988 | if ((N_READY) == 2) \ | |
9989 | swap_reorder (READY, N_READY); \ | |
9990 | else if ((N_READY) > 2) \ | |
9991 | qsort (READY, N_READY, sizeof (rtx), rank_for_reorder); \ | |
9992 | } \ | |
9993 | while (0) | |
58565a33 SKG |
9994 | |
9995 | /* Sort the ready list READY by ascending priority, using the SCHED_REORDER | |
9996 | macro. */ | |
9997 | static void | |
508ea1c5 | 9998 | ready_reorder (rtx *ready, int nready) |
58565a33 SKG |
9999 | { |
10000 | SCHED_REORDER (ready, nready); | |
10001 | } | |
10002 | ||
6fb5fa3c DB |
10003 | /* Count life regions of r0 for a block. */ |
10004 | static int | |
10005 | find_r0_life_regions (basic_block b) | |
10006 | { | |
10007 | rtx end, insn; | |
10008 | rtx pset; | |
10009 | rtx r0_reg; | |
10010 | int live; | |
10011 | int set; | |
10012 | int death = 0; | |
10013 | ||
89a95777 | 10014 | if (REGNO_REG_SET_P (df_get_live_in (b), R0_REG)) |
6fb5fa3c DB |
10015 | { |
10016 | set = 1; | |
10017 | live = 1; | |
10018 | } | |
10019 | else | |
10020 | { | |
10021 | set = 0; | |
10022 | live = 0; | |
10023 | } | |
10024 | ||
10025 | insn = BB_HEAD (b); | |
10026 | end = BB_END (b); | |
10027 | r0_reg = gen_rtx_REG (SImode, R0_REG); | |
10028 | while (1) | |
10029 | { | |
10030 | if (INSN_P (insn)) | |
10031 | { | |
10032 | if (find_regno_note (insn, REG_DEAD, R0_REG)) | |
10033 | { | |
10034 | death++; | |
10035 | live = 0; | |
10036 | } | |
10037 | if (!live | |
10038 | && (pset = single_set (insn)) | |
10039 | && reg_overlap_mentioned_p (r0_reg, SET_DEST (pset)) | |
10040 | && !find_regno_note (insn, REG_UNUSED, R0_REG)) | |
10041 | { | |
10042 | set++; | |
10043 | live = 1; | |
10044 | } | |
10045 | } | |
10046 | if (insn == end) | |
10047 | break; | |
10048 | insn = NEXT_INSN (insn); | |
10049 | } | |
10050 | return set - death; | |
10051 | } | |
10052 | ||
58565a33 SKG |
10053 | /* Calculate regmode weights for all insns of all basic block. */ |
10054 | static void | |
508ea1c5 KH |
10055 | sh_md_init_global (FILE *dump ATTRIBUTE_UNUSED, |
10056 | int verbose ATTRIBUTE_UNUSED, | |
10057 | int old_max_uid) | |
58565a33 SKG |
10058 | { |
10059 | basic_block b; | |
10060 | ||
10061 | regmode_weight[0] = (short *) xcalloc (old_max_uid, sizeof (short)); | |
10062 | regmode_weight[1] = (short *) xcalloc (old_max_uid, sizeof (short)); | |
6fb5fa3c | 10063 | r0_life_regions = 0; |
58565a33 SKG |
10064 | |
10065 | FOR_EACH_BB_REVERSE (b) | |
10066 | { | |
c6d56443 KK |
10067 | find_regmode_weight (b, SImode); |
10068 | find_regmode_weight (b, SFmode); | |
6fb5fa3c DB |
10069 | if (!reload_completed) |
10070 | r0_life_regions += find_r0_life_regions (b); | |
58565a33 SKG |
10071 | } |
10072 | ||
10073 | CURR_REGMODE_PRESSURE (SImode) = 0; | |
10074 | CURR_REGMODE_PRESSURE (SFmode) = 0; | |
10075 | ||
10076 | } | |
10077 | ||
10078 | /* Cleanup. */ | |
10079 | static void | |
508ea1c5 KH |
10080 | sh_md_finish_global (FILE *dump ATTRIBUTE_UNUSED, |
10081 | int verbose ATTRIBUTE_UNUSED) | |
58565a33 SKG |
10082 | { |
10083 | if (regmode_weight[0]) | |
10084 | { | |
10085 | free (regmode_weight[0]); | |
10086 | regmode_weight[0] = NULL; | |
10087 | } | |
10088 | if (regmode_weight[1]) | |
10089 | { | |
10090 | free (regmode_weight[1]); | |
10091 | regmode_weight[1] = NULL; | |
10092 | } | |
10093 | } | |
10094 | ||
b0fbea13 KK |
10095 | /* The scalar modes supported differs from the default version in TImode |
10096 | for 32-bit SHMEDIA. */ | |
10097 | static bool | |
10098 | sh_scalar_mode_supported_p (enum machine_mode mode) | |
10099 | { | |
10100 | if (TARGET_SHMEDIA32 && mode == TImode) | |
10101 | return false; | |
10102 | ||
10103 | return default_scalar_mode_supported_p (mode); | |
10104 | } | |
10105 | ||
58565a33 SKG |
10106 | /* Cache the can_issue_more so that we can return it from reorder2. Also, |
10107 | keep count of register pressures on SImode and SFmode. */ | |
10108 | static int | |
508ea1c5 KH |
10109 | sh_variable_issue (FILE *dump ATTRIBUTE_UNUSED, |
10110 | int sched_verbose ATTRIBUTE_UNUSED, | |
10111 | rtx insn, | |
10112 | int can_issue_more) | |
58565a33 SKG |
10113 | { |
10114 | if (GET_CODE (PATTERN (insn)) != USE | |
10115 | && GET_CODE (PATTERN (insn)) != CLOBBER) | |
10116 | cached_can_issue_more = can_issue_more - 1; | |
10117 | else | |
10118 | cached_can_issue_more = can_issue_more; | |
10119 | ||
10120 | if (reload_completed) | |
10121 | return cached_can_issue_more; | |
10122 | ||
10123 | CURR_REGMODE_PRESSURE (SImode) += INSN_REGMODE_WEIGHT (insn, SImode); | |
10124 | CURR_REGMODE_PRESSURE (SFmode) += INSN_REGMODE_WEIGHT (insn, SFmode); | |
10125 | ||
10126 | return cached_can_issue_more; | |
10127 | } | |
10128 | ||
10129 | static void | |
508ea1c5 KH |
10130 | sh_md_init (FILE *dump ATTRIBUTE_UNUSED, |
10131 | int verbose ATTRIBUTE_UNUSED, | |
10132 | int veclen ATTRIBUTE_UNUSED) | |
58565a33 SKG |
10133 | { |
10134 | CURR_REGMODE_PRESSURE (SImode) = 0; | |
10135 | CURR_REGMODE_PRESSURE (SFmode) = 0; | |
10136 | } | |
10137 | ||
10138 | /* Some magic numbers. */ | |
10139 | /* Pressure on register r0 can lead to spill failures. so avoid sched1 for | |
10140 | functions that already have high pressure on r0. */ | |
10141 | #define R0_MAX_LIFE_REGIONS 2 | |
4ee31f1e | 10142 | /* Register Pressure thresholds for SImode and SFmode registers. */ |
58565a33 SKG |
10143 | #define SIMODE_MAX_WEIGHT 5 |
10144 | #define SFMODE_MAX_WEIGHT 10 | |
10145 | ||
10146 | /* Return true if the pressure is high for MODE. */ | |
10147 | static short | |
508ea1c5 | 10148 | high_pressure (enum machine_mode mode) |
58565a33 SKG |
10149 | { |
10150 | /* Pressure on register r0 can lead to spill failures. so avoid sched1 for | |
10151 | functions that already have high pressure on r0. */ | |
6fb5fa3c DB |
10152 | if (r0_life_regions >= R0_MAX_LIFE_REGIONS) |
10153 | return 1; | |
58565a33 SKG |
10154 | |
10155 | if (mode == SFmode) | |
10156 | return (CURR_REGMODE_PRESSURE (SFmode) > SFMODE_MAX_WEIGHT); | |
10157 | else | |
10158 | return (CURR_REGMODE_PRESSURE (SImode) > SIMODE_MAX_WEIGHT); | |
10159 | } | |
10160 | ||
10161 | /* Reorder ready queue if register pressure is high. */ | |
10162 | static int | |
508ea1c5 KH |
10163 | sh_reorder (FILE *dump ATTRIBUTE_UNUSED, |
10164 | int sched_verbose ATTRIBUTE_UNUSED, | |
10165 | rtx *ready, | |
10166 | int *n_readyp, | |
10167 | int clock_var ATTRIBUTE_UNUSED) | |
58565a33 SKG |
10168 | { |
10169 | if (reload_completed) | |
10170 | return sh_issue_rate (); | |
10171 | ||
10172 | if (high_pressure (SFmode) || high_pressure (SImode)) | |
10173 | { | |
10174 | ready_reorder (ready, *n_readyp); | |
10175 | } | |
10176 | ||
10177 | return sh_issue_rate (); | |
10178 | } | |
10179 | ||
10180 | /* Skip cycles if the current register pressure is high. */ | |
f676971a | 10181 | static int |
508ea1c5 KH |
10182 | sh_reorder2 (FILE *dump ATTRIBUTE_UNUSED, |
10183 | int sched_verbose ATTRIBUTE_UNUSED, | |
10184 | rtx *ready ATTRIBUTE_UNUSED, | |
10185 | int *n_readyp ATTRIBUTE_UNUSED, | |
10186 | int clock_var ATTRIBUTE_UNUSED) | |
58565a33 SKG |
10187 | { |
10188 | if (reload_completed) | |
10189 | return cached_can_issue_more; | |
10190 | ||
f676971a | 10191 | if (high_pressure(SFmode) || high_pressure (SImode)) |
58565a33 SKG |
10192 | skip_cycles = 1; |
10193 | ||
10194 | return cached_can_issue_more; | |
10195 | } | |
10196 | ||
10197 | /* Skip cycles without sorting the ready queue. This will move insn from | |
10198 | Q->R. If this is the last cycle we are skipping; allow sorting of ready | |
f676971a | 10199 | queue by sh_reorder. */ |
58565a33 | 10200 | |
f676971a EC |
10201 | /* Generally, skipping these many cycles are sufficient for all insns to move |
10202 | from Q -> R. */ | |
10203 | #define MAX_SKIPS 8 | |
58565a33 SKG |
10204 | |
10205 | static int | |
508ea1c5 KH |
10206 | sh_dfa_new_cycle (FILE *sched_dump ATTRIBUTE_UNUSED, |
10207 | int sched_verbose ATTRIBUTE_UNUSED, | |
10208 | rtx insn ATTRIBUTE_UNUSED, | |
10209 | int last_clock_var, | |
10210 | int clock_var, | |
10211 | int *sort_p) | |
58565a33 SKG |
10212 | { |
10213 | if (reload_completed) | |
832a3292 | 10214 | return 0; |
58565a33 | 10215 | |
f676971a | 10216 | if (skip_cycles) |
58565a33 | 10217 | { |
832a3292 KH |
10218 | if ((clock_var - last_clock_var) < MAX_SKIPS) |
10219 | { | |
10220 | *sort_p = 0; | |
10221 | return 1; | |
10222 | } | |
10223 | /* If this is the last cycle we are skipping, allow reordering of R. */ | |
10224 | if ((clock_var - last_clock_var) == MAX_SKIPS) | |
10225 | { | |
10226 | *sort_p = 1; | |
10227 | return 1; | |
10228 | } | |
58565a33 | 10229 | } |
58565a33 SKG |
10230 | |
10231 | skip_cycles = 0; | |
10232 | ||
10233 | return 0; | |
10234 | } | |
10235 | ||
fa5322fa AO |
10236 | /* SHmedia requires registers for branches, so we can't generate new |
10237 | branches past reload. */ | |
10238 | static bool | |
cf277499 | 10239 | sh_cannot_modify_jumps_p (void) |
fa5322fa AO |
10240 | { |
10241 | return (TARGET_SHMEDIA && (reload_in_progress || reload_completed)); | |
10242 | } | |
10243 | ||
81f40b79 | 10244 | static enum reg_class |
fe3ad572 SC |
10245 | sh_target_reg_class (void) |
10246 | { | |
10247 | return TARGET_SHMEDIA ? TARGET_REGS : NO_REGS; | |
10248 | } | |
10249 | ||
10250 | static bool | |
10251 | sh_optimize_target_register_callee_saved (bool after_prologue_epilogue_gen) | |
10252 | { | |
73a4d10b | 10253 | HARD_REG_SET dummy; |
cb3d3842 | 10254 | #if 0 |
73a4d10b | 10255 | rtx insn; |
cb3d3842 | 10256 | #endif |
73a4d10b R |
10257 | |
10258 | if (! shmedia_space_reserved_for_target_registers) | |
10259 | return 0; | |
10260 | if (after_prologue_epilogue_gen && ! TARGET_SAVE_ALL_TARGET_REGS) | |
10261 | return 0; | |
10262 | if (calc_live_regs (&dummy) >= 6 * 8) | |
10263 | return 1; | |
73a4d10b | 10264 | return 0; |
fe3ad572 SC |
10265 | } |
10266 | ||
fa5322fa | 10267 | static bool |
3101faab | 10268 | sh_ms_bitfield_layout_p (const_tree record_type ATTRIBUTE_UNUSED) |
fa5322fa | 10269 | { |
61f71b34 | 10270 | return (TARGET_SH5 || TARGET_HITACHI || sh_attr_renesas_p (record_type)); |
fa5322fa | 10271 | } |
ca903bba | 10272 | \f |
f676971a | 10273 | /* |
ca903bba R |
10274 | On the SH1..SH4, the trampoline looks like |
10275 | 2 0002 D202 mov.l l2,r2 | |
10276 | 1 0000 D301 mov.l l1,r3 | |
10277 | 3 0004 422B jmp @r2 | |
10278 | 4 0006 0009 nop | |
10279 | 5 0008 00000000 l1: .long area | |
10280 | 6 000c 00000000 l2: .long function | |
10281 | ||
10282 | SH5 (compact) uses r1 instead of r3 for the static chain. */ | |
10283 | ||
10284 | ||
10285 | /* Emit RTL insns to initialize the variable parts of a trampoline. | |
10286 | FNADDR is an RTX for the address of the function's pure code. | |
10287 | CXT is an RTX for the static chain value for the function. */ | |
10288 | ||
10289 | void | |
cf277499 | 10290 | sh_initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt) |
ca903bba | 10291 | { |
57d38024 R |
10292 | rtx tramp_mem = gen_frame_mem (BLKmode, tramp); |
10293 | ||
ca903bba R |
10294 | if (TARGET_SHMEDIA64) |
10295 | { | |
10296 | rtx tramp_templ; | |
10297 | int fixed_len; | |
10298 | ||
10299 | rtx movi1 = GEN_INT (0xcc000010); | |
10300 | rtx shori1 = GEN_INT (0xc8000010); | |
10301 | rtx src, dst; | |
10302 | ||
10303 | /* The following trampoline works within a +- 128 KB range for cxt: | |
10304 | ptb/u cxt,tr1; movi fnaddr >> 48,r0; shori fnaddr >> 32,r0; | |
10305 | shori fnaddr >> 16,r0; shori fnaddr,r0; ptabs/l r0,tr0 | |
10306 | gettr tr1,r1; blink tr0,r63 */ | |
10307 | /* Address rounding makes it hard to compute the exact bounds of the | |
10308 | offset for this trampoline, but we have a rather generous offset | |
10309 | range, so frame_offset should do fine as an upper bound. */ | |
10310 | if (cxt == virtual_stack_vars_rtx && frame_offset < 0x20000) | |
10311 | { | |
10312 | /* ??? could optimize this trampoline initialization | |
10313 | by writing DImode words with two insns each. */ | |
10314 | rtx mask = force_reg (DImode, GEN_INT (0x3fffc00)); | |
10315 | rtx insn = gen_rtx_MINUS (DImode, cxt, tramp); | |
10316 | insn = gen_rtx_ASHIFT (DImode, insn, GEN_INT (10-2)); | |
10317 | insn = gen_rtx_AND (DImode, insn, mask); | |
10318 | /* Or in ptb/u .,tr1 pattern */ | |
10319 | insn = gen_rtx_IOR (DImode, insn, gen_int_mode (0xec000010, SImode)); | |
10320 | insn = force_operand (insn, NULL_RTX); | |
10321 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 10322 | emit_move_insn (change_address (tramp_mem, SImode, NULL_RTX), insn); |
ca903bba R |
10323 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (38)); |
10324 | insn = gen_rtx_AND (DImode, insn, mask); | |
10325 | insn = force_operand (gen_rtx_IOR (DImode, movi1, insn), NULL_RTX); | |
10326 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 10327 | emit_move_insn (adjust_address (tramp_mem, SImode, 4), insn); |
ca903bba R |
10328 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (22)); |
10329 | insn = gen_rtx_AND (DImode, insn, mask); | |
10330 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
10331 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 10332 | emit_move_insn (adjust_address (tramp_mem, SImode, 8), insn); |
ca903bba R |
10333 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (6)); |
10334 | insn = gen_rtx_AND (DImode, insn, mask); | |
10335 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
10336 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 10337 | emit_move_insn (adjust_address (tramp_mem, SImode, 12), insn); |
ca903bba R |
10338 | insn = gen_rtx_ASHIFT (DImode, fnaddr, GEN_INT (10)); |
10339 | insn = gen_rtx_AND (DImode, insn, mask); | |
10340 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
10341 | insn = gen_lowpart (SImode, insn); | |
57d38024 R |
10342 | emit_move_insn (adjust_address (tramp_mem, SImode, 16), insn); |
10343 | emit_move_insn (adjust_address (tramp_mem, SImode, 20), | |
ca903bba | 10344 | GEN_INT (0x6bf10600)); |
57d38024 | 10345 | emit_move_insn (adjust_address (tramp_mem, SImode, 24), |
ca903bba | 10346 | GEN_INT (0x4415fc10)); |
57d38024 | 10347 | emit_move_insn (adjust_address (tramp_mem, SImode, 28), |
ca903bba R |
10348 | GEN_INT (0x4401fff0)); |
10349 | emit_insn (gen_ic_invalidate_line (tramp)); | |
10350 | return; | |
10351 | } | |
10352 | tramp_templ = gen_rtx_SYMBOL_REF (Pmode,"__GCC_nested_trampoline"); | |
10353 | fixed_len = TRAMPOLINE_SIZE - 2 * GET_MODE_SIZE (Pmode); | |
10354 | ||
10355 | tramp_templ = gen_datalabel_ref (tramp_templ); | |
57d38024 R |
10356 | dst = tramp_mem; |
10357 | src = gen_const_mem (BLKmode, tramp_templ); | |
ca903bba R |
10358 | set_mem_align (dst, 256); |
10359 | set_mem_align (src, 64); | |
44bb111a | 10360 | emit_block_move (dst, src, GEN_INT (fixed_len), BLOCK_OP_NORMAL); |
ca903bba | 10361 | |
57d38024 R |
10362 | emit_move_insn (adjust_address (tramp_mem, Pmode, fixed_len), fnaddr); |
10363 | emit_move_insn (adjust_address (tramp_mem, Pmode, | |
10364 | fixed_len + GET_MODE_SIZE (Pmode)), | |
ca903bba R |
10365 | cxt); |
10366 | emit_insn (gen_ic_invalidate_line (tramp)); | |
10367 | return; | |
10368 | } | |
10369 | else if (TARGET_SHMEDIA) | |
10370 | { | |
10371 | /* movi fnaddr >> 16,r1; shori fnaddr,r1; ptabs/l r1,tr0 | |
10372 | movi cxt >> 16,r1; shori cxt,r1; blink tr0,r63 */ | |
10373 | rtx quad0 = gen_reg_rtx (DImode), cxtload = gen_reg_rtx (DImode); | |
10374 | rtx quad1 = gen_reg_rtx (DImode), quad2 = gen_reg_rtx (DImode); | |
10375 | /* movi 0,r1: 0xcc000010 shori 0,r1: c8000010 concatenated, | |
825db093 | 10376 | rotated 10 right, and higher 16 bit of every 32 selected. */ |
ca903bba R |
10377 | rtx movishori |
10378 | = force_reg (V2HImode, (simplify_gen_subreg | |
10379 | (V2HImode, GEN_INT (0x4330432), SImode, 0))); | |
10380 | rtx ptabs = force_reg (DImode, GEN_INT (0x6bf10600)); | |
10381 | rtx blink = force_reg (DImode, GEN_INT (0x4401fff0)); | |
10382 | ||
10383 | tramp = force_reg (Pmode, tramp); | |
10384 | fnaddr = force_reg (SImode, fnaddr); | |
10385 | cxt = force_reg (SImode, cxt); | |
10386 | emit_insn (gen_mshflo_w_x (gen_rtx_SUBREG (V4HImode, quad0, 0), | |
10387 | gen_rtx_SUBREG (V2HImode, fnaddr, 0), | |
10388 | movishori)); | |
f1e2239f | 10389 | emit_insn (gen_rotrdi3_mextr (quad0, quad0, |
ca903bba | 10390 | GEN_INT (TARGET_LITTLE_ENDIAN ? 24 : 56))); |
a556fd39 | 10391 | emit_insn (gen_ashldi3_media (quad0, quad0, const2_rtx)); |
57d38024 | 10392 | emit_move_insn (change_address (tramp_mem, DImode, NULL_RTX), quad0); |
ca903bba R |
10393 | emit_insn (gen_mshflo_w_x (gen_rtx_SUBREG (V4HImode, cxtload, 0), |
10394 | gen_rtx_SUBREG (V2HImode, cxt, 0), | |
10395 | movishori)); | |
f1e2239f | 10396 | emit_insn (gen_rotrdi3_mextr (cxtload, cxtload, |
ca903bba | 10397 | GEN_INT (TARGET_LITTLE_ENDIAN ? 24 : 56))); |
a556fd39 | 10398 | emit_insn (gen_ashldi3_media (cxtload, cxtload, const2_rtx)); |
ca903bba R |
10399 | if (TARGET_LITTLE_ENDIAN) |
10400 | { | |
10401 | emit_insn (gen_mshflo_l_di (quad1, ptabs, cxtload)); | |
10402 | emit_insn (gen_mextr4 (quad2, cxtload, blink)); | |
10403 | } | |
10404 | else | |
10405 | { | |
10406 | emit_insn (gen_mextr4 (quad1, cxtload, ptabs)); | |
10407 | emit_insn (gen_mshflo_l_di (quad2, blink, cxtload)); | |
10408 | } | |
57d38024 R |
10409 | emit_move_insn (adjust_address (tramp_mem, DImode, 8), quad1); |
10410 | emit_move_insn (adjust_address (tramp_mem, DImode, 16), quad2); | |
ca903bba R |
10411 | emit_insn (gen_ic_invalidate_line (tramp)); |
10412 | return; | |
10413 | } | |
10414 | else if (TARGET_SHCOMPACT) | |
10415 | { | |
10416 | emit_insn (gen_initialize_trampoline (tramp, cxt, fnaddr)); | |
10417 | return; | |
10418 | } | |
57d38024 | 10419 | emit_move_insn (change_address (tramp_mem, SImode, NULL_RTX), |
ca903bba R |
10420 | gen_int_mode (TARGET_LITTLE_ENDIAN ? 0xd301d202 : 0xd202d301, |
10421 | SImode)); | |
57d38024 | 10422 | emit_move_insn (adjust_address (tramp_mem, SImode, 4), |
ca903bba R |
10423 | gen_int_mode (TARGET_LITTLE_ENDIAN ? 0x0009422b : 0x422b0009, |
10424 | SImode)); | |
57d38024 R |
10425 | emit_move_insn (adjust_address (tramp_mem, SImode, 8), cxt); |
10426 | emit_move_insn (adjust_address (tramp_mem, SImode, 12), fnaddr); | |
ca903bba R |
10427 | if (TARGET_HARVARD) |
10428 | { | |
dc557046 | 10429 | if (!TARGET_INLINE_IC_INVALIDATE |
6f8eec29 | 10430 | || (!(TARGET_SH4A_ARCH || TARGET_SH4_300) && TARGET_USERMODE)) |
73a4d10b R |
10431 | emit_library_call (function_symbol (NULL, "__ic_invalidate", |
10432 | FUNCTION_ORDINARY), | |
b757d36f | 10433 | LCT_NORMAL, VOIDmode, 1, tramp, SImode); |
ca903bba R |
10434 | else |
10435 | emit_insn (gen_ic_invalidate_line (tramp)); | |
10436 | } | |
10437 | } | |
10438 | ||
4977bab6 ZW |
10439 | /* FIXME: This is overly conservative. A SHcompact function that |
10440 | receives arguments ``by reference'' will have them stored in its | |
10441 | own stack frame, so it must not pass pointers or references to | |
10442 | these arguments to other functions by means of sibling calls. */ | |
73a4d10b R |
10443 | /* If PIC, we cannot make sibling calls to global functions |
10444 | because the PLT requires r12 to be live. */ | |
4977bab6 | 10445 | static bool |
cf277499 | 10446 | sh_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED) |
4977bab6 | 10447 | { |
73a4d10b | 10448 | return (1 |
4977bab6 | 10449 | && (! TARGET_SHCOMPACT |
38173d38 | 10450 | || crtl->args.info.stack_regs == 0) |
73a4d10b R |
10451 | && ! sh_cfun_interrupt_handler_p () |
10452 | && (! flag_pic | |
10453 | || (decl && ! TREE_PUBLIC (decl)) | |
10454 | || (decl && DECL_VISIBILITY (decl) != VISIBILITY_DEFAULT))); | |
4977bab6 | 10455 | } |
c1b92d09 R |
10456 | \f |
10457 | /* Machine specific built-in functions. */ | |
10458 | ||
10459 | struct builtin_description | |
10460 | { | |
10461 | const enum insn_code icode; | |
10462 | const char *const name; | |
10463 | int signature; | |
10464 | }; | |
10465 | ||
10466 | /* describe number and signedness of arguments; arg[0] == result | |
10467 | (1: unsigned, 2: signed, 4: don't care, 8: pointer 0: no argument */ | |
3b962ea0 | 10468 | /* 9: 64-bit pointer, 10: 32-bit pointer */ |
c1b92d09 R |
10469 | static const char signature_args[][4] = |
10470 | { | |
10471 | #define SH_BLTIN_V2SI2 0 | |
10472 | { 4, 4 }, | |
10473 | #define SH_BLTIN_V4HI2 1 | |
10474 | { 4, 4 }, | |
10475 | #define SH_BLTIN_V2SI3 2 | |
10476 | { 4, 4, 4 }, | |
10477 | #define SH_BLTIN_V4HI3 3 | |
10478 | { 4, 4, 4 }, | |
10479 | #define SH_BLTIN_V8QI3 4 | |
10480 | { 4, 4, 4 }, | |
10481 | #define SH_BLTIN_MAC_HISI 5 | |
10482 | { 1, 4, 4, 1 }, | |
10483 | #define SH_BLTIN_SH_HI 6 | |
10484 | { 4, 4, 1 }, | |
10485 | #define SH_BLTIN_SH_SI 7 | |
10486 | { 4, 4, 1 }, | |
10487 | #define SH_BLTIN_V4HI2V2SI 8 | |
10488 | { 4, 4, 4 }, | |
10489 | #define SH_BLTIN_V4HI2V8QI 9 | |
10490 | { 4, 4, 4 }, | |
10491 | #define SH_BLTIN_SISF 10 | |
10492 | { 4, 2 }, | |
10493 | #define SH_BLTIN_LDUA_L 11 | |
73a4d10b | 10494 | { 2, 10 }, |
c1b92d09 | 10495 | #define SH_BLTIN_LDUA_Q 12 |
73a4d10b | 10496 | { 1, 10 }, |
c1b92d09 | 10497 | #define SH_BLTIN_STUA_L 13 |
73a4d10b | 10498 | { 0, 10, 2 }, |
c1b92d09 | 10499 | #define SH_BLTIN_STUA_Q 14 |
73a4d10b R |
10500 | { 0, 10, 1 }, |
10501 | #define SH_BLTIN_LDUA_L64 15 | |
10502 | { 2, 9 }, | |
10503 | #define SH_BLTIN_LDUA_Q64 16 | |
10504 | { 1, 9 }, | |
10505 | #define SH_BLTIN_STUA_L64 17 | |
10506 | { 0, 9, 2 }, | |
10507 | #define SH_BLTIN_STUA_Q64 18 | |
10508 | { 0, 9, 1 }, | |
10509 | #define SH_BLTIN_NUM_SHARED_SIGNATURES 19 | |
10510 | #define SH_BLTIN_2 19 | |
10511 | #define SH_BLTIN_SU 19 | |
c1b92d09 | 10512 | { 1, 2 }, |
73a4d10b R |
10513 | #define SH_BLTIN_3 20 |
10514 | #define SH_BLTIN_SUS 20 | |
c1b92d09 | 10515 | { 2, 2, 1 }, |
73a4d10b | 10516 | #define SH_BLTIN_PSSV 21 |
c1b92d09 | 10517 | { 0, 8, 2, 2 }, |
73a4d10b R |
10518 | #define SH_BLTIN_XXUU 22 |
10519 | #define SH_BLTIN_UUUU 22 | |
c1b92d09 | 10520 | { 1, 1, 1, 1 }, |
73a4d10b | 10521 | #define SH_BLTIN_PV 23 |
c1b92d09 R |
10522 | { 0, 8 }, |
10523 | }; | |
24746a42 | 10524 | /* mcmv: operands considered unsigned. */ |
c1b92d09 | 10525 | /* mmulsum_wq, msad_ubq: result considered unsigned long long. */ |
24746a42 KH |
10526 | /* mperm: control value considered unsigned int. */ |
10527 | /* mshalds, mshard, mshards, mshlld, mshlrd: shift count is unsigned int. */ | |
c1b92d09 R |
10528 | /* mshards_q: returns signed short. */ |
10529 | /* nsb: takes long long arg, returns unsigned char. */ | |
10530 | static const struct builtin_description bdesc[] = | |
10531 | { | |
10532 | { CODE_FOR_absv2si2, "__builtin_absv2si2", SH_BLTIN_V2SI2 }, | |
10533 | { CODE_FOR_absv4hi2, "__builtin_absv4hi2", SH_BLTIN_V4HI2 }, | |
10534 | { CODE_FOR_addv2si3, "__builtin_addv2si3", SH_BLTIN_V2SI3 }, | |
10535 | { CODE_FOR_addv4hi3, "__builtin_addv4hi3", SH_BLTIN_V4HI3 }, | |
10536 | { CODE_FOR_ssaddv2si3,"__builtin_ssaddv2si3", SH_BLTIN_V2SI3 }, | |
10537 | { CODE_FOR_usaddv8qi3,"__builtin_usaddv8qi3", SH_BLTIN_V8QI3 }, | |
10538 | { CODE_FOR_ssaddv4hi3,"__builtin_ssaddv4hi3", SH_BLTIN_V4HI3 }, | |
73a4d10b | 10539 | { CODE_FOR_alloco_i, "__builtin_sh_media_ALLOCO", SH_BLTIN_PV }, |
c1b92d09 R |
10540 | { CODE_FOR_negcmpeqv8qi,"__builtin_sh_media_MCMPEQ_B", SH_BLTIN_V8QI3 }, |
10541 | { CODE_FOR_negcmpeqv2si,"__builtin_sh_media_MCMPEQ_L", SH_BLTIN_V2SI3 }, | |
10542 | { CODE_FOR_negcmpeqv4hi,"__builtin_sh_media_MCMPEQ_W", SH_BLTIN_V4HI3 }, | |
10543 | { CODE_FOR_negcmpgtuv8qi,"__builtin_sh_media_MCMPGT_UB", SH_BLTIN_V8QI3 }, | |
10544 | { CODE_FOR_negcmpgtv2si,"__builtin_sh_media_MCMPGT_L", SH_BLTIN_V2SI3 }, | |
10545 | { CODE_FOR_negcmpgtv4hi,"__builtin_sh_media_MCMPGT_W", SH_BLTIN_V4HI3 }, | |
10546 | { CODE_FOR_mcmv, "__builtin_sh_media_MCMV", SH_BLTIN_UUUU }, | |
10547 | { CODE_FOR_mcnvs_lw, "__builtin_sh_media_MCNVS_LW", SH_BLTIN_3 }, | |
10548 | { CODE_FOR_mcnvs_wb, "__builtin_sh_media_MCNVS_WB", SH_BLTIN_V4HI2V8QI }, | |
10549 | { CODE_FOR_mcnvs_wub, "__builtin_sh_media_MCNVS_WUB", SH_BLTIN_V4HI2V8QI }, | |
73a4d10b R |
10550 | { CODE_FOR_mextr1, "__builtin_sh_media_MEXTR1", SH_BLTIN_V8QI3 }, |
10551 | { CODE_FOR_mextr2, "__builtin_sh_media_MEXTR2", SH_BLTIN_V8QI3 }, | |
10552 | { CODE_FOR_mextr3, "__builtin_sh_media_MEXTR3", SH_BLTIN_V8QI3 }, | |
10553 | { CODE_FOR_mextr4, "__builtin_sh_media_MEXTR4", SH_BLTIN_V8QI3 }, | |
10554 | { CODE_FOR_mextr5, "__builtin_sh_media_MEXTR5", SH_BLTIN_V8QI3 }, | |
10555 | { CODE_FOR_mextr6, "__builtin_sh_media_MEXTR6", SH_BLTIN_V8QI3 }, | |
10556 | { CODE_FOR_mextr7, "__builtin_sh_media_MEXTR7", SH_BLTIN_V8QI3 }, | |
c1b92d09 R |
10557 | { CODE_FOR_mmacfx_wl, "__builtin_sh_media_MMACFX_WL", SH_BLTIN_MAC_HISI }, |
10558 | { CODE_FOR_mmacnfx_wl,"__builtin_sh_media_MMACNFX_WL", SH_BLTIN_MAC_HISI }, | |
10559 | { CODE_FOR_mulv2si3, "__builtin_mulv2si3", SH_BLTIN_V2SI3, }, | |
10560 | { CODE_FOR_mulv4hi3, "__builtin_mulv4hi3", SH_BLTIN_V4HI3 }, | |
10561 | { CODE_FOR_mmulfx_l, "__builtin_sh_media_MMULFX_L", SH_BLTIN_V2SI3 }, | |
10562 | { CODE_FOR_mmulfx_w, "__builtin_sh_media_MMULFX_W", SH_BLTIN_V4HI3 }, | |
10563 | { CODE_FOR_mmulfxrp_w,"__builtin_sh_media_MMULFXRP_W", SH_BLTIN_V4HI3 }, | |
10564 | { CODE_FOR_mmulhi_wl, "__builtin_sh_media_MMULHI_WL", SH_BLTIN_V4HI2V2SI }, | |
10565 | { CODE_FOR_mmullo_wl, "__builtin_sh_media_MMULLO_WL", SH_BLTIN_V4HI2V2SI }, | |
10566 | { CODE_FOR_mmulsum_wq,"__builtin_sh_media_MMULSUM_WQ", SH_BLTIN_XXUU }, | |
10567 | { CODE_FOR_mperm_w, "__builtin_sh_media_MPERM_W", SH_BLTIN_SH_HI }, | |
10568 | { CODE_FOR_msad_ubq, "__builtin_sh_media_MSAD_UBQ", SH_BLTIN_XXUU }, | |
10569 | { CODE_FOR_mshalds_l, "__builtin_sh_media_MSHALDS_L", SH_BLTIN_SH_SI }, | |
10570 | { CODE_FOR_mshalds_w, "__builtin_sh_media_MSHALDS_W", SH_BLTIN_SH_HI }, | |
10571 | { CODE_FOR_ashrv2si3, "__builtin_ashrv2si3", SH_BLTIN_SH_SI }, | |
10572 | { CODE_FOR_ashrv4hi3, "__builtin_ashrv4hi3", SH_BLTIN_SH_HI }, | |
10573 | { CODE_FOR_mshards_q, "__builtin_sh_media_MSHARDS_Q", SH_BLTIN_SUS }, | |
10574 | { CODE_FOR_mshfhi_b, "__builtin_sh_media_MSHFHI_B", SH_BLTIN_V8QI3 }, | |
10575 | { CODE_FOR_mshfhi_l, "__builtin_sh_media_MSHFHI_L", SH_BLTIN_V2SI3 }, | |
10576 | { CODE_FOR_mshfhi_w, "__builtin_sh_media_MSHFHI_W", SH_BLTIN_V4HI3 }, | |
10577 | { CODE_FOR_mshflo_b, "__builtin_sh_media_MSHFLO_B", SH_BLTIN_V8QI3 }, | |
10578 | { CODE_FOR_mshflo_l, "__builtin_sh_media_MSHFLO_L", SH_BLTIN_V2SI3 }, | |
10579 | { CODE_FOR_mshflo_w, "__builtin_sh_media_MSHFLO_W", SH_BLTIN_V4HI3 }, | |
10580 | { CODE_FOR_ashlv2si3, "__builtin_ashlv2si3", SH_BLTIN_SH_SI }, | |
10581 | { CODE_FOR_ashlv4hi3, "__builtin_ashlv4hi3", SH_BLTIN_SH_HI }, | |
10582 | { CODE_FOR_lshrv2si3, "__builtin_lshrv2si3", SH_BLTIN_SH_SI }, | |
10583 | { CODE_FOR_lshrv4hi3, "__builtin_lshrv4hi3", SH_BLTIN_SH_HI }, | |
10584 | { CODE_FOR_subv2si3, "__builtin_subv2si3", SH_BLTIN_V2SI3 }, | |
10585 | { CODE_FOR_subv4hi3, "__builtin_subv4hi3", SH_BLTIN_V4HI3 }, | |
10586 | { CODE_FOR_sssubv2si3,"__builtin_sssubv2si3", SH_BLTIN_V2SI3 }, | |
10587 | { CODE_FOR_ussubv8qi3,"__builtin_ussubv8qi3", SH_BLTIN_V8QI3 }, | |
10588 | { CODE_FOR_sssubv4hi3,"__builtin_sssubv4hi3", SH_BLTIN_V4HI3 }, | |
10589 | { CODE_FOR_fcosa_s, "__builtin_sh_media_FCOSA_S", SH_BLTIN_SISF }, | |
10590 | { CODE_FOR_fsina_s, "__builtin_sh_media_FSINA_S", SH_BLTIN_SISF }, | |
10591 | { CODE_FOR_fipr, "__builtin_sh_media_FIPR_S", SH_BLTIN_3 }, | |
10592 | { CODE_FOR_ftrv, "__builtin_sh_media_FTRV_S", SH_BLTIN_3 }, | |
73a4d10b R |
10593 | { CODE_FOR_mac_media, "__builtin_sh_media_FMAC_S", SH_BLTIN_3 }, |
10594 | { CODE_FOR_sqrtdf2, "__builtin_sh_media_FSQRT_D", SH_BLTIN_2 }, | |
10595 | { CODE_FOR_sqrtsf2, "__builtin_sh_media_FSQRT_S", SH_BLTIN_2 }, | |
c1b92d09 | 10596 | { CODE_FOR_fsrra_s, "__builtin_sh_media_FSRRA_S", SH_BLTIN_2 }, |
c1b92d09 R |
10597 | { CODE_FOR_ldhi_l, "__builtin_sh_media_LDHI_L", SH_BLTIN_LDUA_L }, |
10598 | { CODE_FOR_ldhi_q, "__builtin_sh_media_LDHI_Q", SH_BLTIN_LDUA_Q }, | |
10599 | { CODE_FOR_ldlo_l, "__builtin_sh_media_LDLO_L", SH_BLTIN_LDUA_L }, | |
10600 | { CODE_FOR_ldlo_q, "__builtin_sh_media_LDLO_Q", SH_BLTIN_LDUA_Q }, | |
10601 | { CODE_FOR_sthi_l, "__builtin_sh_media_STHI_L", SH_BLTIN_STUA_L }, | |
10602 | { CODE_FOR_sthi_q, "__builtin_sh_media_STHI_Q", SH_BLTIN_STUA_Q }, | |
10603 | { CODE_FOR_stlo_l, "__builtin_sh_media_STLO_L", SH_BLTIN_STUA_L }, | |
10604 | { CODE_FOR_stlo_q, "__builtin_sh_media_STLO_Q", SH_BLTIN_STUA_Q }, | |
73a4d10b R |
10605 | { CODE_FOR_ldhi_l64, "__builtin_sh_media_LDHI_L", SH_BLTIN_LDUA_L64 }, |
10606 | { CODE_FOR_ldhi_q64, "__builtin_sh_media_LDHI_Q", SH_BLTIN_LDUA_Q64 }, | |
10607 | { CODE_FOR_ldlo_l64, "__builtin_sh_media_LDLO_L", SH_BLTIN_LDUA_L64 }, | |
10608 | { CODE_FOR_ldlo_q64, "__builtin_sh_media_LDLO_Q", SH_BLTIN_LDUA_Q64 }, | |
10609 | { CODE_FOR_sthi_l64, "__builtin_sh_media_STHI_L", SH_BLTIN_STUA_L64 }, | |
10610 | { CODE_FOR_sthi_q64, "__builtin_sh_media_STHI_Q", SH_BLTIN_STUA_Q64 }, | |
10611 | { CODE_FOR_stlo_l64, "__builtin_sh_media_STLO_L", SH_BLTIN_STUA_L64 }, | |
10612 | { CODE_FOR_stlo_q64, "__builtin_sh_media_STLO_Q", SH_BLTIN_STUA_Q64 }, | |
c1b92d09 R |
10613 | { CODE_FOR_nsb, "__builtin_sh_media_NSB", SH_BLTIN_SU }, |
10614 | { CODE_FOR_byterev, "__builtin_sh_media_BYTEREV", SH_BLTIN_2 }, | |
73a4d10b | 10615 | { CODE_FOR_prefetch, "__builtin_sh_media_PREFO", SH_BLTIN_PSSV }, |
c1b92d09 R |
10616 | }; |
10617 | ||
10618 | static void | |
cf277499 | 10619 | sh_media_init_builtins (void) |
c1b92d09 R |
10620 | { |
10621 | tree shared[SH_BLTIN_NUM_SHARED_SIGNATURES]; | |
10622 | const struct builtin_description *d; | |
10623 | ||
a3acdc0c | 10624 | memset (shared, 0, sizeof shared); |
e97a46ce | 10625 | for (d = bdesc; d - bdesc < (int) ARRAY_SIZE (bdesc); d++) |
c1b92d09 | 10626 | { |
73a4d10b | 10627 | tree type, arg_type = 0; |
c1b92d09 R |
10628 | int signature = d->signature; |
10629 | int i; | |
10630 | ||
10631 | if (signature < SH_BLTIN_NUM_SHARED_SIGNATURES && shared[signature]) | |
10632 | type = shared[signature]; | |
10633 | else | |
10634 | { | |
10635 | int has_result = signature_args[signature][0] != 0; | |
10636 | ||
73a4d10b R |
10637 | if ((signature_args[signature][1] & 8) |
10638 | && (((signature_args[signature][1] & 1) && TARGET_SHMEDIA32) | |
10639 | || ((signature_args[signature][1] & 2) && TARGET_SHMEDIA64))) | |
c1b92d09 R |
10640 | continue; |
10641 | if (! TARGET_FPU_ANY | |
10642 | && FLOAT_MODE_P (insn_data[d->icode].operand[0].mode)) | |
10643 | continue; | |
10644 | type = void_list_node; | |
10645 | for (i = 3; ; i--) | |
10646 | { | |
10647 | int arg = signature_args[signature][i]; | |
10648 | int opno = i - 1 + has_result; | |
10649 | ||
73a4d10b | 10650 | if (arg & 8) |
c1b92d09 R |
10651 | arg_type = ptr_type_node; |
10652 | else if (arg) | |
73a4d10b R |
10653 | arg_type = (*lang_hooks.types.type_for_mode) |
10654 | (insn_data[d->icode].operand[opno].mode, | |
10655 | (arg & 1)); | |
c1b92d09 R |
10656 | else if (i) |
10657 | continue; | |
10658 | else | |
10659 | arg_type = void_type_node; | |
10660 | if (i == 0) | |
10661 | break; | |
10662 | type = tree_cons (NULL_TREE, arg_type, type); | |
10663 | } | |
10664 | type = build_function_type (arg_type, type); | |
10665 | if (signature < SH_BLTIN_NUM_SHARED_SIGNATURES) | |
10666 | shared[signature] = type; | |
10667 | } | |
c79efc4d RÁE |
10668 | add_builtin_function (d->name, type, d - bdesc, BUILT_IN_MD, |
10669 | NULL, NULL_TREE); | |
c1b92d09 R |
10670 | } |
10671 | } | |
10672 | ||
f676971a EC |
10673 | /* Implements target hook vector_mode_supported_p. */ |
10674 | bool | |
10675 | sh_vector_mode_supported_p (enum machine_mode mode) | |
10676 | { | |
10677 | if (TARGET_FPU_ANY | |
10678 | && ((mode == V2SFmode) | |
10679 | || (mode == V4SFmode) | |
10680 | || (mode == V16SFmode))) | |
10681 | return true; | |
10682 | ||
10683 | else if (TARGET_SHMEDIA | |
10684 | && ((mode == V8QImode) | |
10685 | || (mode == V2HImode) | |
10686 | || (mode == V4HImode) | |
10687 | || (mode == V2SImode))) | |
10688 | return true; | |
10689 | ||
10690 | return false; | |
10691 | } | |
10692 | ||
a1c496cb EC |
10693 | /* Implements target hook dwarf_calling_convention. Return an enum |
10694 | of dwarf_calling_convention. */ | |
10695 | int | |
3101faab | 10696 | sh_dwarf_calling_convention (const_tree func) |
a1c496cb EC |
10697 | { |
10698 | if (sh_attr_renesas_p (func)) | |
29832164 | 10699 | return DW_CC_GNU_renesas_sh; |
a1c496cb EC |
10700 | |
10701 | return DW_CC_normal; | |
10702 | } | |
10703 | ||
c1b92d09 | 10704 | static void |
cf277499 | 10705 | sh_init_builtins (void) |
c1b92d09 R |
10706 | { |
10707 | if (TARGET_SHMEDIA) | |
10708 | sh_media_init_builtins (); | |
10709 | } | |
10710 | ||
10711 | /* Expand an expression EXP that calls a built-in function, | |
10712 | with result going to TARGET if that's convenient | |
10713 | (and in mode MODE if that's convenient). | |
10714 | SUBTARGET may be used as the target for computing one of EXP's operands. | |
10715 | IGNORE is nonzero if the value is to be ignored. */ | |
10716 | ||
10717 | static rtx | |
cf277499 SB |
10718 | sh_expand_builtin (tree exp, rtx target, rtx subtarget ATTRIBUTE_UNUSED, |
10719 | enum machine_mode mode ATTRIBUTE_UNUSED, int ignore) | |
c1b92d09 | 10720 | { |
5039610b | 10721 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); |
c1b92d09 R |
10722 | unsigned int fcode = DECL_FUNCTION_CODE (fndecl); |
10723 | const struct builtin_description *d = &bdesc[fcode]; | |
10724 | enum insn_code icode = d->icode; | |
10725 | int signature = d->signature; | |
10726 | enum machine_mode tmode = VOIDmode; | |
10727 | int nop = 0, i; | |
10728 | rtx op[4]; | |
73a4d10b | 10729 | rtx pat = 0; |
c1b92d09 R |
10730 | |
10731 | if (signature_args[signature][0]) | |
10732 | { | |
10733 | if (ignore) | |
10734 | return 0; | |
10735 | ||
10736 | tmode = insn_data[icode].operand[0].mode; | |
10737 | if (! target | |
10738 | || GET_MODE (target) != tmode | |
10739 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
10740 | target = gen_reg_rtx (tmode); | |
10741 | op[nop++] = target; | |
10742 | } | |
10743 | else | |
10744 | target = 0; | |
10745 | ||
10746 | for (i = 1; i <= 3; i++, nop++) | |
10747 | { | |
10748 | tree arg; | |
10749 | enum machine_mode opmode, argmode; | |
73a4d10b | 10750 | tree optype; |
c1b92d09 R |
10751 | |
10752 | if (! signature_args[signature][i]) | |
10753 | break; | |
5039610b | 10754 | arg = CALL_EXPR_ARG (exp, i - 1); |
3578cf63 R |
10755 | if (arg == error_mark_node) |
10756 | return const0_rtx; | |
73a4d10b R |
10757 | if (signature_args[signature][i] & 8) |
10758 | { | |
10759 | opmode = ptr_mode; | |
10760 | optype = ptr_type_node; | |
10761 | } | |
10762 | else | |
10763 | { | |
10764 | opmode = insn_data[icode].operand[nop].mode; | |
10765 | optype = (*lang_hooks.types.type_for_mode) (opmode, 0); | |
10766 | } | |
c1b92d09 R |
10767 | argmode = TYPE_MODE (TREE_TYPE (arg)); |
10768 | if (argmode != opmode) | |
73a4d10b | 10769 | arg = build1 (NOP_EXPR, optype, arg); |
b757d36f | 10770 | op[nop] = expand_expr (arg, NULL_RTX, opmode, EXPAND_NORMAL); |
c1b92d09 R |
10771 | if (! (*insn_data[icode].operand[nop].predicate) (op[nop], opmode)) |
10772 | op[nop] = copy_to_mode_reg (opmode, op[nop]); | |
10773 | } | |
10774 | ||
10775 | switch (nop) | |
10776 | { | |
10777 | case 1: | |
10778 | pat = (*insn_data[d->icode].genfun) (op[0]); | |
10779 | break; | |
10780 | case 2: | |
10781 | pat = (*insn_data[d->icode].genfun) (op[0], op[1]); | |
10782 | break; | |
10783 | case 3: | |
10784 | pat = (*insn_data[d->icode].genfun) (op[0], op[1], op[2]); | |
10785 | break; | |
10786 | case 4: | |
10787 | pat = (*insn_data[d->icode].genfun) (op[0], op[1], op[2], op[3]); | |
10788 | break; | |
95c0af87 | 10789 | default: |
f5b9e7c9 | 10790 | gcc_unreachable (); |
c1b92d09 R |
10791 | } |
10792 | if (! pat) | |
10793 | return 0; | |
10794 | emit_insn (pat); | |
10795 | return target; | |
10796 | } | |
0ac78517 R |
10797 | |
10798 | void | |
cf277499 | 10799 | sh_expand_unop_v2sf (enum rtx_code code, rtx op0, rtx op1) |
0ac78517 R |
10800 | { |
10801 | rtx sel0 = const0_rtx; | |
10802 | rtx sel1 = const1_rtx; | |
cf277499 | 10803 | rtx (*fn) (rtx, rtx, rtx, rtx, rtx) = gen_unary_sf_op; |
0ac78517 R |
10804 | rtx op = gen_rtx_fmt_e (code, SFmode, op1); |
10805 | ||
10806 | emit_insn ((*fn) (op0, op1, op, sel0, sel0)); | |
10807 | emit_insn ((*fn) (op0, op1, op, sel1, sel1)); | |
10808 | } | |
10809 | ||
10810 | void | |
cf277499 | 10811 | sh_expand_binop_v2sf (enum rtx_code code, rtx op0, rtx op1, rtx op2) |
0ac78517 | 10812 | { |
0ac78517 R |
10813 | rtx op = gen_rtx_fmt_ee (code, SFmode, op1, op2); |
10814 | ||
f8be3f56 KK |
10815 | emit_insn (gen_binary_sf_op0 (op0, op1, op2, op)); |
10816 | emit_insn (gen_binary_sf_op1 (op0, op1, op2, op)); | |
0ac78517 R |
10817 | } |
10818 | ||
347146a4 KK |
10819 | /* Return true if hard register REGNO can hold a value of machine-mode MODE. |
10820 | We can allow any mode in any general register. The special registers | |
10821 | only allow SImode. Don't allow any mode in the PR. | |
10822 | ||
10823 | We cannot hold DCmode values in the XD registers because alter_reg | |
10824 | handles subregs of them incorrectly. We could work around this by | |
10825 | spacing the XD registers like the DR registers, but this would require | |
10826 | additional memory in every compilation to hold larger register vectors. | |
10827 | We could hold SFmode / SCmode values in XD registers, but that | |
10828 | would require a tertiary reload when reloading from / to memory, | |
10829 | and a secondary reload to reload from / to general regs; that | |
10830 | seems to be a loosing proposition. | |
10831 | ||
10832 | We want to allow TImode FP regs so that when V4SFmode is loaded as TImode, | |
10833 | it won't be ferried through GP registers first. */ | |
10834 | ||
10835 | bool | |
10836 | sh_hard_regno_mode_ok (unsigned int regno, enum machine_mode mode) | |
10837 | { | |
10838 | if (SPECIAL_REGISTER_P (regno)) | |
10839 | return mode == SImode; | |
10840 | ||
10841 | if (regno == FPUL_REG) | |
10842 | return (mode == SImode || mode == SFmode); | |
10843 | ||
10844 | if (FP_REGISTER_P (regno) && mode == SFmode) | |
10845 | return true; | |
10846 | ||
10847 | if (mode == V2SFmode) | |
10848 | { | |
10849 | if (((FP_REGISTER_P (regno) && (regno - FIRST_FP_REG) % 2 == 0) | |
10850 | || GENERAL_REGISTER_P (regno))) | |
10851 | return true; | |
10852 | else | |
10853 | return false; | |
10854 | } | |
10855 | ||
10856 | if (mode == V4SFmode) | |
10857 | { | |
10858 | if ((FP_REGISTER_P (regno) && (regno - FIRST_FP_REG) % 4 == 0) | |
10859 | || GENERAL_REGISTER_P (regno)) | |
10860 | return true; | |
10861 | else | |
10862 | return false; | |
10863 | } | |
10864 | ||
10865 | if (mode == V16SFmode) | |
10866 | { | |
10867 | if (TARGET_SHMEDIA) | |
10868 | { | |
10869 | if (FP_REGISTER_P (regno) && (regno - FIRST_FP_REG) % 16 == 0) | |
10870 | return true; | |
10871 | else | |
10872 | return false; | |
10873 | } | |
10874 | else | |
10875 | return regno == FIRST_XD_REG; | |
10876 | } | |
10877 | ||
10878 | if (FP_REGISTER_P (regno)) | |
10879 | { | |
10880 | if (mode == SFmode | |
10881 | || mode == SImode | |
10882 | || ((TARGET_SH2E || TARGET_SHMEDIA) && mode == SCmode) | |
10883 | || ((((TARGET_SH4 || TARGET_SH2A_DOUBLE) && mode == DFmode) | |
10884 | || mode == DCmode | |
10885 | || (TARGET_SHMEDIA | |
10886 | && (mode == DFmode || mode == DImode | |
10887 | || mode == V2SFmode || mode == TImode))) | |
10888 | && ((regno - FIRST_FP_REG) & 1) == 0) | |
10889 | || ((TARGET_SH4 || TARGET_SHMEDIA) && mode == TImode | |
10890 | && ((regno - FIRST_FP_REG) & 3) == 0)) | |
10891 | return true; | |
10892 | else | |
10893 | return false; | |
10894 | } | |
10895 | ||
10896 | if (XD_REGISTER_P (regno)) | |
10897 | return mode == DFmode; | |
10898 | ||
10899 | if (TARGET_REGISTER_P (regno)) | |
10900 | return (mode == DImode || mode == SImode || mode == PDImode); | |
10901 | ||
10902 | if (regno == PR_REG) | |
10903 | return mode == SImode; | |
10904 | ||
10905 | if (regno == FPSCR_REG) | |
10906 | return mode == PSImode; | |
10907 | ||
10908 | /* FIXME. This works around PR target/37633 for -O0. */ | |
10909 | if (!optimize && TARGET_SHMEDIA32 && GET_MODE_SIZE (mode) > 4) | |
10910 | { | |
10911 | unsigned int n = GET_MODE_SIZE (mode) / 8; | |
10912 | ||
10913 | if (regno >= FIRST_GENERAL_REG + 10 - n + 1 | |
10914 | && regno <= FIRST_GENERAL_REG + 14) | |
10915 | return false; | |
10916 | } | |
10917 | ||
10918 | return true; | |
10919 | } | |
10920 | ||
cff9f8d5 AH |
10921 | /* Return the class of registers for which a mode change from FROM to TO |
10922 | is invalid. */ | |
b0c42aed | 10923 | bool |
cf277499 | 10924 | sh_cannot_change_mode_class (enum machine_mode from, enum machine_mode to, |
0a2aaacc | 10925 | enum reg_class rclass) |
cff9f8d5 | 10926 | { |
312209c6 AO |
10927 | /* We want to enable the use of SUBREGs as a means to |
10928 | VEC_SELECT a single element of a vector. */ | |
10929 | if (to == SFmode && VECTOR_MODE_P (from) && GET_MODE_INNER (from) == SFmode) | |
0a2aaacc | 10930 | return (reg_classes_intersect_p (GENERAL_REGS, rclass)); |
312209c6 | 10931 | |
cff9f8d5 AH |
10932 | if (GET_MODE_SIZE (from) != GET_MODE_SIZE (to)) |
10933 | { | |
832a3292 KH |
10934 | if (TARGET_LITTLE_ENDIAN) |
10935 | { | |
10936 | if (GET_MODE_SIZE (to) < 8 || GET_MODE_SIZE (from) < 8) | |
0a2aaacc | 10937 | return reg_classes_intersect_p (DF_REGS, rclass); |
832a3292 KH |
10938 | } |
10939 | else | |
10940 | { | |
10941 | if (GET_MODE_SIZE (from) < 8) | |
0a2aaacc | 10942 | return reg_classes_intersect_p (DF_HI_REGS, rclass); |
832a3292 | 10943 | } |
cff9f8d5 | 10944 | } |
b0c42aed | 10945 | return 0; |
cff9f8d5 AH |
10946 | } |
10947 | ||
ea4210ef RS |
10948 | |
10949 | /* If ADDRESS refers to a CODE_LABEL, add NUSES to the number of times | |
10950 | that label is used. */ | |
10951 | ||
10952 | void | |
cf277499 | 10953 | sh_mark_label (rtx address, int nuses) |
ea4210ef RS |
10954 | { |
10955 | if (GOTOFF_P (address)) | |
10956 | { | |
10957 | /* Extract the label or symbol. */ | |
10958 | address = XEXP (address, 0); | |
10959 | if (GET_CODE (address) == PLUS) | |
10960 | address = XEXP (address, 0); | |
10961 | address = XVECEXP (address, 0, 0); | |
10962 | } | |
10963 | if (GET_CODE (address) == LABEL_REF | |
f3536097 | 10964 | && LABEL_P (XEXP (address, 0))) |
ea4210ef RS |
10965 | LABEL_NUSES (XEXP (address, 0)) += nuses; |
10966 | } | |
10967 | ||
95c0af87 JR |
10968 | /* Compute extra cost of moving data between one register class |
10969 | and another. */ | |
10970 | ||
ec555f32 R |
10971 | /* If SECONDARY*_RELOAD_CLASS says something about the src/dst pair, regclass |
10972 | uses this information. Hence, the general register <-> floating point | |
95c0af87 JR |
10973 | register information here is not used for SFmode. */ |
10974 | ||
10975 | int | |
cf277499 SB |
10976 | sh_register_move_cost (enum machine_mode mode, |
10977 | enum reg_class srcclass, enum reg_class dstclass) | |
95c0af87 JR |
10978 | { |
10979 | if (dstclass == T_REGS || dstclass == PR_REGS) | |
10980 | return 10; | |
10981 | ||
b1dcf523 DD |
10982 | if (dstclass == MAC_REGS && srcclass == MAC_REGS) |
10983 | return 4; | |
10984 | ||
ec555f32 R |
10985 | if (mode == SImode && ! TARGET_SHMEDIA && TARGET_FMOVD |
10986 | && REGCLASS_HAS_FP_REG (srcclass) | |
10987 | && REGCLASS_HAS_FP_REG (dstclass)) | |
10988 | return 4; | |
10989 | ||
9c24200c KK |
10990 | if (REGCLASS_HAS_FP_REG (dstclass) && srcclass == T_REGS) |
10991 | return ((TARGET_HARD_SH4 && !optimize_size) ? 10 : 7); | |
10992 | ||
a60a9553 | 10993 | if ((REGCLASS_HAS_FP_REG (dstclass) && srcclass == MAC_REGS) |
832a3292 | 10994 | || (dstclass == MAC_REGS && REGCLASS_HAS_FP_REG (srcclass))) |
a60a9553 R |
10995 | return 9; |
10996 | ||
95c0af87 JR |
10997 | if ((REGCLASS_HAS_FP_REG (dstclass) |
10998 | && REGCLASS_HAS_GENERAL_REG (srcclass)) | |
10999 | || (REGCLASS_HAS_GENERAL_REG (dstclass) | |
11000 | && REGCLASS_HAS_FP_REG (srcclass))) | |
832a3292 KH |
11001 | return ((TARGET_SHMEDIA ? 4 : TARGET_FMOVD ? 8 : 12) |
11002 | * ((GET_MODE_SIZE (mode) + 7) / 8U)); | |
95c0af87 JR |
11003 | |
11004 | if ((dstclass == FPUL_REGS | |
11005 | && REGCLASS_HAS_GENERAL_REG (srcclass)) | |
11006 | || (srcclass == FPUL_REGS | |
11007 | && REGCLASS_HAS_GENERAL_REG (dstclass))) | |
11008 | return 5; | |
11009 | ||
11010 | if ((dstclass == FPUL_REGS | |
11011 | && (srcclass == PR_REGS || srcclass == MAC_REGS || srcclass == T_REGS)) | |
f676971a | 11012 | || (srcclass == FPUL_REGS |
95c0af87 JR |
11013 | && (dstclass == PR_REGS || dstclass == MAC_REGS))) |
11014 | return 7; | |
11015 | ||
11016 | if ((srcclass == TARGET_REGS && ! REGCLASS_HAS_GENERAL_REG (dstclass)) | |
11017 | || ((dstclass) == TARGET_REGS && ! REGCLASS_HAS_GENERAL_REG (srcclass))) | |
11018 | return 20; | |
11019 | ||
73a4d10b R |
11020 | /* ??? ptabs faults on (value & 0x3) == 0x3 */ |
11021 | if (TARGET_SHMEDIA | |
11022 | && ((srcclass) == TARGET_REGS || (srcclass) == SIBCALL_REGS)) | |
11023 | { | |
7dd2f19b RS |
11024 | if (sh_gettrcost >= 0) |
11025 | return sh_gettrcost; | |
73a4d10b R |
11026 | else if (!TARGET_PT_FIXED) |
11027 | return 100; | |
11028 | } | |
11029 | ||
95c0af87 JR |
11030 | if ((srcclass == FPSCR_REGS && ! REGCLASS_HAS_GENERAL_REG (dstclass)) |
11031 | || (dstclass == FPSCR_REGS && ! REGCLASS_HAS_GENERAL_REG (srcclass))) | |
11032 | return 4; | |
11033 | ||
ec555f32 R |
11034 | if (TARGET_SHMEDIA |
11035 | || (TARGET_FMOVD | |
11036 | && ! REGCLASS_HAS_GENERAL_REG (srcclass) | |
11037 | && ! REGCLASS_HAS_GENERAL_REG (dstclass))) | |
11038 | return 2 * ((GET_MODE_SIZE (mode) + 7) / 8U); | |
11039 | ||
11040 | return 2 * ((GET_MODE_SIZE (mode) + 3) / 4U); | |
95c0af87 JR |
11041 | } |
11042 | ||
cf277499 | 11043 | static rtx emit_load_ptr (rtx, rtx); |
1aa03f38 R |
11044 | |
11045 | static rtx | |
cf277499 | 11046 | emit_load_ptr (rtx reg, rtx addr) |
1aa03f38 | 11047 | { |
57d38024 | 11048 | rtx mem = gen_const_mem (ptr_mode, addr); |
1aa03f38 R |
11049 | |
11050 | if (Pmode != ptr_mode) | |
11051 | mem = gen_rtx_SIGN_EXTEND (Pmode, mem); | |
11052 | return emit_move_insn (reg, mem); | |
11053 | } | |
11054 | ||
73a4d10b | 11055 | static void |
cf277499 SB |
11056 | sh_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED, |
11057 | HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset, | |
11058 | tree function) | |
1aa03f38 R |
11059 | { |
11060 | CUMULATIVE_ARGS cum; | |
11061 | int structure_value_byref = 0; | |
0a2aaacc | 11062 | rtx this_rtx, this_value, sibcall, insns, funexp; |
1aa03f38 | 11063 | tree funtype = TREE_TYPE (function); |
735cb76e | 11064 | int simple_add = CONST_OK_FOR_ADD (delta); |
1aa03f38 R |
11065 | int did_load = 0; |
11066 | rtx scratch0, scratch1, scratch2; | |
73a4d10b | 11067 | unsigned i; |
1aa03f38 R |
11068 | |
11069 | reload_completed = 1; | |
fe3ad572 | 11070 | epilogue_completed = 1; |
1aa03f38 R |
11071 | current_function_uses_only_leaf_regs = 1; |
11072 | ||
2e040219 | 11073 | emit_note (NOTE_INSN_PROLOGUE_END); |
1aa03f38 R |
11074 | |
11075 | /* Find the "this" pointer. We have such a wide range of ABIs for the | |
11076 | SH that it's best to do this completely machine independently. | |
f676971a | 11077 | "this" is passed as first argument, unless a structure return pointer |
1aa03f38 | 11078 | comes first, in which case "this" comes second. */ |
0f6937fe | 11079 | INIT_CUMULATIVE_ARGS (cum, funtype, NULL_RTX, 0, 1); |
1aa03f38 | 11080 | #ifndef PCC_STATIC_STRUCT_RETURN |
61f71b34 | 11081 | if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function)) |
1aa03f38 R |
11082 | structure_value_byref = 1; |
11083 | #endif /* not PCC_STATIC_STRUCT_RETURN */ | |
61f71b34 | 11084 | if (structure_value_byref && sh_struct_value_rtx (function, 0) == 0) |
f676971a | 11085 | { |
1aa03f38 R |
11086 | tree ptype = build_pointer_type (TREE_TYPE (funtype)); |
11087 | ||
11088 | FUNCTION_ARG_ADVANCE (cum, Pmode, ptype, 1); | |
11089 | } | |
0a2aaacc | 11090 | this_rtx = FUNCTION_ARG (cum, Pmode, ptr_type_node, 1); |
1aa03f38 R |
11091 | |
11092 | /* For SHcompact, we only have r0 for a scratch register: r1 is the | |
11093 | static chain pointer (even if you can't have nested virtual functions | |
11094 | right now, someone might implement them sometime), and the rest of the | |
11095 | registers are used for argument passing, are callee-saved, or reserved. */ | |
73a4d10b R |
11096 | /* We need to check call_used_regs / fixed_regs in case -fcall_saved-reg / |
11097 | -ffixed-reg has been used. */ | |
11098 | if (! call_used_regs[0] || fixed_regs[0]) | |
11099 | error ("r0 needs to be available as a call-clobbered register"); | |
1aa03f38 R |
11100 | scratch0 = scratch1 = scratch2 = gen_rtx_REG (Pmode, 0); |
11101 | if (! TARGET_SH5) | |
11102 | { | |
73a4d10b R |
11103 | if (call_used_regs[1] && ! fixed_regs[1]) |
11104 | scratch1 = gen_rtx_REG (ptr_mode, 1); | |
1aa03f38 R |
11105 | /* N.B., if not TARGET_HITACHI, register 2 is used to pass the pointer |
11106 | pointing where to return struct values. */ | |
73a4d10b R |
11107 | if (call_used_regs[3] && ! fixed_regs[3]) |
11108 | scratch2 = gen_rtx_REG (Pmode, 3); | |
1aa03f38 R |
11109 | } |
11110 | else if (TARGET_SHMEDIA) | |
11111 | { | |
73a4d10b R |
11112 | for (i = FIRST_GENERAL_REG; i <= LAST_GENERAL_REG; i++) |
11113 | if (i != REGNO (scratch0) && | |
11114 | call_used_regs[i] && ! fixed_regs[i] && ! FUNCTION_ARG_REGNO_P (i)) | |
11115 | { | |
11116 | scratch1 = gen_rtx_REG (ptr_mode, i); | |
11117 | break; | |
11118 | } | |
11119 | if (scratch1 == scratch0) | |
11120 | error ("Need a second call-clobbered general purpose register"); | |
11121 | for (i = FIRST_TARGET_REG; i <= LAST_TARGET_REG; i++) | |
11122 | if (call_used_regs[i] && ! fixed_regs[i]) | |
11123 | { | |
11124 | scratch2 = gen_rtx_REG (Pmode, i); | |
11125 | break; | |
11126 | } | |
11127 | if (scratch2 == scratch0) | |
11128 | error ("Need a call-clobbered target register"); | |
1aa03f38 R |
11129 | } |
11130 | ||
0a2aaacc | 11131 | this_value = plus_constant (this_rtx, delta); |
1aa03f38 R |
11132 | if (vcall_offset |
11133 | && (simple_add || scratch0 != scratch1) | |
11134 | && strict_memory_address_p (ptr_mode, this_value)) | |
11135 | { | |
11136 | emit_load_ptr (scratch0, this_value); | |
11137 | did_load = 1; | |
11138 | } | |
11139 | ||
11140 | if (!delta) | |
11141 | ; /* Do nothing. */ | |
11142 | else if (simple_add) | |
0a2aaacc | 11143 | emit_move_insn (this_rtx, this_value); |
1aa03f38 R |
11144 | else |
11145 | { | |
11146 | emit_move_insn (scratch1, GEN_INT (delta)); | |
0a2aaacc | 11147 | emit_insn (gen_add2_insn (this_rtx, scratch1)); |
1aa03f38 R |
11148 | } |
11149 | ||
11150 | if (vcall_offset) | |
11151 | { | |
11152 | rtx offset_addr; | |
11153 | ||
11154 | if (!did_load) | |
0a2aaacc | 11155 | emit_load_ptr (scratch0, this_rtx); |
1aa03f38 R |
11156 | |
11157 | offset_addr = plus_constant (scratch0, vcall_offset); | |
11158 | if (strict_memory_address_p (ptr_mode, offset_addr)) | |
11159 | ; /* Do nothing. */ | |
73a4d10b | 11160 | else if (! TARGET_SH5 && scratch0 != scratch1) |
1aa03f38 R |
11161 | { |
11162 | /* scratch0 != scratch1, and we have indexed loads. Get better | |
11163 | schedule by loading the offset into r1 and using an indexed | |
11164 | load - then the load of r1 can issue before the load from | |
0a2aaacc | 11165 | (this_rtx + delta) finishes. */ |
1aa03f38 R |
11166 | emit_move_insn (scratch1, GEN_INT (vcall_offset)); |
11167 | offset_addr = gen_rtx_PLUS (Pmode, scratch0, scratch1); | |
11168 | } | |
735cb76e | 11169 | else if (CONST_OK_FOR_ADD (vcall_offset)) |
1aa03f38 R |
11170 | { |
11171 | emit_insn (gen_add2_insn (scratch0, GEN_INT (vcall_offset))); | |
11172 | offset_addr = scratch0; | |
11173 | } | |
11174 | else if (scratch0 != scratch1) | |
11175 | { | |
11176 | emit_move_insn (scratch1, GEN_INT (vcall_offset)); | |
11177 | emit_insn (gen_add2_insn (scratch0, scratch1)); | |
11178 | offset_addr = scratch0; | |
11179 | } | |
11180 | else | |
f5b9e7c9 | 11181 | gcc_unreachable (); /* FIXME */ |
1aa03f38 R |
11182 | emit_load_ptr (scratch0, offset_addr); |
11183 | ||
832a3292 | 11184 | if (Pmode != ptr_mode) |
1aa03f38 | 11185 | scratch0 = gen_rtx_TRUNCATE (ptr_mode, scratch0); |
0a2aaacc | 11186 | emit_insn (gen_add2_insn (this_rtx, scratch0)); |
1aa03f38 R |
11187 | } |
11188 | ||
11189 | /* Generate a tail call to the target function. */ | |
11190 | if (! TREE_USED (function)) | |
11191 | { | |
11192 | assemble_external (function); | |
11193 | TREE_USED (function) = 1; | |
11194 | } | |
11195 | funexp = XEXP (DECL_RTL (function), 0); | |
73a4d10b R |
11196 | /* If the function is overridden, so is the thunk, hence we don't |
11197 | need GOT addressing even if this is a public symbol. */ | |
11198 | #if 0 | |
11199 | if (TARGET_SH1 && ! flag_weak) | |
11200 | sibcall = gen_sibcalli_thunk (funexp, const0_rtx); | |
11201 | else | |
11202 | #endif | |
11203 | if (TARGET_SH2 && flag_pic) | |
11204 | { | |
11205 | sibcall = gen_sibcall_pcrel (funexp, const0_rtx); | |
11206 | XEXP (XVECEXP (sibcall, 0, 2), 0) = scratch2; | |
11207 | } | |
11208 | else | |
11209 | { | |
11210 | if (TARGET_SHMEDIA && flag_pic) | |
11211 | { | |
11212 | funexp = gen_sym2PIC (funexp); | |
11213 | PUT_MODE (funexp, Pmode); | |
11214 | } | |
11215 | emit_move_insn (scratch2, funexp); | |
11216 | funexp = gen_rtx_MEM (FUNCTION_MODE, scratch2); | |
11217 | sibcall = gen_sibcall (funexp, const0_rtx, NULL_RTX); | |
11218 | } | |
11219 | sibcall = emit_call_insn (sibcall); | |
1aa03f38 | 11220 | SIBLING_CALL_P (sibcall) = 1; |
0a2aaacc | 11221 | use_reg (&CALL_INSN_FUNCTION_USAGE (sibcall), this_rtx); |
1aa03f38 R |
11222 | emit_barrier (); |
11223 | ||
a2855205 | 11224 | /* Run just enough of rest_of_compilation to do scheduling and get |
1aa03f38 R |
11225 | the insns emitted. Note that use_thunk calls |
11226 | assemble_start_function and assemble_end_function. */ | |
a2855205 | 11227 | |
55e092c4 | 11228 | insn_locators_alloc (); |
1aa03f38 R |
11229 | insns = get_insns (); |
11230 | ||
6fb5fa3c DB |
11231 | if (optimize > 0) |
11232 | { | |
11233 | if (! cfun->cfg) | |
eb6cbf91 | 11234 | init_flow (cfun); |
6fb5fa3c DB |
11235 | split_all_insns_noflow (); |
11236 | } | |
1aa03f38 | 11237 | |
18dbd950 | 11238 | sh_reorg (); |
1aa03f38 R |
11239 | |
11240 | if (optimize > 0 && flag_delayed_branch) | |
10d22567 | 11241 | dbr_schedule (insns); |
384a8760 | 11242 | |
1aa03f38 R |
11243 | shorten_branches (insns); |
11244 | final_start_function (insns, file, 1); | |
c9d691e9 | 11245 | final (insns, file, 1); |
1aa03f38 | 11246 | final_end_function (); |
2a7428c0 | 11247 | free_after_compilation (cfun); |
1aa03f38 | 11248 | |
1aa03f38 | 11249 | reload_completed = 0; |
fe3ad572 | 11250 | epilogue_completed = 0; |
1aa03f38 R |
11251 | } |
11252 | ||
90534361 | 11253 | rtx |
73a4d10b | 11254 | function_symbol (rtx target, const char *name, enum sh_function_kind kind) |
90534361 | 11255 | { |
73a4d10b R |
11256 | rtx sym; |
11257 | ||
11258 | /* If this is not an ordinary function, the name usually comes from a | |
11259 | string literal or an sprintf buffer. Make sure we use the same | |
11260 | string consistently, so that cse will be able to unify address loads. */ | |
11261 | if (kind != FUNCTION_ORDINARY) | |
11262 | name = IDENTIFIER_POINTER (get_identifier (name)); | |
11263 | sym = gen_rtx_SYMBOL_REF (Pmode, name); | |
90534361 | 11264 | SYMBOL_REF_FLAGS (sym) = SYMBOL_FLAG_FUNCTION; |
73a4d10b R |
11265 | if (flag_pic) |
11266 | switch (kind) | |
11267 | { | |
11268 | case FUNCTION_ORDINARY: | |
11269 | break; | |
11270 | case SFUNC_GOT: | |
11271 | { | |
11272 | rtx reg = target ? target : gen_reg_rtx (Pmode); | |
11273 | ||
11274 | emit_insn (gen_symGOT2reg (reg, sym)); | |
11275 | sym = reg; | |
11276 | break; | |
11277 | } | |
11278 | case SFUNC_STATIC: | |
11279 | { | |
11280 | /* ??? To allow cse to work, we use GOTOFF relocations. | |
11281 | we could add combiner patterns to transform this into | |
11282 | straight pc-relative calls with sym2PIC / bsrf when | |
11283 | label load and function call are still 1:1 and in the | |
11284 | same basic block during combine. */ | |
11285 | rtx reg = target ? target : gen_reg_rtx (Pmode); | |
11286 | ||
11287 | emit_insn (gen_symGOTOFF2reg (reg, sym)); | |
11288 | sym = reg; | |
11289 | break; | |
11290 | } | |
11291 | } | |
11292 | if (target && sym != target) | |
11293 | { | |
11294 | emit_move_insn (target, sym); | |
11295 | return target; | |
11296 | } | |
90534361 R |
11297 | return sym; |
11298 | } | |
11299 | ||
fada1961 R |
11300 | /* Find the number of a general purpose register in S. */ |
11301 | static int | |
11302 | scavenge_reg (HARD_REG_SET *s) | |
11303 | { | |
11304 | int r; | |
11305 | for (r = FIRST_GENERAL_REG; r <= LAST_GENERAL_REG; r++) | |
11306 | if (TEST_HARD_REG_BIT (*s, r)) | |
11307 | return r; | |
11308 | return -1; | |
11309 | } | |
11310 | ||
11311 | rtx | |
11312 | sh_get_pr_initial_val (void) | |
11313 | { | |
7d73a2ba R |
11314 | rtx val; |
11315 | ||
b869f904 R |
11316 | /* ??? Unfortunately, get_hard_reg_initial_val doesn't always work for the |
11317 | PR register on SHcompact, because it might be clobbered by the prologue. | |
7d73a2ba | 11318 | We check first if that is known to be the case. */ |
b869f904 | 11319 | if (TARGET_SHCOMPACT |
38173d38 | 11320 | && ((crtl->args.info.call_cookie |
7d73a2ba | 11321 | & ~ CALL_COOKIE_RET_TRAMP (1)) |
e3b5732b | 11322 | || crtl->saves_all_registers)) |
57d38024 | 11323 | return gen_frame_mem (SImode, return_address_pointer_rtx); |
7d73a2ba R |
11324 | |
11325 | /* If we haven't finished rtl generation, there might be a nonlocal label | |
11326 | that we haven't seen yet. | |
b3a13419 ILT |
11327 | ??? get_hard_reg_initial_val fails if it is called after register |
11328 | allocation has started, unless it has been called before for the | |
11329 | same register. And even then, we end in trouble if we didn't use | |
11330 | the register in the same basic block before. So call | |
11331 | get_hard_reg_initial_val now and wrap it in an unspec if we might | |
11332 | need to replace it. */ | |
dce20bbc R |
11333 | /* ??? We also must do this for TARGET_SH1 in general, because otherwise |
11334 | combine can put the pseudo returned by get_hard_reg_initial_val into | |
11335 | instructions that need a general purpose registers, which will fail to | |
11336 | be recognized when the pseudo becomes allocated to PR. */ | |
7d73a2ba R |
11337 | val |
11338 | = get_hard_reg_initial_val (Pmode, TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG); | |
dce20bbc | 11339 | if (TARGET_SH1) |
7d73a2ba R |
11340 | return gen_rtx_UNSPEC (SImode, gen_rtvec (1, val), UNSPEC_RA); |
11341 | return val; | |
fada1961 R |
11342 | } |
11343 | ||
3db1b434 | 11344 | int |
f90b7a5a | 11345 | sh_expand_t_scc (rtx operands[]) |
3db1b434 | 11346 | { |
f90b7a5a PB |
11347 | enum rtx_code code = GET_CODE (operands[1]); |
11348 | rtx target = operands[0]; | |
11349 | rtx op0 = operands[2]; | |
11350 | rtx op1 = operands[3]; | |
3db1b434 R |
11351 | rtx result = target; |
11352 | HOST_WIDE_INT val; | |
11353 | ||
f3536097 SZ |
11354 | if (!REG_P (op0) || REGNO (op0) != T_REG |
11355 | || !CONST_INT_P (op1)) | |
3db1b434 | 11356 | return 0; |
f3536097 | 11357 | if (!REG_P (result)) |
3db1b434 | 11358 | result = gen_reg_rtx (SImode); |
f90b7a5a | 11359 | val = INTVAL (op1); |
3db1b434 R |
11360 | if ((code == EQ && val == 1) || (code == NE && val == 0)) |
11361 | emit_insn (gen_movt (result)); | |
f326a6cb N |
11362 | else if (TARGET_SH2A && ((code == EQ && val == 0) |
11363 | || (code == NE && val == 1))) | |
c185de40 | 11364 | emit_insn (gen_xorsi3_movrt (result)); |
3db1b434 R |
11365 | else if ((code == EQ && val == 0) || (code == NE && val == 1)) |
11366 | { | |
c41c1387 | 11367 | emit_clobber (result); |
3db1b434 | 11368 | emit_insn (gen_subc (result, result, result)); |
a556fd39 | 11369 | emit_insn (gen_addsi3 (result, result, const1_rtx)); |
3db1b434 R |
11370 | } |
11371 | else if (code == EQ || code == NE) | |
11372 | emit_insn (gen_move_insn (result, GEN_INT (code == NE))); | |
11373 | else | |
11374 | return 0; | |
11375 | if (result != target) | |
11376 | emit_move_insn (target, result); | |
11377 | return 1; | |
11378 | } | |
11379 | ||
07d7d2f4 R |
11380 | /* INSN is an sfunc; return the rtx that describes the address used. */ |
11381 | static rtx | |
11382 | extract_sfunc_addr (rtx insn) | |
11383 | { | |
11384 | rtx pattern, part = NULL_RTX; | |
11385 | int len, i; | |
11386 | ||
11387 | pattern = PATTERN (insn); | |
11388 | len = XVECLEN (pattern, 0); | |
11389 | for (i = 0; i < len; i++) | |
11390 | { | |
11391 | part = XVECEXP (pattern, 0, i); | |
11392 | if (GET_CODE (part) == USE && GET_MODE (XEXP (part, 0)) == Pmode | |
11393 | && GENERAL_REGISTER_P (true_regnum (XEXP (part, 0)))) | |
11394 | return XEXP (part, 0); | |
11395 | } | |
f5b9e7c9 NS |
11396 | gcc_assert (GET_CODE (XVECEXP (pattern, 0, 0)) == UNSPEC_VOLATILE); |
11397 | return XVECEXP (XVECEXP (pattern, 0, 0), 0, 1); | |
07d7d2f4 R |
11398 | } |
11399 | ||
11400 | /* Verify that the register in use_sfunc_addr still agrees with the address | |
11401 | used in the sfunc. This prevents fill_slots_from_thread from changing | |
11402 | use_sfunc_addr. | |
11403 | INSN is the use_sfunc_addr instruction, and REG is the register it | |
11404 | guards. */ | |
11405 | int | |
11406 | check_use_sfunc_addr (rtx insn, rtx reg) | |
11407 | { | |
11408 | /* Search for the sfunc. It should really come right after INSN. */ | |
11409 | while ((insn = NEXT_INSN (insn))) | |
11410 | { | |
f3536097 | 11411 | if (LABEL_P (insn) || JUMP_P (insn)) |
07d7d2f4 R |
11412 | break; |
11413 | if (! INSN_P (insn)) | |
11414 | continue; | |
f676971a | 11415 | |
07d7d2f4 R |
11416 | if (GET_CODE (PATTERN (insn)) == SEQUENCE) |
11417 | insn = XVECEXP (PATTERN (insn), 0, 0); | |
11418 | if (GET_CODE (PATTERN (insn)) != PARALLEL | |
11419 | || get_attr_type (insn) != TYPE_SFUNC) | |
11420 | continue; | |
11421 | return rtx_equal_p (extract_sfunc_addr (insn), reg); | |
11422 | } | |
f5b9e7c9 | 11423 | gcc_unreachable (); |
07d7d2f4 R |
11424 | } |
11425 | ||
312209c6 AO |
11426 | /* This function returns a constant rtx that represents pi / 2**15 in |
11427 | SFmode. it's used to scale SFmode angles, in radians, to a | |
11428 | fixed-point signed 16.16-bit fraction of a full circle, i.e., 2*pi | |
11429 | maps to 0x10000). */ | |
11430 | ||
11431 | static GTY(()) rtx sh_fsca_sf2int_rtx; | |
11432 | ||
11433 | rtx | |
11434 | sh_fsca_sf2int (void) | |
11435 | { | |
11436 | if (! sh_fsca_sf2int_rtx) | |
11437 | { | |
11438 | REAL_VALUE_TYPE rv; | |
11439 | ||
11440 | real_from_string (&rv, "10430.378350470453"); | |
11441 | sh_fsca_sf2int_rtx = const_double_from_real_value (rv, SFmode); | |
11442 | } | |
11443 | ||
11444 | return sh_fsca_sf2int_rtx; | |
11445 | } | |
f676971a | 11446 | |
312209c6 AO |
11447 | /* This function returns a constant rtx that represents pi / 2**15 in |
11448 | DFmode. it's used to scale DFmode angles, in radians, to a | |
11449 | fixed-point signed 16.16-bit fraction of a full circle, i.e., 2*pi | |
11450 | maps to 0x10000). */ | |
11451 | ||
11452 | static GTY(()) rtx sh_fsca_df2int_rtx; | |
11453 | ||
11454 | rtx | |
11455 | sh_fsca_df2int (void) | |
11456 | { | |
11457 | if (! sh_fsca_df2int_rtx) | |
11458 | { | |
11459 | REAL_VALUE_TYPE rv; | |
11460 | ||
11461 | real_from_string (&rv, "10430.378350470453"); | |
11462 | sh_fsca_df2int_rtx = const_double_from_real_value (rv, DFmode); | |
11463 | } | |
11464 | ||
11465 | return sh_fsca_df2int_rtx; | |
11466 | } | |
f676971a | 11467 | |
312209c6 AO |
11468 | /* This function returns a constant rtx that represents 2**15 / pi in |
11469 | SFmode. it's used to scale a fixed-point signed 16.16-bit fraction | |
11470 | of a full circle back to a SFmode value, i.e., 0x10000 maps to | |
11471 | 2*pi). */ | |
11472 | ||
11473 | static GTY(()) rtx sh_fsca_int2sf_rtx; | |
11474 | ||
11475 | rtx | |
11476 | sh_fsca_int2sf (void) | |
11477 | { | |
11478 | if (! sh_fsca_int2sf_rtx) | |
11479 | { | |
11480 | REAL_VALUE_TYPE rv; | |
11481 | ||
11482 | real_from_string (&rv, "9.587379924285257e-5"); | |
11483 | sh_fsca_int2sf_rtx = const_double_from_real_value (rv, SFmode); | |
11484 | } | |
11485 | ||
11486 | return sh_fsca_int2sf_rtx; | |
11487 | } | |
cc15e98f | 11488 | |
8b0add27 | 11489 | /* Initialize the CUMULATIVE_ARGS structure. */ |
cc15e98f NC |
11490 | |
11491 | void | |
11492 | sh_init_cumulative_args (CUMULATIVE_ARGS * pcum, | |
11493 | tree fntype, | |
11494 | rtx libname ATTRIBUTE_UNUSED, | |
11495 | tree fndecl, | |
11496 | signed int n_named_args, | |
11497 | enum machine_mode mode) | |
11498 | { | |
11499 | pcum->arg_count [(int) SH_ARG_FLOAT] = 0; | |
11500 | pcum->free_single_fp_reg = 0; | |
11501 | pcum->stack_regs = 0; | |
11502 | pcum->byref_regs = 0; | |
11503 | pcum->byref = 0; | |
11504 | pcum->outgoing = (n_named_args == -1) ? 0 : 1; | |
11505 | ||
11506 | /* XXX - Should we check TARGET_HITACHI here ??? */ | |
11507 | pcum->renesas_abi = sh_attr_renesas_p (fntype) ? 1 : 0; | |
11508 | ||
11509 | if (fntype) | |
11510 | { | |
11511 | pcum->force_mem = ((TARGET_HITACHI || pcum->renesas_abi) | |
11512 | && aggregate_value_p (TREE_TYPE (fntype), fndecl)); | |
11513 | pcum->prototype_p = TYPE_ARG_TYPES (fntype) ? TRUE : FALSE; | |
11514 | pcum->arg_count [(int) SH_ARG_INT] | |
11515 | = TARGET_SH5 && aggregate_value_p (TREE_TYPE (fntype), fndecl); | |
11516 | ||
11517 | pcum->call_cookie | |
11518 | = CALL_COOKIE_RET_TRAMP (TARGET_SHCOMPACT | |
11519 | && pcum->arg_count [(int) SH_ARG_INT] == 0 | |
11520 | && (TYPE_MODE (TREE_TYPE (fntype)) == BLKmode | |
11521 | ? int_size_in_bytes (TREE_TYPE (fntype)) | |
11522 | : GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (fntype)))) > 4 | |
11523 | && (BASE_RETURN_VALUE_REG (TYPE_MODE (TREE_TYPE (fntype))) | |
11524 | == FIRST_RET_REG)); | |
11525 | } | |
11526 | else | |
11527 | { | |
11528 | pcum->arg_count [(int) SH_ARG_INT] = 0; | |
11529 | pcum->prototype_p = FALSE; | |
11530 | if (mode != VOIDmode) | |
11531 | { | |
29832164 | 11532 | pcum->call_cookie = |
cc15e98f NC |
11533 | CALL_COOKIE_RET_TRAMP (TARGET_SHCOMPACT |
11534 | && GET_MODE_SIZE (mode) > 4 | |
11535 | && BASE_RETURN_VALUE_REG (mode) == FIRST_RET_REG); | |
11536 | ||
11537 | /* If the default ABI is the Renesas ABI then all library | |
11538 | calls must assume that the library will be using the | |
11539 | Renesas ABI. So if the function would return its result | |
11540 | in memory then we must force the address of this memory | |
11541 | block onto the stack. Ideally we would like to call | |
11542 | targetm.calls.return_in_memory() here but we do not have | |
8b0add27 | 11543 | the TYPE or the FNDECL available so we synthesize the |
cc15e98f NC |
11544 | contents of that function as best we can. */ |
11545 | pcum->force_mem = | |
c0fb94d7 | 11546 | (TARGET_DEFAULT & MASK_HITACHI) |
cc15e98f NC |
11547 | && (mode == BLKmode |
11548 | || (GET_MODE_SIZE (mode) > 4 | |
11549 | && !(mode == DFmode | |
11550 | && TARGET_FPU_DOUBLE))); | |
11551 | } | |
11552 | else | |
11553 | { | |
11554 | pcum->call_cookie = 0; | |
11555 | pcum->force_mem = FALSE; | |
11556 | } | |
11557 | } | |
11558 | } | |
11559 | ||
73a4d10b R |
11560 | /* Replace any occurrence of FROM(n) in X with TO(n). The function does |
11561 | not enter into CONST_DOUBLE for the replace. | |
11562 | ||
11563 | Note that copying is not done so X must not be shared unless all copies | |
11564 | are to be modified. | |
11565 | ||
11566 | This is like replace_rtx, except that we operate on N_REPLACEMENTS | |
30dc60c7 | 11567 | replacements simultaneously - FROM(n) is replacements[n*2] and to(n) is |
73a4d10b R |
11568 | replacements[n*2+1] - and that we take mode changes into account. |
11569 | ||
30dc60c7 | 11570 | If a replacement is ambiguous, return NULL_RTX. |
73a4d10b R |
11571 | |
11572 | If MODIFY is zero, don't modify any rtl in place, | |
11573 | just return zero or nonzero for failure / success. */ | |
11574 | ||
11575 | rtx | |
11576 | replace_n_hard_rtx (rtx x, rtx *replacements, int n_replacements, int modify) | |
11577 | { | |
11578 | int i, j; | |
11579 | const char *fmt; | |
11580 | ||
11581 | /* The following prevents loops occurrence when we change MEM in | |
11582 | CONST_DOUBLE onto the same CONST_DOUBLE. */ | |
11583 | if (x != 0 && GET_CODE (x) == CONST_DOUBLE) | |
11584 | return x; | |
11585 | ||
11586 | for (i = n_replacements - 1; i >= 0 ; i--) | |
11587 | if (x == replacements[i*2] && GET_MODE (x) == GET_MODE (replacements[i*2+1])) | |
11588 | return replacements[i*2+1]; | |
11589 | ||
11590 | /* Allow this function to make replacements in EXPR_LISTs. */ | |
11591 | if (x == 0) | |
11592 | return 0; | |
11593 | ||
11594 | if (GET_CODE (x) == SUBREG) | |
11595 | { | |
0a2aaacc | 11596 | rtx new_rtx = replace_n_hard_rtx (SUBREG_REG (x), replacements, |
73a4d10b R |
11597 | n_replacements, modify); |
11598 | ||
f3536097 | 11599 | if (CONST_INT_P (new_rtx)) |
73a4d10b | 11600 | { |
0a2aaacc | 11601 | x = simplify_subreg (GET_MODE (x), new_rtx, |
73a4d10b R |
11602 | GET_MODE (SUBREG_REG (x)), |
11603 | SUBREG_BYTE (x)); | |
11604 | if (! x) | |
11605 | abort (); | |
11606 | } | |
11607 | else if (modify) | |
0a2aaacc | 11608 | SUBREG_REG (x) = new_rtx; |
73a4d10b R |
11609 | |
11610 | return x; | |
11611 | } | |
f3536097 | 11612 | else if (REG_P (x)) |
73a4d10b R |
11613 | { |
11614 | unsigned regno = REGNO (x); | |
11615 | unsigned nregs = (regno < FIRST_PSEUDO_REGISTER | |
11616 | ? HARD_REGNO_NREGS (regno, GET_MODE (x)) : 1); | |
11617 | rtx result = NULL_RTX; | |
11618 | ||
11619 | for (i = n_replacements - 1; i >= 0; i--) | |
11620 | { | |
11621 | rtx from = replacements[i*2]; | |
11622 | rtx to = replacements[i*2+1]; | |
11623 | unsigned from_regno, from_nregs, to_regno, new_regno; | |
11624 | ||
f3536097 | 11625 | if (!REG_P (from)) |
73a4d10b R |
11626 | continue; |
11627 | from_regno = REGNO (from); | |
11628 | from_nregs = (from_regno < FIRST_PSEUDO_REGISTER | |
11629 | ? HARD_REGNO_NREGS (from_regno, GET_MODE (from)) : 1); | |
11630 | if (regno < from_regno + from_nregs && regno + nregs > from_regno) | |
11631 | { | |
11632 | if (regno < from_regno | |
11633 | || regno + nregs > from_regno + nregs | |
f3536097 | 11634 | || !REG_P (to) |
73a4d10b R |
11635 | || result) |
11636 | return NULL_RTX; | |
11637 | to_regno = REGNO (to); | |
11638 | if (to_regno < FIRST_PSEUDO_REGISTER) | |
11639 | { | |
11640 | new_regno = regno + to_regno - from_regno; | |
11641 | if ((unsigned) HARD_REGNO_NREGS (new_regno, GET_MODE (x)) | |
11642 | != nregs) | |
11643 | return NULL_RTX; | |
11644 | result = gen_rtx_REG (GET_MODE (x), new_regno); | |
11645 | } | |
11646 | else if (GET_MODE (x) <= GET_MODE (to)) | |
11647 | result = gen_lowpart_common (GET_MODE (x), to); | |
11648 | else | |
11649 | result = gen_lowpart_SUBREG (GET_MODE (x), to); | |
11650 | } | |
11651 | } | |
11652 | return result ? result : x; | |
11653 | } | |
11654 | else if (GET_CODE (x) == ZERO_EXTEND) | |
11655 | { | |
0a2aaacc | 11656 | rtx new_rtx = replace_n_hard_rtx (XEXP (x, 0), replacements, |
73a4d10b R |
11657 | n_replacements, modify); |
11658 | ||
f3536097 | 11659 | if (CONST_INT_P (new_rtx)) |
73a4d10b R |
11660 | { |
11661 | x = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x), | |
0a2aaacc | 11662 | new_rtx, GET_MODE (XEXP (x, 0))); |
73a4d10b R |
11663 | if (! x) |
11664 | abort (); | |
11665 | } | |
11666 | else if (modify) | |
0a2aaacc | 11667 | XEXP (x, 0) = new_rtx; |
73a4d10b R |
11668 | |
11669 | return x; | |
11670 | } | |
11671 | ||
11672 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
11673 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
11674 | { | |
0a2aaacc | 11675 | rtx new_rtx; |
73a4d10b R |
11676 | |
11677 | if (fmt[i] == 'e') | |
11678 | { | |
0a2aaacc | 11679 | new_rtx = replace_n_hard_rtx (XEXP (x, i), replacements, |
73a4d10b | 11680 | n_replacements, modify); |
0a2aaacc | 11681 | if (!new_rtx) |
73a4d10b R |
11682 | return NULL_RTX; |
11683 | if (modify) | |
0a2aaacc | 11684 | XEXP (x, i) = new_rtx; |
73a4d10b R |
11685 | } |
11686 | else if (fmt[i] == 'E') | |
11687 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
11688 | { | |
0a2aaacc | 11689 | new_rtx = replace_n_hard_rtx (XVECEXP (x, i, j), replacements, |
73a4d10b | 11690 | n_replacements, modify); |
0a2aaacc | 11691 | if (!new_rtx) |
73a4d10b R |
11692 | return NULL_RTX; |
11693 | if (modify) | |
0a2aaacc | 11694 | XVECEXP (x, i, j) = new_rtx; |
73a4d10b R |
11695 | } |
11696 | } | |
11697 | ||
11698 | return x; | |
11699 | } | |
11700 | ||
11701 | rtx | |
11702 | sh_gen_truncate (enum machine_mode mode, rtx x, int need_sign_ext) | |
11703 | { | |
11704 | enum rtx_code code = TRUNCATE; | |
11705 | ||
11706 | if (GET_CODE (x) == ZERO_EXTEND || GET_CODE (x) == SIGN_EXTEND) | |
11707 | { | |
11708 | rtx inner = XEXP (x, 0); | |
11709 | enum machine_mode inner_mode = GET_MODE (inner); | |
11710 | ||
11711 | if (inner_mode == mode) | |
11712 | return inner; | |
11713 | else if (GET_MODE_SIZE (inner_mode) >= GET_MODE_SIZE (mode)) | |
11714 | x = inner; | |
11715 | else if (GET_MODE_SIZE (inner_mode) < GET_MODE_SIZE (mode) | |
11716 | && (! need_sign_ext || GET_CODE (x) == SIGN_EXTEND)) | |
11717 | { | |
11718 | code = GET_CODE (x); | |
11719 | x = inner; | |
11720 | } | |
11721 | } | |
11722 | return gen_rtx_fmt_e (code, mode, x); | |
11723 | } | |
11724 | ||
11725 | /* called via for_each_rtx after reload, to clean up truncates of | |
11726 | registers that span multiple actual hard registers. */ | |
11727 | int | |
11728 | shmedia_cleanup_truncate (rtx *p, void *n_changes) | |
11729 | { | |
11730 | rtx x = *p, reg; | |
11731 | ||
11732 | if (GET_CODE (x) != TRUNCATE) | |
11733 | return 0; | |
11734 | reg = XEXP (x, 0); | |
f3536097 | 11735 | if (GET_MODE_SIZE (GET_MODE (reg)) > 8 && REG_P (reg)) |
73a4d10b R |
11736 | { |
11737 | enum machine_mode reg_mode = GET_MODE (reg); | |
11738 | XEXP (x, 0) = simplify_subreg (DImode, reg, reg_mode, | |
11739 | subreg_lowpart_offset (DImode, reg_mode)); | |
11740 | *(int*) n_changes += 1; | |
11741 | return -1; | |
11742 | } | |
11743 | return 0; | |
11744 | } | |
11745 | ||
11746 | /* Load and store depend on the highpart of the address. However, | |
11747 | set_attr_alternative does not give well-defined results before reload, | |
11748 | so we must look at the rtl ourselves to see if any of the feeding | |
11749 | registers is used in a memref. */ | |
11750 | ||
11751 | /* Called by sh_contains_memref_p via for_each_rtx. */ | |
11752 | static int | |
11753 | sh_contains_memref_p_1 (rtx *loc, void *data ATTRIBUTE_UNUSED) | |
11754 | { | |
f3536097 | 11755 | return (MEM_P (*loc)); |
73a4d10b R |
11756 | } |
11757 | ||
a4d05547 | 11758 | /* Return nonzero iff INSN contains a MEM. */ |
73a4d10b R |
11759 | int |
11760 | sh_contains_memref_p (rtx insn) | |
11761 | { | |
11762 | return for_each_rtx (&PATTERN (insn), &sh_contains_memref_p_1, NULL); | |
11763 | } | |
11764 | ||
5c7cafa8 CB |
11765 | /* Return nonzero iff INSN loads a banked register. */ |
11766 | int | |
11767 | sh_loads_bankedreg_p (rtx insn) | |
11768 | { | |
11769 | if (GET_CODE (PATTERN (insn)) == SET) | |
11770 | { | |
11771 | rtx op = SET_DEST (PATTERN(insn)); | |
11772 | if (REG_P (op) && BANKED_REGISTER_P (REGNO (op))) | |
11773 | return 1; | |
11774 | } | |
11775 | ||
11776 | return 0; | |
11777 | } | |
11778 | ||
73a4d10b R |
11779 | /* FNADDR is the MEM expression from a call expander. Return an address |
11780 | to use in an SHmedia insn pattern. */ | |
11781 | rtx | |
11782 | shmedia_prepare_call_address (rtx fnaddr, int is_sibcall) | |
11783 | { | |
11784 | int is_sym; | |
11785 | ||
11786 | fnaddr = XEXP (fnaddr, 0); | |
11787 | is_sym = GET_CODE (fnaddr) == SYMBOL_REF; | |
11788 | if (flag_pic && is_sym) | |
11789 | { | |
11790 | if (! SYMBOL_REF_LOCAL_P (fnaddr)) | |
11791 | { | |
11792 | rtx reg = gen_reg_rtx (Pmode); | |
11793 | ||
11794 | /* We must not use GOTPLT for sibcalls, because PIC_REG | |
11795 | must be restored before the PLT code gets to run. */ | |
11796 | if (is_sibcall) | |
11797 | emit_insn (gen_symGOT2reg (reg, fnaddr)); | |
11798 | else | |
11799 | emit_insn (gen_symGOTPLT2reg (reg, fnaddr)); | |
11800 | fnaddr = reg; | |
11801 | } | |
11802 | else | |
11803 | { | |
11804 | fnaddr = gen_sym2PIC (fnaddr); | |
11805 | PUT_MODE (fnaddr, Pmode); | |
11806 | } | |
11807 | } | |
11808 | /* If ptabs might trap, make this visible to the rest of the compiler. | |
11809 | We generally assume that symbols pertain to valid locations, but | |
11810 | it is possible to generate invalid symbols with asm or linker tricks. | |
30dc60c7 | 11811 | In a list of functions where each returns its successor, an invalid |
73a4d10b R |
11812 | symbol might denote an empty list. */ |
11813 | if (!TARGET_PT_FIXED | |
11814 | && (!is_sym || TARGET_INVALID_SYMBOLS) | |
11815 | && (!REG_P (fnaddr) || ! TARGET_REGISTER_P (REGNO (fnaddr)))) | |
11816 | { | |
11817 | rtx tr = gen_reg_rtx (PDImode); | |
11818 | ||
11819 | emit_insn (gen_ptabs (tr, fnaddr)); | |
11820 | fnaddr = tr; | |
11821 | } | |
11822 | else if (! target_reg_operand (fnaddr, Pmode)) | |
11823 | fnaddr = copy_to_mode_reg (Pmode, fnaddr); | |
11824 | return fnaddr; | |
11825 | } | |
11826 | ||
8a99f6f9 | 11827 | enum reg_class |
0a2aaacc | 11828 | sh_secondary_reload (bool in_p, rtx x, enum reg_class rclass, |
8a99f6f9 R |
11829 | enum machine_mode mode, secondary_reload_info *sri) |
11830 | { | |
11831 | if (in_p) | |
11832 | { | |
0a2aaacc | 11833 | if (REGCLASS_HAS_FP_REG (rclass) |
8a99f6f9 R |
11834 | && ! TARGET_SHMEDIA |
11835 | && immediate_operand ((x), mode) | |
11836 | && ! ((fp_zero_operand (x) || fp_one_operand (x)) | |
11837 | && mode == SFmode && fldi_ok ())) | |
11838 | switch (mode) | |
11839 | { | |
11840 | case SFmode: | |
11841 | sri->icode = CODE_FOR_reload_insf__frn; | |
11842 | return NO_REGS; | |
11843 | case DFmode: | |
11844 | sri->icode = CODE_FOR_reload_indf__frn; | |
11845 | return NO_REGS; | |
11846 | case SImode: | |
11847 | /* ??? If we knew that we are in the appropriate mode - | |
11848 | single precision - we could use a reload pattern directly. */ | |
11849 | return FPUL_REGS; | |
11850 | default: | |
11851 | abort (); | |
11852 | } | |
0a2aaacc | 11853 | if (rclass == FPUL_REGS |
f3536097 | 11854 | && ((REG_P (x) |
8a99f6f9 R |
11855 | && (REGNO (x) == MACL_REG || REGNO (x) == MACH_REG |
11856 | || REGNO (x) == T_REG)) | |
11857 | || GET_CODE (x) == PLUS)) | |
11858 | return GENERAL_REGS; | |
0a2aaacc | 11859 | if (rclass == FPUL_REGS && immediate_operand (x, mode)) |
8a99f6f9 | 11860 | { |
d973ac5a | 11861 | if (satisfies_constraint_I08 (x) || fp_zero_operand (x)) |
8a99f6f9 | 11862 | return GENERAL_REGS; |
d973ac5a KK |
11863 | else if (mode == SFmode) |
11864 | return FP_REGS; | |
8a99f6f9 R |
11865 | sri->icode = CODE_FOR_reload_insi__i_fpul; |
11866 | return NO_REGS; | |
11867 | } | |
0a2aaacc | 11868 | if (rclass == FPSCR_REGS |
f3536097 SZ |
11869 | && ((REG_P (x) && REGNO (x) >= FIRST_PSEUDO_REGISTER) |
11870 | || (MEM_P (x) && GET_CODE (XEXP (x, 0)) == PLUS))) | |
8a99f6f9 | 11871 | return GENERAL_REGS; |
0a2aaacc | 11872 | if (REGCLASS_HAS_FP_REG (rclass) |
8a99f6f9 R |
11873 | && TARGET_SHMEDIA |
11874 | && immediate_operand (x, mode) | |
11875 | && x != CONST0_RTX (GET_MODE (x)) | |
11876 | && GET_MODE (x) != V4SFmode) | |
11877 | return GENERAL_REGS; | |
11878 | if ((mode == QImode || mode == HImode) | |
11879 | && TARGET_SHMEDIA && inqhi_operand (x, mode)) | |
11880 | { | |
11881 | sri->icode = ((mode == QImode) | |
11882 | ? CODE_FOR_reload_inqi : CODE_FOR_reload_inhi); | |
11883 | return NO_REGS; | |
11884 | } | |
0a2aaacc | 11885 | if (TARGET_SHMEDIA && rclass == GENERAL_REGS |
dc3ba671 | 11886 | && (GET_CODE (x) == LABEL_REF || PIC_ADDR_P (x))) |
8a99f6f9 R |
11887 | return TARGET_REGS; |
11888 | } /* end of input-only processing. */ | |
11889 | ||
0a2aaacc | 11890 | if (((REGCLASS_HAS_FP_REG (rclass) |
f3536097 | 11891 | && (REG_P (x) |
8a99f6f9 R |
11892 | && (GENERAL_OR_AP_REGISTER_P (REGNO (x)) |
11893 | || (FP_REGISTER_P (REGNO (x)) && mode == SImode | |
11894 | && TARGET_FMOVD)))) | |
0a2aaacc | 11895 | || (REGCLASS_HAS_GENERAL_REG (rclass) |
f3536097 | 11896 | && REG_P (x) |
8a99f6f9 R |
11897 | && FP_REGISTER_P (REGNO (x)))) |
11898 | && ! TARGET_SHMEDIA | |
11899 | && (mode == SFmode || mode == SImode)) | |
11900 | return FPUL_REGS; | |
0a2aaacc KG |
11901 | if ((rclass == FPUL_REGS |
11902 | || (REGCLASS_HAS_FP_REG (rclass) | |
8a99f6f9 | 11903 | && ! TARGET_SHMEDIA && mode == SImode)) |
f3536097 SZ |
11904 | && (MEM_P (x) |
11905 | || (REG_P (x) | |
8a99f6f9 R |
11906 | && (REGNO (x) >= FIRST_PSEUDO_REGISTER |
11907 | || REGNO (x) == T_REG | |
11908 | || system_reg_operand (x, VOIDmode))))) | |
11909 | { | |
0a2aaacc | 11910 | if (rclass == FPUL_REGS) |
8a99f6f9 R |
11911 | return GENERAL_REGS; |
11912 | return FPUL_REGS; | |
11913 | } | |
0a2aaacc KG |
11914 | if ((rclass == TARGET_REGS |
11915 | || (TARGET_SHMEDIA && rclass == SIBCALL_REGS)) | |
32a7ab3d | 11916 | && !satisfies_constraint_Csy (x) |
f3536097 | 11917 | && (!REG_P (x) || ! GENERAL_REGISTER_P (REGNO (x)))) |
8a99f6f9 | 11918 | return GENERAL_REGS; |
0a2aaacc | 11919 | if ((rclass == MAC_REGS || rclass == PR_REGS) |
f3536097 | 11920 | && REG_P (x) && ! GENERAL_REGISTER_P (REGNO (x)) |
0a2aaacc | 11921 | && rclass != REGNO_REG_CLASS (REGNO (x))) |
8a99f6f9 | 11922 | return GENERAL_REGS; |
f3536097 | 11923 | if (rclass != GENERAL_REGS && REG_P (x) |
8a99f6f9 R |
11924 | && TARGET_REGISTER_P (REGNO (x))) |
11925 | return GENERAL_REGS; | |
11926 | return NO_REGS; | |
11927 | } | |
11928 | ||
73a4d10b R |
11929 | enum sh_divide_strategy_e sh_div_strategy = SH_DIV_STRATEGY_DEFAULT; |
11930 | ||
e2500fed | 11931 | #include "gt-sh.h" |