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