]>
Commit | Line | Data |
---|---|---|
7b23765d | 1 | /* Output routines for GCC for Renesas / SuperH SH. |
3cebdd3a | 2 | Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001, 2002, |
65b91e5b | 3 | 2003, 2004, 2005 Free Software Foundation, Inc. |
7014838c | 4 | Contributed by Steve Chamberlain (sac@cygnus.com). |
9e7454d0 | 5 | Improved by Jim Wilson (wilson@cygnus.com). |
7208c779 | 6 | |
187b36cf | 7 | This file is part of GCC. |
9b754436 | 8 | |
187b36cf | 9 | GCC is free software; you can redistribute it and/or modify |
9b754436 | 10 | it under the terms of the GNU General Public License as published by |
11 | the Free Software Foundation; either version 2, or (at your option) | |
12 | any later version. | |
13 | ||
187b36cf | 14 | GCC is distributed in the hope that it will be useful, |
9b754436 | 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 | |
187b36cf | 20 | along with GCC; see the file COPYING. If not, write to |
dbddc6c4 | 21 | the Free Software Foundation, 51 Franklin Street, Fifth Floor, |
22 | Boston, MA 02110-1301, USA. */ | |
7208c779 | 23 | |
7208c779 | 24 | #include "config.h" |
7014838c | 25 | #include "system.h" |
805e22b2 | 26 | #include "coretypes.h" |
27 | #include "tm.h" | |
b7dbbdb2 | 28 | #include "insn-config.h" |
7208c779 | 29 | #include "rtl.h" |
7208c779 | 30 | #include "tree.h" |
7208c779 | 31 | #include "flags.h" |
7208c779 | 32 | #include "expr.h" |
d8fc4d0b | 33 | #include "optabs.h" |
0a893c29 | 34 | #include "function.h" |
73401833 | 35 | #include "regs.h" |
36 | #include "hard-reg-set.h" | |
37 | #include "output.h" | |
891b35f2 | 38 | #include "insn-attr.h" |
b7dbbdb2 | 39 | #include "toplev.h" |
40 | #include "recog.h" | |
1fcd08b1 | 41 | #include "c-pragma.h" |
e51fa923 | 42 | #include "integrate.h" |
8ff30ff6 | 43 | #include "dwarf2.h" |
b7dbbdb2 | 44 | #include "tm_p.h" |
a767736d | 45 | #include "target.h" |
46 | #include "target-def.h" | |
30ab56ba | 47 | #include "real.h" |
e3fac27d | 48 | #include "langhooks.h" |
c83ee613 | 49 | #include "basic-block.h" |
356907d3 | 50 | #include "cfglayout.h" |
133479bd | 51 | #include "intl.h" |
f77a5bb0 | 52 | #include "sched-int.h" |
5bf8e2fe | 53 | #include "ggc.h" |
20cdebf8 | 54 | #include "tree-gimple.h" |
59312820 | 55 | #include "cfgloop.h" |
20cdebf8 | 56 | |
7208c779 | 57 | |
8ded0752 | 58 | int code_for_indirect_jump_scratch = CODE_FOR_indirect_jump_scratch; |
59 | ||
3dfc1807 | 60 | #define MSW (TARGET_LITTLE_ENDIAN ? 1 : 0) |
61 | #define LSW (TARGET_LITTLE_ENDIAN ? 0 : 1) | |
62 | ||
87e19636 | 63 | /* These are some macros to abstract register modes. */ |
64 | #define CONST_OK_FOR_ADD(size) \ | |
af2c1324 | 65 | (TARGET_SHMEDIA ? CONST_OK_FOR_I10 (size) : CONST_OK_FOR_I08 (size)) |
87e19636 | 66 | #define GEN_MOV (*(TARGET_SHMEDIA64 ? gen_movdi : gen_movsi)) |
67 | #define GEN_ADD3 (*(TARGET_SHMEDIA64 ? gen_adddi3 : gen_addsi3)) | |
68 | #define GEN_SUB3 (*(TARGET_SHMEDIA64 ? gen_subdi3 : gen_subsi3)) | |
69 | ||
0d687b6d | 70 | /* Set to 1 by expand_prologue() when the function is an interrupt handler. */ |
71 | int current_function_interrupt; | |
72 | ||
0b89dfaf | 73 | /* ??? The pragma interrupt support will not work for SH3. */ |
73401833 | 74 | /* This is set by #pragma interrupt and #pragma trapa, and causes gcc to |
75 | output code for the next function appropriate for an interrupt handler. */ | |
97595bfd | 76 | int pragma_interrupt; |
7208c779 | 77 | |
15cb1799 | 78 | /* This is set by the trap_exit attribute for functions. It specifies |
79 | a trap number to be used in a trapa instruction at function exit | |
80 | (instead of an rte instruction). */ | |
81 | int trap_exit; | |
82 | ||
83 | /* This is used by the sp_switch attribute for functions. It specifies | |
84 | a variable holding the address of the stack the interrupt function | |
85 | should switch to/from at entry/exit. */ | |
86 | rtx sp_switch; | |
87 | ||
73401833 | 88 | /* This is set by #pragma trapa, and is similar to the above, except that |
89 | the compiler doesn't emit code to preserve all registers. */ | |
90 | static int pragma_trapa; | |
91 | ||
891b35f2 | 92 | /* This is set by #pragma nosave_low_regs. This is useful on the SH3, |
93 | which has a separate set of low regs for User and Supervisor modes. | |
94 | This should only be used for the lowest level of interrupts. Higher levels | |
95 | of interrupts must save the registers in case they themselves are | |
96 | interrupted. */ | |
97 | int pragma_nosave_low_regs; | |
98 | ||
6644435d | 99 | /* This is used for communication between TARGET_SETUP_INCOMING_VARARGS and |
73401833 | 100 | sh_expand_prologue. */ |
7208c779 | 101 | int current_function_anonymous_args; |
1504eb7e | 102 | |
9df03d69 | 103 | /* Global variables for machine-dependent things. */ |
7208c779 | 104 | |
1504eb7e | 105 | /* Which cpu are we scheduling for. */ |
106 | enum processor_type sh_cpu; | |
107 | ||
f77a5bb0 | 108 | /* Definitions used in ready queue reordering for first scheduling pass. */ |
109 | ||
110 | /* Reg weights arrays for modes SFmode and SImode, indexed by insn LUID. */ | |
111 | static short *regmode_weight[2]; | |
112 | ||
113 | /* Total SFmode and SImode weights of scheduled insns. */ | |
114 | static int curr_regmode_pressure[2]; | |
115 | ||
116 | /* If true, skip cycles for Q -> R movement. */ | |
117 | static int skip_cycles = 0; | |
118 | ||
119 | /* Cached value of can_issue_more. This is cached in sh_variable_issue hook | |
120 | and returned from sh_reorder2. */ | |
121 | static short cached_can_issue_more; | |
122 | ||
7208c779 | 123 | /* Saved operands from the last compare to use when we generate an scc |
1504eb7e | 124 | or bcc insn. */ |
7208c779 | 125 | |
126 | rtx sh_compare_op0; | |
127 | rtx sh_compare_op1; | |
128 | ||
129 | /* Provides the class number of the smallest class containing | |
1504eb7e | 130 | reg number. */ |
7208c779 | 131 | |
41fafa66 | 132 | enum reg_class regno_reg_class[FIRST_PSEUDO_REGISTER] = |
7208c779 | 133 | { |
2e3d4844 | 134 | R0_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, |
7208c779 | 135 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, |
136 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
137 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
87e19636 | 138 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, |
139 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
140 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
141 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
142 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
143 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
144 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
145 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
146 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
147 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
148 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
149 | GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, | |
b00564da | 150 | FP0_REGS,FP_REGS, FP_REGS, FP_REGS, |
151 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
152 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
153 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
87e19636 | 154 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, |
155 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
156 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
157 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
158 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
159 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
160 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
161 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
162 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
163 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
164 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
165 | FP_REGS, FP_REGS, FP_REGS, FP_REGS, | |
166 | TARGET_REGS, TARGET_REGS, TARGET_REGS, TARGET_REGS, | |
167 | TARGET_REGS, TARGET_REGS, TARGET_REGS, TARGET_REGS, | |
1b61190c | 168 | DF_REGS, DF_REGS, DF_REGS, DF_REGS, |
169 | DF_REGS, DF_REGS, DF_REGS, DF_REGS, | |
87e19636 | 170 | NO_REGS, GENERAL_REGS, PR_REGS, T_REGS, |
171 | MAC_REGS, MAC_REGS, FPUL_REGS, FPSCR_REGS, | |
b9a602c6 | 172 | GENERAL_REGS, GENERAL_REGS, |
1b61190c | 173 | }; |
174 | ||
87e19636 | 175 | char sh_register_names[FIRST_PSEUDO_REGISTER] \ |
176 | [MAX_REGISTER_NAME_LENGTH + 1] = SH_REGISTER_NAMES_INITIALIZER; | |
177 | ||
178 | char sh_additional_register_names[ADDREGNAMES_SIZE] \ | |
179 | [MAX_ADDITIONAL_REGISTER_NAME_LENGTH + 1] | |
180 | = SH_ADDITIONAL_REGISTER_NAMES_INITIALIZER; | |
7208c779 | 181 | |
182 | /* Provide reg_class from a letter such as appears in the machine | |
e04da7b6 | 183 | description. *: target independently reserved letter. |
4aa67c38 | 184 | reg_class_from_letter['e' - 'a'] is set to NO_REGS for TARGET_FMOVD. */ |
7208c779 | 185 | |
e04da7b6 | 186 | enum reg_class reg_class_from_letter[] = |
7208c779 | 187 | { |
e04da7b6 | 188 | /* a */ ALL_REGS, /* b */ TARGET_REGS, /* c */ FPSCR_REGS, /* d */ DF_REGS, |
189 | /* e */ FP_REGS, /* f */ FP_REGS, /* g **/ NO_REGS, /* h */ NO_REGS, | |
190 | /* i **/ NO_REGS, /* j */ NO_REGS, /* k */ SIBCALL_REGS, /* l */ PR_REGS, | |
191 | /* m **/ NO_REGS, /* n **/ NO_REGS, /* o **/ NO_REGS, /* p **/ NO_REGS, | |
192 | /* q */ NO_REGS, /* r **/ NO_REGS, /* s **/ NO_REGS, /* t */ T_REGS, | |
193 | /* u */ NO_REGS, /* v */ NO_REGS, /* w */ FP0_REGS, /* x */ MAC_REGS, | |
b00564da | 194 | /* y */ FPUL_REGS, /* z */ R0_REGS |
7208c779 | 195 | }; |
8ded0752 | 196 | |
1b61190c | 197 | int assembler_dialect; |
198 | ||
8af3db02 | 199 | static bool shmedia_space_reserved_for_target_registers; |
200 | ||
156fac8b | 201 | static bool sh_handle_option (size_t, const char *, int); |
04f04b72 | 202 | static void split_branches (rtx); |
203 | static int branch_dest (rtx); | |
204 | static void force_into (rtx, rtx); | |
205 | static void print_slot (rtx); | |
206 | static rtx add_constant (rtx, enum machine_mode, rtx); | |
51dade95 | 207 | static void dump_table (rtx, rtx); |
04f04b72 | 208 | static int hi_const (rtx); |
209 | static int broken_move (rtx); | |
210 | static int mova_p (rtx); | |
211 | static rtx find_barrier (int, rtx, rtx); | |
212 | static int noncall_uses_reg (rtx, rtx, rtx *); | |
213 | static rtx gen_block_redirect (rtx, int, int); | |
214 | static void sh_reorg (void); | |
41fafa66 | 215 | static void output_stack_adjust (int, rtx, int, HARD_REG_SET *); |
04f04b72 | 216 | static rtx frame_insn (rtx); |
217 | static rtx push (int); | |
218 | static void pop (int); | |
219 | static void push_regs (HARD_REG_SET *, int); | |
220 | static int calc_live_regs (HARD_REG_SET *); | |
221 | static void mark_use (rtx, rtx *); | |
222 | static HOST_WIDE_INT rounded_frame_size (int); | |
223 | static rtx mark_constant_pool_use (rtx); | |
e3c541f0 | 224 | const struct attribute_spec sh_attribute_table[]; |
04f04b72 | 225 | static tree sh_handle_interrupt_handler_attribute (tree *, tree, tree, int, bool *); |
226 | static tree sh_handle_sp_switch_attribute (tree *, tree, tree, int, bool *); | |
227 | static tree sh_handle_trap_exit_attribute (tree *, tree, tree, int, bool *); | |
228 | static tree sh_handle_renesas_attribute (tree *, tree, tree, int, bool *); | |
229 | static void sh_output_function_epilogue (FILE *, HOST_WIDE_INT); | |
230 | static void sh_insert_attributes (tree, tree *); | |
761d9126 | 231 | static const char *sh_check_pch_target_flags (int); |
04f04b72 | 232 | static int sh_adjust_cost (rtx, rtx, rtx, int); |
04f04b72 | 233 | static int sh_issue_rate (void); |
f77a5bb0 | 234 | static int sh_dfa_new_cycle (FILE *, int, rtx, int, int, int *sort_p); |
235 | static short find_set_regmode_weight (rtx, enum machine_mode); | |
236 | static short find_insn_regmode_weight (rtx, enum machine_mode); | |
237 | static void find_regmode_weight (int, enum machine_mode); | |
238 | static void sh_md_init_global (FILE *, int, int); | |
239 | static void sh_md_finish_global (FILE *, int); | |
240 | static int rank_for_reorder (const void *, const void *); | |
241 | static void swap_reorder (rtx *, int); | |
242 | static void ready_reorder (rtx *, int); | |
243 | static short high_pressure (enum machine_mode); | |
244 | static int sh_reorder (FILE *, int, rtx *, int *, int); | |
245 | static int sh_reorder2 (FILE *, int, rtx *, int *, int); | |
246 | static void sh_md_init (FILE *, int, int); | |
247 | static int sh_variable_issue (FILE *, int, rtx, int); | |
9e7454d0 | 248 | |
04f04b72 | 249 | static bool sh_function_ok_for_sibcall (tree, tree); |
250 | ||
251 | static bool sh_cannot_modify_jumps_p (void); | |
e26dbf6d | 252 | static int sh_target_reg_class (void); |
8af3db02 | 253 | static bool sh_optimize_target_register_callee_saved (bool); |
04f04b72 | 254 | static bool sh_ms_bitfield_layout_p (tree); |
255 | ||
256 | static void sh_init_builtins (void); | |
257 | static void sh_media_init_builtins (void); | |
258 | static rtx sh_expand_builtin (tree, rtx, rtx, enum machine_mode, int); | |
259 | static void sh_output_mi_thunk (FILE *, tree, HOST_WIDE_INT, HOST_WIDE_INT, tree); | |
260 | static void sh_file_start (void); | |
261 | static int flow_dependent_p (rtx, rtx); | |
262 | static void flow_dependent_p_1 (rtx, rtx, void *); | |
263 | static int shiftcosts (rtx); | |
264 | static int andcosts (rtx); | |
265 | static int addsubcosts (rtx); | |
266 | static int multcosts (rtx); | |
267 | static bool unspec_caller_rtx_p (rtx); | |
268 | static bool sh_cannot_copy_insn_p (rtx); | |
269 | static bool sh_rtx_costs (rtx, int, int, int *); | |
270 | static int sh_address_cost (rtx); | |
c5b4f25c | 271 | #ifdef TARGET_ADJUST_UNROLL_MAX |
59312820 | 272 | static int sh_adjust_unroll_max (struct loop *, int, int, int, int); |
c5b4f25c | 273 | #endif |
edf54f2a | 274 | static int sh_pr_n_sets (void); |
275 | static rtx sh_allocate_initial_value (rtx); | |
8af3db02 | 276 | static int shmedia_target_regs_stack_space (HARD_REG_SET *); |
277 | static int shmedia_reserve_space_for_target_registers_p (int, HARD_REG_SET *); | |
278 | static int shmedia_target_regs_stack_adjust (HARD_REG_SET *); | |
41fafa66 | 279 | static int scavenge_reg (HARD_REG_SET *s); |
280 | struct save_schedule_s; | |
281 | static struct save_entry_s *sh5_schedule_saves (HARD_REG_SET *, | |
282 | struct save_schedule_s *, int); | |
45550790 | 283 | |
04f04b72 | 284 | static rtx sh_struct_value_rtx (tree, int); |
285 | static bool sh_return_in_memory (tree, tree); | |
286 | static rtx sh_builtin_saveregs (void); | |
287 | static void sh_setup_incoming_varargs (CUMULATIVE_ARGS *, enum machine_mode, tree, int *, int); | |
288 | static bool sh_strict_argument_naming (CUMULATIVE_ARGS *); | |
289 | static bool sh_pretend_outgoing_varargs_named (CUMULATIVE_ARGS *); | |
2e15d750 | 290 | static tree sh_build_builtin_va_list (void); |
20cdebf8 | 291 | static tree sh_gimplify_va_arg_expr (tree, tree, tree *, tree *); |
b981d932 | 292 | static bool sh_pass_by_reference (CUMULATIVE_ARGS *, enum machine_mode, |
293 | tree, bool); | |
13f08ee7 | 294 | static bool sh_callee_copies (CUMULATIVE_ARGS *, enum machine_mode, |
295 | tree, bool); | |
f054eb3c | 296 | static int sh_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode, |
297 | tree, bool); | |
8ff30ff6 | 298 | static int sh_dwarf_calling_convention (tree); |
3cebdd3a | 299 | static int hard_regs_intersect_p (HARD_REG_SET *, HARD_REG_SET *); |
45550790 | 300 | |
a767736d | 301 | \f |
302 | /* Initialize the GCC target structure. */ | |
e3c541f0 | 303 | #undef TARGET_ATTRIBUTE_TABLE |
304 | #define TARGET_ATTRIBUTE_TABLE sh_attribute_table | |
a767736d | 305 | |
58356836 | 306 | /* The next two are used for debug info when compiling with -gdwarf. */ |
307 | #undef TARGET_ASM_UNALIGNED_HI_OP | |
308 | #define TARGET_ASM_UNALIGNED_HI_OP "\t.uaword\t" | |
309 | #undef TARGET_ASM_UNALIGNED_SI_OP | |
310 | #define TARGET_ASM_UNALIGNED_SI_OP "\t.ualong\t" | |
311 | ||
87e19636 | 312 | /* These are NULLed out on non-SH5 in OVERRIDE_OPTIONS. */ |
313 | #undef TARGET_ASM_UNALIGNED_DI_OP | |
314 | #define TARGET_ASM_UNALIGNED_DI_OP "\t.uaquad\t" | |
315 | #undef TARGET_ASM_ALIGNED_DI_OP | |
316 | #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t" | |
317 | ||
17d9b0c3 | 318 | #undef TARGET_ASM_FUNCTION_EPILOGUE |
319 | #define TARGET_ASM_FUNCTION_EPILOGUE sh_output_function_epilogue | |
320 | ||
c83ee613 | 321 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
322 | #define TARGET_ASM_OUTPUT_MI_THUNK sh_output_mi_thunk | |
323 | ||
324 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
325 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_tree_hwi_hwi_tree_true | |
326 | ||
92c473b8 | 327 | #undef TARGET_ASM_FILE_START |
328 | #define TARGET_ASM_FILE_START sh_file_start | |
329 | #undef TARGET_ASM_FILE_START_FILE_DIRECTIVE | |
330 | #define TARGET_ASM_FILE_START_FILE_DIRECTIVE true | |
331 | ||
156fac8b | 332 | #undef TARGET_DEFAULT_TARGET_FLAGS |
333 | #define TARGET_DEFAULT_TARGET_FLAGS TARGET_DEFAULT | |
334 | #undef TARGET_HANDLE_OPTION | |
335 | #define TARGET_HANDLE_OPTION sh_handle_option | |
336 | ||
8ee295a7 | 337 | #undef TARGET_INSERT_ATTRIBUTES |
338 | #define TARGET_INSERT_ATTRIBUTES sh_insert_attributes | |
339 | ||
747af5e7 | 340 | #undef TARGET_SCHED_ADJUST_COST |
341 | #define TARGET_SCHED_ADJUST_COST sh_adjust_cost | |
342 | ||
bea4bad2 | 343 | #undef TARGET_SCHED_ISSUE_RATE |
344 | #define TARGET_SCHED_ISSUE_RATE sh_issue_rate | |
345 | ||
f77a5bb0 | 346 | /* The next 5 hooks have been implemented for reenabling sched1. With the |
347 | help of these macros we are limiting the movement of insns in sched1 to | |
9e7454d0 | 348 | reduce the register pressure. The overall idea is to keep count of SImode |
f77a5bb0 | 349 | and SFmode regs required by already scheduled insns. When these counts |
350 | cross some threshold values; give priority to insns that free registers. | |
351 | The insn that frees registers is most likely to be the insn with lowest | |
9e7454d0 | 352 | LUID (original insn order); but such an insn might be there in the stalled |
f77a5bb0 | 353 | queue (Q) instead of the ready queue (R). To solve this, we skip cycles |
354 | upto a max of 8 cycles so that such insns may move from Q -> R. | |
355 | ||
356 | The description of the hooks are as below: | |
357 | ||
358 | TARGET_SCHED_INIT_GLOBAL: Added a new target hook in the generic | |
359 | scheduler; it is called inside the sched_init function just after | |
360 | find_insn_reg_weights function call. It is used to calculate the SImode | |
7bd28bba | 361 | and SFmode weights of insns of basic blocks; much similar to what |
9e7454d0 | 362 | find_insn_reg_weights does. |
f77a5bb0 | 363 | TARGET_SCHED_FINISH_GLOBAL: Corresponding cleanup hook. |
364 | ||
365 | TARGET_SCHED_DFA_NEW_CYCLE: Skip cycles if high register pressure is | |
366 | indicated by TARGET_SCHED_REORDER2; doing this may move insns from | |
367 | (Q)->(R). | |
368 | ||
369 | TARGET_SCHED_REORDER: If the register pressure for SImode or SFmode is | |
370 | high; reorder the ready queue so that the insn with lowest LUID will be | |
371 | issued next. | |
372 | ||
373 | TARGET_SCHED_REORDER2: If the register pressure is high, indicate to | |
374 | TARGET_SCHED_DFA_NEW_CYCLE to skip cycles. | |
375 | ||
376 | TARGET_SCHED_VARIABLE_ISSUE: Cache the value of can_issue_more so that it | |
377 | can be returned from TARGET_SCHED_REORDER2. | |
378 | ||
379 | TARGET_SCHED_INIT: Reset the register pressure counting variables. */ | |
380 | ||
381 | #undef TARGET_SCHED_DFA_NEW_CYCLE | |
382 | #define TARGET_SCHED_DFA_NEW_CYCLE sh_dfa_new_cycle | |
383 | ||
384 | #undef TARGET_SCHED_INIT_GLOBAL | |
385 | #define TARGET_SCHED_INIT_GLOBAL sh_md_init_global | |
386 | ||
387 | #undef TARGET_SCHED_FINISH_GLOBAL | |
388 | #define TARGET_SCHED_FINISH_GLOBAL sh_md_finish_global | |
389 | ||
390 | #undef TARGET_SCHED_VARIABLE_ISSUE | |
391 | #define TARGET_SCHED_VARIABLE_ISSUE sh_variable_issue | |
392 | ||
393 | #undef TARGET_SCHED_REORDER | |
394 | #define TARGET_SCHED_REORDER sh_reorder | |
395 | ||
396 | #undef TARGET_SCHED_REORDER2 | |
397 | #define TARGET_SCHED_REORDER2 sh_reorder2 | |
398 | ||
399 | #undef TARGET_SCHED_INIT | |
400 | #define TARGET_SCHED_INIT sh_md_init | |
401 | ||
87e19636 | 402 | #undef TARGET_CANNOT_MODIFY_JUMPS_P |
403 | #define TARGET_CANNOT_MODIFY_JUMPS_P sh_cannot_modify_jumps_p | |
8af3db02 | 404 | #undef TARGET_BRANCH_TARGET_REGISTER_CLASS |
405 | #define TARGET_BRANCH_TARGET_REGISTER_CLASS sh_target_reg_class | |
406 | #undef TARGET_BRANCH_TARGET_REGISTER_CALLEE_SAVED | |
407 | #define TARGET_BRANCH_TARGET_REGISTER_CALLEE_SAVED \ | |
408 | sh_optimize_target_register_callee_saved | |
87e19636 | 409 | |
410 | #undef TARGET_MS_BITFIELD_LAYOUT_P | |
411 | #define TARGET_MS_BITFIELD_LAYOUT_P sh_ms_bitfield_layout_p | |
412 | ||
e3fac27d | 413 | #undef TARGET_INIT_BUILTINS |
414 | #define TARGET_INIT_BUILTINS sh_init_builtins | |
415 | #undef TARGET_EXPAND_BUILTIN | |
416 | #define TARGET_EXPAND_BUILTIN sh_expand_builtin | |
417 | ||
805e22b2 | 418 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL |
419 | #define TARGET_FUNCTION_OK_FOR_SIBCALL sh_function_ok_for_sibcall | |
420 | ||
02359f85 | 421 | #undef TARGET_CANNOT_COPY_INSN_P |
422 | #define TARGET_CANNOT_COPY_INSN_P sh_cannot_copy_insn_p | |
fab7adbf | 423 | #undef TARGET_RTX_COSTS |
424 | #define TARGET_RTX_COSTS sh_rtx_costs | |
ec0457a8 | 425 | #undef TARGET_ADDRESS_COST |
426 | #define TARGET_ADDRESS_COST sh_address_cost | |
edf54f2a | 427 | #undef TARGET_ALLOCATE_INITIAL_VALUE |
428 | #define TARGET_ALLOCATE_INITIAL_VALUE sh_allocate_initial_value | |
fab7adbf | 429 | |
2efea8c0 | 430 | #undef TARGET_MACHINE_DEPENDENT_REORG |
431 | #define TARGET_MACHINE_DEPENDENT_REORG sh_reorg | |
432 | ||
cec78553 | 433 | #ifdef HAVE_AS_TLS |
434 | #undef TARGET_HAVE_TLS | |
435 | #define TARGET_HAVE_TLS true | |
436 | #endif | |
437 | ||
45550790 | 438 | #undef TARGET_PROMOTE_PROTOTYPES |
439 | #define TARGET_PROMOTE_PROTOTYPES sh_promote_prototypes | |
440 | #undef TARGET_PROMOTE_FUNCTION_ARGS | |
441 | #define TARGET_PROMOTE_FUNCTION_ARGS sh_promote_prototypes | |
442 | #undef TARGET_PROMOTE_FUNCTION_RETURN | |
443 | #define TARGET_PROMOTE_FUNCTION_RETURN sh_promote_prototypes | |
444 | ||
445 | #undef TARGET_STRUCT_VALUE_RTX | |
446 | #define TARGET_STRUCT_VALUE_RTX sh_struct_value_rtx | |
447 | #undef TARGET_RETURN_IN_MEMORY | |
448 | #define TARGET_RETURN_IN_MEMORY sh_return_in_memory | |
449 | ||
450 | #undef TARGET_EXPAND_BUILTIN_SAVEREGS | |
451 | #define TARGET_EXPAND_BUILTIN_SAVEREGS sh_builtin_saveregs | |
452 | #undef TARGET_SETUP_INCOMING_VARARGS | |
453 | #define TARGET_SETUP_INCOMING_VARARGS sh_setup_incoming_varargs | |
454 | #undef TARGET_STRICT_ARGUMENT_NAMING | |
455 | #define TARGET_STRICT_ARGUMENT_NAMING sh_strict_argument_naming | |
456 | #undef TARGET_PRETEND_OUTGOING_VARARGS_NAMED | |
457 | #define TARGET_PRETEND_OUTGOING_VARARGS_NAMED sh_pretend_outgoing_varargs_named | |
0336f0f0 | 458 | #undef TARGET_MUST_PASS_IN_STACK |
459 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
b981d932 | 460 | #undef TARGET_PASS_BY_REFERENCE |
461 | #define TARGET_PASS_BY_REFERENCE sh_pass_by_reference | |
13f08ee7 | 462 | #undef TARGET_CALLEE_COPIES |
463 | #define TARGET_CALLEE_COPIES sh_callee_copies | |
f054eb3c | 464 | #undef TARGET_ARG_PARTIAL_BYTES |
465 | #define TARGET_ARG_PARTIAL_BYTES sh_arg_partial_bytes | |
45550790 | 466 | |
2e15d750 | 467 | #undef TARGET_BUILD_BUILTIN_VA_LIST |
468 | #define TARGET_BUILD_BUILTIN_VA_LIST sh_build_builtin_va_list | |
20cdebf8 | 469 | #undef TARGET_GIMPLIFY_VA_ARG_EXPR |
470 | #define TARGET_GIMPLIFY_VA_ARG_EXPR sh_gimplify_va_arg_expr | |
2e15d750 | 471 | |
9e7454d0 | 472 | #undef TARGET_VECTOR_MODE_SUPPORTED_P |
473 | #define TARGET_VECTOR_MODE_SUPPORTED_P sh_vector_mode_supported_p | |
474 | ||
761d9126 | 475 | #undef TARGET_CHECK_PCH_TARGET_FLAGS |
476 | #define TARGET_CHECK_PCH_TARGET_FLAGS sh_check_pch_target_flags | |
133479bd | 477 | |
8ff30ff6 | 478 | #undef TARGET_DWARF_CALLING_CONVENTION |
479 | #define TARGET_DWARF_CALLING_CONVENTION sh_dwarf_calling_convention | |
480 | ||
f77a5bb0 | 481 | /* Return regmode weight for insn. */ |
482 | #define INSN_REGMODE_WEIGHT(INSN, MODE) regmode_weight[((MODE) == SImode) ? 0 : 1][INSN_UID (INSN)] | |
483 | ||
484 | /* Return current register pressure for regmode. */ | |
485 | #define CURR_REGMODE_PRESSURE(MODE) curr_regmode_pressure[((MODE) == SImode) ? 0 : 1] | |
486 | ||
db281f5a | 487 | #ifdef SYMBIAN |
488 | ||
489 | #undef TARGET_ENCODE_SECTION_INFO | |
490 | #define TARGET_ENCODE_SECTION_INFO sh_symbian_encode_section_info | |
491 | #undef TARGET_STRIP_NAME_ENCODING | |
492 | #define TARGET_STRIP_NAME_ENCODING sh_symbian_strip_name_encoding | |
493 | #undef TARGET_CXX_IMPORT_EXPORT_CLASS | |
494 | #define TARGET_CXX_IMPORT_EXPORT_CLASS symbian_import_export_class | |
495 | ||
496 | #endif /* SYMBIAN */ | |
497 | ||
59312820 | 498 | #ifdef TARGET_ADJUST_UNROLL_MAX |
499 | #undef TARGET_ADJUST_UNROLL_MAX | |
500 | #define TARGET_ADJUST_UNROLL_MAX sh_adjust_unroll_max | |
501 | #endif | |
502 | ||
57e4bbfb | 503 | struct gcc_target targetm = TARGET_INITIALIZER; |
47c009e5 | 504 | \f |
156fac8b | 505 | /* Implement TARGET_HANDLE_OPTION. */ |
506 | ||
507 | static bool | |
508 | sh_handle_option (size_t code, const char *arg ATTRIBUTE_UNUSED, | |
509 | int value ATTRIBUTE_UNUSED) | |
510 | { | |
511 | switch (code) | |
512 | { | |
513 | case OPT_m1: | |
514 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH1; | |
515 | return true; | |
516 | ||
517 | case OPT_m2: | |
518 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2; | |
519 | return true; | |
520 | ||
521 | case OPT_m2a: | |
522 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A; | |
523 | return true; | |
524 | ||
525 | case OPT_m2a_nofpu: | |
526 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_NOFPU; | |
527 | return true; | |
528 | ||
529 | case OPT_m2a_single: | |
530 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_SINGLE; | |
531 | return true; | |
532 | ||
533 | case OPT_m2a_single_only: | |
534 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2A_SINGLE_ONLY; | |
535 | return true; | |
536 | ||
537 | case OPT_m2e: | |
538 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH2E; | |
539 | return true; | |
540 | ||
541 | case OPT_m3: | |
542 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH3; | |
543 | return true; | |
544 | ||
545 | case OPT_m3e: | |
546 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH3E; | |
547 | return true; | |
548 | ||
549 | case OPT_m4: | |
550 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4; | |
551 | return true; | |
552 | ||
553 | case OPT_m4_nofpu: | |
554 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_NOFPU; | |
555 | return true; | |
556 | ||
557 | case OPT_m4_single: | |
558 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_SINGLE; | |
559 | return true; | |
560 | ||
561 | case OPT_m4_single_only: | |
562 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_SINGLE_ONLY; | |
563 | return true; | |
564 | ||
565 | case OPT_m4a: | |
566 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A; | |
567 | return true; | |
568 | ||
569 | case OPT_m4a_nofpu: | |
570 | case OPT_m4al: | |
571 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_NOFPU; | |
572 | return true; | |
573 | ||
574 | case OPT_m4a_single: | |
575 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_SINGLE; | |
576 | return true; | |
577 | ||
578 | case OPT_m4a_single_only: | |
579 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4A_SINGLE_ONLY; | |
580 | return true; | |
581 | ||
582 | case OPT_m5_32media: | |
583 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_32MEDIA; | |
584 | return true; | |
585 | ||
586 | case OPT_m5_32media_nofpu: | |
587 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_32MEDIA_NOFPU; | |
588 | return true; | |
589 | ||
590 | case OPT_m5_64media: | |
591 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_64MEDIA; | |
592 | return true; | |
593 | ||
594 | case OPT_m5_64media_nofpu: | |
595 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_64MEDIA_NOFPU; | |
596 | return true; | |
597 | ||
598 | case OPT_m5_compact: | |
599 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_COMPACT; | |
600 | return true; | |
601 | ||
602 | case OPT_m5_compact_nofpu: | |
603 | target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH5_COMPACT_NOFPU; | |
604 | return true; | |
605 | ||
606 | default: | |
607 | return true; | |
608 | } | |
609 | } | |
610 | \f | |
1504eb7e | 611 | /* Print the operand address in x to the stream. */ |
7208c779 | 612 | |
613 | void | |
04f04b72 | 614 | print_operand_address (FILE *stream, rtx x) |
7208c779 | 615 | { |
616 | switch (GET_CODE (x)) | |
617 | { | |
618 | case REG: | |
1b61190c | 619 | case SUBREG: |
620 | fprintf (stream, "@%s", reg_names[true_regnum (x)]); | |
7208c779 | 621 | break; |
73401833 | 622 | |
7208c779 | 623 | case PLUS: |
624 | { | |
625 | rtx base = XEXP (x, 0); | |
626 | rtx index = XEXP (x, 1); | |
627 | ||
7208c779 | 628 | switch (GET_CODE (index)) |
629 | { | |
630 | case CONST_INT: | |
26576132 | 631 | fprintf (stream, "@(%d,%s)", (int) INTVAL (index), |
1b61190c | 632 | reg_names[true_regnum (base)]); |
7208c779 | 633 | break; |
634 | ||
635 | case REG: | |
1b61190c | 636 | case SUBREG: |
637 | { | |
638 | int base_num = true_regnum (base); | |
639 | int index_num = true_regnum (index); | |
640 | ||
641 | fprintf (stream, "@(r0,%s)", | |
642 | reg_names[MAX (base_num, index_num)]); | |
643 | break; | |
644 | } | |
7208c779 | 645 | |
646 | default: | |
1a6a7a27 | 647 | gcc_unreachable (); |
7208c779 | 648 | } |
649 | } | |
7208c779 | 650 | break; |
73401833 | 651 | |
7208c779 | 652 | case PRE_DEC: |
1b61190c | 653 | fprintf (stream, "@-%s", reg_names[true_regnum (XEXP (x, 0))]); |
7208c779 | 654 | break; |
655 | ||
656 | case POST_INC: | |
1b61190c | 657 | fprintf (stream, "@%s+", reg_names[true_regnum (XEXP (x, 0))]); |
7208c779 | 658 | break; |
659 | ||
660 | default: | |
ecf6ff7c | 661 | x = mark_constant_pool_use (x); |
b60f0749 | 662 | output_addr_const (stream, x); |
7208c779 | 663 | break; |
664 | } | |
665 | } | |
666 | ||
667 | /* Print operand x (an rtx) in assembler syntax to file stream | |
668 | according to modifier code. | |
669 | ||
47c009e5 | 670 | '.' print a .s if insn needs delay slot |
8ded0752 | 671 | ',' print LOCAL_LABEL_PREFIX |
15cb1799 | 672 | '@' print trap, rte or rts depending upon pragma interruptness |
73401833 | 673 | '#' output a nop if there is nothing to put in the delay slot |
1fc184ee | 674 | ''' print likelihood suffix (/u for unlikely). |
59312820 | 675 | '>' print branch target if -fverbose-asm |
73401833 | 676 | 'O' print a constant without the # |
3dfc1807 | 677 | 'R' print the LSW of a dp value - changes if in little endian |
3dfc1807 | 678 | 'S' print the MSW of a dp value - changes if in little endian |
1b61190c | 679 | 'T' print the next word of a dp value - same as 'R' in big endian mode. |
87e19636 | 680 | 'M' print an `x' if `m' will print `base,index'. |
e3fac27d | 681 | 'N' print 'r63' if the operand is (const_int 0). |
9435e831 | 682 | 'd' print a V2SF reg as dN instead of fpN. |
87e19636 | 683 | 'm' print a pair `base,offset' or `base,index', for LD and ST. |
59312820 | 684 | 'U' Likewise for {LD,ST}{HI,LO}. |
87e19636 | 685 | 'u' prints the lowest 16 bits of CONST_INT, as an unsigned value. |
1b61190c | 686 | 'o' output an operator. */ |
7208c779 | 687 | |
688 | void | |
04f04b72 | 689 | print_operand (FILE *stream, rtx x, int code) |
7208c779 | 690 | { |
59312820 | 691 | int regno; |
692 | enum machine_mode mode; | |
693 | ||
7208c779 | 694 | switch (code) |
695 | { | |
47c009e5 | 696 | case '.': |
3f2613e4 | 697 | if (final_sequence |
36858eff | 698 | && ! INSN_ANNULLED_BRANCH_P (XVECEXP (final_sequence, 0, 0)) |
699 | && get_attr_length (XVECEXP (final_sequence, 0, 1))) | |
8ded0752 | 700 | fprintf (stream, ASSEMBLER_DIALECT ? "/s" : ".s"); |
701 | break; | |
702 | case ',': | |
703 | fprintf (stream, "%s", LOCAL_LABEL_PREFIX); | |
47c009e5 | 704 | break; |
f3d93547 | 705 | case '@': |
15cb1799 | 706 | if (trap_exit) |
707 | fprintf (stream, "trapa #%d", trap_exit); | |
5df3d11f | 708 | else if (sh_cfun_interrupt_handler_p ()) |
2e3d4844 | 709 | fprintf (stream, "rte"); |
f3d93547 | 710 | else |
2e3d4844 | 711 | fprintf (stream, "rts"); |
f3d93547 | 712 | break; |
7208c779 | 713 | case '#': |
1504eb7e | 714 | /* Output a nop if there's nothing in the delay slot. */ |
7208c779 | 715 | if (dbr_sequence_length () == 0) |
73401833 | 716 | fprintf (stream, "\n\tnop"); |
7208c779 | 717 | break; |
3c4bc54e | 718 | case '\'': |
719 | { | |
720 | rtx note = find_reg_note (current_output_insn, REG_BR_PROB, 0); | |
721 | ||
722 | if (note && INTVAL (XEXP (note, 0)) * 2 < REG_BR_PROB_BASE) | |
723 | fputs ("/u", stream); | |
724 | break; | |
725 | } | |
59312820 | 726 | case '>': |
727 | if (flag_verbose_asm && JUMP_LABEL (current_output_insn)) | |
728 | { | |
729 | fputs ("\t! target: ", stream); | |
730 | output_addr_const (stream, JUMP_LABEL (current_output_insn)); | |
731 | } | |
732 | break; | |
47c009e5 | 733 | case 'O': |
ecf6ff7c | 734 | x = mark_constant_pool_use (x); |
97595bfd | 735 | output_addr_const (stream, x); |
7208c779 | 736 | break; |
c8cd1805 | 737 | /* N.B.: %R / %S / %T adjust memory addresses by four. |
738 | For SHMEDIA, that means they can be used to access the first and | |
739 | second 32 bit part of a 64 bit (or larger) value that | |
740 | might be held in floating point registers or memory. | |
741 | While they can be used to access 64 bit parts of a larger value | |
742 | held in general purpose registers, that won't work with memory - | |
743 | neither for fp registers, since the frxx names are used. */ | |
7208c779 | 744 | case 'R': |
c8cd1805 | 745 | if (REG_P (x) || GET_CODE (x) == SUBREG) |
746 | { | |
747 | regno = true_regnum (x); | |
748 | regno += FP_REGISTER_P (regno) ? 1 : LSW; | |
749 | fputs (reg_names[regno], (stream)); | |
750 | } | |
751 | else if (MEM_P (x)) | |
752 | { | |
753 | x = adjust_address (x, SImode, 4 * LSW); | |
754 | print_operand_address (stream, XEXP (x, 0)); | |
755 | } | |
756 | else | |
757 | { | |
758 | rtx sub = NULL_RTX; | |
759 | ||
760 | mode = GET_MODE (x); | |
761 | if (mode == VOIDmode) | |
762 | mode = DImode; | |
763 | if (GET_MODE_SIZE (mode) >= 8) | |
764 | sub = simplify_subreg (SImode, x, mode, 4 * LSW); | |
765 | if (sub) | |
766 | print_operand (stream, sub, 0); | |
767 | else | |
768 | output_operand_lossage ("invalid operand to %%R"); | |
769 | } | |
73401833 | 770 | break; |
771 | case 'S': | |
c8cd1805 | 772 | if (REG_P (x) || GET_CODE (x) == SUBREG) |
773 | { | |
774 | regno = true_regnum (x); | |
775 | regno += FP_REGISTER_P (regno) ? 0 : MSW; | |
776 | fputs (reg_names[regno], (stream)); | |
777 | } | |
778 | else if (MEM_P (x)) | |
779 | { | |
780 | x = adjust_address (x, SImode, 4 * MSW); | |
781 | print_operand_address (stream, XEXP (x, 0)); | |
782 | } | |
783 | else | |
784 | { | |
785 | rtx sub = NULL_RTX; | |
786 | ||
787 | mode = GET_MODE (x); | |
788 | if (mode == VOIDmode) | |
789 | mode = DImode; | |
790 | if (GET_MODE_SIZE (mode) >= 8) | |
791 | sub = simplify_subreg (SImode, x, mode, 4 * MSW); | |
792 | if (sub) | |
793 | print_operand (stream, sub, 0); | |
794 | else | |
795 | output_operand_lossage ("invalid operand to %%S"); | |
796 | } | |
3dfc1807 | 797 | break; |
798 | case 'T': | |
1504eb7e | 799 | /* Next word of a double. */ |
7208c779 | 800 | switch (GET_CODE (x)) |
801 | { | |
802 | case REG: | |
803 | fputs (reg_names[REGNO (x) + 1], (stream)); | |
804 | break; | |
805 | case MEM: | |
1b61190c | 806 | if (GET_CODE (XEXP (x, 0)) != PRE_DEC |
807 | && GET_CODE (XEXP (x, 0)) != POST_INC) | |
eafc6604 | 808 | x = adjust_address (x, SImode, 4); |
1b61190c | 809 | print_operand_address (stream, XEXP (x, 0)); |
7208c779 | 810 | break; |
b7dbbdb2 | 811 | default: |
812 | break; | |
7208c779 | 813 | } |
814 | break; | |
1b61190c | 815 | case 'o': |
816 | switch (GET_CODE (x)) | |
817 | { | |
818 | case PLUS: fputs ("add", stream); break; | |
819 | case MINUS: fputs ("sub", stream); break; | |
820 | case MULT: fputs ("mul", stream); break; | |
821 | case DIV: fputs ("div", stream); break; | |
3c4bc54e | 822 | case EQ: fputs ("eq", stream); break; |
823 | case NE: fputs ("ne", stream); break; | |
824 | case GT: case LT: fputs ("gt", stream); break; | |
825 | case GE: case LE: fputs ("ge", stream); break; | |
826 | case GTU: case LTU: fputs ("gtu", stream); break; | |
827 | case GEU: case LEU: fputs ("geu", stream); break; | |
b7dbbdb2 | 828 | default: |
829 | break; | |
1b61190c | 830 | } |
831 | break; | |
87e19636 | 832 | case 'M': |
833 | if (GET_CODE (x) == MEM | |
834 | && GET_CODE (XEXP (x, 0)) == PLUS | |
835 | && (GET_CODE (XEXP (XEXP (x, 0), 1)) == REG | |
836 | || GET_CODE (XEXP (XEXP (x, 0), 1)) == SUBREG)) | |
837 | fputc ('x', stream); | |
838 | break; | |
839 | ||
840 | case 'm': | |
1a6a7a27 | 841 | gcc_assert (GET_CODE (x) == MEM); |
87e19636 | 842 | x = XEXP (x, 0); |
59312820 | 843 | /* Fall through. */ |
844 | case 'U': | |
87e19636 | 845 | switch (GET_CODE (x)) |
846 | { | |
847 | case REG: | |
848 | case SUBREG: | |
849 | print_operand (stream, x, 0); | |
850 | fputs (", 0", stream); | |
851 | break; | |
852 | ||
853 | case PLUS: | |
854 | print_operand (stream, XEXP (x, 0), 0); | |
855 | fputs (", ", stream); | |
856 | print_operand (stream, XEXP (x, 1), 0); | |
857 | break; | |
858 | ||
859 | default: | |
1a6a7a27 | 860 | gcc_unreachable (); |
87e19636 | 861 | } |
862 | break; | |
863 | ||
9435e831 | 864 | case 'd': |
1a6a7a27 | 865 | gcc_assert (GET_CODE (x) == REG && GET_MODE (x) == V2SFmode); |
9435e831 | 866 | |
867 | fprintf ((stream), "d%s", reg_names[REGNO (x)] + 1); | |
868 | break; | |
9e7454d0 | 869 | |
e3fac27d | 870 | case 'N': |
d3b29bbd | 871 | if (x == CONST0_RTX (GET_MODE (x))) |
e3fac27d | 872 | { |
873 | fprintf ((stream), "r63"); | |
874 | break; | |
875 | } | |
876 | goto default_output; | |
87e19636 | 877 | case 'u': |
878 | if (GET_CODE (x) == CONST_INT) | |
3c4bc54e | 879 | { |
87e19636 | 880 | fprintf ((stream), "%u", (unsigned) INTVAL (x) & (0x10000 - 1)); |
881 | break; | |
882 | } | |
883 | /* Fall through. */ | |
884 | ||
e3fac27d | 885 | default_output: |
7208c779 | 886 | default: |
59312820 | 887 | regno = 0; |
888 | mode = GET_MODE (x); | |
889 | ||
7208c779 | 890 | switch (GET_CODE (x)) |
891 | { | |
59312820 | 892 | case TRUNCATE: |
893 | { | |
894 | rtx inner = XEXP (x, 0); | |
895 | int offset = 0; | |
896 | enum machine_mode inner_mode; | |
897 | ||
898 | /* We might see SUBREGs with vector mode registers inside. */ | |
899 | if (GET_CODE (inner) == SUBREG | |
900 | && (GET_MODE_SIZE (GET_MODE (inner)) | |
901 | == GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner)))) | |
902 | && subreg_lowpart_p (inner)) | |
903 | inner = SUBREG_REG (inner); | |
904 | if (GET_CODE (inner) == CONST_INT) | |
905 | { | |
906 | x = GEN_INT (trunc_int_for_mode (INTVAL (inner), GET_MODE (x))); | |
907 | goto default_output; | |
908 | } | |
909 | inner_mode = GET_MODE (inner); | |
910 | if (GET_CODE (inner) == SUBREG | |
911 | && (GET_MODE_SIZE (GET_MODE (inner)) | |
912 | < GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner)))) | |
913 | && GET_CODE (SUBREG_REG (inner)) == REG) | |
914 | { | |
915 | offset = subreg_regno_offset (REGNO (SUBREG_REG (inner)), | |
916 | GET_MODE (SUBREG_REG (inner)), | |
917 | SUBREG_BYTE (inner), | |
918 | GET_MODE (inner)); | |
919 | inner = SUBREG_REG (inner); | |
920 | } | |
921 | if (GET_CODE (inner) != REG || GET_MODE_SIZE (inner_mode) > 8) | |
922 | abort (); | |
923 | /* Floating point register pairs are always big endian; | |
35fe320d | 924 | general purpose registers are 64 bit wide. */ |
59312820 | 925 | regno = REGNO (inner); |
926 | regno = (HARD_REGNO_NREGS (regno, inner_mode) | |
927 | - HARD_REGNO_NREGS (regno, mode)) | |
928 | + offset; | |
929 | x = inner; | |
930 | goto reg; | |
931 | } | |
932 | case SIGN_EXTEND: | |
933 | x = XEXP (x, 0); | |
934 | goto reg; | |
87e19636 | 935 | /* FIXME: We need this on SHmedia32 because reload generates |
936 | some sign-extended HI or QI loads into DImode registers | |
937 | but, because Pmode is SImode, the address ends up with a | |
938 | subreg:SI of the DImode register. Maybe reload should be | |
939 | fixed so as to apply alter_subreg to such loads? */ | |
59312820 | 940 | case IF_THEN_ELSE: |
941 | gcc_assert (trapping_target_operand (x, VOIDmode)); | |
942 | x = XEXP (XEXP (x, 2), 0); | |
943 | goto default_output; | |
87e19636 | 944 | case SUBREG: |
1a6a7a27 | 945 | gcc_assert (SUBREG_BYTE (x) == 0 |
946 | && GET_CODE (SUBREG_REG (x)) == REG); | |
87e19636 | 947 | |
948 | x = SUBREG_REG (x); | |
949 | /* Fall through. */ | |
950 | ||
59312820 | 951 | reg: |
7208c779 | 952 | case REG: |
59312820 | 953 | regno += REGNO (x); |
954 | if (FP_REGISTER_P (regno) | |
955 | && mode == V16SFmode) | |
956 | fprintf ((stream), "mtrx%s", reg_names[regno] + 2); | |
87e19636 | 957 | else if (FP_REGISTER_P (REGNO (x)) |
59312820 | 958 | && mode == V4SFmode) |
959 | fprintf ((stream), "fv%s", reg_names[regno] + 2); | |
87e19636 | 960 | else if (GET_CODE (x) == REG |
59312820 | 961 | && mode == V2SFmode) |
962 | fprintf ((stream), "fp%s", reg_names[regno] + 2); | |
87e19636 | 963 | else if (FP_REGISTER_P (REGNO (x)) |
59312820 | 964 | && GET_MODE_SIZE (mode) > 4) |
965 | fprintf ((stream), "d%s", reg_names[regno] + 1); | |
1b61190c | 966 | else |
59312820 | 967 | fputs (reg_names[regno], (stream)); |
7208c779 | 968 | break; |
87e19636 | 969 | |
7208c779 | 970 | case MEM: |
971 | output_address (XEXP (x, 0)); | |
972 | break; | |
9e7454d0 | 973 | |
87e19636 | 974 | case CONST: |
975 | if (TARGET_SHMEDIA | |
976 | && GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
59312820 | 977 | && (GET_MODE (XEXP (x, 0)) == DImode |
978 | || GET_MODE (XEXP (x, 0)) == SImode) | |
87e19636 | 979 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == TRUNCATE |
980 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == HImode) | |
981 | { | |
982 | rtx val = XEXP (XEXP (XEXP (x, 0), 0), 0); | |
983 | ||
984 | fputc ('(', stream); | |
985 | if (GET_CODE (val) == ASHIFTRT) | |
986 | { | |
987 | fputc ('(', stream); | |
988 | if (GET_CODE (XEXP (val, 0)) == CONST) | |
989 | fputc ('(', stream); | |
990 | output_addr_const (stream, XEXP (val, 0)); | |
991 | if (GET_CODE (XEXP (val, 0)) == CONST) | |
992 | fputc (')', stream); | |
993 | fputs (" >> ", stream); | |
994 | output_addr_const (stream, XEXP (val, 1)); | |
995 | fputc (')', stream); | |
996 | } | |
997 | else | |
998 | { | |
999 | if (GET_CODE (val) == CONST) | |
1000 | fputc ('(', stream); | |
1001 | output_addr_const (stream, val); | |
1002 | if (GET_CODE (val) == CONST) | |
1003 | fputc (')', stream); | |
1004 | } | |
1005 | fputs (" & 65535)", stream); | |
1006 | break; | |
1007 | } | |
1008 | ||
1009 | /* Fall through. */ | |
7208c779 | 1010 | default: |
87e19636 | 1011 | if (TARGET_SH1) |
1012 | fputc ('#', stream); | |
7208c779 | 1013 | output_addr_const (stream, x); |
1014 | break; | |
7208c779 | 1015 | } |
1016 | break; | |
1017 | } | |
1018 | } | |
7208c779 | 1019 | \f |
eb400f70 | 1020 | /* Like force_operand, but guarantees that VALUE ends up in TARGET. */ |
1021 | static void | |
04f04b72 | 1022 | force_into (rtx value, rtx target) |
eb400f70 | 1023 | { |
1024 | value = force_operand (value, target); | |
1025 | if (! rtx_equal_p (value, target)) | |
1026 | emit_insn (gen_move_insn (target, value)); | |
1027 | } | |
1028 | ||
97595bfd | 1029 | /* Emit code to perform a block move. Choose the best method. |
1030 | ||
1031 | OPERANDS[0] is the destination. | |
1032 | OPERANDS[1] is the source. | |
1033 | OPERANDS[2] is the size. | |
1034 | OPERANDS[3] is the alignment safe to use. */ | |
1035 | ||
97595bfd | 1036 | int |
04f04b72 | 1037 | expand_block_move (rtx *operands) |
97595bfd | 1038 | { |
1039 | int align = INTVAL (operands[3]); | |
1040 | int constp = (GET_CODE (operands[2]) == CONST_INT); | |
1041 | int bytes = (constp ? INTVAL (operands[2]) : 0); | |
97595bfd | 1042 | |
9435e831 | 1043 | if (! constp) |
1044 | return 0; | |
1045 | ||
1046 | /* If we could use mov.l to move words and dest is word-aligned, we | |
1047 | can use movua.l for loads and still generate a relatively short | |
1048 | and efficient sequence. */ | |
1049 | if (TARGET_SH4A_ARCH && align < 4 | |
1050 | && MEM_ALIGN (operands[0]) >= 32 | |
1051 | && can_move_by_pieces (bytes, 32)) | |
1052 | { | |
1053 | rtx dest = copy_rtx (operands[0]); | |
1054 | rtx src = copy_rtx (operands[1]); | |
1055 | /* We could use different pseudos for each copied word, but | |
1056 | since movua can only load into r0, it's kind of | |
1057 | pointless. */ | |
1058 | rtx temp = gen_reg_rtx (SImode); | |
1059 | rtx src_addr = copy_addr_to_reg (XEXP (src, 0)); | |
1060 | int copied = 0; | |
1061 | ||
1062 | while (copied + 4 <= bytes) | |
1063 | { | |
1064 | rtx to = adjust_address (dest, SImode, copied); | |
1065 | rtx from = adjust_automodify_address (src, SImode, src_addr, copied); | |
1066 | ||
1067 | emit_insn (gen_movua (temp, from)); | |
1068 | emit_move_insn (src_addr, plus_constant (src_addr, 4)); | |
1069 | emit_move_insn (to, temp); | |
1070 | copied += 4; | |
1071 | } | |
1072 | ||
1073 | if (copied < bytes) | |
1074 | move_by_pieces (adjust_address (dest, BLKmode, copied), | |
1075 | adjust_automodify_address (src, BLKmode, | |
1076 | src_addr, copied), | |
1077 | bytes - copied, align, 0); | |
1078 | ||
1079 | return 1; | |
1080 | } | |
1081 | ||
73401833 | 1082 | /* If it isn't a constant number of bytes, or if it doesn't have 4 byte |
1083 | alignment, or if it isn't a multiple of 4 bytes, then fail. */ | |
9435e831 | 1084 | if (align < 4 || (bytes % 4 != 0)) |
f3d93547 | 1085 | return 0; |
1086 | ||
1b61190c | 1087 | if (TARGET_HARD_SH4) |
1088 | { | |
1089 | if (bytes < 12) | |
1090 | return 0; | |
1091 | else if (bytes == 12) | |
1092 | { | |
59312820 | 1093 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
407d9153 | 1094 | rtx r4 = gen_rtx_REG (SImode, 4); |
1095 | rtx r5 = gen_rtx_REG (SImode, 5); | |
1b61190c | 1096 | |
59312820 | 1097 | function_symbol (func_addr_rtx, "__movmemSI12_i4", SFUNC_STATIC); |
eb400f70 | 1098 | force_into (XEXP (operands[0], 0), r4); |
1099 | force_into (XEXP (operands[1], 0), r5); | |
1b61190c | 1100 | emit_insn (gen_block_move_real_i4 (func_addr_rtx)); |
1101 | return 1; | |
1102 | } | |
1103 | else if (! TARGET_SMALLCODE) | |
1104 | { | |
59312820 | 1105 | const char *entry_name; |
1106 | rtx func_addr_rtx = gen_reg_rtx (Pmode); | |
1b61190c | 1107 | int dwords; |
407d9153 | 1108 | rtx r4 = gen_rtx_REG (SImode, 4); |
1109 | rtx r5 = gen_rtx_REG (SImode, 5); | |
1110 | rtx r6 = gen_rtx_REG (SImode, 6); | |
1b61190c | 1111 | |
59312820 | 1112 | entry_name = (bytes & 4 ? "__movmem_i4_odd" : "__movmem_i4_even"); |
1113 | function_symbol (func_addr_rtx, entry_name, SFUNC_STATIC); | |
eb400f70 | 1114 | force_into (XEXP (operands[0], 0), r4); |
1115 | force_into (XEXP (operands[1], 0), r5); | |
1b61190c | 1116 | |
1117 | dwords = bytes >> 3; | |
1118 | emit_insn (gen_move_insn (r6, GEN_INT (dwords - 1))); | |
1119 | emit_insn (gen_block_lump_real_i4 (func_addr_rtx)); | |
1120 | return 1; | |
1121 | } | |
1122 | else | |
1123 | return 0; | |
1124 | } | |
73401833 | 1125 | if (bytes < 64) |
97595bfd | 1126 | { |
1127 | char entry[30]; | |
59312820 | 1128 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
7014838c | 1129 | rtx r4 = gen_rtx_REG (SImode, 4); |
1130 | rtx r5 = gen_rtx_REG (SImode, 5); | |
73401833 | 1131 | |
008c057d | 1132 | sprintf (entry, "__movmemSI%d", bytes); |
59312820 | 1133 | function_symbol (func_addr_rtx, entry, SFUNC_STATIC); |
eb400f70 | 1134 | force_into (XEXP (operands[0], 0), r4); |
1135 | force_into (XEXP (operands[1], 0), r5); | |
97595bfd | 1136 | emit_insn (gen_block_move_real (func_addr_rtx)); |
1137 | return 1; | |
1138 | } | |
73401833 | 1139 | |
1140 | /* This is the same number of bytes as a memcpy call, but to a different | |
1141 | less common function name, so this will occasionally use more space. */ | |
1142 | if (! TARGET_SMALLCODE) | |
97595bfd | 1143 | { |
59312820 | 1144 | rtx func_addr_rtx = gen_reg_rtx (Pmode); |
73401833 | 1145 | int final_switch, while_loop; |
7014838c | 1146 | rtx r4 = gen_rtx_REG (SImode, 4); |
1147 | rtx r5 = gen_rtx_REG (SImode, 5); | |
1148 | rtx r6 = gen_rtx_REG (SImode, 6); | |
97595bfd | 1149 | |
59312820 | 1150 | function_symbol (func_addr_rtx, "__movmem", SFUNC_STATIC); |
eb400f70 | 1151 | force_into (XEXP (operands[0], 0), r4); |
1152 | force_into (XEXP (operands[1], 0), r5); | |
97595bfd | 1153 | |
73401833 | 1154 | /* r6 controls the size of the move. 16 is decremented from it |
1155 | for each 64 bytes moved. Then the negative bit left over is used | |
1156 | as an index into a list of move instructions. e.g., a 72 byte move | |
1157 | would be set up with size(r6) = 14, for one iteration through the | |
1158 | big while loop, and a switch of -2 for the last part. */ | |
97595bfd | 1159 | |
73401833 | 1160 | final_switch = 16 - ((bytes / 4) % 16); |
1161 | while_loop = ((bytes / 4) / 16 - 1) * 16; | |
1162 | emit_insn (gen_move_insn (r6, GEN_INT (while_loop + final_switch))); | |
1163 | emit_insn (gen_block_lump_real (func_addr_rtx)); | |
1164 | return 1; | |
97595bfd | 1165 | } |
97595bfd | 1166 | |
f3d93547 | 1167 | return 0; |
97595bfd | 1168 | } |
1169 | ||
7208c779 | 1170 | /* Prepare operands for a move define_expand; specifically, one of the |
73401833 | 1171 | operands must be in a register. */ |
7208c779 | 1172 | |
47c009e5 | 1173 | int |
04f04b72 | 1174 | prepare_move_operands (rtx operands[], enum machine_mode mode) |
7208c779 | 1175 | { |
cec78553 | 1176 | if ((mode == SImode || mode == DImode) |
1177 | && flag_pic | |
1178 | && ! ((mode == Pmode || mode == ptr_mode) | |
1179 | && tls_symbolic_operand (operands[1], Pmode) != 0)) | |
0abf894c | 1180 | { |
1181 | rtx temp; | |
1182 | if (SYMBOLIC_CONST_P (operands[1])) | |
1183 | { | |
1184 | if (GET_CODE (operands[0]) == MEM) | |
1185 | operands[1] = force_reg (Pmode, operands[1]); | |
afef6540 | 1186 | else if (TARGET_SHMEDIA |
1187 | && GET_CODE (operands[1]) == LABEL_REF | |
87e19636 | 1188 | && target_reg_operand (operands[0], mode)) |
1189 | /* It's ok. */; | |
0abf894c | 1190 | else |
1191 | { | |
1192 | temp = no_new_pseudos ? operands[0] : gen_reg_rtx (Pmode); | |
87e19636 | 1193 | operands[1] = legitimize_pic_address (operands[1], mode, temp); |
0abf894c | 1194 | } |
1195 | } | |
1196 | else if (GET_CODE (operands[1]) == CONST | |
1197 | && GET_CODE (XEXP (operands[1], 0)) == PLUS | |
1198 | && SYMBOLIC_CONST_P (XEXP (XEXP (operands[1], 0), 0))) | |
1199 | { | |
6883f2aa | 1200 | temp = no_new_pseudos ? operands[0] : gen_reg_rtx (Pmode); |
0abf894c | 1201 | temp = legitimize_pic_address (XEXP (XEXP (operands[1], 0), 0), |
87e19636 | 1202 | mode, temp); |
1203 | operands[1] = expand_binop (mode, add_optab, temp, | |
0abf894c | 1204 | XEXP (XEXP (operands[1], 0), 1), |
6883f2aa | 1205 | no_new_pseudos ? temp |
1206 | : gen_reg_rtx (Pmode), | |
1207 | 0, OPTAB_LIB_WIDEN); | |
0abf894c | 1208 | } |
1209 | } | |
1210 | ||
5beb0a57 | 1211 | if (! reload_in_progress && ! reload_completed) |
1212 | { | |
1213 | /* Copy the source to a register if both operands aren't registers. */ | |
1214 | if (! register_operand (operands[0], mode) | |
0edea8ba | 1215 | && ! sh_register_operand (operands[1], mode)) |
5beb0a57 | 1216 | operands[1] = copy_to_mode_reg (mode, operands[1]); |
1217 | ||
7bc2d3b3 | 1218 | if (GET_CODE (operands[0]) == MEM && ! memory_operand (operands[0], mode)) |
1219 | { | |
1220 | /* This is like change_address_1 (operands[0], mode, 0, 1) , | |
1221 | except that we can't use that function because it is static. */ | |
1222 | rtx new = change_address (operands[0], mode, 0); | |
1223 | MEM_COPY_ATTRIBUTES (new, operands[0]); | |
1224 | operands[0] = new; | |
1225 | } | |
1226 | ||
5beb0a57 | 1227 | /* This case can happen while generating code to move the result |
1228 | of a library call to the target. Reject `st r0,@(rX,rY)' because | |
1229 | reload will fail to find a spill register for rX, since r0 is already | |
1230 | being used for the source. */ | |
59312820 | 1231 | else if (TARGET_SH1 |
1232 | && refers_to_regno_p (R0_REG, R0_REG + 1, operands[1], (rtx *)0) | |
5beb0a57 | 1233 | && GET_CODE (operands[0]) == MEM |
1234 | && GET_CODE (XEXP (operands[0], 0)) == PLUS | |
1235 | && GET_CODE (XEXP (XEXP (operands[0], 0), 1)) == REG) | |
1236 | operands[1] = copy_to_mode_reg (mode, operands[1]); | |
1237 | } | |
73401833 | 1238 | |
cec78553 | 1239 | if (mode == Pmode || mode == ptr_mode) |
1240 | { | |
1241 | rtx op0, op1; | |
1242 | enum tls_model tls_kind; | |
1243 | ||
1244 | op0 = operands[0]; | |
1245 | op1 = operands[1]; | |
1246 | if ((tls_kind = tls_symbolic_operand (op1, Pmode))) | |
1247 | { | |
1248 | rtx tga_op1, tga_ret, tmp, tmp2; | |
1249 | ||
cec78553 | 1250 | switch (tls_kind) |
1251 | { | |
1252 | case TLS_MODEL_GLOBAL_DYNAMIC: | |
1253 | tga_ret = gen_rtx_REG (Pmode, R0_REG); | |
8cdd3257 | 1254 | emit_call_insn (gen_tls_global_dynamic (tga_ret, op1)); |
cec78553 | 1255 | op1 = tga_ret; |
1256 | break; | |
1257 | ||
1258 | case TLS_MODEL_LOCAL_DYNAMIC: | |
1259 | tga_ret = gen_rtx_REG (Pmode, R0_REG); | |
8cdd3257 | 1260 | emit_call_insn (gen_tls_local_dynamic (tga_ret, op1)); |
cec78553 | 1261 | |
1262 | tmp = gen_reg_rtx (Pmode); | |
1263 | emit_move_insn (tmp, tga_ret); | |
1264 | ||
1265 | if (register_operand (op0, Pmode)) | |
1266 | tmp2 = op0; | |
1267 | else | |
1268 | tmp2 = gen_reg_rtx (Pmode); | |
1269 | ||
1270 | emit_insn (gen_symDTPOFF2reg (tmp2, op1, tmp)); | |
1271 | op1 = tmp2; | |
1272 | break; | |
1273 | ||
1274 | case TLS_MODEL_INITIAL_EXEC: | |
1275 | if (! flag_pic) | |
9e4e504c | 1276 | { |
c8d059f8 | 1277 | /* Don't schedule insns for getting GOT address when |
1278 | the first scheduling is enabled, to avoid spill | |
1279 | failures for R0. */ | |
1280 | if (flag_schedule_insns) | |
1281 | emit_insn (gen_blockage ()); | |
9e4e504c | 1282 | emit_insn (gen_GOTaddr2picreg ()); |
1283 | emit_insn (gen_rtx_USE (VOIDmode, gen_rtx_REG (SImode, | |
1284 | PIC_REG))); | |
c8d059f8 | 1285 | if (flag_schedule_insns) |
1286 | emit_insn (gen_blockage ()); | |
9e4e504c | 1287 | } |
756343c6 | 1288 | tga_op1 = no_new_pseudos ? op0 : gen_reg_rtx (Pmode); |
cec78553 | 1289 | tmp = gen_sym2GOTTPOFF (op1); |
1290 | emit_insn (gen_tls_initial_exec (tga_op1, tmp)); | |
1291 | op1 = tga_op1; | |
1292 | break; | |
1293 | ||
1294 | case TLS_MODEL_LOCAL_EXEC: | |
1295 | tmp2 = gen_reg_rtx (Pmode); | |
1296 | emit_insn (gen_load_gbr (tmp2)); | |
1297 | tmp = gen_reg_rtx (Pmode); | |
1298 | emit_insn (gen_symTPOFF2reg (tmp, op1)); | |
cec78553 | 1299 | |
1300 | if (register_operand (op0, Pmode)) | |
1301 | op1 = op0; | |
1302 | else | |
1303 | op1 = gen_reg_rtx (Pmode); | |
1304 | ||
1305 | emit_insn (gen_addsi3 (op1, tmp, tmp2)); | |
1306 | break; | |
1307 | ||
1308 | default: | |
1a6a7a27 | 1309 | gcc_unreachable (); |
cec78553 | 1310 | } |
1311 | operands[1] = op1; | |
1312 | } | |
1313 | } | |
1314 | ||
73401833 | 1315 | return 0; |
1316 | } | |
1317 | ||
1318 | /* Prepare the operands for an scc instruction; make sure that the | |
1319 | compare has been done. */ | |
1320 | rtx | |
04f04b72 | 1321 | prepare_scc_operands (enum rtx_code code) |
73401833 | 1322 | { |
7014838c | 1323 | rtx t_reg = gen_rtx_REG (SImode, T_REG); |
73401833 | 1324 | enum rtx_code oldcode = code; |
e9646547 | 1325 | enum machine_mode mode; |
73401833 | 1326 | |
1327 | /* First need a compare insn. */ | |
1328 | switch (code) | |
7208c779 | 1329 | { |
73401833 | 1330 | case NE: |
1331 | /* It isn't possible to handle this case. */ | |
1a6a7a27 | 1332 | gcc_unreachable (); |
73401833 | 1333 | case LT: |
1334 | code = GT; | |
1335 | break; | |
1336 | case LE: | |
1337 | code = GE; | |
1338 | break; | |
1339 | case LTU: | |
1340 | code = GTU; | |
1341 | break; | |
1342 | case LEU: | |
1343 | code = GEU; | |
1344 | break; | |
7984f145 | 1345 | default: |
1346 | break; | |
7208c779 | 1347 | } |
73401833 | 1348 | if (code != oldcode) |
47c009e5 | 1349 | { |
73401833 | 1350 | rtx tmp = sh_compare_op0; |
1351 | sh_compare_op0 = sh_compare_op1; | |
1352 | sh_compare_op1 = tmp; | |
97595bfd | 1353 | } |
47c009e5 | 1354 | |
e9646547 | 1355 | mode = GET_MODE (sh_compare_op0); |
1356 | if (mode == VOIDmode) | |
1357 | mode = GET_MODE (sh_compare_op1); | |
1358 | ||
1359 | sh_compare_op0 = force_reg (mode, sh_compare_op0); | |
8ded0752 | 1360 | if ((code != EQ && code != NE |
1361 | && (sh_compare_op1 != const0_rtx | |
1362 | || code == GTU || code == GEU || code == LTU || code == LEU)) | |
e5d518fa | 1363 | || (mode == DImode && sh_compare_op1 != const0_rtx) |
87ed74ef | 1364 | || (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT)) |
e9646547 | 1365 | sh_compare_op1 = force_reg (mode, sh_compare_op1); |
97595bfd | 1366 | |
7105fb72 | 1367 | if ((TARGET_SH4 || TARGET_SH2A) && GET_MODE_CLASS (mode) == MODE_FLOAT) |
1b61190c | 1368 | (mode == SFmode ? emit_sf_insn : emit_df_insn) |
407d9153 | 1369 | (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, |
1370 | gen_rtx_SET (VOIDmode, t_reg, | |
dc55b8a9 | 1371 | gen_rtx_fmt_ee (code, SImode, |
1372 | sh_compare_op0, sh_compare_op1)), | |
407d9153 | 1373 | gen_rtx_USE (VOIDmode, get_fpscr_rtx ())))); |
1b61190c | 1374 | else |
407d9153 | 1375 | emit_insn (gen_rtx_SET (VOIDmode, t_reg, |
dc55b8a9 | 1376 | gen_rtx_fmt_ee (code, SImode, |
1377 | sh_compare_op0, sh_compare_op1))); | |
97595bfd | 1378 | |
73401833 | 1379 | return t_reg; |
7208c779 | 1380 | } |
1381 | ||
73401833 | 1382 | /* Called from the md file, set up the operands of a compare instruction. */ |
1383 | ||
1384 | void | |
04f04b72 | 1385 | from_compare (rtx *operands, int code) |
7208c779 | 1386 | { |
8ded0752 | 1387 | enum machine_mode mode = GET_MODE (sh_compare_op0); |
1388 | rtx insn; | |
1389 | if (mode == VOIDmode) | |
1390 | mode = GET_MODE (sh_compare_op1); | |
1391 | if (code != EQ | |
1392 | || mode == DImode | |
87ed74ef | 1393 | || (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT)) |
7208c779 | 1394 | { |
73401833 | 1395 | /* Force args into regs, since we can't use constants here. */ |
65bc3ae6 | 1396 | sh_compare_op0 = force_reg (mode, sh_compare_op0); |
73401833 | 1397 | if (sh_compare_op1 != const0_rtx |
8ded0752 | 1398 | || code == GTU || code == GEU |
87ed74ef | 1399 | || (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT)) |
65bc3ae6 | 1400 | sh_compare_op1 = force_reg (mode, sh_compare_op1); |
7208c779 | 1401 | } |
87ed74ef | 1402 | if (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT && code == GE) |
8ded0752 | 1403 | { |
1404 | from_compare (operands, GT); | |
1405 | insn = gen_ieee_ccmpeqsf_t (sh_compare_op0, sh_compare_op1); | |
1406 | } | |
1407 | else | |
7014838c | 1408 | insn = gen_rtx_SET (VOIDmode, |
a5bd34f1 | 1409 | gen_rtx_REG (SImode, T_REG), |
dc55b8a9 | 1410 | gen_rtx_fmt_ee (code, SImode, |
1411 | sh_compare_op0, sh_compare_op1)); | |
7105fb72 | 1412 | if ((TARGET_SH4 || TARGET_SH2A) && GET_MODE_CLASS (mode) == MODE_FLOAT) |
1b61190c | 1413 | { |
407d9153 | 1414 | insn = gen_rtx_PARALLEL (VOIDmode, |
1b61190c | 1415 | gen_rtvec (2, insn, |
407d9153 | 1416 | gen_rtx_USE (VOIDmode, get_fpscr_rtx ()))); |
1b61190c | 1417 | (mode == SFmode ? emit_sf_insn : emit_df_insn) (insn); |
1418 | } | |
1419 | else | |
1420 | emit_insn (insn); | |
7208c779 | 1421 | } |
1422 | \f | |
1504eb7e | 1423 | /* Functions to output assembly code. */ |
7208c779 | 1424 | |
47c009e5 | 1425 | /* Return a sequence of instructions to perform DI or DF move. |
7208c779 | 1426 | |
47c009e5 | 1427 | Since the SH cannot move a DI or DF in one instruction, we have |
1504eb7e | 1428 | to take care when we see overlapping source and dest registers. */ |
97595bfd | 1429 | |
b7dbbdb2 | 1430 | const char * |
04f04b72 | 1431 | output_movedouble (rtx insn ATTRIBUTE_UNUSED, rtx operands[], |
1432 | enum machine_mode mode) | |
7208c779 | 1433 | { |
47c009e5 | 1434 | rtx dst = operands[0]; |
1435 | rtx src = operands[1]; | |
47c009e5 | 1436 | |
97595bfd | 1437 | if (GET_CODE (dst) == MEM |
1504eb7e | 1438 | && GET_CODE (XEXP (dst, 0)) == PRE_DEC) |
73401833 | 1439 | return "mov.l %T1,%0\n\tmov.l %1,%0"; |
1440 | ||
47c009e5 | 1441 | if (register_operand (dst, mode) |
1442 | && register_operand (src, mode)) | |
7208c779 | 1443 | { |
47c009e5 | 1444 | if (REGNO (src) == MACH_REG) |
3dfc1807 | 1445 | return "sts mach,%S0\n\tsts macl,%R0"; |
7208c779 | 1446 | |
73401833 | 1447 | /* When mov.d r1,r2 do r2->r3 then r1->r2; |
1448 | when mov.d r1,r0 do r1->r0 then r2->r1. */ | |
47c009e5 | 1449 | |
1450 | if (REGNO (src) + 1 == REGNO (dst)) | |
335e16a0 | 1451 | return "mov %T1,%T0\n\tmov %1,%0"; |
47c009e5 | 1452 | else |
335e16a0 | 1453 | return "mov %1,%0\n\tmov %T1,%T0"; |
47c009e5 | 1454 | } |
1455 | else if (GET_CODE (src) == CONST_INT) | |
7208c779 | 1456 | { |
73401833 | 1457 | if (INTVAL (src) < 0) |
1458 | output_asm_insn ("mov #-1,%S0", operands); | |
7208c779 | 1459 | else |
73401833 | 1460 | output_asm_insn ("mov #0,%S0", operands); |
7208c779 | 1461 | |
73401833 | 1462 | return "mov %1,%R0"; |
97595bfd | 1463 | } |
47c009e5 | 1464 | else if (GET_CODE (src) == MEM) |
7208c779 | 1465 | { |
73401833 | 1466 | int ptrreg = -1; |
47c009e5 | 1467 | int dreg = REGNO (dst); |
1468 | rtx inside = XEXP (src, 0); | |
7208c779 | 1469 | |
1a6a7a27 | 1470 | switch (GET_CODE (inside)) |
7208c779 | 1471 | { |
1a6a7a27 | 1472 | case REG: |
1473 | ptrreg = REGNO (inside); | |
1474 | break; | |
1475 | ||
1476 | case SUBREG: | |
1477 | ptrreg = subreg_regno (inside); | |
1478 | break; | |
1479 | ||
1480 | case PLUS: | |
73401833 | 1481 | ptrreg = REGNO (XEXP (inside, 0)); |
1482 | /* ??? A r0+REG address shouldn't be possible here, because it isn't | |
1483 | an offsettable address. Unfortunately, offsettable addresses use | |
1484 | QImode to check the offset, and a QImode offsettable address | |
1485 | requires r0 for the other operand, which is not currently | |
1486 | supported, so we can't use the 'o' constraint. | |
1487 | Thus we must check for and handle r0+REG addresses here. | |
1488 | We punt for now, since this is likely very rare. */ | |
1a6a7a27 | 1489 | gcc_assert (GET_CODE (XEXP (inside, 1)) != REG); |
1490 | break; | |
1491 | ||
1492 | case LABEL_REF: | |
1493 | return "mov.l %1,%0\n\tmov.l %1+4,%T0"; | |
1494 | case POST_INC: | |
1495 | return "mov.l %1,%0\n\tmov.l %1,%T0"; | |
1496 | default: | |
1497 | gcc_unreachable (); | |
7208c779 | 1498 | } |
7208c779 | 1499 | |
73401833 | 1500 | /* Work out the safe way to copy. Copy into the second half first. */ |
1501 | if (dreg == ptrreg) | |
1502 | return "mov.l %T1,%T0\n\tmov.l %1,%0"; | |
7208c779 | 1503 | } |
1504 | ||
3dfc1807 | 1505 | return "mov.l %1,%0\n\tmov.l %T1,%T0"; |
7208c779 | 1506 | } |
1507 | ||
73401833 | 1508 | /* Print an instruction which would have gone into a delay slot after |
1509 | another instruction, but couldn't because the other instruction expanded | |
1510 | into a sequence where putting the slot insn at the end wouldn't work. */ | |
97595bfd | 1511 | |
73401833 | 1512 | static void |
04f04b72 | 1513 | print_slot (rtx insn) |
73401833 | 1514 | { |
4bf029b0 | 1515 | final_scan_insn (XVECEXP (insn, 0, 1), asm_out_file, optimize, 1, NULL); |
47c009e5 | 1516 | |
73401833 | 1517 | INSN_DELETED_P (XVECEXP (insn, 0, 1)) = 1; |
1518 | } | |
97595bfd | 1519 | |
b7dbbdb2 | 1520 | const char * |
04f04b72 | 1521 | output_far_jump (rtx insn, rtx op) |
97595bfd | 1522 | { |
8ded0752 | 1523 | struct { rtx lab, reg, op; } this; |
e51fa923 | 1524 | rtx braf_base_lab = NULL_RTX; |
b7dbbdb2 | 1525 | const char *jump; |
8ded0752 | 1526 | int far; |
47fc0706 | 1527 | int offset = branch_dest (insn) - INSN_ADDRESSES (INSN_UID (insn)); |
2a6493a5 | 1528 | rtx prev; |
97595bfd | 1529 | |
8ded0752 | 1530 | this.lab = gen_label_rtx (); |
47c009e5 | 1531 | |
15494e01 | 1532 | if (TARGET_SH2 |
1533 | && offset >= -32764 | |
1534 | && offset - get_attr_length (insn) <= 32766) | |
8ded0752 | 1535 | { |
1536 | far = 0; | |
d1665c4c | 1537 | jump = "mov.w %O0,%1; braf %1"; |
8ded0752 | 1538 | } |
1539 | else | |
1540 | { | |
1541 | far = 1; | |
0abf894c | 1542 | if (flag_pic) |
d1665c4c | 1543 | { |
1544 | if (TARGET_SH2) | |
1545 | jump = "mov.l %O0,%1; braf %1"; | |
1546 | else | |
1547 | jump = "mov.l r0,@-r15; mova %O0,r0; mov.l @r0,%1; add r0,%1; mov.l @r15+,r0; jmp @%1"; | |
1548 | } | |
0abf894c | 1549 | else |
d1665c4c | 1550 | jump = "mov.l %O0,%1; jmp @%1"; |
8ded0752 | 1551 | } |
1552 | /* If we have a scratch register available, use it. */ | |
2a6493a5 | 1553 | if (GET_CODE ((prev = prev_nonnote_insn (insn))) == INSN |
1554 | && INSN_CODE (prev) == CODE_FOR_indirect_jump_scratch) | |
8ded0752 | 1555 | { |
2a6493a5 | 1556 | this.reg = SET_DEST (XVECEXP (PATTERN (prev), 0, 0)); |
d1665c4c | 1557 | if (REGNO (this.reg) == R0_REG && flag_pic && ! TARGET_SH2) |
1558 | jump = "mov.l r1,@-r15; mova %O0,r0; mov.l @r0,r1; add r1,r0; mov.l @r15+,r1; jmp @%1"; | |
8ded0752 | 1559 | output_asm_insn (jump, &this.lab); |
1560 | if (dbr_sequence_length ()) | |
1561 | print_slot (final_sequence); | |
1562 | else | |
1563 | output_asm_insn ("nop", 0); | |
1564 | } | |
1565 | else | |
1566 | { | |
1567 | /* Output the delay slot insn first if any. */ | |
1568 | if (dbr_sequence_length ()) | |
1569 | print_slot (final_sequence); | |
1570 | ||
7014838c | 1571 | this.reg = gen_rtx_REG (SImode, 13); |
87e19636 | 1572 | /* We must keep the stack aligned to 8-byte boundaries on SH5. |
1573 | Fortunately, MACL is fixed and call-clobbered, and we never | |
1574 | need its value across jumps, so save r13 in it instead of in | |
1575 | the stack. */ | |
1576 | if (TARGET_SH5) | |
1577 | output_asm_insn ("lds r13, macl", 0); | |
1578 | else | |
1579 | output_asm_insn ("mov.l r13,@-r15", 0); | |
8ded0752 | 1580 | output_asm_insn (jump, &this.lab); |
87e19636 | 1581 | if (TARGET_SH5) |
1582 | output_asm_insn ("sts macl, r13", 0); | |
1583 | else | |
1584 | output_asm_insn ("mov.l @r15+,r13", 0); | |
8ded0752 | 1585 | } |
d1665c4c | 1586 | if (far && flag_pic && TARGET_SH2) |
1587 | { | |
1588 | braf_base_lab = gen_label_rtx (); | |
805e22b2 | 1589 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
d1665c4c | 1590 | CODE_LABEL_NUMBER (braf_base_lab)); |
1591 | } | |
8ded0752 | 1592 | if (far) |
1593 | output_asm_insn (".align 2", 0); | |
805e22b2 | 1594 | (*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (this.lab)); |
8ded0752 | 1595 | this.op = op; |
0abf894c | 1596 | if (far && flag_pic) |
d1665c4c | 1597 | { |
1598 | if (TARGET_SH2) | |
1599 | this.lab = braf_base_lab; | |
1600 | output_asm_insn (".long %O2-%O0", &this.lab); | |
1601 | } | |
0abf894c | 1602 | else |
1603 | output_asm_insn (far ? ".long %O2" : ".word %O2-%O0", &this.lab); | |
97595bfd | 1604 | return ""; |
1605 | } | |
7208c779 | 1606 | |
73401833 | 1607 | /* Local label counter, used for constants in the pool and inside |
1608 | pattern branches. */ | |
1609 | ||
1610 | static int lf = 100; | |
1611 | ||
1612 | /* Output code for ordinary branches. */ | |
1613 | ||
b7dbbdb2 | 1614 | const char * |
04f04b72 | 1615 | output_branch (int logic, rtx insn, rtx *operands) |
7208c779 | 1616 | { |
fb2fe7e1 | 1617 | switch (get_attr_length (insn)) |
7208c779 | 1618 | { |
9eaab178 | 1619 | case 6: |
1620 | /* This can happen if filling the delay slot has caused a forward | |
1621 | branch to exceed its range (we could reverse it, but only | |
1622 | when we know we won't overextend other branches; this should | |
1623 | best be handled by relaxation). | |
1624 | It can also happen when other condbranches hoist delay slot insn | |
8ded0752 | 1625 | from their destination, thus leading to code size increase. |
1626 | But the branch will still be in the range -4092..+4098 bytes. */ | |
7208c779 | 1627 | |
9eaab178 | 1628 | if (! TARGET_RELAX) |
8ded0752 | 1629 | { |
9eaab178 | 1630 | int label = lf++; |
1631 | /* The call to print_slot will clobber the operands. */ | |
1632 | rtx op0 = operands[0]; | |
9e7454d0 | 1633 | |
9eaab178 | 1634 | /* If the instruction in the delay slot is annulled (true), then |
1635 | there is no delay slot where we can put it now. The only safe | |
1636 | place for it is after the label. final will do that by default. */ | |
9e7454d0 | 1637 | |
9eaab178 | 1638 | if (final_sequence |
62111d8e | 1639 | && ! INSN_ANNULLED_BRANCH_P (XVECEXP (final_sequence, 0, 0)) |
1640 | && get_attr_length (XVECEXP (final_sequence, 0, 1))) | |
9eaab178 | 1641 | { |
fb2fe7e1 | 1642 | asm_fprintf (asm_out_file, "\tb%s%ss\t%LLF%d\n", logic ? "f" : "t", |
9eaab178 | 1643 | ASSEMBLER_DIALECT ? "/" : ".", label); |
1644 | print_slot (final_sequence); | |
1645 | } | |
1646 | else | |
fb2fe7e1 | 1647 | asm_fprintf (asm_out_file, "\tb%s\t%LLF%d\n", logic ? "f" : "t", label); |
9e7454d0 | 1648 | |
fb2fe7e1 | 1649 | output_asm_insn ("bra\t%l0", &op0); |
1650 | fprintf (asm_out_file, "\tnop\n"); | |
d767e27e | 1651 | (*targetm.asm_out.internal_label) (asm_out_file, "LF", label); |
9e7454d0 | 1652 | |
9eaab178 | 1653 | return ""; |
8ded0752 | 1654 | } |
9eaab178 | 1655 | /* When relaxing, handle this like a short branch. The linker |
1656 | will fix it up if it still doesn't fit after relaxation. */ | |
1657 | case 2: | |
1658 | return logic ? "bt%.\t%l0" : "bf%.\t%l0"; | |
87ed74ef | 1659 | |
1660 | /* These are for SH2e, in which we have to account for the | |
1661 | extra nop because of the hardware bug in annulled branches. */ | |
1662 | case 8: | |
1663 | if (! TARGET_RELAX) | |
1664 | { | |
1665 | int label = lf++; | |
1666 | ||
1a6a7a27 | 1667 | gcc_assert (!final_sequence |
1668 | || !(INSN_ANNULLED_BRANCH_P | |
1669 | (XVECEXP (final_sequence, 0, 0)))); | |
87ed74ef | 1670 | asm_fprintf (asm_out_file, "b%s%ss\t%LLF%d\n", |
1671 | logic ? "f" : "t", | |
1672 | ASSEMBLER_DIALECT ? "/" : ".", label); | |
1673 | fprintf (asm_out_file, "\tnop\n"); | |
1674 | output_asm_insn ("bra\t%l0", operands); | |
1675 | fprintf (asm_out_file, "\tnop\n"); | |
1676 | (*targetm.asm_out.internal_label) (asm_out_file, "LF", label); | |
1677 | ||
1678 | return ""; | |
1679 | } | |
1680 | /* When relaxing, fall through. */ | |
1681 | case 4: | |
1682 | { | |
1683 | char buffer[10]; | |
9e7454d0 | 1684 | |
87ed74ef | 1685 | sprintf (buffer, "b%s%ss\t%%l0", |
1686 | logic ? "t" : "f", | |
1687 | ASSEMBLER_DIALECT ? "/" : "."); | |
1688 | output_asm_insn (buffer, &operands[0]); | |
1689 | return "nop"; | |
1690 | } | |
1691 | ||
9eaab178 | 1692 | default: |
fb2fe7e1 | 1693 | /* There should be no longer branches now - that would |
1694 | indicate that something has destroyed the branches set | |
1695 | up in machine_dependent_reorg. */ | |
1a6a7a27 | 1696 | gcc_unreachable (); |
8ded0752 | 1697 | } |
8ded0752 | 1698 | } |
7208c779 | 1699 | |
b7dbbdb2 | 1700 | const char * |
04f04b72 | 1701 | output_branchy_insn (enum rtx_code code, const char *template, |
1702 | rtx insn, rtx *operands) | |
8ded0752 | 1703 | { |
1704 | rtx next_insn = NEXT_INSN (insn); | |
3f2613e4 | 1705 | |
8ded0752 | 1706 | if (next_insn && GET_CODE (next_insn) == JUMP_INSN && condjump_p (next_insn)) |
1707 | { | |
1708 | rtx src = SET_SRC (PATTERN (next_insn)); | |
1709 | if (GET_CODE (src) == IF_THEN_ELSE && GET_CODE (XEXP (src, 0)) != code) | |
1710 | { | |
1711 | /* Following branch not taken */ | |
1712 | operands[9] = gen_label_rtx (); | |
1713 | emit_label_after (operands[9], next_insn); | |
47fc0706 | 1714 | INSN_ADDRESSES_NEW (operands[9], |
1715 | INSN_ADDRESSES (INSN_UID (next_insn)) | |
1716 | + get_attr_length (next_insn)); | |
8ded0752 | 1717 | return template; |
1718 | } | |
1719 | else | |
1720 | { | |
9eaab178 | 1721 | int offset = (branch_dest (next_insn) |
47fc0706 | 1722 | - INSN_ADDRESSES (INSN_UID (next_insn)) + 4); |
9eaab178 | 1723 | if (offset >= -252 && offset <= 258) |
8ded0752 | 1724 | { |
1725 | if (GET_CODE (src) == IF_THEN_ELSE) | |
1726 | /* branch_true */ | |
1727 | src = XEXP (src, 1); | |
1728 | operands[9] = src; | |
1729 | return template; | |
1730 | } | |
1731 | } | |
7208c779 | 1732 | } |
8ded0752 | 1733 | operands[9] = gen_label_rtx (); |
1734 | emit_label_after (operands[9], insn); | |
47fc0706 | 1735 | INSN_ADDRESSES_NEW (operands[9], |
1736 | INSN_ADDRESSES (INSN_UID (insn)) | |
1737 | + get_attr_length (insn)); | |
8ded0752 | 1738 | return template; |
1739 | } | |
e6c94d0d | 1740 | |
b7dbbdb2 | 1741 | const char * |
04f04b72 | 1742 | output_ieee_ccmpeq (rtx insn, rtx *operands) |
8ded0752 | 1743 | { |
aacd46a5 | 1744 | return output_branchy_insn (NE, "bt\t%l9\n\tfcmp/eq\t%1,%0", |
1745 | insn, operands); | |
7208c779 | 1746 | } |
2e3d4844 | 1747 | \f |
92c473b8 | 1748 | /* Output the start of the assembler file. */ |
73401833 | 1749 | |
92c473b8 | 1750 | static void |
04f04b72 | 1751 | sh_file_start (void) |
7208c779 | 1752 | { |
92c473b8 | 1753 | default_file_start (); |
1754 | ||
db281f5a | 1755 | #ifdef SYMBIAN |
1756 | /* Declare the .directive section before it is used. */ | |
1757 | fputs ("\t.section .directive, \"SM\", @progbits, 1\n", asm_out_file); | |
1758 | fputs ("\t.asciz \"#<SYMEDIT>#\\n\"\n", asm_out_file); | |
1759 | #endif | |
9e7454d0 | 1760 | |
92c473b8 | 1761 | if (TARGET_ELF) |
1762 | /* We need to show the text section with the proper | |
1763 | attributes as in TEXT_SECTION_ASM_OP, before dwarf2out | |
9fd95d94 | 1764 | emits it without attributes in TEXT_SECTION_ASM_OP, else GAS |
92c473b8 | 1765 | will complain. We can teach GAS specifically about the |
1766 | default attributes for our choice of text section, but | |
1767 | then we would have to change GAS again if/when we change | |
1768 | the text section name. */ | |
1769 | fprintf (asm_out_file, "%s\n", TEXT_SECTION_ASM_OP); | |
1770 | else | |
1771 | /* Switch to the data section so that the coffsem symbol | |
1772 | isn't in the text section. */ | |
1773 | data_section (); | |
47c009e5 | 1774 | |
73401833 | 1775 | if (TARGET_LITTLE_ENDIAN) |
92c473b8 | 1776 | fputs ("\t.little\n", asm_out_file); |
87e19636 | 1777 | |
92c473b8 | 1778 | if (!TARGET_ELF) |
1779 | { | |
1780 | if (TARGET_SHCOMPACT) | |
1781 | fputs ("\t.mode\tSHcompact\n", asm_out_file); | |
1782 | else if (TARGET_SHMEDIA) | |
1783 | fprintf (asm_out_file, "\t.mode\tSHmedia\n\t.abi\t%i\n", | |
1784 | TARGET_SHMEDIA64 ? 64 : 32); | |
1785 | } | |
7208c779 | 1786 | } |
97595bfd | 1787 | \f |
02359f85 | 1788 | /* Check if PAT includes UNSPEC_CALLER unspec pattern. */ |
1789 | ||
1790 | static bool | |
04f04b72 | 1791 | unspec_caller_rtx_p (rtx pat) |
02359f85 | 1792 | { |
1793 | switch (GET_CODE (pat)) | |
1794 | { | |
1795 | case CONST: | |
1796 | return unspec_caller_rtx_p (XEXP (pat, 0)); | |
1797 | case PLUS: | |
1798 | case MINUS: | |
1799 | if (unspec_caller_rtx_p (XEXP (pat, 0))) | |
1800 | return true; | |
1801 | return unspec_caller_rtx_p (XEXP (pat, 1)); | |
1802 | case UNSPEC: | |
1803 | if (XINT (pat, 1) == UNSPEC_CALLER) | |
1804 | return true; | |
1805 | default: | |
1806 | break; | |
1807 | } | |
1808 | ||
1809 | return false; | |
1810 | } | |
1811 | ||
1812 | /* Indicate that INSN cannot be duplicated. This is true for insn | |
77aa6362 | 1813 | that generates a unique label. */ |
02359f85 | 1814 | |
1815 | static bool | |
04f04b72 | 1816 | sh_cannot_copy_insn_p (rtx insn) |
02359f85 | 1817 | { |
1818 | rtx pat; | |
1819 | ||
1820 | if (!reload_completed || !flag_pic) | |
1821 | return false; | |
1822 | ||
1823 | if (GET_CODE (insn) != INSN) | |
1824 | return false; | |
1825 | if (asm_noperands (insn) >= 0) | |
1826 | return false; | |
1827 | ||
1828 | pat = PATTERN (insn); | |
1829 | if (GET_CODE (pat) != SET) | |
1830 | return false; | |
1831 | pat = SET_SRC (pat); | |
1832 | ||
1833 | if (unspec_caller_rtx_p (pat)) | |
1834 | return true; | |
1835 | ||
1836 | return false; | |
1837 | } | |
1838 | \f | |
1504eb7e | 1839 | /* Actual number of instructions used to make a shift by N. */ |
e7c691ac | 1840 | static const char ashiftrt_insns[] = |
1504eb7e | 1841 | { 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}; |
1842 | ||
1843 | /* Left shift and logical right shift are the same. */ | |
e7c691ac | 1844 | static const char shift_insns[] = |
1504eb7e | 1845 | { 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}; |
73401833 | 1846 | |
1504eb7e | 1847 | /* Individual shift amounts needed to get the above length sequences. |
1848 | One bit right shifts clobber the T bit, so when possible, put one bit | |
1849 | shifts in the middle of the sequence, so the ends are eligible for | |
1850 | branch delay slots. */ | |
c8834c5f | 1851 | static const short shift_amounts[32][5] = { |
1504eb7e | 1852 | {0}, {1}, {2}, {2, 1}, |
1853 | {2, 2}, {2, 1, 2}, {2, 2, 2}, {2, 2, 1, 2}, | |
1854 | {8}, {8, 1}, {8, 2}, {8, 1, 2}, | |
1855 | {8, 2, 2}, {8, 2, 1, 2}, {8, -2, 8}, {8, -1, 8}, | |
1856 | {16}, {16, 1}, {16, 2}, {16, 1, 2}, | |
1857 | {16, 2, 2}, {16, 2, 1, 2}, {16, -2, 8}, {16, -1, 8}, | |
1858 | {16, 8}, {16, 1, 8}, {16, 8, 2}, {16, 8, 1, 2}, | |
ad7aadac | 1859 | {16, 8, 2, 2}, {16, -1, -2, 16}, {16, -2, 16}, {16, -1, 16}}; |
1504eb7e | 1860 | |
2ce8df5a | 1861 | /* Likewise, but for shift amounts < 16, up to three highmost bits |
1862 | might be clobbered. This is typically used when combined with some | |
1863 | kind of sign or zero extension. */ | |
9e7454d0 | 1864 | |
e7c691ac | 1865 | static const char ext_shift_insns[] = |
2ce8df5a | 1866 | { 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}; |
1867 | ||
e7c691ac | 1868 | static const short ext_shift_amounts[32][4] = { |
2ce8df5a | 1869 | {0}, {1}, {2}, {2, 1}, |
1870 | {2, 2}, {2, 1, 2}, {8, -2}, {8, -1}, | |
1871 | {8}, {8, 1}, {8, 2}, {8, 1, 2}, | |
1872 | {8, 2, 2}, {16, -2, -1}, {16, -2}, {16, -1}, | |
1873 | {16}, {16, 1}, {16, 2}, {16, 1, 2}, | |
1874 | {16, 2, 2}, {16, 2, 1, 2}, {16, -2, 8}, {16, -1, 8}, | |
1875 | {16, 8}, {16, 1, 8}, {16, 8, 2}, {16, 8, 1, 2}, | |
1876 | {16, 8, 2, 2}, {16, -1, -2, 16}, {16, -2, 16}, {16, -1, 16}}; | |
1877 | ||
69b02875 | 1878 | /* Assuming we have a value that has been sign-extended by at least one bit, |
1879 | can we use the ext_shift_amounts with the last shift turned to an arithmetic shift | |
1880 | to shift it by N without data loss, and quicker than by other means? */ | |
1881 | #define EXT_SHIFT_SIGNED(n) (((n) | 8) == 15) | |
1882 | ||
1504eb7e | 1883 | /* This is used in length attributes in sh.md to help compute the length |
1884 | of arbitrary constant shift instructions. */ | |
7208c779 | 1885 | |
1504eb7e | 1886 | int |
04f04b72 | 1887 | shift_insns_rtx (rtx insn) |
1504eb7e | 1888 | { |
1889 | rtx set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
1890 | int shift_count = INTVAL (XEXP (set_src, 1)); | |
1891 | enum rtx_code shift_code = GET_CODE (set_src); | |
3dfc1807 | 1892 | |
1504eb7e | 1893 | switch (shift_code) |
1894 | { | |
1895 | case ASHIFTRT: | |
1896 | return ashiftrt_insns[shift_count]; | |
1897 | case LSHIFTRT: | |
1898 | case ASHIFT: | |
1899 | return shift_insns[shift_count]; | |
1900 | default: | |
1a6a7a27 | 1901 | gcc_unreachable (); |
1504eb7e | 1902 | } |
1903 | } | |
1904 | ||
1504eb7e | 1905 | /* Return the cost of a shift. */ |
7208c779 | 1906 | |
fab7adbf | 1907 | static inline int |
04f04b72 | 1908 | shiftcosts (rtx x) |
7208c779 | 1909 | { |
d6fe3e0d | 1910 | int value; |
73401833 | 1911 | |
87e19636 | 1912 | if (TARGET_SHMEDIA) |
1913 | return 1; | |
1914 | ||
41f4d177 | 1915 | if (GET_MODE_SIZE (GET_MODE (x)) > UNITS_PER_WORD) |
1916 | { | |
1917 | if (GET_MODE (x) == DImode | |
1918 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
1919 | && INTVAL (XEXP (x, 1)) == 1) | |
1920 | return 2; | |
1921 | ||
1922 | /* Everything else is invalid, because there is no pattern for it. */ | |
1923 | return 10000; | |
1924 | } | |
1504eb7e | 1925 | /* If shift by a non constant, then this will be expensive. */ |
73401833 | 1926 | if (GET_CODE (XEXP (x, 1)) != CONST_INT) |
8ded0752 | 1927 | return SH_DYNAMIC_SHIFT_COST; |
7208c779 | 1928 | |
d6fe3e0d | 1929 | value = INTVAL (XEXP (x, 1)); |
1930 | ||
1504eb7e | 1931 | /* Otherwise, return the true cost in instructions. */ |
73401833 | 1932 | if (GET_CODE (x) == ASHIFTRT) |
79f0f2fe | 1933 | { |
1934 | int cost = ashiftrt_insns[value]; | |
1935 | /* If SH3, then we put the constant in a reg and use shad. */ | |
8ded0752 | 1936 | if (cost > 1 + SH_DYNAMIC_SHIFT_COST) |
1937 | cost = 1 + SH_DYNAMIC_SHIFT_COST; | |
79f0f2fe | 1938 | return cost; |
1939 | } | |
73401833 | 1940 | else |
1941 | return shift_insns[value]; | |
97595bfd | 1942 | } |
47c009e5 | 1943 | |
73401833 | 1944 | /* Return the cost of an AND operation. */ |
1945 | ||
fab7adbf | 1946 | static inline int |
04f04b72 | 1947 | andcosts (rtx x) |
97595bfd | 1948 | { |
1949 | int i; | |
73401833 | 1950 | |
f22b6811 | 1951 | /* Anding with a register is a single cycle and instruction. */ |
73401833 | 1952 | if (GET_CODE (XEXP (x, 1)) != CONST_INT) |
f22b6811 | 1953 | return 1; |
73401833 | 1954 | |
1955 | i = INTVAL (XEXP (x, 1)); | |
87e19636 | 1956 | |
1957 | if (TARGET_SHMEDIA) | |
1958 | { | |
1959 | if ((GET_CODE (XEXP (x, 1)) == CONST_INT | |
af2c1324 | 1960 | && CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1)))) |
1961 | || EXTRA_CONSTRAINT_C16 (XEXP (x, 1))) | |
87e19636 | 1962 | return 1; |
1963 | else | |
1964 | return 2; | |
1965 | } | |
1966 | ||
f22b6811 | 1967 | /* These constants are single cycle extu.[bw] instructions. */ |
97595bfd | 1968 | if (i == 0xff || i == 0xffff) |
f22b6811 | 1969 | return 1; |
af2c1324 | 1970 | /* Constants that can be used in an and immediate instruction in a single |
f22b6811 | 1971 | cycle, but this requires r0, so make it a little more expensive. */ |
af2c1324 | 1972 | if (CONST_OK_FOR_K08 (i)) |
97595bfd | 1973 | return 2; |
f22b6811 | 1974 | /* Constants that can be loaded with a mov immediate and an and. |
1975 | This case is probably unnecessary. */ | |
af2c1324 | 1976 | if (CONST_OK_FOR_I08 (i)) |
f22b6811 | 1977 | return 2; |
1978 | /* Any other constants requires a 2 cycle pc-relative load plus an and. | |
1979 | This case is probably unnecessary. */ | |
1980 | return 3; | |
97595bfd | 1981 | } |
f3d93547 | 1982 | |
d27eb4b1 | 1983 | /* Return the cost of an addition or a subtraction. */ |
1984 | ||
fab7adbf | 1985 | static inline int |
04f04b72 | 1986 | addsubcosts (rtx x) |
d27eb4b1 | 1987 | { |
1988 | /* Adding a register is a single cycle insn. */ | |
0d93aeb0 | 1989 | if (GET_CODE (XEXP (x, 1)) == REG |
1990 | || GET_CODE (XEXP (x, 1)) == SUBREG) | |
d27eb4b1 | 1991 | return 1; |
1992 | ||
1993 | /* Likewise for small constants. */ | |
d6fe3e0d | 1994 | if (GET_CODE (XEXP (x, 1)) == CONST_INT |
87e19636 | 1995 | && CONST_OK_FOR_ADD (INTVAL (XEXP (x, 1)))) |
d27eb4b1 | 1996 | return 1; |
1997 | ||
87e19636 | 1998 | if (TARGET_SHMEDIA) |
1999 | switch (GET_CODE (XEXP (x, 1))) | |
2000 | { | |
2001 | case CONST: | |
2002 | case LABEL_REF: | |
2003 | case SYMBOL_REF: | |
2004 | return TARGET_SHMEDIA64 ? 5 : 3; | |
2005 | ||
2006 | case CONST_INT: | |
af2c1324 | 2007 | if (CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1)))) |
87e19636 | 2008 | return 2; |
af2c1324 | 2009 | else if (CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1)) >> 16)) |
87e19636 | 2010 | return 3; |
af2c1324 | 2011 | else if (CONST_OK_FOR_I16 ((INTVAL (XEXP (x, 1)) >> 16) >> 16)) |
87e19636 | 2012 | return 4; |
2013 | ||
2014 | /* Fall through. */ | |
2015 | default: | |
d767e27e | 2016 | return 5; |
87e19636 | 2017 | } |
2018 | ||
d27eb4b1 | 2019 | /* Any other constant requires a 2 cycle pc-relative load plus an |
2020 | addition. */ | |
2021 | return 3; | |
2022 | } | |
2023 | ||
1504eb7e | 2024 | /* Return the cost of a multiply. */ |
fab7adbf | 2025 | static inline int |
04f04b72 | 2026 | multcosts (rtx x ATTRIBUTE_UNUSED) |
97595bfd | 2027 | { |
4906bcbb | 2028 | if (sh_multcost >= 0) |
2029 | return sh_multcost; | |
87e19636 | 2030 | if (TARGET_SHMEDIA) |
59312820 | 2031 | /* ??? We have a mul insn, but it has a latency of three, and doesn't |
2032 | accept constants. Ideally, we would use a cost of one or two and | |
2033 | add the cost of the operand, but disregard the latter when inside loops | |
2034 | and loop invariant code motion is still to follow. | |
2035 | Using a multiply first and splitting it later if it's a loss | |
2036 | doesn't work because of different sign / zero extension semantics | |
2037 | of multiplies vs. shifts. */ | |
2038 | return TARGET_SMALLCODE ? 2 : 3; | |
87e19636 | 2039 | |
97595bfd | 2040 | if (TARGET_SH2) |
f3d93547 | 2041 | { |
2042 | /* We have a mul insn, so we can never take more than the mul and the | |
8cd86814 | 2043 | read of the mac reg, but count more because of the latency and extra |
1504eb7e | 2044 | reg usage. */ |
f3d93547 | 2045 | if (TARGET_SMALLCODE) |
2e3d4844 | 2046 | return 2; |
8cd86814 | 2047 | return 3; |
f3d93547 | 2048 | } |
2049 | ||
8cd86814 | 2050 | /* If we're aiming at small code, then just count the number of |
1504eb7e | 2051 | insns in a multiply call sequence. */ |
2e3d4844 | 2052 | if (TARGET_SMALLCODE) |
73401833 | 2053 | return 5; |
f3d93547 | 2054 | |
1504eb7e | 2055 | /* Otherwise count all the insns in the routine we'd be calling too. */ |
f3d93547 | 2056 | return 20; |
97595bfd | 2057 | } |
47c009e5 | 2058 | |
fab7adbf | 2059 | /* Compute a (partial) cost for rtx X. Return true if the complete |
2060 | cost has been computed, and false if subexpressions should be | |
2061 | scanned. In either case, *TOTAL contains the cost result. */ | |
2062 | ||
2063 | static bool | |
04f04b72 | 2064 | sh_rtx_costs (rtx x, int code, int outer_code, int *total) |
fab7adbf | 2065 | { |
2066 | switch (code) | |
2067 | { | |
2068 | case CONST_INT: | |
2069 | if (TARGET_SHMEDIA) | |
2070 | { | |
2071 | if (INTVAL (x) == 0) | |
2072 | *total = 0; | |
2073 | else if (outer_code == AND && and_operand ((x), DImode)) | |
2074 | *total = 0; | |
2075 | else if ((outer_code == IOR || outer_code == XOR | |
2076 | || outer_code == PLUS) | |
af2c1324 | 2077 | && CONST_OK_FOR_I10 (INTVAL (x))) |
fab7adbf | 2078 | *total = 0; |
af2c1324 | 2079 | else if (CONST_OK_FOR_I16 (INTVAL (x))) |
fab7adbf | 2080 | *total = COSTS_N_INSNS (outer_code != SET); |
af2c1324 | 2081 | else if (CONST_OK_FOR_I16 (INTVAL (x) >> 16)) |
59312820 | 2082 | *total = COSTS_N_INSNS ((outer_code != SET) + 1); |
af2c1324 | 2083 | else if (CONST_OK_FOR_I16 ((INTVAL (x) >> 16) >> 16)) |
fab7adbf | 2084 | *total = COSTS_N_INSNS (3); |
2085 | else | |
2086 | *total = COSTS_N_INSNS (4); | |
2087 | return true; | |
2088 | } | |
af2c1324 | 2089 | if (CONST_OK_FOR_I08 (INTVAL (x))) |
fab7adbf | 2090 | *total = 0; |
2091 | else if ((outer_code == AND || outer_code == IOR || outer_code == XOR) | |
af2c1324 | 2092 | && CONST_OK_FOR_K08 (INTVAL (x))) |
fab7adbf | 2093 | *total = 1; |
2094 | else | |
2095 | *total = 8; | |
2096 | return true; | |
2097 | ||
2098 | case CONST: | |
2099 | case LABEL_REF: | |
2100 | case SYMBOL_REF: | |
2101 | if (TARGET_SHMEDIA64) | |
2102 | *total = COSTS_N_INSNS (4); | |
2103 | else if (TARGET_SHMEDIA32) | |
2104 | *total = COSTS_N_INSNS (2); | |
2105 | else | |
2106 | *total = 5; | |
2107 | return true; | |
2108 | ||
2109 | case CONST_DOUBLE: | |
2110 | if (TARGET_SHMEDIA) | |
2111 | *total = COSTS_N_INSNS (4); | |
2112 | else | |
2113 | *total = 10; | |
2114 | return true; | |
59312820 | 2115 | case CONST_VECTOR: |
2116 | if (x == CONST0_RTX (GET_MODE (x))) | |
2117 | *total = 0; | |
2118 | else if (sh_1el_vec (x, VOIDmode)) | |
2119 | *total = outer_code != SET; | |
2120 | if (sh_rep_vec (x, VOIDmode)) | |
2121 | *total = ((GET_MODE_UNIT_SIZE (GET_MODE (x)) + 3) / 4 | |
2122 | + (outer_code != SET)); | |
2123 | *total = COSTS_N_INSNS (3) + (outer_code != SET); | |
2124 | return true; | |
fab7adbf | 2125 | |
2126 | case PLUS: | |
59312820 | 2127 | case MINUS: |
fab7adbf | 2128 | *total = COSTS_N_INSNS (addsubcosts (x)); |
2129 | return true; | |
2130 | ||
2131 | case AND: | |
2132 | *total = COSTS_N_INSNS (andcosts (x)); | |
2133 | return true; | |
2134 | ||
2135 | case MULT: | |
2136 | *total = COSTS_N_INSNS (multcosts (x)); | |
2137 | return true; | |
2138 | ||
2139 | case ASHIFT: | |
2140 | case ASHIFTRT: | |
2141 | case LSHIFTRT: | |
2142 | *total = COSTS_N_INSNS (shiftcosts (x)); | |
2143 | return true; | |
2144 | ||
2145 | case DIV: | |
2146 | case UDIV: | |
2147 | case MOD: | |
2148 | case UMOD: | |
2149 | *total = COSTS_N_INSNS (20); | |
2150 | return true; | |
2151 | ||
59312820 | 2152 | case PARALLEL: |
2153 | if (sh_1el_vec (x, VOIDmode)) | |
2154 | *total = outer_code != SET; | |
2155 | if (sh_rep_vec (x, VOIDmode)) | |
2156 | *total = ((GET_MODE_UNIT_SIZE (GET_MODE (x)) + 3) / 4 | |
2157 | + (outer_code != SET)); | |
2158 | *total = COSTS_N_INSNS (3) + (outer_code != SET); | |
2159 | return true; | |
2160 | ||
fab7adbf | 2161 | case FLOAT: |
2162 | case FIX: | |
2163 | *total = 100; | |
2164 | return true; | |
2165 | ||
2166 | default: | |
2167 | return false; | |
2168 | } | |
2169 | } | |
2170 | ||
ec0457a8 | 2171 | /* Compute the cost of an address. For the SH, all valid addresses are |
2172 | the same cost. Use a slightly higher cost for reg + reg addressing, | |
2173 | since it increases pressure on r0. */ | |
2174 | ||
2175 | static int | |
04f04b72 | 2176 | sh_address_cost (rtx X) |
ec0457a8 | 2177 | { |
2178 | return (GET_CODE (X) == PLUS | |
2179 | && ! CONSTANT_P (XEXP (X, 1)) | |
2180 | && ! TARGET_SHMEDIA ? 1 : 0); | |
2181 | } | |
fab7adbf | 2182 | |
1504eb7e | 2183 | /* Code to expand a shift. */ |
47c009e5 | 2184 | |
97595bfd | 2185 | void |
04f04b72 | 2186 | gen_ashift (int type, int n, rtx reg) |
97595bfd | 2187 | { |
1504eb7e | 2188 | /* Negative values here come from the shift_amounts array. */ |
2189 | if (n < 0) | |
2190 | { | |
2191 | if (type == ASHIFT) | |
2192 | type = LSHIFTRT; | |
2193 | else | |
2194 | type = ASHIFT; | |
2195 | n = -n; | |
2196 | } | |
2197 | ||
97595bfd | 2198 | switch (type) |
7208c779 | 2199 | { |
97595bfd | 2200 | case ASHIFTRT: |
2201 | emit_insn (gen_ashrsi3_k (reg, reg, GEN_INT (n))); | |
2202 | break; | |
2203 | case LSHIFTRT: | |
1504eb7e | 2204 | if (n == 1) |
2205 | emit_insn (gen_lshrsi3_m (reg, reg, GEN_INT (n))); | |
2206 | else | |
2207 | emit_insn (gen_lshrsi3_k (reg, reg, GEN_INT (n))); | |
97595bfd | 2208 | break; |
2209 | case ASHIFT: | |
184d8105 | 2210 | emit_insn (gen_ashlsi3_std (reg, reg, GEN_INT (n))); |
97595bfd | 2211 | break; |
7208c779 | 2212 | } |
7208c779 | 2213 | } |
7208c779 | 2214 | |
2ce8df5a | 2215 | /* Same for HImode */ |
2216 | ||
2217 | void | |
04f04b72 | 2218 | gen_ashift_hi (int type, int n, rtx reg) |
2ce8df5a | 2219 | { |
2220 | /* Negative values here come from the shift_amounts array. */ | |
2221 | if (n < 0) | |
2222 | { | |
2223 | if (type == ASHIFT) | |
2224 | type = LSHIFTRT; | |
2225 | else | |
2226 | type = ASHIFT; | |
2227 | n = -n; | |
2228 | } | |
2229 | ||
2230 | switch (type) | |
2231 | { | |
2232 | case ASHIFTRT: | |
2ce8df5a | 2233 | case LSHIFTRT: |
4b9cbc29 | 2234 | /* We don't have HImode right shift operations because using the |
2235 | ordinary 32 bit shift instructions for that doesn't generate proper | |
2236 | zero/sign extension. | |
2237 | gen_ashift_hi is only called in contexts where we know that the | |
2238 | sign extension works out correctly. */ | |
7d175685 | 2239 | { |
701e46d0 | 2240 | int offset = 0; |
7d175685 | 2241 | if (GET_CODE (reg) == SUBREG) |
2242 | { | |
701e46d0 | 2243 | offset = SUBREG_BYTE (reg); |
7d175685 | 2244 | reg = SUBREG_REG (reg); |
2245 | } | |
701e46d0 | 2246 | gen_ashift (type, n, gen_rtx_SUBREG (SImode, reg, offset)); |
7d175685 | 2247 | break; |
2248 | } | |
2ce8df5a | 2249 | case ASHIFT: |
2250 | emit_insn (gen_ashlhi3_k (reg, reg, GEN_INT (n))); | |
2251 | break; | |
2252 | } | |
2253 | } | |
2254 | ||
73401833 | 2255 | /* Output RTL to split a constant shift into its component SH constant |
2256 | shift instructions. */ | |
9e7454d0 | 2257 | |
b7dbbdb2 | 2258 | void |
04f04b72 | 2259 | gen_shifty_op (int code, rtx *operands) |
7208c779 | 2260 | { |
1504eb7e | 2261 | int value = INTVAL (operands[2]); |
73401833 | 2262 | int max, i; |
3dfc1807 | 2263 | |
01515da1 | 2264 | /* Truncate the shift count in case it is out of bounds. */ |
2265 | value = value & 0x1f; | |
9e7454d0 | 2266 | |
73401833 | 2267 | if (value == 31) |
1504eb7e | 2268 | { |
73401833 | 2269 | if (code == LSHIFTRT) |
97595bfd | 2270 | { |
73401833 | 2271 | emit_insn (gen_rotlsi3_1 (operands[0], operands[0])); |
2272 | emit_insn (gen_movt (operands[0])); | |
2273 | return; | |
1504eb7e | 2274 | } |
73401833 | 2275 | else if (code == ASHIFT) |
1504eb7e | 2276 | { |
73401833 | 2277 | /* There is a two instruction sequence for 31 bit left shifts, |
2278 | but it requires r0. */ | |
2279 | if (GET_CODE (operands[0]) == REG && REGNO (operands[0]) == 0) | |
97595bfd | 2280 | { |
73401833 | 2281 | emit_insn (gen_andsi3 (operands[0], operands[0], const1_rtx)); |
2282 | emit_insn (gen_rotlsi3_31 (operands[0], operands[0])); | |
2283 | return; | |
97595bfd | 2284 | } |
1504eb7e | 2285 | } |
73401833 | 2286 | } |
37ad966e | 2287 | else if (value == 0) |
2288 | { | |
59312820 | 2289 | /* This can happen even when optimizing, if there were subregs before |
2290 | reload. Don't output a nop here, as this is never optimized away; | |
2291 | use a no-op move instead. */ | |
2292 | emit_insn (gen_rtx_SET (VOIDmode, operands[0], operands[0])); | |
37ad966e | 2293 | return; |
2294 | } | |
4248bee7 | 2295 | |
73401833 | 2296 | max = shift_insns[value]; |
2297 | for (i = 0; i < max; i++) | |
2298 | gen_ashift (code, shift_amounts[value][i], operands[0]); | |
2299 | } | |
9e7454d0 | 2300 | |
2ce8df5a | 2301 | /* Same as above, but optimized for values where the topmost bits don't |
2302 | matter. */ | |
2303 | ||
b7dbbdb2 | 2304 | void |
04f04b72 | 2305 | gen_shifty_hi_op (int code, rtx *operands) |
2ce8df5a | 2306 | { |
2307 | int value = INTVAL (operands[2]); | |
2308 | int max, i; | |
04f04b72 | 2309 | void (*gen_fun) (int, int, rtx); |
2ce8df5a | 2310 | |
2311 | /* This operation is used by and_shl for SImode values with a few | |
2312 | high bits known to be cleared. */ | |
2313 | value &= 31; | |
2314 | if (value == 0) | |
2315 | { | |
2316 | emit_insn (gen_nop ()); | |
2317 | return; | |
2318 | } | |
2319 | ||
2320 | gen_fun = GET_MODE (operands[0]) == HImode ? gen_ashift_hi : gen_ashift; | |
2321 | if (code == ASHIFT) | |
2322 | { | |
2323 | max = ext_shift_insns[value]; | |
2324 | for (i = 0; i < max; i++) | |
2325 | gen_fun (code, ext_shift_amounts[value][i], operands[0]); | |
2326 | } | |
2327 | else | |
2328 | /* When shifting right, emit the shifts in reverse order, so that | |
2329 | solitary negative values come first. */ | |
2330 | for (i = ext_shift_insns[value] - 1; i >= 0; i--) | |
2331 | gen_fun (code, ext_shift_amounts[value][i], operands[0]); | |
2332 | } | |
73401833 | 2333 | |
2334 | /* Output RTL for an arithmetic right shift. */ | |
2335 | ||
2336 | /* ??? Rewrite to use super-optimizer sequences. */ | |
2337 | ||
2338 | int | |
04f04b72 | 2339 | expand_ashiftrt (rtx *operands) |
73401833 | 2340 | { |
2341 | rtx wrk; | |
2342 | char func[18]; | |
73401833 | 2343 | int value; |
2344 | ||
79f0f2fe | 2345 | if (TARGET_SH3) |
0b89dfaf | 2346 | { |
79f0f2fe | 2347 | if (GET_CODE (operands[2]) != CONST_INT) |
2348 | { | |
2349 | rtx count = copy_to_mode_reg (SImode, operands[2]); | |
2350 | emit_insn (gen_negsi2 (count, count)); | |
2351 | emit_insn (gen_ashrsi3_d (operands[0], operands[1], count)); | |
2352 | return 1; | |
2353 | } | |
8ded0752 | 2354 | else if (ashiftrt_insns[INTVAL (operands[2]) & 31] |
2355 | > 1 + SH_DYNAMIC_SHIFT_COST) | |
79f0f2fe | 2356 | { |
8ded0752 | 2357 | rtx count |
2358 | = force_reg (SImode, GEN_INT (- (INTVAL (operands[2]) & 31))); | |
79f0f2fe | 2359 | emit_insn (gen_ashrsi3_d (operands[0], operands[1], count)); |
2360 | return 1; | |
2361 | } | |
0b89dfaf | 2362 | } |
73401833 | 2363 | if (GET_CODE (operands[2]) != CONST_INT) |
2364 | return 0; | |
2365 | ||
8ded0752 | 2366 | value = INTVAL (operands[2]) & 31; |
73401833 | 2367 | |
2368 | if (value == 31) | |
2369 | { | |
59312820 | 2370 | /* If we are called from abs expansion, arrange things so that we |
2371 | we can use a single MT instruction that doesn't clobber the source, | |
2372 | if LICM can hoist out the load of the constant zero. */ | |
2373 | if (currently_expanding_to_rtl) | |
2374 | { | |
2375 | emit_insn (gen_cmpgtsi_t (force_reg (SImode, CONST0_RTX (SImode)), | |
2376 | operands[1])); | |
2377 | emit_insn (gen_mov_neg_si_t (operands[0])); | |
2378 | return 1; | |
2379 | } | |
73401833 | 2380 | emit_insn (gen_ashrsi2_31 (operands[0], operands[1])); |
2381 | return 1; | |
2382 | } | |
2383 | else if (value >= 16 && value <= 19) | |
2384 | { | |
2385 | wrk = gen_reg_rtx (SImode); | |
2386 | emit_insn (gen_ashrsi2_16 (wrk, operands[1])); | |
2387 | value -= 16; | |
2388 | while (value--) | |
2389 | gen_ashift (ASHIFTRT, 1, wrk); | |
2390 | emit_move_insn (operands[0], wrk); | |
2391 | return 1; | |
4248bee7 | 2392 | } |
73401833 | 2393 | /* Expand a short sequence inline, longer call a magic routine. */ |
2394 | else if (value <= 5) | |
2395 | { | |
2396 | wrk = gen_reg_rtx (SImode); | |
2397 | emit_move_insn (wrk, operands[1]); | |
2398 | while (value--) | |
2399 | gen_ashift (ASHIFTRT, 1, wrk); | |
2400 | emit_move_insn (operands[0], wrk); | |
2401 | return 1; | |
2402 | } | |
2403 | ||
2404 | wrk = gen_reg_rtx (Pmode); | |
2405 | ||
2406 | /* Load the value into an arg reg and call a helper. */ | |
7014838c | 2407 | emit_move_insn (gen_rtx_REG (SImode, 4), operands[1]); |
73401833 | 2408 | sprintf (func, "__ashiftrt_r4_%d", value); |
59312820 | 2409 | function_symbol (wrk, func, SFUNC_STATIC); |
73401833 | 2410 | emit_insn (gen_ashrsi3_n (GEN_INT (value), wrk)); |
7014838c | 2411 | emit_move_insn (operands[0], gen_rtx_REG (SImode, 4)); |
73401833 | 2412 | return 1; |
7208c779 | 2413 | } |
2ce8df5a | 2414 | |
b7dbbdb2 | 2415 | int |
04f04b72 | 2416 | sh_dynamicalize_shift_p (rtx count) |
8ded0752 | 2417 | { |
2418 | return shift_insns[INTVAL (count)] > 1 + SH_DYNAMIC_SHIFT_COST; | |
2419 | } | |
2420 | ||
2ce8df5a | 2421 | /* Try to find a good way to implement the combiner pattern |
2422 | [(set (match_operand:SI 0 "register_operand" "r") | |
2423 | (and:SI (ashift:SI (match_operand:SI 1 "register_operand" "r") | |
2424 | (match_operand:SI 2 "const_int_operand" "n")) | |
2425 | (match_operand:SI 3 "const_int_operand" "n"))) . | |
2426 | LEFT_RTX is operand 2 in the above pattern, and MASK_RTX is operand 3. | |
2427 | return 0 for simple right / left or left/right shift combination. | |
2428 | return 1 for a combination of shifts with zero_extend. | |
2429 | return 2 for a combination of shifts with an AND that needs r0. | |
2430 | return 3 for a combination of shifts with an AND that needs an extra | |
2431 | scratch register, when the three highmost bits of the AND mask are clear. | |
2432 | return 4 for a combination of shifts with an AND that needs an extra | |
2433 | scratch register, when any of the three highmost bits of the AND mask | |
2434 | is set. | |
2435 | If ATTRP is set, store an initial right shift width in ATTRP[0], | |
2436 | and the instruction length in ATTRP[1] . These values are not valid | |
2437 | when returning 0. | |
2438 | When ATTRP is set and returning 1, ATTRP[2] gets set to the index into | |
2439 | shift_amounts for the last shift value that is to be used before the | |
2440 | sign extend. */ | |
2441 | int | |
04f04b72 | 2442 | shl_and_kind (rtx left_rtx, rtx mask_rtx, int *attrp) |
2ce8df5a | 2443 | { |
2444 | unsigned HOST_WIDE_INT mask, lsb, mask2, lsb2; | |
2445 | int left = INTVAL (left_rtx), right; | |
2446 | int best = 0; | |
2447 | int cost, best_cost = 10000; | |
2448 | int best_right = 0, best_len = 0; | |
2449 | int i; | |
2450 | int can_ext; | |
2451 | ||
2452 | if (left < 0 || left > 31) | |
2453 | return 0; | |
2454 | if (GET_CODE (mask_rtx) == CONST_INT) | |
2455 | mask = (unsigned HOST_WIDE_INT) INTVAL (mask_rtx) >> left; | |
2456 | else | |
2457 | mask = (unsigned HOST_WIDE_INT) GET_MODE_MASK (SImode) >> left; | |
d767e27e | 2458 | /* Can this be expressed as a right shift / left shift pair? */ |
2ce8df5a | 2459 | lsb = ((mask ^ (mask - 1)) >> 1) + 1; |
2460 | right = exact_log2 (lsb); | |
2461 | mask2 = ~(mask + lsb - 1); | |
2462 | lsb2 = ((mask2 ^ (mask2 - 1)) >> 1) + 1; | |
2463 | /* mask has no zeroes but trailing zeroes <==> ! mask2 */ | |
2464 | if (! mask2) | |
2465 | best_cost = shift_insns[right] + shift_insns[right + left]; | |
2466 | /* mask has no trailing zeroes <==> ! right */ | |
2467 | else if (! right && mask2 == ~(lsb2 - 1)) | |
2468 | { | |
2469 | int late_right = exact_log2 (lsb2); | |
2470 | best_cost = shift_insns[left + late_right] + shift_insns[late_right]; | |
2471 | } | |
d767e27e | 2472 | /* Try to use zero extend. */ |
2ce8df5a | 2473 | if (mask2 == ~(lsb2 - 1)) |
2474 | { | |
2475 | int width, first; | |
2476 | ||
2477 | for (width = 8; width <= 16; width += 8) | |
2478 | { | |
d767e27e | 2479 | /* Can we zero-extend right away? */ |
2480 | if (lsb2 == (unsigned HOST_WIDE_INT) 1 << width) | |
2ce8df5a | 2481 | { |
2482 | cost | |
2483 | = 1 + ext_shift_insns[right] + ext_shift_insns[left + right]; | |
2484 | if (cost < best_cost) | |
2485 | { | |
2486 | best = 1; | |
2487 | best_cost = cost; | |
2488 | best_right = right; | |
2489 | best_len = cost; | |
2490 | if (attrp) | |
2491 | attrp[2] = -1; | |
2492 | } | |
2493 | continue; | |
2494 | } | |
2495 | /* ??? Could try to put zero extend into initial right shift, | |
9df03d69 | 2496 | or even shift a bit left before the right shift. */ |
2ce8df5a | 2497 | /* Determine value of first part of left shift, to get to the |
2498 | zero extend cut-off point. */ | |
2499 | first = width - exact_log2 (lsb2) + right; | |
2500 | if (first >= 0 && right + left - first >= 0) | |
2501 | { | |
2502 | cost = ext_shift_insns[right] + ext_shift_insns[first] + 1 | |
2503 | + ext_shift_insns[right + left - first]; | |
2504 | if (cost < best_cost) | |
2505 | { | |
2506 | best = 1; | |
2507 | best_cost = cost; | |
2508 | best_right = right; | |
2509 | best_len = cost; | |
2510 | if (attrp) | |
2511 | attrp[2] = first; | |
d767e27e | 2512 | } |
2ce8df5a | 2513 | } |
2514 | } | |
2515 | } | |
2516 | /* Try to use r0 AND pattern */ | |
2517 | for (i = 0; i <= 2; i++) | |
2518 | { | |
2519 | if (i > right) | |
2520 | break; | |
af2c1324 | 2521 | if (! CONST_OK_FOR_K08 (mask >> i)) |
2ce8df5a | 2522 | continue; |
2523 | cost = (i != 0) + 2 + ext_shift_insns[left + i]; | |
2524 | if (cost < best_cost) | |
2525 | { | |
2526 | best = 2; | |
2527 | best_cost = cost; | |
2528 | best_right = i; | |
2529 | best_len = cost - 1; | |
2530 | } | |
2531 | } | |
2532 | /* Try to use a scratch register to hold the AND operand. */ | |
d767e27e | 2533 | can_ext = ((mask << left) & ((unsigned HOST_WIDE_INT) 3 << 30)) == 0; |
2ce8df5a | 2534 | for (i = 0; i <= 2; i++) |
2535 | { | |
2536 | if (i > right) | |
2537 | break; | |
af2c1324 | 2538 | cost = (i != 0) + (CONST_OK_FOR_I08 (mask >> i) ? 2 : 3) |
57ea6605 | 2539 | + (can_ext ? ext_shift_insns : shift_insns)[left + i]; |
2ce8df5a | 2540 | if (cost < best_cost) |
2541 | { | |
2542 | best = 4 - can_ext; | |
2543 | best_cost = cost; | |
2544 | best_right = i; | |
af2c1324 | 2545 | best_len = cost - 1 - ! CONST_OK_FOR_I08 (mask >> i); |
2ce8df5a | 2546 | } |
2547 | } | |
2548 | ||
2549 | if (attrp) | |
2550 | { | |
2551 | attrp[0] = best_right; | |
2552 | attrp[1] = best_len; | |
2553 | } | |
2554 | return best; | |
2555 | } | |
2556 | ||
2557 | /* This is used in length attributes of the unnamed instructions | |
2558 | corresponding to shl_and_kind return values of 1 and 2. */ | |
2559 | int | |
04f04b72 | 2560 | shl_and_length (rtx insn) |
2ce8df5a | 2561 | { |
2562 | rtx set_src, left_rtx, mask_rtx; | |
2563 | int attributes[3]; | |
2564 | ||
2565 | set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
2566 | left_rtx = XEXP (XEXP (set_src, 0), 1); | |
2567 | mask_rtx = XEXP (set_src, 1); | |
2568 | shl_and_kind (left_rtx, mask_rtx, attributes); | |
2569 | return attributes[1]; | |
2570 | } | |
2571 | ||
2572 | /* This is used in length attribute of the and_shl_scratch instruction. */ | |
2573 | ||
2574 | int | |
04f04b72 | 2575 | shl_and_scr_length (rtx insn) |
2ce8df5a | 2576 | { |
2577 | rtx set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
2578 | int len = shift_insns[INTVAL (XEXP (set_src, 1))]; | |
2579 | rtx op = XEXP (set_src, 0); | |
2580 | len += shift_insns[INTVAL (XEXP (op, 1))] + 1; | |
2581 | op = XEXP (XEXP (op, 0), 0); | |
2582 | return len + shift_insns[INTVAL (XEXP (op, 1))]; | |
2583 | } | |
2584 | ||
2ce8df5a | 2585 | /* Generate rtl for instructions for which shl_and_kind advised a particular |
2586 | method of generating them, i.e. returned zero. */ | |
2587 | ||
2588 | int | |
04f04b72 | 2589 | gen_shl_and (rtx dest, rtx left_rtx, rtx mask_rtx, rtx source) |
2ce8df5a | 2590 | { |
2591 | int attributes[3]; | |
2592 | unsigned HOST_WIDE_INT mask; | |
2593 | int kind = shl_and_kind (left_rtx, mask_rtx, attributes); | |
2594 | int right, total_shift; | |
d767e27e | 2595 | void (*shift_gen_fun) (int, rtx *) = gen_shifty_hi_op; |
2ce8df5a | 2596 | |
2597 | right = attributes[0]; | |
2598 | total_shift = INTVAL (left_rtx) + right; | |
2599 | mask = (unsigned HOST_WIDE_INT) INTVAL (mask_rtx) >> total_shift; | |
2600 | switch (kind) | |
2601 | { | |
2602 | default: | |
2603 | return -1; | |
2604 | case 1: | |
2605 | { | |
2606 | int first = attributes[2]; | |
2607 | rtx operands[3]; | |
2608 | ||
2609 | if (first < 0) | |
2610 | { | |
250fc29e | 2611 | emit_insn ((mask << right) <= 0xff |
d767e27e | 2612 | ? gen_zero_extendqisi2 (dest, |
2613 | gen_lowpart (QImode, source)) | |
2614 | : gen_zero_extendhisi2 (dest, | |
2615 | gen_lowpart (HImode, source))); | |
2ce8df5a | 2616 | source = dest; |
2617 | } | |
2618 | if (source != dest) | |
2619 | emit_insn (gen_movsi (dest, source)); | |
2620 | operands[0] = dest; | |
2621 | if (right) | |
2622 | { | |
2623 | operands[2] = GEN_INT (right); | |
2624 | gen_shifty_hi_op (LSHIFTRT, operands); | |
2625 | } | |
2626 | if (first > 0) | |
2627 | { | |
2628 | operands[2] = GEN_INT (first); | |
2629 | gen_shifty_hi_op (ASHIFT, operands); | |
2630 | total_shift -= first; | |
2631 | mask <<= first; | |
2632 | } | |
2633 | if (first >= 0) | |
c721f167 | 2634 | emit_insn (mask <= 0xff |
d767e27e | 2635 | ? gen_zero_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
2636 | : gen_zero_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
2ce8df5a | 2637 | if (total_shift > 0) |
2638 | { | |
2639 | operands[2] = GEN_INT (total_shift); | |
2640 | gen_shifty_hi_op (ASHIFT, operands); | |
2641 | } | |
2642 | break; | |
2643 | } | |
2644 | case 4: | |
2645 | shift_gen_fun = gen_shifty_op; | |
2ce8df5a | 2646 | case 3: |
74605264 | 2647 | /* If the topmost bit that matters is set, set the topmost bits |
2648 | that don't matter. This way, we might be able to get a shorter | |
2649 | signed constant. */ | |
d767e27e | 2650 | if (mask & ((HOST_WIDE_INT) 1 << (31 - total_shift))) |
2651 | mask |= (HOST_WIDE_INT) ~0 << (31 - total_shift); | |
f8641a34 | 2652 | case 2: |
2ce8df5a | 2653 | /* Don't expand fine-grained when combining, because that will |
2654 | make the pattern fail. */ | |
723c0ee7 | 2655 | if (currently_expanding_to_rtl |
2ce8df5a | 2656 | || reload_in_progress || reload_completed) |
2657 | { | |
2658 | rtx operands[3]; | |
9e7454d0 | 2659 | |
f8641a34 | 2660 | /* Cases 3 and 4 should be handled by this split |
2661 | only while combining */ | |
1a6a7a27 | 2662 | gcc_assert (kind <= 2); |
2ce8df5a | 2663 | if (right) |
2664 | { | |
2665 | emit_insn (gen_lshrsi3 (dest, source, GEN_INT (right))); | |
2666 | source = dest; | |
2667 | } | |
2668 | emit_insn (gen_andsi3 (dest, source, GEN_INT (mask))); | |
2d7f129b | 2669 | if (total_shift) |
2670 | { | |
2671 | operands[0] = dest; | |
2672 | operands[1] = dest; | |
2673 | operands[2] = GEN_INT (total_shift); | |
2674 | shift_gen_fun (ASHIFT, operands); | |
2675 | } | |
2ce8df5a | 2676 | break; |
2677 | } | |
2678 | else | |
2679 | { | |
2680 | int neg = 0; | |
2681 | if (kind != 4 && total_shift < 16) | |
2682 | { | |
2683 | neg = -ext_shift_amounts[total_shift][1]; | |
2684 | if (neg > 0) | |
2685 | neg -= ext_shift_amounts[total_shift][2]; | |
2686 | else | |
2687 | neg = 0; | |
2688 | } | |
2689 | emit_insn (gen_and_shl_scratch (dest, source, | |
2690 | GEN_INT (right), | |
2691 | GEN_INT (mask), | |
2692 | GEN_INT (total_shift + neg), | |
2693 | GEN_INT (neg))); | |
2694 | emit_insn (gen_movsi (dest, dest)); | |
2695 | break; | |
2696 | } | |
2697 | } | |
2698 | return 0; | |
2699 | } | |
2700 | ||
2701 | /* Try to find a good way to implement the combiner pattern | |
2702 | [(set (match_operand:SI 0 "register_operand" "=r") | |
2703 | (sign_extract:SI (ashift:SI (match_operand:SI 1 "register_operand" "r") | |
2704 | (match_operand:SI 2 "const_int_operand" "n") | |
2705 | (match_operand:SI 3 "const_int_operand" "n") | |
2706 | (const_int 0))) | |
a5bd34f1 | 2707 | (clobber (reg:SI T_REG))] |
2ce8df5a | 2708 | LEFT_RTX is operand 2 in the above pattern, and SIZE_RTX is operand 3. |
2709 | return 0 for simple left / right shift combination. | |
2710 | return 1 for left shift / 8 bit sign extend / left shift. | |
2711 | return 2 for left shift / 16 bit sign extend / left shift. | |
2712 | return 3 for left shift / 8 bit sign extend / shift / sign extend. | |
2713 | return 4 for left shift / 16 bit sign extend / shift / sign extend. | |
2714 | return 5 for left shift / 16 bit sign extend / right shift | |
2715 | return 6 for < 8 bit sign extend / left shift. | |
2716 | return 7 for < 8 bit sign extend / left shift / single right shift. | |
2717 | If COSTP is nonzero, assign the calculated cost to *COSTP. */ | |
2718 | ||
2719 | int | |
04f04b72 | 2720 | shl_sext_kind (rtx left_rtx, rtx size_rtx, int *costp) |
2ce8df5a | 2721 | { |
2722 | int left, size, insize, ext; | |
87ed74ef | 2723 | int cost = 0, best_cost; |
2ce8df5a | 2724 | int kind; |
2725 | ||
2726 | left = INTVAL (left_rtx); | |
2727 | size = INTVAL (size_rtx); | |
2728 | insize = size - left; | |
1a6a7a27 | 2729 | gcc_assert (insize > 0); |
2ce8df5a | 2730 | /* Default to left / right shift. */ |
2731 | kind = 0; | |
2732 | best_cost = shift_insns[32 - insize] + ashiftrt_insns[32 - size]; | |
2733 | if (size <= 16) | |
2734 | { | |
2735 | /* 16 bit shift / sign extend / 16 bit shift */ | |
2736 | cost = shift_insns[16 - insize] + 1 + ashiftrt_insns[16 - size]; | |
2d7f129b | 2737 | /* If ashiftrt_insns[16 - size] is 8, this choice will be overridden |
2738 | below, by alternative 3 or something even better. */ | |
2ce8df5a | 2739 | if (cost < best_cost) |
2740 | { | |
2741 | kind = 5; | |
2742 | best_cost = cost; | |
2743 | } | |
2744 | } | |
2745 | /* Try a plain sign extend between two shifts. */ | |
2746 | for (ext = 16; ext >= insize; ext -= 8) | |
2747 | { | |
2748 | if (ext <= size) | |
2749 | { | |
2750 | cost = ext_shift_insns[ext - insize] + 1 + shift_insns[size - ext]; | |
2751 | if (cost < best_cost) | |
2752 | { | |
30435bf8 | 2753 | kind = ext / (unsigned) 8; |
2ce8df5a | 2754 | best_cost = cost; |
2755 | } | |
2756 | } | |
69b02875 | 2757 | /* Check if we can do a sloppy shift with a final signed shift |
2758 | restoring the sign. */ | |
2759 | if (EXT_SHIFT_SIGNED (size - ext)) | |
2760 | cost = ext_shift_insns[ext - insize] + ext_shift_insns[size - ext] + 1; | |
2761 | /* If not, maybe it's still cheaper to do the second shift sloppy, | |
2762 | and do a final sign extend? */ | |
2763 | else if (size <= 16) | |
2764 | cost = ext_shift_insns[ext - insize] + 1 | |
2765 | + ext_shift_insns[size > ext ? size - ext : ext - size] + 1; | |
2766 | else | |
2767 | continue; | |
2768 | if (cost < best_cost) | |
2ce8df5a | 2769 | { |
30435bf8 | 2770 | kind = ext / (unsigned) 8 + 2; |
69b02875 | 2771 | best_cost = cost; |
2ce8df5a | 2772 | } |
2773 | } | |
2774 | /* Check if we can sign extend in r0 */ | |
2775 | if (insize < 8) | |
2776 | { | |
2777 | cost = 3 + shift_insns[left]; | |
2778 | if (cost < best_cost) | |
2779 | { | |
2780 | kind = 6; | |
2781 | best_cost = cost; | |
2782 | } | |
2783 | /* Try the same with a final signed shift. */ | |
2784 | if (left < 31) | |
2785 | { | |
2786 | cost = 3 + ext_shift_insns[left + 1] + 1; | |
2787 | if (cost < best_cost) | |
2788 | { | |
2789 | kind = 7; | |
2790 | best_cost = cost; | |
2791 | } | |
2792 | } | |
2793 | } | |
2794 | if (TARGET_SH3) | |
2795 | { | |
2796 | /* Try to use a dynamic shift. */ | |
8ded0752 | 2797 | cost = shift_insns[32 - insize] + 1 + SH_DYNAMIC_SHIFT_COST; |
2ce8df5a | 2798 | if (cost < best_cost) |
2799 | { | |
2800 | kind = 0; | |
2801 | best_cost = cost; | |
2802 | } | |
2803 | } | |
2804 | if (costp) | |
2805 | *costp = cost; | |
2806 | return kind; | |
2807 | } | |
2808 | ||
2809 | /* Function to be used in the length attribute of the instructions | |
2810 | implementing this pattern. */ | |
2811 | ||
2812 | int | |
04f04b72 | 2813 | shl_sext_length (rtx insn) |
2ce8df5a | 2814 | { |
2815 | rtx set_src, left_rtx, size_rtx; | |
2816 | int cost; | |
2817 | ||
2818 | set_src = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
2819 | left_rtx = XEXP (XEXP (set_src, 0), 1); | |
2820 | size_rtx = XEXP (set_src, 1); | |
2821 | shl_sext_kind (left_rtx, size_rtx, &cost); | |
2822 | return cost; | |
2823 | } | |
2824 | ||
2825 | /* Generate rtl for this pattern */ | |
2826 | ||
2827 | int | |
04f04b72 | 2828 | gen_shl_sext (rtx dest, rtx left_rtx, rtx size_rtx, rtx source) |
2ce8df5a | 2829 | { |
2830 | int kind; | |
1a47e85b | 2831 | int left, size, insize, cost; |
2ce8df5a | 2832 | rtx operands[3]; |
2833 | ||
1a47e85b | 2834 | kind = shl_sext_kind (left_rtx, size_rtx, &cost); |
2ce8df5a | 2835 | left = INTVAL (left_rtx); |
2836 | size = INTVAL (size_rtx); | |
2837 | insize = size - left; | |
2838 | switch (kind) | |
2839 | { | |
2840 | case 1: | |
2841 | case 2: | |
2842 | case 3: | |
2843 | case 4: | |
2844 | { | |
2845 | int ext = kind & 1 ? 8 : 16; | |
2846 | int shift2 = size - ext; | |
2847 | ||
2848 | /* Don't expand fine-grained when combining, because that will | |
2849 | make the pattern fail. */ | |
723c0ee7 | 2850 | if (! currently_expanding_to_rtl |
2ce8df5a | 2851 | && ! reload_in_progress && ! reload_completed) |
2852 | { | |
2853 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
2854 | emit_insn (gen_movsi (dest, source)); | |
2855 | break; | |
2856 | } | |
2857 | if (dest != source) | |
2858 | emit_insn (gen_movsi (dest, source)); | |
2859 | operands[0] = dest; | |
2d7f129b | 2860 | if (ext - insize) |
2861 | { | |
2862 | operands[2] = GEN_INT (ext - insize); | |
2863 | gen_shifty_hi_op (ASHIFT, operands); | |
2864 | } | |
2ce8df5a | 2865 | emit_insn (kind & 1 |
d767e27e | 2866 | ? gen_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
2867 | : gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
2ce8df5a | 2868 | if (kind <= 2) |
2869 | { | |
2d7f129b | 2870 | if (shift2) |
2871 | { | |
2872 | operands[2] = GEN_INT (shift2); | |
2873 | gen_shifty_op (ASHIFT, operands); | |
2874 | } | |
2ce8df5a | 2875 | } |
2876 | else | |
2877 | { | |
2d7f129b | 2878 | if (shift2 > 0) |
2ce8df5a | 2879 | { |
69b02875 | 2880 | if (EXT_SHIFT_SIGNED (shift2)) |
2881 | { | |
2882 | operands[2] = GEN_INT (shift2 + 1); | |
2883 | gen_shifty_op (ASHIFT, operands); | |
bcd9bd66 | 2884 | operands[2] = const1_rtx; |
69b02875 | 2885 | gen_shifty_op (ASHIFTRT, operands); |
2886 | break; | |
2887 | } | |
2ce8df5a | 2888 | operands[2] = GEN_INT (shift2); |
2889 | gen_shifty_hi_op (ASHIFT, operands); | |
2890 | } | |
2d7f129b | 2891 | else if (shift2) |
2ce8df5a | 2892 | { |
2893 | operands[2] = GEN_INT (-shift2); | |
2894 | gen_shifty_hi_op (LSHIFTRT, operands); | |
2895 | } | |
2896 | emit_insn (size <= 8 | |
2d7f129b | 2897 | ? gen_extendqisi2 (dest, gen_lowpart (QImode, dest)) |
2898 | : gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
2ce8df5a | 2899 | } |
2900 | break; | |
2901 | } | |
2902 | case 5: | |
2d7f129b | 2903 | { |
2904 | int i = 16 - size; | |
723c0ee7 | 2905 | if (! currently_expanding_to_rtl |
f2bfaeea | 2906 | && ! reload_in_progress && ! reload_completed) |
2907 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
2908 | else | |
2909 | { | |
2910 | operands[0] = dest; | |
2911 | operands[2] = GEN_INT (16 - insize); | |
2912 | gen_shifty_hi_op (ASHIFT, operands); | |
2913 | emit_insn (gen_extendhisi2 (dest, gen_lowpart (HImode, dest))); | |
2914 | } | |
2d7f129b | 2915 | /* Don't use gen_ashrsi3 because it generates new pseudos. */ |
2916 | while (--i >= 0) | |
2917 | gen_ashift (ASHIFTRT, 1, dest); | |
2918 | break; | |
2919 | } | |
2ce8df5a | 2920 | case 6: |
2921 | case 7: | |
2922 | /* Don't expand fine-grained when combining, because that will | |
2923 | make the pattern fail. */ | |
723c0ee7 | 2924 | if (! currently_expanding_to_rtl |
2ce8df5a | 2925 | && ! reload_in_progress && ! reload_completed) |
2926 | { | |
2927 | emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx)); | |
2928 | emit_insn (gen_movsi (dest, source)); | |
2929 | break; | |
2930 | } | |
2931 | emit_insn (gen_andsi3 (dest, source, GEN_INT ((1 << insize) - 1))); | |
2932 | emit_insn (gen_xorsi3 (dest, dest, GEN_INT (1 << (insize - 1)))); | |
2933 | emit_insn (gen_addsi3 (dest, dest, GEN_INT (-1 << (insize - 1)))); | |
2934 | operands[0] = dest; | |
2935 | operands[2] = kind == 7 ? GEN_INT (left + 1) : left_rtx; | |
2936 | gen_shifty_op (ASHIFT, operands); | |
2937 | if (kind == 7) | |
bcd9bd66 | 2938 | emit_insn (gen_ashrsi3_k (dest, dest, const1_rtx)); |
2ce8df5a | 2939 | break; |
2940 | default: | |
2941 | return -1; | |
2942 | } | |
2943 | return 0; | |
2944 | } | |
87e19636 | 2945 | |
2946 | /* Prefix a symbol_ref name with "datalabel". */ | |
2947 | ||
2948 | rtx | |
04f04b72 | 2949 | gen_datalabel_ref (rtx sym) |
87e19636 | 2950 | { |
59312820 | 2951 | const char *str; |
2952 | ||
87e19636 | 2953 | if (GET_CODE (sym) == LABEL_REF) |
2954 | return gen_rtx_CONST (GET_MODE (sym), | |
2955 | gen_rtx_UNSPEC (GET_MODE (sym), | |
2956 | gen_rtvec (1, sym), | |
2957 | UNSPEC_DATALABEL)); | |
9e7454d0 | 2958 | |
1a6a7a27 | 2959 | gcc_assert (GET_CODE (sym) == SYMBOL_REF); |
87e19636 | 2960 | |
59312820 | 2961 | str = XSTR (sym, 0); |
2962 | /* Share all SYMBOL_REF strings with the same value - that is important | |
2963 | for cse. */ | |
2964 | str = IDENTIFIER_POINTER (get_identifier (str)); | |
2965 | XSTR (sym, 0) = str; | |
2966 | ||
87e19636 | 2967 | return sym; |
2968 | } | |
2969 | ||
73401833 | 2970 | \f |
2971 | /* The SH cannot load a large constant into a register, constants have to | |
2972 | come from a pc relative load. The reference of a pc relative load | |
442e3cb9 | 2973 | instruction must be less than 1k in front of the instruction. This |
73401833 | 2974 | means that we often have to dump a constant inside a function, and |
2975 | generate code to branch around it. | |
7208c779 | 2976 | |
73401833 | 2977 | It is important to minimize this, since the branches will slow things |
2978 | down and make things bigger. | |
2979 | ||
2980 | Worst case code looks like: | |
2981 | ||
2982 | mov.l L1,rn | |
2983 | bra L2 | |
2984 | nop | |
2985 | align | |
2986 | L1: .long value | |
2987 | L2: | |
2988 | .. | |
2989 | ||
2990 | mov.l L3,rn | |
2991 | bra L4 | |
2992 | nop | |
2993 | align | |
2994 | L3: .long value | |
2995 | L4: | |
2996 | .. | |
2997 | ||
2998 | We fix this by performing a scan before scheduling, which notices which | |
2999 | instructions need to have their operands fetched from the constant table | |
3000 | and builds the table. | |
3001 | ||
3002 | The algorithm is: | |
3003 | ||
3004 | scan, find an instruction which needs a pcrel move. Look forward, find the | |
3005 | last barrier which is within MAX_COUNT bytes of the requirement. | |
3006 | If there isn't one, make one. Process all the instructions between | |
3007 | the find and the barrier. | |
3008 | ||
3009 | In the above example, we can tell that L3 is within 1k of L1, so | |
3010 | the first move can be shrunk from the 3 insn+constant sequence into | |
3011 | just 1 insn, and the constant moved to L3 to make: | |
3012 | ||
3013 | mov.l L1,rn | |
3014 | .. | |
3015 | mov.l L3,rn | |
3016 | bra L4 | |
3017 | nop | |
3018 | align | |
3019 | L3:.long value | |
3020 | L4:.long value | |
3021 | ||
3022 | Then the second move becomes the target for the shortening process. */ | |
3023 | ||
3024 | typedef struct | |
3025 | { | |
3026 | rtx value; /* Value in table. */ | |
3027 | rtx label; /* Label of value. */ | |
ecf6ff7c | 3028 | rtx wend; /* End of window. */ |
73401833 | 3029 | enum machine_mode mode; /* Mode of value. */ |
c5ea35c4 | 3030 | |
3031 | /* True if this constant is accessed as part of a post-increment | |
3032 | sequence. Note that HImode constants are never accessed in this way. */ | |
3033 | bool part_of_sequence_p; | |
73401833 | 3034 | } pool_node; |
3035 | ||
3036 | /* The maximum number of constants that can fit into one pool, since | |
59312820 | 3037 | constants in the range 0..510 are at least 2 bytes long, and in the |
3038 | range from there to 1018 at least 4 bytes. */ | |
73401833 | 3039 | |
59312820 | 3040 | #define MAX_POOL_SIZE 372 |
73401833 | 3041 | static pool_node pool_vector[MAX_POOL_SIZE]; |
3042 | static int pool_size; | |
ecf6ff7c | 3043 | static rtx pool_window_label; |
3044 | static int pool_window_last; | |
73401833 | 3045 | |
3046 | /* ??? If we need a constant in HImode which is the truncated value of a | |
3047 | constant we need in SImode, we could combine the two entries thus saving | |
3048 | two bytes. Is this common enough to be worth the effort of implementing | |
3049 | it? */ | |
3050 | ||
3051 | /* ??? This stuff should be done at the same time that we shorten branches. | |
3052 | As it is now, we must assume that all branches are the maximum size, and | |
3053 | this causes us to almost always output constant pools sooner than | |
3054 | necessary. */ | |
3055 | ||
3056 | /* Add a constant to the pool and return its label. */ | |
3057 | ||
3058 | static rtx | |
04f04b72 | 3059 | add_constant (rtx x, enum machine_mode mode, rtx last_value) |
97595bfd | 3060 | { |
3061 | int i; | |
ecf6ff7c | 3062 | rtx lab, new, ref, newref; |
73401833 | 3063 | |
3064 | /* First see if we've already got it. */ | |
3065 | for (i = 0; i < pool_size; i++) | |
97595bfd | 3066 | { |
73401833 | 3067 | if (x->code == pool_vector[i].value->code |
3068 | && mode == pool_vector[i].mode) | |
97595bfd | 3069 | { |
73401833 | 3070 | if (x->code == CODE_LABEL) |
3071 | { | |
3072 | if (XINT (x, 3) != XINT (pool_vector[i].value, 3)) | |
3073 | continue; | |
3074 | } | |
3075 | if (rtx_equal_p (x, pool_vector[i].value)) | |
1b61190c | 3076 | { |
ecf6ff7c | 3077 | lab = new = 0; |
1b61190c | 3078 | if (! last_value |
3079 | || ! i | |
3080 | || ! rtx_equal_p (last_value, pool_vector[i-1].value)) | |
3081 | { | |
ecf6ff7c | 3082 | new = gen_label_rtx (); |
3083 | LABEL_REFS (new) = pool_vector[i].label; | |
3084 | pool_vector[i].label = lab = new; | |
1b61190c | 3085 | } |
ecf6ff7c | 3086 | if (lab && pool_window_label) |
3087 | { | |
3088 | newref = gen_rtx_LABEL_REF (VOIDmode, pool_window_label); | |
3089 | ref = pool_vector[pool_window_last].wend; | |
3090 | LABEL_NEXTREF (newref) = ref; | |
3091 | pool_vector[pool_window_last].wend = newref; | |
3092 | } | |
3093 | if (new) | |
3094 | pool_window_label = new; | |
3095 | pool_window_last = i; | |
1b61190c | 3096 | return lab; |
3097 | } | |
97595bfd | 3098 | } |
97595bfd | 3099 | } |
47c009e5 | 3100 | |
73401833 | 3101 | /* Need a new one. */ |
3102 | pool_vector[pool_size].value = x; | |
1b61190c | 3103 | if (last_value && rtx_equal_p (last_value, pool_vector[pool_size - 1].value)) |
c5ea35c4 | 3104 | { |
3105 | lab = 0; | |
3106 | pool_vector[pool_size - 1].part_of_sequence_p = true; | |
3107 | } | |
1b61190c | 3108 | else |
3109 | lab = gen_label_rtx (); | |
73401833 | 3110 | pool_vector[pool_size].mode = mode; |
3111 | pool_vector[pool_size].label = lab; | |
ecf6ff7c | 3112 | pool_vector[pool_size].wend = NULL_RTX; |
c5ea35c4 | 3113 | pool_vector[pool_size].part_of_sequence_p = (lab == 0); |
ecf6ff7c | 3114 | if (lab && pool_window_label) |
3115 | { | |
3116 | newref = gen_rtx_LABEL_REF (VOIDmode, pool_window_label); | |
3117 | ref = pool_vector[pool_window_last].wend; | |
3118 | LABEL_NEXTREF (newref) = ref; | |
3119 | pool_vector[pool_window_last].wend = newref; | |
3120 | } | |
3121 | if (lab) | |
3122 | pool_window_label = lab; | |
3123 | pool_window_last = pool_size; | |
73401833 | 3124 | pool_size++; |
3125 | return lab; | |
97595bfd | 3126 | } |
1504eb7e | 3127 | |
51dade95 | 3128 | /* Output the literal table. START, if nonzero, is the first instruction |
3129 | this table is needed for, and also indicates that there is at least one | |
3130 | casesi_worker_2 instruction; We have to emit the operand3 labels from | |
3131 | these insns at a 4-byte aligned position. BARRIER is the barrier | |
3132 | after which we are to place the table. */ | |
47c009e5 | 3133 | |
47c009e5 | 3134 | static void |
51dade95 | 3135 | dump_table (rtx start, rtx barrier) |
47c009e5 | 3136 | { |
51dade95 | 3137 | rtx scan = barrier; |
97595bfd | 3138 | int i; |
97595bfd | 3139 | int need_align = 1; |
ecf6ff7c | 3140 | rtx lab, ref; |
d1b048f6 | 3141 | int have_df = 0; |
47c009e5 | 3142 | |
1504eb7e | 3143 | /* Do two passes, first time dump out the HI sized constants. */ |
47c009e5 | 3144 | |
97595bfd | 3145 | for (i = 0; i < pool_size; i++) |
47c009e5 | 3146 | { |
73401833 | 3147 | pool_node *p = &pool_vector[i]; |
3148 | ||
97595bfd | 3149 | if (p->mode == HImode) |
3150 | { | |
3151 | if (need_align) | |
3152 | { | |
3153 | scan = emit_insn_after (gen_align_2 (), scan); | |
3154 | need_align = 0; | |
3155 | } | |
ecf6ff7c | 3156 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
3157 | scan = emit_label_after (lab, scan); | |
3158 | scan = emit_insn_after (gen_consttable_2 (p->value, const0_rtx), | |
3159 | scan); | |
3160 | for (ref = p->wend; ref; ref = LABEL_NEXTREF (ref)) | |
3161 | { | |
3162 | lab = XEXP (ref, 0); | |
3163 | scan = emit_insn_after (gen_consttable_window_end (lab), scan); | |
3164 | } | |
97595bfd | 3165 | } |
d1b048f6 | 3166 | else if (p->mode == DFmode) |
3167 | have_df = 1; | |
47c009e5 | 3168 | } |
73401833 | 3169 | |
97595bfd | 3170 | need_align = 1; |
47c009e5 | 3171 | |
51dade95 | 3172 | if (start) |
3173 | { | |
3174 | scan = emit_insn_after (gen_align_4 (), scan); | |
3175 | need_align = 0; | |
3176 | for (; start != barrier; start = NEXT_INSN (start)) | |
3177 | if (GET_CODE (start) == INSN | |
3178 | && recog_memoized (start) == CODE_FOR_casesi_worker_2) | |
3179 | { | |
3180 | rtx src = SET_SRC (XVECEXP (PATTERN (start), 0, 0)); | |
3181 | rtx lab = XEXP (XVECEXP (src, 0, 3), 0); | |
3182 | ||
3183 | scan = emit_label_after (lab, scan); | |
3184 | } | |
3185 | } | |
d1b048f6 | 3186 | if (TARGET_FMOVD && TARGET_ALIGN_DOUBLE && have_df) |
87e19636 | 3187 | { |
3188 | rtx align_insn = NULL_RTX; | |
3189 | ||
3190 | scan = emit_label_after (gen_label_rtx (), scan); | |
3191 | scan = emit_insn_after (gen_align_log (GEN_INT (3)), scan); | |
3192 | need_align = 0; | |
3193 | ||
3194 | for (i = 0; i < pool_size; i++) | |
3195 | { | |
3196 | pool_node *p = &pool_vector[i]; | |
3197 | ||
3198 | switch (p->mode) | |
3199 | { | |
3200 | case HImode: | |
3201 | break; | |
3202 | case SImode: | |
3203 | case SFmode: | |
c5ea35c4 | 3204 | if (align_insn && !p->part_of_sequence_p) |
87e19636 | 3205 | { |
3206 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) | |
3207 | emit_label_before (lab, align_insn); | |
3208 | emit_insn_before (gen_consttable_4 (p->value, const0_rtx), | |
3209 | align_insn); | |
3210 | for (ref = p->wend; ref; ref = LABEL_NEXTREF (ref)) | |
3211 | { | |
3212 | lab = XEXP (ref, 0); | |
3213 | emit_insn_before (gen_consttable_window_end (lab), | |
d767e27e | 3214 | align_insn); |
87e19636 | 3215 | } |
3216 | delete_insn (align_insn); | |
3217 | align_insn = NULL_RTX; | |
3218 | continue; | |
3219 | } | |
3220 | else | |
3221 | { | |
3222 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) | |
3223 | scan = emit_label_after (lab, scan); | |
3224 | scan = emit_insn_after (gen_consttable_4 (p->value, | |
3225 | const0_rtx), scan); | |
3226 | need_align = ! need_align; | |
3227 | } | |
3228 | break; | |
3229 | case DFmode: | |
87e19636 | 3230 | if (need_align) |
3231 | { | |
3232 | scan = emit_insn_after (gen_align_log (GEN_INT (3)), scan); | |
3233 | align_insn = scan; | |
3234 | need_align = 0; | |
3235 | } | |
d1b048f6 | 3236 | case DImode: |
87e19636 | 3237 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
3238 | scan = emit_label_after (lab, scan); | |
3239 | scan = emit_insn_after (gen_consttable_8 (p->value, const0_rtx), | |
3240 | scan); | |
3241 | break; | |
3242 | default: | |
1a6a7a27 | 3243 | gcc_unreachable (); |
87e19636 | 3244 | } |
3245 | ||
3246 | if (p->mode != HImode) | |
3247 | { | |
3248 | for (ref = p->wend; ref; ref = LABEL_NEXTREF (ref)) | |
3249 | { | |
3250 | lab = XEXP (ref, 0); | |
3251 | scan = emit_insn_after (gen_consttable_window_end (lab), | |
3252 | scan); | |
3253 | } | |
3254 | } | |
3255 | } | |
3256 | ||
3257 | pool_size = 0; | |
3258 | } | |
9e7454d0 | 3259 | |
97595bfd | 3260 | for (i = 0; i < pool_size; i++) |
47c009e5 | 3261 | { |
73401833 | 3262 | pool_node *p = &pool_vector[i]; |
47c009e5 | 3263 | |
97595bfd | 3264 | switch (p->mode) |
47c009e5 | 3265 | { |
97595bfd | 3266 | case HImode: |
3267 | break; | |
3268 | case SImode: | |
00f87557 | 3269 | case SFmode: |
97595bfd | 3270 | if (need_align) |
47c009e5 | 3271 | { |
97595bfd | 3272 | need_align = 0; |
f3d93547 | 3273 | scan = emit_label_after (gen_label_rtx (), scan); |
97595bfd | 3274 | scan = emit_insn_after (gen_align_4 (), scan); |
47c009e5 | 3275 | } |
ecf6ff7c | 3276 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
3277 | scan = emit_label_after (lab, scan); | |
3278 | scan = emit_insn_after (gen_consttable_4 (p->value, const0_rtx), | |
3279 | scan); | |
97595bfd | 3280 | break; |
00f87557 | 3281 | case DFmode: |
97595bfd | 3282 | case DImode: |
3283 | if (need_align) | |
3284 | { | |
3285 | need_align = 0; | |
f3d93547 | 3286 | scan = emit_label_after (gen_label_rtx (), scan); |
97595bfd | 3287 | scan = emit_insn_after (gen_align_4 (), scan); |
3288 | } | |
ecf6ff7c | 3289 | for (lab = p->label; lab; lab = LABEL_REFS (lab)) |
3290 | scan = emit_label_after (lab, scan); | |
3291 | scan = emit_insn_after (gen_consttable_8 (p->value, const0_rtx), | |
3292 | scan); | |
97595bfd | 3293 | break; |
3294 | default: | |
1a6a7a27 | 3295 | gcc_unreachable (); |
47c009e5 | 3296 | } |
ecf6ff7c | 3297 | |
3298 | if (p->mode != HImode) | |
3299 | { | |
3300 | for (ref = p->wend; ref; ref = LABEL_NEXTREF (ref)) | |
3301 | { | |
3302 | lab = XEXP (ref, 0); | |
3303 | scan = emit_insn_after (gen_consttable_window_end (lab), scan); | |
3304 | } | |
3305 | } | |
47c009e5 | 3306 | } |
47c009e5 | 3307 | |
97595bfd | 3308 | scan = emit_insn_after (gen_consttable_end (), scan); |
3309 | scan = emit_barrier_after (scan); | |
3310 | pool_size = 0; | |
ecf6ff7c | 3311 | pool_window_label = NULL_RTX; |
3312 | pool_window_last = 0; | |
97595bfd | 3313 | } |
47c009e5 | 3314 | |
808a491c | 3315 | /* Return nonzero if constant would be an ok source for a |
73401833 | 3316 | mov.w instead of a mov.l. */ |
47c009e5 | 3317 | |
1504eb7e | 3318 | static int |
04f04b72 | 3319 | hi_const (rtx src) |
97595bfd | 3320 | { |
73401833 | 3321 | return (GET_CODE (src) == CONST_INT |
3322 | && INTVAL (src) >= -32768 | |
3323 | && INTVAL (src) <= 32767); | |
47c009e5 | 3324 | } |
3325 | ||
808a491c | 3326 | /* Nonzero if the insn is a move instruction which needs to be fixed. */ |
73401833 | 3327 | |
3328 | /* ??? For a DImode/DFmode moves, we don't need to fix it if each half of the | |
af2c1324 | 3329 | CONST_DOUBLE input value is CONST_OK_FOR_I08. For a SFmode move, we don't |
3330 | need to fix it if the input value is CONST_OK_FOR_I08. */ | |
73401833 | 3331 | |
3332 | static int | |
04f04b72 | 3333 | broken_move (rtx insn) |
47c009e5 | 3334 | { |
00f87557 | 3335 | if (GET_CODE (insn) == INSN) |
3336 | { | |
3337 | rtx pat = PATTERN (insn); | |
3338 | if (GET_CODE (pat) == PARALLEL) | |
3339 | pat = XVECEXP (pat, 0, 0); | |
3340 | if (GET_CODE (pat) == SET | |
3341 | /* We can load any 8 bit value if we don't care what the high | |
3342 | order bits end up as. */ | |
3343 | && GET_MODE (SET_DEST (pat)) != QImode | |
393fac7b | 3344 | && (CONSTANT_P (SET_SRC (pat)) |
3345 | /* Match mova_const. */ | |
3346 | || (GET_CODE (SET_SRC (pat)) == UNSPEC | |
3347 | && XINT (SET_SRC (pat), 1) == UNSPEC_MOVA | |
3348 | && GET_CODE (XVECEXP (SET_SRC (pat), 0, 0)) == CONST)) | |
87ed74ef | 3349 | && ! (TARGET_SH2E |
8ded0752 | 3350 | && GET_CODE (SET_SRC (pat)) == CONST_DOUBLE |
00f87557 | 3351 | && (fp_zero_operand (SET_SRC (pat)) |
3352 | || fp_one_operand (SET_SRC (pat))) | |
4ea39690 | 3353 | /* ??? If this is a -m4 or -m4-single compilation, in general |
3354 | we don't know the current setting of fpscr, so disable fldi. | |
3355 | There is an exception if this was a register-register move | |
3356 | before reload - and hence it was ascertained that we have | |
3357 | single precision setting - and in a post-reload optimization | |
3358 | we changed this to do a constant load. In that case | |
3359 | we don't have an r0 clobber, hence we must use fldi. */ | |
3360 | && (! TARGET_SH4 || TARGET_FMOVD | |
3361 | || (GET_CODE (XEXP (XVECEXP (PATTERN (insn), 0, 2), 0)) | |
3362 | == SCRATCH)) | |
00f87557 | 3363 | && GET_CODE (SET_DEST (pat)) == REG |
e2142b04 | 3364 | && FP_REGISTER_P (REGNO (SET_DEST (pat)))) |
7105fb72 | 3365 | && ! (TARGET_SH2A |
3366 | && GET_MODE (SET_DEST (pat)) == SImode | |
3367 | && GET_CODE (SET_SRC (pat)) == CONST_INT | |
3368 | && CONST_OK_FOR_I20 (INTVAL (SET_SRC (pat)))) | |
00f87557 | 3369 | && (GET_CODE (SET_SRC (pat)) != CONST_INT |
af2c1324 | 3370 | || ! CONST_OK_FOR_I08 (INTVAL (SET_SRC (pat))))) |
00f87557 | 3371 | return 1; |
3372 | } | |
f3d93547 | 3373 | |
73401833 | 3374 | return 0; |
47c009e5 | 3375 | } |
47c009e5 | 3376 | |
8ded0752 | 3377 | static int |
04f04b72 | 3378 | mova_p (rtx insn) |
8ded0752 | 3379 | { |
3380 | return (GET_CODE (insn) == INSN | |
3381 | && GET_CODE (PATTERN (insn)) == SET | |
3382 | && GET_CODE (SET_SRC (PATTERN (insn))) == UNSPEC | |
393fac7b | 3383 | && XINT (SET_SRC (PATTERN (insn)), 1) == UNSPEC_MOVA |
3384 | /* Don't match mova_const. */ | |
dc115ee8 | 3385 | && GET_CODE (XVECEXP (SET_SRC (PATTERN (insn)), 0, 0)) == LABEL_REF); |
8ded0752 | 3386 | } |
3387 | ||
51dade95 | 3388 | /* Fix up a mova from a switch that went out of range. */ |
3389 | static void | |
3390 | fixup_mova (rtx mova) | |
3391 | { | |
3392 | if (! flag_pic) | |
3393 | { | |
3394 | SET_SRC (PATTERN (mova)) = XVECEXP (SET_SRC (PATTERN (mova)), 0, 0); | |
3395 | INSN_CODE (mova) = -1; | |
3396 | } | |
3397 | else | |
3398 | { | |
3399 | rtx worker = mova; | |
3400 | rtx lab = gen_label_rtx (); | |
3401 | rtx wpat, wpat0, wpat1, wsrc, diff; | |
3402 | ||
3403 | do | |
3404 | { | |
3405 | worker = NEXT_INSN (worker); | |
1a6a7a27 | 3406 | gcc_assert (worker |
3407 | && GET_CODE (worker) != CODE_LABEL | |
3408 | && GET_CODE (worker) != JUMP_INSN); | |
51dade95 | 3409 | } while (recog_memoized (worker) != CODE_FOR_casesi_worker_1); |
3410 | wpat = PATTERN (worker); | |
3411 | wpat0 = XVECEXP (wpat, 0, 0); | |
3412 | wpat1 = XVECEXP (wpat, 0, 1); | |
3413 | wsrc = SET_SRC (wpat0); | |
3414 | PATTERN (worker) = (gen_casesi_worker_2 | |
3415 | (SET_DEST (wpat0), XVECEXP (wsrc, 0, 1), | |
3416 | XEXP (XVECEXP (wsrc, 0, 2), 0), lab, | |
3417 | XEXP (wpat1, 0))); | |
3418 | INSN_CODE (worker) = -1; | |
3419 | diff = gen_rtx_MINUS (Pmode, XVECEXP (SET_SRC (PATTERN (mova)), 0, 0), | |
3420 | gen_rtx_LABEL_REF (Pmode, lab)); | |
3421 | diff = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, diff), UNSPEC_PIC); | |
3422 | SET_SRC (PATTERN (mova)) = gen_rtx_CONST (Pmode, diff); | |
3423 | INSN_CODE (mova) = -1; | |
3424 | } | |
3425 | } | |
3426 | ||
73401833 | 3427 | /* Find the last barrier from insn FROM which is close enough to hold the |
3428 | constant pool. If we can't find one, then create one near the end of | |
3429 | the range. */ | |
47c009e5 | 3430 | |
73401833 | 3431 | static rtx |
04f04b72 | 3432 | find_barrier (int num_mova, rtx mova, rtx from) |
47c009e5 | 3433 | { |
97595bfd | 3434 | int count_si = 0; |
3435 | int count_hi = 0; | |
3436 | int found_hi = 0; | |
3437 | int found_si = 0; | |
87e19636 | 3438 | int found_di = 0; |
9eaab178 | 3439 | int hi_align = 2; |
3440 | int si_align = 2; | |
8ded0752 | 3441 | int leading_mova = num_mova; |
87ed74ef | 3442 | rtx barrier_before_mova = 0, found_barrier = 0, good_barrier = 0; |
37ad966e | 3443 | int si_limit; |
3444 | int hi_limit; | |
73401833 | 3445 | |
3446 | /* For HImode: range is 510, add 4 because pc counts from address of | |
3447 | second instruction after this one, subtract 2 for the jump instruction | |
9629ee1b | 3448 | that we may need to emit before the table, subtract 2 for the instruction |
3449 | that fills the jump delay slot (in very rare cases, reorg will take an | |
3450 | instruction from after the constant pool or will leave the delay slot | |
3451 | empty). This gives 510. | |
73401833 | 3452 | For SImode: range is 1020, add 4 because pc counts from address of |
3453 | second instruction after this one, subtract 2 in case pc is 2 byte | |
3454 | aligned, subtract 2 for the jump instruction that we may need to emit | |
9629ee1b | 3455 | before the table, subtract 2 for the instruction that fills the jump |
3456 | delay slot. This gives 1018. */ | |
37ad966e | 3457 | |
8ded0752 | 3458 | /* The branch will always be shortened now that the reference address for |
ad87de1e | 3459 | forward branches is the successor address, thus we need no longer make |
8ded0752 | 3460 | adjustments to the [sh]i_limit for -O0. */ |
37ad966e | 3461 | |
8ded0752 | 3462 | si_limit = 1018; |
3463 | hi_limit = 510; | |
e6c94d0d | 3464 | |
37ad966e | 3465 | while (from && count_si < si_limit && count_hi < hi_limit) |
97595bfd | 3466 | { |
9eaab178 | 3467 | int inc = get_attr_length (from); |
3468 | int new_align = 1; | |
8ded0752 | 3469 | |
9eaab178 | 3470 | if (GET_CODE (from) == CODE_LABEL) |
887884c5 | 3471 | { |
3472 | if (optimize) | |
3473 | new_align = 1 << label_to_alignment (from); | |
3474 | else if (GET_CODE (prev_nonnote_insn (from)) == BARRIER) | |
3475 | new_align = 1 << barrier_align (from); | |
3476 | else | |
3477 | new_align = 1; | |
3478 | inc = 0; | |
3479 | } | |
73401833 | 3480 | |
97595bfd | 3481 | if (GET_CODE (from) == BARRIER) |
8ded0752 | 3482 | { |
9eaab178 | 3483 | |
8ded0752 | 3484 | found_barrier = from; |
9eaab178 | 3485 | |
ad87de1e | 3486 | /* If we are at the end of the function, or in front of an alignment |
8ded0752 | 3487 | instruction, we need not insert an extra alignment. We prefer |
3488 | this kind of barrier. */ | |
9eaab178 | 3489 | if (barrier_align (from) > 2) |
8ded0752 | 3490 | good_barrier = from; |
3491 | } | |
73401833 | 3492 | |
73401833 | 3493 | if (broken_move (from)) |
97595bfd | 3494 | { |
8ded0752 | 3495 | rtx pat, src, dst; |
3496 | enum machine_mode mode; | |
3497 | ||
3498 | pat = PATTERN (from); | |
3499 | if (GET_CODE (pat) == PARALLEL) | |
3500 | pat = XVECEXP (pat, 0, 0); | |
3501 | src = SET_SRC (pat); | |
3502 | dst = SET_DEST (pat); | |
3503 | mode = GET_MODE (dst); | |
37ad966e | 3504 | |
3505 | /* We must explicitly check the mode, because sometimes the | |
3506 | front end will generate code to load unsigned constants into | |
3507 | HImode targets without properly sign extending them. */ | |
1b61190c | 3508 | if (mode == HImode |
3509 | || (mode == SImode && hi_const (src) && REGNO (dst) != FPUL_REG)) | |
22f26681 | 3510 | { |
8ded0752 | 3511 | found_hi += 2; |
22f26681 | 3512 | /* We put the short constants before the long constants, so |
3513 | we must count the length of short constants in the range | |
3514 | for the long constants. */ | |
3515 | /* ??? This isn't optimal, but is easy to do. */ | |
8ded0752 | 3516 | si_limit -= 2; |
22f26681 | 3517 | } |
97595bfd | 3518 | else |
8ded0752 | 3519 | { |
87e19636 | 3520 | /* We dump DF/DI constants before SF/SI ones, because |
3521 | the limit is the same, but the alignment requirements | |
3522 | are higher. We may waste up to 4 additional bytes | |
3523 | for alignment, and the DF/DI constant may have | |
ada8cf8b | 3524 | another SF/SI constant placed before it. */ |
87e19636 | 3525 | if (TARGET_SHCOMPACT |
3526 | && ! found_di | |
3527 | && (mode == DFmode || mode == DImode)) | |
3528 | { | |
3529 | found_di = 1; | |
3530 | si_limit -= 8; | |
3531 | } | |
9eaab178 | 3532 | while (si_align > 2 && found_si + si_align - 2 > count_si) |
3533 | si_align >>= 1; | |
8ded0752 | 3534 | if (found_si > count_si) |
3535 | count_si = found_si; | |
3536 | found_si += GET_MODE_SIZE (mode); | |
3537 | if (num_mova) | |
3538 | si_limit -= GET_MODE_SIZE (mode); | |
3539 | } | |
97595bfd | 3540 | } |
335e16a0 | 3541 | |
9eaab178 | 3542 | if (mova_p (from)) |
8ded0752 | 3543 | { |
3544 | if (! num_mova++) | |
3545 | { | |
3546 | leading_mova = 0; | |
3547 | mova = from; | |
3548 | barrier_before_mova = good_barrier ? good_barrier : found_barrier; | |
3549 | } | |
3550 | if (found_si > count_si) | |
3551 | count_si = found_si; | |
3552 | } | |
335e16a0 | 3553 | else if (GET_CODE (from) == JUMP_INSN |
3554 | && (GET_CODE (PATTERN (from)) == ADDR_VEC | |
3555 | || GET_CODE (PATTERN (from)) == ADDR_DIFF_VEC)) | |
8ded0752 | 3556 | { |
3557 | if (num_mova) | |
3558 | num_mova--; | |
2a6493a5 | 3559 | if (barrier_align (next_real_insn (from)) == align_jumps_log) |
8ded0752 | 3560 | { |
3398e91d | 3561 | /* We have just passed the barrier in front of the |
bd967f88 | 3562 | ADDR_DIFF_VEC, which is stored in found_barrier. Since |
3563 | the ADDR_DIFF_VEC is accessed as data, just like our pool | |
3564 | constants, this is a good opportunity to accommodate what | |
3565 | we have gathered so far. | |
8ded0752 | 3566 | If we waited any longer, we could end up at a barrier in |
3567 | front of code, which gives worse cache usage for separated | |
3568 | instruction / data caches. */ | |
bd967f88 | 3569 | good_barrier = found_barrier; |
8ded0752 | 3570 | break; |
3571 | } | |
bd967f88 | 3572 | else |
3573 | { | |
3574 | rtx body = PATTERN (from); | |
3575 | inc = XVECLEN (body, 1) * GET_MODE_SIZE (GET_MODE (body)); | |
3576 | } | |
8ded0752 | 3577 | } |
0d93aeb0 | 3578 | /* For the SH1, we generate alignments even after jumps-around-jumps. */ |
3579 | else if (GET_CODE (from) == JUMP_INSN | |
3580 | && ! TARGET_SH2 | |
3581 | && ! TARGET_SMALLCODE) | |
3582 | new_align = 4; | |
335e16a0 | 3583 | |
97595bfd | 3584 | if (found_si) |
9eaab178 | 3585 | { |
0d93aeb0 | 3586 | count_si += inc; |
9eaab178 | 3587 | if (new_align > si_align) |
3588 | { | |
b7dbbdb2 | 3589 | si_limit -= (count_si - 1) & (new_align - si_align); |
9eaab178 | 3590 | si_align = new_align; |
3591 | } | |
b7dbbdb2 | 3592 | count_si = (count_si + new_align - 1) & -new_align; |
9eaab178 | 3593 | } |
97595bfd | 3594 | if (found_hi) |
9eaab178 | 3595 | { |
0d93aeb0 | 3596 | count_hi += inc; |
9eaab178 | 3597 | if (new_align > hi_align) |
3598 | { | |
b7dbbdb2 | 3599 | hi_limit -= (count_hi - 1) & (new_align - hi_align); |
9eaab178 | 3600 | hi_align = new_align; |
3601 | } | |
b7dbbdb2 | 3602 | count_hi = (count_hi + new_align - 1) & -new_align; |
9eaab178 | 3603 | } |
97595bfd | 3604 | from = NEXT_INSN (from); |
3605 | } | |
3606 | ||
8ded0752 | 3607 | if (num_mova) |
17bec211 | 3608 | { |
3609 | if (leading_mova) | |
3610 | { | |
3611 | /* Try as we might, the leading mova is out of range. Change | |
3612 | it into a load (which will become a pcload) and retry. */ | |
51dade95 | 3613 | fixup_mova (mova); |
17bec211 | 3614 | return find_barrier (0, 0, mova); |
3615 | } | |
3616 | else | |
3617 | { | |
3618 | /* Insert the constant pool table before the mova instruction, | |
3619 | to prevent the mova label reference from going out of range. */ | |
3620 | from = mova; | |
3621 | good_barrier = found_barrier = barrier_before_mova; | |
3622 | } | |
3623 | } | |
335e16a0 | 3624 | |
8ded0752 | 3625 | if (found_barrier) |
3626 | { | |
9eaab178 | 3627 | if (good_barrier && next_real_insn (found_barrier)) |
8ded0752 | 3628 | found_barrier = good_barrier; |
8ded0752 | 3629 | } |
3630 | else | |
47c009e5 | 3631 | { |
73401833 | 3632 | /* We didn't find a barrier in time to dump our stuff, |
3633 | so we'll make one. */ | |
97595bfd | 3634 | rtx label = gen_label_rtx (); |
73401833 | 3635 | |
92e80318 | 3636 | /* If we exceeded the range, then we must back up over the last |
3637 | instruction we looked at. Otherwise, we just need to undo the | |
3638 | NEXT_INSN at the end of the loop. */ | |
37ad966e | 3639 | if (count_hi > hi_limit || count_si > si_limit) |
92e80318 | 3640 | from = PREV_INSN (PREV_INSN (from)); |
3641 | else | |
3642 | from = PREV_INSN (from); | |
3643 | ||
73401833 | 3644 | /* Walk back to be just before any jump or label. |
3645 | Putting it before a label reduces the number of times the branch | |
3646 | around the constant pool table will be hit. Putting it before | |
3647 | a jump makes it more likely that the bra delay slot will be | |
3648 | filled. */ | |
3649 | while (GET_CODE (from) == JUMP_INSN || GET_CODE (from) == NOTE | |
2e3d4844 | 3650 | || GET_CODE (from) == CODE_LABEL) |
73401833 | 3651 | from = PREV_INSN (from); |
3652 | ||
97595bfd | 3653 | from = emit_jump_insn_after (gen_jump (label), from); |
3654 | JUMP_LABEL (from) = label; | |
bdb9edaa | 3655 | LABEL_NUSES (label) = 1; |
97595bfd | 3656 | found_barrier = emit_barrier_after (from); |
3657 | emit_label_after (label, found_barrier); | |
47c009e5 | 3658 | } |
47c009e5 | 3659 | |
73401833 | 3660 | return found_barrier; |
97595bfd | 3661 | } |
47c009e5 | 3662 | |
881ffd32 | 3663 | /* If the instruction INSN is implemented by a special function, and we can |
3664 | positively find the register that is used to call the sfunc, and this | |
3665 | register is not used anywhere else in this instruction - except as the | |
3666 | destination of a set, return this register; else, return 0. */ | |
8ded0752 | 3667 | rtx |
04f04b72 | 3668 | sfunc_uses_reg (rtx insn) |
881ffd32 | 3669 | { |
3670 | int i; | |
3671 | rtx pattern, part, reg_part, reg; | |
3672 | ||
3673 | if (GET_CODE (insn) != INSN) | |
3674 | return 0; | |
3675 | pattern = PATTERN (insn); | |
3676 | if (GET_CODE (pattern) != PARALLEL || get_attr_type (insn) != TYPE_SFUNC) | |
3677 | return 0; | |
3678 | ||
3679 | for (reg_part = 0, i = XVECLEN (pattern, 0) - 1; i >= 1; i--) | |
3680 | { | |
3681 | part = XVECEXP (pattern, 0, i); | |
8ded0752 | 3682 | if (GET_CODE (part) == USE && GET_MODE (XEXP (part, 0)) == SImode) |
881ffd32 | 3683 | reg_part = part; |
3684 | } | |
3685 | if (! reg_part) | |
3686 | return 0; | |
3687 | reg = XEXP (reg_part, 0); | |
3688 | for (i = XVECLEN (pattern, 0) - 1; i >= 0; i--) | |
3689 | { | |
3690 | part = XVECEXP (pattern, 0, i); | |
1b61190c | 3691 | if (part == reg_part || GET_CODE (part) == CLOBBER) |
881ffd32 | 3692 | continue; |
3693 | if (reg_mentioned_p (reg, ((GET_CODE (part) == SET | |
3694 | && GET_CODE (SET_DEST (part)) == REG) | |
3695 | ? SET_SRC (part) : part))) | |
3696 | return 0; | |
3697 | } | |
3698 | return reg; | |
3699 | } | |
3700 | ||
90e84010 | 3701 | /* See if the only way in which INSN uses REG is by calling it, or by |
3702 | setting it while calling it. Set *SET to a SET rtx if the register | |
3703 | is set by INSN. */ | |
3704 | ||
3705 | static int | |
04f04b72 | 3706 | noncall_uses_reg (rtx reg, rtx insn, rtx *set) |
90e84010 | 3707 | { |
881ffd32 | 3708 | rtx pattern, reg2; |
90e84010 | 3709 | |
3710 | *set = NULL_RTX; | |
3711 | ||
881ffd32 | 3712 | reg2 = sfunc_uses_reg (insn); |
3713 | if (reg2 && REGNO (reg2) == REGNO (reg)) | |
3714 | { | |
3715 | pattern = single_set (insn); | |
3716 | if (pattern | |
3717 | && GET_CODE (SET_DEST (pattern)) == REG | |
3718 | && REGNO (reg) == REGNO (SET_DEST (pattern))) | |
3719 | *set = pattern; | |
3720 | return 0; | |
3721 | } | |
90e84010 | 3722 | if (GET_CODE (insn) != CALL_INSN) |
3723 | { | |
3724 | /* We don't use rtx_equal_p because we don't care if the mode is | |
3725 | different. */ | |
3726 | pattern = single_set (insn); | |
3727 | if (pattern | |
3728 | && GET_CODE (SET_DEST (pattern)) == REG | |
3729 | && REGNO (reg) == REGNO (SET_DEST (pattern))) | |
3730 | { | |
881ffd32 | 3731 | rtx par, part; |
3732 | int i; | |
3733 | ||
90e84010 | 3734 | *set = pattern; |
881ffd32 | 3735 | par = PATTERN (insn); |
3736 | if (GET_CODE (par) == PARALLEL) | |
3737 | for (i = XVECLEN (par, 0) - 1; i >= 0; i--) | |
3738 | { | |
3739 | part = XVECEXP (par, 0, i); | |
3740 | if (GET_CODE (part) != SET && reg_mentioned_p (reg, part)) | |
3741 | return 1; | |
3742 | } | |
3743 | return reg_mentioned_p (reg, SET_SRC (pattern)); | |
90e84010 | 3744 | } |
3745 | ||
3746 | return 1; | |
3747 | } | |
3748 | ||
3749 | pattern = PATTERN (insn); | |
3750 | ||
3751 | if (GET_CODE (pattern) == PARALLEL) | |
3752 | { | |
3753 | int i; | |
3754 | ||
3755 | for (i = XVECLEN (pattern, 0) - 1; i >= 1; i--) | |
3756 | if (reg_mentioned_p (reg, XVECEXP (pattern, 0, i))) | |
3757 | return 1; | |
3758 | pattern = XVECEXP (pattern, 0, 0); | |
3759 | } | |
3760 | ||
3761 | if (GET_CODE (pattern) == SET) | |
3762 | { | |
3763 | if (reg_mentioned_p (reg, SET_DEST (pattern))) | |
3764 | { | |
3765 | /* We don't use rtx_equal_p, because we don't care if the | |
3766 | mode is different. */ | |
3767 | if (GET_CODE (SET_DEST (pattern)) != REG | |
3768 | || REGNO (reg) != REGNO (SET_DEST (pattern))) | |
3769 | return 1; | |
3770 | ||
3771 | *set = pattern; | |
3772 | } | |
3773 | ||
3774 | pattern = SET_SRC (pattern); | |
3775 | } | |
3776 | ||
3777 | if (GET_CODE (pattern) != CALL | |
3778 | || GET_CODE (XEXP (pattern, 0)) != MEM | |
3779 | || ! rtx_equal_p (reg, XEXP (XEXP (pattern, 0), 0))) | |
3780 | return 1; | |
3781 | ||
3782 | return 0; | |
3783 | } | |
3784 | ||
8ded0752 | 3785 | /* Given a X, a pattern of an insn or a part of it, return a mask of used |
3786 | general registers. Bits 0..15 mean that the respective registers | |
3787 | are used as inputs in the instruction. Bits 16..31 mean that the | |
3788 | registers 0..15, respectively, are used as outputs, or are clobbered. | |
3789 | IS_DEST should be set to 16 if X is the destination of a SET, else to 0. */ | |
3790 | int | |
04f04b72 | 3791 | regs_used (rtx x, int is_dest) |
8ded0752 | 3792 | { |
3793 | enum rtx_code code; | |
d2ca078f | 3794 | const char *fmt; |
8ded0752 | 3795 | int i, used = 0; |
3796 | ||
3797 | if (! x) | |
3798 | return used; | |
3799 | code = GET_CODE (x); | |
3800 | switch (code) | |
3801 | { | |
3802 | case REG: | |
3803 | if (REGNO (x) < 16) | |
3804 | return (((1 << HARD_REGNO_NREGS (0, GET_MODE (x))) - 1) | |
3805 | << (REGNO (x) + is_dest)); | |
3806 | return 0; | |
3807 | case SUBREG: | |
3808 | { | |
3809 | rtx y = SUBREG_REG (x); | |
9e7454d0 | 3810 | |
8ded0752 | 3811 | if (GET_CODE (y) != REG) |
3812 | break; | |
3813 | if (REGNO (y) < 16) | |
3814 | return (((1 << HARD_REGNO_NREGS (0, GET_MODE (x))) - 1) | |
701e46d0 | 3815 | << (REGNO (y) + |
3816 | subreg_regno_offset (REGNO (y), | |
3817 | GET_MODE (y), | |
3818 | SUBREG_BYTE (x), | |
3819 | GET_MODE (x)) + is_dest)); | |
8ded0752 | 3820 | return 0; |
3821 | } | |
3822 | case SET: | |
3823 | return regs_used (SET_SRC (x), 0) | regs_used (SET_DEST (x), 16); | |
3824 | case RETURN: | |
3825 | /* If there was a return value, it must have been indicated with USE. */ | |
3826 | return 0x00ffff00; | |
3827 | case CLOBBER: | |
3828 | is_dest = 1; | |
3829 | break; | |
3830 | case MEM: | |
3831 | is_dest = 0; | |
3832 | break; | |
3833 | case CALL: | |
3834 | used |= 0x00ff00f0; | |
3835 | break; | |
b7dbbdb2 | 3836 | default: |
3837 | break; | |
8ded0752 | 3838 | } |
3839 | ||
3840 | fmt = GET_RTX_FORMAT (code); | |
3841 | ||
3842 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
3843 | { | |
3844 | if (fmt[i] == 'E') | |
3845 | { | |
3846 | register int j; | |
3847 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
3848 | used |= regs_used (XVECEXP (x, i, j), is_dest); | |
3849 | } | |
3850 | else if (fmt[i] == 'e') | |
3851 | used |= regs_used (XEXP (x, i), is_dest); | |
3852 | } | |
3853 | return used; | |
3854 | } | |
3855 | ||
3856 | /* Create an instruction that prevents redirection of a conditional branch | |
ad87de1e | 3857 | to the destination of the JUMP with address ADDR. |
8ded0752 | 3858 | If the branch needs to be implemented as an indirect jump, try to find |
3859 | a scratch register for it. | |
3860 | If NEED_BLOCK is 0, don't do anything unless we need a scratch register. | |
3861 | If any preceding insn that doesn't fit into a delay slot is good enough, | |
3862 | pass 1. Pass 2 if a definite blocking insn is needed. | |
3863 | -1 is used internally to avoid deep recursion. | |
3864 | If a blocking instruction is made or recognized, return it. */ | |
9e7454d0 | 3865 | |
8ded0752 | 3866 | static rtx |
04f04b72 | 3867 | gen_block_redirect (rtx jump, int addr, int need_block) |
8ded0752 | 3868 | { |
3869 | int dead = 0; | |
3870 | rtx prev = prev_nonnote_insn (jump); | |
3871 | rtx dest; | |
3872 | ||
3873 | /* First, check if we already have an instruction that satisfies our need. */ | |
3874 | if (prev && GET_CODE (prev) == INSN && ! INSN_DELETED_P (prev)) | |
3875 | { | |
3876 | if (INSN_CODE (prev) == CODE_FOR_indirect_jump_scratch) | |
3877 | return prev; | |
3878 | if (GET_CODE (PATTERN (prev)) == USE | |
3879 | || GET_CODE (PATTERN (prev)) == CLOBBER | |
3880 | || get_attr_in_delay_slot (prev) == IN_DELAY_SLOT_YES) | |
3881 | prev = jump; | |
3882 | else if ((need_block &= ~1) < 0) | |
3883 | return prev; | |
3884 | else if (recog_memoized (prev) == CODE_FOR_block_branch_redirect) | |
3885 | need_block = 0; | |
3886 | } | |
b78739b6 | 3887 | if (GET_CODE (PATTERN (jump)) == RETURN) |
3888 | { | |
3889 | if (! need_block) | |
3890 | return prev; | |
3891 | /* Reorg even does nasty things with return insns that cause branches | |
3892 | to go out of range - see find_end_label and callers. */ | |
bcd9bd66 | 3893 | return emit_insn_before (gen_block_branch_redirect (const0_rtx) , jump); |
b78739b6 | 3894 | } |
8ded0752 | 3895 | /* We can't use JUMP_LABEL here because it might be undefined |
3896 | when not optimizing. */ | |
3897 | dest = XEXP (SET_SRC (PATTERN (jump)), 0); | |
3898 | /* If the branch is out of range, try to find a scratch register for it. */ | |
3899 | if (optimize | |
30435bf8 | 3900 | && (INSN_ADDRESSES (INSN_UID (dest)) - addr + (unsigned) 4092 |
3901 | > 4092 + 4098)) | |
8ded0752 | 3902 | { |
3903 | rtx scan; | |
3904 | /* Don't look for the stack pointer as a scratch register, | |
ad87de1e | 3905 | it would cause trouble if an interrupt occurred. */ |
8ded0752 | 3906 | unsigned try = 0x7fff, used; |
3907 | int jump_left = flag_expensive_optimizations + 1; | |
9e7454d0 | 3908 | |
8ded0752 | 3909 | /* It is likely that the most recent eligible instruction is wanted for |
3910 | the delay slot. Therefore, find out which registers it uses, and | |
3911 | try to avoid using them. */ | |
9e7454d0 | 3912 | |
b7dbbdb2 | 3913 | for (scan = jump; (scan = PREV_INSN (scan)); ) |
8ded0752 | 3914 | { |
3915 | enum rtx_code code; | |
3916 | ||
3917 | if (INSN_DELETED_P (scan)) | |
3918 | continue; | |
3919 | code = GET_CODE (scan); | |
3920 | if (code == CODE_LABEL || code == JUMP_INSN) | |
3921 | break; | |
3922 | if (code == INSN | |
3923 | && GET_CODE (PATTERN (scan)) != USE | |
3924 | && GET_CODE (PATTERN (scan)) != CLOBBER | |
3925 | && get_attr_in_delay_slot (scan) == IN_DELAY_SLOT_YES) | |
3926 | { | |
3927 | try &= ~regs_used (PATTERN (scan), 0); | |
3928 | break; | |
3929 | } | |
3930 | } | |
17bec211 | 3931 | for (used = dead = 0, scan = JUMP_LABEL (jump); |
3932 | (scan = NEXT_INSN (scan)); ) | |
8ded0752 | 3933 | { |
3934 | enum rtx_code code; | |
3935 | ||
3936 | if (INSN_DELETED_P (scan)) | |
3937 | continue; | |
3938 | code = GET_CODE (scan); | |
6720e96c | 3939 | if (INSN_P (scan)) |
8ded0752 | 3940 | { |
3941 | used |= regs_used (PATTERN (scan), 0); | |
3942 | if (code == CALL_INSN) | |
3943 | used |= regs_used (CALL_INSN_FUNCTION_USAGE (scan), 0); | |
3944 | dead |= (used >> 16) & ~used; | |
3945 | if (dead & try) | |
3946 | { | |
3947 | dead &= try; | |
3948 | break; | |
3949 | } | |
3950 | if (code == JUMP_INSN) | |
17bec211 | 3951 | { |
3952 | if (jump_left-- && simplejump_p (scan)) | |
3953 | scan = JUMP_LABEL (scan); | |
3954 | else | |
3955 | break; | |
3956 | } | |
8ded0752 | 3957 | } |
3958 | } | |
3959 | /* Mask out the stack pointer again, in case it was | |
3960 | the only 'free' register we have found. */ | |
3961 | dead &= 0x7fff; | |
3962 | } | |
3963 | /* If the immediate destination is still in range, check for possible | |
3964 | threading with a jump beyond the delay slot insn. | |
3965 | Don't check if we are called recursively; the jump has been or will be | |
ad87de1e | 3966 | checked in a different invocation then. */ |
9e7454d0 | 3967 | |
8ded0752 | 3968 | else if (optimize && need_block >= 0) |
3969 | { | |
3970 | rtx next = next_active_insn (next_active_insn (dest)); | |
3971 | if (next && GET_CODE (next) == JUMP_INSN | |
3972 | && GET_CODE (PATTERN (next)) == SET | |
adbcc7c4 | 3973 | && recog_memoized (next) == CODE_FOR_jump_compact) |
8ded0752 | 3974 | { |
3975 | dest = JUMP_LABEL (next); | |
3976 | if (dest | |
30435bf8 | 3977 | && (INSN_ADDRESSES (INSN_UID (dest)) - addr + (unsigned) 4092 |
47fc0706 | 3978 | > 4092 + 4098)) |
3979 | gen_block_redirect (next, INSN_ADDRESSES (INSN_UID (next)), -1); | |
8ded0752 | 3980 | } |
3981 | } | |
3982 | ||
3983 | if (dead) | |
3984 | { | |
7014838c | 3985 | rtx reg = gen_rtx_REG (SImode, exact_log2 (dead & -dead)); |
8ded0752 | 3986 | |
3987 | /* It would be nice if we could convert the jump into an indirect | |
ad87de1e | 3988 | jump / far branch right now, and thus exposing all constituent |
8ded0752 | 3989 | instructions to further optimization. However, reorg uses |
3990 | simplejump_p to determine if there is an unconditional jump where | |
3991 | it should try to schedule instructions from the target of the | |
3992 | branch; simplejump_p fails for indirect jumps even if they have | |
3993 | a JUMP_LABEL. */ | |
3994 | rtx insn = emit_insn_before (gen_indirect_jump_scratch | |
3995 | (reg, GEN_INT (INSN_UID (JUMP_LABEL (jump)))) | |
3996 | , jump); | |
2a6493a5 | 3997 | /* ??? We would like this to have the scope of the jump, but that |
3998 | scope will change when a delay slot insn of an inner scope is added. | |
3999 | Hence, after delay slot scheduling, we'll have to expect | |
4000 | NOTE_INSN_BLOCK_END notes between the indirect_jump_scratch and | |
4001 | the jump. */ | |
9e7454d0 | 4002 | |
2f49e3cb | 4003 | INSN_LOCATOR (insn) = INSN_LOCATOR (jump); |
8ded0752 | 4004 | INSN_CODE (insn) = CODE_FOR_indirect_jump_scratch; |
4005 | return insn; | |
4006 | } | |
4007 | else if (need_block) | |
4008 | /* We can't use JUMP_LABEL here because it might be undefined | |
4009 | when not optimizing. */ | |
4010 | return emit_insn_before (gen_block_branch_redirect | |
4011 | (GEN_INT (INSN_UID (XEXP (SET_SRC (PATTERN (jump)), 0)))) | |
4012 | , jump); | |
4013 | return prev; | |
4014 | } | |
4015 | ||
4016 | #define CONDJUMP_MIN -252 | |
4017 | #define CONDJUMP_MAX 262 | |
4018 | struct far_branch | |
4019 | { | |
4020 | /* A label (to be placed) in front of the jump | |
4021 | that jumps to our ultimate destination. */ | |
4022 | rtx near_label; | |
4023 | /* Where we are going to insert it if we cannot move the jump any farther, | |
4024 | or the jump itself if we have picked up an existing jump. */ | |
4025 | rtx insert_place; | |
4026 | /* The ultimate destination. */ | |
4027 | rtx far_label; | |
4028 | struct far_branch *prev; | |
4029 | /* If the branch has already been created, its address; | |
4030 | else the address of its first prospective user. */ | |
4031 | int address; | |
4032 | }; | |
4033 | ||
04f04b72 | 4034 | static void gen_far_branch (struct far_branch *); |
8ded0752 | 4035 | enum mdep_reorg_phase_e mdep_reorg_phase; |
26576132 | 4036 | static void |
04f04b72 | 4037 | gen_far_branch (struct far_branch *bp) |
8ded0752 | 4038 | { |
4039 | rtx insn = bp->insert_place; | |
4040 | rtx jump; | |
4041 | rtx label = gen_label_rtx (); | |
1a6a7a27 | 4042 | int ok; |
8ded0752 | 4043 | |
4044 | emit_label_after (label, insn); | |
4045 | if (bp->far_label) | |
4046 | { | |
4047 | jump = emit_jump_insn_after (gen_jump (bp->far_label), insn); | |
4048 | LABEL_NUSES (bp->far_label)++; | |
4049 | } | |
4050 | else | |
4051 | jump = emit_jump_insn_after (gen_return (), insn); | |
1b61190c | 4052 | /* Emit a barrier so that reorg knows that any following instructions |
4053 | are not reachable via a fall-through path. | |
1fc184ee | 4054 | But don't do this when not optimizing, since we wouldn't suppress the |
1b61190c | 4055 | alignment for the barrier then, and could end up with out-of-range |
4056 | pc-relative loads. */ | |
4057 | if (optimize) | |
4058 | emit_barrier_after (jump); | |
8ded0752 | 4059 | emit_label_after (bp->near_label, insn); |
4060 | JUMP_LABEL (jump) = bp->far_label; | |
1a6a7a27 | 4061 | ok = invert_jump (insn, label, 1); |
4062 | gcc_assert (ok); | |
4063 | ||
b78739b6 | 4064 | /* If we are branching around a jump (rather than a return), prevent |
4065 | reorg from using an insn from the jump target as the delay slot insn - | |
4066 | when reorg did this, it pessimized code (we rather hide the delay slot) | |
4067 | and it could cause branches to go out of range. */ | |
4068 | if (bp->far_label) | |
4069 | (emit_insn_after | |
4070 | (gen_stuff_delay_slot | |
4071 | (GEN_INT (INSN_UID (XEXP (SET_SRC (PATTERN (jump)), 0))), | |
4072 | GEN_INT (recog_memoized (insn) == CODE_FOR_branch_false)), | |
4073 | insn)); | |
8ded0752 | 4074 | /* Prevent reorg from undoing our splits. */ |
4075 | gen_block_redirect (jump, bp->address += 2, 2); | |
4076 | } | |
4077 | ||
8ded0752 | 4078 | /* Fix up ADDR_DIFF_VECs. */ |
4079 | void | |
04f04b72 | 4080 | fixup_addr_diff_vecs (rtx first) |
8ded0752 | 4081 | { |
4082 | rtx insn; | |
9eaab178 | 4083 | |
8ded0752 | 4084 | for (insn = first; insn; insn = NEXT_INSN (insn)) |
4085 | { | |
21022457 | 4086 | rtx vec_lab, pat, prev, prevpat, x, braf_label; |
8ded0752 | 4087 | |
4088 | if (GET_CODE (insn) != JUMP_INSN | |
4089 | || GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC) | |
4090 | continue; | |
4091 | pat = PATTERN (insn); | |
9eaab178 | 4092 | vec_lab = XEXP (XEXP (pat, 0), 0); |
8ded0752 | 4093 | |
9eaab178 | 4094 | /* Search the matching casesi_jump_2. */ |
4095 | for (prev = vec_lab; ; prev = PREV_INSN (prev)) | |
8ded0752 | 4096 | { |
9eaab178 | 4097 | if (GET_CODE (prev) != JUMP_INSN) |
4098 | continue; | |
4099 | prevpat = PATTERN (prev); | |
4100 | if (GET_CODE (prevpat) != PARALLEL || XVECLEN (prevpat, 0) != 2) | |
4101 | continue; | |
4102 | x = XVECEXP (prevpat, 0, 1); | |
4103 | if (GET_CODE (x) != USE) | |
4104 | continue; | |
4105 | x = XEXP (x, 0); | |
4106 | if (GET_CODE (x) == LABEL_REF && XEXP (x, 0) == vec_lab) | |
4107 | break; | |
8ded0752 | 4108 | } |
7105fb72 | 4109 | /* FIXME: This is a bug in the optimizer, but it seems harmless |
4110 | to just avoid panicing. */ | |
4111 | if (!prev) | |
4112 | continue; | |
21022457 | 4113 | |
4114 | /* Emit the reference label of the braf where it belongs, right after | |
4115 | the casesi_jump_2 (i.e. braf). */ | |
4116 | braf_label = XEXP (XEXP (SET_SRC (XVECEXP (prevpat, 0, 0)), 1), 0); | |
4117 | emit_label_after (braf_label, prev); | |
4118 | ||
9eaab178 | 4119 | /* Fix up the ADDR_DIF_VEC to be relative |
4120 | to the reference address of the braf. */ | |
21022457 | 4121 | XEXP (XEXP (pat, 0), 0) = braf_label; |
8ded0752 | 4122 | } |
8ded0752 | 4123 | } |
4124 | ||
9eaab178 | 4125 | /* BARRIER_OR_LABEL is either a BARRIER or a CODE_LABEL immediately following |
4126 | a barrier. Return the base 2 logarithm of the desired alignment. */ | |
8ded0752 | 4127 | int |
04f04b72 | 4128 | barrier_align (rtx barrier_or_label) |
8ded0752 | 4129 | { |
9eaab178 | 4130 | rtx next = next_real_insn (barrier_or_label), pat, prev; |
87ed74ef | 4131 | int slot, credit, jump_to_next = 0; |
9e7454d0 | 4132 | |
9eaab178 | 4133 | if (! next) |
8ded0752 | 4134 | return 0; |
8ded0752 | 4135 | |
9eaab178 | 4136 | pat = PATTERN (next); |
8ded0752 | 4137 | |
9eaab178 | 4138 | if (GET_CODE (pat) == ADDR_DIFF_VEC) |
4139 | return 2; | |
8ded0752 | 4140 | |
a5bd34f1 | 4141 | if (GET_CODE (pat) == UNSPEC_VOLATILE && XINT (pat, 1) == UNSPECV_ALIGN) |
9eaab178 | 4142 | /* This is a barrier in front of a constant table. */ |
4143 | return 0; | |
4144 | ||
4145 | prev = prev_real_insn (barrier_or_label); | |
4146 | if (GET_CODE (PATTERN (prev)) == ADDR_DIFF_VEC) | |
8ded0752 | 4147 | { |
9eaab178 | 4148 | pat = PATTERN (prev); |
4149 | /* If this is a very small table, we want to keep the alignment after | |
4150 | the table to the minimum for proper code alignment. */ | |
4151 | return ((TARGET_SMALLCODE | |
acda5809 | 4152 | || ((unsigned) XVECLEN (pat, 1) * GET_MODE_SIZE (GET_MODE (pat)) |
d767e27e | 4153 | <= (unsigned) 1 << (CACHE_LOG - 2))) |
2a6493a5 | 4154 | ? 1 << TARGET_SHMEDIA : align_jumps_log); |
8ded0752 | 4155 | } |
9eaab178 | 4156 | |
4157 | if (TARGET_SMALLCODE) | |
4158 | return 0; | |
4159 | ||
ecc15191 | 4160 | if (! TARGET_SH2 || ! optimize) |
2a6493a5 | 4161 | return align_jumps_log; |
9eaab178 | 4162 | |
1b61190c | 4163 | /* When fixing up pcloads, a constant table might be inserted just before |
4164 | the basic block that ends with the barrier. Thus, we can't trust the | |
4165 | instruction lengths before that. */ | |
4166 | if (mdep_reorg_phase > SH_FIXUP_PCLOAD) | |
8ded0752 | 4167 | { |
1b61190c | 4168 | /* Check if there is an immediately preceding branch to the insn beyond |
4169 | the barrier. We must weight the cost of discarding useful information | |
4170 | from the current cache line when executing this branch and there is | |
4171 | an alignment, against that of fetching unneeded insn in front of the | |
4172 | branch target when there is no alignment. */ | |
4173 | ||
9e7454d0 | 4174 | /* There are two delay_slot cases to consider. One is the simple case |
4175 | where the preceding branch is to the insn beyond the barrier (simple | |
4176 | delay slot filling), and the other is where the preceding branch has | |
4177 | a delay slot that is a duplicate of the insn after the barrier | |
4178 | (fill_eager_delay_slots) and the branch is to the insn after the insn | |
fd787fea | 4179 | after the barrier. */ |
4180 | ||
1b61190c | 4181 | /* PREV is presumed to be the JUMP_INSN for the barrier under |
4182 | investigation. Skip to the insn before it. */ | |
4183 | prev = prev_real_insn (prev); | |
4184 | ||
26576132 | 4185 | for (slot = 2, credit = (1 << (CACHE_LOG - 2)) + 2; |
1b61190c | 4186 | credit >= 0 && prev && GET_CODE (prev) == INSN; |
4187 | prev = prev_real_insn (prev)) | |
4188 | { | |
fd787fea | 4189 | jump_to_next = 0; |
1b61190c | 4190 | if (GET_CODE (PATTERN (prev)) == USE |
4191 | || GET_CODE (PATTERN (prev)) == CLOBBER) | |
4192 | continue; | |
4193 | if (GET_CODE (PATTERN (prev)) == SEQUENCE) | |
fd787fea | 4194 | { |
4195 | prev = XVECEXP (PATTERN (prev), 0, 1); | |
9e7454d0 | 4196 | if (INSN_UID (prev) == INSN_UID (next)) |
fd787fea | 4197 | { |
4198 | /* Delay slot was filled with insn at jump target. */ | |
4199 | jump_to_next = 1; | |
4200 | continue; | |
4201 | } | |
4202 | } | |
4203 | ||
1b61190c | 4204 | if (slot && |
4205 | get_attr_in_delay_slot (prev) == IN_DELAY_SLOT_YES) | |
4206 | slot = 0; | |
4207 | credit -= get_attr_length (prev); | |
4208 | } | |
4209 | if (prev | |
4210 | && GET_CODE (prev) == JUMP_INSN | |
24f975d6 | 4211 | && JUMP_LABEL (prev)) |
4212 | { | |
4213 | rtx x; | |
7628b8bf | 4214 | if (jump_to_next |
24f975d6 | 4215 | || next_real_insn (JUMP_LABEL (prev)) == next |
07eac433 | 4216 | /* If relax_delay_slots() decides NEXT was redundant |
4217 | with some previous instruction, it will have | |
4218 | redirected PREV's jump to the following insn. */ | |
67169380 | 4219 | || JUMP_LABEL (prev) == next_nonnote_insn (next) |
7628b8bf | 4220 | /* There is no upper bound on redundant instructions |
4221 | that might have been skipped, but we must not put an | |
24f975d6 | 4222 | alignment where none had been before. */ |
9e7454d0 | 4223 | || (x = (NEXT_INSN (NEXT_INSN (PREV_INSN (prev)))), |
4224 | (INSN_P (x) | |
24f975d6 | 4225 | && (INSN_CODE (x) == CODE_FOR_block_branch_redirect |
72cc4fb4 | 4226 | || INSN_CODE (x) == CODE_FOR_indirect_jump_scratch |
4227 | || INSN_CODE (x) == CODE_FOR_stuff_delay_slot)))) | |
24f975d6 | 4228 | { |
4229 | rtx pat = PATTERN (prev); | |
4230 | if (GET_CODE (pat) == PARALLEL) | |
7628b8bf | 4231 | pat = XVECEXP (pat, 0, 0); |
24f975d6 | 4232 | if (credit - slot >= (GET_CODE (SET_SRC (pat)) == PC ? 2 : 0)) |
4233 | return 0; | |
4234 | } | |
7628b8bf | 4235 | } |
9eaab178 | 4236 | } |
9e7454d0 | 4237 | |
2a6493a5 | 4238 | return align_jumps_log; |
8ded0752 | 4239 | } |
4240 | ||
afd5d59a | 4241 | /* If we are inside a phony loop, almost any kind of label can turn up as the |
4242 | first one in the loop. Aligning a braf label causes incorrect switch | |
4243 | destination addresses; we can detect braf labels because they are | |
4244 | followed by a BARRIER. | |
4245 | Applying loop alignment to small constant or switch tables is a waste | |
4246 | of space, so we suppress this too. */ | |
4247 | int | |
04f04b72 | 4248 | sh_loop_align (rtx label) |
afd5d59a | 4249 | { |
4250 | rtx next = label; | |
4251 | ||
4252 | do | |
4253 | next = next_nonnote_insn (next); | |
4254 | while (next && GET_CODE (next) == CODE_LABEL); | |
4255 | ||
4256 | if (! next | |
9204e736 | 4257 | || ! INSN_P (next) |
afd5d59a | 4258 | || GET_CODE (PATTERN (next)) == ADDR_DIFF_VEC |
4259 | || recog_memoized (next) == CODE_FOR_consttable_2) | |
4260 | return 0; | |
87e19636 | 4261 | |
2a6493a5 | 4262 | return align_loops_log; |
afd5d59a | 4263 | } |
4264 | ||
2efea8c0 | 4265 | /* Do a final pass over the function, just before delayed branch |
90e84010 | 4266 | scheduling. */ |
47c009e5 | 4267 | |
2efea8c0 | 4268 | static void |
04f04b72 | 4269 | sh_reorg (void) |
97595bfd | 4270 | { |
2efea8c0 | 4271 | rtx first, insn, mova = NULL_RTX; |
8ded0752 | 4272 | int num_mova; |
7014838c | 4273 | rtx r0_rtx = gen_rtx_REG (Pmode, 0); |
4274 | rtx r0_inc_rtx = gen_rtx_POST_INC (Pmode, r0_rtx); | |
73401833 | 4275 | |
2efea8c0 | 4276 | first = get_insns (); |
4277 | ||
5acd0683 | 4278 | /* We must split call insns before introducing `mova's. If we're |
4279 | optimizing, they'll have already been split. Otherwise, make | |
4280 | sure we don't split them too late. */ | |
4281 | if (! optimize) | |
161b1647 | 4282 | split_all_insns_noflow (); |
5acd0683 | 4283 | |
87e19636 | 4284 | if (TARGET_SHMEDIA) |
4285 | return; | |
4286 | ||
90e84010 | 4287 | /* If relaxing, generate pseudo-ops to associate function calls with |
4288 | the symbols they call. It does no harm to not generate these | |
4289 | pseudo-ops. However, when we can generate them, it enables to | |
4290 | linker to potentially relax the jsr to a bsr, and eliminate the | |
4291 | register load and, possibly, the constant pool entry. */ | |
4292 | ||
8ded0752 | 4293 | mdep_reorg_phase = SH_INSERT_USES_LABELS; |
90e84010 | 4294 | if (TARGET_RELAX) |
4295 | { | |
4296 | /* Remove all REG_LABEL notes. We want to use them for our own | |
4297 | purposes. This works because none of the remaining passes | |
4298 | need to look at them. | |
4299 | ||
4300 | ??? But it may break in the future. We should use a machine | |
4301 | dependent REG_NOTE, or some other approach entirely. */ | |
4302 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
4303 | { | |
9204e736 | 4304 | if (INSN_P (insn)) |
90e84010 | 4305 | { |
4306 | rtx note; | |
4307 | ||
4308 | while ((note = find_reg_note (insn, REG_LABEL, NULL_RTX)) != 0) | |
4309 | remove_note (insn, note); | |
4310 | } | |
4311 | } | |
4312 | ||
4313 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
4314 | { | |
4315 | rtx pattern, reg, link, set, scan, dies, label; | |
4316 | int rescan = 0, foundinsn = 0; | |
4317 | ||
881ffd32 | 4318 | if (GET_CODE (insn) == CALL_INSN) |
4319 | { | |
4320 | pattern = PATTERN (insn); | |
90e84010 | 4321 | |
881ffd32 | 4322 | if (GET_CODE (pattern) == PARALLEL) |
4323 | pattern = XVECEXP (pattern, 0, 0); | |
4324 | if (GET_CODE (pattern) == SET) | |
4325 | pattern = SET_SRC (pattern); | |
90e84010 | 4326 | |
881ffd32 | 4327 | if (GET_CODE (pattern) != CALL |
4328 | || GET_CODE (XEXP (pattern, 0)) != MEM) | |
4329 | continue; | |
90e84010 | 4330 | |
881ffd32 | 4331 | reg = XEXP (XEXP (pattern, 0), 0); |
4332 | } | |
4333 | else | |
4334 | { | |
4335 | reg = sfunc_uses_reg (insn); | |
4336 | if (! reg) | |
4337 | continue; | |
4338 | } | |
90e84010 | 4339 | |
90e84010 | 4340 | if (GET_CODE (reg) != REG) |
4341 | continue; | |
4342 | ||
4343 | /* This is a function call via REG. If the only uses of REG | |
4344 | between the time that it is set and the time that it dies | |
4345 | are in function calls, then we can associate all the | |
4346 | function calls with the setting of REG. */ | |
4347 | ||
4348 | for (link = LOG_LINKS (insn); link; link = XEXP (link, 1)) | |
4349 | { | |
3730a264 | 4350 | if (REG_NOTE_KIND (link) != 0) |
4351 | continue; | |
90e84010 | 4352 | set = single_set (XEXP (link, 0)); |
4353 | if (set && rtx_equal_p (reg, SET_DEST (set))) | |
4354 | { | |
4355 | link = XEXP (link, 0); | |
4356 | break; | |
4357 | } | |
4358 | } | |
4359 | ||
4360 | if (! link) | |
4361 | { | |
4362 | /* ??? Sometimes global register allocation will have | |
4363 | deleted the insn pointed to by LOG_LINKS. Try | |
4364 | scanning backward to find where the register is set. */ | |
4365 | for (scan = PREV_INSN (insn); | |
4366 | scan && GET_CODE (scan) != CODE_LABEL; | |
4367 | scan = PREV_INSN (scan)) | |
4368 | { | |
9204e736 | 4369 | if (! INSN_P (scan)) |
90e84010 | 4370 | continue; |
4371 | ||
4372 | if (! reg_mentioned_p (reg, scan)) | |
4373 | continue; | |
4374 | ||
4375 | if (noncall_uses_reg (reg, scan, &set)) | |
4376 | break; | |
4377 | ||
4378 | if (set) | |
4379 | { | |
4380 | link = scan; | |
4381 | break; | |
4382 | } | |
4383 | } | |
4384 | } | |
4385 | ||
4386 | if (! link) | |
4387 | continue; | |
4388 | ||
4389 | /* The register is set at LINK. */ | |
4390 | ||
4391 | /* We can only optimize the function call if the register is | |
4392 | being set to a symbol. In theory, we could sometimes | |
4393 | optimize calls to a constant location, but the assembler | |
4394 | and linker do not support that at present. */ | |
4395 | if (GET_CODE (SET_SRC (set)) != SYMBOL_REF | |
4396 | && GET_CODE (SET_SRC (set)) != LABEL_REF) | |
4397 | continue; | |
4398 | ||
4399 | /* Scan forward from LINK to the place where REG dies, and | |
4400 | make sure that the only insns which use REG are | |
4401 | themselves function calls. */ | |
4402 | ||
4fa2e8b8 | 4403 | /* ??? This doesn't work for call targets that were allocated |
4404 | by reload, since there may not be a REG_DEAD note for the | |
4405 | register. */ | |
4406 | ||
90e84010 | 4407 | dies = NULL_RTX; |
4408 | for (scan = NEXT_INSN (link); scan; scan = NEXT_INSN (scan)) | |
4409 | { | |
4410 | rtx scanset; | |
4411 | ||
4fa2e8b8 | 4412 | /* Don't try to trace forward past a CODE_LABEL if we haven't |
4413 | seen INSN yet. Ordinarily, we will only find the setting insn | |
4414 | in LOG_LINKS if it is in the same basic block. However, | |
4415 | cross-jumping can insert code labels in between the load and | |
4416 | the call, and can result in situations where a single call | |
4417 | insn may have two targets depending on where we came from. */ | |
4418 | ||
4419 | if (GET_CODE (scan) == CODE_LABEL && ! foundinsn) | |
4420 | break; | |
4421 | ||
9204e736 | 4422 | if (! INSN_P (scan)) |
90e84010 | 4423 | continue; |
4424 | ||
4425 | /* Don't try to trace forward past a JUMP. To optimize | |
4426 | safely, we would have to check that all the | |
4fa2e8b8 | 4427 | instructions at the jump destination did not use REG. */ |
90e84010 | 4428 | |
4fa2e8b8 | 4429 | if (GET_CODE (scan) == JUMP_INSN) |
90e84010 | 4430 | break; |
4431 | ||
4432 | if (! reg_mentioned_p (reg, scan)) | |
4433 | continue; | |
4434 | ||
4435 | if (noncall_uses_reg (reg, scan, &scanset)) | |
4436 | break; | |
4437 | ||
4438 | if (scan == insn) | |
4439 | foundinsn = 1; | |
4440 | ||
881ffd32 | 4441 | if (scan != insn |
4442 | && (GET_CODE (scan) == CALL_INSN || sfunc_uses_reg (scan))) | |
90e84010 | 4443 | { |
4444 | /* There is a function call to this register other | |
4445 | than the one we are checking. If we optimize | |
4446 | this call, we need to rescan again below. */ | |
4447 | rescan = 1; | |
4448 | } | |
4449 | ||
4450 | /* ??? We shouldn't have to worry about SCANSET here. | |
4451 | We should just be able to check for a REG_DEAD note | |
4452 | on a function call. However, the REG_DEAD notes are | |
4453 | apparently not dependable around libcalls; c-torture | |
4454 | execute/920501-2 is a test case. If SCANSET is set, | |
4455 | then this insn sets the register, so it must have | |
4456 | died earlier. Unfortunately, this will only handle | |
4457 | the cases in which the register is, in fact, set in a | |
4458 | later insn. */ | |
4459 | ||
4460 | /* ??? We shouldn't have to use FOUNDINSN here. | |
4461 | However, the LOG_LINKS fields are apparently not | |
4462 | entirely reliable around libcalls; | |
4463 | newlib/libm/math/e_pow.c is a test case. Sometimes | |
4464 | an insn will appear in LOG_LINKS even though it is | |
9df03d69 | 4465 | not the most recent insn which sets the register. */ |
90e84010 | 4466 | |
4467 | if (foundinsn | |
4468 | && (scanset | |
4469 | || find_reg_note (scan, REG_DEAD, reg))) | |
4470 | { | |
4471 | dies = scan; | |
4472 | break; | |
4473 | } | |
4474 | } | |
4475 | ||
4476 | if (! dies) | |
4477 | { | |
4478 | /* Either there was a branch, or some insn used REG | |
4479 | other than as a function call address. */ | |
4480 | continue; | |
4481 | } | |
4482 | ||
4483 | /* Create a code label, and put it in a REG_LABEL note on | |
4484 | the insn which sets the register, and on each call insn | |
4485 | which uses the register. In final_prescan_insn we look | |
4486 | for the REG_LABEL notes, and output the appropriate label | |
4487 | or pseudo-op. */ | |
4488 | ||
4489 | label = gen_label_rtx (); | |
4bb30577 | 4490 | REG_NOTES (link) = gen_rtx_INSN_LIST (REG_LABEL, label, |
7014838c | 4491 | REG_NOTES (link)); |
4bb30577 | 4492 | REG_NOTES (insn) = gen_rtx_INSN_LIST (REG_LABEL, label, |
7014838c | 4493 | REG_NOTES (insn)); |
90e84010 | 4494 | if (rescan) |
4495 | { | |
4496 | scan = link; | |
4497 | do | |
4498 | { | |
881ffd32 | 4499 | rtx reg2; |
4500 | ||
90e84010 | 4501 | scan = NEXT_INSN (scan); |
4502 | if (scan != insn | |
881ffd32 | 4503 | && ((GET_CODE (scan) == CALL_INSN |
4504 | && reg_mentioned_p (reg, scan)) | |
4505 | || ((reg2 = sfunc_uses_reg (scan)) | |
4506 | && REGNO (reg2) == REGNO (reg)))) | |
7014838c | 4507 | REG_NOTES (scan) |
4bb30577 | 4508 | = gen_rtx_INSN_LIST (REG_LABEL, label, REG_NOTES (scan)); |
90e84010 | 4509 | } |
4510 | while (scan != dies); | |
4511 | } | |
4512 | } | |
4513 | } | |
4514 | ||
9eaab178 | 4515 | if (TARGET_SH2) |
4516 | fixup_addr_diff_vecs (first); | |
8ded0752 | 4517 | |
4518 | if (optimize) | |
4519 | { | |
8ded0752 | 4520 | mdep_reorg_phase = SH_SHORTEN_BRANCHES0; |
4521 | shorten_branches (first); | |
4522 | } | |
90e84010 | 4523 | /* Scan the function looking for move instructions which have to be |
4524 | changed to pc-relative loads and insert the literal tables. */ | |
4525 | ||
8ded0752 | 4526 | mdep_reorg_phase = SH_FIXUP_PCLOAD; |
4527 | for (insn = first, num_mova = 0; insn; insn = NEXT_INSN (insn)) | |
97595bfd | 4528 | { |
8ded0752 | 4529 | if (mova_p (insn)) |
4530 | { | |
51dade95 | 4531 | /* ??? basic block reordering can move a switch table dispatch |
4532 | below the switch table. Check if that has happened. | |
4533 | We only have the addresses available when optimizing; but then, | |
4534 | this check shouldn't be needed when not optimizing. */ | |
4535 | rtx label_ref = XVECEXP (SET_SRC (PATTERN (insn)), 0, 0); | |
4536 | if (optimize | |
4537 | && (INSN_ADDRESSES (INSN_UID (insn)) | |
4538 | > INSN_ADDRESSES (INSN_UID (XEXP (label_ref, 0))))) | |
4539 | { | |
4540 | /* Change the mova into a load. | |
4541 | broken_move will then return true for it. */ | |
4542 | fixup_mova (insn); | |
4543 | } | |
4544 | else if (! num_mova++) | |
8ded0752 | 4545 | mova = insn; |
4546 | } | |
4547 | else if (GET_CODE (insn) == JUMP_INSN | |
4548 | && GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC | |
4549 | && num_mova) | |
4550 | { | |
4551 | rtx scan; | |
4552 | int total; | |
4553 | ||
4554 | num_mova--; | |
4555 | ||
4556 | /* Some code might have been inserted between the mova and | |
4557 | its ADDR_DIFF_VEC. Check if the mova is still in range. */ | |
4558 | for (scan = mova, total = 0; scan != insn; scan = NEXT_INSN (scan)) | |
9eaab178 | 4559 | total += get_attr_length (scan); |
8ded0752 | 4560 | |
4561 | /* range of mova is 1020, add 4 because pc counts from address of | |
4562 | second instruction after this one, subtract 2 in case pc is 2 | |
4563 | byte aligned. Possible alignment needed for the ADDR_DIFF_VEC | |
ad87de1e | 4564 | cancels out with alignment effects of the mova itself. */ |
8ded0752 | 4565 | if (total > 1022) |
4566 | { | |
4567 | /* Change the mova into a load, and restart scanning | |
4568 | there. broken_move will then return true for mova. */ | |
51dade95 | 4569 | fixup_mova (mova); |
8ded0752 | 4570 | insn = mova; |
4571 | } | |
4572 | } | |
51dade95 | 4573 | if (broken_move (insn) |
4574 | || (GET_CODE (insn) == INSN | |
4575 | && recog_memoized (insn) == CODE_FOR_casesi_worker_2)) | |
97595bfd | 4576 | { |
97595bfd | 4577 | rtx scan; |
73401833 | 4578 | /* Scan ahead looking for a barrier to stick the constant table |
4579 | behind. */ | |
8ded0752 | 4580 | rtx barrier = find_barrier (num_mova, mova, insn); |
87ed74ef | 4581 | rtx last_float_move = NULL_RTX, last_float = 0, *last_float_addr = NULL; |
51dade95 | 4582 | int need_aligned_label = 0; |
47c009e5 | 4583 | |
8ded0752 | 4584 | if (num_mova && ! mova_p (mova)) |
4585 | { | |
4586 | /* find_barrier had to change the first mova into a | |
4587 | pcload; thus, we have to start with this new pcload. */ | |
4588 | insn = mova; | |
4589 | num_mova = 0; | |
4590 | } | |
1504eb7e | 4591 | /* Now find all the moves between the points and modify them. */ |
97595bfd | 4592 | for (scan = insn; scan != barrier; scan = NEXT_INSN (scan)) |
4593 | { | |
8ded0752 | 4594 | if (GET_CODE (scan) == CODE_LABEL) |
4595 | last_float = 0; | |
51dade95 | 4596 | if (GET_CODE (scan) == INSN |
4597 | && recog_memoized (scan) == CODE_FOR_casesi_worker_2) | |
4598 | need_aligned_label = 1; | |
97595bfd | 4599 | if (broken_move (scan)) |
4600 | { | |
00f87557 | 4601 | rtx *patp = &PATTERN (scan), pat = *patp; |
4602 | rtx src, dst; | |
97595bfd | 4603 | rtx lab; |
97595bfd | 4604 | rtx newsrc; |
00f87557 | 4605 | enum machine_mode mode; |
4606 | ||
4607 | if (GET_CODE (pat) == PARALLEL) | |
4608 | patp = &XVECEXP (pat, 0, 0), pat = *patp; | |
4609 | src = SET_SRC (pat); | |
4610 | dst = SET_DEST (pat); | |
4611 | mode = GET_MODE (dst); | |
97595bfd | 4612 | |
1b61190c | 4613 | if (mode == SImode && hi_const (src) |
4614 | && REGNO (dst) != FPUL_REG) | |
97595bfd | 4615 | { |
59a5a83f | 4616 | int offset = 0; |
73401833 | 4617 | |
97595bfd | 4618 | mode = HImode; |
f3d93547 | 4619 | while (GET_CODE (dst) == SUBREG) |
59a5a83f | 4620 | { |
701e46d0 | 4621 | offset += subreg_regno_offset (REGNO (SUBREG_REG (dst)), |
4622 | GET_MODE (SUBREG_REG (dst)), | |
4623 | SUBREG_BYTE (dst), | |
4624 | GET_MODE (dst)); | |
59a5a83f | 4625 | dst = SUBREG_REG (dst); |
4626 | } | |
7014838c | 4627 | dst = gen_rtx_REG (HImode, REGNO (dst) + offset); |
97595bfd | 4628 | } |
e2142b04 | 4629 | if (GET_CODE (dst) == REG && FP_ANY_REGISTER_P (REGNO (dst))) |
8ded0752 | 4630 | { |
c961a48b | 4631 | /* This must be an insn that clobbers r0. */ |
adbf1ac5 | 4632 | rtx *clobberp = &XVECEXP (PATTERN (scan), 0, |
4633 | XVECLEN (PATTERN (scan), 0) | |
4634 | - 1); | |
4635 | rtx clobber = *clobberp; | |
c961a48b | 4636 | |
1a6a7a27 | 4637 | gcc_assert (GET_CODE (clobber) == CLOBBER |
4638 | && rtx_equal_p (XEXP (clobber, 0), r0_rtx)); | |
c961a48b | 4639 | |
8ded0752 | 4640 | if (last_float |
4641 | && reg_set_between_p (r0_rtx, last_float_move, scan)) | |
4642 | last_float = 0; | |
c5ea35c4 | 4643 | if (last_float |
4644 | && TARGET_SHCOMPACT | |
4645 | && GET_MODE_SIZE (mode) != 4 | |
4646 | && GET_MODE_SIZE (GET_MODE (last_float)) == 4) | |
4647 | last_float = 0; | |
8ded0752 | 4648 | lab = add_constant (src, mode, last_float); |
4649 | if (lab) | |
4650 | emit_insn_before (gen_mova (lab), scan); | |
4651 | else | |
c961a48b | 4652 | { |
4653 | /* There will be a REG_UNUSED note for r0 on | |
4654 | LAST_FLOAT_MOVE; we have to change it to REG_INC, | |
4655 | lest reorg:mark_target_live_regs will not | |
4656 | consider r0 to be used, and we end up with delay | |
d06d4df0 | 4657 | slot insn in front of SCAN that clobbers r0. */ |
c961a48b | 4658 | rtx note |
4659 | = find_regno_note (last_float_move, REG_UNUSED, 0); | |
4660 | ||
46bc3e81 | 4661 | /* If we are not optimizing, then there may not be |
4662 | a note. */ | |
4663 | if (note) | |
4664 | PUT_MODE (note, REG_INC); | |
c961a48b | 4665 | |
4666 | *last_float_addr = r0_inc_rtx; | |
4667 | } | |
8ded0752 | 4668 | last_float_move = scan; |
4669 | last_float = src; | |
aa7e75f2 | 4670 | newsrc = gen_const_mem (mode, |
b7dbbdb2 | 4671 | (((TARGET_SH4 && ! TARGET_FMOVD) |
1b61190c | 4672 | || REGNO (dst) == FPUL_REG) |
8ded0752 | 4673 | ? r0_inc_rtx |
4674 | : r0_rtx)); | |
4675 | last_float_addr = &XEXP (newsrc, 0); | |
c961a48b | 4676 | |
4677 | /* Remove the clobber of r0. */ | |
adbf1ac5 | 4678 | *clobberp = gen_rtx_CLOBBER (GET_MODE (clobber), |
4679 | gen_rtx_SCRATCH (Pmode)); | |
8ded0752 | 4680 | } |
0abf894c | 4681 | /* This is a mova needing a label. Create it. */ |
393fac7b | 4682 | else if (GET_CODE (src) == UNSPEC |
4683 | && XINT (src, 1) == UNSPEC_MOVA | |
4684 | && GET_CODE (XVECEXP (src, 0, 0)) == CONST) | |
0abf894c | 4685 | { |
393fac7b | 4686 | lab = add_constant (XVECEXP (src, 0, 0), mode, 0); |
0abf894c | 4687 | newsrc = gen_rtx_LABEL_REF (VOIDmode, lab); |
f4dba476 | 4688 | newsrc = gen_rtx_UNSPEC (SImode, |
a5bd34f1 | 4689 | gen_rtvec (1, newsrc), |
4690 | UNSPEC_MOVA); | |
0abf894c | 4691 | } |
8ded0752 | 4692 | else |
4693 | { | |
4694 | lab = add_constant (src, mode, 0); | |
e265a6da | 4695 | newsrc = gen_rtx_LABEL_REF (VOIDmode, lab); |
4696 | newsrc = gen_const_mem (mode, newsrc); | |
8ded0752 | 4697 | } |
7014838c | 4698 | *patp = gen_rtx_SET (VOIDmode, dst, newsrc); |
00f87557 | 4699 | INSN_CODE (scan) = -1; |
97595bfd | 4700 | } |
4701 | } | |
51dade95 | 4702 | dump_table (need_aligned_label ? insn : 0, barrier); |
8ded0752 | 4703 | insn = barrier; |
97595bfd | 4704 | } |
4705 | } | |
8ded0752 | 4706 | |
4707 | mdep_reorg_phase = SH_SHORTEN_BRANCHES1; | |
47fc0706 | 4708 | INSN_ADDRESSES_FREE (); |
8ded0752 | 4709 | split_branches (first); |
4710 | ||
4711 | /* The INSN_REFERENCES_ARE_DELAYED in sh.h is problematic because it | |
f747aa2d | 4712 | also has an effect on the register that holds the address of the sfunc. |
8ded0752 | 4713 | Insert an extra dummy insn in front of each sfunc that pretends to |
4714 | use this register. */ | |
4715 | if (flag_delayed_branch) | |
4716 | { | |
4717 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
4718 | { | |
4719 | rtx reg = sfunc_uses_reg (insn); | |
4720 | ||
4721 | if (! reg) | |
4722 | continue; | |
4723 | emit_insn_before (gen_use_sfunc_addr (reg), insn); | |
4724 | } | |
4725 | } | |
1b61190c | 4726 | #if 0 |
4727 | /* fpscr is not actually a user variable, but we pretend it is for the | |
4728 | sake of the previous optimization passes, since we want it handled like | |
18862b5a | 4729 | one. However, we don't have any debugging information for it, so turn |
1b61190c | 4730 | it into a non-user variable now. */ |
4731 | if (TARGET_SH4) | |
4732 | REG_USERVAR_P (get_fpscr_rtx ()) = 0; | |
4733 | #endif | |
8ded0752 | 4734 | mdep_reorg_phase = SH_AFTER_MDEP_REORG; |
4735 | } | |
4736 | ||
4737 | int | |
04f04b72 | 4738 | get_dest_uid (rtx label, int max_uid) |
8ded0752 | 4739 | { |
4740 | rtx dest = next_real_insn (label); | |
4741 | int dest_uid; | |
4742 | if (! dest) | |
4743 | /* This can happen for an undefined label. */ | |
4744 | return 0; | |
4745 | dest_uid = INSN_UID (dest); | |
4746 | /* If this is a newly created branch redirection blocking instruction, | |
4747 | we cannot index the branch_uid or insn_addresses arrays with its | |
4748 | uid. But then, we won't need to, because the actual destination is | |
4749 | the following branch. */ | |
4750 | while (dest_uid >= max_uid) | |
4751 | { | |
4752 | dest = NEXT_INSN (dest); | |
4753 | dest_uid = INSN_UID (dest); | |
4754 | } | |
4755 | if (GET_CODE (dest) == JUMP_INSN && GET_CODE (PATTERN (dest)) == RETURN) | |
4756 | return 0; | |
4757 | return dest_uid; | |
4758 | } | |
4759 | ||
4760 | /* Split condbranches that are out of range. Also add clobbers for | |
4761 | scratch registers that are needed in far jumps. | |
4762 | We do this before delay slot scheduling, so that it can take our | |
4763 | newly created instructions into account. It also allows us to | |
4764 | find branches with common targets more easily. */ | |
4765 | ||
4766 | static void | |
04f04b72 | 4767 | split_branches (rtx first) |
8ded0752 | 4768 | { |
4769 | rtx insn; | |
4770 | struct far_branch **uid_branch, *far_branch_list = 0; | |
4771 | int max_uid = get_max_uid (); | |
1a6a7a27 | 4772 | int ok; |
8ded0752 | 4773 | |
4774 | /* Find out which branches are out of range. */ | |
8ded0752 | 4775 | shorten_branches (first); |
4776 | ||
4777 | uid_branch = (struct far_branch **) alloca (max_uid * sizeof *uid_branch); | |
8e547276 | 4778 | memset ((char *) uid_branch, 0, max_uid * sizeof *uid_branch); |
8ded0752 | 4779 | |
4780 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
9204e736 | 4781 | if (! INSN_P (insn)) |
8ded0752 | 4782 | continue; |
4783 | else if (INSN_DELETED_P (insn)) | |
4784 | { | |
4785 | /* Shorten_branches would split this instruction again, | |
4786 | so transform it into a note. */ | |
4787 | PUT_CODE (insn, NOTE); | |
4788 | NOTE_LINE_NUMBER (insn) = NOTE_INSN_DELETED; | |
4789 | NOTE_SOURCE_FILE (insn) = 0; | |
4790 | } | |
4791 | else if (GET_CODE (insn) == JUMP_INSN | |
4792 | /* Don't mess with ADDR_DIFF_VEC */ | |
4793 | && (GET_CODE (PATTERN (insn)) == SET | |
4794 | || GET_CODE (PATTERN (insn)) == RETURN)) | |
4795 | { | |
4796 | enum attr_type type = get_attr_type (insn); | |
4797 | if (type == TYPE_CBRANCH) | |
4798 | { | |
4799 | rtx next, beyond; | |
9e7454d0 | 4800 | |
8ded0752 | 4801 | if (get_attr_length (insn) > 4) |
4802 | { | |
4803 | rtx src = SET_SRC (PATTERN (insn)); | |
8ded0752 | 4804 | rtx olabel = XEXP (XEXP (src, 1), 0); |
47fc0706 | 4805 | int addr = INSN_ADDRESSES (INSN_UID (insn)); |
8ded0752 | 4806 | rtx label = 0; |
4807 | int dest_uid = get_dest_uid (olabel, max_uid); | |
4808 | struct far_branch *bp = uid_branch[dest_uid]; | |
9e7454d0 | 4809 | |
8ded0752 | 4810 | /* redirect_jump needs a valid JUMP_LABEL, and it might delete |
40df6fd4 | 4811 | the label if the LABEL_NUSES count drops to zero. There is |
8ded0752 | 4812 | always a jump_optimize pass that sets these values, but it |
4813 | proceeds to delete unreferenced code, and then if not | |
ad87de1e | 4814 | optimizing, to un-delete the deleted instructions, thus |
8ded0752 | 4815 | leaving labels with too low uses counts. */ |
4816 | if (! optimize) | |
4817 | { | |
4818 | JUMP_LABEL (insn) = olabel; | |
4819 | LABEL_NUSES (olabel)++; | |
4820 | } | |
4821 | if (! bp) | |
4822 | { | |
4823 | bp = (struct far_branch *) alloca (sizeof *bp); | |
4824 | uid_branch[dest_uid] = bp; | |
4825 | bp->prev = far_branch_list; | |
4826 | far_branch_list = bp; | |
4827 | bp->far_label | |
4828 | = XEXP (XEXP (SET_SRC (PATTERN (insn)), 1), 0); | |
4829 | LABEL_NUSES (bp->far_label)++; | |
4830 | } | |
4831 | else | |
4832 | { | |
4833 | label = bp->near_label; | |
4834 | if (! label && bp->address - addr >= CONDJUMP_MIN) | |
4835 | { | |
4836 | rtx block = bp->insert_place; | |
4837 | ||
4838 | if (GET_CODE (PATTERN (block)) == RETURN) | |
4839 | block = PREV_INSN (block); | |
4840 | else | |
4841 | block = gen_block_redirect (block, | |
4842 | bp->address, 2); | |
4843 | label = emit_label_after (gen_label_rtx (), | |
4844 | PREV_INSN (block)); | |
4845 | bp->near_label = label; | |
4846 | } | |
4847 | else if (label && ! NEXT_INSN (label)) | |
17bec211 | 4848 | { |
4849 | if (addr + 2 - bp->address <= CONDJUMP_MAX) | |
4850 | bp->insert_place = insn; | |
4851 | else | |
4852 | gen_far_branch (bp); | |
4853 | } | |
8ded0752 | 4854 | } |
4855 | if (! label | |
b7dbbdb2 | 4856 | || (NEXT_INSN (label) && bp->address - addr < CONDJUMP_MIN)) |
8ded0752 | 4857 | { |
4858 | bp->near_label = label = gen_label_rtx (); | |
4859 | bp->insert_place = insn; | |
4860 | bp->address = addr; | |
4861 | } | |
1a6a7a27 | 4862 | ok = redirect_jump (insn, label, 1); |
4863 | gcc_assert (ok); | |
8ded0752 | 4864 | } |
4865 | else | |
4866 | { | |
4867 | /* get_attr_length (insn) == 2 */ | |
4868 | /* Check if we have a pattern where reorg wants to redirect | |
4869 | the branch to a label from an unconditional branch that | |
4870 | is too far away. */ | |
4871 | /* We can't use JUMP_LABEL here because it might be undefined | |
4872 | when not optimizing. */ | |
9eaab178 | 4873 | /* A syntax error might cause beyond to be NULL_RTX. */ |
8ded0752 | 4874 | beyond |
4875 | = next_active_insn (XEXP (XEXP (SET_SRC (PATTERN (insn)), 1), | |
4876 | 0)); | |
9e7454d0 | 4877 | |
9eaab178 | 4878 | if (beyond |
4879 | && (GET_CODE (beyond) == JUMP_INSN | |
b1060622 | 4880 | || ((beyond = next_active_insn (beyond)) |
4881 | && GET_CODE (beyond) == JUMP_INSN)) | |
8ded0752 | 4882 | && GET_CODE (PATTERN (beyond)) == SET |
adbcc7c4 | 4883 | && recog_memoized (beyond) == CODE_FOR_jump_compact |
47fc0706 | 4884 | && ((INSN_ADDRESSES |
4885 | (INSN_UID (XEXP (SET_SRC (PATTERN (beyond)), 0))) | |
30435bf8 | 4886 | - INSN_ADDRESSES (INSN_UID (insn)) + (unsigned) 252) |
8ded0752 | 4887 | > 252 + 258 + 2)) |
4888 | gen_block_redirect (beyond, | |
47fc0706 | 4889 | INSN_ADDRESSES (INSN_UID (beyond)), 1); |
8ded0752 | 4890 | } |
9e7454d0 | 4891 | |
8ded0752 | 4892 | next = next_active_insn (insn); |
4893 | ||
4894 | if ((GET_CODE (next) == JUMP_INSN | |
ee617972 | 4895 | || ((next = next_active_insn (next)) |
4896 | && GET_CODE (next) == JUMP_INSN)) | |
8ded0752 | 4897 | && GET_CODE (PATTERN (next)) == SET |
adbcc7c4 | 4898 | && recog_memoized (next) == CODE_FOR_jump_compact |
47fc0706 | 4899 | && ((INSN_ADDRESSES |
4900 | (INSN_UID (XEXP (SET_SRC (PATTERN (next)), 0))) | |
30435bf8 | 4901 | - INSN_ADDRESSES (INSN_UID (insn)) + (unsigned) 252) |
8ded0752 | 4902 | > 252 + 258 + 2)) |
47fc0706 | 4903 | gen_block_redirect (next, INSN_ADDRESSES (INSN_UID (next)), 1); |
8ded0752 | 4904 | } |
4905 | else if (type == TYPE_JUMP || type == TYPE_RETURN) | |
4906 | { | |
47fc0706 | 4907 | int addr = INSN_ADDRESSES (INSN_UID (insn)); |
8ded0752 | 4908 | rtx far_label = 0; |
4909 | int dest_uid = 0; | |
4910 | struct far_branch *bp; | |
4911 | ||
4912 | if (type == TYPE_JUMP) | |
4913 | { | |
4914 | far_label = XEXP (SET_SRC (PATTERN (insn)), 0); | |
4915 | dest_uid = get_dest_uid (far_label, max_uid); | |
4916 | if (! dest_uid) | |
4917 | { | |
4918 | /* Parse errors can lead to labels outside | |
4919 | the insn stream. */ | |
4920 | if (! NEXT_INSN (far_label)) | |
4921 | continue; | |
4922 | ||
4923 | if (! optimize) | |
4924 | { | |
4925 | JUMP_LABEL (insn) = far_label; | |
4926 | LABEL_NUSES (far_label)++; | |
4927 | } | |
5e20801c | 4928 | redirect_jump (insn, NULL_RTX, 1); |
8ded0752 | 4929 | far_label = 0; |
4930 | } | |
4931 | } | |
4932 | bp = uid_branch[dest_uid]; | |
4933 | if (! bp) | |
4934 | { | |
4935 | bp = (struct far_branch *) alloca (sizeof *bp); | |
4936 | uid_branch[dest_uid] = bp; | |
4937 | bp->prev = far_branch_list; | |
4938 | far_branch_list = bp; | |
4939 | bp->near_label = 0; | |
4940 | bp->far_label = far_label; | |
4941 | if (far_label) | |
4942 | LABEL_NUSES (far_label)++; | |
4943 | } | |
4944 | else if (bp->near_label && ! NEXT_INSN (bp->near_label)) | |
4945 | if (addr - bp->address <= CONDJUMP_MAX) | |
4946 | emit_label_after (bp->near_label, PREV_INSN (insn)); | |
4947 | else | |
4948 | { | |
4949 | gen_far_branch (bp); | |
4950 | bp->near_label = 0; | |
4951 | } | |
4952 | else | |
4953 | bp->near_label = 0; | |
4954 | bp->address = addr; | |
4955 | bp->insert_place = insn; | |
4956 | if (! far_label) | |
4957 | emit_insn_before (gen_block_branch_redirect (const0_rtx), insn); | |
4958 | else | |
4959 | gen_block_redirect (insn, addr, bp->near_label ? 2 : 0); | |
4960 | } | |
4961 | } | |
4962 | /* Generate all pending far branches, | |
4963 | and free our references to the far labels. */ | |
4964 | while (far_branch_list) | |
4965 | { | |
4966 | if (far_branch_list->near_label | |
4967 | && ! NEXT_INSN (far_branch_list->near_label)) | |
4968 | gen_far_branch (far_branch_list); | |
4969 | if (optimize | |
4970 | && far_branch_list->far_label | |
4971 | && ! --LABEL_NUSES (far_branch_list->far_label)) | |
4972 | delete_insn (far_branch_list->far_label); | |
4973 | far_branch_list = far_branch_list->prev; | |
4974 | } | |
bd967f88 | 4975 | |
4976 | /* Instruction length information is no longer valid due to the new | |
4977 | instructions that have been generated. */ | |
4978 | init_insn_lengths (); | |
47c009e5 | 4979 | } |
4980 | ||
73401833 | 4981 | /* Dump out instruction addresses, which is useful for debugging the |
90e84010 | 4982 | constant pool table stuff. |
4983 | ||
4984 | If relaxing, output the label and pseudo-ops used to link together | |
4985 | calls and the instruction which set the registers. */ | |
73401833 | 4986 | |
73401833 | 4987 | /* ??? The addresses printed by this routine for insns are nonsense for |
4988 | insns which are inside of a sequence where none of the inner insns have | |
4989 | variable length. This is because the second pass of shorten_branches | |
4990 | does not bother to update them. */ | |
97595bfd | 4991 | |
2e3d4844 | 4992 | void |
04f04b72 | 4993 | final_prescan_insn (rtx insn, rtx *opvec ATTRIBUTE_UNUSED, |
4994 | int noperands ATTRIBUTE_UNUSED) | |
47c009e5 | 4995 | { |
73401833 | 4996 | if (TARGET_DUMPISIZE) |
47fc0706 | 4997 | fprintf (asm_out_file, "\n! at %04x\n", INSN_ADDRESSES (INSN_UID (insn))); |
90e84010 | 4998 | |
4999 | if (TARGET_RELAX) | |
5000 | { | |
5001 | rtx note; | |
5002 | ||
5003 | note = find_reg_note (insn, REG_LABEL, NULL_RTX); | |
5004 | if (note) | |
5005 | { | |
5006 | rtx pattern; | |
5007 | ||
5008 | pattern = PATTERN (insn); | |
a866e0f2 | 5009 | if (GET_CODE (pattern) == PARALLEL) |
5010 | pattern = XVECEXP (pattern, 0, 0); | |
1a6a7a27 | 5011 | switch (GET_CODE (pattern)) |
5012 | { | |
1a6a7a27 | 5013 | case SET: |
5014 | if (GET_CODE (SET_SRC (pattern)) != CALL | |
5015 | && get_attr_type (insn) != TYPE_SFUNC) | |
5016 | { | |
5017 | targetm.asm_out.internal_label | |
5018 | (asm_out_file, "L", CODE_LABEL_NUMBER (XEXP (note, 0))); | |
5019 | break; | |
5020 | } | |
5021 | /* else FALLTHROUGH */ | |
5022 | case CALL: | |
5023 | asm_fprintf (asm_out_file, "\t.uses %LL%d\n", | |
5024 | CODE_LABEL_NUMBER (XEXP (note, 0))); | |
5025 | break; | |
5026 | ||
5027 | default: | |
5028 | gcc_unreachable (); | |
5029 | } | |
90e84010 | 5030 | } |
5031 | } | |
97595bfd | 5032 | } |
47c009e5 | 5033 | |
73401833 | 5034 | /* Dump out any constants accumulated in the final pass. These will |
3398e91d | 5035 | only be labels. */ |
47c009e5 | 5036 | |
b7dbbdb2 | 5037 | const char * |
04f04b72 | 5038 | output_jump_label_table (void) |
97595bfd | 5039 | { |
73401833 | 5040 | int i; |
5041 | ||
5042 | if (pool_size) | |
5043 | { | |
5044 | fprintf (asm_out_file, "\t.align 2\n"); | |
5045 | for (i = 0; i < pool_size; i++) | |
5046 | { | |
5047 | pool_node *p = &pool_vector[i]; | |
5048 | ||
805e22b2 | 5049 | (*targetm.asm_out.internal_label) (asm_out_file, "L", |
73401833 | 5050 | CODE_LABEL_NUMBER (p->label)); |
5051 | output_asm_insn (".long %O0", &p->value); | |
5052 | } | |
5053 | pool_size = 0; | |
5054 | } | |
47c009e5 | 5055 | |
73401833 | 5056 | return ""; |
5057 | } | |
5058 | \f | |
5059 | /* A full frame looks like: | |
1504eb7e | 5060 | |
5061 | arg-5 | |
5062 | arg-4 | |
5063 | [ if current_function_anonymous_args | |
5064 | arg-3 | |
5065 | arg-2 | |
5066 | arg-1 | |
5067 | arg-0 ] | |
5068 | saved-fp | |
5069 | saved-r10 | |
5070 | saved-r11 | |
5071 | saved-r12 | |
5072 | saved-pr | |
5073 | local-n | |
5074 | .. | |
5075 | local-1 | |
5076 | local-0 <- fp points here. */ | |
2e3d4844 | 5077 | |
73401833 | 5078 | /* Number of bytes pushed for anonymous args, used to pass information |
5079 | between expand_prologue and expand_epilogue. */ | |
5080 | ||
41fafa66 | 5081 | /* Adjust the stack by SIZE bytes. REG holds the rtl of the register to be |
5082 | adjusted. If epilogue_p is zero, this is for a prologue; otherwise, it's | |
66209063 | 5083 | for an epilogue and a negative value means that it's for a sibcall |
5084 | epilogue. If LIVE_REGS_MASK is nonzero, it points to a HARD_REG_SET of | |
5085 | all the registers that are about to be restored, and hence dead. */ | |
73401833 | 5086 | |
5087 | static void | |
41fafa66 | 5088 | output_stack_adjust (int size, rtx reg, int epilogue_p, |
5089 | HARD_REG_SET *live_regs_mask) | |
73401833 | 5090 | { |
41fafa66 | 5091 | rtx (*emit_fn) (rtx) = epilogue_p ? &emit_insn : &frame_insn; |
73401833 | 5092 | if (size) |
5093 | { | |
87e19636 | 5094 | HOST_WIDE_INT align = STACK_BOUNDARY / BITS_PER_UNIT; |
5095 | ||
7105fb72 | 5096 | /* This test is bogus, as output_stack_adjust is used to re-align the |
5097 | stack. */ | |
5098 | #if 0 | |
1a6a7a27 | 5099 | gcc_assert (!(size % align)); |
7105fb72 | 5100 | #endif |
87e19636 | 5101 | |
5102 | if (CONST_OK_FOR_ADD (size)) | |
363600c8 | 5103 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size))); |
b9d4bd5a | 5104 | /* Try to do it with two partial adjustments; however, we must make |
5105 | sure that the stack is properly aligned at all times, in case | |
9df03d69 | 5106 | an interrupt occurs between the two partial adjustments. */ |
87e19636 | 5107 | else if (CONST_OK_FOR_ADD (size / 2 & -align) |
5108 | && CONST_OK_FOR_ADD (size - (size / 2 & -align))) | |
b9d4bd5a | 5109 | { |
363600c8 | 5110 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size / 2 & -align))); |
5111 | emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size - (size / 2 & -align)))); | |
b9d4bd5a | 5112 | } |
5113 | else | |
73401833 | 5114 | { |
b9d4bd5a | 5115 | rtx const_reg; |
363600c8 | 5116 | rtx insn; |
41fafa66 | 5117 | int temp = epilogue_p ? 7 : (TARGET_SH5 ? 0 : 1); |
5118 | int i; | |
b9d4bd5a | 5119 | |
5120 | /* If TEMP is invalid, we could temporarily save a general | |
5121 | register to MACL. However, there is currently no need | |
1a6a7a27 | 5122 | to handle this case, so just die when we see it. */ |
66209063 | 5123 | if (epilogue_p < 0 |
5124 | || current_function_interrupt | |
956dd562 | 5125 | || ! call_really_used_regs[temp] || fixed_regs[temp]) |
41fafa66 | 5126 | temp = -1; |
66209063 | 5127 | if (temp < 0 && ! current_function_interrupt |
5128 | && (TARGET_SHMEDIA || epilogue_p >= 0)) | |
41fafa66 | 5129 | { |
5130 | HARD_REG_SET temps; | |
5131 | COPY_HARD_REG_SET (temps, call_used_reg_set); | |
5132 | AND_COMPL_HARD_REG_SET (temps, call_fixed_reg_set); | |
66209063 | 5133 | if (epilogue_p > 0) |
41fafa66 | 5134 | { |
66209063 | 5135 | int nreg = 0; |
5136 | if (current_function_return_rtx) | |
5137 | { | |
5138 | enum machine_mode mode; | |
5139 | mode = GET_MODE (current_function_return_rtx); | |
5140 | if (BASE_RETURN_VALUE_REG (mode) == FIRST_RET_REG) | |
5141 | nreg = HARD_REGNO_NREGS (FIRST_RET_REG, mode); | |
5142 | } | |
5143 | for (i = 0; i < nreg; i++) | |
41fafa66 | 5144 | CLEAR_HARD_REG_BIT (temps, FIRST_RET_REG + i); |
5145 | if (current_function_calls_eh_return) | |
5146 | { | |
5147 | CLEAR_HARD_REG_BIT (temps, EH_RETURN_STACKADJ_REGNO); | |
5148 | for (i = 0; i <= 3; i++) | |
5149 | CLEAR_HARD_REG_BIT (temps, EH_RETURN_DATA_REGNO (i)); | |
5150 | } | |
5151 | } | |
66209063 | 5152 | if (TARGET_SHMEDIA && epilogue_p < 0) |
5153 | for (i = FIRST_TARGET_REG; i <= LAST_TARGET_REG; i++) | |
5154 | CLEAR_HARD_REG_BIT (temps, i); | |
5155 | if (epilogue_p <= 0) | |
41fafa66 | 5156 | { |
5157 | for (i = FIRST_PARM_REG; | |
5158 | i < FIRST_PARM_REG + NPARM_REGS (SImode); i++) | |
5159 | CLEAR_HARD_REG_BIT (temps, i); | |
4ee9c684 | 5160 | if (cfun->static_chain_decl != NULL) |
41fafa66 | 5161 | CLEAR_HARD_REG_BIT (temps, STATIC_CHAIN_REGNUM); |
5162 | } | |
5163 | temp = scavenge_reg (&temps); | |
5164 | } | |
5165 | if (temp < 0 && live_regs_mask) | |
5166 | temp = scavenge_reg (live_regs_mask); | |
b9d4bd5a | 5167 | if (temp < 0) |
66209063 | 5168 | { |
1a6a7a27 | 5169 | rtx adj_reg, tmp_reg, mem; |
5170 | ||
66209063 | 5171 | /* If we reached here, the most likely case is the (sibcall) |
5172 | epilogue for non SHmedia. Put a special push/pop sequence | |
5173 | for such case as the last resort. This looks lengthy but | |
1a6a7a27 | 5174 | would not be problem because it seems to be very |
5175 | rare. */ | |
5176 | ||
5177 | gcc_assert (!TARGET_SHMEDIA && epilogue_p); | |
5178 | ||
5179 | ||
5180 | /* ??? There is still the slight possibility that r4 or | |
5181 | r5 have been reserved as fixed registers or assigned | |
5182 | as global registers, and they change during an | |
5183 | interrupt. There are possible ways to handle this: | |
5184 | ||
5185 | - If we are adjusting the frame pointer (r14), we can do | |
5186 | with a single temp register and an ordinary push / pop | |
5187 | on the stack. | |
5188 | - Grab any call-used or call-saved registers (i.e. not | |
5189 | fixed or globals) for the temps we need. We might | |
5190 | also grab r14 if we are adjusting the stack pointer. | |
5191 | If we can't find enough available registers, issue | |
5192 | a diagnostic and die - the user must have reserved | |
5193 | way too many registers. | |
5194 | But since all this is rather unlikely to happen and | |
5195 | would require extra testing, we just die if r4 / r5 | |
5196 | are not available. */ | |
5197 | gcc_assert (!fixed_regs[4] && !fixed_regs[5] | |
5198 | && !global_regs[4] && !global_regs[5]); | |
5199 | ||
5200 | adj_reg = gen_rtx_REG (GET_MODE (reg), 4); | |
5201 | tmp_reg = gen_rtx_REG (GET_MODE (reg), 5); | |
aa7e75f2 | 5202 | emit_move_insn (gen_tmp_stack_mem (Pmode, reg), adj_reg); |
1a6a7a27 | 5203 | emit_insn (GEN_MOV (adj_reg, GEN_INT (size))); |
5204 | emit_insn (GEN_ADD3 (adj_reg, adj_reg, reg)); | |
aa7e75f2 | 5205 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_PRE_DEC (Pmode, adj_reg)); |
1a6a7a27 | 5206 | emit_move_insn (mem, tmp_reg); |
aa7e75f2 | 5207 | emit_move_insn (tmp_reg, gen_tmp_stack_mem (Pmode, reg)); |
5208 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_PRE_DEC (Pmode, adj_reg)); | |
5209 | emit_move_insn (mem, tmp_reg); | |
5210 | emit_move_insn (reg, adj_reg); | |
5211 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_POST_INC (Pmode, reg)); | |
5212 | emit_move_insn (adj_reg, mem); | |
5213 | mem = gen_tmp_stack_mem (Pmode, gen_rtx_POST_INC (Pmode, reg)); | |
5214 | emit_move_insn (tmp_reg, mem); | |
5215 | return; | |
66209063 | 5216 | } |
87e19636 | 5217 | const_reg = gen_rtx_REG (GET_MODE (reg), temp); |
b9d4bd5a | 5218 | |
5219 | /* If SIZE is negative, subtract the positive value. | |
5220 | This sometimes allows a constant pool entry to be shared | |
5221 | between prologue and epilogue code. */ | |
5222 | if (size < 0) | |
51143b47 | 5223 | { |
87e19636 | 5224 | emit_insn (GEN_MOV (const_reg, GEN_INT (-size))); |
363600c8 | 5225 | insn = emit_fn (GEN_SUB3 (reg, reg, const_reg)); |
51143b47 | 5226 | } |
5227 | else | |
5228 | { | |
87e19636 | 5229 | emit_insn (GEN_MOV (const_reg, GEN_INT (size))); |
363600c8 | 5230 | insn = emit_fn (GEN_ADD3 (reg, reg, const_reg)); |
51143b47 | 5231 | } |
41fafa66 | 5232 | if (! epilogue_p) |
363600c8 | 5233 | REG_NOTES (insn) |
5234 | = (gen_rtx_EXPR_LIST | |
5235 | (REG_FRAME_RELATED_EXPR, | |
5236 | gen_rtx_SET (VOIDmode, reg, | |
5237 | gen_rtx_PLUS (SImode, reg, GEN_INT (size))), | |
5238 | REG_NOTES (insn))); | |
73401833 | 5239 | } |
73401833 | 5240 | } |
5241 | } | |
5242 | ||
363600c8 | 5243 | static rtx |
04f04b72 | 5244 | frame_insn (rtx x) |
363600c8 | 5245 | { |
5246 | x = emit_insn (x); | |
5247 | RTX_FRAME_RELATED_P (x) = 1; | |
5248 | return x; | |
5249 | } | |
5250 | ||
73401833 | 5251 | /* Output RTL to push register RN onto the stack. */ |
5252 | ||
363600c8 | 5253 | static rtx |
04f04b72 | 5254 | push (int rn) |
73401833 | 5255 | { |
5256 | rtx x; | |
1b61190c | 5257 | if (rn == FPUL_REG) |
5258 | x = gen_push_fpul (); | |
88b7f6c6 | 5259 | else if (rn == FPSCR_REG) |
5260 | x = gen_push_fpscr (); | |
7105fb72 | 5261 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && ! TARGET_FPU_SINGLE |
e2142b04 | 5262 | && FP_OR_XD_REGISTER_P (rn)) |
1b61190c | 5263 | { |
e2142b04 | 5264 | if (FP_REGISTER_P (rn) && (rn - FIRST_FP_REG) & 1) |
acda5809 | 5265 | return NULL_RTX; |
7014838c | 5266 | x = gen_push_4 (gen_rtx_REG (DFmode, rn)); |
1b61190c | 5267 | } |
87ed74ef | 5268 | else if (TARGET_SH2E && FP_REGISTER_P (rn)) |
7014838c | 5269 | x = gen_push_e (gen_rtx_REG (SFmode, rn)); |
b00564da | 5270 | else |
7014838c | 5271 | x = gen_push (gen_rtx_REG (SImode, rn)); |
b00564da | 5272 | |
363600c8 | 5273 | x = frame_insn (x); |
7014838c | 5274 | REG_NOTES (x) |
5275 | = gen_rtx_EXPR_LIST (REG_INC, | |
5276 | gen_rtx_REG (SImode, STACK_POINTER_REGNUM), 0); | |
363600c8 | 5277 | return x; |
73401833 | 5278 | } |
5279 | ||
5280 | /* Output RTL to pop register RN from the stack. */ | |
5281 | ||
5282 | static void | |
04f04b72 | 5283 | pop (int rn) |
73401833 | 5284 | { |
5285 | rtx x; | |
1b61190c | 5286 | if (rn == FPUL_REG) |
5287 | x = gen_pop_fpul (); | |
88b7f6c6 | 5288 | else if (rn == FPSCR_REG) |
5289 | x = gen_pop_fpscr (); | |
7105fb72 | 5290 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && ! TARGET_FPU_SINGLE |
e2142b04 | 5291 | && FP_OR_XD_REGISTER_P (rn)) |
1b61190c | 5292 | { |
e2142b04 | 5293 | if (FP_REGISTER_P (rn) && (rn - FIRST_FP_REG) & 1) |
1b61190c | 5294 | return; |
7014838c | 5295 | x = gen_pop_4 (gen_rtx_REG (DFmode, rn)); |
1b61190c | 5296 | } |
87ed74ef | 5297 | else if (TARGET_SH2E && FP_REGISTER_P (rn)) |
7014838c | 5298 | x = gen_pop_e (gen_rtx_REG (SFmode, rn)); |
b00564da | 5299 | else |
7014838c | 5300 | x = gen_pop (gen_rtx_REG (SImode, rn)); |
9e7454d0 | 5301 | |
8ded0752 | 5302 | x = emit_insn (x); |
7014838c | 5303 | REG_NOTES (x) |
5304 | = gen_rtx_EXPR_LIST (REG_INC, | |
5305 | gen_rtx_REG (SImode, STACK_POINTER_REGNUM), 0); | |
73401833 | 5306 | } |
5307 | ||
8ded0752 | 5308 | /* Generate code to push the regs specified in the mask. */ |
73401833 | 5309 | |
5310 | static void | |
04f04b72 | 5311 | push_regs (HARD_REG_SET *mask, int interrupt_handler) |
73401833 | 5312 | { |
5313 | int i; | |
36a2dfdb | 5314 | int skip_fpscr = 0; |
73401833 | 5315 | |
8ded0752 | 5316 | /* Push PR last; this gives better latencies after the prologue, and |
5317 | candidates for the return delay slot when there are no general | |
5318 | registers pushed. */ | |
87e19636 | 5319 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) |
36a2dfdb | 5320 | { |
5321 | /* If this is an interrupt handler, and the SZ bit varies, | |
5322 | and we have to push any floating point register, we need | |
5323 | to switch to the correct precision first. */ | |
5324 | if (i == FIRST_FP_REG && interrupt_handler && TARGET_FMOVD | |
5325 | && hard_regs_intersect_p (mask, ®_class_contents[DF_REGS])) | |
5326 | { | |
5327 | HARD_REG_SET unsaved; | |
5328 | ||
5329 | push (FPSCR_REG); | |
d767e27e | 5330 | COMPL_HARD_REG_SET (unsaved, *mask); |
36a2dfdb | 5331 | fpscr_set_from_mem (NORMAL_MODE (FP_MODE), unsaved); |
5332 | skip_fpscr = 1; | |
5333 | } | |
5334 | if (i != PR_REG | |
5335 | && (i != FPSCR_REG || ! skip_fpscr) | |
5336 | && TEST_HARD_REG_BIT (*mask, i)) | |
5337 | push (i); | |
5338 | } | |
ed8483c0 | 5339 | if (TEST_HARD_REG_BIT (*mask, PR_REG)) |
8ded0752 | 5340 | push (PR_REG); |
73401833 | 5341 | } |
5342 | ||
8af3db02 | 5343 | /* Calculate how much extra space is needed to save all callee-saved |
5344 | target registers. | |
5345 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
5346 | ||
5347 | static int | |
5348 | shmedia_target_regs_stack_space (HARD_REG_SET *live_regs_mask) | |
5349 | { | |
5350 | int reg; | |
5351 | int stack_space = 0; | |
5352 | int interrupt_handler = sh_cfun_interrupt_handler_p (); | |
5353 | ||
5354 | for (reg = LAST_TARGET_REG; reg >= FIRST_TARGET_REG; reg--) | |
956dd562 | 5355 | if ((! call_really_used_regs[reg] || interrupt_handler) |
8af3db02 | 5356 | && ! TEST_HARD_REG_BIT (*live_regs_mask, reg)) |
5357 | /* Leave space to save this target register on the stack, | |
ada8cf8b | 5358 | in case target register allocation wants to use it. */ |
8af3db02 | 5359 | stack_space += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); |
5360 | return stack_space; | |
5361 | } | |
9e7454d0 | 5362 | |
8af3db02 | 5363 | /* Decide whether we should reserve space for callee-save target registers, |
5364 | in case target register allocation wants to use them. REGS_SAVED is | |
5365 | the space, in bytes, that is already required for register saves. | |
5366 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
5367 | ||
5368 | static int | |
5369 | shmedia_reserve_space_for_target_registers_p (int regs_saved, | |
5370 | HARD_REG_SET *live_regs_mask) | |
5371 | { | |
5372 | if (optimize_size) | |
5373 | return 0; | |
5374 | return shmedia_target_regs_stack_space (live_regs_mask) <= regs_saved; | |
5375 | } | |
5376 | ||
5377 | /* Decide how much space to reserve for callee-save target registers | |
5378 | in case target register allocation wants to use them. | |
5379 | LIVE_REGS_MASK is the register mask calculated by calc_live_regs. */ | |
5380 | ||
5381 | static int | |
5382 | shmedia_target_regs_stack_adjust (HARD_REG_SET *live_regs_mask) | |
5383 | { | |
5384 | if (shmedia_space_reserved_for_target_registers) | |
5385 | return shmedia_target_regs_stack_space (live_regs_mask); | |
5386 | else | |
5387 | return 0; | |
5388 | } | |
5389 | ||
73401833 | 5390 | /* Work out the registers which need to be saved, both as a mask and a |
ed8483c0 | 5391 | count of saved words. Return the count. |
73401833 | 5392 | |
5393 | If doing a pragma interrupt function, then push all regs used by the | |
5394 | function, and if we call another function (we can tell by looking at PR), | |
5395 | make sure that all the regs it clobbers are safe too. */ | |
5396 | ||
ed8483c0 | 5397 | static int |
04f04b72 | 5398 | calc_live_regs (HARD_REG_SET *live_regs_mask) |
73401833 | 5399 | { |
bfa0b1cf | 5400 | unsigned int reg; |
8ded0752 | 5401 | int count; |
49f630e1 | 5402 | int interrupt_handler; |
41fafa66 | 5403 | int pr_live, has_call; |
49f630e1 | 5404 | |
5df3d11f | 5405 | interrupt_handler = sh_cfun_interrupt_handler_p (); |
73401833 | 5406 | |
41fafa66 | 5407 | CLEAR_HARD_REG_SET (*live_regs_mask); |
7105fb72 | 5408 | if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && interrupt_handler |
36a2dfdb | 5409 | && regs_ever_live[FPSCR_REG]) |
156fac8b | 5410 | target_flags &= ~MASK_FPU_SINGLE; |
1b61190c | 5411 | /* If we can save a lot of saves by switching to double mode, do that. */ |
7105fb72 | 5412 | else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && TARGET_FPU_SINGLE) |
1b61190c | 5413 | for (count = 0, reg = FIRST_FP_REG; reg <= LAST_FP_REG; reg += 2) |
5414 | if (regs_ever_live[reg] && regs_ever_live[reg+1] | |
956dd562 | 5415 | && (! call_really_used_regs[reg] |
5416 | || (interrupt_handler && ! pragma_trapa)) | |
1b61190c | 5417 | && ++count > 2) |
5418 | { | |
156fac8b | 5419 | target_flags &= ~MASK_FPU_SINGLE; |
1b61190c | 5420 | break; |
5421 | } | |
0b1c30e1 | 5422 | /* PR_MEDIA_REG is a general purpose register, thus global_alloc already |
5423 | knows how to use it. That means the pseudo originally allocated for | |
5424 | the initial value can become the PR_MEDIA_REG hard register, as seen for | |
5425 | execute/20010122-1.c:test9. */ | |
5426 | if (TARGET_SHMEDIA) | |
ca474d26 | 5427 | /* ??? this function is called from initial_elimination_offset, hence we |
5428 | can't use the result of sh_media_register_for_return here. */ | |
93ea47c4 | 5429 | pr_live = sh_pr_n_sets (); |
0b1c30e1 | 5430 | else |
5431 | { | |
5432 | rtx pr_initial = has_hard_reg_initial_val (Pmode, PR_REG); | |
5433 | pr_live = (pr_initial | |
0a41c525 | 5434 | ? (GET_CODE (pr_initial) != REG |
5435 | || REGNO (pr_initial) != (PR_REG)) | |
5436 | : regs_ever_live[PR_REG]); | |
ca474d26 | 5437 | /* For Shcompact, if not optimizing, we end up with a memory reference |
5438 | using the return address pointer for __builtin_return_address even | |
5439 | though there is no actual need to put the PR register on the stack. */ | |
5440 | pr_live |= regs_ever_live[RETURN_ADDRESS_POINTER_REGNUM]; | |
0b1c30e1 | 5441 | } |
87e19636 | 5442 | /* Force PR to be live if the prologue has to call the SHmedia |
5443 | argument decoder or register saver. */ | |
5444 | if (TARGET_SHCOMPACT | |
5445 | && ((current_function_args_info.call_cookie | |
5446 | & ~ CALL_COOKIE_RET_TRAMP (1)) | |
5447 | || current_function_has_nonlocal_label)) | |
5448 | pr_live = 1; | |
41fafa66 | 5449 | has_call = TARGET_SHMEDIA ? ! leaf_function_p () : pr_live; |
bfa0b1cf | 5450 | for (count = 0, reg = FIRST_PSEUDO_REGISTER; reg-- != 0; ) |
73401833 | 5451 | { |
93ea47c4 | 5452 | if (reg == (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG) |
476444cc | 5453 | ? pr_live |
5454 | : (interrupt_handler && ! pragma_trapa) | |
8ded0752 | 5455 | ? (/* Need to save all the regs ever live. */ |
5456 | (regs_ever_live[reg] | |
956dd562 | 5457 | || (call_really_used_regs[reg] |
5458 | && (! fixed_regs[reg] || reg == MACH_REG || reg == MACL_REG | |
5459 | || reg == PIC_OFFSET_TABLE_REGNUM) | |
41fafa66 | 5460 | && has_call) |
5461 | || (has_call && REGISTER_NATURAL_MODE (reg) == SImode | |
5462 | && (GENERAL_REGISTER_P (reg) || TARGET_REGISTER_P (reg)))) | |
8ded0752 | 5463 | && reg != STACK_POINTER_REGNUM && reg != ARG_POINTER_REGNUM |
5464 | && reg != RETURN_ADDRESS_POINTER_REGNUM | |
88b7f6c6 | 5465 | && reg != T_REG && reg != GBR_REG |
5466 | /* Push fpscr only on targets which have FPU */ | |
5467 | && (reg != FPSCR_REG || TARGET_FPU_ANY)) | |
8ded0752 | 5468 | : (/* Only push those regs which are used and need to be saved. */ |
6977cd64 | 5469 | (TARGET_SHCOMPACT |
5470 | && flag_pic | |
5471 | && current_function_args_info.call_cookie | |
bfa0b1cf | 5472 | && reg == PIC_OFFSET_TABLE_REGNUM) |
956dd562 | 5473 | || (regs_ever_live[reg] && ! call_really_used_regs[reg]) |
805e22b2 | 5474 | || (current_function_calls_eh_return |
bfa0b1cf | 5475 | && (reg == EH_RETURN_DATA_REGNO (0) |
5476 | || reg == EH_RETURN_DATA_REGNO (1) | |
5477 | || reg == EH_RETURN_DATA_REGNO (2) | |
5478 | || reg == EH_RETURN_DATA_REGNO (3))) | |
45550790 | 5479 | || ((reg == MACL_REG || reg == MACH_REG) |
5480 | && regs_ever_live[reg] | |
5481 | && sh_cfun_attr_renesas_p ()) | |
5482 | )) | |
73401833 | 5483 | { |
ed8483c0 | 5484 | SET_HARD_REG_BIT (*live_regs_mask, reg); |
87e19636 | 5485 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); |
5486 | ||
7105fb72 | 5487 | if ((TARGET_SH4 || TARGET_SH2A_DOUBLE || TARGET_SH5) && TARGET_FMOVD |
87e19636 | 5488 | && GET_MODE_CLASS (REGISTER_NATURAL_MODE (reg)) == MODE_FLOAT) |
17bec211 | 5489 | { |
e2142b04 | 5490 | if (FP_REGISTER_P (reg)) |
17bec211 | 5491 | { |
5492 | if (! TARGET_FPU_SINGLE && ! regs_ever_live[reg ^ 1]) | |
5493 | { | |
ed8483c0 | 5494 | SET_HARD_REG_BIT (*live_regs_mask, (reg ^ 1)); |
87e19636 | 5495 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg ^ 1)); |
17bec211 | 5496 | } |
5497 | } | |
87e19636 | 5498 | else if (XD_REGISTER_P (reg)) |
17bec211 | 5499 | { |
5500 | /* Must switch to double mode to access these registers. */ | |
156fac8b | 5501 | target_flags &= ~MASK_FPU_SINGLE; |
17bec211 | 5502 | } |
5503 | } | |
73401833 | 5504 | } |
5505 | } | |
8af3db02 | 5506 | /* If we have a target register optimization pass after prologue / epilogue |
5507 | threading, we need to assume all target registers will be live even if | |
5508 | they aren't now. */ | |
5509 | if (flag_branch_target_load_optimize2 | |
5510 | && TARGET_SAVE_ALL_TARGET_REGS | |
5511 | && shmedia_space_reserved_for_target_registers) | |
5512 | for (reg = LAST_TARGET_REG; reg >= FIRST_TARGET_REG; reg--) | |
956dd562 | 5513 | if ((! call_really_used_regs[reg] || interrupt_handler) |
8af3db02 | 5514 | && ! TEST_HARD_REG_BIT (*live_regs_mask, reg)) |
5515 | { | |
5516 | SET_HARD_REG_BIT (*live_regs_mask, reg); | |
5517 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg)); | |
5518 | } | |
41fafa66 | 5519 | /* If this is an interrupt handler, we don't have any call-clobbered |
5520 | registers we can conveniently use for target register save/restore. | |
5521 | Make sure we save at least one general purpose register when we need | |
5522 | to save target registers. */ | |
5523 | if (interrupt_handler | |
5524 | && hard_regs_intersect_p (live_regs_mask, | |
5525 | ®_class_contents[TARGET_REGS]) | |
5526 | && ! hard_regs_intersect_p (live_regs_mask, | |
5527 | ®_class_contents[GENERAL_REGS])) | |
5528 | { | |
5529 | SET_HARD_REG_BIT (*live_regs_mask, R0_REG); | |
5530 | count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (R0_REG)); | |
5531 | } | |
73401833 | 5532 | |
ed8483c0 | 5533 | return count; |
73401833 | 5534 | } |
5535 | ||
5536 | /* Code to generate prologue and epilogue sequences */ | |
47c009e5 | 5537 | |
4aa67c38 | 5538 | /* PUSHED is the number of bytes that are being pushed on the |
f10edf5a | 5539 | stack for register saves. Return the frame size, padded |
5540 | appropriately so that the stack stays properly aligned. */ | |
9200a51c | 5541 | static HOST_WIDE_INT |
04f04b72 | 5542 | rounded_frame_size (int pushed) |
9200a51c | 5543 | { |
5544 | HOST_WIDE_INT size = get_frame_size (); | |
5545 | HOST_WIDE_INT align = STACK_BOUNDARY / BITS_PER_UNIT; | |
5546 | ||
26576132 | 5547 | return ((size + pushed + align - 1) & -align) - pushed; |
9200a51c | 5548 | } |
5549 | ||
87e19636 | 5550 | /* Choose a call-clobbered target-branch register that remains |
5551 | unchanged along the whole function. We set it up as the return | |
5552 | value in the prologue. */ | |
5553 | int | |
04f04b72 | 5554 | sh_media_register_for_return (void) |
87e19636 | 5555 | { |
5556 | int regno; | |
5557 | int tr0_used; | |
5558 | ||
5559 | if (! current_function_is_leaf) | |
5560 | return -1; | |
41fafa66 | 5561 | if (lookup_attribute ("interrupt_handler", |
5562 | DECL_ATTRIBUTES (current_function_decl))) | |
5563 | return -1; | |
59312820 | 5564 | if (sh_cfun_interrupt_handler_p ()) |
5565 | return -1; | |
87e19636 | 5566 | |
5567 | tr0_used = flag_pic && regs_ever_live[PIC_OFFSET_TABLE_REGNUM]; | |
5568 | ||
5569 | for (regno = FIRST_TARGET_REG + tr0_used; regno <= LAST_TARGET_REG; regno++) | |
956dd562 | 5570 | if (call_really_used_regs[regno] && ! regs_ever_live[regno]) |
87e19636 | 5571 | return regno; |
5572 | ||
5573 | return -1; | |
5574 | } | |
5575 | ||
41fafa66 | 5576 | /* The maximum registers we need to save are: |
5577 | - 62 general purpose registers (r15 is stack pointer, r63 is zero) | |
5578 | - 32 floating point registers (for each pair, we save none, | |
5579 | one single precision value, or a double precision value). | |
5580 | - 8 target registers | |
5581 | - add 1 entry for a delimiter. */ | |
5582 | #define MAX_SAVED_REGS (62+32+8) | |
5583 | ||
5584 | typedef struct save_entry_s | |
5585 | { | |
5586 | unsigned char reg; | |
5587 | unsigned char mode; | |
5588 | short offset; | |
5589 | } save_entry; | |
5590 | ||
5591 | #define MAX_TEMPS 4 | |
5592 | ||
5593 | /* There will be a delimiter entry with VOIDmode both at the start and the | |
5594 | end of a filled in schedule. The end delimiter has the offset of the | |
5595 | save with the smallest (i.e. most negative) offset. */ | |
5596 | typedef struct save_schedule_s | |
5597 | { | |
5598 | save_entry entries[MAX_SAVED_REGS + 2]; | |
5599 | int temps[MAX_TEMPS+1]; | |
5600 | } save_schedule; | |
5601 | ||
5602 | /* Fill in SCHEDULE according to LIVE_REGS_MASK. If RESTORE is nonzero, | |
5603 | use reverse order. Returns the last entry written to (not counting | |
5604 | the delimiter). OFFSET_BASE is a number to be added to all offset | |
5605 | entries. */ | |
9e7454d0 | 5606 | |
41fafa66 | 5607 | static save_entry * |
5608 | sh5_schedule_saves (HARD_REG_SET *live_regs_mask, save_schedule *schedule, | |
5609 | int offset_base) | |
5610 | { | |
5611 | int align, i; | |
5612 | save_entry *entry = schedule->entries; | |
5613 | int tmpx = 0; | |
5614 | int offset; | |
5615 | ||
5616 | if (! current_function_interrupt) | |
5617 | for (i = FIRST_GENERAL_REG; tmpx < MAX_TEMPS && i <= LAST_GENERAL_REG; i++) | |
956dd562 | 5618 | if (call_really_used_regs[i] && ! fixed_regs[i] && i != PR_MEDIA_REG |
41fafa66 | 5619 | && ! FUNCTION_ARG_REGNO_P (i) |
5620 | && i != FIRST_RET_REG | |
4ee9c684 | 5621 | && ! (cfun->static_chain_decl != NULL && i == STATIC_CHAIN_REGNUM) |
41fafa66 | 5622 | && ! (current_function_calls_eh_return |
5623 | && (i == EH_RETURN_STACKADJ_REGNO | |
1b95b63b | 5624 | || ((unsigned) i >= EH_RETURN_DATA_REGNO (0) |
5625 | && (unsigned) i <= EH_RETURN_DATA_REGNO (3))))) | |
41fafa66 | 5626 | schedule->temps[tmpx++] = i; |
5627 | entry->reg = -1; | |
5628 | entry->mode = VOIDmode; | |
5629 | entry->offset = offset_base; | |
5630 | entry++; | |
5631 | /* We loop twice: first, we save 8-byte aligned registers in the | |
5632 | higher addresses, that are known to be aligned. Then, we | |
5633 | proceed to saving 32-bit registers that don't need 8-byte | |
5634 | alignment. | |
5635 | If this is an interrupt function, all registers that need saving | |
5636 | need to be saved in full. moreover, we need to postpone saving | |
5637 | target registers till we have saved some general purpose registers | |
5638 | we can then use as scratch registers. */ | |
5639 | offset = offset_base; | |
5640 | for (align = 1; align >= 0; align--) | |
5641 | { | |
5642 | for (i = FIRST_PSEUDO_REGISTER - 1; i >= 0; i--) | |
5643 | if (TEST_HARD_REG_BIT (*live_regs_mask, i)) | |
5644 | { | |
5645 | enum machine_mode mode = REGISTER_NATURAL_MODE (i); | |
5646 | int reg = i; | |
5647 | ||
5648 | if (current_function_interrupt) | |
5649 | { | |
5650 | if (TARGET_REGISTER_P (i)) | |
5651 | continue; | |
5652 | if (GENERAL_REGISTER_P (i)) | |
5653 | mode = DImode; | |
5654 | } | |
5655 | if (mode == SFmode && (i % 2) == 1 | |
5656 | && ! TARGET_FPU_SINGLE && FP_REGISTER_P (i) | |
5657 | && (TEST_HARD_REG_BIT (*live_regs_mask, (i ^ 1)))) | |
5658 | { | |
5659 | mode = DFmode; | |
5660 | i--; | |
5661 | reg--; | |
5662 | } | |
5663 | ||
5664 | /* If we're doing the aligned pass and this is not aligned, | |
5665 | or we're doing the unaligned pass and this is aligned, | |
5666 | skip it. */ | |
5667 | if ((GET_MODE_SIZE (mode) % (STACK_BOUNDARY / BITS_PER_UNIT) == 0) | |
5668 | != align) | |
5669 | continue; | |
5670 | ||
5671 | if (current_function_interrupt | |
5672 | && GENERAL_REGISTER_P (i) | |
5673 | && tmpx < MAX_TEMPS) | |
5674 | schedule->temps[tmpx++] = i; | |
5675 | ||
5676 | offset -= GET_MODE_SIZE (mode); | |
5677 | entry->reg = i; | |
5678 | entry->mode = mode; | |
5679 | entry->offset = offset; | |
5680 | entry++; | |
5681 | } | |
5682 | if (align && current_function_interrupt) | |
5683 | for (i = LAST_TARGET_REG; i >= FIRST_TARGET_REG; i--) | |
5684 | if (TEST_HARD_REG_BIT (*live_regs_mask, i)) | |
5685 | { | |
5686 | offset -= GET_MODE_SIZE (DImode); | |
5687 | entry->reg = i; | |
5688 | entry->mode = DImode; | |
5689 | entry->offset = offset; | |
5690 | entry++; | |
5691 | } | |
5692 | } | |
5693 | entry->reg = -1; | |
5694 | entry->mode = VOIDmode; | |
5695 | entry->offset = offset; | |
5696 | schedule->temps[tmpx] = -1; | |
5697 | return entry - 1; | |
5698 | } | |
5699 | ||
47c009e5 | 5700 | void |
04f04b72 | 5701 | sh_expand_prologue (void) |
47c009e5 | 5702 | { |
ed8483c0 | 5703 | HARD_REG_SET live_regs_mask; |
0f021c78 | 5704 | int d, i; |
87e19636 | 5705 | int d_rounding = 0; |
1b61190c | 5706 | int save_flags = target_flags; |
09a512b9 | 5707 | int pretend_args; |
47c009e5 | 5708 | |
5df3d11f | 5709 | current_function_interrupt = sh_cfun_interrupt_handler_p (); |
0d687b6d | 5710 | |
97595bfd | 5711 | /* We have pretend args if we had an object sent partially in registers |
73401833 | 5712 | and partially on the stack, e.g. a large structure. */ |
09a512b9 | 5713 | pretend_args = current_function_pretend_args_size; |
5714 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl) | |
5715 | && (NPARM_REGS(SImode) | |
5716 | > current_function_args_info.arg_count[(int) SH_ARG_INT])) | |
5717 | pretend_args = 0; | |
5718 | output_stack_adjust (-pretend_args | |
87e19636 | 5719 | - current_function_args_info.stack_regs * 8, |
41fafa66 | 5720 | stack_pointer_rtx, 0, NULL); |
47c009e5 | 5721 | |
87708ea5 | 5722 | if (TARGET_SHCOMPACT && flag_pic && current_function_args_info.call_cookie) |
5723 | /* We're going to use the PIC register to load the address of the | |
5724 | incoming-argument decoder and/or of the return trampoline from | |
5725 | the GOT, so make sure the PIC register is preserved and | |
5726 | initialized. */ | |
5727 | regs_ever_live[PIC_OFFSET_TABLE_REGNUM] = 1; | |
5728 | ||
87e19636 | 5729 | if (TARGET_SHCOMPACT |
5730 | && (current_function_args_info.call_cookie & ~ CALL_COOKIE_RET_TRAMP(1))) | |
5731 | { | |
5732 | int reg; | |
5733 | ||
5734 | /* First, make all registers with incoming arguments that will | |
5735 | be pushed onto the stack live, so that register renaming | |
5736 | doesn't overwrite them. */ | |
5737 | for (reg = 0; reg < NPARM_REGS (SImode); reg++) | |
5738 | if (CALL_COOKIE_STACKSEQ_GET (current_function_args_info.call_cookie) | |
5739 | >= NPARM_REGS (SImode) - reg) | |
5740 | for (; reg < NPARM_REGS (SImode); reg++) | |
5741 | emit_insn (gen_shcompact_preserve_incoming_args | |
5742 | (gen_rtx_REG (SImode, FIRST_PARM_REG + reg))); | |
5743 | else if (CALL_COOKIE_INT_REG_GET | |
5744 | (current_function_args_info.call_cookie, reg) == 1) | |
5745 | emit_insn (gen_shcompact_preserve_incoming_args | |
5746 | (gen_rtx_REG (SImode, FIRST_PARM_REG + reg))); | |
5747 | ||
5748 | emit_move_insn (gen_rtx_REG (Pmode, MACL_REG), | |
5749 | stack_pointer_rtx); | |
5750 | emit_move_insn (gen_rtx_REG (SImode, R0_REG), | |
5751 | GEN_INT (current_function_args_info.call_cookie)); | |
5752 | emit_move_insn (gen_rtx_REG (SImode, MACH_REG), | |
5753 | gen_rtx_REG (SImode, R0_REG)); | |
5754 | } | |
5755 | else if (TARGET_SHMEDIA) | |
5756 | { | |
5757 | int tr = sh_media_register_for_return (); | |
5758 | ||
5759 | if (tr >= 0) | |
5760 | { | |
5761 | rtx insn = emit_move_insn (gen_rtx_REG (DImode, tr), | |
5762 | gen_rtx_REG (DImode, PR_MEDIA_REG)); | |
5763 | ||
ca474d26 | 5764 | /* ??? We should suppress saving pr when we don't need it, but this |
5765 | is tricky because of builtin_return_address. */ | |
5766 | ||
87e19636 | 5767 | /* If this function only exits with sibcalls, this copy |
5768 | will be flagged as dead. */ | |
5769 | REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, | |
5770 | const0_rtx, | |
5771 | REG_NOTES (insn)); | |
5772 | } | |
5773 | } | |
5774 | ||
bf622d7d | 5775 | /* Emit the code for SETUP_VARARGS. */ |
7ccc713a | 5776 | if (current_function_stdarg) |
47c009e5 | 5777 | { |
09a512b9 | 5778 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl)) |
8ded0752 | 5779 | { |
b00564da | 5780 | /* Push arg regs as if they'd been provided by caller in stack. */ |
5781 | for (i = 0; i < NPARM_REGS(SImode); i++) | |
5782 | { | |
5783 | int rn = NPARM_REGS(SImode) + FIRST_PARM_REG - i - 1; | |
363600c8 | 5784 | rtx insn; |
5785 | ||
9e7454d0 | 5786 | if (i >= (NPARM_REGS(SImode) |
75c2fa85 | 5787 | - current_function_args_info.arg_count[(int) SH_ARG_INT] |
5788 | )) | |
b00564da | 5789 | break; |
363600c8 | 5790 | insn = push (rn); |
5791 | RTX_FRAME_RELATED_P (insn) = 0; | |
b00564da | 5792 | } |
8ded0752 | 5793 | } |
47c009e5 | 5794 | } |
b00564da | 5795 | |
15cb1799 | 5796 | /* If we're supposed to switch stacks at function entry, do so now. */ |
5797 | if (sp_switch) | |
5798 | emit_insn (gen_sp_switch_1 ()); | |
5799 | ||
ed8483c0 | 5800 | d = calc_live_regs (&live_regs_mask); |
1b61190c | 5801 | /* ??? Maybe we could save some switching if we can move a mode switch |
5802 | that already happens to be at the function start into the prologue. */ | |
36a2dfdb | 5803 | if (target_flags != save_flags && ! current_function_interrupt) |
1b61190c | 5804 | emit_insn (gen_toggle_sz ()); |
9e7454d0 | 5805 | |
87e19636 | 5806 | if (TARGET_SH5) |
5807 | { | |
41fafa66 | 5808 | int offset_base, offset; |
5809 | rtx r0 = NULL_RTX; | |
87e19636 | 5810 | int offset_in_r0 = -1; |
5811 | int sp_in_r0 = 0; | |
8af3db02 | 5812 | int tregs_space = shmedia_target_regs_stack_adjust (&live_regs_mask); |
5813 | int total_size, save_size; | |
41fafa66 | 5814 | save_schedule schedule; |
5815 | save_entry *entry; | |
5816 | int *tmp_pnt; | |
5817 | ||
956dd562 | 5818 | if (call_really_used_regs[R0_REG] && ! fixed_regs[R0_REG] |
41fafa66 | 5819 | && ! current_function_interrupt) |
5820 | r0 = gen_rtx_REG (Pmode, R0_REG); | |
8af3db02 | 5821 | |
5822 | /* D is the actual number of bytes that we need for saving registers, | |
5823 | however, in initial_elimination_offset we have committed to using | |
5824 | an additional TREGS_SPACE amount of bytes - in order to keep both | |
5825 | addresses to arguments supplied by the caller and local variables | |
5826 | valid, we must keep this gap. Place it between the incoming | |
5827 | arguments and the actually saved registers in a bid to optimize | |
5828 | locality of reference. */ | |
5829 | total_size = d + tregs_space; | |
5830 | total_size += rounded_frame_size (total_size); | |
5831 | save_size = total_size - rounded_frame_size (d); | |
5832 | if (save_size % (STACK_BOUNDARY / BITS_PER_UNIT)) | |
87e19636 | 5833 | d_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) |
8af3db02 | 5834 | - save_size % (STACK_BOUNDARY / BITS_PER_UNIT)); |
5835 | ||
5836 | /* If adjusting the stack in a single step costs nothing extra, do so. | |
5837 | I.e. either if a single addi is enough, or we need a movi anyway, | |
5838 | and we don't exceed the maximum offset range (the test for the | |
5839 | latter is conservative for simplicity). */ | |
5840 | if (TARGET_SHMEDIA | |
5841 | && (CONST_OK_FOR_I10 (-total_size) | |
5842 | || (! CONST_OK_FOR_I10 (-(save_size + d_rounding)) | |
5843 | && total_size <= 2044))) | |
5844 | d_rounding = total_size - save_size; | |
87e19636 | 5845 | |
41fafa66 | 5846 | offset_base = d + d_rounding; |
8af3db02 | 5847 | |
5848 | output_stack_adjust (-(save_size + d_rounding), stack_pointer_rtx, | |
41fafa66 | 5849 | 0, NULL); |
87e19636 | 5850 | |
41fafa66 | 5851 | sh5_schedule_saves (&live_regs_mask, &schedule, offset_base); |
5852 | tmp_pnt = schedule.temps; | |
5853 | for (entry = &schedule.entries[1]; entry->mode != VOIDmode; entry++) | |
5854 | { | |
5855 | enum machine_mode mode = entry->mode; | |
bfa0b1cf | 5856 | unsigned int reg = entry->reg; |
41fafa66 | 5857 | rtx reg_rtx, mem_rtx, pre_dec = NULL_RTX; |
1685bae3 | 5858 | rtx orig_reg_rtx; |
87e19636 | 5859 | |
41fafa66 | 5860 | offset = entry->offset; |
87e19636 | 5861 | |
41fafa66 | 5862 | reg_rtx = gen_rtx_REG (mode, reg); |
87e19636 | 5863 | |
aa7e75f2 | 5864 | mem_rtx = gen_frame_mem (mode, |
5865 | gen_rtx_PLUS (Pmode, | |
5866 | stack_pointer_rtx, | |
5867 | GEN_INT (offset))); | |
87e19636 | 5868 | |
41fafa66 | 5869 | GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (mem_rtx, 0), try_pre_dec); |
87e19636 | 5870 | |
1a6a7a27 | 5871 | gcc_assert (r0); |
41fafa66 | 5872 | mem_rtx = NULL_RTX; |
87e19636 | 5873 | |
41fafa66 | 5874 | try_pre_dec: |
5875 | do | |
5876 | if (HAVE_PRE_DECREMENT | |
5877 | && (offset_in_r0 - offset == GET_MODE_SIZE (mode) | |
5878 | || mem_rtx == NULL_RTX | |
5879 | || reg == PR_REG || SPECIAL_REGISTER_P (reg))) | |
5880 | { | |
aa7e75f2 | 5881 | pre_dec = gen_frame_mem (mode, gen_rtx_PRE_DEC (Pmode, r0)); |
87e19636 | 5882 | |
41fafa66 | 5883 | GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (pre_dec, 0), |
5884 | pre_dec_ok); | |
87e19636 | 5885 | |
41fafa66 | 5886 | pre_dec = NULL_RTX; |
87e19636 | 5887 | |
41fafa66 | 5888 | break; |
87e19636 | 5889 | |
41fafa66 | 5890 | pre_dec_ok: |
5891 | mem_rtx = NULL_RTX; | |
5892 | offset += GET_MODE_SIZE (mode); | |
5893 | } | |
5894 | while (0); | |
87e19636 | 5895 | |
41fafa66 | 5896 | if (mem_rtx != NULL_RTX) |
5897 | goto addr_ok; | |
87e19636 | 5898 | |
41fafa66 | 5899 | if (offset_in_r0 == -1) |
5900 | { | |
5901 | emit_move_insn (r0, GEN_INT (offset)); | |
5902 | offset_in_r0 = offset; | |
5903 | } | |
5904 | else if (offset != offset_in_r0) | |
5905 | { | |
5906 | emit_move_insn (r0, | |
5907 | gen_rtx_PLUS | |
5908 | (Pmode, r0, | |
5909 | GEN_INT (offset - offset_in_r0))); | |
5910 | offset_in_r0 += offset - offset_in_r0; | |
5911 | } | |
9e7454d0 | 5912 | |
41fafa66 | 5913 | if (pre_dec != NULL_RTX) |
5914 | { | |
5915 | if (! sp_in_r0) | |
87e19636 | 5916 | { |
5917 | emit_move_insn (r0, | |
5918 | gen_rtx_PLUS | |
41fafa66 | 5919 | (Pmode, r0, stack_pointer_rtx)); |
5920 | sp_in_r0 = 1; | |
87e19636 | 5921 | } |
87e19636 | 5922 | |
41fafa66 | 5923 | offset -= GET_MODE_SIZE (mode); |
5924 | offset_in_r0 -= GET_MODE_SIZE (mode); | |
87e19636 | 5925 | |
41fafa66 | 5926 | mem_rtx = pre_dec; |
5927 | } | |
5928 | else if (sp_in_r0) | |
aa7e75f2 | 5929 | mem_rtx = gen_frame_mem (mode, r0); |
41fafa66 | 5930 | else |
aa7e75f2 | 5931 | mem_rtx = gen_frame_mem (mode, |
5932 | gen_rtx_PLUS (Pmode, | |
5933 | stack_pointer_rtx, | |
5934 | r0)); | |
41fafa66 | 5935 | |
5936 | /* We must not use an r0-based address for target-branch | |
5937 | registers or for special registers without pre-dec | |
5938 | memory addresses, since we store their values in r0 | |
5939 | first. */ | |
1a6a7a27 | 5940 | gcc_assert (!TARGET_REGISTER_P (reg) |
5941 | && ((reg != PR_REG && !SPECIAL_REGISTER_P (reg)) | |
5942 | || mem_rtx == pre_dec)); | |
5943 | ||
41fafa66 | 5944 | addr_ok: |
1685bae3 | 5945 | orig_reg_rtx = reg_rtx; |
41fafa66 | 5946 | if (TARGET_REGISTER_P (reg) |
5947 | || ((reg == PR_REG || SPECIAL_REGISTER_P (reg)) | |
5948 | && mem_rtx != pre_dec)) | |
5949 | { | |
5950 | rtx tmp_reg = gen_rtx_REG (GET_MODE (reg_rtx), *tmp_pnt); | |
87e19636 | 5951 | |
41fafa66 | 5952 | emit_move_insn (tmp_reg, reg_rtx); |
87e19636 | 5953 | |
41fafa66 | 5954 | if (REGNO (tmp_reg) == R0_REG) |
5955 | { | |
87e19636 | 5956 | offset_in_r0 = -1; |
5957 | sp_in_r0 = 0; | |
1a6a7a27 | 5958 | gcc_assert (!refers_to_regno_p |
5959 | (R0_REG, R0_REG+1, mem_rtx, (rtx *) 0)); | |
87e19636 | 5960 | } |
5961 | ||
41fafa66 | 5962 | if (*++tmp_pnt <= 0) |
5963 | tmp_pnt = schedule.temps; | |
5964 | ||
5965 | reg_rtx = tmp_reg; | |
87e19636 | 5966 | } |
41fafa66 | 5967 | { |
5968 | rtx insn; | |
5969 | ||
5970 | /* Mark as interesting for dwarf cfi generator */ | |
5971 | insn = emit_move_insn (mem_rtx, reg_rtx); | |
5972 | RTX_FRAME_RELATED_P (insn) = 1; | |
1685bae3 | 5973 | /* If we use an intermediate register for the save, we can't |
5974 | describe this exactly in cfi as a copy of the to-be-saved | |
5975 | register into the temporary register and then the temporary | |
5976 | register on the stack, because the temporary register can | |
5977 | have a different natural size than the to-be-saved register. | |
5978 | Thus, we gloss over the intermediate copy and pretend we do | |
5979 | a direct save from the to-be-saved register. */ | |
5980 | if (REGNO (reg_rtx) != reg) | |
5981 | { | |
5982 | rtx set, note_rtx; | |
5983 | ||
5984 | set = gen_rtx_SET (VOIDmode, mem_rtx, orig_reg_rtx); | |
5985 | note_rtx = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, set, | |
5986 | REG_NOTES (insn)); | |
5987 | REG_NOTES (insn) = note_rtx; | |
5988 | } | |
41fafa66 | 5989 | |
9e7454d0 | 5990 | if (TARGET_SHCOMPACT && (offset_in_r0 != -1)) |
41fafa66 | 5991 | { |
5992 | rtx reg_rtx = gen_rtx_REG (mode, reg); | |
5993 | rtx set, note_rtx; | |
aa7e75f2 | 5994 | rtx mem_rtx = gen_frame_mem (mode, |
5995 | gen_rtx_PLUS (Pmode, | |
5996 | stack_pointer_rtx, | |
5997 | GEN_INT (offset))); | |
41fafa66 | 5998 | |
5999 | set = gen_rtx_SET (VOIDmode, mem_rtx, reg_rtx); | |
6000 | note_rtx = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, set, | |
6001 | REG_NOTES (insn)); | |
6002 | REG_NOTES (insn) = note_rtx; | |
6003 | } | |
6004 | } | |
6005 | } | |
87e19636 | 6006 | |
1a6a7a27 | 6007 | gcc_assert (entry->offset == d_rounding); |
87e19636 | 6008 | } |
6009 | else | |
36a2dfdb | 6010 | push_regs (&live_regs_mask, current_function_interrupt); |
0abf894c | 6011 | |
420261ef | 6012 | if (flag_pic && regs_ever_live[PIC_OFFSET_TABLE_REGNUM]) |
47890aff | 6013 | { |
6014 | rtx insn = get_last_insn (); | |
6015 | rtx last = emit_insn (gen_GOTaddr2picreg ()); | |
6016 | ||
6017 | /* Mark these insns as possibly dead. Sometimes, flow2 may | |
6018 | delete all uses of the PIC register. In this case, let it | |
6019 | delete the initialization too. */ | |
6020 | do | |
6021 | { | |
6022 | insn = NEXT_INSN (insn); | |
6023 | ||
6024 | REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, | |
6025 | const0_rtx, | |
6026 | REG_NOTES (insn)); | |
6027 | } | |
6028 | while (insn != last); | |
6029 | } | |
0abf894c | 6030 | |
87e19636 | 6031 | if (SHMEDIA_REGS_STACK_ADJUST ()) |
6032 | { | |
87e19636 | 6033 | /* This must NOT go through the PLT, otherwise mach and macl |
6034 | may be clobbered. */ | |
59312820 | 6035 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
6036 | (TARGET_FPU_ANY | |
6037 | ? "__GCC_push_shmedia_regs" | |
6038 | : "__GCC_push_shmedia_regs_nofpu"), SFUNC_GOT); | |
87e19636 | 6039 | emit_insn (gen_shmedia_save_restore_regs_compact |
6040 | (GEN_INT (-SHMEDIA_REGS_STACK_ADJUST ()))); | |
6041 | } | |
6042 | ||
36a2dfdb | 6043 | if (target_flags != save_flags && ! current_function_interrupt) |
55a97ca8 | 6044 | { |
6045 | rtx insn = emit_insn (gen_toggle_sz ()); | |
6046 | ||
6047 | /* If we're lucky, a mode switch in the function body will | |
6048 | overwrite fpscr, turning this insn dead. Tell flow this | |
6049 | insn is ok to delete. */ | |
6050 | REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, | |
6051 | const0_rtx, | |
6052 | REG_NOTES (insn)); | |
6053 | } | |
b00564da | 6054 | |
1b61190c | 6055 | target_flags = save_flags; |
6056 | ||
87e19636 | 6057 | output_stack_adjust (-rounded_frame_size (d) + d_rounding, |
41fafa66 | 6058 | stack_pointer_rtx, 0, NULL); |
47c009e5 | 6059 | |
6060 | if (frame_pointer_needed) | |
b9a602c6 | 6061 | frame_insn (GEN_MOV (hard_frame_pointer_rtx, stack_pointer_rtx)); |
87e19636 | 6062 | |
87e19636 | 6063 | if (TARGET_SHCOMPACT |
6064 | && (current_function_args_info.call_cookie & ~ CALL_COOKIE_RET_TRAMP(1))) | |
6065 | { | |
6066 | /* This must NOT go through the PLT, otherwise mach and macl | |
6067 | may be clobbered. */ | |
59312820 | 6068 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
6069 | "__GCC_shcompact_incoming_args", SFUNC_GOT); | |
87e19636 | 6070 | emit_insn (gen_shcompact_incoming_args ()); |
6071 | } | |
47c009e5 | 6072 | } |
6073 | ||
6074 | void | |
66209063 | 6075 | sh_expand_epilogue (bool sibcall_p) |
47c009e5 | 6076 | { |
ed8483c0 | 6077 | HARD_REG_SET live_regs_mask; |
0f021c78 | 6078 | int d, i; |
87e19636 | 6079 | int d_rounding = 0; |
47c009e5 | 6080 | |
1b61190c | 6081 | int save_flags = target_flags; |
8af3db02 | 6082 | int frame_size, save_size; |
36a2dfdb | 6083 | int fpscr_deferred = 0; |
66209063 | 6084 | int e = sibcall_p ? -1 : 1; |
ba46e0eb | 6085 | |
ed8483c0 | 6086 | d = calc_live_regs (&live_regs_mask); |
ba46e0eb | 6087 | |
8af3db02 | 6088 | save_size = d; |
6089 | frame_size = rounded_frame_size (d); | |
6090 | ||
6091 | if (TARGET_SH5) | |
6092 | { | |
6093 | int tregs_space = shmedia_target_regs_stack_adjust (&live_regs_mask); | |
6094 | int total_size; | |
6095 | if (d % (STACK_BOUNDARY / BITS_PER_UNIT)) | |
6096 | d_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) | |
6097 | - d % (STACK_BOUNDARY / BITS_PER_UNIT)); | |
6098 | ||
6099 | total_size = d + tregs_space; | |
6100 | total_size += rounded_frame_size (total_size); | |
6101 | save_size = total_size - frame_size; | |
6102 | ||
6103 | /* If adjusting the stack in a single step costs nothing extra, do so. | |
6104 | I.e. either if a single addi is enough, or we need a movi anyway, | |
6105 | and we don't exceed the maximum offset range (the test for the | |
6106 | latter is conservative for simplicity). */ | |
6107 | if (TARGET_SHMEDIA | |
6108 | && ! frame_pointer_needed | |
6109 | && (CONST_OK_FOR_I10 (total_size) | |
6110 | || (! CONST_OK_FOR_I10 (save_size + d_rounding) | |
6111 | && total_size <= 2044))) | |
6112 | d_rounding = frame_size; | |
87e19636 | 6113 | |
8af3db02 | 6114 | frame_size -= d_rounding; |
6115 | } | |
1504eb7e | 6116 | |
47c009e5 | 6117 | if (frame_pointer_needed) |
64623096 | 6118 | { |
840750c5 | 6119 | /* We must avoid scheduling the epilogue with previous basic blocks |
6120 | when exception handling is enabled. See PR/18032. */ | |
6121 | if (flag_exceptions) | |
6122 | emit_insn (gen_blockage ()); | |
b9a602c6 | 6123 | output_stack_adjust (frame_size, hard_frame_pointer_rtx, e, |
6124 | &live_regs_mask); | |
b5a24772 | 6125 | |
6126 | /* We must avoid moving the stack pointer adjustment past code | |
6127 | which reads from the local frame, else an interrupt could | |
6128 | occur after the SP adjustment and clobber data in the local | |
6129 | frame. */ | |
6130 | emit_insn (gen_blockage ()); | |
b9a602c6 | 6131 | emit_insn (GEN_MOV (stack_pointer_rtx, hard_frame_pointer_rtx)); |
64623096 | 6132 | } |
ba46e0eb | 6133 | else if (frame_size) |
b5a24772 | 6134 | { |
6135 | /* We must avoid moving the stack pointer adjustment past code | |
6136 | which reads from the local frame, else an interrupt could | |
6137 | occur after the SP adjustment and clobber data in the local | |
6138 | frame. */ | |
6139 | emit_insn (gen_blockage ()); | |
66209063 | 6140 | output_stack_adjust (frame_size, stack_pointer_rtx, e, &live_regs_mask); |
b5a24772 | 6141 | } |
47c009e5 | 6142 | |
87e19636 | 6143 | if (SHMEDIA_REGS_STACK_ADJUST ()) |
6144 | { | |
59312820 | 6145 | function_symbol (gen_rtx_REG (Pmode, R0_REG), |
6146 | (TARGET_FPU_ANY | |
6147 | ? "__GCC_pop_shmedia_regs" | |
6148 | : "__GCC_pop_shmedia_regs_nofpu"), SFUNC_GOT); | |
87e19636 | 6149 | /* This must NOT go through the PLT, otherwise mach and macl |
6150 | may be clobbered. */ | |
6151 | emit_insn (gen_shmedia_save_restore_regs_compact | |
6152 | (GEN_INT (SHMEDIA_REGS_STACK_ADJUST ()))); | |
6153 | } | |
6154 | ||
1504eb7e | 6155 | /* Pop all the registers. */ |
97595bfd | 6156 | |
36a2dfdb | 6157 | if (target_flags != save_flags && ! current_function_interrupt) |
1b61190c | 6158 | emit_insn (gen_toggle_sz ()); |
87e19636 | 6159 | if (TARGET_SH5) |
6160 | { | |
41fafa66 | 6161 | int offset_base, offset; |
87e19636 | 6162 | int offset_in_r0 = -1; |
6163 | int sp_in_r0 = 0; | |
87e19636 | 6164 | rtx r0 = gen_rtx_REG (Pmode, R0_REG); |
41fafa66 | 6165 | save_schedule schedule; |
6166 | save_entry *entry; | |
6167 | int *tmp_pnt; | |
ca474d26 | 6168 | |
41fafa66 | 6169 | entry = sh5_schedule_saves (&live_regs_mask, &schedule, d_rounding); |
6170 | offset_base = -entry[1].offset + d_rounding; | |
6171 | tmp_pnt = schedule.temps; | |
6172 | for (; entry->mode != VOIDmode; entry--) | |
6173 | { | |
6174 | enum machine_mode mode = entry->mode; | |
6175 | int reg = entry->reg; | |
6176 | rtx reg_rtx, mem_rtx, post_inc = NULL_RTX, insn; | |
87e19636 | 6177 | |
41fafa66 | 6178 | offset = offset_base + entry->offset; |
6179 | reg_rtx = gen_rtx_REG (mode, reg); | |
87e19636 | 6180 | |
aa7e75f2 | 6181 | mem_rtx = gen_frame_mem (mode, |
6182 | gen_rtx_PLUS (Pmode, | |
6183 | stack_pointer_rtx, | |
6184 | GEN_INT (offset))); | |
87e19636 | 6185 | |
41fafa66 | 6186 | GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (mem_rtx, 0), try_post_inc); |
87e19636 | 6187 | |
41fafa66 | 6188 | mem_rtx = NULL_RTX; |
87e19636 | 6189 | |
41fafa66 | 6190 | try_post_inc: |
6191 | do | |
6192 | if (HAVE_POST_INCREMENT | |
6193 | && (offset == offset_in_r0 | |
6194 | || (offset + GET_MODE_SIZE (mode) != d + d_rounding | |
6195 | && mem_rtx == NULL_RTX) | |
6196 | || reg == PR_REG || SPECIAL_REGISTER_P (reg))) | |
6197 | { | |
aa7e75f2 | 6198 | post_inc = gen_frame_mem (mode, gen_rtx_POST_INC (Pmode, r0)); |
87e19636 | 6199 | |
41fafa66 | 6200 | GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (post_inc, 0), |
6201 | post_inc_ok); | |
87e19636 | 6202 | |
41fafa66 | 6203 | post_inc = NULL_RTX; |
87e19636 | 6204 | |
41fafa66 | 6205 | break; |
9e7454d0 | 6206 | |
41fafa66 | 6207 | post_inc_ok: |
6208 | mem_rtx = NULL_RTX; | |
6209 | } | |
6210 | while (0); | |
9e7454d0 | 6211 | |
41fafa66 | 6212 | if (mem_rtx != NULL_RTX) |
6213 | goto addr_ok; | |
87e19636 | 6214 | |
41fafa66 | 6215 | if (offset_in_r0 == -1) |
6216 | { | |
6217 | emit_move_insn (r0, GEN_INT (offset)); | |
6218 | offset_in_r0 = offset; | |
6219 | } | |
6220 | else if (offset != offset_in_r0) | |
6221 | { | |
6222 | emit_move_insn (r0, | |
6223 | gen_rtx_PLUS | |
6224 | (Pmode, r0, | |
6225 | GEN_INT (offset - offset_in_r0))); | |
6226 | offset_in_r0 += offset - offset_in_r0; | |
6227 | } | |
9e7454d0 | 6228 | |
41fafa66 | 6229 | if (post_inc != NULL_RTX) |
6230 | { | |
6231 | if (! sp_in_r0) | |
87e19636 | 6232 | { |
6233 | emit_move_insn (r0, | |
6234 | gen_rtx_PLUS | |
41fafa66 | 6235 | (Pmode, r0, stack_pointer_rtx)); |
6236 | sp_in_r0 = 1; | |
87e19636 | 6237 | } |
9e7454d0 | 6238 | |
41fafa66 | 6239 | mem_rtx = post_inc; |
87e19636 | 6240 | |
41fafa66 | 6241 | offset_in_r0 += GET_MODE_SIZE (mode); |
6242 | } | |
6243 | else if (sp_in_r0) | |
aa7e75f2 | 6244 | mem_rtx = gen_frame_mem (mode, r0); |
41fafa66 | 6245 | else |
aa7e75f2 | 6246 | mem_rtx = gen_frame_mem (mode, |
6247 | gen_rtx_PLUS (Pmode, | |
6248 | stack_pointer_rtx, | |
6249 | r0)); | |
87e19636 | 6250 | |
1a6a7a27 | 6251 | gcc_assert ((reg != PR_REG && !SPECIAL_REGISTER_P (reg)) |
6252 | || mem_rtx == post_inc); | |
87e19636 | 6253 | |
41fafa66 | 6254 | addr_ok: |
6255 | if ((reg == PR_REG || SPECIAL_REGISTER_P (reg)) | |
6256 | && mem_rtx != post_inc) | |
6257 | { | |
6258 | insn = emit_move_insn (r0, mem_rtx); | |
6259 | mem_rtx = r0; | |
6260 | } | |
6261 | else if (TARGET_REGISTER_P (reg)) | |
6262 | { | |
6263 | rtx tmp_reg = gen_rtx_REG (mode, *tmp_pnt); | |
6264 | ||
6265 | /* Give the scheduler a bit of freedom by using up to | |
6266 | MAX_TEMPS registers in a round-robin fashion. */ | |
6267 | insn = emit_move_insn (tmp_reg, mem_rtx); | |
6268 | mem_rtx = tmp_reg; | |
6269 | if (*++tmp_pnt < 0) | |
6270 | tmp_pnt = schedule.temps; | |
87e19636 | 6271 | } |
6272 | ||
41fafa66 | 6273 | insn = emit_move_insn (reg_rtx, mem_rtx); |
ca474d26 | 6274 | if (reg == PR_MEDIA_REG && sh_media_register_for_return () >= 0) |
6275 | /* This is dead, unless we return with a sibcall. */ | |
6276 | REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, | |
6277 | const0_rtx, | |
6278 | REG_NOTES (insn)); | |
41fafa66 | 6279 | } |
6280 | ||
1a6a7a27 | 6281 | gcc_assert (entry->offset + offset_base == d + d_rounding); |
87e19636 | 6282 | } |
8af3db02 | 6283 | else /* ! TARGET_SH5 */ |
47c009e5 | 6284 | { |
8af3db02 | 6285 | save_size = 0; |
6286 | if (TEST_HARD_REG_BIT (live_regs_mask, PR_REG)) | |
6287 | pop (PR_REG); | |
6288 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
6289 | { | |
6290 | int j = (FIRST_PSEUDO_REGISTER - 1) - i; | |
9e7454d0 | 6291 | |
8af3db02 | 6292 | if (j == FPSCR_REG && current_function_interrupt && TARGET_FMOVD |
6293 | && hard_regs_intersect_p (&live_regs_mask, | |
6294 | ®_class_contents[DF_REGS])) | |
6295 | fpscr_deferred = 1; | |
6296 | else if (j != PR_REG && TEST_HARD_REG_BIT (live_regs_mask, j)) | |
6297 | pop (j); | |
6298 | if (j == FIRST_FP_REG && fpscr_deferred) | |
6299 | pop (FPSCR_REG); | |
87e19636 | 6300 | |
8af3db02 | 6301 | } |
47c009e5 | 6302 | } |
36a2dfdb | 6303 | if (target_flags != save_flags && ! current_function_interrupt) |
1b61190c | 6304 | emit_insn (gen_toggle_sz ()); |
6305 | target_flags = save_flags; | |
47c009e5 | 6306 | |
09a512b9 | 6307 | output_stack_adjust (current_function_pretend_args_size |
8af3db02 | 6308 | + save_size + d_rounding |
87e19636 | 6309 | + current_function_args_info.stack_regs * 8, |
66209063 | 6310 | stack_pointer_rtx, e, NULL); |
15cb1799 | 6311 | |
805e22b2 | 6312 | if (current_function_calls_eh_return) |
6313 | emit_insn (GEN_ADD3 (stack_pointer_rtx, stack_pointer_rtx, | |
6314 | EH_RETURN_STACKADJ_RTX)); | |
6315 | ||
15cb1799 | 6316 | /* Switch back to the normal stack if necessary. */ |
6317 | if (sp_switch) | |
6318 | emit_insn (gen_sp_switch_2 ()); | |
6dc67d01 | 6319 | |
6320 | /* Tell flow the insn that pops PR isn't dead. */ | |
87e19636 | 6321 | /* PR_REG will never be live in SHmedia mode, and we don't need to |
6322 | USE PR_MEDIA_REG, since it will be explicitly copied to TR0_REG | |
6323 | by the return pattern. */ | |
ed8483c0 | 6324 | if (TEST_HARD_REG_BIT (live_regs_mask, PR_REG)) |
6dc67d01 | 6325 | emit_insn (gen_rtx_USE (VOIDmode, gen_rtx_REG (SImode, PR_REG))); |
47c009e5 | 6326 | } |
6327 | ||
303d40e1 | 6328 | static int sh_need_epilogue_known = 0; |
6329 | ||
6330 | int | |
04f04b72 | 6331 | sh_need_epilogue (void) |
303d40e1 | 6332 | { |
6333 | if (! sh_need_epilogue_known) | |
6334 | { | |
6335 | rtx epilogue; | |
6336 | ||
6337 | start_sequence (); | |
66209063 | 6338 | sh_expand_epilogue (0); |
31d3e01c | 6339 | epilogue = get_insns (); |
303d40e1 | 6340 | end_sequence (); |
31d3e01c | 6341 | sh_need_epilogue_known = (epilogue == NULL ? -1 : 1); |
303d40e1 | 6342 | } |
6343 | return sh_need_epilogue_known > 0; | |
6344 | } | |
6345 | ||
805e22b2 | 6346 | /* Emit code to change the current function's return address to RA. |
6347 | TEMP is available as a scratch register, if needed. */ | |
6348 | ||
6349 | void | |
04f04b72 | 6350 | sh_set_return_address (rtx ra, rtx tmp) |
805e22b2 | 6351 | { |
ed8483c0 | 6352 | HARD_REG_SET live_regs_mask; |
805e22b2 | 6353 | int d; |
805e22b2 | 6354 | int pr_reg = TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG; |
6355 | int pr_offset; | |
6356 | ||
ed8483c0 | 6357 | d = calc_live_regs (&live_regs_mask); |
805e22b2 | 6358 | |
6359 | /* If pr_reg isn't life, we can set it (or the register given in | |
6360 | sh_media_register_for_return) directly. */ | |
ed8483c0 | 6361 | if (! TEST_HARD_REG_BIT (live_regs_mask, pr_reg)) |
805e22b2 | 6362 | { |
6363 | rtx rr; | |
6364 | ||
6365 | if (TARGET_SHMEDIA) | |
6366 | { | |
6367 | int rr_regno = sh_media_register_for_return (); | |
6368 | ||
6369 | if (rr_regno < 0) | |
6370 | rr_regno = pr_reg; | |
6371 | ||
6372 | rr = gen_rtx_REG (DImode, rr_regno); | |
6373 | } | |
6374 | else | |
6375 | rr = gen_rtx_REG (SImode, pr_reg); | |
6376 | ||
6377 | emit_insn (GEN_MOV (rr, ra)); | |
6378 | /* Tell flow the register for return isn't dead. */ | |
6379 | emit_insn (gen_rtx_USE (VOIDmode, rr)); | |
6380 | return; | |
6381 | } | |
6382 | ||
6383 | if (TARGET_SH5) | |
6384 | { | |
805e22b2 | 6385 | int offset; |
41fafa66 | 6386 | save_schedule schedule; |
6387 | save_entry *entry; | |
9e7454d0 | 6388 | |
41fafa66 | 6389 | entry = sh5_schedule_saves (&live_regs_mask, &schedule, 0); |
6390 | offset = entry[1].offset; | |
6391 | for (; entry->mode != VOIDmode; entry--) | |
6392 | if (entry->reg == pr_reg) | |
6393 | goto found; | |
805e22b2 | 6394 | |
6395 | /* We can't find pr register. */ | |
1a6a7a27 | 6396 | gcc_unreachable (); |
805e22b2 | 6397 | |
6398 | found: | |
41fafa66 | 6399 | offset = entry->offset - offset; |
6400 | pr_offset = (rounded_frame_size (d) + offset | |
805e22b2 | 6401 | + SHMEDIA_REGS_STACK_ADJUST ()); |
6402 | } | |
6403 | else | |
41fafa66 | 6404 | pr_offset = rounded_frame_size (d); |
805e22b2 | 6405 | |
6406 | emit_insn (GEN_MOV (tmp, GEN_INT (pr_offset))); | |
b9a602c6 | 6407 | emit_insn (GEN_ADD3 (tmp, tmp, hard_frame_pointer_rtx)); |
805e22b2 | 6408 | |
aa7e75f2 | 6409 | tmp = gen_frame_mem (Pmode, tmp); |
805e22b2 | 6410 | emit_insn (GEN_MOV (tmp, ra)); |
6411 | } | |
6412 | ||
73401833 | 6413 | /* Clear variables at function end. */ |
6414 | ||
17d9b0c3 | 6415 | static void |
04f04b72 | 6416 | sh_output_function_epilogue (FILE *file ATTRIBUTE_UNUSED, |
6417 | HOST_WIDE_INT size ATTRIBUTE_UNUSED) | |
73401833 | 6418 | { |
15cb1799 | 6419 | trap_exit = pragma_interrupt = pragma_trapa = pragma_nosave_low_regs = 0; |
303d40e1 | 6420 | sh_need_epilogue_known = 0; |
15cb1799 | 6421 | sp_switch = NULL_RTX; |
73401833 | 6422 | } |
6423 | ||
45550790 | 6424 | static rtx |
04f04b72 | 6425 | sh_builtin_saveregs (void) |
b00564da | 6426 | { |
b00564da | 6427 | /* First unnamed integer register. */ |
6428 | int first_intreg = current_function_args_info.arg_count[(int) SH_ARG_INT]; | |
6429 | /* Number of integer registers we need to save. */ | |
6430 | int n_intregs = MAX (0, NPARM_REGS (SImode) - first_intreg); | |
6431 | /* First unnamed SFmode float reg */ | |
6432 | int first_floatreg = current_function_args_info.arg_count[(int) SH_ARG_FLOAT]; | |
6433 | /* Number of SFmode float regs to save. */ | |
6434 | int n_floatregs = MAX (0, NPARM_REGS (SFmode) - first_floatreg); | |
b7dbbdb2 | 6435 | rtx regbuf, fpregs; |
ab6ab77e | 6436 | int bufsize, regno; |
6437 | HOST_WIDE_INT alias_set; | |
b00564da | 6438 | |
87e19636 | 6439 | if (TARGET_SH5) |
6440 | { | |
6441 | if (n_intregs) | |
6442 | { | |
6443 | int pushregs = n_intregs; | |
6444 | ||
6445 | while (pushregs < NPARM_REGS (SImode) - 1 | |
6446 | && (CALL_COOKIE_INT_REG_GET | |
6447 | (current_function_args_info.call_cookie, | |
6448 | NPARM_REGS (SImode) - pushregs) | |
6449 | == 1)) | |
6450 | { | |
6451 | current_function_args_info.call_cookie | |
6452 | &= ~ CALL_COOKIE_INT_REG (NPARM_REGS (SImode) | |
6453 | - pushregs, 1); | |
6454 | pushregs++; | |
6455 | } | |
6456 | ||
6457 | if (pushregs == NPARM_REGS (SImode)) | |
6458 | current_function_args_info.call_cookie | |
6459 | |= (CALL_COOKIE_INT_REG (0, 1) | |
6460 | | CALL_COOKIE_STACKSEQ (pushregs - 1)); | |
6461 | else | |
6462 | current_function_args_info.call_cookie | |
6463 | |= CALL_COOKIE_STACKSEQ (pushregs); | |
6464 | ||
6465 | current_function_pretend_args_size += 8 * n_intregs; | |
6466 | } | |
6467 | if (TARGET_SHCOMPACT) | |
6468 | return const0_rtx; | |
6469 | } | |
9e7454d0 | 6470 | |
87ed74ef | 6471 | if (! TARGET_SH2E && ! TARGET_SH4 && ! TARGET_SH5) |
87e19636 | 6472 | { |
6473 | error ("__builtin_saveregs not supported by this subtarget"); | |
6474 | return const0_rtx; | |
6475 | } | |
6476 | ||
6477 | if (TARGET_SHMEDIA) | |
6478 | n_floatregs = 0; | |
6479 | ||
9df03d69 | 6480 | /* Allocate block of memory for the regs. */ |
b00564da | 6481 | /* ??? If n_intregs + n_floatregs == 0, should we allocate at least 1 byte? |
6482 | Or can assign_stack_local accept a 0 SIZE argument? */ | |
6483 | bufsize = (n_intregs * UNITS_PER_WORD) + (n_floatregs * UNITS_PER_WORD); | |
6484 | ||
87e19636 | 6485 | if (TARGET_SHMEDIA) |
aa7e75f2 | 6486 | regbuf = gen_frame_mem (BLKmode, gen_rtx_REG (Pmode, ARG_POINTER_REGNUM)); |
65f753eb | 6487 | else if (n_floatregs & 1) |
6488 | { | |
6489 | rtx addr; | |
6490 | ||
6491 | regbuf = assign_stack_local (BLKmode, bufsize + UNITS_PER_WORD, 0); | |
6492 | addr = copy_to_mode_reg (Pmode, XEXP (regbuf, 0)); | |
6493 | emit_insn (gen_iorsi3 (addr, addr, GEN_INT (UNITS_PER_WORD))); | |
6494 | regbuf = change_address (regbuf, BLKmode, addr); | |
6495 | } | |
faffb001 | 6496 | else if (STACK_BOUNDARY < 64 && TARGET_FPU_DOUBLE && n_floatregs) |
6497 | { | |
6498 | rtx addr, mask; | |
6499 | ||
6500 | regbuf = assign_stack_local (BLKmode, bufsize + UNITS_PER_WORD, 0); | |
6501 | addr = copy_to_mode_reg (Pmode, plus_constant (XEXP (regbuf, 0), 4)); | |
6502 | mask = copy_to_mode_reg (Pmode, GEN_INT (-8)); | |
6503 | emit_insn (gen_andsi3 (addr, addr, mask)); | |
6504 | regbuf = change_address (regbuf, BLKmode, addr); | |
6505 | } | |
87e19636 | 6506 | else |
faffb001 | 6507 | regbuf = assign_stack_local (BLKmode, bufsize, TARGET_FPU_DOUBLE ? 64 : 0); |
d1de1e66 | 6508 | alias_set = get_varargs_alias_set (); |
ab6ab77e | 6509 | set_mem_alias_set (regbuf, alias_set); |
b00564da | 6510 | |
6511 | /* Save int args. | |
6512 | This is optimized to only save the regs that are necessary. Explicitly | |
6513 | named args need not be saved. */ | |
6514 | if (n_intregs > 0) | |
6515 | move_block_from_reg (BASE_ARG_REG (SImode) + first_intreg, | |
e513d163 | 6516 | adjust_address (regbuf, BLKmode, |
6517 | n_floatregs * UNITS_PER_WORD), | |
530178a9 | 6518 | n_intregs); |
b00564da | 6519 | |
87e19636 | 6520 | if (TARGET_SHMEDIA) |
6521 | /* Return the address of the regbuf. */ | |
6522 | return XEXP (regbuf, 0); | |
6523 | ||
b00564da | 6524 | /* Save float args. |
6525 | This is optimized to only save the regs that are necessary. Explicitly | |
6526 | named args need not be saved. | |
6527 | We explicitly build a pointer to the buffer because it halves the insn | |
6528 | count when not optimizing (otherwise the pointer is built for each reg | |
8ded0752 | 6529 | saved). |
6530 | We emit the moves in reverse order so that we can use predecrement. */ | |
b00564da | 6531 | |
dcf3f9fe | 6532 | fpregs = copy_to_mode_reg (Pmode, |
6533 | plus_constant (XEXP (regbuf, 0), | |
6534 | n_floatregs * UNITS_PER_WORD)); | |
7105fb72 | 6535 | if (TARGET_SH4 || TARGET_SH2A_DOUBLE) |
1b61190c | 6536 | { |
d1de1e66 | 6537 | rtx mem; |
1b61190c | 6538 | for (regno = NPARM_REGS (DFmode) - 2; regno >= first_floatreg; regno -= 2) |
6539 | { | |
6540 | emit_insn (gen_addsi3 (fpregs, fpregs, | |
6541 | GEN_INT (-2 * UNITS_PER_WORD))); | |
aa7e75f2 | 6542 | mem = change_address (regbuf, DFmode, fpregs); |
9e7454d0 | 6543 | emit_move_insn (mem, |
407d9153 | 6544 | gen_rtx_REG (DFmode, BASE_ARG_REG (DFmode) + regno)); |
1b61190c | 6545 | } |
6546 | regno = first_floatreg; | |
6547 | if (regno & 1) | |
6548 | { | |
d767e27e | 6549 | emit_insn (gen_addsi3 (fpregs, fpregs, GEN_INT (-UNITS_PER_WORD))); |
aa7e75f2 | 6550 | mem = change_address (regbuf, SFmode, fpregs); |
d1de1e66 | 6551 | emit_move_insn (mem, |
407d9153 | 6552 | gen_rtx_REG (SFmode, BASE_ARG_REG (SFmode) + regno |
1b61190c | 6553 | - (TARGET_LITTLE_ENDIAN != 0))); |
6554 | } | |
6555 | } | |
6556 | else | |
8ded0752 | 6557 | for (regno = NPARM_REGS (SFmode) - 1; regno >= first_floatreg; regno--) |
6558 | { | |
d1de1e66 | 6559 | rtx mem; |
ab6ab77e | 6560 | |
d767e27e | 6561 | emit_insn (gen_addsi3 (fpregs, fpregs, GEN_INT (-UNITS_PER_WORD))); |
aa7e75f2 | 6562 | mem = change_address (regbuf, SFmode, fpregs); |
d1de1e66 | 6563 | emit_move_insn (mem, |
7014838c | 6564 | gen_rtx_REG (SFmode, BASE_ARG_REG (SFmode) + regno)); |
8ded0752 | 6565 | } |
b00564da | 6566 | |
6567 | /* Return the address of the regbuf. */ | |
6568 | return XEXP (regbuf, 0); | |
6569 | } | |
6570 | ||
d1de1e66 | 6571 | /* Define the `__builtin_va_list' type for the ABI. */ |
6572 | ||
2e15d750 | 6573 | static tree |
6574 | sh_build_builtin_va_list (void) | |
d1de1e66 | 6575 | { |
6576 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
6577 | tree record; | |
6578 | ||
45550790 | 6579 | if (TARGET_SH5 || (! TARGET_SH2E && ! TARGET_SH4) |
6580 | || TARGET_HITACHI || sh_cfun_attr_renesas_p ()) | |
d1de1e66 | 6581 | return ptr_type_node; |
6582 | ||
89fdb407 | 6583 | record = (*lang_hooks.types.make_type) (RECORD_TYPE); |
d1de1e66 | 6584 | |
6585 | f_next_o = build_decl (FIELD_DECL, get_identifier ("__va_next_o"), | |
6586 | ptr_type_node); | |
6587 | f_next_o_limit = build_decl (FIELD_DECL, | |
6588 | get_identifier ("__va_next_o_limit"), | |
6589 | ptr_type_node); | |
6590 | f_next_fp = build_decl (FIELD_DECL, get_identifier ("__va_next_fp"), | |
6591 | ptr_type_node); | |
6592 | f_next_fp_limit = build_decl (FIELD_DECL, | |
6593 | get_identifier ("__va_next_fp_limit"), | |
6594 | ptr_type_node); | |
6595 | f_next_stack = build_decl (FIELD_DECL, get_identifier ("__va_next_stack"), | |
6596 | ptr_type_node); | |
6597 | ||
6598 | DECL_FIELD_CONTEXT (f_next_o) = record; | |
6599 | DECL_FIELD_CONTEXT (f_next_o_limit) = record; | |
6600 | DECL_FIELD_CONTEXT (f_next_fp) = record; | |
6601 | DECL_FIELD_CONTEXT (f_next_fp_limit) = record; | |
6602 | DECL_FIELD_CONTEXT (f_next_stack) = record; | |
6603 | ||
6604 | TYPE_FIELDS (record) = f_next_o; | |
6605 | TREE_CHAIN (f_next_o) = f_next_o_limit; | |
6606 | TREE_CHAIN (f_next_o_limit) = f_next_fp; | |
6607 | TREE_CHAIN (f_next_fp) = f_next_fp_limit; | |
6608 | TREE_CHAIN (f_next_fp_limit) = f_next_stack; | |
6609 | ||
6610 | layout_type (record); | |
6611 | ||
6612 | return record; | |
6613 | } | |
6614 | ||
6615 | /* Implement `va_start' for varargs and stdarg. */ | |
6616 | ||
6617 | void | |
04f04b72 | 6618 | sh_va_start (tree valist, rtx nextarg) |
d1de1e66 | 6619 | { |
6620 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
6621 | tree next_o, next_o_limit, next_fp, next_fp_limit, next_stack; | |
6622 | tree t, u; | |
6623 | int nfp, nint; | |
6624 | ||
87e19636 | 6625 | if (TARGET_SH5) |
6626 | { | |
6627 | expand_builtin_saveregs (); | |
7df226a2 | 6628 | std_expand_builtin_va_start (valist, nextarg); |
87e19636 | 6629 | return; |
6630 | } | |
6631 | ||
45550790 | 6632 | if ((! TARGET_SH2E && ! TARGET_SH4) |
6633 | || TARGET_HITACHI || sh_cfun_attr_renesas_p ()) | |
d1de1e66 | 6634 | { |
7df226a2 | 6635 | std_expand_builtin_va_start (valist, nextarg); |
d1de1e66 | 6636 | return; |
6637 | } | |
6638 | ||
6639 | f_next_o = TYPE_FIELDS (va_list_type_node); | |
6640 | f_next_o_limit = TREE_CHAIN (f_next_o); | |
6641 | f_next_fp = TREE_CHAIN (f_next_o_limit); | |
6642 | f_next_fp_limit = TREE_CHAIN (f_next_fp); | |
6643 | f_next_stack = TREE_CHAIN (f_next_fp_limit); | |
6644 | ||
6374121b | 6645 | next_o = build (COMPONENT_REF, TREE_TYPE (f_next_o), valist, f_next_o, |
6646 | NULL_TREE); | |
d1de1e66 | 6647 | next_o_limit = build (COMPONENT_REF, TREE_TYPE (f_next_o_limit), |
6374121b | 6648 | valist, f_next_o_limit, NULL_TREE); |
6649 | next_fp = build (COMPONENT_REF, TREE_TYPE (f_next_fp), valist, f_next_fp, | |
6650 | NULL_TREE); | |
d1de1e66 | 6651 | next_fp_limit = build (COMPONENT_REF, TREE_TYPE (f_next_fp_limit), |
6374121b | 6652 | valist, f_next_fp_limit, NULL_TREE); |
d1de1e66 | 6653 | next_stack = build (COMPONENT_REF, TREE_TYPE (f_next_stack), |
6374121b | 6654 | valist, f_next_stack, NULL_TREE); |
d1de1e66 | 6655 | |
6656 | /* Call __builtin_saveregs. */ | |
6657 | u = make_tree (ptr_type_node, expand_builtin_saveregs ()); | |
6658 | t = build (MODIFY_EXPR, ptr_type_node, next_fp, u); | |
6659 | TREE_SIDE_EFFECTS (t) = 1; | |
6660 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
6661 | ||
6662 | nfp = current_function_args_info.arg_count[SH_ARG_FLOAT]; | |
6663 | if (nfp < 8) | |
6664 | nfp = 8 - nfp; | |
6665 | else | |
6666 | nfp = 0; | |
6667 | u = fold (build (PLUS_EXPR, ptr_type_node, u, | |
7016c612 | 6668 | build_int_cst (NULL_TREE, UNITS_PER_WORD * nfp))); |
d1de1e66 | 6669 | t = build (MODIFY_EXPR, ptr_type_node, next_fp_limit, u); |
6670 | TREE_SIDE_EFFECTS (t) = 1; | |
6671 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
6672 | ||
6673 | t = build (MODIFY_EXPR, ptr_type_node, next_o, u); | |
6674 | TREE_SIDE_EFFECTS (t) = 1; | |
6675 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
6676 | ||
6677 | nint = current_function_args_info.arg_count[SH_ARG_INT]; | |
6678 | if (nint < 4) | |
6679 | nint = 4 - nint; | |
6680 | else | |
6681 | nint = 0; | |
6682 | u = fold (build (PLUS_EXPR, ptr_type_node, u, | |
7016c612 | 6683 | build_int_cst (NULL_TREE, UNITS_PER_WORD * nint))); |
d1de1e66 | 6684 | t = build (MODIFY_EXPR, ptr_type_node, next_o_limit, u); |
6685 | TREE_SIDE_EFFECTS (t) = 1; | |
6686 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
6687 | ||
6688 | u = make_tree (ptr_type_node, nextarg); | |
d1de1e66 | 6689 | t = build (MODIFY_EXPR, ptr_type_node, next_stack, u); |
6690 | TREE_SIDE_EFFECTS (t) = 1; | |
6691 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
6692 | } | |
6693 | ||
9c1799a1 | 6694 | /* TYPE is a RECORD_TYPE. If there is only a single non-zero-sized |
6695 | member, return it. */ | |
6696 | static tree | |
6697 | find_sole_member (tree type) | |
6698 | { | |
6699 | tree field, member = NULL_TREE; | |
6700 | ||
6701 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
6702 | { | |
6703 | if (TREE_CODE (field) != FIELD_DECL) | |
6704 | continue; | |
6705 | if (!DECL_SIZE (field)) | |
6706 | return NULL_TREE; | |
6707 | if (integer_zerop (DECL_SIZE (field))) | |
6708 | continue; | |
6709 | if (member) | |
6710 | return NULL_TREE; | |
6711 | member = field; | |
6712 | } | |
6713 | return member; | |
6714 | } | |
d1de1e66 | 6715 | /* Implement `va_arg'. */ |
6716 | ||
20cdebf8 | 6717 | static tree |
6718 | sh_gimplify_va_arg_expr (tree valist, tree type, tree *pre_p, | |
6719 | tree *post_p ATTRIBUTE_UNUSED) | |
d1de1e66 | 6720 | { |
6721 | HOST_WIDE_INT size, rsize; | |
b7dbbdb2 | 6722 | tree tmp, pptr_type_node; |
bfa0b1cf | 6723 | tree addr, lab_over = NULL, result = NULL; |
e5a89c0e | 6724 | int pass_by_ref = targetm.calls.must_pass_in_stack (TYPE_MODE (type), type); |
20cdebf8 | 6725 | |
6726 | if (pass_by_ref) | |
6727 | type = build_pointer_type (type); | |
d1de1e66 | 6728 | |
6729 | size = int_size_in_bytes (type); | |
6730 | rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD; | |
6731 | pptr_type_node = build_pointer_type (ptr_type_node); | |
6732 | ||
45550790 | 6733 | if (! TARGET_SH5 && (TARGET_SH2E || TARGET_SH4) |
6734 | && ! (TARGET_HITACHI || sh_cfun_attr_renesas_p ())) | |
d1de1e66 | 6735 | { |
6736 | tree f_next_o, f_next_o_limit, f_next_fp, f_next_fp_limit, f_next_stack; | |
6737 | tree next_o, next_o_limit, next_fp, next_fp_limit, next_stack; | |
6738 | int pass_as_float; | |
20cdebf8 | 6739 | tree lab_false; |
9c1799a1 | 6740 | tree member; |
d1de1e66 | 6741 | |
6742 | f_next_o = TYPE_FIELDS (va_list_type_node); | |
6743 | f_next_o_limit = TREE_CHAIN (f_next_o); | |
6744 | f_next_fp = TREE_CHAIN (f_next_o_limit); | |
6745 | f_next_fp_limit = TREE_CHAIN (f_next_fp); | |
6746 | f_next_stack = TREE_CHAIN (f_next_fp_limit); | |
6747 | ||
6374121b | 6748 | next_o = build (COMPONENT_REF, TREE_TYPE (f_next_o), valist, f_next_o, |
6749 | NULL_TREE); | |
d1de1e66 | 6750 | next_o_limit = build (COMPONENT_REF, TREE_TYPE (f_next_o_limit), |
6374121b | 6751 | valist, f_next_o_limit, NULL_TREE); |
d1de1e66 | 6752 | next_fp = build (COMPONENT_REF, TREE_TYPE (f_next_fp), |
6374121b | 6753 | valist, f_next_fp, NULL_TREE); |
d1de1e66 | 6754 | next_fp_limit = build (COMPONENT_REF, TREE_TYPE (f_next_fp_limit), |
6374121b | 6755 | valist, f_next_fp_limit, NULL_TREE); |
d1de1e66 | 6756 | next_stack = build (COMPONENT_REF, TREE_TYPE (f_next_stack), |
6374121b | 6757 | valist, f_next_stack, NULL_TREE); |
d1de1e66 | 6758 | |
d8e6ba8f | 6759 | /* Structures with a single member with a distinct mode are passed |
6760 | like their member. This is relevant if the latter has a REAL_TYPE | |
6761 | or COMPLEX_TYPE type. */ | |
9ba7abb0 | 6762 | while (TREE_CODE (type) == RECORD_TYPE |
9c1799a1 | 6763 | && (member = find_sole_member (type)) |
6764 | && (TREE_CODE (TREE_TYPE (member)) == REAL_TYPE | |
6765 | || TREE_CODE (TREE_TYPE (member)) == COMPLEX_TYPE | |
6766 | || TREE_CODE (TREE_TYPE (member)) == RECORD_TYPE)) | |
ba59da21 | 6767 | { |
9c1799a1 | 6768 | tree field_type = TREE_TYPE (member); |
ba59da21 | 6769 | |
6770 | if (TYPE_MODE (type) == TYPE_MODE (field_type)) | |
6771 | type = field_type; | |
6772 | else | |
9ba7abb0 | 6773 | { |
9c1799a1 | 6774 | gcc_assert ((TYPE_ALIGN (type) |
6775 | < GET_MODE_ALIGNMENT (TYPE_MODE (field_type))) | |
6776 | || (TYPE_ALIGN (type) | |
6777 | > GET_MODE_BITSIZE (TYPE_MODE (field_type)))); | |
6778 | break; | |
9ba7abb0 | 6779 | } |
ba59da21 | 6780 | } |
20cdebf8 | 6781 | |
d1de1e66 | 6782 | if (TARGET_SH4) |
6783 | { | |
6784 | pass_as_float = ((TREE_CODE (type) == REAL_TYPE && size <= 8) | |
6785 | || (TREE_CODE (type) == COMPLEX_TYPE | |
6786 | && TREE_CODE (TREE_TYPE (type)) == REAL_TYPE | |
6787 | && size <= 16)); | |
6788 | } | |
6789 | else | |
6790 | { | |
6791 | pass_as_float = (TREE_CODE (type) == REAL_TYPE && size == 4); | |
6792 | } | |
6793 | ||
20cdebf8 | 6794 | addr = create_tmp_var (pptr_type_node, NULL); |
6795 | lab_false = create_artificial_label (); | |
6796 | lab_over = create_artificial_label (); | |
d1de1e66 | 6797 | |
20cdebf8 | 6798 | valist = build1 (INDIRECT_REF, ptr_type_node, addr); |
d8e6ba8f | 6799 | |
d1de1e66 | 6800 | if (pass_as_float) |
6801 | { | |
faffb001 | 6802 | tree next_fp_tmp = create_tmp_var (TREE_TYPE (f_next_fp), NULL); |
6803 | tree cmp; | |
6804 | bool is_double = size == 8 && TREE_CODE (type) == REAL_TYPE; | |
65f753eb | 6805 | |
faffb001 | 6806 | tmp = build1 (ADDR_EXPR, pptr_type_node, next_fp); |
6807 | tmp = build2 (MODIFY_EXPR, void_type_node, addr, tmp); | |
6808 | gimplify_and_add (tmp, pre_p); | |
6809 | ||
6810 | tmp = build2 (MODIFY_EXPR, ptr_type_node, next_fp_tmp, valist); | |
6811 | gimplify_and_add (tmp, pre_p); | |
6812 | tmp = next_fp_limit; | |
6813 | if (size > 4 && !is_double) | |
6814 | tmp = build2 (PLUS_EXPR, TREE_TYPE (tmp), tmp, | |
6815 | fold_convert (TREE_TYPE (tmp), size_int (4 - size))); | |
6816 | tmp = build (GE_EXPR, boolean_type_node, next_fp_tmp, tmp); | |
6817 | cmp = build (COND_EXPR, void_type_node, tmp, | |
20cdebf8 | 6818 | build (GOTO_EXPR, void_type_node, lab_false), |
6819 | NULL); | |
faffb001 | 6820 | if (!is_double) |
6821 | gimplify_and_add (cmp, pre_p); | |
d1de1e66 | 6822 | |
faffb001 | 6823 | if (TYPE_ALIGN (type) > BITS_PER_WORD || (is_double || size == 16)) |
d1de1e66 | 6824 | { |
20cdebf8 | 6825 | tmp = fold_convert (ptr_type_node, size_int (UNITS_PER_WORD)); |
faffb001 | 6826 | tmp = build (BIT_AND_EXPR, ptr_type_node, next_fp_tmp, tmp); |
6827 | tmp = build (PLUS_EXPR, ptr_type_node, next_fp_tmp, tmp); | |
6828 | tmp = build (MODIFY_EXPR, ptr_type_node, next_fp_tmp, tmp); | |
20cdebf8 | 6829 | gimplify_and_add (tmp, pre_p); |
d1de1e66 | 6830 | } |
faffb001 | 6831 | if (is_double) |
6832 | gimplify_and_add (cmp, pre_p); | |
d1de1e66 | 6833 | |
d8e6ba8f | 6834 | #ifdef FUNCTION_ARG_SCmode_WART |
6835 | if (TYPE_MODE (type) == SCmode && TARGET_SH4 && TARGET_LITTLE_ENDIAN) | |
6836 | { | |
d8e6ba8f | 6837 | tree subtype = TREE_TYPE (type); |
20cdebf8 | 6838 | tree real, imag; |
d8e6ba8f | 6839 | |
faffb001 | 6840 | imag |
6841 | = std_gimplify_va_arg_expr (next_fp_tmp, subtype, pre_p, NULL); | |
20cdebf8 | 6842 | imag = get_initialized_tmp_var (imag, pre_p, NULL); |
d8e6ba8f | 6843 | |
faffb001 | 6844 | real |
6845 | = std_gimplify_va_arg_expr (next_fp_tmp, subtype, pre_p, NULL); | |
20cdebf8 | 6846 | real = get_initialized_tmp_var (real, pre_p, NULL); |
6847 | ||
6848 | result = build (COMPLEX_EXPR, type, real, imag); | |
6849 | result = get_initialized_tmp_var (result, pre_p, NULL); | |
d8e6ba8f | 6850 | } |
6851 | #endif /* FUNCTION_ARG_SCmode_WART */ | |
6852 | ||
20cdebf8 | 6853 | tmp = build (GOTO_EXPR, void_type_node, lab_over); |
6854 | gimplify_and_add (tmp, pre_p); | |
6855 | ||
6856 | tmp = build (LABEL_EXPR, void_type_node, lab_false); | |
6857 | gimplify_and_add (tmp, pre_p); | |
d1de1e66 | 6858 | |
6859 | tmp = build1 (ADDR_EXPR, pptr_type_node, next_stack); | |
20cdebf8 | 6860 | tmp = build (MODIFY_EXPR, void_type_node, addr, tmp); |
6861 | gimplify_and_add (tmp, pre_p); | |
faffb001 | 6862 | tmp = build2 (MODIFY_EXPR, ptr_type_node, next_fp_tmp, valist); |
6863 | gimplify_and_add (tmp, pre_p); | |
6864 | ||
6865 | tmp = build2 (MODIFY_EXPR, ptr_type_node, valist, next_fp_tmp); | |
6866 | gimplify_and_add (tmp, post_p); | |
6867 | valist = next_fp_tmp; | |
d1de1e66 | 6868 | } |
6869 | else | |
6870 | { | |
20cdebf8 | 6871 | tmp = fold_convert (ptr_type_node, size_int (rsize)); |
6872 | tmp = build (PLUS_EXPR, ptr_type_node, next_o, tmp); | |
6873 | tmp = build (GT_EXPR, boolean_type_node, tmp, next_o_limit); | |
6874 | tmp = build (COND_EXPR, void_type_node, tmp, | |
6875 | build (GOTO_EXPR, void_type_node, lab_false), | |
6876 | NULL); | |
6877 | gimplify_and_add (tmp, pre_p); | |
d1de1e66 | 6878 | |
6879 | tmp = build1 (ADDR_EXPR, pptr_type_node, next_o); | |
20cdebf8 | 6880 | tmp = build (MODIFY_EXPR, void_type_node, addr, tmp); |
6881 | gimplify_and_add (tmp, pre_p); | |
d1de1e66 | 6882 | |
20cdebf8 | 6883 | tmp = build (GOTO_EXPR, void_type_node, lab_over); |
6884 | gimplify_and_add (tmp, pre_p); | |
6885 | ||
6886 | tmp = build (LABEL_EXPR, void_type_node, lab_false); | |
6887 | gimplify_and_add (tmp, pre_p); | |
d1de1e66 | 6888 | |
6889 | if (size > 4 && ! TARGET_SH4) | |
6890 | { | |
6891 | tmp = build (MODIFY_EXPR, ptr_type_node, next_o, next_o_limit); | |
20cdebf8 | 6892 | gimplify_and_add (tmp, pre_p); |
d1de1e66 | 6893 | } |
6894 | ||
6895 | tmp = build1 (ADDR_EXPR, pptr_type_node, next_stack); | |
20cdebf8 | 6896 | tmp = build (MODIFY_EXPR, void_type_node, addr, tmp); |
6897 | gimplify_and_add (tmp, pre_p); | |
d1de1e66 | 6898 | } |
6899 | ||
20cdebf8 | 6900 | if (!result) |
6901 | { | |
6902 | tmp = build (LABEL_EXPR, void_type_node, lab_over); | |
6903 | gimplify_and_add (tmp, pre_p); | |
6904 | } | |
d1de1e66 | 6905 | } |
6906 | ||
6907 | /* ??? In va-sh.h, there had been code to make values larger than | |
6908 | size 8 indirect. This does not match the FUNCTION_ARG macros. */ | |
6909 | ||
20cdebf8 | 6910 | tmp = std_gimplify_va_arg_expr (valist, type, pre_p, NULL); |
d8e6ba8f | 6911 | if (result) |
6912 | { | |
20cdebf8 | 6913 | tmp = build (MODIFY_EXPR, void_type_node, result, tmp); |
6914 | gimplify_and_add (tmp, pre_p); | |
6915 | ||
6916 | tmp = build (LABEL_EXPR, void_type_node, lab_over); | |
6917 | gimplify_and_add (tmp, pre_p); | |
d8e6ba8f | 6918 | } |
6919 | else | |
20cdebf8 | 6920 | result = tmp; |
d8e6ba8f | 6921 | |
a3e17e8c | 6922 | if (pass_by_ref) |
063f5fdd | 6923 | result = build_va_arg_indirect_ref (result); |
20cdebf8 | 6924 | |
a3e17e8c | 6925 | return result; |
d1de1e66 | 6926 | } |
6927 | ||
f5f12868 | 6928 | bool |
04f04b72 | 6929 | sh_promote_prototypes (tree type) |
45550790 | 6930 | { |
6931 | if (TARGET_HITACHI) | |
6932 | return 0; | |
6933 | if (! type) | |
6934 | return 1; | |
6935 | return ! sh_attr_renesas_p (type); | |
6936 | } | |
6937 | ||
b981d932 | 6938 | /* Whether an argument must be passed by reference. On SHcompact, we |
6939 | pretend arguments wider than 32-bits that would have been passed in | |
6940 | registers are passed by reference, so that an SHmedia trampoline | |
6941 | loads them into the full 64-bits registers. */ | |
6942 | ||
6943 | static int | |
6944 | shcompact_byref (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
6945 | tree type, bool named) | |
6946 | { | |
6947 | unsigned HOST_WIDE_INT size; | |
6948 | ||
6949 | if (type) | |
6950 | size = int_size_in_bytes (type); | |
6951 | else | |
6952 | size = GET_MODE_SIZE (mode); | |
6953 | ||
6954 | if (cum->arg_count[SH_ARG_INT] < NPARM_REGS (SImode) | |
6955 | && (!named | |
6956 | || GET_SH_ARG_CLASS (mode) == SH_ARG_INT | |
6957 | || (GET_SH_ARG_CLASS (mode) == SH_ARG_FLOAT | |
6958 | && cum->arg_count[SH_ARG_FLOAT] >= NPARM_REGS (SFmode))) | |
6959 | && size > 4 | |
6960 | && !SHCOMPACT_FORCE_ON_STACK (mode, type) | |
6961 | && !SH5_WOULD_BE_PARTIAL_NREGS (*cum, mode, type, named)) | |
6962 | return size; | |
6963 | else | |
6964 | return 0; | |
6965 | } | |
6966 | ||
6967 | static bool | |
6968 | sh_pass_by_reference (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
6969 | tree type, bool named) | |
6970 | { | |
6971 | if (targetm.calls.must_pass_in_stack (mode, type)) | |
6972 | return true; | |
6973 | ||
e5a89c0e | 6974 | /* ??? std_gimplify_va_arg_expr passes NULL for cum. That function |
6975 | wants to know about pass-by-reference semantics for incoming | |
6976 | arguments. */ | |
6977 | if (! cum) | |
6978 | return false; | |
6979 | ||
b981d932 | 6980 | if (TARGET_SHCOMPACT) |
6981 | { | |
6982 | cum->byref = shcompact_byref (cum, mode, type, named); | |
6983 | return cum->byref != 0; | |
6984 | } | |
6985 | ||
6986 | return false; | |
6987 | } | |
6988 | ||
13f08ee7 | 6989 | static bool |
6990 | sh_callee_copies (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
6991 | tree type, bool named ATTRIBUTE_UNUSED) | |
6992 | { | |
6993 | /* ??? How can it possibly be correct to return true only on the | |
6994 | caller side of the equation? Is there someplace else in the | |
6995 | sh backend that's magically producing the copies? */ | |
6996 | return (cum->outgoing | |
6997 | && ((mode == BLKmode ? TYPE_ALIGN (type) : GET_MODE_ALIGNMENT (mode)) | |
6998 | % SH_MIN_ALIGN_FOR_CALLEE_COPY == 0)); | |
6999 | } | |
7000 | ||
f054eb3c | 7001 | static int |
7002 | sh_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
7003 | tree type, bool named ATTRIBUTE_UNUSED) | |
7004 | { | |
7005 | int words = 0; | |
7006 | ||
7007 | if (!TARGET_SH5 | |
7008 | && PASS_IN_REG_P (*cum, mode, type) | |
7009 | && !(TARGET_SH4 || TARGET_SH2A_DOUBLE) | |
7010 | && (ROUND_REG (*cum, mode) | |
7011 | + (mode != BLKmode | |
7012 | ? ROUND_ADVANCE (GET_MODE_SIZE (mode)) | |
7013 | : ROUND_ADVANCE (int_size_in_bytes (type))) | |
7014 | > NPARM_REGS (mode))) | |
7015 | words = NPARM_REGS (mode) - ROUND_REG (*cum, mode); | |
7016 | ||
7017 | else if (!TARGET_SHCOMPACT | |
7018 | && SH5_WOULD_BE_PARTIAL_NREGS (*cum, mode, type, named)) | |
7019 | words = NPARM_REGS (SImode) - cum->arg_count[SH_ARG_INT]; | |
7020 | ||
7021 | return words * UNITS_PER_WORD; | |
7022 | } | |
7023 | ||
7024 | ||
45550790 | 7025 | /* Define where to put the arguments to a function. |
7026 | Value is zero to push the argument on the stack, | |
7027 | or a hard register in which to store the argument. | |
7028 | ||
7029 | MODE is the argument's machine mode. | |
7030 | TYPE is the data type of the argument (as a tree). | |
7031 | This is null for libcalls where that information may | |
7032 | not be available. | |
7033 | CUM is a variable of type CUMULATIVE_ARGS which gives info about | |
7034 | the preceding args and about the function being called. | |
7035 | NAMED is nonzero if this argument is a named parameter | |
7036 | (otherwise it is an extra parameter matching an ellipsis). | |
7037 | ||
7038 | On SH the first args are normally in registers | |
7039 | and the rest are pushed. Any arg that starts within the first | |
7040 | NPARM_REGS words is at least partially passed in a register unless | |
7041 | its data type forbids. */ | |
7042 | ||
7043 | ||
7044 | rtx | |
04f04b72 | 7045 | sh_function_arg (CUMULATIVE_ARGS *ca, enum machine_mode mode, |
7046 | tree type, int named) | |
45550790 | 7047 | { |
7048 | if (! TARGET_SH5 && mode == VOIDmode) | |
7049 | return GEN_INT (ca->renesas_abi ? 1 : 0); | |
7050 | ||
7051 | if (! TARGET_SH5 | |
7052 | && PASS_IN_REG_P (*ca, mode, type) | |
7053 | && (named || ! (TARGET_HITACHI || ca->renesas_abi))) | |
7054 | { | |
7055 | int regno; | |
7056 | ||
7057 | if (mode == SCmode && TARGET_SH4 && TARGET_LITTLE_ENDIAN | |
7058 | && (! FUNCTION_ARG_SCmode_WART || (ROUND_REG (*ca, mode) & 1))) | |
7059 | { | |
7060 | rtx r1 = gen_rtx_EXPR_LIST (VOIDmode, | |
7061 | gen_rtx_REG (SFmode, | |
7062 | BASE_ARG_REG (mode) | |
7063 | + (ROUND_REG (*ca, mode) ^ 1)), | |
7064 | const0_rtx); | |
d767e27e | 7065 | rtx r2 = gen_rtx_EXPR_LIST (VOIDmode, |
7066 | gen_rtx_REG (SFmode, | |
7067 | BASE_ARG_REG (mode) | |
7068 | + ((ROUND_REG (*ca, mode) + 1) ^ 1)), | |
7069 | GEN_INT (4)); | |
45550790 | 7070 | return gen_rtx_PARALLEL(SCmode, gen_rtvec(2, r1, r2)); |
7071 | } | |
7072 | ||
7073 | /* If the alignment of a DF value causes an SF register to be | |
7074 | skipped, we will use that skipped register for the next SF | |
7075 | value. */ | |
7076 | if ((TARGET_HITACHI || ca->renesas_abi) | |
7077 | && ca->free_single_fp_reg | |
7078 | && mode == SFmode) | |
7079 | return gen_rtx_REG (mode, ca->free_single_fp_reg); | |
7080 | ||
7081 | regno = (BASE_ARG_REG (mode) + ROUND_REG (*ca, mode)) | |
7082 | ^ (mode == SFmode && TARGET_SH4 | |
7083 | && TARGET_LITTLE_ENDIAN != 0 | |
7084 | && ! TARGET_HITACHI && ! ca->renesas_abi); | |
7085 | return gen_rtx_REG (mode, regno); | |
7086 | ||
7087 | } | |
9e7454d0 | 7088 | |
45550790 | 7089 | if (TARGET_SH5) |
7090 | { | |
7091 | if (mode == VOIDmode && TARGET_SHCOMPACT) | |
7092 | return GEN_INT (ca->call_cookie); | |
7093 | ||
7094 | /* The following test assumes unnamed arguments are promoted to | |
7095 | DFmode. */ | |
7096 | if (mode == SFmode && ca->free_single_fp_reg) | |
7097 | return SH5_PROTOTYPED_FLOAT_ARG (*ca, mode, ca->free_single_fp_reg); | |
7098 | ||
7099 | if ((GET_SH_ARG_CLASS (mode) == SH_ARG_FLOAT) | |
7100 | && (named || ! ca->prototype_p) | |
7101 | && ca->arg_count[(int) SH_ARG_FLOAT] < NPARM_REGS (SFmode)) | |
7102 | { | |
7103 | if (! ca->prototype_p && TARGET_SHMEDIA) | |
7104 | return SH5_PROTOTYPELESS_FLOAT_ARG (*ca, mode); | |
7105 | ||
7106 | return SH5_PROTOTYPED_FLOAT_ARG (*ca, mode, | |
7107 | FIRST_FP_PARM_REG | |
7108 | + ca->arg_count[(int) SH_ARG_FLOAT]); | |
7109 | } | |
7110 | ||
7111 | if (ca->arg_count[(int) SH_ARG_INT] < NPARM_REGS (SImode) | |
7112 | && (! TARGET_SHCOMPACT | |
7113 | || (! SHCOMPACT_FORCE_ON_STACK (mode, type) | |
7114 | && ! SH5_WOULD_BE_PARTIAL_NREGS (*ca, mode, | |
7115 | type, named)))) | |
7116 | { | |
7117 | return gen_rtx_REG (mode, (FIRST_PARM_REG | |
7118 | + ca->arg_count[(int) SH_ARG_INT])); | |
7119 | } | |
7120 | ||
7121 | return 0; | |
7122 | } | |
7123 | ||
7124 | return 0; | |
7125 | } | |
9e7454d0 | 7126 | |
45550790 | 7127 | /* Update the data in CUM to advance over an argument |
7128 | of mode MODE and data type TYPE. | |
7129 | (TYPE is null for libcalls where that information may not be | |
7130 | available.) */ | |
7131 | ||
7132 | void | |
04f04b72 | 7133 | sh_function_arg_advance (CUMULATIVE_ARGS *ca, enum machine_mode mode, |
7134 | tree type, int named) | |
45550790 | 7135 | { |
d767e27e | 7136 | if (ca->force_mem) |
7137 | ca->force_mem = 0; | |
7138 | else if (TARGET_SH5) | |
7139 | { | |
7140 | tree type2 = (ca->byref && type | |
7141 | ? TREE_TYPE (type) | |
7142 | : type); | |
7143 | enum machine_mode mode2 = (ca->byref && type | |
7144 | ? TYPE_MODE (type2) | |
7145 | : mode); | |
7146 | int dwords = ((ca->byref | |
7147 | ? ca->byref | |
7148 | : mode2 == BLKmode | |
7149 | ? int_size_in_bytes (type2) | |
7150 | : GET_MODE_SIZE (mode2)) + 7) / 8; | |
7151 | int numregs = MIN (dwords, NPARM_REGS (SImode) | |
7152 | - ca->arg_count[(int) SH_ARG_INT]); | |
7153 | ||
7154 | if (numregs) | |
7155 | { | |
7156 | ca->arg_count[(int) SH_ARG_INT] += numregs; | |
7157 | if (TARGET_SHCOMPACT | |
7158 | && SHCOMPACT_FORCE_ON_STACK (mode2, type2)) | |
7159 | { | |
7160 | ca->call_cookie | |
7161 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
7162 | - numregs, 1); | |
7163 | /* N.B. We want this also for outgoing. */ | |
7164 | ca->stack_regs += numregs; | |
7165 | } | |
7166 | else if (ca->byref) | |
7167 | { | |
7168 | if (! ca->outgoing) | |
7169 | ca->stack_regs += numregs; | |
7170 | ca->byref_regs += numregs; | |
7171 | ca->byref = 0; | |
7172 | do | |
7173 | ca->call_cookie | |
7174 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
7175 | - numregs, 2); | |
7176 | while (--numregs); | |
7177 | ca->call_cookie | |
7178 | |= CALL_COOKIE_INT_REG (ca->arg_count[(int) SH_ARG_INT] | |
7179 | - 1, 1); | |
7180 | } | |
7181 | else if (dwords > numregs) | |
7182 | { | |
7183 | int pushregs = numregs; | |
7184 | ||
7185 | if (TARGET_SHCOMPACT) | |
7186 | ca->stack_regs += numregs; | |
7187 | while (pushregs < NPARM_REGS (SImode) - 1 | |
7188 | && (CALL_COOKIE_INT_REG_GET | |
7189 | (ca->call_cookie, | |
7190 | NPARM_REGS (SImode) - pushregs) | |
7191 | == 1)) | |
7192 | { | |
7193 | ca->call_cookie | |
7194 | &= ~ CALL_COOKIE_INT_REG (NPARM_REGS (SImode) | |
7195 | - pushregs, 1); | |
7196 | pushregs++; | |
7197 | } | |
7198 | if (numregs == NPARM_REGS (SImode)) | |
7199 | ca->call_cookie | |
7200 | |= CALL_COOKIE_INT_REG (0, 1) | |
7201 | | CALL_COOKIE_STACKSEQ (numregs - 1); | |
7202 | else | |
7203 | ca->call_cookie | |
7204 | |= CALL_COOKIE_STACKSEQ (numregs); | |
7205 | } | |
7206 | } | |
7207 | if (GET_SH_ARG_CLASS (mode2) == SH_ARG_FLOAT | |
7208 | && (named || ! ca->prototype_p)) | |
7209 | { | |
7210 | if (mode2 == SFmode && ca->free_single_fp_reg) | |
7211 | ca->free_single_fp_reg = 0; | |
7212 | else if (ca->arg_count[(int) SH_ARG_FLOAT] | |
7213 | < NPARM_REGS (SFmode)) | |
7214 | { | |
7215 | int numfpregs | |
7216 | = MIN ((GET_MODE_SIZE (mode2) + 7) / 8 * 2, | |
7217 | NPARM_REGS (SFmode) | |
7218 | - ca->arg_count[(int) SH_ARG_FLOAT]); | |
7219 | ||
7220 | ca->arg_count[(int) SH_ARG_FLOAT] += numfpregs; | |
7221 | ||
7222 | if (TARGET_SHCOMPACT && ! ca->prototype_p) | |
7223 | { | |
7224 | if (ca->outgoing && numregs > 0) | |
7225 | do | |
7226 | { | |
7227 | ca->call_cookie | |
7228 | |= (CALL_COOKIE_INT_REG | |
7229 | (ca->arg_count[(int) SH_ARG_INT] | |
7230 | - numregs + ((numfpregs - 2) / 2), | |
7231 | 4 + (ca->arg_count[(int) SH_ARG_FLOAT] | |
7232 | - numfpregs) / 2)); | |
7233 | } | |
7234 | while (numfpregs -= 2); | |
7235 | } | |
7236 | else if (mode2 == SFmode && (named) | |
7237 | && (ca->arg_count[(int) SH_ARG_FLOAT] | |
7238 | < NPARM_REGS (SFmode))) | |
7239 | ca->free_single_fp_reg | |
7240 | = FIRST_FP_PARM_REG - numfpregs | |
7241 | + ca->arg_count[(int) SH_ARG_FLOAT] + 1; | |
7242 | } | |
7243 | } | |
7244 | return; | |
7245 | } | |
7246 | ||
7247 | if ((TARGET_HITACHI || ca->renesas_abi) && TARGET_FPU_DOUBLE) | |
7248 | { | |
7249 | /* Note that we've used the skipped register. */ | |
7250 | if (mode == SFmode && ca->free_single_fp_reg) | |
7251 | { | |
7252 | ca->free_single_fp_reg = 0; | |
7253 | return; | |
7254 | } | |
7255 | /* When we have a DF after an SF, there's an SF register that get | |
7256 | skipped in order to align the DF value. We note this skipped | |
7257 | register, because the next SF value will use it, and not the | |
7258 | SF that follows the DF. */ | |
7259 | if (mode == DFmode | |
7260 | && ROUND_REG (*ca, DFmode) != ROUND_REG (*ca, SFmode)) | |
7261 | { | |
7262 | ca->free_single_fp_reg = (ROUND_REG (*ca, SFmode) | |
7263 | + BASE_ARG_REG (mode)); | |
7264 | } | |
7265 | } | |
7266 | ||
7105fb72 | 7267 | if (! ((TARGET_SH4 || TARGET_SH2A) || ca->renesas_abi) |
d767e27e | 7268 | || PASS_IN_REG_P (*ca, mode, type)) |
7269 | (ca->arg_count[(int) GET_SH_ARG_CLASS (mode)] | |
7270 | = (ROUND_REG (*ca, mode) | |
7271 | + (mode == BLKmode | |
7272 | ? ROUND_ADVANCE (int_size_in_bytes (type)) | |
7273 | : ROUND_ADVANCE (GET_MODE_SIZE (mode))))); | |
45550790 | 7274 | } |
7275 | ||
45550790 | 7276 | /* The Renesas calling convention doesn't quite fit into this scheme since |
7277 | the address is passed like an invisible argument, but one that is always | |
7278 | passed in memory. */ | |
7279 | static rtx | |
04f04b72 | 7280 | sh_struct_value_rtx (tree fndecl, int incoming ATTRIBUTE_UNUSED) |
45550790 | 7281 | { |
7282 | if (TARGET_HITACHI || sh_attr_renesas_p (fndecl)) | |
7283 | return 0; | |
7284 | return gen_rtx_REG (Pmode, 2); | |
7285 | } | |
7286 | ||
6644435d | 7287 | /* Worker function for TARGET_RETURN_IN_MEMORY. */ |
7288 | ||
45550790 | 7289 | static bool |
04f04b72 | 7290 | sh_return_in_memory (tree type, tree fndecl) |
45550790 | 7291 | { |
7292 | if (TARGET_SH5) | |
7293 | { | |
7294 | if (TYPE_MODE (type) == BLKmode) | |
7295 | return ((unsigned HOST_WIDE_INT) int_size_in_bytes (type)) > 8; | |
7296 | else | |
7297 | return GET_MODE_SIZE (TYPE_MODE (type)) > 8; | |
7298 | } | |
7299 | else | |
7300 | { | |
7301 | return (TYPE_MODE (type) == BLKmode | |
7302 | || ((TARGET_HITACHI || sh_attr_renesas_p (fndecl)) | |
7303 | && TREE_CODE (type) == RECORD_TYPE)); | |
7304 | } | |
7305 | } | |
7306 | ||
7307 | /* We actually emit the code in sh_expand_prologue. We used to use | |
7308 | a static variable to flag that we need to emit this code, but that | |
7309 | doesn't when inlining, when functions are deferred and then emitted | |
7310 | later. Fortunately, we already have two flags that are part of struct | |
7311 | function that tell if a function uses varargs or stdarg. */ | |
7312 | static void | |
09a512b9 | 7313 | sh_setup_incoming_varargs (CUMULATIVE_ARGS *ca, |
7314 | enum machine_mode mode, | |
7315 | tree type, | |
7316 | int *pretend_arg_size, | |
04f04b72 | 7317 | int second_time ATTRIBUTE_UNUSED) |
45550790 | 7318 | { |
1a6a7a27 | 7319 | gcc_assert (current_function_stdarg); |
09a512b9 | 7320 | if (TARGET_VARARGS_PRETEND_ARGS (current_function_decl)) |
7321 | { | |
7322 | int named_parm_regs, anon_parm_regs; | |
7323 | ||
7324 | named_parm_regs = (ROUND_REG (*ca, mode) | |
7325 | + (mode == BLKmode | |
7326 | ? ROUND_ADVANCE (int_size_in_bytes (type)) | |
7327 | : ROUND_ADVANCE (GET_MODE_SIZE (mode)))); | |
7328 | anon_parm_regs = NPARM_REGS (SImode) - named_parm_regs; | |
7329 | if (anon_parm_regs > 0) | |
7330 | *pretend_arg_size = anon_parm_regs * 4; | |
7331 | } | |
45550790 | 7332 | } |
7333 | ||
7334 | static bool | |
04f04b72 | 7335 | sh_strict_argument_naming (CUMULATIVE_ARGS *ca ATTRIBUTE_UNUSED) |
45550790 | 7336 | { |
7337 | return TARGET_SH5; | |
7338 | } | |
7339 | ||
7340 | static bool | |
04f04b72 | 7341 | sh_pretend_outgoing_varargs_named (CUMULATIVE_ARGS *ca) |
45550790 | 7342 | { |
7343 | return ! (TARGET_HITACHI || ca->renesas_abi) && ! TARGET_SH5; | |
7344 | } | |
7345 | ||
7346 | ||
97595bfd | 7347 | /* Define the offset between two registers, one to be eliminated, and |
7348 | the other its replacement, at the start of a routine. */ | |
7349 | ||
7350 | int | |
04f04b72 | 7351 | initial_elimination_offset (int from, int to) |
97595bfd | 7352 | { |
7353 | int regs_saved; | |
87e19636 | 7354 | int regs_saved_rounding = 0; |
97595bfd | 7355 | int total_saved_regs_space; |
9200a51c | 7356 | int total_auto_space; |
8ded0752 | 7357 | int save_flags = target_flags; |
87e19636 | 7358 | int copy_flags; |
ed8483c0 | 7359 | HARD_REG_SET live_regs_mask; |
8af3db02 | 7360 | |
7361 | shmedia_space_reserved_for_target_registers = false; | |
ed8483c0 | 7362 | regs_saved = calc_live_regs (&live_regs_mask); |
87e19636 | 7363 | regs_saved += SHMEDIA_REGS_STACK_ADJUST (); |
8af3db02 | 7364 | |
7365 | if (shmedia_reserve_space_for_target_registers_p (regs_saved, &live_regs_mask)) | |
7366 | { | |
7367 | shmedia_space_reserved_for_target_registers = true; | |
7368 | regs_saved += shmedia_target_regs_stack_adjust (&live_regs_mask); | |
7369 | } | |
7370 | ||
87e19636 | 7371 | if (TARGET_SH5 && regs_saved % (STACK_BOUNDARY / BITS_PER_UNIT)) |
7372 | regs_saved_rounding = ((STACK_BOUNDARY / BITS_PER_UNIT) | |
7373 | - regs_saved % (STACK_BOUNDARY / BITS_PER_UNIT)); | |
2e3d4844 | 7374 | |
87e19636 | 7375 | total_auto_space = rounded_frame_size (regs_saved) - regs_saved_rounding; |
7376 | copy_flags = target_flags; | |
8ded0752 | 7377 | target_flags = save_flags; |
b00564da | 7378 | |
87e19636 | 7379 | total_saved_regs_space = regs_saved + regs_saved_rounding; |
47c009e5 | 7380 | |
b9a602c6 | 7381 | if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) |
87e19636 | 7382 | return total_saved_regs_space + total_auto_space |
7383 | + current_function_args_info.byref_regs * 8; | |
73401833 | 7384 | |
97595bfd | 7385 | if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
87e19636 | 7386 | return total_saved_regs_space + total_auto_space |
7387 | + current_function_args_info.byref_regs * 8; | |
73401833 | 7388 | |
7389 | /* Initial gap between fp and sp is 0. */ | |
b9a602c6 | 7390 | if (from == HARD_FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
73401833 | 7391 | return 0; |
7392 | ||
b9a602c6 | 7393 | if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
7394 | return rounded_frame_size (0); | |
7395 | ||
7396 | if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
7397 | return rounded_frame_size (0); | |
7398 | ||
1a6a7a27 | 7399 | gcc_assert (from == RETURN_ADDRESS_POINTER_REGNUM |
b9a602c6 | 7400 | && (to == HARD_FRAME_POINTER_REGNUM |
7401 | || to == STACK_POINTER_REGNUM)); | |
1a6a7a27 | 7402 | if (TARGET_SH5) |
acda5809 | 7403 | { |
1a6a7a27 | 7404 | int n = total_saved_regs_space; |
7405 | int pr_reg = TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG; | |
7406 | save_schedule schedule; | |
7407 | save_entry *entry; | |
7408 | ||
7409 | n += total_auto_space; | |
7410 | ||
7411 | /* If it wasn't saved, there's not much we can do. */ | |
7412 | if (! TEST_HARD_REG_BIT (live_regs_mask, pr_reg)) | |
7413 | return n; | |
7414 | ||
7415 | target_flags = copy_flags; | |
7416 | ||
7417 | sh5_schedule_saves (&live_regs_mask, &schedule, n); | |
7418 | for (entry = &schedule.entries[1]; entry->mode != VOIDmode; entry++) | |
7419 | if (entry->reg == pr_reg) | |
7420 | { | |
7421 | target_flags = save_flags; | |
7422 | return entry->offset; | |
7423 | } | |
7424 | gcc_unreachable (); | |
acda5809 | 7425 | } |
1a6a7a27 | 7426 | else |
7427 | return total_auto_space; | |
97595bfd | 7428 | } |
73401833 | 7429 | \f |
7b23765d | 7430 | /* Handle machine specific pragmas to be semi-compatible with Renesas |
1504eb7e | 7431 | compiler. */ |
47c009e5 | 7432 | |
1fcd08b1 | 7433 | void |
04f04b72 | 7434 | sh_pr_interrupt (struct cpp_reader *pfile ATTRIBUTE_UNUSED) |
47c009e5 | 7435 | { |
1fcd08b1 | 7436 | pragma_interrupt = 1; |
7437 | } | |
b97b38c0 | 7438 | |
1fcd08b1 | 7439 | void |
04f04b72 | 7440 | sh_pr_trapa (struct cpp_reader *pfile ATTRIBUTE_UNUSED) |
1fcd08b1 | 7441 | { |
7442 | pragma_interrupt = pragma_trapa = 1; | |
7443 | } | |
b97b38c0 | 7444 | |
1fcd08b1 | 7445 | void |
04f04b72 | 7446 | sh_pr_nosave_low_regs (struct cpp_reader *pfile ATTRIBUTE_UNUSED) |
1fcd08b1 | 7447 | { |
7448 | pragma_nosave_low_regs = 1; | |
97595bfd | 7449 | } |
49f630e1 | 7450 | |
7451 | /* Generate 'handle_interrupt' attribute for decls */ | |
7452 | ||
8ee295a7 | 7453 | static void |
04f04b72 | 7454 | sh_insert_attributes (tree node, tree *attributes) |
49f630e1 | 7455 | { |
49f630e1 | 7456 | if (! pragma_interrupt |
7457 | || TREE_CODE (node) != FUNCTION_DECL) | |
7458 | return; | |
7459 | ||
7460 | /* We are only interested in fields. */ | |
ce45a448 | 7461 | if (!DECL_P (node)) |
49f630e1 | 7462 | return; |
7463 | ||
7464 | /* Add a 'handle_interrupt' attribute. */ | |
7465 | * attributes = tree_cons (get_identifier ("interrupt_handler"), NULL, * attributes); | |
7466 | ||
7467 | return; | |
7468 | } | |
7469 | ||
e3c541f0 | 7470 | /* Supported attributes: |
15cb1799 | 7471 | |
7472 | interrupt_handler -- specifies this function is an interrupt handler. | |
7473 | ||
7474 | sp_switch -- specifies an alternate stack for an interrupt handler | |
7475 | to run on. | |
7476 | ||
ad87de1e | 7477 | trap_exit -- use a trapa to exit an interrupt function instead of |
45550790 | 7478 | an rte instruction. |
7479 | ||
7480 | renesas -- use Renesas calling/layout conventions (functions and | |
7481 | structures). | |
7482 | ||
7483 | */ | |
15cb1799 | 7484 | |
e3c541f0 | 7485 | const struct attribute_spec sh_attribute_table[] = |
15cb1799 | 7486 | { |
e3c541f0 | 7487 | /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */ |
7488 | { "interrupt_handler", 0, 0, true, false, false, sh_handle_interrupt_handler_attribute }, | |
7489 | { "sp_switch", 1, 1, true, false, false, sh_handle_sp_switch_attribute }, | |
7490 | { "trap_exit", 1, 1, true, false, false, sh_handle_trap_exit_attribute }, | |
45550790 | 7491 | { "renesas", 0, 0, false, true, false, sh_handle_renesas_attribute }, |
db281f5a | 7492 | #ifdef SYMBIAN |
7493 | /* Symbian support adds three new attributes: | |
7494 | dllexport - for exporting a function/variable that will live in a dll | |
7495 | dllimport - for importing a function/variable from a dll | |
9e7454d0 | 7496 | |
db281f5a | 7497 | Microsoft allows multiple declspecs in one __declspec, separating |
7498 | them with spaces. We do NOT support this. Instead, use __declspec | |
7499 | multiple times. */ | |
7500 | { "dllimport", 0, 0, true, false, false, sh_symbian_handle_dll_attribute }, | |
7501 | { "dllexport", 0, 0, true, false, false, sh_symbian_handle_dll_attribute }, | |
7502 | #endif | |
e3c541f0 | 7503 | { NULL, 0, 0, false, false, false, NULL } |
7504 | }; | |
15cb1799 | 7505 | |
e3c541f0 | 7506 | /* Handle an "interrupt_handler" attribute; arguments as in |
7507 | struct attribute_spec.handler. */ | |
7508 | static tree | |
04f04b72 | 7509 | sh_handle_interrupt_handler_attribute (tree *node, tree name, |
7510 | tree args ATTRIBUTE_UNUSED, | |
7511 | int flags ATTRIBUTE_UNUSED, | |
7512 | bool *no_add_attrs) | |
e3c541f0 | 7513 | { |
7514 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
15cb1799 | 7515 | { |
9b2d6d13 | 7516 | warning (OPT_Wattributes, "%qs attribute only applies to functions", |
e3c541f0 | 7517 | IDENTIFIER_POINTER (name)); |
7518 | *no_add_attrs = true; | |
15cb1799 | 7519 | } |
154b29aa | 7520 | else if (TARGET_SHCOMPACT) |
7521 | { | |
7522 | error ("attribute interrupt_handler is not compatible with -m5-compact"); | |
7523 | *no_add_attrs = true; | |
7524 | } | |
15cb1799 | 7525 | |
e3c541f0 | 7526 | return NULL_TREE; |
7527 | } | |
7528 | ||
7529 | /* Handle an "sp_switch" attribute; arguments as in | |
7530 | struct attribute_spec.handler. */ | |
7531 | static tree | |
04f04b72 | 7532 | sh_handle_sp_switch_attribute (tree *node, tree name, tree args, |
7533 | int flags ATTRIBUTE_UNUSED, bool *no_add_attrs) | |
e3c541f0 | 7534 | { |
7535 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
7536 | { | |
9b2d6d13 | 7537 | warning (OPT_Wattributes, "%qs attribute only applies to functions", |
e3c541f0 | 7538 | IDENTIFIER_POINTER (name)); |
7539 | *no_add_attrs = true; | |
7540 | } | |
7541 | else if (!pragma_interrupt) | |
15cb1799 | 7542 | { |
7543 | /* The sp_switch attribute only has meaning for interrupt functions. */ | |
9b2d6d13 | 7544 | warning (OPT_Wattributes, "%qs attribute only applies to " |
7545 | "interrupt functions", IDENTIFIER_POINTER (name)); | |
e3c541f0 | 7546 | *no_add_attrs = true; |
7547 | } | |
7548 | else if (TREE_CODE (TREE_VALUE (args)) != STRING_CST) | |
7549 | { | |
15cb1799 | 7550 | /* The argument must be a constant string. */ |
9b2d6d13 | 7551 | warning (OPT_Wattributes, "%qs attribute argument not a string constant", |
e3c541f0 | 7552 | IDENTIFIER_POINTER (name)); |
7553 | *no_add_attrs = true; | |
7554 | } | |
7555 | else | |
7556 | { | |
bfa0b1cf | 7557 | const char *s = ggc_strdup (TREE_STRING_POINTER (TREE_VALUE (args))); |
d5406300 | 7558 | sp_switch = gen_rtx_SYMBOL_REF (VOIDmode, s); |
15cb1799 | 7559 | } |
7560 | ||
e3c541f0 | 7561 | return NULL_TREE; |
7562 | } | |
7563 | ||
7564 | /* Handle an "trap_exit" attribute; arguments as in | |
7565 | struct attribute_spec.handler. */ | |
7566 | static tree | |
04f04b72 | 7567 | sh_handle_trap_exit_attribute (tree *node, tree name, tree args, |
7568 | int flags ATTRIBUTE_UNUSED, bool *no_add_attrs) | |
e3c541f0 | 7569 | { |
7570 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
7571 | { | |
9b2d6d13 | 7572 | warning (OPT_Wattributes, "%qs attribute only applies to functions", |
e3c541f0 | 7573 | IDENTIFIER_POINTER (name)); |
7574 | *no_add_attrs = true; | |
7575 | } | |
7576 | else if (!pragma_interrupt) | |
15cb1799 | 7577 | { |
7578 | /* The trap_exit attribute only has meaning for interrupt functions. */ | |
9b2d6d13 | 7579 | warning (OPT_Wattributes, "%qs attribute only applies to " |
7580 | "interrupt functions", IDENTIFIER_POINTER (name)); | |
e3c541f0 | 7581 | *no_add_attrs = true; |
7582 | } | |
7583 | else if (TREE_CODE (TREE_VALUE (args)) != INTEGER_CST) | |
7584 | { | |
15cb1799 | 7585 | /* The argument must be a constant integer. */ |
9b2d6d13 | 7586 | warning (OPT_Wattributes, "%qs attribute argument not an " |
7587 | "integer constant", IDENTIFIER_POINTER (name)); | |
e3c541f0 | 7588 | *no_add_attrs = true; |
7589 | } | |
7590 | else | |
7591 | { | |
15cb1799 | 7592 | trap_exit = TREE_INT_CST_LOW (TREE_VALUE (args)); |
15cb1799 | 7593 | } |
b7dbbdb2 | 7594 | |
e3c541f0 | 7595 | return NULL_TREE; |
15cb1799 | 7596 | } |
7597 | ||
45550790 | 7598 | static tree |
04f04b72 | 7599 | sh_handle_renesas_attribute (tree *node ATTRIBUTE_UNUSED, |
7600 | tree name ATTRIBUTE_UNUSED, | |
7601 | tree args ATTRIBUTE_UNUSED, | |
7602 | int flags ATTRIBUTE_UNUSED, | |
7603 | bool *no_add_attrs ATTRIBUTE_UNUSED) | |
45550790 | 7604 | { |
7605 | return NULL_TREE; | |
7606 | } | |
7607 | ||
7608 | /* True if __attribute__((renesas)) or -mrenesas. */ | |
7609 | int | |
04f04b72 | 7610 | sh_attr_renesas_p (tree td) |
45550790 | 7611 | { |
7612 | if (TARGET_HITACHI) | |
7613 | return 1; | |
7614 | if (td == 0) | |
7615 | return 0; | |
7616 | if (DECL_P (td)) | |
7617 | td = TREE_TYPE (td); | |
faffb001 | 7618 | if (td == error_mark_node) |
7619 | return 0; | |
45550790 | 7620 | return (lookup_attribute ("renesas", TYPE_ATTRIBUTES (td)) |
7621 | != NULL_TREE); | |
7622 | } | |
7623 | ||
7624 | /* True if __attribute__((renesas)) or -mrenesas, for the current | |
7625 | function. */ | |
7626 | int | |
04f04b72 | 7627 | sh_cfun_attr_renesas_p (void) |
45550790 | 7628 | { |
7629 | return sh_attr_renesas_p (current_function_decl); | |
7630 | } | |
7631 | ||
5df3d11f | 7632 | int |
04f04b72 | 7633 | sh_cfun_interrupt_handler_p (void) |
5df3d11f | 7634 | { |
7635 | return (lookup_attribute ("interrupt_handler", | |
7636 | DECL_ATTRIBUTES (current_function_decl)) | |
7637 | != NULL_TREE); | |
7638 | } | |
133479bd | 7639 | |
761d9126 | 7640 | /* Implement TARGET_CHECK_PCH_TARGET_FLAGS. */ |
59312820 | 7641 | |
761d9126 | 7642 | static const char * |
7643 | sh_check_pch_target_flags (int old_flags) | |
7644 | { | |
156fac8b | 7645 | if ((old_flags ^ target_flags) & (MASK_SH1 | MASK_SH2 | MASK_SH3 |
7646 | | MASK_SH_E | MASK_HARD_SH4 | |
7647 | | MASK_FPU_SINGLE | MASK_SH4)) | |
7648 | return _("created and used with different architectures / ABIs"); | |
7649 | if ((old_flags ^ target_flags) & MASK_HITACHI) | |
7650 | return _("created and used with different ABIs"); | |
7651 | if ((old_flags ^ target_flags) & MASK_LITTLE_ENDIAN) | |
7652 | return _("created and used with different endianness"); | |
133479bd | 7653 | return NULL; |
133479bd | 7654 | } |
97595bfd | 7655 | \f |
73401833 | 7656 | /* Predicates used by the templates. */ |
97595bfd | 7657 | |
73401833 | 7658 | /* Returns 1 if OP is MACL, MACH or PR. The input must be a REG rtx. |
7659 | Used only in general_movsrc_operand. */ | |
97595bfd | 7660 | |
73401833 | 7661 | int |
04f04b72 | 7662 | system_reg_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
97595bfd | 7663 | { |
73401833 | 7664 | switch (REGNO (op)) |
97595bfd | 7665 | { |
73401833 | 7666 | case PR_REG: |
7667 | case MACL_REG: | |
7668 | case MACH_REG: | |
7669 | return 1; | |
97595bfd | 7670 | } |
73401833 | 7671 | return 0; |
97595bfd | 7672 | } |
97595bfd | 7673 | |
b00564da | 7674 | /* Nonzero if OP is a floating point value with value 0.0. */ |
7675 | ||
7676 | int | |
04f04b72 | 7677 | fp_zero_operand (rtx op) |
f3d93547 | 7678 | { |
b00564da | 7679 | REAL_VALUE_TYPE r; |
f3d93547 | 7680 | |
b00564da | 7681 | if (GET_MODE (op) != SFmode) |
7682 | return 0; | |
7683 | ||
7684 | REAL_VALUE_FROM_CONST_DOUBLE (r, op); | |
e6c94d0d | 7685 | return REAL_VALUES_EQUAL (r, dconst0) && ! REAL_VALUE_MINUS_ZERO (r); |
f3d93547 | 7686 | } |
7687 | ||
b00564da | 7688 | /* Nonzero if OP is a floating point value with value 1.0. */ |
f3d93547 | 7689 | |
7690 | int | |
04f04b72 | 7691 | fp_one_operand (rtx op) |
f3d93547 | 7692 | { |
b00564da | 7693 | REAL_VALUE_TYPE r; |
7694 | ||
7695 | if (GET_MODE (op) != SFmode) | |
7696 | return 0; | |
7697 | ||
7698 | REAL_VALUE_FROM_CONST_DOUBLE (r, op); | |
7699 | return REAL_VALUES_EQUAL (r, dconst1); | |
f3d93547 | 7700 | } |
8ded0752 | 7701 | |
18862b5a | 7702 | /* For -m4 and -m4-single-only, mode switching is used. If we are |
7703 | compiling without -mfmovd, movsf_ie isn't taken into account for | |
7704 | mode switching. We could check in machine_dependent_reorg for | |
7705 | cases where we know we are in single precision mode, but there is | |
7706 | interface to find that out during reload, so we must avoid | |
7707 | choosing an fldi alternative during reload and thus failing to | |
7708 | allocate a scratch register for the constant loading. */ | |
7709 | int | |
04f04b72 | 7710 | fldi_ok (void) |
18862b5a | 7711 | { |
7712 | return ! TARGET_SH4 || TARGET_FMOVD || reload_completed; | |
7713 | } | |
7714 | ||
1b61190c | 7715 | int |
04f04b72 | 7716 | tertiary_reload_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
1b61190c | 7717 | { |
7718 | enum rtx_code code = GET_CODE (op); | |
7719 | return code == MEM || (TARGET_SH4 && code == CONST_DOUBLE); | |
7720 | } | |
7721 | ||
cec78553 | 7722 | /* Return the TLS type for TLS symbols, 0 for otherwise. */ |
7723 | int | |
04f04b72 | 7724 | tls_symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED) |
cec78553 | 7725 | { |
cec78553 | 7726 | if (GET_CODE (op) != SYMBOL_REF) |
7727 | return 0; | |
2fb8cf90 | 7728 | return SYMBOL_REF_TLS_MODEL (op); |
cec78553 | 7729 | } |
8ded0752 | 7730 | \f |
9eaab178 | 7731 | /* Return the destination address of a branch. */ |
9e7454d0 | 7732 | |
b7dbbdb2 | 7733 | static int |
04f04b72 | 7734 | branch_dest (rtx branch) |
8ded0752 | 7735 | { |
9eaab178 | 7736 | rtx dest = SET_SRC (PATTERN (branch)); |
7737 | int dest_uid; | |
8ded0752 | 7738 | |
7739 | if (GET_CODE (dest) == IF_THEN_ELSE) | |
7740 | dest = XEXP (dest, 1); | |
7741 | dest = XEXP (dest, 0); | |
7742 | dest_uid = INSN_UID (dest); | |
47fc0706 | 7743 | return INSN_ADDRESSES (dest_uid); |
8ded0752 | 7744 | } |
72bccaaa | 7745 | \f |
808a491c | 7746 | /* Return nonzero if REG is not used after INSN. |
72bccaaa | 7747 | We assume REG is a reload reg, and therefore does |
e020b473 | 7748 | not live past labels. It may live past calls or jumps though. */ |
72bccaaa | 7749 | int |
04f04b72 | 7750 | reg_unused_after (rtx reg, rtx insn) |
72bccaaa | 7751 | { |
0f7bb88e | 7752 | enum rtx_code code; |
72bccaaa | 7753 | rtx set; |
7754 | ||
7755 | /* If the reg is set by this instruction, then it is safe for our | |
7756 | case. Disregard the case where this is a store to memory, since | |
7757 | we are checking a register used in the store address. */ | |
7758 | set = single_set (insn); | |
7759 | if (set && GET_CODE (SET_DEST (set)) != MEM | |
7760 | && reg_overlap_mentioned_p (reg, SET_DEST (set))) | |
7761 | return 1; | |
7762 | ||
b7dbbdb2 | 7763 | while ((insn = NEXT_INSN (insn))) |
72bccaaa | 7764 | { |
6720e96c | 7765 | rtx set; |
7766 | if (!INSN_P (insn)) | |
7767 | continue; | |
7768 | ||
72bccaaa | 7769 | code = GET_CODE (insn); |
0f7bb88e | 7770 | |
87faf2d3 | 7771 | #if 0 |
7772 | /* If this is a label that existed before reload, then the register | |
7773 | if dead here. However, if this is a label added by reorg, then | |
7774 | the register may still be live here. We can't tell the difference, | |
7775 | so we just ignore labels completely. */ | |
0f7bb88e | 7776 | if (code == CODE_LABEL) |
72bccaaa | 7777 | return 1; |
87faf2d3 | 7778 | /* else */ |
7779 | #endif | |
72bccaaa | 7780 | |
e020b473 | 7781 | if (code == JUMP_INSN) |
7782 | return 0; | |
7783 | ||
0f7bb88e | 7784 | /* If this is a sequence, we must handle them all at once. |
7785 | We could have for instance a call that sets the target register, | |
ebb11c7b | 7786 | and an insn in a delay slot that uses the register. In this case, |
0f7bb88e | 7787 | we must return 0. */ |
e020b473 | 7788 | else if (code == INSN && GET_CODE (PATTERN (insn)) == SEQUENCE) |
72bccaaa | 7789 | { |
0f7bb88e | 7790 | int i; |
7791 | int retval = 0; | |
7792 | ||
7793 | for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++) | |
7794 | { | |
7795 | rtx this_insn = XVECEXP (PATTERN (insn), 0, i); | |
7796 | rtx set = single_set (this_insn); | |
7797 | ||
7798 | if (GET_CODE (this_insn) == CALL_INSN) | |
7799 | code = CALL_INSN; | |
fafe466d | 7800 | else if (GET_CODE (this_insn) == JUMP_INSN) |
7801 | { | |
7802 | if (INSN_ANNULLED_BRANCH_P (this_insn)) | |
7803 | return 0; | |
7804 | code = JUMP_INSN; | |
7805 | } | |
72bccaaa | 7806 | |
0f7bb88e | 7807 | if (set && reg_overlap_mentioned_p (reg, SET_SRC (set))) |
7808 | return 0; | |
7809 | if (set && reg_overlap_mentioned_p (reg, SET_DEST (set))) | |
7810 | { | |
7811 | if (GET_CODE (SET_DEST (set)) != MEM) | |
7812 | retval = 1; | |
7813 | else | |
7814 | return 0; | |
7815 | } | |
7816 | if (set == 0 | |
7817 | && reg_overlap_mentioned_p (reg, PATTERN (this_insn))) | |
7818 | return 0; | |
7819 | } | |
7820 | if (retval == 1) | |
7821 | return 1; | |
fafe466d | 7822 | else if (code == JUMP_INSN) |
7823 | return 0; | |
0f7bb88e | 7824 | } |
72bccaaa | 7825 | |
6720e96c | 7826 | set = single_set (insn); |
7827 | if (set && reg_overlap_mentioned_p (reg, SET_SRC (set))) | |
7828 | return 0; | |
7829 | if (set && reg_overlap_mentioned_p (reg, SET_DEST (set))) | |
7830 | return GET_CODE (SET_DEST (set)) != MEM; | |
7831 | if (set == 0 && reg_overlap_mentioned_p (reg, PATTERN (insn))) | |
7832 | return 0; | |
0f7bb88e | 7833 | |
956dd562 | 7834 | if (code == CALL_INSN && call_really_used_regs[REGNO (reg)]) |
0f7bb88e | 7835 | return 1; |
72bccaaa | 7836 | } |
7837 | return 1; | |
7838 | } | |
1b61190c | 7839 | \f |
26576132 | 7840 | #include "ggc.h" |
1b61190c | 7841 | |
1f3233d1 | 7842 | static GTY(()) rtx fpscr_rtx; |
1b61190c | 7843 | rtx |
04f04b72 | 7844 | get_fpscr_rtx (void) |
1b61190c | 7845 | { |
1b61190c | 7846 | if (! fpscr_rtx) |
7847 | { | |
407d9153 | 7848 | fpscr_rtx = gen_rtx_REG (PSImode, FPSCR_REG); |
1b61190c | 7849 | REG_USERVAR_P (fpscr_rtx) = 1; |
1b61190c | 7850 | mark_user_reg (fpscr_rtx); |
7851 | } | |
7852 | if (! reload_completed || mdep_reorg_phase != SH_AFTER_MDEP_REORG) | |
7853 | mark_user_reg (fpscr_rtx); | |
7854 | return fpscr_rtx; | |
7855 | } | |
7856 | ||
37430745 | 7857 | static GTY(()) tree fpscr_values; |
7858 | ||
7859 | static void | |
7860 | emit_fpu_switch (rtx scratch, int index) | |
7861 | { | |
7862 | rtx dst, src; | |
7863 | ||
7864 | if (fpscr_values == NULL) | |
7865 | { | |
7866 | tree t; | |
7867 | ||
7868 | t = build_index_type (integer_one_node); | |
7869 | t = build_array_type (integer_type_node, t); | |
7870 | t = build_decl (VAR_DECL, get_identifier ("__fpscr_values"), t); | |
7871 | DECL_ARTIFICIAL (t) = 1; | |
7872 | DECL_IGNORED_P (t) = 1; | |
7873 | DECL_EXTERNAL (t) = 1; | |
7874 | TREE_STATIC (t) = 1; | |
7875 | TREE_USED (t) = 1; | |
7876 | ||
7877 | fpscr_values = t; | |
7878 | } | |
7879 | ||
7880 | src = DECL_RTL (fpscr_values); | |
7881 | if (no_new_pseudos) | |
7882 | { | |
7883 | emit_move_insn (scratch, XEXP (src, 0)); | |
7884 | if (index != 0) | |
7885 | emit_insn (gen_addsi3 (scratch, scratch, GEN_INT (index * 4))); | |
7886 | src = adjust_automodify_address (src, PSImode, scratch, index * 4); | |
7887 | } | |
7888 | else | |
7889 | src = adjust_address (src, PSImode, index * 4); | |
7890 | ||
7891 | dst = get_fpscr_rtx (); | |
7892 | emit_move_insn (dst, src); | |
7893 | } | |
7894 | ||
1b61190c | 7895 | void |
04f04b72 | 7896 | emit_sf_insn (rtx pat) |
1b61190c | 7897 | { |
1b61190c | 7898 | emit_insn (pat); |
1b61190c | 7899 | } |
7900 | ||
7901 | void | |
04f04b72 | 7902 | emit_df_insn (rtx pat) |
1b61190c | 7903 | { |
1b61190c | 7904 | emit_insn (pat); |
1b61190c | 7905 | } |
7906 | ||
7907 | void | |
04f04b72 | 7908 | expand_sf_unop (rtx (*fun) (rtx, rtx, rtx), rtx *operands) |
1b61190c | 7909 | { |
7910 | emit_sf_insn ((*fun) (operands[0], operands[1], get_fpscr_rtx ())); | |
7911 | } | |
7912 | ||
7913 | void | |
04f04b72 | 7914 | expand_sf_binop (rtx (*fun) (rtx, rtx, rtx, rtx), rtx *operands) |
1b61190c | 7915 | { |
7916 | emit_sf_insn ((*fun) (operands[0], operands[1], operands[2], | |
7917 | get_fpscr_rtx ())); | |
7918 | } | |
7919 | ||
7920 | void | |
04f04b72 | 7921 | expand_df_unop (rtx (*fun) (rtx, rtx, rtx), rtx *operands) |
1b61190c | 7922 | { |
7923 | emit_df_insn ((*fun) (operands[0], operands[1], get_fpscr_rtx ())); | |
7924 | } | |
7925 | ||
7926 | void | |
04f04b72 | 7927 | expand_df_binop (rtx (*fun) (rtx, rtx, rtx, rtx), rtx *operands) |
1b61190c | 7928 | { |
7929 | emit_df_insn ((*fun) (operands[0], operands[1], operands[2], | |
d767e27e | 7930 | get_fpscr_rtx ())); |
1b61190c | 7931 | } |
1b61190c | 7932 | \f |
7933 | /* ??? gcc does flow analysis strictly after common subexpression | |
0924aa1d | 7934 | elimination. As a result, common subexpression elimination fails |
1b61190c | 7935 | when there are some intervening statements setting the same register. |
7936 | If we did nothing about this, this would hurt the precision switching | |
7937 | for SH4 badly. There is some cse after reload, but it is unable to | |
7938 | undo the extra register pressure from the unused instructions, and | |
7939 | it cannot remove auto-increment loads. | |
7940 | ||
7941 | A C code example that shows this flow/cse weakness for (at least) SH | |
7942 | and sparc (as of gcc ss-970706) is this: | |
7943 | ||
7944 | double | |
7945 | f(double a) | |
7946 | { | |
7947 | double d; | |
7948 | d = 0.1; | |
7949 | a += d; | |
7950 | d = 1.1; | |
7951 | d = 0.1; | |
7952 | a *= d; | |
7953 | return a; | |
7954 | } | |
7955 | ||
7956 | So we add another pass before common subexpression elimination, to | |
7957 | remove assignments that are dead due to a following assignment in the | |
7958 | same basic block. */ | |
7959 | ||
9e7454d0 | 7960 | static void |
04f04b72 | 7961 | mark_use (rtx x, rtx *reg_set_block) |
1b61190c | 7962 | { |
7963 | enum rtx_code code; | |
7964 | ||
7965 | if (! x) | |
7966 | return; | |
7967 | code = GET_CODE (x); | |
7968 | switch (code) | |
7969 | { | |
7970 | case REG: | |
7971 | { | |
7972 | int regno = REGNO (x); | |
7973 | int nregs = (regno < FIRST_PSEUDO_REGISTER | |
7974 | ? HARD_REGNO_NREGS (regno, GET_MODE (x)) | |
7975 | : 1); | |
7976 | do | |
7977 | { | |
7978 | reg_set_block[regno + nregs - 1] = 0; | |
7979 | } | |
7980 | while (--nregs); | |
7981 | break; | |
7982 | } | |
7983 | case SET: | |
7984 | { | |
7985 | rtx dest = SET_DEST (x); | |
7986 | ||
7987 | if (GET_CODE (dest) == SUBREG) | |
7988 | dest = SUBREG_REG (dest); | |
7989 | if (GET_CODE (dest) != REG) | |
7990 | mark_use (dest, reg_set_block); | |
7991 | mark_use (SET_SRC (x), reg_set_block); | |
7992 | break; | |
7993 | } | |
7994 | case CLOBBER: | |
7995 | break; | |
7996 | default: | |
7997 | { | |
d2ca078f | 7998 | const char *fmt = GET_RTX_FORMAT (code); |
1b61190c | 7999 | int i, j; |
8000 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
8001 | { | |
8002 | if (fmt[i] == 'e') | |
8003 | mark_use (XEXP (x, i), reg_set_block); | |
8004 | else if (fmt[i] == 'E') | |
8005 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
8006 | mark_use (XVECEXP (x, i, j), reg_set_block); | |
8007 | } | |
8008 | break; | |
8009 | } | |
8010 | } | |
8011 | } | |
18862b5a | 8012 | \f |
04f04b72 | 8013 | static rtx get_free_reg (HARD_REG_SET); |
18862b5a | 8014 | |
8015 | /* This function returns a register to use to load the address to load | |
8016 | the fpscr from. Currently it always returns r1 or r7, but when we are | |
8017 | able to use pseudo registers after combine, or have a better mechanism | |
8018 | for choosing a register, it should be done here. */ | |
8019 | /* REGS_LIVE is the liveness information for the point for which we | |
8020 | need this allocation. In some bare-bones exit blocks, r1 is live at the | |
8021 | start. We can even have all of r0..r3 being live: | |
8022 | __complex__ long long f (double d) { if (d == 0) return 2; else return 3; } | |
8023 | INSN before which new insns are placed with will clobber the register | |
8024 | we return. If a basic block consists only of setting the return value | |
8025 | register to a pseudo and using that register, the return value is not | |
8026 | live before or after this block, yet we we'll insert our insns right in | |
8027 | the middle. */ | |
1b61190c | 8028 | |
18862b5a | 8029 | static rtx |
04f04b72 | 8030 | get_free_reg (HARD_REG_SET regs_live) |
1b61190c | 8031 | { |
18862b5a | 8032 | if (! TEST_HARD_REG_BIT (regs_live, 1)) |
8033 | return gen_rtx_REG (Pmode, 1); | |
1b61190c | 8034 | |
18862b5a | 8035 | /* Hard reg 1 is live; since this is a SMALL_REGISTER_CLASSES target, |
8036 | there shouldn't be anything but a jump before the function end. */ | |
1a6a7a27 | 8037 | gcc_assert (!TEST_HARD_REG_BIT (regs_live, 7)); |
8038 | return gen_rtx_REG (Pmode, 7); | |
18862b5a | 8039 | } |
1b61190c | 8040 | |
9e7454d0 | 8041 | /* This function will set the fpscr from memory. |
18862b5a | 8042 | MODE is the mode we are setting it to. */ |
8043 | void | |
04f04b72 | 8044 | fpscr_set_from_mem (int mode, HARD_REG_SET regs_live) |
18862b5a | 8045 | { |
8046 | enum attr_fp_mode fp_mode = mode; | |
37430745 | 8047 | enum attr_fp_mode norm_mode = ACTUAL_NORMAL_MODE (FP_MODE); |
18862b5a | 8048 | rtx addr_reg = get_free_reg (regs_live); |
8049 | ||
37430745 | 8050 | emit_fpu_switch (addr_reg, fp_mode == norm_mode); |
1b61190c | 8051 | } |
2522e445 | 8052 | |
8053 | /* Is the given character a logical line separator for the assembler? */ | |
8054 | #ifndef IS_ASM_LOGICAL_LINE_SEPARATOR | |
8055 | #define IS_ASM_LOGICAL_LINE_SEPARATOR(C) ((C) == ';') | |
8056 | #endif | |
8057 | ||
8058 | int | |
04f04b72 | 8059 | sh_insn_length_adjustment (rtx insn) |
2522e445 | 8060 | { |
8061 | /* Instructions with unfilled delay slots take up an extra two bytes for | |
8062 | the nop in the delay slot. */ | |
8063 | if (((GET_CODE (insn) == INSN | |
e3fac27d | 8064 | && GET_CODE (PATTERN (insn)) != USE |
8065 | && GET_CODE (PATTERN (insn)) != CLOBBER) | |
2522e445 | 8066 | || GET_CODE (insn) == CALL_INSN |
8067 | || (GET_CODE (insn) == JUMP_INSN | |
8068 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC | |
8069 | && GET_CODE (PATTERN (insn)) != ADDR_VEC)) | |
8070 | && GET_CODE (PATTERN (NEXT_INSN (PREV_INSN (insn)))) != SEQUENCE | |
8071 | && get_attr_needs_delay_slot (insn) == NEEDS_DELAY_SLOT_YES) | |
8072 | return 2; | |
8073 | ||
87ed74ef | 8074 | /* SH2e has a bug that prevents the use of annulled branches, so if |
8075 | the delay slot is not filled, we'll have to put a NOP in it. */ | |
8076 | if (sh_cpu == CPU_SH2E | |
8077 | && GET_CODE (insn) == JUMP_INSN | |
8078 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC | |
8079 | && GET_CODE (PATTERN (insn)) != ADDR_VEC | |
8080 | && get_attr_type (insn) == TYPE_CBRANCH | |
8081 | && GET_CODE (PATTERN (NEXT_INSN (PREV_INSN (insn)))) != SEQUENCE) | |
8082 | return 2; | |
8083 | ||
2522e445 | 8084 | /* sh-dsp parallel processing insn take four bytes instead of two. */ |
9e7454d0 | 8085 | |
2522e445 | 8086 | if (GET_CODE (insn) == INSN) |
8087 | { | |
8088 | int sum = 0; | |
8089 | rtx body = PATTERN (insn); | |
5e20801c | 8090 | const char *template; |
8091 | char c; | |
2522e445 | 8092 | int maybe_label = 1; |
8093 | ||
8094 | if (GET_CODE (body) == ASM_INPUT) | |
8095 | template = XSTR (body, 0); | |
8096 | else if (asm_noperands (body) >= 0) | |
8097 | template | |
2571646d | 8098 | = decode_asm_operands (body, NULL, NULL, NULL, NULL); |
2522e445 | 8099 | else |
8100 | return 0; | |
8101 | do | |
8102 | { | |
8103 | int ppi_adjust = 0; | |
8104 | ||
8105 | do | |
8106 | c = *template++; | |
8107 | while (c == ' ' || c == '\t'); | |
8108 | /* all sh-dsp parallel-processing insns start with p. | |
8109 | The only non-ppi sh insn starting with p is pref. | |
8110 | The only ppi starting with pr is prnd. */ | |
8111 | if ((c == 'p' || c == 'P') && strncasecmp ("re", template, 2)) | |
8112 | ppi_adjust = 2; | |
8113 | /* The repeat pseudo-insn expands two three insns, a total of | |
8114 | six bytes in size. */ | |
8115 | else if ((c == 'r' || c == 'R') | |
8116 | && ! strncasecmp ("epeat", template, 5)) | |
8117 | ppi_adjust = 4; | |
8118 | while (c && c != '\n' && ! IS_ASM_LOGICAL_LINE_SEPARATOR (c)) | |
8119 | { | |
8120 | /* If this is a label, it is obviously not a ppi insn. */ | |
8121 | if (c == ':' && maybe_label) | |
8122 | { | |
8123 | ppi_adjust = 0; | |
8124 | break; | |
8125 | } | |
8126 | else if (c == '\'' || c == '"') | |
8127 | maybe_label = 0; | |
8128 | c = *template++; | |
8129 | } | |
8130 | sum += ppi_adjust; | |
8131 | maybe_label = c != ':'; | |
8132 | } | |
8133 | while (c); | |
8134 | return sum; | |
8135 | } | |
8136 | return 0; | |
8137 | } | |
0abf894c | 8138 | \f |
55e0f12f | 8139 | /* Return TRUE if X references a SYMBOL_REF or LABEL_REF whose symbol |
8140 | isn't protected by a PIC unspec. */ | |
0abf894c | 8141 | int |
04f04b72 | 8142 | nonpic_symbol_mentioned_p (rtx x) |
0abf894c | 8143 | { |
8144 | register const char *fmt; | |
8145 | register int i; | |
8146 | ||
87e19636 | 8147 | if (GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == LABEL_REF |
8148 | || GET_CODE (x) == PC) | |
0abf894c | 8149 | return 1; |
8150 | ||
87e19636 | 8151 | /* We don't want to look into the possible MEM location of a |
8152 | CONST_DOUBLE, since we're not going to use it, in general. */ | |
8153 | if (GET_CODE (x) == CONST_DOUBLE) | |
8154 | return 0; | |
8155 | ||
0abf894c | 8156 | if (GET_CODE (x) == UNSPEC |
a5bd34f1 | 8157 | && (XINT (x, 1) == UNSPEC_PIC |
8158 | || XINT (x, 1) == UNSPEC_GOT | |
8159 | || XINT (x, 1) == UNSPEC_GOTOFF | |
87e19636 | 8160 | || XINT (x, 1) == UNSPEC_GOTPLT |
cec78553 | 8161 | || XINT (x, 1) == UNSPEC_GOTTPOFF |
8162 | || XINT (x, 1) == UNSPEC_DTPOFF | |
a5bd34f1 | 8163 | || XINT (x, 1) == UNSPEC_PLT)) |
d767e27e | 8164 | return 0; |
0abf894c | 8165 | |
8166 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
8167 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
8168 | { | |
8169 | if (fmt[i] == 'E') | |
8170 | { | |
8171 | register int j; | |
8172 | ||
8173 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
8174 | if (nonpic_symbol_mentioned_p (XVECEXP (x, i, j))) | |
8175 | return 1; | |
8176 | } | |
8177 | else if (fmt[i] == 'e' && nonpic_symbol_mentioned_p (XEXP (x, i))) | |
8178 | return 1; | |
8179 | } | |
8180 | ||
8181 | return 0; | |
8182 | } | |
8183 | ||
8184 | /* Convert a non-PIC address in `orig' to a PIC address using @GOT or | |
9df03d69 | 8185 | @GOTOFF in `reg'. */ |
0abf894c | 8186 | rtx |
04f04b72 | 8187 | legitimize_pic_address (rtx orig, enum machine_mode mode ATTRIBUTE_UNUSED, |
8188 | rtx reg) | |
0abf894c | 8189 | { |
cec78553 | 8190 | if (tls_symbolic_operand (orig, Pmode)) |
8191 | return orig; | |
8192 | ||
0abf894c | 8193 | if (GET_CODE (orig) == LABEL_REF |
2fb8cf90 | 8194 | || (GET_CODE (orig) == SYMBOL_REF && SYMBOL_REF_LOCAL_P (orig))) |
0abf894c | 8195 | { |
8196 | if (reg == 0) | |
8197 | reg = gen_reg_rtx (Pmode); | |
8198 | ||
8199 | emit_insn (gen_symGOTOFF2reg (reg, orig)); | |
8200 | return reg; | |
8201 | } | |
8202 | else if (GET_CODE (orig) == SYMBOL_REF) | |
8203 | { | |
8204 | if (reg == 0) | |
8205 | reg = gen_reg_rtx (Pmode); | |
8206 | ||
8207 | emit_insn (gen_symGOT2reg (reg, orig)); | |
8208 | return reg; | |
8209 | } | |
8210 | return orig; | |
8211 | } | |
ecf6ff7c | 8212 | |
8213 | /* Mark the use of a constant in the literal table. If the constant | |
8214 | has multiple labels, make it unique. */ | |
e3fac27d | 8215 | static rtx |
04f04b72 | 8216 | mark_constant_pool_use (rtx x) |
ecf6ff7c | 8217 | { |
8218 | rtx insn, lab, pattern; | |
8219 | ||
8220 | if (x == NULL) | |
8221 | return x; | |
8222 | ||
8223 | switch (GET_CODE (x)) | |
8224 | { | |
8225 | case LABEL_REF: | |
8226 | x = XEXP (x, 0); | |
8227 | case CODE_LABEL: | |
8228 | break; | |
8229 | default: | |
8230 | return x; | |
8231 | } | |
8232 | ||
8233 | /* Get the first label in the list of labels for the same constant | |
8234 | and delete another labels in the list. */ | |
8235 | lab = x; | |
8236 | for (insn = PREV_INSN (x); insn; insn = PREV_INSN (insn)) | |
8237 | { | |
8238 | if (GET_CODE (insn) != CODE_LABEL | |
8239 | || LABEL_REFS (insn) != NEXT_INSN (insn)) | |
8240 | break; | |
8241 | lab = insn; | |
8242 | } | |
8243 | ||
8244 | for (insn = LABEL_REFS (lab); insn; insn = LABEL_REFS (insn)) | |
8245 | INSN_DELETED_P (insn) = 1; | |
8246 | ||
8247 | /* Mark constants in a window. */ | |
8248 | for (insn = NEXT_INSN (x); insn; insn = NEXT_INSN (insn)) | |
8249 | { | |
8250 | if (GET_CODE (insn) != INSN) | |
8251 | continue; | |
8252 | ||
8253 | pattern = PATTERN (insn); | |
8254 | if (GET_CODE (pattern) != UNSPEC_VOLATILE) | |
8255 | continue; | |
8256 | ||
8257 | switch (XINT (pattern, 1)) | |
8258 | { | |
8259 | case UNSPECV_CONST2: | |
8260 | case UNSPECV_CONST4: | |
8261 | case UNSPECV_CONST8: | |
8262 | XVECEXP (pattern, 0, 1) = const1_rtx; | |
8263 | break; | |
8264 | case UNSPECV_WINDOW_END: | |
8265 | if (XVECEXP (pattern, 0, 0) == x) | |
8266 | return lab; | |
8267 | break; | |
8268 | case UNSPECV_CONST_END: | |
8269 | return lab; | |
8270 | default: | |
8271 | break; | |
8272 | } | |
8273 | } | |
8274 | ||
8275 | return lab; | |
8276 | } | |
e95ebfe0 | 8277 | \f |
8278 | /* Return true if it's possible to redirect BRANCH1 to the destination | |
8279 | of an unconditional jump BRANCH2. We only want to do this if the | |
8280 | resulting branch will have a short displacement. */ | |
9e7454d0 | 8281 | int |
04f04b72 | 8282 | sh_can_redirect_branch (rtx branch1, rtx branch2) |
e95ebfe0 | 8283 | { |
8284 | if (flag_expensive_optimizations && simplejump_p (branch2)) | |
8285 | { | |
8286 | rtx dest = XEXP (SET_SRC (single_set (branch2)), 0); | |
8287 | rtx insn; | |
8288 | int distance; | |
9e7454d0 | 8289 | |
8290 | for (distance = 0, insn = NEXT_INSN (branch1); | |
8291 | insn && distance < 256; | |
e95ebfe0 | 8292 | insn = PREV_INSN (insn)) |
8293 | { | |
9e7454d0 | 8294 | if (insn == dest) |
e95ebfe0 | 8295 | return 1; |
8296 | else | |
8297 | distance += get_attr_length (insn); | |
8298 | } | |
9e7454d0 | 8299 | for (distance = 0, insn = NEXT_INSN (branch1); |
8300 | insn && distance < 256; | |
e95ebfe0 | 8301 | insn = NEXT_INSN (insn)) |
8302 | { | |
9e7454d0 | 8303 | if (insn == dest) |
e95ebfe0 | 8304 | return 1; |
8305 | else | |
8306 | distance += get_attr_length (insn); | |
8307 | } | |
8308 | } | |
8309 | return 0; | |
8310 | } | |
8311 | ||
808a491c | 8312 | /* Return nonzero if register old_reg can be renamed to register new_reg. */ |
c67138a9 | 8313 | int |
04f04b72 | 8314 | sh_hard_regno_rename_ok (unsigned int old_reg ATTRIBUTE_UNUSED, |
8315 | unsigned int new_reg) | |
c67138a9 | 8316 | { |
d767e27e | 8317 | /* Interrupt functions can only use registers that have already been |
8318 | saved by the prologue, even if they would normally be | |
8319 | call-clobbered. */ | |
c67138a9 | 8320 | |
5df3d11f | 8321 | if (sh_cfun_interrupt_handler_p () && !regs_ever_live[new_reg]) |
d767e27e | 8322 | return 0; |
c67138a9 | 8323 | |
d767e27e | 8324 | return 1; |
c67138a9 | 8325 | } |
8326 | ||
83b35a61 | 8327 | /* Function to update the integer COST |
747af5e7 | 8328 | based on the relationship between INSN that is dependent on |
8329 | DEP_INSN through the dependence LINK. The default is to make no | |
8330 | adjustment to COST. This can be used for example to specify to | |
8331 | the scheduler that an output- or anti-dependence does not incur | |
83b35a61 | 8332 | the same cost as a data-dependence. The return value should be |
8333 | the new value for COST. */ | |
747af5e7 | 8334 | static int |
04f04b72 | 8335 | sh_adjust_cost (rtx insn, rtx link ATTRIBUTE_UNUSED, rtx dep_insn, int cost) |
747af5e7 | 8336 | { |
ab55c58f | 8337 | rtx reg, use_pat; |
747af5e7 | 8338 | |
dca8af3e | 8339 | if (TARGET_SHMEDIA) |
8340 | { | |
8341 | /* On SHmedia, if the dependence is an anti-dependence or | |
ada8cf8b | 8342 | output-dependence, there is no cost. */ |
dca8af3e | 8343 | if (REG_NOTE_KIND (link) != 0) |
59312820 | 8344 | { |
8345 | /* However, dependencies between target register loads and | |
8346 | uses of the register in a subsequent block that are separated | |
8347 | by a conditional branch are not modelled - we have to do with | |
8348 | the anti-dependency between the target register load and the | |
8349 | conditional branch that ends the current block. */ | |
8350 | if (REG_NOTE_KIND (link) == REG_DEP_ANTI | |
8351 | && GET_CODE (PATTERN (dep_insn)) == SET | |
8352 | && (get_attr_type (dep_insn) == TYPE_PT_MEDIA | |
8353 | || get_attr_type (dep_insn) == TYPE_PTABS_MEDIA) | |
8354 | && get_attr_type (insn) == TYPE_CBRANCH_MEDIA) | |
8355 | { | |
8356 | int orig_cost = cost; | |
8357 | rtx note = find_reg_note (insn, REG_BR_PROB, 0); | |
8358 | rtx target = ((! note | |
8359 | || INTVAL (XEXP (note, 0)) * 2 < REG_BR_PROB_BASE) | |
8360 | ? insn : JUMP_LABEL (insn)); | |
8361 | /* On the likely path, the branch costs 1, on the unlikely path, | |
8362 | it costs 3. */ | |
8363 | cost--; | |
8364 | do | |
8365 | target = next_active_insn (target); | |
8366 | while (target && ! flow_dependent_p (target, dep_insn) | |
8367 | && --cost > 0); | |
8368 | /* If two branches are executed in immediate succession, with the | |
8369 | first branch properly predicted, this causes a stall at the | |
8370 | second branch, hence we won't need the target for the | |
8371 | second branch for two cycles after the launch of the first | |
8372 | branch. */ | |
8373 | if (cost > orig_cost - 2) | |
8374 | cost = orig_cost - 2; | |
8375 | } | |
8376 | else | |
8377 | cost = 0; | |
8378 | } | |
dca8af3e | 8379 | |
59312820 | 8380 | else if (get_attr_is_mac_media (insn) |
8381 | && get_attr_is_mac_media (dep_insn)) | |
8382 | cost = 1; | |
8383 | ||
8384 | else if (! reload_completed | |
8385 | && GET_CODE (PATTERN (insn)) == SET | |
8386 | && GET_CODE (SET_SRC (PATTERN (insn))) == FLOAT | |
8387 | && GET_CODE (PATTERN (dep_insn)) == SET | |
8388 | && fp_arith_reg_operand (SET_SRC (PATTERN (dep_insn)), VOIDmode) | |
8389 | && cost < 4) | |
8390 | cost = 4; | |
8391 | /* Schedule the ptabs for a casesi_jump_media in preference to stuff | |
8392 | that is needed at the target. */ | |
8393 | else if (get_attr_type (insn) == TYPE_JUMP_MEDIA | |
8394 | && ! flow_dependent_p (insn, dep_insn)) | |
8395 | cost--; | |
dca8af3e | 8396 | } |
ab55c58f | 8397 | else if (REG_NOTE_KIND (link) == 0) |
747af5e7 | 8398 | { |
ab55c58f | 8399 | enum attr_type dep_type, type; |
8400 | ||
8401 | if (recog_memoized (insn) < 0 | |
8402 | || recog_memoized (dep_insn) < 0) | |
83b35a61 | 8403 | return cost; |
ab55c58f | 8404 | |
8405 | dep_type = get_attr_type (dep_insn); | |
8406 | if (dep_type == TYPE_FLOAD || dep_type == TYPE_PCFLOAD) | |
8407 | cost--; | |
8408 | if ((dep_type == TYPE_LOAD_SI || dep_type == TYPE_PCLOAD_SI) | |
8409 | && (type = get_attr_type (insn)) != TYPE_CALL | |
8410 | && type != TYPE_SFUNC) | |
8411 | cost--; | |
8412 | ||
747af5e7 | 8413 | /* The only input for a call that is timing-critical is the |
8414 | function's address. */ | |
ab55c58f | 8415 | if (GET_CODE(insn) == CALL_INSN) |
8416 | { | |
8417 | rtx call = PATTERN (insn); | |
8418 | ||
8419 | if (GET_CODE (call) == PARALLEL) | |
8420 | call = XVECEXP (call, 0 ,0); | |
8421 | if (GET_CODE (call) == SET) | |
8422 | call = SET_SRC (call); | |
8423 | if (GET_CODE (call) == CALL && GET_CODE (XEXP (call, 0)) == MEM | |
59312820 | 8424 | /* sibcalli_thunk uses a symbol_ref in an unspec. */ |
8425 | && (GET_CODE (XEXP (XEXP (call, 0), 0)) == UNSPEC | |
8426 | || ! reg_set_p (XEXP (XEXP (call, 0), 0), dep_insn))) | |
ab55c58f | 8427 | cost = 0; |
8428 | } | |
747af5e7 | 8429 | /* Likewise, the most timing critical input for an sfuncs call |
8430 | is the function address. However, sfuncs typically start | |
8431 | using their arguments pretty quickly. | |
8432 | Assume a four cycle delay before they are needed. */ | |
ab55c58f | 8433 | /* All sfunc calls are parallels with at least four components. |
8434 | Exploit this to avoid unnecessary calls to sfunc_uses_reg. */ | |
8435 | else if (GET_CODE (PATTERN (insn)) == PARALLEL | |
8436 | && XVECLEN (PATTERN (insn), 0) >= 4 | |
8437 | && (reg = sfunc_uses_reg (insn))) | |
8438 | { | |
8439 | if (! reg_set_p (reg, dep_insn)) | |
8440 | cost -= 4; | |
8441 | } | |
8442 | /* When the preceding instruction loads the shift amount of | |
8443 | the following SHAD/SHLD, the latency of the load is increased | |
8444 | by 1 cycle. */ | |
8445 | else if (TARGET_SH4 | |
8446 | && get_attr_type (insn) == TYPE_DYN_SHIFT | |
8447 | && get_attr_any_int_load (dep_insn) == ANY_INT_LOAD_YES | |
8448 | && reg_overlap_mentioned_p (SET_DEST (PATTERN (dep_insn)), | |
d767e27e | 8449 | XEXP (SET_SRC (single_set (insn)), |
ab55c58f | 8450 | 1))) |
8451 | cost++; | |
8452 | /* When an LS group instruction with a latency of less than | |
8453 | 3 cycles is followed by a double-precision floating-point | |
8454 | instruction, FIPR, or FTRV, the latency of the first | |
8455 | instruction is increased to 3 cycles. */ | |
8456 | else if (cost < 3 | |
8457 | && get_attr_insn_class (dep_insn) == INSN_CLASS_LS_GROUP | |
8458 | && get_attr_dfp_comp (insn) == DFP_COMP_YES) | |
8459 | cost = 3; | |
8460 | /* The lsw register of a double-precision computation is ready one | |
8461 | cycle earlier. */ | |
8462 | else if (reload_completed | |
8463 | && get_attr_dfp_comp (dep_insn) == DFP_COMP_YES | |
8464 | && (use_pat = single_set (insn)) | |
8465 | && ! regno_use_in (REGNO (SET_DEST (single_set (dep_insn))), | |
8466 | SET_SRC (use_pat))) | |
8467 | cost -= 1; | |
8468 | ||
8469 | if (get_attr_any_fp_comp (dep_insn) == ANY_FP_COMP_YES | |
8470 | && get_attr_late_fp_use (insn) == LATE_FP_USE_YES) | |
8471 | cost -= 1; | |
8472 | } | |
8473 | /* An anti-dependence penalty of two applies if the first insn is a double | |
8474 | precision fadd / fsub / fmul. */ | |
8475 | else if (REG_NOTE_KIND (link) == REG_DEP_ANTI | |
8476 | && recog_memoized (dep_insn) >= 0 | |
8477 | && get_attr_type (dep_insn) == TYPE_DFP_ARITH | |
8478 | /* A lot of alleged anti-flow dependences are fake, | |
8479 | so check this one is real. */ | |
8480 | && flow_dependent_p (dep_insn, insn)) | |
747af5e7 | 8481 | cost = 2; |
ab55c58f | 8482 | |
747af5e7 | 8483 | |
8484 | return cost; | |
8485 | } | |
476444cc | 8486 | |
ab55c58f | 8487 | /* Check if INSN is flow-dependent on DEP_INSN. Can also be used to check |
8488 | if DEP_INSN is anti-flow dependent on INSN. */ | |
8489 | static int | |
04f04b72 | 8490 | flow_dependent_p (rtx insn, rtx dep_insn) |
ab55c58f | 8491 | { |
8492 | rtx tmp = PATTERN (insn); | |
8493 | ||
8494 | note_stores (PATTERN (dep_insn), flow_dependent_p_1, &tmp); | |
8495 | return tmp == NULL_RTX; | |
8496 | } | |
8497 | ||
8498 | /* A helper function for flow_dependent_p called through note_stores. */ | |
8499 | static void | |
04f04b72 | 8500 | flow_dependent_p_1 (rtx x, rtx pat ATTRIBUTE_UNUSED, void *data) |
ab55c58f | 8501 | { |
8502 | rtx * pinsn = (rtx *) data; | |
8503 | ||
8504 | if (*pinsn && reg_referenced_p (x, *pinsn)) | |
8505 | *pinsn = NULL_RTX; | |
8506 | } | |
8507 | ||
edf54f2a | 8508 | /* For use by sh_allocate_initial_value. Note that sh.md contains some |
476444cc | 8509 | 'special function' patterns (type sfunc) that clobber pr, but that |
8510 | do not look like function calls to leaf_function_p. Hence we must | |
8511 | do this extra check. */ | |
edf54f2a | 8512 | static int |
04f04b72 | 8513 | sh_pr_n_sets (void) |
476444cc | 8514 | { |
4a1a3f67 | 8515 | return REG_N_SETS (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG); |
476444cc | 8516 | } |
87e19636 | 8517 | |
edf54f2a | 8518 | /* Return where to allocate pseudo for a given hard register initial |
8519 | value. */ | |
8520 | static rtx | |
8521 | sh_allocate_initial_value (rtx hard_reg) | |
8522 | { | |
8523 | rtx x; | |
8524 | ||
8525 | if (REGNO (hard_reg) == (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG)) | |
8526 | { | |
8527 | if (current_function_is_leaf | |
8528 | && ! sh_pr_n_sets () | |
8529 | && ! (TARGET_SHCOMPACT | |
8530 | && ((current_function_args_info.call_cookie | |
8531 | & ~ CALL_COOKIE_RET_TRAMP (1)) | |
8532 | || current_function_has_nonlocal_label))) | |
8533 | x = hard_reg; | |
8534 | else | |
aa7e75f2 | 8535 | x = gen_frame_mem (Pmode, return_address_pointer_rtx); |
edf54f2a | 8536 | } |
8537 | else | |
8538 | x = NULL_RTX; | |
8539 | ||
8540 | return x; | |
8541 | } | |
8542 | ||
ab55c58f | 8543 | /* This function returns "2" to indicate dual issue for the SH4 |
8544 | processor. To be used by the DFA pipeline description. */ | |
bea4bad2 | 8545 | static int |
d767e27e | 8546 | sh_issue_rate (void) |
bea4bad2 | 8547 | { |
ab55c58f | 8548 | if (TARGET_SUPERSCALAR) |
8549 | return 2; | |
8550 | else | |
8551 | return 1; | |
bea4bad2 | 8552 | } |
8553 | ||
f77a5bb0 | 8554 | /* Functions for ready queue reordering for sched1. */ |
8555 | ||
8556 | /* Get weight for mode for a set x. */ | |
8557 | static short | |
b60226ef | 8558 | find_set_regmode_weight (rtx x, enum machine_mode mode) |
f77a5bb0 | 8559 | { |
8560 | if (GET_CODE (x) == CLOBBER && register_operand (SET_DEST (x), mode)) | |
8561 | return 1; | |
8562 | if (GET_CODE (x) == SET && register_operand (SET_DEST (x), mode)) | |
8563 | { | |
8564 | if (GET_CODE (SET_DEST (x)) == REG) | |
8565 | { | |
8566 | if (!reg_mentioned_p (SET_DEST (x), SET_SRC (x))) | |
8567 | return 1; | |
8568 | else | |
8569 | return 0; | |
8570 | } | |
8571 | return 1; | |
8572 | } | |
8573 | return 0; | |
8574 | } | |
8575 | ||
8576 | /* Get regmode weight for insn. */ | |
8577 | static short | |
b60226ef | 8578 | find_insn_regmode_weight (rtx insn, enum machine_mode mode) |
f77a5bb0 | 8579 | { |
8580 | short reg_weight = 0; | |
8581 | rtx x; | |
8582 | ||
8583 | /* Increment weight for each register born here. */ | |
8584 | x = PATTERN (insn); | |
8585 | reg_weight += find_set_regmode_weight (x, mode); | |
8586 | if (GET_CODE (x) == PARALLEL) | |
8587 | { | |
8588 | int j; | |
8589 | for (j = XVECLEN (x, 0) - 1; j >= 0; j--) | |
8590 | { | |
8591 | x = XVECEXP (PATTERN (insn), 0, j); | |
8592 | reg_weight += find_set_regmode_weight (x, mode); | |
8593 | } | |
8594 | } | |
8595 | /* Decrement weight for each register that dies here. */ | |
8596 | for (x = REG_NOTES (insn); x; x = XEXP (x, 1)) | |
8597 | { | |
8598 | if (REG_NOTE_KIND (x) == REG_DEAD || REG_NOTE_KIND (x) == REG_UNUSED) | |
8599 | { | |
8600 | rtx note = XEXP (x, 0); | |
8601 | if (GET_CODE (note) == REG && GET_MODE (note) == mode) | |
8602 | reg_weight--; | |
8603 | } | |
8604 | } | |
8605 | return reg_weight; | |
8606 | } | |
8607 | ||
8608 | /* Calculate regmode weights for all insns of a basic block. */ | |
8609 | static void | |
b60226ef | 8610 | find_regmode_weight (int b, enum machine_mode mode) |
f77a5bb0 | 8611 | { |
8612 | rtx insn, next_tail, head, tail; | |
8613 | ||
8614 | get_block_head_tail (b, &head, &tail); | |
8615 | next_tail = NEXT_INSN (tail); | |
8616 | ||
8617 | for (insn = head; insn != next_tail; insn = NEXT_INSN (insn)) | |
8618 | { | |
8619 | /* Handle register life information. */ | |
8620 | if (!INSN_P (insn)) | |
8621 | continue; | |
8622 | ||
8623 | if (mode == SFmode) | |
8624 | INSN_REGMODE_WEIGHT (insn, mode) = | |
8625 | find_insn_regmode_weight (insn, mode) + 2 * find_insn_regmode_weight (insn, DFmode); | |
8626 | else if (mode == SImode) | |
8627 | INSN_REGMODE_WEIGHT (insn, mode) = | |
8628 | find_insn_regmode_weight (insn, mode) + 2 * find_insn_regmode_weight (insn, DImode); | |
8629 | } | |
8630 | } | |
8631 | ||
8632 | /* Comparison function for ready queue sorting. */ | |
8633 | static int | |
b60226ef | 8634 | rank_for_reorder (const void *x, const void *y) |
f77a5bb0 | 8635 | { |
8636 | rtx tmp = *(const rtx *) y; | |
8637 | rtx tmp2 = *(const rtx *) x; | |
8638 | ||
8639 | /* The insn in a schedule group should be issued the first. */ | |
8640 | if (SCHED_GROUP_P (tmp) != SCHED_GROUP_P (tmp2)) | |
8641 | return SCHED_GROUP_P (tmp2) ? 1 : -1; | |
8642 | ||
9e7454d0 | 8643 | /* If insns are equally good, sort by INSN_LUID (original insn order), This |
f77a5bb0 | 8644 | minimizes instruction movement, thus minimizing sched's effect on |
8645 | register pressure. */ | |
8646 | return INSN_LUID (tmp) - INSN_LUID (tmp2); | |
8647 | } | |
8648 | ||
8649 | /* Resort the array A in which only element at index N may be out of order. */ | |
8650 | static void | |
b60226ef | 8651 | swap_reorder (rtx *a, int n) |
f77a5bb0 | 8652 | { |
8653 | rtx insn = a[n - 1]; | |
8654 | int i = n - 2; | |
8655 | ||
8656 | while (i >= 0 && rank_for_reorder (a + i, &insn) >= 0) | |
8657 | { | |
8658 | a[i + 1] = a[i]; | |
8659 | i -= 1; | |
8660 | } | |
8661 | a[i + 1] = insn; | |
8662 | } | |
8663 | ||
d767e27e | 8664 | #define SCHED_REORDER(READY, N_READY) \ |
8665 | do \ | |
8666 | { \ | |
8667 | if ((N_READY) == 2) \ | |
8668 | swap_reorder (READY, N_READY); \ | |
8669 | else if ((N_READY) > 2) \ | |
8670 | qsort (READY, N_READY, sizeof (rtx), rank_for_reorder); \ | |
8671 | } \ | |
8672 | while (0) | |
f77a5bb0 | 8673 | |
8674 | /* Sort the ready list READY by ascending priority, using the SCHED_REORDER | |
8675 | macro. */ | |
8676 | static void | |
b60226ef | 8677 | ready_reorder (rtx *ready, int nready) |
f77a5bb0 | 8678 | { |
8679 | SCHED_REORDER (ready, nready); | |
8680 | } | |
8681 | ||
8682 | /* Calculate regmode weights for all insns of all basic block. */ | |
8683 | static void | |
b60226ef | 8684 | sh_md_init_global (FILE *dump ATTRIBUTE_UNUSED, |
8685 | int verbose ATTRIBUTE_UNUSED, | |
8686 | int old_max_uid) | |
f77a5bb0 | 8687 | { |
8688 | basic_block b; | |
8689 | ||
8690 | regmode_weight[0] = (short *) xcalloc (old_max_uid, sizeof (short)); | |
8691 | regmode_weight[1] = (short *) xcalloc (old_max_uid, sizeof (short)); | |
8692 | ||
8693 | FOR_EACH_BB_REVERSE (b) | |
8694 | { | |
8695 | find_regmode_weight (b->index, SImode); | |
8696 | find_regmode_weight (b->index, SFmode); | |
8697 | } | |
8698 | ||
8699 | CURR_REGMODE_PRESSURE (SImode) = 0; | |
8700 | CURR_REGMODE_PRESSURE (SFmode) = 0; | |
8701 | ||
8702 | } | |
8703 | ||
8704 | /* Cleanup. */ | |
8705 | static void | |
b60226ef | 8706 | sh_md_finish_global (FILE *dump ATTRIBUTE_UNUSED, |
8707 | int verbose ATTRIBUTE_UNUSED) | |
f77a5bb0 | 8708 | { |
8709 | if (regmode_weight[0]) | |
8710 | { | |
8711 | free (regmode_weight[0]); | |
8712 | regmode_weight[0] = NULL; | |
8713 | } | |
8714 | if (regmode_weight[1]) | |
8715 | { | |
8716 | free (regmode_weight[1]); | |
8717 | regmode_weight[1] = NULL; | |
8718 | } | |
8719 | } | |
8720 | ||
8721 | /* Cache the can_issue_more so that we can return it from reorder2. Also, | |
8722 | keep count of register pressures on SImode and SFmode. */ | |
8723 | static int | |
b60226ef | 8724 | sh_variable_issue (FILE *dump ATTRIBUTE_UNUSED, |
8725 | int sched_verbose ATTRIBUTE_UNUSED, | |
8726 | rtx insn, | |
8727 | int can_issue_more) | |
f77a5bb0 | 8728 | { |
8729 | if (GET_CODE (PATTERN (insn)) != USE | |
8730 | && GET_CODE (PATTERN (insn)) != CLOBBER) | |
8731 | cached_can_issue_more = can_issue_more - 1; | |
8732 | else | |
8733 | cached_can_issue_more = can_issue_more; | |
8734 | ||
8735 | if (reload_completed) | |
8736 | return cached_can_issue_more; | |
8737 | ||
8738 | CURR_REGMODE_PRESSURE (SImode) += INSN_REGMODE_WEIGHT (insn, SImode); | |
8739 | CURR_REGMODE_PRESSURE (SFmode) += INSN_REGMODE_WEIGHT (insn, SFmode); | |
8740 | ||
8741 | return cached_can_issue_more; | |
8742 | } | |
8743 | ||
8744 | static void | |
b60226ef | 8745 | sh_md_init (FILE *dump ATTRIBUTE_UNUSED, |
8746 | int verbose ATTRIBUTE_UNUSED, | |
8747 | int veclen ATTRIBUTE_UNUSED) | |
f77a5bb0 | 8748 | { |
8749 | CURR_REGMODE_PRESSURE (SImode) = 0; | |
8750 | CURR_REGMODE_PRESSURE (SFmode) = 0; | |
8751 | } | |
8752 | ||
8753 | /* Some magic numbers. */ | |
8754 | /* Pressure on register r0 can lead to spill failures. so avoid sched1 for | |
8755 | functions that already have high pressure on r0. */ | |
8756 | #define R0_MAX_LIFE_REGIONS 2 | |
8757 | #define R0_MAX_LIVE_LENGTH 12 | |
0aee14a5 | 8758 | /* Register Pressure thresholds for SImode and SFmode registers. */ |
f77a5bb0 | 8759 | #define SIMODE_MAX_WEIGHT 5 |
8760 | #define SFMODE_MAX_WEIGHT 10 | |
8761 | ||
8762 | /* Return true if the pressure is high for MODE. */ | |
8763 | static short | |
b60226ef | 8764 | high_pressure (enum machine_mode mode) |
f77a5bb0 | 8765 | { |
8766 | /* Pressure on register r0 can lead to spill failures. so avoid sched1 for | |
8767 | functions that already have high pressure on r0. */ | |
8768 | if ((REG_N_SETS (0) - REG_N_DEATHS (0)) >= R0_MAX_LIFE_REGIONS | |
8769 | && REG_LIVE_LENGTH (0) >= R0_MAX_LIVE_LENGTH) | |
8770 | return 1; | |
8771 | ||
8772 | if (mode == SFmode) | |
8773 | return (CURR_REGMODE_PRESSURE (SFmode) > SFMODE_MAX_WEIGHT); | |
8774 | else | |
8775 | return (CURR_REGMODE_PRESSURE (SImode) > SIMODE_MAX_WEIGHT); | |
8776 | } | |
8777 | ||
8778 | /* Reorder ready queue if register pressure is high. */ | |
8779 | static int | |
b60226ef | 8780 | sh_reorder (FILE *dump ATTRIBUTE_UNUSED, |
8781 | int sched_verbose ATTRIBUTE_UNUSED, | |
8782 | rtx *ready, | |
8783 | int *n_readyp, | |
8784 | int clock_var ATTRIBUTE_UNUSED) | |
f77a5bb0 | 8785 | { |
8786 | if (reload_completed) | |
8787 | return sh_issue_rate (); | |
8788 | ||
8789 | if (high_pressure (SFmode) || high_pressure (SImode)) | |
8790 | { | |
8791 | ready_reorder (ready, *n_readyp); | |
8792 | } | |
8793 | ||
8794 | return sh_issue_rate (); | |
8795 | } | |
8796 | ||
8797 | /* Skip cycles if the current register pressure is high. */ | |
9e7454d0 | 8798 | static int |
b60226ef | 8799 | sh_reorder2 (FILE *dump ATTRIBUTE_UNUSED, |
8800 | int sched_verbose ATTRIBUTE_UNUSED, | |
8801 | rtx *ready ATTRIBUTE_UNUSED, | |
8802 | int *n_readyp ATTRIBUTE_UNUSED, | |
8803 | int clock_var ATTRIBUTE_UNUSED) | |
f77a5bb0 | 8804 | { |
8805 | if (reload_completed) | |
8806 | return cached_can_issue_more; | |
8807 | ||
9e7454d0 | 8808 | if (high_pressure(SFmode) || high_pressure (SImode)) |
f77a5bb0 | 8809 | skip_cycles = 1; |
8810 | ||
8811 | return cached_can_issue_more; | |
8812 | } | |
8813 | ||
8814 | /* Skip cycles without sorting the ready queue. This will move insn from | |
8815 | Q->R. If this is the last cycle we are skipping; allow sorting of ready | |
9e7454d0 | 8816 | queue by sh_reorder. */ |
f77a5bb0 | 8817 | |
9e7454d0 | 8818 | /* Generally, skipping these many cycles are sufficient for all insns to move |
8819 | from Q -> R. */ | |
8820 | #define MAX_SKIPS 8 | |
f77a5bb0 | 8821 | |
8822 | static int | |
b60226ef | 8823 | sh_dfa_new_cycle (FILE *sched_dump ATTRIBUTE_UNUSED, |
8824 | int sched_verbose ATTRIBUTE_UNUSED, | |
8825 | rtx insn ATTRIBUTE_UNUSED, | |
8826 | int last_clock_var, | |
8827 | int clock_var, | |
8828 | int *sort_p) | |
f77a5bb0 | 8829 | { |
8830 | if (reload_completed) | |
d767e27e | 8831 | return 0; |
f77a5bb0 | 8832 | |
9e7454d0 | 8833 | if (skip_cycles) |
f77a5bb0 | 8834 | { |
d767e27e | 8835 | if ((clock_var - last_clock_var) < MAX_SKIPS) |
8836 | { | |
8837 | *sort_p = 0; | |
8838 | return 1; | |
8839 | } | |
8840 | /* If this is the last cycle we are skipping, allow reordering of R. */ | |
8841 | if ((clock_var - last_clock_var) == MAX_SKIPS) | |
8842 | { | |
8843 | *sort_p = 1; | |
8844 | return 1; | |
8845 | } | |
f77a5bb0 | 8846 | } |
f77a5bb0 | 8847 | |
8848 | skip_cycles = 0; | |
8849 | ||
8850 | return 0; | |
8851 | } | |
8852 | ||
87e19636 | 8853 | /* SHmedia requires registers for branches, so we can't generate new |
8854 | branches past reload. */ | |
8855 | static bool | |
04f04b72 | 8856 | sh_cannot_modify_jumps_p (void) |
87e19636 | 8857 | { |
8858 | return (TARGET_SHMEDIA && (reload_in_progress || reload_completed)); | |
8859 | } | |
8860 | ||
e26dbf6d | 8861 | static int |
8af3db02 | 8862 | sh_target_reg_class (void) |
8863 | { | |
8864 | return TARGET_SHMEDIA ? TARGET_REGS : NO_REGS; | |
8865 | } | |
8866 | ||
8867 | static bool | |
8868 | sh_optimize_target_register_callee_saved (bool after_prologue_epilogue_gen) | |
8869 | { | |
59312820 | 8870 | HARD_REG_SET dummy; |
8871 | rtx insn; | |
8872 | ||
8873 | if (! shmedia_space_reserved_for_target_registers) | |
8874 | return 0; | |
8875 | if (after_prologue_epilogue_gen && ! TARGET_SAVE_ALL_TARGET_REGS) | |
8876 | return 0; | |
8877 | if (calc_live_regs (&dummy) >= 6 * 8) | |
8878 | return 1; | |
8879 | /* This is a borderline case. See if we got a nested loop, or a loop | |
8880 | with a call, or with more than 4 labels inside. */ | |
8881 | for (insn = get_insns(); insn; insn = NEXT_INSN (insn)) | |
8882 | { | |
8883 | if (GET_CODE (insn) == NOTE | |
8884 | && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG) | |
8885 | { | |
8886 | int labels = 0; | |
8887 | ||
8888 | do | |
8889 | { | |
8890 | insn = NEXT_INSN (insn); | |
8891 | if ((GET_CODE (insn) == NOTE | |
8892 | && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG) | |
8893 | || GET_CODE (insn) == CALL_INSN | |
8894 | || (GET_CODE (insn) == CODE_LABEL && ++labels > 4)) | |
8895 | return 1; | |
8896 | } | |
8897 | while (GET_CODE (insn) != NOTE | |
8898 | || NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_END); | |
8899 | } | |
8900 | } | |
8901 | return 0; | |
8af3db02 | 8902 | } |
8903 | ||
87e19636 | 8904 | static bool |
b60226ef | 8905 | sh_ms_bitfield_layout_p (tree record_type ATTRIBUTE_UNUSED) |
87e19636 | 8906 | { |
45550790 | 8907 | return (TARGET_SH5 || TARGET_HITACHI || sh_attr_renesas_p (record_type)); |
87e19636 | 8908 | } |
e40c2d35 | 8909 | \f |
9e7454d0 | 8910 | /* |
e40c2d35 | 8911 | On the SH1..SH4, the trampoline looks like |
8912 | 2 0002 D202 mov.l l2,r2 | |
8913 | 1 0000 D301 mov.l l1,r3 | |
8914 | 3 0004 422B jmp @r2 | |
8915 | 4 0006 0009 nop | |
8916 | 5 0008 00000000 l1: .long area | |
8917 | 6 000c 00000000 l2: .long function | |
8918 | ||
8919 | SH5 (compact) uses r1 instead of r3 for the static chain. */ | |
8920 | ||
8921 | ||
8922 | /* Emit RTL insns to initialize the variable parts of a trampoline. | |
8923 | FNADDR is an RTX for the address of the function's pure code. | |
8924 | CXT is an RTX for the static chain value for the function. */ | |
8925 | ||
8926 | void | |
04f04b72 | 8927 | sh_initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt) |
e40c2d35 | 8928 | { |
aa7e75f2 | 8929 | rtx tramp_mem = gen_frame_mem (BLKmode, tramp); |
8930 | ||
e40c2d35 | 8931 | if (TARGET_SHMEDIA64) |
8932 | { | |
8933 | rtx tramp_templ; | |
8934 | int fixed_len; | |
8935 | ||
8936 | rtx movi1 = GEN_INT (0xcc000010); | |
8937 | rtx shori1 = GEN_INT (0xc8000010); | |
8938 | rtx src, dst; | |
8939 | ||
8940 | /* The following trampoline works within a +- 128 KB range for cxt: | |
8941 | ptb/u cxt,tr1; movi fnaddr >> 48,r0; shori fnaddr >> 32,r0; | |
8942 | shori fnaddr >> 16,r0; shori fnaddr,r0; ptabs/l r0,tr0 | |
8943 | gettr tr1,r1; blink tr0,r63 */ | |
8944 | /* Address rounding makes it hard to compute the exact bounds of the | |
8945 | offset for this trampoline, but we have a rather generous offset | |
8946 | range, so frame_offset should do fine as an upper bound. */ | |
8947 | if (cxt == virtual_stack_vars_rtx && frame_offset < 0x20000) | |
8948 | { | |
8949 | /* ??? could optimize this trampoline initialization | |
8950 | by writing DImode words with two insns each. */ | |
8951 | rtx mask = force_reg (DImode, GEN_INT (0x3fffc00)); | |
8952 | rtx insn = gen_rtx_MINUS (DImode, cxt, tramp); | |
8953 | insn = gen_rtx_ASHIFT (DImode, insn, GEN_INT (10-2)); | |
8954 | insn = gen_rtx_AND (DImode, insn, mask); | |
8955 | /* Or in ptb/u .,tr1 pattern */ | |
8956 | insn = gen_rtx_IOR (DImode, insn, gen_int_mode (0xec000010, SImode)); | |
8957 | insn = force_operand (insn, NULL_RTX); | |
8958 | insn = gen_lowpart (SImode, insn); | |
aa7e75f2 | 8959 | emit_move_insn (change_address (tramp_mem, SImode, NULL_RTX), insn); |
e40c2d35 | 8960 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (38)); |
8961 | insn = gen_rtx_AND (DImode, insn, mask); | |
8962 | insn = force_operand (gen_rtx_IOR (DImode, movi1, insn), NULL_RTX); | |
8963 | insn = gen_lowpart (SImode, insn); | |
aa7e75f2 | 8964 | emit_move_insn (adjust_address (tramp_mem, SImode, 4), insn); |
e40c2d35 | 8965 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (22)); |
8966 | insn = gen_rtx_AND (DImode, insn, mask); | |
8967 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
8968 | insn = gen_lowpart (SImode, insn); | |
aa7e75f2 | 8969 | emit_move_insn (adjust_address (tramp_mem, SImode, 8), insn); |
e40c2d35 | 8970 | insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (6)); |
8971 | insn = gen_rtx_AND (DImode, insn, mask); | |
8972 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
8973 | insn = gen_lowpart (SImode, insn); | |
aa7e75f2 | 8974 | emit_move_insn (adjust_address (tramp_mem, SImode, 12), insn); |
e40c2d35 | 8975 | insn = gen_rtx_ASHIFT (DImode, fnaddr, GEN_INT (10)); |
8976 | insn = gen_rtx_AND (DImode, insn, mask); | |
8977 | insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX); | |
8978 | insn = gen_lowpart (SImode, insn); | |
aa7e75f2 | 8979 | emit_move_insn (adjust_address (tramp_mem, SImode, 16), insn); |
8980 | emit_move_insn (adjust_address (tramp_mem, SImode, 20), | |
e40c2d35 | 8981 | GEN_INT (0x6bf10600)); |
aa7e75f2 | 8982 | emit_move_insn (adjust_address (tramp_mem, SImode, 24), |
e40c2d35 | 8983 | GEN_INT (0x4415fc10)); |
aa7e75f2 | 8984 | emit_move_insn (adjust_address (tramp_mem, SImode, 28), |
e40c2d35 | 8985 | GEN_INT (0x4401fff0)); |
8986 | emit_insn (gen_ic_invalidate_line (tramp)); | |
8987 | return; | |
8988 | } | |
8989 | tramp_templ = gen_rtx_SYMBOL_REF (Pmode,"__GCC_nested_trampoline"); | |
8990 | fixed_len = TRAMPOLINE_SIZE - 2 * GET_MODE_SIZE (Pmode); | |
8991 | ||
8992 | tramp_templ = gen_datalabel_ref (tramp_templ); | |
aa7e75f2 | 8993 | dst = tramp_mem; |
8994 | src = gen_const_mem (BLKmode, tramp_templ); | |
e40c2d35 | 8995 | set_mem_align (dst, 256); |
8996 | set_mem_align (src, 64); | |
0378dbdc | 8997 | emit_block_move (dst, src, GEN_INT (fixed_len), BLOCK_OP_NORMAL); |
e40c2d35 | 8998 | |
aa7e75f2 | 8999 | emit_move_insn (adjust_address (tramp_mem, Pmode, fixed_len), fnaddr); |
9000 | emit_move_insn (adjust_address (tramp_mem, Pmode, | |
9001 | fixed_len + GET_MODE_SIZE (Pmode)), | |
e40c2d35 | 9002 | cxt); |
9003 | emit_insn (gen_ic_invalidate_line (tramp)); | |
9004 | return; | |
9005 | } | |
9006 | else if (TARGET_SHMEDIA) | |
9007 | { | |
9008 | /* movi fnaddr >> 16,r1; shori fnaddr,r1; ptabs/l r1,tr0 | |
9009 | movi cxt >> 16,r1; shori cxt,r1; blink tr0,r63 */ | |
9010 | rtx quad0 = gen_reg_rtx (DImode), cxtload = gen_reg_rtx (DImode); | |
9011 | rtx quad1 = gen_reg_rtx (DImode), quad2 = gen_reg_rtx (DImode); | |
9012 | /* movi 0,r1: 0xcc000010 shori 0,r1: c8000010 concatenated, | |
0924aa1d | 9013 | rotated 10 right, and higher 16 bit of every 32 selected. */ |
e40c2d35 | 9014 | rtx movishori |
9015 | = force_reg (V2HImode, (simplify_gen_subreg | |
9016 | (V2HImode, GEN_INT (0x4330432), SImode, 0))); | |
9017 | rtx ptabs = force_reg (DImode, GEN_INT (0x6bf10600)); | |
9018 | rtx blink = force_reg (DImode, GEN_INT (0x4401fff0)); | |
9019 | ||
9020 | tramp = force_reg (Pmode, tramp); | |
9021 | fnaddr = force_reg (SImode, fnaddr); | |
9022 | cxt = force_reg (SImode, cxt); | |
9023 | emit_insn (gen_mshflo_w_x (gen_rtx_SUBREG (V4HImode, quad0, 0), | |
9024 | gen_rtx_SUBREG (V2HImode, fnaddr, 0), | |
9025 | movishori)); | |
e586c16d | 9026 | emit_insn (gen_rotrdi3_mextr (quad0, quad0, |
e40c2d35 | 9027 | GEN_INT (TARGET_LITTLE_ENDIAN ? 24 : 56))); |
bcd9bd66 | 9028 | emit_insn (gen_ashldi3_media (quad0, quad0, const2_rtx)); |
aa7e75f2 | 9029 | emit_move_insn (change_address (tramp_mem, DImode, NULL_RTX), quad0); |
e40c2d35 | 9030 | emit_insn (gen_mshflo_w_x (gen_rtx_SUBREG (V4HImode, cxtload, 0), |
9031 | gen_rtx_SUBREG (V2HImode, cxt, 0), | |
9032 | movishori)); | |
e586c16d | 9033 | emit_insn (gen_rotrdi3_mextr (cxtload, cxtload, |
e40c2d35 | 9034 | GEN_INT (TARGET_LITTLE_ENDIAN ? 24 : 56))); |
bcd9bd66 | 9035 | emit_insn (gen_ashldi3_media (cxtload, cxtload, const2_rtx)); |
e40c2d35 | 9036 | if (TARGET_LITTLE_ENDIAN) |
9037 | { | |
9038 | emit_insn (gen_mshflo_l_di (quad1, ptabs, cxtload)); | |
9039 | emit_insn (gen_mextr4 (quad2, cxtload, blink)); | |
9040 | } | |
9041 | else | |
9042 | { | |
9043 | emit_insn (gen_mextr4 (quad1, cxtload, ptabs)); | |
9044 | emit_insn (gen_mshflo_l_di (quad2, blink, cxtload)); | |
9045 | } | |
aa7e75f2 | 9046 | emit_move_insn (adjust_address (tramp_mem, DImode, 8), quad1); |
9047 | emit_move_insn (adjust_address (tramp_mem, DImode, 16), quad2); | |
e40c2d35 | 9048 | emit_insn (gen_ic_invalidate_line (tramp)); |
9049 | return; | |
9050 | } | |
9051 | else if (TARGET_SHCOMPACT) | |
9052 | { | |
9053 | emit_insn (gen_initialize_trampoline (tramp, cxt, fnaddr)); | |
9054 | return; | |
9055 | } | |
aa7e75f2 | 9056 | emit_move_insn (change_address (tramp_mem, SImode, NULL_RTX), |
e40c2d35 | 9057 | gen_int_mode (TARGET_LITTLE_ENDIAN ? 0xd301d202 : 0xd202d301, |
9058 | SImode)); | |
aa7e75f2 | 9059 | emit_move_insn (adjust_address (tramp_mem, SImode, 4), |
e40c2d35 | 9060 | gen_int_mode (TARGET_LITTLE_ENDIAN ? 0x0009422b : 0x422b0009, |
9061 | SImode)); | |
aa7e75f2 | 9062 | emit_move_insn (adjust_address (tramp_mem, SImode, 8), cxt); |
9063 | emit_move_insn (adjust_address (tramp_mem, SImode, 12), fnaddr); | |
e40c2d35 | 9064 | if (TARGET_HARVARD) |
9065 | { | |
9066 | if (TARGET_USERMODE) | |
59312820 | 9067 | emit_library_call (function_symbol (NULL, "__ic_invalidate", |
9068 | FUNCTION_ORDINARY), | |
e40c2d35 | 9069 | 0, VOIDmode, 1, tramp, SImode); |
9070 | else | |
9071 | emit_insn (gen_ic_invalidate_line (tramp)); | |
9072 | } | |
9073 | } | |
9074 | ||
805e22b2 | 9075 | /* FIXME: This is overly conservative. A SHcompact function that |
9076 | receives arguments ``by reference'' will have them stored in its | |
9077 | own stack frame, so it must not pass pointers or references to | |
9078 | these arguments to other functions by means of sibling calls. */ | |
59312820 | 9079 | /* If PIC, we cannot make sibling calls to global functions |
9080 | because the PLT requires r12 to be live. */ | |
805e22b2 | 9081 | static bool |
04f04b72 | 9082 | sh_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED) |
805e22b2 | 9083 | { |
59312820 | 9084 | return (1 |
805e22b2 | 9085 | && (! TARGET_SHCOMPACT |
372faa42 | 9086 | || current_function_args_info.stack_regs == 0) |
59312820 | 9087 | && ! sh_cfun_interrupt_handler_p () |
9088 | && (! flag_pic | |
9089 | || (decl && ! TREE_PUBLIC (decl)) | |
9090 | || (decl && DECL_VISIBILITY (decl) != VISIBILITY_DEFAULT))); | |
805e22b2 | 9091 | } |
e3fac27d | 9092 | \f |
9093 | /* Machine specific built-in functions. */ | |
9094 | ||
9095 | struct builtin_description | |
9096 | { | |
9097 | const enum insn_code icode; | |
9098 | const char *const name; | |
9099 | int signature; | |
9100 | }; | |
9101 | ||
9102 | /* describe number and signedness of arguments; arg[0] == result | |
9103 | (1: unsigned, 2: signed, 4: don't care, 8: pointer 0: no argument */ | |
59312820 | 9104 | /* 9: 64 bit pointer, 10: 32 bit pointer */ |
e3fac27d | 9105 | static const char signature_args[][4] = |
9106 | { | |
9107 | #define SH_BLTIN_V2SI2 0 | |
9108 | { 4, 4 }, | |
9109 | #define SH_BLTIN_V4HI2 1 | |
9110 | { 4, 4 }, | |
9111 | #define SH_BLTIN_V2SI3 2 | |
9112 | { 4, 4, 4 }, | |
9113 | #define SH_BLTIN_V4HI3 3 | |
9114 | { 4, 4, 4 }, | |
9115 | #define SH_BLTIN_V8QI3 4 | |
9116 | { 4, 4, 4 }, | |
9117 | #define SH_BLTIN_MAC_HISI 5 | |
9118 | { 1, 4, 4, 1 }, | |
9119 | #define SH_BLTIN_SH_HI 6 | |
9120 | { 4, 4, 1 }, | |
9121 | #define SH_BLTIN_SH_SI 7 | |
9122 | { 4, 4, 1 }, | |
9123 | #define SH_BLTIN_V4HI2V2SI 8 | |
9124 | { 4, 4, 4 }, | |
9125 | #define SH_BLTIN_V4HI2V8QI 9 | |
9126 | { 4, 4, 4 }, | |
9127 | #define SH_BLTIN_SISF 10 | |
9128 | { 4, 2 }, | |
9129 | #define SH_BLTIN_LDUA_L 11 | |
59312820 | 9130 | { 2, 10 }, |
e3fac27d | 9131 | #define SH_BLTIN_LDUA_Q 12 |
59312820 | 9132 | { 1, 10 }, |
e3fac27d | 9133 | #define SH_BLTIN_STUA_L 13 |
59312820 | 9134 | { 0, 10, 2 }, |
e3fac27d | 9135 | #define SH_BLTIN_STUA_Q 14 |
59312820 | 9136 | { 0, 10, 1 }, |
9137 | #define SH_BLTIN_LDUA_L64 15 | |
9138 | { 2, 9 }, | |
9139 | #define SH_BLTIN_LDUA_Q64 16 | |
9140 | { 1, 9 }, | |
9141 | #define SH_BLTIN_STUA_L64 17 | |
9142 | { 0, 9, 2 }, | |
9143 | #define SH_BLTIN_STUA_Q64 18 | |
9144 | { 0, 9, 1 }, | |
9145 | #define SH_BLTIN_NUM_SHARED_SIGNATURES 19 | |
9146 | #define SH_BLTIN_2 19 | |
9147 | #define SH_BLTIN_SU 19 | |
e3fac27d | 9148 | { 1, 2 }, |
59312820 | 9149 | #define SH_BLTIN_3 20 |
9150 | #define SH_BLTIN_SUS 20 | |
e3fac27d | 9151 | { 2, 2, 1 }, |
59312820 | 9152 | #define SH_BLTIN_PSSV 21 |
e3fac27d | 9153 | { 0, 8, 2, 2 }, |
59312820 | 9154 | #define SH_BLTIN_XXUU 22 |
9155 | #define SH_BLTIN_UUUU 22 | |
e3fac27d | 9156 | { 1, 1, 1, 1 }, |
59312820 | 9157 | #define SH_BLTIN_PV 23 |
e3fac27d | 9158 | { 0, 8 }, |
9159 | }; | |
ada8cf8b | 9160 | /* mcmv: operands considered unsigned. */ |
e3fac27d | 9161 | /* mmulsum_wq, msad_ubq: result considered unsigned long long. */ |
ada8cf8b | 9162 | /* mperm: control value considered unsigned int. */ |
9163 | /* mshalds, mshard, mshards, mshlld, mshlrd: shift count is unsigned int. */ | |
e3fac27d | 9164 | /* mshards_q: returns signed short. */ |
9165 | /* nsb: takes long long arg, returns unsigned char. */ | |
9166 | static const struct builtin_description bdesc[] = | |
9167 | { | |
9168 | { CODE_FOR_absv2si2, "__builtin_absv2si2", SH_BLTIN_V2SI2 }, | |
9169 | { CODE_FOR_absv4hi2, "__builtin_absv4hi2", SH_BLTIN_V4HI2 }, | |
9170 | { CODE_FOR_addv2si3, "__builtin_addv2si3", SH_BLTIN_V2SI3 }, | |
9171 | { CODE_FOR_addv4hi3, "__builtin_addv4hi3", SH_BLTIN_V4HI3 }, | |
9172 | { CODE_FOR_ssaddv2si3,"__builtin_ssaddv2si3", SH_BLTIN_V2SI3 }, | |
9173 | { CODE_FOR_usaddv8qi3,"__builtin_usaddv8qi3", SH_BLTIN_V8QI3 }, | |
9174 | { CODE_FOR_ssaddv4hi3,"__builtin_ssaddv4hi3", SH_BLTIN_V4HI3 }, | |
59312820 | 9175 | { CODE_FOR_alloco_i, "__builtin_sh_media_ALLOCO", SH_BLTIN_PV }, |
e3fac27d | 9176 | { CODE_FOR_negcmpeqv8qi,"__builtin_sh_media_MCMPEQ_B", SH_BLTIN_V8QI3 }, |
9177 | { CODE_FOR_negcmpeqv2si,"__builtin_sh_media_MCMPEQ_L", SH_BLTIN_V2SI3 }, | |
9178 | { CODE_FOR_negcmpeqv4hi,"__builtin_sh_media_MCMPEQ_W", SH_BLTIN_V4HI3 }, | |
9179 | { CODE_FOR_negcmpgtuv8qi,"__builtin_sh_media_MCMPGT_UB", SH_BLTIN_V8QI3 }, | |
9180 | { CODE_FOR_negcmpgtv2si,"__builtin_sh_media_MCMPGT_L", SH_BLTIN_V2SI3 }, | |
9181 | { CODE_FOR_negcmpgtv4hi,"__builtin_sh_media_MCMPGT_W", SH_BLTIN_V4HI3 }, | |
9182 | { CODE_FOR_mcmv, "__builtin_sh_media_MCMV", SH_BLTIN_UUUU }, | |
9183 | { CODE_FOR_mcnvs_lw, "__builtin_sh_media_MCNVS_LW", SH_BLTIN_3 }, | |
9184 | { CODE_FOR_mcnvs_wb, "__builtin_sh_media_MCNVS_WB", SH_BLTIN_V4HI2V8QI }, | |
9185 | { CODE_FOR_mcnvs_wub, "__builtin_sh_media_MCNVS_WUB", SH_BLTIN_V4HI2V8QI }, | |
59312820 | 9186 | { CODE_FOR_mextr1, "__builtin_sh_media_MEXTR1", SH_BLTIN_V8QI3 }, |
9187 | { CODE_FOR_mextr2, "__builtin_sh_media_MEXTR2", SH_BLTIN_V8QI3 }, | |
9188 | { CODE_FOR_mextr3, "__builtin_sh_media_MEXTR3", SH_BLTIN_V8QI3 }, | |
9189 | { CODE_FOR_mextr4, "__builtin_sh_media_MEXTR4", SH_BLTIN_V8QI3 }, | |
9190 | { CODE_FOR_mextr5, "__builtin_sh_media_MEXTR5", SH_BLTIN_V8QI3 }, | |
9191 | { CODE_FOR_mextr6, "__builtin_sh_media_MEXTR6", SH_BLTIN_V8QI3 }, | |
9192 | { CODE_FOR_mextr7, "__builtin_sh_media_MEXTR7", SH_BLTIN_V8QI3 }, | |
e3fac27d | 9193 | { CODE_FOR_mmacfx_wl, "__builtin_sh_media_MMACFX_WL", SH_BLTIN_MAC_HISI }, |
9194 | { CODE_FOR_mmacnfx_wl,"__builtin_sh_media_MMACNFX_WL", SH_BLTIN_MAC_HISI }, | |
9195 | { CODE_FOR_mulv2si3, "__builtin_mulv2si3", SH_BLTIN_V2SI3, }, | |
9196 | { CODE_FOR_mulv4hi3, "__builtin_mulv4hi3", SH_BLTIN_V4HI3 }, | |
9197 | { CODE_FOR_mmulfx_l, "__builtin_sh_media_MMULFX_L", SH_BLTIN_V2SI3 }, | |
9198 | { CODE_FOR_mmulfx_w, "__builtin_sh_media_MMULFX_W", SH_BLTIN_V4HI3 }, | |
9199 | { CODE_FOR_mmulfxrp_w,"__builtin_sh_media_MMULFXRP_W", SH_BLTIN_V4HI3 }, | |
9200 | { CODE_FOR_mmulhi_wl, "__builtin_sh_media_MMULHI_WL", SH_BLTIN_V4HI2V2SI }, | |
9201 | { CODE_FOR_mmullo_wl, "__builtin_sh_media_MMULLO_WL", SH_BLTIN_V4HI2V2SI }, | |
9202 | { CODE_FOR_mmulsum_wq,"__builtin_sh_media_MMULSUM_WQ", SH_BLTIN_XXUU }, | |
9203 | { CODE_FOR_mperm_w, "__builtin_sh_media_MPERM_W", SH_BLTIN_SH_HI }, | |
9204 | { CODE_FOR_msad_ubq, "__builtin_sh_media_MSAD_UBQ", SH_BLTIN_XXUU }, | |
9205 | { CODE_FOR_mshalds_l, "__builtin_sh_media_MSHALDS_L", SH_BLTIN_SH_SI }, | |
9206 | { CODE_FOR_mshalds_w, "__builtin_sh_media_MSHALDS_W", SH_BLTIN_SH_HI }, | |
9207 | { CODE_FOR_ashrv2si3, "__builtin_ashrv2si3", SH_BLTIN_SH_SI }, | |
9208 | { CODE_FOR_ashrv4hi3, "__builtin_ashrv4hi3", SH_BLTIN_SH_HI }, | |
9209 | { CODE_FOR_mshards_q, "__builtin_sh_media_MSHARDS_Q", SH_BLTIN_SUS }, | |
9210 | { CODE_FOR_mshfhi_b, "__builtin_sh_media_MSHFHI_B", SH_BLTIN_V8QI3 }, | |
9211 | { CODE_FOR_mshfhi_l, "__builtin_sh_media_MSHFHI_L", SH_BLTIN_V2SI3 }, | |
9212 | { CODE_FOR_mshfhi_w, "__builtin_sh_media_MSHFHI_W", SH_BLTIN_V4HI3 }, | |
9213 | { CODE_FOR_mshflo_b, "__builtin_sh_media_MSHFLO_B", SH_BLTIN_V8QI3 }, | |
9214 | { CODE_FOR_mshflo_l, "__builtin_sh_media_MSHFLO_L", SH_BLTIN_V2SI3 }, | |
9215 | { CODE_FOR_mshflo_w, "__builtin_sh_media_MSHFLO_W", SH_BLTIN_V4HI3 }, | |
9216 | { CODE_FOR_ashlv2si3, "__builtin_ashlv2si3", SH_BLTIN_SH_SI }, | |
9217 | { CODE_FOR_ashlv4hi3, "__builtin_ashlv4hi3", SH_BLTIN_SH_HI }, | |
9218 | { CODE_FOR_lshrv2si3, "__builtin_lshrv2si3", SH_BLTIN_SH_SI }, | |
9219 | { CODE_FOR_lshrv4hi3, "__builtin_lshrv4hi3", SH_BLTIN_SH_HI }, | |
9220 | { CODE_FOR_subv2si3, "__builtin_subv2si3", SH_BLTIN_V2SI3 }, | |
9221 | { CODE_FOR_subv4hi3, "__builtin_subv4hi3", SH_BLTIN_V4HI3 }, | |
9222 | { CODE_FOR_sssubv2si3,"__builtin_sssubv2si3", SH_BLTIN_V2SI3 }, | |
9223 | { CODE_FOR_ussubv8qi3,"__builtin_ussubv8qi3", SH_BLTIN_V8QI3 }, | |
9224 | { CODE_FOR_sssubv4hi3,"__builtin_sssubv4hi3", SH_BLTIN_V4HI3 }, | |
9225 | { CODE_FOR_fcosa_s, "__builtin_sh_media_FCOSA_S", SH_BLTIN_SISF }, | |
9226 | { CODE_FOR_fsina_s, "__builtin_sh_media_FSINA_S", SH_BLTIN_SISF }, | |
9227 | { CODE_FOR_fipr, "__builtin_sh_media_FIPR_S", SH_BLTIN_3 }, | |
9228 | { CODE_FOR_ftrv, "__builtin_sh_media_FTRV_S", SH_BLTIN_3 }, | |
59312820 | 9229 | { CODE_FOR_mac_media, "__builtin_sh_media_FMAC_S", SH_BLTIN_3 }, |
9230 | { CODE_FOR_sqrtdf2, "__builtin_sh_media_FSQRT_D", SH_BLTIN_2 }, | |
9231 | { CODE_FOR_sqrtsf2, "__builtin_sh_media_FSQRT_S", SH_BLTIN_2 }, | |
e3fac27d | 9232 | { CODE_FOR_fsrra_s, "__builtin_sh_media_FSRRA_S", SH_BLTIN_2 }, |
e3fac27d | 9233 | { CODE_FOR_ldhi_l, "__builtin_sh_media_LDHI_L", SH_BLTIN_LDUA_L }, |
9234 | { CODE_FOR_ldhi_q, "__builtin_sh_media_LDHI_Q", SH_BLTIN_LDUA_Q }, | |
9235 | { CODE_FOR_ldlo_l, "__builtin_sh_media_LDLO_L", SH_BLTIN_LDUA_L }, | |
9236 | { CODE_FOR_ldlo_q, "__builtin_sh_media_LDLO_Q", SH_BLTIN_LDUA_Q }, | |
9237 | { CODE_FOR_sthi_l, "__builtin_sh_media_STHI_L", SH_BLTIN_STUA_L }, | |
9238 | { CODE_FOR_sthi_q, "__builtin_sh_media_STHI_Q", SH_BLTIN_STUA_Q }, | |
9239 | { CODE_FOR_stlo_l, "__builtin_sh_media_STLO_L", SH_BLTIN_STUA_L }, | |
9240 | { CODE_FOR_stlo_q, "__builtin_sh_media_STLO_Q", SH_BLTIN_STUA_Q }, | |
59312820 | 9241 | { CODE_FOR_ldhi_l64, "__builtin_sh_media_LDHI_L", SH_BLTIN_LDUA_L64 }, |
9242 | { CODE_FOR_ldhi_q64, "__builtin_sh_media_LDHI_Q", SH_BLTIN_LDUA_Q64 }, | |
9243 | { CODE_FOR_ldlo_l64, "__builtin_sh_media_LDLO_L", SH_BLTIN_LDUA_L64 }, | |
9244 | { CODE_FOR_ldlo_q64, "__builtin_sh_media_LDLO_Q", SH_BLTIN_LDUA_Q64 }, | |
9245 | { CODE_FOR_sthi_l64, "__builtin_sh_media_STHI_L", SH_BLTIN_STUA_L64 }, | |
9246 | { CODE_FOR_sthi_q64, "__builtin_sh_media_STHI_Q", SH_BLTIN_STUA_Q64 }, | |
9247 | { CODE_FOR_stlo_l64, "__builtin_sh_media_STLO_L", SH_BLTIN_STUA_L64 }, | |
9248 | { CODE_FOR_stlo_q64, "__builtin_sh_media_STLO_Q", SH_BLTIN_STUA_Q64 }, | |
e3fac27d | 9249 | { CODE_FOR_nsb, "__builtin_sh_media_NSB", SH_BLTIN_SU }, |
9250 | { CODE_FOR_byterev, "__builtin_sh_media_BYTEREV", SH_BLTIN_2 }, | |
59312820 | 9251 | { CODE_FOR_prefetch, "__builtin_sh_media_PREFO", SH_BLTIN_PSSV }, |
e3fac27d | 9252 | }; |
9253 | ||
9254 | static void | |
04f04b72 | 9255 | sh_media_init_builtins (void) |
e3fac27d | 9256 | { |
9257 | tree shared[SH_BLTIN_NUM_SHARED_SIGNATURES]; | |
9258 | const struct builtin_description *d; | |
9259 | ||
c5006c15 | 9260 | memset (shared, 0, sizeof shared); |
3b9d9482 | 9261 | for (d = bdesc; d - bdesc < (int) ARRAY_SIZE (bdesc); d++) |
e3fac27d | 9262 | { |
59312820 | 9263 | tree type, arg_type = 0; |
e3fac27d | 9264 | int signature = d->signature; |
9265 | int i; | |
9266 | ||
9267 | if (signature < SH_BLTIN_NUM_SHARED_SIGNATURES && shared[signature]) | |
9268 | type = shared[signature]; | |
9269 | else | |
9270 | { | |
9271 | int has_result = signature_args[signature][0] != 0; | |
9272 | ||
59312820 | 9273 | if ((signature_args[signature][1] & 8) |
9274 | && (((signature_args[signature][1] & 1) && TARGET_SHMEDIA32) | |
9275 | || ((signature_args[signature][1] & 2) && TARGET_SHMEDIA64))) | |
e3fac27d | 9276 | continue; |
9277 | if (! TARGET_FPU_ANY | |
9278 | && FLOAT_MODE_P (insn_data[d->icode].operand[0].mode)) | |
9279 | continue; | |
9280 | type = void_list_node; | |
9281 | for (i = 3; ; i--) | |
9282 | { | |
9283 | int arg = signature_args[signature][i]; | |
9284 | int opno = i - 1 + has_result; | |
9285 | ||
59312820 | 9286 | if (arg & 8) |
e3fac27d | 9287 | arg_type = ptr_type_node; |
9288 | else if (arg) | |
59312820 | 9289 | arg_type = (*lang_hooks.types.type_for_mode) |
9290 | (insn_data[d->icode].operand[opno].mode, | |
9291 | (arg & 1)); | |
e3fac27d | 9292 | else if (i) |
9293 | continue; | |
9294 | else | |
9295 | arg_type = void_type_node; | |
9296 | if (i == 0) | |
9297 | break; | |
9298 | type = tree_cons (NULL_TREE, arg_type, type); | |
9299 | } | |
9300 | type = build_function_type (arg_type, type); | |
9301 | if (signature < SH_BLTIN_NUM_SHARED_SIGNATURES) | |
9302 | shared[signature] = type; | |
9303 | } | |
e60d3615 | 9304 | lang_hooks.builtin_function (d->name, type, d - bdesc, BUILT_IN_MD, |
9305 | NULL, NULL_TREE); | |
e3fac27d | 9306 | } |
9307 | } | |
9308 | ||
9e7454d0 | 9309 | /* Implements target hook vector_mode_supported_p. */ |
9310 | bool | |
9311 | sh_vector_mode_supported_p (enum machine_mode mode) | |
9312 | { | |
9313 | if (TARGET_FPU_ANY | |
9314 | && ((mode == V2SFmode) | |
9315 | || (mode == V4SFmode) | |
9316 | || (mode == V16SFmode))) | |
9317 | return true; | |
9318 | ||
9319 | else if (TARGET_SHMEDIA | |
9320 | && ((mode == V8QImode) | |
9321 | || (mode == V2HImode) | |
9322 | || (mode == V4HImode) | |
9323 | || (mode == V2SImode))) | |
9324 | return true; | |
9325 | ||
9326 | return false; | |
9327 | } | |
9328 | ||
8ff30ff6 | 9329 | /* Implements target hook dwarf_calling_convention. Return an enum |
9330 | of dwarf_calling_convention. */ | |
9331 | int | |
9332 | sh_dwarf_calling_convention (tree func) | |
9333 | { | |
9334 | if (sh_attr_renesas_p (func)) | |
7126085a | 9335 | return DW_CC_GNU_renesas_sh; |
8ff30ff6 | 9336 | |
9337 | return DW_CC_normal; | |
9338 | } | |
9339 | ||
e3fac27d | 9340 | static void |
04f04b72 | 9341 | sh_init_builtins (void) |
e3fac27d | 9342 | { |
9343 | if (TARGET_SHMEDIA) | |
9344 | sh_media_init_builtins (); | |
9345 | } | |
9346 | ||
9347 | /* Expand an expression EXP that calls a built-in function, | |
9348 | with result going to TARGET if that's convenient | |
9349 | (and in mode MODE if that's convenient). | |
9350 | SUBTARGET may be used as the target for computing one of EXP's operands. | |
9351 | IGNORE is nonzero if the value is to be ignored. */ | |
9352 | ||
9353 | static rtx | |
04f04b72 | 9354 | sh_expand_builtin (tree exp, rtx target, rtx subtarget ATTRIBUTE_UNUSED, |
9355 | enum machine_mode mode ATTRIBUTE_UNUSED, int ignore) | |
e3fac27d | 9356 | { |
9357 | tree fndecl = TREE_OPERAND (TREE_OPERAND (exp, 0), 0); | |
9358 | tree arglist = TREE_OPERAND (exp, 1); | |
9359 | unsigned int fcode = DECL_FUNCTION_CODE (fndecl); | |
9360 | const struct builtin_description *d = &bdesc[fcode]; | |
9361 | enum insn_code icode = d->icode; | |
9362 | int signature = d->signature; | |
9363 | enum machine_mode tmode = VOIDmode; | |
9364 | int nop = 0, i; | |
9365 | rtx op[4]; | |
59312820 | 9366 | rtx pat = 0; |
e3fac27d | 9367 | |
9368 | if (signature_args[signature][0]) | |
9369 | { | |
9370 | if (ignore) | |
9371 | return 0; | |
9372 | ||
9373 | tmode = insn_data[icode].operand[0].mode; | |
9374 | if (! target | |
9375 | || GET_MODE (target) != tmode | |
9376 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
9377 | target = gen_reg_rtx (tmode); | |
9378 | op[nop++] = target; | |
9379 | } | |
9380 | else | |
9381 | target = 0; | |
9382 | ||
9383 | for (i = 1; i <= 3; i++, nop++) | |
9384 | { | |
9385 | tree arg; | |
9386 | enum machine_mode opmode, argmode; | |
59312820 | 9387 | tree optype; |
e3fac27d | 9388 | |
9389 | if (! signature_args[signature][i]) | |
9390 | break; | |
9391 | arg = TREE_VALUE (arglist); | |
dec9884d | 9392 | if (arg == error_mark_node) |
9393 | return const0_rtx; | |
e3fac27d | 9394 | arglist = TREE_CHAIN (arglist); |
59312820 | 9395 | if (signature_args[signature][i] & 8) |
9396 | { | |
9397 | opmode = ptr_mode; | |
9398 | optype = ptr_type_node; | |
9399 | } | |
9400 | else | |
9401 | { | |
9402 | opmode = insn_data[icode].operand[nop].mode; | |
9403 | optype = (*lang_hooks.types.type_for_mode) (opmode, 0); | |
9404 | } | |
e3fac27d | 9405 | argmode = TYPE_MODE (TREE_TYPE (arg)); |
9406 | if (argmode != opmode) | |
59312820 | 9407 | arg = build1 (NOP_EXPR, optype, arg); |
e3fac27d | 9408 | op[nop] = expand_expr (arg, NULL_RTX, opmode, 0); |
9409 | if (! (*insn_data[icode].operand[nop].predicate) (op[nop], opmode)) | |
9410 | op[nop] = copy_to_mode_reg (opmode, op[nop]); | |
9411 | } | |
9412 | ||
9413 | switch (nop) | |
9414 | { | |
9415 | case 1: | |
9416 | pat = (*insn_data[d->icode].genfun) (op[0]); | |
9417 | break; | |
9418 | case 2: | |
9419 | pat = (*insn_data[d->icode].genfun) (op[0], op[1]); | |
9420 | break; | |
9421 | case 3: | |
9422 | pat = (*insn_data[d->icode].genfun) (op[0], op[1], op[2]); | |
9423 | break; | |
9424 | case 4: | |
9425 | pat = (*insn_data[d->icode].genfun) (op[0], op[1], op[2], op[3]); | |
9426 | break; | |
83b35a61 | 9427 | default: |
1a6a7a27 | 9428 | gcc_unreachable (); |
e3fac27d | 9429 | } |
9430 | if (! pat) | |
9431 | return 0; | |
9432 | emit_insn (pat); | |
9433 | return target; | |
9434 | } | |
d3b29bbd | 9435 | |
9436 | void | |
04f04b72 | 9437 | sh_expand_unop_v2sf (enum rtx_code code, rtx op0, rtx op1) |
d3b29bbd | 9438 | { |
9439 | rtx sel0 = const0_rtx; | |
9440 | rtx sel1 = const1_rtx; | |
04f04b72 | 9441 | rtx (*fn) (rtx, rtx, rtx, rtx, rtx) = gen_unary_sf_op; |
d3b29bbd | 9442 | rtx op = gen_rtx_fmt_e (code, SFmode, op1); |
9443 | ||
9444 | emit_insn ((*fn) (op0, op1, op, sel0, sel0)); | |
9445 | emit_insn ((*fn) (op0, op1, op, sel1, sel1)); | |
9446 | } | |
9447 | ||
9448 | void | |
04f04b72 | 9449 | sh_expand_binop_v2sf (enum rtx_code code, rtx op0, rtx op1, rtx op2) |
d3b29bbd | 9450 | { |
9451 | rtx sel0 = const0_rtx; | |
9452 | rtx sel1 = const1_rtx; | |
04f04b72 | 9453 | rtx (*fn) (rtx, rtx, rtx, rtx, rtx, rtx, rtx, rtx) |
4aa67c38 | 9454 | = gen_binary_sf_op; |
d3b29bbd | 9455 | rtx op = gen_rtx_fmt_ee (code, SFmode, op1, op2); |
9456 | ||
4aa67c38 | 9457 | emit_insn ((*fn) (op0, op1, op2, op, sel0, sel0, sel0, sel1)); |
9458 | emit_insn ((*fn) (op0, op1, op2, op, sel1, sel1, sel1, sel0)); | |
d3b29bbd | 9459 | } |
9460 | ||
897118e8 | 9461 | /* Return the class of registers for which a mode change from FROM to TO |
9462 | is invalid. */ | |
22aae821 | 9463 | bool |
04f04b72 | 9464 | sh_cannot_change_mode_class (enum machine_mode from, enum machine_mode to, |
9465 | enum reg_class class) | |
897118e8 | 9466 | { |
9435e831 | 9467 | /* We want to enable the use of SUBREGs as a means to |
9468 | VEC_SELECT a single element of a vector. */ | |
9469 | if (to == SFmode && VECTOR_MODE_P (from) && GET_MODE_INNER (from) == SFmode) | |
9470 | return (reg_classes_intersect_p (GENERAL_REGS, class)); | |
9471 | ||
897118e8 | 9472 | if (GET_MODE_SIZE (from) != GET_MODE_SIZE (to)) |
9473 | { | |
d767e27e | 9474 | if (TARGET_LITTLE_ENDIAN) |
9475 | { | |
9476 | if (GET_MODE_SIZE (to) < 8 || GET_MODE_SIZE (from) < 8) | |
9477 | return reg_classes_intersect_p (DF_REGS, class); | |
9478 | } | |
9479 | else | |
9480 | { | |
9481 | if (GET_MODE_SIZE (from) < 8) | |
9482 | return reg_classes_intersect_p (DF_HI_REGS, class); | |
9483 | } | |
897118e8 | 9484 | } |
22aae821 | 9485 | return 0; |
897118e8 | 9486 | } |
9487 | ||
33946087 | 9488 | |
9489 | /* If ADDRESS refers to a CODE_LABEL, add NUSES to the number of times | |
9490 | that label is used. */ | |
9491 | ||
9492 | void | |
04f04b72 | 9493 | sh_mark_label (rtx address, int nuses) |
33946087 | 9494 | { |
9495 | if (GOTOFF_P (address)) | |
9496 | { | |
9497 | /* Extract the label or symbol. */ | |
9498 | address = XEXP (address, 0); | |
9499 | if (GET_CODE (address) == PLUS) | |
9500 | address = XEXP (address, 0); | |
9501 | address = XVECEXP (address, 0, 0); | |
9502 | } | |
9503 | if (GET_CODE (address) == LABEL_REF | |
9504 | && GET_CODE (XEXP (address, 0)) == CODE_LABEL) | |
9505 | LABEL_NUSES (XEXP (address, 0)) += nuses; | |
9506 | } | |
9507 | ||
83b35a61 | 9508 | /* Compute extra cost of moving data between one register class |
9509 | and another. */ | |
9510 | ||
e04da7b6 | 9511 | /* If SECONDARY*_RELOAD_CLASS says something about the src/dst pair, regclass |
9512 | uses this information. Hence, the general register <-> floating point | |
83b35a61 | 9513 | register information here is not used for SFmode. */ |
9514 | ||
9515 | int | |
04f04b72 | 9516 | sh_register_move_cost (enum machine_mode mode, |
9517 | enum reg_class srcclass, enum reg_class dstclass) | |
83b35a61 | 9518 | { |
9519 | if (dstclass == T_REGS || dstclass == PR_REGS) | |
9520 | return 10; | |
9521 | ||
f9afdd51 | 9522 | if (dstclass == MAC_REGS && srcclass == MAC_REGS) |
9523 | return 4; | |
9524 | ||
e04da7b6 | 9525 | if (mode == SImode && ! TARGET_SHMEDIA && TARGET_FMOVD |
9526 | && REGCLASS_HAS_FP_REG (srcclass) | |
9527 | && REGCLASS_HAS_FP_REG (dstclass)) | |
9528 | return 4; | |
9529 | ||
b7e76f11 | 9530 | if (REGCLASS_HAS_FP_REG (dstclass) && srcclass == T_REGS) |
9531 | return ((TARGET_HARD_SH4 && !optimize_size) ? 10 : 7); | |
9532 | ||
0ec45dc1 | 9533 | if ((REGCLASS_HAS_FP_REG (dstclass) && srcclass == MAC_REGS) |
d767e27e | 9534 | || (dstclass == MAC_REGS && REGCLASS_HAS_FP_REG (srcclass))) |
0ec45dc1 | 9535 | return 9; |
9536 | ||
83b35a61 | 9537 | if ((REGCLASS_HAS_FP_REG (dstclass) |
9538 | && REGCLASS_HAS_GENERAL_REG (srcclass)) | |
9539 | || (REGCLASS_HAS_GENERAL_REG (dstclass) | |
9540 | && REGCLASS_HAS_FP_REG (srcclass))) | |
d767e27e | 9541 | return ((TARGET_SHMEDIA ? 4 : TARGET_FMOVD ? 8 : 12) |
9542 | * ((GET_MODE_SIZE (mode) + 7) / 8U)); | |
83b35a61 | 9543 | |
9544 | if ((dstclass == FPUL_REGS | |
9545 | && REGCLASS_HAS_GENERAL_REG (srcclass)) | |
9546 | || (srcclass == FPUL_REGS | |
9547 | && REGCLASS_HAS_GENERAL_REG (dstclass))) | |
9548 | return 5; | |
9549 | ||
9550 | if ((dstclass == FPUL_REGS | |
9551 | && (srcclass == PR_REGS || srcclass == MAC_REGS || srcclass == T_REGS)) | |
9e7454d0 | 9552 | || (srcclass == FPUL_REGS |
83b35a61 | 9553 | && (dstclass == PR_REGS || dstclass == MAC_REGS))) |
9554 | return 7; | |
9555 | ||
9556 | if ((srcclass == TARGET_REGS && ! REGCLASS_HAS_GENERAL_REG (dstclass)) | |
9557 | || ((dstclass) == TARGET_REGS && ! REGCLASS_HAS_GENERAL_REG (srcclass))) | |
9558 | return 20; | |
9559 | ||
59312820 | 9560 | /* ??? ptabs faults on (value & 0x3) == 0x3 */ |
9561 | if (TARGET_SHMEDIA | |
9562 | && ((srcclass) == TARGET_REGS || (srcclass) == SIBCALL_REGS)) | |
9563 | { | |
4906bcbb | 9564 | if (sh_gettrcost >= 0) |
9565 | return sh_gettrcost; | |
59312820 | 9566 | else if (!TARGET_PT_FIXED) |
9567 | return 100; | |
9568 | } | |
9569 | ||
83b35a61 | 9570 | if ((srcclass == FPSCR_REGS && ! REGCLASS_HAS_GENERAL_REG (dstclass)) |
9571 | || (dstclass == FPSCR_REGS && ! REGCLASS_HAS_GENERAL_REG (srcclass))) | |
9572 | return 4; | |
9573 | ||
e04da7b6 | 9574 | if (TARGET_SHMEDIA |
9575 | || (TARGET_FMOVD | |
9576 | && ! REGCLASS_HAS_GENERAL_REG (srcclass) | |
9577 | && ! REGCLASS_HAS_GENERAL_REG (dstclass))) | |
9578 | return 2 * ((GET_MODE_SIZE (mode) + 7) / 8U); | |
9579 | ||
9580 | return 2 * ((GET_MODE_SIZE (mode) + 3) / 4U); | |
83b35a61 | 9581 | } |
9582 | ||
04f04b72 | 9583 | static rtx emit_load_ptr (rtx, rtx); |
c83ee613 | 9584 | |
9585 | static rtx | |
04f04b72 | 9586 | emit_load_ptr (rtx reg, rtx addr) |
c83ee613 | 9587 | { |
aa7e75f2 | 9588 | rtx mem = gen_const_mem (ptr_mode, addr); |
c83ee613 | 9589 | |
9590 | if (Pmode != ptr_mode) | |
9591 | mem = gen_rtx_SIGN_EXTEND (Pmode, mem); | |
9592 | return emit_move_insn (reg, mem); | |
9593 | } | |
9594 | ||
59312820 | 9595 | static void |
04f04b72 | 9596 | sh_output_mi_thunk (FILE *file, tree thunk_fndecl ATTRIBUTE_UNUSED, |
9597 | HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset, | |
9598 | tree function) | |
c83ee613 | 9599 | { |
9600 | CUMULATIVE_ARGS cum; | |
9601 | int structure_value_byref = 0; | |
9602 | rtx this, this_value, sibcall, insns, funexp; | |
9603 | tree funtype = TREE_TYPE (function); | |
af2c1324 | 9604 | int simple_add = CONST_OK_FOR_ADD (delta); |
c83ee613 | 9605 | int did_load = 0; |
9606 | rtx scratch0, scratch1, scratch2; | |
59312820 | 9607 | unsigned i; |
c83ee613 | 9608 | |
9609 | reload_completed = 1; | |
8af3db02 | 9610 | epilogue_completed = 1; |
c83ee613 | 9611 | no_new_pseudos = 1; |
9612 | current_function_uses_only_leaf_regs = 1; | |
656047bf | 9613 | reset_block_changes (); |
c83ee613 | 9614 | |
31b97e8f | 9615 | emit_note (NOTE_INSN_PROLOGUE_END); |
c83ee613 | 9616 | |
9617 | /* Find the "this" pointer. We have such a wide range of ABIs for the | |
9618 | SH that it's best to do this completely machine independently. | |
9e7454d0 | 9619 | "this" is passed as first argument, unless a structure return pointer |
c83ee613 | 9620 | comes first, in which case "this" comes second. */ |
30c70355 | 9621 | INIT_CUMULATIVE_ARGS (cum, funtype, NULL_RTX, 0, 1); |
c83ee613 | 9622 | #ifndef PCC_STATIC_STRUCT_RETURN |
45550790 | 9623 | if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function)) |
c83ee613 | 9624 | structure_value_byref = 1; |
9625 | #endif /* not PCC_STATIC_STRUCT_RETURN */ | |
45550790 | 9626 | if (structure_value_byref && sh_struct_value_rtx (function, 0) == 0) |
9e7454d0 | 9627 | { |
c83ee613 | 9628 | tree ptype = build_pointer_type (TREE_TYPE (funtype)); |
9629 | ||
9630 | FUNCTION_ARG_ADVANCE (cum, Pmode, ptype, 1); | |
9631 | } | |
9632 | this = FUNCTION_ARG (cum, Pmode, ptr_type_node, 1); | |
9633 | ||
9634 | /* For SHcompact, we only have r0 for a scratch register: r1 is the | |
9635 | static chain pointer (even if you can't have nested virtual functions | |
9636 | right now, someone might implement them sometime), and the rest of the | |
9637 | registers are used for argument passing, are callee-saved, or reserved. */ | |
59312820 | 9638 | /* We need to check call_used_regs / fixed_regs in case -fcall_saved-reg / |
9639 | -ffixed-reg has been used. */ | |
9640 | if (! call_used_regs[0] || fixed_regs[0]) | |
9641 | error ("r0 needs to be available as a call-clobbered register"); | |
c83ee613 | 9642 | scratch0 = scratch1 = scratch2 = gen_rtx_REG (Pmode, 0); |
9643 | if (! TARGET_SH5) | |
9644 | { | |
59312820 | 9645 | if (call_used_regs[1] && ! fixed_regs[1]) |
9646 | scratch1 = gen_rtx_REG (ptr_mode, 1); | |
c83ee613 | 9647 | /* N.B., if not TARGET_HITACHI, register 2 is used to pass the pointer |
9648 | pointing where to return struct values. */ | |
59312820 | 9649 | if (call_used_regs[3] && ! fixed_regs[3]) |
9650 | scratch2 = gen_rtx_REG (Pmode, 3); | |
c83ee613 | 9651 | } |
9652 | else if (TARGET_SHMEDIA) | |
9653 | { | |
59312820 | 9654 | for (i = FIRST_GENERAL_REG; i <= LAST_GENERAL_REG; i++) |
9655 | if (i != REGNO (scratch0) && | |
9656 | call_used_regs[i] && ! fixed_regs[i] && ! FUNCTION_ARG_REGNO_P (i)) | |
9657 | { | |
9658 | scratch1 = gen_rtx_REG (ptr_mode, i); | |
9659 | break; | |
9660 | } | |
9661 | if (scratch1 == scratch0) | |
9662 | error ("Need a second call-clobbered general purpose register"); | |
9663 | for (i = FIRST_TARGET_REG; i <= LAST_TARGET_REG; i++) | |
9664 | if (call_used_regs[i] && ! fixed_regs[i]) | |
9665 | { | |
9666 | scratch2 = gen_rtx_REG (Pmode, i); | |
9667 | break; | |
9668 | } | |
9669 | if (scratch2 == scratch0) | |
9670 | error ("Need a call-clobbered target register"); | |
c83ee613 | 9671 | } |
9672 | ||
9673 | this_value = plus_constant (this, delta); | |
9674 | if (vcall_offset | |
9675 | && (simple_add || scratch0 != scratch1) | |
9676 | && strict_memory_address_p (ptr_mode, this_value)) | |
9677 | { | |
9678 | emit_load_ptr (scratch0, this_value); | |
9679 | did_load = 1; | |
9680 | } | |
9681 | ||
9682 | if (!delta) | |
9683 | ; /* Do nothing. */ | |
9684 | else if (simple_add) | |
9685 | emit_move_insn (this, this_value); | |
9686 | else | |
9687 | { | |
9688 | emit_move_insn (scratch1, GEN_INT (delta)); | |
9689 | emit_insn (gen_add2_insn (this, scratch1)); | |
9690 | } | |
9691 | ||
9692 | if (vcall_offset) | |
9693 | { | |
9694 | rtx offset_addr; | |
9695 | ||
9696 | if (!did_load) | |
9697 | emit_load_ptr (scratch0, this); | |
9698 | ||
9699 | offset_addr = plus_constant (scratch0, vcall_offset); | |
9700 | if (strict_memory_address_p (ptr_mode, offset_addr)) | |
9701 | ; /* Do nothing. */ | |
59312820 | 9702 | else if (! TARGET_SH5 && scratch0 != scratch1) |
c83ee613 | 9703 | { |
9704 | /* scratch0 != scratch1, and we have indexed loads. Get better | |
9705 | schedule by loading the offset into r1 and using an indexed | |
9706 | load - then the load of r1 can issue before the load from | |
9707 | (this + delta) finishes. */ | |
9708 | emit_move_insn (scratch1, GEN_INT (vcall_offset)); | |
9709 | offset_addr = gen_rtx_PLUS (Pmode, scratch0, scratch1); | |
9710 | } | |
af2c1324 | 9711 | else if (CONST_OK_FOR_ADD (vcall_offset)) |
c83ee613 | 9712 | { |
9713 | emit_insn (gen_add2_insn (scratch0, GEN_INT (vcall_offset))); | |
9714 | offset_addr = scratch0; | |
9715 | } | |
9716 | else if (scratch0 != scratch1) | |
9717 | { | |
9718 | emit_move_insn (scratch1, GEN_INT (vcall_offset)); | |
9719 | emit_insn (gen_add2_insn (scratch0, scratch1)); | |
9720 | offset_addr = scratch0; | |
9721 | } | |
9722 | else | |
1a6a7a27 | 9723 | gcc_unreachable (); /* FIXME */ |
c83ee613 | 9724 | emit_load_ptr (scratch0, offset_addr); |
9725 | ||
d767e27e | 9726 | if (Pmode != ptr_mode) |
c83ee613 | 9727 | scratch0 = gen_rtx_TRUNCATE (ptr_mode, scratch0); |
9728 | emit_insn (gen_add2_insn (this, scratch0)); | |
9729 | } | |
9730 | ||
9731 | /* Generate a tail call to the target function. */ | |
9732 | if (! TREE_USED (function)) | |
9733 | { | |
9734 | assemble_external (function); | |
9735 | TREE_USED (function) = 1; | |
9736 | } | |
9737 | funexp = XEXP (DECL_RTL (function), 0); | |
59312820 | 9738 | /* If the function is overridden, so is the thunk, hence we don't |
9739 | need GOT addressing even if this is a public symbol. */ | |
9740 | #if 0 | |
9741 | if (TARGET_SH1 && ! flag_weak) | |
9742 | sibcall = gen_sibcalli_thunk (funexp, const0_rtx); | |
9743 | else | |
9744 | #endif | |
9745 | if (TARGET_SH2 && flag_pic) | |
9746 | { | |
9747 | sibcall = gen_sibcall_pcrel (funexp, const0_rtx); | |
9748 | XEXP (XVECEXP (sibcall, 0, 2), 0) = scratch2; | |
9749 | } | |
9750 | else | |
9751 | { | |
9752 | if (TARGET_SHMEDIA && flag_pic) | |
9753 | { | |
9754 | funexp = gen_sym2PIC (funexp); | |
9755 | PUT_MODE (funexp, Pmode); | |
9756 | } | |
9757 | emit_move_insn (scratch2, funexp); | |
9758 | funexp = gen_rtx_MEM (FUNCTION_MODE, scratch2); | |
9759 | sibcall = gen_sibcall (funexp, const0_rtx, NULL_RTX); | |
9760 | } | |
9761 | sibcall = emit_call_insn (sibcall); | |
c83ee613 | 9762 | SIBLING_CALL_P (sibcall) = 1; |
9763 | use_reg (&CALL_INSN_FUNCTION_USAGE (sibcall), this); | |
9764 | emit_barrier (); | |
9765 | ||
50494502 | 9766 | /* Run just enough of rest_of_compilation to do scheduling and get |
c83ee613 | 9767 | the insns emitted. Note that use_thunk calls |
9768 | assemble_start_function and assemble_end_function. */ | |
50494502 | 9769 | |
9770 | insn_locators_initialize (); | |
c83ee613 | 9771 | insns = get_insns (); |
9772 | ||
d4569a19 | 9773 | if (optimize > 0) |
c83ee613 | 9774 | { |
5c88dcc7 | 9775 | /* Initialize the bitmap obstacks. */ |
9776 | bitmap_obstack_initialize (NULL); | |
9777 | bitmap_obstack_initialize (®_obstack); | |
f763cd84 | 9778 | if (! cfun->cfg) |
ee617972 | 9779 | init_flow (); |
9780 | rtl_register_cfg_hooks (); | |
d4569a19 | 9781 | init_rtl_bb_info (ENTRY_BLOCK_PTR); |
9782 | init_rtl_bb_info (EXIT_BLOCK_PTR); | |
9783 | ENTRY_BLOCK_PTR->flags |= BB_RTL; | |
9784 | EXIT_BLOCK_PTR->flags |= BB_RTL; | |
92f146bd | 9785 | find_basic_blocks (insns); |
c83ee613 | 9786 | |
d4569a19 | 9787 | if (flag_schedule_insns_after_reload) |
9788 | { | |
9789 | life_analysis (dump_file, PROP_FINAL); | |
9790 | ||
9791 | split_all_insns (1); | |
c83ee613 | 9792 | |
d4569a19 | 9793 | schedule_insns (dump_file); |
9794 | } | |
9795 | /* We must split jmp insn in PIC case. */ | |
9796 | else if (flag_pic) | |
9797 | split_all_insns_noflow (); | |
c83ee613 | 9798 | } |
9799 | ||
2efea8c0 | 9800 | sh_reorg (); |
c83ee613 | 9801 | |
9802 | if (optimize > 0 && flag_delayed_branch) | |
d4569a19 | 9803 | dbr_schedule (insns, dump_file); |
9804 | ||
c83ee613 | 9805 | shorten_branches (insns); |
9806 | final_start_function (insns, file, 1); | |
4bf029b0 | 9807 | final (insns, file, 1); |
c83ee613 | 9808 | final_end_function (); |
9809 | ||
d4569a19 | 9810 | if (optimize > 0) |
c83ee613 | 9811 | { |
9812 | /* Release all memory allocated by flow. */ | |
e086b176 | 9813 | free_basic_block_vars (); |
5c88dcc7 | 9814 | |
9815 | /* Release the bitmap obstacks. */ | |
9816 | bitmap_obstack_release (®_obstack); | |
9817 | bitmap_obstack_release (NULL); | |
c83ee613 | 9818 | } |
9819 | ||
9820 | reload_completed = 0; | |
8af3db02 | 9821 | epilogue_completed = 0; |
c83ee613 | 9822 | no_new_pseudos = 0; |
9823 | } | |
9824 | ||
45271277 | 9825 | rtx |
59312820 | 9826 | function_symbol (rtx target, const char *name, enum sh_function_kind kind) |
45271277 | 9827 | { |
59312820 | 9828 | rtx sym; |
9829 | ||
9830 | /* If this is not an ordinary function, the name usually comes from a | |
9831 | string literal or an sprintf buffer. Make sure we use the same | |
9832 | string consistently, so that cse will be able to unify address loads. */ | |
9833 | if (kind != FUNCTION_ORDINARY) | |
9834 | name = IDENTIFIER_POINTER (get_identifier (name)); | |
9835 | sym = gen_rtx_SYMBOL_REF (Pmode, name); | |
45271277 | 9836 | SYMBOL_REF_FLAGS (sym) = SYMBOL_FLAG_FUNCTION; |
59312820 | 9837 | if (flag_pic) |
9838 | switch (kind) | |
9839 | { | |
9840 | case FUNCTION_ORDINARY: | |
9841 | break; | |
9842 | case SFUNC_GOT: | |
9843 | { | |
9844 | rtx reg = target ? target : gen_reg_rtx (Pmode); | |
9845 | ||
9846 | emit_insn (gen_symGOT2reg (reg, sym)); | |
9847 | sym = reg; | |
9848 | break; | |
9849 | } | |
9850 | case SFUNC_STATIC: | |
9851 | { | |
9852 | /* ??? To allow cse to work, we use GOTOFF relocations. | |
9853 | we could add combiner patterns to transform this into | |
9854 | straight pc-relative calls with sym2PIC / bsrf when | |
9855 | label load and function call are still 1:1 and in the | |
9856 | same basic block during combine. */ | |
9857 | rtx reg = target ? target : gen_reg_rtx (Pmode); | |
9858 | ||
9859 | emit_insn (gen_symGOTOFF2reg (reg, sym)); | |
9860 | sym = reg; | |
9861 | break; | |
9862 | } | |
9863 | } | |
9864 | if (target && sym != target) | |
9865 | { | |
9866 | emit_move_insn (target, sym); | |
9867 | return target; | |
9868 | } | |
45271277 | 9869 | return sym; |
9870 | } | |
9871 | ||
41fafa66 | 9872 | /* Find the number of a general purpose register in S. */ |
9873 | static int | |
9874 | scavenge_reg (HARD_REG_SET *s) | |
9875 | { | |
9876 | int r; | |
9877 | for (r = FIRST_GENERAL_REG; r <= LAST_GENERAL_REG; r++) | |
9878 | if (TEST_HARD_REG_BIT (*s, r)) | |
9879 | return r; | |
9880 | return -1; | |
9881 | } | |
9882 | ||
9883 | rtx | |
9884 | sh_get_pr_initial_val (void) | |
9885 | { | |
ca474d26 | 9886 | rtx val; |
9887 | ||
7a1492b5 | 9888 | /* ??? Unfortunately, get_hard_reg_initial_val doesn't always work for the |
9889 | PR register on SHcompact, because it might be clobbered by the prologue. | |
ca474d26 | 9890 | We check first if that is known to be the case. */ |
7a1492b5 | 9891 | if (TARGET_SHCOMPACT |
ca474d26 | 9892 | && ((current_function_args_info.call_cookie |
9893 | & ~ CALL_COOKIE_RET_TRAMP (1)) | |
7a1492b5 | 9894 | || current_function_has_nonlocal_label)) |
aa7e75f2 | 9895 | return gen_frame_mem (SImode, return_address_pointer_rtx); |
ca474d26 | 9896 | |
9897 | /* If we haven't finished rtl generation, there might be a nonlocal label | |
9898 | that we haven't seen yet. | |
9899 | ??? get_hard_reg_initial_val fails if it is called while no_new_pseudos | |
9900 | is set, unless it has been called before for the same register. And even | |
9901 | then, we end in trouble if we didn't use the register in the same | |
9902 | basic block before. So call get_hard_reg_initial_val now and wrap it | |
9903 | in an unspec if we might need to replace it. */ | |
ce022e7a | 9904 | /* ??? We also must do this for TARGET_SH1 in general, because otherwise |
9905 | combine can put the pseudo returned by get_hard_reg_initial_val into | |
9906 | instructions that need a general purpose registers, which will fail to | |
9907 | be recognized when the pseudo becomes allocated to PR. */ | |
ca474d26 | 9908 | val |
9909 | = get_hard_reg_initial_val (Pmode, TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG); | |
ce022e7a | 9910 | if (TARGET_SH1) |
ca474d26 | 9911 | return gen_rtx_UNSPEC (SImode, gen_rtvec (1, val), UNSPEC_RA); |
9912 | return val; | |
41fafa66 | 9913 | } |
9914 | ||
0c3578e6 | 9915 | int |
9916 | sh_expand_t_scc (enum rtx_code code, rtx target) | |
9917 | { | |
9918 | rtx result = target; | |
9919 | HOST_WIDE_INT val; | |
9920 | ||
9921 | if (GET_CODE (sh_compare_op0) != REG || REGNO (sh_compare_op0) != T_REG | |
9922 | || GET_CODE (sh_compare_op1) != CONST_INT) | |
9923 | return 0; | |
9924 | if (GET_CODE (result) != REG) | |
9925 | result = gen_reg_rtx (SImode); | |
9926 | val = INTVAL (sh_compare_op1); | |
9927 | if ((code == EQ && val == 1) || (code == NE && val == 0)) | |
9928 | emit_insn (gen_movt (result)); | |
9929 | else if ((code == EQ && val == 0) || (code == NE && val == 1)) | |
9930 | { | |
9931 | emit_insn (gen_rtx_CLOBBER (VOIDmode, result)); | |
9932 | emit_insn (gen_subc (result, result, result)); | |
bcd9bd66 | 9933 | emit_insn (gen_addsi3 (result, result, const1_rtx)); |
0c3578e6 | 9934 | } |
9935 | else if (code == EQ || code == NE) | |
9936 | emit_insn (gen_move_insn (result, GEN_INT (code == NE))); | |
9937 | else | |
9938 | return 0; | |
9939 | if (result != target) | |
9940 | emit_move_insn (target, result); | |
9941 | return 1; | |
9942 | } | |
9943 | ||
05e09565 | 9944 | /* INSN is an sfunc; return the rtx that describes the address used. */ |
9945 | static rtx | |
9946 | extract_sfunc_addr (rtx insn) | |
9947 | { | |
9948 | rtx pattern, part = NULL_RTX; | |
9949 | int len, i; | |
9950 | ||
9951 | pattern = PATTERN (insn); | |
9952 | len = XVECLEN (pattern, 0); | |
9953 | for (i = 0; i < len; i++) | |
9954 | { | |
9955 | part = XVECEXP (pattern, 0, i); | |
9956 | if (GET_CODE (part) == USE && GET_MODE (XEXP (part, 0)) == Pmode | |
9957 | && GENERAL_REGISTER_P (true_regnum (XEXP (part, 0)))) | |
9958 | return XEXP (part, 0); | |
9959 | } | |
1a6a7a27 | 9960 | gcc_assert (GET_CODE (XVECEXP (pattern, 0, 0)) == UNSPEC_VOLATILE); |
9961 | return XVECEXP (XVECEXP (pattern, 0, 0), 0, 1); | |
05e09565 | 9962 | } |
9963 | ||
9964 | /* Verify that the register in use_sfunc_addr still agrees with the address | |
9965 | used in the sfunc. This prevents fill_slots_from_thread from changing | |
9966 | use_sfunc_addr. | |
9967 | INSN is the use_sfunc_addr instruction, and REG is the register it | |
9968 | guards. */ | |
9969 | int | |
9970 | check_use_sfunc_addr (rtx insn, rtx reg) | |
9971 | { | |
9972 | /* Search for the sfunc. It should really come right after INSN. */ | |
9973 | while ((insn = NEXT_INSN (insn))) | |
9974 | { | |
9975 | if (GET_CODE (insn) == CODE_LABEL || GET_CODE (insn) == JUMP_INSN) | |
9976 | break; | |
9977 | if (! INSN_P (insn)) | |
9978 | continue; | |
9e7454d0 | 9979 | |
05e09565 | 9980 | if (GET_CODE (PATTERN (insn)) == SEQUENCE) |
9981 | insn = XVECEXP (PATTERN (insn), 0, 0); | |
9982 | if (GET_CODE (PATTERN (insn)) != PARALLEL | |
9983 | || get_attr_type (insn) != TYPE_SFUNC) | |
9984 | continue; | |
9985 | return rtx_equal_p (extract_sfunc_addr (insn), reg); | |
9986 | } | |
1a6a7a27 | 9987 | gcc_unreachable (); |
05e09565 | 9988 | } |
9989 | ||
9435e831 | 9990 | /* This function returns a constant rtx that represents pi / 2**15 in |
9991 | SFmode. it's used to scale SFmode angles, in radians, to a | |
9992 | fixed-point signed 16.16-bit fraction of a full circle, i.e., 2*pi | |
9993 | maps to 0x10000). */ | |
9994 | ||
9995 | static GTY(()) rtx sh_fsca_sf2int_rtx; | |
9996 | ||
9997 | rtx | |
9998 | sh_fsca_sf2int (void) | |
9999 | { | |
10000 | if (! sh_fsca_sf2int_rtx) | |
10001 | { | |
10002 | REAL_VALUE_TYPE rv; | |
10003 | ||
10004 | real_from_string (&rv, "10430.378350470453"); | |
10005 | sh_fsca_sf2int_rtx = const_double_from_real_value (rv, SFmode); | |
10006 | } | |
10007 | ||
10008 | return sh_fsca_sf2int_rtx; | |
10009 | } | |
9e7454d0 | 10010 | |
9435e831 | 10011 | /* This function returns a constant rtx that represents pi / 2**15 in |
10012 | DFmode. it's used to scale DFmode angles, in radians, to a | |
10013 | fixed-point signed 16.16-bit fraction of a full circle, i.e., 2*pi | |
10014 | maps to 0x10000). */ | |
10015 | ||
10016 | static GTY(()) rtx sh_fsca_df2int_rtx; | |
10017 | ||
10018 | rtx | |
10019 | sh_fsca_df2int (void) | |
10020 | { | |
10021 | if (! sh_fsca_df2int_rtx) | |
10022 | { | |
10023 | REAL_VALUE_TYPE rv; | |
10024 | ||
10025 | real_from_string (&rv, "10430.378350470453"); | |
10026 | sh_fsca_df2int_rtx = const_double_from_real_value (rv, DFmode); | |
10027 | } | |
10028 | ||
10029 | return sh_fsca_df2int_rtx; | |
10030 | } | |
9e7454d0 | 10031 | |
9435e831 | 10032 | /* This function returns a constant rtx that represents 2**15 / pi in |
10033 | SFmode. it's used to scale a fixed-point signed 16.16-bit fraction | |
10034 | of a full circle back to a SFmode value, i.e., 0x10000 maps to | |
10035 | 2*pi). */ | |
10036 | ||
10037 | static GTY(()) rtx sh_fsca_int2sf_rtx; | |
10038 | ||
10039 | rtx | |
10040 | sh_fsca_int2sf (void) | |
10041 | { | |
10042 | if (! sh_fsca_int2sf_rtx) | |
10043 | { | |
10044 | REAL_VALUE_TYPE rv; | |
10045 | ||
10046 | real_from_string (&rv, "9.587379924285257e-5"); | |
10047 | sh_fsca_int2sf_rtx = const_double_from_real_value (rv, SFmode); | |
10048 | } | |
10049 | ||
10050 | return sh_fsca_int2sf_rtx; | |
10051 | } | |
43b3dabf | 10052 | |
f8f84db9 | 10053 | /* Initialize the CUMULATIVE_ARGS structure. */ |
43b3dabf | 10054 | |
10055 | void | |
10056 | sh_init_cumulative_args (CUMULATIVE_ARGS * pcum, | |
10057 | tree fntype, | |
10058 | rtx libname ATTRIBUTE_UNUSED, | |
10059 | tree fndecl, | |
10060 | signed int n_named_args, | |
10061 | enum machine_mode mode) | |
10062 | { | |
10063 | pcum->arg_count [(int) SH_ARG_FLOAT] = 0; | |
10064 | pcum->free_single_fp_reg = 0; | |
10065 | pcum->stack_regs = 0; | |
10066 | pcum->byref_regs = 0; | |
10067 | pcum->byref = 0; | |
10068 | pcum->outgoing = (n_named_args == -1) ? 0 : 1; | |
10069 | ||
10070 | /* XXX - Should we check TARGET_HITACHI here ??? */ | |
10071 | pcum->renesas_abi = sh_attr_renesas_p (fntype) ? 1 : 0; | |
10072 | ||
10073 | if (fntype) | |
10074 | { | |
10075 | pcum->force_mem = ((TARGET_HITACHI || pcum->renesas_abi) | |
10076 | && aggregate_value_p (TREE_TYPE (fntype), fndecl)); | |
10077 | pcum->prototype_p = TYPE_ARG_TYPES (fntype) ? TRUE : FALSE; | |
10078 | pcum->arg_count [(int) SH_ARG_INT] | |
10079 | = TARGET_SH5 && aggregate_value_p (TREE_TYPE (fntype), fndecl); | |
10080 | ||
10081 | pcum->call_cookie | |
10082 | = CALL_COOKIE_RET_TRAMP (TARGET_SHCOMPACT | |
10083 | && pcum->arg_count [(int) SH_ARG_INT] == 0 | |
10084 | && (TYPE_MODE (TREE_TYPE (fntype)) == BLKmode | |
10085 | ? int_size_in_bytes (TREE_TYPE (fntype)) | |
10086 | : GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (fntype)))) > 4 | |
10087 | && (BASE_RETURN_VALUE_REG (TYPE_MODE (TREE_TYPE (fntype))) | |
10088 | == FIRST_RET_REG)); | |
10089 | } | |
10090 | else | |
10091 | { | |
10092 | pcum->arg_count [(int) SH_ARG_INT] = 0; | |
10093 | pcum->prototype_p = FALSE; | |
10094 | if (mode != VOIDmode) | |
10095 | { | |
7126085a | 10096 | pcum->call_cookie = |
43b3dabf | 10097 | CALL_COOKIE_RET_TRAMP (TARGET_SHCOMPACT |
10098 | && GET_MODE_SIZE (mode) > 4 | |
10099 | && BASE_RETURN_VALUE_REG (mode) == FIRST_RET_REG); | |
10100 | ||
10101 | /* If the default ABI is the Renesas ABI then all library | |
10102 | calls must assume that the library will be using the | |
10103 | Renesas ABI. So if the function would return its result | |
10104 | in memory then we must force the address of this memory | |
10105 | block onto the stack. Ideally we would like to call | |
10106 | targetm.calls.return_in_memory() here but we do not have | |
f8f84db9 | 10107 | the TYPE or the FNDECL available so we synthesize the |
43b3dabf | 10108 | contents of that function as best we can. */ |
10109 | pcum->force_mem = | |
156fac8b | 10110 | (TARGET_DEFAULT & MASK_HITACHI) |
43b3dabf | 10111 | && (mode == BLKmode |
10112 | || (GET_MODE_SIZE (mode) > 4 | |
10113 | && !(mode == DFmode | |
10114 | && TARGET_FPU_DOUBLE))); | |
10115 | } | |
10116 | else | |
10117 | { | |
10118 | pcum->call_cookie = 0; | |
10119 | pcum->force_mem = FALSE; | |
10120 | } | |
10121 | } | |
10122 | } | |
10123 | ||
3cebdd3a | 10124 | /* Determine if two hard register sets intersect. |
10125 | Return 1 if they do. */ | |
10126 | ||
10127 | static int | |
10128 | hard_regs_intersect_p (HARD_REG_SET *a, HARD_REG_SET *b) | |
10129 | { | |
10130 | HARD_REG_SET c; | |
10131 | COPY_HARD_REG_SET (c, *a); | |
10132 | AND_HARD_REG_SET (c, *b); | |
10133 | GO_IF_HARD_REG_SUBSET (c, reg_class_contents[(int) NO_REGS], lose); | |
10134 | return 1; | |
10135 | lose: | |
10136 | return 0; | |
10137 | } | |
10138 | ||
59312820 | 10139 | #ifdef TARGET_ADJUST_UNROLL_MAX |
10140 | static int | |
10141 | sh_adjust_unroll_max (struct loop * loop, int insn_count, | |
10142 | int max_unrolled_insns, int strength_reduce_p, | |
10143 | int unroll_type) | |
10144 | { | |
10145 | /* This doesn't work in 4.0 because the old unroller & loop.h is gone. */ | |
10146 | if (TARGET_ADJUST_UNROLL && TARGET_SHMEDIA) | |
10147 | { | |
10148 | /* Throttle back loop unrolling so that the costs of using more | |
10149 | targets than the eight target register we have don't outweigh | |
10150 | the benefits of unrolling. */ | |
10151 | rtx insn; | |
10152 | int n_labels = 0, n_calls = 0, n_exit_dest = 0, n_inner_loops = -1; | |
10153 | int n_barriers = 0; | |
10154 | rtx dest; | |
10155 | int i; | |
10156 | rtx exit_dest[8]; | |
10157 | int threshold; | |
10158 | int unroll_benefit = 0, mem_latency = 0; | |
10159 | int base_cost, best_cost, cost; | |
10160 | int factor, best_factor; | |
10161 | int n_dest; | |
10162 | unsigned max_iterations = 32767; | |
10163 | int n_iterations; | |
10164 | int need_precond = 0, precond = 0; | |
10165 | basic_block * bbs = get_loop_body (loop); | |
10166 | struct niter_desc *desc; | |
10167 | ||
10168 | /* Assume that all labels inside the loop are used from inside the | |
10169 | loop. If the loop has multiple entry points, it is unlikely to | |
10170 | be unrolled anyways. | |
10171 | Also assume that all calls are to different functions. That is | |
10172 | somewhat pessimistic, but if you have lots of calls, unrolling the | |
10173 | loop is not likely to gain you much in the first place. */ | |
10174 | i = loop->num_nodes - 1; | |
10175 | for (insn = BB_HEAD (bbs[i]); ; ) | |
10176 | { | |
10177 | if (GET_CODE (insn) == CODE_LABEL) | |
10178 | n_labels++; | |
10179 | else if (GET_CODE (insn) == CALL_INSN) | |
10180 | n_calls++; | |
10181 | else if (GET_CODE (insn) == NOTE | |
10182 | && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG) | |
10183 | n_inner_loops++; | |
10184 | else if (GET_CODE (insn) == BARRIER) | |
10185 | n_barriers++; | |
10186 | if (insn != BB_END (bbs[i])) | |
10187 | insn = NEXT_INSN (insn); | |
10188 | else if (--i >= 0) | |
10189 | insn = BB_HEAD (bbs[i]); | |
10190 | else | |
10191 | break; | |
10192 | } | |
10193 | free (bbs); | |
10194 | /* One label for the loop top is normal, and it won't be duplicated by | |
10195 | unrolling. */ | |
10196 | if (n_labels <= 1) | |
10197 | return max_unrolled_insns; | |
10198 | if (n_inner_loops > 0) | |
10199 | return 0; | |
10200 | for (dest = loop->exit_labels; dest && n_exit_dest < 8; | |
10201 | dest = LABEL_NEXTREF (dest)) | |
10202 | { | |
10203 | for (i = n_exit_dest - 1; | |
10204 | i >= 0 && XEXP (dest, 0) != XEXP (exit_dest[i], 0); i--); | |
10205 | if (i < 0) | |
10206 | exit_dest[n_exit_dest++] = dest; | |
10207 | } | |
10208 | /* If the loop top and call and exit destinations are enough to fill up | |
10209 | the target registers, we're unlikely to do any more damage by | |
10210 | unrolling. */ | |
10211 | if (n_calls + n_exit_dest >= 7) | |
10212 | return max_unrolled_insns; | |
10213 | ||
10214 | /* ??? In the new loop unroller, there is no longer any strength | |
10215 | reduction information available. Thus, when it comes to unrolling, | |
10216 | we know the cost of everything, but we know the value of nothing. */ | |
10217 | #if 0 | |
10218 | if (strength_reduce_p | |
10219 | && (unroll_type == LPT_UNROLL_RUNTIME | |
10220 | || unroll_type == LPT_UNROLL_CONSTANT | |
10221 | || unroll_type == LPT_PEEL_COMPLETELY)) | |
10222 | { | |
10223 | struct loop_ivs *ivs = LOOP_IVS (loop); | |
10224 | struct iv_class *bl; | |
10225 | ||
10226 | /* We'll save one compare-and-branch in each loop body copy | |
10227 | but the last one. */ | |
10228 | unroll_benefit = 1; | |
10229 | /* Assess the benefit of removing biv & giv updates. */ | |
10230 | for (bl = ivs->list; bl; bl = bl->next) | |
10231 | { | |
10232 | rtx increment = biv_total_increment (bl); | |
10233 | struct induction *v; | |
10234 | ||
10235 | if (increment && GET_CODE (increment) == CONST_INT) | |
10236 | { | |
10237 | unroll_benefit++; | |
10238 | for (v = bl->giv; v; v = v->next_iv) | |
10239 | { | |
10240 | if (! v->ignore && v->same == 0 | |
10241 | && GET_CODE (v->mult_val) == CONST_INT) | |
10242 | unroll_benefit++; | |
10243 | /* If this giv uses an array, try to determine | |
10244 | a maximum iteration count from the size of the | |
10245 | array. This need not be correct all the time, | |
10246 | but should not be too far off the mark too often. */ | |
10247 | while (v->giv_type == DEST_ADDR) | |
10248 | { | |
10249 | rtx mem = PATTERN (v->insn); | |
10250 | tree mem_expr, type, size_tree; | |
10251 | ||
10252 | if (GET_CODE (SET_SRC (mem)) == MEM) | |
10253 | mem = SET_SRC (mem); | |
10254 | else if (GET_CODE (SET_DEST (mem)) == MEM) | |
10255 | mem = SET_DEST (mem); | |
10256 | else | |
10257 | break; | |
10258 | mem_expr = MEM_EXPR (mem); | |
10259 | if (! mem_expr) | |
10260 | break; | |
10261 | type = TREE_TYPE (mem_expr); | |
10262 | if (TREE_CODE (type) != ARRAY_TYPE | |
10263 | || ! TYPE_SIZE (type) || ! TYPE_SIZE_UNIT (type)) | |
10264 | break; | |
10265 | size_tree = fold (build (TRUNC_DIV_EXPR, | |
10266 | bitsizetype, | |
10267 | TYPE_SIZE (type), | |
10268 | TYPE_SIZE_UNIT (type))); | |
10269 | if (TREE_CODE (size_tree) == INTEGER_CST | |
10270 | && ! TREE_INT_CST_HIGH (size_tree) | |
10271 | && TREE_INT_CST_LOW (size_tree) < max_iterations) | |
10272 | max_iterations = TREE_INT_CST_LOW (size_tree); | |
10273 | break; | |
10274 | } | |
10275 | } | |
10276 | } | |
10277 | } | |
10278 | } | |
10279 | #else /* 0 */ | |
10280 | /* Assume there is at least some benefit. */ | |
10281 | unroll_benefit = 1; | |
10282 | #endif /* 0 */ | |
10283 | ||
10284 | desc = get_simple_loop_desc (loop); | |
10285 | n_iterations = desc->const_iter ? desc->niter : 0; | |
10286 | max_iterations | |
10287 | = max_iterations < desc->niter_max ? max_iterations : desc->niter_max; | |
10288 | ||
10289 | if (! strength_reduce_p || ! n_iterations) | |
10290 | need_precond = 1; | |
10291 | if (! n_iterations) | |
10292 | { | |
10293 | n_iterations | |
10294 | = max_iterations < 3 ? max_iterations : max_iterations * 3 / 4; | |
10295 | if (! n_iterations) | |
10296 | return 0; | |
10297 | } | |
10298 | #if 0 /* ??? See above - missing induction variable information. */ | |
10299 | while (unroll_benefit > 1) /* no loop */ | |
10300 | { | |
10301 | /* We include the benefit of biv/ giv updates. Check if some or | |
10302 | all of these updates are likely to fit into a scheduling | |
10303 | bubble of a load. | |
10304 | We check for the following case: | |
10305 | - All the insns leading to the first JUMP_INSN are in a strict | |
10306 | dependency chain. | |
10307 | - there is at least one memory reference in them. | |
10308 | ||
10309 | When we find such a pattern, we assume that we can hide as many | |
10310 | updates as the total of the load latency is, if we have an | |
10311 | unroll factor of at least two. We might or might not also do | |
10312 | this without unrolling, so rather than considering this as an | |
10313 | extra unroll benefit, discount it in the unroll benefits of unroll | |
10314 | factors higher than two. */ | |
10315 | ||
10316 | rtx set, last_set; | |
10317 | ||
10318 | insn = next_active_insn (loop->start); | |
10319 | last_set = single_set (insn); | |
10320 | if (! last_set) | |
10321 | break; | |
10322 | if (GET_CODE (SET_SRC (last_set)) == MEM) | |
10323 | mem_latency += 2; | |
10324 | for (insn = NEXT_INSN (insn); insn != end; insn = NEXT_INSN (insn)) | |
10325 | { | |
10326 | if (! INSN_P (insn)) | |
10327 | continue; | |
10328 | if (GET_CODE (insn) == JUMP_INSN) | |
10329 | break; | |
10330 | if (! reg_referenced_p (SET_DEST (last_set), PATTERN (insn))) | |
10331 | { | |
10332 | /* Check if this is a to-be-reduced giv insn. */ | |
10333 | struct loop_ivs *ivs = LOOP_IVS (loop); | |
10334 | struct iv_class *bl; | |
10335 | struct induction *v; | |
10336 | for (bl = ivs->list; bl; bl = bl->next) | |
10337 | { | |
10338 | if (bl->biv->insn == insn) | |
10339 | goto is_biv; | |
10340 | for (v = bl->giv; v; v = v->next_iv) | |
10341 | if (v->insn == insn) | |
10342 | goto is_giv; | |
10343 | } | |
10344 | mem_latency--; | |
10345 | is_biv: | |
10346 | is_giv: | |
10347 | continue; | |
10348 | } | |
10349 | set = single_set (insn); | |
10350 | if (! set) | |
10351 | continue; | |
10352 | if (GET_CODE (SET_SRC (set)) == MEM) | |
10353 | mem_latency += 2; | |
10354 | last_set = set; | |
10355 | } | |
10356 | if (mem_latency < 0) | |
10357 | mem_latency = 0; | |
10358 | else if (mem_latency > unroll_benefit - 1) | |
10359 | mem_latency = unroll_benefit - 1; | |
10360 | break; | |
10361 | } | |
10362 | #endif /* 0 */ | |
10363 | if (n_labels + (unroll_benefit + n_labels * 8) / n_iterations | |
10364 | <= unroll_benefit) | |
10365 | return max_unrolled_insns; | |
10366 | ||
10367 | n_dest = n_labels + n_calls + n_exit_dest; | |
10368 | base_cost = n_dest <= 8 ? 0 : n_dest - 7; | |
10369 | best_cost = 0; | |
10370 | best_factor = 1; | |
10371 | if (n_barriers * 2 > n_labels - 1) | |
10372 | n_barriers = (n_labels - 1) / 2; | |
10373 | for (factor = 2; factor <= 8; factor++) | |
10374 | { | |
10375 | /* Bump up preconditioning cost for each power of two. */ | |
10376 | if (! (factor & (factor-1))) | |
10377 | precond += 4; | |
10378 | /* When preconditioning, only powers of two will be considered. */ | |
10379 | else if (need_precond) | |
10380 | continue; | |
10381 | n_dest = ((unroll_type != LPT_PEEL_COMPLETELY) | |
10382 | + (n_labels - 1) * factor + n_calls + n_exit_dest | |
10383 | - (n_barriers * factor >> 1) | |
10384 | + need_precond); | |
10385 | cost | |
10386 | = ((n_dest <= 8 ? 0 : n_dest - 7) | |
10387 | - base_cost * factor | |
10388 | - ((factor > 2 ? unroll_benefit - mem_latency : unroll_benefit) | |
10389 | * (factor - (unroll_type != LPT_PEEL_COMPLETELY))) | |
10390 | + ((unroll_benefit + 1 + (n_labels - 1) * factor) | |
10391 | / n_iterations)); | |
10392 | if (need_precond) | |
10393 | cost += (precond + unroll_benefit * factor / 2) / n_iterations; | |
10394 | if (cost < best_cost) | |
10395 | { | |
10396 | best_cost = cost; | |
10397 | best_factor = factor; | |
10398 | } | |
10399 | } | |
10400 | threshold = best_factor * insn_count; | |
10401 | if (max_unrolled_insns > threshold) | |
10402 | max_unrolled_insns = threshold; | |
10403 | } | |
10404 | return max_unrolled_insns; | |
10405 | } | |
10406 | #endif /* TARGET_ADJUST_UNROLL_MAX */ | |
10407 | ||
10408 | /* Replace any occurrence of FROM(n) in X with TO(n). The function does | |
10409 | not enter into CONST_DOUBLE for the replace. | |
10410 | ||
10411 | Note that copying is not done so X must not be shared unless all copies | |
10412 | are to be modified. | |
10413 | ||
10414 | This is like replace_rtx, except that we operate on N_REPLACEMENTS | |
35fe320d | 10415 | replacements simultaneously - FROM(n) is replacements[n*2] and to(n) is |
59312820 | 10416 | replacements[n*2+1] - and that we take mode changes into account. |
10417 | ||
35fe320d | 10418 | If a replacement is ambiguous, return NULL_RTX. |
59312820 | 10419 | |
10420 | If MODIFY is zero, don't modify any rtl in place, | |
10421 | just return zero or nonzero for failure / success. */ | |
10422 | ||
10423 | rtx | |
10424 | replace_n_hard_rtx (rtx x, rtx *replacements, int n_replacements, int modify) | |
10425 | { | |
10426 | int i, j; | |
10427 | const char *fmt; | |
10428 | ||
10429 | /* The following prevents loops occurrence when we change MEM in | |
10430 | CONST_DOUBLE onto the same CONST_DOUBLE. */ | |
10431 | if (x != 0 && GET_CODE (x) == CONST_DOUBLE) | |
10432 | return x; | |
10433 | ||
10434 | for (i = n_replacements - 1; i >= 0 ; i--) | |
10435 | if (x == replacements[i*2] && GET_MODE (x) == GET_MODE (replacements[i*2+1])) | |
10436 | return replacements[i*2+1]; | |
10437 | ||
10438 | /* Allow this function to make replacements in EXPR_LISTs. */ | |
10439 | if (x == 0) | |
10440 | return 0; | |
10441 | ||
10442 | if (GET_CODE (x) == SUBREG) | |
10443 | { | |
10444 | rtx new = replace_n_hard_rtx (SUBREG_REG (x), replacements, | |
10445 | n_replacements, modify); | |
10446 | ||
10447 | if (GET_CODE (new) == CONST_INT) | |
10448 | { | |
10449 | x = simplify_subreg (GET_MODE (x), new, | |
10450 | GET_MODE (SUBREG_REG (x)), | |
10451 | SUBREG_BYTE (x)); | |
10452 | if (! x) | |
10453 | abort (); | |
10454 | } | |
10455 | else if (modify) | |
10456 | SUBREG_REG (x) = new; | |
10457 | ||
10458 | return x; | |
10459 | } | |
10460 | else if (GET_CODE (x) == REG) | |
10461 | { | |
10462 | unsigned regno = REGNO (x); | |
10463 | unsigned nregs = (regno < FIRST_PSEUDO_REGISTER | |
10464 | ? HARD_REGNO_NREGS (regno, GET_MODE (x)) : 1); | |
10465 | rtx result = NULL_RTX; | |
10466 | ||
10467 | for (i = n_replacements - 1; i >= 0; i--) | |
10468 | { | |
10469 | rtx from = replacements[i*2]; | |
10470 | rtx to = replacements[i*2+1]; | |
10471 | unsigned from_regno, from_nregs, to_regno, new_regno; | |
10472 | ||
10473 | if (GET_CODE (from) != REG) | |
10474 | continue; | |
10475 | from_regno = REGNO (from); | |
10476 | from_nregs = (from_regno < FIRST_PSEUDO_REGISTER | |
10477 | ? HARD_REGNO_NREGS (from_regno, GET_MODE (from)) : 1); | |
10478 | if (regno < from_regno + from_nregs && regno + nregs > from_regno) | |
10479 | { | |
10480 | if (regno < from_regno | |
10481 | || regno + nregs > from_regno + nregs | |
10482 | || GET_CODE (to) != REG | |
10483 | || result) | |
10484 | return NULL_RTX; | |
10485 | to_regno = REGNO (to); | |
10486 | if (to_regno < FIRST_PSEUDO_REGISTER) | |
10487 | { | |
10488 | new_regno = regno + to_regno - from_regno; | |
10489 | if ((unsigned) HARD_REGNO_NREGS (new_regno, GET_MODE (x)) | |
10490 | != nregs) | |
10491 | return NULL_RTX; | |
10492 | result = gen_rtx_REG (GET_MODE (x), new_regno); | |
10493 | } | |
10494 | else if (GET_MODE (x) <= GET_MODE (to)) | |
10495 | result = gen_lowpart_common (GET_MODE (x), to); | |
10496 | else | |
10497 | result = gen_lowpart_SUBREG (GET_MODE (x), to); | |
10498 | } | |
10499 | } | |
10500 | return result ? result : x; | |
10501 | } | |
10502 | else if (GET_CODE (x) == ZERO_EXTEND) | |
10503 | { | |
10504 | rtx new = replace_n_hard_rtx (XEXP (x, 0), replacements, | |
10505 | n_replacements, modify); | |
10506 | ||
10507 | if (GET_CODE (new) == CONST_INT) | |
10508 | { | |
10509 | x = simplify_unary_operation (ZERO_EXTEND, GET_MODE (x), | |
10510 | new, GET_MODE (XEXP (x, 0))); | |
10511 | if (! x) | |
10512 | abort (); | |
10513 | } | |
10514 | else if (modify) | |
10515 | XEXP (x, 0) = new; | |
10516 | ||
10517 | return x; | |
10518 | } | |
10519 | ||
10520 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
10521 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
10522 | { | |
10523 | rtx new; | |
10524 | ||
10525 | if (fmt[i] == 'e') | |
10526 | { | |
10527 | new = replace_n_hard_rtx (XEXP (x, i), replacements, | |
10528 | n_replacements, modify); | |
10529 | if (!new) | |
10530 | return NULL_RTX; | |
10531 | if (modify) | |
10532 | XEXP (x, i) = new; | |
10533 | } | |
10534 | else if (fmt[i] == 'E') | |
10535 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
10536 | { | |
10537 | new = replace_n_hard_rtx (XVECEXP (x, i, j), replacements, | |
10538 | n_replacements, modify); | |
10539 | if (!new) | |
10540 | return NULL_RTX; | |
10541 | if (modify) | |
10542 | XVECEXP (x, i, j) = new; | |
10543 | } | |
10544 | } | |
10545 | ||
10546 | return x; | |
10547 | } | |
10548 | ||
10549 | rtx | |
10550 | sh_gen_truncate (enum machine_mode mode, rtx x, int need_sign_ext) | |
10551 | { | |
10552 | enum rtx_code code = TRUNCATE; | |
10553 | ||
10554 | if (GET_CODE (x) == ZERO_EXTEND || GET_CODE (x) == SIGN_EXTEND) | |
10555 | { | |
10556 | rtx inner = XEXP (x, 0); | |
10557 | enum machine_mode inner_mode = GET_MODE (inner); | |
10558 | ||
10559 | if (inner_mode == mode) | |
10560 | return inner; | |
10561 | else if (GET_MODE_SIZE (inner_mode) >= GET_MODE_SIZE (mode)) | |
10562 | x = inner; | |
10563 | else if (GET_MODE_SIZE (inner_mode) < GET_MODE_SIZE (mode) | |
10564 | && (! need_sign_ext || GET_CODE (x) == SIGN_EXTEND)) | |
10565 | { | |
10566 | code = GET_CODE (x); | |
10567 | x = inner; | |
10568 | } | |
10569 | } | |
10570 | return gen_rtx_fmt_e (code, mode, x); | |
10571 | } | |
10572 | ||
10573 | /* called via for_each_rtx after reload, to clean up truncates of | |
10574 | registers that span multiple actual hard registers. */ | |
10575 | int | |
10576 | shmedia_cleanup_truncate (rtx *p, void *n_changes) | |
10577 | { | |
10578 | rtx x = *p, reg; | |
10579 | ||
10580 | if (GET_CODE (x) != TRUNCATE) | |
10581 | return 0; | |
10582 | reg = XEXP (x, 0); | |
10583 | if (GET_MODE_SIZE (GET_MODE (reg)) > 8 && GET_CODE (reg) == REG) | |
10584 | { | |
10585 | enum machine_mode reg_mode = GET_MODE (reg); | |
10586 | XEXP (x, 0) = simplify_subreg (DImode, reg, reg_mode, | |
10587 | subreg_lowpart_offset (DImode, reg_mode)); | |
10588 | *(int*) n_changes += 1; | |
10589 | return -1; | |
10590 | } | |
10591 | return 0; | |
10592 | } | |
10593 | ||
10594 | /* Load and store depend on the highpart of the address. However, | |
10595 | set_attr_alternative does not give well-defined results before reload, | |
10596 | so we must look at the rtl ourselves to see if any of the feeding | |
10597 | registers is used in a memref. */ | |
10598 | ||
10599 | /* Called by sh_contains_memref_p via for_each_rtx. */ | |
10600 | static int | |
10601 | sh_contains_memref_p_1 (rtx *loc, void *data ATTRIBUTE_UNUSED) | |
10602 | { | |
10603 | return (GET_CODE (*loc) == MEM); | |
10604 | } | |
10605 | ||
3ce7ff97 | 10606 | /* Return nonzero iff INSN contains a MEM. */ |
59312820 | 10607 | int |
10608 | sh_contains_memref_p (rtx insn) | |
10609 | { | |
10610 | return for_each_rtx (&PATTERN (insn), &sh_contains_memref_p_1, NULL); | |
10611 | } | |
10612 | ||
10613 | /* FNADDR is the MEM expression from a call expander. Return an address | |
10614 | to use in an SHmedia insn pattern. */ | |
10615 | rtx | |
10616 | shmedia_prepare_call_address (rtx fnaddr, int is_sibcall) | |
10617 | { | |
10618 | int is_sym; | |
10619 | ||
10620 | fnaddr = XEXP (fnaddr, 0); | |
10621 | is_sym = GET_CODE (fnaddr) == SYMBOL_REF; | |
10622 | if (flag_pic && is_sym) | |
10623 | { | |
10624 | if (! SYMBOL_REF_LOCAL_P (fnaddr)) | |
10625 | { | |
10626 | rtx reg = gen_reg_rtx (Pmode); | |
10627 | ||
10628 | /* We must not use GOTPLT for sibcalls, because PIC_REG | |
10629 | must be restored before the PLT code gets to run. */ | |
10630 | if (is_sibcall) | |
10631 | emit_insn (gen_symGOT2reg (reg, fnaddr)); | |
10632 | else | |
10633 | emit_insn (gen_symGOTPLT2reg (reg, fnaddr)); | |
10634 | fnaddr = reg; | |
10635 | } | |
10636 | else | |
10637 | { | |
10638 | fnaddr = gen_sym2PIC (fnaddr); | |
10639 | PUT_MODE (fnaddr, Pmode); | |
10640 | } | |
10641 | } | |
10642 | /* If ptabs might trap, make this visible to the rest of the compiler. | |
10643 | We generally assume that symbols pertain to valid locations, but | |
10644 | it is possible to generate invalid symbols with asm or linker tricks. | |
35fe320d | 10645 | In a list of functions where each returns its successor, an invalid |
59312820 | 10646 | symbol might denote an empty list. */ |
10647 | if (!TARGET_PT_FIXED | |
10648 | && (!is_sym || TARGET_INVALID_SYMBOLS) | |
10649 | && (!REG_P (fnaddr) || ! TARGET_REGISTER_P (REGNO (fnaddr)))) | |
10650 | { | |
10651 | rtx tr = gen_reg_rtx (PDImode); | |
10652 | ||
10653 | emit_insn (gen_ptabs (tr, fnaddr)); | |
10654 | fnaddr = tr; | |
10655 | } | |
10656 | else if (! target_reg_operand (fnaddr, Pmode)) | |
10657 | fnaddr = copy_to_mode_reg (Pmode, fnaddr); | |
10658 | return fnaddr; | |
10659 | } | |
10660 | ||
59312820 | 10661 | enum sh_divide_strategy_e sh_div_strategy = SH_DIV_STRATEGY_DEFAULT; |
10662 | ||
10663 | /* This defines the storage for the variable part of a -mboard= option. | |
10664 | It is only required when using the sh-superh-elf target */ | |
10665 | #ifdef _SUPERH_H | |
10666 | const char * boardtype = "7750p2"; | |
10667 | const char * osruntime = "bare"; | |
10668 | #endif | |
10669 | ||
1f3233d1 | 10670 | #include "gt-sh.h" |