]>
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, |
2acdc1fb | 3 | 2003, 2004, 2005, 2006, 2007, 2008 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" | |
8b97c5f8 | 40 | #include "c-pragma.h" |
950a3816 | 41 | #include "integrate.h" |
a1c496cb | 42 | #include "dwarf2.h" |
318881c0 | 43 | #include "tm_p.h" |
672a6f42 NB |
44 | #include "target.h" |
45 | #include "target-def.h" | |
62526ec3 | 46 | #include "real.h" |
c1b92d09 | 47 | #include "langhooks.h" |
1aa03f38 | 48 | #include "basic-block.h" |
6fb5fa3c | 49 | #include "df.h" |
117dca74 | 50 | #include "cfglayout.h" |
bcc8cc82 | 51 | #include "intl.h" |
58565a33 | 52 | #include "sched-int.h" |
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 | /* Saved operands from the last compare to use when we generate an scc |
16bea517 | 110 | or bcc insn. */ |
bc45ade3 SC |
111 | |
112 | rtx sh_compare_op0; | |
113 | rtx sh_compare_op1; | |
114 | ||
115 | /* Provides the class number of the smallest class containing | |
16bea517 | 116 | reg number. */ |
bc45ade3 | 117 | |
fada1961 | 118 | enum reg_class regno_reg_class[FIRST_PSEUDO_REGISTER] = |
bc45ade3 | 119 | { |
8e87e161 | 120 | R0_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, |
bc45ade3 SC |
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, | |
fa5322fa AO |
124 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, |
125 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
126 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
127 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
128 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
129 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
130 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
131 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
132 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
133 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
134 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
135 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
1a95a963 JW |
136 | FP0_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, | |
fa5322fa AO |
140 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, |
141 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
142 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
143 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
144 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
145 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
146 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
147 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
148 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
149 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
150 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
151 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
152 | TARGET_REGS, TARGET_REGS, TARGET_REGS, TARGET_REGS, | |
153 | TARGET_REGS, TARGET_REGS, TARGET_REGS, TARGET_REGS, | |
225e4f43 R |
154 | DF_REGS, DF_REGS, DF_REGS, DF_REGS, |
155 | DF_REGS, DF_REGS, DF_REGS, DF_REGS, | |
fa5322fa AO |
156 | NO_REGS, GENERAL_REGS, PR_REGS, T_REGS, |
157 | MAC_REGS, MAC_REGS, FPUL_REGS, FPSCR_REGS, | |
96a2347e | 158 | GENERAL_REGS, GENERAL_REGS, |
225e4f43 R |
159 | }; |
160 | ||
fa5322fa AO |
161 | char sh_register_names[FIRST_PSEUDO_REGISTER] \ |
162 | [MAX_REGISTER_NAME_LENGTH + 1] = SH_REGISTER_NAMES_INITIALIZER; | |
163 | ||
164 | char sh_additional_register_names[ADDREGNAMES_SIZE] \ | |
165 | [MAX_ADDITIONAL_REGISTER_NAME_LENGTH + 1] | |
166 | = SH_ADDITIONAL_REGISTER_NAMES_INITIALIZER; | |
bc45ade3 | 167 | |
225e4f43 R |
168 | int assembler_dialect; |
169 | ||
fe3ad572 SC |
170 | static bool shmedia_space_reserved_for_target_registers; |
171 | ||
c0fb94d7 | 172 | static bool sh_handle_option (size_t, const char *, int); |
cf277499 SB |
173 | static void split_branches (rtx); |
174 | static int branch_dest (rtx); | |
175 | static void force_into (rtx, rtx); | |
176 | static void print_slot (rtx); | |
177 | static rtx add_constant (rtx, enum machine_mode, rtx); | |
078c8b08 | 178 | static void dump_table (rtx, rtx); |
cf277499 SB |
179 | static int hi_const (rtx); |
180 | static int broken_move (rtx); | |
181 | static int mova_p (rtx); | |
182 | static rtx find_barrier (int, rtx, rtx); | |
183 | static int noncall_uses_reg (rtx, rtx, rtx *); | |
184 | static rtx gen_block_redirect (rtx, int, int); | |
185 | static void sh_reorg (void); | |
fada1961 | 186 | static void output_stack_adjust (int, rtx, int, HARD_REG_SET *); |
cf277499 SB |
187 | static rtx frame_insn (rtx); |
188 | static rtx push (int); | |
189 | static void pop (int); | |
190 | static void push_regs (HARD_REG_SET *, int); | |
191 | static int calc_live_regs (HARD_REG_SET *); | |
cf277499 SB |
192 | static HOST_WIDE_INT rounded_frame_size (int); |
193 | static rtx mark_constant_pool_use (rtx); | |
91d231cb | 194 | const struct attribute_spec sh_attribute_table[]; |
cf277499 | 195 | static tree sh_handle_interrupt_handler_attribute (tree *, tree, tree, int, bool *); |
561642fa AP |
196 | static tree sh_handle_resbank_handler_attribute (tree *, tree, |
197 | tree, int, bool *); | |
198 | static tree sh2a_handle_function_vector_handler_attribute (tree *, tree, | |
199 | tree, int, bool *); | |
cf277499 SB |
200 | static tree sh_handle_sp_switch_attribute (tree *, tree, tree, int, bool *); |
201 | static tree sh_handle_trap_exit_attribute (tree *, tree, tree, int, bool *); | |
202 | static tree sh_handle_renesas_attribute (tree *, tree, tree, int, bool *); | |
203 | static void sh_output_function_epilogue (FILE *, HOST_WIDE_INT); | |
204 | static void sh_insert_attributes (tree, tree *); | |
8d932be3 | 205 | static const char *sh_check_pch_target_flags (int); |
cf277499 | 206 | static int sh_adjust_cost (rtx, rtx, rtx, int); |
cf277499 | 207 | static int sh_issue_rate (void); |
58565a33 SKG |
208 | static int sh_dfa_new_cycle (FILE *, int, rtx, int, int, int *sort_p); |
209 | static short find_set_regmode_weight (rtx, enum machine_mode); | |
210 | static short find_insn_regmode_weight (rtx, enum machine_mode); | |
c6d56443 | 211 | static void find_regmode_weight (basic_block, enum machine_mode); |
6fb5fa3c | 212 | static int find_r0_life_regions (basic_block); |
58565a33 SKG |
213 | static void sh_md_init_global (FILE *, int, int); |
214 | static void sh_md_finish_global (FILE *, int); | |
215 | static int rank_for_reorder (const void *, const void *); | |
216 | static void swap_reorder (rtx *, int); | |
217 | static void ready_reorder (rtx *, int); | |
218 | static short high_pressure (enum machine_mode); | |
219 | static int sh_reorder (FILE *, int, rtx *, int *, int); | |
220 | static int sh_reorder2 (FILE *, int, rtx *, int *, int); | |
221 | static void sh_md_init (FILE *, int, int); | |
222 | static int sh_variable_issue (FILE *, int, rtx, int); | |
f676971a | 223 | |
cf277499 SB |
224 | static bool sh_function_ok_for_sibcall (tree, tree); |
225 | ||
226 | static bool sh_cannot_modify_jumps_p (void); | |
a3424f5c | 227 | static int sh_target_reg_class (void); |
fe3ad572 | 228 | static bool sh_optimize_target_register_callee_saved (bool); |
3101faab | 229 | static bool sh_ms_bitfield_layout_p (const_tree); |
cf277499 SB |
230 | |
231 | static void sh_init_builtins (void); | |
232 | static void sh_media_init_builtins (void); | |
233 | static rtx sh_expand_builtin (tree, rtx, rtx, enum machine_mode, int); | |
234 | static void sh_output_mi_thunk (FILE *, tree, HOST_WIDE_INT, HOST_WIDE_INT, tree); | |
235 | static void sh_file_start (void); | |
236 | static int flow_dependent_p (rtx, rtx); | |
7bc980e1 | 237 | static void flow_dependent_p_1 (rtx, const_rtx, void *); |
cf277499 SB |
238 | static int shiftcosts (rtx); |
239 | static int andcosts (rtx); | |
240 | static int addsubcosts (rtx); | |
241 | static int multcosts (rtx); | |
242 | static bool unspec_caller_rtx_p (rtx); | |
243 | static bool sh_cannot_copy_insn_p (rtx); | |
244 | static bool sh_rtx_costs (rtx, int, int, int *); | |
245 | static int sh_address_cost (rtx); | |
b48f503c KK |
246 | static int sh_pr_n_sets (void); |
247 | static rtx sh_allocate_initial_value (rtx); | |
fe3ad572 SC |
248 | static int shmedia_target_regs_stack_space (HARD_REG_SET *); |
249 | static int shmedia_reserve_space_for_target_registers_p (int, HARD_REG_SET *); | |
250 | static int shmedia_target_regs_stack_adjust (HARD_REG_SET *); | |
fada1961 R |
251 | static int scavenge_reg (HARD_REG_SET *s); |
252 | struct save_schedule_s; | |
253 | static struct save_entry_s *sh5_schedule_saves (HARD_REG_SET *, | |
254 | struct save_schedule_s *, int); | |
61f71b34 | 255 | |
cf277499 | 256 | static rtx sh_struct_value_rtx (tree, int); |
586de218 | 257 | static bool sh_return_in_memory (const_tree, const_tree); |
cf277499 SB |
258 | static rtx sh_builtin_saveregs (void); |
259 | static void sh_setup_incoming_varargs (CUMULATIVE_ARGS *, enum machine_mode, tree, int *, int); | |
260 | static bool sh_strict_argument_naming (CUMULATIVE_ARGS *); | |
261 | static bool sh_pretend_outgoing_varargs_named (CUMULATIVE_ARGS *); | |
c35d187f | 262 | static tree sh_build_builtin_va_list (void); |
28f16d05 | 263 | static tree sh_canonical_va_list_type (tree); |
d7bd8aeb | 264 | static void sh_va_start (tree, rtx); |
726a989a | 265 | static tree sh_gimplify_va_arg_expr (tree, tree, gimple_seq *, gimple_seq *); |
8cd5a4e0 | 266 | static bool sh_pass_by_reference (CUMULATIVE_ARGS *, enum machine_mode, |
586de218 | 267 | const_tree, bool); |
6cdd5672 | 268 | static bool sh_callee_copies (CUMULATIVE_ARGS *, enum machine_mode, |
586de218 | 269 | const_tree, bool); |
78a52f11 RH |
270 | static int sh_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode, |
271 | tree, bool); | |
b0fbea13 | 272 | static bool sh_scalar_mode_supported_p (enum machine_mode); |
3101faab | 273 | static int sh_dwarf_calling_convention (const_tree); |
561642fa AP |
274 | static void sh_encode_section_info (tree, rtx, int); |
275 | static int sh2a_function_vector_p (tree); | |
61f71b34 | 276 | |
672a6f42 NB |
277 | \f |
278 | /* Initialize the GCC target structure. */ | |
91d231cb JM |
279 | #undef TARGET_ATTRIBUTE_TABLE |
280 | #define TARGET_ATTRIBUTE_TABLE sh_attribute_table | |
672a6f42 | 281 | |
301d03af RS |
282 | /* The next two are used for debug info when compiling with -gdwarf. */ |
283 | #undef TARGET_ASM_UNALIGNED_HI_OP | |
284 | #define TARGET_ASM_UNALIGNED_HI_OP "\t.uaword\t" | |
285 | #undef TARGET_ASM_UNALIGNED_SI_OP | |
286 | #define TARGET_ASM_UNALIGNED_SI_OP "\t.ualong\t" | |
287 | ||
fa5322fa AO |
288 | /* These are NULLed out on non-SH5 in OVERRIDE_OPTIONS. */ |
289 | #undef TARGET_ASM_UNALIGNED_DI_OP | |
290 | #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaquad\t" | |
291 | #undef TARGET_ASM_ALIGNED_DI_OP | |
292 | #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t" | |
293 | ||
08c148a8 NB |
294 | #undef TARGET_ASM_FUNCTION_EPILOGUE |
295 | #define TARGET_ASM_FUNCTION_EPILOGUE sh_output_function_epilogue | |
296 | ||
1aa03f38 R |
297 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
298 | #define TARGET_ASM_OUTPUT_MI_THUNK sh_output_mi_thunk | |
299 | ||
300 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
3101faab | 301 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true |
1aa03f38 | 302 | |
1bc7c5b6 ZW |
303 | #undef TARGET_ASM_FILE_START |
304 | #define TARGET_ASM_FILE_START sh_file_start | |
305 | #undef TARGET_ASM_FILE_START_FILE_DIRECTIVE | |
306 | #define TARGET_ASM_FILE_START_FILE_DIRECTIVE true | |
307 | ||
c0fb94d7 RS |
308 | #undef TARGET_DEFAULT_TARGET_FLAGS |
309 | #define TARGET_DEFAULT_TARGET_FLAGS TARGET_DEFAULT | |
310 | #undef TARGET_HANDLE_OPTION | |
311 | #define TARGET_HANDLE_OPTION sh_handle_option | |
312 | ||
12a68f1f JM |
313 | #undef TARGET_INSERT_ATTRIBUTES |
314 | #define TARGET_INSERT_ATTRIBUTES sh_insert_attributes | |
315 | ||
c237e94a ZW |
316 | #undef TARGET_SCHED_ADJUST_COST |
317 | #define TARGET_SCHED_ADJUST_COST sh_adjust_cost | |
318 | ||
fae15c93 VM |
319 | #undef TARGET_SCHED_ISSUE_RATE |
320 | #define TARGET_SCHED_ISSUE_RATE sh_issue_rate | |
321 | ||
58565a33 SKG |
322 | /* The next 5 hooks have been implemented for reenabling sched1. With the |
323 | help of these macros we are limiting the movement of insns in sched1 to | |
f676971a | 324 | reduce the register pressure. The overall idea is to keep count of SImode |
58565a33 SKG |
325 | and SFmode regs required by already scheduled insns. When these counts |
326 | cross some threshold values; give priority to insns that free registers. | |
327 | The insn that frees registers is most likely to be the insn with lowest | |
f676971a | 328 | LUID (original insn order); but such an insn might be there in the stalled |
58565a33 SKG |
329 | queue (Q) instead of the ready queue (R). To solve this, we skip cycles |
330 | upto a max of 8 cycles so that such insns may move from Q -> R. | |
331 | ||
332 | The description of the hooks are as below: | |
333 | ||
334 | TARGET_SCHED_INIT_GLOBAL: Added a new target hook in the generic | |
335 | scheduler; it is called inside the sched_init function just after | |
336 | find_insn_reg_weights function call. It is used to calculate the SImode | |
1ae58c30 | 337 | and SFmode weights of insns of basic blocks; much similar to what |
f676971a | 338 | find_insn_reg_weights does. |
58565a33 SKG |
339 | TARGET_SCHED_FINISH_GLOBAL: Corresponding cleanup hook. |
340 | ||
341 | TARGET_SCHED_DFA_NEW_CYCLE: Skip cycles if high register pressure is | |
342 | indicated by TARGET_SCHED_REORDER2; doing this may move insns from | |
343 | (Q)->(R). | |
344 | ||
345 | TARGET_SCHED_REORDER: If the register pressure for SImode or SFmode is | |
346 | high; reorder the ready queue so that the insn with lowest LUID will be | |
347 | issued next. | |
348 | ||
349 | TARGET_SCHED_REORDER2: If the register pressure is high, indicate to | |
350 | TARGET_SCHED_DFA_NEW_CYCLE to skip cycles. | |
351 | ||
352 | TARGET_SCHED_VARIABLE_ISSUE: Cache the value of can_issue_more so that it | |
353 | can be returned from TARGET_SCHED_REORDER2. | |
354 | ||
355 | TARGET_SCHED_INIT: Reset the register pressure counting variables. */ | |
356 | ||
357 | #undef TARGET_SCHED_DFA_NEW_CYCLE | |
358 | #define TARGET_SCHED_DFA_NEW_CYCLE sh_dfa_new_cycle | |
359 | ||
360 | #undef TARGET_SCHED_INIT_GLOBAL | |
361 | #define TARGET_SCHED_INIT_GLOBAL sh_md_init_global | |
362 | ||
363 | #undef TARGET_SCHED_FINISH_GLOBAL | |
364 | #define TARGET_SCHED_FINISH_GLOBAL sh_md_finish_global | |
365 | ||
366 | #undef TARGET_SCHED_VARIABLE_ISSUE | |
367 | #define TARGET_SCHED_VARIABLE_ISSUE sh_variable_issue | |
368 | ||
369 | #undef TARGET_SCHED_REORDER | |
370 | #define TARGET_SCHED_REORDER sh_reorder | |
371 | ||
372 | #undef TARGET_SCHED_REORDER2 | |
373 | #define TARGET_SCHED_REORDER2 sh_reorder2 | |
374 | ||
375 | #undef TARGET_SCHED_INIT | |
376 | #define TARGET_SCHED_INIT sh_md_init | |
377 | ||
fa5322fa AO |
378 | #undef TARGET_CANNOT_MODIFY_JUMPS_P |
379 | #define TARGET_CANNOT_MODIFY_JUMPS_P sh_cannot_modify_jumps_p | |
fe3ad572 SC |
380 | #undef TARGET_BRANCH_TARGET_REGISTER_CLASS |
381 | #define TARGET_BRANCH_TARGET_REGISTER_CLASS sh_target_reg_class | |
382 | #undef TARGET_BRANCH_TARGET_REGISTER_CALLEE_SAVED | |
383 | #define TARGET_BRANCH_TARGET_REGISTER_CALLEE_SAVED \ | |
384 | sh_optimize_target_register_callee_saved | |
fa5322fa AO |
385 | |
386 | #undef TARGET_MS_BITFIELD_LAYOUT_P | |
387 | #define TARGET_MS_BITFIELD_LAYOUT_P sh_ms_bitfield_layout_p | |
388 | ||
c1b92d09 R |
389 | #undef TARGET_INIT_BUILTINS |
390 | #define TARGET_INIT_BUILTINS sh_init_builtins | |
391 | #undef TARGET_EXPAND_BUILTIN | |
392 | #define TARGET_EXPAND_BUILTIN sh_expand_builtin | |
393 | ||
4977bab6 ZW |
394 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL |
395 | #define TARGET_FUNCTION_OK_FOR_SIBCALL sh_function_ok_for_sibcall | |
396 | ||
9f3a9a08 KK |
397 | #undef TARGET_CANNOT_COPY_INSN_P |
398 | #define TARGET_CANNOT_COPY_INSN_P sh_cannot_copy_insn_p | |
3c50106f RH |
399 | #undef TARGET_RTX_COSTS |
400 | #define TARGET_RTX_COSTS sh_rtx_costs | |
dcefdf67 RH |
401 | #undef TARGET_ADDRESS_COST |
402 | #define TARGET_ADDRESS_COST sh_address_cost | |
b48f503c KK |
403 | #undef TARGET_ALLOCATE_INITIAL_VALUE |
404 | #define TARGET_ALLOCATE_INITIAL_VALUE sh_allocate_initial_value | |
3c50106f | 405 | |
18dbd950 RS |
406 | #undef TARGET_MACHINE_DEPENDENT_REORG |
407 | #define TARGET_MACHINE_DEPENDENT_REORG sh_reorg | |
408 | ||
463f02cd KK |
409 | #ifdef HAVE_AS_TLS |
410 | #undef TARGET_HAVE_TLS | |
411 | #define TARGET_HAVE_TLS true | |
412 | #endif | |
413 | ||
61f71b34 DD |
414 | #undef TARGET_PROMOTE_PROTOTYPES |
415 | #define TARGET_PROMOTE_PROTOTYPES sh_promote_prototypes | |
416 | #undef TARGET_PROMOTE_FUNCTION_ARGS | |
417 | #define TARGET_PROMOTE_FUNCTION_ARGS sh_promote_prototypes | |
418 | #undef TARGET_PROMOTE_FUNCTION_RETURN | |
419 | #define TARGET_PROMOTE_FUNCTION_RETURN sh_promote_prototypes | |
420 | ||
421 | #undef TARGET_STRUCT_VALUE_RTX | |
422 | #define TARGET_STRUCT_VALUE_RTX sh_struct_value_rtx | |
423 | #undef TARGET_RETURN_IN_MEMORY | |
424 | #define TARGET_RETURN_IN_MEMORY sh_return_in_memory | |
425 | ||
426 | #undef TARGET_EXPAND_BUILTIN_SAVEREGS | |
427 | #define TARGET_EXPAND_BUILTIN_SAVEREGS sh_builtin_saveregs | |
428 | #undef TARGET_SETUP_INCOMING_VARARGS | |
429 | #define TARGET_SETUP_INCOMING_VARARGS sh_setup_incoming_varargs | |
430 | #undef TARGET_STRICT_ARGUMENT_NAMING | |
431 | #define TARGET_STRICT_ARGUMENT_NAMING sh_strict_argument_naming | |
432 | #undef TARGET_PRETEND_OUTGOING_VARARGS_NAMED | |
433 | #define TARGET_PRETEND_OUTGOING_VARARGS_NAMED sh_pretend_outgoing_varargs_named | |
fe984136 RH |
434 | #undef TARGET_MUST_PASS_IN_STACK |
435 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
8cd5a4e0 RH |
436 | #undef TARGET_PASS_BY_REFERENCE |
437 | #define TARGET_PASS_BY_REFERENCE sh_pass_by_reference | |
6cdd5672 RH |
438 | #undef TARGET_CALLEE_COPIES |
439 | #define TARGET_CALLEE_COPIES sh_callee_copies | |
78a52f11 RH |
440 | #undef TARGET_ARG_PARTIAL_BYTES |
441 | #define TARGET_ARG_PARTIAL_BYTES sh_arg_partial_bytes | |
61f71b34 | 442 | |
c35d187f RH |
443 | #undef TARGET_BUILD_BUILTIN_VA_LIST |
444 | #define TARGET_BUILD_BUILTIN_VA_LIST sh_build_builtin_va_list | |
28f16d05 KK |
445 | #undef TARGET_CANONICAL_VA_LIST_TYPE |
446 | #define TARGET_CANONICAL_VA_LIST_TYPE sh_canonical_va_list_type | |
d7bd8aeb JJ |
447 | #undef TARGET_EXPAND_BUILTIN_VA_START |
448 | #define TARGET_EXPAND_BUILTIN_VA_START sh_va_start | |
5fac2ed2 RH |
449 | #undef TARGET_GIMPLIFY_VA_ARG_EXPR |
450 | #define TARGET_GIMPLIFY_VA_ARG_EXPR sh_gimplify_va_arg_expr | |
c35d187f | 451 | |
b0fbea13 KK |
452 | #undef TARGET_SCALAR_MODE_SUPPORTED_P |
453 | #define TARGET_SCALAR_MODE_SUPPORTED_P sh_scalar_mode_supported_p | |
f676971a EC |
454 | #undef TARGET_VECTOR_MODE_SUPPORTED_P |
455 | #define TARGET_VECTOR_MODE_SUPPORTED_P sh_vector_mode_supported_p | |
456 | ||
8d932be3 RS |
457 | #undef TARGET_CHECK_PCH_TARGET_FLAGS |
458 | #define TARGET_CHECK_PCH_TARGET_FLAGS sh_check_pch_target_flags | |
bcc8cc82 | 459 | |
a1c496cb EC |
460 | #undef TARGET_DWARF_CALLING_CONVENTION |
461 | #define TARGET_DWARF_CALLING_CONVENTION sh_dwarf_calling_convention | |
462 | ||
58565a33 SKG |
463 | /* Return regmode weight for insn. */ |
464 | #define INSN_REGMODE_WEIGHT(INSN, MODE) regmode_weight[((MODE) == SImode) ? 0 : 1][INSN_UID (INSN)] | |
465 | ||
466 | /* Return current register pressure for regmode. */ | |
467 | #define CURR_REGMODE_PRESSURE(MODE) curr_regmode_pressure[((MODE) == SImode) ? 0 : 1] | |
468 | ||
561642fa AP |
469 | #undef TARGET_ENCODE_SECTION_INFO |
470 | #define TARGET_ENCODE_SECTION_INFO sh_encode_section_info | |
471 | ||
b55e3aad NC |
472 | #ifdef SYMBIAN |
473 | ||
474 | #undef TARGET_ENCODE_SECTION_INFO | |
475 | #define TARGET_ENCODE_SECTION_INFO sh_symbian_encode_section_info | |
476 | #undef TARGET_STRIP_NAME_ENCODING | |
477 | #define TARGET_STRIP_NAME_ENCODING sh_symbian_strip_name_encoding | |
478 | #undef TARGET_CXX_IMPORT_EXPORT_CLASS | |
479 | #define TARGET_CXX_IMPORT_EXPORT_CLASS symbian_import_export_class | |
480 | ||
481 | #endif /* SYMBIAN */ | |
482 | ||
8a99f6f9 R |
483 | #undef TARGET_SECONDARY_RELOAD |
484 | #define TARGET_SECONDARY_RELOAD sh_secondary_reload | |
485 | ||
561642fa AP |
486 | /* Machine-specific symbol_ref flags. */ |
487 | #define SYMBOL_FLAG_FUNCVEC_FUNCTION (SYMBOL_FLAG_MACH_DEP << 0) | |
488 | ||
f6897b10 | 489 | struct gcc_target targetm = TARGET_INITIALIZER; |
b9654711 | 490 | \f |
c0fb94d7 RS |
491 | /* Implement TARGET_HANDLE_OPTION. */ |
492 | ||
493 | static bool | |
494 | sh_handle_option (size_t code, const char *arg ATTRIBUTE_UNUSED, | |
495 | int value ATTRIBUTE_UNUSED) | |
496 | { | |
497 | switch (code) | |
498 | { | |
499 | case OPT_m1: | |
500 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH1; | |
501 | return true; | |
502 | ||
503 | case OPT_m2: | |
504 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2; | |
505 | return true; | |
506 | ||
507 | case OPT_m2a: | |
508 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A; | |
509 | return true; | |
510 | ||
511 | case OPT_m2a_nofpu: | |
512 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_NOFPU; | |
513 | return true; | |
514 | ||
515 | case OPT_m2a_single: | |
516 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_SINGLE; | |
517 | return true; | |
518 | ||
519 | case OPT_m2a_single_only: | |
520 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_SINGLE_ONLY; | |
521 | return true; | |
522 | ||
523 | case OPT_m2e: | |
524 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2E; | |
525 | return true; | |
526 | ||
527 | case OPT_m3: | |
528 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH3; | |
529 | return true; | |
530 | ||
531 | case OPT_m3e: | |
532 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH3E; | |
533 | return true; | |
534 | ||
535 | case OPT_m4: | |
25e651ca AS |
536 | case OPT_m4_100: |
537 | case OPT_m4_200: | |
78d310c2 | 538 | case OPT_m4_300: |
c0fb94d7 RS |
539 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4; |
540 | return true; | |
541 | ||
542 | case OPT_m4_nofpu: | |
78d310c2 R |
543 | case OPT_m4_100_nofpu: |
544 | case OPT_m4_200_nofpu: | |
545 | case OPT_m4_300_nofpu: | |
546 | case OPT_m4_340: | |
25e651ca AS |
547 | case OPT_m4_400: |
548 | case OPT_m4_500: | |
c0fb94d7 RS |
549 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_NOFPU; |
550 | return true; | |
551 | ||
552 | case OPT_m4_single: | |
25e651ca AS |
553 | case OPT_m4_100_single: |
554 | case OPT_m4_200_single: | |
78d310c2 | 555 | case OPT_m4_300_single: |
c0fb94d7 RS |
556 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_SINGLE; |
557 | return true; | |
558 | ||
559 | case OPT_m4_single_only: | |
25e651ca AS |
560 | case OPT_m4_100_single_only: |
561 | case OPT_m4_200_single_only: | |
78d310c2 | 562 | case OPT_m4_300_single_only: |
c0fb94d7 RS |
563 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_SINGLE_ONLY; |
564 | return true; | |
565 | ||
566 | case OPT_m4a: | |
567 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A; | |
568 | return true; | |
569 | ||
570 | case OPT_m4a_nofpu: | |
571 | case OPT_m4al: | |
572 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_NOFPU; | |
573 | return true; | |
574 | ||
575 | case OPT_m4a_single: | |
576 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_SINGLE; | |
577 | return true; | |
578 | ||
579 | case OPT_m4a_single_only: | |
580 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_SINGLE_ONLY; | |
581 | return true; | |
582 | ||
583 | case OPT_m5_32media: | |
584 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_32MEDIA; | |
585 | return true; | |
586 | ||
587 | case OPT_m5_32media_nofpu: | |
588 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_32MEDIA_NOFPU; | |
589 | return true; | |
590 | ||
591 | case OPT_m5_64media: | |
592 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_64MEDIA; | |
593 | return true; | |
594 | ||
595 | case OPT_m5_64media_nofpu: | |
596 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_64MEDIA_NOFPU; | |
597 | return true; | |
598 | ||
599 | case OPT_m5_compact: | |
600 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_COMPACT; | |
601 | return true; | |
602 | ||
603 | case OPT_m5_compact_nofpu: | |
604 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_COMPACT_NOFPU; | |
605 | return true; | |
606 | ||
607 | default: | |
608 | return true; | |
609 | } | |
610 | } | |
611 | \f | |
16bea517 | 612 | /* Print the operand address in x to the stream. */ |
bc45ade3 SC |
613 | |
614 | void | |
cf277499 | 615 | print_operand_address (FILE *stream, rtx x) |
bc45ade3 SC |
616 | { |
617 | switch (GET_CODE (x)) | |
618 | { | |
619 | case REG: | |
225e4f43 R |
620 | case SUBREG: |
621 | fprintf (stream, "@%s", reg_names[true_regnum (x)]); | |
bc45ade3 | 622 | break; |
8aa2a305 | 623 | |
bc45ade3 SC |
624 | case PLUS: |
625 | { | |
626 | rtx base = XEXP (x, 0); | |
627 | rtx index = XEXP (x, 1); | |
628 | ||
bc45ade3 SC |
629 | switch (GET_CODE (index)) |
630 | { | |
631 | case CONST_INT: | |
c5b9ef02 | 632 | fprintf (stream, "@(%d,%s)", (int) INTVAL (index), |
225e4f43 | 633 | reg_names[true_regnum (base)]); |
bc45ade3 SC |
634 | break; |
635 | ||
636 | case REG: | |
225e4f43 R |
637 | case SUBREG: |
638 | { | |
639 | int base_num = true_regnum (base); | |
640 | int index_num = true_regnum (index); | |
641 | ||
642 | fprintf (stream, "@(r0,%s)", | |
643 | reg_names[MAX (base_num, index_num)]); | |
644 | break; | |
645 | } | |
bc45ade3 SC |
646 | |
647 | default: | |
f5b9e7c9 | 648 | gcc_unreachable (); |
bc45ade3 SC |
649 | } |
650 | } | |
bc45ade3 | 651 | break; |
8aa2a305 | 652 | |
bc45ade3 | 653 | case PRE_DEC: |
225e4f43 | 654 | fprintf (stream, "@-%s", reg_names[true_regnum (XEXP (x, 0))]); |
bc45ade3 SC |
655 | break; |
656 | ||
657 | case POST_INC: | |
225e4f43 | 658 | fprintf (stream, "@%s+", reg_names[true_regnum (XEXP (x, 0))]); |
bc45ade3 SC |
659 | break; |
660 | ||
661 | default: | |
b91455de | 662 | x = mark_constant_pool_use (x); |
422be3c3 | 663 | output_addr_const (stream, x); |
bc45ade3 SC |
664 | break; |
665 | } | |
666 | } | |
667 | ||
668 | /* Print operand x (an rtx) in assembler syntax to file stream | |
669 | according to modifier code. | |
670 | ||
b9654711 | 671 | '.' print a .s if insn needs delay slot |
1245df60 | 672 | ',' print LOCAL_LABEL_PREFIX |
4408efce | 673 | '@' print trap, rte or rts depending upon pragma interruptness |
8aa2a305 | 674 | '#' output a nop if there is nothing to put in the delay slot |
fae778eb | 675 | ''' print likelihood suffix (/u for unlikely). |
73a4d10b | 676 | '>' print branch target if -fverbose-asm |
8aa2a305 | 677 | 'O' print a constant without the # |
00f8ff66 | 678 | 'R' print the LSW of a dp value - changes if in little endian |
00f8ff66 | 679 | 'S' print the MSW of a dp value - changes if in little endian |
225e4f43 | 680 | 'T' print the next word of a dp value - same as 'R' in big endian mode. |
31b6f0ae R |
681 | 'M' SHMEDIA: print an `x' if `m' will print `base,index'. |
682 | otherwise: print .b / .w / .l / .s / .d suffix if operand is a MEM. | |
c1b92d09 | 683 | 'N' print 'r63' if the operand is (const_int 0). |
312209c6 | 684 | 'd' print a V2SF reg as dN instead of fpN. |
fa5322fa | 685 | 'm' print a pair `base,offset' or `base,index', for LD and ST. |
73a4d10b | 686 | 'U' Likewise for {LD,ST}{HI,LO}. |
de6adfa2 N |
687 | 'V' print the position of a single bit set. |
688 | 'W' print the position of a single bit cleared. | |
9eb3a0dd | 689 | 't' print a memory address which is a register. |
fa5322fa | 690 | 'u' prints the lowest 16 bits of CONST_INT, as an unsigned value. |
225e4f43 | 691 | 'o' output an operator. */ |
bc45ade3 SC |
692 | |
693 | void | |
cf277499 | 694 | print_operand (FILE *stream, rtx x, int code) |
bc45ade3 | 695 | { |
73a4d10b R |
696 | int regno; |
697 | enum machine_mode mode; | |
698 | ||
bc45ade3 SC |
699 | switch (code) |
700 | { | |
a6ab9fc0 R |
701 | tree trapa_attr; |
702 | ||
b9654711 | 703 | case '.': |
79b2746a | 704 | if (final_sequence |
c608a684 R |
705 | && ! INSN_ANNULLED_BRANCH_P (XVECEXP (final_sequence, 0, 0)) |
706 | && get_attr_length (XVECEXP (final_sequence, 0, 1))) | |
1245df60 R |
707 | fprintf (stream, ASSEMBLER_DIALECT ? "/s" : ".s"); |
708 | break; | |
709 | case ',': | |
710 | fprintf (stream, "%s", LOCAL_LABEL_PREFIX); | |
b9654711 | 711 | break; |
d3ae8277 | 712 | case '@': |
a6ab9fc0 R |
713 | trapa_attr = lookup_attribute ("trap_exit", |
714 | DECL_ATTRIBUTES (current_function_decl)); | |
715 | if (trapa_attr) | |
716 | fprintf (stream, "trapa #%ld", | |
717 | (long) TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (trapa_attr)))); | |
e3ba8d11 | 718 | else if (sh_cfun_interrupt_handler_p ()) |
561642fa AP |
719 | { |
720 | if (sh_cfun_resbank_handler_p ()) | |
721 | fprintf (stream, "resbank\n"); | |
722 | fprintf (stream, "rte"); | |
723 | } | |
d3ae8277 | 724 | else |
8e87e161 | 725 | fprintf (stream, "rts"); |
d3ae8277 | 726 | break; |
bc45ade3 | 727 | case '#': |
16bea517 | 728 | /* Output a nop if there's nothing in the delay slot. */ |
bc45ade3 | 729 | if (dbr_sequence_length () == 0) |
8aa2a305 | 730 | fprintf (stream, "\n\tnop"); |
bc45ade3 | 731 | break; |
c8cc4417 R |
732 | case '\'': |
733 | { | |
734 | rtx note = find_reg_note (current_output_insn, REG_BR_PROB, 0); | |
735 | ||
736 | if (note && INTVAL (XEXP (note, 0)) * 2 < REG_BR_PROB_BASE) | |
737 | fputs ("/u", stream); | |
738 | break; | |
739 | } | |
73a4d10b R |
740 | case '>': |
741 | if (flag_verbose_asm && JUMP_LABEL (current_output_insn)) | |
742 | { | |
743 | fputs ("\t! target: ", stream); | |
744 | output_addr_const (stream, JUMP_LABEL (current_output_insn)); | |
745 | } | |
746 | break; | |
b9654711 | 747 | case 'O': |
b91455de | 748 | x = mark_constant_pool_use (x); |
0d7e008e | 749 | output_addr_const (stream, x); |
bc45ade3 | 750 | break; |
4802062c R |
751 | /* N.B.: %R / %S / %T adjust memory addresses by four. |
752 | For SHMEDIA, that means they can be used to access the first and | |
753 | second 32 bit part of a 64 bit (or larger) value that | |
754 | might be held in floating point registers or memory. | |
755 | While they can be used to access 64 bit parts of a larger value | |
756 | held in general purpose registers, that won't work with memory - | |
757 | neither for fp registers, since the frxx names are used. */ | |
bc45ade3 | 758 | case 'R': |
4802062c R |
759 | if (REG_P (x) || GET_CODE (x) == SUBREG) |
760 | { | |
761 | regno = true_regnum (x); | |
762 | regno += FP_REGISTER_P (regno) ? 1 : LSW; | |
763 | fputs (reg_names[regno], (stream)); | |
764 | } | |
765 | else if (MEM_P (x)) | |
766 | { | |
767 | x = adjust_address (x, SImode, 4 * LSW); | |
768 | print_operand_address (stream, XEXP (x, 0)); | |
769 | } | |
770 | else | |
771 | { | |
772 | rtx sub = NULL_RTX; | |
773 | ||
774 | mode = GET_MODE (x); | |
775 | if (mode == VOIDmode) | |
776 | mode = DImode; | |
777 | if (GET_MODE_SIZE (mode) >= 8) | |
778 | sub = simplify_subreg (SImode, x, mode, 4 * LSW); | |
779 | if (sub) | |
780 | print_operand (stream, sub, 0); | |
781 | else | |
782 | output_operand_lossage ("invalid operand to %%R"); | |
783 | } | |
8aa2a305 JW |
784 | break; |
785 | case 'S': | |
4802062c R |
786 | if (REG_P (x) || GET_CODE (x) == SUBREG) |
787 | { | |
788 | regno = true_regnum (x); | |
789 | regno += FP_REGISTER_P (regno) ? 0 : MSW; | |
790 | fputs (reg_names[regno], (stream)); | |
791 | } | |
792 | else if (MEM_P (x)) | |
793 | { | |
794 | x = adjust_address (x, SImode, 4 * MSW); | |
795 | print_operand_address (stream, XEXP (x, 0)); | |
796 | } | |
797 | else | |
798 | { | |
799 | rtx sub = NULL_RTX; | |
800 | ||
801 | mode = GET_MODE (x); | |
802 | if (mode == VOIDmode) | |
803 | mode = DImode; | |
804 | if (GET_MODE_SIZE (mode) >= 8) | |
805 | sub = simplify_subreg (SImode, x, mode, 4 * MSW); | |
806 | if (sub) | |
807 | print_operand (stream, sub, 0); | |
808 | else | |
809 | output_operand_lossage ("invalid operand to %%S"); | |
810 | } | |
00f8ff66 SC |
811 | break; |
812 | case 'T': | |
16bea517 | 813 | /* Next word of a double. */ |
bc45ade3 SC |
814 | switch (GET_CODE (x)) |
815 | { | |
816 | case REG: | |
817 | fputs (reg_names[REGNO (x) + 1], (stream)); | |
818 | break; | |
819 | case MEM: | |
225e4f43 R |
820 | if (GET_CODE (XEXP (x, 0)) != PRE_DEC |
821 | && GET_CODE (XEXP (x, 0)) != POST_INC) | |
b72f00af | 822 | x = adjust_address (x, SImode, 4); |
225e4f43 | 823 | print_operand_address (stream, XEXP (x, 0)); |
bc45ade3 | 824 | break; |
318881c0 KG |
825 | default: |
826 | break; | |
bc45ade3 SC |
827 | } |
828 | break; | |
9eb3a0dd N |
829 | |
830 | case 't': | |
831 | gcc_assert (GET_CODE (x) == MEM); | |
832 | x = XEXP (x, 0); | |
833 | switch (GET_CODE (x)) | |
834 | { | |
835 | case REG: | |
836 | case SUBREG: | |
837 | print_operand (stream, x, 0); | |
838 | break; | |
839 | default: | |
840 | break; | |
841 | } | |
842 | break; | |
843 | ||
225e4f43 R |
844 | case 'o': |
845 | switch (GET_CODE (x)) | |
846 | { | |
847 | case PLUS: fputs ("add", stream); break; | |
848 | case MINUS: fputs ("sub", stream); break; | |
849 | case MULT: fputs ("mul", stream); break; | |
850 | case DIV: fputs ("div", stream); break; | |
c8cc4417 R |
851 | case EQ: fputs ("eq", stream); break; |
852 | case NE: fputs ("ne", stream); break; | |
853 | case GT: case LT: fputs ("gt", stream); break; | |
854 | case GE: case LE: fputs ("ge", stream); break; | |
855 | case GTU: case LTU: fputs ("gtu", stream); break; | |
856 | case GEU: case LEU: fputs ("geu", stream); break; | |
318881c0 KG |
857 | default: |
858 | break; | |
225e4f43 R |
859 | } |
860 | break; | |
fa5322fa | 861 | case 'M': |
31b6f0ae R |
862 | if (TARGET_SHMEDIA) |
863 | { | |
864 | if (GET_CODE (x) == MEM | |
865 | && GET_CODE (XEXP (x, 0)) == PLUS | |
866 | && (GET_CODE (XEXP (XEXP (x, 0), 1)) == REG | |
867 | || GET_CODE (XEXP (XEXP (x, 0), 1)) == SUBREG)) | |
868 | fputc ('x', stream); | |
869 | } | |
870 | else | |
871 | { | |
872 | if (GET_CODE (x) == MEM) | |
873 | { | |
874 | switch (GET_MODE (x)) | |
875 | { | |
876 | case QImode: fputs (".b", stream); break; | |
877 | case HImode: fputs (".w", stream); break; | |
878 | case SImode: fputs (".l", stream); break; | |
879 | case SFmode: fputs (".s", stream); break; | |
880 | case DFmode: fputs (".d", stream); break; | |
881 | default: gcc_unreachable (); | |
882 | } | |
883 | } | |
884 | } | |
fa5322fa AO |
885 | break; |
886 | ||
887 | case 'm': | |
f5b9e7c9 | 888 | gcc_assert (GET_CODE (x) == MEM); |
fa5322fa | 889 | x = XEXP (x, 0); |
73a4d10b R |
890 | /* Fall through. */ |
891 | case 'U': | |
fa5322fa AO |
892 | switch (GET_CODE (x)) |
893 | { | |
894 | case REG: | |
895 | case SUBREG: | |
896 | print_operand (stream, x, 0); | |
897 | fputs (", 0", stream); | |
898 | break; | |
899 | ||
900 | case PLUS: | |
901 | print_operand (stream, XEXP (x, 0), 0); | |
902 | fputs (", ", stream); | |
903 | print_operand (stream, XEXP (x, 1), 0); | |
904 | break; | |
905 | ||
906 | default: | |
f5b9e7c9 | 907 | gcc_unreachable (); |
fa5322fa AO |
908 | } |
909 | break; | |
910 | ||
de6adfa2 N |
911 | case 'V': |
912 | { | |
913 | int num = exact_log2 (INTVAL (x)); | |
914 | gcc_assert (num >= 0); | |
915 | fprintf (stream, "#%d", num); | |
916 | } | |
917 | break; | |
918 | ||
919 | case 'W': | |
920 | { | |
921 | int num = exact_log2 (~INTVAL (x)); | |
922 | gcc_assert (num >= 0); | |
923 | fprintf (stream, "#%d", num); | |
924 | } | |
925 | break; | |
926 | ||
312209c6 | 927 | case 'd': |
f5b9e7c9 | 928 | gcc_assert (GET_CODE (x) == REG && GET_MODE (x) == V2SFmode); |
312209c6 AO |
929 | |
930 | fprintf ((stream), "d%s", reg_names[REGNO (x)] + 1); | |
931 | break; | |
f676971a | 932 | |
c1b92d09 | 933 | case 'N': |
0ac78517 | 934 | if (x == CONST0_RTX (GET_MODE (x))) |
c1b92d09 R |
935 | { |
936 | fprintf ((stream), "r63"); | |
937 | break; | |
938 | } | |
939 | goto default_output; | |
fa5322fa AO |
940 | case 'u': |
941 | if (GET_CODE (x) == CONST_INT) | |
c8cc4417 | 942 | { |
fa5322fa AO |
943 | fprintf ((stream), "%u", (unsigned) INTVAL (x) & (0x10000 - 1)); |
944 | break; | |
945 | } | |
946 | /* Fall through. */ | |
947 | ||
c1b92d09 | 948 | default_output: |
bc45ade3 | 949 | default: |
73a4d10b R |
950 | regno = 0; |
951 | mode = GET_MODE (x); | |
952 | ||
bc45ade3 SC |
953 | switch (GET_CODE (x)) |
954 | { | |
73a4d10b R |
955 | case TRUNCATE: |
956 | { | |
957 | rtx inner = XEXP (x, 0); | |
958 | int offset = 0; | |
959 | enum machine_mode inner_mode; | |
960 | ||
961 | /* We might see SUBREGs with vector mode registers inside. */ | |
962 | if (GET_CODE (inner) == SUBREG | |
963 | && (GET_MODE_SIZE (GET_MODE (inner)) | |
964 | == GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner)))) | |
965 | && subreg_lowpart_p (inner)) | |
966 | inner = SUBREG_REG (inner); | |
967 | if (GET_CODE (inner) == CONST_INT) | |
968 | { | |
969 | x = GEN_INT (trunc_int_for_mode (INTVAL (inner), GET_MODE (x))); | |
970 | goto default_output; | |
971 | } | |
972 | inner_mode = GET_MODE (inner); | |
973 | if (GET_CODE (inner) == SUBREG | |
974 | && (GET_MODE_SIZE (GET_MODE (inner)) | |
975 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner)))) | |
976 | && GET_CODE (SUBREG_REG (inner)) == REG) | |
977 | { | |
978 | offset = subreg_regno_offset (REGNO (SUBREG_REG (inner)), | |
979 | GET_MODE (SUBREG_REG (inner)), | |
980 | SUBREG_BYTE (inner), | |
981 | GET_MODE (inner)); | |
982 | inner = SUBREG_REG (inner); | |
983 | } | |
984 | if (GET_CODE (inner) != REG || GET_MODE_SIZE (inner_mode) > 8) | |
985 | abort (); | |
986 | /* Floating point register pairs are always big endian; | |
30dc60c7 | 987 | general purpose registers are 64 bit wide. */ |
73a4d10b R |
988 | regno = REGNO (inner); |
989 | regno = (HARD_REGNO_NREGS (regno, inner_mode) | |
990 | - HARD_REGNO_NREGS (regno, mode)) | |
991 | + offset; | |
992 | x = inner; | |
993 | goto reg; | |
994 | } | |
995 | case SIGN_EXTEND: | |
996 | x = XEXP (x, 0); | |
997 | goto reg; | |
fa5322fa AO |
998 | /* FIXME: We need this on SHmedia32 because reload generates |
999 | some sign-extended HI or QI loads into DImode registers | |
1000 | but, because Pmode is SImode, the address ends up with a | |
1001 | subreg:SI of the DImode register. Maybe reload should be | |
1002 | fixed so as to apply alter_subreg to such loads? */ | |
73a4d10b R |
1003 | case IF_THEN_ELSE: |
1004 | gcc_assert (trapping_target_operand (x, VOIDmode)); | |
1005 | x = XEXP (XEXP (x, 2), 0); | |
1006 | goto default_output; | |
fa5322fa | 1007 | case SUBREG: |
f5b9e7c9 NS |
1008 | gcc_assert (SUBREG_BYTE (x) == 0 |
1009 | && GET_CODE (SUBREG_REG (x)) == REG); | |
fa5322fa AO |
1010 | |
1011 | x = SUBREG_REG (x); | |
1012 | /* Fall through. */ | |
1013 | ||
73a4d10b | 1014 | reg: |
bc45ade3 | 1015 | case REG: |
73a4d10b R |
1016 | regno += REGNO (x); |
1017 | if (FP_REGISTER_P (regno) | |
1018 | && mode == V16SFmode) | |
1019 | fprintf ((stream), "mtrx%s", reg_names[regno] + 2); | |
fa5322fa | 1020 | else if (FP_REGISTER_P (REGNO (x)) |
73a4d10b R |
1021 | && mode == V4SFmode) |
1022 | fprintf ((stream), "fv%s", reg_names[regno] + 2); | |
fa5322fa | 1023 | else if (GET_CODE (x) == REG |
73a4d10b R |
1024 | && mode == V2SFmode) |
1025 | fprintf ((stream), "fp%s", reg_names[regno] + 2); | |
fa5322fa | 1026 | else if (FP_REGISTER_P (REGNO (x)) |
73a4d10b R |
1027 | && GET_MODE_SIZE (mode) > 4) |
1028 | fprintf ((stream), "d%s", reg_names[regno] + 1); | |
225e4f43 | 1029 | else |
73a4d10b | 1030 | fputs (reg_names[regno], (stream)); |
bc45ade3 | 1031 | break; |
fa5322fa | 1032 | |
bc45ade3 SC |
1033 | case MEM: |
1034 | output_address (XEXP (x, 0)); | |
1035 | break; | |
f676971a | 1036 | |
fa5322fa AO |
1037 | case CONST: |
1038 | if (TARGET_SHMEDIA | |
42201282 R |
1039 | && (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND |
1040 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
73a4d10b R |
1041 | && (GET_MODE (XEXP (x, 0)) == DImode |
1042 | || GET_MODE (XEXP (x, 0)) == SImode) | |
fa5322fa AO |
1043 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == TRUNCATE |
1044 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == HImode) | |
1045 | { | |
1046 | rtx val = XEXP (XEXP (XEXP (x, 0), 0), 0); | |
42201282 R |
1047 | rtx val2 = val; |
1048 | bool nested_expr = false; | |
fa5322fa AO |
1049 | |
1050 | fputc ('(', stream); | |
1051 | if (GET_CODE (val) == ASHIFTRT) | |
1052 | { | |
1053 | fputc ('(', stream); | |
42201282 R |
1054 | val2 = XEXP (val, 0); |
1055 | } | |
1056 | if (GET_CODE (val2) == CONST | |
1057 | || GET_RTX_CLASS (GET_CODE (val2)) != RTX_OBJ) | |
1058 | { | |
1059 | fputc ('(', stream); | |
1060 | nested_expr = true; | |
1061 | } | |
1062 | output_addr_const (stream, val2); | |
1063 | if (nested_expr) | |
1064 | fputc (')', stream); | |
1065 | if (GET_CODE (val) == ASHIFTRT) | |
1066 | { | |
fa5322fa AO |
1067 | fputs (" >> ", stream); |
1068 | output_addr_const (stream, XEXP (val, 1)); | |
1069 | fputc (')', stream); | |
1070 | } | |
fa5322fa AO |
1071 | fputs (" & 65535)", stream); |
1072 | break; | |
1073 | } | |
1074 | ||
1075 | /* Fall through. */ | |
bc45ade3 | 1076 | default: |
fa5322fa AO |
1077 | if (TARGET_SH1) |
1078 | fputc ('#', stream); | |
bc45ade3 SC |
1079 | output_addr_const (stream, x); |
1080 | break; | |
bc45ade3 SC |
1081 | } |
1082 | break; | |
1083 | } | |
1084 | } | |
bc45ade3 | 1085 | \f |
561642fa AP |
1086 | |
1087 | /* Encode symbol attributes of a SYMBOL_REF into its | |
1088 | SYMBOL_REF_FLAGS. */ | |
1089 | static void | |
1090 | sh_encode_section_info (tree decl, rtx rtl, int first) | |
1091 | { | |
1092 | default_encode_section_info (decl, rtl, first); | |
1093 | ||
1094 | if (TREE_CODE (decl) == FUNCTION_DECL | |
1095 | && sh2a_function_vector_p (decl) && TARGET_SH2A) | |
1096 | SYMBOL_REF_FLAGS (XEXP (rtl, 0)) |= SYMBOL_FLAG_FUNCVEC_FUNCTION; | |
1097 | } | |
1098 | ||
0aa692fc R |
1099 | /* Like force_operand, but guarantees that VALUE ends up in TARGET. */ |
1100 | static void | |
cf277499 | 1101 | force_into (rtx value, rtx target) |
0aa692fc R |
1102 | { |
1103 | value = force_operand (value, target); | |
1104 | if (! rtx_equal_p (value, target)) | |
1105 | emit_insn (gen_move_insn (target, value)); | |
1106 | } | |
1107 | ||
0d7e008e SC |
1108 | /* Emit code to perform a block move. Choose the best method. |
1109 | ||
1110 | OPERANDS[0] is the destination. | |
1111 | OPERANDS[1] is the source. | |
1112 | OPERANDS[2] is the size. | |
1113 | OPERANDS[3] is the alignment safe to use. */ | |
1114 | ||
0d7e008e | 1115 | int |
cf277499 | 1116 | expand_block_move (rtx *operands) |
0d7e008e SC |
1117 | { |
1118 | int align = INTVAL (operands[3]); | |
1119 | int constp = (GET_CODE (operands[2]) == CONST_INT); | |
1120 | int bytes = (constp ? INTVAL (operands[2]) : 0); | |
0d7e008e | 1121 | |
312209c6 AO |
1122 | if (! constp) |
1123 | return 0; | |
1124 | ||
1125 | /* If we could use mov.l to move words and dest is word-aligned, we | |
1126 | can use movua.l for loads and still generate a relatively short | |
1127 | and efficient sequence. */ | |
1128 | if (TARGET_SH4A_ARCH && align < 4 | |
1129 | && MEM_ALIGN (operands[0]) >= 32 | |
1130 | && can_move_by_pieces (bytes, 32)) | |
1131 | { | |
1132 | rtx dest = copy_rtx (operands[0]); | |
1133 | rtx src = copy_rtx (operands[1]); | |
1134 | /* We could use different pseudos for each copied word, but | |
1135 | since movua can only load into r0, it's kind of | |
1136 | pointless. */ | |
1137 | rtx temp = gen_reg_rtx (SImode); | |
1138 | rtx src_addr = copy_addr_to_reg (XEXP (src, 0)); | |
1139 | int copied = 0; | |
1140 | ||
1141 | while (copied + 4 <= bytes) | |
1142 | { | |
1143 | rtx to = adjust_address (dest, SImode, copied); | |
85144c5c RS |
1144 | rtx from = adjust_automodify_address (src, BLKmode, |
1145 | src_addr, copied); | |
312209c6 | 1146 | |
85144c5c | 1147 | set_mem_size (from, GEN_INT (4)); |
312209c6 AO |
1148 | emit_insn (gen_movua (temp, from)); |
1149 | emit_move_insn (src_addr, plus_constant (src_addr, 4)); | |
1150 | emit_move_insn (to, temp); | |
1151 | copied += 4; | |
1152 | } | |
1153 | ||
1154 | if (copied < bytes) | |
1155 | move_by_pieces (adjust_address (dest, BLKmode, copied), | |
1156 | adjust_automodify_address (src, BLKmode, | |
1157 | src_addr, copied), | |
1158 | bytes - copied, align, 0); | |
1159 | ||
1160 | return 1; | |
1161 | } | |
1162 | ||
8aa2a305 JW |
1163 | /* If it isn't a constant number of bytes, or if it doesn't have 4 byte |
1164 | alignment, or if it isn't a multiple of 4 bytes, then fail. */ | |
312209c6 | 1165 | if (align < 4 || (bytes % 4 != 0)) |
d3ae8277 SC |
1166 | return 0; |
1167 | ||
225e4f43 R |
1168 | if (TARGET_HARD_SH4) |
1169 | { | |
1170 | if (bytes < 12) | |
1171 | return 0; | |
1172 | else if (bytes == 12) | |
1173 | { | |
73a4d10b | 1174 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
c0d4e710 KH |
1175 | rtx r4 = gen_rtx_REG (SImode, 4); |
1176 | rtx r5 = gen_rtx_REG (SImode, 5); | |
225e4f43 | 1177 | |
73a4d10b | 1178 | function_symbol (func_addr_rtx, "__movmemSI12_i4", SFUNC_STATIC); |
0aa692fc R |
1179 | force_into (XEXP (operands[0], 0), r4); |
1180 | force_into (XEXP (operands[1], 0), r5); | |
225e4f43 R |
1181 | emit_insn (gen_block_move_real_i4 (func_addr_rtx)); |
1182 | return 1; | |
1183 | } | |
1184 | else if (! TARGET_SMALLCODE) | |
1185 | { | |
73a4d10b R |
1186 | const char *entry_name; |
1187 | rtx func_addr_rtx = gen_reg_rtx (Pmode); | |
225e4f43 | 1188 | int dwords; |
c0d4e710 KH |
1189 | rtx r4 = gen_rtx_REG (SImode, 4); |
1190 | rtx r5 = gen_rtx_REG (SImode, 5); | |
1191 | rtx r6 = gen_rtx_REG (SImode, 6); | |
225e4f43 | 1192 | |
73a4d10b R |
1193 | entry_name = (bytes & 4 ? "__movmem_i4_odd" : "__movmem_i4_even"); |
1194 | function_symbol (func_addr_rtx, entry_name, SFUNC_STATIC); | |
0aa692fc R |
1195 | force_into (XEXP (operands[0], 0), r4); |
1196 | force_into (XEXP (operands[1], 0), r5); | |
225e4f43 R |
1197 | |
1198 | dwords = bytes >> 3; | |
1199 | emit_insn (gen_move_insn (r6, GEN_INT (dwords - 1))); | |
1200 | emit_insn (gen_block_lump_real_i4 (func_addr_rtx)); | |
1201 | return 1; | |
1202 | } | |
1203 | else | |
1204 | return 0; | |
1205 | } | |
8aa2a305 | 1206 | if (bytes < 64) |
0d7e008e SC |
1207 | { |
1208 | char entry[30]; | |
73a4d10b | 1209 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
c5c76735 JL |
1210 | rtx r4 = gen_rtx_REG (SImode, 4); |
1211 | rtx r5 = gen_rtx_REG (SImode, 5); | |
8aa2a305 | 1212 | |
70128ad9 | 1213 | sprintf (entry, "__movmemSI%d", bytes); |
73a4d10b | 1214 | function_symbol (func_addr_rtx, entry, SFUNC_STATIC); |
0aa692fc R |
1215 | force_into (XEXP (operands[0], 0), r4); |
1216 | force_into (XEXP (operands[1], 0), r5); | |
0d7e008e SC |
1217 | emit_insn (gen_block_move_real (func_addr_rtx)); |
1218 | return 1; | |
1219 | } | |
8aa2a305 JW |
1220 | |
1221 | /* This is the same number of bytes as a memcpy call, but to a different | |
1222 | less common function name, so this will occasionally use more space. */ | |
1223 | if (! TARGET_SMALLCODE) | |
0d7e008e | 1224 | { |
73a4d10b | 1225 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
8aa2a305 | 1226 | int final_switch, while_loop; |
c5c76735 JL |
1227 | rtx r4 = gen_rtx_REG (SImode, 4); |
1228 | rtx r5 = gen_rtx_REG (SImode, 5); | |
1229 | rtx r6 = gen_rtx_REG (SImode, 6); | |
0d7e008e | 1230 | |
73a4d10b | 1231 | function_symbol (func_addr_rtx, "__movmem", SFUNC_STATIC); |
0aa692fc R |
1232 | force_into (XEXP (operands[0], 0), r4); |
1233 | force_into (XEXP (operands[1], 0), r5); | |
0d7e008e | 1234 | |
8aa2a305 JW |
1235 | /* r6 controls the size of the move. 16 is decremented from it |
1236 | for each 64 bytes moved. Then the negative bit left over is used | |
1237 | as an index into a list of move instructions. e.g., a 72 byte move | |
1238 | would be set up with size(r6) = 14, for one iteration through the | |
1239 | big while loop, and a switch of -2 for the last part. */ | |
0d7e008e | 1240 | |
8aa2a305 JW |
1241 | final_switch = 16 - ((bytes / 4) % 16); |
1242 | while_loop = ((bytes / 4) / 16 - 1) * 16; | |
1243 | emit_insn (gen_move_insn (r6, GEN_INT (while_loop + final_switch))); | |
1244 | emit_insn (gen_block_lump_real (func_addr_rtx)); | |
1245 | return 1; | |
0d7e008e | 1246 | } |
0d7e008e | 1247 | |
d3ae8277 | 1248 | return 0; |
0d7e008e SC |
1249 | } |
1250 | ||
bc45ade3 | 1251 | /* Prepare operands for a move define_expand; specifically, one of the |
8aa2a305 | 1252 | operands must be in a register. */ |
bc45ade3 | 1253 | |
b9654711 | 1254 | int |
cf277499 | 1255 | prepare_move_operands (rtx operands[], enum machine_mode mode) |
bc45ade3 | 1256 | { |
463f02cd KK |
1257 | if ((mode == SImode || mode == DImode) |
1258 | && flag_pic | |
1259 | && ! ((mode == Pmode || mode == ptr_mode) | |
1260 | && tls_symbolic_operand (operands[1], Pmode) != 0)) | |
1a66cd67 AO |
1261 | { |
1262 | rtx temp; | |
1263 | if (SYMBOLIC_CONST_P (operands[1])) | |
1264 | { | |
1265 | if (GET_CODE (operands[0]) == MEM) | |
1266 | operands[1] = force_reg (Pmode, operands[1]); | |
5615d8fd SC |
1267 | else if (TARGET_SHMEDIA |
1268 | && GET_CODE (operands[1]) == LABEL_REF | |
fa5322fa AO |
1269 | && target_reg_operand (operands[0], mode)) |
1270 | /* It's ok. */; | |
1a66cd67 AO |
1271 | else |
1272 | { | |
b3a13419 ILT |
1273 | temp = (!can_create_pseudo_p () |
1274 | ? operands[0] | |
1275 | : gen_reg_rtx (Pmode)); | |
fa5322fa | 1276 | operands[1] = legitimize_pic_address (operands[1], mode, temp); |
1a66cd67 AO |
1277 | } |
1278 | } | |
1279 | else if (GET_CODE (operands[1]) == CONST | |
1280 | && GET_CODE (XEXP (operands[1], 0)) == PLUS | |
1281 | && SYMBOLIC_CONST_P (XEXP (XEXP (operands[1], 0), 0))) | |
1282 | { | |
b3a13419 | 1283 | temp = !can_create_pseudo_p () ? operands[0] : gen_reg_rtx (Pmode); |
1a66cd67 | 1284 | temp = legitimize_pic_address (XEXP (XEXP (operands[1], 0), 0), |
fa5322fa AO |
1285 | mode, temp); |
1286 | operands[1] = expand_binop (mode, add_optab, temp, | |
1a66cd67 | 1287 | XEXP (XEXP (operands[1], 0), 1), |
b3a13419 ILT |
1288 | (!can_create_pseudo_p () |
1289 | ? temp | |
1290 | : gen_reg_rtx (Pmode)), | |
68c37dc9 | 1291 | 0, OPTAB_LIB_WIDEN); |
1a66cd67 AO |
1292 | } |
1293 | } | |
1294 | ||
e22856e8 JW |
1295 | if (! reload_in_progress && ! reload_completed) |
1296 | { | |
1297 | /* Copy the source to a register if both operands aren't registers. */ | |
1298 | if (! register_operand (operands[0], mode) | |
d9da94a1 | 1299 | && ! sh_register_operand (operands[1], mode)) |
e22856e8 JW |
1300 | operands[1] = copy_to_mode_reg (mode, operands[1]); |
1301 | ||
266a2732 R |
1302 | if (GET_CODE (operands[0]) == MEM && ! memory_operand (operands[0], mode)) |
1303 | { | |
1304 | /* This is like change_address_1 (operands[0], mode, 0, 1) , | |
1305 | except that we can't use that function because it is static. */ | |
1306 | rtx new = change_address (operands[0], mode, 0); | |
1307 | MEM_COPY_ATTRIBUTES (new, operands[0]); | |
1308 | operands[0] = new; | |
1309 | } | |
1310 | ||
e22856e8 JW |
1311 | /* This case can happen while generating code to move the result |
1312 | of a library call to the target. Reject `st r0,@(rX,rY)' because | |
1313 | reload will fail to find a spill register for rX, since r0 is already | |
1314 | being used for the source. */ | |
73a4d10b R |
1315 | else if (TARGET_SH1 |
1316 | && refers_to_regno_p (R0_REG, R0_REG + 1, operands[1], (rtx *)0) | |
e22856e8 JW |
1317 | && GET_CODE (operands[0]) == MEM |
1318 | && GET_CODE (XEXP (operands[0], 0)) == PLUS | |
1319 | && GET_CODE (XEXP (XEXP (operands[0], 0), 1)) == REG) | |
1320 | operands[1] = copy_to_mode_reg (mode, operands[1]); | |
1321 | } | |
8aa2a305 | 1322 | |
463f02cd KK |
1323 | if (mode == Pmode || mode == ptr_mode) |
1324 | { | |
6058b2fa | 1325 | rtx op0, op1, opc; |
463f02cd KK |
1326 | enum tls_model tls_kind; |
1327 | ||
1328 | op0 = operands[0]; | |
1329 | op1 = operands[1]; | |
6058b2fa KK |
1330 | if (GET_CODE (op1) == CONST |
1331 | && GET_CODE (XEXP (op1, 0)) == PLUS | |
1332 | && tls_symbolic_operand (XEXP (XEXP (op1, 0), 0), Pmode)) | |
1333 | { | |
1334 | opc = XEXP (XEXP (op1, 0), 1); | |
1335 | op1 = XEXP (XEXP (op1, 0), 0); | |
1336 | } | |
1337 | else | |
1338 | opc = NULL_RTX; | |
1339 | ||
463f02cd KK |
1340 | if ((tls_kind = tls_symbolic_operand (op1, Pmode))) |
1341 | { | |
1342 | rtx tga_op1, tga_ret, tmp, tmp2; | |
1343 | ||
463f02cd KK |
1344 | switch (tls_kind) |
1345 | { | |
1346 | case TLS_MODEL_GLOBAL_DYNAMIC: | |
1347 | tga_ret = gen_rtx_REG (Pmode, R0_REG); | |
13ecc9e0 | 1348 | emit_call_insn (gen_tls_global_dynamic (tga_ret, op1)); |
463f02cd KK |
1349 | op1 = tga_ret; |
1350 | break; | |
1351 | ||
1352 | case TLS_MODEL_LOCAL_DYNAMIC: | |
1353 | tga_ret = gen_rtx_REG (Pmode, R0_REG); | |
13ecc9e0 | 1354 | emit_call_insn (gen_tls_local_dynamic (tga_ret, op1)); |
463f02cd KK |
1355 | |
1356 | tmp = gen_reg_rtx (Pmode); | |
1357 | emit_move_insn (tmp, tga_ret); | |
1358 | ||
1359 | if (register_operand (op0, Pmode)) | |
1360 | tmp2 = op0; | |
1361 | else | |
1362 | tmp2 = gen_reg_rtx (Pmode); | |
1363 | ||
1364 | emit_insn (gen_symDTPOFF2reg (tmp2, op1, tmp)); | |
1365 | op1 = tmp2; | |
1366 | break; | |
1367 | ||
1368 | case TLS_MODEL_INITIAL_EXEC: | |
1369 | if (! flag_pic) | |
b741fa00 | 1370 | { |
b92c3264 KK |
1371 | /* Don't schedule insns for getting GOT address when |
1372 | the first scheduling is enabled, to avoid spill | |
1373 | failures for R0. */ | |
1374 | if (flag_schedule_insns) | |
1375 | emit_insn (gen_blockage ()); | |
b741fa00 | 1376 | emit_insn (gen_GOTaddr2picreg ()); |
c41c1387 | 1377 | emit_use (gen_rtx_REG (SImode, PIC_REG)); |
b92c3264 KK |
1378 | if (flag_schedule_insns) |
1379 | emit_insn (gen_blockage ()); | |
b741fa00 | 1380 | } |
b3a13419 | 1381 | tga_op1 = !can_create_pseudo_p () ? op0 : gen_reg_rtx (Pmode); |
463f02cd KK |
1382 | tmp = gen_sym2GOTTPOFF (op1); |
1383 | emit_insn (gen_tls_initial_exec (tga_op1, tmp)); | |
1384 | op1 = tga_op1; | |
1385 | break; | |
1386 | ||
1387 | case TLS_MODEL_LOCAL_EXEC: | |
1388 | tmp2 = gen_reg_rtx (Pmode); | |
1389 | emit_insn (gen_load_gbr (tmp2)); | |
1390 | tmp = gen_reg_rtx (Pmode); | |
1391 | emit_insn (gen_symTPOFF2reg (tmp, op1)); | |
463f02cd KK |
1392 | |
1393 | if (register_operand (op0, Pmode)) | |
1394 | op1 = op0; | |
1395 | else | |
1396 | op1 = gen_reg_rtx (Pmode); | |
1397 | ||
1398 | emit_insn (gen_addsi3 (op1, tmp, tmp2)); | |
1399 | break; | |
1400 | ||
1401 | default: | |
f5b9e7c9 | 1402 | gcc_unreachable (); |
463f02cd | 1403 | } |
6058b2fa KK |
1404 | if (opc) |
1405 | emit_insn (gen_addsi3 (op1, op1, force_reg (SImode, opc))); | |
463f02cd KK |
1406 | operands[1] = op1; |
1407 | } | |
1408 | } | |
1409 | ||
8aa2a305 JW |
1410 | return 0; |
1411 | } | |
1412 | ||
78d310c2 R |
1413 | enum rtx_code |
1414 | prepare_cbranch_operands (rtx *operands, enum machine_mode mode, | |
1415 | enum rtx_code comparison) | |
1416 | { | |
1417 | rtx op1; | |
1418 | rtx scratch = NULL_RTX; | |
1419 | ||
1420 | if (comparison == CODE_FOR_nothing) | |
1421 | comparison = GET_CODE (operands[0]); | |
1422 | else | |
1423 | scratch = operands[4]; | |
1424 | if (GET_CODE (operands[1]) == CONST_INT | |
1425 | && GET_CODE (operands[2]) != CONST_INT) | |
1426 | { | |
1427 | rtx tmp = operands[1]; | |
1428 | ||
1429 | operands[1] = operands[2]; | |
1430 | operands[2] = tmp; | |
1431 | comparison = swap_condition (comparison); | |
1432 | } | |
1433 | if (GET_CODE (operands[2]) == CONST_INT) | |
1434 | { | |
1435 | HOST_WIDE_INT val = INTVAL (operands[2]); | |
1436 | if ((val == -1 || val == -0x81) | |
1437 | && (comparison == GT || comparison == LE)) | |
1438 | { | |
1439 | comparison = (comparison == GT) ? GE : LT; | |
1440 | operands[2] = gen_int_mode (val + 1, mode); | |
1441 | } | |
1442 | else if ((val == 1 || val == 0x80) | |
1443 | && (comparison == GE || comparison == LT)) | |
1444 | { | |
1445 | comparison = (comparison == GE) ? GT : LE; | |
1446 | operands[2] = gen_int_mode (val - 1, mode); | |
1447 | } | |
1448 | else if (val == 1 && (comparison == GEU || comparison == LTU)) | |
1449 | { | |
1450 | comparison = (comparison == GEU) ? NE : EQ; | |
1451 | operands[2] = CONST0_RTX (mode); | |
1452 | } | |
1453 | else if (val == 0x80 && (comparison == GEU || comparison == LTU)) | |
1454 | { | |
1455 | comparison = (comparison == GEU) ? GTU : LEU; | |
1456 | operands[2] = gen_int_mode (val - 1, mode); | |
1457 | } | |
1458 | else if (val == 0 && (comparison == GTU || comparison == LEU)) | |
1459 | comparison = (comparison == GTU) ? NE : EQ; | |
1460 | else if (mode == SImode | |
1461 | && ((val == 0x7fffffff | |
1462 | && (comparison == GTU || comparison == LEU)) | |
1463 | || ((unsigned HOST_WIDE_INT) val | |
1464 | == (unsigned HOST_WIDE_INT) 0x7fffffff + 1 | |
1465 | && (comparison == GEU || comparison == LTU)))) | |
1466 | { | |
1467 | comparison = (comparison == GTU || comparison == GEU) ? LT : GE; | |
1468 | operands[2] = CONST0_RTX (mode); | |
1469 | } | |
1470 | } | |
1471 | op1 = operands[1]; | |
b3a13419 | 1472 | if (can_create_pseudo_p ()) |
78d310c2 R |
1473 | operands[1] = force_reg (mode, op1); |
1474 | /* When we are handling DImode comparisons, we want to keep constants so | |
1475 | that we can optimize the component comparisons; however, memory loads | |
1476 | are better issued as a whole so that they can be scheduled well. | |
1477 | SImode equality comparisons allow I08 constants, but only when they | |
1478 | compare r0. Hence, if operands[1] has to be loaded from somewhere else | |
1479 | into a register, that register might as well be r0, and we allow the | |
1480 | constant. If it is already in a register, this is likely to be | |
2f8e468b | 1481 | allocated to a different hard register, thus we load the constant into |
78d310c2 R |
1482 | a register unless it is zero. */ |
1483 | if (!REG_P (operands[2]) | |
1484 | && (GET_CODE (operands[2]) != CONST_INT | |
1485 | || (mode == SImode && operands[2] != CONST0_RTX (SImode) | |
1486 | && ((comparison != EQ && comparison != NE) | |
1487 | || (REG_P (op1) && REGNO (op1) != R0_REG) | |
32a7ab3d | 1488 | || !satisfies_constraint_I08 (operands[2]))))) |
78d310c2 R |
1489 | { |
1490 | if (scratch && GET_MODE (scratch) == mode) | |
1491 | { | |
1492 | emit_move_insn (scratch, operands[2]); | |
1493 | operands[2] = scratch; | |
1494 | } | |
b3a13419 | 1495 | else if (can_create_pseudo_p ()) |
78d310c2 R |
1496 | operands[2] = force_reg (mode, operands[2]); |
1497 | } | |
1498 | return comparison; | |
1499 | } | |
1500 | ||
1501 | void | |
1502 | expand_cbranchsi4 (rtx *operands, enum rtx_code comparison, int probability) | |
1503 | { | |
1504 | rtx (*branch_expander) (rtx) = gen_branch_true; | |
1505 | rtx jump; | |
1506 | ||
1507 | comparison = prepare_cbranch_operands (operands, SImode, comparison); | |
1508 | switch (comparison) | |
1509 | { | |
1510 | case NE: case LT: case LE: case LTU: case LEU: | |
1511 | comparison = reverse_condition (comparison); | |
1512 | branch_expander = gen_branch_false; | |
1513 | default: ; | |
1514 | } | |
1515 | emit_insn (gen_rtx_SET (VOIDmode, gen_rtx_REG (SImode, T_REG), | |
1516 | gen_rtx_fmt_ee (comparison, SImode, | |
1517 | operands[1], operands[2]))); | |
1518 | jump = emit_jump_insn (branch_expander (operands[3])); | |
1519 | if (probability >= 0) | |
1520 | REG_NOTES (jump) | |
1521 | = gen_rtx_EXPR_LIST (REG_BR_PROB, GEN_INT (probability), | |
1522 | REG_NOTES (jump)); | |
1523 | ||
1524 | } | |
1525 | ||
1526 | /* ??? How should we distribute probabilities when more than one branch | |
1527 | is generated. So far we only have soem ad-hoc observations: | |
1528 | - If the operands are random, they are likely to differ in both parts. | |
1529 | - If comparing items in a hash chain, the operands are random or equal; | |
1530 | operation should be EQ or NE. | |
1531 | - If items are searched in an ordered tree from the root, we can expect | |
1532 | the highpart to be unequal about half of the time; operation should be | |
2f8e468b | 1533 | an inequality comparison, operands non-constant, and overall probability |
78d310c2 R |
1534 | about 50%. Likewise for quicksort. |
1535 | - Range checks will be often made against constants. Even if we assume for | |
1536 | simplicity an even distribution of the non-constant operand over a | |
1537 | sub-range here, the same probability could be generated with differently | |
1538 | wide sub-ranges - as long as the ratio of the part of the subrange that | |
1539 | is before the threshold to the part that comes after the threshold stays | |
1540 | the same. Thus, we can't really tell anything here; | |
1541 | assuming random distribution is at least simple. | |
1542 | */ | |
1543 | ||
1544 | bool | |
1545 | expand_cbranchdi4 (rtx *operands, enum rtx_code comparison) | |
1546 | { | |
1547 | enum rtx_code msw_taken, msw_skip, lsw_taken; | |
cb3d3842 | 1548 | rtx skip_label = NULL_RTX; |
78d310c2 R |
1549 | rtx op1h, op1l, op2h, op2l; |
1550 | int num_branches; | |
1551 | int prob, rev_prob; | |
1552 | int msw_taken_prob = -1, msw_skip_prob = -1, lsw_taken_prob = -1; | |
6f8eec29 | 1553 | rtx scratch = operands[4]; |
78d310c2 R |
1554 | |
1555 | comparison = prepare_cbranch_operands (operands, DImode, comparison); | |
1556 | op1h = gen_highpart_mode (SImode, DImode, operands[1]); | |
1557 | op2h = gen_highpart_mode (SImode, DImode, operands[2]); | |
1558 | op1l = gen_lowpart (SImode, operands[1]); | |
1559 | op2l = gen_lowpart (SImode, operands[2]); | |
1560 | msw_taken = msw_skip = lsw_taken = CODE_FOR_nothing; | |
1561 | prob = split_branch_probability; | |
1562 | rev_prob = REG_BR_PROB_BASE - prob; | |
1563 | switch (comparison) | |
1564 | { | |
1565 | /* ??? Should we use the cmpeqdi_t pattern for equality comparisons? | |
1566 | That costs 1 cycle more when the first branch can be predicted taken, | |
1567 | but saves us mispredicts because only one branch needs prediction. | |
1568 | It also enables generating the cmpeqdi_t-1 pattern. */ | |
1569 | case EQ: | |
1570 | if (TARGET_CMPEQDI_T) | |
1571 | { | |
1572 | emit_insn (gen_cmpeqdi_t (operands[1], operands[2])); | |
1573 | emit_jump_insn (gen_branch_true (operands[3])); | |
1574 | return true; | |
1575 | } | |
1576 | msw_skip = NE; | |
1577 | lsw_taken = EQ; | |
1578 | if (prob >= 0) | |
1579 | { | |
1580 | /* If we had more precision, we'd use rev_prob - (rev_prob >> 32) . | |
1581 | */ | |
1582 | msw_skip_prob = rev_prob; | |
1583 | if (REG_BR_PROB_BASE <= 65535) | |
1584 | lsw_taken_prob = prob ? REG_BR_PROB_BASE : 0; | |
1585 | else | |
1586 | { | |
1587 | gcc_assert (HOST_BITS_PER_WIDEST_INT >= 64); | |
1588 | lsw_taken_prob | |
1589 | = (prob | |
1590 | ? (REG_BR_PROB_BASE | |
1591 | - ((HOST_WIDEST_INT) REG_BR_PROB_BASE * rev_prob | |
1592 | / ((HOST_WIDEST_INT) prob << 32))) | |
1593 | : 0); | |
1594 | } | |
1595 | } | |
1596 | break; | |
1597 | case NE: | |
1598 | if (TARGET_CMPEQDI_T) | |
1599 | { | |
1600 | emit_insn (gen_cmpeqdi_t (operands[1], operands[2])); | |
1601 | emit_jump_insn (gen_branch_false (operands[3])); | |
1602 | return true; | |
1603 | } | |
1604 | msw_taken = NE; | |
6f8eec29 | 1605 | msw_taken_prob = prob; |
78d310c2 R |
1606 | lsw_taken = NE; |
1607 | lsw_taken_prob = 0; | |
1608 | break; | |
1609 | case GTU: case GT: | |
1610 | msw_taken = comparison; | |
1611 | if (GET_CODE (op2l) == CONST_INT && INTVAL (op2l) == -1) | |
1612 | break; | |
1613 | if (comparison != GTU || op2h != CONST0_RTX (SImode)) | |
1614 | msw_skip = swap_condition (msw_taken); | |
1615 | lsw_taken = GTU; | |
1616 | break; | |
1617 | case GEU: case GE: | |
1618 | if (op2l == CONST0_RTX (SImode)) | |
1619 | msw_taken = comparison; | |
1620 | else | |
1621 | { | |
1622 | msw_taken = comparison == GE ? GT : GTU; | |
1623 | msw_skip = swap_condition (msw_taken); | |
1624 | lsw_taken = GEU; | |
1625 | } | |
1626 | break; | |
1627 | case LTU: case LT: | |
1628 | msw_taken = comparison; | |
1629 | if (op2l == CONST0_RTX (SImode)) | |
1630 | break; | |
1631 | msw_skip = swap_condition (msw_taken); | |
1632 | lsw_taken = LTU; | |
1633 | break; | |
1634 | case LEU: case LE: | |
1635 | if (GET_CODE (op2l) == CONST_INT && INTVAL (op2l) == -1) | |
1636 | msw_taken = comparison; | |
1637 | else | |
1638 | { | |
1639 | lsw_taken = LEU; | |
1640 | if (comparison == LE) | |
1641 | msw_taken = LT; | |
1642 | else if (op2h != CONST0_RTX (SImode)) | |
1643 | msw_taken = LTU; | |
1644 | else | |
1645 | break; | |
1646 | msw_skip = swap_condition (msw_taken); | |
1647 | } | |
1648 | break; | |
1649 | default: return false; | |
1650 | } | |
1651 | num_branches = ((msw_taken != CODE_FOR_nothing) | |
1652 | + (msw_skip != CODE_FOR_nothing) | |
1653 | + (lsw_taken != CODE_FOR_nothing)); | |
1654 | if (comparison != EQ && comparison != NE && num_branches > 1) | |
1655 | { | |
1656 | if (!CONSTANT_P (operands[2]) | |
1657 | && prob >= (int) (REG_BR_PROB_BASE * 3 / 8U) | |
1658 | && prob <= (int) (REG_BR_PROB_BASE * 5 / 8U)) | |
1659 | { | |
1660 | msw_taken_prob = prob / 2U; | |
1661 | msw_skip_prob | |
1662 | = REG_BR_PROB_BASE * rev_prob / (REG_BR_PROB_BASE + rev_prob); | |
1663 | lsw_taken_prob = prob; | |
1664 | } | |
1665 | else | |
1666 | { | |
1667 | msw_taken_prob = prob; | |
1668 | msw_skip_prob = REG_BR_PROB_BASE; | |
1669 | /* ??? If we have a constant op2h, should we use that when | |
1670 | calculating lsw_taken_prob? */ | |
1671 | lsw_taken_prob = prob; | |
1672 | } | |
1673 | } | |
1674 | operands[1] = op1h; | |
1675 | operands[2] = op2h; | |
1676 | operands[4] = NULL_RTX; | |
6f8eec29 R |
1677 | if (reload_completed |
1678 | && ! arith_reg_or_0_operand (op2h, SImode) && true_regnum (op1h) | |
1679 | && (msw_taken != CODE_FOR_nothing || msw_skip != CODE_FOR_nothing)) | |
1680 | { | |
1681 | emit_move_insn (scratch, operands[2]); | |
1682 | operands[2] = scratch; | |
1683 | } | |
78d310c2 R |
1684 | if (msw_taken != CODE_FOR_nothing) |
1685 | expand_cbranchsi4 (operands, msw_taken, msw_taken_prob); | |
1686 | if (msw_skip != CODE_FOR_nothing) | |
1687 | { | |
1688 | rtx taken_label = operands[3]; | |
1689 | ||
d60819f8 CB |
1690 | /* Operands were possibly modified, but msw_skip doesn't expect this. |
1691 | Always use the original ones. */ | |
1692 | if (msw_taken != CODE_FOR_nothing) | |
1693 | { | |
1694 | operands[1] = op1h; | |
1695 | operands[2] = op2h; | |
1696 | } | |
1697 | ||
78d310c2 R |
1698 | operands[3] = skip_label = gen_label_rtx (); |
1699 | expand_cbranchsi4 (operands, msw_skip, msw_skip_prob); | |
1700 | operands[3] = taken_label; | |
1701 | } | |
1702 | operands[1] = op1l; | |
1703 | operands[2] = op2l; | |
1704 | if (lsw_taken != CODE_FOR_nothing) | |
6f8eec29 R |
1705 | { |
1706 | if (reload_completed | |
1707 | && ! arith_reg_or_0_operand (op2l, SImode) && true_regnum (op1l)) | |
1708 | operands[4] = scratch; | |
1709 | expand_cbranchsi4 (operands, lsw_taken, lsw_taken_prob); | |
1710 | } | |
78d310c2 R |
1711 | if (msw_skip != CODE_FOR_nothing) |
1712 | emit_label (skip_label); | |
1713 | return true; | |
1714 | } | |
1715 | ||
8aa2a305 JW |
1716 | /* Prepare the operands for an scc instruction; make sure that the |
1717 | compare has been done. */ | |
1718 | rtx | |
cf277499 | 1719 | prepare_scc_operands (enum rtx_code code) |
8aa2a305 | 1720 | { |
c5c76735 | 1721 | rtx t_reg = gen_rtx_REG (SImode, T_REG); |
8aa2a305 | 1722 | enum rtx_code oldcode = code; |
9374bd85 | 1723 | enum machine_mode mode; |
8aa2a305 JW |
1724 | |
1725 | /* First need a compare insn. */ | |
1726 | switch (code) | |
bc45ade3 | 1727 | { |
8aa2a305 JW |
1728 | case NE: |
1729 | /* It isn't possible to handle this case. */ | |
f5b9e7c9 | 1730 | gcc_unreachable (); |
8aa2a305 JW |
1731 | case LT: |
1732 | code = GT; | |
1733 | break; | |
1734 | case LE: | |
1735 | code = GE; | |
1736 | break; | |
1737 | case LTU: | |
1738 | code = GTU; | |
1739 | break; | |
1740 | case LEU: | |
1741 | code = GEU; | |
1742 | break; | |
10db12e0 R |
1743 | default: |
1744 | break; | |
bc45ade3 | 1745 | } |
8aa2a305 | 1746 | if (code != oldcode) |
b9654711 | 1747 | { |
8aa2a305 JW |
1748 | rtx tmp = sh_compare_op0; |
1749 | sh_compare_op0 = sh_compare_op1; | |
1750 | sh_compare_op1 = tmp; | |
0d7e008e | 1751 | } |
b9654711 | 1752 | |
9374bd85 RK |
1753 | mode = GET_MODE (sh_compare_op0); |
1754 | if (mode == VOIDmode) | |
1755 | mode = GET_MODE (sh_compare_op1); | |
1756 | ||
1757 | sh_compare_op0 = force_reg (mode, sh_compare_op0); | |
1245df60 R |
1758 | if ((code != EQ && code != NE |
1759 | && (sh_compare_op1 != const0_rtx | |
1760 | || code == GTU || code == GEU || code == LTU || code == LEU)) | |
89e43e33 | 1761 | || (mode == DImode && sh_compare_op1 != const0_rtx) |
3a8699c7 | 1762 | || (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT)) |
9374bd85 | 1763 | sh_compare_op1 = force_reg (mode, sh_compare_op1); |
0d7e008e | 1764 | |
157371cf | 1765 | if ((TARGET_SH4 || TARGET_SH2A) && GET_MODE_CLASS (mode) == MODE_FLOAT) |
225e4f43 | 1766 | (mode == SFmode ? emit_sf_insn : emit_df_insn) |
c0d4e710 KH |
1767 | (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, |
1768 | gen_rtx_SET (VOIDmode, t_reg, | |
0f4c242b KH |
1769 | gen_rtx_fmt_ee (code, SImode, |
1770 | sh_compare_op0, sh_compare_op1)), | |
c0d4e710 | 1771 | gen_rtx_USE (VOIDmode, get_fpscr_rtx ())))); |
225e4f43 | 1772 | else |
c0d4e710 | 1773 | emit_insn (gen_rtx_SET (VOIDmode, t_reg, |
0f4c242b KH |
1774 | gen_rtx_fmt_ee (code, SImode, |
1775 | sh_compare_op0, sh_compare_op1))); | |
0d7e008e | 1776 | |
8aa2a305 | 1777 | return t_reg; |
bc45ade3 SC |
1778 | } |
1779 | ||
8aa2a305 JW |
1780 | /* Called from the md file, set up the operands of a compare instruction. */ |
1781 | ||
1782 | void | |
cf277499 | 1783 | from_compare (rtx *operands, int code) |
bc45ade3 | 1784 | { |
1245df60 R |
1785 | enum machine_mode mode = GET_MODE (sh_compare_op0); |
1786 | rtx insn; | |
1787 | if (mode == VOIDmode) | |
1788 | mode = GET_MODE (sh_compare_op1); | |
1789 | if (code != EQ | |
1790 | || mode == DImode | |
3a8699c7 | 1791 | || (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT)) |
bc45ade3 | 1792 | { |
8aa2a305 | 1793 | /* Force args into regs, since we can't use constants here. */ |
24a25d45 | 1794 | sh_compare_op0 = force_reg (mode, sh_compare_op0); |
8aa2a305 | 1795 | if (sh_compare_op1 != const0_rtx |
1245df60 | 1796 | || code == GTU || code == GEU |
3a8699c7 | 1797 | || (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT)) |
24a25d45 | 1798 | sh_compare_op1 = force_reg (mode, sh_compare_op1); |
bc45ade3 | 1799 | } |
3a8699c7 | 1800 | if (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT && code == GE) |
1245df60 R |
1801 | { |
1802 | from_compare (operands, GT); | |
1803 | insn = gen_ieee_ccmpeqsf_t (sh_compare_op0, sh_compare_op1); | |
1804 | } | |
1805 | else | |
c5c76735 | 1806 | insn = gen_rtx_SET (VOIDmode, |
4773afa4 | 1807 | gen_rtx_REG (SImode, T_REG), |
0f4c242b KH |
1808 | gen_rtx_fmt_ee (code, SImode, |
1809 | sh_compare_op0, sh_compare_op1)); | |
157371cf | 1810 | if ((TARGET_SH4 || TARGET_SH2A) && GET_MODE_CLASS (mode) == MODE_FLOAT) |
225e4f43 | 1811 | { |
c0d4e710 | 1812 | insn = gen_rtx_PARALLEL (VOIDmode, |
225e4f43 | 1813 | gen_rtvec (2, insn, |
c0d4e710 | 1814 | gen_rtx_USE (VOIDmode, get_fpscr_rtx ()))); |
225e4f43 R |
1815 | (mode == SFmode ? emit_sf_insn : emit_df_insn) (insn); |
1816 | } | |
1817 | else | |
1818 | emit_insn (insn); | |
bc45ade3 SC |
1819 | } |
1820 | \f | |
16bea517 | 1821 | /* Functions to output assembly code. */ |
bc45ade3 | 1822 | |
b9654711 | 1823 | /* Return a sequence of instructions to perform DI or DF move. |
bc45ade3 | 1824 | |
b9654711 | 1825 | Since the SH cannot move a DI or DF in one instruction, we have |
16bea517 | 1826 | to take care when we see overlapping source and dest registers. */ |
0d7e008e | 1827 | |
318881c0 | 1828 | const char * |
cf277499 SB |
1829 | output_movedouble (rtx insn ATTRIBUTE_UNUSED, rtx operands[], |
1830 | enum machine_mode mode) | |
bc45ade3 | 1831 | { |
b9654711 SC |
1832 | rtx dst = operands[0]; |
1833 | rtx src = operands[1]; | |
b9654711 | 1834 | |
0d7e008e | 1835 | if (GET_CODE (dst) == MEM |
16bea517 | 1836 | && GET_CODE (XEXP (dst, 0)) == PRE_DEC) |
8aa2a305 JW |
1837 | return "mov.l %T1,%0\n\tmov.l %1,%0"; |
1838 | ||
b9654711 SC |
1839 | if (register_operand (dst, mode) |
1840 | && register_operand (src, mode)) | |
bc45ade3 | 1841 | { |
b9654711 | 1842 | if (REGNO (src) == MACH_REG) |
00f8ff66 | 1843 | return "sts mach,%S0\n\tsts macl,%R0"; |
bc45ade3 | 1844 | |
8aa2a305 JW |
1845 | /* When mov.d r1,r2 do r2->r3 then r1->r2; |
1846 | when mov.d r1,r0 do r1->r0 then r2->r1. */ | |
b9654711 SC |
1847 | |
1848 | if (REGNO (src) + 1 == REGNO (dst)) | |
5325c0fa | 1849 | return "mov %T1,%T0\n\tmov %1,%0"; |
b9654711 | 1850 | else |
5325c0fa | 1851 | return "mov %1,%0\n\tmov %T1,%T0"; |
b9654711 SC |
1852 | } |
1853 | else if (GET_CODE (src) == CONST_INT) | |
bc45ade3 | 1854 | { |
8aa2a305 JW |
1855 | if (INTVAL (src) < 0) |
1856 | output_asm_insn ("mov #-1,%S0", operands); | |
bc45ade3 | 1857 | else |
8aa2a305 | 1858 | output_asm_insn ("mov #0,%S0", operands); |
bc45ade3 | 1859 | |
8aa2a305 | 1860 | return "mov %1,%R0"; |
0d7e008e | 1861 | } |
b9654711 | 1862 | else if (GET_CODE (src) == MEM) |
bc45ade3 | 1863 | { |
8aa2a305 | 1864 | int ptrreg = -1; |
b9654711 SC |
1865 | int dreg = REGNO (dst); |
1866 | rtx inside = XEXP (src, 0); | |
bc45ade3 | 1867 | |
f5b9e7c9 | 1868 | switch (GET_CODE (inside)) |
bc45ade3 | 1869 | { |
f5b9e7c9 NS |
1870 | case REG: |
1871 | ptrreg = REGNO (inside); | |
1872 | break; | |
1873 | ||
1874 | case SUBREG: | |
1875 | ptrreg = subreg_regno (inside); | |
1876 | break; | |
1877 | ||
1878 | case PLUS: | |
8aa2a305 JW |
1879 | ptrreg = REGNO (XEXP (inside, 0)); |
1880 | /* ??? A r0+REG address shouldn't be possible here, because it isn't | |
1881 | an offsettable address. Unfortunately, offsettable addresses use | |
1882 | QImode to check the offset, and a QImode offsettable address | |
1883 | requires r0 for the other operand, which is not currently | |
1884 | supported, so we can't use the 'o' constraint. | |
1885 | Thus we must check for and handle r0+REG addresses here. | |
1886 | We punt for now, since this is likely very rare. */ | |
f5b9e7c9 NS |
1887 | gcc_assert (GET_CODE (XEXP (inside, 1)) != REG); |
1888 | break; | |
1889 | ||
1890 | case LABEL_REF: | |
1891 | return "mov.l %1,%0\n\tmov.l %1+4,%T0"; | |
1892 | case POST_INC: | |
1893 | return "mov.l %1,%0\n\tmov.l %1,%T0"; | |
1894 | default: | |
1895 | gcc_unreachable (); | |
bc45ade3 | 1896 | } |
bc45ade3 | 1897 | |
8aa2a305 JW |
1898 | /* Work out the safe way to copy. Copy into the second half first. */ |
1899 | if (dreg == ptrreg) | |
1900 | return "mov.l %T1,%T0\n\tmov.l %1,%0"; | |
bc45ade3 SC |
1901 | } |
1902 | ||
00f8ff66 | 1903 | return "mov.l %1,%0\n\tmov.l %T1,%T0"; |
bc45ade3 SC |
1904 | } |
1905 | ||
8aa2a305 JW |
1906 | /* Print an instruction which would have gone into a delay slot after |
1907 | another instruction, but couldn't because the other instruction expanded | |
1908 | into a sequence where putting the slot insn at the end wouldn't work. */ | |
0d7e008e | 1909 | |
8aa2a305 | 1910 | static void |
cf277499 | 1911 | print_slot (rtx insn) |
8aa2a305 | 1912 | { |
c9d691e9 | 1913 | final_scan_insn (XVECEXP (insn, 0, 1), asm_out_file, optimize, 1, NULL); |
b9654711 | 1914 | |
8aa2a305 JW |
1915 | INSN_DELETED_P (XVECEXP (insn, 0, 1)) = 1; |
1916 | } | |
0d7e008e | 1917 | |
318881c0 | 1918 | const char * |
cf277499 | 1919 | output_far_jump (rtx insn, rtx op) |
0d7e008e | 1920 | { |
1245df60 | 1921 | struct { rtx lab, reg, op; } this; |
950a3816 | 1922 | rtx braf_base_lab = NULL_RTX; |
318881c0 | 1923 | const char *jump; |
1245df60 | 1924 | int far; |
9d98a694 | 1925 | int offset = branch_dest (insn) - INSN_ADDRESSES (INSN_UID (insn)); |
10f4f635 | 1926 | rtx prev; |
0d7e008e | 1927 | |
1245df60 | 1928 | this.lab = gen_label_rtx (); |
b9654711 | 1929 | |
1a953d0f R |
1930 | if (TARGET_SH2 |
1931 | && offset >= -32764 | |
1932 | && offset - get_attr_length (insn) <= 32766) | |
1245df60 R |
1933 | { |
1934 | far = 0; | |
e6dfd05f | 1935 | jump = "mov.w %O0,%1; braf %1"; |
1245df60 R |
1936 | } |
1937 | else | |
1938 | { | |
1939 | far = 1; | |
1a66cd67 | 1940 | if (flag_pic) |
e6dfd05f AO |
1941 | { |
1942 | if (TARGET_SH2) | |
1943 | jump = "mov.l %O0,%1; braf %1"; | |
1944 | else | |
1945 | jump = "mov.l r0,@-r15; mova %O0,r0; mov.l @r0,%1; add r0,%1; mov.l @r15+,r0; jmp @%1"; | |
1946 | } | |
1a66cd67 | 1947 | else |
e6dfd05f | 1948 | jump = "mov.l %O0,%1; jmp @%1"; |
1245df60 R |
1949 | } |
1950 | /* If we have a scratch register available, use it. */ | |
10f4f635 R |
1951 | if (GET_CODE ((prev = prev_nonnote_insn (insn))) == INSN |
1952 | && INSN_CODE (prev) == CODE_FOR_indirect_jump_scratch) | |
1245df60 | 1953 | { |
10f4f635 | 1954 | this.reg = SET_DEST (XVECEXP (PATTERN (prev), 0, 0)); |
e6dfd05f AO |
1955 | if (REGNO (this.reg) == R0_REG && flag_pic && ! TARGET_SH2) |
1956 | jump = "mov.l r1,@-r15; mova %O0,r0; mov.l @r0,r1; add r1,r0; mov.l @r15+,r1; jmp @%1"; | |
1245df60 R |
1957 | output_asm_insn (jump, &this.lab); |
1958 | if (dbr_sequence_length ()) | |
1959 | print_slot (final_sequence); | |
1960 | else | |
1961 | output_asm_insn ("nop", 0); | |
1962 | } | |
1963 | else | |
1964 | { | |
1965 | /* Output the delay slot insn first if any. */ | |
1966 | if (dbr_sequence_length ()) | |
1967 | print_slot (final_sequence); | |
1968 | ||
c5c76735 | 1969 | this.reg = gen_rtx_REG (SImode, 13); |
fa5322fa AO |
1970 | /* We must keep the stack aligned to 8-byte boundaries on SH5. |
1971 | Fortunately, MACL is fixed and call-clobbered, and we never | |
1972 | need its value across jumps, so save r13 in it instead of in | |
1973 | the stack. */ | |
1974 | if (TARGET_SH5) | |
1975 | output_asm_insn ("lds r13, macl", 0); | |
1976 | else | |
1977 | output_asm_insn ("mov.l r13,@-r15", 0); | |
1245df60 | 1978 | output_asm_insn (jump, &this.lab); |
fa5322fa AO |
1979 | if (TARGET_SH5) |
1980 | output_asm_insn ("sts macl, r13", 0); | |
1981 | else | |
1982 | output_asm_insn ("mov.l @r15+,r13", 0); | |
1245df60 | 1983 | } |
e6dfd05f AO |
1984 | if (far && flag_pic && TARGET_SH2) |
1985 | { | |
1986 | braf_base_lab = gen_label_rtx (); | |
4977bab6 | 1987 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
e6dfd05f AO |
1988 | CODE_LABEL_NUMBER (braf_base_lab)); |
1989 | } | |
1245df60 R |
1990 | if (far) |
1991 | output_asm_insn (".align 2", 0); | |
4977bab6 | 1992 | (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (this.lab)); |
1245df60 | 1993 | this.op = op; |
1a66cd67 | 1994 | if (far && flag_pic) |
e6dfd05f AO |
1995 | { |
1996 | if (TARGET_SH2) | |
1997 | this.lab = braf_base_lab; | |
1998 | output_asm_insn (".long %O2-%O0", &this.lab); | |
1999 | } | |
1a66cd67 AO |
2000 | else |
2001 | output_asm_insn (far ? ".long %O2" : ".word %O2-%O0", &this.lab); | |
0d7e008e SC |
2002 | return ""; |
2003 | } | |
bc45ade3 | 2004 | |
8aa2a305 JW |
2005 | /* Local label counter, used for constants in the pool and inside |
2006 | pattern branches. */ | |
2007 | ||
2008 | static int lf = 100; | |
2009 | ||
2010 | /* Output code for ordinary branches. */ | |
2011 | ||
318881c0 | 2012 | const char * |
cf277499 | 2013 | output_branch (int logic, rtx insn, rtx *operands) |
bc45ade3 | 2014 | { |
afbc5d86 | 2015 | switch (get_attr_length (insn)) |
bc45ade3 | 2016 | { |
33f7f353 JR |
2017 | case 6: |
2018 | /* This can happen if filling the delay slot has caused a forward | |
2019 | branch to exceed its range (we could reverse it, but only | |
2020 | when we know we won't overextend other branches; this should | |
2021 | best be handled by relaxation). | |
2022 | It can also happen when other condbranches hoist delay slot insn | |
1245df60 R |
2023 | from their destination, thus leading to code size increase. |
2024 | But the branch will still be in the range -4092..+4098 bytes. */ | |
bc45ade3 | 2025 | |
33f7f353 | 2026 | if (! TARGET_RELAX) |
1245df60 | 2027 | { |
33f7f353 JR |
2028 | int label = lf++; |
2029 | /* The call to print_slot will clobber the operands. */ | |
2030 | rtx op0 = operands[0]; | |
f676971a | 2031 | |
33f7f353 JR |
2032 | /* If the instruction in the delay slot is annulled (true), then |
2033 | there is no delay slot where we can put it now. The only safe | |
2034 | place for it is after the label. final will do that by default. */ | |
f676971a | 2035 | |
33f7f353 | 2036 | if (final_sequence |
bee9efbc KK |
2037 | && ! INSN_ANNULLED_BRANCH_P (XVECEXP (final_sequence, 0, 0)) |
2038 | && get_attr_length (XVECEXP (final_sequence, 0, 1))) | |
33f7f353 | 2039 | { |
afbc5d86 | 2040 | asm_fprintf (asm_out_file, "\tb%s%ss\t%LLF%d\n", logic ? "f" : "t", |
33f7f353 JR |
2041 | ASSEMBLER_DIALECT ? "/" : ".", label); |
2042 | print_slot (final_sequence); | |
2043 | } | |
2044 | else | |
afbc5d86 | 2045 | asm_fprintf (asm_out_file, "\tb%s\t%LLF%d\n", logic ? "f" : "t", label); |
f676971a | 2046 | |
afbc5d86 R |
2047 | output_asm_insn ("bra\t%l0", &op0); |
2048 | fprintf (asm_out_file, "\tnop\n"); | |
832a3292 | 2049 | (*targetm.asm_out.internal_label) (asm_out_file, "LF", label); |
f676971a | 2050 | |
33f7f353 | 2051 | return ""; |
1245df60 | 2052 | } |
33f7f353 JR |
2053 | /* When relaxing, handle this like a short branch. The linker |
2054 | will fix it up if it still doesn't fit after relaxation. */ | |
2055 | case 2: | |
2056 | return logic ? "bt%.\t%l0" : "bf%.\t%l0"; | |
3a8699c7 AO |
2057 | |
2058 | /* These are for SH2e, in which we have to account for the | |
2059 | extra nop because of the hardware bug in annulled branches. */ | |
2060 | case 8: | |
2061 | if (! TARGET_RELAX) | |
2062 | { | |
2063 | int label = lf++; | |
2064 | ||
f5b9e7c9 NS |
2065 | gcc_assert (!final_sequence |
2066 | || !(INSN_ANNULLED_BRANCH_P | |
2067 | (XVECEXP (final_sequence, 0, 0)))); | |
3a8699c7 AO |
2068 | asm_fprintf (asm_out_file, "b%s%ss\t%LLF%d\n", |
2069 | logic ? "f" : "t", | |
2070 | ASSEMBLER_DIALECT ? "/" : ".", label); | |
2071 | fprintf (asm_out_file, "\tnop\n"); | |
2072 | output_asm_insn ("bra\t%l0", operands); | |
2073 | fprintf (asm_out_file, "\tnop\n"); | |
2074 | (*targetm.asm_out.internal_label) (asm_out_file, "LF", label); | |
2075 | ||
2076 | return ""; | |
2077 | } | |
2078 | /* When relaxing, fall through. */ | |
2079 | case 4: | |
2080 | { | |
2081 | char buffer[10]; | |
f676971a | 2082 | |
3a8699c7 AO |
2083 | sprintf (buffer, "b%s%ss\t%%l0", |
2084 | logic ? "t" : "f", | |
2085 | ASSEMBLER_DIALECT ? "/" : "."); | |
2086 | output_asm_insn (buffer, &operands[0]); | |
2087 | return "nop"; | |
2088 | } | |
2089 | ||
33f7f353 | 2090 | default: |
afbc5d86 R |
2091 | /* There should be no longer branches now - that would |
2092 | indicate that something has destroyed the branches set | |
2093 | up in machine_dependent_reorg. */ | |
f5b9e7c9 | 2094 | gcc_unreachable (); |
1245df60 | 2095 | } |
1245df60 | 2096 | } |
bc45ade3 | 2097 | |
78d310c2 R |
2098 | /* Output a code sequence for INSN using TEMPLATE with OPERANDS; but before, |
2099 | fill in operands 9 as a label to the successor insn. | |
2100 | We try to use jump threading where possible. | |
2101 | IF CODE matches the comparison in the IF_THEN_ELSE of a following jump, | |
2102 | we assume the jump is taken. I.e. EQ means follow jmp and bf, NE means | |
2103 | follow jmp and bt, if the address is in range. */ | |
318881c0 | 2104 | const char * |
cf277499 SB |
2105 | output_branchy_insn (enum rtx_code code, const char *template, |
2106 | rtx insn, rtx *operands) | |
1245df60 R |
2107 | { |
2108 | rtx next_insn = NEXT_INSN (insn); | |
79b2746a | 2109 | |
1245df60 R |
2110 | if (next_insn && GET_CODE (next_insn) == JUMP_INSN && condjump_p (next_insn)) |
2111 | { | |
2112 | rtx src = SET_SRC (PATTERN (next_insn)); | |
2113 | if (GET_CODE (src) == IF_THEN_ELSE && GET_CODE (XEXP (src, 0)) != code) | |
2114 | { | |
2115 | /* Following branch not taken */ | |
2116 | operands[9] = gen_label_rtx (); | |
2117 | emit_label_after (operands[9], next_insn); | |
9d98a694 AO |
2118 | INSN_ADDRESSES_NEW (operands[9], |
2119 | INSN_ADDRESSES (INSN_UID (next_insn)) | |
2120 | + get_attr_length (next_insn)); | |
1245df60 R |
2121 | return template; |
2122 | } | |
2123 | else | |
2124 | { | |
33f7f353 | 2125 | int offset = (branch_dest (next_insn) |
9d98a694 | 2126 | - INSN_ADDRESSES (INSN_UID (next_insn)) + 4); |
33f7f353 | 2127 | if (offset >= -252 && offset <= 258) |
1245df60 R |
2128 | { |
2129 | if (GET_CODE (src) == IF_THEN_ELSE) | |
2130 | /* branch_true */ | |
2131 | src = XEXP (src, 1); | |
2132 | operands[9] = src; | |
2133 | return template; | |
2134 | } | |
2135 | } | |
bc45ade3 | 2136 | } |
1245df60 R |
2137 | operands[9] = gen_label_rtx (); |
2138 | emit_label_after (operands[9], insn); | |
9d98a694 AO |
2139 | INSN_ADDRESSES_NEW (operands[9], |
2140 | INSN_ADDRESSES (INSN_UID (insn)) | |
2141 | + get_attr_length (insn)); | |
1245df60 R |
2142 | return template; |
2143 | } | |
e4fa6b06 | 2144 | |
318881c0 | 2145 | const char * |
cf277499 | 2146 | output_ieee_ccmpeq (rtx insn, rtx *operands) |
1245df60 | 2147 | { |
9fd0c419 R |
2148 | return output_branchy_insn (NE, "bt\t%l9\n\tfcmp/eq\t%1,%0", |
2149 | insn, operands); | |
bc45ade3 | 2150 | } |
8e87e161 | 2151 | \f |
1bc7c5b6 | 2152 | /* Output the start of the assembler file. */ |
8aa2a305 | 2153 | |
1bc7c5b6 | 2154 | static void |
cf277499 | 2155 | sh_file_start (void) |
bc45ade3 | 2156 | { |
1bc7c5b6 ZW |
2157 | default_file_start (); |
2158 | ||
b55e3aad NC |
2159 | #ifdef SYMBIAN |
2160 | /* Declare the .directive section before it is used. */ | |
2161 | fputs ("\t.section .directive, \"SM\", @progbits, 1\n", asm_out_file); | |
2162 | fputs ("\t.asciz \"#<SYMEDIT>#\\n\"\n", asm_out_file); | |
2163 | #endif | |
f676971a | 2164 | |
1bc7c5b6 ZW |
2165 | if (TARGET_ELF) |
2166 | /* We need to show the text section with the proper | |
2167 | attributes as in TEXT_SECTION_ASM_OP, before dwarf2out | |
0fe7abcc | 2168 | emits it without attributes in TEXT_SECTION_ASM_OP, else GAS |
1bc7c5b6 ZW |
2169 | will complain. We can teach GAS specifically about the |
2170 | default attributes for our choice of text section, but | |
2171 | then we would have to change GAS again if/when we change | |
2172 | the text section name. */ | |
2173 | fprintf (asm_out_file, "%s\n", TEXT_SECTION_ASM_OP); | |
2174 | else | |
2175 | /* Switch to the data section so that the coffsem symbol | |
2176 | isn't in the text section. */ | |
d6b5193b | 2177 | switch_to_section (data_section); |
b9654711 | 2178 | |
8aa2a305 | 2179 | if (TARGET_LITTLE_ENDIAN) |
1bc7c5b6 | 2180 | fputs ("\t.little\n", asm_out_file); |
fa5322fa | 2181 | |
1bc7c5b6 ZW |
2182 | if (!TARGET_ELF) |
2183 | { | |
2184 | if (TARGET_SHCOMPACT) | |
2185 | fputs ("\t.mode\tSHcompact\n", asm_out_file); | |
2186 | else if (TARGET_SHMEDIA) | |
2187 | fprintf (asm_out_file, "\t.mode\tSHmedia\n\t.abi\t%i\n", | |
2188 | TARGET_SHMEDIA64 ? 64 : 32); | |
2189 | } | |
bc45ade3 | 2190 | } |
0d7e008e | 2191 | \f |
9f3a9a08 KK |
2192 | /* Check if PAT includes UNSPEC_CALLER unspec pattern. */ |
2193 | ||
2194 | static bool | |
cf277499 | 2195 | unspec_caller_rtx_p (rtx pat) |
9f3a9a08 KK |
2196 | { |
2197 | switch (GET_CODE (pat)) | |
2198 | { | |
2199 | case CONST: | |
2200 | return unspec_caller_rtx_p (XEXP (pat, 0)); | |
2201 | case PLUS: | |
2202 | case MINUS: | |
2203 | if (unspec_caller_rtx_p (XEXP (pat, 0))) | |
2204 | return true; | |
2205 | return unspec_caller_rtx_p (XEXP (pat, 1)); | |
2206 | case UNSPEC: | |
2207 | if (XINT (pat, 1) == UNSPEC_CALLER) | |
2208 | return true; | |
2209 | default: | |
2210 | break; | |
2211 | } | |
2212 | ||
2213 | return false; | |
2214 | } | |
2215 | ||
2216 | /* Indicate that INSN cannot be duplicated. This is true for insn | |
569b7f6a | 2217 | that generates a unique label. */ |
9f3a9a08 KK |
2218 | |
2219 | static bool | |
cf277499 | 2220 | sh_cannot_copy_insn_p (rtx insn) |
9f3a9a08 KK |
2221 | { |
2222 | rtx pat; | |
2223 | ||
2224 | if (!reload_completed || !flag_pic) | |
2225 | return false; | |
2226 | ||
2227 | if (GET_CODE (insn) != INSN) | |
2228 | return false; | |
2229 | if (asm_noperands (insn) >= 0) | |
2230 | return false; | |
2231 | ||
2232 | pat = PATTERN (insn); | |
2233 | if (GET_CODE (pat) != SET) | |
2234 | return false; | |
2235 | pat = SET_SRC (pat); | |
2236 | ||
2237 | if (unspec_caller_rtx_p (pat)) | |
2238 | return true; | |
2239 | ||
2240 | return false; | |
2241 | } | |
2242 | \f | |
16bea517 | 2243 | /* Actual number of instructions used to make a shift by N. */ |
0b5826ac | 2244 | static const char ashiftrt_insns[] = |
16bea517 JW |
2245 | { 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}; |
2246 | ||
2247 | /* Left shift and logical right shift are the same. */ | |
0b5826ac | 2248 | static const char shift_insns[] = |
16bea517 | 2249 | { 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 | 2250 | |
16bea517 JW |
2251 | /* Individual shift amounts needed to get the above length sequences. |
2252 | One bit right shifts clobber the T bit, so when possible, put one bit | |
2253 | shifts in the middle of the sequence, so the ends are eligible for | |
2254 | branch delay slots. */ | |
0139adca | 2255 | static const short shift_amounts[32][5] = { |
16bea517 JW |
2256 | {0}, {1}, {2}, {2, 1}, |
2257 | {2, 2}, {2, 1, 2}, {2, 2, 2}, {2, 2, 1, 2}, | |
2258 | {8}, {8, 1}, {8, 2}, {8, 1, 2}, | |
2259 | {8, 2, 2}, {8, 2, 1, 2}, {8, -2, 8}, {8, -1, 8}, | |
2260 | {16}, {16, 1}, {16, 2}, {16, 1, 2}, | |
2261 | {16, 2, 2}, {16, 2, 1, 2}, {16, -2, 8}, {16, -1, 8}, | |
2262 | {16, 8}, {16, 1, 8}, {16, 8, 2}, {16, 8, 1, 2}, | |
1d3534c0 | 2263 | {16, 8, 2, 2}, {16, -1, -2, 16}, {16, -2, 16}, {16, -1, 16}}; |
16bea517 | 2264 | |
8d481241 JW |
2265 | /* Likewise, but for shift amounts < 16, up to three highmost bits |
2266 | might be clobbered. This is typically used when combined with some | |
2267 | kind of sign or zero extension. */ | |
f676971a | 2268 | |
0b5826ac | 2269 | static const char ext_shift_insns[] = |
8d481241 JW |
2270 | { 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}; |
2271 | ||
0b5826ac | 2272 | static const short ext_shift_amounts[32][4] = { |
8d481241 JW |
2273 | {0}, {1}, {2}, {2, 1}, |
2274 | {2, 2}, {2, 1, 2}, {8, -2}, {8, -1}, | |
2275 | {8}, {8, 1}, {8, 2}, {8, 1, 2}, | |
2276 | {8, 2, 2}, {16, -2, -1}, {16, -2}, {16, -1}, | |
2277 | {16}, {16, 1}, {16, 2}, {16, 1, 2}, | |
2278 | {16, 2, 2}, {16, 2, 1, 2}, {16, -2, 8}, {16, -1, 8}, | |
2279 | {16, 8}, {16, 1, 8}, {16, 8, 2}, {16, 8, 1, 2}, | |
2280 | {16, 8, 2, 2}, {16, -1, -2, 16}, {16, -2, 16}, {16, -1, 16}}; | |
2281 | ||
ae9d19c0 JR |
2282 | /* Assuming we have a value that has been sign-extended by at least one bit, |
2283 | can we use the ext_shift_amounts with the last shift turned to an arithmetic shift | |
2284 | to shift it by N without data loss, and quicker than by other means? */ | |
2285 | #define EXT_SHIFT_SIGNED(n) (((n) | 8) == 15) | |
2286 | ||
16bea517 JW |
2287 | /* This is used in length attributes in sh.md to help compute the length |
2288 | of arbitrary constant shift instructions. */ | |
bc45ade3 | 2289 | |
16bea517 | 2290 | int |
cf277499 | 2291 | shift_insns_rtx (rtx insn) |
16bea517 JW |
2292 | { |
2293 | rtx set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
2294 | int shift_count = INTVAL (XEXP (set_src, 1)); | |
2295 | enum rtx_code shift_code = GET_CODE (set_src); | |
00f8ff66 | 2296 | |
16bea517 JW |
2297 | switch (shift_code) |
2298 | { | |
2299 | case ASHIFTRT: | |
2300 | return ashiftrt_insns[shift_count]; | |
2301 | case LSHIFTRT: | |
2302 | case ASHIFT: | |
2303 | return shift_insns[shift_count]; | |
2304 | default: | |
f5b9e7c9 | 2305 | gcc_unreachable (); |
16bea517 JW |
2306 | } |
2307 | } | |
2308 | ||
16bea517 | 2309 | /* Return the cost of a shift. */ |
bc45ade3 | 2310 | |
3c50106f | 2311 | static inline int |
cf277499 | 2312 | shiftcosts (rtx x) |
bc45ade3 | 2313 | { |
3d422cb1 | 2314 | int value; |
8aa2a305 | 2315 | |
fa5322fa AO |
2316 | if (TARGET_SHMEDIA) |
2317 | return 1; | |
2318 | ||
c762ab6e R |
2319 | if (GET_MODE_SIZE (GET_MODE (x)) > UNITS_PER_WORD) |
2320 | { | |
2321 | if (GET_MODE (x) == DImode | |
2322 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
2323 | && INTVAL (XEXP (x, 1)) == 1) | |
2324 | return 2; | |
2325 | ||
2326 | /* Everything else is invalid, because there is no pattern for it. */ | |
cb2eb96f | 2327 | return MAX_COST; |
c762ab6e | 2328 | } |
16bea517 | 2329 | /* If shift by a non constant, then this will be expensive. */ |
8aa2a305 | 2330 | if (GET_CODE (XEXP (x, 1)) != CONST_INT) |
1245df60 | 2331 | return SH_DYNAMIC_SHIFT_COST; |
bc45ade3 | 2332 | |
3d422cb1 AO |
2333 | value = INTVAL (XEXP (x, 1)); |
2334 | ||
16bea517 | 2335 | /* Otherwise, return the true cost in instructions. */ |
8aa2a305 | 2336 | if (GET_CODE (x) == ASHIFTRT) |
49b6d06b JW |
2337 | { |
2338 | int cost = ashiftrt_insns[value]; | |
2339 | /* If SH3, then we put the constant in a reg and use shad. */ | |
1245df60 R |
2340 | if (cost > 1 + SH_DYNAMIC_SHIFT_COST) |
2341 | cost = 1 + SH_DYNAMIC_SHIFT_COST; | |
49b6d06b JW |
2342 | return cost; |
2343 | } | |
8aa2a305 JW |
2344 | else |
2345 | return shift_insns[value]; | |
0d7e008e | 2346 | } |
b9654711 | 2347 | |
8aa2a305 JW |
2348 | /* Return the cost of an AND operation. */ |
2349 | ||
3c50106f | 2350 | static inline int |
cf277499 | 2351 | andcosts (rtx x) |
0d7e008e SC |
2352 | { |
2353 | int i; | |
8aa2a305 | 2354 | |
a6f71af5 | 2355 | /* Anding with a register is a single cycle and instruction. */ |
8aa2a305 | 2356 | if (GET_CODE (XEXP (x, 1)) != CONST_INT) |
a6f71af5 | 2357 | return 1; |
8aa2a305 JW |
2358 | |
2359 | i = INTVAL (XEXP (x, 1)); | |
fa5322fa AO |
2360 | |
2361 | if (TARGET_SHMEDIA) | |
2362 | { | |
32a7ab3d KK |
2363 | if (satisfies_constraint_I10 (XEXP (x, 1)) |
2364 | || satisfies_constraint_J16 (XEXP (x, 1))) | |
fa5322fa AO |
2365 | return 1; |
2366 | else | |
42201282 | 2367 | return 1 + rtx_cost (XEXP (x, 1), AND); |
fa5322fa AO |
2368 | } |
2369 | ||
a6f71af5 | 2370 | /* These constants are single cycle extu.[bw] instructions. */ |
0d7e008e | 2371 | if (i == 0xff || i == 0xffff) |
a6f71af5 | 2372 | return 1; |
735cb76e | 2373 | /* Constants that can be used in an and immediate instruction in a single |
a6f71af5 | 2374 | cycle, but this requires r0, so make it a little more expensive. */ |
735cb76e | 2375 | if (CONST_OK_FOR_K08 (i)) |
0d7e008e | 2376 | return 2; |
a6f71af5 JW |
2377 | /* Constants that can be loaded with a mov immediate and an and. |
2378 | This case is probably unnecessary. */ | |
735cb76e | 2379 | if (CONST_OK_FOR_I08 (i)) |
a6f71af5 JW |
2380 | return 2; |
2381 | /* Any other constants requires a 2 cycle pc-relative load plus an and. | |
2382 | This case is probably unnecessary. */ | |
2383 | return 3; | |
0d7e008e | 2384 | } |
d3ae8277 | 2385 | |
630c79be BS |
2386 | /* Return the cost of an addition or a subtraction. */ |
2387 | ||
3c50106f | 2388 | static inline int |
cf277499 | 2389 | addsubcosts (rtx x) |
630c79be BS |
2390 | { |
2391 | /* Adding a register is a single cycle insn. */ | |
50ceefc2 R |
2392 | if (GET_CODE (XEXP (x, 1)) == REG |
2393 | || GET_CODE (XEXP (x, 1)) == SUBREG) | |
630c79be BS |
2394 | return 1; |
2395 | ||
2396 | /* Likewise for small constants. */ | |
3d422cb1 | 2397 | if (GET_CODE (XEXP (x, 1)) == CONST_INT |
fa5322fa | 2398 | && CONST_OK_FOR_ADD (INTVAL (XEXP (x, 1)))) |
630c79be BS |
2399 | return 1; |
2400 | ||
fa5322fa AO |
2401 | if (TARGET_SHMEDIA) |
2402 | switch (GET_CODE (XEXP (x, 1))) | |
2403 | { | |
2404 | case CONST: | |
2405 | case LABEL_REF: | |
2406 | case SYMBOL_REF: | |
2407 | return TARGET_SHMEDIA64 ? 5 : 3; | |
2408 | ||
2409 | case CONST_INT: | |
735cb76e | 2410 | if (CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1)))) |
fa5322fa | 2411 | return 2; |
735cb76e | 2412 | else if (CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1)) >> 16)) |
fa5322fa | 2413 | return 3; |
735cb76e | 2414 | else if (CONST_OK_FOR_I16 ((INTVAL (XEXP (x, 1)) >> 16) >> 16)) |
fa5322fa AO |
2415 | return 4; |
2416 | ||
2417 | /* Fall through. */ | |
2418 | default: | |
832a3292 | 2419 | return 5; |
fa5322fa AO |
2420 | } |
2421 | ||
630c79be BS |
2422 | /* Any other constant requires a 2 cycle pc-relative load plus an |
2423 | addition. */ | |
2424 | return 3; | |
2425 | } | |
2426 | ||
16bea517 | 2427 | /* Return the cost of a multiply. */ |
3c50106f | 2428 | static inline int |
cf277499 | 2429 | multcosts (rtx x ATTRIBUTE_UNUSED) |
0d7e008e | 2430 | { |
7dd2f19b RS |
2431 | if (sh_multcost >= 0) |
2432 | return sh_multcost; | |
fa5322fa | 2433 | if (TARGET_SHMEDIA) |
73a4d10b R |
2434 | /* ??? We have a mul insn, but it has a latency of three, and doesn't |
2435 | accept constants. Ideally, we would use a cost of one or two and | |
2436 | add the cost of the operand, but disregard the latter when inside loops | |
2437 | and loop invariant code motion is still to follow. | |
2438 | Using a multiply first and splitting it later if it's a loss | |
2439 | doesn't work because of different sign / zero extension semantics | |
2440 | of multiplies vs. shifts. */ | |
2441 | return TARGET_SMALLCODE ? 2 : 3; | |
fa5322fa | 2442 | |
0d7e008e | 2443 | if (TARGET_SH2) |
d3ae8277 SC |
2444 | { |
2445 | /* We have a mul insn, so we can never take more than the mul and the | |
a7771f78 | 2446 | read of the mac reg, but count more because of the latency and extra |
16bea517 | 2447 | reg usage. */ |
d3ae8277 | 2448 | if (TARGET_SMALLCODE) |
8e87e161 | 2449 | return 2; |
a7771f78 | 2450 | return 3; |
d3ae8277 SC |
2451 | } |
2452 | ||
a7771f78 | 2453 | /* If we're aiming at small code, then just count the number of |
16bea517 | 2454 | insns in a multiply call sequence. */ |
8e87e161 | 2455 | if (TARGET_SMALLCODE) |
8aa2a305 | 2456 | return 5; |
d3ae8277 | 2457 | |
16bea517 | 2458 | /* Otherwise count all the insns in the routine we'd be calling too. */ |
d3ae8277 | 2459 | return 20; |
0d7e008e | 2460 | } |
b9654711 | 2461 | |
3c50106f RH |
2462 | /* Compute a (partial) cost for rtx X. Return true if the complete |
2463 | cost has been computed, and false if subexpressions should be | |
2464 | scanned. In either case, *TOTAL contains the cost result. */ | |
2465 | ||
2466 | static bool | |
cf277499 | 2467 | sh_rtx_costs (rtx x, int code, int outer_code, int *total) |
3c50106f RH |
2468 | { |
2469 | switch (code) | |
2470 | { | |
2471 | case CONST_INT: | |
2472 | if (TARGET_SHMEDIA) | |
2473 | { | |
2474 | if (INTVAL (x) == 0) | |
2475 | *total = 0; | |
2476 | else if (outer_code == AND && and_operand ((x), DImode)) | |
2477 | *total = 0; | |
2478 | else if ((outer_code == IOR || outer_code == XOR | |
2479 | || outer_code == PLUS) | |
735cb76e | 2480 | && CONST_OK_FOR_I10 (INTVAL (x))) |
3c50106f | 2481 | *total = 0; |
735cb76e | 2482 | else if (CONST_OK_FOR_I16 (INTVAL (x))) |
3c50106f | 2483 | *total = COSTS_N_INSNS (outer_code != SET); |
735cb76e | 2484 | else if (CONST_OK_FOR_I16 (INTVAL (x) >> 16)) |
73a4d10b | 2485 | *total = COSTS_N_INSNS ((outer_code != SET) + 1); |
735cb76e | 2486 | else if (CONST_OK_FOR_I16 ((INTVAL (x) >> 16) >> 16)) |
42201282 | 2487 | *total = COSTS_N_INSNS ((outer_code != SET) + 2); |
3c50106f | 2488 | else |
42201282 | 2489 | *total = COSTS_N_INSNS ((outer_code != SET) + 3); |
3c50106f RH |
2490 | return true; |
2491 | } | |
735cb76e | 2492 | if (CONST_OK_FOR_I08 (INTVAL (x))) |
3c50106f RH |
2493 | *total = 0; |
2494 | else if ((outer_code == AND || outer_code == IOR || outer_code == XOR) | |
735cb76e | 2495 | && CONST_OK_FOR_K08 (INTVAL (x))) |
3c50106f | 2496 | *total = 1; |
78d310c2 | 2497 | /* prepare_cmp_insn will force costly constants int registers before |
35912544 | 2498 | the cbranch[sd]i4 patterns can see them, so preserve potentially |
78d310c2 R |
2499 | interesting ones not covered by I08 above. */ |
2500 | else if (outer_code == COMPARE | |
2501 | && ((unsigned HOST_WIDE_INT) INTVAL (x) | |
2502 | == (unsigned HOST_WIDE_INT) 0x7fffffff + 1 | |
2503 | || INTVAL (x) == 0x7fffffff | |
2504 | || INTVAL (x) == 0x80 || INTVAL (x) == -0x81)) | |
2505 | *total = 1; | |
3c50106f RH |
2506 | else |
2507 | *total = 8; | |
2508 | return true; | |
2509 | ||
2510 | case CONST: | |
2511 | case LABEL_REF: | |
2512 | case SYMBOL_REF: | |
2513 | if (TARGET_SHMEDIA64) | |
2514 | *total = COSTS_N_INSNS (4); | |
2515 | else if (TARGET_SHMEDIA32) | |
2516 | *total = COSTS_N_INSNS (2); | |
2517 | else | |
2518 | *total = 5; | |
2519 | return true; | |
2520 | ||
2521 | case CONST_DOUBLE: | |
2522 | if (TARGET_SHMEDIA) | |
2523 | *total = COSTS_N_INSNS (4); | |
78d310c2 | 2524 | /* prepare_cmp_insn will force costly constants int registers before |
35912544 | 2525 | the cbranchdi4 pattern can see them, so preserve potentially |
78d310c2 R |
2526 | interesting ones. */ |
2527 | else if (outer_code == COMPARE && GET_MODE (x) == DImode) | |
2528 | *total = 1; | |
3c50106f RH |
2529 | else |
2530 | *total = 10; | |
2531 | return true; | |
73a4d10b R |
2532 | case CONST_VECTOR: |
2533 | if (x == CONST0_RTX (GET_MODE (x))) | |
2534 | *total = 0; | |
2535 | else if (sh_1el_vec (x, VOIDmode)) | |
2536 | *total = outer_code != SET; | |
2537 | if (sh_rep_vec (x, VOIDmode)) | |
2538 | *total = ((GET_MODE_UNIT_SIZE (GET_MODE (x)) + 3) / 4 | |
2539 | + (outer_code != SET)); | |
2540 | *total = COSTS_N_INSNS (3) + (outer_code != SET); | |
2541 | return true; | |
3c50106f RH |
2542 | |
2543 | case PLUS: | |
73a4d10b | 2544 | case MINUS: |
3c50106f RH |
2545 | *total = COSTS_N_INSNS (addsubcosts (x)); |
2546 | return true; | |
2547 | ||
2548 | case AND: | |
2549 | *total = COSTS_N_INSNS (andcosts (x)); | |
2550 | return true; | |
2551 | ||
2552 | case MULT: | |
2553 | *total = COSTS_N_INSNS (multcosts (x)); | |
2554 | return true; | |
2555 | ||
2556 | case ASHIFT: | |
2557 | case ASHIFTRT: | |
2558 | case LSHIFTRT: | |
2559 | *total = COSTS_N_INSNS (shiftcosts (x)); | |
2560 | return true; | |
2561 | ||
2562 | case DIV: | |
2563 | case UDIV: | |
2564 | case MOD: | |
2565 | case UMOD: | |
2566 | *total = COSTS_N_INSNS (20); | |
2567 | return true; | |
2568 | ||
73a4d10b R |
2569 | case PARALLEL: |
2570 | if (sh_1el_vec (x, VOIDmode)) | |
2571 | *total = outer_code != SET; | |
2572 | if (sh_rep_vec (x, VOIDmode)) | |
2573 | *total = ((GET_MODE_UNIT_SIZE (GET_MODE (x)) + 3) / 4 | |
2574 | + (outer_code != SET)); | |
2575 | *total = COSTS_N_INSNS (3) + (outer_code != SET); | |
2576 | return true; | |
2577 | ||
3c50106f RH |
2578 | case FLOAT: |
2579 | case FIX: | |
2580 | *total = 100; | |
2581 | return true; | |
2582 | ||
2583 | default: | |
2584 | return false; | |
2585 | } | |
2586 | } | |
2587 | ||
dcefdf67 RH |
2588 | /* Compute the cost of an address. For the SH, all valid addresses are |
2589 | the same cost. Use a slightly higher cost for reg + reg addressing, | |
2590 | since it increases pressure on r0. */ | |
2591 | ||
2592 | static int | |
cf277499 | 2593 | sh_address_cost (rtx X) |
dcefdf67 RH |
2594 | { |
2595 | return (GET_CODE (X) == PLUS | |
2596 | && ! CONSTANT_P (XEXP (X, 1)) | |
2597 | && ! TARGET_SHMEDIA ? 1 : 0); | |
2598 | } | |
3c50106f | 2599 | |
16bea517 | 2600 | /* Code to expand a shift. */ |
b9654711 | 2601 | |
0d7e008e | 2602 | void |
cf277499 | 2603 | gen_ashift (int type, int n, rtx reg) |
0d7e008e | 2604 | { |
16bea517 JW |
2605 | /* Negative values here come from the shift_amounts array. */ |
2606 | if (n < 0) | |
2607 | { | |
2608 | if (type == ASHIFT) | |
2609 | type = LSHIFTRT; | |
2610 | else | |
2611 | type = ASHIFT; | |
2612 | n = -n; | |
2613 | } | |
2614 | ||
0d7e008e | 2615 | switch (type) |
bc45ade3 | 2616 | { |
0d7e008e SC |
2617 | case ASHIFTRT: |
2618 | emit_insn (gen_ashrsi3_k (reg, reg, GEN_INT (n))); | |
2619 | break; | |
2620 | case LSHIFTRT: | |
16bea517 JW |
2621 | if (n == 1) |
2622 | emit_insn (gen_lshrsi3_m (reg, reg, GEN_INT (n))); | |
2623 | else | |
2624 | emit_insn (gen_lshrsi3_k (reg, reg, GEN_INT (n))); | |
0d7e008e SC |
2625 | break; |
2626 | case ASHIFT: | |
7e2fda6e | 2627 | emit_insn (gen_ashlsi3_std (reg, reg, GEN_INT (n))); |
0d7e008e | 2628 | break; |
bc45ade3 | 2629 | } |
bc45ade3 | 2630 | } |
bc45ade3 | 2631 | |
8d481241 JW |
2632 | /* Same for HImode */ |
2633 | ||
2634 | void | |
cf277499 | 2635 | gen_ashift_hi (int type, int n, rtx reg) |
8d481241 JW |
2636 | { |
2637 | /* Negative values here come from the shift_amounts array. */ | |
2638 | if (n < 0) | |
2639 | { | |
2640 | if (type == ASHIFT) | |
2641 | type = LSHIFTRT; | |
2642 | else | |
2643 | type = ASHIFT; | |
2644 | n = -n; | |
2645 | } | |
2646 | ||
2647 | switch (type) | |
2648 | { | |
2649 | case ASHIFTRT: | |
8d481241 | 2650 | case LSHIFTRT: |
d0c42859 R |
2651 | /* We don't have HImode right shift operations because using the |
2652 | ordinary 32 bit shift instructions for that doesn't generate proper | |
2653 | zero/sign extension. | |
2654 | gen_ashift_hi is only called in contexts where we know that the | |
2655 | sign extension works out correctly. */ | |
97d6fd65 | 2656 | { |
ddef6bc7 | 2657 | int offset = 0; |
97d6fd65 R |
2658 | if (GET_CODE (reg) == SUBREG) |
2659 | { | |
ddef6bc7 | 2660 | offset = SUBREG_BYTE (reg); |
97d6fd65 R |
2661 | reg = SUBREG_REG (reg); |
2662 | } | |
ddef6bc7 | 2663 | gen_ashift (type, n, gen_rtx_SUBREG (SImode, reg, offset)); |
97d6fd65 R |
2664 | break; |
2665 | } | |
8d481241 JW |
2666 | case ASHIFT: |
2667 | emit_insn (gen_ashlhi3_k (reg, reg, GEN_INT (n))); | |
2668 | break; | |
2669 | } | |
2670 | } | |
2671 | ||
8aa2a305 JW |
2672 | /* Output RTL to split a constant shift into its component SH constant |
2673 | shift instructions. */ | |
f676971a | 2674 | |
318881c0 | 2675 | void |
cf277499 | 2676 | gen_shifty_op (int code, rtx *operands) |
bc45ade3 | 2677 | { |
16bea517 | 2678 | int value = INTVAL (operands[2]); |
8aa2a305 | 2679 | int max, i; |
00f8ff66 | 2680 | |
cff3d762 JW |
2681 | /* Truncate the shift count in case it is out of bounds. */ |
2682 | value = value & 0x1f; | |
f676971a | 2683 | |
8aa2a305 | 2684 | if (value == 31) |
16bea517 | 2685 | { |
8aa2a305 | 2686 | if (code == LSHIFTRT) |
0d7e008e | 2687 | { |
8aa2a305 JW |
2688 | emit_insn (gen_rotlsi3_1 (operands[0], operands[0])); |
2689 | emit_insn (gen_movt (operands[0])); | |
2690 | return; | |
16bea517 | 2691 | } |
8aa2a305 | 2692 | else if (code == ASHIFT) |
16bea517 | 2693 | { |
8aa2a305 JW |
2694 | /* There is a two instruction sequence for 31 bit left shifts, |
2695 | but it requires r0. */ | |
2696 | if (GET_CODE (operands[0]) == REG && REGNO (operands[0]) == 0) | |
0d7e008e | 2697 | { |
8aa2a305 JW |
2698 | emit_insn (gen_andsi3 (operands[0], operands[0], const1_rtx)); |
2699 | emit_insn (gen_rotlsi3_31 (operands[0], operands[0])); | |
2700 | return; | |
0d7e008e | 2701 | } |
16bea517 | 2702 | } |
8aa2a305 | 2703 | } |
c17f53a8 JW |
2704 | else if (value == 0) |
2705 | { | |
73a4d10b R |
2706 | /* This can happen even when optimizing, if there were subregs before |
2707 | reload. Don't output a nop here, as this is never optimized away; | |
2708 | use a no-op move instead. */ | |
2709 | emit_insn (gen_rtx_SET (VOIDmode, operands[0], operands[0])); | |
c17f53a8 JW |
2710 | return; |
2711 | } | |
a9f71ad8 | 2712 | |
8aa2a305 JW |
2713 | max = shift_insns[value]; |
2714 | for (i = 0; i < max; i++) | |
2715 | gen_ashift (code, shift_amounts[value][i], operands[0]); | |
2716 | } | |
f676971a | 2717 | |
8d481241 JW |
2718 | /* Same as above, but optimized for values where the topmost bits don't |
2719 | matter. */ | |
2720 | ||
318881c0 | 2721 | void |
cf277499 | 2722 | gen_shifty_hi_op (int code, rtx *operands) |
8d481241 JW |
2723 | { |
2724 | int value = INTVAL (operands[2]); | |
2725 | int max, i; | |
cf277499 | 2726 | void (*gen_fun) (int, int, rtx); |
8d481241 JW |
2727 | |
2728 | /* This operation is used by and_shl for SImode values with a few | |
2729 | high bits known to be cleared. */ | |
2730 | value &= 31; | |
2731 | if (value == 0) | |
2732 | { | |
2733 | emit_insn (gen_nop ()); | |
2734 | return; | |
2735 | } | |
2736 | ||
2737 | gen_fun = GET_MODE (operands[0]) == HImode ? gen_ashift_hi : gen_ashift; | |
2738 | if (code == ASHIFT) | |
2739 | { | |
2740 | max = ext_shift_insns[value]; | |
2741 | for (i = 0; i < max; i++) | |
2742 | gen_fun (code, ext_shift_amounts[value][i], operands[0]); | |
2743 | } | |
2744 | else | |
2745 | /* When shifting right, emit the shifts in reverse order, so that | |
2746 | solitary negative values come first. */ | |
2747 | for (i = ext_shift_insns[value] - 1; i >= 0; i--) | |
2748 | gen_fun (code, ext_shift_amounts[value][i], operands[0]); | |
2749 | } | |
8aa2a305 JW |
2750 | |
2751 | /* Output RTL for an arithmetic right shift. */ | |
2752 | ||
2753 | /* ??? Rewrite to use super-optimizer sequences. */ | |
2754 | ||
2755 | int | |
cf277499 | 2756 | expand_ashiftrt (rtx *operands) |
8aa2a305 JW |
2757 | { |
2758 | rtx wrk; | |
2759 | char func[18]; | |
8aa2a305 JW |
2760 | int value; |
2761 | ||
49b6d06b | 2762 | if (TARGET_SH3) |
20b04867 | 2763 | { |
49b6d06b JW |
2764 | if (GET_CODE (operands[2]) != CONST_INT) |
2765 | { | |
2766 | rtx count = copy_to_mode_reg (SImode, operands[2]); | |
2767 | emit_insn (gen_negsi2 (count, count)); | |
2768 | emit_insn (gen_ashrsi3_d (operands[0], operands[1], count)); | |
2769 | return 1; | |
2770 | } | |
1245df60 R |
2771 | else if (ashiftrt_insns[INTVAL (operands[2]) & 31] |
2772 | > 1 + SH_DYNAMIC_SHIFT_COST) | |
49b6d06b | 2773 | { |
1245df60 R |
2774 | rtx count |
2775 | = force_reg (SImode, GEN_INT (- (INTVAL (operands[2]) & 31))); | |
49b6d06b JW |
2776 | emit_insn (gen_ashrsi3_d (operands[0], operands[1], count)); |
2777 | return 1; | |
2778 | } | |
20b04867 | 2779 | } |
8aa2a305 JW |
2780 | if (GET_CODE (operands[2]) != CONST_INT) |
2781 | return 0; | |
2782 | ||
1245df60 | 2783 | value = INTVAL (operands[2]) & 31; |
8aa2a305 JW |
2784 | |
2785 | if (value == 31) | |
2786 | { | |
73a4d10b R |
2787 | /* If we are called from abs expansion, arrange things so that we |
2788 | we can use a single MT instruction that doesn't clobber the source, | |
2789 | if LICM can hoist out the load of the constant zero. */ | |
2790 | if (currently_expanding_to_rtl) | |
2791 | { | |
2792 | emit_insn (gen_cmpgtsi_t (force_reg (SImode, CONST0_RTX (SImode)), | |
2793 | operands[1])); | |
2794 | emit_insn (gen_mov_neg_si_t (operands[0])); | |
2795 | return 1; | |
2796 | } | |
8aa2a305 JW |
2797 | emit_insn (gen_ashrsi2_31 (operands[0], operands[1])); |
2798 | return 1; | |
2799 | } | |
2800 | else if (value >= 16 && value <= 19) | |
2801 | { | |
2802 | wrk = gen_reg_rtx (SImode); | |
2803 | emit_insn (gen_ashrsi2_16 (wrk, operands[1])); | |
2804 | value -= 16; | |
2805 | while (value--) | |
2806 | gen_ashift (ASHIFTRT, 1, wrk); | |
2807 | emit_move_insn (operands[0], wrk); | |
2808 | return 1; | |
a9f71ad8 | 2809 | } |
8aa2a305 JW |
2810 | /* Expand a short sequence inline, longer call a magic routine. */ |
2811 | else if (value <= 5) | |
2812 | { | |
2813 | wrk = gen_reg_rtx (SImode); | |
2814 | emit_move_insn (wrk, operands[1]); | |
2815 | while (value--) | |
2816 | gen_ashift (ASHIFTRT, 1, wrk); | |
2817 | emit_move_insn (operands[0], wrk); | |
2818 | return 1; | |
2819 | } | |
2820 | ||
2821 | wrk = gen_reg_rtx (Pmode); | |
2822 | ||
2823 | /* Load the value into an arg reg and call a helper. */ | |
c5c76735 | 2824 | emit_move_insn (gen_rtx_REG (SImode, 4), operands[1]); |
8aa2a305 | 2825 | sprintf (func, "__ashiftrt_r4_%d", value); |
73a4d10b | 2826 | function_symbol (wrk, func, SFUNC_STATIC); |
8aa2a305 | 2827 | emit_insn (gen_ashrsi3_n (GEN_INT (value), wrk)); |
c5c76735 | 2828 | emit_move_insn (operands[0], gen_rtx_REG (SImode, 4)); |
8aa2a305 | 2829 | return 1; |
bc45ade3 | 2830 | } |
8d481241 | 2831 | |
318881c0 | 2832 | int |
cf277499 | 2833 | sh_dynamicalize_shift_p (rtx count) |
1245df60 R |
2834 | { |
2835 | return shift_insns[INTVAL (count)] > 1 + SH_DYNAMIC_SHIFT_COST; | |
2836 | } | |
2837 | ||
8d481241 JW |
2838 | /* Try to find a good way to implement the combiner pattern |
2839 | [(set (match_operand:SI 0 "register_operand" "r") | |
2840 | (and:SI (ashift:SI (match_operand:SI 1 "register_operand" "r") | |
2841 | (match_operand:SI 2 "const_int_operand" "n")) | |
2842 | (match_operand:SI 3 "const_int_operand" "n"))) . | |
2843 | LEFT_RTX is operand 2 in the above pattern, and MASK_RTX is operand 3. | |
2844 | return 0 for simple right / left or left/right shift combination. | |
2845 | return 1 for a combination of shifts with zero_extend. | |
2846 | return 2 for a combination of shifts with an AND that needs r0. | |
2847 | return 3 for a combination of shifts with an AND that needs an extra | |
2848 | scratch register, when the three highmost bits of the AND mask are clear. | |
2849 | return 4 for a combination of shifts with an AND that needs an extra | |
2850 | scratch register, when any of the three highmost bits of the AND mask | |
2851 | is set. | |
2852 | If ATTRP is set, store an initial right shift width in ATTRP[0], | |
2853 | and the instruction length in ATTRP[1] . These values are not valid | |
2854 | when returning 0. | |
2855 | When ATTRP is set and returning 1, ATTRP[2] gets set to the index into | |
2856 | shift_amounts for the last shift value that is to be used before the | |
2857 | sign extend. */ | |
2858 | int | |
cf277499 | 2859 | shl_and_kind (rtx left_rtx, rtx mask_rtx, int *attrp) |
8d481241 JW |
2860 | { |
2861 | unsigned HOST_WIDE_INT mask, lsb, mask2, lsb2; | |
2862 | int left = INTVAL (left_rtx), right; | |
2863 | int best = 0; | |
2864 | int cost, best_cost = 10000; | |
2865 | int best_right = 0, best_len = 0; | |
2866 | int i; | |
2867 | int can_ext; | |
2868 | ||
2869 | if (left < 0 || left > 31) | |
2870 | return 0; | |
2871 | if (GET_CODE (mask_rtx) == CONST_INT) | |
2872 | mask = (unsigned HOST_WIDE_INT) INTVAL (mask_rtx) >> left; | |
2873 | else | |
2874 | mask = (unsigned HOST_WIDE_INT) GET_MODE_MASK (SImode) >> left; | |
832a3292 | 2875 | /* Can this be expressed as a right shift / left shift pair? */ |
8d481241 JW |
2876 | lsb = ((mask ^ (mask - 1)) >> 1) + 1; |
2877 | right = exact_log2 (lsb); | |
2878 | mask2 = ~(mask + lsb - 1); | |
2879 | lsb2 = ((mask2 ^ (mask2 - 1)) >> 1) + 1; | |
2880 | /* mask has no zeroes but trailing zeroes <==> ! mask2 */ | |
2881 | if (! mask2) | |
2882 | best_cost = shift_insns[right] + shift_insns[right + left]; | |
2883 | /* mask has no trailing zeroes <==> ! right */ | |
2884 | else if (! right && mask2 == ~(lsb2 - 1)) | |
2885 | { | |
2886 | int late_right = exact_log2 (lsb2); | |
2887 | best_cost = shift_insns[left + late_right] + shift_insns[late_right]; | |
2888 | } | |
832a3292 | 2889 | /* Try to use zero extend. */ |
8d481241 JW |
2890 | if (mask2 == ~(lsb2 - 1)) |
2891 | { | |
2892 | int width, first; | |
2893 | ||
2894 | for (width = 8; width <= 16; width += 8) | |
2895 | { | |
832a3292 KH |
2896 | /* Can we zero-extend right away? */ |
2897 | if (lsb2 == (unsigned HOST_WIDE_INT) 1 << width) | |
8d481241 JW |
2898 | { |
2899 | cost | |
2900 | = 1 + ext_shift_insns[right] + ext_shift_insns[left + right]; | |
2901 | if (cost < best_cost) | |
2902 | { | |
2903 | best = 1; | |
2904 | best_cost = cost; | |
2905 | best_right = right; | |
2906 | best_len = cost; | |
2907 | if (attrp) | |
2908 | attrp[2] = -1; | |
2909 | } | |
2910 | continue; | |
2911 | } | |
2912 | /* ??? Could try to put zero extend into initial right shift, | |
6f317ef3 | 2913 | or even shift a bit left before the right shift. */ |
8d481241 JW |
2914 | /* Determine value of first part of left shift, to get to the |
2915 | zero extend cut-off point. */ | |
2916 | first = width - exact_log2 (lsb2) + right; | |
2917 | if (first >= 0 && right + left - first >= 0) | |
2918 | { | |
2919 | cost = ext_shift_insns[right] + ext_shift_insns[first] + 1 | |
2920 | + ext_shift_insns[right + left - first]; | |
2921 | if (cost < best_cost) | |
2922 | { | |
2923 | best = 1; | |
2924 | best_cost = cost; | |
2925 | best_right = right; | |
2926 | best_len = cost; | |
2927 | if (attrp) | |
2928 | attrp[2] = first; | |
832a3292 | 2929 | } |
8d481241 JW |
2930 | } |
2931 | } | |
2932 | } | |
2933 | /* Try to use r0 AND pattern */ | |
2934 | for (i = 0; i <= 2; i++) | |
2935 | { | |
2936 | if (i > right) | |
2937 | break; | |
735cb76e | 2938 | if (! CONST_OK_FOR_K08 (mask >> i)) |
8d481241 JW |
2939 | continue; |
2940 | cost = (i != 0) + 2 + ext_shift_insns[left + i]; | |
2941 | if (cost < best_cost) | |
2942 | { | |
2943 | best = 2; | |
2944 | best_cost = cost; | |
2945 | best_right = i; | |
2946 | best_len = cost - 1; | |
2947 | } | |
2948 | } | |
2949 | /* Try to use a scratch register to hold the AND operand. */ | |
832a3292 | 2950 | can_ext = ((mask << left) & ((unsigned HOST_WIDE_INT) 3 << 30)) == 0; |
8d481241 JW |
2951 | for (i = 0; i <= 2; i++) |
2952 | { | |
2953 | if (i > right) | |
2954 | break; | |
735cb76e | 2955 | cost = (i != 0) + (CONST_OK_FOR_I08 (mask >> i) ? 2 : 3) |
6ab911bb | 2956 | + (can_ext ? ext_shift_insns : shift_insns)[left + i]; |
8d481241 JW |
2957 | if (cost < best_cost) |
2958 | { | |
2959 | best = 4 - can_ext; | |
2960 | best_cost = cost; | |
2961 | best_right = i; | |
735cb76e | 2962 | best_len = cost - 1 - ! CONST_OK_FOR_I08 (mask >> i); |
8d481241 JW |
2963 | } |
2964 | } | |
2965 | ||
2966 | if (attrp) | |
2967 | { | |
2968 | attrp[0] = best_right; | |
2969 | attrp[1] = best_len; | |
2970 | } | |
2971 | return best; | |
2972 | } | |
2973 | ||
2974 | /* This is used in length attributes of the unnamed instructions | |
2975 | corresponding to shl_and_kind return values of 1 and 2. */ | |
2976 | int | |
cf277499 | 2977 | shl_and_length (rtx insn) |
8d481241 JW |
2978 | { |
2979 | rtx set_src, left_rtx, mask_rtx; | |
2980 | int attributes[3]; | |
2981 | ||
2982 | set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
2983 | left_rtx = XEXP (XEXP (set_src, 0), 1); | |
2984 | mask_rtx = XEXP (set_src, 1); | |
2985 | shl_and_kind (left_rtx, mask_rtx, attributes); | |
2986 | return attributes[1]; | |
2987 | } | |
2988 | ||
2989 | /* This is used in length attribute of the and_shl_scratch instruction. */ | |
2990 | ||
2991 | int | |
cf277499 | 2992 | shl_and_scr_length (rtx insn) |
8d481241 JW |
2993 | { |
2994 | rtx set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
2995 | int len = shift_insns[INTVAL (XEXP (set_src, 1))]; | |
2996 | rtx op = XEXP (set_src, 0); | |
2997 | len += shift_insns[INTVAL (XEXP (op, 1))] + 1; | |
2998 | op = XEXP (XEXP (op, 0), 0); | |
2999 | return len + shift_insns[INTVAL (XEXP (op, 1))]; | |
3000 | } | |
3001 | ||
8d481241 JW |
3002 | /* Generate rtl for instructions for which shl_and_kind advised a particular |
3003 | method of generating them, i.e. returned zero. */ | |
3004 | ||
3005 | int | |
cf277499 | 3006 | gen_shl_and (rtx dest, rtx left_rtx, rtx mask_rtx, rtx source) |
8d481241 JW |
3007 | { |
3008 | int attributes[3]; | |
3009 | unsigned HOST_WIDE_INT mask; | |
3010 | int kind = shl_and_kind (left_rtx, mask_rtx, attributes); | |
3011 | int right, total_shift; | |
832a3292 | 3012 | void (*shift_gen_fun) (int, rtx *) = gen_shifty_hi_op; |
8d481241 JW |
3013 | |
3014 | right = attributes[0]; | |
3015 | total_shift = INTVAL (left_rtx) + right; | |
3016 | mask = (unsigned HOST_WIDE_INT) INTVAL (mask_rtx) >> total_shift; | |
3017 | switch (kind) | |
3018 | { | |
3019 | default: | |
3020 | return -1; | |
3021 | case 1: | |
3022 | { | |
3023 | int first = attributes[2]; | |
3024 | rtx operands[3]; | |
3025 | ||
3026 | if (first < 0) | |
3027 | { | |
7174c937 | 3028 | emit_insn ((mask << right) <= 0xff |
832a3292 KH |
3029 | ? gen_zero_extendqisi2 (dest, |
3030 | gen_lowpart (QImode, source)) | |
3031 | : gen_zero_extendhisi2 (dest, | |
3032 | gen_lowpart (HImode, source))); | |
8d481241 JW |
3033 | source = dest; |
3034 | } | |
3035 | if (source != dest) | |
3036 | emit_insn (gen_movsi (dest, source)); | |
3037 | operands[0] = dest; | |
3038 | if (right) | |
3039 | { | |
3040 | operands[2] = GEN_INT (right); | |
3041 | gen_shifty_hi_op (LSHIFTRT, operands); | |
3042 | } | |
3043 | if (first > 0) | |
3044 | { | |
3045 | operands[2] = GEN_INT (first); | |
3046 | gen_shifty_hi_op (ASHIFT, operands); | |
3047 | total_shift -= first; | |
3048 | mask <<= first; | |
3049 | } | |
3050 | if (first >= 0) | |
85af47b9 | 3051 | emit_insn (mask <= 0xff |
832a3292 KH |
3052 | ? gen_zero_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
3053 | : gen_zero_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
8d481241 JW |
3054 | if (total_shift > 0) |
3055 | { | |
3056 | operands[2] = GEN_INT (total_shift); | |
3057 | gen_shifty_hi_op (ASHIFT, operands); | |
3058 | } | |
3059 | break; | |
3060 | } | |
3061 | case 4: | |
3062 | shift_gen_fun = gen_shifty_op; | |
8d481241 | 3063 | case 3: |
24c50999 JR |
3064 | /* If the topmost bit that matters is set, set the topmost bits |
3065 | that don't matter. This way, we might be able to get a shorter | |
3066 | signed constant. */ | |
832a3292 KH |
3067 | if (mask & ((HOST_WIDE_INT) 1 << (31 - total_shift))) |
3068 | mask |= (HOST_WIDE_INT) ~0 << (31 - total_shift); | |
61fb6bac | 3069 | case 2: |
8d481241 JW |
3070 | /* Don't expand fine-grained when combining, because that will |
3071 | make the pattern fail. */ | |
4586b4ca | 3072 | if (currently_expanding_to_rtl |
8d481241 JW |
3073 | || reload_in_progress || reload_completed) |
3074 | { | |
3075 | rtx operands[3]; | |
f676971a | 3076 | |
61fb6bac R |
3077 | /* Cases 3 and 4 should be handled by this split |
3078 | only while combining */ | |
f5b9e7c9 | 3079 | gcc_assert (kind <= 2); |
8d481241 JW |
3080 | if (right) |
3081 | { | |
3082 | emit_insn (gen_lshrsi3 (dest, source, GEN_INT (right))); | |
3083 | source = dest; | |
3084 | } | |
3085 | emit_insn (gen_andsi3 (dest, source, GEN_INT (mask))); | |
afad3d2c JW |
3086 | if (total_shift) |
3087 | { | |
3088 | operands[0] = dest; | |
3089 | operands[1] = dest; | |
3090 | operands[2] = GEN_INT (total_shift); | |
3091 | shift_gen_fun (ASHIFT, operands); | |
3092 | } | |
8d481241 JW |
3093 | break; |
3094 | } | |
3095 | else | |
3096 | { | |
3097 | int neg = 0; | |
3098 | if (kind != 4 && total_shift < 16) | |
3099 | { | |
3100 | neg = -ext_shift_amounts[total_shift][1]; | |
3101 | if (neg > 0) | |
3102 | neg -= ext_shift_amounts[total_shift][2]; | |
3103 | else | |
3104 | neg = 0; | |
3105 | } | |
3106 | emit_insn (gen_and_shl_scratch (dest, source, | |
3107 | GEN_INT (right), | |
3108 | GEN_INT (mask), | |
3109 | GEN_INT (total_shift + neg), | |
3110 | GEN_INT (neg))); | |
3111 | emit_insn (gen_movsi (dest, dest)); | |
3112 | break; | |
3113 | } | |
3114 | } | |
3115 | return 0; | |
3116 | } | |
3117 | ||
3118 | /* Try to find a good way to implement the combiner pattern | |
3119 | [(set (match_operand:SI 0 "register_operand" "=r") | |
3120 | (sign_extract:SI (ashift:SI (match_operand:SI 1 "register_operand" "r") | |
3121 | (match_operand:SI 2 "const_int_operand" "n") | |
3122 | (match_operand:SI 3 "const_int_operand" "n") | |
3123 | (const_int 0))) | |
4773afa4 | 3124 | (clobber (reg:SI T_REG))] |
8d481241 JW |
3125 | LEFT_RTX is operand 2 in the above pattern, and SIZE_RTX is operand 3. |
3126 | return 0 for simple left / right shift combination. | |
3127 | return 1 for left shift / 8 bit sign extend / left shift. | |
3128 | return 2 for left shift / 16 bit sign extend / left shift. | |
3129 | return 3 for left shift / 8 bit sign extend / shift / sign extend. | |
3130 | return 4 for left shift / 16 bit sign extend / shift / sign extend. | |
3131 | return 5 for left shift / 16 bit sign extend / right shift | |
3132 | return 6 for < 8 bit sign extend / left shift. | |
3133 | return 7 for < 8 bit sign extend / left shift / single right shift. | |
3134 | If COSTP is nonzero, assign the calculated cost to *COSTP. */ | |
3135 | ||
3136 | int | |
cf277499 | 3137 | shl_sext_kind (rtx left_rtx, rtx size_rtx, int *costp) |
8d481241 JW |
3138 | { |
3139 | int left, size, insize, ext; | |
3a8699c7 | 3140 | int cost = 0, best_cost; |
8d481241 JW |
3141 | int kind; |
3142 | ||
3143 | left = INTVAL (left_rtx); | |
3144 | size = INTVAL (size_rtx); | |
3145 | insize = size - left; | |
f5b9e7c9 | 3146 | gcc_assert (insize > 0); |
8d481241 JW |
3147 | /* Default to left / right shift. */ |
3148 | kind = 0; | |
3149 | best_cost = shift_insns[32 - insize] + ashiftrt_insns[32 - size]; | |
3150 | if (size <= 16) | |
3151 | { | |
3152 | /* 16 bit shift / sign extend / 16 bit shift */ | |
3153 | cost = shift_insns[16 - insize] + 1 + ashiftrt_insns[16 - size]; | |
afad3d2c JW |
3154 | /* If ashiftrt_insns[16 - size] is 8, this choice will be overridden |
3155 | below, by alternative 3 or something even better. */ | |
8d481241 JW |
3156 | if (cost < best_cost) |
3157 | { | |
3158 | kind = 5; | |
3159 | best_cost = cost; | |
3160 | } | |
3161 | } | |
3162 | /* Try a plain sign extend between two shifts. */ | |
3163 | for (ext = 16; ext >= insize; ext -= 8) | |
3164 | { | |
3165 | if (ext <= size) | |
3166 | { | |
3167 | cost = ext_shift_insns[ext - insize] + 1 + shift_insns[size - ext]; | |
3168 | if (cost < best_cost) | |
3169 | { | |
11f9ed1a | 3170 | kind = ext / (unsigned) 8; |
8d481241 JW |
3171 | best_cost = cost; |
3172 | } | |
3173 | } | |
ae9d19c0 JR |
3174 | /* Check if we can do a sloppy shift with a final signed shift |
3175 | restoring the sign. */ | |
3176 | if (EXT_SHIFT_SIGNED (size - ext)) | |
3177 | cost = ext_shift_insns[ext - insize] + ext_shift_insns[size - ext] + 1; | |
3178 | /* If not, maybe it's still cheaper to do the second shift sloppy, | |
3179 | and do a final sign extend? */ | |
3180 | else if (size <= 16) | |
3181 | cost = ext_shift_insns[ext - insize] + 1 | |
3182 | + ext_shift_insns[size > ext ? size - ext : ext - size] + 1; | |
3183 | else | |
3184 | continue; | |
3185 | if (cost < best_cost) | |
8d481241 | 3186 | { |
11f9ed1a | 3187 | kind = ext / (unsigned) 8 + 2; |
ae9d19c0 | 3188 | best_cost = cost; |
8d481241 JW |
3189 | } |
3190 | } | |
3191 | /* Check if we can sign extend in r0 */ | |
3192 | if (insize < 8) | |
3193 | { | |
3194 | cost = 3 + shift_insns[left]; | |
3195 | if (cost < best_cost) | |
3196 | { | |
3197 | kind = 6; | |
3198 | best_cost = cost; | |
3199 | } | |
3200 | /* Try the same with a final signed shift. */ | |
3201 | if (left < 31) | |
3202 | { | |
3203 | cost = 3 + ext_shift_insns[left + 1] + 1; | |
3204 | if (cost < best_cost) | |
3205 | { | |
3206 | kind = 7; | |
3207 | best_cost = cost; | |
3208 | } | |
3209 | } | |
3210 | } | |
3211 | if (TARGET_SH3) | |
3212 | { | |
3213 | /* Try to use a dynamic shift. */ | |
1245df60 | 3214 | cost = shift_insns[32 - insize] + 1 + SH_DYNAMIC_SHIFT_COST; |
8d481241 JW |
3215 | if (cost < best_cost) |
3216 | { | |
3217 | kind = 0; | |
3218 | best_cost = cost; | |
3219 | } | |
3220 | } | |
3221 | if (costp) | |
3222 | *costp = cost; | |
3223 | return kind; | |
3224 | } | |
3225 | ||
3226 | /* Function to be used in the length attribute of the instructions | |
3227 | implementing this pattern. */ | |
3228 | ||
3229 | int | |
cf277499 | 3230 | shl_sext_length (rtx insn) |
8d481241 JW |
3231 | { |
3232 | rtx set_src, left_rtx, size_rtx; | |
3233 | int cost; | |
3234 | ||
3235 | set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
3236 | left_rtx = XEXP (XEXP (set_src, 0), 1); | |
3237 | size_rtx = XEXP (set_src, 1); | |
3238 | shl_sext_kind (left_rtx, size_rtx, &cost); | |
3239 | return cost; | |
3240 | } | |
3241 | ||
3242 | /* Generate rtl for this pattern */ | |
3243 | ||
3244 | int | |
cf277499 | 3245 | gen_shl_sext (rtx dest, rtx left_rtx, rtx size_rtx, rtx source) |
8d481241 JW |
3246 | { |
3247 | int kind; | |
d00d338c | 3248 | int left, size, insize, cost; |
8d481241 JW |
3249 | rtx operands[3]; |
3250 | ||
d00d338c | 3251 | kind = shl_sext_kind (left_rtx, size_rtx, &cost); |
8d481241 JW |
3252 | left = INTVAL (left_rtx); |
3253 | size = INTVAL (size_rtx); | |
3254 | insize = size - left; | |
3255 | switch (kind) | |
3256 | { | |
3257 | case 1: | |
3258 | case 2: | |
3259 | case 3: | |
3260 | case 4: | |
3261 | { | |
3262 | int ext = kind & 1 ? 8 : 16; | |
3263 | int shift2 = size - ext; | |
3264 | ||
3265 | /* Don't expand fine-grained when combining, because that will | |
3266 | make the pattern fail. */ | |
4586b4ca | 3267 | if (! currently_expanding_to_rtl |
8d481241 JW |
3268 | && ! reload_in_progress && ! reload_completed) |
3269 | { | |
3270 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
3271 | emit_insn (gen_movsi (dest, source)); | |
3272 | break; | |
3273 | } | |
3274 | if (dest != source) | |
3275 | emit_insn (gen_movsi (dest, source)); | |
3276 | operands[0] = dest; | |
afad3d2c JW |
3277 | if (ext - insize) |
3278 | { | |
3279 | operands[2] = GEN_INT (ext - insize); | |
3280 | gen_shifty_hi_op (ASHIFT, operands); | |
3281 | } | |
8d481241 | 3282 | emit_insn (kind & 1 |
832a3292 KH |
3283 | ? gen_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
3284 | : gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
8d481241 JW |
3285 | if (kind <= 2) |
3286 | { | |
afad3d2c JW |
3287 | if (shift2) |
3288 | { | |
3289 | operands[2] = GEN_INT (shift2); | |
3290 | gen_shifty_op (ASHIFT, operands); | |
3291 | } | |
8d481241 JW |
3292 | } |
3293 | else | |
3294 | { | |
afad3d2c | 3295 | if (shift2 > 0) |
8d481241 | 3296 | { |
ae9d19c0 JR |
3297 | if (EXT_SHIFT_SIGNED (shift2)) |
3298 | { | |
3299 | operands[2] = GEN_INT (shift2 + 1); | |
3300 | gen_shifty_op (ASHIFT, operands); | |
a556fd39 | 3301 | operands[2] = const1_rtx; |
ae9d19c0 JR |
3302 | gen_shifty_op (ASHIFTRT, operands); |
3303 | break; | |
3304 | } | |
8d481241 JW |
3305 | operands[2] = GEN_INT (shift2); |
3306 | gen_shifty_hi_op (ASHIFT, operands); | |
3307 | } | |
afad3d2c | 3308 | else if (shift2) |
8d481241 JW |
3309 | { |
3310 | operands[2] = GEN_INT (-shift2); | |
3311 | gen_shifty_hi_op (LSHIFTRT, operands); | |
3312 | } | |
3313 | emit_insn (size <= 8 | |
afad3d2c JW |
3314 | ? gen_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
3315 | : gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
8d481241 JW |
3316 | } |
3317 | break; | |
3318 | } | |
3319 | case 5: | |
afad3d2c JW |
3320 | { |
3321 | int i = 16 - size; | |
4586b4ca | 3322 | if (! currently_expanding_to_rtl |
913d8e13 R |
3323 | && ! reload_in_progress && ! reload_completed) |
3324 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
3325 | else | |
3326 | { | |
3327 | operands[0] = dest; | |
3328 | operands[2] = GEN_INT (16 - insize); | |
3329 | gen_shifty_hi_op (ASHIFT, operands); | |
3330 | emit_insn (gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
3331 | } | |
afad3d2c JW |
3332 | /* Don't use gen_ashrsi3 because it generates new pseudos. */ |
3333 | while (--i >= 0) | |
3334 | gen_ashift (ASHIFTRT, 1, dest); | |
3335 | break; | |
3336 | } | |
8d481241 JW |
3337 | case 6: |
3338 | case 7: | |
3339 | /* Don't expand fine-grained when combining, because that will | |
3340 | make the pattern fail. */ | |
4586b4ca | 3341 | if (! currently_expanding_to_rtl |
8d481241 JW |
3342 | && ! reload_in_progress && ! reload_completed) |
3343 | { | |
3344 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
3345 | emit_insn (gen_movsi (dest, source)); | |
3346 | break; | |
3347 | } | |
3348 | emit_insn (gen_andsi3 (dest, source, GEN_INT ((1 << insize) - 1))); | |
3349 | emit_insn (gen_xorsi3 (dest, dest, GEN_INT (1 << (insize - 1)))); | |
3350 | emit_insn (gen_addsi3 (dest, dest, GEN_INT (-1 << (insize - 1)))); | |
3351 | operands[0] = dest; | |
3352 | operands[2] = kind == 7 ? GEN_INT (left + 1) : left_rtx; | |
3353 | gen_shifty_op (ASHIFT, operands); | |
3354 | if (kind == 7) | |
a556fd39 | 3355 | emit_insn (gen_ashrsi3_k (dest, dest, const1_rtx)); |
8d481241 JW |
3356 | break; |
3357 | default: | |
3358 | return -1; | |
3359 | } | |
3360 | return 0; | |
3361 | } | |
fa5322fa AO |
3362 | |
3363 | /* Prefix a symbol_ref name with "datalabel". */ | |
3364 | ||
3365 | rtx | |
cf277499 | 3366 | gen_datalabel_ref (rtx sym) |
fa5322fa | 3367 | { |
73a4d10b R |
3368 | const char *str; |
3369 | ||
fa5322fa AO |
3370 | if (GET_CODE (sym) == LABEL_REF) |
3371 | return gen_rtx_CONST (GET_MODE (sym), | |
3372 | gen_rtx_UNSPEC (GET_MODE (sym), | |
3373 | gen_rtvec (1, sym), | |
3374 | UNSPEC_DATALABEL)); | |
f676971a | 3375 | |
f5b9e7c9 | 3376 | gcc_assert (GET_CODE (sym) == SYMBOL_REF); |
fa5322fa | 3377 | |
73a4d10b R |
3378 | str = XSTR (sym, 0); |
3379 | /* Share all SYMBOL_REF strings with the same value - that is important | |
3380 | for cse. */ | |
3381 | str = IDENTIFIER_POINTER (get_identifier (str)); | |
3382 | XSTR (sym, 0) = str; | |
3383 | ||
fa5322fa AO |
3384 | return sym; |
3385 | } | |
3386 | ||
8aa2a305 | 3387 | \f |
6dd3c0a5 SB |
3388 | static alloc_pool label_ref_list_pool; |
3389 | ||
3390 | typedef struct label_ref_list_d | |
3391 | { | |
3392 | rtx label; | |
3393 | struct label_ref_list_d *next; | |
3394 | } *label_ref_list_t; | |
3395 | ||
8aa2a305 JW |
3396 | /* The SH cannot load a large constant into a register, constants have to |
3397 | come from a pc relative load. The reference of a pc relative load | |
0fa2e4df | 3398 | instruction must be less than 1k in front of the instruction. This |
8aa2a305 JW |
3399 | means that we often have to dump a constant inside a function, and |
3400 | generate code to branch around it. | |
bc45ade3 | 3401 | |
8aa2a305 JW |
3402 | It is important to minimize this, since the branches will slow things |
3403 | down and make things bigger. | |
3404 | ||
3405 | Worst case code looks like: | |
3406 | ||
3407 | mov.l L1,rn | |
3408 | bra L2 | |
3409 | nop | |
3410 | align | |
3411 | L1: .long value | |
3412 | L2: | |
3413 | .. | |
3414 | ||
3415 | mov.l L3,rn | |
3416 | bra L4 | |
3417 | nop | |
3418 | align | |
3419 | L3: .long value | |
3420 | L4: | |
3421 | .. | |
3422 | ||
3423 | We fix this by performing a scan before scheduling, which notices which | |
3424 | instructions need to have their operands fetched from the constant table | |
3425 | and builds the table. | |
3426 | ||
3427 | The algorithm is: | |
3428 | ||
3429 | scan, find an instruction which needs a pcrel move. Look forward, find the | |
3430 | last barrier which is within MAX_COUNT bytes of the requirement. | |
3431 | If there isn't one, make one. Process all the instructions between | |
3432 | the find and the barrier. | |
3433 | ||
3434 | In the above example, we can tell that L3 is within 1k of L1, so | |
3435 | the first move can be shrunk from the 3 insn+constant sequence into | |
3436 | just 1 insn, and the constant moved to L3 to make: | |
3437 | ||
3438 | mov.l L1,rn | |
3439 | .. | |
3440 | mov.l L3,rn | |
3441 | bra L4 | |
3442 | nop | |
3443 | align | |
3444 | L3:.long value | |
3445 | L4:.long value | |
3446 | ||
3447 | Then the second move becomes the target for the shortening process. */ | |
3448 | ||
3449 | typedef struct | |
3450 | { | |
3451 | rtx value; /* Value in table. */ | |
3452 | rtx label; /* Label of value. */ | |
6dd3c0a5 | 3453 | label_ref_list_t wend; /* End of window. */ |
8aa2a305 | 3454 | enum machine_mode mode; /* Mode of value. */ |
3503150c RS |
3455 | |
3456 | /* True if this constant is accessed as part of a post-increment | |
3457 | sequence. Note that HImode constants are never accessed in this way. */ | |
3458 | bool part_of_sequence_p; | |
8aa2a305 JW |
3459 | } pool_node; |
3460 | ||
3461 | /* The maximum number of constants that can fit into one pool, since | |
73a4d10b R |
3462 | constants in the range 0..510 are at least 2 bytes long, and in the |
3463 | range from there to 1018 at least 4 bytes. */ | |
8aa2a305 | 3464 | |
73a4d10b | 3465 | #define MAX_POOL_SIZE 372 |
8aa2a305 JW |
3466 | static pool_node pool_vector[MAX_POOL_SIZE]; |
3467 | static int pool_size; | |
b91455de KK |
3468 | static rtx pool_window_label; |
3469 | static int pool_window_last; | |
8aa2a305 | 3470 | |
4241ecb0 R |
3471 | static int max_labelno_before_reorg; |
3472 | ||
8aa2a305 JW |
3473 | /* ??? If we need a constant in HImode which is the truncated value of a |
3474 | constant we need in SImode, we could combine the two entries thus saving | |
3475 | two bytes. Is this common enough to be worth the effort of implementing | |
3476 | it? */ | |
3477 | ||
3478 | /* ??? This stuff should be done at the same time that we shorten branches. | |
3479 | As it is now, we must assume that all branches are the maximum size, and | |
3480 | this causes us to almost always output constant pools sooner than | |
3481 | necessary. */ | |
3482 | ||
3483 | /* Add a constant to the pool and return its label. */ | |
3484 | ||
3485 | static rtx | |
cf277499 | 3486 | add_constant (rtx x, enum machine_mode mode, rtx last_value) |
0d7e008e SC |
3487 | { |
3488 | int i; | |
6dd3c0a5 SB |
3489 | rtx lab, new; |
3490 | label_ref_list_t ref, newref; | |
8aa2a305 JW |
3491 | |
3492 | /* First see if we've already got it. */ | |
3493 | for (i = 0; i < pool_size; i++) | |
0d7e008e | 3494 | { |
8aa2a305 JW |
3495 | if (x->code == pool_vector[i].value->code |
3496 | && mode == pool_vector[i].mode) | |
0d7e008e | 3497 | { |
8aa2a305 JW |
3498 | if (x->code == CODE_LABEL) |
3499 | { | |
3500 | if (XINT (x, 3) != XINT (pool_vector[i].value, 3)) | |
3501 | continue; | |
3502 | } | |
3503 | if (rtx_equal_p (x, pool_vector[i].value)) | |
225e4f43 | 3504 | { |
b91455de | 3505 | lab = new = 0; |
225e4f43 R |
3506 | if (! last_value |
3507 | || ! i | |
3508 | || ! rtx_equal_p (last_value, pool_vector[i-1].value)) | |
3509 | { | |
b91455de KK |
3510 | new = gen_label_rtx (); |
3511 | LABEL_REFS (new) = pool_vector[i].label; | |
3512 | pool_vector[i].label = lab = new; | |
225e4f43 | 3513 | } |
b91455de KK |
3514 | if (lab && pool_window_label) |
3515 | { | |
6dd3c0a5 SB |
3516 | newref = (label_ref_list_t) pool_alloc (label_ref_list_pool); |
3517 | newref->label = pool_window_label; | |
b91455de | 3518 | ref = pool_vector[pool_window_last].wend; |
6dd3c0a5 | 3519 | newref->next = ref; |
b91455de KK |
3520 | pool_vector[pool_window_last].wend = newref; |
3521 | } | |
3522 | if (new) | |
3523 | pool_window_label = new; | |
3524 | pool_window_last = i; | |
225e4f43 R |
3525 | return lab; |
3526 | } | |
0d7e008e | 3527 | } |
0d7e008e | 3528 | } |
b9654711 | 3529 | |
8aa2a305 JW |
3530 | /* Need a new one. */ |
3531 | pool_vector[pool_size].value = x; | |
225e4f43 | 3532 | if (last_value && rtx_equal_p (last_value, pool_vector[pool_size - 1].value)) |
3503150c RS |
3533 | { |
3534 | lab = 0; | |
3535 | pool_vector[pool_size - 1].part_of_sequence_p = true; | |
3536 | } | |
225e4f43 R |
3537 | else |
3538 | lab = gen_label_rtx (); | |
8aa2a305 JW |
3539 | pool_vector[pool_size].mode = mode; |
3540 | pool_vector[pool_size].label = lab; | |
6dd3c0a5 | 3541 | pool_vector[pool_size].wend = NULL; |
3503150c | 3542 | pool_vector[pool_size].part_of_sequence_p = (lab == 0); |
b91455de KK |
3543 | if (lab && pool_window_label) |
3544 | { | |
6dd3c0a5 SB |
3545 | newref = (label_ref_list_t) pool_alloc (label_ref_list_pool); |
3546 | newref->label = pool_window_label; | |
b91455de | 3547 | ref = pool_vector[pool_window_last].wend; |
6dd3c0a5 | 3548 | newref->next = ref; |
b91455de KK |
3549 | pool_vector[pool_window_last].wend = newref; |
3550 | } | |
3551 | if (lab) | |
3552 | pool_window_label = lab; | |
3553 | pool_window_last = pool_size; | |
8aa2a305 JW |
3554 | pool_size++; |
3555 | return lab; | |
0d7e008e | 3556 | } |
16bea517 | 3557 | |
078c8b08 R |
3558 | /* Output the literal table. START, if nonzero, is the first instruction |
3559 | this table is needed for, and also indicates that there is at least one | |
3560 | casesi_worker_2 instruction; We have to emit the operand3 labels from | |
3561 | these insns at a 4-byte aligned position. BARRIER is the barrier | |
3562 | after which we are to place the table. */ | |
b9654711 | 3563 | |
b9654711 | 3564 | static void |
078c8b08 | 3565 | dump_table (rtx start, rtx barrier) |
b9654711 | 3566 | { |
078c8b08 | 3567 | rtx scan = barrier; |
0d7e008e | 3568 | int i; |
0d7e008e | 3569 | int need_align = 1; |
6dd3c0a5 SB |
3570 | rtx lab; |
3571 | label_ref_list_t ref; | |
aa06e8f5 | 3572 | int have_df = 0; |
b9654711 | 3573 | |
16bea517 | 3574 | /* Do two passes, first time dump out the HI sized constants. */ |
b9654711 | 3575 | |
0d7e008e | 3576 | for (i = 0; i < pool_size; i++) |
b9654711 | 3577 | { |
8aa2a305 JW |
3578 | pool_node *p = &pool_vector[i]; |
3579 | ||
0d7e008e SC |
3580 | if (p->mode == HImode) |
3581 | { | |
3582 | if (need_align) | |
3583 | { | |
3584 | scan = emit_insn_after (gen_align_2 (), scan); | |
3585 | need_align = 0; | |
3586 | } | |
b91455de KK |
3587 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
3588 | scan = emit_label_after (lab, scan); | |
3589 | scan = emit_insn_after (gen_consttable_2 (p->value, const0_rtx), | |
3590 | scan); | |
6dd3c0a5 | 3591 | for (ref = p->wend; ref; ref = ref->next) |
b91455de | 3592 | { |
6dd3c0a5 | 3593 | lab = ref->label; |
b91455de KK |
3594 | scan = emit_insn_after (gen_consttable_window_end (lab), scan); |
3595 | } | |
0d7e008e | 3596 | } |
aa06e8f5 R |
3597 | else if (p->mode == DFmode) |
3598 | have_df = 1; | |
b9654711 | 3599 | } |
8aa2a305 | 3600 | |
0d7e008e | 3601 | need_align = 1; |
b9654711 | 3602 | |
078c8b08 R |
3603 | if (start) |
3604 | { | |
3605 | scan = emit_insn_after (gen_align_4 (), scan); | |
3606 | need_align = 0; | |
3607 | for (; start != barrier; start = NEXT_INSN (start)) | |
3608 | if (GET_CODE (start) == INSN | |
3609 | && recog_memoized (start) == CODE_FOR_casesi_worker_2) | |
3610 | { | |
3611 | rtx src = SET_SRC (XVECEXP (PATTERN (start), 0, 0)); | |
3612 | rtx lab = XEXP (XVECEXP (src, 0, 3), 0); | |
3613 | ||
3614 | scan = emit_label_after (lab, scan); | |
3615 | } | |
3616 | } | |
aa06e8f5 | 3617 | if (TARGET_FMOVD && TARGET_ALIGN_DOUBLE && have_df) |
fa5322fa AO |
3618 | { |
3619 | rtx align_insn = NULL_RTX; | |
3620 | ||
3621 | scan = emit_label_after (gen_label_rtx (), scan); | |
3622 | scan = emit_insn_after (gen_align_log (GEN_INT (3)), scan); | |
3623 | need_align = 0; | |
3624 | ||
3625 | for (i = 0; i < pool_size; i++) | |
3626 | { | |
3627 | pool_node *p = &pool_vector[i]; | |
3628 | ||
3629 | switch (p->mode) | |
3630 | { | |
3631 | case HImode: | |
3632 | break; | |
3633 | case SImode: | |
3634 | case SFmode: | |
3503150c | 3635 | if (align_insn && !p->part_of_sequence_p) |
fa5322fa AO |
3636 | { |
3637 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) | |
3638 | emit_label_before (lab, align_insn); | |
3639 | emit_insn_before (gen_consttable_4 (p->value, const0_rtx), | |
3640 | align_insn); | |
6dd3c0a5 | 3641 | for (ref = p->wend; ref; ref = ref->next) |
fa5322fa | 3642 | { |
6dd3c0a5 | 3643 | lab = ref->label; |
fa5322fa | 3644 | emit_insn_before (gen_consttable_window_end (lab), |
832a3292 | 3645 | align_insn); |
fa5322fa AO |
3646 | } |
3647 | delete_insn (align_insn); | |
3648 | align_insn = NULL_RTX; | |
3649 | continue; | |
3650 | } | |
3651 | else | |
3652 | { | |
3653 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) | |
3654 | scan = emit_label_after (lab, scan); | |
3655 | scan = emit_insn_after (gen_consttable_4 (p->value, | |
3656 | const0_rtx), scan); | |
3657 | need_align = ! need_align; | |
3658 | } | |
3659 | break; | |
3660 | case DFmode: | |
fa5322fa AO |
3661 | if (need_align) |
3662 | { | |
3663 | scan = emit_insn_after (gen_align_log (GEN_INT (3)), scan); | |
3664 | align_insn = scan; | |
3665 | need_align = 0; | |
3666 | } | |
aa06e8f5 | 3667 | case DImode: |
fa5322fa AO |
3668 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
3669 | scan = emit_label_after (lab, scan); | |
3670 | scan = emit_insn_after (gen_consttable_8 (p->value, const0_rtx), | |
3671 | scan); | |
3672 | break; | |
3673 | default: | |
f5b9e7c9 | 3674 | gcc_unreachable (); |
fa5322fa AO |
3675 | } |
3676 | ||
3677 | if (p->mode != HImode) | |
3678 | { | |
6dd3c0a5 | 3679 | for (ref = p->wend; ref; ref = ref->next) |
fa5322fa | 3680 | { |
6dd3c0a5 | 3681 | lab = ref->label; |
fa5322fa AO |
3682 | scan = emit_insn_after (gen_consttable_window_end (lab), |
3683 | scan); | |
3684 | } | |
3685 | } | |
3686 | } | |
3687 | ||
3688 | pool_size = 0; | |
3689 | } | |
f676971a | 3690 | |
0d7e008e | 3691 | for (i = 0; i < pool_size; i++) |
b9654711 | 3692 | { |
8aa2a305 | 3693 | pool_node *p = &pool_vector[i]; |
b9654711 | 3694 | |
0d7e008e | 3695 | switch (p->mode) |
b9654711 | 3696 | { |
0d7e008e SC |
3697 | case HImode: |
3698 | break; | |
3699 | case SImode: | |
e577c183 | 3700 | case SFmode: |
0d7e008e | 3701 | if (need_align) |
b9654711 | 3702 | { |
0d7e008e | 3703 | need_align = 0; |
d3ae8277 | 3704 | scan = emit_label_after (gen_label_rtx (), scan); |
0d7e008e | 3705 | scan = emit_insn_after (gen_align_4 (), scan); |
b9654711 | 3706 | } |
b91455de KK |
3707 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
3708 | scan = emit_label_after (lab, scan); | |
3709 | scan = emit_insn_after (gen_consttable_4 (p->value, const0_rtx), | |
3710 | scan); | |
0d7e008e | 3711 | break; |
e577c183 | 3712 | case DFmode: |
0d7e008e SC |
3713 | case DImode: |
3714 | if (need_align) | |
3715 | { | |
3716 | need_align = 0; | |
d3ae8277 | 3717 | scan = emit_label_after (gen_label_rtx (), scan); |
0d7e008e SC |
3718 | scan = emit_insn_after (gen_align_4 (), scan); |
3719 | } | |
b91455de KK |
3720 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
3721 | scan = emit_label_after (lab, scan); | |
3722 | scan = emit_insn_after (gen_consttable_8 (p->value, const0_rtx), | |
3723 | scan); | |
0d7e008e SC |
3724 | break; |
3725 | default: | |
f5b9e7c9 | 3726 | gcc_unreachable (); |
b9654711 | 3727 | } |
b91455de KK |
3728 | |
3729 | if (p->mode != HImode) | |
3730 | { | |
6dd3c0a5 | 3731 | for (ref = p->wend; ref; ref = ref->next) |
b91455de | 3732 | { |
6dd3c0a5 | 3733 | lab = ref->label; |
b91455de KK |
3734 | scan = emit_insn_after (gen_consttable_window_end (lab), scan); |
3735 | } | |
3736 | } | |
b9654711 | 3737 | } |
b9654711 | 3738 | |
0d7e008e SC |
3739 | scan = emit_insn_after (gen_consttable_end (), scan); |
3740 | scan = emit_barrier_after (scan); | |
3741 | pool_size = 0; | |
b91455de KK |
3742 | pool_window_label = NULL_RTX; |
3743 | pool_window_last = 0; | |
0d7e008e | 3744 | } |
b9654711 | 3745 | |
5e7a8ee0 | 3746 | /* Return nonzero if constant would be an ok source for a |
8aa2a305 | 3747 | mov.w instead of a mov.l. */ |
b9654711 | 3748 | |
16bea517 | 3749 | static int |
cf277499 | 3750 | hi_const (rtx src) |
0d7e008e | 3751 | { |
8aa2a305 JW |
3752 | return (GET_CODE (src) == CONST_INT |
3753 | && INTVAL (src) >= -32768 | |
3754 | && INTVAL (src) <= 32767); | |
b9654711 SC |
3755 | } |
3756 | ||
4241ecb0 R |
3757 | #define MOVA_LABELREF(mova) XVECEXP (SET_SRC (PATTERN (mova)), 0, 0) |
3758 | ||
5e7a8ee0 | 3759 | /* Nonzero if the insn is a move instruction which needs to be fixed. */ |
8aa2a305 JW |
3760 | |
3761 | /* ??? For a DImode/DFmode moves, we don't need to fix it if each half of the | |
735cb76e R |
3762 | CONST_DOUBLE input value is CONST_OK_FOR_I08. For a SFmode move, we don't |
3763 | need to fix it if the input value is CONST_OK_FOR_I08. */ | |
8aa2a305 JW |
3764 | |
3765 | static int | |
cf277499 | 3766 | broken_move (rtx insn) |
b9654711 | 3767 | { |
e577c183 JR |
3768 | if (GET_CODE (insn) == INSN) |
3769 | { | |
3770 | rtx pat = PATTERN (insn); | |
3771 | if (GET_CODE (pat) == PARALLEL) | |
3772 | pat = XVECEXP (pat, 0, 0); | |
3773 | if (GET_CODE (pat) == SET | |
3b962ea0 | 3774 | /* We can load any 8-bit value if we don't care what the high |
e577c183 JR |
3775 | order bits end up as. */ |
3776 | && GET_MODE (SET_DEST (pat)) != QImode | |
43c05634 AO |
3777 | && (CONSTANT_P (SET_SRC (pat)) |
3778 | /* Match mova_const. */ | |
3779 | || (GET_CODE (SET_SRC (pat)) == UNSPEC | |
3780 | && XINT (SET_SRC (pat), 1) == UNSPEC_MOVA | |
3781 | && GET_CODE (XVECEXP (SET_SRC (pat), 0, 0)) == CONST)) | |
3a8699c7 | 3782 | && ! (TARGET_SH2E |
1245df60 | 3783 | && GET_CODE (SET_SRC (pat)) == CONST_DOUBLE |
e577c183 JR |
3784 | && (fp_zero_operand (SET_SRC (pat)) |
3785 | || fp_one_operand (SET_SRC (pat))) | |
18778292 R |
3786 | /* ??? If this is a -m4 or -m4-single compilation, in general |
3787 | we don't know the current setting of fpscr, so disable fldi. | |
3788 | There is an exception if this was a register-register move | |
3789 | before reload - and hence it was ascertained that we have | |
3790 | single precision setting - and in a post-reload optimization | |
3791 | we changed this to do a constant load. In that case | |
3792 | we don't have an r0 clobber, hence we must use fldi. */ | |
3793 | && (! TARGET_SH4 || TARGET_FMOVD | |
3794 | || (GET_CODE (XEXP (XVECEXP (PATTERN (insn), 0, 2), 0)) | |
3795 | == SCRATCH)) | |
e577c183 | 3796 | && GET_CODE (SET_DEST (pat)) == REG |
104ee20b | 3797 | && FP_REGISTER_P (REGNO (SET_DEST (pat)))) |
157371cf AO |
3798 | && ! (TARGET_SH2A |
3799 | && GET_MODE (SET_DEST (pat)) == SImode | |
e990551a JS |
3800 | && (satisfies_constraint_I20 (SET_SRC (pat)) |
3801 | || satisfies_constraint_I28 (SET_SRC (pat)))) | |
32a7ab3d | 3802 | && ! satisfies_constraint_I08 (SET_SRC (pat))) |
e577c183 JR |
3803 | return 1; |
3804 | } | |
d3ae8277 | 3805 | |
8aa2a305 | 3806 | return 0; |
b9654711 | 3807 | } |
b9654711 | 3808 | |
1245df60 | 3809 | static int |
cf277499 | 3810 | mova_p (rtx insn) |
1245df60 R |
3811 | { |
3812 | return (GET_CODE (insn) == INSN | |
3813 | && GET_CODE (PATTERN (insn)) == SET | |
3814 | && GET_CODE (SET_SRC (PATTERN (insn))) == UNSPEC | |
43c05634 AO |
3815 | && XINT (SET_SRC (PATTERN (insn)), 1) == UNSPEC_MOVA |
3816 | /* Don't match mova_const. */ | |
4241ecb0 | 3817 | && GET_CODE (MOVA_LABELREF (insn)) == LABEL_REF); |
1245df60 R |
3818 | } |
3819 | ||
078c8b08 R |
3820 | /* Fix up a mova from a switch that went out of range. */ |
3821 | static void | |
3822 | fixup_mova (rtx mova) | |
3823 | { | |
4241ecb0 | 3824 | PUT_MODE (XEXP (MOVA_LABELREF (mova), 0), QImode); |
078c8b08 R |
3825 | if (! flag_pic) |
3826 | { | |
4241ecb0 | 3827 | SET_SRC (PATTERN (mova)) = MOVA_LABELREF (mova); |
078c8b08 R |
3828 | INSN_CODE (mova) = -1; |
3829 | } | |
3830 | else | |
3831 | { | |
3832 | rtx worker = mova; | |
3833 | rtx lab = gen_label_rtx (); | |
3834 | rtx wpat, wpat0, wpat1, wsrc, diff; | |
3835 | ||
3836 | do | |
3837 | { | |
3838 | worker = NEXT_INSN (worker); | |
f5b9e7c9 NS |
3839 | gcc_assert (worker |
3840 | && GET_CODE (worker) != CODE_LABEL | |
3841 | && GET_CODE (worker) != JUMP_INSN); | |
0bf72672 KK |
3842 | } while (GET_CODE (worker) == NOTE |
3843 | || recog_memoized (worker) != CODE_FOR_casesi_worker_1); | |
078c8b08 R |
3844 | wpat = PATTERN (worker); |
3845 | wpat0 = XVECEXP (wpat, 0, 0); | |
3846 | wpat1 = XVECEXP (wpat, 0, 1); | |
3847 | wsrc = SET_SRC (wpat0); | |
3848 | PATTERN (worker) = (gen_casesi_worker_2 | |
3849 | (SET_DEST (wpat0), XVECEXP (wsrc, 0, 1), | |
3850 | XEXP (XVECEXP (wsrc, 0, 2), 0), lab, | |
3851 | XEXP (wpat1, 0))); | |
3852 | INSN_CODE (worker) = -1; | |
3853 | diff = gen_rtx_MINUS (Pmode, XVECEXP (SET_SRC (PATTERN (mova)), 0, 0), | |
3854 | gen_rtx_LABEL_REF (Pmode, lab)); | |
3855 | diff = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, diff), UNSPEC_PIC); | |
3856 | SET_SRC (PATTERN (mova)) = gen_rtx_CONST (Pmode, diff); | |
3857 | INSN_CODE (mova) = -1; | |
3858 | } | |
3859 | } | |
3860 | ||
4241ecb0 R |
3861 | /* NEW_MOVA is a mova we've just encountered while scanning forward. Update |
3862 | *num_mova, and check if the new mova is not nested within the first one. | |
3863 | return 0 if *first_mova was replaced, 1 if new_mova was replaced, | |
3864 | 2 if new_mova has been assigned to *first_mova, -1 otherwise.. */ | |
3865 | static int | |
3866 | untangle_mova (int *num_mova, rtx *first_mova, rtx new_mova) | |
3867 | { | |
c6d56443 KK |
3868 | int n_addr = 0; /* Initialization to shut up spurious warning. */ |
3869 | int f_target, n_target = 0; /* Likewise. */ | |
4241ecb0 R |
3870 | |
3871 | if (optimize) | |
3872 | { | |
3873 | n_addr = INSN_ADDRESSES (INSN_UID (new_mova)); | |
3874 | n_target = INSN_ADDRESSES (INSN_UID (XEXP (MOVA_LABELREF (new_mova), 0))); | |
3875 | if (n_addr > n_target || n_addr + 1022 < n_target) | |
3876 | { | |
3877 | /* Change the mova into a load. | |
3878 | broken_move will then return true for it. */ | |
3879 | fixup_mova (new_mova); | |
3880 | return 1; | |
3881 | } | |
3882 | } | |
3883 | if (!(*num_mova)++) | |
3884 | { | |
3885 | *first_mova = new_mova; | |
3886 | return 2; | |
3887 | } | |
3888 | if (!optimize | |
3889 | || ((f_target | |
3890 | = INSN_ADDRESSES (INSN_UID (XEXP (MOVA_LABELREF (*first_mova), 0)))) | |
3891 | >= n_target)) | |
3892 | return -1; | |
3893 | ||
3894 | (*num_mova)--; | |
3895 | if (f_target - INSN_ADDRESSES (INSN_UID (*first_mova)) | |
3896 | > n_target - n_addr) | |
3897 | { | |
3898 | fixup_mova (*first_mova); | |
3899 | return 0; | |
3900 | } | |
3901 | else | |
3902 | { | |
3903 | fixup_mova (new_mova); | |
3904 | return 1; | |
3905 | } | |
3906 | } | |
3907 | ||
8aa2a305 JW |
3908 | /* Find the last barrier from insn FROM which is close enough to hold the |
3909 | constant pool. If we can't find one, then create one near the end of | |
3910 | the range. */ | |
b9654711 | 3911 | |
8aa2a305 | 3912 | static rtx |
cf277499 | 3913 | find_barrier (int num_mova, rtx mova, rtx from) |
b9654711 | 3914 | { |
0d7e008e SC |
3915 | int count_si = 0; |
3916 | int count_hi = 0; | |
3917 | int found_hi = 0; | |
3918 | int found_si = 0; | |
fa5322fa | 3919 | int found_di = 0; |
33f7f353 JR |
3920 | int hi_align = 2; |
3921 | int si_align = 2; | |
1245df60 | 3922 | int leading_mova = num_mova; |
3a8699c7 | 3923 | rtx barrier_before_mova = 0, found_barrier = 0, good_barrier = 0; |
c17f53a8 JW |
3924 | int si_limit; |
3925 | int hi_limit; | |
7a3cb3ab | 3926 | rtx orig = from; |
8aa2a305 JW |
3927 | |
3928 | /* For HImode: range is 510, add 4 because pc counts from address of | |
3929 | second instruction after this one, subtract 2 for the jump instruction | |
3adbde60 JW |
3930 | that we may need to emit before the table, subtract 2 for the instruction |
3931 | that fills the jump delay slot (in very rare cases, reorg will take an | |
3932 | instruction from after the constant pool or will leave the delay slot | |
3933 | empty). This gives 510. | |
8aa2a305 JW |
3934 | For SImode: range is 1020, add 4 because pc counts from address of |
3935 | second instruction after this one, subtract 2 in case pc is 2 byte | |
3936 | aligned, subtract 2 for the jump instruction that we may need to emit | |
3adbde60 JW |
3937 | before the table, subtract 2 for the instruction that fills the jump |
3938 | delay slot. This gives 1018. */ | |
c17f53a8 | 3939 | |
1245df60 | 3940 | /* The branch will always be shortened now that the reference address for |
956d6950 | 3941 | forward branches is the successor address, thus we need no longer make |
1245df60 | 3942 | adjustments to the [sh]i_limit for -O0. */ |
c17f53a8 | 3943 | |
1245df60 R |
3944 | si_limit = 1018; |
3945 | hi_limit = 510; | |
e4fa6b06 | 3946 | |
c17f53a8 | 3947 | while (from && count_si < si_limit && count_hi < hi_limit) |
0d7e008e | 3948 | { |
33f7f353 JR |
3949 | int inc = get_attr_length (from); |
3950 | int new_align = 1; | |
1245df60 | 3951 | |
4241ecb0 R |
3952 | /* If this is a label that existed at the time of the compute_alignments |
3953 | call, determine the alignment. N.B. When find_barrier recurses for | |
3954 | an out-of-reach mova, we might see labels at the start of previously | |
3955 | inserted constant tables. */ | |
3956 | if (GET_CODE (from) == CODE_LABEL | |
3957 | && CODE_LABEL_NUMBER (from) <= max_labelno_before_reorg) | |
77008a44 R |
3958 | { |
3959 | if (optimize) | |
3960 | new_align = 1 << label_to_alignment (from); | |
3961 | else if (GET_CODE (prev_nonnote_insn (from)) == BARRIER) | |
3962 | new_align = 1 << barrier_align (from); | |
3963 | else | |
3964 | new_align = 1; | |
3965 | inc = 0; | |
3966 | } | |
4241ecb0 R |
3967 | /* In case we are scanning a constant table because of recursion, check |
3968 | for explicit alignments. If the table is long, we might be forced | |
3969 | to emit the new table in front of it; the length of the alignment | |
3970 | might be the last straw. */ | |
3971 | else if (GET_CODE (from) == INSN | |
3972 | && GET_CODE (PATTERN (from)) == UNSPEC_VOLATILE | |
3973 | && XINT (PATTERN (from), 1) == UNSPECV_ALIGN) | |
3974 | new_align = INTVAL (XVECEXP (PATTERN (from), 0, 0)); | |
3975 | /* When we find the end of a constant table, paste the new constant | |
3976 | at the end. That is better than putting it in front because | |
3977 | this way, we don't need extra alignment for adding a 4-byte-aligned | |
3978 | mov(a) label to a 2/4 or 8/4 byte aligned table. */ | |
3979 | else if (GET_CODE (from) == INSN | |
3980 | && GET_CODE (PATTERN (from)) == UNSPEC_VOLATILE | |
3981 | && XINT (PATTERN (from), 1) == UNSPECV_CONST_END) | |
3982 | return from; | |
8aa2a305 | 3983 | |
0d7e008e | 3984 | if (GET_CODE (from) == BARRIER) |
1245df60 | 3985 | { |
6d4d15b8 | 3986 | rtx next; |
33f7f353 | 3987 | |
1245df60 | 3988 | found_barrier = from; |
33f7f353 | 3989 | |
956d6950 | 3990 | /* If we are at the end of the function, or in front of an alignment |
1245df60 R |
3991 | instruction, we need not insert an extra alignment. We prefer |
3992 | this kind of barrier. */ | |
33f7f353 | 3993 | if (barrier_align (from) > 2) |
1245df60 | 3994 | good_barrier = from; |
6d4d15b8 KK |
3995 | |
3996 | /* If we are at the end of a hot/cold block, dump the constants | |
3997 | here. */ | |
3998 | next = NEXT_INSN (from); | |
3999 | if (next | |
4000 | && NOTE_P (next) | |
4001 | && NOTE_KIND (next) == NOTE_INSN_SWITCH_TEXT_SECTIONS) | |
4002 | break; | |
1245df60 | 4003 | } |
8aa2a305 | 4004 | |
8aa2a305 | 4005 | if (broken_move (from)) |
0d7e008e | 4006 | { |
1245df60 R |
4007 | rtx pat, src, dst; |
4008 | enum machine_mode mode; | |
4009 | ||
4010 | pat = PATTERN (from); | |
4011 | if (GET_CODE (pat) == PARALLEL) | |
4012 | pat = XVECEXP (pat, 0, 0); | |
4013 | src = SET_SRC (pat); | |
4014 | dst = SET_DEST (pat); | |
4015 | mode = GET_MODE (dst); | |
c17f53a8 JW |
4016 | |
4017 | /* We must explicitly check the mode, because sometimes the | |
4018 | front end will generate code to load unsigned constants into | |
4019 | HImode targets without properly sign extending them. */ | |
225e4f43 R |
4020 | if (mode == HImode |
4021 | || (mode == SImode && hi_const (src) && REGNO (dst) != FPUL_REG)) | |
00e94d65 | 4022 | { |
1245df60 | 4023 | found_hi += 2; |
00e94d65 RK |
4024 | /* We put the short constants before the long constants, so |
4025 | we must count the length of short constants in the range | |
4026 | for the long constants. */ | |
4027 | /* ??? This isn't optimal, but is easy to do. */ | |
1245df60 | 4028 | si_limit -= 2; |
00e94d65 | 4029 | } |
0d7e008e | 4030 | else |
1245df60 | 4031 | { |
fa5322fa AO |
4032 | /* We dump DF/DI constants before SF/SI ones, because |
4033 | the limit is the same, but the alignment requirements | |
4034 | are higher. We may waste up to 4 additional bytes | |
4035 | for alignment, and the DF/DI constant may have | |
24746a42 | 4036 | another SF/SI constant placed before it. */ |
fa5322fa AO |
4037 | if (TARGET_SHCOMPACT |
4038 | && ! found_di | |
4039 | && (mode == DFmode || mode == DImode)) | |
4040 | { | |
4041 | found_di = 1; | |
4042 | si_limit -= 8; | |
4043 | } | |
33f7f353 JR |
4044 | while (si_align > 2 && found_si + si_align - 2 > count_si) |
4045 | si_align >>= 1; | |
1245df60 R |
4046 | if (found_si > count_si) |
4047 | count_si = found_si; | |
4048 | found_si += GET_MODE_SIZE (mode); | |
4049 | if (num_mova) | |
4050 | si_limit -= GET_MODE_SIZE (mode); | |
4051 | } | |
0d7e008e | 4052 | } |
5325c0fa | 4053 | |
33f7f353 | 4054 | if (mova_p (from)) |
1245df60 | 4055 | { |
4241ecb0 | 4056 | switch (untangle_mova (&num_mova, &mova, from)) |
1245df60 | 4057 | { |
4241ecb0 R |
4058 | case 0: return find_barrier (0, 0, mova); |
4059 | case 2: | |
4060 | { | |
4061 | leading_mova = 0; | |
4062 | barrier_before_mova | |
4063 | = good_barrier ? good_barrier : found_barrier; | |
4064 | } | |
4065 | default: break; | |
1245df60 R |
4066 | } |
4067 | if (found_si > count_si) | |
4068 | count_si = found_si; | |
4069 | } | |
5325c0fa JW |
4070 | else if (GET_CODE (from) == JUMP_INSN |
4071 | && (GET_CODE (PATTERN (from)) == ADDR_VEC | |
4072 | || GET_CODE (PATTERN (from)) == ADDR_DIFF_VEC)) | |
1245df60 | 4073 | { |
4241ecb0 R |
4074 | if ((num_mova > 1 && GET_MODE (prev_nonnote_insn (from)) == VOIDmode) |
4075 | || (num_mova | |
4076 | && (prev_nonnote_insn (from) | |
4077 | == XEXP (MOVA_LABELREF (mova), 0)))) | |
1245df60 | 4078 | num_mova--; |
10f4f635 | 4079 | if (barrier_align (next_real_insn (from)) == align_jumps_log) |
1245df60 | 4080 | { |
38e01259 | 4081 | /* We have just passed the barrier in front of the |
a0798779 R |
4082 | ADDR_DIFF_VEC, which is stored in found_barrier. Since |
4083 | the ADDR_DIFF_VEC is accessed as data, just like our pool | |
4084 | constants, this is a good opportunity to accommodate what | |
4085 | we have gathered so far. | |
1245df60 R |
4086 | If we waited any longer, we could end up at a barrier in |
4087 | front of code, which gives worse cache usage for separated | |
4088 | instruction / data caches. */ | |
a0798779 | 4089 | good_barrier = found_barrier; |
1245df60 R |
4090 | break; |
4091 | } | |
a0798779 R |
4092 | else |
4093 | { | |
4094 | rtx body = PATTERN (from); | |
4095 | inc = XVECLEN (body, 1) * GET_MODE_SIZE (GET_MODE (body)); | |
4096 | } | |
1245df60 | 4097 | } |
50ceefc2 R |
4098 | /* For the SH1, we generate alignments even after jumps-around-jumps. */ |
4099 | else if (GET_CODE (from) == JUMP_INSN | |
4100 | && ! TARGET_SH2 | |
4101 | && ! TARGET_SMALLCODE) | |
4102 | new_align = 4; | |
5325c0fa | 4103 | |
0d7e008e | 4104 | if (found_si) |
33f7f353 | 4105 | { |
50ceefc2 | 4106 | count_si += inc; |
33f7f353 JR |
4107 | if (new_align > si_align) |
4108 | { | |
318881c0 | 4109 | si_limit -= (count_si - 1) & (new_align - si_align); |
33f7f353 JR |
4110 | si_align = new_align; |
4111 | } | |
318881c0 | 4112 | count_si = (count_si + new_align - 1) & -new_align; |
33f7f353 | 4113 | } |
0d7e008e | 4114 | if (found_hi) |
33f7f353 | 4115 | { |
50ceefc2 | 4116 | count_hi += inc; |
33f7f353 JR |
4117 | if (new_align > hi_align) |
4118 | { | |
318881c0 | 4119 | hi_limit -= (count_hi - 1) & (new_align - hi_align); |
33f7f353 JR |
4120 | hi_align = new_align; |
4121 | } | |
318881c0 | 4122 | count_hi = (count_hi + new_align - 1) & -new_align; |
33f7f353 | 4123 | } |
0d7e008e SC |
4124 | from = NEXT_INSN (from); |
4125 | } | |
4126 | ||
1245df60 | 4127 | if (num_mova) |
eda44c78 R |
4128 | { |
4129 | if (leading_mova) | |
4130 | { | |
4131 | /* Try as we might, the leading mova is out of range. Change | |
4132 | it into a load (which will become a pcload) and retry. */ | |
078c8b08 | 4133 | fixup_mova (mova); |
eda44c78 R |
4134 | return find_barrier (0, 0, mova); |
4135 | } | |
4136 | else | |
4137 | { | |
4138 | /* Insert the constant pool table before the mova instruction, | |
4139 | to prevent the mova label reference from going out of range. */ | |
4140 | from = mova; | |
4141 | good_barrier = found_barrier = barrier_before_mova; | |
4142 | } | |
4143 | } | |
5325c0fa | 4144 | |
1245df60 R |
4145 | if (found_barrier) |
4146 | { | |
33f7f353 | 4147 | if (good_barrier && next_real_insn (found_barrier)) |
1245df60 | 4148 | found_barrier = good_barrier; |
1245df60 R |
4149 | } |
4150 | else | |
b9654711 | 4151 | { |
8aa2a305 JW |
4152 | /* We didn't find a barrier in time to dump our stuff, |
4153 | so we'll make one. */ | |
0d7e008e | 4154 | rtx label = gen_label_rtx (); |
8aa2a305 | 4155 | |
5dbcc9c0 RK |
4156 | /* If we exceeded the range, then we must back up over the last |
4157 | instruction we looked at. Otherwise, we just need to undo the | |
4158 | NEXT_INSN at the end of the loop. */ | |
7a3cb3ab KK |
4159 | if (PREV_INSN (from) != orig |
4160 | && (count_hi > hi_limit || count_si > si_limit)) | |
5dbcc9c0 RK |
4161 | from = PREV_INSN (PREV_INSN (from)); |
4162 | else | |
4163 | from = PREV_INSN (from); | |
4164 | ||
8aa2a305 JW |
4165 | /* Walk back to be just before any jump or label. |
4166 | Putting it before a label reduces the number of times the branch | |
4167 | around the constant pool table will be hit. Putting it before | |
4168 | a jump makes it more likely that the bra delay slot will be | |
4169 | filled. */ | |
4170 | while (GET_CODE (from) == JUMP_INSN || GET_CODE (from) == NOTE | |
8e87e161 | 4171 | || GET_CODE (from) == CODE_LABEL) |
8aa2a305 JW |
4172 | from = PREV_INSN (from); |
4173 | ||
0d7e008e SC |
4174 | from = emit_jump_insn_after (gen_jump (label), from); |
4175 | JUMP_LABEL (from) = label; | |
e081ed26 | 4176 | LABEL_NUSES (label) = 1; |
0d7e008e SC |
4177 | found_barrier = emit_barrier_after (from); |
4178 | emit_label_after (label, found_barrier); | |
b9654711 | 4179 | } |
b9654711 | 4180 | |
8aa2a305 | 4181 | return found_barrier; |
0d7e008e | 4182 | } |
b9654711 | 4183 | |
4787bce0 JW |
4184 | /* If the instruction INSN is implemented by a special function, and we can |
4185 | positively find the register that is used to call the sfunc, and this | |
4186 | register is not used anywhere else in this instruction - except as the | |
4187 | destination of a set, return this register; else, return 0. */ | |
1245df60 | 4188 | rtx |
cf277499 | 4189 | sfunc_uses_reg (rtx insn) |
4787bce0 JW |
4190 | { |
4191 | int i; | |
4192 | rtx pattern, part, reg_part, reg; | |
4193 | ||
4194 | if (GET_CODE (insn) != INSN) | |
4195 | return 0; | |
4196 | pattern = PATTERN (insn); | |
4197 | if (GET_CODE (pattern) != PARALLEL || get_attr_type (insn) != TYPE_SFUNC) | |
4198 | return 0; | |
4199 | ||
4200 | for (reg_part = 0, i = XVECLEN (pattern, 0) - 1; i >= 1; i--) | |
4201 | { | |
4202 | part = XVECEXP (pattern, 0, i); | |
1245df60 | 4203 | if (GET_CODE (part) == USE && GET_MODE (XEXP (part, 0)) == SImode) |
4787bce0 JW |
4204 | reg_part = part; |
4205 | } | |
4206 | if (! reg_part) | |
4207 | return 0; | |
4208 | reg = XEXP (reg_part, 0); | |
4209 | for (i = XVECLEN (pattern, 0) - 1; i >= 0; i--) | |
4210 | { | |
4211 | part = XVECEXP (pattern, 0, i); | |
225e4f43 | 4212 | if (part == reg_part || GET_CODE (part) == CLOBBER) |
4787bce0 JW |
4213 | continue; |
4214 | if (reg_mentioned_p (reg, ((GET_CODE (part) == SET | |
4215 | && GET_CODE (SET_DEST (part)) == REG) | |
4216 | ? SET_SRC (part) : part))) | |
4217 | return 0; | |
4218 | } | |
4219 | return reg; | |
4220 | } | |
4221 | ||
933c3ba3 JW |
4222 | /* See if the only way in which INSN uses REG is by calling it, or by |
4223 | setting it while calling it. Set *SET to a SET rtx if the register | |
4224 | is set by INSN. */ | |
4225 | ||
4226 | static int | |
cf277499 | 4227 | noncall_uses_reg (rtx reg, rtx insn, rtx *set) |
933c3ba3 | 4228 | { |
4787bce0 | 4229 | rtx pattern, reg2; |
933c3ba3 JW |
4230 | |
4231 | *set = NULL_RTX; | |
4232 | ||
4787bce0 JW |
4233 | reg2 = sfunc_uses_reg (insn); |
4234 | if (reg2 && REGNO (reg2) == REGNO (reg)) | |
4235 | { | |
4236 | pattern = single_set (insn); | |
4237 | if (pattern | |
4238 | && GET_CODE (SET_DEST (pattern)) == REG | |
4239 | && REGNO (reg) == REGNO (SET_DEST (pattern))) | |
4240 | *set = pattern; | |
4241 | return 0; | |
4242 | } | |
933c3ba3 JW |
4243 | if (GET_CODE (insn) != CALL_INSN) |
4244 | { | |
4245 | /* We don't use rtx_equal_p because we don't care if the mode is | |
4246 | different. */ | |
4247 | pattern = single_set (insn); | |
4248 | if (pattern | |
4249 | && GET_CODE (SET_DEST (pattern)) == REG | |
4250 | && REGNO (reg) == REGNO (SET_DEST (pattern))) | |
4251 | { | |
4787bce0 JW |
4252 | rtx par, part; |
4253 | int i; | |
4254 | ||
933c3ba3 | 4255 | *set = pattern; |
4787bce0 JW |
4256 | par = PATTERN (insn); |
4257 | if (GET_CODE (par) == PARALLEL) | |
4258 | for (i = XVECLEN (par, 0) - 1; i >= 0; i--) | |
4259 | { | |
4260 | part = XVECEXP (par, 0, i); | |
4261 | if (GET_CODE (part) != SET && reg_mentioned_p (reg, part)) | |
4262 | return 1; | |
4263 | } | |
4264 | return reg_mentioned_p (reg, SET_SRC (pattern)); | |
933c3ba3 JW |
4265 | } |
4266 | ||
4267 | return 1; | |
4268 | } | |
4269 | ||
4270 | pattern = PATTERN (insn); | |
4271 | ||
4272 | if (GET_CODE (pattern) == PARALLEL) | |
4273 | { | |
4274 | int i; | |
4275 | ||
4276 | for (i = XVECLEN (pattern, 0) - 1; i >= 1; i--) | |
4277 | if (reg_mentioned_p (reg, XVECEXP (pattern, 0, i))) | |
4278 | return 1; | |
4279 | pattern = XVECEXP (pattern, 0, 0); | |
4280 | } | |
4281 | ||
4282 | if (GET_CODE (pattern) == SET) | |
4283 | { | |
4284 | if (reg_mentioned_p (reg, SET_DEST (pattern))) | |
4285 | { | |
4286 | /* We don't use rtx_equal_p, because we don't care if the | |
4287 | mode is different. */ | |
4288 | if (GET_CODE (SET_DEST (pattern)) != REG | |
4289 | || REGNO (reg) != REGNO (SET_DEST (pattern))) | |
4290 | return 1; | |
4291 | ||
4292 | *set = pattern; | |
4293 | } | |
4294 | ||
4295 | pattern = SET_SRC (pattern); | |
4296 | } | |
4297 | ||
4298 | if (GET_CODE (pattern) != CALL | |
4299 | || GET_CODE (XEXP (pattern, 0)) != MEM | |
4300 | || ! rtx_equal_p (reg, XEXP (XEXP (pattern, 0), 0))) | |
4301 | return 1; | |
4302 | ||
4303 | return 0; | |
4304 | } | |
4305 | ||
1245df60 R |
4306 | /* Given a X, a pattern of an insn or a part of it, return a mask of used |
4307 | general registers. Bits 0..15 mean that the respective registers | |
4308 | are used as inputs in the instruction. Bits 16..31 mean that the | |
4309 | registers 0..15, respectively, are used as outputs, or are clobbered. | |
4310 | IS_DEST should be set to 16 if X is the destination of a SET, else to 0. */ | |
4311 | int | |
cf277499 | 4312 | regs_used (rtx x, int is_dest) |
1245df60 R |
4313 | { |
4314 | enum rtx_code code; | |
6f7d635c | 4315 | const char *fmt; |
1245df60 R |
4316 | int i, used = 0; |
4317 | ||
4318 | if (! x) | |
4319 | return used; | |
4320 | code = GET_CODE (x); | |
4321 | switch (code) | |
4322 | { | |
4323 | case REG: | |
4324 | if (REGNO (x) < 16) | |
4325 | return (((1 << HARD_REGNO_NREGS (0, GET_MODE (x))) - 1) | |
4326 | << (REGNO (x) + is_dest)); | |
4327 | return 0; | |
4328 | case SUBREG: | |
4329 | { | |
4330 | rtx y = SUBREG_REG (x); | |
f676971a | 4331 | |
1245df60 R |
4332 | if (GET_CODE (y) != REG) |
4333 | break; | |
4334 | if (REGNO (y) < 16) | |
4335 | return (((1 << HARD_REGNO_NREGS (0, GET_MODE (x))) - 1) | |
ddef6bc7 JJ |
4336 | << (REGNO (y) + |
4337 | subreg_regno_offset (REGNO (y), | |
4338 | GET_MODE (y), | |
4339 | SUBREG_BYTE (x), | |
4340 | GET_MODE (x)) + is_dest)); | |
1245df60 R |
4341 | return 0; |
4342 | } | |
4343 | case SET: | |
4344 | return regs_used (SET_SRC (x), 0) | regs_used (SET_DEST (x), 16); | |
4345 | case RETURN: | |
4346 | /* If there was a return value, it must have been indicated with USE. */ | |
4347 | return 0x00ffff00; | |
4348 | case CLOBBER: | |
4349 | is_dest = 1; | |
4350 | break; | |
4351 | case MEM: | |
4352 | is_dest = 0; | |
4353 | break; | |
4354 | case CALL: | |
4355 | used |= 0x00ff00f0; | |
4356 | break; | |
318881c0 KG |
4357 | default: |
4358 | break; | |
1245df60 R |
4359 | } |
4360 | ||
4361 | fmt = GET_RTX_FORMAT (code); | |
4362 | ||
4363 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
4364 | { | |
4365 | if (fmt[i] == 'E') | |
4366 | { | |
4367 | register int j; | |
4368 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
4369 | used |= regs_used (XVECEXP (x, i, j), is_dest); | |
4370 | } | |
4371 | else if (fmt[i] == 'e') | |
4372 | used |= regs_used (XEXP (x, i), is_dest); | |
4373 | } | |
4374 | return used; | |
4375 | } | |
4376 | ||
4377 | /* Create an instruction that prevents redirection of a conditional branch | |
956d6950 | 4378 | to the destination of the JUMP with address ADDR. |
1245df60 R |
4379 | If the branch needs to be implemented as an indirect jump, try to find |
4380 | a scratch register for it. | |
4381 | If NEED_BLOCK is 0, don't do anything unless we need a scratch register. | |
4382 | If any preceding insn that doesn't fit into a delay slot is good enough, | |
4383 | pass 1. Pass 2 if a definite blocking insn is needed. | |
4384 | -1 is used internally to avoid deep recursion. | |
4385 | If a blocking instruction is made or recognized, return it. */ | |
f676971a | 4386 | |
1245df60 | 4387 | static rtx |
cf277499 | 4388 | gen_block_redirect (rtx jump, int addr, int need_block) |
1245df60 R |
4389 | { |
4390 | int dead = 0; | |
4391 | rtx prev = prev_nonnote_insn (jump); | |
4392 | rtx dest; | |
4393 | ||
4394 | /* First, check if we already have an instruction that satisfies our need. */ | |
4395 | if (prev && GET_CODE (prev) == INSN && ! INSN_DELETED_P (prev)) | |
4396 | { | |
4397 | if (INSN_CODE (prev) == CODE_FOR_indirect_jump_scratch) | |
4398 | return prev; | |
4399 | if (GET_CODE (PATTERN (prev)) == USE | |
4400 | || GET_CODE (PATTERN (prev)) == CLOBBER | |
4401 | || get_attr_in_delay_slot (prev) == IN_DELAY_SLOT_YES) | |
4402 | prev = jump; | |
4403 | else if ((need_block &= ~1) < 0) | |
4404 | return prev; | |
4405 | else if (recog_memoized (prev) == CODE_FOR_block_branch_redirect) | |
4406 | need_block = 0; | |
4407 | } | |
f9654065 R |
4408 | if (GET_CODE (PATTERN (jump)) == RETURN) |
4409 | { | |
4410 | if (! need_block) | |
4411 | return prev; | |
4412 | /* Reorg even does nasty things with return insns that cause branches | |
4413 | to go out of range - see find_end_label and callers. */ | |
a556fd39 | 4414 | return emit_insn_before (gen_block_branch_redirect (const0_rtx) , jump); |
f9654065 | 4415 | } |
1245df60 R |
4416 | /* We can't use JUMP_LABEL here because it might be undefined |
4417 | when not optimizing. */ | |
4418 | dest = XEXP (SET_SRC (PATTERN (jump)), 0); | |
4419 | /* If the branch is out of range, try to find a scratch register for it. */ | |
4420 | if (optimize | |
11f9ed1a KG |
4421 | && (INSN_ADDRESSES (INSN_UID (dest)) - addr + (unsigned) 4092 |
4422 | > 4092 + 4098)) | |
1245df60 R |
4423 | { |
4424 | rtx scan; | |
4425 | /* Don't look for the stack pointer as a scratch register, | |
956d6950 | 4426 | it would cause trouble if an interrupt occurred. */ |
1245df60 R |
4427 | unsigned try = 0x7fff, used; |
4428 | int jump_left = flag_expensive_optimizations + 1; | |
f676971a | 4429 | |
1245df60 R |
4430 | /* It is likely that the most recent eligible instruction is wanted for |
4431 | the delay slot. Therefore, find out which registers it uses, and | |
4432 | try to avoid using them. */ | |
f676971a | 4433 | |
318881c0 | 4434 | for (scan = jump; (scan = PREV_INSN (scan)); ) |
1245df60 R |
4435 | { |
4436 | enum rtx_code code; | |
4437 | ||
4438 | if (INSN_DELETED_P (scan)) | |
4439 | continue; | |
4440 | code = GET_CODE (scan); | |
4441 | if (code == CODE_LABEL || code == JUMP_INSN) | |
4442 | break; | |
4443 | if (code == INSN | |
4444 | && GET_CODE (PATTERN (scan)) != USE | |
4445 | && GET_CODE (PATTERN (scan)) != CLOBBER | |
4446 | && get_attr_in_delay_slot (scan) == IN_DELAY_SLOT_YES) | |
4447 | { | |
4448 | try &= ~regs_used (PATTERN (scan), 0); | |
4449 | break; | |
4450 | } | |
4451 | } | |
eda44c78 R |
4452 | for (used = dead = 0, scan = JUMP_LABEL (jump); |
4453 | (scan = NEXT_INSN (scan)); ) | |
1245df60 R |
4454 | { |
4455 | enum rtx_code code; | |
4456 | ||
4457 | if (INSN_DELETED_P (scan)) | |
4458 | continue; | |
4459 | code = GET_CODE (scan); | |
ec8e098d | 4460 | if (INSN_P (scan)) |
1245df60 R |
4461 | { |
4462 | used |= regs_used (PATTERN (scan), 0); | |
4463 | if (code == CALL_INSN) | |
4464 | used |= regs_used (CALL_INSN_FUNCTION_USAGE (scan), 0); | |
4465 | dead |= (used >> 16) & ~used; | |
4466 | if (dead & try) | |
4467 | { | |
4468 | dead &= try; | |
4469 | break; | |
4470 | } | |
4471 | if (code == JUMP_INSN) | |
eda44c78 R |
4472 | { |
4473 | if (jump_left-- && simplejump_p (scan)) | |
4474 | scan = JUMP_LABEL (scan); | |
4475 | else | |
4476 | break; | |
4477 | } | |
1245df60 R |
4478 | } |
4479 | } | |
4480 | /* Mask out the stack pointer again, in case it was | |
4481 | the only 'free' register we have found. */ | |
4482 | dead &= 0x7fff; | |
4483 | } | |
4484 | /* If the immediate destination is still in range, check for possible | |
4485 | threading with a jump beyond the delay slot insn. | |
4486 | Don't check if we are called recursively; the jump has been or will be | |
956d6950 | 4487 | checked in a different invocation then. */ |
f676971a | 4488 | |
1245df60 R |
4489 | else if (optimize && need_block >= 0) |
4490 | { | |
4491 | rtx next = next_active_insn (next_active_insn (dest)); | |
4492 | if (next && GET_CODE (next) == JUMP_INSN | |
4493 | && GET_CODE (PATTERN (next)) == SET | |
ddf8ec7e | 4494 | && recog_memoized (next) == CODE_FOR_jump_compact) |
1245df60 R |
4495 | { |
4496 | dest = JUMP_LABEL (next); | |
4497 | if (dest | |
11f9ed1a | 4498 | && (INSN_ADDRESSES (INSN_UID (dest)) - addr + (unsigned) 4092 |
9d98a694 AO |
4499 | > 4092 + 4098)) |
4500 | gen_block_redirect (next, INSN_ADDRESSES (INSN_UID (next)), -1); | |
1245df60 R |
4501 | } |
4502 | } | |
4503 | ||
4504 | if (dead) | |
4505 | { | |
c5c76735 | 4506 | rtx reg = gen_rtx_REG (SImode, exact_log2 (dead & -dead)); |
1245df60 R |
4507 | |
4508 | /* It would be nice if we could convert the jump into an indirect | |
956d6950 | 4509 | jump / far branch right now, and thus exposing all constituent |
1245df60 R |
4510 | instructions to further optimization. However, reorg uses |
4511 | simplejump_p to determine if there is an unconditional jump where | |
4512 | it should try to schedule instructions from the target of the | |
4513 | branch; simplejump_p fails for indirect jumps even if they have | |
4514 | a JUMP_LABEL. */ | |
4515 | rtx insn = emit_insn_before (gen_indirect_jump_scratch | |
4516 | (reg, GEN_INT (INSN_UID (JUMP_LABEL (jump)))) | |
4517 | , jump); | |
10f4f635 R |
4518 | /* ??? We would like this to have the scope of the jump, but that |
4519 | scope will change when a delay slot insn of an inner scope is added. | |
4520 | Hence, after delay slot scheduling, we'll have to expect | |
4521 | NOTE_INSN_BLOCK_END notes between the indirect_jump_scratch and | |
4522 | the jump. */ | |
f676971a | 4523 | |
86143cb4 | 4524 | INSN_LOCATOR (insn) = INSN_LOCATOR (jump); |
1245df60 R |
4525 | INSN_CODE (insn) = CODE_FOR_indirect_jump_scratch; |
4526 | return insn; | |
4527 | } | |
4528 | else if (need_block) | |
4529 | /* We can't use JUMP_LABEL here because it might be undefined | |
4530 | when not optimizing. */ | |
4531 | return emit_insn_before (gen_block_branch_redirect | |
4532 | (GEN_INT (INSN_UID (XEXP (SET_SRC (PATTERN (jump)), 0)))) | |
4533 | , jump); | |
4534 | return prev; | |
4535 | } | |
4536 | ||
4537 | #define CONDJUMP_MIN -252 | |
4538 | #define CONDJUMP_MAX 262 | |
4539 | struct far_branch | |
4540 | { | |
4541 | /* A label (to be placed) in front of the jump | |
4542 | that jumps to our ultimate destination. */ | |
4543 | rtx near_label; | |
4544 | /* Where we are going to insert it if we cannot move the jump any farther, | |
4545 | or the jump itself if we have picked up an existing jump. */ | |
4546 | rtx insert_place; | |
4547 | /* The ultimate destination. */ | |
4548 | rtx far_label; | |
4549 | struct far_branch *prev; | |
4550 | /* If the branch has already been created, its address; | |
4551 | else the address of its first prospective user. */ | |
4552 | int address; | |
4553 | }; | |
4554 | ||
cf277499 | 4555 | static void gen_far_branch (struct far_branch *); |
1245df60 | 4556 | enum mdep_reorg_phase_e mdep_reorg_phase; |
c5b9ef02 | 4557 | static void |
cf277499 | 4558 | gen_far_branch (struct far_branch *bp) |
1245df60 R |
4559 | { |
4560 | rtx insn = bp->insert_place; | |
4561 | rtx jump; | |
4562 | rtx label = gen_label_rtx (); | |
f5b9e7c9 | 4563 | int ok; |
1245df60 R |
4564 | |
4565 | emit_label_after (label, insn); | |
4566 | if (bp->far_label) | |
4567 | { | |
4568 | jump = emit_jump_insn_after (gen_jump (bp->far_label), insn); | |
4569 | LABEL_NUSES (bp->far_label)++; | |
4570 | } | |
4571 | else | |
4572 | jump = emit_jump_insn_after (gen_return (), insn); | |
225e4f43 R |
4573 | /* Emit a barrier so that reorg knows that any following instructions |
4574 | are not reachable via a fall-through path. | |
fae778eb | 4575 | But don't do this when not optimizing, since we wouldn't suppress the |
225e4f43 R |
4576 | alignment for the barrier then, and could end up with out-of-range |
4577 | pc-relative loads. */ | |
4578 | if (optimize) | |
4579 | emit_barrier_after (jump); | |
1245df60 R |
4580 | emit_label_after (bp->near_label, insn); |
4581 | JUMP_LABEL (jump) = bp->far_label; | |
f5b9e7c9 NS |
4582 | ok = invert_jump (insn, label, 1); |
4583 | gcc_assert (ok); | |
4584 | ||
f9654065 R |
4585 | /* If we are branching around a jump (rather than a return), prevent |
4586 | reorg from using an insn from the jump target as the delay slot insn - | |
4587 | when reorg did this, it pessimized code (we rather hide the delay slot) | |
4588 | and it could cause branches to go out of range. */ | |
4589 | if (bp->far_label) | |
4590 | (emit_insn_after | |
4591 | (gen_stuff_delay_slot | |
4592 | (GEN_INT (INSN_UID (XEXP (SET_SRC (PATTERN (jump)), 0))), | |
4593 | GEN_INT (recog_memoized (insn) == CODE_FOR_branch_false)), | |
4594 | insn)); | |
1245df60 R |
4595 | /* Prevent reorg from undoing our splits. */ |
4596 | gen_block_redirect (jump, bp->address += 2, 2); | |
4597 | } | |
4598 | ||
1245df60 R |
4599 | /* Fix up ADDR_DIFF_VECs. */ |
4600 | void | |
cf277499 | 4601 | fixup_addr_diff_vecs (rtx first) |
1245df60 R |
4602 | { |
4603 | rtx insn; | |
33f7f353 | 4604 | |
1245df60 R |
4605 | for (insn = first; insn; insn = NEXT_INSN (insn)) |
4606 | { | |
eb3881bf | 4607 | rtx vec_lab, pat, prev, prevpat, x, braf_label; |
1245df60 R |
4608 | |
4609 | if (GET_CODE (insn) != JUMP_INSN | |
4610 | || GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC) | |
4611 | continue; | |
4612 | pat = PATTERN (insn); | |
33f7f353 | 4613 | vec_lab = XEXP (XEXP (pat, 0), 0); |
1245df60 | 4614 | |
33f7f353 JR |
4615 | /* Search the matching casesi_jump_2. */ |
4616 | for (prev = vec_lab; ; prev = PREV_INSN (prev)) | |
1245df60 | 4617 | { |
33f7f353 JR |
4618 | if (GET_CODE (prev) != JUMP_INSN) |
4619 | continue; | |
4620 | prevpat = PATTERN (prev); | |
4621 | if (GET_CODE (prevpat) != PARALLEL || XVECLEN (prevpat, 0) != 2) | |
4622 | continue; | |
4623 | x = XVECEXP (prevpat, 0, 1); | |
4624 | if (GET_CODE (x) != USE) | |
4625 | continue; | |
4626 | x = XEXP (x, 0); | |
4627 | if (GET_CODE (x) == LABEL_REF && XEXP (x, 0) == vec_lab) | |
4628 | break; | |
1245df60 | 4629 | } |
157371cf AO |
4630 | /* FIXME: This is a bug in the optimizer, but it seems harmless |
4631 | to just avoid panicing. */ | |
4632 | if (!prev) | |
4633 | continue; | |
eb3881bf R |
4634 | |
4635 | /* Emit the reference label of the braf where it belongs, right after | |
4636 | the casesi_jump_2 (i.e. braf). */ | |
4637 | braf_label = XEXP (XEXP (SET_SRC (XVECEXP (prevpat, 0, 0)), 1), 0); | |
4638 | emit_label_after (braf_label, prev); | |
4639 | ||
33f7f353 JR |
4640 | /* Fix up the ADDR_DIF_VEC to be relative |
4641 | to the reference address of the braf. */ | |
eb3881bf | 4642 | XEXP (XEXP (pat, 0), 0) = braf_label; |
1245df60 | 4643 | } |
1245df60 R |
4644 | } |
4645 | ||
33f7f353 JR |
4646 | /* BARRIER_OR_LABEL is either a BARRIER or a CODE_LABEL immediately following |
4647 | a barrier. Return the base 2 logarithm of the desired alignment. */ | |
1245df60 | 4648 | int |
cf277499 | 4649 | barrier_align (rtx barrier_or_label) |
1245df60 | 4650 | { |
33f7f353 | 4651 | rtx next = next_real_insn (barrier_or_label), pat, prev; |
3a8699c7 | 4652 | int slot, credit, jump_to_next = 0; |
f676971a | 4653 | |
33f7f353 | 4654 | if (! next) |
1245df60 | 4655 | return 0; |
1245df60 | 4656 | |
33f7f353 | 4657 | pat = PATTERN (next); |
1245df60 | 4658 | |
33f7f353 JR |
4659 | if (GET_CODE (pat) == ADDR_DIFF_VEC) |
4660 | return 2; | |
1245df60 | 4661 | |
4773afa4 | 4662 | if (GET_CODE (pat) == UNSPEC_VOLATILE && XINT (pat, 1) == UNSPECV_ALIGN) |
33f7f353 JR |
4663 | /* This is a barrier in front of a constant table. */ |
4664 | return 0; | |
4665 | ||
4666 | prev = prev_real_insn (barrier_or_label); | |
4667 | if (GET_CODE (PATTERN (prev)) == ADDR_DIFF_VEC) | |
1245df60 | 4668 | { |
33f7f353 JR |
4669 | pat = PATTERN (prev); |
4670 | /* If this is a very small table, we want to keep the alignment after | |
4671 | the table to the minimum for proper code alignment. */ | |
4672 | return ((TARGET_SMALLCODE | |
d042370e | 4673 | || ((unsigned) XVECLEN (pat, 1) * GET_MODE_SIZE (GET_MODE (pat)) |
832a3292 | 4674 | <= (unsigned) 1 << (CACHE_LOG - 2))) |
10f4f635 | 4675 | ? 1 << TARGET_SHMEDIA : align_jumps_log); |
1245df60 | 4676 | } |
33f7f353 JR |
4677 | |
4678 | if (TARGET_SMALLCODE) | |
4679 | return 0; | |
4680 | ||
4d070fd3 | 4681 | if (! TARGET_SH2 || ! optimize) |
10f4f635 | 4682 | return align_jumps_log; |
33f7f353 | 4683 | |
225e4f43 R |
4684 | /* When fixing up pcloads, a constant table might be inserted just before |
4685 | the basic block that ends with the barrier. Thus, we can't trust the | |
4686 | instruction lengths before that. */ | |
4687 | if (mdep_reorg_phase > SH_FIXUP_PCLOAD) | |
1245df60 | 4688 | { |
225e4f43 R |
4689 | /* Check if there is an immediately preceding branch to the insn beyond |
4690 | the barrier. We must weight the cost of discarding useful information | |
4691 | from the current cache line when executing this branch and there is | |
4692 | an alignment, against that of fetching unneeded insn in front of the | |
4693 | branch target when there is no alignment. */ | |
4694 | ||
f676971a EC |
4695 | /* There are two delay_slot cases to consider. One is the simple case |
4696 | where the preceding branch is to the insn beyond the barrier (simple | |
4697 | delay slot filling), and the other is where the preceding branch has | |
4698 | a delay slot that is a duplicate of the insn after the barrier | |
4699 | (fill_eager_delay_slots) and the branch is to the insn after the insn | |
b1b41fe4 CP |
4700 | after the barrier. */ |
4701 | ||
225e4f43 R |
4702 | /* PREV is presumed to be the JUMP_INSN for the barrier under |
4703 | investigation. Skip to the insn before it. */ | |
4704 | prev = prev_real_insn (prev); | |
4705 | ||
c5b9ef02 | 4706 | for (slot = 2, credit = (1 << (CACHE_LOG - 2)) + 2; |
225e4f43 R |
4707 | credit >= 0 && prev && GET_CODE (prev) == INSN; |
4708 | prev = prev_real_insn (prev)) | |
4709 | { | |
b1b41fe4 | 4710 | jump_to_next = 0; |
225e4f43 R |
4711 | if (GET_CODE (PATTERN (prev)) == USE |
4712 | || GET_CODE (PATTERN (prev)) == CLOBBER) | |
4713 | continue; | |
4714 | if (GET_CODE (PATTERN (prev)) == SEQUENCE) | |
b1b41fe4 CP |
4715 | { |
4716 | prev = XVECEXP (PATTERN (prev), 0, 1); | |
f676971a | 4717 | if (INSN_UID (prev) == INSN_UID (next)) |
b1b41fe4 CP |
4718 | { |
4719 | /* Delay slot was filled with insn at jump target. */ | |
4720 | jump_to_next = 1; | |
4721 | continue; | |
4722 | } | |
4723 | } | |
4724 | ||
225e4f43 R |
4725 | if (slot && |
4726 | get_attr_in_delay_slot (prev) == IN_DELAY_SLOT_YES) | |
4727 | slot = 0; | |
4728 | credit -= get_attr_length (prev); | |
4729 | } | |
4730 | if (prev | |
4731 | && GET_CODE (prev) == JUMP_INSN | |
2b9d17c7 AH |
4732 | && JUMP_LABEL (prev)) |
4733 | { | |
4734 | rtx x; | |
300c5453 | 4735 | if (jump_to_next |
2b9d17c7 | 4736 | || next_real_insn (JUMP_LABEL (prev)) == next |
1594c6bc AO |
4737 | /* If relax_delay_slots() decides NEXT was redundant |
4738 | with some previous instruction, it will have | |
4739 | redirected PREV's jump to the following insn. */ | |
5d291213 | 4740 | || JUMP_LABEL (prev) == next_nonnote_insn (next) |
300c5453 AO |
4741 | /* There is no upper bound on redundant instructions |
4742 | that might have been skipped, but we must not put an | |
2b9d17c7 | 4743 | alignment where none had been before. */ |
f676971a EC |
4744 | || (x = (NEXT_INSN (NEXT_INSN (PREV_INSN (prev)))), |
4745 | (INSN_P (x) | |
2b9d17c7 | 4746 | && (INSN_CODE (x) == CODE_FOR_block_branch_redirect |
9959c8c7 R |
4747 | || INSN_CODE (x) == CODE_FOR_indirect_jump_scratch |
4748 | || INSN_CODE (x) == CODE_FOR_stuff_delay_slot)))) | |
2b9d17c7 AH |
4749 | { |
4750 | rtx pat = PATTERN (prev); | |
4751 | if (GET_CODE (pat) == PARALLEL) | |
300c5453 | 4752 | pat = XVECEXP (pat, 0, 0); |
2b9d17c7 AH |
4753 | if (credit - slot >= (GET_CODE (SET_SRC (pat)) == PC ? 2 : 0)) |
4754 | return 0; | |
4755 | } | |
300c5453 | 4756 | } |
33f7f353 | 4757 | } |
f676971a | 4758 | |
10f4f635 | 4759 | return align_jumps_log; |
1245df60 R |
4760 | } |
4761 | ||
2148624a R |
4762 | /* If we are inside a phony loop, almost any kind of label can turn up as the |
4763 | first one in the loop. Aligning a braf label causes incorrect switch | |
4764 | destination addresses; we can detect braf labels because they are | |
4765 | followed by a BARRIER. | |
4766 | Applying loop alignment to small constant or switch tables is a waste | |
4767 | of space, so we suppress this too. */ | |
4768 | int | |
cf277499 | 4769 | sh_loop_align (rtx label) |
2148624a R |
4770 | { |
4771 | rtx next = label; | |
4772 | ||
4773 | do | |
4774 | next = next_nonnote_insn (next); | |
4775 | while (next && GET_CODE (next) == CODE_LABEL); | |
4776 | ||
4777 | if (! next | |
2c3c49de | 4778 | || ! INSN_P (next) |
2148624a R |
4779 | || GET_CODE (PATTERN (next)) == ADDR_DIFF_VEC |
4780 | || recog_memoized (next) == CODE_FOR_consttable_2) | |
4781 | return 0; | |
fa5322fa | 4782 | |
10f4f635 | 4783 | return align_loops_log; |
2148624a R |
4784 | } |
4785 | ||
18dbd950 | 4786 | /* Do a final pass over the function, just before delayed branch |
933c3ba3 | 4787 | scheduling. */ |
b9654711 | 4788 | |
18dbd950 | 4789 | static void |
cf277499 | 4790 | sh_reorg (void) |
0d7e008e | 4791 | { |
18dbd950 | 4792 | rtx first, insn, mova = NULL_RTX; |
1245df60 | 4793 | int num_mova; |
c5c76735 JL |
4794 | rtx r0_rtx = gen_rtx_REG (Pmode, 0); |
4795 | rtx r0_inc_rtx = gen_rtx_POST_INC (Pmode, r0_rtx); | |
8aa2a305 | 4796 | |
18dbd950 | 4797 | first = get_insns (); |
4241ecb0 | 4798 | max_labelno_before_reorg = max_label_num (); |
18dbd950 | 4799 | |
cb51ecd2 AO |
4800 | /* We must split call insns before introducing `mova's. If we're |
4801 | optimizing, they'll have already been split. Otherwise, make | |
4802 | sure we don't split them too late. */ | |
4803 | if (! optimize) | |
6f862f2f | 4804 | split_all_insns_noflow (); |
cb51ecd2 | 4805 | |
fa5322fa AO |
4806 | if (TARGET_SHMEDIA) |
4807 | return; | |
4808 | ||
933c3ba3 JW |
4809 | /* If relaxing, generate pseudo-ops to associate function calls with |
4810 | the symbols they call. It does no harm to not generate these | |
4811 | pseudo-ops. However, when we can generate them, it enables to | |
4812 | linker to potentially relax the jsr to a bsr, and eliminate the | |
4813 | register load and, possibly, the constant pool entry. */ | |
4814 | ||
1245df60 | 4815 | mdep_reorg_phase = SH_INSERT_USES_LABELS; |
933c3ba3 JW |
4816 | if (TARGET_RELAX) |
4817 | { | |
cf7c4aa6 HPN |
4818 | /* Remove all REG_LABEL_OPERAND notes. We want to use them for our |
4819 | own purposes. This works because none of the remaining passes | |
933c3ba3 JW |
4820 | need to look at them. |
4821 | ||
4822 | ??? But it may break in the future. We should use a machine | |
4823 | dependent REG_NOTE, or some other approach entirely. */ | |
4824 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
4825 | { | |
2c3c49de | 4826 | if (INSN_P (insn)) |
933c3ba3 JW |
4827 | { |
4828 | rtx note; | |
4829 | ||
cf7c4aa6 HPN |
4830 | while ((note = find_reg_note (insn, REG_LABEL_OPERAND, |
4831 | NULL_RTX)) != 0) | |
933c3ba3 JW |
4832 | remove_note (insn, note); |
4833 | } | |
4834 | } | |
4835 | ||
4836 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
4837 | { | |
4838 | rtx pattern, reg, link, set, scan, dies, label; | |
4839 | int rescan = 0, foundinsn = 0; | |
4840 | ||
4787bce0 JW |
4841 | if (GET_CODE (insn) == CALL_INSN) |
4842 | { | |
4843 | pattern = PATTERN (insn); | |
933c3ba3 | 4844 | |
4787bce0 JW |
4845 | if (GET_CODE (pattern) == PARALLEL) |
4846 | pattern = XVECEXP (pattern, 0, 0); | |
4847 | if (GET_CODE (pattern) == SET) | |
4848 | pattern = SET_SRC (pattern); | |
933c3ba3 | 4849 | |
4787bce0 JW |
4850 | if (GET_CODE (pattern) != CALL |
4851 | || GET_CODE (XEXP (pattern, 0)) != MEM) | |
4852 | continue; | |
933c3ba3 | 4853 | |
4787bce0 JW |
4854 | reg = XEXP (XEXP (pattern, 0), 0); |
4855 | } | |
4856 | else | |
4857 | { | |
4858 | reg = sfunc_uses_reg (insn); | |
4859 | if (! reg) | |
4860 | continue; | |
4861 | } | |
933c3ba3 | 4862 | |
933c3ba3 JW |
4863 | if (GET_CODE (reg) != REG) |
4864 | continue; | |
4865 | ||
6fb5fa3c DB |
4866 | /* Try scanning backward to find where the register is set. */ |
4867 | link = NULL; | |
4868 | for (scan = PREV_INSN (insn); | |
4869 | scan && GET_CODE (scan) != CODE_LABEL; | |
4870 | scan = PREV_INSN (scan)) | |
933c3ba3 | 4871 | { |
6fb5fa3c DB |
4872 | if (! INSN_P (scan)) |
4873 | continue; | |
933c3ba3 | 4874 | |
6fb5fa3c DB |
4875 | if (! reg_mentioned_p (reg, scan)) |
4876 | continue; | |
933c3ba3 | 4877 | |
6fb5fa3c DB |
4878 | if (noncall_uses_reg (reg, scan, &set)) |
4879 | break; | |
933c3ba3 | 4880 | |
6fb5fa3c DB |
4881 | if (set) |
4882 | { | |
4883 | link = scan; | |
4884 | break; | |
933c3ba3 JW |
4885 | } |
4886 | } | |
4887 | ||
4888 | if (! link) | |
4889 | continue; | |
4890 | ||
4891 | /* The register is set at LINK. */ | |
4892 | ||
4893 | /* We can only optimize the function call if the register is | |
4894 | being set to a symbol. In theory, we could sometimes | |
4895 | optimize calls to a constant location, but the assembler | |
4896 | and linker do not support that at present. */ | |
4897 | if (GET_CODE (SET_SRC (set)) != SYMBOL_REF | |
4898 | && GET_CODE (SET_SRC (set)) != LABEL_REF) | |
4899 | continue; | |
4900 | ||
4901 | /* Scan forward from LINK to the place where REG dies, and | |
4902 | make sure that the only insns which use REG are | |
4903 | themselves function calls. */ | |
4904 | ||
c1e8e6bd JW |
4905 | /* ??? This doesn't work for call targets that were allocated |
4906 | by reload, since there may not be a REG_DEAD note for the | |
4907 | register. */ | |
4908 | ||
933c3ba3 JW |
4909 | dies = NULL_RTX; |
4910 | for (scan = NEXT_INSN (link); scan; scan = NEXT_INSN (scan)) | |
4911 | { | |
4912 | rtx scanset; | |
4913 | ||
c1e8e6bd JW |
4914 | /* Don't try to trace forward past a CODE_LABEL if we haven't |
4915 | seen INSN yet. Ordinarily, we will only find the setting insn | |
6fb5fa3c | 4916 | if it is in the same basic block. However, |
c1e8e6bd JW |
4917 | cross-jumping can insert code labels in between the load and |
4918 | the call, and can result in situations where a single call | |
4919 | insn may have two targets depending on where we came from. */ | |
4920 | ||
4921 | if (GET_CODE (scan) == CODE_LABEL && ! foundinsn) | |
4922 | break; | |
4923 | ||
2c3c49de | 4924 | if (! INSN_P (scan)) |
933c3ba3 JW |
4925 | continue; |
4926 | ||
4927 | /* Don't try to trace forward past a JUMP. To optimize | |
4928 | safely, we would have to check that all the | |
c1e8e6bd | 4929 | instructions at the jump destination did not use REG. */ |
933c3ba3 | 4930 | |
c1e8e6bd | 4931 | if (GET_CODE (scan) == JUMP_INSN) |
933c3ba3 JW |
4932 | break; |
4933 | ||
4934 | if (! reg_mentioned_p (reg, scan)) | |
4935 | continue; | |
4936 | ||
4937 | if (noncall_uses_reg (reg, scan, &scanset)) | |
4938 | break; | |
4939 | ||
4940 | if (scan == insn) | |
4941 | foundinsn = 1; | |
4942 | ||
4787bce0 JW |
4943 | if (scan != insn |
4944 | && (GET_CODE (scan) == CALL_INSN || sfunc_uses_reg (scan))) | |
933c3ba3 JW |
4945 | { |
4946 | /* There is a function call to this register other | |
4947 | than the one we are checking. If we optimize | |
4948 | this call, we need to rescan again below. */ | |
4949 | rescan = 1; | |
4950 | } | |
4951 | ||
4952 | /* ??? We shouldn't have to worry about SCANSET here. | |
4953 | We should just be able to check for a REG_DEAD note | |
4954 | on a function call. However, the REG_DEAD notes are | |
4955 | apparently not dependable around libcalls; c-torture | |
4956 | execute/920501-2 is a test case. If SCANSET is set, | |
4957 | then this insn sets the register, so it must have | |
4958 | died earlier. Unfortunately, this will only handle | |
4959 | the cases in which the register is, in fact, set in a | |
4960 | later insn. */ | |
4961 | ||
4962 | /* ??? We shouldn't have to use FOUNDINSN here. | |
6fb5fa3c DB |
4963 | This dates back to when we used LOG_LINKS to find |
4964 | the most recent insn which sets the register. */ | |
933c3ba3 JW |
4965 | |
4966 | if (foundinsn | |
4967 | && (scanset | |
4968 | || find_reg_note (scan, REG_DEAD, reg))) | |
4969 | { | |
4970 | dies = scan; | |
4971 | break; | |
4972 | } | |
4973 | } | |
4974 | ||
4975 | if (! dies) | |
4976 | { | |
4977 | /* Either there was a branch, or some insn used REG | |
4978 | other than as a function call address. */ | |
4979 | continue; | |
4980 | } | |
4981 | ||
cf7c4aa6 HPN |
4982 | /* Create a code label, and put it in a REG_LABEL_OPERAND note |
4983 | on the insn which sets the register, and on each call insn | |
4984 | which uses the register. In final_prescan_insn we look for | |
4985 | the REG_LABEL_OPERAND notes, and output the appropriate label | |
933c3ba3 JW |
4986 | or pseudo-op. */ |
4987 | ||
4988 | label = gen_label_rtx (); | |
65c5f2a6 ILT |
4989 | add_reg_note (link, REG_LABEL_OPERAND, label); |
4990 | add_reg_note (insn, REG_LABEL_OPERAND, label); | |
933c3ba3 JW |
4991 | if (rescan) |
4992 | { | |
4993 | scan = link; | |
4994 | do | |
4995 | { | |
4787bce0 JW |
4996 | rtx reg2; |
4997 | ||
933c3ba3 JW |
4998 | scan = NEXT_INSN (scan); |
4999 | if (scan != insn | |
4787bce0 JW |
5000 | && ((GET_CODE (scan) == CALL_INSN |
5001 | && reg_mentioned_p (reg, scan)) | |
5002 | || ((reg2 = sfunc_uses_reg (scan)) | |
5003 | && REGNO (reg2) == REGNO (reg)))) | |
65c5f2a6 | 5004 | add_reg_note (scan, REG_LABEL_OPERAND, label); |
933c3ba3 JW |
5005 | } |
5006 | while (scan != dies); | |
5007 | } | |
5008 | } | |
5009 | } | |
5010 | ||
33f7f353 JR |
5011 | if (TARGET_SH2) |
5012 | fixup_addr_diff_vecs (first); | |
1245df60 R |
5013 | |
5014 | if (optimize) | |
5015 | { | |
1245df60 R |
5016 | mdep_reorg_phase = SH_SHORTEN_BRANCHES0; |
5017 | shorten_branches (first); | |
5018 | } | |
6dd3c0a5 | 5019 | |
933c3ba3 JW |
5020 | /* Scan the function looking for move instructions which have to be |
5021 | changed to pc-relative loads and insert the literal tables. */ | |
6dd3c0a5 SB |
5022 | label_ref_list_pool = create_alloc_pool ("label references list", |
5023 | sizeof (struct label_ref_list_d), | |
5024 | 30); | |
1245df60 R |
5025 | mdep_reorg_phase = SH_FIXUP_PCLOAD; |
5026 | for (insn = first, num_mova = 0; insn; insn = NEXT_INSN (insn)) | |
0d7e008e | 5027 | { |
1245df60 R |
5028 | if (mova_p (insn)) |
5029 | { | |
078c8b08 R |
5030 | /* ??? basic block reordering can move a switch table dispatch |
5031 | below the switch table. Check if that has happened. | |
5032 | We only have the addresses available when optimizing; but then, | |
5033 | this check shouldn't be needed when not optimizing. */ | |
4241ecb0 | 5034 | if (!untangle_mova (&num_mova, &mova, insn)) |
078c8b08 | 5035 | { |
4241ecb0 R |
5036 | insn = mova; |
5037 | num_mova = 0; | |
078c8b08 | 5038 | } |
1245df60 R |
5039 | } |
5040 | else if (GET_CODE (insn) == JUMP_INSN | |
5041 | && GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC | |
4241ecb0 R |
5042 | && num_mova |
5043 | /* ??? loop invariant motion can also move a mova out of a | |
5044 | loop. Since loop does this code motion anyway, maybe we | |
5045 | should wrap UNSPEC_MOVA into a CONST, so that reload can | |
5046 | move it back. */ | |
5047 | && ((num_mova > 1 | |
5048 | && GET_MODE (prev_nonnote_insn (insn)) == VOIDmode) | |
5049 | || (prev_nonnote_insn (insn) | |
5050 | == XEXP (MOVA_LABELREF (mova), 0)))) | |
1245df60 R |
5051 | { |
5052 | rtx scan; | |
5053 | int total; | |
5054 | ||
5055 | num_mova--; | |
5056 | ||
5057 | /* Some code might have been inserted between the mova and | |
5058 | its ADDR_DIFF_VEC. Check if the mova is still in range. */ | |
5059 | for (scan = mova, total = 0; scan != insn; scan = NEXT_INSN (scan)) | |
33f7f353 | 5060 | total += get_attr_length (scan); |
1245df60 R |
5061 | |
5062 | /* range of mova is 1020, add 4 because pc counts from address of | |
5063 | second instruction after this one, subtract 2 in case pc is 2 | |
5064 | byte aligned. Possible alignment needed for the ADDR_DIFF_VEC | |
956d6950 | 5065 | cancels out with alignment effects of the mova itself. */ |
1245df60 R |
5066 | if (total > 1022) |
5067 | { | |
5068 | /* Change the mova into a load, and restart scanning | |
5069 | there. broken_move will then return true for mova. */ | |
078c8b08 | 5070 | fixup_mova (mova); |
1245df60 R |
5071 | insn = mova; |
5072 | } | |
5073 | } | |
078c8b08 R |
5074 | if (broken_move (insn) |
5075 | || (GET_CODE (insn) == INSN | |
5076 | && recog_memoized (insn) == CODE_FOR_casesi_worker_2)) | |
0d7e008e | 5077 | { |
0d7e008e | 5078 | rtx scan; |
8aa2a305 JW |
5079 | /* Scan ahead looking for a barrier to stick the constant table |
5080 | behind. */ | |
1245df60 | 5081 | rtx barrier = find_barrier (num_mova, mova, insn); |
3a8699c7 | 5082 | rtx last_float_move = NULL_RTX, last_float = 0, *last_float_addr = NULL; |
078c8b08 | 5083 | int need_aligned_label = 0; |
b9654711 | 5084 | |
1245df60 R |
5085 | if (num_mova && ! mova_p (mova)) |
5086 | { | |
5087 | /* find_barrier had to change the first mova into a | |
5088 | pcload; thus, we have to start with this new pcload. */ | |
5089 | insn = mova; | |
5090 | num_mova = 0; | |
5091 | } | |
16bea517 | 5092 | /* Now find all the moves between the points and modify them. */ |
0d7e008e SC |
5093 | for (scan = insn; scan != barrier; scan = NEXT_INSN (scan)) |
5094 | { | |
1245df60 R |
5095 | if (GET_CODE (scan) == CODE_LABEL) |
5096 | last_float = 0; | |
078c8b08 R |
5097 | if (GET_CODE (scan) == INSN |
5098 | && recog_memoized (scan) == CODE_FOR_casesi_worker_2) | |
5099 | need_aligned_label = 1; | |
0d7e008e SC |
5100 | if (broken_move (scan)) |
5101 | { | |
e577c183 JR |
5102 | rtx *patp = &PATTERN (scan), pat = *patp; |
5103 | rtx src, dst; | |
0d7e008e | 5104 | rtx lab; |
0d7e008e | 5105 | rtx newsrc; |
e577c183 JR |
5106 | enum machine_mode mode; |
5107 | ||
5108 | if (GET_CODE (pat) == PARALLEL) | |
5109 | patp = &XVECEXP (pat, 0, 0), pat = *patp; | |
5110 | src = SET_SRC (pat); | |
5111 | dst = SET_DEST (pat); | |
5112 | mode = GET_MODE (dst); | |
0d7e008e | 5113 | |
225e4f43 R |
5114 | if (mode == SImode && hi_const (src) |
5115 | && REGNO (dst) != FPUL_REG) | |
0d7e008e | 5116 | { |
23ed92ca | 5117 | int offset = 0; |
8aa2a305 | 5118 | |
0d7e008e | 5119 | mode = HImode; |
d3ae8277 | 5120 | while (GET_CODE (dst) == SUBREG) |
23ed92ca | 5121 | { |
ddef6bc7 JJ |
5122 | offset += subreg_regno_offset (REGNO (SUBREG_REG (dst)), |
5123 | GET_MODE (SUBREG_REG (dst)), | |
5124 | SUBREG_BYTE (dst), | |
5125 | GET_MODE (dst)); | |
23ed92ca JW |
5126 | dst = SUBREG_REG (dst); |
5127 | } | |
c5c76735 | 5128 | dst = gen_rtx_REG (HImode, REGNO (dst) + offset); |
0d7e008e | 5129 | } |
104ee20b | 5130 | if (GET_CODE (dst) == REG && FP_ANY_REGISTER_P (REGNO (dst))) |
1245df60 | 5131 | { |
b8794099 | 5132 | /* This must be an insn that clobbers r0. */ |
0f2b41b2 AO |
5133 | rtx *clobberp = &XVECEXP (PATTERN (scan), 0, |
5134 | XVECLEN (PATTERN (scan), 0) | |
5135 | - 1); | |
5136 | rtx clobber = *clobberp; | |
b8794099 | 5137 | |
f5b9e7c9 NS |
5138 | gcc_assert (GET_CODE (clobber) == CLOBBER |
5139 | && rtx_equal_p (XEXP (clobber, 0), r0_rtx)); | |
b8794099 | 5140 | |
1245df60 R |
5141 | if (last_float |
5142 | && reg_set_between_p (r0_rtx, last_float_move, scan)) | |
5143 | last_float = 0; | |
3503150c RS |
5144 | if (last_float |
5145 | && TARGET_SHCOMPACT | |
5146 | && GET_MODE_SIZE (mode) != 4 | |
5147 | && GET_MODE_SIZE (GET_MODE (last_float)) == 4) | |
5148 | last_float = 0; | |
1245df60 R |
5149 | lab = add_constant (src, mode, last_float); |
5150 | if (lab) | |
5151 | emit_insn_before (gen_mova (lab), scan); | |
5152 | else | |
b8794099 R |
5153 | { |
5154 | /* There will be a REG_UNUSED note for r0 on | |
5155 | LAST_FLOAT_MOVE; we have to change it to REG_INC, | |
5156 | lest reorg:mark_target_live_regs will not | |
5157 | consider r0 to be used, and we end up with delay | |
475f26dd | 5158 | slot insn in front of SCAN that clobbers r0. */ |
b8794099 R |
5159 | rtx note |
5160 | = find_regno_note (last_float_move, REG_UNUSED, 0); | |
5161 | ||
ab1672a3 JW |
5162 | /* If we are not optimizing, then there may not be |
5163 | a note. */ | |
5164 | if (note) | |
5165 | PUT_MODE (note, REG_INC); | |
b8794099 R |
5166 | |
5167 | *last_float_addr = r0_inc_rtx; | |
5168 | } | |
1245df60 R |
5169 | last_float_move = scan; |
5170 | last_float = src; | |
57d38024 | 5171 | newsrc = gen_const_mem (mode, |
318881c0 | 5172 | (((TARGET_SH4 && ! TARGET_FMOVD) |
225e4f43 | 5173 | || REGNO (dst) == FPUL_REG) |
1245df60 R |
5174 | ? r0_inc_rtx |
5175 | : r0_rtx)); | |
5176 | last_float_addr = &XEXP (newsrc, 0); | |
b8794099 R |
5177 | |
5178 | /* Remove the clobber of r0. */ | |
0f2b41b2 AO |
5179 | *clobberp = gen_rtx_CLOBBER (GET_MODE (clobber), |
5180 | gen_rtx_SCRATCH (Pmode)); | |
1245df60 | 5181 | } |
1a66cd67 | 5182 | /* This is a mova needing a label. Create it. */ |
43c05634 AO |
5183 | else if (GET_CODE (src) == UNSPEC |
5184 | && XINT (src, 1) == UNSPEC_MOVA | |
5185 | && GET_CODE (XVECEXP (src, 0, 0)) == CONST) | |
1a66cd67 | 5186 | { |
43c05634 | 5187 | lab = add_constant (XVECEXP (src, 0, 0), mode, 0); |
1a66cd67 | 5188 | newsrc = gen_rtx_LABEL_REF (VOIDmode, lab); |
e69d1422 | 5189 | newsrc = gen_rtx_UNSPEC (SImode, |
4773afa4 AO |
5190 | gen_rtvec (1, newsrc), |
5191 | UNSPEC_MOVA); | |
1a66cd67 | 5192 | } |
1245df60 R |
5193 | else |
5194 | { | |
5195 | lab = add_constant (src, mode, 0); | |
542a8afa RH |
5196 | newsrc = gen_rtx_LABEL_REF (VOIDmode, lab); |
5197 | newsrc = gen_const_mem (mode, newsrc); | |
1245df60 | 5198 | } |
c5c76735 | 5199 | *patp = gen_rtx_SET (VOIDmode, dst, newsrc); |
e577c183 | 5200 | INSN_CODE (scan) = -1; |
0d7e008e SC |
5201 | } |
5202 | } | |
078c8b08 | 5203 | dump_table (need_aligned_label ? insn : 0, barrier); |
1245df60 | 5204 | insn = barrier; |
0d7e008e SC |
5205 | } |
5206 | } | |
6dd3c0a5 | 5207 | free_alloc_pool (label_ref_list_pool); |
4241ecb0 R |
5208 | for (insn = first; insn; insn = NEXT_INSN (insn)) |
5209 | PUT_MODE (insn, VOIDmode); | |
5210 | ||
1245df60 | 5211 | mdep_reorg_phase = SH_SHORTEN_BRANCHES1; |
9d98a694 | 5212 | INSN_ADDRESSES_FREE (); |
1245df60 R |
5213 | split_branches (first); |
5214 | ||
5215 | /* The INSN_REFERENCES_ARE_DELAYED in sh.h is problematic because it | |
a920aefe | 5216 | also has an effect on the register that holds the address of the sfunc. |
1245df60 R |
5217 | Insert an extra dummy insn in front of each sfunc that pretends to |
5218 | use this register. */ | |
5219 | if (flag_delayed_branch) | |
5220 | { | |
5221 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
5222 | { | |
5223 | rtx reg = sfunc_uses_reg (insn); | |
5224 | ||
5225 | if (! reg) | |
5226 | continue; | |
5227 | emit_insn_before (gen_use_sfunc_addr (reg), insn); | |
5228 | } | |
5229 | } | |
225e4f43 R |
5230 | #if 0 |
5231 | /* fpscr is not actually a user variable, but we pretend it is for the | |
5232 | sake of the previous optimization passes, since we want it handled like | |
9f09b1f2 | 5233 | one. However, we don't have any debugging information for it, so turn |
225e4f43 R |
5234 | it into a non-user variable now. */ |
5235 | if (TARGET_SH4) | |
5236 | REG_USERVAR_P (get_fpscr_rtx ()) = 0; | |
5237 | #endif | |
1245df60 R |
5238 | mdep_reorg_phase = SH_AFTER_MDEP_REORG; |
5239 | } | |
5240 | ||
5241 | int | |
cf277499 | 5242 | get_dest_uid (rtx label, int max_uid) |
1245df60 R |
5243 | { |
5244 | rtx dest = next_real_insn (label); | |
5245 | int dest_uid; | |
5246 | if (! dest) | |
5247 | /* This can happen for an undefined label. */ | |
5248 | return 0; | |
5249 | dest_uid = INSN_UID (dest); | |
5250 | /* If this is a newly created branch redirection blocking instruction, | |
5251 | we cannot index the branch_uid or insn_addresses arrays with its | |
5252 | uid. But then, we won't need to, because the actual destination is | |
5253 | the following branch. */ | |
5254 | while (dest_uid >= max_uid) | |
5255 | { | |
5256 | dest = NEXT_INSN (dest); | |
5257 | dest_uid = INSN_UID (dest); | |
5258 | } | |
5259 | if (GET_CODE (dest) == JUMP_INSN && GET_CODE (PATTERN (dest)) == RETURN) | |
5260 | return 0; | |
5261 | return dest_uid; | |
5262 | } | |
5263 | ||
5264 | /* Split condbranches that are out of range. Also add clobbers for | |
5265 | scratch registers that are needed in far jumps. | |
5266 | We do this before delay slot scheduling, so that it can take our | |
5267 | newly created instructions into account. It also allows us to | |
5268 | find branches with common targets more easily. */ | |
5269 | ||
5270 | static void | |
cf277499 | 5271 | split_branches (rtx first) |
1245df60 R |
5272 | { |
5273 | rtx insn; | |
5274 | struct far_branch **uid_branch, *far_branch_list = 0; | |
5275 | int max_uid = get_max_uid (); | |
f5b9e7c9 | 5276 | int ok; |
1245df60 R |
5277 | |
5278 | /* Find out which branches are out of range. */ | |
1245df60 R |
5279 | shorten_branches (first); |
5280 | ||
5281 | uid_branch = (struct far_branch **) alloca (max_uid * sizeof *uid_branch); | |
2e09e75a | 5282 | memset ((char *) uid_branch, 0, max_uid * sizeof *uid_branch); |
1245df60 R |
5283 | |
5284 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
2c3c49de | 5285 | if (! INSN_P (insn)) |
1245df60 R |
5286 | continue; |
5287 | else if (INSN_DELETED_P (insn)) | |
5288 | { | |
5289 | /* Shorten_branches would split this instruction again, | |
5290 | so transform it into a note. */ | |
a38e7aa5 | 5291 | SET_INSN_DELETED (insn); |
1245df60 R |
5292 | } |
5293 | else if (GET_CODE (insn) == JUMP_INSN | |
5294 | /* Don't mess with ADDR_DIFF_VEC */ | |
5295 | && (GET_CODE (PATTERN (insn)) == SET | |
5296 | || GET_CODE (PATTERN (insn)) == RETURN)) | |
5297 | { | |
5298 | enum attr_type type = get_attr_type (insn); | |
5299 | if (type == TYPE_CBRANCH) | |
5300 | { | |
5301 | rtx next, beyond; | |
f676971a | 5302 | |
1245df60 R |
5303 | if (get_attr_length (insn) > 4) |
5304 | { | |
5305 | rtx src = SET_SRC (PATTERN (insn)); | |
1245df60 | 5306 | rtx olabel = XEXP (XEXP (src, 1), 0); |
9d98a694 | 5307 | int addr = INSN_ADDRESSES (INSN_UID (insn)); |
1245df60 R |
5308 | rtx label = 0; |
5309 | int dest_uid = get_dest_uid (olabel, max_uid); | |
5310 | struct far_branch *bp = uid_branch[dest_uid]; | |
f676971a | 5311 | |
1245df60 | 5312 | /* redirect_jump needs a valid JUMP_LABEL, and it might delete |
25938114 | 5313 | the label if the LABEL_NUSES count drops to zero. There is |
1245df60 R |
5314 | always a jump_optimize pass that sets these values, but it |
5315 | proceeds to delete unreferenced code, and then if not | |
956d6950 | 5316 | optimizing, to un-delete the deleted instructions, thus |
1245df60 R |
5317 | leaving labels with too low uses counts. */ |
5318 | if (! optimize) | |
5319 | { | |
5320 | JUMP_LABEL (insn) = olabel; | |
5321 | LABEL_NUSES (olabel)++; | |
5322 | } | |
5323 | if (! bp) | |
5324 | { | |
5325 | bp = (struct far_branch *) alloca (sizeof *bp); | |
5326 | uid_branch[dest_uid] = bp; | |
5327 | bp->prev = far_branch_list; | |
5328 | far_branch_list = bp; | |
5329 | bp->far_label | |
5330 | = XEXP (XEXP (SET_SRC (PATTERN (insn)), 1), 0); | |
5331 | LABEL_NUSES (bp->far_label)++; | |
5332 | } | |
5333 | else | |
5334 | { | |
5335 | label = bp->near_label; | |
5336 | if (! label && bp->address - addr >= CONDJUMP_MIN) | |
5337 | { | |
5338 | rtx block = bp->insert_place; | |
5339 | ||
5340 | if (GET_CODE (PATTERN (block)) == RETURN) | |
5341 | block = PREV_INSN (block); | |
5342 | else | |
5343 | block = gen_block_redirect (block, | |
5344 | bp->address, 2); | |
5345 | label = emit_label_after (gen_label_rtx (), | |
5346 | PREV_INSN (block)); | |
5347 | bp->near_label = label; | |
5348 | } | |
5349 | else if (label && ! NEXT_INSN (label)) | |
eda44c78 R |
5350 | { |
5351 | if (addr + 2 - bp->address <= CONDJUMP_MAX) | |
5352 | bp->insert_place = insn; | |
5353 | else | |
5354 | gen_far_branch (bp); | |
5355 | } | |
1245df60 R |
5356 | } |
5357 | if (! label | |
318881c0 | 5358 | || (NEXT_INSN (label) && bp->address - addr < CONDJUMP_MIN)) |
1245df60 R |
5359 | { |
5360 | bp->near_label = label = gen_label_rtx (); | |
5361 | bp->insert_place = insn; | |
5362 | bp->address = addr; | |
5363 | } | |
2acdc1fb | 5364 | ok = redirect_jump (insn, label, 0); |
f5b9e7c9 | 5365 | gcc_assert (ok); |
1245df60 R |
5366 | } |
5367 | else | |
5368 | { | |
5369 | /* get_attr_length (insn) == 2 */ | |
5370 | /* Check if we have a pattern where reorg wants to redirect | |
5371 | the branch to a label from an unconditional branch that | |
5372 | is too far away. */ | |
5373 | /* We can't use JUMP_LABEL here because it might be undefined | |
5374 | when not optimizing. */ | |
33f7f353 | 5375 | /* A syntax error might cause beyond to be NULL_RTX. */ |
1245df60 R |
5376 | beyond |
5377 | = next_active_insn (XEXP (XEXP (SET_SRC (PATTERN (insn)), 1), | |
5378 | 0)); | |
f676971a | 5379 | |
33f7f353 JR |
5380 | if (beyond |
5381 | && (GET_CODE (beyond) == JUMP_INSN | |
b94544e4 AO |
5382 | || ((beyond = next_active_insn (beyond)) |
5383 | && GET_CODE (beyond) == JUMP_INSN)) | |
1245df60 | 5384 | && GET_CODE (PATTERN (beyond)) == SET |
ddf8ec7e | 5385 | && recog_memoized (beyond) == CODE_FOR_jump_compact |
9d98a694 AO |
5386 | && ((INSN_ADDRESSES |
5387 | (INSN_UID (XEXP (SET_SRC (PATTERN (beyond)), 0))) | |
11f9ed1a | 5388 | - INSN_ADDRESSES (INSN_UID (insn)) + (unsigned) 252) |
1245df60 R |
5389 | > 252 + 258 + 2)) |
5390 | gen_block_redirect (beyond, | |
9d98a694 | 5391 | INSN_ADDRESSES (INSN_UID (beyond)), 1); |
1245df60 | 5392 | } |
f676971a | 5393 | |
1245df60 R |
5394 | next = next_active_insn (insn); |
5395 | ||
5396 | if ((GET_CODE (next) == JUMP_INSN | |
03904fb3 KK |
5397 | || ((next = next_active_insn (next)) |
5398 | && GET_CODE (next) == JUMP_INSN)) | |
1245df60 | 5399 | && GET_CODE (PATTERN (next)) == SET |
ddf8ec7e | 5400 | && recog_memoized (next) == CODE_FOR_jump_compact |
9d98a694 AO |
5401 | && ((INSN_ADDRESSES |
5402 | (INSN_UID (XEXP (SET_SRC (PATTERN (next)), 0))) | |
11f9ed1a | 5403 | - INSN_ADDRESSES (INSN_UID (insn)) + (unsigned) 252) |
1245df60 | 5404 | > 252 + 258 + 2)) |
9d98a694 | 5405 | gen_block_redirect (next, INSN_ADDRESSES (INSN_UID (next)), 1); |
1245df60 R |
5406 | } |
5407 | else if (type == TYPE_JUMP || type == TYPE_RETURN) | |
5408 | { | |
9d98a694 | 5409 | int addr = INSN_ADDRESSES (INSN_UID (insn)); |
1245df60 R |
5410 | rtx far_label = 0; |
5411 | int dest_uid = 0; | |
5412 | struct far_branch *bp; | |
5413 | ||
5414 | if (type == TYPE_JUMP) | |
5415 | { | |
5416 | far_label = XEXP (SET_SRC (PATTERN (insn)), 0); | |
5417 | dest_uid = get_dest_uid (far_label, max_uid); | |
5418 | if (! dest_uid) | |
5419 | { | |
5420 | /* Parse errors can lead to labels outside | |
5421 | the insn stream. */ | |
5422 | if (! NEXT_INSN (far_label)) | |
5423 | continue; | |
5424 | ||
5425 | if (! optimize) | |
5426 | { | |
5427 | JUMP_LABEL (insn) = far_label; | |
5428 | LABEL_NUSES (far_label)++; | |
5429 | } | |
8ceaac3c | 5430 | redirect_jump (insn, NULL_RTX, 1); |
1245df60 R |
5431 | far_label = 0; |
5432 | } | |
5433 | } | |
5434 | bp = uid_branch[dest_uid]; | |
5435 | if (! bp) | |
5436 | { | |
5437 | bp = (struct far_branch *) alloca (sizeof *bp); | |
5438 | uid_branch[dest_uid] = bp; | |
5439 | bp->prev = far_branch_list; | |
5440 | far_branch_list = bp; | |
5441 | bp->near_label = 0; | |
5442 | bp->far_label = far_label; | |
5443 | if (far_label) | |
5444 | LABEL_NUSES (far_label)++; | |
5445 | } | |
5446 | else if (bp->near_label && ! NEXT_INSN (bp->near_label)) | |
5447 | if (addr - bp->address <= CONDJUMP_MAX) | |
5448 | emit_label_after (bp->near_label, PREV_INSN (insn)); | |
5449 | else | |
5450 | { | |
5451 | gen_far_branch (bp); | |
5452 | bp->near_label = 0; | |
5453 | } | |
5454 | else | |
5455 | bp->near_label = 0; | |
5456 | bp->address = addr; | |
5457 | bp->insert_place = insn; | |
5458 | if (! far_label) | |
5459 | emit_insn_before (gen_block_branch_redirect (const0_rtx), insn); | |
5460 | else | |
5461 | gen_block_redirect (insn, addr, bp->near_label ? 2 : 0); | |
5462 | } | |
5463 | } | |
5464 | /* Generate all pending far branches, | |
5465 | and free our references to the far labels. */ | |
5466 | while (far_branch_list) | |
5467 | { | |
5468 | if (far_branch_list->near_label | |
5469 | && ! NEXT_INSN (far_branch_list->near_label)) | |
5470 | gen_far_branch (far_branch_list); | |
5471 | if (optimize | |
5472 | && far_branch_list->far_label | |
5473 | && ! --LABEL_NUSES (far_branch_list->far_label)) | |
5474 | delete_insn (far_branch_list->far_label); | |
5475 | far_branch_list = far_branch_list->prev; | |
5476 | } | |
a0798779 R |
5477 | |
5478 | /* Instruction length information is no longer valid due to the new | |
5479 | instructions that have been generated. */ | |
5480 | init_insn_lengths (); | |
b9654711 SC |
5481 | } |
5482 | ||
8aa2a305 | 5483 | /* Dump out instruction addresses, which is useful for debugging the |
933c3ba3 JW |
5484 | constant pool table stuff. |
5485 | ||
5486 | If relaxing, output the label and pseudo-ops used to link together | |
5487 | calls and the instruction which set the registers. */ | |
8aa2a305 | 5488 | |
8aa2a305 JW |
5489 | /* ??? The addresses printed by this routine for insns are nonsense for |
5490 | insns which are inside of a sequence where none of the inner insns have | |
5491 | variable length. This is because the second pass of shorten_branches | |
5492 | does not bother to update them. */ | |
0d7e008e | 5493 | |
8e87e161 | 5494 | void |
cf277499 SB |
5495 | final_prescan_insn (rtx insn, rtx *opvec ATTRIBUTE_UNUSED, |
5496 | int noperands ATTRIBUTE_UNUSED) | |
b9654711 | 5497 | { |
8aa2a305 | 5498 | if (TARGET_DUMPISIZE) |
9d98a694 | 5499 | fprintf (asm_out_file, "\n! at %04x\n", INSN_ADDRESSES (INSN_UID (insn))); |
933c3ba3 JW |
5500 | |
5501 | if (TARGET_RELAX) | |
5502 | { | |
5503 | rtx note; | |
5504 | ||
cf7c4aa6 | 5505 | note = find_reg_note (insn, REG_LABEL_OPERAND, NULL_RTX); |
933c3ba3 JW |
5506 | if (note) |
5507 | { | |
5508 | rtx pattern; | |
5509 | ||
5510 | pattern = PATTERN (insn); | |
4b7d2f07 R |
5511 | if (GET_CODE (pattern) == PARALLEL) |
5512 | pattern = XVECEXP (pattern, 0, 0); | |
f5b9e7c9 NS |
5513 | switch (GET_CODE (pattern)) |
5514 | { | |
f5b9e7c9 NS |
5515 | case SET: |
5516 | if (GET_CODE (SET_SRC (pattern)) != CALL | |
5517 | && get_attr_type (insn) != TYPE_SFUNC) | |
5518 | { | |
5519 | targetm.asm_out.internal_label | |
5520 | (asm_out_file, "L", CODE_LABEL_NUMBER (XEXP (note, 0))); | |
5521 | break; | |
5522 | } | |
5523 | /* else FALLTHROUGH */ | |
5524 | case CALL: | |
5525 | asm_fprintf (asm_out_file, "\t.uses %LL%d\n", | |
5526 | CODE_LABEL_NUMBER (XEXP (note, 0))); | |
5527 | break; | |
5528 | ||
5529 | default: | |
5530 | gcc_unreachable (); | |
5531 | } | |
933c3ba3 JW |
5532 | } |
5533 | } | |
0d7e008e | 5534 | } |
b9654711 | 5535 | |
8aa2a305 | 5536 | /* Dump out any constants accumulated in the final pass. These will |
38e01259 | 5537 | only be labels. */ |
b9654711 | 5538 | |
318881c0 | 5539 | const char * |
cf277499 | 5540 | output_jump_label_table (void) |
0d7e008e | 5541 | { |
8aa2a305 JW |
5542 | int i; |
5543 | ||
5544 | if (pool_size) | |
5545 | { | |
5546 | fprintf (asm_out_file, "\t.align 2\n"); | |
5547 | for (i = 0; i < pool_size; i++) | |
5548 | { | |
5549 | pool_node *p = &pool_vector[i]; | |
5550 | ||
4977bab6 | 5551 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
8aa2a305 JW |
5552 | CODE_LABEL_NUMBER (p->label)); |
5553 | output_asm_insn (".long %O0", &p->value); | |
5554 | } | |
5555 | pool_size = 0; | |
5556 | } | |
b9654711 | 5557 | |
8aa2a305 JW |
5558 | return ""; |
5559 | } | |
5560 | \f | |
5561 | /* A full frame looks like: | |
16bea517 JW |
5562 | |
5563 | arg-5 | |
5564 | arg-4 | |
5565 | [ if current_function_anonymous_args | |
5566 | arg-3 | |
5567 | arg-2 | |
5568 | arg-1 | |
5569 | arg-0 ] | |
5570 | saved-fp | |
5571 | saved-r10 | |
5572 | saved-r11 | |
5573 | saved-r12 | |
5574 | saved-pr | |
5575 | local-n | |
5576 | .. | |
5577 | local-1 | |
5578 | local-0 <- fp points here. */ | |
8e87e161 | 5579 | |
8aa2a305 JW |
5580 | /* Number of bytes pushed for anonymous args, used to pass information |
5581 | between expand_prologue and expand_epilogue. */ | |
5582 | ||
fada1961 R |
5583 | /* Adjust the stack by SIZE bytes. REG holds the rtl of the register to be |
5584 | adjusted. If epilogue_p is zero, this is for a prologue; otherwise, it's | |
726d4cb7 KK |
5585 | for an epilogue and a negative value means that it's for a sibcall |
5586 | epilogue. If LIVE_REGS_MASK is nonzero, it points to a HARD_REG_SET of | |
5587 | all the registers that are about to be restored, and hence dead. */ | |
8aa2a305 JW |
5588 | |
5589 | static void | |
fada1961 R |
5590 | output_stack_adjust (int size, rtx reg, int epilogue_p, |
5591 | HARD_REG_SET *live_regs_mask) | |
8aa2a305 | 5592 | { |
fada1961 | 5593 | rtx (*emit_fn) (rtx) = epilogue_p ? &emit_insn : &frame_insn; |
8aa2a305 JW |
5594 | if (size) |
5595 | { | |
fa5322fa AO |
5596 | HOST_WIDE_INT align = STACK_BOUNDARY / BITS_PER_UNIT; |
5597 | ||
157371cf AO |
5598 | /* This test is bogus, as output_stack_adjust is used to re-align the |
5599 | stack. */ | |
5600 | #if 0 | |
f5b9e7c9 | 5601 | gcc_assert (!(size % align)); |
157371cf | 5602 | #endif |
fa5322fa AO |
5603 | |
5604 | if (CONST_OK_FOR_ADD (size)) | |
2754d3c5 | 5605 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size))); |
f3cd5375 JW |
5606 | /* Try to do it with two partial adjustments; however, we must make |
5607 | sure that the stack is properly aligned at all times, in case | |
6f317ef3 | 5608 | an interrupt occurs between the two partial adjustments. */ |
fa5322fa AO |
5609 | else if (CONST_OK_FOR_ADD (size / 2 & -align) |
5610 | && CONST_OK_FOR_ADD (size - (size / 2 & -align))) | |
f3cd5375 | 5611 | { |
2754d3c5 R |
5612 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size / 2 & -align))); |
5613 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size - (size / 2 & -align)))); | |
f3cd5375 JW |
5614 | } |
5615 | else | |
8aa2a305 | 5616 | { |
f3cd5375 | 5617 | rtx const_reg; |
2754d3c5 | 5618 | rtx insn; |
fada1961 R |
5619 | int temp = epilogue_p ? 7 : (TARGET_SH5 ? 0 : 1); |
5620 | int i; | |
f3cd5375 JW |
5621 | |
5622 | /* If TEMP is invalid, we could temporarily save a general | |
5623 | register to MACL. However, there is currently no need | |
f5b9e7c9 | 5624 | to handle this case, so just die when we see it. */ |
726d4cb7 KK |
5625 | if (epilogue_p < 0 |
5626 | || current_function_interrupt | |
62164eb4 | 5627 | || ! call_really_used_regs[temp] || fixed_regs[temp]) |
fada1961 | 5628 | temp = -1; |
726d4cb7 KK |
5629 | if (temp < 0 && ! current_function_interrupt |
5630 | && (TARGET_SHMEDIA || epilogue_p >= 0)) | |
fada1961 R |
5631 | { |
5632 | HARD_REG_SET temps; | |
5633 | COPY_HARD_REG_SET (temps, call_used_reg_set); | |
5634 | AND_COMPL_HARD_REG_SET (temps, call_fixed_reg_set); | |
726d4cb7 | 5635 | if (epilogue_p > 0) |
fada1961 | 5636 | { |
726d4cb7 | 5637 | int nreg = 0; |
38173d38 | 5638 | if (crtl->return_rtx) |
726d4cb7 KK |
5639 | { |
5640 | enum machine_mode mode; | |
38173d38 | 5641 | mode = GET_MODE (crtl->return_rtx); |
726d4cb7 KK |
5642 | if (BASE_RETURN_VALUE_REG (mode) == FIRST_RET_REG) |
5643 | nreg = HARD_REGNO_NREGS (FIRST_RET_REG, mode); | |
5644 | } | |
5645 | for (i = 0; i < nreg; i++) | |
fada1961 | 5646 | CLEAR_HARD_REG_BIT (temps, FIRST_RET_REG + i); |
e3b5732b | 5647 | if (crtl->calls_eh_return) |
fada1961 R |
5648 | { |
5649 | CLEAR_HARD_REG_BIT (temps, EH_RETURN_STACKADJ_REGNO); | |
5650 | for (i = 0; i <= 3; i++) | |
5651 | CLEAR_HARD_REG_BIT (temps, EH_RETURN_DATA_REGNO (i)); | |
5652 | } | |
5653 | } | |
726d4cb7 KK |
5654 | if (TARGET_SHMEDIA && epilogue_p < 0) |
5655 | for (i = FIRST_TARGET_REG; i <= LAST_TARGET_REG; i++) | |
5656 | CLEAR_HARD_REG_BIT (temps, i); | |
5657 | if (epilogue_p <= 0) | |
fada1961 R |
5658 | { |
5659 | for (i = FIRST_PARM_REG; | |
5660 | i < FIRST_PARM_REG + NPARM_REGS (SImode); i++) | |
5661 | CLEAR_HARD_REG_BIT (temps, i); | |
6de9cd9a | 5662 | if (cfun->static_chain_decl != NULL) |
fada1961 R |
5663 | CLEAR_HARD_REG_BIT (temps, STATIC_CHAIN_REGNUM); |
5664 | } | |
5665 | temp = scavenge_reg (&temps); | |
5666 | } | |
5667 | if (temp < 0 && live_regs_mask) | |
eb996a4a KK |
5668 | { |
5669 | HARD_REG_SET temps; | |
5670 | ||
5671 | COPY_HARD_REG_SET (temps, *live_regs_mask); | |
5672 | CLEAR_HARD_REG_BIT (temps, REGNO (reg)); | |
5673 | temp = scavenge_reg (&temps); | |
5674 | } | |
f3cd5375 | 5675 | if (temp < 0) |
726d4cb7 | 5676 | { |
f5b9e7c9 NS |
5677 | rtx adj_reg, tmp_reg, mem; |
5678 | ||
726d4cb7 KK |
5679 | /* If we reached here, the most likely case is the (sibcall) |
5680 | epilogue for non SHmedia. Put a special push/pop sequence | |
5681 | for such case as the last resort. This looks lengthy but | |
f5b9e7c9 NS |
5682 | would not be problem because it seems to be very |
5683 | rare. */ | |
5684 | ||
5685 | gcc_assert (!TARGET_SHMEDIA && epilogue_p); | |
5686 | ||
5687 | ||
5688 | /* ??? There is still the slight possibility that r4 or | |
5689 | r5 have been reserved as fixed registers or assigned | |
5690 | as global registers, and they change during an | |
5691 | interrupt. There are possible ways to handle this: | |
5692 | ||
5693 | - If we are adjusting the frame pointer (r14), we can do | |
5694 | with a single temp register and an ordinary push / pop | |
5695 | on the stack. | |
5696 | - Grab any call-used or call-saved registers (i.e. not | |
5697 | fixed or globals) for the temps we need. We might | |
5698 | also grab r14 if we are adjusting the stack pointer. | |
5699 | If we can't find enough available registers, issue | |
5700 | a diagnostic and die - the user must have reserved | |
5701 | way too many registers. | |
5702 | But since all this is rather unlikely to happen and | |
5703 | would require extra testing, we just die if r4 / r5 | |
5704 | are not available. */ | |
5705 | gcc_assert (!fixed_regs[4] && !fixed_regs[5] | |
5706 | && !global_regs[4] && !global_regs[5]); | |
5707 | ||
5708 | adj_reg = gen_rtx_REG (GET_MODE (reg), 4); | |
5709 | tmp_reg = gen_rtx_REG (GET_MODE (reg), 5); | |
57d38024 | 5710 | emit_move_insn (gen_tmp_stack_mem (Pmode, reg), adj_reg); |
f5b9e7c9 NS |
5711 | emit_insn (GEN_MOV (adj_reg, GEN_INT (size))); |
5712 | emit_insn (GEN_ADD3 (adj_reg, adj_reg, reg)); | |
57d38024 | 5713 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_PRE_DEC (Pmode, adj_reg)); |
f5b9e7c9 | 5714 | emit_move_insn (mem, tmp_reg); |
57d38024 R |
5715 | emit_move_insn (tmp_reg, gen_tmp_stack_mem (Pmode, reg)); |
5716 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_PRE_DEC (Pmode, adj_reg)); | |
5717 | emit_move_insn (mem, tmp_reg); | |
5718 | emit_move_insn (reg, adj_reg); | |
5719 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_POST_INC (Pmode, reg)); | |
5720 | emit_move_insn (adj_reg, mem); | |
5721 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_POST_INC (Pmode, reg)); | |
5722 | emit_move_insn (tmp_reg, mem); | |
eb996a4a | 5723 | /* Tell flow the insns that pop r4/r5 aren't dead. */ |
c41c1387 RS |
5724 | emit_use (tmp_reg); |
5725 | emit_use (adj_reg); | |
57d38024 | 5726 | return; |
726d4cb7 | 5727 | } |
fa5322fa | 5728 | const_reg = gen_rtx_REG (GET_MODE (reg), temp); |
f3cd5375 JW |
5729 | |
5730 | /* If SIZE is negative, subtract the positive value. | |
5731 | This sometimes allows a constant pool entry to be shared | |
5732 | between prologue and epilogue code. */ | |
5733 | if (size < 0) | |
885c9a39 | 5734 | { |
fa5322fa | 5735 | emit_insn (GEN_MOV (const_reg, GEN_INT (-size))); |
2754d3c5 | 5736 | insn = emit_fn (GEN_SUB3 (reg, reg, const_reg)); |
885c9a39 JR |
5737 | } |
5738 | else | |
5739 | { | |
fa5322fa | 5740 | emit_insn (GEN_MOV (const_reg, GEN_INT (size))); |
2754d3c5 | 5741 | insn = emit_fn (GEN_ADD3 (reg, reg, const_reg)); |
885c9a39 | 5742 | } |
fada1961 | 5743 | if (! epilogue_p) |
2754d3c5 R |
5744 | REG_NOTES (insn) |
5745 | = (gen_rtx_EXPR_LIST | |
5746 | (REG_FRAME_RELATED_EXPR, | |
5747 | gen_rtx_SET (VOIDmode, reg, | |
5748 | gen_rtx_PLUS (SImode, reg, GEN_INT (size))), | |
5749 | REG_NOTES (insn))); | |
8aa2a305 | 5750 | } |
8aa2a305 JW |
5751 | } |
5752 | } | |
5753 | ||
2754d3c5 | 5754 | static rtx |
cf277499 | 5755 | frame_insn (rtx x) |
2754d3c5 R |
5756 | { |
5757 | x = emit_insn (x); | |
5758 | RTX_FRAME_RELATED_P (x) = 1; | |
5759 | return x; | |
5760 | } | |
5761 | ||
8aa2a305 JW |
5762 | /* Output RTL to push register RN onto the stack. */ |
5763 | ||
2754d3c5 | 5764 | static rtx |
cf277499 | 5765 | push (int rn) |
8aa2a305 JW |
5766 | { |
5767 | rtx x; | |
225e4f43 R |
5768 | if (rn == FPUL_REG) |
5769 | x = gen_push_fpul (); | |
7144b2d8 DD |
5770 | else if (rn == FPSCR_REG) |
5771 | x = gen_push_fpscr (); | |
157371cf | 5772 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && ! TARGET_FPU_SINGLE |
104ee20b | 5773 | && FP_OR_XD_REGISTER_P (rn)) |
225e4f43 | 5774 | { |
104ee20b | 5775 | if (FP_REGISTER_P (rn) && (rn - FIRST_FP_REG) & 1) |
d042370e | 5776 | return NULL_RTX; |
c5c76735 | 5777 | x = gen_push_4 (gen_rtx_REG (DFmode, rn)); |
225e4f43 | 5778 | } |
3a8699c7 | 5779 | else if (TARGET_SH2E && FP_REGISTER_P (rn)) |
c5c76735 | 5780 | x = gen_push_e (gen_rtx_REG (SFmode, rn)); |
1a95a963 | 5781 | else |
c5c76735 | 5782 | x = gen_push (gen_rtx_REG (SImode, rn)); |
1a95a963 | 5783 | |
2754d3c5 | 5784 | x = frame_insn (x); |
c5c76735 JL |
5785 | REG_NOTES (x) |
5786 | = gen_rtx_EXPR_LIST (REG_INC, | |
5787 | gen_rtx_REG (SImode, STACK_POINTER_REGNUM), 0); | |
2754d3c5 | 5788 | return x; |
8aa2a305 JW |
5789 | } |
5790 | ||
5791 | /* Output RTL to pop register RN from the stack. */ | |
5792 | ||
5793 | static void | |
cf277499 | 5794 | pop (int rn) |
8aa2a305 JW |
5795 | { |
5796 | rtx x; | |
225e4f43 R |
5797 | if (rn == FPUL_REG) |
5798 | x = gen_pop_fpul (); | |
7144b2d8 DD |
5799 | else if (rn == FPSCR_REG) |
5800 | x = gen_pop_fpscr (); | |
157371cf | 5801 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && ! TARGET_FPU_SINGLE |
104ee20b | 5802 | && FP_OR_XD_REGISTER_P (rn)) |
225e4f43 | 5803 | { |
104ee20b | 5804 | if (FP_REGISTER_P (rn) && (rn - FIRST_FP_REG) & 1) |
225e4f43 | 5805 | return; |
c5c76735 | 5806 | x = gen_pop_4 (gen_rtx_REG (DFmode, rn)); |
225e4f43 | 5807 | } |
3a8699c7 | 5808 | else if (TARGET_SH2E && FP_REGISTER_P (rn)) |
c5c76735 | 5809 | x = gen_pop_e (gen_rtx_REG (SFmode, rn)); |
1a95a963 | 5810 | else |
c5c76735 | 5811 | x = gen_pop (gen_rtx_REG (SImode, rn)); |
f676971a | 5812 | |
1245df60 | 5813 | x = emit_insn (x); |
c5c76735 JL |
5814 | REG_NOTES (x) |
5815 | = gen_rtx_EXPR_LIST (REG_INC, | |
5816 | gen_rtx_REG (SImode, STACK_POINTER_REGNUM), 0); | |
8aa2a305 JW |
5817 | } |
5818 | ||
1245df60 | 5819 | /* Generate code to push the regs specified in the mask. */ |
8aa2a305 JW |
5820 | |
5821 | static void | |
cf277499 | 5822 | push_regs (HARD_REG_SET *mask, int interrupt_handler) |
8aa2a305 | 5823 | { |
5c7cafa8 | 5824 | int i = interrupt_handler ? LAST_BANKED_REG + 1 : 0; |
ac55736a | 5825 | int skip_fpscr = 0; |
8aa2a305 | 5826 | |
1245df60 R |
5827 | /* Push PR last; this gives better latencies after the prologue, and |
5828 | candidates for the return delay slot when there are no general | |
5829 | registers pushed. */ | |
5c7cafa8 | 5830 | for (; i < FIRST_PSEUDO_REGISTER; i++) |
ac55736a R |
5831 | { |
5832 | /* If this is an interrupt handler, and the SZ bit varies, | |
5833 | and we have to push any floating point register, we need | |
5834 | to switch to the correct precision first. */ | |
5835 | if (i == FIRST_FP_REG && interrupt_handler && TARGET_FMOVD | |
56b138ae | 5836 | && hard_reg_set_intersect_p (*mask, reg_class_contents[DF_REGS])) |
ac55736a R |
5837 | { |
5838 | HARD_REG_SET unsaved; | |
5839 | ||
5840 | push (FPSCR_REG); | |
832a3292 | 5841 | COMPL_HARD_REG_SET (unsaved, *mask); |
ac55736a R |
5842 | fpscr_set_from_mem (NORMAL_MODE (FP_MODE), unsaved); |
5843 | skip_fpscr = 1; | |
5844 | } | |
5845 | if (i != PR_REG | |
5846 | && (i != FPSCR_REG || ! skip_fpscr) | |
5847 | && TEST_HARD_REG_BIT (*mask, i)) | |
561642fa AP |
5848 | { |
5849 | /* If the ISR has RESBANK attribute assigned, don't push any of | |
5850 | the following registers - R0-R14, MACH, MACL and GBR. */ | |
5851 | if (! (sh_cfun_resbank_handler_p () | |
5852 | && ((i >= FIRST_GENERAL_REG && i < LAST_GENERAL_REG) | |
5853 | || i == MACH_REG | |
5854 | || i == MACL_REG | |
5855 | || i == GBR_REG))) | |
5856 | push (i); | |
5857 | } | |
ac55736a | 5858 | } |
5c7cafa8 CB |
5859 | |
5860 | /* Push banked registers last to improve delay slot opportunities. */ | |
5861 | if (interrupt_handler) | |
5862 | for (i = FIRST_BANKED_REG; i <= LAST_BANKED_REG; i++) | |
5863 | if (TEST_HARD_REG_BIT (*mask, i)) | |
5864 | push (i); | |
5865 | ||
561642fa AP |
5866 | /* Don't push PR register for an ISR with RESBANK attribute assigned. */ |
5867 | if (TEST_HARD_REG_BIT (*mask, PR_REG) && !sh_cfun_resbank_handler_p ()) | |
1245df60 | 5868 | push (PR_REG); |
8aa2a305 JW |
5869 | } |
5870 | ||
fe3ad572 SC |
5871 | /* Calculate how much extra space is needed to save all callee-saved |
5872 | target registers. | |
5873 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
5874 | ||
5875 | static int | |
5876 | shmedia_target_regs_stack_space (HARD_REG_SET *live_regs_mask) | |
5877 | { | |
5878 | int reg; | |
5879 | int stack_space = 0; | |
5880 | int interrupt_handler = sh_cfun_interrupt_handler_p (); | |
5881 | ||
5882 | for (reg = LAST_TARGET_REG; reg >= FIRST_TARGET_REG; reg--) | |
62164eb4 | 5883 | if ((! call_really_used_regs[reg] || interrupt_handler) |
fe3ad572 SC |
5884 | && ! TEST_HARD_REG_BIT (*live_regs_mask, reg)) |
5885 | /* Leave space to save this target register on the stack, | |
24746a42 | 5886 | in case target register allocation wants to use it. */ |
fe3ad572 SC |
5887 | stack_space += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); |
5888 | return stack_space; | |
5889 | } | |
f676971a | 5890 | |
fe3ad572 SC |
5891 | /* Decide whether we should reserve space for callee-save target registers, |
5892 | in case target register allocation wants to use them. REGS_SAVED is | |
5893 | the space, in bytes, that is already required for register saves. | |
5894 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
5895 | ||
5896 | static int | |
5897 | shmedia_reserve_space_for_target_registers_p (int regs_saved, | |
5898 | HARD_REG_SET *live_regs_mask) | |
5899 | { | |
5900 | if (optimize_size) | |
5901 | return 0; | |
5902 | return shmedia_target_regs_stack_space (live_regs_mask) <= regs_saved; | |
5903 | } | |
5904 | ||
5905 | /* Decide how much space to reserve for callee-save target registers | |
5906 | in case target register allocation wants to use them. | |
5907 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
5908 | ||
5909 | static int | |
5910 | shmedia_target_regs_stack_adjust (HARD_REG_SET *live_regs_mask) | |
5911 | { | |
5912 | if (shmedia_space_reserved_for_target_registers) | |
5913 | return shmedia_target_regs_stack_space (live_regs_mask); | |
5914 | else | |
5915 | return 0; | |
5916 | } | |
5917 | ||
8aa2a305 | 5918 | /* Work out the registers which need to be saved, both as a mask and a |
490dbac7 | 5919 | count of saved words. Return the count. |
8aa2a305 JW |
5920 | |
5921 | If doing a pragma interrupt function, then push all regs used by the | |
5922 | function, and if we call another function (we can tell by looking at PR), | |
5923 | make sure that all the regs it clobbers are safe too. */ | |
5924 | ||
490dbac7 | 5925 | static int |
cf277499 | 5926 | calc_live_regs (HARD_REG_SET *live_regs_mask) |
8aa2a305 | 5927 | { |
0ef2c525 | 5928 | unsigned int reg; |
1245df60 | 5929 | int count; |
a6ab9fc0 R |
5930 | tree attrs; |
5931 | bool interrupt_or_trapa_handler, trapa_handler, interrupt_handler; | |
5932 | bool nosave_low_regs; | |
fada1961 | 5933 | int pr_live, has_call; |
157131d7 | 5934 | |
a6ab9fc0 R |
5935 | attrs = DECL_ATTRIBUTES (current_function_decl); |
5936 | interrupt_or_trapa_handler = sh_cfun_interrupt_handler_p (); | |
5937 | trapa_handler = lookup_attribute ("trapa_handler", attrs) != NULL_TREE; | |
5938 | interrupt_handler = interrupt_or_trapa_handler && ! trapa_handler; | |
5939 | nosave_low_regs = lookup_attribute ("nosave_low_regs", attrs) != NULL_TREE; | |
8aa2a305 | 5940 | |
fada1961 | 5941 | CLEAR_HARD_REG_SET (*live_regs_mask); |
157371cf | 5942 | if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && interrupt_handler |
6fb5fa3c | 5943 | && df_regs_ever_live_p (FPSCR_REG)) |
c0fb94d7 | 5944 | target_flags &= ~MASK_FPU_SINGLE; |
225e4f43 | 5945 | /* If we can save a lot of saves by switching to double mode, do that. */ |
157371cf | 5946 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && TARGET_FPU_SINGLE) |
225e4f43 | 5947 | for (count = 0, reg = FIRST_FP_REG; reg <= LAST_FP_REG; reg += 2) |
6fb5fa3c | 5948 | if (df_regs_ever_live_p (reg) && df_regs_ever_live_p (reg+1) |
62164eb4 | 5949 | && (! call_really_used_regs[reg] |
a6ab9fc0 | 5950 | || interrupt_handler) |
225e4f43 R |
5951 | && ++count > 2) |
5952 | { | |
c0fb94d7 | 5953 | target_flags &= ~MASK_FPU_SINGLE; |
225e4f43 R |
5954 | break; |
5955 | } | |
9d7ed806 R |
5956 | /* PR_MEDIA_REG is a general purpose register, thus global_alloc already |
5957 | knows how to use it. That means the pseudo originally allocated for | |
5958 | the initial value can become the PR_MEDIA_REG hard register, as seen for | |
5959 | execute/20010122-1.c:test9. */ | |
5960 | if (TARGET_SHMEDIA) | |
7d73a2ba R |
5961 | /* ??? this function is called from initial_elimination_offset, hence we |
5962 | can't use the result of sh_media_register_for_return here. */ | |
97b3d0f0 | 5963 | pr_live = sh_pr_n_sets (); |
9d7ed806 R |
5964 | else |
5965 | { | |
5966 | rtx pr_initial = has_hard_reg_initial_val (Pmode, PR_REG); | |
5967 | pr_live = (pr_initial | |
29db2a1e R |
5968 | ? (GET_CODE (pr_initial) != REG |
5969 | || REGNO (pr_initial) != (PR_REG)) | |
6fb5fa3c | 5970 | : df_regs_ever_live_p (PR_REG)); |
7d73a2ba R |
5971 | /* For Shcompact, if not optimizing, we end up with a memory reference |
5972 | using the return address pointer for __builtin_return_address even | |
5973 | though there is no actual need to put the PR register on the stack. */ | |
6fb5fa3c | 5974 | pr_live |= df_regs_ever_live_p (RETURN_ADDRESS_POINTER_REGNUM); |
9d7ed806 | 5975 | } |
fa5322fa AO |
5976 | /* Force PR to be live if the prologue has to call the SHmedia |
5977 | argument decoder or register saver. */ | |
5978 | if (TARGET_SHCOMPACT | |
38173d38 | 5979 | && ((crtl->args.info.call_cookie |
fa5322fa | 5980 | & ~ CALL_COOKIE_RET_TRAMP (1)) |
e3b5732b | 5981 | || crtl->saves_all_registers)) |
fa5322fa | 5982 | pr_live = 1; |
fada1961 | 5983 | has_call = TARGET_SHMEDIA ? ! leaf_function_p () : pr_live; |
0ef2c525 | 5984 | for (count = 0, reg = FIRST_PSEUDO_REGISTER; reg-- != 0; ) |
8aa2a305 | 5985 | { |
97b3d0f0 | 5986 | if (reg == (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG) |
3007d592 | 5987 | ? pr_live |
a6ab9fc0 | 5988 | : interrupt_handler |
1245df60 | 5989 | ? (/* Need to save all the regs ever live. */ |
6fb5fa3c | 5990 | (df_regs_ever_live_p (reg) |
62164eb4 R |
5991 | || (call_really_used_regs[reg] |
5992 | && (! fixed_regs[reg] || reg == MACH_REG || reg == MACL_REG | |
5993 | || reg == PIC_OFFSET_TABLE_REGNUM) | |
fada1961 | 5994 | && has_call) |
a6ab9fc0 R |
5995 | || (TARGET_SHMEDIA && has_call |
5996 | && REGISTER_NATURAL_MODE (reg) == SImode | |
fada1961 | 5997 | && (GENERAL_REGISTER_P (reg) || TARGET_REGISTER_P (reg)))) |
1245df60 R |
5998 | && reg != STACK_POINTER_REGNUM && reg != ARG_POINTER_REGNUM |
5999 | && reg != RETURN_ADDRESS_POINTER_REGNUM | |
7144b2d8 DD |
6000 | && reg != T_REG && reg != GBR_REG |
6001 | /* Push fpscr only on targets which have FPU */ | |
6002 | && (reg != FPSCR_REG || TARGET_FPU_ANY)) | |
1245df60 | 6003 | : (/* Only push those regs which are used and need to be saved. */ |
180bde4f RS |
6004 | (TARGET_SHCOMPACT |
6005 | && flag_pic | |
38173d38 | 6006 | && crtl->args.info.call_cookie |
0ef2c525 | 6007 | && reg == PIC_OFFSET_TABLE_REGNUM) |
6fb5fa3c | 6008 | || (df_regs_ever_live_p (reg) |
a6ab9fc0 R |
6009 | && (!call_really_used_regs[reg] |
6010 | || (trapa_handler && reg == FPSCR_REG && TARGET_FPU_ANY))) | |
e3b5732b | 6011 | || (crtl->calls_eh_return |
0ef2c525 KK |
6012 | && (reg == EH_RETURN_DATA_REGNO (0) |
6013 | || reg == EH_RETURN_DATA_REGNO (1) | |
6014 | || reg == EH_RETURN_DATA_REGNO (2) | |
6015 | || reg == EH_RETURN_DATA_REGNO (3))) | |
61f71b34 | 6016 | || ((reg == MACL_REG || reg == MACH_REG) |
6fb5fa3c | 6017 | && df_regs_ever_live_p (reg) |
61f71b34 DD |
6018 | && sh_cfun_attr_renesas_p ()) |
6019 | )) | |
8aa2a305 | 6020 | { |
490dbac7 | 6021 | SET_HARD_REG_BIT (*live_regs_mask, reg); |
fa5322fa AO |
6022 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); |
6023 | ||
157371cf | 6024 | if ((TARGET_SH4 || TARGET_SH2A_DOUBLE || TARGET_SH5) && TARGET_FMOVD |
fa5322fa | 6025 | && GET_MODE_CLASS (REGISTER_NATURAL_MODE (reg)) == MODE_FLOAT) |
eda44c78 | 6026 | { |
104ee20b | 6027 | if (FP_REGISTER_P (reg)) |
eda44c78 | 6028 | { |
6fb5fa3c | 6029 | if (! TARGET_FPU_SINGLE && ! df_regs_ever_live_p (reg ^ 1)) |
eda44c78 | 6030 | { |
490dbac7 | 6031 | SET_HARD_REG_BIT (*live_regs_mask, (reg ^ 1)); |
fa5322fa | 6032 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg ^ 1)); |
eda44c78 R |
6033 | } |
6034 | } | |
fa5322fa | 6035 | else if (XD_REGISTER_P (reg)) |
eda44c78 R |
6036 | { |
6037 | /* Must switch to double mode to access these registers. */ | |
c0fb94d7 | 6038 | target_flags &= ~MASK_FPU_SINGLE; |
eda44c78 R |
6039 | } |
6040 | } | |
8aa2a305 | 6041 | } |
a6ab9fc0 R |
6042 | if (nosave_low_regs && reg == R8_REG) |
6043 | break; | |
8aa2a305 | 6044 | } |
fe3ad572 SC |
6045 | /* If we have a target register optimization pass after prologue / epilogue |
6046 | threading, we need to assume all target registers will be live even if | |
6047 | they aren't now. */ | |
6048 | if (flag_branch_target_load_optimize2 | |
6049 | && TARGET_SAVE_ALL_TARGET_REGS | |
6050 | && shmedia_space_reserved_for_target_registers) | |
6051 | for (reg = LAST_TARGET_REG; reg >= FIRST_TARGET_REG; reg--) | |
62164eb4 | 6052 | if ((! call_really_used_regs[reg] || interrupt_handler) |
fe3ad572 SC |
6053 | && ! TEST_HARD_REG_BIT (*live_regs_mask, reg)) |
6054 | { | |
6055 | SET_HARD_REG_BIT (*live_regs_mask, reg); | |
6056 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); | |
6057 | } | |
fada1961 R |
6058 | /* If this is an interrupt handler, we don't have any call-clobbered |
6059 | registers we can conveniently use for target register save/restore. | |
6060 | Make sure we save at least one general purpose register when we need | |
6061 | to save target registers. */ | |
6062 | if (interrupt_handler | |
56b138ae RS |
6063 | && hard_reg_set_intersect_p (*live_regs_mask, |
6064 | reg_class_contents[TARGET_REGS]) | |
6065 | && ! hard_reg_set_intersect_p (*live_regs_mask, | |
6066 | reg_class_contents[GENERAL_REGS])) | |
fada1961 R |
6067 | { |
6068 | SET_HARD_REG_BIT (*live_regs_mask, R0_REG); | |
6069 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (R0_REG)); | |
6070 | } | |
8aa2a305 | 6071 | |
490dbac7 | 6072 | return count; |
8aa2a305 JW |
6073 | } |
6074 | ||
6075 | /* Code to generate prologue and epilogue sequences */ | |
b9654711 | 6076 | |
a93d1ba2 | 6077 | /* PUSHED is the number of bytes that are being pushed on the |
0ad913af R |
6078 | stack for register saves. Return the frame size, padded |
6079 | appropriately so that the stack stays properly aligned. */ | |
e5f8d4c0 | 6080 | static HOST_WIDE_INT |
cf277499 | 6081 | rounded_frame_size (int pushed) |
e5f8d4c0 R |
6082 | { |
6083 | HOST_WIDE_INT size = get_frame_size (); | |
6084 | HOST_WIDE_INT align = STACK_BOUNDARY / BITS_PER_UNIT; | |
6085 | ||
c5b9ef02 | 6086 | return ((size + pushed + align - 1) & -align) - pushed; |
e5f8d4c0 R |
6087 | } |
6088 | ||
fa5322fa AO |
6089 | /* Choose a call-clobbered target-branch register that remains |
6090 | unchanged along the whole function. We set it up as the return | |
6091 | value in the prologue. */ | |
6092 | int | |
cf277499 | 6093 | sh_media_register_for_return (void) |
fa5322fa AO |
6094 | { |
6095 | int regno; | |
6096 | int tr0_used; | |
6097 | ||
6098 | if (! current_function_is_leaf) | |
6099 | return -1; | |
fada1961 R |
6100 | if (lookup_attribute ("interrupt_handler", |
6101 | DECL_ATTRIBUTES (current_function_decl))) | |
6102 | return -1; | |
73a4d10b R |
6103 | if (sh_cfun_interrupt_handler_p ()) |
6104 | return -1; | |
fa5322fa | 6105 | |
6fb5fa3c | 6106 | tr0_used = flag_pic && df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM); |
fa5322fa AO |
6107 | |
6108 | for (regno = FIRST_TARGET_REG + tr0_used; regno <= LAST_TARGET_REG; regno++) | |
6fb5fa3c | 6109 | if (call_really_used_regs[regno] && ! df_regs_ever_live_p (regno)) |
fa5322fa AO |
6110 | return regno; |
6111 | ||
6112 | return -1; | |
6113 | } | |
6114 | ||
fada1961 R |
6115 | /* The maximum registers we need to save are: |
6116 | - 62 general purpose registers (r15 is stack pointer, r63 is zero) | |
6117 | - 32 floating point registers (for each pair, we save none, | |
6118 | one single precision value, or a double precision value). | |
6119 | - 8 target registers | |
6120 | - add 1 entry for a delimiter. */ | |
6121 | #define MAX_SAVED_REGS (62+32+8) | |
6122 | ||
6123 | typedef struct save_entry_s | |
6124 | { | |
6125 | unsigned char reg; | |
6126 | unsigned char mode; | |
6127 | short offset; | |
6128 | } save_entry; | |
6129 | ||
6130 | #define MAX_TEMPS 4 | |
6131 | ||
6132 | /* There will be a delimiter entry with VOIDmode both at the start and the | |
6133 | end of a filled in schedule. The end delimiter has the offset of the | |
6134 | save with the smallest (i.e. most negative) offset. */ | |
6135 | typedef struct save_schedule_s | |
6136 | { | |
6137 | save_entry entries[MAX_SAVED_REGS + 2]; | |
6138 | int temps[MAX_TEMPS+1]; | |
6139 | } save_schedule; | |
6140 | ||
6141 | /* Fill in SCHEDULE according to LIVE_REGS_MASK. If RESTORE is nonzero, | |
6142 | use reverse order. Returns the last entry written to (not counting | |
6143 | the delimiter). OFFSET_BASE is a number to be added to all offset | |
6144 | entries. */ | |
f676971a | 6145 | |
fada1961 R |
6146 | static save_entry * |
6147 | sh5_schedule_saves (HARD_REG_SET *live_regs_mask, save_schedule *schedule, | |
6148 | int offset_base) | |
6149 | { | |
6150 | int align, i; | |
6151 | save_entry *entry = schedule->entries; | |
6152 | int tmpx = 0; | |
6153 | int offset; | |
6154 | ||
6155 | if (! current_function_interrupt) | |
6156 | for (i = FIRST_GENERAL_REG; tmpx < MAX_TEMPS && i <= LAST_GENERAL_REG; i++) | |
62164eb4 | 6157 | if (call_really_used_regs[i] && ! fixed_regs[i] && i != PR_MEDIA_REG |
fada1961 R |
6158 | && ! FUNCTION_ARG_REGNO_P (i) |
6159 | && i != FIRST_RET_REG | |
6de9cd9a | 6160 | && ! (cfun->static_chain_decl != NULL && i == STATIC_CHAIN_REGNUM) |
e3b5732b | 6161 | && ! (crtl->calls_eh_return |
fada1961 | 6162 | && (i == EH_RETURN_STACKADJ_REGNO |
f94cef0a KK |
6163 | || ((unsigned) i >= EH_RETURN_DATA_REGNO (0) |
6164 | && (unsigned) i <= EH_RETURN_DATA_REGNO (3))))) | |
fada1961 R |
6165 | schedule->temps[tmpx++] = i; |
6166 | entry->reg = -1; | |
6167 | entry->mode = VOIDmode; | |
6168 | entry->offset = offset_base; | |
6169 | entry++; | |
6170 | /* We loop twice: first, we save 8-byte aligned registers in the | |
6171 | higher addresses, that are known to be aligned. Then, we | |
6172 | proceed to saving 32-bit registers that don't need 8-byte | |
6173 | alignment. | |
6174 | If this is an interrupt function, all registers that need saving | |
6175 | need to be saved in full. moreover, we need to postpone saving | |
6176 | target registers till we have saved some general purpose registers | |
6177 | we can then use as scratch registers. */ | |
6178 | offset = offset_base; | |
6179 | for (align = 1; align >= 0; align--) | |
6180 | { | |
6181 | for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; i--) | |
6182 | if (TEST_HARD_REG_BIT (*live_regs_mask, i)) | |
6183 | { | |
6184 | enum machine_mode mode = REGISTER_NATURAL_MODE (i); | |
6185 | int reg = i; | |
6186 | ||
6187 | if (current_function_interrupt) | |
6188 | { | |
6189 | if (TARGET_REGISTER_P (i)) | |
6190 | continue; | |
6191 | if (GENERAL_REGISTER_P (i)) | |
6192 | mode = DImode; | |
6193 | } | |
6194 | if (mode == SFmode && (i % 2) == 1 | |
6195 | && ! TARGET_FPU_SINGLE && FP_REGISTER_P (i) | |
6196 | && (TEST_HARD_REG_BIT (*live_regs_mask, (i ^ 1)))) | |
6197 | { | |
6198 | mode = DFmode; | |
6199 | i--; | |
6200 | reg--; | |
6201 | } | |
6202 | ||
6203 | /* If we're doing the aligned pass and this is not aligned, | |
6204 | or we're doing the unaligned pass and this is aligned, | |
6205 | skip it. */ | |
6206 | if ((GET_MODE_SIZE (mode) % (STACK_BOUNDARY / BITS_PER_UNIT) == 0) | |
6207 | != align) | |
6208 | continue; | |
6209 | ||
6210 | if (current_function_interrupt | |
6211 | && GENERAL_REGISTER_P (i) | |
6212 | && tmpx < MAX_TEMPS) | |
6213 | schedule->temps[tmpx++] = i; | |
6214 | ||
6215 | offset -= GET_MODE_SIZE (mode); | |
6216 | entry->reg = i; | |
6217 | entry->mode = mode; | |
6218 | entry->offset = offset; | |
6219 | entry++; | |
6220 | } | |
6221 | if (align && current_function_interrupt) | |
6222 | for (i = LAST_TARGET_REG; i >= FIRST_TARGET_REG; i--) | |
6223 | if (TEST_HARD_REG_BIT (*live_regs_mask, i)) | |
6224 | { | |
6225 | offset -= GET_MODE_SIZE (DImode); | |
6226 | entry->reg = i; | |
6227 | entry->mode = DImode; | |
6228 | entry->offset = offset; | |
6229 | entry++; | |
6230 | } | |
6231 | } | |
6232 | entry->reg = -1; | |
6233 | entry->mode = VOIDmode; | |
6234 | entry->offset = offset; | |
6235 | schedule->temps[tmpx] = -1; | |
6236 | return entry - 1; | |
6237 | } | |
6238 | ||
b9654711 | 6239 | void |
cf277499 | 6240 | sh_expand_prologue (void) |
b9654711 | 6241 | { |
490dbac7 | 6242 | HARD_REG_SET live_regs_mask; |
40d2032b | 6243 | int d, i; |
fa5322fa | 6244 | int d_rounding = 0; |
225e4f43 | 6245 | int save_flags = target_flags; |
49616835 | 6246 | int pretend_args; |
a6ab9fc0 R |
6247 | tree sp_switch_attr |
6248 | = lookup_attribute ("sp_switch", DECL_ATTRIBUTES (current_function_decl)); | |
b9654711 | 6249 | |
e3ba8d11 | 6250 | current_function_interrupt = sh_cfun_interrupt_handler_p (); |
552ecbd9 | 6251 | |
0d7e008e | 6252 | /* We have pretend args if we had an object sent partially in registers |
8aa2a305 | 6253 | and partially on the stack, e.g. a large structure. */ |
38173d38 | 6254 | pretend_args = crtl->args.pretend_args_size; |
49616835 JR |
6255 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl) |
6256 | && (NPARM_REGS(SImode) | |
38173d38 | 6257 | > crtl->args.info.arg_count[(int) SH_ARG_INT])) |
49616835 JR |
6258 | pretend_args = 0; |
6259 | output_stack_adjust (-pretend_args | |
38173d38 | 6260 | - crtl->args.info.stack_regs * 8, |
fada1961 | 6261 | stack_pointer_rtx, 0, NULL); |
b9654711 | 6262 | |
38173d38 | 6263 | if (TARGET_SHCOMPACT && flag_pic && crtl->args.info.call_cookie) |
2b1fd83f AO |
6264 | /* We're going to use the PIC register to load the address of the |
6265 | incoming-argument decoder and/or of the return trampoline from | |
6266 | the GOT, so make sure the PIC register is preserved and | |
6267 | initialized. */ | |
6fb5fa3c | 6268 | df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true); |
2b1fd83f | 6269 | |
fa5322fa | 6270 | if (TARGET_SHCOMPACT |
38173d38 | 6271 | && (crtl->args.info.call_cookie & ~ CALL_COOKIE_RET_TRAMP(1))) |
fa5322fa AO |
6272 | { |
6273 | int reg; | |
6274 | ||
6275 | /* First, make all registers with incoming arguments that will | |
6276 | be pushed onto the stack live, so that register renaming | |
6277 | doesn't overwrite them. */ | |
6278 | for (reg = 0; reg < NPARM_REGS (SImode); reg++) | |
38173d38 | 6279 | if (CALL_COOKIE_STACKSEQ_GET (crtl->args.info.call_cookie) |
fa5322fa AO |
6280 | >= NPARM_REGS (SImode) - reg) |
6281 | for (; reg < NPARM_REGS (SImode); reg++) | |
6282 | emit_insn (gen_shcompact_preserve_incoming_args | |
6283 | (gen_rtx_REG (SImode, FIRST_PARM_REG + reg))); | |
6284 | else if (CALL_COOKIE_INT_REG_GET | |
38173d38 | 6285 | (crtl->args.info.call_cookie, reg) == 1) |
fa5322fa AO |
6286 | emit_insn (gen_shcompact_preserve_incoming_args |
6287 | (gen_rtx_REG (SImode, FIRST_PARM_REG + reg))); | |
6288 | ||
6289 | emit_move_insn (gen_rtx_REG (Pmode, MACL_REG), | |
6290 | stack_pointer_rtx); | |
6291 | emit_move_insn (gen_rtx_REG (SImode, R0_REG), | |
38173d38 | 6292 | GEN_INT (crtl->args.info.call_cookie)); |
fa5322fa AO |
6293 | emit_move_insn (gen_rtx_REG (SImode, MACH_REG), |
6294 | gen_rtx_REG (SImode, R0_REG)); | |
6295 | } | |
6296 | else if (TARGET_SHMEDIA) | |
6297 | { | |
6298 | int tr = sh_media_register_for_return (); | |
6299 | ||
6300 | if (tr >= 0) | |
6fb5fa3c DB |
6301 | emit_move_insn (gen_rtx_REG (DImode, tr), |
6302 | gen_rtx_REG (DImode, PR_MEDIA_REG)); | |
fa5322fa AO |
6303 | } |
6304 | ||
7ab56274 | 6305 | /* Emit the code for SETUP_VARARGS. */ |
e3b5732b | 6306 | if (cfun->stdarg) |
b9654711 | 6307 | { |
49616835 | 6308 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl)) |
1245df60 | 6309 | { |
1a95a963 JW |
6310 | /* Push arg regs as if they'd been provided by caller in stack. */ |
6311 | for (i = 0; i < NPARM_REGS(SImode); i++) | |
6312 | { | |
6313 | int rn = NPARM_REGS(SImode) + FIRST_PARM_REG - i - 1; | |
2754d3c5 R |
6314 | rtx insn; |
6315 | ||
f676971a | 6316 | if (i >= (NPARM_REGS(SImode) |
38173d38 | 6317 | - crtl->args.info.arg_count[(int) SH_ARG_INT] |
aec373f1 | 6318 | )) |
1a95a963 | 6319 | break; |
2754d3c5 | 6320 | insn = push (rn); |
1a95a963 | 6321 | } |
1245df60 | 6322 | } |
b9654711 | 6323 | } |
1a95a963 | 6324 | |
4408efce | 6325 | /* If we're supposed to switch stacks at function entry, do so now. */ |
a6ab9fc0 R |
6326 | if (sp_switch_attr) |
6327 | { | |
6328 | /* The argument specifies a variable holding the address of the | |
6329 | stack the interrupt function should switch to/from at entry/exit. */ | |
6330 | const char *s | |
6331 | = ggc_strdup (TREE_STRING_POINTER (TREE_VALUE (sp_switch_attr))); | |
6332 | rtx sp_switch = gen_rtx_SYMBOL_REF (Pmode, s); | |
6333 | ||
6334 | emit_insn (gen_sp_switch_1 (sp_switch)); | |
6335 | } | |
4408efce | 6336 | |
490dbac7 | 6337 | d = calc_live_regs (&live_regs_mask); |
225e4f43 R |
6338 | /* ??? Maybe we could save some switching if we can move a mode switch |
6339 | that already happens to be at the function start into the prologue. */ | |
ac55736a | 6340 | if (target_flags != save_flags && ! current_function_interrupt) |
225e4f43 | 6341 | emit_insn (gen_toggle_sz ()); |
f676971a | 6342 | |
fa5322fa AO |
6343 | if (TARGET_SH5) |
6344 | { | |
fada1961 R |
6345 | int offset_base, offset; |
6346 | rtx r0 = NULL_RTX; | |
fa5322fa AO |
6347 | int offset_in_r0 = -1; |
6348 | int sp_in_r0 = 0; | |
fe3ad572 SC |
6349 | int tregs_space = shmedia_target_regs_stack_adjust (&live_regs_mask); |
6350 | int total_size, save_size; | |
fada1961 R |
6351 | save_schedule schedule; |
6352 | save_entry *entry; | |
6353 | int *tmp_pnt; | |
6354 | ||
62164eb4 | 6355 | if (call_really_used_regs[R0_REG] && ! fixed_regs[R0_REG] |
fada1961 R |
6356 | && ! current_function_interrupt) |
6357 | r0 = gen_rtx_REG (Pmode, R0_REG); | |
fe3ad572 SC |
6358 | |
6359 | /* D is the actual number of bytes that we need for saving registers, | |
6360 | however, in initial_elimination_offset we have committed to using | |
6361 | an additional TREGS_SPACE amount of bytes - in order to keep both | |
6362 | addresses to arguments supplied by the caller and local variables | |
6363 | valid, we must keep this gap. Place it between the incoming | |
6364 | arguments and the actually saved registers in a bid to optimize | |
6365 | locality of reference. */ | |
6366 | total_size = d + tregs_space; | |
6367 | total_size += rounded_frame_size (total_size); | |
6368 | save_size = total_size - rounded_frame_size (d); | |
6369 | if (save_size % (STACK_BOUNDARY / BITS_PER_UNIT)) | |
fa5322fa | 6370 | d_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) |
fe3ad572 SC |
6371 | - save_size % (STACK_BOUNDARY / BITS_PER_UNIT)); |
6372 | ||
6373 | /* If adjusting the stack in a single step costs nothing extra, do so. | |
6374 | I.e. either if a single addi is enough, or we need a movi anyway, | |
6375 | and we don't exceed the maximum offset range (the test for the | |
6376 | latter is conservative for simplicity). */ | |
6377 | if (TARGET_SHMEDIA | |
6378 | && (CONST_OK_FOR_I10 (-total_size) | |
6379 | || (! CONST_OK_FOR_I10 (-(save_size + d_rounding)) | |
6380 | && total_size <= 2044))) | |
6381 | d_rounding = total_size - save_size; | |
fa5322fa | 6382 | |
fada1961 | 6383 | offset_base = d + d_rounding; |
fe3ad572 SC |
6384 | |
6385 | output_stack_adjust (-(save_size + d_rounding), stack_pointer_rtx, | |
fada1961 | 6386 | 0, NULL); |
fa5322fa | 6387 | |
fada1961 R |
6388 | sh5_schedule_saves (&live_regs_mask, &schedule, offset_base); |
6389 | tmp_pnt = schedule.temps; | |
6390 | for (entry = &schedule.entries[1]; entry->mode != VOIDmode; entry++) | |
6391 | { | |
6392 | enum machine_mode mode = entry->mode; | |
0ef2c525 | 6393 | unsigned int reg = entry->reg; |
fada1961 | 6394 | rtx reg_rtx, mem_rtx, pre_dec = NULL_RTX; |
455786b6 | 6395 | rtx orig_reg_rtx; |
fa5322fa | 6396 | |
fada1961 | 6397 | offset = entry->offset; |
fa5322fa | 6398 | |
fada1961 | 6399 | reg_rtx = gen_rtx_REG (mode, reg); |
fa5322fa | 6400 | |
57d38024 R |
6401 | mem_rtx = gen_frame_mem (mode, |
6402 | gen_rtx_PLUS (Pmode, | |
6403 | stack_pointer_rtx, | |
6404 | GEN_INT (offset))); | |
fa5322fa | 6405 | |
fada1961 | 6406 | GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (mem_rtx, 0), try_pre_dec); |
fa5322fa | 6407 | |
f5b9e7c9 | 6408 | gcc_assert (r0); |
fada1961 | 6409 | mem_rtx = NULL_RTX; |
fa5322fa | 6410 | |
fada1961 R |
6411 | try_pre_dec: |
6412 | do | |
6413 | if (HAVE_PRE_DECREMENT | |
6414 | && (offset_in_r0 - offset == GET_MODE_SIZE (mode) | |
6415 | || mem_rtx == NULL_RTX | |
6416 | || reg == PR_REG || SPECIAL_REGISTER_P (reg))) | |
6417 | { | |
57d38024 | 6418 | pre_dec = gen_frame_mem (mode, gen_rtx_PRE_DEC (Pmode, r0)); |
fa5322fa | 6419 | |
fada1961 R |
6420 | GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (pre_dec, 0), |
6421 | pre_dec_ok); | |
fa5322fa | 6422 | |
fada1961 | 6423 | pre_dec = NULL_RTX; |
fa5322fa | 6424 | |
fada1961 | 6425 | break; |
fa5322fa | 6426 | |
fada1961 R |
6427 | pre_dec_ok: |
6428 | mem_rtx = NULL_RTX; | |
6429 | offset += GET_MODE_SIZE (mode); | |
6430 | } | |
6431 | while (0); | |
fa5322fa | 6432 | |
fada1961 R |
6433 | if (mem_rtx != NULL_RTX) |
6434 | goto addr_ok; | |
fa5322fa | 6435 | |
fada1961 R |
6436 | if (offset_in_r0 == -1) |
6437 | { | |
6438 | emit_move_insn (r0, GEN_INT (offset)); | |
6439 | offset_in_r0 = offset; | |
6440 | } | |
6441 | else if (offset != offset_in_r0) | |
6442 | { | |
6443 | emit_move_insn (r0, | |
6444 | gen_rtx_PLUS | |
6445 | (Pmode, r0, | |
6446 | GEN_INT (offset - offset_in_r0))); | |
6447 | offset_in_r0 += offset - offset_in_r0; | |
6448 | } | |
f676971a | 6449 | |
fada1961 R |
6450 | if (pre_dec != NULL_RTX) |
6451 | { | |
6452 | if (! sp_in_r0) | |
fa5322fa AO |
6453 | { |
6454 | emit_move_insn (r0, | |
6455 | gen_rtx_PLUS | |
fada1961 R |
6456 | (Pmode, r0, stack_pointer_rtx)); |
6457 | sp_in_r0 = 1; | |
fa5322fa | 6458 | } |
fa5322fa | 6459 | |
fada1961 R |
6460 | offset -= GET_MODE_SIZE (mode); |
6461 | offset_in_r0 -= GET_MODE_SIZE (mode); | |
fa5322fa | 6462 | |
fada1961 R |
6463 | mem_rtx = pre_dec; |
6464 | } | |
6465 | else if (sp_in_r0) | |
57d38024 | 6466 | mem_rtx = gen_frame_mem (mode, r0); |
fada1961 | 6467 | else |
57d38024 R |
6468 | mem_rtx = gen_frame_mem (mode, |
6469 | gen_rtx_PLUS (Pmode, | |
6470 | stack_pointer_rtx, | |
6471 | r0)); | |
fada1961 R |
6472 | |
6473 | /* We must not use an r0-based address for target-branch | |
6474 | registers or for special registers without pre-dec | |
6475 | memory addresses, since we store their values in r0 | |
6476 | first. */ | |
f5b9e7c9 NS |
6477 | gcc_assert (!TARGET_REGISTER_P (reg) |
6478 | && ((reg != PR_REG && !SPECIAL_REGISTER_P (reg)) | |
6479 | || mem_rtx == pre_dec)); | |
6480 | ||
fada1961 | 6481 | addr_ok: |
455786b6 | 6482 | orig_reg_rtx = reg_rtx; |
fada1961 R |
6483 | if (TARGET_REGISTER_P (reg) |
6484 | || ((reg == PR_REG || SPECIAL_REGISTER_P (reg)) | |
6485 | && mem_rtx != pre_dec)) | |
6486 | { | |
6487 | rtx tmp_reg = gen_rtx_REG (GET_MODE (reg_rtx), *tmp_pnt); | |
fa5322fa | 6488 | |
fada1961 | 6489 | emit_move_insn (tmp_reg, reg_rtx); |
fa5322fa | 6490 | |
fada1961 R |
6491 | if (REGNO (tmp_reg) == R0_REG) |
6492 | { | |
fa5322fa AO |
6493 | offset_in_r0 = -1; |
6494 | sp_in_r0 = 0; | |
f5b9e7c9 NS |
6495 | gcc_assert (!refers_to_regno_p |
6496 | (R0_REG, R0_REG+1, mem_rtx, (rtx *) 0)); | |
fa5322fa AO |
6497 | } |
6498 | ||
fada1961 R |
6499 | if (*++tmp_pnt <= 0) |
6500 | tmp_pnt = schedule.temps; | |
6501 | ||
6502 | reg_rtx = tmp_reg; | |
fa5322fa | 6503 | } |
fada1961 R |
6504 | { |
6505 | rtx insn; | |
6506 | ||
6507 | /* Mark as interesting for dwarf cfi generator */ | |
6508 | insn = emit_move_insn (mem_rtx, reg_rtx); | |
6509 | RTX_FRAME_RELATED_P (insn) = 1; | |
455786b6 KK |
6510 | /* If we use an intermediate register for the save, we can't |
6511 | describe this exactly in cfi as a copy of the to-be-saved | |
6512 | register into the temporary register and then the temporary | |
6513 | register on the stack, because the temporary register can | |
6514 | have a different natural size than the to-be-saved register. | |
6515 | Thus, we gloss over the intermediate copy and pretend we do | |
6516 | a direct save from the to-be-saved register. */ | |
6517 | if (REGNO (reg_rtx) != reg) | |
6518 | { | |
6519 | rtx set, note_rtx; | |
6520 | ||
6521 | set = gen_rtx_SET (VOIDmode, mem_rtx, orig_reg_rtx); | |
6522 | note_rtx = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, set, | |
6523 | REG_NOTES (insn)); | |
6524 | REG_NOTES (insn) = note_rtx; | |
6525 | } | |
fada1961 | 6526 | |
f676971a | 6527 | if (TARGET_SHCOMPACT && (offset_in_r0 != -1)) |
fada1961 R |
6528 | { |
6529 | rtx reg_rtx = gen_rtx_REG (mode, reg); | |
6530 | rtx set, note_rtx; | |
57d38024 R |
6531 | rtx mem_rtx = gen_frame_mem (mode, |
6532 | gen_rtx_PLUS (Pmode, | |
6533 | stack_pointer_rtx, | |
6534 | GEN_INT (offset))); | |
fada1961 R |
6535 | |
6536 | set = gen_rtx_SET (VOIDmode, mem_rtx, reg_rtx); | |
6537 | note_rtx = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, set, | |
6538 | REG_NOTES (insn)); | |
6539 | REG_NOTES (insn) = note_rtx; | |
6540 | } | |
6541 | } | |
6542 | } | |
fa5322fa | 6543 | |
f5b9e7c9 | 6544 | gcc_assert (entry->offset == d_rounding); |
fa5322fa AO |
6545 | } |
6546 | else | |
ac55736a | 6547 | push_regs (&live_regs_mask, current_function_interrupt); |
1a66cd67 | 6548 | |
6fb5fa3c DB |
6549 | if (flag_pic && df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM)) |
6550 | emit_insn (gen_GOTaddr2picreg ()); | |
1a66cd67 | 6551 | |
fa5322fa AO |
6552 | if (SHMEDIA_REGS_STACK_ADJUST ()) |
6553 | { | |
fa5322fa AO |
6554 | /* This must NOT go through the PLT, otherwise mach and macl |
6555 | may be clobbered. */ | |
73a4d10b R |
6556 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
6557 | (TARGET_FPU_ANY | |
6558 | ? "__GCC_push_shmedia_regs" | |
6559 | : "__GCC_push_shmedia_regs_nofpu"), SFUNC_GOT); | |
fa5322fa AO |
6560 | emit_insn (gen_shmedia_save_restore_regs_compact |
6561 | (GEN_INT (-SHMEDIA_REGS_STACK_ADJUST ()))); | |
6562 | } | |
6563 | ||
ac55736a | 6564 | if (target_flags != save_flags && ! current_function_interrupt) |
6fb5fa3c | 6565 | emit_insn (gen_toggle_sz ()); |
1a95a963 | 6566 | |
225e4f43 R |
6567 | target_flags = save_flags; |
6568 | ||
fa5322fa | 6569 | output_stack_adjust (-rounded_frame_size (d) + d_rounding, |
fada1961 | 6570 | stack_pointer_rtx, 0, NULL); |
b9654711 SC |
6571 | |
6572 | if (frame_pointer_needed) | |
96a2347e | 6573 | frame_insn (GEN_MOV (hard_frame_pointer_rtx, stack_pointer_rtx)); |
fa5322fa | 6574 | |
fa5322fa | 6575 | if (TARGET_SHCOMPACT |
38173d38 | 6576 | && (crtl->args.info.call_cookie & ~ CALL_COOKIE_RET_TRAMP(1))) |
fa5322fa AO |
6577 | { |
6578 | /* This must NOT go through the PLT, otherwise mach and macl | |
6579 | may be clobbered. */ | |
73a4d10b R |
6580 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
6581 | "__GCC_shcompact_incoming_args", SFUNC_GOT); | |
fa5322fa AO |
6582 | emit_insn (gen_shcompact_incoming_args ()); |
6583 | } | |
b9654711 SC |
6584 | } |
6585 | ||
6586 | void | |
726d4cb7 | 6587 | sh_expand_epilogue (bool sibcall_p) |
b9654711 | 6588 | { |
490dbac7 | 6589 | HARD_REG_SET live_regs_mask; |
40d2032b | 6590 | int d, i; |
fa5322fa | 6591 | int d_rounding = 0; |
b9654711 | 6592 | |
225e4f43 | 6593 | int save_flags = target_flags; |
fe3ad572 | 6594 | int frame_size, save_size; |
ac55736a | 6595 | int fpscr_deferred = 0; |
726d4cb7 | 6596 | int e = sibcall_p ? -1 : 1; |
3d5a0820 | 6597 | |
490dbac7 | 6598 | d = calc_live_regs (&live_regs_mask); |
3d5a0820 | 6599 | |
fe3ad572 SC |
6600 | save_size = d; |
6601 | frame_size = rounded_frame_size (d); | |
6602 | ||
6603 | if (TARGET_SH5) | |
6604 | { | |
6605 | int tregs_space = shmedia_target_regs_stack_adjust (&live_regs_mask); | |
6606 | int total_size; | |
6607 | if (d % (STACK_BOUNDARY / BITS_PER_UNIT)) | |
6608 | d_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) | |
6609 | - d % (STACK_BOUNDARY / BITS_PER_UNIT)); | |
6610 | ||
6611 | total_size = d + tregs_space; | |
6612 | total_size += rounded_frame_size (total_size); | |
6613 | save_size = total_size - frame_size; | |
6614 | ||
6615 | /* If adjusting the stack in a single step costs nothing extra, do so. | |
6616 | I.e. either if a single addi is enough, or we need a movi anyway, | |
6617 | and we don't exceed the maximum offset range (the test for the | |
6618 | latter is conservative for simplicity). */ | |
6619 | if (TARGET_SHMEDIA | |
6620 | && ! frame_pointer_needed | |
6621 | && (CONST_OK_FOR_I10 (total_size) | |
6622 | || (! CONST_OK_FOR_I10 (save_size + d_rounding) | |
6623 | && total_size <= 2044))) | |
6624 | d_rounding = frame_size; | |
fa5322fa | 6625 | |
fe3ad572 SC |
6626 | frame_size -= d_rounding; |
6627 | } | |
16bea517 | 6628 | |
b9654711 | 6629 | if (frame_pointer_needed) |
46d81ffa | 6630 | { |
d0edbeec KK |
6631 | /* We must avoid scheduling the epilogue with previous basic blocks |
6632 | when exception handling is enabled. See PR/18032. */ | |
6633 | if (flag_exceptions) | |
6634 | emit_insn (gen_blockage ()); | |
96a2347e KK |
6635 | output_stack_adjust (frame_size, hard_frame_pointer_rtx, e, |
6636 | &live_regs_mask); | |
07f5b9aa JL |
6637 | |
6638 | /* We must avoid moving the stack pointer adjustment past code | |
6639 | which reads from the local frame, else an interrupt could | |
6640 | occur after the SP adjustment and clobber data in the local | |
6641 | frame. */ | |
6642 | emit_insn (gen_blockage ()); | |
96a2347e | 6643 | emit_insn (GEN_MOV (stack_pointer_rtx, hard_frame_pointer_rtx)); |
46d81ffa | 6644 | } |
3d5a0820 | 6645 | else if (frame_size) |
07f5b9aa JL |
6646 | { |
6647 | /* We must avoid moving the stack pointer adjustment past code | |
6648 | which reads from the local frame, else an interrupt could | |
6649 | occur after the SP adjustment and clobber data in the local | |
6650 | frame. */ | |
6651 | emit_insn (gen_blockage ()); | |
726d4cb7 | 6652 | output_stack_adjust (frame_size, stack_pointer_rtx, e, &live_regs_mask); |
07f5b9aa | 6653 | } |
b9654711 | 6654 | |
fa5322fa AO |
6655 | if (SHMEDIA_REGS_STACK_ADJUST ()) |
6656 | { | |
73a4d10b R |
6657 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
6658 | (TARGET_FPU_ANY | |
6659 | ? "__GCC_pop_shmedia_regs" | |
6660 | : "__GCC_pop_shmedia_regs_nofpu"), SFUNC_GOT); | |
fa5322fa AO |
6661 | /* This must NOT go through the PLT, otherwise mach and macl |
6662 | may be clobbered. */ | |
6663 | emit_insn (gen_shmedia_save_restore_regs_compact | |
6664 | (GEN_INT (SHMEDIA_REGS_STACK_ADJUST ()))); | |
6665 | } | |
6666 | ||
16bea517 | 6667 | /* Pop all the registers. */ |
0d7e008e | 6668 | |
ac55736a | 6669 | if (target_flags != save_flags && ! current_function_interrupt) |
225e4f43 | 6670 | emit_insn (gen_toggle_sz ()); |
fa5322fa AO |
6671 | if (TARGET_SH5) |
6672 | { | |
fada1961 | 6673 | int offset_base, offset; |
fa5322fa AO |
6674 | int offset_in_r0 = -1; |
6675 | int sp_in_r0 = 0; | |
fa5322fa | 6676 | rtx r0 = gen_rtx_REG (Pmode, R0_REG); |
fada1961 R |
6677 | save_schedule schedule; |
6678 | save_entry *entry; | |
6679 | int *tmp_pnt; | |
7d73a2ba | 6680 | |
fada1961 R |
6681 | entry = sh5_schedule_saves (&live_regs_mask, &schedule, d_rounding); |
6682 | offset_base = -entry[1].offset + d_rounding; | |
6683 | tmp_pnt = schedule.temps; | |
6684 | for (; entry->mode != VOIDmode; entry--) | |
6685 | { | |
6686 | enum machine_mode mode = entry->mode; | |
6687 | int reg = entry->reg; | |
6688 | rtx reg_rtx, mem_rtx, post_inc = NULL_RTX, insn; | |
fa5322fa | 6689 | |
fada1961 R |
6690 | offset = offset_base + entry->offset; |
6691 | reg_rtx = gen_rtx_REG (mode, reg); | |
fa5322fa | 6692 | |
57d38024 R |
6693 | mem_rtx = gen_frame_mem (mode, |
6694 | gen_rtx_PLUS (Pmode, | |
6695 | stack_pointer_rtx, | |
6696 | GEN_INT (offset))); | |
fa5322fa | 6697 | |
fada1961 | 6698 | GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (mem_rtx, 0), try_post_inc); |
fa5322fa | 6699 | |
fada1961 | 6700 | mem_rtx = NULL_RTX; |
fa5322fa | 6701 | |
fada1961 R |
6702 | try_post_inc: |
6703 | do | |
6704 | if (HAVE_POST_INCREMENT | |
6705 | && (offset == offset_in_r0 | |
6706 | || (offset + GET_MODE_SIZE (mode) != d + d_rounding | |
6707 | && mem_rtx == NULL_RTX) | |
6708 | || reg == PR_REG || SPECIAL_REGISTER_P (reg))) | |
6709 | { | |
57d38024 | 6710 | post_inc = gen_frame_mem (mode, gen_rtx_POST_INC (Pmode, r0)); |
fa5322fa | 6711 | |
fada1961 R |
6712 | GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (post_inc, 0), |
6713 | post_inc_ok); | |
fa5322fa | 6714 | |
fada1961 | 6715 | post_inc = NULL_RTX; |
fa5322fa | 6716 | |
fada1961 | 6717 | break; |
f676971a | 6718 | |
fada1961 R |
6719 | post_inc_ok: |
6720 | mem_rtx = NULL_RTX; | |
6721 | } | |
6722 | while (0); | |
f676971a | 6723 | |
fada1961 R |
6724 | if (mem_rtx != NULL_RTX) |
6725 | goto addr_ok; | |
fa5322fa | 6726 | |
fada1961 R |
6727 | if (offset_in_r0 == -1) |
6728 | { | |
6729 | emit_move_insn (r0, GEN_INT (offset)); | |
6730 | offset_in_r0 = offset; | |
6731 | } | |
6732 | else if (offset != offset_in_r0) | |
6733 | { | |
6734 | emit_move_insn (r0, | |
6735 | gen_rtx_PLUS | |
6736 | (Pmode, r0, | |
6737 | GEN_INT (offset - offset_in_r0))); | |
6738 | offset_in_r0 += offset - offset_in_r0; | |
6739 | } | |
f676971a | 6740 | |
fada1961 R |
6741 | if (post_inc != NULL_RTX) |
6742 | { | |
6743 | if (! sp_in_r0) | |
fa5322fa AO |
6744 | { |
6745 | emit_move_insn (r0, | |
6746 | gen_rtx_PLUS | |
fada1961 R |
6747 | (Pmode, r0, stack_pointer_rtx)); |
6748 | sp_in_r0 = 1; | |
fa5322fa | 6749 | } |
f676971a | 6750 | |
fada1961 | 6751 | mem_rtx = post_inc; |
fa5322fa | 6752 | |
fada1961 R |
6753 | offset_in_r0 += GET_MODE_SIZE (mode); |
6754 | } | |
6755 | else if (sp_in_r0) | |
57d38024 | 6756 | mem_rtx = gen_frame_mem (mode, r0); |
fada1961 | 6757 | else |
57d38024 R |
6758 | mem_rtx = gen_frame_mem (mode, |
6759 | gen_rtx_PLUS (Pmode, | |
6760 | stack_pointer_rtx, | |
6761 | r0)); | |
fa5322fa | 6762 | |
f5b9e7c9 NS |
6763 | gcc_assert ((reg != PR_REG && !SPECIAL_REGISTER_P (reg)) |
6764 | || mem_rtx == post_inc); | |
fa5322fa | 6765 | |
fada1961 R |
6766 | addr_ok: |
6767 | if ((reg == PR_REG || SPECIAL_REGISTER_P (reg)) | |
6768 | && mem_rtx != post_inc) | |
6769 | { | |
6770 | insn = emit_move_insn (r0, mem_rtx); | |
6771 | mem_rtx = r0; | |
6772 | } | |
6773 | else if (TARGET_REGISTER_P (reg)) | |
6774 | { | |
6775 | rtx tmp_reg = gen_rtx_REG (mode, *tmp_pnt); | |
6776 | ||
6777 | /* Give the scheduler a bit of freedom by using up to | |
6778 | MAX_TEMPS registers in a round-robin fashion. */ | |
6779 | insn = emit_move_insn (tmp_reg, mem_rtx); | |
6780 | mem_rtx = tmp_reg; | |
6781 | if (*++tmp_pnt < 0) | |
6782 | tmp_pnt = schedule.temps; | |
fa5322fa AO |
6783 | } |
6784 | ||
fada1961 | 6785 | insn = emit_move_insn (reg_rtx, mem_rtx); |
fada1961 R |
6786 | } |
6787 | ||
f5b9e7c9 | 6788 | gcc_assert (entry->offset + offset_base == d + d_rounding); |
fa5322fa | 6789 | } |
fe3ad572 | 6790 | else /* ! TARGET_SH5 */ |
b9654711 | 6791 | { |
5c7cafa8 CB |
6792 | int last_reg; |
6793 | ||
fe3ad572 | 6794 | save_size = 0; |
561642fa AP |
6795 | /* For an ISR with RESBANK attribute assigned, don't pop PR |
6796 | register. */ | |
6797 | if (TEST_HARD_REG_BIT (live_regs_mask, PR_REG) | |
6798 | && !sh_cfun_resbank_handler_p ()) | |
6fb5fa3c DB |
6799 | { |
6800 | if (!frame_pointer_needed) | |
6801 | emit_insn (gen_blockage ()); | |
6802 | pop (PR_REG); | |
6803 | } | |
5c7cafa8 CB |
6804 | |
6805 | /* Banked registers are poped first to avoid being scheduled in the | |
6806 | delay slot. RTE switches banks before the ds instruction. */ | |
6807 | if (current_function_interrupt) | |
6808 | { | |
6809 | for (i = FIRST_BANKED_REG; i <= LAST_BANKED_REG; i++) | |
6810 | if (TEST_HARD_REG_BIT (live_regs_mask, i)) | |
6811 | pop (LAST_BANKED_REG - i); | |
6812 | ||
6813 | last_reg = FIRST_PSEUDO_REGISTER - LAST_BANKED_REG - 1; | |
6814 | } | |
6815 | else | |
6816 | last_reg = FIRST_PSEUDO_REGISTER; | |
6817 | ||
6818 | for (i = 0; i < last_reg; i++) | |
fe3ad572 SC |
6819 | { |
6820 | int j = (FIRST_PSEUDO_REGISTER - 1) - i; | |
f676971a | 6821 | |
fe3ad572 | 6822 | if (j == FPSCR_REG && current_function_interrupt && TARGET_FMOVD |
56b138ae RS |
6823 | && hard_reg_set_intersect_p (live_regs_mask, |
6824 | reg_class_contents[DF_REGS])) | |
fe3ad572 | 6825 | fpscr_deferred = 1; |
561642fa AP |
6826 | /* For an ISR with RESBANK attribute assigned, don't pop |
6827 | following registers, R0-R14, MACH, MACL and GBR. */ | |
6828 | else if (j != PR_REG && TEST_HARD_REG_BIT (live_regs_mask, j) | |
6829 | && ! (sh_cfun_resbank_handler_p () | |
6830 | && ((j >= FIRST_GENERAL_REG | |
6831 | && j < LAST_GENERAL_REG) | |
6832 | || j == MACH_REG | |
6833 | || j == MACL_REG | |
6834 | || j == GBR_REG))) | |
fe3ad572 | 6835 | pop (j); |
5c7cafa8 | 6836 | |
fe3ad572 SC |
6837 | if (j == FIRST_FP_REG && fpscr_deferred) |
6838 | pop (FPSCR_REG); | |
fe3ad572 | 6839 | } |
b9654711 | 6840 | } |
ac55736a | 6841 | if (target_flags != save_flags && ! current_function_interrupt) |
225e4f43 R |
6842 | emit_insn (gen_toggle_sz ()); |
6843 | target_flags = save_flags; | |
b9654711 | 6844 | |
38173d38 | 6845 | output_stack_adjust (crtl->args.pretend_args_size |
fe3ad572 | 6846 | + save_size + d_rounding |
38173d38 | 6847 | + crtl->args.info.stack_regs * 8, |
726d4cb7 | 6848 | stack_pointer_rtx, e, NULL); |
4408efce | 6849 | |
e3b5732b | 6850 | if (crtl->calls_eh_return) |
4977bab6 ZW |
6851 | emit_insn (GEN_ADD3 (stack_pointer_rtx, stack_pointer_rtx, |
6852 | EH_RETURN_STACKADJ_RTX)); | |
6853 | ||
4408efce | 6854 | /* Switch back to the normal stack if necessary. */ |
a6ab9fc0 | 6855 | if (lookup_attribute ("sp_switch", DECL_ATTRIBUTES (current_function_decl))) |
4408efce | 6856 | emit_insn (gen_sp_switch_2 ()); |
89d4bc35 AO |
6857 | |
6858 | /* Tell flow the insn that pops PR isn't dead. */ | |
fa5322fa AO |
6859 | /* PR_REG will never be live in SHmedia mode, and we don't need to |
6860 | USE PR_MEDIA_REG, since it will be explicitly copied to TR0_REG | |
6861 | by the return pattern. */ | |
490dbac7 | 6862 | if (TEST_HARD_REG_BIT (live_regs_mask, PR_REG)) |
c41c1387 | 6863 | emit_use (gen_rtx_REG (SImode, PR_REG)); |
b9654711 SC |
6864 | } |
6865 | ||
afbc2905 R |
6866 | static int sh_need_epilogue_known = 0; |
6867 | ||
6868 | int | |
cf277499 | 6869 | sh_need_epilogue (void) |
afbc2905 R |
6870 | { |
6871 | if (! sh_need_epilogue_known) | |
6872 | { | |
6873 | rtx epilogue; | |
6874 | ||
6875 | start_sequence (); | |
726d4cb7 | 6876 | sh_expand_epilogue (0); |
2f937369 | 6877 | epilogue = get_insns (); |
afbc2905 | 6878 | end_sequence (); |
2f937369 | 6879 | sh_need_epilogue_known = (epilogue == NULL ? -1 : 1); |
afbc2905 R |
6880 | } |
6881 | return sh_need_epilogue_known > 0; | |
6882 | } | |
6883 | ||
4977bab6 ZW |
6884 | /* Emit code to change the current function's return address to RA. |
6885 | TEMP is available as a scratch register, if needed. */ | |
6886 | ||
6887 | void | |
cf277499 | 6888 | sh_set_return_address (rtx ra, rtx tmp) |
4977bab6 | 6889 | { |
490dbac7 | 6890 | HARD_REG_SET live_regs_mask; |
4977bab6 | 6891 | int d; |
4977bab6 ZW |
6892 | int pr_reg = TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG; |
6893 | int pr_offset; | |
6894 | ||
490dbac7 | 6895 | d = calc_live_regs (&live_regs_mask); |
4977bab6 ZW |
6896 | |
6897 | /* If pr_reg isn't life, we can set it (or the register given in | |
6898 | sh_media_register_for_return) directly. */ | |
490dbac7 | 6899 | if (! TEST_HARD_REG_BIT (live_regs_mask, pr_reg)) |
4977bab6 ZW |
6900 | { |
6901 | rtx rr; | |
6902 | ||
6903 | if (TARGET_SHMEDIA) | |
6904 | { | |
6905 | int rr_regno = sh_media_register_for_return (); | |
6906 | ||
6907 | if (rr_regno < 0) | |
6908 | rr_regno = pr_reg; | |
6909 | ||
6910 | rr = gen_rtx_REG (DImode, rr_regno); | |
6911 | } | |
6912 | else | |
6913 | rr = gen_rtx_REG (SImode, pr_reg); | |
6914 | ||
6915 | emit_insn (GEN_MOV (rr, ra)); | |
6916 | /* Tell flow the register for return isn't dead. */ | |
c41c1387 | 6917 | emit_use (rr); |
4977bab6 ZW |
6918 | return; |
6919 | } | |
6920 | ||
6921 | if (TARGET_SH5) | |
6922 | { | |
4977bab6 | 6923 | int offset; |
fada1961 R |
6924 | save_schedule schedule; |
6925 | save_entry *entry; | |
f676971a | 6926 | |
fada1961 R |
6927 | entry = sh5_schedule_saves (&live_regs_mask, &schedule, 0); |
6928 | offset = entry[1].offset; | |
6929 | for (; entry->mode != VOIDmode; entry--) | |
6930 | if (entry->reg == pr_reg) | |
6931 | goto found; | |
4977bab6 ZW |
6932 | |
6933 | /* We can't find pr register. */ | |
f5b9e7c9 | 6934 | gcc_unreachable (); |
4977bab6 ZW |
6935 | |
6936 | found: | |
fada1961 R |
6937 | offset = entry->offset - offset; |
6938 | pr_offset = (rounded_frame_size (d) + offset | |
4977bab6 ZW |
6939 | + SHMEDIA_REGS_STACK_ADJUST ()); |
6940 | } | |
6941 | else | |
fada1961 | 6942 | pr_offset = rounded_frame_size (d); |
4977bab6 ZW |
6943 | |
6944 | emit_insn (GEN_MOV (tmp, GEN_INT (pr_offset))); | |
96a2347e | 6945 | emit_insn (GEN_ADD3 (tmp, tmp, hard_frame_pointer_rtx)); |
4977bab6 | 6946 | |
57d38024 | 6947 | tmp = gen_frame_mem (Pmode, tmp); |
4977bab6 ZW |
6948 | emit_insn (GEN_MOV (tmp, ra)); |
6949 | } | |
6950 | ||
8aa2a305 JW |
6951 | /* Clear variables at function end. */ |
6952 | ||
08c148a8 | 6953 | static void |
cf277499 SB |
6954 | sh_output_function_epilogue (FILE *file ATTRIBUTE_UNUSED, |
6955 | HOST_WIDE_INT size ATTRIBUTE_UNUSED) | |
8aa2a305 | 6956 | { |
afbc2905 | 6957 | sh_need_epilogue_known = 0; |
8aa2a305 JW |
6958 | } |
6959 | ||
61f71b34 | 6960 | static rtx |
cf277499 | 6961 | sh_builtin_saveregs (void) |
1a95a963 | 6962 | { |
1a95a963 | 6963 | /* First unnamed integer register. */ |
38173d38 | 6964 | int first_intreg = crtl->args.info.arg_count[(int) SH_ARG_INT]; |
1a95a963 JW |
6965 | /* Number of integer registers we need to save. */ |
6966 | int n_intregs = MAX (0, NPARM_REGS (SImode) - first_intreg); | |
6967 | /* First unnamed SFmode float reg */ | |
38173d38 | 6968 | int first_floatreg = crtl->args.info.arg_count[(int) SH_ARG_FLOAT]; |
1a95a963 JW |
6969 | /* Number of SFmode float regs to save. */ |
6970 | int n_floatregs = MAX (0, NPARM_REGS (SFmode) - first_floatreg); | |
318881c0 | 6971 | rtx regbuf, fpregs; |
ba4828e0 | 6972 | int bufsize, regno; |
4862826d | 6973 | alias_set_type alias_set; |
1a95a963 | 6974 | |
fa5322fa AO |
6975 | if (TARGET_SH5) |
6976 | { | |
6977 | if (n_intregs) | |
6978 | { | |
6979 | int pushregs = n_intregs; | |
6980 | ||
6981 | while (pushregs < NPARM_REGS (SImode) - 1 | |
6982 | && (CALL_COOKIE_INT_REG_GET | |
38173d38 | 6983 | (crtl->args.info.call_cookie, |
fa5322fa AO |
6984 | NPARM_REGS (SImode) - pushregs) |
6985 | == 1)) | |
6986 | { | |
38173d38 | 6987 | crtl->args.info.call_cookie |
fa5322fa AO |
6988 | &= ~ CALL_COOKIE_INT_REG (NPARM_REGS (SImode) |
6989 | - pushregs, 1); | |
6990 | pushregs++; | |
6991 | } | |
6992 | ||
6993 | if (pushregs == NPARM_REGS (SImode)) | |
38173d38 | 6994 | crtl->args.info.call_cookie |
fa5322fa AO |
6995 | |= (CALL_COOKIE_INT_REG (0, 1) |
6996 | | CALL_COOKIE_STACKSEQ (pushregs - 1)); | |
6997 | else | |
38173d38 | 6998 | crtl->args.info.call_cookie |
fa5322fa AO |
6999 | |= CALL_COOKIE_STACKSEQ (pushregs); |
7000 | ||
38173d38 | 7001 | crtl->args.pretend_args_size += 8 * n_intregs; |
fa5322fa AO |
7002 | } |
7003 | if (TARGET_SHCOMPACT) | |
7004 | return const0_rtx; | |
7005 | } | |
f676971a | 7006 | |
3a8699c7 | 7007 | if (! TARGET_SH2E && ! TARGET_SH4 && ! TARGET_SH5) |
fa5322fa AO |
7008 | { |
7009 | error ("__builtin_saveregs not supported by this subtarget"); | |
7010 | return const0_rtx; | |
7011 | } | |
7012 | ||
7013 | if (TARGET_SHMEDIA) | |
7014 | n_floatregs = 0; | |
7015 | ||
6f317ef3 | 7016 | /* Allocate block of memory for the regs. */ |
1a95a963 JW |
7017 | /* ??? If n_intregs + n_floatregs == 0, should we allocate at least 1 byte? |
7018 | Or can assign_stack_local accept a 0 SIZE argument? */ | |
7019 | bufsize = (n_intregs * UNITS_PER_WORD) + (n_floatregs * UNITS_PER_WORD); | |
7020 | ||
fa5322fa | 7021 | if (TARGET_SHMEDIA) |
57d38024 | 7022 | regbuf = gen_frame_mem (BLKmode, gen_rtx_REG (Pmode, ARG_POINTER_REGNUM)); |
f78ec36a R |
7023 | else if (n_floatregs & 1) |
7024 | { | |
7025 | rtx addr; | |
7026 | ||
7027 | regbuf = assign_stack_local (BLKmode, bufsize + UNITS_PER_WORD, 0); | |
7028 | addr = copy_to_mode_reg (Pmode, XEXP (regbuf, 0)); | |
7029 | emit_insn (gen_iorsi3 (addr, addr, GEN_INT (UNITS_PER_WORD))); | |
7030 | regbuf = change_address (regbuf, BLKmode, addr); | |
7031 | } | |
a11c004b R |
7032 | else if (STACK_BOUNDARY < 64 && TARGET_FPU_DOUBLE && n_floatregs) |
7033 | { | |
7034 | rtx addr, mask; | |
7035 | ||
7036 | regbuf = assign_stack_local (BLKmode, bufsize + UNITS_PER_WORD, 0); | |
7037 | addr = copy_to_mode_reg (Pmode, plus_constant (XEXP (regbuf, 0), 4)); | |
7038 | mask = copy_to_mode_reg (Pmode, GEN_INT (-8)); | |
7039 | emit_insn (gen_andsi3 (addr, addr, mask)); | |
7040 | regbuf = change_address (regbuf, BLKmode, addr); | |
7041 | } | |
fa5322fa | 7042 | else |
a11c004b | 7043 | regbuf = assign_stack_local (BLKmode, bufsize, TARGET_FPU_DOUBLE ? 64 : 0); |
514066a1 | 7044 | alias_set = get_varargs_alias_set (); |
ba4828e0 | 7045 | set_mem_alias_set (regbuf, alias_set); |
1a95a963 JW |
7046 | |
7047 | /* Save int args. | |
7048 | This is optimized to only save the regs that are necessary. Explicitly | |
7049 | named args need not be saved. */ | |
7050 | if (n_intregs > 0) | |
7051 | move_block_from_reg (BASE_ARG_REG (SImode) + first_intreg, | |
f4ef873c RK |
7052 | adjust_address (regbuf, BLKmode, |
7053 | n_floatregs * UNITS_PER_WORD), | |
c6b97fac | 7054 | n_intregs); |
1a95a963 | 7055 | |
fa5322fa AO |
7056 | if (TARGET_SHMEDIA) |
7057 | /* Return the address of the regbuf. */ | |
7058 | return XEXP (regbuf, 0); | |
7059 | ||
1a95a963 JW |
7060 | /* Save float args. |
7061 | This is optimized to only save the regs that are necessary. Explicitly | |
7062 | named args need not be saved. | |
7063 | We explicitly build a pointer to the buffer because it halves the insn | |
7064 | count when not optimizing (otherwise the pointer is built for each reg | |
1245df60 R |
7065 | saved). |
7066 | We emit the moves in reverse order so that we can use predecrement. */ | |
1a95a963 | 7067 | |
7d9c5aa1 R |
7068 | fpregs = copy_to_mode_reg (Pmode, |
7069 | plus_constant (XEXP (regbuf, 0), | |
7070 | n_floatregs * UNITS_PER_WORD)); | |
157371cf | 7071 | if (TARGET_SH4 || TARGET_SH2A_DOUBLE) |
225e4f43 | 7072 | { |
514066a1 | 7073 | rtx mem; |
225e4f43 R |
7074 | for (regno = NPARM_REGS (DFmode) - 2; regno >= first_floatreg; regno -= 2) |
7075 | { | |
7076 | emit_insn (gen_addsi3 (fpregs, fpregs, | |
7077 | GEN_INT (-2 * UNITS_PER_WORD))); | |
57d38024 | 7078 | mem = change_address (regbuf, DFmode, fpregs); |
f676971a | 7079 | emit_move_insn (mem, |
c0d4e710 | 7080 | gen_rtx_REG (DFmode, BASE_ARG_REG (DFmode) + regno)); |
225e4f43 R |
7081 | } |
7082 | regno = first_floatreg; | |
7083 | if (regno & 1) | |
7084 | { | |
832a3292 | 7085 | emit_insn (gen_addsi3 (fpregs, fpregs, GEN_INT (-UNITS_PER_WORD))); |
57d38024 | 7086 | mem = change_address (regbuf, SFmode, fpregs); |
514066a1 | 7087 | emit_move_insn (mem, |
c0d4e710 | 7088 | gen_rtx_REG (SFmode, BASE_ARG_REG (SFmode) + regno |
225e4f43 R |
7089 | - (TARGET_LITTLE_ENDIAN != 0))); |
7090 | } | |
7091 | } | |
7092 | else | |
1245df60 R |
7093 | for (regno = NPARM_REGS (SFmode) - 1; regno >= first_floatreg; regno--) |
7094 | { | |
514066a1 | 7095 | rtx mem; |
ba4828e0 | 7096 | |
832a3292 | 7097 | emit_insn (gen_addsi3 (fpregs, fpregs, GEN_INT (-UNITS_PER_WORD))); |
57d38024 | 7098 | mem = change_address (regbuf, SFmode, fpregs); |
514066a1 | 7099 | emit_move_insn (mem, |
c5c76735 | 7100 | gen_rtx_REG (SFmode, BASE_ARG_REG (SFmode) + regno)); |
1245df60 | 7101 | } |
1a95a963 JW |
7102 | |
7103 | /* Return the address of the regbuf. */ | |
7104 | return XEXP (regbuf, 0); | |
7105 | } | |
7106 | ||
514066a1 RH |
7107 | /* Define the `__builtin_va_list' type for the ABI. */ |
7108 | ||
c35d187f RH |
7109 | static tree |
7110 | sh_build_builtin_va_list (void) | |
514066a1 RH |
7111 | { |
7112 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
7113 | tree record; | |
7114 | ||
61f71b34 DD |
7115 | if (TARGET_SH5 || (! TARGET_SH2E && ! TARGET_SH4) |
7116 | || TARGET_HITACHI || sh_cfun_attr_renesas_p ()) | |
514066a1 RH |
7117 | return ptr_type_node; |
7118 | ||
5d27ef94 | 7119 | record = (*lang_hooks.types.make_type) (RECORD_TYPE); |
514066a1 RH |
7120 | |
7121 | f_next_o = build_decl (FIELD_DECL, get_identifier ("__va_next_o"), | |
7122 | ptr_type_node); | |
7123 | f_next_o_limit = build_decl (FIELD_DECL, | |
7124 | get_identifier ("__va_next_o_limit"), | |
7125 | ptr_type_node); | |
7126 | f_next_fp = build_decl (FIELD_DECL, get_identifier ("__va_next_fp"), | |
7127 | ptr_type_node); | |
7128 | f_next_fp_limit = build_decl (FIELD_DECL, | |
7129 | get_identifier ("__va_next_fp_limit"), | |
7130 | ptr_type_node); | |
7131 | f_next_stack = build_decl (FIELD_DECL, get_identifier ("__va_next_stack"), | |
7132 | ptr_type_node); | |
7133 | ||
7134 | DECL_FIELD_CONTEXT (f_next_o) = record; | |
7135 | DECL_FIELD_CONTEXT (f_next_o_limit) = record; | |
7136 | DECL_FIELD_CONTEXT (f_next_fp) = record; | |
7137 | DECL_FIELD_CONTEXT (f_next_fp_limit) = record; | |
7138 | DECL_FIELD_CONTEXT (f_next_stack) = record; | |
7139 | ||
7140 | TYPE_FIELDS (record) = f_next_o; | |
7141 | TREE_CHAIN (f_next_o) = f_next_o_limit; | |
7142 | TREE_CHAIN (f_next_o_limit) = f_next_fp; | |
7143 | TREE_CHAIN (f_next_fp) = f_next_fp_limit; | |
7144 | TREE_CHAIN (f_next_fp_limit) = f_next_stack; | |
7145 | ||
7146 | layout_type (record); | |
7147 | ||
7148 | return record; | |
7149 | } | |
7150 | ||
28f16d05 KK |
7151 | /* Return always va_list_type_node. */ |
7152 | ||
7153 | static tree | |
7154 | sh_canonical_va_list_type (tree type ATTRIBUTE_UNUSED) | |
7155 | { | |
7156 | return va_list_type_node; | |
7157 | } | |
7158 | ||
514066a1 RH |
7159 | /* Implement `va_start' for varargs and stdarg. */ |
7160 | ||
d7bd8aeb | 7161 | static void |
cf277499 | 7162 | sh_va_start (tree valist, rtx nextarg) |
514066a1 RH |
7163 | { |
7164 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
7165 | tree next_o, next_o_limit, next_fp, next_fp_limit, next_stack; | |
7166 | tree t, u; | |
7167 | int nfp, nint; | |
7168 | ||
fa5322fa AO |
7169 | if (TARGET_SH5) |
7170 | { | |
7171 | expand_builtin_saveregs (); | |
e5faf155 | 7172 | std_expand_builtin_va_start (valist, nextarg); |
fa5322fa AO |
7173 | return; |
7174 | } | |
7175 | ||
61f71b34 DD |
7176 | if ((! TARGET_SH2E && ! TARGET_SH4) |
7177 | || TARGET_HITACHI || sh_cfun_attr_renesas_p ()) | |
514066a1 | 7178 | { |
e5faf155 | 7179 | std_expand_builtin_va_start (valist, nextarg); |
514066a1 RH |
7180 | return; |
7181 | } | |
7182 | ||
7183 | f_next_o = TYPE_FIELDS (va_list_type_node); | |
7184 | f_next_o_limit = TREE_CHAIN (f_next_o); | |
7185 | f_next_fp = TREE_CHAIN (f_next_o_limit); | |
7186 | f_next_fp_limit = TREE_CHAIN (f_next_fp); | |
7187 | f_next_stack = TREE_CHAIN (f_next_fp_limit); | |
7188 | ||
47a25a46 | 7189 | next_o = build3 (COMPONENT_REF, TREE_TYPE (f_next_o), valist, f_next_o, |
44de5aeb | 7190 | NULL_TREE); |
47a25a46 RG |
7191 | next_o_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_o_limit), |
7192 | valist, f_next_o_limit, NULL_TREE); | |
7193 | next_fp = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp), valist, f_next_fp, | |
7194 | NULL_TREE); | |
7195 | next_fp_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp_limit), | |
7196 | valist, f_next_fp_limit, NULL_TREE); | |
7197 | next_stack = build3 (COMPONENT_REF, TREE_TYPE (f_next_stack), | |
7198 | valist, f_next_stack, NULL_TREE); | |
514066a1 RH |
7199 | |
7200 | /* Call __builtin_saveregs. */ | |
5be014d5 AP |
7201 | u = make_tree (sizetype, expand_builtin_saveregs ()); |
7202 | u = fold_convert (ptr_type_node, u); | |
726a989a | 7203 | t = build2 (MODIFY_EXPR, ptr_type_node, next_fp, u); |
514066a1 RH |
7204 | TREE_SIDE_EFFECTS (t) = 1; |
7205 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7206 | ||
38173d38 | 7207 | nfp = crtl->args.info.arg_count[SH_ARG_FLOAT]; |
514066a1 RH |
7208 | if (nfp < 8) |
7209 | nfp = 8 - nfp; | |
7210 | else | |
7211 | nfp = 0; | |
5be014d5 AP |
7212 | u = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, u, |
7213 | size_int (UNITS_PER_WORD * nfp)); | |
726a989a | 7214 | t = build2 (MODIFY_EXPR, ptr_type_node, next_fp_limit, u); |
514066a1 RH |
7215 | TREE_SIDE_EFFECTS (t) = 1; |
7216 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7217 | ||
726a989a | 7218 | t = build2 (MODIFY_EXPR, ptr_type_node, next_o, u); |
514066a1 RH |
7219 | TREE_SIDE_EFFECTS (t) = 1; |
7220 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7221 | ||
38173d38 | 7222 | nint = crtl->args.info.arg_count[SH_ARG_INT]; |
514066a1 RH |
7223 | if (nint < 4) |
7224 | nint = 4 - nint; | |
7225 | else | |
7226 | nint = 0; | |
5be014d5 AP |
7227 | u = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, u, |
7228 | size_int (UNITS_PER_WORD * nint)); | |
726a989a | 7229 | t = build2 (MODIFY_EXPR, ptr_type_node, next_o_limit, u); |
514066a1 RH |
7230 | TREE_SIDE_EFFECTS (t) = 1; |
7231 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7232 | ||
7233 | u = make_tree (ptr_type_node, nextarg); | |
726a989a | 7234 | t = build2 (MODIFY_EXPR, ptr_type_node, next_stack, u); |
514066a1 RH |
7235 | TREE_SIDE_EFFECTS (t) = 1; |
7236 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7237 | } | |
7238 | ||
917f1b7e | 7239 | /* TYPE is a RECORD_TYPE. If there is only a single nonzero-sized |
21bdb106 JR |
7240 | member, return it. */ |
7241 | static tree | |
7242 | find_sole_member (tree type) | |
7243 | { | |
7244 | tree field, member = NULL_TREE; | |
7245 | ||
7246 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
7247 | { | |
7248 | if (TREE_CODE (field) != FIELD_DECL) | |
7249 | continue; | |
7250 | if (!DECL_SIZE (field)) | |
7251 | return NULL_TREE; | |
7252 | if (integer_zerop (DECL_SIZE (field))) | |
7253 | continue; | |
7254 | if (member) | |
7255 | return NULL_TREE; | |
7256 | member = field; | |
7257 | } | |
7258 | return member; | |
7259 | } | |
514066a1 RH |
7260 | /* Implement `va_arg'. */ |
7261 | ||
5fac2ed2 | 7262 | static tree |
726a989a RB |
7263 | sh_gimplify_va_arg_expr (tree valist, tree type, gimple_seq *pre_p, |
7264 | gimple_seq *post_p ATTRIBUTE_UNUSED) | |
514066a1 RH |
7265 | { |
7266 | HOST_WIDE_INT size, rsize; | |
318881c0 | 7267 | tree tmp, pptr_type_node; |
0ef2c525 | 7268 | tree addr, lab_over = NULL, result = NULL; |
03603192 | 7269 | int pass_by_ref = targetm.calls.must_pass_in_stack (TYPE_MODE (type), type); |
11160b58 | 7270 | tree eff_type; |
5fac2ed2 RH |
7271 | |
7272 | if (pass_by_ref) | |
7273 | type = build_pointer_type (type); | |
514066a1 RH |
7274 | |
7275 | size = int_size_in_bytes (type); | |
7276 | rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD; | |
7277 | pptr_type_node = build_pointer_type (ptr_type_node); | |
7278 | ||
61f71b34 DD |
7279 | if (! TARGET_SH5 && (TARGET_SH2E || TARGET_SH4) |
7280 | && ! (TARGET_HITACHI || sh_cfun_attr_renesas_p ())) | |
514066a1 RH |
7281 | { |
7282 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
7283 | tree next_o, next_o_limit, next_fp, next_fp_limit, next_stack; | |
7284 | int pass_as_float; | |
5fac2ed2 | 7285 | tree lab_false; |
21bdb106 | 7286 | tree member; |
514066a1 RH |
7287 | |
7288 | f_next_o = TYPE_FIELDS (va_list_type_node); | |
7289 | f_next_o_limit = TREE_CHAIN (f_next_o); | |
7290 | f_next_fp = TREE_CHAIN (f_next_o_limit); | |
7291 | f_next_fp_limit = TREE_CHAIN (f_next_fp); | |
7292 | f_next_stack = TREE_CHAIN (f_next_fp_limit); | |
7293 | ||
47a25a46 RG |
7294 | next_o = build3 (COMPONENT_REF, TREE_TYPE (f_next_o), valist, f_next_o, |
7295 | NULL_TREE); | |
7296 | next_o_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_o_limit), | |
7297 | valist, f_next_o_limit, NULL_TREE); | |
7298 | next_fp = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp), | |
7299 | valist, f_next_fp, NULL_TREE); | |
7300 | next_fp_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp_limit), | |
7301 | valist, f_next_fp_limit, NULL_TREE); | |
7302 | next_stack = build3 (COMPONENT_REF, TREE_TYPE (f_next_stack), | |
7303 | valist, f_next_stack, NULL_TREE); | |
514066a1 | 7304 | |
6d1cb95f R |
7305 | /* Structures with a single member with a distinct mode are passed |
7306 | like their member. This is relevant if the latter has a REAL_TYPE | |
7307 | or COMPLEX_TYPE type. */ | |
11160b58 R |
7308 | eff_type = type; |
7309 | while (TREE_CODE (eff_type) == RECORD_TYPE | |
7310 | && (member = find_sole_member (eff_type)) | |
21bdb106 JR |
7311 | && (TREE_CODE (TREE_TYPE (member)) == REAL_TYPE |
7312 | || TREE_CODE (TREE_TYPE (member)) == COMPLEX_TYPE | |
7313 | || TREE_CODE (TREE_TYPE (member)) == RECORD_TYPE)) | |
c26104e4 | 7314 | { |
21bdb106 | 7315 | tree field_type = TREE_TYPE (member); |
c26104e4 | 7316 | |
11160b58 R |
7317 | if (TYPE_MODE (eff_type) == TYPE_MODE (field_type)) |
7318 | eff_type = field_type; | |
c26104e4 | 7319 | else |
8a2209bf | 7320 | { |
11160b58 | 7321 | gcc_assert ((TYPE_ALIGN (eff_type) |
21bdb106 | 7322 | < GET_MODE_ALIGNMENT (TYPE_MODE (field_type))) |
11160b58 | 7323 | || (TYPE_ALIGN (eff_type) |
21bdb106 JR |
7324 | > GET_MODE_BITSIZE (TYPE_MODE (field_type)))); |
7325 | break; | |
8a2209bf | 7326 | } |
c26104e4 | 7327 | } |
5fac2ed2 | 7328 | |
bdf57c40 | 7329 | if (TARGET_SH4 || TARGET_SH2A_DOUBLE) |
514066a1 | 7330 | { |
11160b58 R |
7331 | pass_as_float = ((TREE_CODE (eff_type) == REAL_TYPE && size <= 8) |
7332 | || (TREE_CODE (eff_type) == COMPLEX_TYPE | |
7333 | && TREE_CODE (TREE_TYPE (eff_type)) == REAL_TYPE | |
514066a1 RH |
7334 | && size <= 16)); |
7335 | } | |
7336 | else | |
7337 | { | |
11160b58 | 7338 | pass_as_float = (TREE_CODE (eff_type) == REAL_TYPE && size == 4); |
514066a1 RH |
7339 | } |
7340 | ||
5fac2ed2 RH |
7341 | addr = create_tmp_var (pptr_type_node, NULL); |
7342 | lab_false = create_artificial_label (); | |
7343 | lab_over = create_artificial_label (); | |
514066a1 | 7344 | |
5fac2ed2 | 7345 | valist = build1 (INDIRECT_REF, ptr_type_node, addr); |
6d1cb95f | 7346 | |
514066a1 RH |
7347 | if (pass_as_float) |
7348 | { | |
a11c004b R |
7349 | tree next_fp_tmp = create_tmp_var (TREE_TYPE (f_next_fp), NULL); |
7350 | tree cmp; | |
11160b58 | 7351 | bool is_double = size == 8 && TREE_CODE (eff_type) == REAL_TYPE; |
f78ec36a | 7352 | |
a11c004b | 7353 | tmp = build1 (ADDR_EXPR, pptr_type_node, next_fp); |
726a989a | 7354 | gimplify_assign (addr, tmp, pre_p); |
a11c004b | 7355 | |
726a989a | 7356 | gimplify_assign (next_fp_tmp, valist, pre_p); |
a11c004b R |
7357 | tmp = next_fp_limit; |
7358 | if (size > 4 && !is_double) | |
5be014d5 AP |
7359 | tmp = build2 (POINTER_PLUS_EXPR, TREE_TYPE (tmp), tmp, |
7360 | size_int (4 - size)); | |
47a25a46 RG |
7361 | tmp = build2 (GE_EXPR, boolean_type_node, next_fp_tmp, tmp); |
7362 | cmp = build3 (COND_EXPR, void_type_node, tmp, | |
7363 | build1 (GOTO_EXPR, void_type_node, lab_false), | |
7364 | NULL_TREE); | |
a11c004b R |
7365 | if (!is_double) |
7366 | gimplify_and_add (cmp, pre_p); | |
514066a1 | 7367 | |
11160b58 R |
7368 | if (TYPE_ALIGN (eff_type) > BITS_PER_WORD |
7369 | || (is_double || size == 16)) | |
514066a1 | 7370 | { |
5be014d5 AP |
7371 | tmp = fold_convert (sizetype, next_fp_tmp); |
7372 | tmp = build2 (BIT_AND_EXPR, sizetype, tmp, | |
7373 | size_int (UNITS_PER_WORD)); | |
7374 | tmp = build2 (POINTER_PLUS_EXPR, ptr_type_node, | |
7375 | next_fp_tmp, tmp); | |
726a989a | 7376 | gimplify_assign (next_fp_tmp, tmp, pre_p); |
514066a1 | 7377 | } |
a11c004b R |
7378 | if (is_double) |
7379 | gimplify_and_add (cmp, pre_p); | |
514066a1 | 7380 | |
6d1cb95f | 7381 | #ifdef FUNCTION_ARG_SCmode_WART |
11160b58 R |
7382 | if (TYPE_MODE (eff_type) == SCmode |
7383 | && TARGET_SH4 && TARGET_LITTLE_ENDIAN) | |
6d1cb95f | 7384 | { |
11160b58 | 7385 | tree subtype = TREE_TYPE (eff_type); |
5fac2ed2 | 7386 | tree real, imag; |
6d1cb95f | 7387 | |
a11c004b R |
7388 | imag |
7389 | = std_gimplify_va_arg_expr (next_fp_tmp, subtype, pre_p, NULL); | |
5fac2ed2 | 7390 | imag = get_initialized_tmp_var (imag, pre_p, NULL); |
6d1cb95f | 7391 | |
a11c004b R |
7392 | real |
7393 | = std_gimplify_va_arg_expr (next_fp_tmp, subtype, pre_p, NULL); | |
5fac2ed2 RH |
7394 | real = get_initialized_tmp_var (real, pre_p, NULL); |
7395 | ||
47a25a46 | 7396 | result = build2 (COMPLEX_EXPR, type, real, imag); |
5fac2ed2 | 7397 | result = get_initialized_tmp_var (result, pre_p, NULL); |
6d1cb95f R |
7398 | } |
7399 | #endif /* FUNCTION_ARG_SCmode_WART */ | |
7400 | ||
47a25a46 | 7401 | tmp = build1 (GOTO_EXPR, void_type_node, lab_over); |
5fac2ed2 RH |
7402 | gimplify_and_add (tmp, pre_p); |
7403 | ||
47a25a46 | 7404 | tmp = build1 (LABEL_EXPR, void_type_node, lab_false); |
5fac2ed2 | 7405 | gimplify_and_add (tmp, pre_p); |
514066a1 RH |
7406 | |
7407 | tmp = build1 (ADDR_EXPR, pptr_type_node, next_stack); | |
726a989a RB |
7408 | gimplify_assign (addr, tmp, pre_p); |
7409 | gimplify_assign (next_fp_tmp, valist, pre_p); | |
a11c004b | 7410 | |
726a989a | 7411 | gimplify_assign (valist, next_fp_tmp, post_p); |
a11c004b | 7412 | valist = next_fp_tmp; |
514066a1 RH |
7413 | } |
7414 | else | |
7415 | { | |
5be014d5 AP |
7416 | tmp = build2 (POINTER_PLUS_EXPR, ptr_type_node, next_o, |
7417 | size_int (rsize)); | |
47a25a46 RG |
7418 | tmp = build2 (GT_EXPR, boolean_type_node, tmp, next_o_limit); |
7419 | tmp = build3 (COND_EXPR, void_type_node, tmp, | |
7420 | build1 (GOTO_EXPR, void_type_node, lab_false), | |
7421 | NULL_TREE); | |
5fac2ed2 | 7422 | gimplify_and_add (tmp, pre_p); |
514066a1 RH |
7423 | |
7424 | tmp = build1 (ADDR_EXPR, pptr_type_node, next_o); | |
726a989a | 7425 | gimplify_assign (addr, tmp, pre_p); |
514066a1 | 7426 | |
47a25a46 | 7427 | tmp = build1 (GOTO_EXPR, void_type_node, lab_over); |
5fac2ed2 RH |
7428 | gimplify_and_add (tmp, pre_p); |
7429 | ||
47a25a46 | 7430 | tmp = build1 (LABEL_EXPR, void_type_node, lab_false); |
5fac2ed2 | 7431 | gimplify_and_add (tmp, pre_p); |
514066a1 | 7432 | |
bdf57c40 | 7433 | if (size > 4 && ! (TARGET_SH4 || TARGET_SH2A)) |
726a989a | 7434 | gimplify_assign (next_o, next_o_limit, pre_p); |
514066a1 RH |
7435 | |
7436 | tmp = build1 (ADDR_EXPR, pptr_type_node, next_stack); | |
726a989a | 7437 | gimplify_assign (addr, tmp, pre_p); |
514066a1 RH |
7438 | } |
7439 | ||
5fac2ed2 RH |
7440 | if (!result) |
7441 | { | |
47a25a46 | 7442 | tmp = build1 (LABEL_EXPR, void_type_node, lab_over); |
5fac2ed2 RH |
7443 | gimplify_and_add (tmp, pre_p); |
7444 | } | |
514066a1 RH |
7445 | } |
7446 | ||
7447 | /* ??? In va-sh.h, there had been code to make values larger than | |
7448 | size 8 indirect. This does not match the FUNCTION_ARG macros. */ | |
7449 | ||
5fac2ed2 | 7450 | tmp = std_gimplify_va_arg_expr (valist, type, pre_p, NULL); |
6d1cb95f R |
7451 | if (result) |
7452 | { | |
726a989a | 7453 | gimplify_assign (result, tmp, pre_p); |
5fac2ed2 | 7454 | |
47a25a46 | 7455 | tmp = build1 (LABEL_EXPR, void_type_node, lab_over); |
5fac2ed2 | 7456 | gimplify_and_add (tmp, pre_p); |
6d1cb95f R |
7457 | } |
7458 | else | |
5fac2ed2 | 7459 | result = tmp; |
6d1cb95f | 7460 | |
5e039e6b | 7461 | if (pass_by_ref) |
d6e9821f | 7462 | result = build_va_arg_indirect_ref (result); |
5fac2ed2 | 7463 | |
5e039e6b | 7464 | return result; |
514066a1 RH |
7465 | } |
7466 | ||
150e407a | 7467 | bool |
586de218 | 7468 | sh_promote_prototypes (const_tree type) |
61f71b34 DD |
7469 | { |
7470 | if (TARGET_HITACHI) | |
7471 | return 0; | |
7472 | if (! type) | |
7473 | return 1; | |
7474 | return ! sh_attr_renesas_p (type); | |
7475 | } | |
7476 | ||
8cd5a4e0 RH |
7477 | /* Whether an argument must be passed by reference. On SHcompact, we |
7478 | pretend arguments wider than 32-bits that would have been passed in | |
7479 | registers are passed by reference, so that an SHmedia trampoline | |
7480 | loads them into the full 64-bits registers. */ | |
7481 | ||
7482 | static int | |
586de218 KG |
7483 | shcompact_byref (const CUMULATIVE_ARGS *cum, enum machine_mode mode, |
7484 | const_tree type, bool named) | |
8cd5a4e0 RH |
7485 | { |
7486 | unsigned HOST_WIDE_INT size; | |
7487 | ||
7488 | if (type) | |
7489 | size = int_size_in_bytes (type); | |
7490 | else | |
7491 | size = GET_MODE_SIZE (mode); | |
7492 | ||
7493 | if (cum->arg_count[SH_ARG_INT] < NPARM_REGS (SImode) | |
7494 | && (!named | |
7495 | || GET_SH_ARG_CLASS (mode) == SH_ARG_INT | |
7496 | || (GET_SH_ARG_CLASS (mode) == SH_ARG_FLOAT | |
7497 | && cum->arg_count[SH_ARG_FLOAT] >= NPARM_REGS (SFmode))) | |
7498 | && size > 4 | |
7499 | && !SHCOMPACT_FORCE_ON_STACK (mode, type) | |
7500 | && !SH5_WOULD_BE_PARTIAL_NREGS (*cum, mode, type, named)) | |
7501 | return size; | |
7502 | else | |
7503 | return 0; | |
7504 | } | |
7505 | ||
7506 | static bool | |
7507 | sh_pass_by_reference (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
586de218 | 7508 | const_tree type, bool named) |
8cd5a4e0 RH |
7509 | { |
7510 | if (targetm.calls.must_pass_in_stack (mode, type)) | |
7511 | return true; | |
7512 | ||
03603192 JR |
7513 | /* ??? std_gimplify_va_arg_expr passes NULL for cum. That function |
7514 | wants to know about pass-by-reference semantics for incoming | |
7515 | arguments. */ | |
7516 | if (! cum) | |
7517 | return false; | |
7518 | ||
8cd5a4e0 RH |
7519 | if (TARGET_SHCOMPACT) |
7520 | { | |
7521 | cum->byref = shcompact_byref (cum, mode, type, named); | |
7522 | return cum->byref != 0; | |
7523 | } | |
7524 | ||
7525 | return false; | |
7526 | } | |
7527 | ||
6cdd5672 RH |
7528 | static bool |
7529 | sh_callee_copies (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
586de218 | 7530 | const_tree type, bool named ATTRIBUTE_UNUSED) |
6cdd5672 RH |
7531 | { |
7532 | /* ??? How can it possibly be correct to return true only on the | |
7533 | caller side of the equation? Is there someplace else in the | |
7534 | sh backend that's magically producing the copies? */ | |
7535 | return (cum->outgoing | |
7536 | && ((mode == BLKmode ? TYPE_ALIGN (type) : GET_MODE_ALIGNMENT (mode)) | |
7537 | % SH_MIN_ALIGN_FOR_CALLEE_COPY == 0)); | |
7538 | } | |
7539 | ||
78a52f11 RH |
7540 | static int |
7541 | sh_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
7542 | tree type, bool named ATTRIBUTE_UNUSED) | |
7543 | { | |
7544 | int words = 0; | |
7545 | ||
7546 | if (!TARGET_SH5 | |
7547 | && PASS_IN_REG_P (*cum, mode, type) | |
7548 | && !(TARGET_SH4 || TARGET_SH2A_DOUBLE) | |
7549 | && (ROUND_REG (*cum, mode) | |
7550 | + (mode != BLKmode | |
7551 | ? ROUND_ADVANCE (GET_MODE_SIZE (mode)) | |
7552 | : ROUND_ADVANCE (int_size_in_bytes (type))) | |
7553 | > NPARM_REGS (mode))) | |
7554 | words = NPARM_REGS (mode) - ROUND_REG (*cum, mode); | |
7555 | ||
7556 | else if (!TARGET_SHCOMPACT | |
7557 | && SH5_WOULD_BE_PARTIAL_NREGS (*cum, mode, type, named)) | |
7558 | words = NPARM_REGS (SImode) - cum->arg_count[SH_ARG_INT]; | |
7559 | ||
7560 | return words * UNITS_PER_WORD; | |
7561 | } | |
7562 | ||
7563 | ||
61f71b34 DD |
7564 | /* Define where to put the arguments to a function. |
7565 | Value is zero to push the argument on the stack, | |
7566 | or a hard register in which to store the argument. | |
7567 | ||
7568 | MODE is the argument's machine mode. | |
7569 | TYPE is the data type of the argument (as a tree). | |
7570 | This is null for libcalls where that information may | |
7571 | not be available. | |
7572 | CUM is a variable of type CUMULATIVE_ARGS which gives info about | |
7573 | the preceding args and about the function being called. | |
7574 | NAMED is nonzero if this argument is a named parameter | |
7575 | (otherwise it is an extra parameter matching an ellipsis). | |
7576 | ||
7577 | On SH the first args are normally in registers | |
7578 | and the rest are pushed. Any arg that starts within the first | |
7579 | NPARM_REGS words is at least partially passed in a register unless | |
7580 | its data type forbids. */ | |
7581 | ||
7582 | ||
7583 | rtx | |
cf277499 SB |
7584 | sh_function_arg (CUMULATIVE_ARGS *ca, enum machine_mode mode, |
7585 | tree type, int named) | |
61f71b34 DD |
7586 | { |
7587 | if (! TARGET_SH5 && mode == VOIDmode) | |
7588 | return GEN_INT (ca->renesas_abi ? 1 : 0); | |
7589 | ||
7590 | if (! TARGET_SH5 | |
7591 | && PASS_IN_REG_P (*ca, mode, type) | |
7592 | && (named || ! (TARGET_HITACHI || ca->renesas_abi))) | |
7593 | { | |
7594 | int regno; | |
7595 | ||
7596 | if (mode == SCmode && TARGET_SH4 && TARGET_LITTLE_ENDIAN | |
7597 | && (! FUNCTION_ARG_SCmode_WART || (ROUND_REG (*ca, mode) & 1))) | |
7598 | { | |
7599 | rtx r1 = gen_rtx_EXPR_LIST (VOIDmode, | |
7600 | gen_rtx_REG (SFmode, | |
7601 | BASE_ARG_REG (mode) | |
7602 | + (ROUND_REG (*ca, mode) ^ 1)), | |
7603 | const0_rtx); | |
832a3292 KH |
7604 | rtx r2 = gen_rtx_EXPR_LIST (VOIDmode, |
7605 | gen_rtx_REG (SFmode, | |
7606 | BASE_ARG_REG (mode) | |
7607 | + ((ROUND_REG (*ca, mode) + 1) ^ 1)), | |
7608 | GEN_INT (4)); | |
61f71b34 DD |
7609 | return gen_rtx_PARALLEL(SCmode, gen_rtvec(2, r1, r2)); |
7610 | } | |
7611 | ||
7612 | /* If the alignment of a DF value causes an SF register to be | |
7613 | skipped, we will use that skipped register for the next SF | |
7614 | value. */ | |
7615 | if ((TARGET_HITACHI || ca->renesas_abi) | |
7616 | && ca->free_single_fp_reg | |
7617 | && mode == SFmode) | |
7618 | return gen_rtx_REG (mode, ca->free_single_fp_reg); | |
7619 | ||
7620 | regno = (BASE_ARG_REG (mode) + ROUND_REG (*ca, mode)) | |
7621 | ^ (mode == SFmode && TARGET_SH4 | |
7622 | && TARGET_LITTLE_ENDIAN != 0 | |
7623 | && ! TARGET_HITACHI && ! ca->renesas_abi); | |
7624 | return gen_rtx_REG (mode, regno); | |
7625 | ||
7626 | } | |
f676971a | 7627 | |
61f71b34 DD |
7628 | if (TARGET_SH5) |
7629 | { | |
7630 | if (mode == VOIDmode && TARGET_SHCOMPACT) | |
7631 | return GEN_INT (ca->call_cookie); | |
7632 | ||
7633 | /* The following test assumes unnamed arguments are promoted to | |
7634 | DFmode. */ | |
7635 | if (mode == SFmode && ca->free_single_fp_reg) | |
7636 | return SH5_PROTOTYPED_FLOAT_ARG (*ca, mode, ca->free_single_fp_reg); | |
7637 | ||
7638 | if ((GET_SH_ARG_CLASS (mode) == SH_ARG_FLOAT) | |
7639 | && (named || ! ca->prototype_p) | |
7640 | && ca->arg_count[(int) SH_ARG_FLOAT] < NPARM_REGS (SFmode)) | |
7641 | { | |
7642 | if (! ca->prototype_p && TARGET_SHMEDIA) | |
7643 | return SH5_PROTOTYPELESS_FLOAT_ARG (*ca, mode); | |
7644 | ||
7645 | return SH5_PROTOTYPED_FLOAT_ARG (*ca, mode, | |
7646 | FIRST_FP_PARM_REG | |
7647 | + ca->arg_count[(int) SH_ARG_FLOAT]); | |
7648 | } | |
7649 | ||
7650 | if (ca->arg_count[(int) SH_ARG_INT] < NPARM_REGS (SImode) | |
7651 | && (! TARGET_SHCOMPACT | |
7652 | || (! SHCOMPACT_FORCE_ON_STACK (mode, type) | |
7653 | && ! SH5_WOULD_BE_PARTIAL_NREGS (*ca, mode, | |
7654 | type, named)))) | |
7655 | { | |
7656 | return gen_rtx_REG (mode, (FIRST_PARM_REG | |
7657 | + ca->arg_count[(int) SH_ARG_INT])); | |
7658 | } | |
7659 | ||
7660 | return 0; | |
7661 | } | |
7662 | ||
7663 | return 0; | |
7664 | } | |
f676971a | 7665 | |
61f71b34 DD |
7666 | /* Update the data in CUM to advance over an argument |
7667 | of mode MODE and data type TYPE. | |
7668 | (TYPE is null for libcalls where that information may not be | |
7669 | available.) */ | |
7670 | ||
7671 | void | |
cf277499 SB |
7672 | sh_function_arg_advance (CUMULATIVE_ARGS *ca, enum machine_mode mode, |
7673 | tree type, int named) | |
61f71b34 | 7674 | { |
832a3292 KH |
7675 | if (ca->force_mem) |
7676 | ca->force_mem = 0; | |
7677 | else if (TARGET_SH5) | |
7678 | { | |
7679 | tree type2 = (ca->byref && type | |
7680 | ? TREE_TYPE (type) | |
7681 | : type); | |
7682 | enum machine_mode mode2 = (ca->byref && type | |
7683 | ? TYPE_MODE (type2) | |
7684 | : mode); | |
7685 | int dwords = ((ca->byref | |
7686 | ? ca->byref | |
7687 | : mode2 == BLKmode | |
7688 | ? int_size_in_bytes (type2) | |
7689 | : GET_MODE_SIZE (mode2)) + 7) / 8; | |
7690 | int numregs = MIN (dwords, NPARM_REGS (SImode) | |
7691 | - ca->arg_count[(int) SH_ARG_INT]); | |
7692 | ||
7693 | if (numregs) | |
7694 | { | |
7695 | ca->arg_count[(int) SH_ARG_INT] += numregs; | |
7696 | if (TARGET_SHCOMPACT | |
7697 | && SHCOMPACT_FORCE_ON_STACK (mode2, type2)) | |
7698 | { | |
7699 | ca->call_cookie | |
7700 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
7701 | - numregs, 1); | |
7702 | /* N.B. We want this also for outgoing. */ | |
7703 | ca->stack_regs += numregs; | |
7704 | } | |
7705 | else if (ca->byref) | |
7706 | { | |
7707 | if (! ca->outgoing) | |
7708 | ca->stack_regs += numregs; | |
7709 | ca->byref_regs += numregs; | |
7710 | ca->byref = 0; | |
7711 | do | |
7712 | ca->call_cookie | |
7713 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
7714 | - numregs, 2); | |
7715 | while (--numregs); | |
7716 | ca->call_cookie | |
7717 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
7718 | - 1, 1); | |
7719 | } | |
7720 | else if (dwords > numregs) | |
7721 | { | |
7722 | int pushregs = numregs; | |
7723 | ||
7724 | if (TARGET_SHCOMPACT) | |
7725 | ca->stack_regs += numregs; | |
7726 | while (pushregs < NPARM_REGS (SImode) - 1 | |
7727 | && (CALL_COOKIE_INT_REG_GET | |
7728 | (ca->call_cookie, | |
7729 | NPARM_REGS (SImode) - pushregs) | |
7730 | == 1)) | |
7731 | { | |
7732 | ca->call_cookie | |
7733 | &= ~ CALL_COOKIE_INT_REG (NPARM_REGS (SImode) | |
7734 | - pushregs, 1); | |
7735 | pushregs++; | |
7736 | } | |
7737 | if (numregs == NPARM_REGS (SImode)) | |
7738 | ca->call_cookie | |
7739 | |= CALL_COOKIE_INT_REG (0, 1) | |
7740 | | CALL_COOKIE_STACKSEQ (numregs - 1); | |
7741 | else | |
7742 | ca->call_cookie | |
7743 | |= CALL_COOKIE_STACKSEQ (numregs); | |
7744 | } | |
7745 | } | |
7746 | if (GET_SH_ARG_CLASS (mode2) == SH_ARG_FLOAT | |
7747 | && (named || ! ca->prototype_p)) | |
7748 | { | |
7749 | if (mode2 == SFmode && ca->free_single_fp_reg) | |
7750 | ca->free_single_fp_reg = 0; | |
7751 | else if (ca->arg_count[(int) SH_ARG_FLOAT] | |
7752 | < NPARM_REGS (SFmode)) | |
7753 | { | |
7754 | int numfpregs | |
7755 | = MIN ((GET_MODE_SIZE (mode2) + 7) / 8 * 2, | |
7756 | NPARM_REGS (SFmode) | |
7757 | - ca->arg_count[(int) SH_ARG_FLOAT]); | |
7758 | ||
7759 | ca->arg_count[(int) SH_ARG_FLOAT] += numfpregs; | |
7760 | ||
7761 | if (TARGET_SHCOMPACT && ! ca->prototype_p) | |
7762 | { | |
7763 | if (ca->outgoing && numregs > 0) | |
7764 | do | |
7765 | { | |
7766 | ca->call_cookie | |
7767 | |= (CALL_COOKIE_INT_REG | |
7768 | (ca->arg_count[(int) SH_ARG_INT] | |
7769 | - numregs + ((numfpregs - 2) / 2), | |
7770 | 4 + (ca->arg_count[(int) SH_ARG_FLOAT] | |
7771 | - numfpregs) / 2)); | |
7772 | } | |
7773 | while (numfpregs -= 2); | |
7774 | } | |
7775 | else if (mode2 == SFmode && (named) | |
7776 | && (ca->arg_count[(int) SH_ARG_FLOAT] | |
7777 | < NPARM_REGS (SFmode))) | |
7778 | ca->free_single_fp_reg | |
7779 | = FIRST_FP_PARM_REG - numfpregs | |
7780 | + ca->arg_count[(int) SH_ARG_FLOAT] + 1; | |
7781 | } | |
7782 | } | |
7783 | return; | |
7784 | } | |
7785 | ||
7786 | if ((TARGET_HITACHI || ca->renesas_abi) && TARGET_FPU_DOUBLE) | |
7787 | { | |
7788 | /* Note that we've used the skipped register. */ | |
7789 | if (mode == SFmode && ca->free_single_fp_reg) | |
7790 | { | |
7791 | ca->free_single_fp_reg = 0; | |
7792 | return; | |
7793 | } | |
7794 | /* When we have a DF after an SF, there's an SF register that get | |
7795 | skipped in order to align the DF value. We note this skipped | |
7796 | register, because the next SF value will use it, and not the | |
7797 | SF that follows the DF. */ | |
7798 | if (mode == DFmode | |
7799 | && ROUND_REG (*ca, DFmode) != ROUND_REG (*ca, SFmode)) | |
7800 | { | |
7801 | ca->free_single_fp_reg = (ROUND_REG (*ca, SFmode) | |
7802 | + BASE_ARG_REG (mode)); | |
7803 | } | |
7804 | } | |
7805 | ||
157371cf | 7806 | if (! ((TARGET_SH4 || TARGET_SH2A) || ca->renesas_abi) |
832a3292 KH |
7807 | || PASS_IN_REG_P (*ca, mode, type)) |
7808 | (ca->arg_count[(int) GET_SH_ARG_CLASS (mode)] | |
7809 | = (ROUND_REG (*ca, mode) | |
7810 | + (mode == BLKmode | |
7811 | ? ROUND_ADVANCE (int_size_in_bytes (type)) | |
7812 | : ROUND_ADVANCE (GET_MODE_SIZE (mode))))); | |
61f71b34 DD |
7813 | } |
7814 | ||
61f71b34 DD |
7815 | /* The Renesas calling convention doesn't quite fit into this scheme since |
7816 | the address is passed like an invisible argument, but one that is always | |
7817 | passed in memory. */ | |
7818 | static rtx | |
cf277499 | 7819 | sh_struct_value_rtx (tree fndecl, int incoming ATTRIBUTE_UNUSED) |
61f71b34 DD |
7820 | { |
7821 | if (TARGET_HITACHI || sh_attr_renesas_p (fndecl)) | |
7822 | return 0; | |
7823 | return gen_rtx_REG (Pmode, 2); | |
7824 | } | |
7825 | ||
bd5bd7ac KH |
7826 | /* Worker function for TARGET_RETURN_IN_MEMORY. */ |
7827 | ||
61f71b34 | 7828 | static bool |
586de218 | 7829 | sh_return_in_memory (const_tree type, const_tree fndecl) |
61f71b34 DD |
7830 | { |
7831 | if (TARGET_SH5) | |
7832 | { | |
7833 | if (TYPE_MODE (type) == BLKmode) | |
7834 | return ((unsigned HOST_WIDE_INT) int_size_in_bytes (type)) > 8; | |
7835 | else | |
7836 | return GET_MODE_SIZE (TYPE_MODE (type)) > 8; | |
7837 | } | |
7838 | else | |
7839 | { | |
7840 | return (TYPE_MODE (type) == BLKmode | |
7841 | || ((TARGET_HITACHI || sh_attr_renesas_p (fndecl)) | |
7842 | && TREE_CODE (type) == RECORD_TYPE)); | |
7843 | } | |
7844 | } | |
7845 | ||
7846 | /* We actually emit the code in sh_expand_prologue. We used to use | |
7847 | a static variable to flag that we need to emit this code, but that | |
7848 | doesn't when inlining, when functions are deferred and then emitted | |
7849 | later. Fortunately, we already have two flags that are part of struct | |
7850 | function that tell if a function uses varargs or stdarg. */ | |
7851 | static void | |
49616835 JR |
7852 | sh_setup_incoming_varargs (CUMULATIVE_ARGS *ca, |
7853 | enum machine_mode mode, | |
7854 | tree type, | |
7855 | int *pretend_arg_size, | |
cf277499 | 7856 | int second_time ATTRIBUTE_UNUSED) |
61f71b34 | 7857 | { |
e3b5732b | 7858 | gcc_assert (cfun->stdarg); |
49616835 JR |
7859 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl)) |
7860 | { | |
7861 | int named_parm_regs, anon_parm_regs; | |
7862 | ||
7863 | named_parm_regs = (ROUND_REG (*ca, mode) | |
7864 | + (mode == BLKmode | |
7865 | ? ROUND_ADVANCE (int_size_in_bytes (type)) | |
7866 | : ROUND_ADVANCE (GET_MODE_SIZE (mode)))); | |
7867 | anon_parm_regs = NPARM_REGS (SImode) - named_parm_regs; | |
7868 | if (anon_parm_regs > 0) | |
7869 | *pretend_arg_size = anon_parm_regs * 4; | |
7870 | } | |
61f71b34 DD |
7871 | } |
7872 | ||
7873 | static bool | |
cf277499 | 7874 | sh_strict_argument_naming (CUMULATIVE_ARGS *ca ATTRIBUTE_UNUSED) |
61f71b34 DD |
7875 | { |
7876 | return TARGET_SH5; | |
7877 | } | |
7878 | ||
7879 | static bool | |
cf277499 | 7880 | sh_pretend_outgoing_varargs_named (CUMULATIVE_ARGS *ca) |
61f71b34 DD |
7881 | { |
7882 | return ! (TARGET_HITACHI || ca->renesas_abi) && ! TARGET_SH5; | |
7883 | } | |
7884 | ||
7885 | ||
0d7e008e SC |
7886 | /* Define the offset between two registers, one to be eliminated, and |
7887 | the other its replacement, at the start of a routine. */ | |
7888 | ||
7889 | int | |
cf277499 | 7890 | initial_elimination_offset (int from, int to) |
0d7e008e SC |
7891 | { |
7892 | int regs_saved; | |
fa5322fa | 7893 | int regs_saved_rounding = 0; |
0d7e008e | 7894 | int total_saved_regs_space; |
e5f8d4c0 | 7895 | int total_auto_space; |
1245df60 | 7896 | int save_flags = target_flags; |
fa5322fa | 7897 | int copy_flags; |
490dbac7 | 7898 | HARD_REG_SET live_regs_mask; |
fe3ad572 SC |
7899 | |
7900 | shmedia_space_reserved_for_target_registers = false; | |
490dbac7 | 7901 | regs_saved = calc_live_regs (&live_regs_mask); |
fa5322fa | 7902 | regs_saved += SHMEDIA_REGS_STACK_ADJUST (); |
fe3ad572 SC |
7903 | |
7904 | if (shmedia_reserve_space_for_target_registers_p (regs_saved, &live_regs_mask)) | |
7905 | { | |
7906 | shmedia_space_reserved_for_target_registers = true; | |
7907 | regs_saved += shmedia_target_regs_stack_adjust (&live_regs_mask); | |
7908 | } | |
7909 | ||
fa5322fa AO |
7910 | if (TARGET_SH5 && regs_saved % (STACK_BOUNDARY / BITS_PER_UNIT)) |
7911 | regs_saved_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) | |
7912 | - regs_saved % (STACK_BOUNDARY / BITS_PER_UNIT)); | |
8e87e161 | 7913 | |
fa5322fa AO |
7914 | total_auto_space = rounded_frame_size (regs_saved) - regs_saved_rounding; |
7915 | copy_flags = target_flags; | |
1245df60 | 7916 | target_flags = save_flags; |
1a95a963 | 7917 | |
fa5322fa | 7918 | total_saved_regs_space = regs_saved + regs_saved_rounding; |
b9654711 | 7919 | |
96a2347e | 7920 | if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) |
fa5322fa | 7921 | return total_saved_regs_space + total_auto_space |
38173d38 | 7922 | + crtl->args.info.byref_regs * 8; |
8aa2a305 | 7923 | |
0d7e008e | 7924 | if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
fa5322fa | 7925 | return total_saved_regs_space + total_auto_space |
38173d38 | 7926 | + crtl->args.info.byref_regs * 8; |
8aa2a305 JW |
7927 | |
7928 | /* Initial gap between fp and sp is 0. */ | |
96a2347e | 7929 | if (from == HARD_FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
8aa2a305 JW |
7930 | return 0; |
7931 | ||
96a2347e KK |
7932 | if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
7933 | return rounded_frame_size (0); | |
7934 | ||
7935 | if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
7936 | return rounded_frame_size (0); | |
7937 | ||
f5b9e7c9 | 7938 | gcc_assert (from == RETURN_ADDRESS_POINTER_REGNUM |
96a2347e KK |
7939 | && (to == HARD_FRAME_POINTER_REGNUM |
7940 | || to == STACK_POINTER_REGNUM)); | |
f5b9e7c9 | 7941 | if (TARGET_SH5) |
d042370e | 7942 | { |
f5b9e7c9 NS |
7943 | int n = total_saved_regs_space; |
7944 | int pr_reg = TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG; | |
7945 | save_schedule schedule; | |
7946 | save_entry *entry; | |
7947 | ||
7948 | n += total_auto_space; | |
7949 | ||
7950 | /* If it wasn't saved, there's not much we can do. */ | |
7951 | if (! TEST_HARD_REG_BIT (live_regs_mask, pr_reg)) | |
7952 | return n; | |
7953 | ||
7954 | target_flags = copy_flags; | |
7955 | ||
7956 | sh5_schedule_saves (&live_regs_mask, &schedule, n); | |
7957 | for (entry = &schedule.entries[1]; entry->mode != VOIDmode; entry++) | |
7958 | if (entry->reg == pr_reg) | |
7959 | { | |
7960 | target_flags = save_flags; | |
7961 | return entry->offset; | |
7962 | } | |
7963 | gcc_unreachable (); | |
d042370e | 7964 | } |
f5b9e7c9 NS |
7965 | else |
7966 | return total_auto_space; | |
0d7e008e | 7967 | } |
3217af3e KK |
7968 | |
7969 | /* Parse the -mfixed-range= option string. */ | |
7970 | void | |
7971 | sh_fix_range (const char *const_str) | |
7972 | { | |
7973 | int i, first, last; | |
7974 | char *str, *dash, *comma; | |
7975 | ||
7976 | /* str must be of the form REG1'-'REG2{,REG1'-'REG} where REG1 and | |
7977 | REG2 are either register names or register numbers. The effect | |
7978 | of this option is to mark the registers in the range from REG1 to | |
7979 | REG2 as ``fixed'' so they won't be used by the compiler. */ | |
7980 | ||
7981 | i = strlen (const_str); | |
7982 | str = (char *) alloca (i + 1); | |
7983 | memcpy (str, const_str, i + 1); | |
7984 | ||
7985 | while (1) | |
7986 | { | |
7987 | dash = strchr (str, '-'); | |
7988 | if (!dash) | |
7989 | { | |
7990 | warning (0, "value of -mfixed-range must have form REG1-REG2"); | |
7991 | return; | |
7992 | } | |
7993 | *dash = '\0'; | |
7994 | comma = strchr (dash + 1, ','); | |
7995 | if (comma) | |
7996 | *comma = '\0'; | |
7997 | ||
7998 | first = decode_reg_name (str); | |
7999 | if (first < 0) | |
8000 | { | |
8001 | warning (0, "unknown register name: %s", str); | |
8002 | return; | |
8003 | } | |
8004 | ||
8005 | last = decode_reg_name (dash + 1); | |
8006 | if (last < 0) | |
8007 | { | |
8008 | warning (0, "unknown register name: %s", dash + 1); | |
8009 | return; | |
8010 | } | |
8011 | ||
8012 | *dash = '-'; | |
8013 | ||
8014 | if (first > last) | |
8015 | { | |
8016 | warning (0, "%s-%s is an empty range", str, dash + 1); | |
8017 | return; | |
8018 | } | |
8019 | ||
8020 | for (i = first; i <= last; ++i) | |
8021 | fixed_regs[i] = call_used_regs[i] = 1; | |
8022 | ||
8023 | if (!comma) | |
8024 | break; | |
8025 | ||
8026 | *comma = ','; | |
8027 | str = comma + 1; | |
8028 | } | |
8029 | } | |
8aa2a305 | 8030 | \f |
a6ab9fc0 | 8031 | /* Insert any deferred function attributes from earlier pragmas. */ |
12a68f1f | 8032 | static void |
cf277499 | 8033 | sh_insert_attributes (tree node, tree *attributes) |
157131d7 | 8034 | { |
a6ab9fc0 R |
8035 | tree attrs; |
8036 | ||
8037 | if (TREE_CODE (node) != FUNCTION_DECL) | |
157131d7 SC |
8038 | return; |
8039 | ||
8040 | /* We are only interested in fields. */ | |
6615c446 | 8041 | if (!DECL_P (node)) |
157131d7 SC |
8042 | return; |
8043 | ||
a6ab9fc0 R |
8044 | /* Append the attributes to the deferred attributes. */ |
8045 | *sh_deferred_function_attributes_tail = *attributes; | |
8046 | attrs = sh_deferred_function_attributes; | |
8047 | if (!attrs) | |
8048 | return; | |
8049 | ||
8050 | /* Some attributes imply or require the interrupt attribute. */ | |
8051 | if (!lookup_attribute ("interrupt_handler", attrs) | |
8052 | && !lookup_attribute ("interrupt_handler", DECL_ATTRIBUTES (node))) | |
8053 | { | |
8054 | /* If we have a trapa_handler, but no interrupt_handler attribute, | |
8055 | insert an interrupt_handler attribute. */ | |
8056 | if (lookup_attribute ("trapa_handler", attrs) != NULL_TREE) | |
8057 | /* We can't use sh_pr_interrupt here because that's not in the | |
8058 | java frontend. */ | |
8059 | attrs | |
8060 | = tree_cons (get_identifier("interrupt_handler"), NULL_TREE, attrs); | |
561642fa AP |
8061 | /* However, for sp_switch, trap_exit, nosave_low_regs and resbank, |
8062 | if the interrupt attribute is missing, we ignore the attribute | |
8063 | and warn. */ | |
a6ab9fc0 R |
8064 | else if (lookup_attribute ("sp_switch", attrs) |
8065 | || lookup_attribute ("trap_exit", attrs) | |
561642fa AP |
8066 | || lookup_attribute ("nosave_low_regs", attrs) |
8067 | || lookup_attribute ("resbank", attrs)) | |
a6ab9fc0 R |
8068 | { |
8069 | tree *tail; | |
8070 | ||
8071 | for (tail = attributes; attrs; attrs = TREE_CHAIN (attrs)) | |
8072 | { | |
8073 | if (is_attribute_p ("sp_switch", TREE_PURPOSE (attrs)) | |
8074 | || is_attribute_p ("trap_exit", TREE_PURPOSE (attrs)) | |
561642fa AP |
8075 | || is_attribute_p ("nosave_low_regs", TREE_PURPOSE (attrs)) |
8076 | || is_attribute_p ("resbank", TREE_PURPOSE (attrs))) | |
a6ab9fc0 R |
8077 | warning (OPT_Wattributes, |
8078 | "%qs attribute only applies to interrupt functions", | |
8079 | IDENTIFIER_POINTER (TREE_PURPOSE (attrs))); | |
8080 | else | |
8081 | { | |
8082 | *tail = tree_cons (TREE_PURPOSE (attrs), NULL_TREE, | |
8083 | NULL_TREE); | |
8084 | tail = &TREE_CHAIN (*tail); | |
8085 | } | |
8086 | } | |
8087 | attrs = *attributes; | |
8088 | } | |
8089 | } | |
8090 | ||
8091 | /* Install the processed list. */ | |
8092 | *attributes = attrs; | |
8093 | ||
8094 | /* Clear deferred attributes. */ | |
8095 | sh_deferred_function_attributes = NULL_TREE; | |
8096 | sh_deferred_function_attributes_tail = &sh_deferred_function_attributes; | |
157131d7 SC |
8097 | |
8098 | return; | |
8099 | } | |
8100 | ||
91d231cb | 8101 | /* Supported attributes: |
4408efce JL |
8102 | |
8103 | interrupt_handler -- specifies this function is an interrupt handler. | |
8104 | ||
a6ab9fc0 R |
8105 | trapa_handler - like above, but don't save all registers. |
8106 | ||
4408efce JL |
8107 | sp_switch -- specifies an alternate stack for an interrupt handler |
8108 | to run on. | |
8109 | ||
956d6950 | 8110 | trap_exit -- use a trapa to exit an interrupt function instead of |
61f71b34 DD |
8111 | an rte instruction. |
8112 | ||
a6ab9fc0 R |
8113 | nosave_low_regs - don't save r0..r7 in an interrupt handler. |
8114 | This is useful on the SH3 and upwards, | |
8115 | which has a separate set of low regs for User and Supervisor modes. | |
8116 | This should only be used for the lowest level of interrupts. Higher levels | |
8117 | of interrupts must save the registers in case they themselves are | |
8118 | interrupted. | |
8119 | ||
61f71b34 DD |
8120 | renesas -- use Renesas calling/layout conventions (functions and |
8121 | structures). | |
8122 | ||
561642fa AP |
8123 | resbank -- In case of an ISR, use a register bank to save registers |
8124 | R0-R14, MACH, MACL, GBR and PR. This is useful only on SH2A targets. | |
61f71b34 | 8125 | */ |
4408efce | 8126 | |
91d231cb | 8127 | const struct attribute_spec sh_attribute_table[] = |
4408efce | 8128 | { |
91d231cb JM |
8129 | /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */ |
8130 | { "interrupt_handler", 0, 0, true, false, false, sh_handle_interrupt_handler_attribute }, | |
8131 | { "sp_switch", 1, 1, true, false, false, sh_handle_sp_switch_attribute }, | |
8132 | { "trap_exit", 1, 1, true, false, false, sh_handle_trap_exit_attribute }, | |
61f71b34 | 8133 | { "renesas", 0, 0, false, true, false, sh_handle_renesas_attribute }, |
a6ab9fc0 R |
8134 | { "trapa_handler", 0, 0, true, false, false, sh_handle_interrupt_handler_attribute }, |
8135 | { "nosave_low_regs", 0, 0, true, false, false, sh_handle_interrupt_handler_attribute }, | |
561642fa AP |
8136 | { "resbank", 0, 0, true, false, false, sh_handle_resbank_handler_attribute }, |
8137 | { "function_vector", 1, 1, true, false, false, sh2a_handle_function_vector_handler_attribute }, | |
b55e3aad NC |
8138 | #ifdef SYMBIAN |
8139 | /* Symbian support adds three new attributes: | |
8140 | dllexport - for exporting a function/variable that will live in a dll | |
8141 | dllimport - for importing a function/variable from a dll | |
f676971a | 8142 | |
b55e3aad NC |
8143 | Microsoft allows multiple declspecs in one __declspec, separating |
8144 | them with spaces. We do NOT support this. Instead, use __declspec | |
8145 | multiple times. */ | |
8146 | { "dllimport", 0, 0, true, false, false, sh_symbian_handle_dll_attribute }, | |
8147 | { "dllexport", 0, 0, true, false, false, sh_symbian_handle_dll_attribute }, | |
8148 | #endif | |
91d231cb JM |
8149 | { NULL, 0, 0, false, false, false, NULL } |
8150 | }; | |
4408efce | 8151 | |
561642fa AP |
8152 | /* Handle a 'resbank' attribute. */ |
8153 | static tree | |
8154 | sh_handle_resbank_handler_attribute (tree * node, tree name, | |
8155 | tree args ATTRIBUTE_UNUSED, | |
8156 | int flags ATTRIBUTE_UNUSED, | |
8157 | bool * no_add_attrs) | |
8158 | { | |
8159 | if (!TARGET_SH2A) | |
8160 | { | |
8161 | warning (OPT_Wattributes, "%qs attribute is supported only for SH2A", | |
8162 | IDENTIFIER_POINTER (name)); | |
8163 | *no_add_attrs = true; | |
8164 | } | |
8165 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
8166 | { | |
8167 | warning (OPT_Wattributes, "%qs attribute only applies to functions", | |
8168 | IDENTIFIER_POINTER (name)); | |
8169 | *no_add_attrs = true; | |
8170 | } | |
8171 | ||
8172 | return NULL_TREE; | |
8173 | } | |
8174 | ||
91d231cb JM |
8175 | /* Handle an "interrupt_handler" attribute; arguments as in |
8176 | struct attribute_spec.handler. */ | |
8177 | static tree | |
cf277499 | 8178 | sh_handle_interrupt_handler_attribute (tree *node, tree name, |
561642fa AP |
8179 | tree args ATTRIBUTE_UNUSED, |
8180 | int flags ATTRIBUTE_UNUSED, | |
8181 | bool *no_add_attrs) | |
91d231cb JM |
8182 | { |
8183 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
4408efce | 8184 | { |
5c498b10 | 8185 | warning (OPT_Wattributes, "%qs attribute only applies to functions", |
561642fa | 8186 | IDENTIFIER_POINTER (name)); |
91d231cb | 8187 | *no_add_attrs = true; |
4408efce | 8188 | } |
58ab7171 R |
8189 | else if (TARGET_SHCOMPACT) |
8190 | { | |
8191 | error ("attribute interrupt_handler is not compatible with -m5-compact"); | |
8192 | *no_add_attrs = true; | |
8193 | } | |
4408efce | 8194 | |
91d231cb JM |
8195 | return NULL_TREE; |
8196 | } | |
8197 | ||
561642fa AP |
8198 | /* Handle an 'function_vector' attribute; arguments as in |
8199 | struct attribute_spec.handler. */ | |
8200 | static tree | |
8201 | sh2a_handle_function_vector_handler_attribute (tree * node, tree name, | |
8202 | tree args ATTRIBUTE_UNUSED, | |
8203 | int flags ATTRIBUTE_UNUSED, | |
8204 | bool * no_add_attrs) | |
8205 | { | |
8206 | if (!TARGET_SH2A) | |
8207 | { | |
8208 | warning (OPT_Wattributes, "%qs attribute only applies to SH2A", | |
8209 | IDENTIFIER_POINTER (name)); | |
8210 | *no_add_attrs = true; | |
8211 | } | |
8212 | else if (TREE_CODE (*node) != FUNCTION_DECL) | |
8213 | { | |
8214 | warning (OPT_Wattributes, "%qs attribute only applies to functions", | |
8215 | IDENTIFIER_POINTER (name)); | |
8216 | *no_add_attrs = true; | |
8217 | } | |
8218 | else if (TREE_CODE (TREE_VALUE (args)) != INTEGER_CST) | |
8219 | { | |
8220 | /* The argument must be a constant integer. */ | |
8221 | warning (OPT_Wattributes, | |
8222 | "`%s' attribute argument not an integer constant", | |
8223 | IDENTIFIER_POINTER (name)); | |
8224 | *no_add_attrs = true; | |
8225 | } | |
8226 | else if (TREE_INT_CST_LOW (TREE_VALUE (args)) > 255) | |
8227 | { | |
8228 | /* The argument value must be between 0 to 255. */ | |
8229 | warning (OPT_Wattributes, | |
8230 | "`%s' attribute argument should be between 0 to 255", | |
8231 | IDENTIFIER_POINTER (name)); | |
8232 | *no_add_attrs = true; | |
8233 | } | |
8234 | return NULL_TREE; | |
8235 | } | |
8236 | ||
8237 | /* Returns 1 if current function has been assigned the attribute | |
8238 | 'function_vector'. */ | |
8239 | int | |
8240 | sh2a_is_function_vector_call (rtx x) | |
8241 | { | |
8242 | if (GET_CODE (x) == SYMBOL_REF | |
8243 | && (SYMBOL_REF_FLAGS (x) & SYMBOL_FLAG_FUNCVEC_FUNCTION)) | |
8244 | { | |
8245 | tree tr = SYMBOL_REF_DECL (x); | |
8246 | ||
8247 | if (sh2a_function_vector_p (tr)) | |
8248 | return 1; | |
8249 | } | |
8250 | ||
8251 | return 0; | |
8252 | } | |
8253 | ||
8254 | /* Returns the function vector number, if the the attribute | |
8255 | 'function_vector' is assigned, otherwise returns zero. */ | |
8256 | int | |
8257 | sh2a_get_function_vector_number (rtx x) | |
8258 | { | |
8259 | int num; | |
8260 | tree list, t; | |
8261 | ||
8262 | if ((GET_CODE (x) == SYMBOL_REF) | |
8263 | && (SYMBOL_REF_FLAGS (x) & SYMBOL_FLAG_FUNCVEC_FUNCTION)) | |
8264 | { | |
8265 | t = SYMBOL_REF_DECL (x); | |
8266 | ||
8267 | if (TREE_CODE (t) != FUNCTION_DECL) | |
8268 | return 0; | |
8269 | ||
8270 | list = SH_ATTRIBUTES (t); | |
8271 | while (list) | |
8272 | { | |
8273 | if (is_attribute_p ("function_vector", TREE_PURPOSE (list))) | |
8274 | { | |
8275 | num = TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (list))); | |
8276 | return num; | |
8277 | } | |
8278 | ||
8279 | list = TREE_CHAIN (list); | |
8280 | } | |
8281 | ||
8282 | return 0; | |
8283 | } | |
8284 | else | |
8285 | return 0; | |
8286 | } | |
8287 | ||
91d231cb JM |
8288 | /* Handle an "sp_switch" attribute; arguments as in |
8289 | struct attribute_spec.handler. */ | |
8290 | static tree | |
cf277499 SB |
8291 | sh_handle_sp_switch_attribute (tree *node, tree name, tree args, |
8292 | int flags ATTRIBUTE_UNUSED, bool *no_add_attrs) | |
91d231cb JM |
8293 | { |
8294 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
8295 | { | |
5c498b10 | 8296 | warning (OPT_Wattributes, "%qs attribute only applies to functions", |
91d231cb JM |
8297 | IDENTIFIER_POINTER (name)); |
8298 | *no_add_attrs = true; | |
8299 | } | |
91d231cb JM |
8300 | else if (TREE_CODE (TREE_VALUE (args)) != STRING_CST) |
8301 | { | |
4408efce | 8302 | /* The argument must be a constant string. */ |
5c498b10 | 8303 | warning (OPT_Wattributes, "%qs attribute argument not a string constant", |
91d231cb JM |
8304 | IDENTIFIER_POINTER (name)); |
8305 | *no_add_attrs = true; | |
8306 | } | |
4408efce | 8307 | |
91d231cb JM |
8308 | return NULL_TREE; |
8309 | } | |
8310 | ||
8311 | /* Handle an "trap_exit" attribute; arguments as in | |
8312 | struct attribute_spec.handler. */ | |
8313 | static tree | |
cf277499 SB |
8314 | sh_handle_trap_exit_attribute (tree *node, tree name, tree args, |
8315 | int flags ATTRIBUTE_UNUSED, bool *no_add_attrs) | |
91d231cb JM |
8316 | { |
8317 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
8318 | { | |
5c498b10 | 8319 | warning (OPT_Wattributes, "%qs attribute only applies to functions", |
91d231cb JM |
8320 | IDENTIFIER_POINTER (name)); |
8321 | *no_add_attrs = true; | |
8322 | } | |
a6ab9fc0 R |
8323 | /* The argument specifies a trap number to be used in a trapa instruction |
8324 | at function exit (instead of an rte instruction). */ | |
91d231cb JM |
8325 | else if (TREE_CODE (TREE_VALUE (args)) != INTEGER_CST) |
8326 | { | |
4408efce | 8327 | /* The argument must be a constant integer. */ |
5c498b10 DD |
8328 | warning (OPT_Wattributes, "%qs attribute argument not an " |
8329 | "integer constant", IDENTIFIER_POINTER (name)); | |
91d231cb JM |
8330 | *no_add_attrs = true; |
8331 | } | |
318881c0 | 8332 | |
91d231cb | 8333 | return NULL_TREE; |
4408efce JL |
8334 | } |
8335 | ||
61f71b34 | 8336 | static tree |
cf277499 SB |
8337 | sh_handle_renesas_attribute (tree *node ATTRIBUTE_UNUSED, |
8338 | tree name ATTRIBUTE_UNUSED, | |
8339 | tree args ATTRIBUTE_UNUSED, | |
8340 | int flags ATTRIBUTE_UNUSED, | |
8341 | bool *no_add_attrs ATTRIBUTE_UNUSED) | |
61f71b34 DD |
8342 | { |
8343 | return NULL_TREE; | |
8344 | } | |
8345 | ||
8346 | /* True if __attribute__((renesas)) or -mrenesas. */ | |
8347 | int | |
586de218 | 8348 | sh_attr_renesas_p (const_tree td) |
61f71b34 DD |
8349 | { |
8350 | if (TARGET_HITACHI) | |
8351 | return 1; | |
8352 | if (td == 0) | |
8353 | return 0; | |
8354 | if (DECL_P (td)) | |
8355 | td = TREE_TYPE (td); | |
a11c004b R |
8356 | if (td == error_mark_node) |
8357 | return 0; | |
61f71b34 DD |
8358 | return (lookup_attribute ("renesas", TYPE_ATTRIBUTES (td)) |
8359 | != NULL_TREE); | |
8360 | } | |
8361 | ||
8362 | /* True if __attribute__((renesas)) or -mrenesas, for the current | |
8363 | function. */ | |
8364 | int | |
cf277499 | 8365 | sh_cfun_attr_renesas_p (void) |
61f71b34 DD |
8366 | { |
8367 | return sh_attr_renesas_p (current_function_decl); | |
8368 | } | |
8369 | ||
e3ba8d11 | 8370 | int |
cf277499 | 8371 | sh_cfun_interrupt_handler_p (void) |
e3ba8d11 R |
8372 | { |
8373 | return (lookup_attribute ("interrupt_handler", | |
8374 | DECL_ATTRIBUTES (current_function_decl)) | |
8375 | != NULL_TREE); | |
8376 | } | |
bcc8cc82 | 8377 | |
561642fa AP |
8378 | /* Returns 1 if FUNC has been assigned the attribute |
8379 | "function_vector". */ | |
8380 | int | |
8381 | sh2a_function_vector_p (tree func) | |
8382 | { | |
8383 | tree list; | |
8384 | if (TREE_CODE (func) != FUNCTION_DECL) | |
8385 | return 0; | |
8386 | ||
8387 | list = SH_ATTRIBUTES (func); | |
8388 | while (list) | |
8389 | { | |
8390 | if (is_attribute_p ("function_vector", TREE_PURPOSE (list))) | |
8391 | return 1; | |
8392 | ||
8393 | list = TREE_CHAIN (list); | |
8394 | } | |
8395 | return 0; | |
8396 | } | |
8397 | ||
8398 | /* Returns TRUE if given tree has the "resbank" attribute. */ | |
8399 | ||
8400 | int | |
8401 | sh_cfun_resbank_handler_p (void) | |
8402 | { | |
8403 | return ((lookup_attribute ("resbank", | |
8404 | DECL_ATTRIBUTES (current_function_decl)) | |
8405 | != NULL_TREE) | |
8406 | && (lookup_attribute ("interrupt_handler", | |
8407 | DECL_ATTRIBUTES (current_function_decl)) | |
8408 | != NULL_TREE) && TARGET_SH2A); | |
8409 | } | |
8410 | ||
8d932be3 | 8411 | /* Implement TARGET_CHECK_PCH_TARGET_FLAGS. */ |
73a4d10b | 8412 | |
8d932be3 RS |
8413 | static const char * |
8414 | sh_check_pch_target_flags (int old_flags) | |
8415 | { | |
c0fb94d7 RS |
8416 | if ((old_flags ^ target_flags) & (MASK_SH1 | MASK_SH2 | MASK_SH3 |
8417 | | MASK_SH_E | MASK_HARD_SH4 | |
8418 | | MASK_FPU_SINGLE | MASK_SH4)) | |
8419 | return _("created and used with different architectures / ABIs"); | |
8420 | if ((old_flags ^ target_flags) & MASK_HITACHI) | |
8421 | return _("created and used with different ABIs"); | |
8422 | if ((old_flags ^ target_flags) & MASK_LITTLE_ENDIAN) | |
8423 | return _("created and used with different endianness"); | |
bcc8cc82 | 8424 | return NULL; |
bcc8cc82 | 8425 | } |
0d7e008e | 8426 | \f |
8aa2a305 | 8427 | /* Predicates used by the templates. */ |
0d7e008e | 8428 | |
8aa2a305 JW |
8429 | /* Returns 1 if OP is MACL, MACH or PR. The input must be a REG rtx. |
8430 | Used only in general_movsrc_operand. */ | |
0d7e008e | 8431 | |
8aa2a305 | 8432 | int |
cf277499 | 8433 | system_reg_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
0d7e008e | 8434 | { |
8aa2a305 | 8435 | switch (REGNO (op)) |
0d7e008e | 8436 | { |
8aa2a305 JW |
8437 | case PR_REG: |
8438 | case MACL_REG: | |
8439 | case MACH_REG: | |
8440 | return 1; | |
0d7e008e | 8441 | } |
8aa2a305 | 8442 | return 0; |
0d7e008e | 8443 | } |
0d7e008e | 8444 | |
1a95a963 JW |
8445 | /* Nonzero if OP is a floating point value with value 0.0. */ |
8446 | ||
8447 | int | |
cf277499 | 8448 | fp_zero_operand (rtx op) |
d3ae8277 | 8449 | { |
1a95a963 | 8450 | REAL_VALUE_TYPE r; |
d3ae8277 | 8451 | |
1a95a963 JW |
8452 | if (GET_MODE (op) != SFmode) |
8453 | return 0; | |
8454 | ||
8455 | REAL_VALUE_FROM_CONST_DOUBLE (r, op); | |
e4fa6b06 | 8456 | return REAL_VALUES_EQUAL (r, dconst0) && ! REAL_VALUE_MINUS_ZERO (r); |
d3ae8277 SC |
8457 | } |
8458 | ||
1a95a963 | 8459 | /* Nonzero if OP is a floating point value with value 1.0. */ |
d3ae8277 SC |
8460 | |
8461 | int | |
cf277499 | 8462 | fp_one_operand (rtx op) |
d3ae8277 | 8463 | { |
1a95a963 JW |
8464 | REAL_VALUE_TYPE r; |
8465 | ||
8466 | if (GET_MODE (op) != SFmode) | |
8467 | return 0; | |
8468 | ||
8469 | REAL_VALUE_FROM_CONST_DOUBLE (r, op); | |
8470 | return REAL_VALUES_EQUAL (r, dconst1); | |
d3ae8277 | 8471 | } |
1245df60 | 8472 | |
9f09b1f2 R |
8473 | /* For -m4 and -m4-single-only, mode switching is used. If we are |
8474 | compiling without -mfmovd, movsf_ie isn't taken into account for | |
8475 | mode switching. We could check in machine_dependent_reorg for | |
8476 | cases where we know we are in single precision mode, but there is | |
8477 | interface to find that out during reload, so we must avoid | |
8478 | choosing an fldi alternative during reload and thus failing to | |
8479 | allocate a scratch register for the constant loading. */ | |
8480 | int | |
cf277499 | 8481 | fldi_ok (void) |
9f09b1f2 R |
8482 | { |
8483 | return ! TARGET_SH4 || TARGET_FMOVD || reload_completed; | |
8484 | } | |
8485 | ||
225e4f43 | 8486 | int |
cf277499 | 8487 | tertiary_reload_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
225e4f43 R |
8488 | { |
8489 | enum rtx_code code = GET_CODE (op); | |
8490 | return code == MEM || (TARGET_SH4 && code == CONST_DOUBLE); | |
8491 | } | |
8492 | ||
463f02cd KK |
8493 | /* Return the TLS type for TLS symbols, 0 for otherwise. */ |
8494 | int | |
cf277499 | 8495 | tls_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
463f02cd | 8496 | { |
463f02cd KK |
8497 | if (GET_CODE (op) != SYMBOL_REF) |
8498 | return 0; | |
675ff4c7 | 8499 | return SYMBOL_REF_TLS_MODEL (op); |
463f02cd | 8500 | } |
1245df60 | 8501 | \f |
33f7f353 | 8502 | /* Return the destination address of a branch. */ |
f676971a | 8503 | |
318881c0 | 8504 | static int |
cf277499 | 8505 | branch_dest (rtx branch) |
1245df60 | 8506 | { |
33f7f353 JR |
8507 | rtx dest = SET_SRC (PATTERN (branch)); |
8508 | int dest_uid; | |
1245df60 R |
8509 | |
8510 | if (GET_CODE (dest) == IF_THEN_ELSE) | |
8511 | dest = XEXP (dest, 1); | |
8512 | dest = XEXP (dest, 0); | |
8513 | dest_uid = INSN_UID (dest); | |
9d98a694 | 8514 | return INSN_ADDRESSES (dest_uid); |
1245df60 | 8515 | } |
a55e9d2b | 8516 | \f |
5e7a8ee0 | 8517 | /* Return nonzero if REG is not used after INSN. |
a55e9d2b | 8518 | We assume REG is a reload reg, and therefore does |
8b760293 | 8519 | not live past labels. It may live past calls or jumps though. */ |
a55e9d2b | 8520 | int |
cf277499 | 8521 | reg_unused_after (rtx reg, rtx insn) |
a55e9d2b | 8522 | { |
8783b15e | 8523 | enum rtx_code code; |
a55e9d2b RK |
8524 | rtx set; |
8525 | ||
8526 | /* If the reg is set by this instruction, then it is safe for our | |
8527 | case. Disregard the case where this is a store to memory, since | |
8528 | we are checking a register used in the store address. */ | |
8529 | set = single_set (insn); | |
8530 | if (set && GET_CODE (SET_DEST (set)) != MEM | |
8531 | && reg_overlap_mentioned_p (reg, SET_DEST (set))) | |
8532 | return 1; | |
8533 | ||
318881c0 | 8534 | while ((insn = NEXT_INSN (insn))) |
a55e9d2b | 8535 | { |
ec8e098d PB |
8536 | rtx set; |
8537 | if (!INSN_P (insn)) | |
8538 | continue; | |
8539 | ||
a55e9d2b | 8540 | code = GET_CODE (insn); |
8783b15e | 8541 | |
c8f6f18d RK |
8542 | #if 0 |
8543 | /* If this is a label that existed before reload, then the register | |
8544 | if dead here. However, if this is a label added by reorg, then | |
8545 | the register may still be live here. We can't tell the difference, | |
8546 | so we just ignore labels completely. */ | |
8783b15e | 8547 | if (code == CODE_LABEL) |
a55e9d2b | 8548 | return 1; |
c8f6f18d RK |
8549 | /* else */ |
8550 | #endif | |
a55e9d2b | 8551 | |
8b760293 JW |
8552 | if (code == JUMP_INSN) |
8553 | return 0; | |
8554 | ||
8783b15e RK |
8555 | /* If this is a sequence, we must handle them all at once. |
8556 | We could have for instance a call that sets the target register, | |
e03f5d43 | 8557 | and an insn in a delay slot that uses the register. In this case, |
8783b15e | 8558 | we must return 0. */ |
8b760293 | 8559 | else if (code == INSN && GET_CODE (PATTERN (insn)) == SEQUENCE) |
a55e9d2b | 8560 | { |
8783b15e RK |
8561 | int i; |
8562 | int retval = 0; | |
8563 | ||
8564 | for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++) | |
8565 | { | |
8566 | rtx this_insn = XVECEXP (PATTERN (insn), 0, i); | |
8567 | rtx set = single_set (this_insn); | |
8568 | ||
8569 | if (GET_CODE (this_insn) == CALL_INSN) | |
8570 | code = CALL_INSN; | |
38f35781 JW |
8571 | else if (GET_CODE (this_insn) == JUMP_INSN) |
8572 | { | |
8573 | if (INSN_ANNULLED_BRANCH_P (this_insn)) | |
8574 | return 0; | |
8575 | code = JUMP_INSN; | |
8576 | } | |
a55e9d2b | 8577 | |
8783b15e RK |
8578 | if (set && reg_overlap_mentioned_p (reg, SET_SRC (set))) |
8579 | return 0; | |
8580 | if (set && reg_overlap_mentioned_p (reg, SET_DEST (set))) | |
8581 | { | |
8582 | if (GET_CODE (SET_DEST (set)) != MEM) | |
8583 | retval = 1; | |
8584 | else | |
8585 | return 0; | |
8586 | } | |
8587 | if (set == 0 | |
8588 | && reg_overlap_mentioned_p (reg, PATTERN (this_insn))) | |
8589 | return 0; | |
8590 | } | |
8591 | if (retval == 1) | |
8592 | return 1; | |
38f35781 JW |
8593 | else if (code == JUMP_INSN) |
8594 | return 0; | |
8783b15e | 8595 | } |
a55e9d2b | 8596 | |
ec8e098d PB |
8597 | set = single_set (insn); |
8598 | if (set && reg_overlap_mentioned_p (reg, SET_SRC (set))) | |
8599 | return 0; | |
8600 | if (set && reg_overlap_mentioned_p (reg, SET_DEST (set))) | |
8601 | return GET_CODE (SET_DEST (set)) != MEM; | |
8602 | if (set == 0 && reg_overlap_mentioned_p (reg, PATTERN (insn))) | |
8603 | return 0; | |
8783b15e | 8604 | |
62164eb4 | 8605 | if (code == CALL_INSN && call_really_used_regs[REGNO (reg)]) |
8783b15e | 8606 | return 1; |
a55e9d2b RK |
8607 | } |
8608 | return 1; | |
8609 | } | |
225e4f43 | 8610 | \f |
c5b9ef02 | 8611 | #include "ggc.h" |
225e4f43 | 8612 | |
e2500fed | 8613 | static GTY(()) rtx fpscr_rtx; |
225e4f43 | 8614 | rtx |
cf277499 | 8615 | get_fpscr_rtx (void) |
225e4f43 | 8616 | { |
225e4f43 R |
8617 | if (! fpscr_rtx) |
8618 | { | |
c0d4e710 | 8619 | fpscr_rtx = gen_rtx_REG (PSImode, FPSCR_REG); |
225e4f43 | 8620 | REG_USERVAR_P (fpscr_rtx) = 1; |
225e4f43 R |
8621 | mark_user_reg (fpscr_rtx); |
8622 | } | |
8623 | if (! reload_completed || mdep_reorg_phase != SH_AFTER_MDEP_REORG) | |
8624 | mark_user_reg (fpscr_rtx); | |
8625 | return fpscr_rtx; | |
8626 | } | |
8627 | ||
24c2fde2 RH |
8628 | static GTY(()) tree fpscr_values; |
8629 | ||
8630 | static void | |
8631 | emit_fpu_switch (rtx scratch, int index) | |
8632 | { | |
8633 | rtx dst, src; | |
8634 | ||
8635 | if (fpscr_values == NULL) | |
8636 | { | |
8637 | tree t; | |
8638 | ||
8639 | t = build_index_type (integer_one_node); | |
8640 | t = build_array_type (integer_type_node, t); | |
8641 | t = build_decl (VAR_DECL, get_identifier ("__fpscr_values"), t); | |
8642 | DECL_ARTIFICIAL (t) = 1; | |
8643 | DECL_IGNORED_P (t) = 1; | |
8644 | DECL_EXTERNAL (t) = 1; | |
8645 | TREE_STATIC (t) = 1; | |
a5f5c6be | 8646 | TREE_PUBLIC (t) = 1; |
24c2fde2 RH |
8647 | TREE_USED (t) = 1; |
8648 | ||
8649 | fpscr_values = t; | |
8650 | } | |
8651 | ||
8652 | src = DECL_RTL (fpscr_values); | |
b3a13419 | 8653 | if (!can_create_pseudo_p ()) |
24c2fde2 RH |
8654 | { |
8655 | emit_move_insn (scratch, XEXP (src, 0)); | |
8656 | if (index != 0) | |
8657 | emit_insn (gen_addsi3 (scratch, scratch, GEN_INT (index * 4))); | |
8658 | src = adjust_automodify_address (src, PSImode, scratch, index * 4); | |
8659 | } | |
8660 | else | |
8661 | src = adjust_address (src, PSImode, index * 4); | |
8662 | ||
8663 | dst = get_fpscr_rtx (); | |
8664 | emit_move_insn (dst, src); | |
8665 | } | |
8666 | ||
225e4f43 | 8667 | void |
cf277499 | 8668 | emit_sf_insn (rtx pat) |
225e4f43 | 8669 | { |
225e4f43 | 8670 | emit_insn (pat); |
225e4f43 R |
8671 | } |
8672 | ||
8673 | void | |
cf277499 | 8674 | emit_df_insn (rtx pat) |
225e4f43 | 8675 | { |
225e4f43 | 8676 | emit_insn (pat); |
225e4f43 R |
8677 | } |
8678 | ||
8679 | void | |
cf277499 | 8680 | expand_sf_unop (rtx (*fun) (rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
8681 | { |
8682 | emit_sf_insn ((*fun) (operands[0], operands[1], get_fpscr_rtx ())); | |
8683 | } | |
8684 | ||
8685 | void | |
cf277499 | 8686 | expand_sf_binop (rtx (*fun) (rtx, rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
8687 | { |
8688 | emit_sf_insn ((*fun) (operands[0], operands[1], operands[2], | |
8689 | get_fpscr_rtx ())); | |
8690 | } | |
8691 | ||
8692 | void | |
cf277499 | 8693 | expand_df_unop (rtx (*fun) (rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
8694 | { |
8695 | emit_df_insn ((*fun) (operands[0], operands[1], get_fpscr_rtx ())); | |
8696 | } | |
8697 | ||
8698 | void | |
cf277499 | 8699 | expand_df_binop (rtx (*fun) (rtx, rtx, rtx, rtx), rtx *operands) |
225e4f43 R |
8700 | { |
8701 | emit_df_insn ((*fun) (operands[0], operands[1], operands[2], | |
832a3292 | 8702 | get_fpscr_rtx ())); |
225e4f43 | 8703 | } |
225e4f43 | 8704 | \f |
cf277499 | 8705 | static rtx get_free_reg (HARD_REG_SET); |
9f09b1f2 R |
8706 | |
8707 | /* This function returns a register to use to load the address to load | |
8708 | the fpscr from. Currently it always returns r1 or r7, but when we are | |
8709 | able to use pseudo registers after combine, or have a better mechanism | |
8710 | for choosing a register, it should be done here. */ | |
8711 | /* REGS_LIVE is the liveness information for the point for which we | |
8712 | need this allocation. In some bare-bones exit blocks, r1 is live at the | |
8713 | start. We can even have all of r0..r3 being live: | |
8714 | __complex__ long long f (double d) { if (d == 0) return 2; else return 3; } | |
8715 | INSN before which new insns are placed with will clobber the register | |
8716 | we return. If a basic block consists only of setting the return value | |
8717 | register to a pseudo and using that register, the return value is not | |
8718 | live before or after this block, yet we we'll insert our insns right in | |
8719 | the middle. */ | |
225e4f43 | 8720 | |
9f09b1f2 | 8721 | static rtx |
cf277499 | 8722 | get_free_reg (HARD_REG_SET regs_live) |
225e4f43 | 8723 | { |
9f09b1f2 R |
8724 | if (! TEST_HARD_REG_BIT (regs_live, 1)) |
8725 | return gen_rtx_REG (Pmode, 1); | |
225e4f43 | 8726 | |
9f09b1f2 R |
8727 | /* Hard reg 1 is live; since this is a SMALL_REGISTER_CLASSES target, |
8728 | there shouldn't be anything but a jump before the function end. */ | |
f5b9e7c9 NS |
8729 | gcc_assert (!TEST_HARD_REG_BIT (regs_live, 7)); |
8730 | return gen_rtx_REG (Pmode, 7); | |
9f09b1f2 | 8731 | } |
225e4f43 | 8732 | |
f676971a | 8733 | /* This function will set the fpscr from memory. |
9f09b1f2 R |
8734 | MODE is the mode we are setting it to. */ |
8735 | void | |
cf277499 | 8736 | fpscr_set_from_mem (int mode, HARD_REG_SET regs_live) |
9f09b1f2 R |
8737 | { |
8738 | enum attr_fp_mode fp_mode = mode; | |
24c2fde2 | 8739 | enum attr_fp_mode norm_mode = ACTUAL_NORMAL_MODE (FP_MODE); |
5e77c6f9 | 8740 | rtx addr_reg; |
9f09b1f2 | 8741 | |
b3a13419 | 8742 | addr_reg = !can_create_pseudo_p () ? get_free_reg (regs_live) : NULL_RTX; |
24c2fde2 | 8743 | emit_fpu_switch (addr_reg, fp_mode == norm_mode); |
225e4f43 | 8744 | } |
9342c0c4 R |
8745 | |
8746 | /* Is the given character a logical line separator for the assembler? */ | |
8747 | #ifndef IS_ASM_LOGICAL_LINE_SEPARATOR | |
980d8882 | 8748 | #define IS_ASM_LOGICAL_LINE_SEPARATOR(C, STR) ((C) == ';') |
9342c0c4 R |
8749 | #endif |
8750 | ||
8751 | int | |
cf277499 | 8752 | sh_insn_length_adjustment (rtx insn) |
9342c0c4 R |
8753 | { |
8754 | /* Instructions with unfilled delay slots take up an extra two bytes for | |
8755 | the nop in the delay slot. */ | |
8756 | if (((GET_CODE (insn) == INSN | |
c1b92d09 R |
8757 | && GET_CODE (PATTERN (insn)) != USE |
8758 | && GET_CODE (PATTERN (insn)) != CLOBBER) | |
9342c0c4 R |
8759 | || GET_CODE (insn) == CALL_INSN |
8760 | || (GET_CODE (insn) == JUMP_INSN | |
8761 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC | |
8762 | && GET_CODE (PATTERN (insn)) != ADDR_VEC)) | |
8763 | && GET_CODE (PATTERN (NEXT_INSN (PREV_INSN (insn)))) != SEQUENCE | |
8764 | && get_attr_needs_delay_slot (insn) == NEEDS_DELAY_SLOT_YES) | |
8765 | return 2; | |
8766 | ||
3a8699c7 AO |
8767 | /* SH2e has a bug that prevents the use of annulled branches, so if |
8768 | the delay slot is not filled, we'll have to put a NOP in it. */ | |
8769 | if (sh_cpu == CPU_SH2E | |
8770 | && GET_CODE (insn) == JUMP_INSN | |
8771 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC | |
8772 | && GET_CODE (PATTERN (insn)) != ADDR_VEC | |
8773 | && get_attr_type (insn) == TYPE_CBRANCH | |
8774 | && GET_CODE (PATTERN (NEXT_INSN (PREV_INSN (insn)))) != SEQUENCE) | |
8775 | return 2; | |
8776 | ||
9342c0c4 | 8777 | /* sh-dsp parallel processing insn take four bytes instead of two. */ |
f676971a | 8778 | |
9342c0c4 R |
8779 | if (GET_CODE (insn) == INSN) |
8780 | { | |
8781 | int sum = 0; | |
8782 | rtx body = PATTERN (insn); | |
8ceaac3c CP |
8783 | const char *template; |
8784 | char c; | |
9342c0c4 R |
8785 | int maybe_label = 1; |
8786 | ||
8787 | if (GET_CODE (body) == ASM_INPUT) | |
8788 | template = XSTR (body, 0); | |
8789 | else if (asm_noperands (body) >= 0) | |
8790 | template | |
343d66ad | 8791 | = decode_asm_operands (body, NULL, NULL, NULL, NULL, NULL); |
9342c0c4 R |
8792 | else |
8793 | return 0; | |
8794 | do | |
8795 | { | |
8796 | int ppi_adjust = 0; | |
8797 | ||
8798 | do | |
8799 | c = *template++; | |
8800 | while (c == ' ' || c == '\t'); | |
8801 | /* all sh-dsp parallel-processing insns start with p. | |
8802 | The only non-ppi sh insn starting with p is pref. | |
8803 | The only ppi starting with pr is prnd. */ | |
8804 | if ((c == 'p' || c == 'P') && strncasecmp ("re", template, 2)) | |
8805 | ppi_adjust = 2; | |
8806 | /* The repeat pseudo-insn expands two three insns, a total of | |
8807 | six bytes in size. */ | |
8808 | else if ((c == 'r' || c == 'R') | |
8809 | && ! strncasecmp ("epeat", template, 5)) | |
8810 | ppi_adjust = 4; | |
980d8882 BS |
8811 | while (c && c != '\n' |
8812 | && ! IS_ASM_LOGICAL_LINE_SEPARATOR (c, template)) | |
9342c0c4 R |
8813 | { |
8814 | /* If this is a label, it is obviously not a ppi insn. */ | |
8815 | if (c == ':' && maybe_label) | |
8816 | { | |
8817 | ppi_adjust = 0; | |
8818 | break; | |
8819 | } | |
8820 | else if (c == '\'' || c == '"') | |
8821 | maybe_label = 0; | |
8822 | c = *template++; | |
8823 | } | |
8824 | sum += ppi_adjust; | |
8825 | maybe_label = c != ':'; | |
8826 | } | |
8827 | while (c); | |
8828 | return sum; | |
8829 | } | |
8830 | return 0; | |
8831 | } | |
1a66cd67 | 8832 | \f |
55eec660 AO |
8833 | /* Return TRUE if X references a SYMBOL_REF or LABEL_REF whose symbol |
8834 | isn't protected by a PIC unspec. */ | |
1a66cd67 | 8835 | int |
cf277499 | 8836 | nonpic_symbol_mentioned_p (rtx x) |
1a66cd67 AO |
8837 | { |
8838 | register const char *fmt; | |
8839 | register int i; | |
8840 | ||
fa5322fa AO |
8841 | if (GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == LABEL_REF |
8842 | || GET_CODE (x) == PC) | |
1a66cd67 AO |
8843 | return 1; |
8844 | ||
fa5322fa AO |
8845 | /* We don't want to look into the possible MEM location of a |
8846 | CONST_DOUBLE, since we're not going to use it, in general. */ | |
8847 | if (GET_CODE (x) == CONST_DOUBLE) | |
8848 | return 0; | |
8849 | ||
1a66cd67 | 8850 | if (GET_CODE (x) == UNSPEC |
4773afa4 AO |
8851 | && (XINT (x, 1) == UNSPEC_PIC |
8852 | || XINT (x, 1) == UNSPEC_GOT | |
8853 | || XINT (x, 1) == UNSPEC_GOTOFF | |
fa5322fa | 8854 | || XINT (x, 1) == UNSPEC_GOTPLT |
463f02cd KK |
8855 | || XINT (x, 1) == UNSPEC_GOTTPOFF |
8856 | || XINT (x, 1) == UNSPEC_DTPOFF | |
4773afa4 | 8857 | || XINT (x, 1) == UNSPEC_PLT)) |
832a3292 | 8858 | return 0; |
1a66cd67 AO |
8859 | |
8860 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
8861 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
8862 | { | |
8863 | if (fmt[i] == 'E') | |
8864 | { | |
8865 | register int j; | |
8866 | ||
8867 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
8868 | if (nonpic_symbol_mentioned_p (XVECEXP (x, i, j))) | |
8869 | return 1; | |
8870 | } | |
8871 | else if (fmt[i] == 'e' && nonpic_symbol_mentioned_p (XEXP (x, i))) | |
8872 | return 1; | |
8873 | } | |
8874 | ||
8875 | return 0; | |
8876 | } | |
8877 | ||
8878 | /* Convert a non-PIC address in `orig' to a PIC address using @GOT or | |
6f317ef3 | 8879 | @GOTOFF in `reg'. */ |
1a66cd67 | 8880 | rtx |
cf277499 SB |
8881 | legitimize_pic_address (rtx orig, enum machine_mode mode ATTRIBUTE_UNUSED, |
8882 | rtx reg) | |
1a66cd67 | 8883 | { |
463f02cd KK |
8884 | if (tls_symbolic_operand (orig, Pmode)) |
8885 | return orig; | |
8886 | ||
1a66cd67 | 8887 | if (GET_CODE (orig) == LABEL_REF |
675ff4c7 | 8888 | || (GET_CODE (orig) == SYMBOL_REF && SYMBOL_REF_LOCAL_P (orig))) |
1a66cd67 AO |
8889 | { |
8890 | if (reg == 0) | |
8891 | reg = gen_reg_rtx (Pmode); | |
8892 | ||
8893 | emit_insn (gen_symGOTOFF2reg (reg, orig)); | |
8894 | return reg; | |
8895 | } | |
8896 | else if (GET_CODE (orig) == SYMBOL_REF) | |
8897 | { | |
8898 | if (reg == 0) | |
8899 | reg = gen_reg_rtx (Pmode); | |
8900 | ||
8901 | emit_insn (gen_symGOT2reg (reg, orig)); | |
8902 | return reg; | |
8903 | } | |
8904 | return orig; | |
8905 | } | |
b91455de KK |
8906 | |
8907 | /* Mark the use of a constant in the literal table. If the constant | |
8908 | has multiple labels, make it unique. */ | |
c1b92d09 | 8909 | static rtx |
cf277499 | 8910 | mark_constant_pool_use (rtx x) |
b91455de KK |
8911 | { |
8912 | rtx insn, lab, pattern; | |
8913 | ||
8914 | if (x == NULL) | |
8915 | return x; | |
8916 | ||
8917 | switch (GET_CODE (x)) | |
8918 | { | |
8919 | case LABEL_REF: | |
8920 | x = XEXP (x, 0); | |
8921 | case CODE_LABEL: | |
8922 | break; | |
8923 | default: | |
8924 | return x; | |
8925 | } | |
8926 | ||
8927 | /* Get the first label in the list of labels for the same constant | |
8928 | and delete another labels in the list. */ | |
8929 | lab = x; | |
8930 | for (insn = PREV_INSN (x); insn; insn = PREV_INSN (insn)) | |
8931 | { | |
8932 | if (GET_CODE (insn) != CODE_LABEL | |
8933 | || LABEL_REFS (insn) != NEXT_INSN (insn)) | |
8934 | break; | |
8935 | lab = insn; | |
8936 | } | |
8937 | ||
8938 | for (insn = LABEL_REFS (lab); insn; insn = LABEL_REFS (insn)) | |
8939 | INSN_DELETED_P (insn) = 1; | |
8940 | ||
8941 | /* Mark constants in a window. */ | |
8942 | for (insn = NEXT_INSN (x); insn; insn = NEXT_INSN (insn)) | |
8943 | { | |
8944 | if (GET_CODE (insn) != INSN) | |
8945 | continue; | |
8946 | ||
8947 | pattern = PATTERN (insn); | |
8948 | if (GET_CODE (pattern) != UNSPEC_VOLATILE) | |
8949 | continue; | |
8950 | ||
8951 | switch (XINT (pattern, 1)) | |
8952 | { | |
8953 | case UNSPECV_CONST2: | |
8954 | case UNSPECV_CONST4: | |
8955 | case UNSPECV_CONST8: | |
8956 | XVECEXP (pattern, 0, 1) = const1_rtx; | |
8957 | break; | |
8958 | case UNSPECV_WINDOW_END: | |
8959 | if (XVECEXP (pattern, 0, 0) == x) | |
8960 | return lab; | |
8961 | break; | |
8962 | case UNSPECV_CONST_END: | |
8963 | return lab; | |
8964 | default: | |
8965 | break; | |
8966 | } | |
8967 | } | |
8968 | ||
8969 | return lab; | |
8970 | } | |
6e7b03e1 AH |
8971 | \f |
8972 | /* Return true if it's possible to redirect BRANCH1 to the destination | |
8973 | of an unconditional jump BRANCH2. We only want to do this if the | |
8974 | resulting branch will have a short displacement. */ | |
f676971a | 8975 | int |
cf277499 | 8976 | sh_can_redirect_branch (rtx branch1, rtx branch2) |
6e7b03e1 AH |
8977 | { |
8978 | if (flag_expensive_optimizations && simplejump_p (branch2)) | |
8979 | { | |
8980 | rtx dest = XEXP (SET_SRC (single_set (branch2)), 0); | |
8981 | rtx insn; | |
8982 | int distance; | |
f676971a EC |
8983 | |
8984 | for (distance = 0, insn = NEXT_INSN (branch1); | |
8985 | insn && distance < 256; | |
6e7b03e1 AH |
8986 | insn = PREV_INSN (insn)) |
8987 | { | |
f676971a | 8988 | if (insn == dest) |
6e7b03e1 AH |
8989 | return 1; |
8990 | else | |
8991 | distance += get_attr_length (insn); | |
8992 | } | |
f676971a EC |
8993 | for (distance = 0, insn = NEXT_INSN (branch1); |
8994 | insn && distance < 256; | |
6e7b03e1 AH |
8995 | insn = NEXT_INSN (insn)) |
8996 | { | |
f676971a | 8997 | if (insn == dest) |
6e7b03e1 AH |
8998 | return 1; |
8999 | else | |
9000 | distance += get_attr_length (insn); | |
9001 | } | |
9002 | } | |
9003 | return 0; | |
9004 | } | |
9005 | ||
5e7a8ee0 | 9006 | /* Return nonzero if register old_reg can be renamed to register new_reg. */ |
03b8ec29 | 9007 | int |
cf277499 SB |
9008 | sh_hard_regno_rename_ok (unsigned int old_reg ATTRIBUTE_UNUSED, |
9009 | unsigned int new_reg) | |
03b8ec29 | 9010 | { |
832a3292 KH |
9011 | /* Interrupt functions can only use registers that have already been |
9012 | saved by the prologue, even if they would normally be | |
9013 | call-clobbered. */ | |
03b8ec29 | 9014 | |
6fb5fa3c | 9015 | if (sh_cfun_interrupt_handler_p () && !df_regs_ever_live_p (new_reg)) |
832a3292 | 9016 | return 0; |
03b8ec29 | 9017 | |
832a3292 | 9018 | return 1; |
03b8ec29 DD |
9019 | } |
9020 | ||
95c0af87 | 9021 | /* Function to update the integer COST |
c237e94a ZW |
9022 | based on the relationship between INSN that is dependent on |
9023 | DEP_INSN through the dependence LINK. The default is to make no | |
9024 | adjustment to COST. This can be used for example to specify to | |
9025 | the scheduler that an output- or anti-dependence does not incur | |
95c0af87 JR |
9026 | the same cost as a data-dependence. The return value should be |
9027 | the new value for COST. */ | |
c237e94a | 9028 | static int |
cf277499 | 9029 | sh_adjust_cost (rtx insn, rtx link ATTRIBUTE_UNUSED, rtx dep_insn, int cost) |
c237e94a | 9030 | { |
c49439f1 | 9031 | rtx reg, use_pat; |
c237e94a | 9032 | |
2ad65b0e SC |
9033 | if (TARGET_SHMEDIA) |
9034 | { | |
9035 | /* On SHmedia, if the dependence is an anti-dependence or | |
24746a42 | 9036 | output-dependence, there is no cost. */ |
2ad65b0e | 9037 | if (REG_NOTE_KIND (link) != 0) |
73a4d10b R |
9038 | { |
9039 | /* However, dependencies between target register loads and | |
9040 | uses of the register in a subsequent block that are separated | |
9041 | by a conditional branch are not modelled - we have to do with | |
9042 | the anti-dependency between the target register load and the | |
9043 | conditional branch that ends the current block. */ | |
9044 | if (REG_NOTE_KIND (link) == REG_DEP_ANTI | |
9045 | && GET_CODE (PATTERN (dep_insn)) == SET | |
9046 | && (get_attr_type (dep_insn) == TYPE_PT_MEDIA | |
9047 | || get_attr_type (dep_insn) == TYPE_PTABS_MEDIA) | |
9048 | && get_attr_type (insn) == TYPE_CBRANCH_MEDIA) | |
9049 | { | |
9050 | int orig_cost = cost; | |
9051 | rtx note = find_reg_note (insn, REG_BR_PROB, 0); | |
9052 | rtx target = ((! note | |
9053 | || INTVAL (XEXP (note, 0)) * 2 < REG_BR_PROB_BASE) | |
9054 | ? insn : JUMP_LABEL (insn)); | |
9055 | /* On the likely path, the branch costs 1, on the unlikely path, | |
9056 | it costs 3. */ | |
9057 | cost--; | |
9058 | do | |
9059 | target = next_active_insn (target); | |
9060 | while (target && ! flow_dependent_p (target, dep_insn) | |
9061 | && --cost > 0); | |
9062 | /* If two branches are executed in immediate succession, with the | |
9063 | first branch properly predicted, this causes a stall at the | |
9064 | second branch, hence we won't need the target for the | |
9065 | second branch for two cycles after the launch of the first | |
9066 | branch. */ | |
9067 | if (cost > orig_cost - 2) | |
9068 | cost = orig_cost - 2; | |
9069 | } | |
9070 | else | |
9071 | cost = 0; | |
9072 | } | |
2ad65b0e | 9073 | |
73a4d10b R |
9074 | else if (get_attr_is_mac_media (insn) |
9075 | && get_attr_is_mac_media (dep_insn)) | |
9076 | cost = 1; | |
9077 | ||
9078 | else if (! reload_completed | |
9079 | && GET_CODE (PATTERN (insn)) == SET | |
9080 | && GET_CODE (SET_SRC (PATTERN (insn))) == FLOAT | |
9081 | && GET_CODE (PATTERN (dep_insn)) == SET | |
9082 | && fp_arith_reg_operand (SET_SRC (PATTERN (dep_insn)), VOIDmode) | |
9083 | && cost < 4) | |
9084 | cost = 4; | |
9085 | /* Schedule the ptabs for a casesi_jump_media in preference to stuff | |
9086 | that is needed at the target. */ | |
9087 | else if (get_attr_type (insn) == TYPE_JUMP_MEDIA | |
9088 | && ! flow_dependent_p (insn, dep_insn)) | |
9089 | cost--; | |
2ad65b0e | 9090 | } |
c49439f1 | 9091 | else if (REG_NOTE_KIND (link) == 0) |
c237e94a | 9092 | { |
78d310c2 R |
9093 | enum attr_type type; |
9094 | rtx dep_set; | |
c49439f1 R |
9095 | |
9096 | if (recog_memoized (insn) < 0 | |
9097 | || recog_memoized (dep_insn) < 0) | |
95c0af87 | 9098 | return cost; |
c49439f1 | 9099 | |
78d310c2 | 9100 | dep_set = single_set (dep_insn); |
c49439f1 | 9101 | |
78d310c2 R |
9102 | /* The latency that we specify in the scheduling description refers |
9103 | to the actual output, not to an auto-increment register; for that, | |
9104 | the latency is one. */ | |
9105 | if (dep_set && MEM_P (SET_SRC (dep_set)) && cost > 1) | |
9106 | { | |
9107 | rtx set = single_set (insn); | |
9108 | ||
9109 | if (set | |
9110 | && !reg_mentioned_p (SET_DEST (dep_set), SET_SRC (set)) | |
9111 | && (!MEM_P (SET_DEST (set)) | |
9112 | || !reg_mentioned_p (SET_DEST (dep_set), | |
9113 | XEXP (SET_DEST (set), 0)))) | |
9114 | cost = 1; | |
9115 | } | |
c237e94a ZW |
9116 | /* The only input for a call that is timing-critical is the |
9117 | function's address. */ | |
78d310c2 | 9118 | if (GET_CODE (insn) == CALL_INSN) |
c49439f1 R |
9119 | { |
9120 | rtx call = PATTERN (insn); | |
9121 | ||
9122 | if (GET_CODE (call) == PARALLEL) | |
9123 | call = XVECEXP (call, 0 ,0); | |
9124 | if (GET_CODE (call) == SET) | |
9125 | call = SET_SRC (call); | |
9126 | if (GET_CODE (call) == CALL && GET_CODE (XEXP (call, 0)) == MEM | |
73a4d10b R |
9127 | /* sibcalli_thunk uses a symbol_ref in an unspec. */ |
9128 | && (GET_CODE (XEXP (XEXP (call, 0), 0)) == UNSPEC | |
9129 | || ! reg_set_p (XEXP (XEXP (call, 0), 0), dep_insn))) | |
78d310c2 | 9130 | cost -= TARGET_SH4_300 ? 3 : 6; |
c49439f1 | 9131 | } |
c237e94a ZW |
9132 | /* Likewise, the most timing critical input for an sfuncs call |
9133 | is the function address. However, sfuncs typically start | |
9134 | using their arguments pretty quickly. | |
78d310c2 R |
9135 | Assume a four cycle delay for SH4 before they are needed. |
9136 | Cached ST40-300 calls are quicker, so assume only a one | |
9137 | cycle delay there. | |
9138 | ??? Maybe we should encode the delays till input registers | |
9139 | are needed by sfuncs into the sfunc call insn. */ | |
c49439f1 R |
9140 | /* All sfunc calls are parallels with at least four components. |
9141 | Exploit this to avoid unnecessary calls to sfunc_uses_reg. */ | |
9142 | else if (GET_CODE (PATTERN (insn)) == PARALLEL | |
9143 | && XVECLEN (PATTERN (insn), 0) >= 4 | |
9144 | && (reg = sfunc_uses_reg (insn))) | |
9145 | { | |
9146 | if (! reg_set_p (reg, dep_insn)) | |
78d310c2 R |
9147 | cost -= TARGET_SH4_300 ? 1 : 4; |
9148 | } | |
9149 | if (TARGET_HARD_SH4 && !TARGET_SH4_300) | |
9150 | { | |
9151 | enum attr_type dep_type = get_attr_type (dep_insn); | |
9152 | ||
9153 | if (dep_type == TYPE_FLOAD || dep_type == TYPE_PCFLOAD) | |
9154 | cost--; | |
9155 | else if ((dep_type == TYPE_LOAD_SI || dep_type == TYPE_PCLOAD_SI) | |
9156 | && (type = get_attr_type (insn)) != TYPE_CALL | |
9157 | && type != TYPE_SFUNC) | |
9158 | cost--; | |
9159 | /* When the preceding instruction loads the shift amount of | |
9160 | the following SHAD/SHLD, the latency of the load is increased | |
9161 | by 1 cycle. */ | |
9162 | if (get_attr_type (insn) == TYPE_DYN_SHIFT | |
9163 | && get_attr_any_int_load (dep_insn) == ANY_INT_LOAD_YES | |
c6ac6e82 | 9164 | && reg_overlap_mentioned_p (SET_DEST (dep_set), |
78d310c2 R |
9165 | XEXP (SET_SRC (single_set (insn)), |
9166 | 1))) | |
9167 | cost++; | |
9168 | /* When an LS group instruction with a latency of less than | |
9169 | 3 cycles is followed by a double-precision floating-point | |
9170 | instruction, FIPR, or FTRV, the latency of the first | |
9171 | instruction is increased to 3 cycles. */ | |
9172 | else if (cost < 3 | |
9173 | && get_attr_insn_class (dep_insn) == INSN_CLASS_LS_GROUP | |
9174 | && get_attr_dfp_comp (insn) == DFP_COMP_YES) | |
9175 | cost = 3; | |
9176 | /* The lsw register of a double-precision computation is ready one | |
9177 | cycle earlier. */ | |
9178 | else if (reload_completed | |
9179 | && get_attr_dfp_comp (dep_insn) == DFP_COMP_YES | |
9180 | && (use_pat = single_set (insn)) | |
9181 | && ! regno_use_in (REGNO (SET_DEST (single_set (dep_insn))), | |
9182 | SET_SRC (use_pat))) | |
9183 | cost -= 1; | |
9184 | ||
9185 | if (get_attr_any_fp_comp (dep_insn) == ANY_FP_COMP_YES | |
9186 | && get_attr_late_fp_use (insn) == LATE_FP_USE_YES) | |
9187 | cost -= 1; | |
9188 | } | |
9189 | else if (TARGET_SH4_300) | |
9190 | { | |
9191 | /* Stores need their input register two cycles later. */ | |
9192 | if (dep_set && cost >= 1 | |
9193 | && ((type = get_attr_type (insn)) == TYPE_STORE | |
9194 | || type == TYPE_PSTORE | |
9195 | || type == TYPE_FSTORE || type == TYPE_MAC_MEM)) | |
9196 | { | |
9197 | rtx set = single_set (insn); | |
9198 | ||
9199 | if (!reg_mentioned_p (SET_SRC (set), XEXP (SET_DEST (set), 0)) | |
9200 | && rtx_equal_p (SET_SRC (set), SET_DEST (dep_set))) | |
9201 | { | |
9202 | cost -= 2; | |
9203 | /* But don't reduce the cost below 1 if the address depends | |
9204 | on a side effect of dep_insn. */ | |
9205 | if (cost < 1 | |
9206 | && modified_in_p (XEXP (SET_DEST (set), 0), dep_insn)) | |
9207 | cost = 1; | |
9208 | } | |
9209 | } | |
9210 | } | |
c49439f1 R |
9211 | } |
9212 | /* An anti-dependence penalty of two applies if the first insn is a double | |
9213 | precision fadd / fsub / fmul. */ | |
78d310c2 R |
9214 | else if (!TARGET_SH4_300 |
9215 | && REG_NOTE_KIND (link) == REG_DEP_ANTI | |
c49439f1 | 9216 | && recog_memoized (dep_insn) >= 0 |
78d310c2 R |
9217 | && (get_attr_type (dep_insn) == TYPE_DFP_ARITH |
9218 | || get_attr_type (dep_insn) == TYPE_DFP_MUL) | |
c49439f1 R |
9219 | /* A lot of alleged anti-flow dependences are fake, |
9220 | so check this one is real. */ | |
9221 | && flow_dependent_p (dep_insn, insn)) | |
c237e94a | 9222 | cost = 2; |
c49439f1 | 9223 | |
c237e94a ZW |
9224 | return cost; |
9225 | } | |
3007d592 | 9226 | |
c49439f1 R |
9227 | /* Check if INSN is flow-dependent on DEP_INSN. Can also be used to check |
9228 | if DEP_INSN is anti-flow dependent on INSN. */ | |
9229 | static int | |
cf277499 | 9230 | flow_dependent_p (rtx insn, rtx dep_insn) |
c49439f1 R |
9231 | { |
9232 | rtx tmp = PATTERN (insn); | |
9233 | ||
9234 | note_stores (PATTERN (dep_insn), flow_dependent_p_1, &tmp); | |
9235 | return tmp == NULL_RTX; | |
9236 | } | |
9237 | ||
9238 | /* A helper function for flow_dependent_p called through note_stores. */ | |
9239 | static void | |
7bc980e1 | 9240 | flow_dependent_p_1 (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) |
c49439f1 R |
9241 | { |
9242 | rtx * pinsn = (rtx *) data; | |
9243 | ||
9244 | if (*pinsn && reg_referenced_p (x, *pinsn)) | |
9245 | *pinsn = NULL_RTX; | |
9246 | } | |
9247 | ||
b48f503c | 9248 | /* For use by sh_allocate_initial_value. Note that sh.md contains some |
3007d592 R |
9249 | 'special function' patterns (type sfunc) that clobber pr, but that |
9250 | do not look like function calls to leaf_function_p. Hence we must | |
9251 | do this extra check. */ | |
b48f503c | 9252 | static int |
cf277499 | 9253 | sh_pr_n_sets (void) |
3007d592 | 9254 | { |
6fb5fa3c | 9255 | return DF_REG_DEF_COUNT (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG); |
3007d592 | 9256 | } |
fa5322fa | 9257 | |
b48f503c KK |
9258 | /* Return where to allocate pseudo for a given hard register initial |
9259 | value. */ | |
9260 | static rtx | |
9261 | sh_allocate_initial_value (rtx hard_reg) | |
9262 | { | |
9263 | rtx x; | |
9264 | ||
9265 | if (REGNO (hard_reg) == (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG)) | |
9266 | { | |
9267 | if (current_function_is_leaf | |
9268 | && ! sh_pr_n_sets () | |
9269 | && ! (TARGET_SHCOMPACT | |
38173d38 | 9270 | && ((crtl->args.info.call_cookie |
b48f503c | 9271 | & ~ CALL_COOKIE_RET_TRAMP (1)) |
e3b5732b | 9272 | || crtl->saves_all_registers))) |
b48f503c KK |
9273 | x = hard_reg; |
9274 | else | |
57d38024 | 9275 | x = gen_frame_mem (Pmode, return_address_pointer_rtx); |
b48f503c KK |
9276 | } |
9277 | else | |
9278 | x = NULL_RTX; | |
9279 | ||
9280 | return x; | |
9281 | } | |
9282 | ||
c49439f1 R |
9283 | /* This function returns "2" to indicate dual issue for the SH4 |
9284 | processor. To be used by the DFA pipeline description. */ | |
fae15c93 | 9285 | static int |
832a3292 | 9286 | sh_issue_rate (void) |
fae15c93 | 9287 | { |
c49439f1 R |
9288 | if (TARGET_SUPERSCALAR) |
9289 | return 2; | |
9290 | else | |
9291 | return 1; | |
fae15c93 VM |
9292 | } |
9293 | ||
58565a33 SKG |
9294 | /* Functions for ready queue reordering for sched1. */ |
9295 | ||
9296 | /* Get weight for mode for a set x. */ | |
9297 | static short | |
508ea1c5 | 9298 | find_set_regmode_weight (rtx x, enum machine_mode mode) |
58565a33 SKG |
9299 | { |
9300 | if (GET_CODE (x) == CLOBBER && register_operand (SET_DEST (x), mode)) | |
9301 | return 1; | |
9302 | if (GET_CODE (x) == SET && register_operand (SET_DEST (x), mode)) | |
9303 | { | |
9304 | if (GET_CODE (SET_DEST (x)) == REG) | |
9305 | { | |
9306 | if (!reg_mentioned_p (SET_DEST (x), SET_SRC (x))) | |
9307 | return 1; | |
9308 | else | |
9309 | return 0; | |
9310 | } | |
9311 | return 1; | |
9312 | } | |
9313 | return 0; | |
9314 | } | |
9315 | ||
9316 | /* Get regmode weight for insn. */ | |
9317 | static short | |
508ea1c5 | 9318 | find_insn_regmode_weight (rtx insn, enum machine_mode mode) |
58565a33 SKG |
9319 | { |
9320 | short reg_weight = 0; | |
9321 | rtx x; | |
9322 | ||
9323 | /* Increment weight for each register born here. */ | |
9324 | x = PATTERN (insn); | |
9325 | reg_weight += find_set_regmode_weight (x, mode); | |
9326 | if (GET_CODE (x) == PARALLEL) | |
9327 | { | |
9328 | int j; | |
9329 | for (j = XVECLEN (x, 0) - 1; j >= 0; j--) | |
9330 | { | |
9331 | x = XVECEXP (PATTERN (insn), 0, j); | |
9332 | reg_weight += find_set_regmode_weight (x, mode); | |
9333 | } | |
9334 | } | |
9335 | /* Decrement weight for each register that dies here. */ | |
9336 | for (x = REG_NOTES (insn); x; x = XEXP (x, 1)) | |
9337 | { | |
9338 | if (REG_NOTE_KIND (x) == REG_DEAD || REG_NOTE_KIND (x) == REG_UNUSED) | |
9339 | { | |
9340 | rtx note = XEXP (x, 0); | |
9341 | if (GET_CODE (note) == REG && GET_MODE (note) == mode) | |
9342 | reg_weight--; | |
9343 | } | |
9344 | } | |
9345 | return reg_weight; | |
9346 | } | |
9347 | ||
9348 | /* Calculate regmode weights for all insns of a basic block. */ | |
9349 | static void | |
c6d56443 | 9350 | find_regmode_weight (basic_block b, enum machine_mode mode) |
58565a33 SKG |
9351 | { |
9352 | rtx insn, next_tail, head, tail; | |
9353 | ||
c6d56443 | 9354 | get_ebb_head_tail (b, b, &head, &tail); |
58565a33 SKG |
9355 | next_tail = NEXT_INSN (tail); |
9356 | ||
9357 | for (insn = head; insn != next_tail; insn = NEXT_INSN (insn)) | |
9358 | { | |
9359 | /* Handle register life information. */ | |
9360 | if (!INSN_P (insn)) | |
9361 | continue; | |
9362 | ||
9363 | if (mode == SFmode) | |
9364 | INSN_REGMODE_WEIGHT (insn, mode) = | |
9365 | find_insn_regmode_weight (insn, mode) + 2 * find_insn_regmode_weight (insn, DFmode); | |
9366 | else if (mode == SImode) | |
9367 | INSN_REGMODE_WEIGHT (insn, mode) = | |
9368 | find_insn_regmode_weight (insn, mode) + 2 * find_insn_regmode_weight (insn, DImode); | |
9369 | } | |
9370 | } | |
9371 | ||
9372 | /* Comparison function for ready queue sorting. */ | |
9373 | static int | |
508ea1c5 | 9374 | rank_for_reorder (const void *x, const void *y) |
58565a33 SKG |
9375 | { |
9376 | rtx tmp = *(const rtx *) y; | |
9377 | rtx tmp2 = *(const rtx *) x; | |
9378 | ||
9379 | /* The insn in a schedule group should be issued the first. */ | |
9380 | if (SCHED_GROUP_P (tmp) != SCHED_GROUP_P (tmp2)) | |
9381 | return SCHED_GROUP_P (tmp2) ? 1 : -1; | |
9382 | ||
f676971a | 9383 | /* If insns are equally good, sort by INSN_LUID (original insn order), This |
58565a33 SKG |
9384 | minimizes instruction movement, thus minimizing sched's effect on |
9385 | register pressure. */ | |
9386 | return INSN_LUID (tmp) - INSN_LUID (tmp2); | |
9387 | } | |
9388 | ||
9389 | /* Resort the array A in which only element at index N may be out of order. */ | |
9390 | static void | |
508ea1c5 | 9391 | swap_reorder (rtx *a, int n) |
58565a33 SKG |
9392 | { |
9393 | rtx insn = a[n - 1]; | |
9394 | int i = n - 2; | |
9395 | ||
9396 | while (i >= 0 && rank_for_reorder (a + i, &insn) >= 0) | |
9397 | { | |
9398 | a[i + 1] = a[i]; | |
9399 | i -= 1; | |
9400 | } | |
9401 | a[i + 1] = insn; | |
9402 | } | |
9403 | ||
832a3292 KH |
9404 | #define SCHED_REORDER(READY, N_READY) \ |
9405 | do \ | |
9406 | { \ | |
9407 | if ((N_READY) == 2) \ | |
9408 | swap_reorder (READY, N_READY); \ | |
9409 | else if ((N_READY) > 2) \ | |
9410 | qsort (READY, N_READY, sizeof (rtx), rank_for_reorder); \ | |
9411 | } \ | |
9412 | while (0) | |
58565a33 SKG |
9413 | |
9414 | /* Sort the ready list READY by ascending priority, using the SCHED_REORDER | |
9415 | macro. */ | |
9416 | static void | |
508ea1c5 | 9417 | ready_reorder (rtx *ready, int nready) |
58565a33 SKG |
9418 | { |
9419 | SCHED_REORDER (ready, nready); | |
9420 | } | |
9421 | ||
6fb5fa3c DB |
9422 | /* Count life regions of r0 for a block. */ |
9423 | static int | |
9424 | find_r0_life_regions (basic_block b) | |
9425 | { | |
9426 | rtx end, insn; | |
9427 | rtx pset; | |
9428 | rtx r0_reg; | |
9429 | int live; | |
9430 | int set; | |
9431 | int death = 0; | |
9432 | ||
89a95777 | 9433 | if (REGNO_REG_SET_P (df_get_live_in (b), R0_REG)) |
6fb5fa3c DB |
9434 | { |
9435 | set = 1; | |
9436 | live = 1; | |
9437 | } | |
9438 | else | |
9439 | { | |
9440 | set = 0; | |
9441 | live = 0; | |
9442 | } | |
9443 | ||
9444 | insn = BB_HEAD (b); | |
9445 | end = BB_END (b); | |
9446 | r0_reg = gen_rtx_REG (SImode, R0_REG); | |
9447 | while (1) | |
9448 | { | |
9449 | if (INSN_P (insn)) | |
9450 | { | |
9451 | if (find_regno_note (insn, REG_DEAD, R0_REG)) | |
9452 | { | |
9453 | death++; | |
9454 | live = 0; | |
9455 | } | |
9456 | if (!live | |
9457 | && (pset = single_set (insn)) | |
9458 | && reg_overlap_mentioned_p (r0_reg, SET_DEST (pset)) | |
9459 | && !find_regno_note (insn, REG_UNUSED, R0_REG)) | |
9460 | { | |
9461 | set++; | |
9462 | live = 1; | |
9463 | } | |
9464 | } | |
9465 | if (insn == end) | |
9466 | break; | |
9467 | insn = NEXT_INSN (insn); | |
9468 | } | |
9469 | return set - death; | |
9470 | } | |
9471 | ||
58565a33 SKG |
9472 | /* Calculate regmode weights for all insns of all basic block. */ |
9473 | static void | |
508ea1c5 KH |
9474 | sh_md_init_global (FILE *dump ATTRIBUTE_UNUSED, |
9475 | int verbose ATTRIBUTE_UNUSED, | |
9476 | int old_max_uid) | |
58565a33 SKG |
9477 | { |
9478 | basic_block b; | |
9479 | ||
9480 | regmode_weight[0] = (short *) xcalloc (old_max_uid, sizeof (short)); | |
9481 | regmode_weight[1] = (short *) xcalloc (old_max_uid, sizeof (short)); | |
6fb5fa3c | 9482 | r0_life_regions = 0; |
58565a33 SKG |
9483 | |
9484 | FOR_EACH_BB_REVERSE (b) | |
9485 | { | |
c6d56443 KK |
9486 | find_regmode_weight (b, SImode); |
9487 | find_regmode_weight (b, SFmode); | |
6fb5fa3c DB |
9488 | if (!reload_completed) |
9489 | r0_life_regions += find_r0_life_regions (b); | |
58565a33 SKG |
9490 | } |
9491 | ||
9492 | CURR_REGMODE_PRESSURE (SImode) = 0; | |
9493 | CURR_REGMODE_PRESSURE (SFmode) = 0; | |
9494 | ||
9495 | } | |
9496 | ||
9497 | /* Cleanup. */ | |
9498 | static void | |
508ea1c5 KH |
9499 | sh_md_finish_global (FILE *dump ATTRIBUTE_UNUSED, |
9500 | int verbose ATTRIBUTE_UNUSED) | |
58565a33 SKG |
9501 | { |
9502 | if (regmode_weight[0]) | |
9503 | { | |
9504 | free (regmode_weight[0]); | |
9505 | regmode_weight[0] = NULL; | |
9506 | } | |
9507 | if (regmode_weight[1]) | |
9508 | { | |
9509 | free (regmode_weight[1]); | |
9510 | regmode_weight[1] = NULL; | |
9511 | } | |
9512 | } | |
9513 | ||
b0fbea13 KK |
9514 | /* The scalar modes supported differs from the default version in TImode |
9515 | for 32-bit SHMEDIA. */ | |
9516 | static bool | |
9517 | sh_scalar_mode_supported_p (enum machine_mode mode) | |
9518 | { | |
9519 | if (TARGET_SHMEDIA32 && mode == TImode) | |
9520 | return false; | |
9521 | ||
9522 | return default_scalar_mode_supported_p (mode); | |
9523 | } | |
9524 | ||
58565a33 SKG |
9525 | /* Cache the can_issue_more so that we can return it from reorder2. Also, |
9526 | keep count of register pressures on SImode and SFmode. */ | |
9527 | static int | |
508ea1c5 KH |
9528 | sh_variable_issue (FILE *dump ATTRIBUTE_UNUSED, |
9529 | int sched_verbose ATTRIBUTE_UNUSED, | |
9530 | rtx insn, | |
9531 | int can_issue_more) | |
58565a33 SKG |
9532 | { |
9533 | if (GET_CODE (PATTERN (insn)) != USE | |
9534 | && GET_CODE (PATTERN (insn)) != CLOBBER) | |
9535 | cached_can_issue_more = can_issue_more - 1; | |
9536 | else | |
9537 | cached_can_issue_more = can_issue_more; | |
9538 | ||
9539 | if (reload_completed) | |
9540 | return cached_can_issue_more; | |
9541 | ||
9542 | CURR_REGMODE_PRESSURE (SImode) += INSN_REGMODE_WEIGHT (insn, SImode); | |
9543 | CURR_REGMODE_PRESSURE (SFmode) += INSN_REGMODE_WEIGHT (insn, SFmode); | |
9544 | ||
9545 | return cached_can_issue_more; | |
9546 | } | |
9547 | ||
9548 | static void | |
508ea1c5 KH |
9549 | sh_md_init (FILE *dump ATTRIBUTE_UNUSED, |
9550 | int verbose ATTRIBUTE_UNUSED, | |
9551 | int veclen ATTRIBUTE_UNUSED) | |
58565a33 SKG |
9552 | { |
9553 | CURR_REGMODE_PRESSURE (SImode) = 0; | |
9554 | CURR_REGMODE_PRESSURE (SFmode) = 0; | |
9555 | } | |
9556 | ||
9557 | /* Some magic numbers. */ | |
9558 | /* Pressure on register r0 can lead to spill failures. so avoid sched1 for | |
9559 | functions that already have high pressure on r0. */ | |
9560 | #define R0_MAX_LIFE_REGIONS 2 | |
4ee31f1e | 9561 | /* Register Pressure thresholds for SImode and SFmode registers. */ |
58565a33 SKG |
9562 | #define SIMODE_MAX_WEIGHT 5 |
9563 | #define SFMODE_MAX_WEIGHT 10 | |
9564 | ||
9565 | /* Return true if the pressure is high for MODE. */ | |
9566 | static short | |
508ea1c5 | 9567 | high_pressure (enum machine_mode mode) |
58565a33 SKG |
9568 | { |
9569 | /* Pressure on register r0 can lead to spill failures. so avoid sched1 for | |
9570 | functions that already have high pressure on r0. */ | |
6fb5fa3c DB |
9571 | if (r0_life_regions >= R0_MAX_LIFE_REGIONS) |
9572 | return 1; | |
58565a33 SKG |
9573 | |
9574 | if (mode == SFmode) | |
9575 | return (CURR_REGMODE_PRESSURE (SFmode) > SFMODE_MAX_WEIGHT); | |
9576 | else | |
9577 | return (CURR_REGMODE_PRESSURE (SImode) > SIMODE_MAX_WEIGHT); | |
9578 | } | |
9579 | ||
9580 | /* Reorder ready queue if register pressure is high. */ | |
9581 | static int | |
508ea1c5 KH |
9582 | sh_reorder (FILE *dump ATTRIBUTE_UNUSED, |
9583 | int sched_verbose ATTRIBUTE_UNUSED, | |
9584 | rtx *ready, | |
9585 | int *n_readyp, | |
9586 | int clock_var ATTRIBUTE_UNUSED) | |
58565a33 SKG |
9587 | { |
9588 | if (reload_completed) | |
9589 | return sh_issue_rate (); | |
9590 | ||
9591 | if (high_pressure (SFmode) || high_pressure (SImode)) | |
9592 | { | |
9593 | ready_reorder (ready, *n_readyp); | |
9594 | } | |
9595 | ||
9596 | return sh_issue_rate (); | |
9597 | } | |
9598 | ||
9599 | /* Skip cycles if the current register pressure is high. */ | |
f676971a | 9600 | static int |
508ea1c5 KH |
9601 | sh_reorder2 (FILE *dump ATTRIBUTE_UNUSED, |
9602 | int sched_verbose ATTRIBUTE_UNUSED, | |
9603 | rtx *ready ATTRIBUTE_UNUSED, | |
9604 | int *n_readyp ATTRIBUTE_UNUSED, | |
9605 | int clock_var ATTRIBUTE_UNUSED) | |
58565a33 SKG |
9606 | { |
9607 | if (reload_completed) | |
9608 | return cached_can_issue_more; | |
9609 | ||
f676971a | 9610 | if (high_pressure(SFmode) || high_pressure (SImode)) |
58565a33 SKG |
9611 | skip_cycles = 1; |
9612 | ||
9613 | return cached_can_issue_more; | |
9614 | } | |
9615 | ||
9616 | /* Skip cycles without sorting the ready queue. This will move insn from | |
9617 | Q->R. If this is the last cycle we are skipping; allow sorting of ready | |
f676971a | 9618 | queue by sh_reorder. */ |
58565a33 | 9619 | |
f676971a EC |
9620 | /* Generally, skipping these many cycles are sufficient for all insns to move |
9621 | from Q -> R. */ | |
9622 | #define MAX_SKIPS 8 | |
58565a33 SKG |
9623 | |
9624 | static int | |
508ea1c5 KH |
9625 | sh_dfa_new_cycle (FILE *sched_dump ATTRIBUTE_UNUSED, |
9626 | int sched_verbose ATTRIBUTE_UNUSED, | |
9627 | rtx insn ATTRIBUTE_UNUSED, | |
9628 | int last_clock_var, | |
9629 | int clock_var, | |
9630 | int *sort_p) | |
58565a33 SKG |
9631 | { |
9632 | if (reload_completed) | |
832a3292 | 9633 | return 0; |
58565a33 | 9634 | |
f676971a | 9635 | if (skip_cycles) |
58565a33 | 9636 | { |
832a3292 KH |
9637 | if ((clock_var - last_clock_var) < MAX_SKIPS) |
9638 | { | |
9639 | *sort_p = 0; | |
9640 | return 1; | |
9641 | } | |
9642 | /* If this is the last cycle we are skipping, allow reordering of R. */ | |
9643 | if ((clock_var - last_clock_var) == MAX_SKIPS) | |
9644 | { | |
9645 | *sort_p = 1; | |
9646 | return 1; | |
9647 | } | |
58565a33 | 9648 | } |
58565a33 SKG |
9649 | |
9650 | skip_cycles = 0; | |
9651 | ||
9652 | return 0; | |
9653 | } | |
9654 | ||
fa5322fa AO |
9655 | /* SHmedia requires registers for branches, so we can't generate new |
9656 | branches past reload. */ | |
9657 | static bool | |
cf277499 | 9658 | sh_cannot_modify_jumps_p (void) |
fa5322fa AO |
9659 | { |
9660 | return (TARGET_SHMEDIA && (reload_in_progress || reload_completed)); | |
9661 | } | |
9662 | ||
a3424f5c | 9663 | static int |
fe3ad572 SC |
9664 | sh_target_reg_class (void) |
9665 | { | |
9666 | return TARGET_SHMEDIA ? TARGET_REGS : NO_REGS; | |
9667 | } | |
9668 | ||
9669 | static bool | |
9670 | sh_optimize_target_register_callee_saved (bool after_prologue_epilogue_gen) | |
9671 | { | |
73a4d10b | 9672 | HARD_REG_SET dummy; |
cb3d3842 | 9673 | #if 0 |
73a4d10b | 9674 | rtx insn; |
cb3d3842 | 9675 | #endif |
73a4d10b R |
9676 | |
9677 | if (! shmedia_space_reserved_for_target_registers) | |
9678 | return 0; | |
9679 | if (after_prologue_epilogue_gen && ! TARGET_SAVE_ALL_TARGET_REGS) | |
9680 | return 0; | |
9681 | if (calc_live_regs (&dummy) >= 6 * 8) | |
9682 | return 1; | |
73a4d10b | 9683 | return 0; |
fe3ad572 SC |
9684 | } |
9685 | ||
fa5322fa | 9686 | static bool |
3101faab | 9687 | sh_ms_bitfield_layout_p (const_tree record_type ATTRIBUTE_UNUSED) |
fa5322fa | 9688 | { |
61f71b34 | 9689 | return (TARGET_SH5 || TARGET_HITACHI || sh_attr_renesas_p (record_type)); |
fa5322fa | 9690 | } |
ca903bba | 9691 | \f |
f676971a | 9692 | /* |
ca903bba R |
9693 | On the SH1..SH4, the trampoline looks like |
9694 | 2 0002 D202 mov.l l2,r2 | |
9695 | 1 0000 D301 mov.l l1,r3 | |
9696 | 3 0004 422B jmp @r2 | |
9697 | 4 0006 0009 nop | |
9698 | 5 0008 00000000 l1: .long area | |
9699 | 6 000c 00000000 l2: .long function | |
9700 | ||
9701 | SH5 (compact) uses r1 instead of r3 for the static chain. */ | |
9702 | ||
9703 | ||
9704 | /* Emit RTL insns to initialize the variable parts of a trampoline. | |
9705 | FNADDR is an RTX for the address of the function's pure code. | |
9706 | CXT is an RTX for the static chain value for the function. */ | |
9707 | ||
9708 | void | |
cf277499 | 9709 | sh_initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt) |
ca903bba | 9710 | { |
57d38024 R |
9711 | rtx tramp_mem = gen_frame_mem (BLKmode, tramp); |
9712 | ||
ca903bba R |
9713 | if (TARGET_SHMEDIA64) |
9714 | { | |
9715 | rtx tramp_templ; | |
9716 | int fixed_len; | |
9717 | ||
9718 | rtx movi1 = GEN_INT (0xcc000010); | |
9719 | rtx shori1 = GEN_INT (0xc8000010); | |
9720 | rtx src, dst; | |
9721 | ||
9722 | /* The following trampoline works within a +- 128 KB range for cxt: | |
9723 | ptb/u cxt,tr1; movi fnaddr >> 48,r0; shori fnaddr >> 32,r0; | |
9724 | shori fnaddr >> 16,r0; shori fnaddr,r0; ptabs/l r0,tr0 | |
9725 | gettr tr1,r1; blink tr0,r63 */ | |
9726 | /* Address rounding makes it hard to compute the exact bounds of the | |
9727 | offset for this trampoline, but we have a rather generous offset | |
9728 | range, so frame_offset should do fine as an upper bound. */ | |
9729 | if (cxt == virtual_stack_vars_rtx && frame_offset < 0x20000) | |
9730 | { | |
9731 | /* ??? could optimize this trampoline initialization | |
9732 | by writing DImode words with two insns each. */ | |
9733 | rtx mask = force_reg (DImode, GEN_INT (0x3fffc00)); | |
9734 | rtx insn = gen_rtx_MINUS (DImode, cxt, tramp); | |
9735 | insn = gen_rtx_ASHIFT (DImode, insn, GEN_INT (10-2)); | |
9736 | insn = gen_rtx_AND (DImode, insn, mask); | |
9737 | /* Or in ptb/u .,tr1 pattern */ | |
9738 | insn = gen_rtx_IOR (DImode, insn, gen_int_mode (0xec000010, SImode)); | |
9739 | insn = force_operand (insn, NULL_RTX); | |
9740 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 9741 | emit_move_insn (change_address (tramp_mem, SImode, NULL_RTX), insn); |
ca903bba R |
9742 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (38)); |
9743 | insn = gen_rtx_AND (DImode, insn, mask); | |
9744 | insn = force_operand (gen_rtx_IOR (DImode, movi1, insn), NULL_RTX); | |
9745 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 9746 | emit_move_insn (adjust_address (tramp_mem, SImode, 4), insn); |
ca903bba R |
9747 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (22)); |
9748 | insn = gen_rtx_AND (DImode, insn, mask); | |
9749 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
9750 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 9751 | emit_move_insn (adjust_address (tramp_mem, SImode, 8), insn); |
ca903bba R |
9752 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (6)); |
9753 | insn = gen_rtx_AND (DImode, insn, mask); | |
9754 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
9755 | insn = gen_lowpart (SImode, insn); | |
57d38024 | 9756 | emit_move_insn (adjust_address (tramp_mem, SImode, 12), insn); |
ca903bba R |
9757 | insn = gen_rtx_ASHIFT (DImode, fnaddr, GEN_INT (10)); |
9758 | insn = gen_rtx_AND (DImode, insn, mask); | |
9759 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
9760 | insn = gen_lowpart (SImode, insn); | |
57d38024 R |
9761 | emit_move_insn (adjust_address (tramp_mem, SImode, 16), insn); |
9762 | emit_move_insn (adjust_address (tramp_mem, SImode, 20), | |
ca903bba | 9763 | GEN_INT (0x6bf10600)); |
57d38024 | 9764 | emit_move_insn (adjust_address (tramp_mem, SImode, 24), |
ca903bba | 9765 | GEN_INT (0x4415fc10)); |
57d38024 | 9766 | emit_move_insn (adjust_address (tramp_mem, SImode, 28), |
ca903bba R |
9767 | GEN_INT (0x4401fff0)); |
9768 | emit_insn (gen_ic_invalidate_line (tramp)); | |
9769 | return; | |
9770 | } | |
9771 | tramp_templ = gen_rtx_SYMBOL_REF (Pmode,"__GCC_nested_trampoline"); | |
9772 | fixed_len = TRAMPOLINE_SIZE - 2 * GET_MODE_SIZE (Pmode); | |
9773 | ||
9774 | tramp_templ = gen_datalabel_ref (tramp_templ); | |
57d38024 R |
9775 | dst = tramp_mem; |
9776 | src = gen_const_mem (BLKmode, tramp_templ); | |
ca903bba R |
9777 | set_mem_align (dst, 256); |
9778 | set_mem_align (src, 64); | |
44bb111a | 9779 | emit_block_move (dst, src, GEN_INT (fixed_len), BLOCK_OP_NORMAL); |
ca903bba | 9780 | |
57d38024 R |
9781 | emit_move_insn (adjust_address (tramp_mem, Pmode, fixed_len), fnaddr); |
9782 | emit_move_insn (adjust_address (tramp_mem, Pmode, | |
9783 | fixed_len + GET_MODE_SIZE (Pmode)), | |
ca903bba R |
9784 | cxt); |
9785 | emit_insn (gen_ic_invalidate_line (tramp)); | |
9786 | return; | |
9787 | } | |
9788 | else if (TARGET_SHMEDIA) | |
9789 | { | |
9790 | /* movi fnaddr >> 16,r1; shori fnaddr,r1; ptabs/l r1,tr0 | |
9791 | movi cxt >> 16,r1; shori cxt,r1; blink tr0,r63 */ | |
9792 | rtx quad0 = gen_reg_rtx (DImode), cxtload = gen_reg_rtx (DImode); | |
9793 | rtx quad1 = gen_reg_rtx (DImode), quad2 = gen_reg_rtx (DImode); | |
9794 | /* movi 0,r1: 0xcc000010 shori 0,r1: c8000010 concatenated, | |
825db093 | 9795 | rotated 10 right, and higher 16 bit of every 32 selected. */ |
ca903bba R |
9796 | rtx movishori |
9797 | = force_reg (V2HImode, (simplify_gen_subreg | |
9798 | (V2HImode, GEN_INT (0x4330432), SImode, 0))); | |
9799 | rtx ptabs = force_reg (DImode, GEN_INT (0x6bf10600)); | |
9800 | rtx blink = force_reg (DImode, GEN_INT (0x4401fff0)); | |
9801 | ||
9802 | tramp = force_reg (Pmode, tramp); | |
9803 | fnaddr = force_reg (SImode, fnaddr); | |
9804 | cxt = force_reg (SImode, cxt); | |
9805 | emit_insn (gen_mshflo_w_x (gen_rtx_SUBREG (V4HImode, quad0, 0), | |
9806 | gen_rtx_SUBREG (V2HImode, fnaddr, 0), | |
9807 | movishori)); | |
f1e2239f | 9808 | emit_insn (gen_rotrdi3_mextr (quad0, quad0, |
ca903bba | 9809 | GEN_INT (TARGET_LITTLE_ENDIAN ? 24 : 56))); |
a556fd39 | 9810 | emit_insn (gen_ashldi3_media (quad0, quad0, const2_rtx)); |
57d38024 | 9811 | emit_move_insn (change_address (tramp_mem, DImode, NULL_RTX), quad0); |
ca903bba R |
9812 | emit_insn (gen_mshflo_w_x (gen_rtx_SUBREG (V4HImode, cxtload, 0), |
9813 | gen_rtx_SUBREG (V2HImode, cxt, 0), | |
9814 | movishori)); | |
f1e2239f | 9815 | emit_insn (gen_rotrdi3_mextr (cxtload, cxtload, |
ca903bba | 9816 | GEN_INT (TARGET_LITTLE_ENDIAN ? 24 : 56))); |
a556fd39 | 9817 | emit_insn (gen_ashldi3_media (cxtload, cxtload, const2_rtx)); |
ca903bba R |
9818 | if (TARGET_LITTLE_ENDIAN) |
9819 | { | |
9820 | emit_insn (gen_mshflo_l_di (quad1, ptabs, cxtload)); | |
9821 | emit_insn (gen_mextr4 (quad2, cxtload, blink)); | |
9822 | } | |
9823 | else | |
9824 | { | |
9825 | emit_insn (gen_mextr4 (quad1, cxtload, ptabs)); | |
9826 | emit_insn (gen_mshflo_l_di (quad2, blink, cxtload)); | |
9827 | } | |
57d38024 R |
9828 | emit_move_insn (adjust_address (tramp_mem, DImode, 8), quad1); |
9829 | emit_move_insn (adjust_address (tramp_mem, DImode, 16), quad2); | |
ca903bba R |
9830 | emit_insn (gen_ic_invalidate_line (tramp)); |
9831 | return; | |
9832 | } | |
9833 | else if (TARGET_SHCOMPACT) | |
9834 | { | |
9835 | emit_insn (gen_initialize_trampoline (tramp, cxt, fnaddr)); | |
9836 | return; | |
9837 | } | |
57d38024 | 9838 | emit_move_insn (change_address (tramp_mem, SImode, NULL_RTX), |
ca903bba R |
9839 | gen_int_mode (TARGET_LITTLE_ENDIAN ? 0xd301d202 : 0xd202d301, |
9840 | SImode)); | |
57d38024 | 9841 | emit_move_insn (adjust_address (tramp_mem, SImode, 4), |
ca903bba R |
9842 | gen_int_mode (TARGET_LITTLE_ENDIAN ? 0x0009422b : 0x422b0009, |
9843 | SImode)); | |
57d38024 R |
9844 | emit_move_insn (adjust_address (tramp_mem, SImode, 8), cxt); |
9845 | emit_move_insn (adjust_address (tramp_mem, SImode, 12), fnaddr); | |
ca903bba R |
9846 | if (TARGET_HARVARD) |
9847 | { | |
dc557046 | 9848 | if (!TARGET_INLINE_IC_INVALIDATE |
6f8eec29 | 9849 | || (!(TARGET_SH4A_ARCH || TARGET_SH4_300) && TARGET_USERMODE)) |
73a4d10b R |
9850 | emit_library_call (function_symbol (NULL, "__ic_invalidate", |
9851 | FUNCTION_ORDINARY), | |
ca903bba R |
9852 | 0, VOIDmode, 1, tramp, SImode); |
9853 | else | |
9854 | emit_insn (gen_ic_invalidate_line (tramp)); | |
9855 | } | |
9856 | } | |
9857 | ||
4977bab6 ZW |
9858 | /* FIXME: This is overly conservative. A SHcompact function that |
9859 | receives arguments ``by reference'' will have them stored in its | |
9860 | own stack frame, so it must not pass pointers or references to | |
9861 | these arguments to other functions by means of sibling calls. */ | |
73a4d10b R |
9862 | /* If PIC, we cannot make sibling calls to global functions |
9863 | because the PLT requires r12 to be live. */ | |
4977bab6 | 9864 | static bool |
cf277499 | 9865 | sh_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED) |
4977bab6 | 9866 | { |
73a4d10b | 9867 | return (1 |
4977bab6 | 9868 | && (! TARGET_SHCOMPACT |
38173d38 | 9869 | || crtl->args.info.stack_regs == 0) |
73a4d10b R |
9870 | && ! sh_cfun_interrupt_handler_p () |
9871 | && (! flag_pic | |
9872 | || (decl && ! TREE_PUBLIC (decl)) | |
9873 | || (decl && DECL_VISIBILITY (decl) != VISIBILITY_DEFAULT))); | |
4977bab6 | 9874 | } |
c1b92d09 R |
9875 | \f |
9876 | /* Machine specific built-in functions. */ | |
9877 | ||
9878 | struct builtin_description | |
9879 | { | |
9880 | const enum insn_code icode; | |
9881 | const char *const name; | |
9882 | int signature; | |
9883 | }; | |
9884 | ||
9885 | /* describe number and signedness of arguments; arg[0] == result | |
9886 | (1: unsigned, 2: signed, 4: don't care, 8: pointer 0: no argument */ | |
3b962ea0 | 9887 | /* 9: 64-bit pointer, 10: 32-bit pointer */ |
c1b92d09 R |
9888 | static const char signature_args[][4] = |
9889 | { | |
9890 | #define SH_BLTIN_V2SI2 0 | |
9891 | { 4, 4 }, | |
9892 | #define SH_BLTIN_V4HI2 1 | |
9893 | { 4, 4 }, | |
9894 | #define SH_BLTIN_V2SI3 2 | |
9895 | { 4, 4, 4 }, | |
9896 | #define SH_BLTIN_V4HI3 3 | |
9897 | { 4, 4, 4 }, | |
9898 | #define SH_BLTIN_V8QI3 4 | |
9899 | { 4, 4, 4 }, | |
9900 | #define SH_BLTIN_MAC_HISI 5 | |
9901 | { 1, 4, 4, 1 }, | |
9902 | #define SH_BLTIN_SH_HI 6 | |
9903 | { 4, 4, 1 }, | |
9904 | #define SH_BLTIN_SH_SI 7 | |
9905 | { 4, 4, 1 }, | |
9906 | #define SH_BLTIN_V4HI2V2SI 8 | |
9907 | { 4, 4, 4 }, | |
9908 | #define SH_BLTIN_V4HI2V8QI 9 | |
9909 | { 4, 4, 4 }, | |
9910 | #define SH_BLTIN_SISF 10 | |
9911 | { 4, 2 }, | |
9912 | #define SH_BLTIN_LDUA_L 11 | |
73a4d10b | 9913 | { 2, 10 }, |
c1b92d09 | 9914 | #define SH_BLTIN_LDUA_Q 12 |
73a4d10b | 9915 | { 1, 10 }, |
c1b92d09 | 9916 | #define SH_BLTIN_STUA_L 13 |
73a4d10b | 9917 | { 0, 10, 2 }, |
c1b92d09 | 9918 | #define SH_BLTIN_STUA_Q 14 |
73a4d10b R |
9919 | { 0, 10, 1 }, |
9920 | #define SH_BLTIN_LDUA_L64 15 | |
9921 | { 2, 9 }, | |
9922 | #define SH_BLTIN_LDUA_Q64 16 | |
9923 | { 1, 9 }, | |
9924 | #define SH_BLTIN_STUA_L64 17 | |
9925 | { 0, 9, 2 }, | |
9926 | #define SH_BLTIN_STUA_Q64 18 | |
9927 | { 0, 9, 1 }, | |
9928 | #define SH_BLTIN_NUM_SHARED_SIGNATURES 19 | |
9929 | #define SH_BLTIN_2 19 | |
9930 | #define SH_BLTIN_SU 19 | |
c1b92d09 | 9931 | { 1, 2 }, |
73a4d10b R |
9932 | #define SH_BLTIN_3 20 |
9933 | #define SH_BLTIN_SUS 20 | |
c1b92d09 | 9934 | { 2, 2, 1 }, |
73a4d10b | 9935 | #define SH_BLTIN_PSSV 21 |
c1b92d09 | 9936 | { 0, 8, 2, 2 }, |
73a4d10b R |
9937 | #define SH_BLTIN_XXUU 22 |
9938 | #define SH_BLTIN_UUUU 22 | |
c1b92d09 | 9939 | { 1, 1, 1, 1 }, |
73a4d10b | 9940 | #define SH_BLTIN_PV 23 |
c1b92d09 R |
9941 | { 0, 8 }, |
9942 | }; | |
24746a42 | 9943 | /* mcmv: operands considered unsigned. */ |
c1b92d09 | 9944 | /* mmulsum_wq, msad_ubq: result considered unsigned long long. */ |
24746a42 KH |
9945 | /* mperm: control value considered unsigned int. */ |
9946 | /* mshalds, mshard, mshards, mshlld, mshlrd: shift count is unsigned int. */ | |
c1b92d09 R |
9947 | /* mshards_q: returns signed short. */ |
9948 | /* nsb: takes long long arg, returns unsigned char. */ | |
9949 | static const struct builtin_description bdesc[] = | |
9950 | { | |
9951 | { CODE_FOR_absv2si2, "__builtin_absv2si2", SH_BLTIN_V2SI2 }, | |
9952 | { CODE_FOR_absv4hi2, "__builtin_absv4hi2", SH_BLTIN_V4HI2 }, | |
9953 | { CODE_FOR_addv2si3, "__builtin_addv2si3", SH_BLTIN_V2SI3 }, | |
9954 | { CODE_FOR_addv4hi3, "__builtin_addv4hi3", SH_BLTIN_V4HI3 }, | |
9955 | { CODE_FOR_ssaddv2si3,"__builtin_ssaddv2si3", SH_BLTIN_V2SI3 }, | |
9956 | { CODE_FOR_usaddv8qi3,"__builtin_usaddv8qi3", SH_BLTIN_V8QI3 }, | |
9957 | { CODE_FOR_ssaddv4hi3,"__builtin_ssaddv4hi3", SH_BLTIN_V4HI3 }, | |
73a4d10b | 9958 | { CODE_FOR_alloco_i, "__builtin_sh_media_ALLOCO", SH_BLTIN_PV }, |
c1b92d09 R |
9959 | { CODE_FOR_negcmpeqv8qi,"__builtin_sh_media_MCMPEQ_B", SH_BLTIN_V8QI3 }, |
9960 | { CODE_FOR_negcmpeqv2si,"__builtin_sh_media_MCMPEQ_L", SH_BLTIN_V2SI3 }, | |
9961 | { CODE_FOR_negcmpeqv4hi,"__builtin_sh_media_MCMPEQ_W", SH_BLTIN_V4HI3 }, | |
9962 | { CODE_FOR_negcmpgtuv8qi,"__builtin_sh_media_MCMPGT_UB", SH_BLTIN_V8QI3 }, | |
9963 | { CODE_FOR_negcmpgtv2si,"__builtin_sh_media_MCMPGT_L", SH_BLTIN_V2SI3 }, | |
9964 | { CODE_FOR_negcmpgtv4hi,"__builtin_sh_media_MCMPGT_W", SH_BLTIN_V4HI3 }, | |
9965 | { CODE_FOR_mcmv, "__builtin_sh_media_MCMV", SH_BLTIN_UUUU }, | |
9966 | { CODE_FOR_mcnvs_lw, "__builtin_sh_media_MCNVS_LW", SH_BLTIN_3 }, | |
9967 | { CODE_FOR_mcnvs_wb, "__builtin_sh_media_MCNVS_WB", SH_BLTIN_V4HI2V8QI }, | |
9968 | { CODE_FOR_mcnvs_wub, "__builtin_sh_media_MCNVS_WUB", SH_BLTIN_V4HI2V8QI }, | |
73a4d10b R |
9969 | { CODE_FOR_mextr1, "__builtin_sh_media_MEXTR1", SH_BLTIN_V8QI3 }, |
9970 | { CODE_FOR_mextr2, "__builtin_sh_media_MEXTR2", SH_BLTIN_V8QI3 }, | |
9971 | { CODE_FOR_mextr3, "__builtin_sh_media_MEXTR3", SH_BLTIN_V8QI3 }, | |
9972 | { CODE_FOR_mextr4, "__builtin_sh_media_MEXTR4", SH_BLTIN_V8QI3 }, | |
9973 | { CODE_FOR_mextr5, "__builtin_sh_media_MEXTR5", SH_BLTIN_V8QI3 }, | |
9974 | { CODE_FOR_mextr6, "__builtin_sh_media_MEXTR6", SH_BLTIN_V8QI3 }, | |
9975 | { CODE_FOR_mextr7, "__builtin_sh_media_MEXTR7", SH_BLTIN_V8QI3 }, | |
c1b92d09 R |
9976 | { CODE_FOR_mmacfx_wl, "__builtin_sh_media_MMACFX_WL", SH_BLTIN_MAC_HISI }, |
9977 | { CODE_FOR_mmacnfx_wl,"__builtin_sh_media_MMACNFX_WL", SH_BLTIN_MAC_HISI }, | |
9978 | { CODE_FOR_mulv2si3, "__builtin_mulv2si3", SH_BLTIN_V2SI3, }, | |
9979 | { CODE_FOR_mulv4hi3, "__builtin_mulv4hi3", SH_BLTIN_V4HI3 }, | |
9980 | { CODE_FOR_mmulfx_l, "__builtin_sh_media_MMULFX_L", SH_BLTIN_V2SI3 }, | |
9981 | { CODE_FOR_mmulfx_w, "__builtin_sh_media_MMULFX_W", SH_BLTIN_V4HI3 }, | |
9982 | { CODE_FOR_mmulfxrp_w,"__builtin_sh_media_MMULFXRP_W", SH_BLTIN_V4HI3 }, | |
9983 | { CODE_FOR_mmulhi_wl, "__builtin_sh_media_MMULHI_WL", SH_BLTIN_V4HI2V2SI }, | |
9984 | { CODE_FOR_mmullo_wl, "__builtin_sh_media_MMULLO_WL", SH_BLTIN_V4HI2V2SI }, | |
9985 | { CODE_FOR_mmulsum_wq,"__builtin_sh_media_MMULSUM_WQ", SH_BLTIN_XXUU }, | |
9986 | { CODE_FOR_mperm_w, "__builtin_sh_media_MPERM_W", SH_BLTIN_SH_HI }, | |
9987 | { CODE_FOR_msad_ubq, "__builtin_sh_media_MSAD_UBQ", SH_BLTIN_XXUU }, | |
9988 | { CODE_FOR_mshalds_l, "__builtin_sh_media_MSHALDS_L", SH_BLTIN_SH_SI }, | |
9989 | { CODE_FOR_mshalds_w, "__builtin_sh_media_MSHALDS_W", SH_BLTIN_SH_HI }, | |
9990 | { CODE_FOR_ashrv2si3, "__builtin_ashrv2si3", SH_BLTIN_SH_SI }, | |
9991 | { CODE_FOR_ashrv4hi3, "__builtin_ashrv4hi3", SH_BLTIN_SH_HI }, | |
9992 | { CODE_FOR_mshards_q, "__builtin_sh_media_MSHARDS_Q", SH_BLTIN_SUS }, | |
9993 | { CODE_FOR_mshfhi_b, "__builtin_sh_media_MSHFHI_B", SH_BLTIN_V8QI3 }, | |
9994 | { CODE_FOR_mshfhi_l, "__builtin_sh_media_MSHFHI_L", SH_BLTIN_V2SI3 }, | |
9995 | { CODE_FOR_mshfhi_w, "__builtin_sh_media_MSHFHI_W", SH_BLTIN_V4HI3 }, | |
9996 | { CODE_FOR_mshflo_b, "__builtin_sh_media_MSHFLO_B", SH_BLTIN_V8QI3 }, | |
9997 | { CODE_FOR_mshflo_l, "__builtin_sh_media_MSHFLO_L", SH_BLTIN_V2SI3 }, | |
9998 | { CODE_FOR_mshflo_w, "__builtin_sh_media_MSHFLO_W", SH_BLTIN_V4HI3 }, | |
9999 | { CODE_FOR_ashlv2si3, "__builtin_ashlv2si3", SH_BLTIN_SH_SI }, | |
10000 | { CODE_FOR_ashlv4hi3, "__builtin_ashlv4hi3", SH_BLTIN_SH_HI }, | |
10001 | { CODE_FOR_lshrv2si3, "__builtin_lshrv2si3", SH_BLTIN_SH_SI }, | |
10002 | { CODE_FOR_lshrv4hi3, "__builtin_lshrv4hi3", SH_BLTIN_SH_HI }, | |
10003 | { CODE_FOR_subv2si3, "__builtin_subv2si3", SH_BLTIN_V2SI3 }, | |
10004 | { CODE_FOR_subv4hi3, "__builtin_subv4hi3", SH_BLTIN_V4HI3 }, | |
10005 | { CODE_FOR_sssubv2si3,"__builtin_sssubv2si3", SH_BLTIN_V2SI3 }, | |
10006 | { CODE_FOR_ussubv8qi3,"__builtin_ussubv8qi3", SH_BLTIN_V8QI3 }, | |
10007 | { CODE_FOR_sssubv4hi3,"__builtin_sssubv4hi3", SH_BLTIN_V4HI3 }, | |
10008 | { CODE_FOR_fcosa_s, "__builtin_sh_media_FCOSA_S", SH_BLTIN_SISF }, | |
10009 | { CODE_FOR_fsina_s, "__builtin_sh_media_FSINA_S", SH_BLTIN_SISF }, | |
10010 | { CODE_FOR_fipr, "__builtin_sh_media_FIPR_S", SH_BLTIN_3 }, | |
10011 | { CODE_FOR_ftrv, "__builtin_sh_media_FTRV_S", SH_BLTIN_3 }, | |
73a4d10b R |
10012 | { CODE_FOR_mac_media, "__builtin_sh_media_FMAC_S", SH_BLTIN_3 }, |
10013 | { CODE_FOR_sqrtdf2, "__builtin_sh_media_FSQRT_D", SH_BLTIN_2 }, | |
10014 | { CODE_FOR_sqrtsf2, "__builtin_sh_media_FSQRT_S", SH_BLTIN_2 }, | |
c1b92d09 | 10015 | { CODE_FOR_fsrra_s, "__builtin_sh_media_FSRRA_S", SH_BLTIN_2 }, |
c1b92d09 R |
10016 | { CODE_FOR_ldhi_l, "__builtin_sh_media_LDHI_L", SH_BLTIN_LDUA_L }, |
10017 | { CODE_FOR_ldhi_q, "__builtin_sh_media_LDHI_Q", SH_BLTIN_LDUA_Q }, | |
10018 | { CODE_FOR_ldlo_l, "__builtin_sh_media_LDLO_L", SH_BLTIN_LDUA_L }, | |
10019 | { CODE_FOR_ldlo_q, "__builtin_sh_media_LDLO_Q", SH_BLTIN_LDUA_Q }, | |
10020 | { CODE_FOR_sthi_l, "__builtin_sh_media_STHI_L", SH_BLTIN_STUA_L }, | |
10021 | { CODE_FOR_sthi_q, "__builtin_sh_media_STHI_Q", SH_BLTIN_STUA_Q }, | |
10022 | { CODE_FOR_stlo_l, "__builtin_sh_media_STLO_L", SH_BLTIN_STUA_L }, | |
10023 | { CODE_FOR_stlo_q, "__builtin_sh_media_STLO_Q", SH_BLTIN_STUA_Q }, | |
73a4d10b R |
10024 | { CODE_FOR_ldhi_l64, "__builtin_sh_media_LDHI_L", SH_BLTIN_LDUA_L64 }, |
10025 | { CODE_FOR_ldhi_q64, "__builtin_sh_media_LDHI_Q", SH_BLTIN_LDUA_Q64 }, | |
10026 | { CODE_FOR_ldlo_l64, "__builtin_sh_media_LDLO_L", SH_BLTIN_LDUA_L64 }, | |
10027 | { CODE_FOR_ldlo_q64, "__builtin_sh_media_LDLO_Q", SH_BLTIN_LDUA_Q64 }, | |
10028 | { CODE_FOR_sthi_l64, "__builtin_sh_media_STHI_L", SH_BLTIN_STUA_L64 }, | |
10029 | { CODE_FOR_sthi_q64, "__builtin_sh_media_STHI_Q", SH_BLTIN_STUA_Q64 }, | |
10030 | { CODE_FOR_stlo_l64, "__builtin_sh_media_STLO_L", SH_BLTIN_STUA_L64 }, | |
10031 | { CODE_FOR_stlo_q64, "__builtin_sh_media_STLO_Q", SH_BLTIN_STUA_Q64 }, | |
c1b92d09 R |
10032 | { CODE_FOR_nsb, "__builtin_sh_media_NSB", SH_BLTIN_SU }, |
10033 | { CODE_FOR_byterev, "__builtin_sh_media_BYTEREV", SH_BLTIN_2 }, | |
73a4d10b | 10034 | { CODE_FOR_prefetch, "__builtin_sh_media_PREFO", SH_BLTIN_PSSV }, |
c1b92d09 R |
10035 | }; |
10036 | ||
10037 | static void | |
cf277499 | 10038 | sh_media_init_builtins (void) |
c1b92d09 R |
10039 | { |
10040 | tree shared[SH_BLTIN_NUM_SHARED_SIGNATURES]; | |
10041 | const struct builtin_description *d; | |
10042 | ||
a3acdc0c | 10043 | memset (shared, 0, sizeof shared); |
e97a46ce | 10044 | for (d = bdesc; d - bdesc < (int) ARRAY_SIZE (bdesc); d++) |
c1b92d09 | 10045 | { |
73a4d10b | 10046 | tree type, arg_type = 0; |
c1b92d09 R |
10047 | int signature = d->signature; |
10048 | int i; | |
10049 | ||
10050 | if (signature < SH_BLTIN_NUM_SHARED_SIGNATURES && shared[signature]) | |
10051 | type = shared[signature]; | |
10052 | else | |
10053 | { | |
10054 | int has_result = signature_args[signature][0] != 0; | |
10055 | ||
73a4d10b R |
10056 | if ((signature_args[signature][1] & 8) |
10057 | && (((signature_args[signature][1] & 1) && TARGET_SHMEDIA32) | |
10058 | || ((signature_args[signature][1] & 2) && TARGET_SHMEDIA64))) | |
c1b92d09 R |
10059 | continue; |
10060 | if (! TARGET_FPU_ANY | |
10061 | && FLOAT_MODE_P (insn_data[d->icode].operand[0].mode)) | |
10062 | continue; | |
10063 | type = void_list_node; | |
10064 | for (i = 3; ; i--) | |
10065 | { | |
10066 | int arg = signature_args[signature][i]; | |
10067 | int opno = i - 1 + has_result; | |
10068 | ||
73a4d10b | 10069 | if (arg & 8) |
c1b92d09 R |
10070 | arg_type = ptr_type_node; |
10071 | else if (arg) | |
73a4d10b R |
10072 | arg_type = (*lang_hooks.types.type_for_mode) |
10073 | (insn_data[d->icode].operand[opno].mode, | |
10074 | (arg & 1)); | |
c1b92d09 R |
10075 | else if (i) |
10076 | continue; | |
10077 | else | |
10078 | arg_type = void_type_node; | |
10079 | if (i == 0) | |
10080 | break; | |
10081 | type = tree_cons (NULL_TREE, arg_type, type); | |
10082 | } | |
10083 | type = build_function_type (arg_type, type); | |
10084 | if (signature < SH_BLTIN_NUM_SHARED_SIGNATURES) | |
10085 | shared[signature] = type; | |
10086 | } | |
c79efc4d RÁE |
10087 | add_builtin_function (d->name, type, d - bdesc, BUILT_IN_MD, |
10088 | NULL, NULL_TREE); | |
c1b92d09 R |
10089 | } |
10090 | } | |
10091 | ||
f676971a EC |
10092 | /* Implements target hook vector_mode_supported_p. */ |
10093 | bool | |
10094 | sh_vector_mode_supported_p (enum machine_mode mode) | |
10095 | { | |
10096 | if (TARGET_FPU_ANY | |
10097 | && ((mode == V2SFmode) | |
10098 | || (mode == V4SFmode) | |
10099 | || (mode == V16SFmode))) | |
10100 | return true; | |
10101 | ||
10102 | else if (TARGET_SHMEDIA | |
10103 | && ((mode == V8QImode) | |
10104 | || (mode == V2HImode) | |
10105 | || (mode == V4HImode) | |
10106 | || (mode == V2SImode))) | |
10107 | return true; | |
10108 | ||
10109 | return false; | |
10110 | } | |
10111 | ||
a1c496cb EC |
10112 | /* Implements target hook dwarf_calling_convention. Return an enum |
10113 | of dwarf_calling_convention. */ | |
10114 | int | |
3101faab | 10115 | sh_dwarf_calling_convention (const_tree func) |
a1c496cb EC |
10116 | { |
10117 | if (sh_attr_renesas_p (func)) | |
29832164 | 10118 | return DW_CC_GNU_renesas_sh; |
a1c496cb EC |
10119 | |
10120 | return DW_CC_normal; | |
10121 | } | |
10122 | ||
c1b92d09 | 10123 | static void |
cf277499 | 10124 | sh_init_builtins (void) |
c1b92d09 R |
10125 | { |
10126 | if (TARGET_SHMEDIA) | |
10127 | sh_media_init_builtins (); | |
10128 | } | |
10129 | ||
10130 | /* Expand an expression EXP that calls a built-in function, | |
10131 | with result going to TARGET if that's convenient | |
10132 | (and in mode MODE if that's convenient). | |
10133 | SUBTARGET may be used as the target for computing one of EXP's operands. | |
10134 | IGNORE is nonzero if the value is to be ignored. */ | |
10135 | ||
10136 | static rtx | |
cf277499 SB |
10137 | sh_expand_builtin (tree exp, rtx target, rtx subtarget ATTRIBUTE_UNUSED, |
10138 | enum machine_mode mode ATTRIBUTE_UNUSED, int ignore) | |
c1b92d09 | 10139 | { |
5039610b | 10140 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); |
c1b92d09 R |
10141 | unsigned int fcode = DECL_FUNCTION_CODE (fndecl); |
10142 | const struct builtin_description *d = &bdesc[fcode]; | |
10143 | enum insn_code icode = d->icode; | |
10144 | int signature = d->signature; | |
10145 | enum machine_mode tmode = VOIDmode; | |
10146 | int nop = 0, i; | |
10147 | rtx op[4]; | |
73a4d10b | 10148 | rtx pat = 0; |
c1b92d09 R |
10149 | |
10150 | if (signature_args[signature][0]) | |
10151 | { | |
10152 | if (ignore) | |
10153 | return 0; | |
10154 | ||
10155 | tmode = insn_data[icode].operand[0].mode; | |
10156 | if (! target | |
10157 | || GET_MODE (target) != tmode | |
10158 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
10159 | target = gen_reg_rtx (tmode); | |
10160 | op[nop++] = target; | |
10161 | } | |
10162 | else | |
10163 | target = 0; | |
10164 | ||
10165 | for (i = 1; i <= 3; i++, nop++) | |
10166 | { | |
10167 | tree arg; | |
10168 | enum machine_mode opmode, argmode; | |
73a4d10b | 10169 | tree optype; |
c1b92d09 R |
10170 | |
10171 | if (! signature_args[signature][i]) | |
10172 | break; | |
5039610b | 10173 | arg = CALL_EXPR_ARG (exp, i - 1); |
3578cf63 R |
10174 | if (arg == error_mark_node) |
10175 | return const0_rtx; | |
73a4d10b R |
10176 | if (signature_args[signature][i] & 8) |
10177 | { | |
10178 | opmode = ptr_mode; | |
10179 | optype = ptr_type_node; | |
10180 | } | |
10181 | else | |
10182 | { | |
10183 | opmode = insn_data[icode].operand[nop].mode; | |
10184 | optype = (*lang_hooks.types.type_for_mode) (opmode, 0); | |
10185 | } | |
c1b92d09 R |
10186 | argmode = TYPE_MODE (TREE_TYPE (arg)); |
10187 | if (argmode != opmode) | |
73a4d10b | 10188 | arg = build1 (NOP_EXPR, optype, arg); |
c1b92d09 R |
10189 | op[nop] = expand_expr (arg, NULL_RTX, opmode, 0); |
10190 | if (! (*insn_data[icode].operand[nop].predicate) (op[nop], opmode)) | |
10191 | op[nop] = copy_to_mode_reg (opmode, op[nop]); | |
10192 | } | |
10193 | ||
10194 | switch (nop) | |
10195 | { | |
10196 | case 1: | |
10197 | pat = (*insn_data[d->icode].genfun) (op[0]); | |
10198 | break; | |
10199 | case 2: | |
10200 | pat = (*insn_data[d->icode].genfun) (op[0], op[1]); | |
10201 | break; | |
10202 | case 3: | |
10203 | pat = (*insn_data[d->icode].genfun) (op[0], op[1], op[2]); | |
10204 | break; | |
10205 | case 4: | |
10206 | pat = (*insn_data[d->icode].genfun) (op[0], op[1], op[2], op[3]); | |
10207 | break; | |
95c0af87 | 10208 | default: |
f5b9e7c9 | 10209 | gcc_unreachable (); |
c1b92d09 R |
10210 | } |
10211 | if (! pat) | |
10212 | return 0; | |
10213 | emit_insn (pat); | |
10214 | return target; | |
10215 | } | |
0ac78517 R |
10216 | |
10217 | void | |
cf277499 | 10218 | sh_expand_unop_v2sf (enum rtx_code code, rtx op0, rtx op1) |
0ac78517 R |
10219 | { |
10220 | rtx sel0 = const0_rtx; | |
10221 | rtx sel1 = const1_rtx; | |
cf277499 | 10222 | rtx (*fn) (rtx, rtx, rtx, rtx, rtx) = gen_unary_sf_op; |
0ac78517 R |
10223 | rtx op = gen_rtx_fmt_e (code, SFmode, op1); |
10224 | ||
10225 | emit_insn ((*fn) (op0, op1, op, sel0, sel0)); | |
10226 | emit_insn ((*fn) (op0, op1, op, sel1, sel1)); | |
10227 | } | |
10228 | ||
10229 | void | |
cf277499 | 10230 | sh_expand_binop_v2sf (enum rtx_code code, rtx op0, rtx op1, rtx op2) |
0ac78517 | 10231 | { |
0ac78517 R |
10232 | rtx op = gen_rtx_fmt_ee (code, SFmode, op1, op2); |
10233 | ||
f8be3f56 KK |
10234 | emit_insn (gen_binary_sf_op0 (op0, op1, op2, op)); |
10235 | emit_insn (gen_binary_sf_op1 (op0, op1, op2, op)); | |
0ac78517 R |
10236 | } |
10237 | ||
cff9f8d5 AH |
10238 | /* Return the class of registers for which a mode change from FROM to TO |
10239 | is invalid. */ | |
b0c42aed | 10240 | bool |
cf277499 SB |
10241 | sh_cannot_change_mode_class (enum machine_mode from, enum machine_mode to, |
10242 | enum reg_class class) | |
cff9f8d5 | 10243 | { |
312209c6 AO |
10244 | /* We want to enable the use of SUBREGs as a means to |
10245 | VEC_SELECT a single element of a vector. */ | |
10246 | if (to == SFmode && VECTOR_MODE_P (from) && GET_MODE_INNER (from) == SFmode) | |
10247 | return (reg_classes_intersect_p (GENERAL_REGS, class)); | |
10248 | ||
cff9f8d5 AH |
10249 | if (GET_MODE_SIZE (from) != GET_MODE_SIZE (to)) |
10250 | { | |
832a3292 KH |
10251 | if (TARGET_LITTLE_ENDIAN) |
10252 | { | |
10253 | if (GET_MODE_SIZE (to) < 8 || GET_MODE_SIZE (from) < 8) | |
10254 | return reg_classes_intersect_p (DF_REGS, class); | |
10255 | } | |
10256 | else | |
10257 | { | |
10258 | if (GET_MODE_SIZE (from) < 8) | |
10259 | return reg_classes_intersect_p (DF_HI_REGS, class); | |
10260 | } | |
cff9f8d5 | 10261 | } |
b0c42aed | 10262 | return 0; |
cff9f8d5 AH |
10263 | } |
10264 | ||
ea4210ef RS |
10265 | |
10266 | /* If ADDRESS refers to a CODE_LABEL, add NUSES to the number of times | |
10267 | that label is used. */ | |
10268 | ||
10269 | void | |
cf277499 | 10270 | sh_mark_label (rtx address, int nuses) |
ea4210ef RS |
10271 | { |
10272 | if (GOTOFF_P (address)) | |
10273 | { | |
10274 | /* Extract the label or symbol. */ | |
10275 | address = XEXP (address, 0); | |
10276 | if (GET_CODE (address) == PLUS) | |
10277 | address = XEXP (address, 0); | |
10278 | address = XVECEXP (address, 0, 0); | |
10279 | } | |
10280 | if (GET_CODE (address) == LABEL_REF | |
10281 | && GET_CODE (XEXP (address, 0)) == CODE_LABEL) | |
10282 | LABEL_NUSES (XEXP (address, 0)) += nuses; | |
10283 | } | |
10284 | ||
95c0af87 JR |
10285 | /* Compute extra cost of moving data between one register class |
10286 | and another. */ | |
10287 | ||
ec555f32 R |
10288 | /* If SECONDARY*_RELOAD_CLASS says something about the src/dst pair, regclass |
10289 | uses this information. Hence, the general register <-> floating point | |
95c0af87 JR |
10290 | register information here is not used for SFmode. */ |
10291 | ||
10292 | int | |
cf277499 SB |
10293 | sh_register_move_cost (enum machine_mode mode, |
10294 | enum reg_class srcclass, enum reg_class dstclass) | |
95c0af87 JR |
10295 | { |
10296 | if (dstclass == T_REGS || dstclass == PR_REGS) | |
10297 | return 10; | |
10298 | ||
b1dcf523 DD |
10299 | if (dstclass == MAC_REGS && srcclass == MAC_REGS) |
10300 | return 4; | |
10301 | ||
ec555f32 R |
10302 | if (mode == SImode && ! TARGET_SHMEDIA && TARGET_FMOVD |
10303 | && REGCLASS_HAS_FP_REG (srcclass) | |
10304 | && REGCLASS_HAS_FP_REG (dstclass)) | |
10305 | return 4; | |
10306 | ||
9c24200c KK |
10307 | if (REGCLASS_HAS_FP_REG (dstclass) && srcclass == T_REGS) |
10308 | return ((TARGET_HARD_SH4 && !optimize_size) ? 10 : 7); | |
10309 | ||
a60a9553 | 10310 | if ((REGCLASS_HAS_FP_REG (dstclass) && srcclass == MAC_REGS) |
832a3292 | 10311 | || (dstclass == MAC_REGS && REGCLASS_HAS_FP_REG (srcclass))) |
a60a9553 R |
10312 | return 9; |
10313 | ||
95c0af87 JR |
10314 | if ((REGCLASS_HAS_FP_REG (dstclass) |
10315 | && REGCLASS_HAS_GENERAL_REG (srcclass)) | |
10316 | || (REGCLASS_HAS_GENERAL_REG (dstclass) | |
10317 | && REGCLASS_HAS_FP_REG (srcclass))) | |
832a3292 KH |
10318 | return ((TARGET_SHMEDIA ? 4 : TARGET_FMOVD ? 8 : 12) |
10319 | * ((GET_MODE_SIZE (mode) + 7) / 8U)); | |
95c0af87 JR |
10320 | |
10321 | if ((dstclass == FPUL_REGS | |
10322 | && REGCLASS_HAS_GENERAL_REG (srcclass)) | |
10323 | || (srcclass == FPUL_REGS | |
10324 | && REGCLASS_HAS_GENERAL_REG (dstclass))) | |
10325 | return 5; | |
10326 | ||
10327 | if ((dstclass == FPUL_REGS | |
10328 | && (srcclass == PR_REGS || srcclass == MAC_REGS || srcclass == T_REGS)) | |
f676971a | 10329 | || (srcclass == FPUL_REGS |
95c0af87 JR |
10330 | && (dstclass == PR_REGS || dstclass == MAC_REGS))) |
10331 | return 7; | |
10332 | ||
10333 | if ((srcclass == TARGET_REGS && ! REGCLASS_HAS_GENERAL_REG (dstclass)) | |
10334 | || ((dstclass) == TARGET_REGS && ! REGCLASS_HAS_GENERAL_REG (srcclass))) | |
10335 | return 20; | |
10336 | ||
73a4d10b R |
10337 | /* ??? ptabs faults on (value & 0x3) == 0x3 */ |
10338 | if (TARGET_SHMEDIA | |
10339 | && ((srcclass) == TARGET_REGS || (srcclass) == SIBCALL_REGS)) | |
10340 | { | |
7dd2f19b RS |
10341 | if (sh_gettrcost >= 0) |
10342 | return sh_gettrcost; | |
73a4d10b R |
10343 | else if (!TARGET_PT_FIXED) |
10344 | return 100; | |
10345 | } | |
10346 | ||
95c0af87 JR |
10347 | if ((srcclass == FPSCR_REGS && ! REGCLASS_HAS_GENERAL_REG (dstclass)) |
10348 | || (dstclass == FPSCR_REGS && ! REGCLASS_HAS_GENERAL_REG (srcclass))) | |
10349 | return 4; | |
10350 | ||
ec555f32 R |
10351 | if (TARGET_SHMEDIA |
10352 | || (TARGET_FMOVD | |
10353 | && ! REGCLASS_HAS_GENERAL_REG (srcclass) | |
10354 | && ! REGCLASS_HAS_GENERAL_REG (dstclass))) | |
10355 | return 2 * ((GET_MODE_SIZE (mode) + 7) / 8U); | |
10356 | ||
10357 | return 2 * ((GET_MODE_SIZE (mode) + 3) / 4U); | |
95c0af87 JR |
10358 | } |
10359 | ||
cf277499 | 10360 | static rtx emit_load_ptr (rtx, rtx); |
1aa03f38 R |
10361 | |
10362 | static rtx | |
cf277499 | 10363 | emit_load_ptr (rtx reg, rtx addr) |
1aa03f38 | 10364 | { |
57d38024 | 10365 | rtx mem = gen_const_mem (ptr_mode, addr); |
1aa03f38 R |
10366 | |
10367 | if (Pmode != ptr_mode) | |
10368 | mem = gen_rtx_SIGN_EXTEND (Pmode, mem); | |
10369 | return emit_move_insn (reg, mem); | |
10370 | } | |
10371 | ||
73a4d10b | 10372 | static void |
cf277499 SB |
10373 | sh_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED, |
10374 | HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset, | |
10375 | tree function) | |
1aa03f38 R |
10376 | { |
10377 | CUMULATIVE_ARGS cum; | |
10378 | int structure_value_byref = 0; | |
10379 | rtx this, this_value, sibcall, insns, funexp; | |
10380 | tree funtype = TREE_TYPE (function); | |
735cb76e | 10381 | int simple_add = CONST_OK_FOR_ADD (delta); |
1aa03f38 R |
10382 | int did_load = 0; |
10383 | rtx scratch0, scratch1, scratch2; | |
73a4d10b | 10384 | unsigned i; |
1aa03f38 R |
10385 | |
10386 | reload_completed = 1; | |
fe3ad572 | 10387 | epilogue_completed = 1; |
1aa03f38 R |
10388 | current_function_uses_only_leaf_regs = 1; |
10389 | ||
2e040219 | 10390 | emit_note (NOTE_INSN_PROLOGUE_END); |
1aa03f38 R |
10391 | |
10392 | /* Find the "this" pointer. We have such a wide range of ABIs for the | |
10393 | SH that it's best to do this completely machine independently. | |
f676971a | 10394 | "this" is passed as first argument, unless a structure return pointer |
1aa03f38 | 10395 | comes first, in which case "this" comes second. */ |
0f6937fe | 10396 | INIT_CUMULATIVE_ARGS (cum, funtype, NULL_RTX, 0, 1); |
1aa03f38 | 10397 | #ifndef PCC_STATIC_STRUCT_RETURN |
61f71b34 | 10398 | if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function)) |
1aa03f38 R |
10399 | structure_value_byref = 1; |
10400 | #endif /* not PCC_STATIC_STRUCT_RETURN */ | |
61f71b34 | 10401 | if (structure_value_byref && sh_struct_value_rtx (function, 0) == 0) |
f676971a | 10402 | { |
1aa03f38 R |
10403 | tree ptype = build_pointer_type (TREE_TYPE (funtype)); |
10404 | ||
10405 | FUNCTION_ARG_ADVANCE (cum, Pmode, ptype, 1); | |
10406 | } | |
10407 | this = FUNCTION_ARG (cum, Pmode, ptr_type_node, 1); | |
10408 | ||
10409 | /* For SHcompact, we only have r0 for a scratch register: r1 is the | |
10410 | static chain pointer (even if you can't have nested virtual functions | |
10411 | right now, someone might implement them sometime), and the rest of the | |
10412 | registers are used for argument passing, are callee-saved, or reserved. */ | |
73a4d10b R |
10413 | /* We need to check call_used_regs / fixed_regs in case -fcall_saved-reg / |
10414 | -ffixed-reg has been used. */ | |
10415 | if (! call_used_regs[0] || fixed_regs[0]) | |
10416 | error ("r0 needs to be available as a call-clobbered register"); | |
1aa03f38 R |
10417 | scratch0 = scratch1 = scratch2 = gen_rtx_REG (Pmode, 0); |
10418 | if (! TARGET_SH5) | |
10419 | { | |
73a4d10b R |
10420 | if (call_used_regs[1] && ! fixed_regs[1]) |
10421 | scratch1 = gen_rtx_REG (ptr_mode, 1); | |
1aa03f38 R |
10422 | /* N.B., if not TARGET_HITACHI, register 2 is used to pass the pointer |
10423 | pointing where to return struct values. */ | |
73a4d10b R |
10424 | if (call_used_regs[3] && ! fixed_regs[3]) |
10425 | scratch2 = gen_rtx_REG (Pmode, 3); | |
1aa03f38 R |
10426 | } |
10427 | else if (TARGET_SHMEDIA) | |
10428 | { | |
73a4d10b R |
10429 | for (i = FIRST_GENERAL_REG; i <= LAST_GENERAL_REG; i++) |
10430 | if (i != REGNO (scratch0) && | |
10431 | call_used_regs[i] && ! fixed_regs[i] && ! FUNCTION_ARG_REGNO_P (i)) | |
10432 | { | |
10433 | scratch1 = gen_rtx_REG (ptr_mode, i); | |
10434 | break; | |
10435 | } | |
10436 | if (scratch1 == scratch0) | |
10437 | error ("Need a second call-clobbered general purpose register"); | |
10438 | for (i = FIRST_TARGET_REG; i <= LAST_TARGET_REG; i++) | |
10439 | if (call_used_regs[i] && ! fixed_regs[i]) | |
10440 | { | |
10441 | scratch2 = gen_rtx_REG (Pmode, i); | |
10442 | break; | |
10443 | } | |
10444 | if (scratch2 == scratch0) | |
10445 | error ("Need a call-clobbered target register"); | |
1aa03f38 R |
10446 | } |
10447 | ||
10448 | this_value = plus_constant (this, delta); | |
10449 | if (vcall_offset | |
10450 | && (simple_add || scratch0 != scratch1) | |
10451 | && strict_memory_address_p (ptr_mode, this_value)) | |
10452 | { | |
10453 | emit_load_ptr (scratch0, this_value); | |
10454 | did_load = 1; | |
10455 | } | |
10456 | ||
10457 | if (!delta) | |
10458 | ; /* Do nothing. */ | |
10459 | else if (simple_add) | |
10460 | emit_move_insn (this, this_value); | |
10461 | else | |
10462 | { | |
10463 | emit_move_insn (scratch1, GEN_INT (delta)); | |
10464 | emit_insn (gen_add2_insn (this, scratch1)); | |
10465 | } | |
10466 | ||
10467 | if (vcall_offset) | |
10468 | { | |
10469 | rtx offset_addr; | |
10470 | ||
10471 | if (!did_load) | |
10472 | emit_load_ptr (scratch0, this); | |
10473 | ||
10474 | offset_addr = plus_constant (scratch0, vcall_offset); | |
10475 | if (strict_memory_address_p (ptr_mode, offset_addr)) | |
10476 | ; /* Do nothing. */ | |
73a4d10b | 10477 | else if (! TARGET_SH5 && scratch0 != scratch1) |
1aa03f38 R |
10478 | { |
10479 | /* scratch0 != scratch1, and we have indexed loads. Get better | |
10480 | schedule by loading the offset into r1 and using an indexed | |
10481 | load - then the load of r1 can issue before the load from | |
10482 | (this + delta) finishes. */ | |
10483 | emit_move_insn (scratch1, GEN_INT (vcall_offset)); | |
10484 | offset_addr = gen_rtx_PLUS (Pmode, scratch0, scratch1); | |
10485 | } | |
735cb76e | 10486 | else if (CONST_OK_FOR_ADD (vcall_offset)) |
1aa03f38 R |
10487 | { |
10488 | emit_insn (gen_add2_insn (scratch0, GEN_INT (vcall_offset))); | |
10489 | offset_addr = scratch0; | |
10490 | } | |
10491 | else if (scratch0 != scratch1) | |
10492 | { | |
10493 | emit_move_insn (scratch1, GEN_INT (vcall_offset)); | |
10494 | emit_insn (gen_add2_insn (scratch0, scratch1)); | |
10495 | offset_addr = scratch0; | |
10496 | } | |
10497 | else | |
f5b9e7c9 | 10498 | gcc_unreachable (); /* FIXME */ |
1aa03f38 R |
10499 | emit_load_ptr (scratch0, offset_addr); |
10500 | ||
832a3292 | 10501 | if (Pmode != ptr_mode) |
1aa03f38 R |
10502 | scratch0 = gen_rtx_TRUNCATE (ptr_mode, scratch0); |
10503 | emit_insn (gen_add2_insn (this, scratch0)); | |
10504 | } | |
10505 | ||
10506 | /* Generate a tail call to the target function. */ | |
10507 | if (! TREE_USED (function)) | |
10508 | { | |
10509 | assemble_external (function); | |
10510 | TREE_USED (function) = 1; | |
10511 | } | |
10512 | funexp = XEXP (DECL_RTL (function), 0); | |
73a4d10b R |
10513 | /* If the function is overridden, so is the thunk, hence we don't |
10514 | need GOT addressing even if this is a public symbol. */ | |
10515 | #if 0 | |
10516 | if (TARGET_SH1 && ! flag_weak) | |
10517 | sibcall = gen_sibcalli_thunk (funexp, const0_rtx); | |
10518 | else | |
10519 | #endif | |
10520 | if (TARGET_SH2 && flag_pic) | |
10521 | { | |
10522 | sibcall = gen_sibcall_pcrel (funexp, const0_rtx); | |
10523 | XEXP (XVECEXP (sibcall, 0, 2), 0) = scratch2; | |
10524 | } | |
10525 | else | |
10526 | { | |
10527 | if (TARGET_SHMEDIA && flag_pic) | |
10528 | { | |
10529 | funexp = gen_sym2PIC (funexp); | |
10530 | PUT_MODE (funexp, Pmode); | |
10531 | } | |
10532 | emit_move_insn (scratch2, funexp); | |
10533 | funexp = gen_rtx_MEM (FUNCTION_MODE, scratch2); | |
10534 | sibcall = gen_sibcall (funexp, const0_rtx, NULL_RTX); | |
10535 | } | |
10536 | sibcall = emit_call_insn (sibcall); | |
1aa03f38 R |
10537 | SIBLING_CALL_P (sibcall) = 1; |
10538 | use_reg (&CALL_INSN_FUNCTION_USAGE (sibcall), this); | |
10539 | emit_barrier (); | |
10540 | ||
a2855205 | 10541 | /* Run just enough of rest_of_compilation to do scheduling and get |
1aa03f38 R |
10542 | the insns emitted. Note that use_thunk calls |
10543 | assemble_start_function and assemble_end_function. */ | |
a2855205 | 10544 | |
55e092c4 | 10545 | insn_locators_alloc (); |
1aa03f38 R |
10546 | insns = get_insns (); |
10547 | ||
6fb5fa3c | 10548 | #if 0 |
384a8760 | 10549 | if (optimize > 0) |
1aa03f38 | 10550 | { |
c9ffaa8d KK |
10551 | /* Initialize the bitmap obstacks. */ |
10552 | bitmap_obstack_initialize (NULL); | |
10553 | bitmap_obstack_initialize (®_obstack); | |
652ad062 | 10554 | if (! cfun->cfg) |
03904fb3 KK |
10555 | init_flow (); |
10556 | rtl_register_cfg_hooks (); | |
384a8760 KK |
10557 | init_rtl_bb_info (ENTRY_BLOCK_PTR); |
10558 | init_rtl_bb_info (EXIT_BLOCK_PTR); | |
10559 | ENTRY_BLOCK_PTR->flags |= BB_RTL; | |
10560 | EXIT_BLOCK_PTR->flags |= BB_RTL; | |
de332a85 | 10561 | find_basic_blocks (insns); |
1aa03f38 | 10562 | |
384a8760 KK |
10563 | if (flag_schedule_insns_after_reload) |
10564 | { | |
10d22567 | 10565 | life_analysis (PROP_FINAL); |
384a8760 KK |
10566 | |
10567 | split_all_insns (1); | |
1aa03f38 | 10568 | |
10d22567 | 10569 | schedule_insns (); |
384a8760 KK |
10570 | } |
10571 | /* We must split jmp insn in PIC case. */ | |
10572 | else if (flag_pic) | |
10573 | split_all_insns_noflow (); | |
1aa03f38 | 10574 | } |
6fb5fa3c DB |
10575 | #else |
10576 | if (optimize > 0) | |
10577 | { | |
10578 | if (! cfun->cfg) | |
eb6cbf91 | 10579 | init_flow (cfun); |
6fb5fa3c DB |
10580 | split_all_insns_noflow (); |
10581 | } | |
10582 | #endif | |
1aa03f38 | 10583 | |
18dbd950 | 10584 | sh_reorg (); |
1aa03f38 R |
10585 | |
10586 | if (optimize > 0 && flag_delayed_branch) | |
10d22567 | 10587 | dbr_schedule (insns); |
384a8760 | 10588 | |
1aa03f38 R |
10589 | shorten_branches (insns); |
10590 | final_start_function (insns, file, 1); | |
c9d691e9 | 10591 | final (insns, file, 1); |
1aa03f38 | 10592 | final_end_function (); |
2a7428c0 | 10593 | free_after_compilation (cfun); |
1aa03f38 | 10594 | |
1aa03f38 | 10595 | reload_completed = 0; |
fe3ad572 | 10596 | epilogue_completed = 0; |
1aa03f38 R |
10597 | } |
10598 | ||
90534361 | 10599 | rtx |
73a4d10b | 10600 | function_symbol (rtx target, const char *name, enum sh_function_kind kind) |
90534361 | 10601 | { |
73a4d10b R |
10602 | rtx sym; |
10603 | ||
10604 | /* If this is not an ordinary function, the name usually comes from a | |
10605 | string literal or an sprintf buffer. Make sure we use the same | |
10606 | string consistently, so that cse will be able to unify address loads. */ | |
10607 | if (kind != FUNCTION_ORDINARY) | |
10608 | name = IDENTIFIER_POINTER (get_identifier (name)); | |
10609 | sym = gen_rtx_SYMBOL_REF (Pmode, name); | |
90534361 | 10610 | SYMBOL_REF_FLAGS (sym) = SYMBOL_FLAG_FUNCTION; |
73a4d10b R |
10611 | if (flag_pic) |
10612 | switch (kind) | |
10613 | { | |
10614 | case FUNCTION_ORDINARY: | |
10615 | break; | |
10616 | case SFUNC_GOT: | |
10617 | { | |
10618 | rtx reg = target ? target : gen_reg_rtx (Pmode); | |
10619 | ||
10620 | emit_insn (gen_symGOT2reg (reg, sym)); | |
10621 | sym = reg; | |
10622 | break; | |
10623 | } | |
10624 | case SFUNC_STATIC: | |
10625 | { | |
10626 | /* ??? To allow cse to work, we use GOTOFF relocations. | |
10627 | we could add combiner patterns to transform this into | |
10628 | straight pc-relative calls with sym2PIC / bsrf when | |
10629 | label load and function call are still 1:1 and in the | |
10630 | same basic block during combine. */ | |
10631 | rtx reg = target ? target : gen_reg_rtx (Pmode); | |
10632 | ||
10633 | emit_insn (gen_symGOTOFF2reg (reg, sym)); | |
10634 | sym = reg; | |
10635 | break; | |
10636 | } | |
10637 | } | |
10638 | if (target && sym != target) | |
10639 | { | |
10640 | emit_move_insn (target, sym); | |
10641 | return target; | |
10642 | } | |
90534361 R |
10643 | return sym; |
10644 | } | |
10645 | ||
fada1961 R |
10646 | /* Find the number of a general purpose register in S. */ |
10647 | static int | |
10648 | scavenge_reg (HARD_REG_SET *s) | |
10649 | { | |
10650 | int r; | |
10651 | for (r = FIRST_GENERAL_REG; r <= LAST_GENERAL_REG; r++) | |
10652 | if (TEST_HARD_REG_BIT (*s, r)) | |
10653 | return r; | |
10654 | return -1; | |
10655 | } | |
10656 | ||
10657 | rtx | |
10658 | sh_get_pr_initial_val (void) | |
10659 | { | |
7d73a2ba R |
10660 | rtx val; |
10661 | ||
b869f904 R |
10662 | /* ??? Unfortunately, get_hard_reg_initial_val doesn't always work for the |
10663 | PR register on SHcompact, because it might be clobbered by the prologue. | |
7d73a2ba | 10664 | We check first if that is known to be the case. */ |
b869f904 | 10665 | if (TARGET_SHCOMPACT |
38173d38 | 10666 | && ((crtl->args.info.call_cookie |
7d73a2ba | 10667 | & ~ CALL_COOKIE_RET_TRAMP (1)) |
e3b5732b | 10668 | || crtl->saves_all_registers)) |
57d38024 | 10669 | return gen_frame_mem (SImode, return_address_pointer_rtx); |
7d73a2ba R |
10670 | |
10671 | /* If we haven't finished rtl generation, there might be a nonlocal label | |
10672 | that we haven't seen yet. | |
b3a13419 ILT |
10673 | ??? get_hard_reg_initial_val fails if it is called after register |
10674 | allocation has started, unless it has been called before for the | |
10675 | same register. And even then, we end in trouble if we didn't use | |
10676 | the register in the same basic block before. So call | |
10677 | get_hard_reg_initial_val now and wrap it in an unspec if we might | |
10678 | need to replace it. */ | |
dce20bbc R |
10679 | /* ??? We also must do this for TARGET_SH1 in general, because otherwise |
10680 | combine can put the pseudo returned by get_hard_reg_initial_val into | |
10681 | instructions that need a general purpose registers, which will fail to | |
10682 | be recognized when the pseudo becomes allocated to PR. */ | |
7d73a2ba R |
10683 | val |
10684 | = get_hard_reg_initial_val (Pmode, TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG); | |
dce20bbc | 10685 | if (TARGET_SH1) |
7d73a2ba R |
10686 | return gen_rtx_UNSPEC (SImode, gen_rtvec (1, val), UNSPEC_RA); |
10687 | return val; | |
fada1961 R |
10688 | } |
10689 | ||
3db1b434 R |
10690 | int |
10691 | sh_expand_t_scc (enum rtx_code code, rtx target) | |
10692 | { | |
10693 | rtx result = target; | |
10694 | HOST_WIDE_INT val; | |
10695 | ||
10696 | if (GET_CODE (sh_compare_op0) != REG || REGNO (sh_compare_op0) != T_REG | |
10697 | || GET_CODE (sh_compare_op1) != CONST_INT) | |
10698 | return 0; | |
10699 | if (GET_CODE (result) != REG) | |
10700 | result = gen_reg_rtx (SImode); | |
10701 | val = INTVAL (sh_compare_op1); | |
10702 | if ((code == EQ && val == 1) || (code == NE && val == 0)) | |
10703 | emit_insn (gen_movt (result)); | |
f326a6cb N |
10704 | else if (TARGET_SH2A && ((code == EQ && val == 0) |
10705 | || (code == NE && val == 1))) | |
10706 | emit_insn (gen_movrt (result)); | |
3db1b434 R |
10707 | else if ((code == EQ && val == 0) || (code == NE && val == 1)) |
10708 | { | |
c41c1387 | 10709 | emit_clobber (result); |
3db1b434 | 10710 | emit_insn (gen_subc (result, result, result)); |
a556fd39 | 10711 | emit_insn (gen_addsi3 (result, result, const1_rtx)); |
3db1b434 R |
10712 | } |
10713 | else if (code == EQ || code == NE) | |
10714 | emit_insn (gen_move_insn (result, GEN_INT (code == NE))); | |
10715 | else | |
10716 | return 0; | |
10717 | if (result != target) | |
10718 | emit_move_insn (target, result); | |
10719 | return 1; | |
10720 | } | |
10721 | ||
07d7d2f4 R |
10722 | /* INSN is an sfunc; return the rtx that describes the address used. */ |
10723 | static rtx | |
10724 | extract_sfunc_addr (rtx insn) | |
10725 | { | |
10726 | rtx pattern, part = NULL_RTX; | |
10727 | int len, i; | |
10728 | ||
10729 | pattern = PATTERN (insn); | |
10730 | len = XVECLEN (pattern, 0); | |
10731 | for (i = 0; i < len; i++) | |
10732 | { | |
10733 | part = XVECEXP (pattern, 0, i); | |
10734 | if (GET_CODE (part) == USE && GET_MODE (XEXP (part, 0)) == Pmode | |
10735 | && GENERAL_REGISTER_P (true_regnum (XEXP (part, 0)))) | |
10736 | return XEXP (part, 0); | |
10737 | } | |
f5b9e7c9 NS |
10738 | gcc_assert (GET_CODE (XVECEXP (pattern, 0, 0)) == UNSPEC_VOLATILE); |
10739 | return XVECEXP (XVECEXP (pattern, 0, 0), 0, 1); | |
07d7d2f4 R |
10740 | } |
10741 | ||
10742 | /* Verify that the register in use_sfunc_addr still agrees with the address | |
10743 | used in the sfunc. This prevents fill_slots_from_thread from changing | |
10744 | use_sfunc_addr. | |
10745 | INSN is the use_sfunc_addr instruction, and REG is the register it | |
10746 | guards. */ | |
10747 | int | |
10748 | check_use_sfunc_addr (rtx insn, rtx reg) | |
10749 | { | |
10750 | /* Search for the sfunc. It should really come right after INSN. */ | |
10751 | while ((insn = NEXT_INSN (insn))) | |
10752 | { | |
10753 | if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN) | |
10754 | break; | |
10755 | if (! INSN_P (insn)) | |
10756 | continue; | |
f676971a | 10757 | |
07d7d2f4 R |
10758 | if (GET_CODE (PATTERN (insn)) == SEQUENCE) |
10759 | insn = XVECEXP (PATTERN (insn), 0, 0); | |
10760 | if (GET_CODE (PATTERN (insn)) != PARALLEL | |
10761 | || get_attr_type (insn) != TYPE_SFUNC) | |
10762 | continue; | |
10763 | return rtx_equal_p (extract_sfunc_addr (insn), reg); | |
10764 | } | |
f5b9e7c9 | 10765 | gcc_unreachable (); |
07d7d2f4 R |
10766 | } |
10767 | ||
312209c6 AO |
10768 | /* This function returns a constant rtx that represents pi / 2**15 in |
10769 | SFmode. it's used to scale SFmode angles, in radians, to a | |
10770 | fixed-point signed 16.16-bit fraction of a full circle, i.e., 2*pi | |
10771 | maps to 0x10000). */ | |
10772 | ||
10773 | static GTY(()) rtx sh_fsca_sf2int_rtx; | |
10774 | ||
10775 | rtx | |
10776 | sh_fsca_sf2int (void) | |
10777 | { | |
10778 | if (! sh_fsca_sf2int_rtx) | |
10779 | { | |
10780 | REAL_VALUE_TYPE rv; | |
10781 | ||
10782 | real_from_string (&rv, "10430.378350470453"); | |
10783 | sh_fsca_sf2int_rtx = const_double_from_real_value (rv, SFmode); | |
10784 | } | |
10785 | ||
10786 | return sh_fsca_sf2int_rtx; | |
10787 | } | |
f676971a | 10788 | |
312209c6 AO |
10789 | /* This function returns a constant rtx that represents pi / 2**15 in |
10790 | DFmode. it's used to scale DFmode angles, in radians, to a | |
10791 | fixed-point signed 16.16-bit fraction of a full circle, i.e., 2*pi | |
10792 | maps to 0x10000). */ | |
10793 | ||
10794 | static GTY(()) rtx sh_fsca_df2int_rtx; | |
10795 | ||
10796 | rtx | |
10797 | sh_fsca_df2int (void) | |
10798 | { | |
10799 | if (! sh_fsca_df2int_rtx) | |
10800 | { | |
10801 | REAL_VALUE_TYPE rv; | |
10802 | ||
10803 | real_from_string (&rv, "10430.378350470453"); | |
10804 | sh_fsca_df2int_rtx = const_double_from_real_value (rv, DFmode); | |
10805 | } | |
10806 | ||
10807 | return sh_fsca_df2int_rtx; | |
10808 | } | |
f676971a | 10809 | |
312209c6 AO |
10810 | /* This function returns a constant rtx that represents 2**15 / pi in |
10811 | SFmode. it's used to scale a fixed-point signed 16.16-bit fraction | |
10812 | of a full circle back to a SFmode value, i.e., 0x10000 maps to | |
10813 | 2*pi). */ | |
10814 | ||
10815 | static GTY(()) rtx sh_fsca_int2sf_rtx; | |
10816 | ||
10817 | rtx | |
10818 | sh_fsca_int2sf (void) | |
10819 | { | |
10820 | if (! sh_fsca_int2sf_rtx) | |
10821 | { | |
10822 | REAL_VALUE_TYPE rv; | |
10823 | ||
10824 | real_from_string (&rv, "9.587379924285257e-5"); | |
10825 | sh_fsca_int2sf_rtx = const_double_from_real_value (rv, SFmode); | |
10826 | } | |
10827 | ||
10828 | return sh_fsca_int2sf_rtx; | |
10829 | } | |
cc15e98f | 10830 | |
8b0add27 | 10831 | /* Initialize the CUMULATIVE_ARGS structure. */ |
cc15e98f NC |
10832 | |
10833 | void | |
10834 | sh_init_cumulative_args (CUMULATIVE_ARGS * pcum, | |
10835 | tree fntype, | |
10836 | rtx libname ATTRIBUTE_UNUSED, | |
10837 | tree fndecl, | |
10838 | signed int n_named_args, | |
10839 | enum machine_mode mode) | |
10840 | { | |
10841 | pcum->arg_count [(int) SH_ARG_FLOAT] = 0; | |
10842 | pcum->free_single_fp_reg = 0; | |
10843 | pcum->stack_regs = 0; | |
10844 | pcum->byref_regs = 0; | |
10845 | pcum->byref = 0; | |
10846 | pcum->outgoing = (n_named_args == -1) ? 0 : 1; | |
10847 | ||
10848 | /* XXX - Should we check TARGET_HITACHI here ??? */ | |
10849 | pcum->renesas_abi = sh_attr_renesas_p (fntype) ? 1 : 0; | |
10850 | ||
10851 | if (fntype) | |
10852 | { | |
10853 | pcum->force_mem = ((TARGET_HITACHI || pcum->renesas_abi) | |
10854 | && aggregate_value_p (TREE_TYPE (fntype), fndecl)); | |
10855 | pcum->prototype_p = TYPE_ARG_TYPES (fntype) ? TRUE : FALSE; | |
10856 | pcum->arg_count [(int) SH_ARG_INT] | |
10857 | = TARGET_SH5 && aggregate_value_p (TREE_TYPE (fntype), fndecl); | |
10858 | ||
10859 | pcum->call_cookie | |
10860 | = CALL_COOKIE_RET_TRAMP (TARGET_SHCOMPACT | |
10861 | && pcum->arg_count [(int) SH_ARG_INT] == 0 | |
10862 | && (TYPE_MODE (TREE_TYPE (fntype)) == BLKmode | |
10863 | ? int_size_in_bytes (TREE_TYPE (fntype)) | |
10864 | : GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (fntype)))) > 4 | |
10865 | && (BASE_RETURN_VALUE_REG (TYPE_MODE (TREE_TYPE (fntype))) | |
10866 | == FIRST_RET_REG)); | |
10867 | } | |
10868 | else | |
10869 | { | |
10870 | pcum->arg_count [(int) SH_ARG_INT] = 0; | |
10871 | pcum->prototype_p = FALSE; | |
10872 | if (mode != VOIDmode) | |
10873 | { | |
29832164 | 10874 | pcum->call_cookie = |
cc15e98f NC |
10875 | CALL_COOKIE_RET_TRAMP (TARGET_SHCOMPACT |
10876 | && GET_MODE_SIZE (mode) > 4 | |
10877 | && BASE_RETURN_VALUE_REG (mode) == FIRST_RET_REG); | |
10878 | ||
10879 | /* If the default ABI is the Renesas ABI then all library | |
10880 | calls must assume that the library will be using the | |
10881 | Renesas ABI. So if the function would return its result | |
10882 | in memory then we must force the address of this memory | |
10883 | block onto the stack. Ideally we would like to call | |
10884 | targetm.calls.return_in_memory() here but we do not have | |
8b0add27 | 10885 | the TYPE or the FNDECL available so we synthesize the |
cc15e98f NC |
10886 | contents of that function as best we can. */ |
10887 | pcum->force_mem = | |
c0fb94d7 | 10888 | (TARGET_DEFAULT & MASK_HITACHI) |
cc15e98f NC |
10889 | && (mode == BLKmode |
10890 | || (GET_MODE_SIZE (mode) > 4 | |
10891 | && !(mode == DFmode | |
10892 | && TARGET_FPU_DOUBLE))); | |
10893 | } | |
10894 | else | |
10895 | { | |
10896 | pcum->call_cookie = 0; | |
10897 | pcum->force_mem = FALSE; | |
10898 | } | |
10899 | } | |
10900 | } | |
10901 | ||
73a4d10b R |
10902 | /* Replace any occurrence of FROM(n) in X with TO(n). The function does |
10903 | not enter into CONST_DOUBLE for the replace. | |
10904 | ||
10905 | Note that copying is not done so X must not be shared unless all copies | |
10906 | are to be modified. | |
10907 | ||
10908 | This is like replace_rtx, except that we operate on N_REPLACEMENTS | |
30dc60c7 | 10909 | replacements simultaneously - FROM(n) is replacements[n*2] and to(n) is |
73a4d10b R |
10910 | replacements[n*2+1] - and that we take mode changes into account. |
10911 | ||
30dc60c7 | 10912 | If a replacement is ambiguous, return NULL_RTX. |
73a4d10b R |
10913 | |
10914 | If MODIFY is zero, don't modify any rtl in place, | |
10915 | just return zero or nonzero for failure / success. */ | |
10916 | ||
10917 | rtx | |
10918 | replace_n_hard_rtx (rtx x, rtx *replacements, int n_replacements, int modify) | |
10919 | { | |
10920 | int i, j; | |
10921 | const char *fmt; | |
10922 | ||
10923 | /* The following prevents loops occurrence when we change MEM in | |
10924 | CONST_DOUBLE onto the same CONST_DOUBLE. */ | |
10925 | if (x != 0 && GET_CODE (x) == CONST_DOUBLE) | |
10926 | return x; | |
10927 | ||
10928 | for (i = n_replacements - 1; i >= 0 ; i--) | |
10929 | if (x == replacements[i*2] && GET_MODE (x) == GET_MODE (replacements[i*2+1])) | |
10930 | return replacements[i*2+1]; | |
10931 | ||
10932 | /* Allow this function to make replacements in EXPR_LISTs. */ | |
10933 | if (x == 0) | |
10934 | return 0; | |
10935 | ||
10936 | if (GET_CODE (x) == SUBREG) | |
10937 | { | |
10938 | rtx new = replace_n_hard_rtx (SUBREG_REG (x), replacements, | |
10939 | n_replacements, modify); | |
10940 | ||
10941 | if (GET_CODE (new) == CONST_INT) | |
10942 | { | |
10943 | x = simplify_subreg (GET_MODE (x), new, | |
10944 | GET_MODE (SUBREG_REG (x)), | |
10945 | SUBREG_BYTE (x)); | |
10946 | if (! x) | |
10947 | abort (); | |
10948 | } | |
10949 | else if (modify) | |
10950 | SUBREG_REG (x) = new; | |
10951 | ||
10952 | return x; | |
10953 | } | |
10954 | else if (GET_CODE (x) == REG) | |
10955 | { | |
10956 | unsigned regno = REGNO (x); | |
10957 | unsigned nregs = (regno < FIRST_PSEUDO_REGISTER | |
10958 | ? HARD_REGNO_NREGS (regno, GET_MODE (x)) : 1); | |
10959 | rtx result = NULL_RTX; | |
10960 | ||
10961 | for (i = n_replacements - 1; i >= 0; i--) | |
10962 | { | |
10963 | rtx from = replacements[i*2]; | |
10964 | rtx to = replacements[i*2+1]; | |
10965 | unsigned from_regno, from_nregs, to_regno, new_regno; | |
10966 | ||
10967 | if (GET_CODE (from) != REG) | |
10968 | continue; | |
10969 | from_regno = REGNO (from); | |
10970 | from_nregs = (from_regno < FIRST_PSEUDO_REGISTER | |
10971 | ? HARD_REGNO_NREGS (from_regno, GET_MODE (from)) : 1); | |
10972 | if (regno < from_regno + from_nregs && regno + nregs > from_regno) | |
10973 | { | |
10974 | if (regno < from_regno | |
10975 | || regno + nregs > from_regno + nregs | |
10976 | || GET_CODE (to) != REG | |
10977 | || result) | |
10978 | return NULL_RTX; | |
10979 | to_regno = REGNO (to); | |
10980 | if (to_regno < FIRST_PSEUDO_REGISTER) | |
10981 | { | |
10982 | new_regno = regno + to_regno - from_regno; | |
10983 | if ((unsigned) HARD_REGNO_NREGS (new_regno, GET_MODE (x)) | |
10984 | != nregs) | |
10985 | return NULL_RTX; | |
10986 | result = gen_rtx_REG (GET_MODE (x), new_regno); | |
10987 | } | |
10988 | else if (GET_MODE (x) <= GET_MODE (to)) | |
10989 | result = gen_lowpart_common (GET_MODE (x), to); | |
10990 | else | |
10991 | result = gen_lowpart_SUBREG (GET_MODE (x), to); | |
10992 | } | |
10993 | } | |
10994 | return result ? result : x; | |
10995 | } | |
10996 | else if (GET_CODE (x) == ZERO_EXTEND) | |
10997 | { | |
10998 | rtx new = replace_n_hard_rtx (XEXP (x, 0), replacements, | |
10999 | n_replacements, modify); | |
11000 | ||
11001 | if (GET_CODE (new) == CONST_INT) | |
11002 | { | |
11003 | x = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x), | |
11004 | new, GET_MODE (XEXP (x, 0))); | |
11005 | if (! x) | |
11006 | abort (); | |
11007 | } | |
11008 | else if (modify) | |
11009 | XEXP (x, 0) = new; | |
11010 | ||
11011 | return x; | |
11012 | } | |
11013 | ||
11014 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
11015 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
11016 | { | |
11017 | rtx new; | |
11018 | ||
11019 | if (fmt[i] == 'e') | |
11020 | { | |
11021 | new = replace_n_hard_rtx (XEXP (x, i), replacements, | |
11022 | n_replacements, modify); | |
11023 | if (!new) | |
11024 | return NULL_RTX; | |
11025 | if (modify) | |
11026 | XEXP (x, i) = new; | |
11027 | } | |
11028 | else if (fmt[i] == 'E') | |
11029 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
11030 | { | |
11031 | new = replace_n_hard_rtx (XVECEXP (x, i, j), replacements, | |
11032 | n_replacements, modify); | |
11033 | if (!new) | |
11034 | return NULL_RTX; | |
11035 | if (modify) | |
11036 | XVECEXP (x, i, j) = new; | |
11037 | } | |
11038 | } | |
11039 | ||
11040 | return x; | |
11041 | } | |
11042 | ||
11043 | rtx | |
11044 | sh_gen_truncate (enum machine_mode mode, rtx x, int need_sign_ext) | |
11045 | { | |
11046 | enum rtx_code code = TRUNCATE; | |
11047 | ||
11048 | if (GET_CODE (x) == ZERO_EXTEND || GET_CODE (x) == SIGN_EXTEND) | |
11049 | { | |
11050 | rtx inner = XEXP (x, 0); | |
11051 | enum machine_mode inner_mode = GET_MODE (inner); | |
11052 | ||
11053 | if (inner_mode == mode) | |
11054 | return inner; | |
11055 | else if (GET_MODE_SIZE (inner_mode) >= GET_MODE_SIZE (mode)) | |
11056 | x = inner; | |
11057 | else if (GET_MODE_SIZE (inner_mode) < GET_MODE_SIZE (mode) | |
11058 | && (! need_sign_ext || GET_CODE (x) == SIGN_EXTEND)) | |
11059 | { | |
11060 | code = GET_CODE (x); | |
11061 | x = inner; | |
11062 | } | |
11063 | } | |
11064 | return gen_rtx_fmt_e (code, mode, x); | |
11065 | } | |
11066 | ||
11067 | /* called via for_each_rtx after reload, to clean up truncates of | |
11068 | registers that span multiple actual hard registers. */ | |
11069 | int | |
11070 | shmedia_cleanup_truncate (rtx *p, void *n_changes) | |
11071 | { | |
11072 | rtx x = *p, reg; | |
11073 | ||
11074 | if (GET_CODE (x) != TRUNCATE) | |
11075 | return 0; | |
11076 | reg = XEXP (x, 0); | |
11077 | if (GET_MODE_SIZE (GET_MODE (reg)) > 8 && GET_CODE (reg) == REG) | |
11078 | { | |
11079 | enum machine_mode reg_mode = GET_MODE (reg); | |
11080 | XEXP (x, 0) = simplify_subreg (DImode, reg, reg_mode, | |
11081 | subreg_lowpart_offset (DImode, reg_mode)); | |
11082 | *(int*) n_changes += 1; | |
11083 | return -1; | |
11084 | } | |
11085 | return 0; | |
11086 | } | |
11087 | ||
11088 | /* Load and store depend on the highpart of the address. However, | |
11089 | set_attr_alternative does not give well-defined results before reload, | |
11090 | so we must look at the rtl ourselves to see if any of the feeding | |
11091 | registers is used in a memref. */ | |
11092 | ||
11093 | /* Called by sh_contains_memref_p via for_each_rtx. */ | |
11094 | static int | |
11095 | sh_contains_memref_p_1 (rtx *loc, void *data ATTRIBUTE_UNUSED) | |
11096 | { | |
11097 | return (GET_CODE (*loc) == MEM); | |
11098 | } | |
11099 | ||
a4d05547 | 11100 | /* Return nonzero iff INSN contains a MEM. */ |
73a4d10b R |
11101 | int |
11102 | sh_contains_memref_p (rtx insn) | |
11103 | { | |
11104 | return for_each_rtx (&PATTERN (insn), &sh_contains_memref_p_1, NULL); | |
11105 | } | |
11106 | ||
5c7cafa8 CB |
11107 | /* Return nonzero iff INSN loads a banked register. */ |
11108 | int | |
11109 | sh_loads_bankedreg_p (rtx insn) | |
11110 | { | |
11111 | if (GET_CODE (PATTERN (insn)) == SET) | |
11112 | { | |
11113 | rtx op = SET_DEST (PATTERN(insn)); | |
11114 | if (REG_P (op) && BANKED_REGISTER_P (REGNO (op))) | |
11115 | return 1; | |
11116 | } | |
11117 | ||
11118 | return 0; | |
11119 | } | |
11120 | ||
73a4d10b R |
11121 | /* FNADDR is the MEM expression from a call expander. Return an address |
11122 | to use in an SHmedia insn pattern. */ | |
11123 | rtx | |
11124 | shmedia_prepare_call_address (rtx fnaddr, int is_sibcall) | |
11125 | { | |
11126 | int is_sym; | |
11127 | ||
11128 | fnaddr = XEXP (fnaddr, 0); | |
11129 | is_sym = GET_CODE (fnaddr) == SYMBOL_REF; | |
11130 | if (flag_pic && is_sym) | |
11131 | { | |
11132 | if (! SYMBOL_REF_LOCAL_P (fnaddr)) | |
11133 | { | |
11134 | rtx reg = gen_reg_rtx (Pmode); | |
11135 | ||
11136 | /* We must not use GOTPLT for sibcalls, because PIC_REG | |
11137 | must be restored before the PLT code gets to run. */ | |
11138 | if (is_sibcall) | |
11139 | emit_insn (gen_symGOT2reg (reg, fnaddr)); | |
11140 | else | |
11141 | emit_insn (gen_symGOTPLT2reg (reg, fnaddr)); | |
11142 | fnaddr = reg; | |
11143 | } | |
11144 | else | |
11145 | { | |
11146 | fnaddr = gen_sym2PIC (fnaddr); | |
11147 | PUT_MODE (fnaddr, Pmode); | |
11148 | } | |
11149 | } | |
11150 | /* If ptabs might trap, make this visible to the rest of the compiler. | |
11151 | We generally assume that symbols pertain to valid locations, but | |
11152 | it is possible to generate invalid symbols with asm or linker tricks. | |
30dc60c7 | 11153 | In a list of functions where each returns its successor, an invalid |
73a4d10b R |
11154 | symbol might denote an empty list. */ |
11155 | if (!TARGET_PT_FIXED | |
11156 | && (!is_sym || TARGET_INVALID_SYMBOLS) | |
11157 | && (!REG_P (fnaddr) || ! TARGET_REGISTER_P (REGNO (fnaddr)))) | |
11158 | { | |
11159 | rtx tr = gen_reg_rtx (PDImode); | |
11160 | ||
11161 | emit_insn (gen_ptabs (tr, fnaddr)); | |
11162 | fnaddr = tr; | |
11163 | } | |
11164 | else if (! target_reg_operand (fnaddr, Pmode)) | |
11165 | fnaddr = copy_to_mode_reg (Pmode, fnaddr); | |
11166 | return fnaddr; | |
11167 | } | |
11168 | ||
8a99f6f9 R |
11169 | enum reg_class |
11170 | sh_secondary_reload (bool in_p, rtx x, enum reg_class class, | |
11171 | enum machine_mode mode, secondary_reload_info *sri) | |
11172 | { | |
11173 | if (in_p) | |
11174 | { | |
11175 | if (REGCLASS_HAS_FP_REG (class) | |
11176 | && ! TARGET_SHMEDIA | |
11177 | && immediate_operand ((x), mode) | |
11178 | && ! ((fp_zero_operand (x) || fp_one_operand (x)) | |
11179 | && mode == SFmode && fldi_ok ())) | |
11180 | switch (mode) | |
11181 | { | |
11182 | case SFmode: | |
11183 | sri->icode = CODE_FOR_reload_insf__frn; | |
11184 | return NO_REGS; | |
11185 | case DFmode: | |
11186 | sri->icode = CODE_FOR_reload_indf__frn; | |
11187 | return NO_REGS; | |
11188 | case SImode: | |
11189 | /* ??? If we knew that we are in the appropriate mode - | |
11190 | single precision - we could use a reload pattern directly. */ | |
11191 | return FPUL_REGS; | |
11192 | default: | |
11193 | abort (); | |
11194 | } | |
11195 | if (class == FPUL_REGS | |
11196 | && ((GET_CODE (x) == REG | |
11197 | && (REGNO (x) == MACL_REG || REGNO (x) == MACH_REG | |
11198 | || REGNO (x) == T_REG)) | |
11199 | || GET_CODE (x) == PLUS)) | |
11200 | return GENERAL_REGS; | |
11201 | if (class == FPUL_REGS && immediate_operand (x, mode)) | |
11202 | { | |
d973ac5a | 11203 | if (satisfies_constraint_I08 (x) || fp_zero_operand (x)) |
8a99f6f9 | 11204 | return GENERAL_REGS; |
d973ac5a KK |
11205 | else if (mode == SFmode) |
11206 | return FP_REGS; | |
8a99f6f9 R |
11207 | sri->icode = CODE_FOR_reload_insi__i_fpul; |
11208 | return NO_REGS; | |
11209 | } | |
11210 | if (class == FPSCR_REGS | |
11211 | && ((GET_CODE (x) == REG && REGNO (x) >= FIRST_PSEUDO_REGISTER) | |
11212 | || (GET_CODE (x) == MEM && GET_CODE (XEXP (x, 0)) == PLUS))) | |
11213 | return GENERAL_REGS; | |
11214 | if (REGCLASS_HAS_FP_REG (class) | |
11215 | && TARGET_SHMEDIA | |
11216 | && immediate_operand (x, mode) | |
11217 | && x != CONST0_RTX (GET_MODE (x)) | |
11218 | && GET_MODE (x) != V4SFmode) | |
11219 | return GENERAL_REGS; | |
11220 | if ((mode == QImode || mode == HImode) | |
11221 | && TARGET_SHMEDIA && inqhi_operand (x, mode)) | |
11222 | { | |
11223 | sri->icode = ((mode == QImode) | |
11224 | ? CODE_FOR_reload_inqi : CODE_FOR_reload_inhi); | |
11225 | return NO_REGS; | |
11226 | } | |
11227 | if (TARGET_SHMEDIA && class == GENERAL_REGS | |
11228 | && (GET_CODE (x) == LABEL_REF || PIC_DIRECT_ADDR_P (x))) | |
11229 | return TARGET_REGS; | |
11230 | } /* end of input-only processing. */ | |
11231 | ||
11232 | if (((REGCLASS_HAS_FP_REG (class) | |
11233 | && (GET_CODE (x) == REG | |
11234 | && (GENERAL_OR_AP_REGISTER_P (REGNO (x)) | |
11235 | || (FP_REGISTER_P (REGNO (x)) && mode == SImode | |
11236 | && TARGET_FMOVD)))) | |
11237 | || (REGCLASS_HAS_GENERAL_REG (class) | |
11238 | && GET_CODE (x) == REG | |
11239 | && FP_REGISTER_P (REGNO (x)))) | |
11240 | && ! TARGET_SHMEDIA | |
11241 | && (mode == SFmode || mode == SImode)) | |
11242 | return FPUL_REGS; | |
11243 | if ((class == FPUL_REGS | |
11244 | || (REGCLASS_HAS_FP_REG (class) | |
11245 | && ! TARGET_SHMEDIA && mode == SImode)) | |
11246 | && (GET_CODE (x) == MEM | |
11247 | || (GET_CODE (x) == REG | |
11248 | && (REGNO (x) >= FIRST_PSEUDO_REGISTER | |
11249 | || REGNO (x) == T_REG | |
11250 | || system_reg_operand (x, VOIDmode))))) | |
11251 | { | |
11252 | if (class == FPUL_REGS) | |
11253 | return GENERAL_REGS; | |
11254 | return FPUL_REGS; | |
11255 | } | |
11256 | if ((class == TARGET_REGS | |
11257 | || (TARGET_SHMEDIA && class == SIBCALL_REGS)) | |
32a7ab3d | 11258 | && !satisfies_constraint_Csy (x) |
8a99f6f9 R |
11259 | && (GET_CODE (x) != REG || ! GENERAL_REGISTER_P (REGNO (x)))) |
11260 | return GENERAL_REGS; | |
11261 | if ((class == MAC_REGS || class == PR_REGS) | |
11262 | && GET_CODE (x) == REG && ! GENERAL_REGISTER_P (REGNO (x)) | |
11263 | && class != REGNO_REG_CLASS (REGNO (x))) | |
11264 | return GENERAL_REGS; | |
11265 | if (class != GENERAL_REGS && GET_CODE (x) == REG | |
11266 | && TARGET_REGISTER_P (REGNO (x))) | |
11267 | return GENERAL_REGS; | |
11268 | return NO_REGS; | |
11269 | } | |
11270 | ||
73a4d10b R |
11271 | enum sh_divide_strategy_e sh_div_strategy = SH_DIV_STRATEGY_DEFAULT; |
11272 | ||
e2500fed | 11273 | #include "gt-sh.h" |