]>
Commit | Line | Data |
---|---|---|
7cf0dbf3 | 1 | /* Copyright (C) 2006, 2007, 2008, 2009, 2010 Free Software Foundation, Inc. |
644459d0 | 2 | |
3 | This file is free software; you can redistribute it and/or modify it under | |
4 | the terms of the GNU General Public License as published by the Free | |
038d1e19 | 5 | Software Foundation; either version 3 of the License, or (at your option) |
644459d0 | 6 | any later version. |
7 | ||
8 | This file is distributed in the hope that it will be useful, but WITHOUT | |
9 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
10 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
11 | for more details. | |
12 | ||
13 | You should have received a copy of the GNU General Public License | |
038d1e19 | 14 | along with GCC; see the file COPYING3. If not see |
15 | <http://www.gnu.org/licenses/>. */ | |
644459d0 | 16 | |
17 | #include "config.h" | |
18 | #include "system.h" | |
19 | #include "coretypes.h" | |
20 | #include "tm.h" | |
21 | #include "rtl.h" | |
22 | #include "regs.h" | |
23 | #include "hard-reg-set.h" | |
644459d0 | 24 | #include "insn-config.h" |
25 | #include "conditions.h" | |
26 | #include "insn-attr.h" | |
27 | #include "flags.h" | |
28 | #include "recog.h" | |
29 | #include "obstack.h" | |
30 | #include "tree.h" | |
31 | #include "expr.h" | |
32 | #include "optabs.h" | |
33 | #include "except.h" | |
34 | #include "function.h" | |
35 | #include "output.h" | |
36 | #include "basic-block.h" | |
37 | #include "integrate.h" | |
0b205f4c | 38 | #include "diagnostic-core.h" |
644459d0 | 39 | #include "ggc.h" |
40 | #include "hashtab.h" | |
41 | #include "tm_p.h" | |
42 | #include "target.h" | |
43 | #include "target-def.h" | |
44 | #include "langhooks.h" | |
45 | #include "reload.h" | |
46 | #include "cfglayout.h" | |
47 | #include "sched-int.h" | |
48 | #include "params.h" | |
644459d0 | 49 | #include "machmode.h" |
75a70cf9 | 50 | #include "gimple.h" |
644459d0 | 51 | #include "tm-constrs.h" |
d52fd16a | 52 | #include "ddg.h" |
5a976006 | 53 | #include "sbitmap.h" |
54 | #include "timevar.h" | |
55 | #include "df.h" | |
6352eedf | 56 | |
57 | /* Builtin types, data and prototypes. */ | |
c2233b46 | 58 | |
59 | enum spu_builtin_type_index | |
60 | { | |
61 | SPU_BTI_END_OF_PARAMS, | |
62 | ||
63 | /* We create new type nodes for these. */ | |
64 | SPU_BTI_V16QI, | |
65 | SPU_BTI_V8HI, | |
66 | SPU_BTI_V4SI, | |
67 | SPU_BTI_V2DI, | |
68 | SPU_BTI_V4SF, | |
69 | SPU_BTI_V2DF, | |
70 | SPU_BTI_UV16QI, | |
71 | SPU_BTI_UV8HI, | |
72 | SPU_BTI_UV4SI, | |
73 | SPU_BTI_UV2DI, | |
74 | ||
75 | /* A 16-byte type. (Implemented with V16QI_type_node) */ | |
76 | SPU_BTI_QUADWORD, | |
77 | ||
78 | /* These all correspond to intSI_type_node */ | |
79 | SPU_BTI_7, | |
80 | SPU_BTI_S7, | |
81 | SPU_BTI_U7, | |
82 | SPU_BTI_S10, | |
83 | SPU_BTI_S10_4, | |
84 | SPU_BTI_U14, | |
85 | SPU_BTI_16, | |
86 | SPU_BTI_S16, | |
87 | SPU_BTI_S16_2, | |
88 | SPU_BTI_U16, | |
89 | SPU_BTI_U16_2, | |
90 | SPU_BTI_U18, | |
91 | ||
92 | /* These correspond to the standard types */ | |
93 | SPU_BTI_INTQI, | |
94 | SPU_BTI_INTHI, | |
95 | SPU_BTI_INTSI, | |
96 | SPU_BTI_INTDI, | |
97 | ||
98 | SPU_BTI_UINTQI, | |
99 | SPU_BTI_UINTHI, | |
100 | SPU_BTI_UINTSI, | |
101 | SPU_BTI_UINTDI, | |
102 | ||
103 | SPU_BTI_FLOAT, | |
104 | SPU_BTI_DOUBLE, | |
105 | ||
106 | SPU_BTI_VOID, | |
107 | SPU_BTI_PTR, | |
108 | ||
109 | SPU_BTI_MAX | |
110 | }; | |
111 | ||
112 | #define V16QI_type_node (spu_builtin_types[SPU_BTI_V16QI]) | |
113 | #define V8HI_type_node (spu_builtin_types[SPU_BTI_V8HI]) | |
114 | #define V4SI_type_node (spu_builtin_types[SPU_BTI_V4SI]) | |
115 | #define V2DI_type_node (spu_builtin_types[SPU_BTI_V2DI]) | |
116 | #define V4SF_type_node (spu_builtin_types[SPU_BTI_V4SF]) | |
117 | #define V2DF_type_node (spu_builtin_types[SPU_BTI_V2DF]) | |
118 | #define unsigned_V16QI_type_node (spu_builtin_types[SPU_BTI_UV16QI]) | |
119 | #define unsigned_V8HI_type_node (spu_builtin_types[SPU_BTI_UV8HI]) | |
120 | #define unsigned_V4SI_type_node (spu_builtin_types[SPU_BTI_UV4SI]) | |
121 | #define unsigned_V2DI_type_node (spu_builtin_types[SPU_BTI_UV2DI]) | |
122 | ||
123 | static GTY(()) tree spu_builtin_types[SPU_BTI_MAX]; | |
124 | ||
6352eedf | 125 | struct spu_builtin_range |
126 | { | |
127 | int low, high; | |
128 | }; | |
129 | ||
130 | static struct spu_builtin_range spu_builtin_range[] = { | |
131 | {-0x40ll, 0x7fll}, /* SPU_BTI_7 */ | |
132 | {-0x40ll, 0x3fll}, /* SPU_BTI_S7 */ | |
133 | {0ll, 0x7fll}, /* SPU_BTI_U7 */ | |
134 | {-0x200ll, 0x1ffll}, /* SPU_BTI_S10 */ | |
135 | {-0x2000ll, 0x1fffll}, /* SPU_BTI_S10_4 */ | |
136 | {0ll, 0x3fffll}, /* SPU_BTI_U14 */ | |
137 | {-0x8000ll, 0xffffll}, /* SPU_BTI_16 */ | |
138 | {-0x8000ll, 0x7fffll}, /* SPU_BTI_S16 */ | |
139 | {-0x20000ll, 0x1ffffll}, /* SPU_BTI_S16_2 */ | |
140 | {0ll, 0xffffll}, /* SPU_BTI_U16 */ | |
141 | {0ll, 0x3ffffll}, /* SPU_BTI_U16_2 */ | |
142 | {0ll, 0x3ffffll}, /* SPU_BTI_U18 */ | |
143 | }; | |
144 | ||
644459d0 | 145 | \f |
146 | /* Target specific attribute specifications. */ | |
147 | char regs_ever_allocated[FIRST_PSEUDO_REGISTER]; | |
148 | ||
149 | /* Prototypes and external defs. */ | |
4c834714 | 150 | static void spu_option_override (void); |
cc07c468 | 151 | static void spu_option_init_struct (struct gcc_options *opts); |
686e2769 | 152 | static void spu_option_default_params (void); |
644459d0 | 153 | static void spu_init_builtins (void); |
e6925042 | 154 | static tree spu_builtin_decl (unsigned, bool); |
b62e30b8 | 155 | static bool spu_scalar_mode_supported_p (enum machine_mode mode); |
156 | static bool spu_vector_mode_supported_p (enum machine_mode mode); | |
fd50b071 | 157 | static bool spu_legitimate_address_p (enum machine_mode, rtx, bool); |
6cf5579e | 158 | static bool spu_addr_space_legitimate_address_p (enum machine_mode, rtx, |
159 | bool, addr_space_t); | |
644459d0 | 160 | static rtx adjust_operand (rtx op, HOST_WIDE_INT * start); |
161 | static rtx get_pic_reg (void); | |
162 | static int need_to_save_reg (int regno, int saving); | |
163 | static rtx frame_emit_store (int regno, rtx addr, HOST_WIDE_INT offset); | |
164 | static rtx frame_emit_load (int regno, rtx addr, HOST_WIDE_INT offset); | |
165 | static rtx frame_emit_add_imm (rtx dst, rtx src, HOST_WIDE_INT imm, | |
166 | rtx scratch); | |
167 | static void emit_nop_for_insn (rtx insn); | |
168 | static bool insn_clobbers_hbr (rtx insn); | |
169 | static void spu_emit_branch_hint (rtx before, rtx branch, rtx target, | |
5a976006 | 170 | int distance, sbitmap blocks); |
5474166e | 171 | static rtx spu_emit_vector_compare (enum rtx_code rcode, rtx op0, rtx op1, |
172 | enum machine_mode dmode); | |
644459d0 | 173 | static rtx get_branch_target (rtx branch); |
644459d0 | 174 | static void spu_machine_dependent_reorg (void); |
175 | static int spu_sched_issue_rate (void); | |
176 | static int spu_sched_variable_issue (FILE * dump, int verbose, rtx insn, | |
177 | int can_issue_more); | |
178 | static int get_pipe (rtx insn); | |
644459d0 | 179 | static int spu_sched_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost); |
5a976006 | 180 | static void spu_sched_init_global (FILE *, int, int); |
181 | static void spu_sched_init (FILE *, int, int); | |
182 | static int spu_sched_reorder (FILE *, int, rtx *, int *, int); | |
644459d0 | 183 | static tree spu_handle_fndecl_attribute (tree * node, tree name, tree args, |
184 | int flags, | |
b62e30b8 | 185 | bool *no_add_attrs); |
644459d0 | 186 | static tree spu_handle_vector_attribute (tree * node, tree name, tree args, |
187 | int flags, | |
b62e30b8 | 188 | bool *no_add_attrs); |
644459d0 | 189 | static int spu_naked_function_p (tree func); |
b62e30b8 | 190 | static bool spu_pass_by_reference (CUMULATIVE_ARGS *cum, enum machine_mode mode, |
191 | const_tree type, bool named); | |
ee9034d4 | 192 | static rtx spu_function_arg (CUMULATIVE_ARGS *cum, enum machine_mode mode, |
193 | const_tree type, bool named); | |
194 | static void spu_function_arg_advance (CUMULATIVE_ARGS *cum, enum machine_mode mode, | |
195 | const_tree type, bool named); | |
644459d0 | 196 | static tree spu_build_builtin_va_list (void); |
8a58ed0a | 197 | static void spu_va_start (tree, rtx); |
75a70cf9 | 198 | static tree spu_gimplify_va_arg_expr (tree valist, tree type, |
199 | gimple_seq * pre_p, gimple_seq * post_p); | |
644459d0 | 200 | static int store_with_one_insn_p (rtx mem); |
644459d0 | 201 | static int mem_is_padded_component_ref (rtx x); |
9d98604b | 202 | static int reg_aligned_for_addr (rtx x); |
644459d0 | 203 | static bool spu_assemble_integer (rtx x, unsigned int size, int aligned_p); |
204 | static void spu_asm_globalize_label (FILE * file, const char *name); | |
b62e30b8 | 205 | static bool spu_rtx_costs (rtx x, int code, int outer_code, |
206 | int *total, bool speed); | |
207 | static bool spu_function_ok_for_sibcall (tree decl, tree exp); | |
644459d0 | 208 | static void spu_init_libfuncs (void); |
fb80456a | 209 | static bool spu_return_in_memory (const_tree type, const_tree fntype); |
c7b91b14 | 210 | static void fix_range (const char *); |
69ced2d6 | 211 | static void spu_encode_section_info (tree, rtx, int); |
41e3a0c7 | 212 | static rtx spu_legitimize_address (rtx, rtx, enum machine_mode); |
6cf5579e | 213 | static rtx spu_addr_space_legitimize_address (rtx, rtx, enum machine_mode, |
214 | addr_space_t); | |
e99f512d | 215 | static tree spu_builtin_mul_widen_even (tree); |
216 | static tree spu_builtin_mul_widen_odd (tree); | |
a76866d3 | 217 | static tree spu_builtin_mask_for_load (void); |
0822b158 | 218 | static int spu_builtin_vectorization_cost (enum vect_cost_for_stmt, tree, int); |
a9f1838b | 219 | static bool spu_vector_alignment_reachable (const_tree, bool); |
a0515226 | 220 | static tree spu_builtin_vec_perm (tree, tree *); |
6cf5579e | 221 | static enum machine_mode spu_addr_space_pointer_mode (addr_space_t); |
222 | static enum machine_mode spu_addr_space_address_mode (addr_space_t); | |
223 | static bool spu_addr_space_subset_p (addr_space_t, addr_space_t); | |
224 | static rtx spu_addr_space_convert (rtx, tree, tree); | |
d52fd16a | 225 | static int spu_sms_res_mii (struct ddg *g); |
5a976006 | 226 | static void asm_file_start (void); |
a08dfd55 | 227 | static unsigned int spu_section_type_flags (tree, const char *, int); |
6cf5579e | 228 | static section *spu_select_section (tree, int, unsigned HOST_WIDE_INT); |
229 | static void spu_unique_section (tree, int); | |
9d98604b | 230 | static rtx spu_expand_load (rtx, rtx, rtx, int); |
e96f2783 | 231 | static void spu_trampoline_init (rtx, tree, rtx); |
b2d7ede1 | 232 | static void spu_conditional_register_usage (void); |
329c1e4e | 233 | static bool spu_ref_may_alias_errno (ao_ref *); |
644459d0 | 234 | |
5474166e | 235 | /* Which instruction set architecture to use. */ |
236 | int spu_arch; | |
237 | /* Which cpu are we tuning for. */ | |
238 | int spu_tune; | |
239 | ||
5a976006 | 240 | /* The hardware requires 8 insns between a hint and the branch it |
241 | effects. This variable describes how many rtl instructions the | |
242 | compiler needs to see before inserting a hint, and then the compiler | |
243 | will insert enough nops to make it at least 8 insns. The default is | |
244 | for the compiler to allow up to 2 nops be emitted. The nops are | |
245 | inserted in pairs, so we round down. */ | |
246 | int spu_hint_dist = (8*4) - (2*4); | |
247 | ||
248 | /* Determines whether we run variable tracking in machine dependent | |
249 | reorganization. */ | |
250 | static int spu_flag_var_tracking; | |
251 | ||
644459d0 | 252 | enum spu_immediate { |
253 | SPU_NONE, | |
254 | SPU_IL, | |
255 | SPU_ILA, | |
256 | SPU_ILH, | |
257 | SPU_ILHU, | |
258 | SPU_ORI, | |
259 | SPU_ORHI, | |
260 | SPU_ORBI, | |
99369027 | 261 | SPU_IOHL |
644459d0 | 262 | }; |
dea01258 | 263 | enum immediate_class |
264 | { | |
265 | IC_POOL, /* constant pool */ | |
266 | IC_IL1, /* one il* instruction */ | |
267 | IC_IL2, /* both ilhu and iohl instructions */ | |
268 | IC_IL1s, /* one il* instruction */ | |
269 | IC_IL2s, /* both ilhu and iohl instructions */ | |
270 | IC_FSMBI, /* the fsmbi instruction */ | |
271 | IC_CPAT, /* one of the c*d instructions */ | |
5df189be | 272 | IC_FSMBI2 /* fsmbi plus 1 other instruction */ |
dea01258 | 273 | }; |
644459d0 | 274 | |
275 | static enum spu_immediate which_immediate_load (HOST_WIDE_INT val); | |
276 | static enum spu_immediate which_logical_immediate (HOST_WIDE_INT val); | |
dea01258 | 277 | static int cpat_info(unsigned char *arr, int size, int *prun, int *pstart); |
278 | static enum immediate_class classify_immediate (rtx op, | |
279 | enum machine_mode mode); | |
644459d0 | 280 | |
1bd43494 | 281 | static enum machine_mode spu_unwind_word_mode (void); |
282 | ||
ea32e033 | 283 | static enum machine_mode |
284 | spu_libgcc_cmp_return_mode (void); | |
285 | ||
286 | static enum machine_mode | |
287 | spu_libgcc_shift_count_mode (void); | |
6cf5579e | 288 | |
289 | /* Pointer mode for __ea references. */ | |
290 | #define EAmode (spu_ea_model != 32 ? DImode : SImode) | |
291 | ||
ef51d1e3 | 292 | \f |
293 | /* Table of machine attributes. */ | |
294 | static const struct attribute_spec spu_attribute_table[] = | |
295 | { | |
ac86af5d | 296 | /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler, |
297 | affects_type_identity } */ | |
298 | { "naked", 0, 0, true, false, false, spu_handle_fndecl_attribute, | |
299 | false }, | |
300 | { "spu_vector", 0, 0, false, true, false, spu_handle_vector_attribute, | |
301 | false }, | |
302 | { NULL, 0, 0, false, false, false, NULL, false } | |
ef51d1e3 | 303 | }; |
644459d0 | 304 | \f |
305 | /* TARGET overrides. */ | |
306 | ||
6cf5579e | 307 | #undef TARGET_ADDR_SPACE_POINTER_MODE |
308 | #define TARGET_ADDR_SPACE_POINTER_MODE spu_addr_space_pointer_mode | |
309 | ||
310 | #undef TARGET_ADDR_SPACE_ADDRESS_MODE | |
311 | #define TARGET_ADDR_SPACE_ADDRESS_MODE spu_addr_space_address_mode | |
312 | ||
313 | #undef TARGET_ADDR_SPACE_LEGITIMATE_ADDRESS_P | |
314 | #define TARGET_ADDR_SPACE_LEGITIMATE_ADDRESS_P \ | |
315 | spu_addr_space_legitimate_address_p | |
316 | ||
317 | #undef TARGET_ADDR_SPACE_LEGITIMIZE_ADDRESS | |
318 | #define TARGET_ADDR_SPACE_LEGITIMIZE_ADDRESS spu_addr_space_legitimize_address | |
319 | ||
320 | #undef TARGET_ADDR_SPACE_SUBSET_P | |
321 | #define TARGET_ADDR_SPACE_SUBSET_P spu_addr_space_subset_p | |
322 | ||
323 | #undef TARGET_ADDR_SPACE_CONVERT | |
324 | #define TARGET_ADDR_SPACE_CONVERT spu_addr_space_convert | |
325 | ||
644459d0 | 326 | #undef TARGET_INIT_BUILTINS |
327 | #define TARGET_INIT_BUILTINS spu_init_builtins | |
e6925042 | 328 | #undef TARGET_BUILTIN_DECL |
329 | #define TARGET_BUILTIN_DECL spu_builtin_decl | |
644459d0 | 330 | |
644459d0 | 331 | #undef TARGET_EXPAND_BUILTIN |
332 | #define TARGET_EXPAND_BUILTIN spu_expand_builtin | |
333 | ||
1bd43494 | 334 | #undef TARGET_UNWIND_WORD_MODE |
335 | #define TARGET_UNWIND_WORD_MODE spu_unwind_word_mode | |
644459d0 | 336 | |
41e3a0c7 | 337 | #undef TARGET_LEGITIMIZE_ADDRESS |
338 | #define TARGET_LEGITIMIZE_ADDRESS spu_legitimize_address | |
339 | ||
6cf5579e | 340 | /* The current assembler doesn't like .4byte foo@ppu, so use the normal .long |
341 | and .quad for the debugger. When it is known that the assembler is fixed, | |
342 | these can be removed. */ | |
343 | #undef TARGET_ASM_UNALIGNED_SI_OP | |
344 | #define TARGET_ASM_UNALIGNED_SI_OP "\t.long\t" | |
345 | ||
346 | #undef TARGET_ASM_ALIGNED_DI_OP | |
347 | #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t" | |
348 | ||
644459d0 | 349 | /* The .8byte directive doesn't seem to work well for a 32 bit |
350 | architecture. */ | |
351 | #undef TARGET_ASM_UNALIGNED_DI_OP | |
352 | #define TARGET_ASM_UNALIGNED_DI_OP NULL | |
353 | ||
354 | #undef TARGET_RTX_COSTS | |
355 | #define TARGET_RTX_COSTS spu_rtx_costs | |
356 | ||
357 | #undef TARGET_ADDRESS_COST | |
f529eb25 | 358 | #define TARGET_ADDRESS_COST hook_int_rtx_bool_0 |
644459d0 | 359 | |
360 | #undef TARGET_SCHED_ISSUE_RATE | |
361 | #define TARGET_SCHED_ISSUE_RATE spu_sched_issue_rate | |
362 | ||
5a976006 | 363 | #undef TARGET_SCHED_INIT_GLOBAL |
364 | #define TARGET_SCHED_INIT_GLOBAL spu_sched_init_global | |
365 | ||
366 | #undef TARGET_SCHED_INIT | |
367 | #define TARGET_SCHED_INIT spu_sched_init | |
368 | ||
644459d0 | 369 | #undef TARGET_SCHED_VARIABLE_ISSUE |
370 | #define TARGET_SCHED_VARIABLE_ISSUE spu_sched_variable_issue | |
371 | ||
5a976006 | 372 | #undef TARGET_SCHED_REORDER |
373 | #define TARGET_SCHED_REORDER spu_sched_reorder | |
374 | ||
375 | #undef TARGET_SCHED_REORDER2 | |
376 | #define TARGET_SCHED_REORDER2 spu_sched_reorder | |
644459d0 | 377 | |
378 | #undef TARGET_SCHED_ADJUST_COST | |
379 | #define TARGET_SCHED_ADJUST_COST spu_sched_adjust_cost | |
380 | ||
644459d0 | 381 | #undef TARGET_ATTRIBUTE_TABLE |
382 | #define TARGET_ATTRIBUTE_TABLE spu_attribute_table | |
383 | ||
384 | #undef TARGET_ASM_INTEGER | |
385 | #define TARGET_ASM_INTEGER spu_assemble_integer | |
386 | ||
387 | #undef TARGET_SCALAR_MODE_SUPPORTED_P | |
388 | #define TARGET_SCALAR_MODE_SUPPORTED_P spu_scalar_mode_supported_p | |
389 | ||
390 | #undef TARGET_VECTOR_MODE_SUPPORTED_P | |
391 | #define TARGET_VECTOR_MODE_SUPPORTED_P spu_vector_mode_supported_p | |
392 | ||
393 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL | |
394 | #define TARGET_FUNCTION_OK_FOR_SIBCALL spu_function_ok_for_sibcall | |
395 | ||
396 | #undef TARGET_ASM_GLOBALIZE_LABEL | |
397 | #define TARGET_ASM_GLOBALIZE_LABEL spu_asm_globalize_label | |
398 | ||
399 | #undef TARGET_PASS_BY_REFERENCE | |
400 | #define TARGET_PASS_BY_REFERENCE spu_pass_by_reference | |
401 | ||
ee9034d4 | 402 | #undef TARGET_FUNCTION_ARG |
403 | #define TARGET_FUNCTION_ARG spu_function_arg | |
404 | ||
405 | #undef TARGET_FUNCTION_ARG_ADVANCE | |
406 | #define TARGET_FUNCTION_ARG_ADVANCE spu_function_arg_advance | |
407 | ||
644459d0 | 408 | #undef TARGET_MUST_PASS_IN_STACK |
409 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
410 | ||
411 | #undef TARGET_BUILD_BUILTIN_VA_LIST | |
412 | #define TARGET_BUILD_BUILTIN_VA_LIST spu_build_builtin_va_list | |
413 | ||
8a58ed0a | 414 | #undef TARGET_EXPAND_BUILTIN_VA_START |
415 | #define TARGET_EXPAND_BUILTIN_VA_START spu_va_start | |
416 | ||
644459d0 | 417 | #undef TARGET_SETUP_INCOMING_VARARGS |
418 | #define TARGET_SETUP_INCOMING_VARARGS spu_setup_incoming_varargs | |
419 | ||
420 | #undef TARGET_MACHINE_DEPENDENT_REORG | |
421 | #define TARGET_MACHINE_DEPENDENT_REORG spu_machine_dependent_reorg | |
422 | ||
423 | #undef TARGET_GIMPLIFY_VA_ARG_EXPR | |
424 | #define TARGET_GIMPLIFY_VA_ARG_EXPR spu_gimplify_va_arg_expr | |
425 | ||
426 | #undef TARGET_DEFAULT_TARGET_FLAGS | |
427 | #define TARGET_DEFAULT_TARGET_FLAGS (TARGET_DEFAULT) | |
428 | ||
429 | #undef TARGET_INIT_LIBFUNCS | |
430 | #define TARGET_INIT_LIBFUNCS spu_init_libfuncs | |
431 | ||
432 | #undef TARGET_RETURN_IN_MEMORY | |
433 | #define TARGET_RETURN_IN_MEMORY spu_return_in_memory | |
434 | ||
69ced2d6 | 435 | #undef TARGET_ENCODE_SECTION_INFO |
436 | #define TARGET_ENCODE_SECTION_INFO spu_encode_section_info | |
437 | ||
e99f512d | 438 | #undef TARGET_VECTORIZE_BUILTIN_MUL_WIDEN_EVEN |
439 | #define TARGET_VECTORIZE_BUILTIN_MUL_WIDEN_EVEN spu_builtin_mul_widen_even | |
440 | ||
441 | #undef TARGET_VECTORIZE_BUILTIN_MUL_WIDEN_ODD | |
442 | #define TARGET_VECTORIZE_BUILTIN_MUL_WIDEN_ODD spu_builtin_mul_widen_odd | |
443 | ||
a76866d3 | 444 | #undef TARGET_VECTORIZE_BUILTIN_MASK_FOR_LOAD |
445 | #define TARGET_VECTORIZE_BUILTIN_MASK_FOR_LOAD spu_builtin_mask_for_load | |
446 | ||
a28df51d | 447 | #undef TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST |
448 | #define TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST spu_builtin_vectorization_cost | |
449 | ||
202d6e5f | 450 | #undef TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE |
451 | #define TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE spu_vector_alignment_reachable | |
0e87db76 | 452 | |
a0515226 | 453 | #undef TARGET_VECTORIZE_BUILTIN_VEC_PERM |
454 | #define TARGET_VECTORIZE_BUILTIN_VEC_PERM spu_builtin_vec_perm | |
455 | ||
ea32e033 | 456 | #undef TARGET_LIBGCC_CMP_RETURN_MODE |
457 | #define TARGET_LIBGCC_CMP_RETURN_MODE spu_libgcc_cmp_return_mode | |
458 | ||
459 | #undef TARGET_LIBGCC_SHIFT_COUNT_MODE | |
460 | #define TARGET_LIBGCC_SHIFT_COUNT_MODE spu_libgcc_shift_count_mode | |
461 | ||
d52fd16a | 462 | #undef TARGET_SCHED_SMS_RES_MII |
463 | #define TARGET_SCHED_SMS_RES_MII spu_sms_res_mii | |
464 | ||
5a976006 | 465 | #undef TARGET_ASM_FILE_START |
466 | #define TARGET_ASM_FILE_START asm_file_start | |
467 | ||
a08dfd55 | 468 | #undef TARGET_SECTION_TYPE_FLAGS |
469 | #define TARGET_SECTION_TYPE_FLAGS spu_section_type_flags | |
470 | ||
6cf5579e | 471 | #undef TARGET_ASM_SELECT_SECTION |
472 | #define TARGET_ASM_SELECT_SECTION spu_select_section | |
473 | ||
474 | #undef TARGET_ASM_UNIQUE_SECTION | |
475 | #define TARGET_ASM_UNIQUE_SECTION spu_unique_section | |
476 | ||
fd50b071 | 477 | #undef TARGET_LEGITIMATE_ADDRESS_P |
478 | #define TARGET_LEGITIMATE_ADDRESS_P spu_legitimate_address_p | |
479 | ||
e96f2783 | 480 | #undef TARGET_TRAMPOLINE_INIT |
481 | #define TARGET_TRAMPOLINE_INIT spu_trampoline_init | |
482 | ||
4c834714 | 483 | #undef TARGET_OPTION_OVERRIDE |
484 | #define TARGET_OPTION_OVERRIDE spu_option_override | |
485 | ||
cc07c468 | 486 | #undef TARGET_OPTION_INIT_STRUCT |
487 | #define TARGET_OPTION_INIT_STRUCT spu_option_init_struct | |
02e53c17 | 488 | |
686e2769 | 489 | #undef TARGET_OPTION_DEFAULT_PARAMS |
490 | #define TARGET_OPTION_DEFAULT_PARAMS spu_option_default_params | |
491 | ||
ed6befa5 | 492 | #undef TARGET_EXCEPT_UNWIND_INFO |
493 | #define TARGET_EXCEPT_UNWIND_INFO sjlj_except_unwind_info | |
494 | ||
b2d7ede1 | 495 | #undef TARGET_CONDITIONAL_REGISTER_USAGE |
496 | #define TARGET_CONDITIONAL_REGISTER_USAGE spu_conditional_register_usage | |
497 | ||
329c1e4e | 498 | #undef TARGET_REF_MAY_ALIAS_ERRNO |
499 | #define TARGET_REF_MAY_ALIAS_ERRNO spu_ref_may_alias_errno | |
500 | ||
644459d0 | 501 | struct gcc_target targetm = TARGET_INITIALIZER; |
502 | ||
02e53c17 | 503 | static void |
cc07c468 | 504 | spu_option_init_struct (struct gcc_options *opts) |
5df189be | 505 | { |
5df189be | 506 | /* With so many registers this is better on by default. */ |
cc07c468 | 507 | opts->x_flag_rename_registers = 1; |
5df189be | 508 | } |
509 | ||
686e2769 | 510 | /* Implement TARGET_OPTION_DEFAULT_PARAMS. */ |
511 | static void | |
512 | spu_option_default_params (void) | |
513 | { | |
514 | /* Override some of the default param values. With so many registers | |
515 | larger values are better for these params. */ | |
516 | set_default_param_value (PARAM_MAX_PENDING_LIST_LENGTH, 128); | |
517 | } | |
518 | ||
4c834714 | 519 | /* Implement TARGET_OPTION_OVERRIDE. */ |
520 | static void | |
521 | spu_option_override (void) | |
644459d0 | 522 | { |
14d408d9 | 523 | /* Small loops will be unpeeled at -O3. For SPU it is more important |
524 | to keep code small by default. */ | |
686e2769 | 525 | if (!flag_unroll_loops && !flag_peel_loops) |
e0b840fc | 526 | maybe_set_param_value (PARAM_MAX_COMPLETELY_PEEL_TIMES, 4, |
56f280c4 | 527 | global_options.x_param_values, |
528 | global_options_set.x_param_values); | |
14d408d9 | 529 | |
644459d0 | 530 | flag_omit_frame_pointer = 1; |
531 | ||
5a976006 | 532 | /* Functions must be 8 byte aligned so we correctly handle dual issue */ |
644459d0 | 533 | if (align_functions < 8) |
534 | align_functions = 8; | |
c7b91b14 | 535 | |
5a976006 | 536 | spu_hint_dist = 8*4 - spu_max_nops*4; |
537 | if (spu_hint_dist < 0) | |
538 | spu_hint_dist = 0; | |
539 | ||
c7b91b14 | 540 | if (spu_fixed_range_string) |
541 | fix_range (spu_fixed_range_string); | |
5474166e | 542 | |
543 | /* Determine processor architectural level. */ | |
544 | if (spu_arch_string) | |
545 | { | |
546 | if (strcmp (&spu_arch_string[0], "cell") == 0) | |
547 | spu_arch = PROCESSOR_CELL; | |
548 | else if (strcmp (&spu_arch_string[0], "celledp") == 0) | |
549 | spu_arch = PROCESSOR_CELLEDP; | |
550 | else | |
8e181c9d | 551 | error ("bad value (%s) for -march= switch", spu_arch_string); |
5474166e | 552 | } |
553 | ||
554 | /* Determine processor to tune for. */ | |
555 | if (spu_tune_string) | |
556 | { | |
557 | if (strcmp (&spu_tune_string[0], "cell") == 0) | |
558 | spu_tune = PROCESSOR_CELL; | |
559 | else if (strcmp (&spu_tune_string[0], "celledp") == 0) | |
560 | spu_tune = PROCESSOR_CELLEDP; | |
561 | else | |
8e181c9d | 562 | error ("bad value (%s) for -mtune= switch", spu_tune_string); |
5474166e | 563 | } |
98bbec1e | 564 | |
13684256 | 565 | /* Change defaults according to the processor architecture. */ |
566 | if (spu_arch == PROCESSOR_CELLEDP) | |
567 | { | |
568 | /* If no command line option has been otherwise specified, change | |
569 | the default to -mno-safe-hints on celledp -- only the original | |
570 | Cell/B.E. processors require this workaround. */ | |
571 | if (!(target_flags_explicit & MASK_SAFE_HINTS)) | |
572 | target_flags &= ~MASK_SAFE_HINTS; | |
573 | } | |
574 | ||
98bbec1e | 575 | REAL_MODE_FORMAT (SFmode) = &spu_single_format; |
644459d0 | 576 | } |
577 | \f | |
578 | /* Handle an attribute requiring a FUNCTION_DECL; arguments as in | |
579 | struct attribute_spec.handler. */ | |
580 | ||
644459d0 | 581 | /* True if MODE is valid for the target. By "valid", we mean able to |
582 | be manipulated in non-trivial ways. In particular, this means all | |
583 | the arithmetic is supported. */ | |
584 | static bool | |
585 | spu_scalar_mode_supported_p (enum machine_mode mode) | |
586 | { | |
587 | switch (mode) | |
588 | { | |
589 | case QImode: | |
590 | case HImode: | |
591 | case SImode: | |
592 | case SFmode: | |
593 | case DImode: | |
594 | case TImode: | |
595 | case DFmode: | |
596 | return true; | |
597 | ||
598 | default: | |
599 | return false; | |
600 | } | |
601 | } | |
602 | ||
603 | /* Similarly for vector modes. "Supported" here is less strict. At | |
604 | least some operations are supported; need to check optabs or builtins | |
605 | for further details. */ | |
606 | static bool | |
607 | spu_vector_mode_supported_p (enum machine_mode mode) | |
608 | { | |
609 | switch (mode) | |
610 | { | |
611 | case V16QImode: | |
612 | case V8HImode: | |
613 | case V4SImode: | |
614 | case V2DImode: | |
615 | case V4SFmode: | |
616 | case V2DFmode: | |
617 | return true; | |
618 | ||
619 | default: | |
620 | return false; | |
621 | } | |
622 | } | |
623 | ||
624 | /* GCC assumes that in a paradoxical SUBREG the inner mode occupies the | |
625 | least significant bytes of the outer mode. This function returns | |
626 | TRUE for the SUBREG's where this is correct. */ | |
627 | int | |
628 | valid_subreg (rtx op) | |
629 | { | |
630 | enum machine_mode om = GET_MODE (op); | |
631 | enum machine_mode im = GET_MODE (SUBREG_REG (op)); | |
632 | return om != VOIDmode && im != VOIDmode | |
633 | && (GET_MODE_SIZE (im) == GET_MODE_SIZE (om) | |
38aca5eb | 634 | || (GET_MODE_SIZE (im) <= 4 && GET_MODE_SIZE (om) <= 4) |
635 | || (GET_MODE_SIZE (im) >= 16 && GET_MODE_SIZE (om) >= 16)); | |
644459d0 | 636 | } |
637 | ||
638 | /* When insv and ext[sz]v ar passed a TI SUBREG, we want to strip it off | |
fa7637bd | 639 | and adjust the start offset. */ |
644459d0 | 640 | static rtx |
641 | adjust_operand (rtx op, HOST_WIDE_INT * start) | |
642 | { | |
643 | enum machine_mode mode; | |
644 | int op_size; | |
38aca5eb | 645 | /* Strip any paradoxical SUBREG. */ |
646 | if (GET_CODE (op) == SUBREG | |
647 | && (GET_MODE_BITSIZE (GET_MODE (op)) | |
648 | > GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op))))) | |
644459d0 | 649 | { |
650 | if (start) | |
651 | *start -= | |
652 | GET_MODE_BITSIZE (GET_MODE (op)) - | |
653 | GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op))); | |
654 | op = SUBREG_REG (op); | |
655 | } | |
656 | /* If it is smaller than SI, assure a SUBREG */ | |
657 | op_size = GET_MODE_BITSIZE (GET_MODE (op)); | |
658 | if (op_size < 32) | |
659 | { | |
660 | if (start) | |
661 | *start += 32 - op_size; | |
662 | op_size = 32; | |
663 | } | |
664 | /* If it is not a MODE_INT (and/or it is smaller than SI) add a SUBREG. */ | |
665 | mode = mode_for_size (op_size, MODE_INT, 0); | |
666 | if (mode != GET_MODE (op)) | |
667 | op = gen_rtx_SUBREG (mode, op, 0); | |
668 | return op; | |
669 | } | |
670 | ||
671 | void | |
672 | spu_expand_extv (rtx ops[], int unsignedp) | |
673 | { | |
9d98604b | 674 | rtx dst = ops[0], src = ops[1]; |
644459d0 | 675 | HOST_WIDE_INT width = INTVAL (ops[2]); |
676 | HOST_WIDE_INT start = INTVAL (ops[3]); | |
9d98604b | 677 | HOST_WIDE_INT align_mask; |
678 | rtx s0, s1, mask, r0; | |
644459d0 | 679 | |
9d98604b | 680 | gcc_assert (REG_P (dst) && GET_MODE (dst) == TImode); |
644459d0 | 681 | |
9d98604b | 682 | if (MEM_P (src)) |
644459d0 | 683 | { |
9d98604b | 684 | /* First, determine if we need 1 TImode load or 2. We need only 1 |
685 | if the bits being extracted do not cross the alignment boundary | |
686 | as determined by the MEM and its address. */ | |
687 | ||
688 | align_mask = -MEM_ALIGN (src); | |
689 | if ((start & align_mask) == ((start + width - 1) & align_mask)) | |
644459d0 | 690 | { |
9d98604b | 691 | /* Alignment is sufficient for 1 load. */ |
692 | s0 = gen_reg_rtx (TImode); | |
693 | r0 = spu_expand_load (s0, 0, src, start / 8); | |
694 | start &= 7; | |
695 | if (r0) | |
696 | emit_insn (gen_rotqby_ti (s0, s0, r0)); | |
644459d0 | 697 | } |
9d98604b | 698 | else |
699 | { | |
700 | /* Need 2 loads. */ | |
701 | s0 = gen_reg_rtx (TImode); | |
702 | s1 = gen_reg_rtx (TImode); | |
703 | r0 = spu_expand_load (s0, s1, src, start / 8); | |
704 | start &= 7; | |
705 | ||
706 | gcc_assert (start + width <= 128); | |
707 | if (r0) | |
708 | { | |
709 | rtx r1 = gen_reg_rtx (SImode); | |
710 | mask = gen_reg_rtx (TImode); | |
711 | emit_move_insn (mask, GEN_INT (-1)); | |
712 | emit_insn (gen_rotqby_ti (s0, s0, r0)); | |
713 | emit_insn (gen_rotqby_ti (s1, s1, r0)); | |
714 | if (GET_CODE (r0) == CONST_INT) | |
715 | r1 = GEN_INT (INTVAL (r0) & 15); | |
716 | else | |
717 | emit_insn (gen_andsi3 (r1, r0, GEN_INT (15))); | |
718 | emit_insn (gen_shlqby_ti (mask, mask, r1)); | |
719 | emit_insn (gen_selb (s0, s1, s0, mask)); | |
720 | } | |
721 | } | |
722 | ||
723 | } | |
724 | else if (GET_CODE (src) == SUBREG) | |
725 | { | |
726 | rtx r = SUBREG_REG (src); | |
727 | gcc_assert (REG_P (r) && SCALAR_INT_MODE_P (GET_MODE (r))); | |
728 | s0 = gen_reg_rtx (TImode); | |
729 | if (GET_MODE_SIZE (GET_MODE (r)) < GET_MODE_SIZE (TImode)) | |
730 | emit_insn (gen_rtx_SET (VOIDmode, s0, gen_rtx_ZERO_EXTEND (TImode, r))); | |
731 | else | |
732 | emit_move_insn (s0, src); | |
733 | } | |
734 | else | |
735 | { | |
736 | gcc_assert (REG_P (src) && GET_MODE (src) == TImode); | |
737 | s0 = gen_reg_rtx (TImode); | |
738 | emit_move_insn (s0, src); | |
644459d0 | 739 | } |
740 | ||
9d98604b | 741 | /* Now s0 is TImode and contains the bits to extract at start. */ |
742 | ||
743 | if (start) | |
744 | emit_insn (gen_rotlti3 (s0, s0, GEN_INT (start))); | |
745 | ||
746 | if (128 - width) | |
644459d0 | 747 | { |
9d98604b | 748 | tree c = build_int_cst (NULL_TREE, 128 - width); |
749 | s0 = expand_shift (RSHIFT_EXPR, TImode, s0, c, s0, unsignedp); | |
644459d0 | 750 | } |
751 | ||
9d98604b | 752 | emit_move_insn (dst, s0); |
644459d0 | 753 | } |
754 | ||
755 | void | |
756 | spu_expand_insv (rtx ops[]) | |
757 | { | |
758 | HOST_WIDE_INT width = INTVAL (ops[1]); | |
759 | HOST_WIDE_INT start = INTVAL (ops[2]); | |
760 | HOST_WIDE_INT maskbits; | |
4cbad5bb | 761 | enum machine_mode dst_mode; |
644459d0 | 762 | rtx dst = ops[0], src = ops[3]; |
4cbad5bb | 763 | int dst_size; |
644459d0 | 764 | rtx mask; |
765 | rtx shift_reg; | |
766 | int shift; | |
767 | ||
768 | ||
769 | if (GET_CODE (ops[0]) == MEM) | |
770 | dst = gen_reg_rtx (TImode); | |
771 | else | |
772 | dst = adjust_operand (dst, &start); | |
773 | dst_mode = GET_MODE (dst); | |
774 | dst_size = GET_MODE_BITSIZE (GET_MODE (dst)); | |
775 | ||
776 | if (CONSTANT_P (src)) | |
777 | { | |
778 | enum machine_mode m = | |
779 | (width <= 32 ? SImode : width <= 64 ? DImode : TImode); | |
780 | src = force_reg (m, convert_to_mode (m, src, 0)); | |
781 | } | |
782 | src = adjust_operand (src, 0); | |
644459d0 | 783 | |
784 | mask = gen_reg_rtx (dst_mode); | |
785 | shift_reg = gen_reg_rtx (dst_mode); | |
786 | shift = dst_size - start - width; | |
787 | ||
788 | /* It's not safe to use subreg here because the compiler assumes | |
789 | that the SUBREG_REG is right justified in the SUBREG. */ | |
790 | convert_move (shift_reg, src, 1); | |
791 | ||
792 | if (shift > 0) | |
793 | { | |
794 | switch (dst_mode) | |
795 | { | |
796 | case SImode: | |
797 | emit_insn (gen_ashlsi3 (shift_reg, shift_reg, GEN_INT (shift))); | |
798 | break; | |
799 | case DImode: | |
800 | emit_insn (gen_ashldi3 (shift_reg, shift_reg, GEN_INT (shift))); | |
801 | break; | |
802 | case TImode: | |
803 | emit_insn (gen_ashlti3 (shift_reg, shift_reg, GEN_INT (shift))); | |
804 | break; | |
805 | default: | |
806 | abort (); | |
807 | } | |
808 | } | |
809 | else if (shift < 0) | |
810 | abort (); | |
811 | ||
812 | switch (dst_size) | |
813 | { | |
814 | case 32: | |
815 | maskbits = (-1ll << (32 - width - start)); | |
816 | if (start) | |
817 | maskbits += (1ll << (32 - start)); | |
818 | emit_move_insn (mask, GEN_INT (maskbits)); | |
819 | break; | |
820 | case 64: | |
821 | maskbits = (-1ll << (64 - width - start)); | |
822 | if (start) | |
823 | maskbits += (1ll << (64 - start)); | |
824 | emit_move_insn (mask, GEN_INT (maskbits)); | |
825 | break; | |
826 | case 128: | |
827 | { | |
828 | unsigned char arr[16]; | |
829 | int i = start / 8; | |
830 | memset (arr, 0, sizeof (arr)); | |
831 | arr[i] = 0xff >> (start & 7); | |
832 | for (i++; i <= (start + width - 1) / 8; i++) | |
833 | arr[i] = 0xff; | |
834 | arr[i - 1] &= 0xff << (7 - ((start + width - 1) & 7)); | |
835 | emit_move_insn (mask, array_to_constant (TImode, arr)); | |
836 | } | |
837 | break; | |
838 | default: | |
839 | abort (); | |
840 | } | |
841 | if (GET_CODE (ops[0]) == MEM) | |
842 | { | |
644459d0 | 843 | rtx low = gen_reg_rtx (SImode); |
644459d0 | 844 | rtx rotl = gen_reg_rtx (SImode); |
845 | rtx mask0 = gen_reg_rtx (TImode); | |
9d98604b | 846 | rtx addr; |
847 | rtx addr0; | |
848 | rtx addr1; | |
644459d0 | 849 | rtx mem; |
850 | ||
9d98604b | 851 | addr = force_reg (Pmode, XEXP (ops[0], 0)); |
852 | addr0 = gen_rtx_AND (Pmode, addr, GEN_INT (-16)); | |
644459d0 | 853 | emit_insn (gen_andsi3 (low, addr, GEN_INT (15))); |
854 | emit_insn (gen_negsi2 (rotl, low)); | |
855 | emit_insn (gen_rotqby_ti (shift_reg, shift_reg, rotl)); | |
856 | emit_insn (gen_rotqmby_ti (mask0, mask, rotl)); | |
9d98604b | 857 | mem = change_address (ops[0], TImode, addr0); |
644459d0 | 858 | set_mem_alias_set (mem, 0); |
859 | emit_move_insn (dst, mem); | |
860 | emit_insn (gen_selb (dst, dst, shift_reg, mask0)); | |
644459d0 | 861 | if (start + width > MEM_ALIGN (ops[0])) |
862 | { | |
863 | rtx shl = gen_reg_rtx (SImode); | |
864 | rtx mask1 = gen_reg_rtx (TImode); | |
865 | rtx dst1 = gen_reg_rtx (TImode); | |
866 | rtx mem1; | |
9d98604b | 867 | addr1 = plus_constant (addr, 16); |
868 | addr1 = gen_rtx_AND (Pmode, addr1, GEN_INT (-16)); | |
644459d0 | 869 | emit_insn (gen_subsi3 (shl, GEN_INT (16), low)); |
870 | emit_insn (gen_shlqby_ti (mask1, mask, shl)); | |
9d98604b | 871 | mem1 = change_address (ops[0], TImode, addr1); |
644459d0 | 872 | set_mem_alias_set (mem1, 0); |
873 | emit_move_insn (dst1, mem1); | |
874 | emit_insn (gen_selb (dst1, dst1, shift_reg, mask1)); | |
875 | emit_move_insn (mem1, dst1); | |
876 | } | |
9d98604b | 877 | emit_move_insn (mem, dst); |
644459d0 | 878 | } |
879 | else | |
71cd778d | 880 | emit_insn (gen_selb (dst, copy_rtx (dst), shift_reg, mask)); |
644459d0 | 881 | } |
882 | ||
883 | ||
884 | int | |
885 | spu_expand_block_move (rtx ops[]) | |
886 | { | |
887 | HOST_WIDE_INT bytes, align, offset; | |
888 | rtx src, dst, sreg, dreg, target; | |
889 | int i; | |
890 | if (GET_CODE (ops[2]) != CONST_INT | |
891 | || GET_CODE (ops[3]) != CONST_INT | |
48eb4342 | 892 | || INTVAL (ops[2]) > (HOST_WIDE_INT) (MOVE_RATIO (optimize_insn_for_speed_p ()) * 8)) |
644459d0 | 893 | return 0; |
894 | ||
895 | bytes = INTVAL (ops[2]); | |
896 | align = INTVAL (ops[3]); | |
897 | ||
898 | if (bytes <= 0) | |
899 | return 1; | |
900 | ||
901 | dst = ops[0]; | |
902 | src = ops[1]; | |
903 | ||
904 | if (align == 16) | |
905 | { | |
906 | for (offset = 0; offset + 16 <= bytes; offset += 16) | |
907 | { | |
908 | dst = adjust_address (ops[0], V16QImode, offset); | |
909 | src = adjust_address (ops[1], V16QImode, offset); | |
910 | emit_move_insn (dst, src); | |
911 | } | |
912 | if (offset < bytes) | |
913 | { | |
914 | rtx mask; | |
915 | unsigned char arr[16] = { 0 }; | |
916 | for (i = 0; i < bytes - offset; i++) | |
917 | arr[i] = 0xff; | |
918 | dst = adjust_address (ops[0], V16QImode, offset); | |
919 | src = adjust_address (ops[1], V16QImode, offset); | |
920 | mask = gen_reg_rtx (V16QImode); | |
921 | sreg = gen_reg_rtx (V16QImode); | |
922 | dreg = gen_reg_rtx (V16QImode); | |
923 | target = gen_reg_rtx (V16QImode); | |
924 | emit_move_insn (mask, array_to_constant (V16QImode, arr)); | |
925 | emit_move_insn (dreg, dst); | |
926 | emit_move_insn (sreg, src); | |
927 | emit_insn (gen_selb (target, dreg, sreg, mask)); | |
928 | emit_move_insn (dst, target); | |
929 | } | |
930 | return 1; | |
931 | } | |
932 | return 0; | |
933 | } | |
934 | ||
935 | enum spu_comp_code | |
936 | { SPU_EQ, SPU_GT, SPU_GTU }; | |
937 | ||
5474166e | 938 | int spu_comp_icode[12][3] = { |
939 | {CODE_FOR_ceq_qi, CODE_FOR_cgt_qi, CODE_FOR_clgt_qi}, | |
940 | {CODE_FOR_ceq_hi, CODE_FOR_cgt_hi, CODE_FOR_clgt_hi}, | |
941 | {CODE_FOR_ceq_si, CODE_FOR_cgt_si, CODE_FOR_clgt_si}, | |
942 | {CODE_FOR_ceq_di, CODE_FOR_cgt_di, CODE_FOR_clgt_di}, | |
943 | {CODE_FOR_ceq_ti, CODE_FOR_cgt_ti, CODE_FOR_clgt_ti}, | |
944 | {CODE_FOR_ceq_sf, CODE_FOR_cgt_sf, 0}, | |
945 | {CODE_FOR_ceq_df, CODE_FOR_cgt_df, 0}, | |
946 | {CODE_FOR_ceq_v16qi, CODE_FOR_cgt_v16qi, CODE_FOR_clgt_v16qi}, | |
947 | {CODE_FOR_ceq_v8hi, CODE_FOR_cgt_v8hi, CODE_FOR_clgt_v8hi}, | |
948 | {CODE_FOR_ceq_v4si, CODE_FOR_cgt_v4si, CODE_FOR_clgt_v4si}, | |
949 | {CODE_FOR_ceq_v4sf, CODE_FOR_cgt_v4sf, 0}, | |
950 | {CODE_FOR_ceq_v2df, CODE_FOR_cgt_v2df, 0}, | |
644459d0 | 951 | }; |
952 | ||
953 | /* Generate a compare for CODE. Return a brand-new rtx that represents | |
954 | the result of the compare. GCC can figure this out too if we don't | |
955 | provide all variations of compares, but GCC always wants to use | |
956 | WORD_MODE, we can generate better code in most cases if we do it | |
957 | ourselves. */ | |
958 | void | |
74f4459c | 959 | spu_emit_branch_or_set (int is_set, rtx cmp, rtx operands[]) |
644459d0 | 960 | { |
961 | int reverse_compare = 0; | |
962 | int reverse_test = 0; | |
5d70b918 | 963 | rtx compare_result, eq_result; |
964 | rtx comp_rtx, eq_rtx; | |
644459d0 | 965 | enum machine_mode comp_mode; |
966 | enum machine_mode op_mode; | |
b9c74b4d | 967 | enum spu_comp_code scode, eq_code; |
968 | enum insn_code ior_code; | |
74f4459c | 969 | enum rtx_code code = GET_CODE (cmp); |
970 | rtx op0 = XEXP (cmp, 0); | |
971 | rtx op1 = XEXP (cmp, 1); | |
644459d0 | 972 | int index; |
5d70b918 | 973 | int eq_test = 0; |
644459d0 | 974 | |
74f4459c | 975 | /* When op1 is a CONST_INT change (X >= C) to (X > C-1), |
644459d0 | 976 | and so on, to keep the constant in operand 1. */ |
74f4459c | 977 | if (GET_CODE (op1) == CONST_INT) |
644459d0 | 978 | { |
74f4459c | 979 | HOST_WIDE_INT val = INTVAL (op1) - 1; |
980 | if (trunc_int_for_mode (val, GET_MODE (op0)) == val) | |
644459d0 | 981 | switch (code) |
982 | { | |
983 | case GE: | |
74f4459c | 984 | op1 = GEN_INT (val); |
644459d0 | 985 | code = GT; |
986 | break; | |
987 | case LT: | |
74f4459c | 988 | op1 = GEN_INT (val); |
644459d0 | 989 | code = LE; |
990 | break; | |
991 | case GEU: | |
74f4459c | 992 | op1 = GEN_INT (val); |
644459d0 | 993 | code = GTU; |
994 | break; | |
995 | case LTU: | |
74f4459c | 996 | op1 = GEN_INT (val); |
644459d0 | 997 | code = LEU; |
998 | break; | |
999 | default: | |
1000 | break; | |
1001 | } | |
1002 | } | |
1003 | ||
5d70b918 | 1004 | comp_mode = SImode; |
74f4459c | 1005 | op_mode = GET_MODE (op0); |
5d70b918 | 1006 | |
644459d0 | 1007 | switch (code) |
1008 | { | |
1009 | case GE: | |
644459d0 | 1010 | scode = SPU_GT; |
07027691 | 1011 | if (HONOR_NANS (op_mode)) |
5d70b918 | 1012 | { |
1013 | reverse_compare = 0; | |
1014 | reverse_test = 0; | |
1015 | eq_test = 1; | |
1016 | eq_code = SPU_EQ; | |
1017 | } | |
1018 | else | |
1019 | { | |
1020 | reverse_compare = 1; | |
1021 | reverse_test = 1; | |
1022 | } | |
644459d0 | 1023 | break; |
1024 | case LE: | |
644459d0 | 1025 | scode = SPU_GT; |
07027691 | 1026 | if (HONOR_NANS (op_mode)) |
5d70b918 | 1027 | { |
1028 | reverse_compare = 1; | |
1029 | reverse_test = 0; | |
1030 | eq_test = 1; | |
1031 | eq_code = SPU_EQ; | |
1032 | } | |
1033 | else | |
1034 | { | |
1035 | reverse_compare = 0; | |
1036 | reverse_test = 1; | |
1037 | } | |
644459d0 | 1038 | break; |
1039 | case LT: | |
1040 | reverse_compare = 1; | |
1041 | reverse_test = 0; | |
1042 | scode = SPU_GT; | |
1043 | break; | |
1044 | case GEU: | |
1045 | reverse_compare = 1; | |
1046 | reverse_test = 1; | |
1047 | scode = SPU_GTU; | |
1048 | break; | |
1049 | case LEU: | |
1050 | reverse_compare = 0; | |
1051 | reverse_test = 1; | |
1052 | scode = SPU_GTU; | |
1053 | break; | |
1054 | case LTU: | |
1055 | reverse_compare = 1; | |
1056 | reverse_test = 0; | |
1057 | scode = SPU_GTU; | |
1058 | break; | |
1059 | case NE: | |
1060 | reverse_compare = 0; | |
1061 | reverse_test = 1; | |
1062 | scode = SPU_EQ; | |
1063 | break; | |
1064 | ||
1065 | case EQ: | |
1066 | scode = SPU_EQ; | |
1067 | break; | |
1068 | case GT: | |
1069 | scode = SPU_GT; | |
1070 | break; | |
1071 | case GTU: | |
1072 | scode = SPU_GTU; | |
1073 | break; | |
1074 | default: | |
1075 | scode = SPU_EQ; | |
1076 | break; | |
1077 | } | |
1078 | ||
644459d0 | 1079 | switch (op_mode) |
1080 | { | |
1081 | case QImode: | |
1082 | index = 0; | |
1083 | comp_mode = QImode; | |
1084 | break; | |
1085 | case HImode: | |
1086 | index = 1; | |
1087 | comp_mode = HImode; | |
1088 | break; | |
1089 | case SImode: | |
1090 | index = 2; | |
1091 | break; | |
1092 | case DImode: | |
1093 | index = 3; | |
1094 | break; | |
1095 | case TImode: | |
1096 | index = 4; | |
1097 | break; | |
1098 | case SFmode: | |
1099 | index = 5; | |
1100 | break; | |
1101 | case DFmode: | |
1102 | index = 6; | |
1103 | break; | |
1104 | case V16QImode: | |
5474166e | 1105 | index = 7; |
1106 | comp_mode = op_mode; | |
1107 | break; | |
644459d0 | 1108 | case V8HImode: |
5474166e | 1109 | index = 8; |
1110 | comp_mode = op_mode; | |
1111 | break; | |
644459d0 | 1112 | case V4SImode: |
5474166e | 1113 | index = 9; |
1114 | comp_mode = op_mode; | |
1115 | break; | |
644459d0 | 1116 | case V4SFmode: |
5474166e | 1117 | index = 10; |
1118 | comp_mode = V4SImode; | |
1119 | break; | |
644459d0 | 1120 | case V2DFmode: |
5474166e | 1121 | index = 11; |
1122 | comp_mode = V2DImode; | |
644459d0 | 1123 | break; |
5474166e | 1124 | case V2DImode: |
644459d0 | 1125 | default: |
1126 | abort (); | |
1127 | } | |
1128 | ||
74f4459c | 1129 | if (GET_MODE (op1) == DFmode |
07027691 | 1130 | && (scode != SPU_GT && scode != SPU_EQ)) |
1131 | abort (); | |
644459d0 | 1132 | |
74f4459c | 1133 | if (is_set == 0 && op1 == const0_rtx |
1134 | && (GET_MODE (op0) == SImode | |
1135 | || GET_MODE (op0) == HImode) && scode == SPU_EQ) | |
644459d0 | 1136 | { |
1137 | /* Don't need to set a register with the result when we are | |
1138 | comparing against zero and branching. */ | |
1139 | reverse_test = !reverse_test; | |
74f4459c | 1140 | compare_result = op0; |
644459d0 | 1141 | } |
1142 | else | |
1143 | { | |
1144 | compare_result = gen_reg_rtx (comp_mode); | |
1145 | ||
1146 | if (reverse_compare) | |
1147 | { | |
74f4459c | 1148 | rtx t = op1; |
1149 | op1 = op0; | |
1150 | op0 = t; | |
644459d0 | 1151 | } |
1152 | ||
1153 | if (spu_comp_icode[index][scode] == 0) | |
1154 | abort (); | |
1155 | ||
1156 | if (!(*insn_data[spu_comp_icode[index][scode]].operand[1].predicate) | |
74f4459c | 1157 | (op0, op_mode)) |
1158 | op0 = force_reg (op_mode, op0); | |
644459d0 | 1159 | if (!(*insn_data[spu_comp_icode[index][scode]].operand[2].predicate) |
74f4459c | 1160 | (op1, op_mode)) |
1161 | op1 = force_reg (op_mode, op1); | |
644459d0 | 1162 | comp_rtx = GEN_FCN (spu_comp_icode[index][scode]) (compare_result, |
74f4459c | 1163 | op0, op1); |
644459d0 | 1164 | if (comp_rtx == 0) |
1165 | abort (); | |
1166 | emit_insn (comp_rtx); | |
1167 | ||
5d70b918 | 1168 | if (eq_test) |
1169 | { | |
1170 | eq_result = gen_reg_rtx (comp_mode); | |
1171 | eq_rtx = GEN_FCN (spu_comp_icode[index][eq_code]) (eq_result, | |
74f4459c | 1172 | op0, op1); |
5d70b918 | 1173 | if (eq_rtx == 0) |
1174 | abort (); | |
1175 | emit_insn (eq_rtx); | |
d6bf3b14 | 1176 | ior_code = optab_handler (ior_optab, comp_mode); |
5d70b918 | 1177 | gcc_assert (ior_code != CODE_FOR_nothing); |
1178 | emit_insn (GEN_FCN (ior_code) | |
1179 | (compare_result, compare_result, eq_result)); | |
1180 | } | |
644459d0 | 1181 | } |
1182 | ||
1183 | if (is_set == 0) | |
1184 | { | |
1185 | rtx bcomp; | |
1186 | rtx loc_ref; | |
1187 | ||
1188 | /* We don't have branch on QI compare insns, so we convert the | |
1189 | QI compare result to a HI result. */ | |
1190 | if (comp_mode == QImode) | |
1191 | { | |
1192 | rtx old_res = compare_result; | |
1193 | compare_result = gen_reg_rtx (HImode); | |
1194 | comp_mode = HImode; | |
1195 | emit_insn (gen_extendqihi2 (compare_result, old_res)); | |
1196 | } | |
1197 | ||
1198 | if (reverse_test) | |
1199 | bcomp = gen_rtx_EQ (comp_mode, compare_result, const0_rtx); | |
1200 | else | |
1201 | bcomp = gen_rtx_NE (comp_mode, compare_result, const0_rtx); | |
1202 | ||
74f4459c | 1203 | loc_ref = gen_rtx_LABEL_REF (VOIDmode, operands[3]); |
644459d0 | 1204 | emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, |
1205 | gen_rtx_IF_THEN_ELSE (VOIDmode, bcomp, | |
1206 | loc_ref, pc_rtx))); | |
1207 | } | |
1208 | else if (is_set == 2) | |
1209 | { | |
74f4459c | 1210 | rtx target = operands[0]; |
644459d0 | 1211 | int compare_size = GET_MODE_BITSIZE (comp_mode); |
1212 | int target_size = GET_MODE_BITSIZE (GET_MODE (target)); | |
1213 | enum machine_mode mode = mode_for_size (target_size, MODE_INT, 0); | |
1214 | rtx select_mask; | |
1215 | rtx op_t = operands[2]; | |
1216 | rtx op_f = operands[3]; | |
1217 | ||
1218 | /* The result of the comparison can be SI, HI or QI mode. Create a | |
1219 | mask based on that result. */ | |
1220 | if (target_size > compare_size) | |
1221 | { | |
1222 | select_mask = gen_reg_rtx (mode); | |
1223 | emit_insn (gen_extend_compare (select_mask, compare_result)); | |
1224 | } | |
1225 | else if (target_size < compare_size) | |
1226 | select_mask = | |
1227 | gen_rtx_SUBREG (mode, compare_result, | |
1228 | (compare_size - target_size) / BITS_PER_UNIT); | |
1229 | else if (comp_mode != mode) | |
1230 | select_mask = gen_rtx_SUBREG (mode, compare_result, 0); | |
1231 | else | |
1232 | select_mask = compare_result; | |
1233 | ||
1234 | if (GET_MODE (target) != GET_MODE (op_t) | |
1235 | || GET_MODE (target) != GET_MODE (op_f)) | |
1236 | abort (); | |
1237 | ||
1238 | if (reverse_test) | |
1239 | emit_insn (gen_selb (target, op_t, op_f, select_mask)); | |
1240 | else | |
1241 | emit_insn (gen_selb (target, op_f, op_t, select_mask)); | |
1242 | } | |
1243 | else | |
1244 | { | |
74f4459c | 1245 | rtx target = operands[0]; |
644459d0 | 1246 | if (reverse_test) |
1247 | emit_insn (gen_rtx_SET (VOIDmode, compare_result, | |
1248 | gen_rtx_NOT (comp_mode, compare_result))); | |
1249 | if (GET_MODE (target) == SImode && GET_MODE (compare_result) == HImode) | |
1250 | emit_insn (gen_extendhisi2 (target, compare_result)); | |
1251 | else if (GET_MODE (target) == SImode | |
1252 | && GET_MODE (compare_result) == QImode) | |
1253 | emit_insn (gen_extend_compare (target, compare_result)); | |
1254 | else | |
1255 | emit_move_insn (target, compare_result); | |
1256 | } | |
1257 | } | |
1258 | ||
1259 | HOST_WIDE_INT | |
1260 | const_double_to_hwint (rtx x) | |
1261 | { | |
1262 | HOST_WIDE_INT val; | |
1263 | REAL_VALUE_TYPE rv; | |
1264 | if (GET_MODE (x) == SFmode) | |
1265 | { | |
1266 | REAL_VALUE_FROM_CONST_DOUBLE (rv, x); | |
1267 | REAL_VALUE_TO_TARGET_SINGLE (rv, val); | |
1268 | } | |
1269 | else if (GET_MODE (x) == DFmode) | |
1270 | { | |
1271 | long l[2]; | |
1272 | REAL_VALUE_FROM_CONST_DOUBLE (rv, x); | |
1273 | REAL_VALUE_TO_TARGET_DOUBLE (rv, l); | |
1274 | val = l[0]; | |
1275 | val = (val << 32) | (l[1] & 0xffffffff); | |
1276 | } | |
1277 | else | |
1278 | abort (); | |
1279 | return val; | |
1280 | } | |
1281 | ||
1282 | rtx | |
1283 | hwint_to_const_double (enum machine_mode mode, HOST_WIDE_INT v) | |
1284 | { | |
1285 | long tv[2]; | |
1286 | REAL_VALUE_TYPE rv; | |
1287 | gcc_assert (mode == SFmode || mode == DFmode); | |
1288 | ||
1289 | if (mode == SFmode) | |
1290 | tv[0] = (v << 32) >> 32; | |
1291 | else if (mode == DFmode) | |
1292 | { | |
1293 | tv[1] = (v << 32) >> 32; | |
1294 | tv[0] = v >> 32; | |
1295 | } | |
1296 | real_from_target (&rv, tv, mode); | |
1297 | return CONST_DOUBLE_FROM_REAL_VALUE (rv, mode); | |
1298 | } | |
1299 | ||
1300 | void | |
1301 | print_operand_address (FILE * file, register rtx addr) | |
1302 | { | |
1303 | rtx reg; | |
1304 | rtx offset; | |
1305 | ||
e04cf423 | 1306 | if (GET_CODE (addr) == AND |
1307 | && GET_CODE (XEXP (addr, 1)) == CONST_INT | |
1308 | && INTVAL (XEXP (addr, 1)) == -16) | |
1309 | addr = XEXP (addr, 0); | |
1310 | ||
644459d0 | 1311 | switch (GET_CODE (addr)) |
1312 | { | |
1313 | case REG: | |
1314 | fprintf (file, "0(%s)", reg_names[REGNO (addr)]); | |
1315 | break; | |
1316 | ||
1317 | case PLUS: | |
1318 | reg = XEXP (addr, 0); | |
1319 | offset = XEXP (addr, 1); | |
1320 | if (GET_CODE (offset) == REG) | |
1321 | { | |
1322 | fprintf (file, "%s,%s", reg_names[REGNO (reg)], | |
1323 | reg_names[REGNO (offset)]); | |
1324 | } | |
1325 | else if (GET_CODE (offset) == CONST_INT) | |
1326 | { | |
1327 | fprintf (file, HOST_WIDE_INT_PRINT_DEC "(%s)", | |
1328 | INTVAL (offset), reg_names[REGNO (reg)]); | |
1329 | } | |
1330 | else | |
1331 | abort (); | |
1332 | break; | |
1333 | ||
1334 | case CONST: | |
1335 | case LABEL_REF: | |
1336 | case SYMBOL_REF: | |
1337 | case CONST_INT: | |
1338 | output_addr_const (file, addr); | |
1339 | break; | |
1340 | ||
1341 | default: | |
1342 | debug_rtx (addr); | |
1343 | abort (); | |
1344 | } | |
1345 | } | |
1346 | ||
1347 | void | |
1348 | print_operand (FILE * file, rtx x, int code) | |
1349 | { | |
1350 | enum machine_mode mode = GET_MODE (x); | |
1351 | HOST_WIDE_INT val; | |
1352 | unsigned char arr[16]; | |
1353 | int xcode = GET_CODE (x); | |
dea01258 | 1354 | int i, info; |
644459d0 | 1355 | if (GET_MODE (x) == VOIDmode) |
1356 | switch (code) | |
1357 | { | |
644459d0 | 1358 | case 'L': /* 128 bits, signed */ |
1359 | case 'm': /* 128 bits, signed */ | |
1360 | case 'T': /* 128 bits, signed */ | |
1361 | case 't': /* 128 bits, signed */ | |
1362 | mode = TImode; | |
1363 | break; | |
644459d0 | 1364 | case 'K': /* 64 bits, signed */ |
1365 | case 'k': /* 64 bits, signed */ | |
1366 | case 'D': /* 64 bits, signed */ | |
1367 | case 'd': /* 64 bits, signed */ | |
1368 | mode = DImode; | |
1369 | break; | |
644459d0 | 1370 | case 'J': /* 32 bits, signed */ |
1371 | case 'j': /* 32 bits, signed */ | |
1372 | case 's': /* 32 bits, signed */ | |
1373 | case 'S': /* 32 bits, signed */ | |
1374 | mode = SImode; | |
1375 | break; | |
1376 | } | |
1377 | switch (code) | |
1378 | { | |
1379 | ||
1380 | case 'j': /* 32 bits, signed */ | |
1381 | case 'k': /* 64 bits, signed */ | |
1382 | case 'm': /* 128 bits, signed */ | |
1383 | if (xcode == CONST_INT | |
1384 | || xcode == CONST_DOUBLE || xcode == CONST_VECTOR) | |
1385 | { | |
1386 | gcc_assert (logical_immediate_p (x, mode)); | |
1387 | constant_to_array (mode, x, arr); | |
1388 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
1389 | val = trunc_int_for_mode (val, SImode); | |
1390 | switch (which_logical_immediate (val)) | |
1391 | { | |
1392 | case SPU_ORI: | |
1393 | break; | |
1394 | case SPU_ORHI: | |
1395 | fprintf (file, "h"); | |
1396 | break; | |
1397 | case SPU_ORBI: | |
1398 | fprintf (file, "b"); | |
1399 | break; | |
1400 | default: | |
1401 | gcc_unreachable(); | |
1402 | } | |
1403 | } | |
1404 | else | |
1405 | gcc_unreachable(); | |
1406 | return; | |
1407 | ||
1408 | case 'J': /* 32 bits, signed */ | |
1409 | case 'K': /* 64 bits, signed */ | |
1410 | case 'L': /* 128 bits, signed */ | |
1411 | if (xcode == CONST_INT | |
1412 | || xcode == CONST_DOUBLE || xcode == CONST_VECTOR) | |
1413 | { | |
1414 | gcc_assert (logical_immediate_p (x, mode) | |
1415 | || iohl_immediate_p (x, mode)); | |
1416 | constant_to_array (mode, x, arr); | |
1417 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
1418 | val = trunc_int_for_mode (val, SImode); | |
1419 | switch (which_logical_immediate (val)) | |
1420 | { | |
1421 | case SPU_ORI: | |
1422 | case SPU_IOHL: | |
1423 | break; | |
1424 | case SPU_ORHI: | |
1425 | val = trunc_int_for_mode (val, HImode); | |
1426 | break; | |
1427 | case SPU_ORBI: | |
1428 | val = trunc_int_for_mode (val, QImode); | |
1429 | break; | |
1430 | default: | |
1431 | gcc_unreachable(); | |
1432 | } | |
1433 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, val); | |
1434 | } | |
1435 | else | |
1436 | gcc_unreachable(); | |
1437 | return; | |
1438 | ||
1439 | case 't': /* 128 bits, signed */ | |
1440 | case 'd': /* 64 bits, signed */ | |
1441 | case 's': /* 32 bits, signed */ | |
dea01258 | 1442 | if (CONSTANT_P (x)) |
644459d0 | 1443 | { |
dea01258 | 1444 | enum immediate_class c = classify_immediate (x, mode); |
1445 | switch (c) | |
1446 | { | |
1447 | case IC_IL1: | |
1448 | constant_to_array (mode, x, arr); | |
1449 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
1450 | val = trunc_int_for_mode (val, SImode); | |
1451 | switch (which_immediate_load (val)) | |
1452 | { | |
1453 | case SPU_IL: | |
1454 | break; | |
1455 | case SPU_ILA: | |
1456 | fprintf (file, "a"); | |
1457 | break; | |
1458 | case SPU_ILH: | |
1459 | fprintf (file, "h"); | |
1460 | break; | |
1461 | case SPU_ILHU: | |
1462 | fprintf (file, "hu"); | |
1463 | break; | |
1464 | default: | |
1465 | gcc_unreachable (); | |
1466 | } | |
1467 | break; | |
1468 | case IC_CPAT: | |
1469 | constant_to_array (mode, x, arr); | |
1470 | cpat_info (arr, GET_MODE_SIZE (mode), &info, 0); | |
1471 | if (info == 1) | |
1472 | fprintf (file, "b"); | |
1473 | else if (info == 2) | |
1474 | fprintf (file, "h"); | |
1475 | else if (info == 4) | |
1476 | fprintf (file, "w"); | |
1477 | else if (info == 8) | |
1478 | fprintf (file, "d"); | |
1479 | break; | |
1480 | case IC_IL1s: | |
1481 | if (xcode == CONST_VECTOR) | |
1482 | { | |
1483 | x = CONST_VECTOR_ELT (x, 0); | |
1484 | xcode = GET_CODE (x); | |
1485 | } | |
1486 | if (xcode == SYMBOL_REF || xcode == LABEL_REF || xcode == CONST) | |
1487 | fprintf (file, "a"); | |
1488 | else if (xcode == HIGH) | |
1489 | fprintf (file, "hu"); | |
1490 | break; | |
1491 | case IC_FSMBI: | |
5df189be | 1492 | case IC_FSMBI2: |
dea01258 | 1493 | case IC_IL2: |
1494 | case IC_IL2s: | |
1495 | case IC_POOL: | |
1496 | abort (); | |
1497 | } | |
644459d0 | 1498 | } |
644459d0 | 1499 | else |
1500 | gcc_unreachable (); | |
1501 | return; | |
1502 | ||
1503 | case 'T': /* 128 bits, signed */ | |
1504 | case 'D': /* 64 bits, signed */ | |
1505 | case 'S': /* 32 bits, signed */ | |
dea01258 | 1506 | if (CONSTANT_P (x)) |
644459d0 | 1507 | { |
dea01258 | 1508 | enum immediate_class c = classify_immediate (x, mode); |
1509 | switch (c) | |
644459d0 | 1510 | { |
dea01258 | 1511 | case IC_IL1: |
1512 | constant_to_array (mode, x, arr); | |
1513 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
1514 | val = trunc_int_for_mode (val, SImode); | |
1515 | switch (which_immediate_load (val)) | |
1516 | { | |
1517 | case SPU_IL: | |
1518 | case SPU_ILA: | |
1519 | break; | |
1520 | case SPU_ILH: | |
1521 | case SPU_ILHU: | |
1522 | val = trunc_int_for_mode (((arr[0] << 8) | arr[1]), HImode); | |
1523 | break; | |
1524 | default: | |
1525 | gcc_unreachable (); | |
1526 | } | |
1527 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, val); | |
1528 | break; | |
1529 | case IC_FSMBI: | |
1530 | constant_to_array (mode, x, arr); | |
1531 | val = 0; | |
1532 | for (i = 0; i < 16; i++) | |
1533 | { | |
1534 | val <<= 1; | |
1535 | val |= arr[i] & 1; | |
1536 | } | |
1537 | print_operand (file, GEN_INT (val), 0); | |
1538 | break; | |
1539 | case IC_CPAT: | |
1540 | constant_to_array (mode, x, arr); | |
1541 | cpat_info (arr, GET_MODE_SIZE (mode), 0, &info); | |
1542 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, (HOST_WIDE_INT)info); | |
644459d0 | 1543 | break; |
dea01258 | 1544 | case IC_IL1s: |
dea01258 | 1545 | if (xcode == HIGH) |
5df189be | 1546 | x = XEXP (x, 0); |
1547 | if (GET_CODE (x) == CONST_VECTOR) | |
1548 | x = CONST_VECTOR_ELT (x, 0); | |
1549 | output_addr_const (file, x); | |
1550 | if (xcode == HIGH) | |
1551 | fprintf (file, "@h"); | |
644459d0 | 1552 | break; |
dea01258 | 1553 | case IC_IL2: |
1554 | case IC_IL2s: | |
5df189be | 1555 | case IC_FSMBI2: |
dea01258 | 1556 | case IC_POOL: |
1557 | abort (); | |
644459d0 | 1558 | } |
c8befdb9 | 1559 | } |
644459d0 | 1560 | else |
1561 | gcc_unreachable (); | |
1562 | return; | |
1563 | ||
644459d0 | 1564 | case 'C': |
1565 | if (xcode == CONST_INT) | |
1566 | { | |
1567 | /* Only 4 least significant bits are relevant for generate | |
1568 | control word instructions. */ | |
1569 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (x) & 15); | |
1570 | return; | |
1571 | } | |
1572 | break; | |
1573 | ||
1574 | case 'M': /* print code for c*d */ | |
1575 | if (GET_CODE (x) == CONST_INT) | |
1576 | switch (INTVAL (x)) | |
1577 | { | |
1578 | case 1: | |
1579 | fprintf (file, "b"); | |
1580 | break; | |
1581 | case 2: | |
1582 | fprintf (file, "h"); | |
1583 | break; | |
1584 | case 4: | |
1585 | fprintf (file, "w"); | |
1586 | break; | |
1587 | case 8: | |
1588 | fprintf (file, "d"); | |
1589 | break; | |
1590 | default: | |
1591 | gcc_unreachable(); | |
1592 | } | |
1593 | else | |
1594 | gcc_unreachable(); | |
1595 | return; | |
1596 | ||
1597 | case 'N': /* Negate the operand */ | |
1598 | if (xcode == CONST_INT) | |
1599 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, -INTVAL (x)); | |
1600 | else if (xcode == CONST_VECTOR) | |
1601 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, | |
1602 | -INTVAL (CONST_VECTOR_ELT (x, 0))); | |
1603 | return; | |
1604 | ||
1605 | case 'I': /* enable/disable interrupts */ | |
1606 | if (xcode == CONST_INT) | |
1607 | fprintf (file, "%s", INTVAL (x) == 0 ? "d" : "e"); | |
1608 | return; | |
1609 | ||
1610 | case 'b': /* branch modifiers */ | |
1611 | if (xcode == REG) | |
1612 | fprintf (file, "%s", GET_MODE (x) == HImode ? "h" : ""); | |
1613 | else if (COMPARISON_P (x)) | |
1614 | fprintf (file, "%s", xcode == NE ? "n" : ""); | |
1615 | return; | |
1616 | ||
1617 | case 'i': /* indirect call */ | |
1618 | if (xcode == MEM) | |
1619 | { | |
1620 | if (GET_CODE (XEXP (x, 0)) == REG) | |
1621 | /* Used in indirect function calls. */ | |
1622 | fprintf (file, "%s", reg_names[REGNO (XEXP (x, 0))]); | |
1623 | else | |
1624 | output_address (XEXP (x, 0)); | |
1625 | } | |
1626 | return; | |
1627 | ||
1628 | case 'p': /* load/store */ | |
1629 | if (xcode == MEM) | |
1630 | { | |
1631 | x = XEXP (x, 0); | |
1632 | xcode = GET_CODE (x); | |
1633 | } | |
e04cf423 | 1634 | if (xcode == AND) |
1635 | { | |
1636 | x = XEXP (x, 0); | |
1637 | xcode = GET_CODE (x); | |
1638 | } | |
644459d0 | 1639 | if (xcode == REG) |
1640 | fprintf (file, "d"); | |
1641 | else if (xcode == CONST_INT) | |
1642 | fprintf (file, "a"); | |
1643 | else if (xcode == CONST || xcode == SYMBOL_REF || xcode == LABEL_REF) | |
1644 | fprintf (file, "r"); | |
1645 | else if (xcode == PLUS || xcode == LO_SUM) | |
1646 | { | |
1647 | if (GET_CODE (XEXP (x, 1)) == REG) | |
1648 | fprintf (file, "x"); | |
1649 | else | |
1650 | fprintf (file, "d"); | |
1651 | } | |
1652 | return; | |
1653 | ||
5df189be | 1654 | case 'e': |
1655 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1656 | val &= 0x7; | |
1657 | output_addr_const (file, GEN_INT (val)); | |
1658 | return; | |
1659 | ||
1660 | case 'f': | |
1661 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1662 | val &= 0x1f; | |
1663 | output_addr_const (file, GEN_INT (val)); | |
1664 | return; | |
1665 | ||
1666 | case 'g': | |
1667 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1668 | val &= 0x3f; | |
1669 | output_addr_const (file, GEN_INT (val)); | |
1670 | return; | |
1671 | ||
1672 | case 'h': | |
1673 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1674 | val = (val >> 3) & 0x1f; | |
1675 | output_addr_const (file, GEN_INT (val)); | |
1676 | return; | |
1677 | ||
1678 | case 'E': | |
1679 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1680 | val = -val; | |
1681 | val &= 0x7; | |
1682 | output_addr_const (file, GEN_INT (val)); | |
1683 | return; | |
1684 | ||
1685 | case 'F': | |
1686 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1687 | val = -val; | |
1688 | val &= 0x1f; | |
1689 | output_addr_const (file, GEN_INT (val)); | |
1690 | return; | |
1691 | ||
1692 | case 'G': | |
1693 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1694 | val = -val; | |
1695 | val &= 0x3f; | |
1696 | output_addr_const (file, GEN_INT (val)); | |
1697 | return; | |
1698 | ||
1699 | case 'H': | |
1700 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1701 | val = -(val & -8ll); | |
1702 | val = (val >> 3) & 0x1f; | |
1703 | output_addr_const (file, GEN_INT (val)); | |
1704 | return; | |
1705 | ||
56c7bfc2 | 1706 | case 'v': |
1707 | case 'w': | |
1708 | constant_to_array (mode, x, arr); | |
1709 | val = (((arr[0] << 1) + (arr[1] >> 7)) & 0xff) - 127; | |
1710 | output_addr_const (file, GEN_INT (code == 'w' ? -val : val)); | |
1711 | return; | |
1712 | ||
644459d0 | 1713 | case 0: |
1714 | if (xcode == REG) | |
1715 | fprintf (file, "%s", reg_names[REGNO (x)]); | |
1716 | else if (xcode == MEM) | |
1717 | output_address (XEXP (x, 0)); | |
1718 | else if (xcode == CONST_VECTOR) | |
dea01258 | 1719 | print_operand (file, CONST_VECTOR_ELT (x, 0), 0); |
644459d0 | 1720 | else |
1721 | output_addr_const (file, x); | |
1722 | return; | |
1723 | ||
f6a0d06f | 1724 | /* unused letters |
56c7bfc2 | 1725 | o qr u yz |
5df189be | 1726 | AB OPQR UVWXYZ */ |
644459d0 | 1727 | default: |
1728 | output_operand_lossage ("invalid %%xn code"); | |
1729 | } | |
1730 | gcc_unreachable (); | |
1731 | } | |
1732 | ||
644459d0 | 1733 | /* For PIC mode we've reserved PIC_OFFSET_TABLE_REGNUM, which is a |
1734 | caller saved register. For leaf functions it is more efficient to | |
1735 | use a volatile register because we won't need to save and restore the | |
1736 | pic register. This routine is only valid after register allocation | |
1737 | is completed, so we can pick an unused register. */ | |
1738 | static rtx | |
1739 | get_pic_reg (void) | |
1740 | { | |
1741 | rtx pic_reg = pic_offset_table_rtx; | |
1742 | if (!reload_completed && !reload_in_progress) | |
1743 | abort (); | |
87a95921 | 1744 | if (current_function_is_leaf && !df_regs_ever_live_p (LAST_ARG_REGNUM)) |
1745 | pic_reg = gen_rtx_REG (SImode, LAST_ARG_REGNUM); | |
644459d0 | 1746 | return pic_reg; |
1747 | } | |
1748 | ||
5df189be | 1749 | /* Split constant addresses to handle cases that are too large. |
1750 | Add in the pic register when in PIC mode. | |
1751 | Split immediates that require more than 1 instruction. */ | |
dea01258 | 1752 | int |
1753 | spu_split_immediate (rtx * ops) | |
c8befdb9 | 1754 | { |
dea01258 | 1755 | enum machine_mode mode = GET_MODE (ops[0]); |
1756 | enum immediate_class c = classify_immediate (ops[1], mode); | |
1757 | ||
1758 | switch (c) | |
c8befdb9 | 1759 | { |
dea01258 | 1760 | case IC_IL2: |
1761 | { | |
1762 | unsigned char arrhi[16]; | |
1763 | unsigned char arrlo[16]; | |
98bbec1e | 1764 | rtx to, temp, hi, lo; |
dea01258 | 1765 | int i; |
98bbec1e | 1766 | enum machine_mode imode = mode; |
1767 | /* We need to do reals as ints because the constant used in the | |
1768 | IOR might not be a legitimate real constant. */ | |
1769 | imode = int_mode_for_mode (mode); | |
dea01258 | 1770 | constant_to_array (mode, ops[1], arrhi); |
98bbec1e | 1771 | if (imode != mode) |
1772 | to = simplify_gen_subreg (imode, ops[0], mode, 0); | |
1773 | else | |
1774 | to = ops[0]; | |
1775 | temp = !can_create_pseudo_p () ? to : gen_reg_rtx (imode); | |
dea01258 | 1776 | for (i = 0; i < 16; i += 4) |
1777 | { | |
1778 | arrlo[i + 2] = arrhi[i + 2]; | |
1779 | arrlo[i + 3] = arrhi[i + 3]; | |
1780 | arrlo[i + 0] = arrlo[i + 1] = 0; | |
1781 | arrhi[i + 2] = arrhi[i + 3] = 0; | |
1782 | } | |
98bbec1e | 1783 | hi = array_to_constant (imode, arrhi); |
1784 | lo = array_to_constant (imode, arrlo); | |
1785 | emit_move_insn (temp, hi); | |
dea01258 | 1786 | emit_insn (gen_rtx_SET |
98bbec1e | 1787 | (VOIDmode, to, gen_rtx_IOR (imode, temp, lo))); |
dea01258 | 1788 | return 1; |
1789 | } | |
5df189be | 1790 | case IC_FSMBI2: |
1791 | { | |
1792 | unsigned char arr_fsmbi[16]; | |
1793 | unsigned char arr_andbi[16]; | |
1794 | rtx to, reg_fsmbi, reg_and; | |
1795 | int i; | |
1796 | enum machine_mode imode = mode; | |
1797 | /* We need to do reals as ints because the constant used in the | |
1798 | * AND might not be a legitimate real constant. */ | |
1799 | imode = int_mode_for_mode (mode); | |
1800 | constant_to_array (mode, ops[1], arr_fsmbi); | |
1801 | if (imode != mode) | |
1802 | to = simplify_gen_subreg(imode, ops[0], GET_MODE (ops[0]), 0); | |
1803 | else | |
1804 | to = ops[0]; | |
1805 | for (i = 0; i < 16; i++) | |
1806 | if (arr_fsmbi[i] != 0) | |
1807 | { | |
1808 | arr_andbi[0] = arr_fsmbi[i]; | |
1809 | arr_fsmbi[i] = 0xff; | |
1810 | } | |
1811 | for (i = 1; i < 16; i++) | |
1812 | arr_andbi[i] = arr_andbi[0]; | |
1813 | reg_fsmbi = array_to_constant (imode, arr_fsmbi); | |
1814 | reg_and = array_to_constant (imode, arr_andbi); | |
1815 | emit_move_insn (to, reg_fsmbi); | |
1816 | emit_insn (gen_rtx_SET | |
1817 | (VOIDmode, to, gen_rtx_AND (imode, to, reg_and))); | |
1818 | return 1; | |
1819 | } | |
dea01258 | 1820 | case IC_POOL: |
1821 | if (reload_in_progress || reload_completed) | |
1822 | { | |
1823 | rtx mem = force_const_mem (mode, ops[1]); | |
1824 | if (TARGET_LARGE_MEM) | |
1825 | { | |
1826 | rtx addr = gen_rtx_REG (Pmode, REGNO (ops[0])); | |
1827 | emit_move_insn (addr, XEXP (mem, 0)); | |
1828 | mem = replace_equiv_address (mem, addr); | |
1829 | } | |
1830 | emit_move_insn (ops[0], mem); | |
1831 | return 1; | |
1832 | } | |
1833 | break; | |
1834 | case IC_IL1s: | |
1835 | case IC_IL2s: | |
1836 | if (reload_completed && GET_CODE (ops[1]) != HIGH) | |
1837 | { | |
1838 | if (c == IC_IL2s) | |
1839 | { | |
5df189be | 1840 | emit_move_insn (ops[0], gen_rtx_HIGH (mode, ops[1])); |
1841 | emit_move_insn (ops[0], gen_rtx_LO_SUM (mode, ops[0], ops[1])); | |
dea01258 | 1842 | } |
1843 | else if (flag_pic) | |
1844 | emit_insn (gen_pic (ops[0], ops[1])); | |
1845 | if (flag_pic) | |
1846 | { | |
1847 | rtx pic_reg = get_pic_reg (); | |
1848 | emit_insn (gen_addsi3 (ops[0], ops[0], pic_reg)); | |
18d50ae6 | 1849 | crtl->uses_pic_offset_table = 1; |
dea01258 | 1850 | } |
1851 | return flag_pic || c == IC_IL2s; | |
1852 | } | |
1853 | break; | |
1854 | case IC_IL1: | |
1855 | case IC_FSMBI: | |
1856 | case IC_CPAT: | |
1857 | break; | |
c8befdb9 | 1858 | } |
dea01258 | 1859 | return 0; |
c8befdb9 | 1860 | } |
1861 | ||
644459d0 | 1862 | /* SAVING is TRUE when we are generating the actual load and store |
1863 | instructions for REGNO. When determining the size of the stack | |
1864 | needed for saving register we must allocate enough space for the | |
1865 | worst case, because we don't always have the information early enough | |
1866 | to not allocate it. But we can at least eliminate the actual loads | |
1867 | and stores during the prologue/epilogue. */ | |
1868 | static int | |
1869 | need_to_save_reg (int regno, int saving) | |
1870 | { | |
3072d30e | 1871 | if (df_regs_ever_live_p (regno) && !call_used_regs[regno]) |
644459d0 | 1872 | return 1; |
1873 | if (flag_pic | |
1874 | && regno == PIC_OFFSET_TABLE_REGNUM | |
18d50ae6 | 1875 | && (!saving || crtl->uses_pic_offset_table) |
644459d0 | 1876 | && (!saving |
3072d30e | 1877 | || !current_function_is_leaf || df_regs_ever_live_p (LAST_ARG_REGNUM))) |
644459d0 | 1878 | return 1; |
1879 | return 0; | |
1880 | } | |
1881 | ||
1882 | /* This function is only correct starting with local register | |
1883 | allocation */ | |
1884 | int | |
1885 | spu_saved_regs_size (void) | |
1886 | { | |
1887 | int reg_save_size = 0; | |
1888 | int regno; | |
1889 | ||
1890 | for (regno = FIRST_PSEUDO_REGISTER - 1; regno >= 0; --regno) | |
1891 | if (need_to_save_reg (regno, 0)) | |
1892 | reg_save_size += 0x10; | |
1893 | return reg_save_size; | |
1894 | } | |
1895 | ||
1896 | static rtx | |
1897 | frame_emit_store (int regno, rtx addr, HOST_WIDE_INT offset) | |
1898 | { | |
1899 | rtx reg = gen_rtx_REG (V4SImode, regno); | |
1900 | rtx mem = | |
1901 | gen_frame_mem (V4SImode, gen_rtx_PLUS (Pmode, addr, GEN_INT (offset))); | |
1902 | return emit_insn (gen_movv4si (mem, reg)); | |
1903 | } | |
1904 | ||
1905 | static rtx | |
1906 | frame_emit_load (int regno, rtx addr, HOST_WIDE_INT offset) | |
1907 | { | |
1908 | rtx reg = gen_rtx_REG (V4SImode, regno); | |
1909 | rtx mem = | |
1910 | gen_frame_mem (V4SImode, gen_rtx_PLUS (Pmode, addr, GEN_INT (offset))); | |
1911 | return emit_insn (gen_movv4si (reg, mem)); | |
1912 | } | |
1913 | ||
1914 | /* This happens after reload, so we need to expand it. */ | |
1915 | static rtx | |
1916 | frame_emit_add_imm (rtx dst, rtx src, HOST_WIDE_INT imm, rtx scratch) | |
1917 | { | |
1918 | rtx insn; | |
1919 | if (satisfies_constraint_K (GEN_INT (imm))) | |
1920 | { | |
1921 | insn = emit_insn (gen_addsi3 (dst, src, GEN_INT (imm))); | |
1922 | } | |
1923 | else | |
1924 | { | |
3072d30e | 1925 | emit_insn (gen_movsi (scratch, gen_int_mode (imm, SImode))); |
644459d0 | 1926 | insn = emit_insn (gen_addsi3 (dst, src, scratch)); |
1927 | if (REGNO (src) == REGNO (scratch)) | |
1928 | abort (); | |
1929 | } | |
644459d0 | 1930 | return insn; |
1931 | } | |
1932 | ||
1933 | /* Return nonzero if this function is known to have a null epilogue. */ | |
1934 | ||
1935 | int | |
1936 | direct_return (void) | |
1937 | { | |
1938 | if (reload_completed) | |
1939 | { | |
1940 | if (cfun->static_chain_decl == 0 | |
1941 | && (spu_saved_regs_size () | |
1942 | + get_frame_size () | |
abe32cce | 1943 | + crtl->outgoing_args_size |
1944 | + crtl->args.pretend_args_size == 0) | |
644459d0 | 1945 | && current_function_is_leaf) |
1946 | return 1; | |
1947 | } | |
1948 | return 0; | |
1949 | } | |
1950 | ||
1951 | /* | |
1952 | The stack frame looks like this: | |
1953 | +-------------+ | |
1954 | | incoming | | |
a8e019fa | 1955 | | args | |
1956 | AP -> +-------------+ | |
644459d0 | 1957 | | $lr save | |
1958 | +-------------+ | |
1959 | prev SP | back chain | | |
1960 | +-------------+ | |
1961 | | var args | | |
abe32cce | 1962 | | reg save | crtl->args.pretend_args_size bytes |
644459d0 | 1963 | +-------------+ |
1964 | | ... | | |
1965 | | saved regs | spu_saved_regs_size() bytes | |
a8e019fa | 1966 | FP -> +-------------+ |
644459d0 | 1967 | | ... | |
a8e019fa | 1968 | | vars | get_frame_size() bytes |
1969 | HFP -> +-------------+ | |
644459d0 | 1970 | | ... | |
1971 | | outgoing | | |
abe32cce | 1972 | | args | crtl->outgoing_args_size bytes |
644459d0 | 1973 | +-------------+ |
1974 | | $lr of next | | |
1975 | | frame | | |
1976 | +-------------+ | |
a8e019fa | 1977 | | back chain | |
1978 | SP -> +-------------+ | |
644459d0 | 1979 | |
1980 | */ | |
1981 | void | |
1982 | spu_expand_prologue (void) | |
1983 | { | |
1984 | HOST_WIDE_INT size = get_frame_size (), offset, regno; | |
1985 | HOST_WIDE_INT total_size; | |
1986 | HOST_WIDE_INT saved_regs_size; | |
1987 | rtx sp_reg = gen_rtx_REG (Pmode, STACK_POINTER_REGNUM); | |
1988 | rtx scratch_reg_0, scratch_reg_1; | |
1989 | rtx insn, real; | |
1990 | ||
644459d0 | 1991 | if (flag_pic && optimize == 0) |
18d50ae6 | 1992 | crtl->uses_pic_offset_table = 1; |
644459d0 | 1993 | |
1994 | if (spu_naked_function_p (current_function_decl)) | |
1995 | return; | |
1996 | ||
1997 | scratch_reg_0 = gen_rtx_REG (SImode, LAST_ARG_REGNUM + 1); | |
1998 | scratch_reg_1 = gen_rtx_REG (SImode, LAST_ARG_REGNUM + 2); | |
1999 | ||
2000 | saved_regs_size = spu_saved_regs_size (); | |
2001 | total_size = size + saved_regs_size | |
abe32cce | 2002 | + crtl->outgoing_args_size |
2003 | + crtl->args.pretend_args_size; | |
644459d0 | 2004 | |
2005 | if (!current_function_is_leaf | |
18d50ae6 | 2006 | || cfun->calls_alloca || total_size > 0) |
644459d0 | 2007 | total_size += STACK_POINTER_OFFSET; |
2008 | ||
2009 | /* Save this first because code after this might use the link | |
2010 | register as a scratch register. */ | |
2011 | if (!current_function_is_leaf) | |
2012 | { | |
2013 | insn = frame_emit_store (LINK_REGISTER_REGNUM, sp_reg, 16); | |
2014 | RTX_FRAME_RELATED_P (insn) = 1; | |
2015 | } | |
2016 | ||
2017 | if (total_size > 0) | |
2018 | { | |
abe32cce | 2019 | offset = -crtl->args.pretend_args_size; |
644459d0 | 2020 | for (regno = 0; regno < FIRST_PSEUDO_REGISTER; ++regno) |
2021 | if (need_to_save_reg (regno, 1)) | |
2022 | { | |
2023 | offset -= 16; | |
2024 | insn = frame_emit_store (regno, sp_reg, offset); | |
2025 | RTX_FRAME_RELATED_P (insn) = 1; | |
2026 | } | |
2027 | } | |
2028 | ||
18d50ae6 | 2029 | if (flag_pic && crtl->uses_pic_offset_table) |
644459d0 | 2030 | { |
2031 | rtx pic_reg = get_pic_reg (); | |
2032 | insn = emit_insn (gen_load_pic_offset (pic_reg, scratch_reg_0)); | |
644459d0 | 2033 | insn = emit_insn (gen_subsi3 (pic_reg, pic_reg, scratch_reg_0)); |
644459d0 | 2034 | } |
2035 | ||
2036 | if (total_size > 0) | |
2037 | { | |
2038 | if (flag_stack_check) | |
2039 | { | |
d819917f | 2040 | /* We compare against total_size-1 because |
644459d0 | 2041 | ($sp >= total_size) <=> ($sp > total_size-1) */ |
2042 | rtx scratch_v4si = gen_rtx_REG (V4SImode, REGNO (scratch_reg_0)); | |
2043 | rtx sp_v4si = gen_rtx_REG (V4SImode, STACK_POINTER_REGNUM); | |
2044 | rtx size_v4si = spu_const (V4SImode, total_size - 1); | |
2045 | if (!satisfies_constraint_K (GEN_INT (total_size - 1))) | |
2046 | { | |
2047 | emit_move_insn (scratch_v4si, size_v4si); | |
2048 | size_v4si = scratch_v4si; | |
2049 | } | |
2050 | emit_insn (gen_cgt_v4si (scratch_v4si, sp_v4si, size_v4si)); | |
2051 | emit_insn (gen_vec_extractv4si | |
2052 | (scratch_reg_0, scratch_v4si, GEN_INT (1))); | |
2053 | emit_insn (gen_spu_heq (scratch_reg_0, GEN_INT (0))); | |
2054 | } | |
2055 | ||
2056 | /* Adjust the stack pointer, and make sure scratch_reg_0 contains | |
2057 | the value of the previous $sp because we save it as the back | |
2058 | chain. */ | |
2059 | if (total_size <= 2000) | |
2060 | { | |
2061 | /* In this case we save the back chain first. */ | |
2062 | insn = frame_emit_store (STACK_POINTER_REGNUM, sp_reg, -total_size); | |
644459d0 | 2063 | insn = |
2064 | frame_emit_add_imm (sp_reg, sp_reg, -total_size, scratch_reg_0); | |
2065 | } | |
644459d0 | 2066 | else |
2067 | { | |
2068 | insn = emit_move_insn (scratch_reg_0, sp_reg); | |
644459d0 | 2069 | insn = |
2070 | frame_emit_add_imm (sp_reg, sp_reg, -total_size, scratch_reg_1); | |
2071 | } | |
2072 | RTX_FRAME_RELATED_P (insn) = 1; | |
2073 | real = gen_addsi3 (sp_reg, sp_reg, GEN_INT (-total_size)); | |
b9c74b4d | 2074 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, real); |
644459d0 | 2075 | |
2076 | if (total_size > 2000) | |
2077 | { | |
2078 | /* Save the back chain ptr */ | |
2079 | insn = frame_emit_store (REGNO (scratch_reg_0), sp_reg, 0); | |
644459d0 | 2080 | } |
2081 | ||
2082 | if (frame_pointer_needed) | |
2083 | { | |
2084 | rtx fp_reg = gen_rtx_REG (Pmode, HARD_FRAME_POINTER_REGNUM); | |
2085 | HOST_WIDE_INT fp_offset = STACK_POINTER_OFFSET | |
abe32cce | 2086 | + crtl->outgoing_args_size; |
644459d0 | 2087 | /* Set the new frame_pointer */ |
d8dfeb55 | 2088 | insn = frame_emit_add_imm (fp_reg, sp_reg, fp_offset, scratch_reg_0); |
2089 | RTX_FRAME_RELATED_P (insn) = 1; | |
2090 | real = gen_addsi3 (fp_reg, sp_reg, GEN_INT (fp_offset)); | |
b9c74b4d | 2091 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, real); |
5df189be | 2092 | REGNO_POINTER_ALIGN (HARD_FRAME_POINTER_REGNUM) = STACK_BOUNDARY; |
644459d0 | 2093 | } |
2094 | } | |
2095 | ||
a512540d | 2096 | if (flag_stack_usage) |
2097 | current_function_static_stack_size = total_size; | |
644459d0 | 2098 | } |
2099 | ||
2100 | void | |
2101 | spu_expand_epilogue (bool sibcall_p) | |
2102 | { | |
2103 | int size = get_frame_size (), offset, regno; | |
2104 | HOST_WIDE_INT saved_regs_size, total_size; | |
2105 | rtx sp_reg = gen_rtx_REG (Pmode, STACK_POINTER_REGNUM); | |
2106 | rtx jump, scratch_reg_0; | |
2107 | ||
644459d0 | 2108 | if (spu_naked_function_p (current_function_decl)) |
2109 | return; | |
2110 | ||
2111 | scratch_reg_0 = gen_rtx_REG (SImode, LAST_ARG_REGNUM + 1); | |
2112 | ||
2113 | saved_regs_size = spu_saved_regs_size (); | |
2114 | total_size = size + saved_regs_size | |
abe32cce | 2115 | + crtl->outgoing_args_size |
2116 | + crtl->args.pretend_args_size; | |
644459d0 | 2117 | |
2118 | if (!current_function_is_leaf | |
18d50ae6 | 2119 | || cfun->calls_alloca || total_size > 0) |
644459d0 | 2120 | total_size += STACK_POINTER_OFFSET; |
2121 | ||
2122 | if (total_size > 0) | |
2123 | { | |
18d50ae6 | 2124 | if (cfun->calls_alloca) |
644459d0 | 2125 | frame_emit_load (STACK_POINTER_REGNUM, sp_reg, 0); |
2126 | else | |
2127 | frame_emit_add_imm (sp_reg, sp_reg, total_size, scratch_reg_0); | |
2128 | ||
2129 | ||
2130 | if (saved_regs_size > 0) | |
2131 | { | |
abe32cce | 2132 | offset = -crtl->args.pretend_args_size; |
644459d0 | 2133 | for (regno = 0; regno < FIRST_PSEUDO_REGISTER; ++regno) |
2134 | if (need_to_save_reg (regno, 1)) | |
2135 | { | |
2136 | offset -= 0x10; | |
2137 | frame_emit_load (regno, sp_reg, offset); | |
2138 | } | |
2139 | } | |
2140 | } | |
2141 | ||
2142 | if (!current_function_is_leaf) | |
2143 | frame_emit_load (LINK_REGISTER_REGNUM, sp_reg, 16); | |
2144 | ||
2145 | if (!sibcall_p) | |
2146 | { | |
18b42941 | 2147 | emit_use (gen_rtx_REG (SImode, LINK_REGISTER_REGNUM)); |
644459d0 | 2148 | jump = emit_jump_insn (gen__return ()); |
2149 | emit_barrier_after (jump); | |
2150 | } | |
2151 | ||
644459d0 | 2152 | } |
2153 | ||
2154 | rtx | |
2155 | spu_return_addr (int count, rtx frame ATTRIBUTE_UNUSED) | |
2156 | { | |
2157 | if (count != 0) | |
2158 | return 0; | |
2159 | /* This is inefficient because it ends up copying to a save-register | |
2160 | which then gets saved even though $lr has already been saved. But | |
2161 | it does generate better code for leaf functions and we don't need | |
2162 | to use RETURN_ADDRESS_POINTER_REGNUM to get it working. It's only | |
2163 | used for __builtin_return_address anyway, so maybe we don't care if | |
2164 | it's inefficient. */ | |
2165 | return get_hard_reg_initial_val (Pmode, LINK_REGISTER_REGNUM); | |
2166 | } | |
2167 | \f | |
2168 | ||
2169 | /* Given VAL, generate a constant appropriate for MODE. | |
2170 | If MODE is a vector mode, every element will be VAL. | |
2171 | For TImode, VAL will be zero extended to 128 bits. */ | |
2172 | rtx | |
2173 | spu_const (enum machine_mode mode, HOST_WIDE_INT val) | |
2174 | { | |
2175 | rtx inner; | |
2176 | rtvec v; | |
2177 | int units, i; | |
2178 | ||
2179 | gcc_assert (GET_MODE_CLASS (mode) == MODE_INT | |
2180 | || GET_MODE_CLASS (mode) == MODE_FLOAT | |
2181 | || GET_MODE_CLASS (mode) == MODE_VECTOR_INT | |
2182 | || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT); | |
2183 | ||
2184 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
2185 | return immed_double_const (val, 0, mode); | |
2186 | ||
2187 | /* val is the bit representation of the float */ | |
2188 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
2189 | return hwint_to_const_double (mode, val); | |
2190 | ||
2191 | if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT) | |
2192 | inner = immed_double_const (val, 0, GET_MODE_INNER (mode)); | |
2193 | else | |
2194 | inner = hwint_to_const_double (GET_MODE_INNER (mode), val); | |
2195 | ||
2196 | units = GET_MODE_NUNITS (mode); | |
2197 | ||
2198 | v = rtvec_alloc (units); | |
2199 | ||
2200 | for (i = 0; i < units; ++i) | |
2201 | RTVEC_ELT (v, i) = inner; | |
2202 | ||
2203 | return gen_rtx_CONST_VECTOR (mode, v); | |
2204 | } | |
644459d0 | 2205 | |
5474166e | 2206 | /* Create a MODE vector constant from 4 ints. */ |
2207 | rtx | |
2208 | spu_const_from_ints(enum machine_mode mode, int a, int b, int c, int d) | |
2209 | { | |
2210 | unsigned char arr[16]; | |
2211 | arr[0] = (a >> 24) & 0xff; | |
2212 | arr[1] = (a >> 16) & 0xff; | |
2213 | arr[2] = (a >> 8) & 0xff; | |
2214 | arr[3] = (a >> 0) & 0xff; | |
2215 | arr[4] = (b >> 24) & 0xff; | |
2216 | arr[5] = (b >> 16) & 0xff; | |
2217 | arr[6] = (b >> 8) & 0xff; | |
2218 | arr[7] = (b >> 0) & 0xff; | |
2219 | arr[8] = (c >> 24) & 0xff; | |
2220 | arr[9] = (c >> 16) & 0xff; | |
2221 | arr[10] = (c >> 8) & 0xff; | |
2222 | arr[11] = (c >> 0) & 0xff; | |
2223 | arr[12] = (d >> 24) & 0xff; | |
2224 | arr[13] = (d >> 16) & 0xff; | |
2225 | arr[14] = (d >> 8) & 0xff; | |
2226 | arr[15] = (d >> 0) & 0xff; | |
2227 | return array_to_constant(mode, arr); | |
2228 | } | |
5a976006 | 2229 | \f |
2230 | /* branch hint stuff */ | |
5474166e | 2231 | |
644459d0 | 2232 | /* An array of these is used to propagate hints to predecessor blocks. */ |
2233 | struct spu_bb_info | |
2234 | { | |
5a976006 | 2235 | rtx prop_jump; /* propagated from another block */ |
2236 | int bb_index; /* the original block. */ | |
644459d0 | 2237 | }; |
5a976006 | 2238 | static struct spu_bb_info *spu_bb_info; |
644459d0 | 2239 | |
5a976006 | 2240 | #define STOP_HINT_P(INSN) \ |
2241 | (GET_CODE(INSN) == CALL_INSN \ | |
2242 | || INSN_CODE(INSN) == CODE_FOR_divmodsi4 \ | |
2243 | || INSN_CODE(INSN) == CODE_FOR_udivmodsi4) | |
2244 | ||
2245 | /* 1 when RTX is a hinted branch or its target. We keep track of | |
2246 | what has been hinted so the safe-hint code can test it easily. */ | |
2247 | #define HINTED_P(RTX) \ | |
2248 | (RTL_FLAG_CHECK3("HINTED_P", (RTX), CODE_LABEL, JUMP_INSN, CALL_INSN)->unchanging) | |
2249 | ||
2250 | /* 1 when RTX is an insn that must be scheduled on an even boundary. */ | |
2251 | #define SCHED_ON_EVEN_P(RTX) \ | |
2252 | (RTL_FLAG_CHECK2("SCHED_ON_EVEN_P", (RTX), JUMP_INSN, CALL_INSN)->in_struct) | |
2253 | ||
2254 | /* Emit a nop for INSN such that the two will dual issue. This assumes | |
2255 | INSN is 8-byte aligned. When INSN is inline asm we emit an lnop. | |
2256 | We check for TImode to handle a MULTI1 insn which has dual issued its | |
2257 | first instruction. get_pipe returns -1 for MULTI0, inline asm, or | |
2258 | ADDR_VEC insns. */ | |
2259 | static void | |
2260 | emit_nop_for_insn (rtx insn) | |
644459d0 | 2261 | { |
5a976006 | 2262 | int p; |
2263 | rtx new_insn; | |
2264 | p = get_pipe (insn); | |
2265 | if ((CALL_P (insn) || JUMP_P (insn)) && SCHED_ON_EVEN_P (insn)) | |
2266 | new_insn = emit_insn_after (gen_lnop (), insn); | |
2267 | else if (p == 1 && GET_MODE (insn) == TImode) | |
644459d0 | 2268 | { |
5a976006 | 2269 | new_insn = emit_insn_before (gen_nopn (GEN_INT (127)), insn); |
2270 | PUT_MODE (new_insn, TImode); | |
2271 | PUT_MODE (insn, VOIDmode); | |
2272 | } | |
2273 | else | |
2274 | new_insn = emit_insn_after (gen_lnop (), insn); | |
2275 | recog_memoized (new_insn); | |
2fbdf9ef | 2276 | INSN_LOCATOR (new_insn) = INSN_LOCATOR (insn); |
5a976006 | 2277 | } |
2278 | ||
2279 | /* Insert nops in basic blocks to meet dual issue alignment | |
2280 | requirements. Also make sure hbrp and hint instructions are at least | |
2281 | one cycle apart, possibly inserting a nop. */ | |
2282 | static void | |
2283 | pad_bb(void) | |
2284 | { | |
2285 | rtx insn, next_insn, prev_insn, hbr_insn = 0; | |
2286 | int length; | |
2287 | int addr; | |
2288 | ||
2289 | /* This sets up INSN_ADDRESSES. */ | |
2290 | shorten_branches (get_insns ()); | |
2291 | ||
2292 | /* Keep track of length added by nops. */ | |
2293 | length = 0; | |
2294 | ||
2295 | prev_insn = 0; | |
2296 | insn = get_insns (); | |
2297 | if (!active_insn_p (insn)) | |
2298 | insn = next_active_insn (insn); | |
2299 | for (; insn; insn = next_insn) | |
2300 | { | |
2301 | next_insn = next_active_insn (insn); | |
2302 | if (INSN_CODE (insn) == CODE_FOR_iprefetch | |
2303 | || INSN_CODE (insn) == CODE_FOR_hbr) | |
644459d0 | 2304 | { |
5a976006 | 2305 | if (hbr_insn) |
2306 | { | |
2307 | int a0 = INSN_ADDRESSES (INSN_UID (hbr_insn)); | |
2308 | int a1 = INSN_ADDRESSES (INSN_UID (insn)); | |
2309 | if ((a1 - a0 == 8 && GET_MODE (insn) != TImode) | |
2310 | || (a1 - a0 == 4)) | |
2311 | { | |
2312 | prev_insn = emit_insn_before (gen_lnop (), insn); | |
2313 | PUT_MODE (prev_insn, GET_MODE (insn)); | |
2314 | PUT_MODE (insn, TImode); | |
2fbdf9ef | 2315 | INSN_LOCATOR (prev_insn) = INSN_LOCATOR (insn); |
5a976006 | 2316 | length += 4; |
2317 | } | |
2318 | } | |
2319 | hbr_insn = insn; | |
2320 | } | |
2321 | if (INSN_CODE (insn) == CODE_FOR_blockage) | |
2322 | { | |
2323 | if (GET_MODE (insn) == TImode) | |
2324 | PUT_MODE (next_insn, TImode); | |
2325 | insn = next_insn; | |
2326 | next_insn = next_active_insn (insn); | |
2327 | } | |
2328 | addr = INSN_ADDRESSES (INSN_UID (insn)); | |
2329 | if ((CALL_P (insn) || JUMP_P (insn)) && SCHED_ON_EVEN_P (insn)) | |
2330 | { | |
2331 | if (((addr + length) & 7) != 0) | |
2332 | { | |
2333 | emit_nop_for_insn (prev_insn); | |
2334 | length += 4; | |
2335 | } | |
644459d0 | 2336 | } |
5a976006 | 2337 | else if (GET_MODE (insn) == TImode |
2338 | && ((next_insn && GET_MODE (next_insn) != TImode) | |
2339 | || get_attr_type (insn) == TYPE_MULTI0) | |
2340 | && ((addr + length) & 7) != 0) | |
2341 | { | |
2342 | /* prev_insn will always be set because the first insn is | |
2343 | always 8-byte aligned. */ | |
2344 | emit_nop_for_insn (prev_insn); | |
2345 | length += 4; | |
2346 | } | |
2347 | prev_insn = insn; | |
644459d0 | 2348 | } |
644459d0 | 2349 | } |
2350 | ||
5a976006 | 2351 | \f |
2352 | /* Routines for branch hints. */ | |
2353 | ||
644459d0 | 2354 | static void |
5a976006 | 2355 | spu_emit_branch_hint (rtx before, rtx branch, rtx target, |
2356 | int distance, sbitmap blocks) | |
644459d0 | 2357 | { |
5a976006 | 2358 | rtx branch_label = 0; |
2359 | rtx hint; | |
2360 | rtx insn; | |
2361 | rtx table; | |
644459d0 | 2362 | |
2363 | if (before == 0 || branch == 0 || target == 0) | |
2364 | return; | |
2365 | ||
5a976006 | 2366 | /* While scheduling we require hints to be no further than 600, so |
2367 | we need to enforce that here too */ | |
644459d0 | 2368 | if (distance > 600) |
2369 | return; | |
2370 | ||
5a976006 | 2371 | /* If we have a Basic block note, emit it after the basic block note. */ |
37534923 | 2372 | if (NOTE_INSN_BASIC_BLOCK_P (before)) |
5a976006 | 2373 | before = NEXT_INSN (before); |
644459d0 | 2374 | |
2375 | branch_label = gen_label_rtx (); | |
2376 | LABEL_NUSES (branch_label)++; | |
2377 | LABEL_PRESERVE_P (branch_label) = 1; | |
2378 | insn = emit_label_before (branch_label, branch); | |
2379 | branch_label = gen_rtx_LABEL_REF (VOIDmode, branch_label); | |
5a976006 | 2380 | SET_BIT (blocks, BLOCK_FOR_INSN (branch)->index); |
2381 | ||
2382 | hint = emit_insn_before (gen_hbr (branch_label, target), before); | |
2383 | recog_memoized (hint); | |
2fbdf9ef | 2384 | INSN_LOCATOR (hint) = INSN_LOCATOR (branch); |
5a976006 | 2385 | HINTED_P (branch) = 1; |
644459d0 | 2386 | |
5a976006 | 2387 | if (GET_CODE (target) == LABEL_REF) |
2388 | HINTED_P (XEXP (target, 0)) = 1; | |
2389 | else if (tablejump_p (branch, 0, &table)) | |
644459d0 | 2390 | { |
5a976006 | 2391 | rtvec vec; |
2392 | int j; | |
2393 | if (GET_CODE (PATTERN (table)) == ADDR_VEC) | |
2394 | vec = XVEC (PATTERN (table), 0); | |
2395 | else | |
2396 | vec = XVEC (PATTERN (table), 1); | |
2397 | for (j = GET_NUM_ELEM (vec) - 1; j >= 0; --j) | |
2398 | HINTED_P (XEXP (RTVEC_ELT (vec, j), 0)) = 1; | |
644459d0 | 2399 | } |
5a976006 | 2400 | |
2401 | if (distance >= 588) | |
644459d0 | 2402 | { |
5a976006 | 2403 | /* Make sure the hint isn't scheduled any earlier than this point, |
2404 | which could make it too far for the branch offest to fit */ | |
2fbdf9ef | 2405 | insn = emit_insn_before (gen_blockage (), hint); |
2406 | recog_memoized (insn); | |
2407 | INSN_LOCATOR (insn) = INSN_LOCATOR (hint); | |
5a976006 | 2408 | } |
2409 | else if (distance <= 8 * 4) | |
2410 | { | |
2411 | /* To guarantee at least 8 insns between the hint and branch we | |
2412 | insert nops. */ | |
2413 | int d; | |
2414 | for (d = distance; d < 8 * 4; d += 4) | |
2415 | { | |
2416 | insn = | |
2417 | emit_insn_after (gen_nopn_nv (gen_rtx_REG (SImode, 127)), hint); | |
2418 | recog_memoized (insn); | |
2fbdf9ef | 2419 | INSN_LOCATOR (insn) = INSN_LOCATOR (hint); |
5a976006 | 2420 | } |
2421 | ||
2422 | /* Make sure any nops inserted aren't scheduled before the hint. */ | |
2fbdf9ef | 2423 | insn = emit_insn_after (gen_blockage (), hint); |
2424 | recog_memoized (insn); | |
2425 | INSN_LOCATOR (insn) = INSN_LOCATOR (hint); | |
5a976006 | 2426 | |
2427 | /* Make sure any nops inserted aren't scheduled after the call. */ | |
2428 | if (CALL_P (branch) && distance < 8 * 4) | |
2fbdf9ef | 2429 | { |
2430 | insn = emit_insn_before (gen_blockage (), branch); | |
2431 | recog_memoized (insn); | |
2432 | INSN_LOCATOR (insn) = INSN_LOCATOR (branch); | |
2433 | } | |
644459d0 | 2434 | } |
644459d0 | 2435 | } |
2436 | ||
2437 | /* Returns 0 if we don't want a hint for this branch. Otherwise return | |
2438 | the rtx for the branch target. */ | |
2439 | static rtx | |
2440 | get_branch_target (rtx branch) | |
2441 | { | |
2442 | if (GET_CODE (branch) == JUMP_INSN) | |
2443 | { | |
2444 | rtx set, src; | |
2445 | ||
2446 | /* Return statements */ | |
2447 | if (GET_CODE (PATTERN (branch)) == RETURN) | |
2448 | return gen_rtx_REG (SImode, LINK_REGISTER_REGNUM); | |
2449 | ||
2450 | /* jump table */ | |
2451 | if (GET_CODE (PATTERN (branch)) == ADDR_VEC | |
2452 | || GET_CODE (PATTERN (branch)) == ADDR_DIFF_VEC) | |
2453 | return 0; | |
2454 | ||
fcc31b99 | 2455 | /* ASM GOTOs. */ |
604157f6 | 2456 | if (extract_asm_operands (PATTERN (branch)) != NULL) |
fcc31b99 | 2457 | return NULL; |
2458 | ||
644459d0 | 2459 | set = single_set (branch); |
2460 | src = SET_SRC (set); | |
2461 | if (GET_CODE (SET_DEST (set)) != PC) | |
2462 | abort (); | |
2463 | ||
2464 | if (GET_CODE (src) == IF_THEN_ELSE) | |
2465 | { | |
2466 | rtx lab = 0; | |
2467 | rtx note = find_reg_note (branch, REG_BR_PROB, 0); | |
2468 | if (note) | |
2469 | { | |
2470 | /* If the more probable case is not a fall through, then | |
2471 | try a branch hint. */ | |
2472 | HOST_WIDE_INT prob = INTVAL (XEXP (note, 0)); | |
2473 | if (prob > (REG_BR_PROB_BASE * 6 / 10) | |
2474 | && GET_CODE (XEXP (src, 1)) != PC) | |
2475 | lab = XEXP (src, 1); | |
2476 | else if (prob < (REG_BR_PROB_BASE * 4 / 10) | |
2477 | && GET_CODE (XEXP (src, 2)) != PC) | |
2478 | lab = XEXP (src, 2); | |
2479 | } | |
2480 | if (lab) | |
2481 | { | |
2482 | if (GET_CODE (lab) == RETURN) | |
2483 | return gen_rtx_REG (SImode, LINK_REGISTER_REGNUM); | |
2484 | return lab; | |
2485 | } | |
2486 | return 0; | |
2487 | } | |
2488 | ||
2489 | return src; | |
2490 | } | |
2491 | else if (GET_CODE (branch) == CALL_INSN) | |
2492 | { | |
2493 | rtx call; | |
2494 | /* All of our call patterns are in a PARALLEL and the CALL is | |
2495 | the first pattern in the PARALLEL. */ | |
2496 | if (GET_CODE (PATTERN (branch)) != PARALLEL) | |
2497 | abort (); | |
2498 | call = XVECEXP (PATTERN (branch), 0, 0); | |
2499 | if (GET_CODE (call) == SET) | |
2500 | call = SET_SRC (call); | |
2501 | if (GET_CODE (call) != CALL) | |
2502 | abort (); | |
2503 | return XEXP (XEXP (call, 0), 0); | |
2504 | } | |
2505 | return 0; | |
2506 | } | |
2507 | ||
5a976006 | 2508 | /* The special $hbr register is used to prevent the insn scheduler from |
2509 | moving hbr insns across instructions which invalidate them. It | |
2510 | should only be used in a clobber, and this function searches for | |
2511 | insns which clobber it. */ | |
2512 | static bool | |
2513 | insn_clobbers_hbr (rtx insn) | |
2514 | { | |
2515 | if (INSN_P (insn) | |
2516 | && GET_CODE (PATTERN (insn)) == PARALLEL) | |
2517 | { | |
2518 | rtx parallel = PATTERN (insn); | |
2519 | rtx clobber; | |
2520 | int j; | |
2521 | for (j = XVECLEN (parallel, 0) - 1; j >= 0; j--) | |
2522 | { | |
2523 | clobber = XVECEXP (parallel, 0, j); | |
2524 | if (GET_CODE (clobber) == CLOBBER | |
2525 | && GET_CODE (XEXP (clobber, 0)) == REG | |
2526 | && REGNO (XEXP (clobber, 0)) == HBR_REGNUM) | |
2527 | return 1; | |
2528 | } | |
2529 | } | |
2530 | return 0; | |
2531 | } | |
2532 | ||
2533 | /* Search up to 32 insns starting at FIRST: | |
2534 | - at any kind of hinted branch, just return | |
2535 | - at any unconditional branch in the first 15 insns, just return | |
2536 | - at a call or indirect branch, after the first 15 insns, force it to | |
2537 | an even address and return | |
2538 | - at any unconditional branch, after the first 15 insns, force it to | |
2539 | an even address. | |
2540 | At then end of the search, insert an hbrp within 4 insns of FIRST, | |
2541 | and an hbrp within 16 instructions of FIRST. | |
2542 | */ | |
644459d0 | 2543 | static void |
5a976006 | 2544 | insert_hbrp_for_ilb_runout (rtx first) |
644459d0 | 2545 | { |
5a976006 | 2546 | rtx insn, before_4 = 0, before_16 = 0; |
2547 | int addr = 0, length, first_addr = -1; | |
2548 | int hbrp_addr0 = 128 * 4, hbrp_addr1 = 128 * 4; | |
2549 | int insert_lnop_after = 0; | |
2550 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
2551 | if (INSN_P (insn)) | |
2552 | { | |
2553 | if (first_addr == -1) | |
2554 | first_addr = INSN_ADDRESSES (INSN_UID (insn)); | |
2555 | addr = INSN_ADDRESSES (INSN_UID (insn)) - first_addr; | |
2556 | length = get_attr_length (insn); | |
2557 | ||
2558 | if (before_4 == 0 && addr + length >= 4 * 4) | |
2559 | before_4 = insn; | |
2560 | /* We test for 14 instructions because the first hbrp will add | |
2561 | up to 2 instructions. */ | |
2562 | if (before_16 == 0 && addr + length >= 14 * 4) | |
2563 | before_16 = insn; | |
2564 | ||
2565 | if (INSN_CODE (insn) == CODE_FOR_hbr) | |
2566 | { | |
2567 | /* Make sure an hbrp is at least 2 cycles away from a hint. | |
2568 | Insert an lnop after the hbrp when necessary. */ | |
2569 | if (before_4 == 0 && addr > 0) | |
2570 | { | |
2571 | before_4 = insn; | |
2572 | insert_lnop_after |= 1; | |
2573 | } | |
2574 | else if (before_4 && addr <= 4 * 4) | |
2575 | insert_lnop_after |= 1; | |
2576 | if (before_16 == 0 && addr > 10 * 4) | |
2577 | { | |
2578 | before_16 = insn; | |
2579 | insert_lnop_after |= 2; | |
2580 | } | |
2581 | else if (before_16 && addr <= 14 * 4) | |
2582 | insert_lnop_after |= 2; | |
2583 | } | |
644459d0 | 2584 | |
5a976006 | 2585 | if (INSN_CODE (insn) == CODE_FOR_iprefetch) |
2586 | { | |
2587 | if (addr < hbrp_addr0) | |
2588 | hbrp_addr0 = addr; | |
2589 | else if (addr < hbrp_addr1) | |
2590 | hbrp_addr1 = addr; | |
2591 | } | |
644459d0 | 2592 | |
5a976006 | 2593 | if (CALL_P (insn) || JUMP_P (insn)) |
2594 | { | |
2595 | if (HINTED_P (insn)) | |
2596 | return; | |
2597 | ||
2598 | /* Any branch after the first 15 insns should be on an even | |
2599 | address to avoid a special case branch. There might be | |
2600 | some nops and/or hbrps inserted, so we test after 10 | |
2601 | insns. */ | |
2602 | if (addr > 10 * 4) | |
2603 | SCHED_ON_EVEN_P (insn) = 1; | |
2604 | } | |
644459d0 | 2605 | |
5a976006 | 2606 | if (CALL_P (insn) || tablejump_p (insn, 0, 0)) |
2607 | return; | |
2608 | ||
2609 | ||
2610 | if (addr + length >= 32 * 4) | |
644459d0 | 2611 | { |
5a976006 | 2612 | gcc_assert (before_4 && before_16); |
2613 | if (hbrp_addr0 > 4 * 4) | |
644459d0 | 2614 | { |
5a976006 | 2615 | insn = |
2616 | emit_insn_before (gen_iprefetch (GEN_INT (1)), before_4); | |
2617 | recog_memoized (insn); | |
2fbdf9ef | 2618 | INSN_LOCATOR (insn) = INSN_LOCATOR (before_4); |
5a976006 | 2619 | INSN_ADDRESSES_NEW (insn, |
2620 | INSN_ADDRESSES (INSN_UID (before_4))); | |
2621 | PUT_MODE (insn, GET_MODE (before_4)); | |
2622 | PUT_MODE (before_4, TImode); | |
2623 | if (insert_lnop_after & 1) | |
644459d0 | 2624 | { |
5a976006 | 2625 | insn = emit_insn_before (gen_lnop (), before_4); |
2626 | recog_memoized (insn); | |
2fbdf9ef | 2627 | INSN_LOCATOR (insn) = INSN_LOCATOR (before_4); |
5a976006 | 2628 | INSN_ADDRESSES_NEW (insn, |
2629 | INSN_ADDRESSES (INSN_UID (before_4))); | |
2630 | PUT_MODE (insn, TImode); | |
644459d0 | 2631 | } |
644459d0 | 2632 | } |
5a976006 | 2633 | if ((hbrp_addr0 <= 4 * 4 || hbrp_addr0 > 16 * 4) |
2634 | && hbrp_addr1 > 16 * 4) | |
644459d0 | 2635 | { |
5a976006 | 2636 | insn = |
2637 | emit_insn_before (gen_iprefetch (GEN_INT (2)), before_16); | |
2638 | recog_memoized (insn); | |
2fbdf9ef | 2639 | INSN_LOCATOR (insn) = INSN_LOCATOR (before_16); |
5a976006 | 2640 | INSN_ADDRESSES_NEW (insn, |
2641 | INSN_ADDRESSES (INSN_UID (before_16))); | |
2642 | PUT_MODE (insn, GET_MODE (before_16)); | |
2643 | PUT_MODE (before_16, TImode); | |
2644 | if (insert_lnop_after & 2) | |
644459d0 | 2645 | { |
5a976006 | 2646 | insn = emit_insn_before (gen_lnop (), before_16); |
2647 | recog_memoized (insn); | |
2fbdf9ef | 2648 | INSN_LOCATOR (insn) = INSN_LOCATOR (before_16); |
5a976006 | 2649 | INSN_ADDRESSES_NEW (insn, |
2650 | INSN_ADDRESSES (INSN_UID | |
2651 | (before_16))); | |
2652 | PUT_MODE (insn, TImode); | |
644459d0 | 2653 | } |
2654 | } | |
5a976006 | 2655 | return; |
644459d0 | 2656 | } |
644459d0 | 2657 | } |
5a976006 | 2658 | else if (BARRIER_P (insn)) |
2659 | return; | |
644459d0 | 2660 | |
644459d0 | 2661 | } |
5a976006 | 2662 | |
2663 | /* The SPU might hang when it executes 48 inline instructions after a | |
2664 | hinted branch jumps to its hinted target. The beginning of a | |
851d9296 | 2665 | function and the return from a call might have been hinted, and |
2666 | must be handled as well. To prevent a hang we insert 2 hbrps. The | |
2667 | first should be within 6 insns of the branch target. The second | |
2668 | should be within 22 insns of the branch target. When determining | |
2669 | if hbrps are necessary, we look for only 32 inline instructions, | |
2670 | because up to 12 nops and 4 hbrps could be inserted. Similarily, | |
2671 | when inserting new hbrps, we insert them within 4 and 16 insns of | |
2672 | the target. */ | |
644459d0 | 2673 | static void |
5a976006 | 2674 | insert_hbrp (void) |
644459d0 | 2675 | { |
5a976006 | 2676 | rtx insn; |
2677 | if (TARGET_SAFE_HINTS) | |
644459d0 | 2678 | { |
5a976006 | 2679 | shorten_branches (get_insns ()); |
2680 | /* Insert hbrp at beginning of function */ | |
2681 | insn = next_active_insn (get_insns ()); | |
2682 | if (insn) | |
2683 | insert_hbrp_for_ilb_runout (insn); | |
2684 | /* Insert hbrp after hinted targets. */ | |
2685 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
2686 | if ((LABEL_P (insn) && HINTED_P (insn)) || CALL_P (insn)) | |
2687 | insert_hbrp_for_ilb_runout (next_active_insn (insn)); | |
644459d0 | 2688 | } |
644459d0 | 2689 | } |
2690 | ||
5a976006 | 2691 | static int in_spu_reorg; |
2692 | ||
2693 | /* Insert branch hints. There are no branch optimizations after this | |
2694 | pass, so it's safe to set our branch hints now. */ | |
644459d0 | 2695 | static void |
5a976006 | 2696 | spu_machine_dependent_reorg (void) |
644459d0 | 2697 | { |
5a976006 | 2698 | sbitmap blocks; |
2699 | basic_block bb; | |
2700 | rtx branch, insn; | |
2701 | rtx branch_target = 0; | |
2702 | int branch_addr = 0, insn_addr, required_dist = 0; | |
2703 | int i; | |
2704 | unsigned int j; | |
644459d0 | 2705 | |
5a976006 | 2706 | if (!TARGET_BRANCH_HINTS || optimize == 0) |
2707 | { | |
2708 | /* We still do it for unoptimized code because an external | |
2709 | function might have hinted a call or return. */ | |
2710 | insert_hbrp (); | |
2711 | pad_bb (); | |
2712 | return; | |
2713 | } | |
644459d0 | 2714 | |
5a976006 | 2715 | blocks = sbitmap_alloc (last_basic_block); |
2716 | sbitmap_zero (blocks); | |
644459d0 | 2717 | |
5a976006 | 2718 | in_spu_reorg = 1; |
2719 | compute_bb_for_insn (); | |
2720 | ||
2721 | compact_blocks (); | |
2722 | ||
2723 | spu_bb_info = | |
2724 | (struct spu_bb_info *) xcalloc (n_basic_blocks, | |
2725 | sizeof (struct spu_bb_info)); | |
2726 | ||
2727 | /* We need exact insn addresses and lengths. */ | |
2728 | shorten_branches (get_insns ()); | |
2729 | ||
2730 | for (i = n_basic_blocks - 1; i >= 0; i--) | |
644459d0 | 2731 | { |
5a976006 | 2732 | bb = BASIC_BLOCK (i); |
2733 | branch = 0; | |
2734 | if (spu_bb_info[i].prop_jump) | |
644459d0 | 2735 | { |
5a976006 | 2736 | branch = spu_bb_info[i].prop_jump; |
2737 | branch_target = get_branch_target (branch); | |
2738 | branch_addr = INSN_ADDRESSES (INSN_UID (branch)); | |
2739 | required_dist = spu_hint_dist; | |
2740 | } | |
2741 | /* Search from end of a block to beginning. In this loop, find | |
2742 | jumps which need a branch and emit them only when: | |
2743 | - it's an indirect branch and we're at the insn which sets | |
2744 | the register | |
2745 | - we're at an insn that will invalidate the hint. e.g., a | |
2746 | call, another hint insn, inline asm that clobbers $hbr, and | |
2747 | some inlined operations (divmodsi4). Don't consider jumps | |
2748 | because they are only at the end of a block and are | |
2749 | considered when we are deciding whether to propagate | |
2750 | - we're getting too far away from the branch. The hbr insns | |
2751 | only have a signed 10 bit offset | |
2752 | We go back as far as possible so the branch will be considered | |
2753 | for propagation when we get to the beginning of the block. */ | |
2754 | for (insn = BB_END (bb); insn; insn = PREV_INSN (insn)) | |
2755 | { | |
2756 | if (INSN_P (insn)) | |
2757 | { | |
2758 | insn_addr = INSN_ADDRESSES (INSN_UID (insn)); | |
2759 | if (branch | |
2760 | && ((GET_CODE (branch_target) == REG | |
2761 | && set_of (branch_target, insn) != NULL_RTX) | |
2762 | || insn_clobbers_hbr (insn) | |
2763 | || branch_addr - insn_addr > 600)) | |
2764 | { | |
2765 | rtx next = NEXT_INSN (insn); | |
2766 | int next_addr = INSN_ADDRESSES (INSN_UID (next)); | |
2767 | if (insn != BB_END (bb) | |
2768 | && branch_addr - next_addr >= required_dist) | |
2769 | { | |
2770 | if (dump_file) | |
2771 | fprintf (dump_file, | |
2772 | "hint for %i in block %i before %i\n", | |
2773 | INSN_UID (branch), bb->index, | |
2774 | INSN_UID (next)); | |
2775 | spu_emit_branch_hint (next, branch, branch_target, | |
2776 | branch_addr - next_addr, blocks); | |
2777 | } | |
2778 | branch = 0; | |
2779 | } | |
2780 | ||
2781 | /* JUMP_P will only be true at the end of a block. When | |
2782 | branch is already set it means we've previously decided | |
2783 | to propagate a hint for that branch into this block. */ | |
2784 | if (CALL_P (insn) || (JUMP_P (insn) && !branch)) | |
2785 | { | |
2786 | branch = 0; | |
2787 | if ((branch_target = get_branch_target (insn))) | |
2788 | { | |
2789 | branch = insn; | |
2790 | branch_addr = insn_addr; | |
2791 | required_dist = spu_hint_dist; | |
2792 | } | |
2793 | } | |
2794 | } | |
2795 | if (insn == BB_HEAD (bb)) | |
2796 | break; | |
2797 | } | |
2798 | ||
2799 | if (branch) | |
2800 | { | |
2801 | /* If we haven't emitted a hint for this branch yet, it might | |
2802 | be profitable to emit it in one of the predecessor blocks, | |
2803 | especially for loops. */ | |
2804 | rtx bbend; | |
2805 | basic_block prev = 0, prop = 0, prev2 = 0; | |
2806 | int loop_exit = 0, simple_loop = 0; | |
2807 | int next_addr = INSN_ADDRESSES (INSN_UID (NEXT_INSN (insn))); | |
2808 | ||
2809 | for (j = 0; j < EDGE_COUNT (bb->preds); j++) | |
2810 | if (EDGE_PRED (bb, j)->flags & EDGE_FALLTHRU) | |
2811 | prev = EDGE_PRED (bb, j)->src; | |
2812 | else | |
2813 | prev2 = EDGE_PRED (bb, j)->src; | |
2814 | ||
2815 | for (j = 0; j < EDGE_COUNT (bb->succs); j++) | |
2816 | if (EDGE_SUCC (bb, j)->flags & EDGE_LOOP_EXIT) | |
2817 | loop_exit = 1; | |
2818 | else if (EDGE_SUCC (bb, j)->dest == bb) | |
2819 | simple_loop = 1; | |
2820 | ||
2821 | /* If this branch is a loop exit then propagate to previous | |
2822 | fallthru block. This catches the cases when it is a simple | |
2823 | loop or when there is an initial branch into the loop. */ | |
2824 | if (prev && (loop_exit || simple_loop) | |
2825 | && prev->loop_depth <= bb->loop_depth) | |
2826 | prop = prev; | |
2827 | ||
2828 | /* If there is only one adjacent predecessor. Don't propagate | |
2829 | outside this loop. This loop_depth test isn't perfect, but | |
2830 | I'm not sure the loop_father member is valid at this point. */ | |
2831 | else if (prev && single_pred_p (bb) | |
2832 | && prev->loop_depth == bb->loop_depth) | |
2833 | prop = prev; | |
2834 | ||
2835 | /* If this is the JOIN block of a simple IF-THEN then | |
2836 | propogate the hint to the HEADER block. */ | |
2837 | else if (prev && prev2 | |
2838 | && EDGE_COUNT (bb->preds) == 2 | |
2839 | && EDGE_COUNT (prev->preds) == 1 | |
2840 | && EDGE_PRED (prev, 0)->src == prev2 | |
2841 | && prev2->loop_depth == bb->loop_depth | |
2842 | && GET_CODE (branch_target) != REG) | |
2843 | prop = prev; | |
2844 | ||
2845 | /* Don't propagate when: | |
2846 | - this is a simple loop and the hint would be too far | |
2847 | - this is not a simple loop and there are 16 insns in | |
2848 | this block already | |
2849 | - the predecessor block ends in a branch that will be | |
2850 | hinted | |
2851 | - the predecessor block ends in an insn that invalidates | |
2852 | the hint */ | |
2853 | if (prop | |
2854 | && prop->index >= 0 | |
2855 | && (bbend = BB_END (prop)) | |
2856 | && branch_addr - INSN_ADDRESSES (INSN_UID (bbend)) < | |
2857 | (simple_loop ? 600 : 16 * 4) && get_branch_target (bbend) == 0 | |
2858 | && (JUMP_P (bbend) || !insn_clobbers_hbr (bbend))) | |
2859 | { | |
2860 | if (dump_file) | |
2861 | fprintf (dump_file, "propagate from %i to %i (loop depth %i) " | |
2862 | "for %i (loop_exit %i simple_loop %i dist %i)\n", | |
2863 | bb->index, prop->index, bb->loop_depth, | |
2864 | INSN_UID (branch), loop_exit, simple_loop, | |
2865 | branch_addr - INSN_ADDRESSES (INSN_UID (bbend))); | |
2866 | ||
2867 | spu_bb_info[prop->index].prop_jump = branch; | |
2868 | spu_bb_info[prop->index].bb_index = i; | |
2869 | } | |
2870 | else if (branch_addr - next_addr >= required_dist) | |
2871 | { | |
2872 | if (dump_file) | |
2873 | fprintf (dump_file, "hint for %i in block %i before %i\n", | |
2874 | INSN_UID (branch), bb->index, | |
2875 | INSN_UID (NEXT_INSN (insn))); | |
2876 | spu_emit_branch_hint (NEXT_INSN (insn), branch, branch_target, | |
2877 | branch_addr - next_addr, blocks); | |
2878 | } | |
2879 | branch = 0; | |
644459d0 | 2880 | } |
644459d0 | 2881 | } |
5a976006 | 2882 | free (spu_bb_info); |
644459d0 | 2883 | |
5a976006 | 2884 | if (!sbitmap_empty_p (blocks)) |
2885 | find_many_sub_basic_blocks (blocks); | |
2886 | ||
2887 | /* We have to schedule to make sure alignment is ok. */ | |
2888 | FOR_EACH_BB (bb) bb->flags &= ~BB_DISABLE_SCHEDULE; | |
2889 | ||
2890 | /* The hints need to be scheduled, so call it again. */ | |
2891 | schedule_insns (); | |
2fbdf9ef | 2892 | df_finish_pass (true); |
5a976006 | 2893 | |
2894 | insert_hbrp (); | |
2895 | ||
2896 | pad_bb (); | |
2897 | ||
8f1d58ad | 2898 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
2899 | if (NONJUMP_INSN_P (insn) && INSN_CODE (insn) == CODE_FOR_hbr) | |
2900 | { | |
2901 | /* Adjust the LABEL_REF in a hint when we have inserted a nop | |
2902 | between its branch label and the branch . We don't move the | |
2903 | label because GCC expects it at the beginning of the block. */ | |
2904 | rtx unspec = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
2905 | rtx label_ref = XVECEXP (unspec, 0, 0); | |
2906 | rtx label = XEXP (label_ref, 0); | |
2907 | rtx branch; | |
2908 | int offset = 0; | |
2909 | for (branch = NEXT_INSN (label); | |
2910 | !JUMP_P (branch) && !CALL_P (branch); | |
2911 | branch = NEXT_INSN (branch)) | |
2912 | if (NONJUMP_INSN_P (branch)) | |
2913 | offset += get_attr_length (branch); | |
2914 | if (offset > 0) | |
2915 | XVECEXP (unspec, 0, 0) = plus_constant (label_ref, offset); | |
2916 | } | |
5a976006 | 2917 | |
2918 | if (spu_flag_var_tracking) | |
644459d0 | 2919 | { |
5a976006 | 2920 | df_analyze (); |
2921 | timevar_push (TV_VAR_TRACKING); | |
2922 | variable_tracking_main (); | |
2923 | timevar_pop (TV_VAR_TRACKING); | |
2924 | df_finish_pass (false); | |
644459d0 | 2925 | } |
5a976006 | 2926 | |
2927 | free_bb_for_insn (); | |
2928 | ||
2929 | in_spu_reorg = 0; | |
644459d0 | 2930 | } |
2931 | \f | |
2932 | ||
2933 | /* Insn scheduling routines, primarily for dual issue. */ | |
2934 | static int | |
2935 | spu_sched_issue_rate (void) | |
2936 | { | |
2937 | return 2; | |
2938 | } | |
2939 | ||
2940 | static int | |
5a976006 | 2941 | uses_ls_unit(rtx insn) |
644459d0 | 2942 | { |
5a976006 | 2943 | rtx set = single_set (insn); |
2944 | if (set != 0 | |
2945 | && (GET_CODE (SET_DEST (set)) == MEM | |
2946 | || GET_CODE (SET_SRC (set)) == MEM)) | |
2947 | return 1; | |
2948 | return 0; | |
644459d0 | 2949 | } |
2950 | ||
2951 | static int | |
2952 | get_pipe (rtx insn) | |
2953 | { | |
2954 | enum attr_type t; | |
2955 | /* Handle inline asm */ | |
2956 | if (INSN_CODE (insn) == -1) | |
2957 | return -1; | |
2958 | t = get_attr_type (insn); | |
2959 | switch (t) | |
2960 | { | |
2961 | case TYPE_CONVERT: | |
2962 | return -2; | |
2963 | case TYPE_MULTI0: | |
2964 | return -1; | |
2965 | ||
2966 | case TYPE_FX2: | |
2967 | case TYPE_FX3: | |
2968 | case TYPE_SPR: | |
2969 | case TYPE_NOP: | |
2970 | case TYPE_FXB: | |
2971 | case TYPE_FPD: | |
2972 | case TYPE_FP6: | |
2973 | case TYPE_FP7: | |
644459d0 | 2974 | return 0; |
2975 | ||
2976 | case TYPE_LNOP: | |
2977 | case TYPE_SHUF: | |
2978 | case TYPE_LOAD: | |
2979 | case TYPE_STORE: | |
2980 | case TYPE_BR: | |
2981 | case TYPE_MULTI1: | |
2982 | case TYPE_HBR: | |
5a976006 | 2983 | case TYPE_IPREFETCH: |
644459d0 | 2984 | return 1; |
2985 | default: | |
2986 | abort (); | |
2987 | } | |
2988 | } | |
2989 | ||
5a976006 | 2990 | |
2991 | /* haifa-sched.c has a static variable that keeps track of the current | |
2992 | cycle. It is passed to spu_sched_reorder, and we record it here for | |
2993 | use by spu_sched_variable_issue. It won't be accurate if the | |
2994 | scheduler updates it's clock_var between the two calls. */ | |
2995 | static int clock_var; | |
2996 | ||
2997 | /* This is used to keep track of insn alignment. Set to 0 at the | |
2998 | beginning of each block and increased by the "length" attr of each | |
2999 | insn scheduled. */ | |
3000 | static int spu_sched_length; | |
3001 | ||
3002 | /* Record when we've issued pipe0 and pipe1 insns so we can reorder the | |
3003 | ready list appropriately in spu_sched_reorder(). */ | |
3004 | static int pipe0_clock; | |
3005 | static int pipe1_clock; | |
3006 | ||
3007 | static int prev_clock_var; | |
3008 | ||
3009 | static int prev_priority; | |
3010 | ||
3011 | /* The SPU needs to load the next ilb sometime during the execution of | |
3012 | the previous ilb. There is a potential conflict if every cycle has a | |
3013 | load or store. To avoid the conflict we make sure the load/store | |
3014 | unit is free for at least one cycle during the execution of insns in | |
3015 | the previous ilb. */ | |
3016 | static int spu_ls_first; | |
3017 | static int prev_ls_clock; | |
3018 | ||
3019 | static void | |
3020 | spu_sched_init_global (FILE *file ATTRIBUTE_UNUSED, int verbose ATTRIBUTE_UNUSED, | |
3021 | int max_ready ATTRIBUTE_UNUSED) | |
3022 | { | |
3023 | spu_sched_length = 0; | |
3024 | } | |
3025 | ||
3026 | static void | |
3027 | spu_sched_init (FILE *file ATTRIBUTE_UNUSED, int verbose ATTRIBUTE_UNUSED, | |
3028 | int max_ready ATTRIBUTE_UNUSED) | |
3029 | { | |
3030 | if (align_labels > 4 || align_loops > 4 || align_jumps > 4) | |
3031 | { | |
3032 | /* When any block might be at least 8-byte aligned, assume they | |
3033 | will all be at least 8-byte aligned to make sure dual issue | |
3034 | works out correctly. */ | |
3035 | spu_sched_length = 0; | |
3036 | } | |
3037 | spu_ls_first = INT_MAX; | |
3038 | clock_var = -1; | |
3039 | prev_ls_clock = -1; | |
3040 | pipe0_clock = -1; | |
3041 | pipe1_clock = -1; | |
3042 | prev_clock_var = -1; | |
3043 | prev_priority = -1; | |
3044 | } | |
3045 | ||
644459d0 | 3046 | static int |
5a976006 | 3047 | spu_sched_variable_issue (FILE *file ATTRIBUTE_UNUSED, |
3048 | int verbose ATTRIBUTE_UNUSED, rtx insn, int more) | |
644459d0 | 3049 | { |
5a976006 | 3050 | int len; |
3051 | int p; | |
644459d0 | 3052 | if (GET_CODE (PATTERN (insn)) == USE |
3053 | || GET_CODE (PATTERN (insn)) == CLOBBER | |
5a976006 | 3054 | || (len = get_attr_length (insn)) == 0) |
3055 | return more; | |
3056 | ||
3057 | spu_sched_length += len; | |
3058 | ||
3059 | /* Reset on inline asm */ | |
3060 | if (INSN_CODE (insn) == -1) | |
3061 | { | |
3062 | spu_ls_first = INT_MAX; | |
3063 | pipe0_clock = -1; | |
3064 | pipe1_clock = -1; | |
3065 | return 0; | |
3066 | } | |
3067 | p = get_pipe (insn); | |
3068 | if (p == 0) | |
3069 | pipe0_clock = clock_var; | |
3070 | else | |
3071 | pipe1_clock = clock_var; | |
3072 | ||
3073 | if (in_spu_reorg) | |
3074 | { | |
3075 | if (clock_var - prev_ls_clock > 1 | |
3076 | || INSN_CODE (insn) == CODE_FOR_iprefetch) | |
3077 | spu_ls_first = INT_MAX; | |
3078 | if (uses_ls_unit (insn)) | |
3079 | { | |
3080 | if (spu_ls_first == INT_MAX) | |
3081 | spu_ls_first = spu_sched_length; | |
3082 | prev_ls_clock = clock_var; | |
3083 | } | |
3084 | ||
3085 | /* The scheduler hasn't inserted the nop, but we will later on. | |
3086 | Include those nops in spu_sched_length. */ | |
3087 | if (prev_clock_var == clock_var && (spu_sched_length & 7)) | |
3088 | spu_sched_length += 4; | |
3089 | prev_clock_var = clock_var; | |
3090 | ||
3091 | /* more is -1 when called from spu_sched_reorder for new insns | |
3092 | that don't have INSN_PRIORITY */ | |
3093 | if (more >= 0) | |
3094 | prev_priority = INSN_PRIORITY (insn); | |
3095 | } | |
3096 | ||
3097 | /* Always try issueing more insns. spu_sched_reorder will decide | |
3098 | when the cycle should be advanced. */ | |
3099 | return 1; | |
3100 | } | |
3101 | ||
3102 | /* This function is called for both TARGET_SCHED_REORDER and | |
3103 | TARGET_SCHED_REORDER2. */ | |
3104 | static int | |
3105 | spu_sched_reorder (FILE *file ATTRIBUTE_UNUSED, int verbose ATTRIBUTE_UNUSED, | |
3106 | rtx *ready, int *nreadyp, int clock) | |
3107 | { | |
3108 | int i, nready = *nreadyp; | |
3109 | int pipe_0, pipe_1, pipe_hbrp, pipe_ls, schedule_i; | |
3110 | rtx insn; | |
3111 | ||
3112 | clock_var = clock; | |
3113 | ||
3114 | if (nready <= 0 || pipe1_clock >= clock) | |
3115 | return 0; | |
3116 | ||
3117 | /* Find any rtl insns that don't generate assembly insns and schedule | |
3118 | them first. */ | |
3119 | for (i = nready - 1; i >= 0; i--) | |
3120 | { | |
3121 | insn = ready[i]; | |
3122 | if (INSN_CODE (insn) == -1 | |
3123 | || INSN_CODE (insn) == CODE_FOR_blockage | |
9d98604b | 3124 | || (INSN_P (insn) && get_attr_length (insn) == 0)) |
5a976006 | 3125 | { |
3126 | ready[i] = ready[nready - 1]; | |
3127 | ready[nready - 1] = insn; | |
3128 | return 1; | |
3129 | } | |
3130 | } | |
3131 | ||
3132 | pipe_0 = pipe_1 = pipe_hbrp = pipe_ls = schedule_i = -1; | |
3133 | for (i = 0; i < nready; i++) | |
3134 | if (INSN_CODE (ready[i]) != -1) | |
3135 | { | |
3136 | insn = ready[i]; | |
3137 | switch (get_attr_type (insn)) | |
3138 | { | |
3139 | default: | |
3140 | case TYPE_MULTI0: | |
3141 | case TYPE_CONVERT: | |
3142 | case TYPE_FX2: | |
3143 | case TYPE_FX3: | |
3144 | case TYPE_SPR: | |
3145 | case TYPE_NOP: | |
3146 | case TYPE_FXB: | |
3147 | case TYPE_FPD: | |
3148 | case TYPE_FP6: | |
3149 | case TYPE_FP7: | |
3150 | pipe_0 = i; | |
3151 | break; | |
3152 | case TYPE_LOAD: | |
3153 | case TYPE_STORE: | |
3154 | pipe_ls = i; | |
3155 | case TYPE_LNOP: | |
3156 | case TYPE_SHUF: | |
3157 | case TYPE_BR: | |
3158 | case TYPE_MULTI1: | |
3159 | case TYPE_HBR: | |
3160 | pipe_1 = i; | |
3161 | break; | |
3162 | case TYPE_IPREFETCH: | |
3163 | pipe_hbrp = i; | |
3164 | break; | |
3165 | } | |
3166 | } | |
3167 | ||
3168 | /* In the first scheduling phase, schedule loads and stores together | |
3169 | to increase the chance they will get merged during postreload CSE. */ | |
3170 | if (!reload_completed && pipe_ls >= 0) | |
3171 | { | |
3172 | insn = ready[pipe_ls]; | |
3173 | ready[pipe_ls] = ready[nready - 1]; | |
3174 | ready[nready - 1] = insn; | |
3175 | return 1; | |
3176 | } | |
3177 | ||
3178 | /* If there is an hbrp ready, prefer it over other pipe 1 insns. */ | |
3179 | if (pipe_hbrp >= 0) | |
3180 | pipe_1 = pipe_hbrp; | |
3181 | ||
3182 | /* When we have loads/stores in every cycle of the last 15 insns and | |
3183 | we are about to schedule another load/store, emit an hbrp insn | |
3184 | instead. */ | |
3185 | if (in_spu_reorg | |
3186 | && spu_sched_length - spu_ls_first >= 4 * 15 | |
3187 | && !(pipe0_clock < clock && pipe_0 >= 0) && pipe_1 == pipe_ls) | |
3188 | { | |
3189 | insn = sched_emit_insn (gen_iprefetch (GEN_INT (3))); | |
3190 | recog_memoized (insn); | |
3191 | if (pipe0_clock < clock) | |
3192 | PUT_MODE (insn, TImode); | |
3193 | spu_sched_variable_issue (file, verbose, insn, -1); | |
3194 | return 0; | |
3195 | } | |
3196 | ||
3197 | /* In general, we want to emit nops to increase dual issue, but dual | |
3198 | issue isn't faster when one of the insns could be scheduled later | |
3199 | without effecting the critical path. We look at INSN_PRIORITY to | |
3200 | make a good guess, but it isn't perfect so -mdual-nops=n can be | |
3201 | used to effect it. */ | |
3202 | if (in_spu_reorg && spu_dual_nops < 10) | |
3203 | { | |
3204 | /* When we are at an even address and we are not issueing nops to | |
3205 | improve scheduling then we need to advance the cycle. */ | |
3206 | if ((spu_sched_length & 7) == 0 && prev_clock_var == clock | |
3207 | && (spu_dual_nops == 0 | |
3208 | || (pipe_1 != -1 | |
3209 | && prev_priority > | |
3210 | INSN_PRIORITY (ready[pipe_1]) + spu_dual_nops))) | |
3211 | return 0; | |
3212 | ||
3213 | /* When at an odd address, schedule the highest priority insn | |
3214 | without considering pipeline. */ | |
3215 | if ((spu_sched_length & 7) == 4 && prev_clock_var != clock | |
3216 | && (spu_dual_nops == 0 | |
3217 | || (prev_priority > | |
3218 | INSN_PRIORITY (ready[nready - 1]) + spu_dual_nops))) | |
3219 | return 1; | |
3220 | } | |
3221 | ||
3222 | ||
3223 | /* We haven't issued a pipe0 insn yet this cycle, if there is a | |
3224 | pipe0 insn in the ready list, schedule it. */ | |
3225 | if (pipe0_clock < clock && pipe_0 >= 0) | |
3226 | schedule_i = pipe_0; | |
3227 | ||
3228 | /* Either we've scheduled a pipe0 insn already or there is no pipe0 | |
3229 | insn to schedule. Put a pipe1 insn at the front of the ready list. */ | |
3230 | else | |
3231 | schedule_i = pipe_1; | |
3232 | ||
3233 | if (schedule_i > -1) | |
3234 | { | |
3235 | insn = ready[schedule_i]; | |
3236 | ready[schedule_i] = ready[nready - 1]; | |
3237 | ready[nready - 1] = insn; | |
3238 | return 1; | |
3239 | } | |
3240 | return 0; | |
644459d0 | 3241 | } |
3242 | ||
3243 | /* INSN is dependent on DEP_INSN. */ | |
3244 | static int | |
5a976006 | 3245 | spu_sched_adjust_cost (rtx insn, rtx link, rtx dep_insn, int cost) |
644459d0 | 3246 | { |
5a976006 | 3247 | rtx set; |
3248 | ||
3249 | /* The blockage pattern is used to prevent instructions from being | |
3250 | moved across it and has no cost. */ | |
3251 | if (INSN_CODE (insn) == CODE_FOR_blockage | |
3252 | || INSN_CODE (dep_insn) == CODE_FOR_blockage) | |
3253 | return 0; | |
3254 | ||
9d98604b | 3255 | if ((INSN_P (insn) && get_attr_length (insn) == 0) |
3256 | || (INSN_P (dep_insn) && get_attr_length (dep_insn) == 0)) | |
5a976006 | 3257 | return 0; |
3258 | ||
3259 | /* Make sure hbrps are spread out. */ | |
3260 | if (INSN_CODE (insn) == CODE_FOR_iprefetch | |
3261 | && INSN_CODE (dep_insn) == CODE_FOR_iprefetch) | |
3262 | return 8; | |
3263 | ||
3264 | /* Make sure hints and hbrps are 2 cycles apart. */ | |
3265 | if ((INSN_CODE (insn) == CODE_FOR_iprefetch | |
3266 | || INSN_CODE (insn) == CODE_FOR_hbr) | |
3267 | && (INSN_CODE (dep_insn) == CODE_FOR_iprefetch | |
3268 | || INSN_CODE (dep_insn) == CODE_FOR_hbr)) | |
3269 | return 2; | |
3270 | ||
3271 | /* An hbrp has no real dependency on other insns. */ | |
3272 | if (INSN_CODE (insn) == CODE_FOR_iprefetch | |
3273 | || INSN_CODE (dep_insn) == CODE_FOR_iprefetch) | |
3274 | return 0; | |
3275 | ||
3276 | /* Assuming that it is unlikely an argument register will be used in | |
3277 | the first cycle of the called function, we reduce the cost for | |
3278 | slightly better scheduling of dep_insn. When not hinted, the | |
3279 | mispredicted branch would hide the cost as well. */ | |
3280 | if (CALL_P (insn)) | |
3281 | { | |
3282 | rtx target = get_branch_target (insn); | |
3283 | if (GET_CODE (target) != REG || !set_of (target, insn)) | |
3284 | return cost - 2; | |
3285 | return cost; | |
3286 | } | |
3287 | ||
3288 | /* And when returning from a function, let's assume the return values | |
3289 | are completed sooner too. */ | |
3290 | if (CALL_P (dep_insn)) | |
644459d0 | 3291 | return cost - 2; |
5a976006 | 3292 | |
3293 | /* Make sure an instruction that loads from the back chain is schedule | |
3294 | away from the return instruction so a hint is more likely to get | |
3295 | issued. */ | |
3296 | if (INSN_CODE (insn) == CODE_FOR__return | |
3297 | && (set = single_set (dep_insn)) | |
3298 | && GET_CODE (SET_DEST (set)) == REG | |
3299 | && REGNO (SET_DEST (set)) == LINK_REGISTER_REGNUM) | |
3300 | return 20; | |
3301 | ||
644459d0 | 3302 | /* The dfa scheduler sets cost to 0 for all anti-dependencies and the |
3303 | scheduler makes every insn in a block anti-dependent on the final | |
3304 | jump_insn. We adjust here so higher cost insns will get scheduled | |
3305 | earlier. */ | |
5a976006 | 3306 | if (JUMP_P (insn) && REG_NOTE_KIND (link) == REG_DEP_ANTI) |
9997bd27 | 3307 | return insn_cost (dep_insn) - 3; |
5a976006 | 3308 | |
644459d0 | 3309 | return cost; |
3310 | } | |
3311 | \f | |
3312 | /* Create a CONST_DOUBLE from a string. */ | |
3313 | struct rtx_def * | |
3314 | spu_float_const (const char *string, enum machine_mode mode) | |
3315 | { | |
3316 | REAL_VALUE_TYPE value; | |
3317 | value = REAL_VALUE_ATOF (string, mode); | |
3318 | return CONST_DOUBLE_FROM_REAL_VALUE (value, mode); | |
3319 | } | |
3320 | ||
644459d0 | 3321 | int |
3322 | spu_constant_address_p (rtx x) | |
3323 | { | |
3324 | return (GET_CODE (x) == LABEL_REF || GET_CODE (x) == SYMBOL_REF | |
3325 | || GET_CODE (x) == CONST_INT || GET_CODE (x) == CONST | |
3326 | || GET_CODE (x) == HIGH); | |
3327 | } | |
3328 | ||
3329 | static enum spu_immediate | |
3330 | which_immediate_load (HOST_WIDE_INT val) | |
3331 | { | |
3332 | gcc_assert (val == trunc_int_for_mode (val, SImode)); | |
3333 | ||
3334 | if (val >= -0x8000 && val <= 0x7fff) | |
3335 | return SPU_IL; | |
3336 | if (val >= 0 && val <= 0x3ffff) | |
3337 | return SPU_ILA; | |
3338 | if ((val & 0xffff) == ((val >> 16) & 0xffff)) | |
3339 | return SPU_ILH; | |
3340 | if ((val & 0xffff) == 0) | |
3341 | return SPU_ILHU; | |
3342 | ||
3343 | return SPU_NONE; | |
3344 | } | |
3345 | ||
dea01258 | 3346 | /* Return true when OP can be loaded by one of the il instructions, or |
3347 | when flow2 is not completed and OP can be loaded using ilhu and iohl. */ | |
644459d0 | 3348 | int |
3349 | immediate_load_p (rtx op, enum machine_mode mode) | |
dea01258 | 3350 | { |
3351 | if (CONSTANT_P (op)) | |
3352 | { | |
3353 | enum immediate_class c = classify_immediate (op, mode); | |
5df189be | 3354 | return c == IC_IL1 || c == IC_IL1s |
3072d30e | 3355 | || (!epilogue_completed && (c == IC_IL2 || c == IC_IL2s)); |
dea01258 | 3356 | } |
3357 | return 0; | |
3358 | } | |
3359 | ||
3360 | /* Return true if the first SIZE bytes of arr is a constant that can be | |
3361 | generated with cbd, chd, cwd or cdd. When non-NULL, PRUN and PSTART | |
3362 | represent the size and offset of the instruction to use. */ | |
3363 | static int | |
3364 | cpat_info(unsigned char *arr, int size, int *prun, int *pstart) | |
3365 | { | |
3366 | int cpat, run, i, start; | |
3367 | cpat = 1; | |
3368 | run = 0; | |
3369 | start = -1; | |
3370 | for (i = 0; i < size && cpat; i++) | |
3371 | if (arr[i] != i+16) | |
3372 | { | |
3373 | if (!run) | |
3374 | { | |
3375 | start = i; | |
3376 | if (arr[i] == 3) | |
3377 | run = 1; | |
3378 | else if (arr[i] == 2 && arr[i+1] == 3) | |
3379 | run = 2; | |
3380 | else if (arr[i] == 0) | |
3381 | { | |
3382 | while (arr[i+run] == run && i+run < 16) | |
3383 | run++; | |
3384 | if (run != 4 && run != 8) | |
3385 | cpat = 0; | |
3386 | } | |
3387 | else | |
3388 | cpat = 0; | |
3389 | if ((i & (run-1)) != 0) | |
3390 | cpat = 0; | |
3391 | i += run; | |
3392 | } | |
3393 | else | |
3394 | cpat = 0; | |
3395 | } | |
b01a6dc3 | 3396 | if (cpat && (run || size < 16)) |
dea01258 | 3397 | { |
3398 | if (run == 0) | |
3399 | run = 1; | |
3400 | if (prun) | |
3401 | *prun = run; | |
3402 | if (pstart) | |
3403 | *pstart = start == -1 ? 16-run : start; | |
3404 | return 1; | |
3405 | } | |
3406 | return 0; | |
3407 | } | |
3408 | ||
3409 | /* OP is a CONSTANT_P. Determine what instructions can be used to load | |
d819917f | 3410 | it into a register. MODE is only valid when OP is a CONST_INT. */ |
dea01258 | 3411 | static enum immediate_class |
3412 | classify_immediate (rtx op, enum machine_mode mode) | |
644459d0 | 3413 | { |
3414 | HOST_WIDE_INT val; | |
3415 | unsigned char arr[16]; | |
5df189be | 3416 | int i, j, repeated, fsmbi, repeat; |
dea01258 | 3417 | |
3418 | gcc_assert (CONSTANT_P (op)); | |
3419 | ||
644459d0 | 3420 | if (GET_MODE (op) != VOIDmode) |
3421 | mode = GET_MODE (op); | |
3422 | ||
dea01258 | 3423 | /* A V4SI const_vector with all identical symbols is ok. */ |
5df189be | 3424 | if (!flag_pic |
3425 | && mode == V4SImode | |
dea01258 | 3426 | && GET_CODE (op) == CONST_VECTOR |
3427 | && GET_CODE (CONST_VECTOR_ELT (op, 0)) != CONST_INT | |
3428 | && GET_CODE (CONST_VECTOR_ELT (op, 0)) != CONST_DOUBLE | |
3429 | && CONST_VECTOR_ELT (op, 0) == CONST_VECTOR_ELT (op, 1) | |
3430 | && CONST_VECTOR_ELT (op, 1) == CONST_VECTOR_ELT (op, 2) | |
3431 | && CONST_VECTOR_ELT (op, 2) == CONST_VECTOR_ELT (op, 3)) | |
3432 | op = CONST_VECTOR_ELT (op, 0); | |
644459d0 | 3433 | |
dea01258 | 3434 | switch (GET_CODE (op)) |
3435 | { | |
3436 | case SYMBOL_REF: | |
3437 | case LABEL_REF: | |
3438 | return TARGET_LARGE_MEM ? IC_IL2s : IC_IL1s; | |
644459d0 | 3439 | |
dea01258 | 3440 | case CONST: |
0cfc65d4 | 3441 | /* We can never know if the resulting address fits in 18 bits and can be |
3442 | loaded with ila. For now, assume the address will not overflow if | |
3443 | the displacement is "small" (fits 'K' constraint). */ | |
3444 | if (!TARGET_LARGE_MEM && GET_CODE (XEXP (op, 0)) == PLUS) | |
3445 | { | |
3446 | rtx sym = XEXP (XEXP (op, 0), 0); | |
3447 | rtx cst = XEXP (XEXP (op, 0), 1); | |
3448 | ||
3449 | if (GET_CODE (sym) == SYMBOL_REF | |
3450 | && GET_CODE (cst) == CONST_INT | |
3451 | && satisfies_constraint_K (cst)) | |
3452 | return IC_IL1s; | |
3453 | } | |
3454 | return IC_IL2s; | |
644459d0 | 3455 | |
dea01258 | 3456 | case HIGH: |
3457 | return IC_IL1s; | |
3458 | ||
3459 | case CONST_VECTOR: | |
3460 | for (i = 0; i < GET_MODE_NUNITS (mode); i++) | |
3461 | if (GET_CODE (CONST_VECTOR_ELT (op, i)) != CONST_INT | |
3462 | && GET_CODE (CONST_VECTOR_ELT (op, i)) != CONST_DOUBLE) | |
3463 | return IC_POOL; | |
3464 | /* Fall through. */ | |
3465 | ||
3466 | case CONST_INT: | |
3467 | case CONST_DOUBLE: | |
3468 | constant_to_array (mode, op, arr); | |
644459d0 | 3469 | |
dea01258 | 3470 | /* Check that each 4-byte slot is identical. */ |
3471 | repeated = 1; | |
3472 | for (i = 4; i < 16; i += 4) | |
3473 | for (j = 0; j < 4; j++) | |
3474 | if (arr[j] != arr[i + j]) | |
3475 | repeated = 0; | |
3476 | ||
3477 | if (repeated) | |
3478 | { | |
3479 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
3480 | val = trunc_int_for_mode (val, SImode); | |
3481 | ||
3482 | if (which_immediate_load (val) != SPU_NONE) | |
3483 | return IC_IL1; | |
3484 | } | |
3485 | ||
3486 | /* Any mode of 2 bytes or smaller can be loaded with an il | |
3487 | instruction. */ | |
3488 | gcc_assert (GET_MODE_SIZE (mode) > 2); | |
3489 | ||
3490 | fsmbi = 1; | |
5df189be | 3491 | repeat = 0; |
dea01258 | 3492 | for (i = 0; i < 16 && fsmbi; i++) |
5df189be | 3493 | if (arr[i] != 0 && repeat == 0) |
3494 | repeat = arr[i]; | |
3495 | else if (arr[i] != 0 && arr[i] != repeat) | |
dea01258 | 3496 | fsmbi = 0; |
3497 | if (fsmbi) | |
5df189be | 3498 | return repeat == 0xff ? IC_FSMBI : IC_FSMBI2; |
dea01258 | 3499 | |
3500 | if (cpat_info (arr, GET_MODE_SIZE (mode), 0, 0)) | |
3501 | return IC_CPAT; | |
3502 | ||
3503 | if (repeated) | |
3504 | return IC_IL2; | |
3505 | ||
3506 | return IC_POOL; | |
3507 | default: | |
3508 | break; | |
3509 | } | |
3510 | gcc_unreachable (); | |
644459d0 | 3511 | } |
3512 | ||
3513 | static enum spu_immediate | |
3514 | which_logical_immediate (HOST_WIDE_INT val) | |
3515 | { | |
3516 | gcc_assert (val == trunc_int_for_mode (val, SImode)); | |
3517 | ||
3518 | if (val >= -0x200 && val <= 0x1ff) | |
3519 | return SPU_ORI; | |
3520 | if (val >= 0 && val <= 0xffff) | |
3521 | return SPU_IOHL; | |
3522 | if ((val & 0xffff) == ((val >> 16) & 0xffff)) | |
3523 | { | |
3524 | val = trunc_int_for_mode (val, HImode); | |
3525 | if (val >= -0x200 && val <= 0x1ff) | |
3526 | return SPU_ORHI; | |
3527 | if ((val & 0xff) == ((val >> 8) & 0xff)) | |
3528 | { | |
3529 | val = trunc_int_for_mode (val, QImode); | |
3530 | if (val >= -0x200 && val <= 0x1ff) | |
3531 | return SPU_ORBI; | |
3532 | } | |
3533 | } | |
3534 | return SPU_NONE; | |
3535 | } | |
3536 | ||
5df189be | 3537 | /* Return TRUE when X, a CONST_VECTOR, only contains CONST_INTs or |
3538 | CONST_DOUBLEs. */ | |
3539 | static int | |
3540 | const_vector_immediate_p (rtx x) | |
3541 | { | |
3542 | int i; | |
3543 | gcc_assert (GET_CODE (x) == CONST_VECTOR); | |
3544 | for (i = 0; i < GET_MODE_NUNITS (GET_MODE (x)); i++) | |
3545 | if (GET_CODE (CONST_VECTOR_ELT (x, i)) != CONST_INT | |
3546 | && GET_CODE (CONST_VECTOR_ELT (x, i)) != CONST_DOUBLE) | |
3547 | return 0; | |
3548 | return 1; | |
3549 | } | |
3550 | ||
644459d0 | 3551 | int |
3552 | logical_immediate_p (rtx op, enum machine_mode mode) | |
3553 | { | |
3554 | HOST_WIDE_INT val; | |
3555 | unsigned char arr[16]; | |
3556 | int i, j; | |
3557 | ||
3558 | gcc_assert (GET_CODE (op) == CONST_INT || GET_CODE (op) == CONST_DOUBLE | |
3559 | || GET_CODE (op) == CONST_VECTOR); | |
3560 | ||
5df189be | 3561 | if (GET_CODE (op) == CONST_VECTOR |
3562 | && !const_vector_immediate_p (op)) | |
3563 | return 0; | |
3564 | ||
644459d0 | 3565 | if (GET_MODE (op) != VOIDmode) |
3566 | mode = GET_MODE (op); | |
3567 | ||
3568 | constant_to_array (mode, op, arr); | |
3569 | ||
3570 | /* Check that bytes are repeated. */ | |
3571 | for (i = 4; i < 16; i += 4) | |
3572 | for (j = 0; j < 4; j++) | |
3573 | if (arr[j] != arr[i + j]) | |
3574 | return 0; | |
3575 | ||
3576 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
3577 | val = trunc_int_for_mode (val, SImode); | |
3578 | ||
3579 | i = which_logical_immediate (val); | |
3580 | return i != SPU_NONE && i != SPU_IOHL; | |
3581 | } | |
3582 | ||
3583 | int | |
3584 | iohl_immediate_p (rtx op, enum machine_mode mode) | |
3585 | { | |
3586 | HOST_WIDE_INT val; | |
3587 | unsigned char arr[16]; | |
3588 | int i, j; | |
3589 | ||
3590 | gcc_assert (GET_CODE (op) == CONST_INT || GET_CODE (op) == CONST_DOUBLE | |
3591 | || GET_CODE (op) == CONST_VECTOR); | |
3592 | ||
5df189be | 3593 | if (GET_CODE (op) == CONST_VECTOR |
3594 | && !const_vector_immediate_p (op)) | |
3595 | return 0; | |
3596 | ||
644459d0 | 3597 | if (GET_MODE (op) != VOIDmode) |
3598 | mode = GET_MODE (op); | |
3599 | ||
3600 | constant_to_array (mode, op, arr); | |
3601 | ||
3602 | /* Check that bytes are repeated. */ | |
3603 | for (i = 4; i < 16; i += 4) | |
3604 | for (j = 0; j < 4; j++) | |
3605 | if (arr[j] != arr[i + j]) | |
3606 | return 0; | |
3607 | ||
3608 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
3609 | val = trunc_int_for_mode (val, SImode); | |
3610 | ||
3611 | return val >= 0 && val <= 0xffff; | |
3612 | } | |
3613 | ||
3614 | int | |
3615 | arith_immediate_p (rtx op, enum machine_mode mode, | |
3616 | HOST_WIDE_INT low, HOST_WIDE_INT high) | |
3617 | { | |
3618 | HOST_WIDE_INT val; | |
3619 | unsigned char arr[16]; | |
3620 | int bytes, i, j; | |
3621 | ||
3622 | gcc_assert (GET_CODE (op) == CONST_INT || GET_CODE (op) == CONST_DOUBLE | |
3623 | || GET_CODE (op) == CONST_VECTOR); | |
3624 | ||
5df189be | 3625 | if (GET_CODE (op) == CONST_VECTOR |
3626 | && !const_vector_immediate_p (op)) | |
3627 | return 0; | |
3628 | ||
644459d0 | 3629 | if (GET_MODE (op) != VOIDmode) |
3630 | mode = GET_MODE (op); | |
3631 | ||
3632 | constant_to_array (mode, op, arr); | |
3633 | ||
3634 | if (VECTOR_MODE_P (mode)) | |
3635 | mode = GET_MODE_INNER (mode); | |
3636 | ||
3637 | bytes = GET_MODE_SIZE (mode); | |
3638 | mode = mode_for_size (GET_MODE_BITSIZE (mode), MODE_INT, 0); | |
3639 | ||
3640 | /* Check that bytes are repeated. */ | |
3641 | for (i = bytes; i < 16; i += bytes) | |
3642 | for (j = 0; j < bytes; j++) | |
3643 | if (arr[j] != arr[i + j]) | |
3644 | return 0; | |
3645 | ||
3646 | val = arr[0]; | |
3647 | for (j = 1; j < bytes; j++) | |
3648 | val = (val << 8) | arr[j]; | |
3649 | ||
3650 | val = trunc_int_for_mode (val, mode); | |
3651 | ||
3652 | return val >= low && val <= high; | |
3653 | } | |
3654 | ||
56c7bfc2 | 3655 | /* TRUE when op is an immediate and an exact power of 2, and given that |
3656 | OP is 2^scale, scale >= LOW && scale <= HIGH. When OP is a vector, | |
3657 | all entries must be the same. */ | |
3658 | bool | |
3659 | exp2_immediate_p (rtx op, enum machine_mode mode, int low, int high) | |
3660 | { | |
3661 | enum machine_mode int_mode; | |
3662 | HOST_WIDE_INT val; | |
3663 | unsigned char arr[16]; | |
3664 | int bytes, i, j; | |
3665 | ||
3666 | gcc_assert (GET_CODE (op) == CONST_INT || GET_CODE (op) == CONST_DOUBLE | |
3667 | || GET_CODE (op) == CONST_VECTOR); | |
3668 | ||
3669 | if (GET_CODE (op) == CONST_VECTOR | |
3670 | && !const_vector_immediate_p (op)) | |
3671 | return 0; | |
3672 | ||
3673 | if (GET_MODE (op) != VOIDmode) | |
3674 | mode = GET_MODE (op); | |
3675 | ||
3676 | constant_to_array (mode, op, arr); | |
3677 | ||
3678 | if (VECTOR_MODE_P (mode)) | |
3679 | mode = GET_MODE_INNER (mode); | |
3680 | ||
3681 | bytes = GET_MODE_SIZE (mode); | |
3682 | int_mode = mode_for_size (GET_MODE_BITSIZE (mode), MODE_INT, 0); | |
3683 | ||
3684 | /* Check that bytes are repeated. */ | |
3685 | for (i = bytes; i < 16; i += bytes) | |
3686 | for (j = 0; j < bytes; j++) | |
3687 | if (arr[j] != arr[i + j]) | |
3688 | return 0; | |
3689 | ||
3690 | val = arr[0]; | |
3691 | for (j = 1; j < bytes; j++) | |
3692 | val = (val << 8) | arr[j]; | |
3693 | ||
3694 | val = trunc_int_for_mode (val, int_mode); | |
3695 | ||
3696 | /* Currently, we only handle SFmode */ | |
3697 | gcc_assert (mode == SFmode); | |
3698 | if (mode == SFmode) | |
3699 | { | |
3700 | int exp = (val >> 23) - 127; | |
3701 | return val > 0 && (val & 0x007fffff) == 0 | |
3702 | && exp >= low && exp <= high; | |
3703 | } | |
3704 | return FALSE; | |
3705 | } | |
3706 | ||
6cf5579e | 3707 | /* Return true if X is a SYMBOL_REF to an __ea qualified variable. */ |
3708 | ||
3709 | static int | |
3710 | ea_symbol_ref (rtx *px, void *data ATTRIBUTE_UNUSED) | |
3711 | { | |
3712 | rtx x = *px; | |
3713 | tree decl; | |
3714 | ||
3715 | if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS) | |
3716 | { | |
3717 | rtx plus = XEXP (x, 0); | |
3718 | rtx op0 = XEXP (plus, 0); | |
3719 | rtx op1 = XEXP (plus, 1); | |
3720 | if (GET_CODE (op1) == CONST_INT) | |
3721 | x = op0; | |
3722 | } | |
3723 | ||
3724 | return (GET_CODE (x) == SYMBOL_REF | |
3725 | && (decl = SYMBOL_REF_DECL (x)) != 0 | |
3726 | && TREE_CODE (decl) == VAR_DECL | |
3727 | && TYPE_ADDR_SPACE (TREE_TYPE (decl))); | |
3728 | } | |
3729 | ||
644459d0 | 3730 | /* We accept: |
5b865faf | 3731 | - any 32-bit constant (SImode, SFmode) |
644459d0 | 3732 | - any constant that can be generated with fsmbi (any mode) |
5b865faf | 3733 | - a 64-bit constant where the high and low bits are identical |
644459d0 | 3734 | (DImode, DFmode) |
5b865faf | 3735 | - a 128-bit constant where the four 32-bit words match. */ |
644459d0 | 3736 | int |
3737 | spu_legitimate_constant_p (rtx x) | |
3738 | { | |
5df189be | 3739 | if (GET_CODE (x) == HIGH) |
3740 | x = XEXP (x, 0); | |
6cf5579e | 3741 | |
3742 | /* Reject any __ea qualified reference. These can't appear in | |
3743 | instructions but must be forced to the constant pool. */ | |
3744 | if (for_each_rtx (&x, ea_symbol_ref, 0)) | |
3745 | return 0; | |
3746 | ||
644459d0 | 3747 | /* V4SI with all identical symbols is valid. */ |
5df189be | 3748 | if (!flag_pic |
3749 | && GET_MODE (x) == V4SImode | |
644459d0 | 3750 | && (GET_CODE (CONST_VECTOR_ELT (x, 0)) == SYMBOL_REF |
3751 | || GET_CODE (CONST_VECTOR_ELT (x, 0)) == LABEL_REF | |
5df189be | 3752 | || GET_CODE (CONST_VECTOR_ELT (x, 0)) == CONST)) |
644459d0 | 3753 | return CONST_VECTOR_ELT (x, 0) == CONST_VECTOR_ELT (x, 1) |
3754 | && CONST_VECTOR_ELT (x, 1) == CONST_VECTOR_ELT (x, 2) | |
3755 | && CONST_VECTOR_ELT (x, 2) == CONST_VECTOR_ELT (x, 3); | |
3756 | ||
5df189be | 3757 | if (GET_CODE (x) == CONST_VECTOR |
3758 | && !const_vector_immediate_p (x)) | |
3759 | return 0; | |
644459d0 | 3760 | return 1; |
3761 | } | |
3762 | ||
3763 | /* Valid address are: | |
3764 | - symbol_ref, label_ref, const | |
3765 | - reg | |
9d98604b | 3766 | - reg + const_int, where const_int is 16 byte aligned |
644459d0 | 3767 | - reg + reg, alignment doesn't matter |
3768 | The alignment matters in the reg+const case because lqd and stqd | |
9d98604b | 3769 | ignore the 4 least significant bits of the const. We only care about |
3770 | 16 byte modes because the expand phase will change all smaller MEM | |
3771 | references to TImode. */ | |
3772 | static bool | |
3773 | spu_legitimate_address_p (enum machine_mode mode, | |
fd50b071 | 3774 | rtx x, bool reg_ok_strict) |
644459d0 | 3775 | { |
9d98604b | 3776 | int aligned = GET_MODE_SIZE (mode) >= 16; |
3777 | if (aligned | |
3778 | && GET_CODE (x) == AND | |
644459d0 | 3779 | && GET_CODE (XEXP (x, 1)) == CONST_INT |
9d98604b | 3780 | && INTVAL (XEXP (x, 1)) == (HOST_WIDE_INT) - 16) |
644459d0 | 3781 | x = XEXP (x, 0); |
3782 | switch (GET_CODE (x)) | |
3783 | { | |
644459d0 | 3784 | case LABEL_REF: |
6cf5579e | 3785 | return !TARGET_LARGE_MEM; |
3786 | ||
9d98604b | 3787 | case SYMBOL_REF: |
644459d0 | 3788 | case CONST: |
6cf5579e | 3789 | /* Keep __ea references until reload so that spu_expand_mov can see them |
3790 | in MEMs. */ | |
3791 | if (ea_symbol_ref (&x, 0)) | |
3792 | return !reload_in_progress && !reload_completed; | |
9d98604b | 3793 | return !TARGET_LARGE_MEM; |
644459d0 | 3794 | |
3795 | case CONST_INT: | |
3796 | return INTVAL (x) >= 0 && INTVAL (x) <= 0x3ffff; | |
3797 | ||
3798 | case SUBREG: | |
3799 | x = XEXP (x, 0); | |
9d98604b | 3800 | if (REG_P (x)) |
3801 | return 0; | |
644459d0 | 3802 | |
3803 | case REG: | |
3804 | return INT_REG_OK_FOR_BASE_P (x, reg_ok_strict); | |
3805 | ||
3806 | case PLUS: | |
3807 | case LO_SUM: | |
3808 | { | |
3809 | rtx op0 = XEXP (x, 0); | |
3810 | rtx op1 = XEXP (x, 1); | |
3811 | if (GET_CODE (op0) == SUBREG) | |
3812 | op0 = XEXP (op0, 0); | |
3813 | if (GET_CODE (op1) == SUBREG) | |
3814 | op1 = XEXP (op1, 0); | |
644459d0 | 3815 | if (GET_CODE (op0) == REG |
3816 | && INT_REG_OK_FOR_BASE_P (op0, reg_ok_strict) | |
3817 | && GET_CODE (op1) == CONST_INT | |
3818 | && INTVAL (op1) >= -0x2000 | |
3819 | && INTVAL (op1) <= 0x1fff | |
9d98604b | 3820 | && (!aligned || (INTVAL (op1) & 15) == 0)) |
3821 | return TRUE; | |
644459d0 | 3822 | if (GET_CODE (op0) == REG |
3823 | && INT_REG_OK_FOR_BASE_P (op0, reg_ok_strict) | |
3824 | && GET_CODE (op1) == REG | |
3825 | && INT_REG_OK_FOR_INDEX_P (op1, reg_ok_strict)) | |
9d98604b | 3826 | return TRUE; |
644459d0 | 3827 | } |
3828 | break; | |
3829 | ||
3830 | default: | |
3831 | break; | |
3832 | } | |
9d98604b | 3833 | return FALSE; |
644459d0 | 3834 | } |
3835 | ||
6cf5579e | 3836 | /* Like spu_legitimate_address_p, except with named addresses. */ |
3837 | static bool | |
3838 | spu_addr_space_legitimate_address_p (enum machine_mode mode, rtx x, | |
3839 | bool reg_ok_strict, addr_space_t as) | |
3840 | { | |
3841 | if (as == ADDR_SPACE_EA) | |
3842 | return (REG_P (x) && (GET_MODE (x) == EAmode)); | |
3843 | ||
3844 | else if (as != ADDR_SPACE_GENERIC) | |
3845 | gcc_unreachable (); | |
3846 | ||
3847 | return spu_legitimate_address_p (mode, x, reg_ok_strict); | |
3848 | } | |
3849 | ||
644459d0 | 3850 | /* When the address is reg + const_int, force the const_int into a |
fa7637bd | 3851 | register. */ |
644459d0 | 3852 | rtx |
3853 | spu_legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED, | |
41e3a0c7 | 3854 | enum machine_mode mode ATTRIBUTE_UNUSED) |
644459d0 | 3855 | { |
3856 | rtx op0, op1; | |
3857 | /* Make sure both operands are registers. */ | |
3858 | if (GET_CODE (x) == PLUS) | |
3859 | { | |
3860 | op0 = XEXP (x, 0); | |
3861 | op1 = XEXP (x, 1); | |
3862 | if (ALIGNED_SYMBOL_REF_P (op0)) | |
3863 | { | |
3864 | op0 = force_reg (Pmode, op0); | |
3865 | mark_reg_pointer (op0, 128); | |
3866 | } | |
3867 | else if (GET_CODE (op0) != REG) | |
3868 | op0 = force_reg (Pmode, op0); | |
3869 | if (ALIGNED_SYMBOL_REF_P (op1)) | |
3870 | { | |
3871 | op1 = force_reg (Pmode, op1); | |
3872 | mark_reg_pointer (op1, 128); | |
3873 | } | |
3874 | else if (GET_CODE (op1) != REG) | |
3875 | op1 = force_reg (Pmode, op1); | |
3876 | x = gen_rtx_PLUS (Pmode, op0, op1); | |
644459d0 | 3877 | } |
41e3a0c7 | 3878 | return x; |
644459d0 | 3879 | } |
3880 | ||
6cf5579e | 3881 | /* Like spu_legitimate_address, except with named address support. */ |
3882 | static rtx | |
3883 | spu_addr_space_legitimize_address (rtx x, rtx oldx, enum machine_mode mode, | |
3884 | addr_space_t as) | |
3885 | { | |
3886 | if (as != ADDR_SPACE_GENERIC) | |
3887 | return x; | |
3888 | ||
3889 | return spu_legitimize_address (x, oldx, mode); | |
3890 | } | |
3891 | ||
644459d0 | 3892 | /* Handle an attribute requiring a FUNCTION_DECL; arguments as in |
3893 | struct attribute_spec.handler. */ | |
3894 | static tree | |
3895 | spu_handle_fndecl_attribute (tree * node, | |
3896 | tree name, | |
3897 | tree args ATTRIBUTE_UNUSED, | |
3898 | int flags ATTRIBUTE_UNUSED, bool * no_add_attrs) | |
3899 | { | |
3900 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
3901 | { | |
67a779df | 3902 | warning (0, "%qE attribute only applies to functions", |
3903 | name); | |
644459d0 | 3904 | *no_add_attrs = true; |
3905 | } | |
3906 | ||
3907 | return NULL_TREE; | |
3908 | } | |
3909 | ||
3910 | /* Handle the "vector" attribute. */ | |
3911 | static tree | |
3912 | spu_handle_vector_attribute (tree * node, tree name, | |
3913 | tree args ATTRIBUTE_UNUSED, | |
3914 | int flags ATTRIBUTE_UNUSED, bool * no_add_attrs) | |
3915 | { | |
3916 | tree type = *node, result = NULL_TREE; | |
3917 | enum machine_mode mode; | |
3918 | int unsigned_p; | |
3919 | ||
3920 | while (POINTER_TYPE_P (type) | |
3921 | || TREE_CODE (type) == FUNCTION_TYPE | |
3922 | || TREE_CODE (type) == METHOD_TYPE || TREE_CODE (type) == ARRAY_TYPE) | |
3923 | type = TREE_TYPE (type); | |
3924 | ||
3925 | mode = TYPE_MODE (type); | |
3926 | ||
3927 | unsigned_p = TYPE_UNSIGNED (type); | |
3928 | switch (mode) | |
3929 | { | |
3930 | case DImode: | |
3931 | result = (unsigned_p ? unsigned_V2DI_type_node : V2DI_type_node); | |
3932 | break; | |
3933 | case SImode: | |
3934 | result = (unsigned_p ? unsigned_V4SI_type_node : V4SI_type_node); | |
3935 | break; | |
3936 | case HImode: | |
3937 | result = (unsigned_p ? unsigned_V8HI_type_node : V8HI_type_node); | |
3938 | break; | |
3939 | case QImode: | |
3940 | result = (unsigned_p ? unsigned_V16QI_type_node : V16QI_type_node); | |
3941 | break; | |
3942 | case SFmode: | |
3943 | result = V4SF_type_node; | |
3944 | break; | |
3945 | case DFmode: | |
3946 | result = V2DF_type_node; | |
3947 | break; | |
3948 | default: | |
3949 | break; | |
3950 | } | |
3951 | ||
3952 | /* Propagate qualifiers attached to the element type | |
3953 | onto the vector type. */ | |
3954 | if (result && result != type && TYPE_QUALS (type)) | |
3955 | result = build_qualified_type (result, TYPE_QUALS (type)); | |
3956 | ||
3957 | *no_add_attrs = true; /* No need to hang on to the attribute. */ | |
3958 | ||
3959 | if (!result) | |
67a779df | 3960 | warning (0, "%qE attribute ignored", name); |
644459d0 | 3961 | else |
d991e6e8 | 3962 | *node = lang_hooks.types.reconstruct_complex_type (*node, result); |
644459d0 | 3963 | |
3964 | return NULL_TREE; | |
3965 | } | |
3966 | ||
f2b32076 | 3967 | /* Return nonzero if FUNC is a naked function. */ |
644459d0 | 3968 | static int |
3969 | spu_naked_function_p (tree func) | |
3970 | { | |
3971 | tree a; | |
3972 | ||
3973 | if (TREE_CODE (func) != FUNCTION_DECL) | |
3974 | abort (); | |
3975 | ||
3976 | a = lookup_attribute ("naked", DECL_ATTRIBUTES (func)); | |
3977 | return a != NULL_TREE; | |
3978 | } | |
3979 | ||
3980 | int | |
3981 | spu_initial_elimination_offset (int from, int to) | |
3982 | { | |
3983 | int saved_regs_size = spu_saved_regs_size (); | |
3984 | int sp_offset = 0; | |
abe32cce | 3985 | if (!current_function_is_leaf || crtl->outgoing_args_size |
644459d0 | 3986 | || get_frame_size () || saved_regs_size) |
3987 | sp_offset = STACK_POINTER_OFFSET; | |
3988 | if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM) | |
a8e019fa | 3989 | return get_frame_size () + crtl->outgoing_args_size + sp_offset; |
644459d0 | 3990 | else if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) |
a8e019fa | 3991 | return get_frame_size (); |
644459d0 | 3992 | else if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
abe32cce | 3993 | return sp_offset + crtl->outgoing_args_size |
644459d0 | 3994 | + get_frame_size () + saved_regs_size + STACK_POINTER_OFFSET; |
3995 | else if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
3996 | return get_frame_size () + saved_regs_size + sp_offset; | |
a8e019fa | 3997 | else |
3998 | gcc_unreachable (); | |
644459d0 | 3999 | } |
4000 | ||
4001 | rtx | |
fb80456a | 4002 | spu_function_value (const_tree type, const_tree func ATTRIBUTE_UNUSED) |
644459d0 | 4003 | { |
4004 | enum machine_mode mode = TYPE_MODE (type); | |
4005 | int byte_size = ((mode == BLKmode) | |
4006 | ? int_size_in_bytes (type) : GET_MODE_SIZE (mode)); | |
4007 | ||
4008 | /* Make sure small structs are left justified in a register. */ | |
4009 | if ((mode == BLKmode || (type && AGGREGATE_TYPE_P (type))) | |
4010 | && byte_size <= UNITS_PER_WORD * MAX_REGISTER_RETURN && byte_size > 0) | |
4011 | { | |
4012 | enum machine_mode smode; | |
4013 | rtvec v; | |
4014 | int i; | |
4015 | int nregs = (byte_size + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
4016 | int n = byte_size / UNITS_PER_WORD; | |
4017 | v = rtvec_alloc (nregs); | |
4018 | for (i = 0; i < n; i++) | |
4019 | { | |
4020 | RTVEC_ELT (v, i) = gen_rtx_EXPR_LIST (VOIDmode, | |
4021 | gen_rtx_REG (TImode, | |
4022 | FIRST_RETURN_REGNUM | |
4023 | + i), | |
4024 | GEN_INT (UNITS_PER_WORD * i)); | |
4025 | byte_size -= UNITS_PER_WORD; | |
4026 | } | |
4027 | ||
4028 | if (n < nregs) | |
4029 | { | |
4030 | if (byte_size < 4) | |
4031 | byte_size = 4; | |
4032 | smode = | |
4033 | smallest_mode_for_size (byte_size * BITS_PER_UNIT, MODE_INT); | |
4034 | RTVEC_ELT (v, n) = | |
4035 | gen_rtx_EXPR_LIST (VOIDmode, | |
4036 | gen_rtx_REG (smode, FIRST_RETURN_REGNUM + n), | |
4037 | GEN_INT (UNITS_PER_WORD * n)); | |
4038 | } | |
4039 | return gen_rtx_PARALLEL (mode, v); | |
4040 | } | |
4041 | return gen_rtx_REG (mode, FIRST_RETURN_REGNUM); | |
4042 | } | |
4043 | ||
ee9034d4 | 4044 | static rtx |
4045 | spu_function_arg (CUMULATIVE_ARGS *cum, | |
644459d0 | 4046 | enum machine_mode mode, |
ee9034d4 | 4047 | const_tree type, bool named ATTRIBUTE_UNUSED) |
644459d0 | 4048 | { |
4049 | int byte_size; | |
4050 | ||
a08c5dd0 | 4051 | if (*cum >= MAX_REGISTER_ARGS) |
644459d0 | 4052 | return 0; |
4053 | ||
4054 | byte_size = ((mode == BLKmode) | |
4055 | ? int_size_in_bytes (type) : GET_MODE_SIZE (mode)); | |
4056 | ||
4057 | /* The ABI does not allow parameters to be passed partially in | |
4058 | reg and partially in stack. */ | |
a08c5dd0 | 4059 | if ((*cum + (byte_size + 15) / 16) > MAX_REGISTER_ARGS) |
644459d0 | 4060 | return 0; |
4061 | ||
4062 | /* Make sure small structs are left justified in a register. */ | |
4063 | if ((mode == BLKmode || (type && AGGREGATE_TYPE_P (type))) | |
4064 | && byte_size < UNITS_PER_WORD && byte_size > 0) | |
4065 | { | |
4066 | enum machine_mode smode; | |
4067 | rtx gr_reg; | |
4068 | if (byte_size < 4) | |
4069 | byte_size = 4; | |
4070 | smode = smallest_mode_for_size (byte_size * BITS_PER_UNIT, MODE_INT); | |
4071 | gr_reg = gen_rtx_EXPR_LIST (VOIDmode, | |
a08c5dd0 | 4072 | gen_rtx_REG (smode, FIRST_ARG_REGNUM + *cum), |
644459d0 | 4073 | const0_rtx); |
4074 | return gen_rtx_PARALLEL (mode, gen_rtvec (1, gr_reg)); | |
4075 | } | |
4076 | else | |
a08c5dd0 | 4077 | return gen_rtx_REG (mode, FIRST_ARG_REGNUM + *cum); |
644459d0 | 4078 | } |
4079 | ||
ee9034d4 | 4080 | static void |
4081 | spu_function_arg_advance (CUMULATIVE_ARGS * cum, enum machine_mode mode, | |
4082 | const_tree type, bool named ATTRIBUTE_UNUSED) | |
4083 | { | |
4084 | *cum += (type && TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST | |
4085 | ? 1 | |
4086 | : mode == BLKmode | |
4087 | ? ((int_size_in_bytes (type) + 15) / 16) | |
4088 | : mode == VOIDmode | |
4089 | ? 1 | |
4090 | : HARD_REGNO_NREGS (cum, mode)); | |
4091 | } | |
4092 | ||
644459d0 | 4093 | /* Variable sized types are passed by reference. */ |
4094 | static bool | |
4095 | spu_pass_by_reference (CUMULATIVE_ARGS * cum ATTRIBUTE_UNUSED, | |
4096 | enum machine_mode mode ATTRIBUTE_UNUSED, | |
fb80456a | 4097 | const_tree type, bool named ATTRIBUTE_UNUSED) |
644459d0 | 4098 | { |
4099 | return type && TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST; | |
4100 | } | |
4101 | \f | |
4102 | ||
4103 | /* Var args. */ | |
4104 | ||
4105 | /* Create and return the va_list datatype. | |
4106 | ||
4107 | On SPU, va_list is an array type equivalent to | |
4108 | ||
4109 | typedef struct __va_list_tag | |
4110 | { | |
4111 | void *__args __attribute__((__aligned(16))); | |
4112 | void *__skip __attribute__((__aligned(16))); | |
4113 | ||
4114 | } va_list[1]; | |
4115 | ||
fa7637bd | 4116 | where __args points to the arg that will be returned by the next |
644459d0 | 4117 | va_arg(), and __skip points to the previous stack frame such that |
4118 | when __args == __skip we should advance __args by 32 bytes. */ | |
4119 | static tree | |
4120 | spu_build_builtin_va_list (void) | |
4121 | { | |
4122 | tree f_args, f_skip, record, type_decl; | |
4123 | bool owp; | |
4124 | ||
4125 | record = (*lang_hooks.types.make_type) (RECORD_TYPE); | |
4126 | ||
4127 | type_decl = | |
54e46243 | 4128 | build_decl (BUILTINS_LOCATION, |
4129 | TYPE_DECL, get_identifier ("__va_list_tag"), record); | |
644459d0 | 4130 | |
54e46243 | 4131 | f_args = build_decl (BUILTINS_LOCATION, |
4132 | FIELD_DECL, get_identifier ("__args"), ptr_type_node); | |
4133 | f_skip = build_decl (BUILTINS_LOCATION, | |
4134 | FIELD_DECL, get_identifier ("__skip"), ptr_type_node); | |
644459d0 | 4135 | |
4136 | DECL_FIELD_CONTEXT (f_args) = record; | |
4137 | DECL_ALIGN (f_args) = 128; | |
4138 | DECL_USER_ALIGN (f_args) = 1; | |
4139 | ||
4140 | DECL_FIELD_CONTEXT (f_skip) = record; | |
4141 | DECL_ALIGN (f_skip) = 128; | |
4142 | DECL_USER_ALIGN (f_skip) = 1; | |
4143 | ||
bc907808 | 4144 | TYPE_STUB_DECL (record) = type_decl; |
644459d0 | 4145 | TYPE_NAME (record) = type_decl; |
4146 | TYPE_FIELDS (record) = f_args; | |
1767a056 | 4147 | DECL_CHAIN (f_args) = f_skip; |
644459d0 | 4148 | |
4149 | /* We know this is being padded and we want it too. It is an internal | |
4150 | type so hide the warnings from the user. */ | |
4151 | owp = warn_padded; | |
4152 | warn_padded = false; | |
4153 | ||
4154 | layout_type (record); | |
4155 | ||
4156 | warn_padded = owp; | |
4157 | ||
4158 | /* The correct type is an array type of one element. */ | |
4159 | return build_array_type (record, build_index_type (size_zero_node)); | |
4160 | } | |
4161 | ||
4162 | /* Implement va_start by filling the va_list structure VALIST. | |
4163 | NEXTARG points to the first anonymous stack argument. | |
4164 | ||
4165 | The following global variables are used to initialize | |
4166 | the va_list structure: | |
4167 | ||
abe32cce | 4168 | crtl->args.info; |
644459d0 | 4169 | the CUMULATIVE_ARGS for this function |
4170 | ||
abe32cce | 4171 | crtl->args.arg_offset_rtx: |
644459d0 | 4172 | holds the offset of the first anonymous stack argument |
4173 | (relative to the virtual arg pointer). */ | |
4174 | ||
8a58ed0a | 4175 | static void |
644459d0 | 4176 | spu_va_start (tree valist, rtx nextarg) |
4177 | { | |
4178 | tree f_args, f_skip; | |
4179 | tree args, skip, t; | |
4180 | ||
4181 | f_args = TYPE_FIELDS (TREE_TYPE (va_list_type_node)); | |
1767a056 | 4182 | f_skip = DECL_CHAIN (f_args); |
644459d0 | 4183 | |
170efcd4 | 4184 | valist = build_simple_mem_ref (valist); |
644459d0 | 4185 | args = |
4186 | build3 (COMPONENT_REF, TREE_TYPE (f_args), valist, f_args, NULL_TREE); | |
4187 | skip = | |
4188 | build3 (COMPONENT_REF, TREE_TYPE (f_skip), valist, f_skip, NULL_TREE); | |
4189 | ||
4190 | /* Find the __args area. */ | |
4191 | t = make_tree (TREE_TYPE (args), nextarg); | |
abe32cce | 4192 | if (crtl->args.pretend_args_size > 0) |
0de36bdb | 4193 | t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (args), t, |
4194 | size_int (-STACK_POINTER_OFFSET)); | |
75a70cf9 | 4195 | t = build2 (MODIFY_EXPR, TREE_TYPE (args), args, t); |
644459d0 | 4196 | TREE_SIDE_EFFECTS (t) = 1; |
4197 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
4198 | ||
4199 | /* Find the __skip area. */ | |
4200 | t = make_tree (TREE_TYPE (skip), virtual_incoming_args_rtx); | |
0de36bdb | 4201 | t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (skip), t, |
abe32cce | 4202 | size_int (crtl->args.pretend_args_size |
0de36bdb | 4203 | - STACK_POINTER_OFFSET)); |
75a70cf9 | 4204 | t = build2 (MODIFY_EXPR, TREE_TYPE (skip), skip, t); |
644459d0 | 4205 | TREE_SIDE_EFFECTS (t) = 1; |
4206 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
4207 | } | |
4208 | ||
4209 | /* Gimplify va_arg by updating the va_list structure | |
4210 | VALIST as required to retrieve an argument of type | |
4211 | TYPE, and returning that argument. | |
4212 | ||
4213 | ret = va_arg(VALIST, TYPE); | |
4214 | ||
4215 | generates code equivalent to: | |
4216 | ||
4217 | paddedsize = (sizeof(TYPE) + 15) & -16; | |
4218 | if (VALIST.__args + paddedsize > VALIST.__skip | |
4219 | && VALIST.__args <= VALIST.__skip) | |
4220 | addr = VALIST.__skip + 32; | |
4221 | else | |
4222 | addr = VALIST.__args; | |
4223 | VALIST.__args = addr + paddedsize; | |
4224 | ret = *(TYPE *)addr; | |
4225 | */ | |
4226 | static tree | |
75a70cf9 | 4227 | spu_gimplify_va_arg_expr (tree valist, tree type, gimple_seq * pre_p, |
4228 | gimple_seq * post_p ATTRIBUTE_UNUSED) | |
644459d0 | 4229 | { |
4230 | tree f_args, f_skip; | |
4231 | tree args, skip; | |
4232 | HOST_WIDE_INT size, rsize; | |
4233 | tree paddedsize, addr, tmp; | |
4234 | bool pass_by_reference_p; | |
4235 | ||
4236 | f_args = TYPE_FIELDS (TREE_TYPE (va_list_type_node)); | |
1767a056 | 4237 | f_skip = DECL_CHAIN (f_args); |
644459d0 | 4238 | |
182cf5a9 | 4239 | valist = build_simple_mem_ref (valist); |
644459d0 | 4240 | args = |
4241 | build3 (COMPONENT_REF, TREE_TYPE (f_args), valist, f_args, NULL_TREE); | |
4242 | skip = | |
4243 | build3 (COMPONENT_REF, TREE_TYPE (f_skip), valist, f_skip, NULL_TREE); | |
4244 | ||
4245 | addr = create_tmp_var (ptr_type_node, "va_arg"); | |
644459d0 | 4246 | |
4247 | /* if an object is dynamically sized, a pointer to it is passed | |
4248 | instead of the object itself. */ | |
4249 | pass_by_reference_p = spu_pass_by_reference (NULL, TYPE_MODE (type), type, | |
4250 | false); | |
4251 | if (pass_by_reference_p) | |
4252 | type = build_pointer_type (type); | |
4253 | size = int_size_in_bytes (type); | |
4254 | rsize = ((size + UNITS_PER_WORD - 1) / UNITS_PER_WORD) * UNITS_PER_WORD; | |
4255 | ||
4256 | /* build conditional expression to calculate addr. The expression | |
4257 | will be gimplified later. */ | |
0de36bdb | 4258 | paddedsize = size_int (rsize); |
75a70cf9 | 4259 | tmp = build2 (POINTER_PLUS_EXPR, ptr_type_node, unshare_expr (args), paddedsize); |
644459d0 | 4260 | tmp = build2 (TRUTH_AND_EXPR, boolean_type_node, |
75a70cf9 | 4261 | build2 (GT_EXPR, boolean_type_node, tmp, unshare_expr (skip)), |
4262 | build2 (LE_EXPR, boolean_type_node, unshare_expr (args), | |
4263 | unshare_expr (skip))); | |
644459d0 | 4264 | |
4265 | tmp = build3 (COND_EXPR, ptr_type_node, tmp, | |
75a70cf9 | 4266 | build2 (POINTER_PLUS_EXPR, ptr_type_node, unshare_expr (skip), |
4267 | size_int (32)), unshare_expr (args)); | |
644459d0 | 4268 | |
75a70cf9 | 4269 | gimplify_assign (addr, tmp, pre_p); |
644459d0 | 4270 | |
4271 | /* update VALIST.__args */ | |
0de36bdb | 4272 | tmp = build2 (POINTER_PLUS_EXPR, ptr_type_node, addr, paddedsize); |
75a70cf9 | 4273 | gimplify_assign (unshare_expr (args), tmp, pre_p); |
644459d0 | 4274 | |
8115f0af | 4275 | addr = fold_convert (build_pointer_type_for_mode (type, ptr_mode, true), |
4276 | addr); | |
644459d0 | 4277 | |
4278 | if (pass_by_reference_p) | |
4279 | addr = build_va_arg_indirect_ref (addr); | |
4280 | ||
4281 | return build_va_arg_indirect_ref (addr); | |
4282 | } | |
4283 | ||
4284 | /* Save parameter registers starting with the register that corresponds | |
4285 | to the first unnamed parameters. If the first unnamed parameter is | |
4286 | in the stack then save no registers. Set pretend_args_size to the | |
4287 | amount of space needed to save the registers. */ | |
4288 | void | |
4289 | spu_setup_incoming_varargs (CUMULATIVE_ARGS * cum, enum machine_mode mode, | |
4290 | tree type, int *pretend_size, int no_rtl) | |
4291 | { | |
4292 | if (!no_rtl) | |
4293 | { | |
4294 | rtx tmp; | |
4295 | int regno; | |
4296 | int offset; | |
4297 | int ncum = *cum; | |
4298 | ||
4299 | /* cum currently points to the last named argument, we want to | |
4300 | start at the next argument. */ | |
ee9034d4 | 4301 | spu_function_arg_advance (&ncum, mode, type, true); |
644459d0 | 4302 | |
4303 | offset = -STACK_POINTER_OFFSET; | |
4304 | for (regno = ncum; regno < MAX_REGISTER_ARGS; regno++) | |
4305 | { | |
4306 | tmp = gen_frame_mem (V4SImode, | |
4307 | plus_constant (virtual_incoming_args_rtx, | |
4308 | offset)); | |
4309 | emit_move_insn (tmp, | |
4310 | gen_rtx_REG (V4SImode, FIRST_ARG_REGNUM + regno)); | |
4311 | offset += 16; | |
4312 | } | |
4313 | *pretend_size = offset + STACK_POINTER_OFFSET; | |
4314 | } | |
4315 | } | |
4316 | \f | |
b2d7ede1 | 4317 | static void |
644459d0 | 4318 | spu_conditional_register_usage (void) |
4319 | { | |
4320 | if (flag_pic) | |
4321 | { | |
4322 | fixed_regs[PIC_OFFSET_TABLE_REGNUM] = 1; | |
4323 | call_used_regs[PIC_OFFSET_TABLE_REGNUM] = 1; | |
4324 | } | |
644459d0 | 4325 | } |
4326 | ||
9d98604b | 4327 | /* This is called any time we inspect the alignment of a register for |
4328 | addresses. */ | |
644459d0 | 4329 | static int |
9d98604b | 4330 | reg_aligned_for_addr (rtx x) |
644459d0 | 4331 | { |
9d98604b | 4332 | int regno = |
4333 | REGNO (x) < FIRST_PSEUDO_REGISTER ? ORIGINAL_REGNO (x) : REGNO (x); | |
4334 | return REGNO_POINTER_ALIGN (regno) >= 128; | |
644459d0 | 4335 | } |
4336 | ||
69ced2d6 | 4337 | /* Encode symbol attributes (local vs. global, tls model) of a SYMBOL_REF |
4338 | into its SYMBOL_REF_FLAGS. */ | |
4339 | static void | |
4340 | spu_encode_section_info (tree decl, rtx rtl, int first) | |
4341 | { | |
4342 | default_encode_section_info (decl, rtl, first); | |
4343 | ||
4344 | /* If a variable has a forced alignment to < 16 bytes, mark it with | |
4345 | SYMBOL_FLAG_ALIGN1. */ | |
4346 | if (TREE_CODE (decl) == VAR_DECL | |
4347 | && DECL_USER_ALIGN (decl) && DECL_ALIGN (decl) < 128) | |
4348 | SYMBOL_REF_FLAGS (XEXP (rtl, 0)) |= SYMBOL_FLAG_ALIGN1; | |
4349 | } | |
4350 | ||
644459d0 | 4351 | /* Return TRUE if we are certain the mem refers to a complete object |
4352 | which is both 16-byte aligned and padded to a 16-byte boundary. This | |
4353 | would make it safe to store with a single instruction. | |
4354 | We guarantee the alignment and padding for static objects by aligning | |
4355 | all of them to 16-bytes. (DATA_ALIGNMENT and CONSTANT_ALIGNMENT.) | |
4356 | FIXME: We currently cannot guarantee this for objects on the stack | |
4357 | because assign_parm_setup_stack calls assign_stack_local with the | |
4358 | alignment of the parameter mode and in that case the alignment never | |
4359 | gets adjusted by LOCAL_ALIGNMENT. */ | |
4360 | static int | |
4361 | store_with_one_insn_p (rtx mem) | |
4362 | { | |
9d98604b | 4363 | enum machine_mode mode = GET_MODE (mem); |
644459d0 | 4364 | rtx addr = XEXP (mem, 0); |
9d98604b | 4365 | if (mode == BLKmode) |
644459d0 | 4366 | return 0; |
9d98604b | 4367 | if (GET_MODE_SIZE (mode) >= 16) |
4368 | return 1; | |
644459d0 | 4369 | /* Only static objects. */ |
4370 | if (GET_CODE (addr) == SYMBOL_REF) | |
4371 | { | |
4372 | /* We use the associated declaration to make sure the access is | |
fa7637bd | 4373 | referring to the whole object. |
851d9296 | 4374 | We check both MEM_EXPR and SYMBOL_REF_DECL. I'm not sure |
644459d0 | 4375 | if it is necessary. Will there be cases where one exists, and |
4376 | the other does not? Will there be cases where both exist, but | |
4377 | have different types? */ | |
4378 | tree decl = MEM_EXPR (mem); | |
4379 | if (decl | |
4380 | && TREE_CODE (decl) == VAR_DECL | |
4381 | && GET_MODE (mem) == TYPE_MODE (TREE_TYPE (decl))) | |
4382 | return 1; | |
4383 | decl = SYMBOL_REF_DECL (addr); | |
4384 | if (decl | |
4385 | && TREE_CODE (decl) == VAR_DECL | |
4386 | && GET_MODE (mem) == TYPE_MODE (TREE_TYPE (decl))) | |
4387 | return 1; | |
4388 | } | |
4389 | return 0; | |
4390 | } | |
4391 | ||
9d98604b | 4392 | /* Return 1 when the address is not valid for a simple load and store as |
4393 | required by the '_mov*' patterns. We could make this less strict | |
4394 | for loads, but we prefer mem's to look the same so they are more | |
4395 | likely to be merged. */ | |
4396 | static int | |
4397 | address_needs_split (rtx mem) | |
4398 | { | |
4399 | if (GET_MODE_SIZE (GET_MODE (mem)) < 16 | |
4400 | && (GET_MODE_SIZE (GET_MODE (mem)) < 4 | |
4401 | || !(store_with_one_insn_p (mem) | |
4402 | || mem_is_padded_component_ref (mem)))) | |
4403 | return 1; | |
4404 | ||
4405 | return 0; | |
4406 | } | |
4407 | ||
6cf5579e | 4408 | static GTY(()) rtx cache_fetch; /* __cache_fetch function */ |
4409 | static GTY(()) rtx cache_fetch_dirty; /* __cache_fetch_dirty function */ | |
4410 | static alias_set_type ea_alias_set = -1; /* alias set for __ea memory */ | |
4411 | ||
4412 | /* MEM is known to be an __ea qualified memory access. Emit a call to | |
4413 | fetch the ppu memory to local store, and return its address in local | |
4414 | store. */ | |
4415 | ||
4416 | static void | |
4417 | ea_load_store (rtx mem, bool is_store, rtx ea_addr, rtx data_addr) | |
4418 | { | |
4419 | if (is_store) | |
4420 | { | |
4421 | rtx ndirty = GEN_INT (GET_MODE_SIZE (GET_MODE (mem))); | |
4422 | if (!cache_fetch_dirty) | |
4423 | cache_fetch_dirty = init_one_libfunc ("__cache_fetch_dirty"); | |
4424 | emit_library_call_value (cache_fetch_dirty, data_addr, LCT_NORMAL, Pmode, | |
4425 | 2, ea_addr, EAmode, ndirty, SImode); | |
4426 | } | |
4427 | else | |
4428 | { | |
4429 | if (!cache_fetch) | |
4430 | cache_fetch = init_one_libfunc ("__cache_fetch"); | |
4431 | emit_library_call_value (cache_fetch, data_addr, LCT_NORMAL, Pmode, | |
4432 | 1, ea_addr, EAmode); | |
4433 | } | |
4434 | } | |
4435 | ||
4436 | /* Like ea_load_store, but do the cache tag comparison and, for stores, | |
4437 | dirty bit marking, inline. | |
4438 | ||
4439 | The cache control data structure is an array of | |
4440 | ||
4441 | struct __cache_tag_array | |
4442 | { | |
4443 | unsigned int tag_lo[4]; | |
4444 | unsigned int tag_hi[4]; | |
4445 | void *data_pointer[4]; | |
4446 | int reserved[4]; | |
4447 | vector unsigned short dirty_bits[4]; | |
4448 | } */ | |
4449 | ||
4450 | static void | |
4451 | ea_load_store_inline (rtx mem, bool is_store, rtx ea_addr, rtx data_addr) | |
4452 | { | |
4453 | rtx ea_addr_si; | |
4454 | HOST_WIDE_INT v; | |
4455 | rtx tag_size_sym = gen_rtx_SYMBOL_REF (Pmode, "__cache_tag_array_size"); | |
4456 | rtx tag_arr_sym = gen_rtx_SYMBOL_REF (Pmode, "__cache_tag_array"); | |
4457 | rtx index_mask = gen_reg_rtx (SImode); | |
4458 | rtx tag_arr = gen_reg_rtx (Pmode); | |
4459 | rtx splat_mask = gen_reg_rtx (TImode); | |
4460 | rtx splat = gen_reg_rtx (V4SImode); | |
4461 | rtx splat_hi = NULL_RTX; | |
4462 | rtx tag_index = gen_reg_rtx (Pmode); | |
4463 | rtx block_off = gen_reg_rtx (SImode); | |
4464 | rtx tag_addr = gen_reg_rtx (Pmode); | |
4465 | rtx tag = gen_reg_rtx (V4SImode); | |
4466 | rtx cache_tag = gen_reg_rtx (V4SImode); | |
4467 | rtx cache_tag_hi = NULL_RTX; | |
4468 | rtx cache_ptrs = gen_reg_rtx (TImode); | |
4469 | rtx cache_ptrs_si = gen_reg_rtx (SImode); | |
4470 | rtx tag_equal = gen_reg_rtx (V4SImode); | |
4471 | rtx tag_equal_hi = NULL_RTX; | |
4472 | rtx tag_eq_pack = gen_reg_rtx (V4SImode); | |
4473 | rtx tag_eq_pack_si = gen_reg_rtx (SImode); | |
4474 | rtx eq_index = gen_reg_rtx (SImode); | |
4475 | rtx bcomp, hit_label, hit_ref, cont_label, insn; | |
4476 | ||
4477 | if (spu_ea_model != 32) | |
4478 | { | |
4479 | splat_hi = gen_reg_rtx (V4SImode); | |
4480 | cache_tag_hi = gen_reg_rtx (V4SImode); | |
4481 | tag_equal_hi = gen_reg_rtx (V4SImode); | |
4482 | } | |
4483 | ||
4484 | emit_move_insn (index_mask, plus_constant (tag_size_sym, -128)); | |
4485 | emit_move_insn (tag_arr, tag_arr_sym); | |
4486 | v = 0x0001020300010203LL; | |
4487 | emit_move_insn (splat_mask, immed_double_const (v, v, TImode)); | |
4488 | ea_addr_si = ea_addr; | |
4489 | if (spu_ea_model != 32) | |
4490 | ea_addr_si = convert_to_mode (SImode, ea_addr, 1); | |
4491 | ||
4492 | /* tag_index = ea_addr & (tag_array_size - 128) */ | |
4493 | emit_insn (gen_andsi3 (tag_index, ea_addr_si, index_mask)); | |
4494 | ||
4495 | /* splat ea_addr to all 4 slots. */ | |
4496 | emit_insn (gen_shufb (splat, ea_addr_si, ea_addr_si, splat_mask)); | |
4497 | /* Similarly for high 32 bits of ea_addr. */ | |
4498 | if (spu_ea_model != 32) | |
4499 | emit_insn (gen_shufb (splat_hi, ea_addr, ea_addr, splat_mask)); | |
4500 | ||
4501 | /* block_off = ea_addr & 127 */ | |
4502 | emit_insn (gen_andsi3 (block_off, ea_addr_si, spu_const (SImode, 127))); | |
4503 | ||
4504 | /* tag_addr = tag_arr + tag_index */ | |
4505 | emit_insn (gen_addsi3 (tag_addr, tag_arr, tag_index)); | |
4506 | ||
4507 | /* Read cache tags. */ | |
4508 | emit_move_insn (cache_tag, gen_rtx_MEM (V4SImode, tag_addr)); | |
4509 | if (spu_ea_model != 32) | |
4510 | emit_move_insn (cache_tag_hi, gen_rtx_MEM (V4SImode, | |
4511 | plus_constant (tag_addr, 16))); | |
4512 | ||
4513 | /* tag = ea_addr & -128 */ | |
4514 | emit_insn (gen_andv4si3 (tag, splat, spu_const (V4SImode, -128))); | |
4515 | ||
4516 | /* Read all four cache data pointers. */ | |
4517 | emit_move_insn (cache_ptrs, gen_rtx_MEM (TImode, | |
4518 | plus_constant (tag_addr, 32))); | |
4519 | ||
4520 | /* Compare tags. */ | |
4521 | emit_insn (gen_ceq_v4si (tag_equal, tag, cache_tag)); | |
4522 | if (spu_ea_model != 32) | |
4523 | { | |
4524 | emit_insn (gen_ceq_v4si (tag_equal_hi, splat_hi, cache_tag_hi)); | |
4525 | emit_insn (gen_andv4si3 (tag_equal, tag_equal, tag_equal_hi)); | |
4526 | } | |
4527 | ||
4528 | /* At most one of the tags compare equal, so tag_equal has one | |
4529 | 32-bit slot set to all 1's, with the other slots all zero. | |
4530 | gbb picks off low bit from each byte in the 128-bit registers, | |
4531 | so tag_eq_pack is one of 0xf000, 0x0f00, 0x00f0, 0x000f, assuming | |
4532 | we have a hit. */ | |
4533 | emit_insn (gen_spu_gbb (tag_eq_pack, spu_gen_subreg (V16QImode, tag_equal))); | |
4534 | emit_insn (gen_spu_convert (tag_eq_pack_si, tag_eq_pack)); | |
4535 | ||
4536 | /* So counting leading zeros will set eq_index to 16, 20, 24 or 28. */ | |
4537 | emit_insn (gen_clzsi2 (eq_index, tag_eq_pack_si)); | |
4538 | ||
4539 | /* Allowing us to rotate the corresponding cache data pointer to slot0. | |
4540 | (rotating eq_index mod 16 bytes). */ | |
4541 | emit_insn (gen_rotqby_ti (cache_ptrs, cache_ptrs, eq_index)); | |
4542 | emit_insn (gen_spu_convert (cache_ptrs_si, cache_ptrs)); | |
4543 | ||
4544 | /* Add block offset to form final data address. */ | |
4545 | emit_insn (gen_addsi3 (data_addr, cache_ptrs_si, block_off)); | |
4546 | ||
4547 | /* Check that we did hit. */ | |
4548 | hit_label = gen_label_rtx (); | |
4549 | hit_ref = gen_rtx_LABEL_REF (VOIDmode, hit_label); | |
4550 | bcomp = gen_rtx_NE (SImode, tag_eq_pack_si, const0_rtx); | |
4551 | insn = emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, | |
4552 | gen_rtx_IF_THEN_ELSE (VOIDmode, bcomp, | |
4553 | hit_ref, pc_rtx))); | |
4554 | /* Say that this branch is very likely to happen. */ | |
4555 | v = REG_BR_PROB_BASE - REG_BR_PROB_BASE / 100 - 1; | |
02501f7f | 4556 | add_reg_note (insn, REG_BR_PROB, GEN_INT (v)); |
6cf5579e | 4557 | |
4558 | ea_load_store (mem, is_store, ea_addr, data_addr); | |
4559 | cont_label = gen_label_rtx (); | |
4560 | emit_jump_insn (gen_jump (cont_label)); | |
4561 | emit_barrier (); | |
4562 | ||
4563 | emit_label (hit_label); | |
4564 | ||
4565 | if (is_store) | |
4566 | { | |
4567 | HOST_WIDE_INT v_hi; | |
4568 | rtx dirty_bits = gen_reg_rtx (TImode); | |
4569 | rtx dirty_off = gen_reg_rtx (SImode); | |
4570 | rtx dirty_128 = gen_reg_rtx (TImode); | |
4571 | rtx neg_block_off = gen_reg_rtx (SImode); | |
4572 | ||
4573 | /* Set up mask with one dirty bit per byte of the mem we are | |
4574 | writing, starting from top bit. */ | |
4575 | v_hi = v = -1; | |
4576 | v <<= (128 - GET_MODE_SIZE (GET_MODE (mem))) & 63; | |
4577 | if ((128 - GET_MODE_SIZE (GET_MODE (mem))) >= 64) | |
4578 | { | |
4579 | v_hi = v; | |
4580 | v = 0; | |
4581 | } | |
4582 | emit_move_insn (dirty_bits, immed_double_const (v, v_hi, TImode)); | |
4583 | ||
4584 | /* Form index into cache dirty_bits. eq_index is one of | |
4585 | 0x10, 0x14, 0x18 or 0x1c. Multiplying by 4 gives us | |
4586 | 0x40, 0x50, 0x60 or 0x70 which just happens to be the | |
4587 | offset to each of the four dirty_bits elements. */ | |
4588 | emit_insn (gen_ashlsi3 (dirty_off, eq_index, spu_const (SImode, 2))); | |
4589 | ||
4590 | emit_insn (gen_spu_lqx (dirty_128, tag_addr, dirty_off)); | |
4591 | ||
4592 | /* Rotate bit mask to proper bit. */ | |
4593 | emit_insn (gen_negsi2 (neg_block_off, block_off)); | |
4594 | emit_insn (gen_rotqbybi_ti (dirty_bits, dirty_bits, neg_block_off)); | |
4595 | emit_insn (gen_rotqbi_ti (dirty_bits, dirty_bits, neg_block_off)); | |
4596 | ||
4597 | /* Or in the new dirty bits. */ | |
4598 | emit_insn (gen_iorti3 (dirty_128, dirty_bits, dirty_128)); | |
4599 | ||
4600 | /* Store. */ | |
4601 | emit_insn (gen_spu_stqx (dirty_128, tag_addr, dirty_off)); | |
4602 | } | |
4603 | ||
4604 | emit_label (cont_label); | |
4605 | } | |
4606 | ||
4607 | static rtx | |
4608 | expand_ea_mem (rtx mem, bool is_store) | |
4609 | { | |
4610 | rtx ea_addr; | |
4611 | rtx data_addr = gen_reg_rtx (Pmode); | |
4612 | rtx new_mem; | |
4613 | ||
4614 | ea_addr = force_reg (EAmode, XEXP (mem, 0)); | |
4615 | if (optimize_size || optimize == 0) | |
4616 | ea_load_store (mem, is_store, ea_addr, data_addr); | |
4617 | else | |
4618 | ea_load_store_inline (mem, is_store, ea_addr, data_addr); | |
4619 | ||
4620 | if (ea_alias_set == -1) | |
4621 | ea_alias_set = new_alias_set (); | |
4622 | ||
4623 | /* We generate a new MEM RTX to refer to the copy of the data | |
4624 | in the cache. We do not copy memory attributes (except the | |
4625 | alignment) from the original MEM, as they may no longer apply | |
4626 | to the cache copy. */ | |
4627 | new_mem = gen_rtx_MEM (GET_MODE (mem), data_addr); | |
4628 | set_mem_alias_set (new_mem, ea_alias_set); | |
4629 | set_mem_align (new_mem, MIN (MEM_ALIGN (mem), 128 * 8)); | |
4630 | ||
4631 | return new_mem; | |
4632 | } | |
4633 | ||
644459d0 | 4634 | int |
4635 | spu_expand_mov (rtx * ops, enum machine_mode mode) | |
4636 | { | |
4637 | if (GET_CODE (ops[0]) == SUBREG && !valid_subreg (ops[0])) | |
abe960bb | 4638 | { |
4639 | /* Perform the move in the destination SUBREG's inner mode. */ | |
4640 | ops[0] = SUBREG_REG (ops[0]); | |
4641 | mode = GET_MODE (ops[0]); | |
4642 | ops[1] = gen_lowpart_common (mode, ops[1]); | |
4643 | gcc_assert (ops[1]); | |
4644 | } | |
644459d0 | 4645 | |
4646 | if (GET_CODE (ops[1]) == SUBREG && !valid_subreg (ops[1])) | |
4647 | { | |
4648 | rtx from = SUBREG_REG (ops[1]); | |
8d72495d | 4649 | enum machine_mode imode = int_mode_for_mode (GET_MODE (from)); |
644459d0 | 4650 | |
4651 | gcc_assert (GET_MODE_CLASS (mode) == MODE_INT | |
4652 | && GET_MODE_CLASS (imode) == MODE_INT | |
4653 | && subreg_lowpart_p (ops[1])); | |
4654 | ||
4655 | if (GET_MODE_SIZE (imode) < 4) | |
8d72495d | 4656 | imode = SImode; |
4657 | if (imode != GET_MODE (from)) | |
4658 | from = gen_rtx_SUBREG (imode, from, 0); | |
644459d0 | 4659 | |
4660 | if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (imode)) | |
4661 | { | |
d6bf3b14 | 4662 | enum insn_code icode = convert_optab_handler (trunc_optab, |
4663 | mode, imode); | |
644459d0 | 4664 | emit_insn (GEN_FCN (icode) (ops[0], from)); |
4665 | } | |
4666 | else | |
4667 | emit_insn (gen_extend_insn (ops[0], from, mode, imode, 1)); | |
4668 | return 1; | |
4669 | } | |
4670 | ||
4671 | /* At least one of the operands needs to be a register. */ | |
4672 | if ((reload_in_progress | reload_completed) == 0 | |
4673 | && !register_operand (ops[0], mode) && !register_operand (ops[1], mode)) | |
4674 | { | |
4675 | rtx temp = force_reg (mode, ops[1]); | |
4676 | emit_move_insn (ops[0], temp); | |
4677 | return 1; | |
4678 | } | |
4679 | if (reload_in_progress || reload_completed) | |
4680 | { | |
dea01258 | 4681 | if (CONSTANT_P (ops[1])) |
4682 | return spu_split_immediate (ops); | |
644459d0 | 4683 | return 0; |
4684 | } | |
9d98604b | 4685 | |
4686 | /* Catch the SImode immediates greater than 0x7fffffff, and sign | |
4687 | extend them. */ | |
4688 | if (GET_CODE (ops[1]) == CONST_INT) | |
644459d0 | 4689 | { |
9d98604b | 4690 | HOST_WIDE_INT val = trunc_int_for_mode (INTVAL (ops[1]), mode); |
4691 | if (val != INTVAL (ops[1])) | |
644459d0 | 4692 | { |
9d98604b | 4693 | emit_move_insn (ops[0], GEN_INT (val)); |
4694 | return 1; | |
644459d0 | 4695 | } |
4696 | } | |
9d98604b | 4697 | if (MEM_P (ops[0])) |
6cf5579e | 4698 | { |
4699 | if (MEM_ADDR_SPACE (ops[0])) | |
4700 | ops[0] = expand_ea_mem (ops[0], true); | |
4701 | return spu_split_store (ops); | |
4702 | } | |
9d98604b | 4703 | if (MEM_P (ops[1])) |
6cf5579e | 4704 | { |
4705 | if (MEM_ADDR_SPACE (ops[1])) | |
4706 | ops[1] = expand_ea_mem (ops[1], false); | |
4707 | return spu_split_load (ops); | |
4708 | } | |
9d98604b | 4709 | |
644459d0 | 4710 | return 0; |
4711 | } | |
4712 | ||
9d98604b | 4713 | static void |
4714 | spu_convert_move (rtx dst, rtx src) | |
644459d0 | 4715 | { |
9d98604b | 4716 | enum machine_mode mode = GET_MODE (dst); |
4717 | enum machine_mode int_mode = mode_for_size (GET_MODE_BITSIZE (mode), MODE_INT, 0); | |
4718 | rtx reg; | |
4719 | gcc_assert (GET_MODE (src) == TImode); | |
4720 | reg = int_mode != mode ? gen_reg_rtx (int_mode) : dst; | |
4721 | emit_insn (gen_rtx_SET (VOIDmode, reg, | |
4722 | gen_rtx_TRUNCATE (int_mode, | |
4723 | gen_rtx_LSHIFTRT (TImode, src, | |
4724 | GEN_INT (int_mode == DImode ? 64 : 96))))); | |
4725 | if (int_mode != mode) | |
4726 | { | |
4727 | reg = simplify_gen_subreg (mode, reg, int_mode, 0); | |
4728 | emit_move_insn (dst, reg); | |
4729 | } | |
4730 | } | |
644459d0 | 4731 | |
9d98604b | 4732 | /* Load TImode values into DST0 and DST1 (when it is non-NULL) using |
4733 | the address from SRC and SRC+16. Return a REG or CONST_INT that | |
4734 | specifies how many bytes to rotate the loaded registers, plus any | |
4735 | extra from EXTRA_ROTQBY. The address and rotate amounts are | |
4736 | normalized to improve merging of loads and rotate computations. */ | |
4737 | static rtx | |
4738 | spu_expand_load (rtx dst0, rtx dst1, rtx src, int extra_rotby) | |
4739 | { | |
4740 | rtx addr = XEXP (src, 0); | |
4741 | rtx p0, p1, rot, addr0, addr1; | |
4742 | int rot_amt; | |
644459d0 | 4743 | |
4744 | rot = 0; | |
4745 | rot_amt = 0; | |
9d98604b | 4746 | |
4747 | if (MEM_ALIGN (src) >= 128) | |
4748 | /* Address is already aligned; simply perform a TImode load. */ ; | |
4749 | else if (GET_CODE (addr) == PLUS) | |
644459d0 | 4750 | { |
4751 | /* 8 cases: | |
4752 | aligned reg + aligned reg => lqx | |
4753 | aligned reg + unaligned reg => lqx, rotqby | |
4754 | aligned reg + aligned const => lqd | |
4755 | aligned reg + unaligned const => lqd, rotqbyi | |
4756 | unaligned reg + aligned reg => lqx, rotqby | |
4757 | unaligned reg + unaligned reg => lqx, a, rotqby (1 scratch) | |
4758 | unaligned reg + aligned const => lqd, rotqby | |
4759 | unaligned reg + unaligned const -> not allowed by legitimate address | |
4760 | */ | |
4761 | p0 = XEXP (addr, 0); | |
4762 | p1 = XEXP (addr, 1); | |
9d98604b | 4763 | if (!reg_aligned_for_addr (p0)) |
644459d0 | 4764 | { |
9d98604b | 4765 | if (REG_P (p1) && !reg_aligned_for_addr (p1)) |
644459d0 | 4766 | { |
9d98604b | 4767 | rot = gen_reg_rtx (SImode); |
4768 | emit_insn (gen_addsi3 (rot, p0, p1)); | |
4769 | } | |
4770 | else if (GET_CODE (p1) == CONST_INT && (INTVAL (p1) & 15)) | |
4771 | { | |
4772 | if (INTVAL (p1) > 0 | |
4773 | && REG_POINTER (p0) | |
4774 | && INTVAL (p1) * BITS_PER_UNIT | |
4775 | < REGNO_POINTER_ALIGN (REGNO (p0))) | |
4776 | { | |
4777 | rot = gen_reg_rtx (SImode); | |
4778 | emit_insn (gen_addsi3 (rot, p0, p1)); | |
4779 | addr = p0; | |
4780 | } | |
4781 | else | |
4782 | { | |
4783 | rtx x = gen_reg_rtx (SImode); | |
4784 | emit_move_insn (x, p1); | |
4785 | if (!spu_arith_operand (p1, SImode)) | |
4786 | p1 = x; | |
4787 | rot = gen_reg_rtx (SImode); | |
4788 | emit_insn (gen_addsi3 (rot, p0, p1)); | |
4789 | addr = gen_rtx_PLUS (Pmode, p0, x); | |
4790 | } | |
644459d0 | 4791 | } |
4792 | else | |
4793 | rot = p0; | |
4794 | } | |
4795 | else | |
4796 | { | |
4797 | if (GET_CODE (p1) == CONST_INT && (INTVAL (p1) & 15)) | |
4798 | { | |
4799 | rot_amt = INTVAL (p1) & 15; | |
9d98604b | 4800 | if (INTVAL (p1) & -16) |
4801 | { | |
4802 | p1 = GEN_INT (INTVAL (p1) & -16); | |
4803 | addr = gen_rtx_PLUS (SImode, p0, p1); | |
4804 | } | |
4805 | else | |
4806 | addr = p0; | |
644459d0 | 4807 | } |
9d98604b | 4808 | else if (REG_P (p1) && !reg_aligned_for_addr (p1)) |
644459d0 | 4809 | rot = p1; |
4810 | } | |
4811 | } | |
9d98604b | 4812 | else if (REG_P (addr)) |
644459d0 | 4813 | { |
9d98604b | 4814 | if (!reg_aligned_for_addr (addr)) |
644459d0 | 4815 | rot = addr; |
4816 | } | |
4817 | else if (GET_CODE (addr) == CONST) | |
4818 | { | |
4819 | if (GET_CODE (XEXP (addr, 0)) == PLUS | |
4820 | && ALIGNED_SYMBOL_REF_P (XEXP (XEXP (addr, 0), 0)) | |
4821 | && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT) | |
4822 | { | |
4823 | rot_amt = INTVAL (XEXP (XEXP (addr, 0), 1)); | |
4824 | if (rot_amt & -16) | |
4825 | addr = gen_rtx_CONST (Pmode, | |
4826 | gen_rtx_PLUS (Pmode, | |
4827 | XEXP (XEXP (addr, 0), 0), | |
4828 | GEN_INT (rot_amt & -16))); | |
4829 | else | |
4830 | addr = XEXP (XEXP (addr, 0), 0); | |
4831 | } | |
4832 | else | |
9d98604b | 4833 | { |
4834 | rot = gen_reg_rtx (Pmode); | |
4835 | emit_move_insn (rot, addr); | |
4836 | } | |
644459d0 | 4837 | } |
4838 | else if (GET_CODE (addr) == CONST_INT) | |
4839 | { | |
4840 | rot_amt = INTVAL (addr); | |
4841 | addr = GEN_INT (rot_amt & -16); | |
4842 | } | |
4843 | else if (!ALIGNED_SYMBOL_REF_P (addr)) | |
9d98604b | 4844 | { |
4845 | rot = gen_reg_rtx (Pmode); | |
4846 | emit_move_insn (rot, addr); | |
4847 | } | |
644459d0 | 4848 | |
9d98604b | 4849 | rot_amt += extra_rotby; |
644459d0 | 4850 | |
4851 | rot_amt &= 15; | |
4852 | ||
4853 | if (rot && rot_amt) | |
4854 | { | |
9d98604b | 4855 | rtx x = gen_reg_rtx (SImode); |
4856 | emit_insn (gen_addsi3 (x, rot, GEN_INT (rot_amt))); | |
4857 | rot = x; | |
644459d0 | 4858 | rot_amt = 0; |
4859 | } | |
9d98604b | 4860 | if (!rot && rot_amt) |
4861 | rot = GEN_INT (rot_amt); | |
4862 | ||
4863 | addr0 = copy_rtx (addr); | |
4864 | addr0 = gen_rtx_AND (SImode, copy_rtx (addr), GEN_INT (-16)); | |
4865 | emit_insn (gen__movti (dst0, change_address (src, TImode, addr0))); | |
4866 | ||
4867 | if (dst1) | |
4868 | { | |
4869 | addr1 = plus_constant (copy_rtx (addr), 16); | |
4870 | addr1 = gen_rtx_AND (SImode, addr1, GEN_INT (-16)); | |
4871 | emit_insn (gen__movti (dst1, change_address (src, TImode, addr1))); | |
4872 | } | |
644459d0 | 4873 | |
9d98604b | 4874 | return rot; |
4875 | } | |
4876 | ||
4877 | int | |
4878 | spu_split_load (rtx * ops) | |
4879 | { | |
4880 | enum machine_mode mode = GET_MODE (ops[0]); | |
4881 | rtx addr, load, rot; | |
4882 | int rot_amt; | |
644459d0 | 4883 | |
9d98604b | 4884 | if (GET_MODE_SIZE (mode) >= 16) |
4885 | return 0; | |
644459d0 | 4886 | |
9d98604b | 4887 | addr = XEXP (ops[1], 0); |
4888 | gcc_assert (GET_CODE (addr) != AND); | |
4889 | ||
4890 | if (!address_needs_split (ops[1])) | |
4891 | { | |
4892 | ops[1] = change_address (ops[1], TImode, addr); | |
4893 | load = gen_reg_rtx (TImode); | |
4894 | emit_insn (gen__movti (load, ops[1])); | |
4895 | spu_convert_move (ops[0], load); | |
4896 | return 1; | |
4897 | } | |
4898 | ||
4899 | rot_amt = GET_MODE_SIZE (mode) < 4 ? GET_MODE_SIZE (mode) - 4 : 0; | |
4900 | ||
4901 | load = gen_reg_rtx (TImode); | |
4902 | rot = spu_expand_load (load, 0, ops[1], rot_amt); | |
644459d0 | 4903 | |
4904 | if (rot) | |
4905 | emit_insn (gen_rotqby_ti (load, load, rot)); | |
644459d0 | 4906 | |
9d98604b | 4907 | spu_convert_move (ops[0], load); |
4908 | return 1; | |
644459d0 | 4909 | } |
4910 | ||
9d98604b | 4911 | int |
644459d0 | 4912 | spu_split_store (rtx * ops) |
4913 | { | |
4914 | enum machine_mode mode = GET_MODE (ops[0]); | |
9d98604b | 4915 | rtx reg; |
644459d0 | 4916 | rtx addr, p0, p1, p1_lo, smem; |
4917 | int aform; | |
4918 | int scalar; | |
4919 | ||
9d98604b | 4920 | if (GET_MODE_SIZE (mode) >= 16) |
4921 | return 0; | |
4922 | ||
644459d0 | 4923 | addr = XEXP (ops[0], 0); |
9d98604b | 4924 | gcc_assert (GET_CODE (addr) != AND); |
4925 | ||
4926 | if (!address_needs_split (ops[0])) | |
4927 | { | |
4928 | reg = gen_reg_rtx (TImode); | |
4929 | emit_insn (gen_spu_convert (reg, ops[1])); | |
4930 | ops[0] = change_address (ops[0], TImode, addr); | |
4931 | emit_move_insn (ops[0], reg); | |
4932 | return 1; | |
4933 | } | |
644459d0 | 4934 | |
4935 | if (GET_CODE (addr) == PLUS) | |
4936 | { | |
4937 | /* 8 cases: | |
4938 | aligned reg + aligned reg => lqx, c?x, shuf, stqx | |
4939 | aligned reg + unaligned reg => lqx, c?x, shuf, stqx | |
4940 | aligned reg + aligned const => lqd, c?d, shuf, stqx | |
4941 | aligned reg + unaligned const => lqd, c?d, shuf, stqx | |
4942 | unaligned reg + aligned reg => lqx, c?x, shuf, stqx | |
4943 | unaligned reg + unaligned reg => lqx, c?x, shuf, stqx | |
4944 | unaligned reg + aligned const => lqd, c?d, shuf, stqx | |
9d98604b | 4945 | unaligned reg + unaligned const -> lqx, c?d, shuf, stqx |
644459d0 | 4946 | */ |
4947 | aform = 0; | |
4948 | p0 = XEXP (addr, 0); | |
4949 | p1 = p1_lo = XEXP (addr, 1); | |
9d98604b | 4950 | if (REG_P (p0) && GET_CODE (p1) == CONST_INT) |
644459d0 | 4951 | { |
4952 | p1_lo = GEN_INT (INTVAL (p1) & 15); | |
9d98604b | 4953 | if (reg_aligned_for_addr (p0)) |
4954 | { | |
4955 | p1 = GEN_INT (INTVAL (p1) & -16); | |
4956 | if (p1 == const0_rtx) | |
4957 | addr = p0; | |
4958 | else | |
4959 | addr = gen_rtx_PLUS (SImode, p0, p1); | |
4960 | } | |
4961 | else | |
4962 | { | |
4963 | rtx x = gen_reg_rtx (SImode); | |
4964 | emit_move_insn (x, p1); | |
4965 | addr = gen_rtx_PLUS (SImode, p0, x); | |
4966 | } | |
644459d0 | 4967 | } |
4968 | } | |
9d98604b | 4969 | else if (REG_P (addr)) |
644459d0 | 4970 | { |
4971 | aform = 0; | |
4972 | p0 = addr; | |
4973 | p1 = p1_lo = const0_rtx; | |
4974 | } | |
4975 | else | |
4976 | { | |
4977 | aform = 1; | |
4978 | p0 = gen_rtx_REG (SImode, STACK_POINTER_REGNUM); | |
4979 | p1 = 0; /* aform doesn't use p1 */ | |
4980 | p1_lo = addr; | |
4981 | if (ALIGNED_SYMBOL_REF_P (addr)) | |
4982 | p1_lo = const0_rtx; | |
9d98604b | 4983 | else if (GET_CODE (addr) == CONST |
4984 | && GET_CODE (XEXP (addr, 0)) == PLUS | |
4985 | && ALIGNED_SYMBOL_REF_P (XEXP (XEXP (addr, 0), 0)) | |
4986 | && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT) | |
644459d0 | 4987 | { |
9d98604b | 4988 | HOST_WIDE_INT v = INTVAL (XEXP (XEXP (addr, 0), 1)); |
4989 | if ((v & -16) != 0) | |
4990 | addr = gen_rtx_CONST (Pmode, | |
4991 | gen_rtx_PLUS (Pmode, | |
4992 | XEXP (XEXP (addr, 0), 0), | |
4993 | GEN_INT (v & -16))); | |
4994 | else | |
4995 | addr = XEXP (XEXP (addr, 0), 0); | |
4996 | p1_lo = GEN_INT (v & 15); | |
644459d0 | 4997 | } |
4998 | else if (GET_CODE (addr) == CONST_INT) | |
4999 | { | |
5000 | p1_lo = GEN_INT (INTVAL (addr) & 15); | |
5001 | addr = GEN_INT (INTVAL (addr) & -16); | |
5002 | } | |
9d98604b | 5003 | else |
5004 | { | |
5005 | p1_lo = gen_reg_rtx (SImode); | |
5006 | emit_move_insn (p1_lo, addr); | |
5007 | } | |
644459d0 | 5008 | } |
5009 | ||
4cbad5bb | 5010 | gcc_assert (aform == 0 || aform == 1); |
9d98604b | 5011 | reg = gen_reg_rtx (TImode); |
e04cf423 | 5012 | |
644459d0 | 5013 | scalar = store_with_one_insn_p (ops[0]); |
5014 | if (!scalar) | |
5015 | { | |
5016 | /* We could copy the flags from the ops[0] MEM to mem here, | |
5017 | We don't because we want this load to be optimized away if | |
5018 | possible, and copying the flags will prevent that in certain | |
5019 | cases, e.g. consider the volatile flag. */ | |
5020 | ||
9d98604b | 5021 | rtx pat = gen_reg_rtx (TImode); |
e04cf423 | 5022 | rtx lmem = change_address (ops[0], TImode, copy_rtx (addr)); |
5023 | set_mem_alias_set (lmem, 0); | |
5024 | emit_insn (gen_movti (reg, lmem)); | |
644459d0 | 5025 | |
9d98604b | 5026 | if (!p0 || reg_aligned_for_addr (p0)) |
644459d0 | 5027 | p0 = stack_pointer_rtx; |
5028 | if (!p1_lo) | |
5029 | p1_lo = const0_rtx; | |
5030 | ||
5031 | emit_insn (gen_cpat (pat, p0, p1_lo, GEN_INT (GET_MODE_SIZE (mode)))); | |
5032 | emit_insn (gen_shufb (reg, ops[1], reg, pat)); | |
5033 | } | |
644459d0 | 5034 | else |
5035 | { | |
5036 | if (GET_CODE (ops[1]) == REG) | |
5037 | emit_insn (gen_spu_convert (reg, ops[1])); | |
5038 | else if (GET_CODE (ops[1]) == SUBREG) | |
5039 | emit_insn (gen_spu_convert (reg, SUBREG_REG (ops[1]))); | |
5040 | else | |
5041 | abort (); | |
5042 | } | |
5043 | ||
5044 | if (GET_MODE_SIZE (mode) < 4 && scalar) | |
9d98604b | 5045 | emit_insn (gen_ashlti3 |
5046 | (reg, reg, GEN_INT (32 - GET_MODE_BITSIZE (mode)))); | |
644459d0 | 5047 | |
9d98604b | 5048 | smem = change_address (ops[0], TImode, copy_rtx (addr)); |
644459d0 | 5049 | /* We can't use the previous alias set because the memory has changed |
5050 | size and can potentially overlap objects of other types. */ | |
5051 | set_mem_alias_set (smem, 0); | |
5052 | ||
e04cf423 | 5053 | emit_insn (gen_movti (smem, reg)); |
9d98604b | 5054 | return 1; |
644459d0 | 5055 | } |
5056 | ||
5057 | /* Return TRUE if X is MEM which is a struct member reference | |
5058 | and the member can safely be loaded and stored with a single | |
5059 | instruction because it is padded. */ | |
5060 | static int | |
5061 | mem_is_padded_component_ref (rtx x) | |
5062 | { | |
5063 | tree t = MEM_EXPR (x); | |
5064 | tree r; | |
5065 | if (!t || TREE_CODE (t) != COMPONENT_REF) | |
5066 | return 0; | |
5067 | t = TREE_OPERAND (t, 1); | |
5068 | if (!t || TREE_CODE (t) != FIELD_DECL | |
5069 | || DECL_ALIGN (t) < 128 || AGGREGATE_TYPE_P (TREE_TYPE (t))) | |
5070 | return 0; | |
5071 | /* Only do this for RECORD_TYPEs, not UNION_TYPEs. */ | |
5072 | r = DECL_FIELD_CONTEXT (t); | |
5073 | if (!r || TREE_CODE (r) != RECORD_TYPE) | |
5074 | return 0; | |
5075 | /* Make sure they are the same mode */ | |
5076 | if (GET_MODE (x) != TYPE_MODE (TREE_TYPE (t))) | |
5077 | return 0; | |
5078 | /* If there are no following fields then the field alignment assures | |
fa7637bd | 5079 | the structure is padded to the alignment which means this field is |
5080 | padded too. */ | |
644459d0 | 5081 | if (TREE_CHAIN (t) == 0) |
5082 | return 1; | |
5083 | /* If the following field is also aligned then this field will be | |
5084 | padded. */ | |
5085 | t = TREE_CHAIN (t); | |
5086 | if (TREE_CODE (t) == FIELD_DECL && DECL_ALIGN (t) >= 128) | |
5087 | return 1; | |
5088 | return 0; | |
5089 | } | |
5090 | ||
c7b91b14 | 5091 | /* Parse the -mfixed-range= option string. */ |
5092 | static void | |
5093 | fix_range (const char *const_str) | |
5094 | { | |
5095 | int i, first, last; | |
5096 | char *str, *dash, *comma; | |
5097 | ||
5098 | /* str must be of the form REG1'-'REG2{,REG1'-'REG} where REG1 and | |
5099 | REG2 are either register names or register numbers. The effect | |
5100 | of this option is to mark the registers in the range from REG1 to | |
5101 | REG2 as ``fixed'' so they won't be used by the compiler. */ | |
5102 | ||
5103 | i = strlen (const_str); | |
5104 | str = (char *) alloca (i + 1); | |
5105 | memcpy (str, const_str, i + 1); | |
5106 | ||
5107 | while (1) | |
5108 | { | |
5109 | dash = strchr (str, '-'); | |
5110 | if (!dash) | |
5111 | { | |
5112 | warning (0, "value of -mfixed-range must have form REG1-REG2"); | |
5113 | return; | |
5114 | } | |
5115 | *dash = '\0'; | |
5116 | comma = strchr (dash + 1, ','); | |
5117 | if (comma) | |
5118 | *comma = '\0'; | |
5119 | ||
5120 | first = decode_reg_name (str); | |
5121 | if (first < 0) | |
5122 | { | |
5123 | warning (0, "unknown register name: %s", str); | |
5124 | return; | |
5125 | } | |
5126 | ||
5127 | last = decode_reg_name (dash + 1); | |
5128 | if (last < 0) | |
5129 | { | |
5130 | warning (0, "unknown register name: %s", dash + 1); | |
5131 | return; | |
5132 | } | |
5133 | ||
5134 | *dash = '-'; | |
5135 | ||
5136 | if (first > last) | |
5137 | { | |
5138 | warning (0, "%s-%s is an empty range", str, dash + 1); | |
5139 | return; | |
5140 | } | |
5141 | ||
5142 | for (i = first; i <= last; ++i) | |
5143 | fixed_regs[i] = call_used_regs[i] = 1; | |
5144 | ||
5145 | if (!comma) | |
5146 | break; | |
5147 | ||
5148 | *comma = ','; | |
5149 | str = comma + 1; | |
5150 | } | |
5151 | } | |
5152 | ||
644459d0 | 5153 | /* Return TRUE if x is a CONST_INT, CONST_DOUBLE or CONST_VECTOR that |
5154 | can be generated using the fsmbi instruction. */ | |
5155 | int | |
5156 | fsmbi_const_p (rtx x) | |
5157 | { | |
dea01258 | 5158 | if (CONSTANT_P (x)) |
5159 | { | |
5df189be | 5160 | /* We can always choose TImode for CONST_INT because the high bits |
dea01258 | 5161 | of an SImode will always be all 1s, i.e., valid for fsmbi. */ |
5df189be | 5162 | enum immediate_class c = classify_immediate (x, TImode); |
3072d30e | 5163 | return c == IC_FSMBI || (!epilogue_completed && c == IC_FSMBI2); |
dea01258 | 5164 | } |
5165 | return 0; | |
5166 | } | |
5167 | ||
5168 | /* Return TRUE if x is a CONST_INT, CONST_DOUBLE or CONST_VECTOR that | |
5169 | can be generated using the cbd, chd, cwd or cdd instruction. */ | |
5170 | int | |
5171 | cpat_const_p (rtx x, enum machine_mode mode) | |
5172 | { | |
5173 | if (CONSTANT_P (x)) | |
5174 | { | |
5175 | enum immediate_class c = classify_immediate (x, mode); | |
5176 | return c == IC_CPAT; | |
5177 | } | |
5178 | return 0; | |
5179 | } | |
644459d0 | 5180 | |
dea01258 | 5181 | rtx |
5182 | gen_cpat_const (rtx * ops) | |
5183 | { | |
5184 | unsigned char dst[16]; | |
5185 | int i, offset, shift, isize; | |
5186 | if (GET_CODE (ops[3]) != CONST_INT | |
5187 | || GET_CODE (ops[2]) != CONST_INT | |
5188 | || (GET_CODE (ops[1]) != CONST_INT | |
5189 | && GET_CODE (ops[1]) != REG)) | |
5190 | return 0; | |
5191 | if (GET_CODE (ops[1]) == REG | |
5192 | && (!REG_POINTER (ops[1]) | |
5193 | || REGNO_POINTER_ALIGN (ORIGINAL_REGNO (ops[1])) < 128)) | |
5194 | return 0; | |
644459d0 | 5195 | |
5196 | for (i = 0; i < 16; i++) | |
dea01258 | 5197 | dst[i] = i + 16; |
5198 | isize = INTVAL (ops[3]); | |
5199 | if (isize == 1) | |
5200 | shift = 3; | |
5201 | else if (isize == 2) | |
5202 | shift = 2; | |
5203 | else | |
5204 | shift = 0; | |
5205 | offset = (INTVAL (ops[2]) + | |
5206 | (GET_CODE (ops[1]) == | |
5207 | CONST_INT ? INTVAL (ops[1]) : 0)) & 15; | |
5208 | for (i = 0; i < isize; i++) | |
5209 | dst[offset + i] = i + shift; | |
5210 | return array_to_constant (TImode, dst); | |
644459d0 | 5211 | } |
5212 | ||
5213 | /* Convert a CONST_INT, CONST_DOUBLE, or CONST_VECTOR into a 16 byte | |
5214 | array. Use MODE for CONST_INT's. When the constant's mode is smaller | |
5215 | than 16 bytes, the value is repeated across the rest of the array. */ | |
5216 | void | |
5217 | constant_to_array (enum machine_mode mode, rtx x, unsigned char arr[16]) | |
5218 | { | |
5219 | HOST_WIDE_INT val; | |
5220 | int i, j, first; | |
5221 | ||
5222 | memset (arr, 0, 16); | |
5223 | mode = GET_MODE (x) != VOIDmode ? GET_MODE (x) : mode; | |
5224 | if (GET_CODE (x) == CONST_INT | |
5225 | || (GET_CODE (x) == CONST_DOUBLE | |
5226 | && (mode == SFmode || mode == DFmode))) | |
5227 | { | |
5228 | gcc_assert (mode != VOIDmode && mode != BLKmode); | |
5229 | ||
5230 | if (GET_CODE (x) == CONST_DOUBLE) | |
5231 | val = const_double_to_hwint (x); | |
5232 | else | |
5233 | val = INTVAL (x); | |
5234 | first = GET_MODE_SIZE (mode) - 1; | |
5235 | for (i = first; i >= 0; i--) | |
5236 | { | |
5237 | arr[i] = val & 0xff; | |
5238 | val >>= 8; | |
5239 | } | |
5240 | /* Splat the constant across the whole array. */ | |
5241 | for (j = 0, i = first + 1; i < 16; i++) | |
5242 | { | |
5243 | arr[i] = arr[j]; | |
5244 | j = (j == first) ? 0 : j + 1; | |
5245 | } | |
5246 | } | |
5247 | else if (GET_CODE (x) == CONST_DOUBLE) | |
5248 | { | |
5249 | val = CONST_DOUBLE_LOW (x); | |
5250 | for (i = 15; i >= 8; i--) | |
5251 | { | |
5252 | arr[i] = val & 0xff; | |
5253 | val >>= 8; | |
5254 | } | |
5255 | val = CONST_DOUBLE_HIGH (x); | |
5256 | for (i = 7; i >= 0; i--) | |
5257 | { | |
5258 | arr[i] = val & 0xff; | |
5259 | val >>= 8; | |
5260 | } | |
5261 | } | |
5262 | else if (GET_CODE (x) == CONST_VECTOR) | |
5263 | { | |
5264 | int units; | |
5265 | rtx elt; | |
5266 | mode = GET_MODE_INNER (mode); | |
5267 | units = CONST_VECTOR_NUNITS (x); | |
5268 | for (i = 0; i < units; i++) | |
5269 | { | |
5270 | elt = CONST_VECTOR_ELT (x, i); | |
5271 | if (GET_CODE (elt) == CONST_INT || GET_CODE (elt) == CONST_DOUBLE) | |
5272 | { | |
5273 | if (GET_CODE (elt) == CONST_DOUBLE) | |
5274 | val = const_double_to_hwint (elt); | |
5275 | else | |
5276 | val = INTVAL (elt); | |
5277 | first = GET_MODE_SIZE (mode) - 1; | |
5278 | if (first + i * GET_MODE_SIZE (mode) > 16) | |
5279 | abort (); | |
5280 | for (j = first; j >= 0; j--) | |
5281 | { | |
5282 | arr[j + i * GET_MODE_SIZE (mode)] = val & 0xff; | |
5283 | val >>= 8; | |
5284 | } | |
5285 | } | |
5286 | } | |
5287 | } | |
5288 | else | |
5289 | gcc_unreachable(); | |
5290 | } | |
5291 | ||
5292 | /* Convert a 16 byte array to a constant of mode MODE. When MODE is | |
5293 | smaller than 16 bytes, use the bytes that would represent that value | |
5294 | in a register, e.g., for QImode return the value of arr[3]. */ | |
5295 | rtx | |
e96f2783 | 5296 | array_to_constant (enum machine_mode mode, const unsigned char arr[16]) |
644459d0 | 5297 | { |
5298 | enum machine_mode inner_mode; | |
5299 | rtvec v; | |
5300 | int units, size, i, j, k; | |
5301 | HOST_WIDE_INT val; | |
5302 | ||
5303 | if (GET_MODE_CLASS (mode) == MODE_INT | |
5304 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT) | |
5305 | { | |
5306 | j = GET_MODE_SIZE (mode); | |
5307 | i = j < 4 ? 4 - j : 0; | |
5308 | for (val = 0; i < j; i++) | |
5309 | val = (val << 8) | arr[i]; | |
5310 | val = trunc_int_for_mode (val, mode); | |
5311 | return GEN_INT (val); | |
5312 | } | |
5313 | ||
5314 | if (mode == TImode) | |
5315 | { | |
5316 | HOST_WIDE_INT high; | |
5317 | for (i = high = 0; i < 8; i++) | |
5318 | high = (high << 8) | arr[i]; | |
5319 | for (i = 8, val = 0; i < 16; i++) | |
5320 | val = (val << 8) | arr[i]; | |
5321 | return immed_double_const (val, high, TImode); | |
5322 | } | |
5323 | if (mode == SFmode) | |
5324 | { | |
5325 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
5326 | val = trunc_int_for_mode (val, SImode); | |
171b6d22 | 5327 | return hwint_to_const_double (SFmode, val); |
644459d0 | 5328 | } |
5329 | if (mode == DFmode) | |
5330 | { | |
1f915911 | 5331 | for (i = 0, val = 0; i < 8; i++) |
5332 | val = (val << 8) | arr[i]; | |
171b6d22 | 5333 | return hwint_to_const_double (DFmode, val); |
644459d0 | 5334 | } |
5335 | ||
5336 | if (!VECTOR_MODE_P (mode)) | |
5337 | abort (); | |
5338 | ||
5339 | units = GET_MODE_NUNITS (mode); | |
5340 | size = GET_MODE_UNIT_SIZE (mode); | |
5341 | inner_mode = GET_MODE_INNER (mode); | |
5342 | v = rtvec_alloc (units); | |
5343 | ||
5344 | for (k = i = 0; i < units; ++i) | |
5345 | { | |
5346 | val = 0; | |
5347 | for (j = 0; j < size; j++, k++) | |
5348 | val = (val << 8) | arr[k]; | |
5349 | ||
5350 | if (GET_MODE_CLASS (inner_mode) == MODE_FLOAT) | |
5351 | RTVEC_ELT (v, i) = hwint_to_const_double (inner_mode, val); | |
5352 | else | |
5353 | RTVEC_ELT (v, i) = GEN_INT (trunc_int_for_mode (val, inner_mode)); | |
5354 | } | |
5355 | if (k > 16) | |
5356 | abort (); | |
5357 | ||
5358 | return gen_rtx_CONST_VECTOR (mode, v); | |
5359 | } | |
5360 | ||
5361 | static void | |
5362 | reloc_diagnostic (rtx x) | |
5363 | { | |
712d2297 | 5364 | tree decl = 0; |
644459d0 | 5365 | if (!flag_pic || !(TARGET_WARN_RELOC || TARGET_ERROR_RELOC)) |
5366 | return; | |
5367 | ||
5368 | if (GET_CODE (x) == SYMBOL_REF) | |
5369 | decl = SYMBOL_REF_DECL (x); | |
5370 | else if (GET_CODE (x) == CONST | |
5371 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF) | |
5372 | decl = SYMBOL_REF_DECL (XEXP (XEXP (x, 0), 0)); | |
5373 | ||
5374 | /* SYMBOL_REF_DECL is not necessarily a DECL. */ | |
5375 | if (decl && !DECL_P (decl)) | |
5376 | decl = 0; | |
5377 | ||
644459d0 | 5378 | /* The decl could be a string constant. */ |
5379 | if (decl && DECL_P (decl)) | |
712d2297 | 5380 | { |
5381 | location_t loc; | |
5382 | /* We use last_assemble_variable_decl to get line information. It's | |
5383 | not always going to be right and might not even be close, but will | |
5384 | be right for the more common cases. */ | |
5385 | if (!last_assemble_variable_decl || in_section == ctors_section) | |
5386 | loc = DECL_SOURCE_LOCATION (decl); | |
5387 | else | |
5388 | loc = DECL_SOURCE_LOCATION (last_assemble_variable_decl); | |
644459d0 | 5389 | |
712d2297 | 5390 | if (TARGET_WARN_RELOC) |
5391 | warning_at (loc, 0, | |
5392 | "creating run-time relocation for %qD", decl); | |
5393 | else | |
5394 | error_at (loc, | |
5395 | "creating run-time relocation for %qD", decl); | |
5396 | } | |
5397 | else | |
5398 | { | |
5399 | if (TARGET_WARN_RELOC) | |
5400 | warning_at (input_location, 0, "creating run-time relocation"); | |
5401 | else | |
5402 | error_at (input_location, "creating run-time relocation"); | |
5403 | } | |
644459d0 | 5404 | } |
5405 | ||
5406 | /* Hook into assemble_integer so we can generate an error for run-time | |
5407 | relocations. The SPU ABI disallows them. */ | |
5408 | static bool | |
5409 | spu_assemble_integer (rtx x, unsigned int size, int aligned_p) | |
5410 | { | |
5411 | /* By default run-time relocations aren't supported, but we allow them | |
5412 | in case users support it in their own run-time loader. And we provide | |
5413 | a warning for those users that don't. */ | |
5414 | if ((GET_CODE (x) == SYMBOL_REF) | |
5415 | || GET_CODE (x) == LABEL_REF || GET_CODE (x) == CONST) | |
5416 | reloc_diagnostic (x); | |
5417 | ||
5418 | return default_assemble_integer (x, size, aligned_p); | |
5419 | } | |
5420 | ||
5421 | static void | |
5422 | spu_asm_globalize_label (FILE * file, const char *name) | |
5423 | { | |
5424 | fputs ("\t.global\t", file); | |
5425 | assemble_name (file, name); | |
5426 | fputs ("\n", file); | |
5427 | } | |
5428 | ||
5429 | static bool | |
f529eb25 | 5430 | spu_rtx_costs (rtx x, int code, int outer_code ATTRIBUTE_UNUSED, int *total, |
5431 | bool speed ATTRIBUTE_UNUSED) | |
644459d0 | 5432 | { |
5433 | enum machine_mode mode = GET_MODE (x); | |
5434 | int cost = COSTS_N_INSNS (2); | |
5435 | ||
5436 | /* Folding to a CONST_VECTOR will use extra space but there might | |
5437 | be only a small savings in cycles. We'd like to use a CONST_VECTOR | |
9505a73b | 5438 | only if it allows us to fold away multiple insns. Changing the cost |
644459d0 | 5439 | of a CONST_VECTOR here (or in CONST_COSTS) doesn't help though |
5440 | because this cost will only be compared against a single insn. | |
5441 | if (code == CONST_VECTOR) | |
5442 | return (LEGITIMATE_CONSTANT_P(x)) ? cost : COSTS_N_INSNS(6); | |
5443 | */ | |
5444 | ||
5445 | /* Use defaults for float operations. Not accurate but good enough. */ | |
5446 | if (mode == DFmode) | |
5447 | { | |
5448 | *total = COSTS_N_INSNS (13); | |
5449 | return true; | |
5450 | } | |
5451 | if (mode == SFmode) | |
5452 | { | |
5453 | *total = COSTS_N_INSNS (6); | |
5454 | return true; | |
5455 | } | |
5456 | switch (code) | |
5457 | { | |
5458 | case CONST_INT: | |
5459 | if (satisfies_constraint_K (x)) | |
5460 | *total = 0; | |
5461 | else if (INTVAL (x) >= -0x80000000ll && INTVAL (x) <= 0xffffffffll) | |
5462 | *total = COSTS_N_INSNS (1); | |
5463 | else | |
5464 | *total = COSTS_N_INSNS (3); | |
5465 | return true; | |
5466 | ||
5467 | case CONST: | |
5468 | *total = COSTS_N_INSNS (3); | |
5469 | return true; | |
5470 | ||
5471 | case LABEL_REF: | |
5472 | case SYMBOL_REF: | |
5473 | *total = COSTS_N_INSNS (0); | |
5474 | return true; | |
5475 | ||
5476 | case CONST_DOUBLE: | |
5477 | *total = COSTS_N_INSNS (5); | |
5478 | return true; | |
5479 | ||
5480 | case FLOAT_EXTEND: | |
5481 | case FLOAT_TRUNCATE: | |
5482 | case FLOAT: | |
5483 | case UNSIGNED_FLOAT: | |
5484 | case FIX: | |
5485 | case UNSIGNED_FIX: | |
5486 | *total = COSTS_N_INSNS (7); | |
5487 | return true; | |
5488 | ||
5489 | case PLUS: | |
5490 | if (mode == TImode) | |
5491 | { | |
5492 | *total = COSTS_N_INSNS (9); | |
5493 | return true; | |
5494 | } | |
5495 | break; | |
5496 | ||
5497 | case MULT: | |
5498 | cost = | |
5499 | GET_CODE (XEXP (x, 0)) == | |
5500 | REG ? COSTS_N_INSNS (12) : COSTS_N_INSNS (7); | |
5501 | if (mode == SImode && GET_CODE (XEXP (x, 0)) == REG) | |
5502 | { | |
5503 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
5504 | { | |
5505 | HOST_WIDE_INT val = INTVAL (XEXP (x, 1)); | |
5506 | cost = COSTS_N_INSNS (14); | |
5507 | if ((val & 0xffff) == 0) | |
5508 | cost = COSTS_N_INSNS (9); | |
5509 | else if (val > 0 && val < 0x10000) | |
5510 | cost = COSTS_N_INSNS (11); | |
5511 | } | |
5512 | } | |
5513 | *total = cost; | |
5514 | return true; | |
5515 | case DIV: | |
5516 | case UDIV: | |
5517 | case MOD: | |
5518 | case UMOD: | |
5519 | *total = COSTS_N_INSNS (20); | |
5520 | return true; | |
5521 | case ROTATE: | |
5522 | case ROTATERT: | |
5523 | case ASHIFT: | |
5524 | case ASHIFTRT: | |
5525 | case LSHIFTRT: | |
5526 | *total = COSTS_N_INSNS (4); | |
5527 | return true; | |
5528 | case UNSPEC: | |
5529 | if (XINT (x, 1) == UNSPEC_CONVERT) | |
5530 | *total = COSTS_N_INSNS (0); | |
5531 | else | |
5532 | *total = COSTS_N_INSNS (4); | |
5533 | return true; | |
5534 | } | |
5535 | /* Scale cost by mode size. Except when initializing (cfun->decl == 0). */ | |
5536 | if (GET_MODE_CLASS (mode) == MODE_INT | |
5537 | && GET_MODE_SIZE (mode) > GET_MODE_SIZE (SImode) && cfun && cfun->decl) | |
5538 | cost = cost * (GET_MODE_SIZE (mode) / GET_MODE_SIZE (SImode)) | |
5539 | * (GET_MODE_SIZE (mode) / GET_MODE_SIZE (SImode)); | |
5540 | *total = cost; | |
5541 | return true; | |
5542 | } | |
5543 | ||
1bd43494 | 5544 | static enum machine_mode |
5545 | spu_unwind_word_mode (void) | |
644459d0 | 5546 | { |
1bd43494 | 5547 | return SImode; |
644459d0 | 5548 | } |
5549 | ||
5550 | /* Decide whether we can make a sibling call to a function. DECL is the | |
5551 | declaration of the function being targeted by the call and EXP is the | |
5552 | CALL_EXPR representing the call. */ | |
5553 | static bool | |
5554 | spu_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED) | |
5555 | { | |
5556 | return decl && !TARGET_LARGE_MEM; | |
5557 | } | |
5558 | ||
5559 | /* We need to correctly update the back chain pointer and the Available | |
5560 | Stack Size (which is in the second slot of the sp register.) */ | |
5561 | void | |
5562 | spu_allocate_stack (rtx op0, rtx op1) | |
5563 | { | |
5564 | HOST_WIDE_INT v; | |
5565 | rtx chain = gen_reg_rtx (V4SImode); | |
5566 | rtx stack_bot = gen_frame_mem (V4SImode, stack_pointer_rtx); | |
5567 | rtx sp = gen_reg_rtx (V4SImode); | |
5568 | rtx splatted = gen_reg_rtx (V4SImode); | |
5569 | rtx pat = gen_reg_rtx (TImode); | |
5570 | ||
5571 | /* copy the back chain so we can save it back again. */ | |
5572 | emit_move_insn (chain, stack_bot); | |
5573 | ||
5574 | op1 = force_reg (SImode, op1); | |
5575 | ||
5576 | v = 0x1020300010203ll; | |
5577 | emit_move_insn (pat, immed_double_const (v, v, TImode)); | |
5578 | emit_insn (gen_shufb (splatted, op1, op1, pat)); | |
5579 | ||
5580 | emit_insn (gen_spu_convert (sp, stack_pointer_rtx)); | |
5581 | emit_insn (gen_subv4si3 (sp, sp, splatted)); | |
5582 | ||
5583 | if (flag_stack_check) | |
5584 | { | |
5585 | rtx avail = gen_reg_rtx(SImode); | |
5586 | rtx result = gen_reg_rtx(SImode); | |
5587 | emit_insn (gen_vec_extractv4si (avail, sp, GEN_INT (1))); | |
5588 | emit_insn (gen_cgt_si(result, avail, GEN_INT (-1))); | |
5589 | emit_insn (gen_spu_heq (result, GEN_INT(0) )); | |
5590 | } | |
5591 | ||
5592 | emit_insn (gen_spu_convert (stack_pointer_rtx, sp)); | |
5593 | ||
5594 | emit_move_insn (stack_bot, chain); | |
5595 | ||
5596 | emit_move_insn (op0, virtual_stack_dynamic_rtx); | |
5597 | } | |
5598 | ||
5599 | void | |
5600 | spu_restore_stack_nonlocal (rtx op0 ATTRIBUTE_UNUSED, rtx op1) | |
5601 | { | |
5602 | static unsigned char arr[16] = | |
5603 | { 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3 }; | |
5604 | rtx temp = gen_reg_rtx (SImode); | |
5605 | rtx temp2 = gen_reg_rtx (SImode); | |
5606 | rtx temp3 = gen_reg_rtx (V4SImode); | |
5607 | rtx temp4 = gen_reg_rtx (V4SImode); | |
5608 | rtx pat = gen_reg_rtx (TImode); | |
5609 | rtx sp = gen_rtx_REG (V4SImode, STACK_POINTER_REGNUM); | |
5610 | ||
5611 | /* Restore the backchain from the first word, sp from the second. */ | |
5612 | emit_move_insn (temp2, adjust_address_nv (op1, SImode, 0)); | |
5613 | emit_move_insn (temp, adjust_address_nv (op1, SImode, 4)); | |
5614 | ||
5615 | emit_move_insn (pat, array_to_constant (TImode, arr)); | |
5616 | ||
5617 | /* Compute Available Stack Size for sp */ | |
5618 | emit_insn (gen_subsi3 (temp, temp, stack_pointer_rtx)); | |
5619 | emit_insn (gen_shufb (temp3, temp, temp, pat)); | |
5620 | ||
5621 | /* Compute Available Stack Size for back chain */ | |
5622 | emit_insn (gen_subsi3 (temp2, temp2, stack_pointer_rtx)); | |
5623 | emit_insn (gen_shufb (temp4, temp2, temp2, pat)); | |
5624 | emit_insn (gen_addv4si3 (temp4, sp, temp4)); | |
5625 | ||
5626 | emit_insn (gen_addv4si3 (sp, sp, temp3)); | |
5627 | emit_move_insn (gen_frame_mem (V4SImode, stack_pointer_rtx), temp4); | |
5628 | } | |
5629 | ||
5630 | static void | |
5631 | spu_init_libfuncs (void) | |
5632 | { | |
5633 | set_optab_libfunc (smul_optab, DImode, "__muldi3"); | |
5634 | set_optab_libfunc (sdiv_optab, DImode, "__divdi3"); | |
5635 | set_optab_libfunc (smod_optab, DImode, "__moddi3"); | |
5636 | set_optab_libfunc (udiv_optab, DImode, "__udivdi3"); | |
5637 | set_optab_libfunc (umod_optab, DImode, "__umoddi3"); | |
5638 | set_optab_libfunc (udivmod_optab, DImode, "__udivmoddi4"); | |
5639 | set_optab_libfunc (ffs_optab, DImode, "__ffsdi2"); | |
5640 | set_optab_libfunc (clz_optab, DImode, "__clzdi2"); | |
5641 | set_optab_libfunc (ctz_optab, DImode, "__ctzdi2"); | |
5642 | set_optab_libfunc (popcount_optab, DImode, "__popcountdi2"); | |
5643 | set_optab_libfunc (parity_optab, DImode, "__paritydi2"); | |
5644 | ||
5645 | set_conv_libfunc (ufloat_optab, DFmode, SImode, "__float_unssidf"); | |
5646 | set_conv_libfunc (ufloat_optab, DFmode, DImode, "__float_unsdidf"); | |
19a53068 | 5647 | |
5825ec3f | 5648 | set_optab_libfunc (addv_optab, SImode, "__addvsi3"); |
5649 | set_optab_libfunc (subv_optab, SImode, "__subvsi3"); | |
5650 | set_optab_libfunc (smulv_optab, SImode, "__mulvsi3"); | |
5651 | set_optab_libfunc (sdivv_optab, SImode, "__divvsi3"); | |
5652 | set_optab_libfunc (negv_optab, SImode, "__negvsi2"); | |
5653 | set_optab_libfunc (absv_optab, SImode, "__absvsi2"); | |
5654 | set_optab_libfunc (addv_optab, DImode, "__addvdi3"); | |
5655 | set_optab_libfunc (subv_optab, DImode, "__subvdi3"); | |
5656 | set_optab_libfunc (smulv_optab, DImode, "__mulvdi3"); | |
5657 | set_optab_libfunc (sdivv_optab, DImode, "__divvdi3"); | |
5658 | set_optab_libfunc (negv_optab, DImode, "__negvdi2"); | |
5659 | set_optab_libfunc (absv_optab, DImode, "__absvdi2"); | |
5660 | ||
19a53068 | 5661 | set_optab_libfunc (smul_optab, TImode, "__multi3"); |
5662 | set_optab_libfunc (sdiv_optab, TImode, "__divti3"); | |
5663 | set_optab_libfunc (smod_optab, TImode, "__modti3"); | |
5664 | set_optab_libfunc (udiv_optab, TImode, "__udivti3"); | |
5665 | set_optab_libfunc (umod_optab, TImode, "__umodti3"); | |
5666 | set_optab_libfunc (udivmod_optab, TImode, "__udivmodti4"); | |
644459d0 | 5667 | } |
5668 | ||
5669 | /* Make a subreg, stripping any existing subreg. We could possibly just | |
5670 | call simplify_subreg, but in this case we know what we want. */ | |
5671 | rtx | |
5672 | spu_gen_subreg (enum machine_mode mode, rtx x) | |
5673 | { | |
5674 | if (GET_CODE (x) == SUBREG) | |
5675 | x = SUBREG_REG (x); | |
5676 | if (GET_MODE (x) == mode) | |
5677 | return x; | |
5678 | return gen_rtx_SUBREG (mode, x, 0); | |
5679 | } | |
5680 | ||
5681 | static bool | |
fb80456a | 5682 | spu_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED) |
644459d0 | 5683 | { |
5684 | return (TYPE_MODE (type) == BLKmode | |
5685 | && ((type) == 0 | |
5686 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST | |
5687 | || int_size_in_bytes (type) > | |
5688 | (MAX_REGISTER_RETURN * UNITS_PER_WORD))); | |
5689 | } | |
5690 | \f | |
5691 | /* Create the built-in types and functions */ | |
5692 | ||
c2233b46 | 5693 | enum spu_function_code |
5694 | { | |
5695 | #define DEF_BUILTIN(fcode, icode, name, type, params) fcode, | |
5696 | #include "spu-builtins.def" | |
5697 | #undef DEF_BUILTIN | |
5698 | NUM_SPU_BUILTINS | |
5699 | }; | |
5700 | ||
5701 | extern GTY(()) struct spu_builtin_description spu_builtins[NUM_SPU_BUILTINS]; | |
5702 | ||
644459d0 | 5703 | struct spu_builtin_description spu_builtins[] = { |
5704 | #define DEF_BUILTIN(fcode, icode, name, type, params) \ | |
0c5c4d59 | 5705 | {fcode, icode, name, type, params}, |
644459d0 | 5706 | #include "spu-builtins.def" |
5707 | #undef DEF_BUILTIN | |
5708 | }; | |
5709 | ||
0c5c4d59 | 5710 | static GTY(()) tree spu_builtin_decls[NUM_SPU_BUILTINS]; |
5711 | ||
5712 | /* Returns the spu builtin decl for CODE. */ | |
e6925042 | 5713 | |
5714 | static tree | |
5715 | spu_builtin_decl (unsigned code, bool initialize_p ATTRIBUTE_UNUSED) | |
5716 | { | |
5717 | if (code >= NUM_SPU_BUILTINS) | |
5718 | return error_mark_node; | |
5719 | ||
0c5c4d59 | 5720 | return spu_builtin_decls[code]; |
e6925042 | 5721 | } |
5722 | ||
5723 | ||
644459d0 | 5724 | static void |
5725 | spu_init_builtins (void) | |
5726 | { | |
5727 | struct spu_builtin_description *d; | |
5728 | unsigned int i; | |
5729 | ||
5730 | V16QI_type_node = build_vector_type (intQI_type_node, 16); | |
5731 | V8HI_type_node = build_vector_type (intHI_type_node, 8); | |
5732 | V4SI_type_node = build_vector_type (intSI_type_node, 4); | |
5733 | V2DI_type_node = build_vector_type (intDI_type_node, 2); | |
5734 | V4SF_type_node = build_vector_type (float_type_node, 4); | |
5735 | V2DF_type_node = build_vector_type (double_type_node, 2); | |
5736 | ||
5737 | unsigned_V16QI_type_node = build_vector_type (unsigned_intQI_type_node, 16); | |
5738 | unsigned_V8HI_type_node = build_vector_type (unsigned_intHI_type_node, 8); | |
5739 | unsigned_V4SI_type_node = build_vector_type (unsigned_intSI_type_node, 4); | |
5740 | unsigned_V2DI_type_node = build_vector_type (unsigned_intDI_type_node, 2); | |
5741 | ||
c4ecce0c | 5742 | spu_builtin_types[SPU_BTI_QUADWORD] = V16QI_type_node; |
644459d0 | 5743 | |
5744 | spu_builtin_types[SPU_BTI_7] = global_trees[TI_INTSI_TYPE]; | |
5745 | spu_builtin_types[SPU_BTI_S7] = global_trees[TI_INTSI_TYPE]; | |
5746 | spu_builtin_types[SPU_BTI_U7] = global_trees[TI_INTSI_TYPE]; | |
5747 | spu_builtin_types[SPU_BTI_S10] = global_trees[TI_INTSI_TYPE]; | |
5748 | spu_builtin_types[SPU_BTI_S10_4] = global_trees[TI_INTSI_TYPE]; | |
5749 | spu_builtin_types[SPU_BTI_U14] = global_trees[TI_INTSI_TYPE]; | |
5750 | spu_builtin_types[SPU_BTI_16] = global_trees[TI_INTSI_TYPE]; | |
5751 | spu_builtin_types[SPU_BTI_S16] = global_trees[TI_INTSI_TYPE]; | |
5752 | spu_builtin_types[SPU_BTI_S16_2] = global_trees[TI_INTSI_TYPE]; | |
5753 | spu_builtin_types[SPU_BTI_U16] = global_trees[TI_INTSI_TYPE]; | |
5754 | spu_builtin_types[SPU_BTI_U16_2] = global_trees[TI_INTSI_TYPE]; | |
5755 | spu_builtin_types[SPU_BTI_U18] = global_trees[TI_INTSI_TYPE]; | |
5756 | ||
5757 | spu_builtin_types[SPU_BTI_INTQI] = global_trees[TI_INTQI_TYPE]; | |
5758 | spu_builtin_types[SPU_BTI_INTHI] = global_trees[TI_INTHI_TYPE]; | |
5759 | spu_builtin_types[SPU_BTI_INTSI] = global_trees[TI_INTSI_TYPE]; | |
5760 | spu_builtin_types[SPU_BTI_INTDI] = global_trees[TI_INTDI_TYPE]; | |
5761 | spu_builtin_types[SPU_BTI_UINTQI] = global_trees[TI_UINTQI_TYPE]; | |
5762 | spu_builtin_types[SPU_BTI_UINTHI] = global_trees[TI_UINTHI_TYPE]; | |
5763 | spu_builtin_types[SPU_BTI_UINTSI] = global_trees[TI_UINTSI_TYPE]; | |
5764 | spu_builtin_types[SPU_BTI_UINTDI] = global_trees[TI_UINTDI_TYPE]; | |
5765 | ||
5766 | spu_builtin_types[SPU_BTI_FLOAT] = global_trees[TI_FLOAT_TYPE]; | |
5767 | spu_builtin_types[SPU_BTI_DOUBLE] = global_trees[TI_DOUBLE_TYPE]; | |
5768 | ||
5769 | spu_builtin_types[SPU_BTI_VOID] = global_trees[TI_VOID_TYPE]; | |
5770 | ||
5771 | spu_builtin_types[SPU_BTI_PTR] = | |
5772 | build_pointer_type (build_qualified_type | |
5773 | (void_type_node, | |
5774 | TYPE_QUAL_CONST | TYPE_QUAL_VOLATILE)); | |
5775 | ||
5776 | /* For each builtin we build a new prototype. The tree code will make | |
5777 | sure nodes are shared. */ | |
5778 | for (i = 0, d = spu_builtins; i < NUM_SPU_BUILTINS; i++, d++) | |
5779 | { | |
5780 | tree p; | |
5781 | char name[64]; /* build_function will make a copy. */ | |
5782 | int parm; | |
5783 | ||
5784 | if (d->name == 0) | |
5785 | continue; | |
5786 | ||
5dfbd18f | 5787 | /* Find last parm. */ |
644459d0 | 5788 | for (parm = 1; d->parm[parm] != SPU_BTI_END_OF_PARAMS; parm++) |
5dfbd18f | 5789 | ; |
644459d0 | 5790 | |
5791 | p = void_list_node; | |
5792 | while (parm > 1) | |
5793 | p = tree_cons (NULL_TREE, spu_builtin_types[d->parm[--parm]], p); | |
5794 | ||
5795 | p = build_function_type (spu_builtin_types[d->parm[0]], p); | |
5796 | ||
5797 | sprintf (name, "__builtin_%s", d->name); | |
0c5c4d59 | 5798 | spu_builtin_decls[i] = |
3726fe5e | 5799 | add_builtin_function (name, p, i, BUILT_IN_MD, NULL, NULL_TREE); |
a76866d3 | 5800 | if (d->fcode == SPU_MASK_FOR_LOAD) |
0c5c4d59 | 5801 | TREE_READONLY (spu_builtin_decls[i]) = 1; |
5dfbd18f | 5802 | |
5803 | /* These builtins don't throw. */ | |
0c5c4d59 | 5804 | TREE_NOTHROW (spu_builtin_decls[i]) = 1; |
644459d0 | 5805 | } |
5806 | } | |
5807 | ||
cf31d486 | 5808 | void |
5809 | spu_restore_stack_block (rtx op0 ATTRIBUTE_UNUSED, rtx op1) | |
5810 | { | |
5811 | static unsigned char arr[16] = | |
5812 | { 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3 }; | |
5813 | ||
5814 | rtx temp = gen_reg_rtx (Pmode); | |
5815 | rtx temp2 = gen_reg_rtx (V4SImode); | |
5816 | rtx temp3 = gen_reg_rtx (V4SImode); | |
5817 | rtx pat = gen_reg_rtx (TImode); | |
5818 | rtx sp = gen_rtx_REG (V4SImode, STACK_POINTER_REGNUM); | |
5819 | ||
5820 | emit_move_insn (pat, array_to_constant (TImode, arr)); | |
5821 | ||
5822 | /* Restore the sp. */ | |
5823 | emit_move_insn (temp, op1); | |
5824 | emit_move_insn (temp2, gen_frame_mem (V4SImode, stack_pointer_rtx)); | |
5825 | ||
5826 | /* Compute available stack size for sp. */ | |
5827 | emit_insn (gen_subsi3 (temp, temp, stack_pointer_rtx)); | |
5828 | emit_insn (gen_shufb (temp3, temp, temp, pat)); | |
5829 | ||
5830 | emit_insn (gen_addv4si3 (sp, sp, temp3)); | |
5831 | emit_move_insn (gen_frame_mem (V4SImode, stack_pointer_rtx), temp2); | |
5832 | } | |
5833 | ||
644459d0 | 5834 | int |
5835 | spu_safe_dma (HOST_WIDE_INT channel) | |
5836 | { | |
006e4b96 | 5837 | return TARGET_SAFE_DMA && channel >= 21 && channel <= 27; |
644459d0 | 5838 | } |
5839 | ||
5840 | void | |
5841 | spu_builtin_splats (rtx ops[]) | |
5842 | { | |
5843 | enum machine_mode mode = GET_MODE (ops[0]); | |
5844 | if (GET_CODE (ops[1]) == CONST_INT || GET_CODE (ops[1]) == CONST_DOUBLE) | |
5845 | { | |
5846 | unsigned char arr[16]; | |
5847 | constant_to_array (GET_MODE_INNER (mode), ops[1], arr); | |
5848 | emit_move_insn (ops[0], array_to_constant (mode, arr)); | |
5849 | } | |
644459d0 | 5850 | else |
5851 | { | |
5852 | rtx reg = gen_reg_rtx (TImode); | |
5853 | rtx shuf; | |
5854 | if (GET_CODE (ops[1]) != REG | |
5855 | && GET_CODE (ops[1]) != SUBREG) | |
5856 | ops[1] = force_reg (GET_MODE_INNER (mode), ops[1]); | |
5857 | switch (mode) | |
5858 | { | |
5859 | case V2DImode: | |
5860 | case V2DFmode: | |
5861 | shuf = | |
5862 | immed_double_const (0x0001020304050607ll, 0x1011121314151617ll, | |
5863 | TImode); | |
5864 | break; | |
5865 | case V4SImode: | |
5866 | case V4SFmode: | |
5867 | shuf = | |
5868 | immed_double_const (0x0001020300010203ll, 0x0001020300010203ll, | |
5869 | TImode); | |
5870 | break; | |
5871 | case V8HImode: | |
5872 | shuf = | |
5873 | immed_double_const (0x0203020302030203ll, 0x0203020302030203ll, | |
5874 | TImode); | |
5875 | break; | |
5876 | case V16QImode: | |
5877 | shuf = | |
5878 | immed_double_const (0x0303030303030303ll, 0x0303030303030303ll, | |
5879 | TImode); | |
5880 | break; | |
5881 | default: | |
5882 | abort (); | |
5883 | } | |
5884 | emit_move_insn (reg, shuf); | |
5885 | emit_insn (gen_shufb (ops[0], ops[1], ops[1], reg)); | |
5886 | } | |
5887 | } | |
5888 | ||
5889 | void | |
5890 | spu_builtin_extract (rtx ops[]) | |
5891 | { | |
5892 | enum machine_mode mode; | |
5893 | rtx rot, from, tmp; | |
5894 | ||
5895 | mode = GET_MODE (ops[1]); | |
5896 | ||
5897 | if (GET_CODE (ops[2]) == CONST_INT) | |
5898 | { | |
5899 | switch (mode) | |
5900 | { | |
5901 | case V16QImode: | |
5902 | emit_insn (gen_vec_extractv16qi (ops[0], ops[1], ops[2])); | |
5903 | break; | |
5904 | case V8HImode: | |
5905 | emit_insn (gen_vec_extractv8hi (ops[0], ops[1], ops[2])); | |
5906 | break; | |
5907 | case V4SFmode: | |
5908 | emit_insn (gen_vec_extractv4sf (ops[0], ops[1], ops[2])); | |
5909 | break; | |
5910 | case V4SImode: | |
5911 | emit_insn (gen_vec_extractv4si (ops[0], ops[1], ops[2])); | |
5912 | break; | |
5913 | case V2DImode: | |
5914 | emit_insn (gen_vec_extractv2di (ops[0], ops[1], ops[2])); | |
5915 | break; | |
5916 | case V2DFmode: | |
5917 | emit_insn (gen_vec_extractv2df (ops[0], ops[1], ops[2])); | |
5918 | break; | |
5919 | default: | |
5920 | abort (); | |
5921 | } | |
5922 | return; | |
5923 | } | |
5924 | ||
5925 | from = spu_gen_subreg (TImode, ops[1]); | |
5926 | rot = gen_reg_rtx (TImode); | |
5927 | tmp = gen_reg_rtx (SImode); | |
5928 | ||
5929 | switch (mode) | |
5930 | { | |
5931 | case V16QImode: | |
5932 | emit_insn (gen_addsi3 (tmp, ops[2], GEN_INT (-3))); | |
5933 | break; | |
5934 | case V8HImode: | |
5935 | emit_insn (gen_addsi3 (tmp, ops[2], ops[2])); | |
5936 | emit_insn (gen_addsi3 (tmp, tmp, GEN_INT (-2))); | |
5937 | break; | |
5938 | case V4SFmode: | |
5939 | case V4SImode: | |
5940 | emit_insn (gen_ashlsi3 (tmp, ops[2], GEN_INT (2))); | |
5941 | break; | |
5942 | case V2DImode: | |
5943 | case V2DFmode: | |
5944 | emit_insn (gen_ashlsi3 (tmp, ops[2], GEN_INT (3))); | |
5945 | break; | |
5946 | default: | |
5947 | abort (); | |
5948 | } | |
5949 | emit_insn (gen_rotqby_ti (rot, from, tmp)); | |
5950 | ||
5951 | emit_insn (gen_spu_convert (ops[0], rot)); | |
5952 | } | |
5953 | ||
5954 | void | |
5955 | spu_builtin_insert (rtx ops[]) | |
5956 | { | |
5957 | enum machine_mode mode = GET_MODE (ops[0]); | |
5958 | enum machine_mode imode = GET_MODE_INNER (mode); | |
5959 | rtx mask = gen_reg_rtx (TImode); | |
5960 | rtx offset; | |
5961 | ||
5962 | if (GET_CODE (ops[3]) == CONST_INT) | |
5963 | offset = GEN_INT (INTVAL (ops[3]) * GET_MODE_SIZE (imode)); | |
5964 | else | |
5965 | { | |
5966 | offset = gen_reg_rtx (SImode); | |
5967 | emit_insn (gen_mulsi3 | |
5968 | (offset, ops[3], GEN_INT (GET_MODE_SIZE (imode)))); | |
5969 | } | |
5970 | emit_insn (gen_cpat | |
5971 | (mask, stack_pointer_rtx, offset, | |
5972 | GEN_INT (GET_MODE_SIZE (imode)))); | |
5973 | emit_insn (gen_shufb (ops[0], ops[1], ops[2], mask)); | |
5974 | } | |
5975 | ||
5976 | void | |
5977 | spu_builtin_promote (rtx ops[]) | |
5978 | { | |
5979 | enum machine_mode mode, imode; | |
5980 | rtx rot, from, offset; | |
5981 | HOST_WIDE_INT pos; | |
5982 | ||
5983 | mode = GET_MODE (ops[0]); | |
5984 | imode = GET_MODE_INNER (mode); | |
5985 | ||
5986 | from = gen_reg_rtx (TImode); | |
5987 | rot = spu_gen_subreg (TImode, ops[0]); | |
5988 | ||
5989 | emit_insn (gen_spu_convert (from, ops[1])); | |
5990 | ||
5991 | if (GET_CODE (ops[2]) == CONST_INT) | |
5992 | { | |
5993 | pos = -GET_MODE_SIZE (imode) * INTVAL (ops[2]); | |
5994 | if (GET_MODE_SIZE (imode) < 4) | |
5995 | pos += 4 - GET_MODE_SIZE (imode); | |
5996 | offset = GEN_INT (pos & 15); | |
5997 | } | |
5998 | else | |
5999 | { | |
6000 | offset = gen_reg_rtx (SImode); | |
6001 | switch (mode) | |
6002 | { | |
6003 | case V16QImode: | |
6004 | emit_insn (gen_subsi3 (offset, GEN_INT (3), ops[2])); | |
6005 | break; | |
6006 | case V8HImode: | |
6007 | emit_insn (gen_subsi3 (offset, GEN_INT (1), ops[2])); | |
6008 | emit_insn (gen_addsi3 (offset, offset, offset)); | |
6009 | break; | |
6010 | case V4SFmode: | |
6011 | case V4SImode: | |
6012 | emit_insn (gen_subsi3 (offset, GEN_INT (0), ops[2])); | |
6013 | emit_insn (gen_ashlsi3 (offset, offset, GEN_INT (2))); | |
6014 | break; | |
6015 | case V2DImode: | |
6016 | case V2DFmode: | |
6017 | emit_insn (gen_ashlsi3 (offset, ops[2], GEN_INT (3))); | |
6018 | break; | |
6019 | default: | |
6020 | abort (); | |
6021 | } | |
6022 | } | |
6023 | emit_insn (gen_rotqby_ti (rot, from, offset)); | |
6024 | } | |
6025 | ||
e96f2783 | 6026 | static void |
6027 | spu_trampoline_init (rtx m_tramp, tree fndecl, rtx cxt) | |
644459d0 | 6028 | { |
e96f2783 | 6029 | rtx fnaddr = XEXP (DECL_RTL (fndecl), 0); |
644459d0 | 6030 | rtx shuf = gen_reg_rtx (V4SImode); |
6031 | rtx insn = gen_reg_rtx (V4SImode); | |
6032 | rtx shufc; | |
6033 | rtx insnc; | |
6034 | rtx mem; | |
6035 | ||
6036 | fnaddr = force_reg (SImode, fnaddr); | |
6037 | cxt = force_reg (SImode, cxt); | |
6038 | ||
6039 | if (TARGET_LARGE_MEM) | |
6040 | { | |
6041 | rtx rotl = gen_reg_rtx (V4SImode); | |
6042 | rtx mask = gen_reg_rtx (V4SImode); | |
6043 | rtx bi = gen_reg_rtx (SImode); | |
e96f2783 | 6044 | static unsigned char const shufa[16] = { |
644459d0 | 6045 | 2, 3, 0, 1, 18, 19, 16, 17, |
6046 | 0, 1, 2, 3, 16, 17, 18, 19 | |
6047 | }; | |
e96f2783 | 6048 | static unsigned char const insna[16] = { |
644459d0 | 6049 | 0x41, 0, 0, 79, |
6050 | 0x41, 0, 0, STATIC_CHAIN_REGNUM, | |
6051 | 0x60, 0x80, 0, 79, | |
6052 | 0x60, 0x80, 0, STATIC_CHAIN_REGNUM | |
6053 | }; | |
6054 | ||
6055 | shufc = force_reg (TImode, array_to_constant (TImode, shufa)); | |
6056 | insnc = force_reg (V4SImode, array_to_constant (V4SImode, insna)); | |
6057 | ||
6058 | emit_insn (gen_shufb (shuf, fnaddr, cxt, shufc)); | |
4d54df85 | 6059 | emit_insn (gen_vrotlv4si3 (rotl, shuf, spu_const (V4SImode, 7))); |
644459d0 | 6060 | emit_insn (gen_movv4si (mask, spu_const (V4SImode, 0xffff << 7))); |
6061 | emit_insn (gen_selb (insn, insnc, rotl, mask)); | |
6062 | ||
e96f2783 | 6063 | mem = adjust_address (m_tramp, V4SImode, 0); |
6064 | emit_move_insn (mem, insn); | |
644459d0 | 6065 | |
6066 | emit_move_insn (bi, GEN_INT (0x35000000 + (79 << 7))); | |
e96f2783 | 6067 | mem = adjust_address (m_tramp, Pmode, 16); |
6068 | emit_move_insn (mem, bi); | |
644459d0 | 6069 | } |
6070 | else | |
6071 | { | |
6072 | rtx scxt = gen_reg_rtx (SImode); | |
6073 | rtx sfnaddr = gen_reg_rtx (SImode); | |
e96f2783 | 6074 | static unsigned char const insna[16] = { |
644459d0 | 6075 | 0x42, 0, 0, STATIC_CHAIN_REGNUM, |
6076 | 0x30, 0, 0, 0, | |
6077 | 0, 0, 0, 0, | |
6078 | 0, 0, 0, 0 | |
6079 | }; | |
6080 | ||
6081 | shufc = gen_reg_rtx (TImode); | |
6082 | insnc = force_reg (V4SImode, array_to_constant (V4SImode, insna)); | |
6083 | ||
6084 | /* By or'ing all of cxt with the ila opcode we are assuming cxt | |
6085 | fits 18 bits and the last 4 are zeros. This will be true if | |
6086 | the stack pointer is initialized to 0x3fff0 at program start, | |
6087 | otherwise the ila instruction will be garbage. */ | |
6088 | ||
6089 | emit_insn (gen_ashlsi3 (scxt, cxt, GEN_INT (7))); | |
6090 | emit_insn (gen_ashlsi3 (sfnaddr, fnaddr, GEN_INT (5))); | |
6091 | emit_insn (gen_cpat | |
6092 | (shufc, stack_pointer_rtx, GEN_INT (4), GEN_INT (4))); | |
6093 | emit_insn (gen_shufb (shuf, sfnaddr, scxt, shufc)); | |
6094 | emit_insn (gen_iorv4si3 (insn, insnc, shuf)); | |
6095 | ||
e96f2783 | 6096 | mem = adjust_address (m_tramp, V4SImode, 0); |
6097 | emit_move_insn (mem, insn); | |
644459d0 | 6098 | } |
6099 | emit_insn (gen_sync ()); | |
6100 | } | |
6101 | ||
6102 | void | |
6103 | spu_expand_sign_extend (rtx ops[]) | |
6104 | { | |
6105 | unsigned char arr[16]; | |
6106 | rtx pat = gen_reg_rtx (TImode); | |
6107 | rtx sign, c; | |
6108 | int i, last; | |
6109 | last = GET_MODE (ops[0]) == DImode ? 7 : 15; | |
6110 | if (GET_MODE (ops[1]) == QImode) | |
6111 | { | |
6112 | sign = gen_reg_rtx (HImode); | |
6113 | emit_insn (gen_extendqihi2 (sign, ops[1])); | |
6114 | for (i = 0; i < 16; i++) | |
6115 | arr[i] = 0x12; | |
6116 | arr[last] = 0x13; | |
6117 | } | |
6118 | else | |
6119 | { | |
6120 | for (i = 0; i < 16; i++) | |
6121 | arr[i] = 0x10; | |
6122 | switch (GET_MODE (ops[1])) | |
6123 | { | |
6124 | case HImode: | |
6125 | sign = gen_reg_rtx (SImode); | |
6126 | emit_insn (gen_extendhisi2 (sign, ops[1])); | |
6127 | arr[last] = 0x03; | |
6128 | arr[last - 1] = 0x02; | |
6129 | break; | |
6130 | case SImode: | |
6131 | sign = gen_reg_rtx (SImode); | |
6132 | emit_insn (gen_ashrsi3 (sign, ops[1], GEN_INT (31))); | |
6133 | for (i = 0; i < 4; i++) | |
6134 | arr[last - i] = 3 - i; | |
6135 | break; | |
6136 | case DImode: | |
6137 | sign = gen_reg_rtx (SImode); | |
6138 | c = gen_reg_rtx (SImode); | |
6139 | emit_insn (gen_spu_convert (c, ops[1])); | |
6140 | emit_insn (gen_ashrsi3 (sign, c, GEN_INT (31))); | |
6141 | for (i = 0; i < 8; i++) | |
6142 | arr[last - i] = 7 - i; | |
6143 | break; | |
6144 | default: | |
6145 | abort (); | |
6146 | } | |
6147 | } | |
6148 | emit_move_insn (pat, array_to_constant (TImode, arr)); | |
6149 | emit_insn (gen_shufb (ops[0], ops[1], sign, pat)); | |
6150 | } | |
6151 | ||
6152 | /* expand vector initialization. If there are any constant parts, | |
6153 | load constant parts first. Then load any non-constant parts. */ | |
6154 | void | |
6155 | spu_expand_vector_init (rtx target, rtx vals) | |
6156 | { | |
6157 | enum machine_mode mode = GET_MODE (target); | |
6158 | int n_elts = GET_MODE_NUNITS (mode); | |
6159 | int n_var = 0; | |
6160 | bool all_same = true; | |
790c536c | 6161 | rtx first, x = NULL_RTX, first_constant = NULL_RTX; |
644459d0 | 6162 | int i; |
6163 | ||
6164 | first = XVECEXP (vals, 0, 0); | |
6165 | for (i = 0; i < n_elts; ++i) | |
6166 | { | |
6167 | x = XVECEXP (vals, 0, i); | |
e442af0b | 6168 | if (!(CONST_INT_P (x) |
6169 | || GET_CODE (x) == CONST_DOUBLE | |
6170 | || GET_CODE (x) == CONST_FIXED)) | |
644459d0 | 6171 | ++n_var; |
6172 | else | |
6173 | { | |
6174 | if (first_constant == NULL_RTX) | |
6175 | first_constant = x; | |
6176 | } | |
6177 | if (i > 0 && !rtx_equal_p (x, first)) | |
6178 | all_same = false; | |
6179 | } | |
6180 | ||
6181 | /* if all elements are the same, use splats to repeat elements */ | |
6182 | if (all_same) | |
6183 | { | |
6184 | if (!CONSTANT_P (first) | |
6185 | && !register_operand (first, GET_MODE (x))) | |
6186 | first = force_reg (GET_MODE (first), first); | |
6187 | emit_insn (gen_spu_splats (target, first)); | |
6188 | return; | |
6189 | } | |
6190 | ||
6191 | /* load constant parts */ | |
6192 | if (n_var != n_elts) | |
6193 | { | |
6194 | if (n_var == 0) | |
6195 | { | |
6196 | emit_move_insn (target, | |
6197 | gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0))); | |
6198 | } | |
6199 | else | |
6200 | { | |
6201 | rtx constant_parts_rtx = copy_rtx (vals); | |
6202 | ||
6203 | gcc_assert (first_constant != NULL_RTX); | |
6204 | /* fill empty slots with the first constant, this increases | |
6205 | our chance of using splats in the recursive call below. */ | |
6206 | for (i = 0; i < n_elts; ++i) | |
e442af0b | 6207 | { |
6208 | x = XVECEXP (constant_parts_rtx, 0, i); | |
6209 | if (!(CONST_INT_P (x) | |
6210 | || GET_CODE (x) == CONST_DOUBLE | |
6211 | || GET_CODE (x) == CONST_FIXED)) | |
6212 | XVECEXP (constant_parts_rtx, 0, i) = first_constant; | |
6213 | } | |
644459d0 | 6214 | |
6215 | spu_expand_vector_init (target, constant_parts_rtx); | |
6216 | } | |
6217 | } | |
6218 | ||
6219 | /* load variable parts */ | |
6220 | if (n_var != 0) | |
6221 | { | |
6222 | rtx insert_operands[4]; | |
6223 | ||
6224 | insert_operands[0] = target; | |
6225 | insert_operands[2] = target; | |
6226 | for (i = 0; i < n_elts; ++i) | |
6227 | { | |
6228 | x = XVECEXP (vals, 0, i); | |
e442af0b | 6229 | if (!(CONST_INT_P (x) |
6230 | || GET_CODE (x) == CONST_DOUBLE | |
6231 | || GET_CODE (x) == CONST_FIXED)) | |
644459d0 | 6232 | { |
6233 | if (!register_operand (x, GET_MODE (x))) | |
6234 | x = force_reg (GET_MODE (x), x); | |
6235 | insert_operands[1] = x; | |
6236 | insert_operands[3] = GEN_INT (i); | |
6237 | spu_builtin_insert (insert_operands); | |
6238 | } | |
6239 | } | |
6240 | } | |
6241 | } | |
6352eedf | 6242 | |
5474166e | 6243 | /* Return insn index for the vector compare instruction for given CODE, |
6244 | and DEST_MODE, OP_MODE. Return -1 if valid insn is not available. */ | |
6245 | ||
6246 | static int | |
6247 | get_vec_cmp_insn (enum rtx_code code, | |
6248 | enum machine_mode dest_mode, | |
6249 | enum machine_mode op_mode) | |
6250 | ||
6251 | { | |
6252 | switch (code) | |
6253 | { | |
6254 | case EQ: | |
6255 | if (dest_mode == V16QImode && op_mode == V16QImode) | |
6256 | return CODE_FOR_ceq_v16qi; | |
6257 | if (dest_mode == V8HImode && op_mode == V8HImode) | |
6258 | return CODE_FOR_ceq_v8hi; | |
6259 | if (dest_mode == V4SImode && op_mode == V4SImode) | |
6260 | return CODE_FOR_ceq_v4si; | |
6261 | if (dest_mode == V4SImode && op_mode == V4SFmode) | |
6262 | return CODE_FOR_ceq_v4sf; | |
6263 | if (dest_mode == V2DImode && op_mode == V2DFmode) | |
6264 | return CODE_FOR_ceq_v2df; | |
6265 | break; | |
6266 | case GT: | |
6267 | if (dest_mode == V16QImode && op_mode == V16QImode) | |
6268 | return CODE_FOR_cgt_v16qi; | |
6269 | if (dest_mode == V8HImode && op_mode == V8HImode) | |
6270 | return CODE_FOR_cgt_v8hi; | |
6271 | if (dest_mode == V4SImode && op_mode == V4SImode) | |
6272 | return CODE_FOR_cgt_v4si; | |
6273 | if (dest_mode == V4SImode && op_mode == V4SFmode) | |
6274 | return CODE_FOR_cgt_v4sf; | |
6275 | if (dest_mode == V2DImode && op_mode == V2DFmode) | |
6276 | return CODE_FOR_cgt_v2df; | |
6277 | break; | |
6278 | case GTU: | |
6279 | if (dest_mode == V16QImode && op_mode == V16QImode) | |
6280 | return CODE_FOR_clgt_v16qi; | |
6281 | if (dest_mode == V8HImode && op_mode == V8HImode) | |
6282 | return CODE_FOR_clgt_v8hi; | |
6283 | if (dest_mode == V4SImode && op_mode == V4SImode) | |
6284 | return CODE_FOR_clgt_v4si; | |
6285 | break; | |
6286 | default: | |
6287 | break; | |
6288 | } | |
6289 | return -1; | |
6290 | } | |
6291 | ||
6292 | /* Emit vector compare for operands OP0 and OP1 using code RCODE. | |
6293 | DMODE is expected destination mode. This is a recursive function. */ | |
6294 | ||
6295 | static rtx | |
6296 | spu_emit_vector_compare (enum rtx_code rcode, | |
6297 | rtx op0, rtx op1, | |
6298 | enum machine_mode dmode) | |
6299 | { | |
6300 | int vec_cmp_insn; | |
6301 | rtx mask; | |
6302 | enum machine_mode dest_mode; | |
6303 | enum machine_mode op_mode = GET_MODE (op1); | |
6304 | ||
6305 | gcc_assert (GET_MODE (op0) == GET_MODE (op1)); | |
6306 | ||
6307 | /* Floating point vector compare instructions uses destination V4SImode. | |
6308 | Double floating point vector compare instructions uses destination V2DImode. | |
6309 | Move destination to appropriate mode later. */ | |
6310 | if (dmode == V4SFmode) | |
6311 | dest_mode = V4SImode; | |
6312 | else if (dmode == V2DFmode) | |
6313 | dest_mode = V2DImode; | |
6314 | else | |
6315 | dest_mode = dmode; | |
6316 | ||
6317 | mask = gen_reg_rtx (dest_mode); | |
6318 | vec_cmp_insn = get_vec_cmp_insn (rcode, dest_mode, op_mode); | |
6319 | ||
6320 | if (vec_cmp_insn == -1) | |
6321 | { | |
6322 | bool swap_operands = false; | |
6323 | bool try_again = false; | |
6324 | switch (rcode) | |
6325 | { | |
6326 | case LT: | |
6327 | rcode = GT; | |
6328 | swap_operands = true; | |
6329 | try_again = true; | |
6330 | break; | |
6331 | case LTU: | |
6332 | rcode = GTU; | |
6333 | swap_operands = true; | |
6334 | try_again = true; | |
6335 | break; | |
6336 | case NE: | |
6337 | /* Treat A != B as ~(A==B). */ | |
6338 | { | |
6339 | enum insn_code nor_code; | |
6340 | rtx eq_rtx = spu_emit_vector_compare (EQ, op0, op1, dest_mode); | |
d6bf3b14 | 6341 | nor_code = optab_handler (one_cmpl_optab, dest_mode); |
5474166e | 6342 | gcc_assert (nor_code != CODE_FOR_nothing); |
6343 | emit_insn (GEN_FCN (nor_code) (mask, eq_rtx)); | |
6344 | if (dmode != dest_mode) | |
6345 | { | |
6346 | rtx temp = gen_reg_rtx (dest_mode); | |
6347 | convert_move (temp, mask, 0); | |
6348 | return temp; | |
6349 | } | |
6350 | return mask; | |
6351 | } | |
6352 | break; | |
6353 | case GE: | |
6354 | case GEU: | |
6355 | case LE: | |
6356 | case LEU: | |
6357 | /* Try GT/GTU/LT/LTU OR EQ */ | |
6358 | { | |
6359 | rtx c_rtx, eq_rtx; | |
6360 | enum insn_code ior_code; | |
6361 | enum rtx_code new_code; | |
6362 | ||
6363 | switch (rcode) | |
6364 | { | |
6365 | case GE: new_code = GT; break; | |
6366 | case GEU: new_code = GTU; break; | |
6367 | case LE: new_code = LT; break; | |
6368 | case LEU: new_code = LTU; break; | |
6369 | default: | |
6370 | gcc_unreachable (); | |
6371 | } | |
6372 | ||
6373 | c_rtx = spu_emit_vector_compare (new_code, op0, op1, dest_mode); | |
6374 | eq_rtx = spu_emit_vector_compare (EQ, op0, op1, dest_mode); | |
6375 | ||
d6bf3b14 | 6376 | ior_code = optab_handler (ior_optab, dest_mode); |
5474166e | 6377 | gcc_assert (ior_code != CODE_FOR_nothing); |
6378 | emit_insn (GEN_FCN (ior_code) (mask, c_rtx, eq_rtx)); | |
6379 | if (dmode != dest_mode) | |
6380 | { | |
6381 | rtx temp = gen_reg_rtx (dest_mode); | |
6382 | convert_move (temp, mask, 0); | |
6383 | return temp; | |
6384 | } | |
6385 | return mask; | |
6386 | } | |
6387 | break; | |
6388 | default: | |
6389 | gcc_unreachable (); | |
6390 | } | |
6391 | ||
6392 | /* You only get two chances. */ | |
6393 | if (try_again) | |
6394 | vec_cmp_insn = get_vec_cmp_insn (rcode, dest_mode, op_mode); | |
6395 | ||
6396 | gcc_assert (vec_cmp_insn != -1); | |
6397 | ||
6398 | if (swap_operands) | |
6399 | { | |
6400 | rtx tmp; | |
6401 | tmp = op0; | |
6402 | op0 = op1; | |
6403 | op1 = tmp; | |
6404 | } | |
6405 | } | |
6406 | ||
6407 | emit_insn (GEN_FCN (vec_cmp_insn) (mask, op0, op1)); | |
6408 | if (dmode != dest_mode) | |
6409 | { | |
6410 | rtx temp = gen_reg_rtx (dest_mode); | |
6411 | convert_move (temp, mask, 0); | |
6412 | return temp; | |
6413 | } | |
6414 | return mask; | |
6415 | } | |
6416 | ||
6417 | ||
6418 | /* Emit vector conditional expression. | |
6419 | DEST is destination. OP1 and OP2 are two VEC_COND_EXPR operands. | |
6420 | CC_OP0 and CC_OP1 are the two operands for the relation operation COND. */ | |
6421 | ||
6422 | int | |
6423 | spu_emit_vector_cond_expr (rtx dest, rtx op1, rtx op2, | |
6424 | rtx cond, rtx cc_op0, rtx cc_op1) | |
6425 | { | |
6426 | enum machine_mode dest_mode = GET_MODE (dest); | |
6427 | enum rtx_code rcode = GET_CODE (cond); | |
6428 | rtx mask; | |
6429 | ||
6430 | /* Get the vector mask for the given relational operations. */ | |
6431 | mask = spu_emit_vector_compare (rcode, cc_op0, cc_op1, dest_mode); | |
6432 | ||
6433 | emit_insn(gen_selb (dest, op2, op1, mask)); | |
6434 | ||
6435 | return 1; | |
6436 | } | |
6437 | ||
6352eedf | 6438 | static rtx |
6439 | spu_force_reg (enum machine_mode mode, rtx op) | |
6440 | { | |
6441 | rtx x, r; | |
6442 | if (GET_MODE (op) == VOIDmode || GET_MODE (op) == BLKmode) | |
6443 | { | |
6444 | if ((SCALAR_INT_MODE_P (mode) && GET_CODE (op) == CONST_INT) | |
6445 | || GET_MODE (op) == BLKmode) | |
6446 | return force_reg (mode, convert_to_mode (mode, op, 0)); | |
6447 | abort (); | |
6448 | } | |
6449 | ||
6450 | r = force_reg (GET_MODE (op), op); | |
6451 | if (GET_MODE_SIZE (GET_MODE (op)) == GET_MODE_SIZE (mode)) | |
6452 | { | |
6453 | x = simplify_gen_subreg (mode, r, GET_MODE (op), 0); | |
6454 | if (x) | |
6455 | return x; | |
6456 | } | |
6457 | ||
6458 | x = gen_reg_rtx (mode); | |
6459 | emit_insn (gen_spu_convert (x, r)); | |
6460 | return x; | |
6461 | } | |
6462 | ||
6463 | static void | |
6464 | spu_check_builtin_parm (struct spu_builtin_description *d, rtx op, int p) | |
6465 | { | |
6466 | HOST_WIDE_INT v = 0; | |
6467 | int lsbits; | |
6468 | /* Check the range of immediate operands. */ | |
6469 | if (p >= SPU_BTI_7 && p <= SPU_BTI_U18) | |
6470 | { | |
6471 | int range = p - SPU_BTI_7; | |
5df189be | 6472 | |
6473 | if (!CONSTANT_P (op)) | |
bf776685 | 6474 | error ("%s expects an integer literal in the range [%d, %d]", |
6352eedf | 6475 | d->name, |
6476 | spu_builtin_range[range].low, spu_builtin_range[range].high); | |
6477 | ||
6478 | if (GET_CODE (op) == CONST | |
6479 | && (GET_CODE (XEXP (op, 0)) == PLUS | |
6480 | || GET_CODE (XEXP (op, 0)) == MINUS)) | |
6481 | { | |
6482 | v = INTVAL (XEXP (XEXP (op, 0), 1)); | |
6483 | op = XEXP (XEXP (op, 0), 0); | |
6484 | } | |
6485 | else if (GET_CODE (op) == CONST_INT) | |
6486 | v = INTVAL (op); | |
5df189be | 6487 | else if (GET_CODE (op) == CONST_VECTOR |
6488 | && GET_CODE (CONST_VECTOR_ELT (op, 0)) == CONST_INT) | |
6489 | v = INTVAL (CONST_VECTOR_ELT (op, 0)); | |
6490 | ||
6491 | /* The default for v is 0 which is valid in every range. */ | |
6492 | if (v < spu_builtin_range[range].low | |
6493 | || v > spu_builtin_range[range].high) | |
bf776685 | 6494 | error ("%s expects an integer literal in the range [%d, %d]. (%wd)", |
5df189be | 6495 | d->name, |
6496 | spu_builtin_range[range].low, spu_builtin_range[range].high, | |
6497 | v); | |
6352eedf | 6498 | |
6499 | switch (p) | |
6500 | { | |
6501 | case SPU_BTI_S10_4: | |
6502 | lsbits = 4; | |
6503 | break; | |
6504 | case SPU_BTI_U16_2: | |
6505 | /* This is only used in lqa, and stqa. Even though the insns | |
6506 | encode 16 bits of the address (all but the 2 least | |
6507 | significant), only 14 bits are used because it is masked to | |
6508 | be 16 byte aligned. */ | |
6509 | lsbits = 4; | |
6510 | break; | |
6511 | case SPU_BTI_S16_2: | |
6512 | /* This is used for lqr and stqr. */ | |
6513 | lsbits = 2; | |
6514 | break; | |
6515 | default: | |
6516 | lsbits = 0; | |
6517 | } | |
6518 | ||
6519 | if (GET_CODE (op) == LABEL_REF | |
6520 | || (GET_CODE (op) == SYMBOL_REF | |
6521 | && SYMBOL_REF_FUNCTION_P (op)) | |
5df189be | 6522 | || (v & ((1 << lsbits) - 1)) != 0) |
bf776685 | 6523 | warning (0, "%d least significant bits of %s are ignored", lsbits, |
6352eedf | 6524 | d->name); |
6525 | } | |
6526 | } | |
6527 | ||
6528 | ||
70ca06f8 | 6529 | static int |
5df189be | 6530 | expand_builtin_args (struct spu_builtin_description *d, tree exp, |
6352eedf | 6531 | rtx target, rtx ops[]) |
6532 | { | |
bc620c5c | 6533 | enum insn_code icode = (enum insn_code) d->icode; |
5df189be | 6534 | int i = 0, a; |
6352eedf | 6535 | |
6536 | /* Expand the arguments into rtl. */ | |
6537 | ||
6538 | if (d->parm[0] != SPU_BTI_VOID) | |
6539 | ops[i++] = target; | |
6540 | ||
70ca06f8 | 6541 | for (a = 0; d->parm[a+1] != SPU_BTI_END_OF_PARAMS; i++, a++) |
6352eedf | 6542 | { |
5df189be | 6543 | tree arg = CALL_EXPR_ARG (exp, a); |
6352eedf | 6544 | if (arg == 0) |
6545 | abort (); | |
b9c74b4d | 6546 | ops[i] = expand_expr (arg, NULL_RTX, VOIDmode, EXPAND_NORMAL); |
6352eedf | 6547 | } |
70ca06f8 | 6548 | |
32f79657 | 6549 | gcc_assert (i == insn_data[icode].n_generator_args); |
70ca06f8 | 6550 | return i; |
6352eedf | 6551 | } |
6552 | ||
6553 | static rtx | |
6554 | spu_expand_builtin_1 (struct spu_builtin_description *d, | |
5df189be | 6555 | tree exp, rtx target) |
6352eedf | 6556 | { |
6557 | rtx pat; | |
6558 | rtx ops[8]; | |
bc620c5c | 6559 | enum insn_code icode = (enum insn_code) d->icode; |
6352eedf | 6560 | enum machine_mode mode, tmode; |
6561 | int i, p; | |
70ca06f8 | 6562 | int n_operands; |
6352eedf | 6563 | tree return_type; |
6564 | ||
6565 | /* Set up ops[] with values from arglist. */ | |
70ca06f8 | 6566 | n_operands = expand_builtin_args (d, exp, target, ops); |
6352eedf | 6567 | |
6568 | /* Handle the target operand which must be operand 0. */ | |
6569 | i = 0; | |
6570 | if (d->parm[0] != SPU_BTI_VOID) | |
6571 | { | |
6572 | ||
6573 | /* We prefer the mode specified for the match_operand otherwise | |
6574 | use the mode from the builtin function prototype. */ | |
6575 | tmode = insn_data[d->icode].operand[0].mode; | |
6576 | if (tmode == VOIDmode) | |
6577 | tmode = TYPE_MODE (spu_builtin_types[d->parm[0]]); | |
6578 | ||
6579 | /* Try to use target because not using it can lead to extra copies | |
6580 | and when we are using all of the registers extra copies leads | |
6581 | to extra spills. */ | |
6582 | if (target && GET_CODE (target) == REG && GET_MODE (target) == tmode) | |
6583 | ops[0] = target; | |
6584 | else | |
6585 | target = ops[0] = gen_reg_rtx (tmode); | |
6586 | ||
6587 | if (!(*insn_data[icode].operand[0].predicate) (ops[0], tmode)) | |
6588 | abort (); | |
6589 | ||
6590 | i++; | |
6591 | } | |
6592 | ||
a76866d3 | 6593 | if (d->fcode == SPU_MASK_FOR_LOAD) |
6594 | { | |
6595 | enum machine_mode mode = insn_data[icode].operand[1].mode; | |
6596 | tree arg; | |
6597 | rtx addr, op, pat; | |
6598 | ||
6599 | /* get addr */ | |
5df189be | 6600 | arg = CALL_EXPR_ARG (exp, 0); |
4b8ee66a | 6601 | gcc_assert (POINTER_TYPE_P (TREE_TYPE (arg))); |
a76866d3 | 6602 | op = expand_expr (arg, NULL_RTX, Pmode, EXPAND_NORMAL); |
6603 | addr = memory_address (mode, op); | |
6604 | ||
6605 | /* negate addr */ | |
6606 | op = gen_reg_rtx (GET_MODE (addr)); | |
6607 | emit_insn (gen_rtx_SET (VOIDmode, op, | |
6608 | gen_rtx_NEG (GET_MODE (addr), addr))); | |
6609 | op = gen_rtx_MEM (mode, op); | |
6610 | ||
6611 | pat = GEN_FCN (icode) (target, op); | |
6612 | if (!pat) | |
6613 | return 0; | |
6614 | emit_insn (pat); | |
6615 | return target; | |
6616 | } | |
6617 | ||
6352eedf | 6618 | /* Ignore align_hint, but still expand it's args in case they have |
6619 | side effects. */ | |
6620 | if (icode == CODE_FOR_spu_align_hint) | |
6621 | return 0; | |
6622 | ||
6623 | /* Handle the rest of the operands. */ | |
70ca06f8 | 6624 | for (p = 1; i < n_operands; i++, p++) |
6352eedf | 6625 | { |
6626 | if (insn_data[d->icode].operand[i].mode != VOIDmode) | |
6627 | mode = insn_data[d->icode].operand[i].mode; | |
6628 | else | |
6629 | mode = TYPE_MODE (spu_builtin_types[d->parm[i]]); | |
6630 | ||
6631 | /* mode can be VOIDmode here for labels */ | |
6632 | ||
6633 | /* For specific intrinsics with an immediate operand, e.g., | |
6634 | si_ai(), we sometimes need to convert the scalar argument to a | |
6635 | vector argument by splatting the scalar. */ | |
6636 | if (VECTOR_MODE_P (mode) | |
6637 | && (GET_CODE (ops[i]) == CONST_INT | |
6638 | || GET_MODE_CLASS (GET_MODE (ops[i])) == MODE_INT | |
3b442530 | 6639 | || GET_MODE_CLASS (GET_MODE (ops[i])) == MODE_FLOAT)) |
6352eedf | 6640 | { |
6641 | if (GET_CODE (ops[i]) == CONST_INT) | |
6642 | ops[i] = spu_const (mode, INTVAL (ops[i])); | |
6643 | else | |
6644 | { | |
6645 | rtx reg = gen_reg_rtx (mode); | |
6646 | enum machine_mode imode = GET_MODE_INNER (mode); | |
6647 | if (!spu_nonmem_operand (ops[i], GET_MODE (ops[i]))) | |
6648 | ops[i] = force_reg (GET_MODE (ops[i]), ops[i]); | |
6649 | if (imode != GET_MODE (ops[i])) | |
6650 | ops[i] = convert_to_mode (imode, ops[i], | |
6651 | TYPE_UNSIGNED (spu_builtin_types | |
6652 | [d->parm[i]])); | |
6653 | emit_insn (gen_spu_splats (reg, ops[i])); | |
6654 | ops[i] = reg; | |
6655 | } | |
6656 | } | |
6657 | ||
5df189be | 6658 | spu_check_builtin_parm (d, ops[i], d->parm[p]); |
6659 | ||
6352eedf | 6660 | if (!(*insn_data[icode].operand[i].predicate) (ops[i], mode)) |
6661 | ops[i] = spu_force_reg (mode, ops[i]); | |
6352eedf | 6662 | } |
6663 | ||
70ca06f8 | 6664 | switch (n_operands) |
6352eedf | 6665 | { |
6666 | case 0: | |
6667 | pat = GEN_FCN (icode) (0); | |
6668 | break; | |
6669 | case 1: | |
6670 | pat = GEN_FCN (icode) (ops[0]); | |
6671 | break; | |
6672 | case 2: | |
6673 | pat = GEN_FCN (icode) (ops[0], ops[1]); | |
6674 | break; | |
6675 | case 3: | |
6676 | pat = GEN_FCN (icode) (ops[0], ops[1], ops[2]); | |
6677 | break; | |
6678 | case 4: | |
6679 | pat = GEN_FCN (icode) (ops[0], ops[1], ops[2], ops[3]); | |
6680 | break; | |
6681 | case 5: | |
6682 | pat = GEN_FCN (icode) (ops[0], ops[1], ops[2], ops[3], ops[4]); | |
6683 | break; | |
6684 | case 6: | |
6685 | pat = GEN_FCN (icode) (ops[0], ops[1], ops[2], ops[3], ops[4], ops[5]); | |
6686 | break; | |
6687 | default: | |
6688 | abort (); | |
6689 | } | |
6690 | ||
6691 | if (!pat) | |
6692 | abort (); | |
6693 | ||
6694 | if (d->type == B_CALL || d->type == B_BISLED) | |
6695 | emit_call_insn (pat); | |
6696 | else if (d->type == B_JUMP) | |
6697 | { | |
6698 | emit_jump_insn (pat); | |
6699 | emit_barrier (); | |
6700 | } | |
6701 | else | |
6702 | emit_insn (pat); | |
6703 | ||
6704 | return_type = spu_builtin_types[d->parm[0]]; | |
6705 | if (d->parm[0] != SPU_BTI_VOID | |
6706 | && GET_MODE (target) != TYPE_MODE (return_type)) | |
6707 | { | |
6708 | /* target is the return value. It should always be the mode of | |
6709 | the builtin function prototype. */ | |
6710 | target = spu_force_reg (TYPE_MODE (return_type), target); | |
6711 | } | |
6712 | ||
6713 | return target; | |
6714 | } | |
6715 | ||
6716 | rtx | |
6717 | spu_expand_builtin (tree exp, | |
6718 | rtx target, | |
6719 | rtx subtarget ATTRIBUTE_UNUSED, | |
6720 | enum machine_mode mode ATTRIBUTE_UNUSED, | |
6721 | int ignore ATTRIBUTE_UNUSED) | |
6722 | { | |
5df189be | 6723 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); |
3726fe5e | 6724 | unsigned int fcode = DECL_FUNCTION_CODE (fndecl); |
6352eedf | 6725 | struct spu_builtin_description *d; |
6726 | ||
6727 | if (fcode < NUM_SPU_BUILTINS) | |
6728 | { | |
6729 | d = &spu_builtins[fcode]; | |
6730 | ||
5df189be | 6731 | return spu_expand_builtin_1 (d, exp, target); |
6352eedf | 6732 | } |
6733 | abort (); | |
6734 | } | |
6735 | ||
e99f512d | 6736 | /* Implement targetm.vectorize.builtin_mul_widen_even. */ |
6737 | static tree | |
6738 | spu_builtin_mul_widen_even (tree type) | |
6739 | { | |
e99f512d | 6740 | switch (TYPE_MODE (type)) |
6741 | { | |
6742 | case V8HImode: | |
6743 | if (TYPE_UNSIGNED (type)) | |
0c5c4d59 | 6744 | return spu_builtin_decls[SPU_MULE_0]; |
e99f512d | 6745 | else |
0c5c4d59 | 6746 | return spu_builtin_decls[SPU_MULE_1]; |
e99f512d | 6747 | break; |
6748 | default: | |
6749 | return NULL_TREE; | |
6750 | } | |
6751 | } | |
6752 | ||
6753 | /* Implement targetm.vectorize.builtin_mul_widen_odd. */ | |
6754 | static tree | |
6755 | spu_builtin_mul_widen_odd (tree type) | |
6756 | { | |
6757 | switch (TYPE_MODE (type)) | |
6758 | { | |
6759 | case V8HImode: | |
6760 | if (TYPE_UNSIGNED (type)) | |
0c5c4d59 | 6761 | return spu_builtin_decls[SPU_MULO_1]; |
e99f512d | 6762 | else |
0c5c4d59 | 6763 | return spu_builtin_decls[SPU_MULO_0]; |
e99f512d | 6764 | break; |
6765 | default: | |
6766 | return NULL_TREE; | |
6767 | } | |
6768 | } | |
6769 | ||
a76866d3 | 6770 | /* Implement targetm.vectorize.builtin_mask_for_load. */ |
6771 | static tree | |
6772 | spu_builtin_mask_for_load (void) | |
6773 | { | |
0c5c4d59 | 6774 | return spu_builtin_decls[SPU_MASK_FOR_LOAD]; |
a76866d3 | 6775 | } |
5df189be | 6776 | |
a28df51d | 6777 | /* Implement targetm.vectorize.builtin_vectorization_cost. */ |
6778 | static int | |
0822b158 | 6779 | spu_builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost, |
6780 | tree vectype ATTRIBUTE_UNUSED, | |
6781 | int misalign ATTRIBUTE_UNUSED) | |
559093aa | 6782 | { |
6783 | switch (type_of_cost) | |
6784 | { | |
6785 | case scalar_stmt: | |
6786 | case vector_stmt: | |
6787 | case vector_load: | |
6788 | case vector_store: | |
6789 | case vec_to_scalar: | |
6790 | case scalar_to_vec: | |
6791 | case cond_branch_not_taken: | |
6792 | case vec_perm: | |
6793 | return 1; | |
6794 | ||
6795 | case scalar_store: | |
6796 | return 10; | |
6797 | ||
6798 | case scalar_load: | |
6799 | /* Load + rotate. */ | |
6800 | return 2; | |
6801 | ||
6802 | case unaligned_load: | |
6803 | return 2; | |
6804 | ||
6805 | case cond_branch_taken: | |
6806 | return 6; | |
6807 | ||
6808 | default: | |
6809 | gcc_unreachable (); | |
6810 | } | |
a28df51d | 6811 | } |
6812 | ||
0e87db76 | 6813 | /* Return true iff, data reference of TYPE can reach vector alignment (16) |
6814 | after applying N number of iterations. This routine does not determine | |
6815 | how may iterations are required to reach desired alignment. */ | |
6816 | ||
6817 | static bool | |
a9f1838b | 6818 | spu_vector_alignment_reachable (const_tree type ATTRIBUTE_UNUSED, bool is_packed) |
0e87db76 | 6819 | { |
6820 | if (is_packed) | |
6821 | return false; | |
6822 | ||
6823 | /* All other types are naturally aligned. */ | |
6824 | return true; | |
6825 | } | |
6826 | ||
a0515226 | 6827 | /* Implement targetm.vectorize.builtin_vec_perm. */ |
6828 | tree | |
6829 | spu_builtin_vec_perm (tree type, tree *mask_element_type) | |
6830 | { | |
a0515226 | 6831 | *mask_element_type = unsigned_char_type_node; |
6832 | ||
6833 | switch (TYPE_MODE (type)) | |
6834 | { | |
6835 | case V16QImode: | |
6836 | if (TYPE_UNSIGNED (type)) | |
0c5c4d59 | 6837 | return spu_builtin_decls[SPU_SHUFFLE_0]; |
a0515226 | 6838 | else |
0c5c4d59 | 6839 | return spu_builtin_decls[SPU_SHUFFLE_1]; |
a0515226 | 6840 | |
6841 | case V8HImode: | |
6842 | if (TYPE_UNSIGNED (type)) | |
0c5c4d59 | 6843 | return spu_builtin_decls[SPU_SHUFFLE_2]; |
a0515226 | 6844 | else |
0c5c4d59 | 6845 | return spu_builtin_decls[SPU_SHUFFLE_3]; |
a0515226 | 6846 | |
6847 | case V4SImode: | |
6848 | if (TYPE_UNSIGNED (type)) | |
0c5c4d59 | 6849 | return spu_builtin_decls[SPU_SHUFFLE_4]; |
a0515226 | 6850 | else |
0c5c4d59 | 6851 | return spu_builtin_decls[SPU_SHUFFLE_5]; |
a0515226 | 6852 | |
6853 | case V2DImode: | |
6854 | if (TYPE_UNSIGNED (type)) | |
0c5c4d59 | 6855 | return spu_builtin_decls[SPU_SHUFFLE_6]; |
a0515226 | 6856 | else |
0c5c4d59 | 6857 | return spu_builtin_decls[SPU_SHUFFLE_7]; |
a0515226 | 6858 | |
6859 | case V4SFmode: | |
0c5c4d59 | 6860 | return spu_builtin_decls[SPU_SHUFFLE_8]; |
a0515226 | 6861 | |
6862 | case V2DFmode: | |
0c5c4d59 | 6863 | return spu_builtin_decls[SPU_SHUFFLE_9]; |
a0515226 | 6864 | |
6865 | default: | |
6866 | return NULL_TREE; | |
6867 | } | |
a0515226 | 6868 | } |
6869 | ||
6cf5579e | 6870 | /* Return the appropriate mode for a named address pointer. */ |
6871 | static enum machine_mode | |
6872 | spu_addr_space_pointer_mode (addr_space_t addrspace) | |
6873 | { | |
6874 | switch (addrspace) | |
6875 | { | |
6876 | case ADDR_SPACE_GENERIC: | |
6877 | return ptr_mode; | |
6878 | case ADDR_SPACE_EA: | |
6879 | return EAmode; | |
6880 | default: | |
6881 | gcc_unreachable (); | |
6882 | } | |
6883 | } | |
6884 | ||
6885 | /* Return the appropriate mode for a named address address. */ | |
6886 | static enum machine_mode | |
6887 | spu_addr_space_address_mode (addr_space_t addrspace) | |
6888 | { | |
6889 | switch (addrspace) | |
6890 | { | |
6891 | case ADDR_SPACE_GENERIC: | |
6892 | return Pmode; | |
6893 | case ADDR_SPACE_EA: | |
6894 | return EAmode; | |
6895 | default: | |
6896 | gcc_unreachable (); | |
6897 | } | |
6898 | } | |
6899 | ||
6900 | /* Determine if one named address space is a subset of another. */ | |
6901 | ||
6902 | static bool | |
6903 | spu_addr_space_subset_p (addr_space_t subset, addr_space_t superset) | |
6904 | { | |
6905 | gcc_assert (subset == ADDR_SPACE_GENERIC || subset == ADDR_SPACE_EA); | |
6906 | gcc_assert (superset == ADDR_SPACE_GENERIC || superset == ADDR_SPACE_EA); | |
6907 | ||
6908 | if (subset == superset) | |
6909 | return true; | |
6910 | ||
6911 | /* If we have -mno-address-space-conversion, treat __ea and generic as not | |
6912 | being subsets but instead as disjoint address spaces. */ | |
6913 | else if (!TARGET_ADDRESS_SPACE_CONVERSION) | |
6914 | return false; | |
6915 | ||
6916 | else | |
6917 | return (subset == ADDR_SPACE_GENERIC && superset == ADDR_SPACE_EA); | |
6918 | } | |
6919 | ||
6920 | /* Convert from one address space to another. */ | |
6921 | static rtx | |
6922 | spu_addr_space_convert (rtx op, tree from_type, tree to_type) | |
6923 | { | |
6924 | addr_space_t from_as = TYPE_ADDR_SPACE (TREE_TYPE (from_type)); | |
6925 | addr_space_t to_as = TYPE_ADDR_SPACE (TREE_TYPE (to_type)); | |
6926 | ||
6927 | gcc_assert (from_as == ADDR_SPACE_GENERIC || from_as == ADDR_SPACE_EA); | |
6928 | gcc_assert (to_as == ADDR_SPACE_GENERIC || to_as == ADDR_SPACE_EA); | |
6929 | ||
6930 | if (to_as == ADDR_SPACE_GENERIC && from_as == ADDR_SPACE_EA) | |
6931 | { | |
6932 | rtx result, ls; | |
6933 | ||
6934 | ls = gen_const_mem (DImode, | |
6935 | gen_rtx_SYMBOL_REF (Pmode, "__ea_local_store")); | |
6936 | set_mem_align (ls, 128); | |
6937 | ||
6938 | result = gen_reg_rtx (Pmode); | |
6939 | ls = force_reg (Pmode, convert_modes (Pmode, DImode, ls, 1)); | |
6940 | op = force_reg (Pmode, convert_modes (Pmode, EAmode, op, 1)); | |
6941 | ls = emit_conditional_move (ls, NE, op, const0_rtx, Pmode, | |
6942 | ls, const0_rtx, Pmode, 1); | |
6943 | ||
6944 | emit_insn (gen_subsi3 (result, op, ls)); | |
6945 | ||
6946 | return result; | |
6947 | } | |
6948 | ||
6949 | else if (to_as == ADDR_SPACE_EA && from_as == ADDR_SPACE_GENERIC) | |
6950 | { | |
6951 | rtx result, ls; | |
6952 | ||
6953 | ls = gen_const_mem (DImode, | |
6954 | gen_rtx_SYMBOL_REF (Pmode, "__ea_local_store")); | |
6955 | set_mem_align (ls, 128); | |
6956 | ||
6957 | result = gen_reg_rtx (EAmode); | |
6958 | ls = force_reg (EAmode, convert_modes (EAmode, DImode, ls, 1)); | |
6959 | op = force_reg (Pmode, op); | |
6960 | ls = emit_conditional_move (ls, NE, op, const0_rtx, Pmode, | |
6961 | ls, const0_rtx, EAmode, 1); | |
6962 | op = force_reg (EAmode, convert_modes (EAmode, Pmode, op, 1)); | |
6963 | ||
6964 | if (EAmode == SImode) | |
6965 | emit_insn (gen_addsi3 (result, op, ls)); | |
6966 | else | |
6967 | emit_insn (gen_adddi3 (result, op, ls)); | |
6968 | ||
6969 | return result; | |
6970 | } | |
6971 | ||
6972 | else | |
6973 | gcc_unreachable (); | |
6974 | } | |
6975 | ||
6976 | ||
d52fd16a | 6977 | /* Count the total number of instructions in each pipe and return the |
6978 | maximum, which is used as the Minimum Iteration Interval (MII) | |
6979 | in the modulo scheduler. get_pipe() will return -2, -1, 0, or 1. | |
6980 | -2 are instructions that can go in pipe0 or pipe1. */ | |
6981 | static int | |
6982 | spu_sms_res_mii (struct ddg *g) | |
6983 | { | |
6984 | int i; | |
6985 | unsigned t[4] = {0, 0, 0, 0}; | |
6986 | ||
6987 | for (i = 0; i < g->num_nodes; i++) | |
6988 | { | |
6989 | rtx insn = g->nodes[i].insn; | |
6990 | int p = get_pipe (insn) + 2; | |
6991 | ||
1e944a0b | 6992 | gcc_assert (p >= 0); |
6993 | gcc_assert (p < 4); | |
d52fd16a | 6994 | |
6995 | t[p]++; | |
6996 | if (dump_file && INSN_P (insn)) | |
6997 | fprintf (dump_file, "i%d %s %d %d\n", | |
6998 | INSN_UID (insn), | |
6999 | insn_data[INSN_CODE(insn)].name, | |
7000 | p, t[p]); | |
7001 | } | |
7002 | if (dump_file) | |
7003 | fprintf (dump_file, "%d %d %d %d\n", t[0], t[1], t[2], t[3]); | |
7004 | ||
7005 | return MAX ((t[0] + t[2] + t[3] + 1) / 2, MAX (t[2], t[3])); | |
7006 | } | |
7007 | ||
7008 | ||
5df189be | 7009 | void |
7010 | spu_init_expanders (void) | |
9d98604b | 7011 | { |
5df189be | 7012 | if (cfun) |
9d98604b | 7013 | { |
7014 | rtx r0, r1; | |
7015 | /* HARD_FRAME_REGISTER is only 128 bit aligned when | |
7016 | frame_pointer_needed is true. We don't know that until we're | |
7017 | expanding the prologue. */ | |
7018 | REGNO_POINTER_ALIGN (HARD_FRAME_POINTER_REGNUM) = 8; | |
7019 | ||
7020 | /* A number of passes use LAST_VIRTUAL_REGISTER+1 and | |
7021 | LAST_VIRTUAL_REGISTER+2 to test the back-end. We want them | |
7022 | to be treated as aligned, so generate them here. */ | |
7023 | r0 = gen_reg_rtx (SImode); | |
7024 | r1 = gen_reg_rtx (SImode); | |
7025 | mark_reg_pointer (r0, 128); | |
7026 | mark_reg_pointer (r1, 128); | |
7027 | gcc_assert (REGNO (r0) == LAST_VIRTUAL_REGISTER + 1 | |
7028 | && REGNO (r1) == LAST_VIRTUAL_REGISTER + 2); | |
7029 | } | |
ea32e033 | 7030 | } |
7031 | ||
7032 | static enum machine_mode | |
7033 | spu_libgcc_cmp_return_mode (void) | |
7034 | { | |
7035 | ||
7036 | /* For SPU word mode is TI mode so it is better to use SImode | |
7037 | for compare returns. */ | |
7038 | return SImode; | |
7039 | } | |
7040 | ||
7041 | static enum machine_mode | |
7042 | spu_libgcc_shift_count_mode (void) | |
7043 | { | |
7044 | /* For SPU word mode is TI mode so it is better to use SImode | |
7045 | for shift counts. */ | |
7046 | return SImode; | |
7047 | } | |
5a976006 | 7048 | |
7049 | /* An early place to adjust some flags after GCC has finished processing | |
7050 | * them. */ | |
7051 | static void | |
7052 | asm_file_start (void) | |
7053 | { | |
7054 | /* Variable tracking should be run after all optimizations which | |
0ef14db8 | 7055 | change order of insns. It also needs a valid CFG. Therefore, |
7056 | *if* we make nontrivial changes in machine-dependent reorg, | |
7057 | run variable tracking after those. However, if we do not run | |
7058 | our machine-dependent reorg pass, we must still run the normal | |
7059 | variable tracking pass (or else we will ICE in final since | |
7060 | debug insns have not been removed). */ | |
7061 | if (TARGET_BRANCH_HINTS && optimize) | |
7062 | { | |
7063 | spu_flag_var_tracking = flag_var_tracking; | |
7064 | flag_var_tracking = 0; | |
7065 | } | |
5a976006 | 7066 | |
7067 | default_file_start (); | |
7068 | } | |
7069 | ||
a08dfd55 | 7070 | /* Implement targetm.section_type_flags. */ |
7071 | static unsigned int | |
7072 | spu_section_type_flags (tree decl, const char *name, int reloc) | |
7073 | { | |
7074 | /* .toe needs to have type @nobits. */ | |
7075 | if (strcmp (name, ".toe") == 0) | |
7076 | return SECTION_BSS; | |
6cf5579e | 7077 | /* Don't load _ea into the current address space. */ |
7078 | if (strcmp (name, "._ea") == 0) | |
7079 | return SECTION_WRITE | SECTION_DEBUG; | |
a08dfd55 | 7080 | return default_section_type_flags (decl, name, reloc); |
7081 | } | |
c2233b46 | 7082 | |
6cf5579e | 7083 | /* Implement targetm.select_section. */ |
7084 | static section * | |
7085 | spu_select_section (tree decl, int reloc, unsigned HOST_WIDE_INT align) | |
7086 | { | |
7087 | /* Variables and constants defined in the __ea address space | |
7088 | go into a special section named "._ea". */ | |
7089 | if (TREE_TYPE (decl) != error_mark_node | |
7090 | && TYPE_ADDR_SPACE (TREE_TYPE (decl)) == ADDR_SPACE_EA) | |
7091 | { | |
7092 | /* We might get called with string constants, but get_named_section | |
7093 | doesn't like them as they are not DECLs. Also, we need to set | |
7094 | flags in that case. */ | |
7095 | if (!DECL_P (decl)) | |
7096 | return get_section ("._ea", SECTION_WRITE | SECTION_DEBUG, NULL); | |
7097 | ||
7098 | return get_named_section (decl, "._ea", reloc); | |
7099 | } | |
7100 | ||
7101 | return default_elf_select_section (decl, reloc, align); | |
7102 | } | |
7103 | ||
7104 | /* Implement targetm.unique_section. */ | |
7105 | static void | |
7106 | spu_unique_section (tree decl, int reloc) | |
7107 | { | |
7108 | /* We don't support unique section names in the __ea address | |
7109 | space for now. */ | |
7110 | if (TREE_TYPE (decl) != error_mark_node | |
7111 | && TYPE_ADDR_SPACE (TREE_TYPE (decl)) != 0) | |
7112 | return; | |
7113 | ||
7114 | default_unique_section (decl, reloc); | |
7115 | } | |
7116 | ||
56c7bfc2 | 7117 | /* Generate a constant or register which contains 2^SCALE. We assume |
7118 | the result is valid for MODE. Currently, MODE must be V4SFmode and | |
7119 | SCALE must be SImode. */ | |
7120 | rtx | |
7121 | spu_gen_exp2 (enum machine_mode mode, rtx scale) | |
7122 | { | |
7123 | gcc_assert (mode == V4SFmode); | |
7124 | gcc_assert (GET_MODE (scale) == SImode || GET_CODE (scale) == CONST_INT); | |
7125 | if (GET_CODE (scale) != CONST_INT) | |
7126 | { | |
7127 | /* unsigned int exp = (127 + scale) << 23; | |
7128 | __vector float m = (__vector float) spu_splats (exp); */ | |
7129 | rtx reg = force_reg (SImode, scale); | |
7130 | rtx exp = gen_reg_rtx (SImode); | |
7131 | rtx mul = gen_reg_rtx (mode); | |
7132 | emit_insn (gen_addsi3 (exp, reg, GEN_INT (127))); | |
7133 | emit_insn (gen_ashlsi3 (exp, exp, GEN_INT (23))); | |
7134 | emit_insn (gen_spu_splats (mul, gen_rtx_SUBREG (GET_MODE_INNER (mode), exp, 0))); | |
7135 | return mul; | |
7136 | } | |
7137 | else | |
7138 | { | |
7139 | HOST_WIDE_INT exp = 127 + INTVAL (scale); | |
7140 | unsigned char arr[16]; | |
7141 | arr[0] = arr[4] = arr[8] = arr[12] = exp >> 1; | |
7142 | arr[1] = arr[5] = arr[9] = arr[13] = exp << 7; | |
7143 | arr[2] = arr[6] = arr[10] = arr[14] = 0; | |
7144 | arr[3] = arr[7] = arr[11] = arr[15] = 0; | |
7145 | return array_to_constant (mode, arr); | |
7146 | } | |
7147 | } | |
7148 | ||
9d98604b | 7149 | /* After reload, just change the convert into a move instruction |
7150 | or a dead instruction. */ | |
7151 | void | |
7152 | spu_split_convert (rtx ops[]) | |
7153 | { | |
7154 | if (REGNO (ops[0]) == REGNO (ops[1])) | |
7155 | emit_note (NOTE_INSN_DELETED); | |
7156 | else | |
7157 | { | |
7158 | /* Use TImode always as this might help hard reg copyprop. */ | |
7159 | rtx op0 = gen_rtx_REG (TImode, REGNO (ops[0])); | |
7160 | rtx op1 = gen_rtx_REG (TImode, REGNO (ops[1])); | |
7161 | emit_insn (gen_move_insn (op0, op1)); | |
7162 | } | |
7163 | } | |
7164 | ||
b3878a6c | 7165 | void |
4cbad5bb | 7166 | spu_function_profiler (FILE * file, int labelno ATTRIBUTE_UNUSED) |
b3878a6c | 7167 | { |
7168 | fprintf (file, "# profile\n"); | |
7169 | fprintf (file, "brsl $75, _mcount\n"); | |
7170 | } | |
7171 | ||
329c1e4e | 7172 | /* Implement targetm.ref_may_alias_errno. */ |
7173 | static bool | |
7174 | spu_ref_may_alias_errno (ao_ref *ref) | |
7175 | { | |
7176 | tree base = ao_ref_base (ref); | |
7177 | ||
7178 | /* With SPU newlib, errno is defined as something like | |
7179 | _impure_data._errno | |
7180 | The default implementation of this target macro does not | |
7181 | recognize such expressions, so special-code for it here. */ | |
7182 | ||
7183 | if (TREE_CODE (base) == VAR_DECL | |
7184 | && !TREE_STATIC (base) | |
7185 | && DECL_EXTERNAL (base) | |
7186 | && TREE_CODE (TREE_TYPE (base)) == RECORD_TYPE | |
7187 | && strcmp (IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (base)), | |
7188 | "_impure_data") == 0 | |
7189 | /* _errno is the first member of _impure_data. */ | |
7190 | && ref->offset == 0) | |
7191 | return true; | |
7192 | ||
7193 | return default_ref_may_alias_errno (ref); | |
7194 | } | |
7195 | ||
c2233b46 | 7196 | #include "gt-spu.h" |