]>
Commit | Line | Data |
---|---|---|
f1717362 | 1 | /* Copyright (C) 2006-2016 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" | |
9ef16211 | 20 | #include "backend.h" |
c1eb80de | 21 | #include "target.h" |
22 | #include "rtl.h" | |
9ef16211 | 23 | #include "tree.h" |
24 | #include "gimple.h" | |
c1eb80de | 25 | #include "cfghooks.h" |
26 | #include "cfgloop.h" | |
9ef16211 | 27 | #include "df.h" |
c1eb80de | 28 | #include "tm_p.h" |
29 | #include "stringpool.h" | |
30 | #include "expmed.h" | |
31 | #include "optabs.h" | |
644459d0 | 32 | #include "regs.h" |
c1eb80de | 33 | #include "emit-rtl.h" |
644459d0 | 34 | #include "recog.h" |
c1eb80de | 35 | #include "diagnostic-core.h" |
36 | #include "insn-attr.h" | |
b20a8bb4 | 37 | #include "alias.h" |
b20a8bb4 | 38 | #include "fold-const.h" |
9ed99284 | 39 | #include "stor-layout.h" |
40 | #include "calls.h" | |
41 | #include "varasm.h" | |
d53441c8 | 42 | #include "explow.h" |
644459d0 | 43 | #include "expr.h" |
644459d0 | 44 | #include "output.h" |
94ea8568 | 45 | #include "cfgrtl.h" |
94ea8568 | 46 | #include "cfgbuild.h" |
644459d0 | 47 | #include "langhooks.h" |
48 | #include "reload.h" | |
644459d0 | 49 | #include "sched-int.h" |
50 | #include "params.h" | |
a8783bee | 51 | #include "gimplify.h" |
644459d0 | 52 | #include "tm-constrs.h" |
94ea8568 | 53 | #include "ddg.h" |
b9ed1410 | 54 | #include "dumpfile.h" |
f7715905 | 55 | #include "builtins.h" |
6f4e40cd | 56 | #include "rtl-iter.h" |
6352eedf | 57 | |
0c71fb4f | 58 | /* This file should be included last. */ |
4b498588 | 59 | #include "target-def.h" |
60 | ||
6352eedf | 61 | /* Builtin types, data and prototypes. */ |
c2233b46 | 62 | |
63 | enum spu_builtin_type_index | |
64 | { | |
65 | SPU_BTI_END_OF_PARAMS, | |
66 | ||
67 | /* We create new type nodes for these. */ | |
68 | SPU_BTI_V16QI, | |
69 | SPU_BTI_V8HI, | |
70 | SPU_BTI_V4SI, | |
71 | SPU_BTI_V2DI, | |
72 | SPU_BTI_V4SF, | |
73 | SPU_BTI_V2DF, | |
74 | SPU_BTI_UV16QI, | |
75 | SPU_BTI_UV8HI, | |
76 | SPU_BTI_UV4SI, | |
77 | SPU_BTI_UV2DI, | |
78 | ||
79 | /* A 16-byte type. (Implemented with V16QI_type_node) */ | |
80 | SPU_BTI_QUADWORD, | |
81 | ||
82 | /* These all correspond to intSI_type_node */ | |
83 | SPU_BTI_7, | |
84 | SPU_BTI_S7, | |
85 | SPU_BTI_U7, | |
86 | SPU_BTI_S10, | |
87 | SPU_BTI_S10_4, | |
88 | SPU_BTI_U14, | |
89 | SPU_BTI_16, | |
90 | SPU_BTI_S16, | |
91 | SPU_BTI_S16_2, | |
92 | SPU_BTI_U16, | |
93 | SPU_BTI_U16_2, | |
94 | SPU_BTI_U18, | |
95 | ||
96 | /* These correspond to the standard types */ | |
97 | SPU_BTI_INTQI, | |
98 | SPU_BTI_INTHI, | |
99 | SPU_BTI_INTSI, | |
100 | SPU_BTI_INTDI, | |
101 | ||
102 | SPU_BTI_UINTQI, | |
103 | SPU_BTI_UINTHI, | |
104 | SPU_BTI_UINTSI, | |
105 | SPU_BTI_UINTDI, | |
106 | ||
107 | SPU_BTI_FLOAT, | |
108 | SPU_BTI_DOUBLE, | |
109 | ||
110 | SPU_BTI_VOID, | |
111 | SPU_BTI_PTR, | |
112 | ||
113 | SPU_BTI_MAX | |
114 | }; | |
115 | ||
116 | #define V16QI_type_node (spu_builtin_types[SPU_BTI_V16QI]) | |
117 | #define V8HI_type_node (spu_builtin_types[SPU_BTI_V8HI]) | |
118 | #define V4SI_type_node (spu_builtin_types[SPU_BTI_V4SI]) | |
119 | #define V2DI_type_node (spu_builtin_types[SPU_BTI_V2DI]) | |
120 | #define V4SF_type_node (spu_builtin_types[SPU_BTI_V4SF]) | |
121 | #define V2DF_type_node (spu_builtin_types[SPU_BTI_V2DF]) | |
122 | #define unsigned_V16QI_type_node (spu_builtin_types[SPU_BTI_UV16QI]) | |
123 | #define unsigned_V8HI_type_node (spu_builtin_types[SPU_BTI_UV8HI]) | |
124 | #define unsigned_V4SI_type_node (spu_builtin_types[SPU_BTI_UV4SI]) | |
125 | #define unsigned_V2DI_type_node (spu_builtin_types[SPU_BTI_UV2DI]) | |
126 | ||
127 | static GTY(()) tree spu_builtin_types[SPU_BTI_MAX]; | |
128 | ||
6352eedf | 129 | struct spu_builtin_range |
130 | { | |
131 | int low, high; | |
132 | }; | |
133 | ||
134 | static struct spu_builtin_range spu_builtin_range[] = { | |
135 | {-0x40ll, 0x7fll}, /* SPU_BTI_7 */ | |
136 | {-0x40ll, 0x3fll}, /* SPU_BTI_S7 */ | |
137 | {0ll, 0x7fll}, /* SPU_BTI_U7 */ | |
138 | {-0x200ll, 0x1ffll}, /* SPU_BTI_S10 */ | |
139 | {-0x2000ll, 0x1fffll}, /* SPU_BTI_S10_4 */ | |
140 | {0ll, 0x3fffll}, /* SPU_BTI_U14 */ | |
141 | {-0x8000ll, 0xffffll}, /* SPU_BTI_16 */ | |
142 | {-0x8000ll, 0x7fffll}, /* SPU_BTI_S16 */ | |
143 | {-0x20000ll, 0x1ffffll}, /* SPU_BTI_S16_2 */ | |
144 | {0ll, 0xffffll}, /* SPU_BTI_U16 */ | |
145 | {0ll, 0x3ffffll}, /* SPU_BTI_U16_2 */ | |
146 | {0ll, 0x3ffffll}, /* SPU_BTI_U18 */ | |
147 | }; | |
148 | ||
644459d0 | 149 | \f |
150 | /* Target specific attribute specifications. */ | |
151 | char regs_ever_allocated[FIRST_PSEUDO_REGISTER]; | |
152 | ||
153 | /* Prototypes and external defs. */ | |
0af56f80 | 154 | static int get_pipe (rtx_insn *insn); |
644459d0 | 155 | static int spu_naked_function_p (tree func); |
644459d0 | 156 | static int mem_is_padded_component_ref (rtx x); |
c7b91b14 | 157 | static void fix_range (const char *); |
9d98604b | 158 | static rtx spu_expand_load (rtx, rtx, rtx, int); |
644459d0 | 159 | |
5474166e | 160 | /* Which instruction set architecture to use. */ |
161 | int spu_arch; | |
162 | /* Which cpu are we tuning for. */ | |
163 | int spu_tune; | |
164 | ||
5a976006 | 165 | /* The hardware requires 8 insns between a hint and the branch it |
166 | effects. This variable describes how many rtl instructions the | |
167 | compiler needs to see before inserting a hint, and then the compiler | |
168 | will insert enough nops to make it at least 8 insns. The default is | |
169 | for the compiler to allow up to 2 nops be emitted. The nops are | |
170 | inserted in pairs, so we round down. */ | |
171 | int spu_hint_dist = (8*4) - (2*4); | |
172 | ||
644459d0 | 173 | enum spu_immediate { |
174 | SPU_NONE, | |
175 | SPU_IL, | |
176 | SPU_ILA, | |
177 | SPU_ILH, | |
178 | SPU_ILHU, | |
179 | SPU_ORI, | |
180 | SPU_ORHI, | |
181 | SPU_ORBI, | |
99369027 | 182 | SPU_IOHL |
644459d0 | 183 | }; |
dea01258 | 184 | enum immediate_class |
185 | { | |
186 | IC_POOL, /* constant pool */ | |
187 | IC_IL1, /* one il* instruction */ | |
188 | IC_IL2, /* both ilhu and iohl instructions */ | |
189 | IC_IL1s, /* one il* instruction */ | |
190 | IC_IL2s, /* both ilhu and iohl instructions */ | |
191 | IC_FSMBI, /* the fsmbi instruction */ | |
192 | IC_CPAT, /* one of the c*d instructions */ | |
5df189be | 193 | IC_FSMBI2 /* fsmbi plus 1 other instruction */ |
dea01258 | 194 | }; |
644459d0 | 195 | |
196 | static enum spu_immediate which_immediate_load (HOST_WIDE_INT val); | |
197 | static enum spu_immediate which_logical_immediate (HOST_WIDE_INT val); | |
dea01258 | 198 | static int cpat_info(unsigned char *arr, int size, int *prun, int *pstart); |
199 | static enum immediate_class classify_immediate (rtx op, | |
3754d046 | 200 | machine_mode mode); |
644459d0 | 201 | |
6cf5579e | 202 | /* Pointer mode for __ea references. */ |
203 | #define EAmode (spu_ea_model != 32 ? DImode : SImode) | |
204 | ||
ef51d1e3 | 205 | \f |
5eb28709 | 206 | /* Define the structure for the machine field in struct function. */ |
207 | struct GTY(()) machine_function | |
208 | { | |
209 | /* Register to use for PIC accesses. */ | |
210 | rtx pic_reg; | |
211 | }; | |
212 | ||
213 | /* How to allocate a 'struct machine_function'. */ | |
214 | static struct machine_function * | |
215 | spu_init_machine_status (void) | |
216 | { | |
25a27413 | 217 | return ggc_cleared_alloc<machine_function> (); |
5eb28709 | 218 | } |
219 | ||
4c834714 | 220 | /* Implement TARGET_OPTION_OVERRIDE. */ |
221 | static void | |
222 | spu_option_override (void) | |
644459d0 | 223 | { |
5eb28709 | 224 | /* Set up function hooks. */ |
225 | init_machine_status = spu_init_machine_status; | |
226 | ||
14d408d9 | 227 | /* Small loops will be unpeeled at -O3. For SPU it is more important |
228 | to keep code small by default. */ | |
686e2769 | 229 | if (!flag_unroll_loops && !flag_peel_loops) |
e0b840fc | 230 | maybe_set_param_value (PARAM_MAX_COMPLETELY_PEEL_TIMES, 4, |
56f280c4 | 231 | global_options.x_param_values, |
232 | global_options_set.x_param_values); | |
14d408d9 | 233 | |
644459d0 | 234 | flag_omit_frame_pointer = 1; |
235 | ||
5a976006 | 236 | /* Functions must be 8 byte aligned so we correctly handle dual issue */ |
644459d0 | 237 | if (align_functions < 8) |
238 | align_functions = 8; | |
c7b91b14 | 239 | |
5a976006 | 240 | spu_hint_dist = 8*4 - spu_max_nops*4; |
241 | if (spu_hint_dist < 0) | |
242 | spu_hint_dist = 0; | |
243 | ||
c7b91b14 | 244 | if (spu_fixed_range_string) |
245 | fix_range (spu_fixed_range_string); | |
5474166e | 246 | |
247 | /* Determine processor architectural level. */ | |
248 | if (spu_arch_string) | |
249 | { | |
250 | if (strcmp (&spu_arch_string[0], "cell") == 0) | |
251 | spu_arch = PROCESSOR_CELL; | |
252 | else if (strcmp (&spu_arch_string[0], "celledp") == 0) | |
253 | spu_arch = PROCESSOR_CELLEDP; | |
254 | else | |
8e181c9d | 255 | error ("bad value (%s) for -march= switch", spu_arch_string); |
5474166e | 256 | } |
257 | ||
258 | /* Determine processor to tune for. */ | |
259 | if (spu_tune_string) | |
260 | { | |
261 | if (strcmp (&spu_tune_string[0], "cell") == 0) | |
262 | spu_tune = PROCESSOR_CELL; | |
263 | else if (strcmp (&spu_tune_string[0], "celledp") == 0) | |
264 | spu_tune = PROCESSOR_CELLEDP; | |
265 | else | |
8e181c9d | 266 | error ("bad value (%s) for -mtune= switch", spu_tune_string); |
5474166e | 267 | } |
98bbec1e | 268 | |
13684256 | 269 | /* Change defaults according to the processor architecture. */ |
270 | if (spu_arch == PROCESSOR_CELLEDP) | |
271 | { | |
272 | /* If no command line option has been otherwise specified, change | |
273 | the default to -mno-safe-hints on celledp -- only the original | |
274 | Cell/B.E. processors require this workaround. */ | |
275 | if (!(target_flags_explicit & MASK_SAFE_HINTS)) | |
276 | target_flags &= ~MASK_SAFE_HINTS; | |
277 | } | |
278 | ||
98bbec1e | 279 | REAL_MODE_FORMAT (SFmode) = &spu_single_format; |
644459d0 | 280 | } |
281 | \f | |
282 | /* Handle an attribute requiring a FUNCTION_DECL; arguments as in | |
283 | struct attribute_spec.handler. */ | |
284 | ||
644459d0 | 285 | /* True if MODE is valid for the target. By "valid", we mean able to |
286 | be manipulated in non-trivial ways. In particular, this means all | |
287 | the arithmetic is supported. */ | |
288 | static bool | |
3754d046 | 289 | spu_scalar_mode_supported_p (machine_mode mode) |
644459d0 | 290 | { |
291 | switch (mode) | |
292 | { | |
293 | case QImode: | |
294 | case HImode: | |
295 | case SImode: | |
296 | case SFmode: | |
297 | case DImode: | |
298 | case TImode: | |
299 | case DFmode: | |
300 | return true; | |
301 | ||
302 | default: | |
303 | return false; | |
304 | } | |
305 | } | |
306 | ||
307 | /* Similarly for vector modes. "Supported" here is less strict. At | |
308 | least some operations are supported; need to check optabs or builtins | |
309 | for further details. */ | |
310 | static bool | |
3754d046 | 311 | spu_vector_mode_supported_p (machine_mode mode) |
644459d0 | 312 | { |
313 | switch (mode) | |
314 | { | |
315 | case V16QImode: | |
316 | case V8HImode: | |
317 | case V4SImode: | |
318 | case V2DImode: | |
319 | case V4SFmode: | |
320 | case V2DFmode: | |
321 | return true; | |
322 | ||
323 | default: | |
324 | return false; | |
325 | } | |
326 | } | |
327 | ||
328 | /* GCC assumes that in a paradoxical SUBREG the inner mode occupies the | |
329 | least significant bytes of the outer mode. This function returns | |
330 | TRUE for the SUBREG's where this is correct. */ | |
331 | int | |
332 | valid_subreg (rtx op) | |
333 | { | |
3754d046 | 334 | machine_mode om = GET_MODE (op); |
335 | machine_mode im = GET_MODE (SUBREG_REG (op)); | |
644459d0 | 336 | return om != VOIDmode && im != VOIDmode |
337 | && (GET_MODE_SIZE (im) == GET_MODE_SIZE (om) | |
38aca5eb | 338 | || (GET_MODE_SIZE (im) <= 4 && GET_MODE_SIZE (om) <= 4) |
339 | || (GET_MODE_SIZE (im) >= 16 && GET_MODE_SIZE (om) >= 16)); | |
644459d0 | 340 | } |
341 | ||
342 | /* When insv and ext[sz]v ar passed a TI SUBREG, we want to strip it off | |
fa7637bd | 343 | and adjust the start offset. */ |
644459d0 | 344 | static rtx |
345 | adjust_operand (rtx op, HOST_WIDE_INT * start) | |
346 | { | |
3754d046 | 347 | machine_mode mode; |
644459d0 | 348 | int op_size; |
38aca5eb | 349 | /* Strip any paradoxical SUBREG. */ |
350 | if (GET_CODE (op) == SUBREG | |
351 | && (GET_MODE_BITSIZE (GET_MODE (op)) | |
352 | > GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op))))) | |
644459d0 | 353 | { |
354 | if (start) | |
355 | *start -= | |
356 | GET_MODE_BITSIZE (GET_MODE (op)) - | |
357 | GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (op))); | |
358 | op = SUBREG_REG (op); | |
359 | } | |
360 | /* If it is smaller than SI, assure a SUBREG */ | |
361 | op_size = GET_MODE_BITSIZE (GET_MODE (op)); | |
362 | if (op_size < 32) | |
363 | { | |
364 | if (start) | |
365 | *start += 32 - op_size; | |
366 | op_size = 32; | |
367 | } | |
368 | /* If it is not a MODE_INT (and/or it is smaller than SI) add a SUBREG. */ | |
369 | mode = mode_for_size (op_size, MODE_INT, 0); | |
370 | if (mode != GET_MODE (op)) | |
371 | op = gen_rtx_SUBREG (mode, op, 0); | |
372 | return op; | |
373 | } | |
374 | ||
375 | void | |
376 | spu_expand_extv (rtx ops[], int unsignedp) | |
377 | { | |
9d98604b | 378 | rtx dst = ops[0], src = ops[1]; |
644459d0 | 379 | HOST_WIDE_INT width = INTVAL (ops[2]); |
380 | HOST_WIDE_INT start = INTVAL (ops[3]); | |
9d98604b | 381 | HOST_WIDE_INT align_mask; |
382 | rtx s0, s1, mask, r0; | |
644459d0 | 383 | |
9d98604b | 384 | gcc_assert (REG_P (dst) && GET_MODE (dst) == TImode); |
644459d0 | 385 | |
9d98604b | 386 | if (MEM_P (src)) |
644459d0 | 387 | { |
9d98604b | 388 | /* First, determine if we need 1 TImode load or 2. We need only 1 |
389 | if the bits being extracted do not cross the alignment boundary | |
390 | as determined by the MEM and its address. */ | |
391 | ||
392 | align_mask = -MEM_ALIGN (src); | |
393 | if ((start & align_mask) == ((start + width - 1) & align_mask)) | |
644459d0 | 394 | { |
9d98604b | 395 | /* Alignment is sufficient for 1 load. */ |
396 | s0 = gen_reg_rtx (TImode); | |
397 | r0 = spu_expand_load (s0, 0, src, start / 8); | |
398 | start &= 7; | |
399 | if (r0) | |
400 | emit_insn (gen_rotqby_ti (s0, s0, r0)); | |
644459d0 | 401 | } |
9d98604b | 402 | else |
403 | { | |
404 | /* Need 2 loads. */ | |
405 | s0 = gen_reg_rtx (TImode); | |
406 | s1 = gen_reg_rtx (TImode); | |
407 | r0 = spu_expand_load (s0, s1, src, start / 8); | |
408 | start &= 7; | |
409 | ||
410 | gcc_assert (start + width <= 128); | |
411 | if (r0) | |
412 | { | |
413 | rtx r1 = gen_reg_rtx (SImode); | |
414 | mask = gen_reg_rtx (TImode); | |
415 | emit_move_insn (mask, GEN_INT (-1)); | |
416 | emit_insn (gen_rotqby_ti (s0, s0, r0)); | |
417 | emit_insn (gen_rotqby_ti (s1, s1, r0)); | |
418 | if (GET_CODE (r0) == CONST_INT) | |
419 | r1 = GEN_INT (INTVAL (r0) & 15); | |
420 | else | |
421 | emit_insn (gen_andsi3 (r1, r0, GEN_INT (15))); | |
422 | emit_insn (gen_shlqby_ti (mask, mask, r1)); | |
423 | emit_insn (gen_selb (s0, s1, s0, mask)); | |
424 | } | |
425 | } | |
426 | ||
427 | } | |
428 | else if (GET_CODE (src) == SUBREG) | |
429 | { | |
430 | rtx r = SUBREG_REG (src); | |
431 | gcc_assert (REG_P (r) && SCALAR_INT_MODE_P (GET_MODE (r))); | |
432 | s0 = gen_reg_rtx (TImode); | |
433 | if (GET_MODE_SIZE (GET_MODE (r)) < GET_MODE_SIZE (TImode)) | |
d1f9b275 | 434 | emit_insn (gen_rtx_SET (s0, gen_rtx_ZERO_EXTEND (TImode, r))); |
9d98604b | 435 | else |
436 | emit_move_insn (s0, src); | |
437 | } | |
438 | else | |
439 | { | |
440 | gcc_assert (REG_P (src) && GET_MODE (src) == TImode); | |
441 | s0 = gen_reg_rtx (TImode); | |
442 | emit_move_insn (s0, src); | |
644459d0 | 443 | } |
444 | ||
9d98604b | 445 | /* Now s0 is TImode and contains the bits to extract at start. */ |
446 | ||
447 | if (start) | |
448 | emit_insn (gen_rotlti3 (s0, s0, GEN_INT (start))); | |
449 | ||
450 | if (128 - width) | |
f5ff0b21 | 451 | s0 = expand_shift (RSHIFT_EXPR, TImode, s0, 128 - width, s0, unsignedp); |
644459d0 | 452 | |
9d98604b | 453 | emit_move_insn (dst, s0); |
644459d0 | 454 | } |
455 | ||
456 | void | |
457 | spu_expand_insv (rtx ops[]) | |
458 | { | |
459 | HOST_WIDE_INT width = INTVAL (ops[1]); | |
460 | HOST_WIDE_INT start = INTVAL (ops[2]); | |
a8cf79e6 | 461 | unsigned HOST_WIDE_INT maskbits; |
3754d046 | 462 | machine_mode dst_mode; |
644459d0 | 463 | rtx dst = ops[0], src = ops[3]; |
4cbad5bb | 464 | int dst_size; |
644459d0 | 465 | rtx mask; |
466 | rtx shift_reg; | |
467 | int shift; | |
468 | ||
469 | ||
470 | if (GET_CODE (ops[0]) == MEM) | |
471 | dst = gen_reg_rtx (TImode); | |
472 | else | |
473 | dst = adjust_operand (dst, &start); | |
474 | dst_mode = GET_MODE (dst); | |
475 | dst_size = GET_MODE_BITSIZE (GET_MODE (dst)); | |
476 | ||
477 | if (CONSTANT_P (src)) | |
478 | { | |
3754d046 | 479 | machine_mode m = |
644459d0 | 480 | (width <= 32 ? SImode : width <= 64 ? DImode : TImode); |
481 | src = force_reg (m, convert_to_mode (m, src, 0)); | |
482 | } | |
483 | src = adjust_operand (src, 0); | |
644459d0 | 484 | |
485 | mask = gen_reg_rtx (dst_mode); | |
486 | shift_reg = gen_reg_rtx (dst_mode); | |
487 | shift = dst_size - start - width; | |
488 | ||
489 | /* It's not safe to use subreg here because the compiler assumes | |
490 | that the SUBREG_REG is right justified in the SUBREG. */ | |
491 | convert_move (shift_reg, src, 1); | |
492 | ||
493 | if (shift > 0) | |
494 | { | |
495 | switch (dst_mode) | |
496 | { | |
497 | case SImode: | |
498 | emit_insn (gen_ashlsi3 (shift_reg, shift_reg, GEN_INT (shift))); | |
499 | break; | |
500 | case DImode: | |
501 | emit_insn (gen_ashldi3 (shift_reg, shift_reg, GEN_INT (shift))); | |
502 | break; | |
503 | case TImode: | |
504 | emit_insn (gen_ashlti3 (shift_reg, shift_reg, GEN_INT (shift))); | |
505 | break; | |
506 | default: | |
507 | abort (); | |
508 | } | |
509 | } | |
510 | else if (shift < 0) | |
511 | abort (); | |
512 | ||
513 | switch (dst_size) | |
514 | { | |
515 | case 32: | |
a8cf79e6 | 516 | maskbits = (~(unsigned HOST_WIDE_INT)0 << (32 - width - start)); |
644459d0 | 517 | if (start) |
a8cf79e6 | 518 | maskbits += ((unsigned HOST_WIDE_INT)1 << (32 - start)); |
644459d0 | 519 | emit_move_insn (mask, GEN_INT (maskbits)); |
520 | break; | |
521 | case 64: | |
a8cf79e6 | 522 | maskbits = (~(unsigned HOST_WIDE_INT)0 << (64 - width - start)); |
644459d0 | 523 | if (start) |
a8cf79e6 | 524 | maskbits += ((unsigned HOST_WIDE_INT)1 << (64 - start)); |
644459d0 | 525 | emit_move_insn (mask, GEN_INT (maskbits)); |
526 | break; | |
527 | case 128: | |
528 | { | |
529 | unsigned char arr[16]; | |
530 | int i = start / 8; | |
531 | memset (arr, 0, sizeof (arr)); | |
532 | arr[i] = 0xff >> (start & 7); | |
533 | for (i++; i <= (start + width - 1) / 8; i++) | |
534 | arr[i] = 0xff; | |
535 | arr[i - 1] &= 0xff << (7 - ((start + width - 1) & 7)); | |
536 | emit_move_insn (mask, array_to_constant (TImode, arr)); | |
537 | } | |
538 | break; | |
539 | default: | |
540 | abort (); | |
541 | } | |
542 | if (GET_CODE (ops[0]) == MEM) | |
543 | { | |
644459d0 | 544 | rtx low = gen_reg_rtx (SImode); |
644459d0 | 545 | rtx rotl = gen_reg_rtx (SImode); |
546 | rtx mask0 = gen_reg_rtx (TImode); | |
9d98604b | 547 | rtx addr; |
548 | rtx addr0; | |
549 | rtx addr1; | |
644459d0 | 550 | rtx mem; |
551 | ||
9d98604b | 552 | addr = force_reg (Pmode, XEXP (ops[0], 0)); |
553 | addr0 = gen_rtx_AND (Pmode, addr, GEN_INT (-16)); | |
644459d0 | 554 | emit_insn (gen_andsi3 (low, addr, GEN_INT (15))); |
555 | emit_insn (gen_negsi2 (rotl, low)); | |
556 | emit_insn (gen_rotqby_ti (shift_reg, shift_reg, rotl)); | |
557 | emit_insn (gen_rotqmby_ti (mask0, mask, rotl)); | |
9d98604b | 558 | mem = change_address (ops[0], TImode, addr0); |
644459d0 | 559 | set_mem_alias_set (mem, 0); |
560 | emit_move_insn (dst, mem); | |
561 | emit_insn (gen_selb (dst, dst, shift_reg, mask0)); | |
644459d0 | 562 | if (start + width > MEM_ALIGN (ops[0])) |
563 | { | |
564 | rtx shl = gen_reg_rtx (SImode); | |
565 | rtx mask1 = gen_reg_rtx (TImode); | |
566 | rtx dst1 = gen_reg_rtx (TImode); | |
567 | rtx mem1; | |
29c05e22 | 568 | addr1 = plus_constant (Pmode, addr, 16); |
9d98604b | 569 | addr1 = gen_rtx_AND (Pmode, addr1, GEN_INT (-16)); |
644459d0 | 570 | emit_insn (gen_subsi3 (shl, GEN_INT (16), low)); |
571 | emit_insn (gen_shlqby_ti (mask1, mask, shl)); | |
9d98604b | 572 | mem1 = change_address (ops[0], TImode, addr1); |
644459d0 | 573 | set_mem_alias_set (mem1, 0); |
574 | emit_move_insn (dst1, mem1); | |
575 | emit_insn (gen_selb (dst1, dst1, shift_reg, mask1)); | |
576 | emit_move_insn (mem1, dst1); | |
577 | } | |
9d98604b | 578 | emit_move_insn (mem, dst); |
644459d0 | 579 | } |
580 | else | |
71cd778d | 581 | emit_insn (gen_selb (dst, copy_rtx (dst), shift_reg, mask)); |
644459d0 | 582 | } |
583 | ||
584 | ||
585 | int | |
586 | spu_expand_block_move (rtx ops[]) | |
587 | { | |
588 | HOST_WIDE_INT bytes, align, offset; | |
589 | rtx src, dst, sreg, dreg, target; | |
590 | int i; | |
591 | if (GET_CODE (ops[2]) != CONST_INT | |
592 | || GET_CODE (ops[3]) != CONST_INT | |
48eb4342 | 593 | || INTVAL (ops[2]) > (HOST_WIDE_INT) (MOVE_RATIO (optimize_insn_for_speed_p ()) * 8)) |
644459d0 | 594 | return 0; |
595 | ||
596 | bytes = INTVAL (ops[2]); | |
597 | align = INTVAL (ops[3]); | |
598 | ||
599 | if (bytes <= 0) | |
600 | return 1; | |
601 | ||
602 | dst = ops[0]; | |
603 | src = ops[1]; | |
604 | ||
605 | if (align == 16) | |
606 | { | |
607 | for (offset = 0; offset + 16 <= bytes; offset += 16) | |
608 | { | |
609 | dst = adjust_address (ops[0], V16QImode, offset); | |
610 | src = adjust_address (ops[1], V16QImode, offset); | |
611 | emit_move_insn (dst, src); | |
612 | } | |
613 | if (offset < bytes) | |
614 | { | |
615 | rtx mask; | |
616 | unsigned char arr[16] = { 0 }; | |
617 | for (i = 0; i < bytes - offset; i++) | |
618 | arr[i] = 0xff; | |
619 | dst = adjust_address (ops[0], V16QImode, offset); | |
620 | src = adjust_address (ops[1], V16QImode, offset); | |
621 | mask = gen_reg_rtx (V16QImode); | |
622 | sreg = gen_reg_rtx (V16QImode); | |
623 | dreg = gen_reg_rtx (V16QImode); | |
624 | target = gen_reg_rtx (V16QImode); | |
625 | emit_move_insn (mask, array_to_constant (V16QImode, arr)); | |
626 | emit_move_insn (dreg, dst); | |
627 | emit_move_insn (sreg, src); | |
628 | emit_insn (gen_selb (target, dreg, sreg, mask)); | |
629 | emit_move_insn (dst, target); | |
630 | } | |
631 | return 1; | |
632 | } | |
633 | return 0; | |
634 | } | |
635 | ||
636 | enum spu_comp_code | |
637 | { SPU_EQ, SPU_GT, SPU_GTU }; | |
638 | ||
5474166e | 639 | int spu_comp_icode[12][3] = { |
640 | {CODE_FOR_ceq_qi, CODE_FOR_cgt_qi, CODE_FOR_clgt_qi}, | |
641 | {CODE_FOR_ceq_hi, CODE_FOR_cgt_hi, CODE_FOR_clgt_hi}, | |
642 | {CODE_FOR_ceq_si, CODE_FOR_cgt_si, CODE_FOR_clgt_si}, | |
643 | {CODE_FOR_ceq_di, CODE_FOR_cgt_di, CODE_FOR_clgt_di}, | |
644 | {CODE_FOR_ceq_ti, CODE_FOR_cgt_ti, CODE_FOR_clgt_ti}, | |
645 | {CODE_FOR_ceq_sf, CODE_FOR_cgt_sf, 0}, | |
646 | {CODE_FOR_ceq_df, CODE_FOR_cgt_df, 0}, | |
647 | {CODE_FOR_ceq_v16qi, CODE_FOR_cgt_v16qi, CODE_FOR_clgt_v16qi}, | |
648 | {CODE_FOR_ceq_v8hi, CODE_FOR_cgt_v8hi, CODE_FOR_clgt_v8hi}, | |
649 | {CODE_FOR_ceq_v4si, CODE_FOR_cgt_v4si, CODE_FOR_clgt_v4si}, | |
650 | {CODE_FOR_ceq_v4sf, CODE_FOR_cgt_v4sf, 0}, | |
651 | {CODE_FOR_ceq_v2df, CODE_FOR_cgt_v2df, 0}, | |
644459d0 | 652 | }; |
653 | ||
654 | /* Generate a compare for CODE. Return a brand-new rtx that represents | |
655 | the result of the compare. GCC can figure this out too if we don't | |
656 | provide all variations of compares, but GCC always wants to use | |
657 | WORD_MODE, we can generate better code in most cases if we do it | |
658 | ourselves. */ | |
659 | void | |
74f4459c | 660 | spu_emit_branch_or_set (int is_set, rtx cmp, rtx operands[]) |
644459d0 | 661 | { |
662 | int reverse_compare = 0; | |
663 | int reverse_test = 0; | |
5d70b918 | 664 | rtx compare_result, eq_result; |
665 | rtx comp_rtx, eq_rtx; | |
3754d046 | 666 | machine_mode comp_mode; |
667 | machine_mode op_mode; | |
b9c74b4d | 668 | enum spu_comp_code scode, eq_code; |
669 | enum insn_code ior_code; | |
74f4459c | 670 | enum rtx_code code = GET_CODE (cmp); |
671 | rtx op0 = XEXP (cmp, 0); | |
672 | rtx op1 = XEXP (cmp, 1); | |
644459d0 | 673 | int index; |
5d70b918 | 674 | int eq_test = 0; |
644459d0 | 675 | |
74f4459c | 676 | /* When op1 is a CONST_INT change (X >= C) to (X > C-1), |
644459d0 | 677 | and so on, to keep the constant in operand 1. */ |
74f4459c | 678 | if (GET_CODE (op1) == CONST_INT) |
644459d0 | 679 | { |
74f4459c | 680 | HOST_WIDE_INT val = INTVAL (op1) - 1; |
681 | if (trunc_int_for_mode (val, GET_MODE (op0)) == val) | |
644459d0 | 682 | switch (code) |
683 | { | |
684 | case GE: | |
74f4459c | 685 | op1 = GEN_INT (val); |
644459d0 | 686 | code = GT; |
687 | break; | |
688 | case LT: | |
74f4459c | 689 | op1 = GEN_INT (val); |
644459d0 | 690 | code = LE; |
691 | break; | |
692 | case GEU: | |
74f4459c | 693 | op1 = GEN_INT (val); |
644459d0 | 694 | code = GTU; |
695 | break; | |
696 | case LTU: | |
74f4459c | 697 | op1 = GEN_INT (val); |
644459d0 | 698 | code = LEU; |
699 | break; | |
700 | default: | |
701 | break; | |
702 | } | |
703 | } | |
704 | ||
686195ea | 705 | /* However, if we generate an integer result, performing a reverse test |
706 | would require an extra negation, so avoid that where possible. */ | |
707 | if (GET_CODE (op1) == CONST_INT && is_set == 1) | |
708 | { | |
709 | HOST_WIDE_INT val = INTVAL (op1) + 1; | |
710 | if (trunc_int_for_mode (val, GET_MODE (op0)) == val) | |
711 | switch (code) | |
712 | { | |
713 | case LE: | |
714 | op1 = GEN_INT (val); | |
715 | code = LT; | |
716 | break; | |
717 | case LEU: | |
718 | op1 = GEN_INT (val); | |
719 | code = LTU; | |
720 | break; | |
721 | default: | |
722 | break; | |
723 | } | |
724 | } | |
725 | ||
5d70b918 | 726 | comp_mode = SImode; |
74f4459c | 727 | op_mode = GET_MODE (op0); |
5d70b918 | 728 | |
644459d0 | 729 | switch (code) |
730 | { | |
731 | case GE: | |
644459d0 | 732 | scode = SPU_GT; |
07027691 | 733 | if (HONOR_NANS (op_mode)) |
5d70b918 | 734 | { |
735 | reverse_compare = 0; | |
736 | reverse_test = 0; | |
737 | eq_test = 1; | |
738 | eq_code = SPU_EQ; | |
739 | } | |
740 | else | |
741 | { | |
742 | reverse_compare = 1; | |
743 | reverse_test = 1; | |
744 | } | |
644459d0 | 745 | break; |
746 | case LE: | |
644459d0 | 747 | scode = SPU_GT; |
07027691 | 748 | if (HONOR_NANS (op_mode)) |
5d70b918 | 749 | { |
750 | reverse_compare = 1; | |
751 | reverse_test = 0; | |
752 | eq_test = 1; | |
753 | eq_code = SPU_EQ; | |
754 | } | |
755 | else | |
756 | { | |
757 | reverse_compare = 0; | |
758 | reverse_test = 1; | |
759 | } | |
644459d0 | 760 | break; |
761 | case LT: | |
762 | reverse_compare = 1; | |
763 | reverse_test = 0; | |
764 | scode = SPU_GT; | |
765 | break; | |
766 | case GEU: | |
767 | reverse_compare = 1; | |
768 | reverse_test = 1; | |
769 | scode = SPU_GTU; | |
770 | break; | |
771 | case LEU: | |
772 | reverse_compare = 0; | |
773 | reverse_test = 1; | |
774 | scode = SPU_GTU; | |
775 | break; | |
776 | case LTU: | |
777 | reverse_compare = 1; | |
778 | reverse_test = 0; | |
779 | scode = SPU_GTU; | |
780 | break; | |
781 | case NE: | |
782 | reverse_compare = 0; | |
783 | reverse_test = 1; | |
784 | scode = SPU_EQ; | |
785 | break; | |
786 | ||
787 | case EQ: | |
788 | scode = SPU_EQ; | |
789 | break; | |
790 | case GT: | |
791 | scode = SPU_GT; | |
792 | break; | |
793 | case GTU: | |
794 | scode = SPU_GTU; | |
795 | break; | |
796 | default: | |
797 | scode = SPU_EQ; | |
798 | break; | |
799 | } | |
800 | ||
644459d0 | 801 | switch (op_mode) |
802 | { | |
803 | case QImode: | |
804 | index = 0; | |
805 | comp_mode = QImode; | |
806 | break; | |
807 | case HImode: | |
808 | index = 1; | |
809 | comp_mode = HImode; | |
810 | break; | |
811 | case SImode: | |
812 | index = 2; | |
813 | break; | |
814 | case DImode: | |
815 | index = 3; | |
816 | break; | |
817 | case TImode: | |
818 | index = 4; | |
819 | break; | |
820 | case SFmode: | |
821 | index = 5; | |
822 | break; | |
823 | case DFmode: | |
824 | index = 6; | |
825 | break; | |
826 | case V16QImode: | |
5474166e | 827 | index = 7; |
828 | comp_mode = op_mode; | |
829 | break; | |
644459d0 | 830 | case V8HImode: |
5474166e | 831 | index = 8; |
832 | comp_mode = op_mode; | |
833 | break; | |
644459d0 | 834 | case V4SImode: |
5474166e | 835 | index = 9; |
836 | comp_mode = op_mode; | |
837 | break; | |
644459d0 | 838 | case V4SFmode: |
5474166e | 839 | index = 10; |
840 | comp_mode = V4SImode; | |
841 | break; | |
644459d0 | 842 | case V2DFmode: |
5474166e | 843 | index = 11; |
844 | comp_mode = V2DImode; | |
644459d0 | 845 | break; |
5474166e | 846 | case V2DImode: |
644459d0 | 847 | default: |
848 | abort (); | |
849 | } | |
850 | ||
74f4459c | 851 | if (GET_MODE (op1) == DFmode |
07027691 | 852 | && (scode != SPU_GT && scode != SPU_EQ)) |
853 | abort (); | |
644459d0 | 854 | |
74f4459c | 855 | if (is_set == 0 && op1 == const0_rtx |
856 | && (GET_MODE (op0) == SImode | |
686195ea | 857 | || GET_MODE (op0) == HImode |
858 | || GET_MODE (op0) == QImode) && scode == SPU_EQ) | |
644459d0 | 859 | { |
860 | /* Don't need to set a register with the result when we are | |
861 | comparing against zero and branching. */ | |
862 | reverse_test = !reverse_test; | |
74f4459c | 863 | compare_result = op0; |
644459d0 | 864 | } |
865 | else | |
866 | { | |
867 | compare_result = gen_reg_rtx (comp_mode); | |
868 | ||
869 | if (reverse_compare) | |
870 | { | |
74f4459c | 871 | rtx t = op1; |
872 | op1 = op0; | |
873 | op0 = t; | |
644459d0 | 874 | } |
875 | ||
876 | if (spu_comp_icode[index][scode] == 0) | |
877 | abort (); | |
878 | ||
879 | if (!(*insn_data[spu_comp_icode[index][scode]].operand[1].predicate) | |
74f4459c | 880 | (op0, op_mode)) |
881 | op0 = force_reg (op_mode, op0); | |
644459d0 | 882 | if (!(*insn_data[spu_comp_icode[index][scode]].operand[2].predicate) |
74f4459c | 883 | (op1, op_mode)) |
884 | op1 = force_reg (op_mode, op1); | |
644459d0 | 885 | comp_rtx = GEN_FCN (spu_comp_icode[index][scode]) (compare_result, |
74f4459c | 886 | op0, op1); |
644459d0 | 887 | if (comp_rtx == 0) |
888 | abort (); | |
889 | emit_insn (comp_rtx); | |
890 | ||
5d70b918 | 891 | if (eq_test) |
892 | { | |
893 | eq_result = gen_reg_rtx (comp_mode); | |
894 | eq_rtx = GEN_FCN (spu_comp_icode[index][eq_code]) (eq_result, | |
74f4459c | 895 | op0, op1); |
5d70b918 | 896 | if (eq_rtx == 0) |
897 | abort (); | |
898 | emit_insn (eq_rtx); | |
d6bf3b14 | 899 | ior_code = optab_handler (ior_optab, comp_mode); |
5d70b918 | 900 | gcc_assert (ior_code != CODE_FOR_nothing); |
901 | emit_insn (GEN_FCN (ior_code) | |
902 | (compare_result, compare_result, eq_result)); | |
903 | } | |
644459d0 | 904 | } |
905 | ||
906 | if (is_set == 0) | |
907 | { | |
908 | rtx bcomp; | |
909 | rtx loc_ref; | |
910 | ||
911 | /* We don't have branch on QI compare insns, so we convert the | |
912 | QI compare result to a HI result. */ | |
913 | if (comp_mode == QImode) | |
914 | { | |
915 | rtx old_res = compare_result; | |
916 | compare_result = gen_reg_rtx (HImode); | |
917 | comp_mode = HImode; | |
918 | emit_insn (gen_extendqihi2 (compare_result, old_res)); | |
919 | } | |
920 | ||
921 | if (reverse_test) | |
922 | bcomp = gen_rtx_EQ (comp_mode, compare_result, const0_rtx); | |
923 | else | |
924 | bcomp = gen_rtx_NE (comp_mode, compare_result, const0_rtx); | |
925 | ||
74f4459c | 926 | loc_ref = gen_rtx_LABEL_REF (VOIDmode, operands[3]); |
d1f9b275 | 927 | emit_jump_insn (gen_rtx_SET (pc_rtx, |
644459d0 | 928 | gen_rtx_IF_THEN_ELSE (VOIDmode, bcomp, |
929 | loc_ref, pc_rtx))); | |
930 | } | |
931 | else if (is_set == 2) | |
932 | { | |
74f4459c | 933 | rtx target = operands[0]; |
644459d0 | 934 | int compare_size = GET_MODE_BITSIZE (comp_mode); |
935 | int target_size = GET_MODE_BITSIZE (GET_MODE (target)); | |
3754d046 | 936 | machine_mode mode = mode_for_size (target_size, MODE_INT, 0); |
644459d0 | 937 | rtx select_mask; |
938 | rtx op_t = operands[2]; | |
939 | rtx op_f = operands[3]; | |
940 | ||
941 | /* The result of the comparison can be SI, HI or QI mode. Create a | |
942 | mask based on that result. */ | |
943 | if (target_size > compare_size) | |
944 | { | |
945 | select_mask = gen_reg_rtx (mode); | |
946 | emit_insn (gen_extend_compare (select_mask, compare_result)); | |
947 | } | |
948 | else if (target_size < compare_size) | |
949 | select_mask = | |
950 | gen_rtx_SUBREG (mode, compare_result, | |
951 | (compare_size - target_size) / BITS_PER_UNIT); | |
952 | else if (comp_mode != mode) | |
953 | select_mask = gen_rtx_SUBREG (mode, compare_result, 0); | |
954 | else | |
955 | select_mask = compare_result; | |
956 | ||
957 | if (GET_MODE (target) != GET_MODE (op_t) | |
958 | || GET_MODE (target) != GET_MODE (op_f)) | |
959 | abort (); | |
960 | ||
961 | if (reverse_test) | |
962 | emit_insn (gen_selb (target, op_t, op_f, select_mask)); | |
963 | else | |
964 | emit_insn (gen_selb (target, op_f, op_t, select_mask)); | |
965 | } | |
966 | else | |
967 | { | |
74f4459c | 968 | rtx target = operands[0]; |
644459d0 | 969 | if (reverse_test) |
d1f9b275 | 970 | emit_insn (gen_rtx_SET (compare_result, |
644459d0 | 971 | gen_rtx_NOT (comp_mode, compare_result))); |
972 | if (GET_MODE (target) == SImode && GET_MODE (compare_result) == HImode) | |
973 | emit_insn (gen_extendhisi2 (target, compare_result)); | |
974 | else if (GET_MODE (target) == SImode | |
975 | && GET_MODE (compare_result) == QImode) | |
976 | emit_insn (gen_extend_compare (target, compare_result)); | |
977 | else | |
978 | emit_move_insn (target, compare_result); | |
979 | } | |
980 | } | |
981 | ||
982 | HOST_WIDE_INT | |
983 | const_double_to_hwint (rtx x) | |
984 | { | |
985 | HOST_WIDE_INT val; | |
644459d0 | 986 | if (GET_MODE (x) == SFmode) |
945f7b03 | 987 | REAL_VALUE_TO_TARGET_SINGLE (*CONST_DOUBLE_REAL_VALUE (x), val); |
644459d0 | 988 | else if (GET_MODE (x) == DFmode) |
989 | { | |
990 | long l[2]; | |
945f7b03 | 991 | REAL_VALUE_TO_TARGET_DOUBLE (*CONST_DOUBLE_REAL_VALUE (x), l); |
644459d0 | 992 | val = l[0]; |
993 | val = (val << 32) | (l[1] & 0xffffffff); | |
994 | } | |
995 | else | |
996 | abort (); | |
997 | return val; | |
998 | } | |
999 | ||
1000 | rtx | |
3754d046 | 1001 | hwint_to_const_double (machine_mode mode, HOST_WIDE_INT v) |
644459d0 | 1002 | { |
1003 | long tv[2]; | |
1004 | REAL_VALUE_TYPE rv; | |
1005 | gcc_assert (mode == SFmode || mode == DFmode); | |
1006 | ||
1007 | if (mode == SFmode) | |
1008 | tv[0] = (v << 32) >> 32; | |
1009 | else if (mode == DFmode) | |
1010 | { | |
1011 | tv[1] = (v << 32) >> 32; | |
1012 | tv[0] = v >> 32; | |
1013 | } | |
1014 | real_from_target (&rv, tv, mode); | |
d5f9611d | 1015 | return const_double_from_real_value (rv, mode); |
644459d0 | 1016 | } |
1017 | ||
1018 | void | |
1019 | print_operand_address (FILE * file, register rtx addr) | |
1020 | { | |
1021 | rtx reg; | |
1022 | rtx offset; | |
1023 | ||
e04cf423 | 1024 | if (GET_CODE (addr) == AND |
1025 | && GET_CODE (XEXP (addr, 1)) == CONST_INT | |
1026 | && INTVAL (XEXP (addr, 1)) == -16) | |
1027 | addr = XEXP (addr, 0); | |
1028 | ||
644459d0 | 1029 | switch (GET_CODE (addr)) |
1030 | { | |
1031 | case REG: | |
1032 | fprintf (file, "0(%s)", reg_names[REGNO (addr)]); | |
1033 | break; | |
1034 | ||
1035 | case PLUS: | |
1036 | reg = XEXP (addr, 0); | |
1037 | offset = XEXP (addr, 1); | |
1038 | if (GET_CODE (offset) == REG) | |
1039 | { | |
1040 | fprintf (file, "%s,%s", reg_names[REGNO (reg)], | |
1041 | reg_names[REGNO (offset)]); | |
1042 | } | |
1043 | else if (GET_CODE (offset) == CONST_INT) | |
1044 | { | |
1045 | fprintf (file, HOST_WIDE_INT_PRINT_DEC "(%s)", | |
1046 | INTVAL (offset), reg_names[REGNO (reg)]); | |
1047 | } | |
1048 | else | |
1049 | abort (); | |
1050 | break; | |
1051 | ||
1052 | case CONST: | |
1053 | case LABEL_REF: | |
1054 | case SYMBOL_REF: | |
1055 | case CONST_INT: | |
1056 | output_addr_const (file, addr); | |
1057 | break; | |
1058 | ||
1059 | default: | |
1060 | debug_rtx (addr); | |
1061 | abort (); | |
1062 | } | |
1063 | } | |
1064 | ||
1065 | void | |
1066 | print_operand (FILE * file, rtx x, int code) | |
1067 | { | |
3754d046 | 1068 | machine_mode mode = GET_MODE (x); |
644459d0 | 1069 | HOST_WIDE_INT val; |
1070 | unsigned char arr[16]; | |
1071 | int xcode = GET_CODE (x); | |
dea01258 | 1072 | int i, info; |
644459d0 | 1073 | if (GET_MODE (x) == VOIDmode) |
1074 | switch (code) | |
1075 | { | |
644459d0 | 1076 | case 'L': /* 128 bits, signed */ |
1077 | case 'm': /* 128 bits, signed */ | |
1078 | case 'T': /* 128 bits, signed */ | |
1079 | case 't': /* 128 bits, signed */ | |
1080 | mode = TImode; | |
1081 | break; | |
644459d0 | 1082 | case 'K': /* 64 bits, signed */ |
1083 | case 'k': /* 64 bits, signed */ | |
1084 | case 'D': /* 64 bits, signed */ | |
1085 | case 'd': /* 64 bits, signed */ | |
1086 | mode = DImode; | |
1087 | break; | |
644459d0 | 1088 | case 'J': /* 32 bits, signed */ |
1089 | case 'j': /* 32 bits, signed */ | |
1090 | case 's': /* 32 bits, signed */ | |
1091 | case 'S': /* 32 bits, signed */ | |
1092 | mode = SImode; | |
1093 | break; | |
1094 | } | |
1095 | switch (code) | |
1096 | { | |
1097 | ||
1098 | case 'j': /* 32 bits, signed */ | |
1099 | case 'k': /* 64 bits, signed */ | |
1100 | case 'm': /* 128 bits, signed */ | |
1101 | if (xcode == CONST_INT | |
1102 | || xcode == CONST_DOUBLE || xcode == CONST_VECTOR) | |
1103 | { | |
1104 | gcc_assert (logical_immediate_p (x, mode)); | |
1105 | constant_to_array (mode, x, arr); | |
1106 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
1107 | val = trunc_int_for_mode (val, SImode); | |
1108 | switch (which_logical_immediate (val)) | |
1109 | { | |
1110 | case SPU_ORI: | |
1111 | break; | |
1112 | case SPU_ORHI: | |
1113 | fprintf (file, "h"); | |
1114 | break; | |
1115 | case SPU_ORBI: | |
1116 | fprintf (file, "b"); | |
1117 | break; | |
1118 | default: | |
1119 | gcc_unreachable(); | |
1120 | } | |
1121 | } | |
1122 | else | |
1123 | gcc_unreachable(); | |
1124 | return; | |
1125 | ||
1126 | case 'J': /* 32 bits, signed */ | |
1127 | case 'K': /* 64 bits, signed */ | |
1128 | case 'L': /* 128 bits, signed */ | |
1129 | if (xcode == CONST_INT | |
1130 | || xcode == CONST_DOUBLE || xcode == CONST_VECTOR) | |
1131 | { | |
1132 | gcc_assert (logical_immediate_p (x, mode) | |
1133 | || iohl_immediate_p (x, mode)); | |
1134 | constant_to_array (mode, x, arr); | |
1135 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
1136 | val = trunc_int_for_mode (val, SImode); | |
1137 | switch (which_logical_immediate (val)) | |
1138 | { | |
1139 | case SPU_ORI: | |
1140 | case SPU_IOHL: | |
1141 | break; | |
1142 | case SPU_ORHI: | |
1143 | val = trunc_int_for_mode (val, HImode); | |
1144 | break; | |
1145 | case SPU_ORBI: | |
1146 | val = trunc_int_for_mode (val, QImode); | |
1147 | break; | |
1148 | default: | |
1149 | gcc_unreachable(); | |
1150 | } | |
1151 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, val); | |
1152 | } | |
1153 | else | |
1154 | gcc_unreachable(); | |
1155 | return; | |
1156 | ||
1157 | case 't': /* 128 bits, signed */ | |
1158 | case 'd': /* 64 bits, signed */ | |
1159 | case 's': /* 32 bits, signed */ | |
dea01258 | 1160 | if (CONSTANT_P (x)) |
644459d0 | 1161 | { |
dea01258 | 1162 | enum immediate_class c = classify_immediate (x, mode); |
1163 | switch (c) | |
1164 | { | |
1165 | case IC_IL1: | |
1166 | constant_to_array (mode, x, arr); | |
1167 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
1168 | val = trunc_int_for_mode (val, SImode); | |
1169 | switch (which_immediate_load (val)) | |
1170 | { | |
1171 | case SPU_IL: | |
1172 | break; | |
1173 | case SPU_ILA: | |
1174 | fprintf (file, "a"); | |
1175 | break; | |
1176 | case SPU_ILH: | |
1177 | fprintf (file, "h"); | |
1178 | break; | |
1179 | case SPU_ILHU: | |
1180 | fprintf (file, "hu"); | |
1181 | break; | |
1182 | default: | |
1183 | gcc_unreachable (); | |
1184 | } | |
1185 | break; | |
1186 | case IC_CPAT: | |
1187 | constant_to_array (mode, x, arr); | |
1188 | cpat_info (arr, GET_MODE_SIZE (mode), &info, 0); | |
1189 | if (info == 1) | |
1190 | fprintf (file, "b"); | |
1191 | else if (info == 2) | |
1192 | fprintf (file, "h"); | |
1193 | else if (info == 4) | |
1194 | fprintf (file, "w"); | |
1195 | else if (info == 8) | |
1196 | fprintf (file, "d"); | |
1197 | break; | |
1198 | case IC_IL1s: | |
1199 | if (xcode == CONST_VECTOR) | |
1200 | { | |
1201 | x = CONST_VECTOR_ELT (x, 0); | |
1202 | xcode = GET_CODE (x); | |
1203 | } | |
1204 | if (xcode == SYMBOL_REF || xcode == LABEL_REF || xcode == CONST) | |
1205 | fprintf (file, "a"); | |
1206 | else if (xcode == HIGH) | |
1207 | fprintf (file, "hu"); | |
1208 | break; | |
1209 | case IC_FSMBI: | |
5df189be | 1210 | case IC_FSMBI2: |
dea01258 | 1211 | case IC_IL2: |
1212 | case IC_IL2s: | |
1213 | case IC_POOL: | |
1214 | abort (); | |
1215 | } | |
644459d0 | 1216 | } |
644459d0 | 1217 | else |
1218 | gcc_unreachable (); | |
1219 | return; | |
1220 | ||
1221 | case 'T': /* 128 bits, signed */ | |
1222 | case 'D': /* 64 bits, signed */ | |
1223 | case 'S': /* 32 bits, signed */ | |
dea01258 | 1224 | if (CONSTANT_P (x)) |
644459d0 | 1225 | { |
dea01258 | 1226 | enum immediate_class c = classify_immediate (x, mode); |
1227 | switch (c) | |
644459d0 | 1228 | { |
dea01258 | 1229 | case IC_IL1: |
1230 | constant_to_array (mode, x, arr); | |
1231 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
1232 | val = trunc_int_for_mode (val, SImode); | |
1233 | switch (which_immediate_load (val)) | |
1234 | { | |
1235 | case SPU_IL: | |
1236 | case SPU_ILA: | |
1237 | break; | |
1238 | case SPU_ILH: | |
1239 | case SPU_ILHU: | |
1240 | val = trunc_int_for_mode (((arr[0] << 8) | arr[1]), HImode); | |
1241 | break; | |
1242 | default: | |
1243 | gcc_unreachable (); | |
1244 | } | |
1245 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, val); | |
1246 | break; | |
1247 | case IC_FSMBI: | |
1248 | constant_to_array (mode, x, arr); | |
1249 | val = 0; | |
1250 | for (i = 0; i < 16; i++) | |
1251 | { | |
1252 | val <<= 1; | |
1253 | val |= arr[i] & 1; | |
1254 | } | |
1255 | print_operand (file, GEN_INT (val), 0); | |
1256 | break; | |
1257 | case IC_CPAT: | |
1258 | constant_to_array (mode, x, arr); | |
1259 | cpat_info (arr, GET_MODE_SIZE (mode), 0, &info); | |
1260 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, (HOST_WIDE_INT)info); | |
644459d0 | 1261 | break; |
dea01258 | 1262 | case IC_IL1s: |
dea01258 | 1263 | if (xcode == HIGH) |
5df189be | 1264 | x = XEXP (x, 0); |
1265 | if (GET_CODE (x) == CONST_VECTOR) | |
1266 | x = CONST_VECTOR_ELT (x, 0); | |
1267 | output_addr_const (file, x); | |
1268 | if (xcode == HIGH) | |
1269 | fprintf (file, "@h"); | |
644459d0 | 1270 | break; |
dea01258 | 1271 | case IC_IL2: |
1272 | case IC_IL2s: | |
5df189be | 1273 | case IC_FSMBI2: |
dea01258 | 1274 | case IC_POOL: |
1275 | abort (); | |
644459d0 | 1276 | } |
c8befdb9 | 1277 | } |
644459d0 | 1278 | else |
1279 | gcc_unreachable (); | |
1280 | return; | |
1281 | ||
644459d0 | 1282 | case 'C': |
1283 | if (xcode == CONST_INT) | |
1284 | { | |
1285 | /* Only 4 least significant bits are relevant for generate | |
1286 | control word instructions. */ | |
1287 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (x) & 15); | |
1288 | return; | |
1289 | } | |
1290 | break; | |
1291 | ||
1292 | case 'M': /* print code for c*d */ | |
1293 | if (GET_CODE (x) == CONST_INT) | |
1294 | switch (INTVAL (x)) | |
1295 | { | |
1296 | case 1: | |
1297 | fprintf (file, "b"); | |
1298 | break; | |
1299 | case 2: | |
1300 | fprintf (file, "h"); | |
1301 | break; | |
1302 | case 4: | |
1303 | fprintf (file, "w"); | |
1304 | break; | |
1305 | case 8: | |
1306 | fprintf (file, "d"); | |
1307 | break; | |
1308 | default: | |
1309 | gcc_unreachable(); | |
1310 | } | |
1311 | else | |
1312 | gcc_unreachable(); | |
1313 | return; | |
1314 | ||
1315 | case 'N': /* Negate the operand */ | |
1316 | if (xcode == CONST_INT) | |
1317 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, -INTVAL (x)); | |
1318 | else if (xcode == CONST_VECTOR) | |
1319 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, | |
1320 | -INTVAL (CONST_VECTOR_ELT (x, 0))); | |
1321 | return; | |
1322 | ||
1323 | case 'I': /* enable/disable interrupts */ | |
1324 | if (xcode == CONST_INT) | |
1325 | fprintf (file, "%s", INTVAL (x) == 0 ? "d" : "e"); | |
1326 | return; | |
1327 | ||
1328 | case 'b': /* branch modifiers */ | |
1329 | if (xcode == REG) | |
1330 | fprintf (file, "%s", GET_MODE (x) == HImode ? "h" : ""); | |
1331 | else if (COMPARISON_P (x)) | |
1332 | fprintf (file, "%s", xcode == NE ? "n" : ""); | |
1333 | return; | |
1334 | ||
1335 | case 'i': /* indirect call */ | |
1336 | if (xcode == MEM) | |
1337 | { | |
1338 | if (GET_CODE (XEXP (x, 0)) == REG) | |
1339 | /* Used in indirect function calls. */ | |
1340 | fprintf (file, "%s", reg_names[REGNO (XEXP (x, 0))]); | |
1341 | else | |
3c047fe9 | 1342 | output_address (GET_MODE (x), XEXP (x, 0)); |
644459d0 | 1343 | } |
1344 | return; | |
1345 | ||
1346 | case 'p': /* load/store */ | |
1347 | if (xcode == MEM) | |
1348 | { | |
1349 | x = XEXP (x, 0); | |
1350 | xcode = GET_CODE (x); | |
1351 | } | |
e04cf423 | 1352 | if (xcode == AND) |
1353 | { | |
1354 | x = XEXP (x, 0); | |
1355 | xcode = GET_CODE (x); | |
1356 | } | |
644459d0 | 1357 | if (xcode == REG) |
1358 | fprintf (file, "d"); | |
1359 | else if (xcode == CONST_INT) | |
1360 | fprintf (file, "a"); | |
1361 | else if (xcode == CONST || xcode == SYMBOL_REF || xcode == LABEL_REF) | |
1362 | fprintf (file, "r"); | |
1363 | else if (xcode == PLUS || xcode == LO_SUM) | |
1364 | { | |
1365 | if (GET_CODE (XEXP (x, 1)) == REG) | |
1366 | fprintf (file, "x"); | |
1367 | else | |
1368 | fprintf (file, "d"); | |
1369 | } | |
1370 | return; | |
1371 | ||
5df189be | 1372 | case 'e': |
1373 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1374 | val &= 0x7; | |
1375 | output_addr_const (file, GEN_INT (val)); | |
1376 | return; | |
1377 | ||
1378 | case 'f': | |
1379 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1380 | val &= 0x1f; | |
1381 | output_addr_const (file, GEN_INT (val)); | |
1382 | return; | |
1383 | ||
1384 | case 'g': | |
1385 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1386 | val &= 0x3f; | |
1387 | output_addr_const (file, GEN_INT (val)); | |
1388 | return; | |
1389 | ||
1390 | case 'h': | |
1391 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1392 | val = (val >> 3) & 0x1f; | |
1393 | output_addr_const (file, GEN_INT (val)); | |
1394 | return; | |
1395 | ||
1396 | case 'E': | |
1397 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1398 | val = -val; | |
1399 | val &= 0x7; | |
1400 | output_addr_const (file, GEN_INT (val)); | |
1401 | return; | |
1402 | ||
1403 | case 'F': | |
1404 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1405 | val = -val; | |
1406 | val &= 0x1f; | |
1407 | output_addr_const (file, GEN_INT (val)); | |
1408 | return; | |
1409 | ||
1410 | case 'G': | |
1411 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1412 | val = -val; | |
1413 | val &= 0x3f; | |
1414 | output_addr_const (file, GEN_INT (val)); | |
1415 | return; | |
1416 | ||
1417 | case 'H': | |
1418 | val = xcode == CONST_INT ? INTVAL (x) : INTVAL (CONST_VECTOR_ELT (x, 0)); | |
1419 | val = -(val & -8ll); | |
1420 | val = (val >> 3) & 0x1f; | |
1421 | output_addr_const (file, GEN_INT (val)); | |
1422 | return; | |
1423 | ||
56c7bfc2 | 1424 | case 'v': |
1425 | case 'w': | |
1426 | constant_to_array (mode, x, arr); | |
1427 | val = (((arr[0] << 1) + (arr[1] >> 7)) & 0xff) - 127; | |
1428 | output_addr_const (file, GEN_INT (code == 'w' ? -val : val)); | |
1429 | return; | |
1430 | ||
644459d0 | 1431 | case 0: |
1432 | if (xcode == REG) | |
1433 | fprintf (file, "%s", reg_names[REGNO (x)]); | |
1434 | else if (xcode == MEM) | |
3c047fe9 | 1435 | output_address (GET_MODE (x), XEXP (x, 0)); |
644459d0 | 1436 | else if (xcode == CONST_VECTOR) |
dea01258 | 1437 | print_operand (file, CONST_VECTOR_ELT (x, 0), 0); |
644459d0 | 1438 | else |
1439 | output_addr_const (file, x); | |
1440 | return; | |
1441 | ||
f6a0d06f | 1442 | /* unused letters |
56c7bfc2 | 1443 | o qr u yz |
5df189be | 1444 | AB OPQR UVWXYZ */ |
644459d0 | 1445 | default: |
1446 | output_operand_lossage ("invalid %%xn code"); | |
1447 | } | |
1448 | gcc_unreachable (); | |
1449 | } | |
1450 | ||
644459d0 | 1451 | /* For PIC mode we've reserved PIC_OFFSET_TABLE_REGNUM, which is a |
1452 | caller saved register. For leaf functions it is more efficient to | |
1453 | use a volatile register because we won't need to save and restore the | |
1454 | pic register. This routine is only valid after register allocation | |
1455 | is completed, so we can pick an unused register. */ | |
1456 | static rtx | |
1457 | get_pic_reg (void) | |
1458 | { | |
644459d0 | 1459 | if (!reload_completed && !reload_in_progress) |
1460 | abort (); | |
5eb28709 | 1461 | |
1462 | /* If we've already made the decision, we need to keep with it. Once we've | |
1463 | decided to use LAST_ARG_REGNUM, future calls to df_regs_ever_live_p may | |
1464 | return true since the register is now live; this should not cause us to | |
1465 | "switch back" to using pic_offset_table_rtx. */ | |
1466 | if (!cfun->machine->pic_reg) | |
1467 | { | |
d5bf7b64 | 1468 | if (crtl->is_leaf && !df_regs_ever_live_p (LAST_ARG_REGNUM)) |
5eb28709 | 1469 | cfun->machine->pic_reg = gen_rtx_REG (SImode, LAST_ARG_REGNUM); |
1470 | else | |
1471 | cfun->machine->pic_reg = pic_offset_table_rtx; | |
1472 | } | |
1473 | ||
1474 | return cfun->machine->pic_reg; | |
644459d0 | 1475 | } |
1476 | ||
5df189be | 1477 | /* Split constant addresses to handle cases that are too large. |
1478 | Add in the pic register when in PIC mode. | |
1479 | Split immediates that require more than 1 instruction. */ | |
dea01258 | 1480 | int |
1481 | spu_split_immediate (rtx * ops) | |
c8befdb9 | 1482 | { |
3754d046 | 1483 | machine_mode mode = GET_MODE (ops[0]); |
dea01258 | 1484 | enum immediate_class c = classify_immediate (ops[1], mode); |
1485 | ||
1486 | switch (c) | |
c8befdb9 | 1487 | { |
dea01258 | 1488 | case IC_IL2: |
1489 | { | |
1490 | unsigned char arrhi[16]; | |
1491 | unsigned char arrlo[16]; | |
98bbec1e | 1492 | rtx to, temp, hi, lo; |
dea01258 | 1493 | int i; |
3754d046 | 1494 | machine_mode imode = mode; |
98bbec1e | 1495 | /* We need to do reals as ints because the constant used in the |
1496 | IOR might not be a legitimate real constant. */ | |
1497 | imode = int_mode_for_mode (mode); | |
dea01258 | 1498 | constant_to_array (mode, ops[1], arrhi); |
98bbec1e | 1499 | if (imode != mode) |
1500 | to = simplify_gen_subreg (imode, ops[0], mode, 0); | |
1501 | else | |
1502 | to = ops[0]; | |
1503 | temp = !can_create_pseudo_p () ? to : gen_reg_rtx (imode); | |
dea01258 | 1504 | for (i = 0; i < 16; i += 4) |
1505 | { | |
1506 | arrlo[i + 2] = arrhi[i + 2]; | |
1507 | arrlo[i + 3] = arrhi[i + 3]; | |
1508 | arrlo[i + 0] = arrlo[i + 1] = 0; | |
1509 | arrhi[i + 2] = arrhi[i + 3] = 0; | |
1510 | } | |
98bbec1e | 1511 | hi = array_to_constant (imode, arrhi); |
1512 | lo = array_to_constant (imode, arrlo); | |
1513 | emit_move_insn (temp, hi); | |
d1f9b275 | 1514 | emit_insn (gen_rtx_SET (to, gen_rtx_IOR (imode, temp, lo))); |
dea01258 | 1515 | return 1; |
1516 | } | |
5df189be | 1517 | case IC_FSMBI2: |
1518 | { | |
1519 | unsigned char arr_fsmbi[16]; | |
1520 | unsigned char arr_andbi[16]; | |
1521 | rtx to, reg_fsmbi, reg_and; | |
1522 | int i; | |
3754d046 | 1523 | machine_mode imode = mode; |
5df189be | 1524 | /* We need to do reals as ints because the constant used in the |
1525 | * AND might not be a legitimate real constant. */ | |
1526 | imode = int_mode_for_mode (mode); | |
1527 | constant_to_array (mode, ops[1], arr_fsmbi); | |
1528 | if (imode != mode) | |
1529 | to = simplify_gen_subreg(imode, ops[0], GET_MODE (ops[0]), 0); | |
1530 | else | |
1531 | to = ops[0]; | |
1532 | for (i = 0; i < 16; i++) | |
1533 | if (arr_fsmbi[i] != 0) | |
1534 | { | |
1535 | arr_andbi[0] = arr_fsmbi[i]; | |
1536 | arr_fsmbi[i] = 0xff; | |
1537 | } | |
1538 | for (i = 1; i < 16; i++) | |
1539 | arr_andbi[i] = arr_andbi[0]; | |
1540 | reg_fsmbi = array_to_constant (imode, arr_fsmbi); | |
1541 | reg_and = array_to_constant (imode, arr_andbi); | |
1542 | emit_move_insn (to, reg_fsmbi); | |
d1f9b275 | 1543 | emit_insn (gen_rtx_SET (to, gen_rtx_AND (imode, to, reg_and))); |
5df189be | 1544 | return 1; |
1545 | } | |
dea01258 | 1546 | case IC_POOL: |
1547 | if (reload_in_progress || reload_completed) | |
1548 | { | |
1549 | rtx mem = force_const_mem (mode, ops[1]); | |
1550 | if (TARGET_LARGE_MEM) | |
1551 | { | |
1552 | rtx addr = gen_rtx_REG (Pmode, REGNO (ops[0])); | |
1553 | emit_move_insn (addr, XEXP (mem, 0)); | |
1554 | mem = replace_equiv_address (mem, addr); | |
1555 | } | |
1556 | emit_move_insn (ops[0], mem); | |
1557 | return 1; | |
1558 | } | |
1559 | break; | |
1560 | case IC_IL1s: | |
1561 | case IC_IL2s: | |
1562 | if (reload_completed && GET_CODE (ops[1]) != HIGH) | |
1563 | { | |
1564 | if (c == IC_IL2s) | |
1565 | { | |
5df189be | 1566 | emit_move_insn (ops[0], gen_rtx_HIGH (mode, ops[1])); |
1567 | emit_move_insn (ops[0], gen_rtx_LO_SUM (mode, ops[0], ops[1])); | |
dea01258 | 1568 | } |
1569 | else if (flag_pic) | |
1570 | emit_insn (gen_pic (ops[0], ops[1])); | |
1571 | if (flag_pic) | |
1572 | { | |
1573 | rtx pic_reg = get_pic_reg (); | |
1574 | emit_insn (gen_addsi3 (ops[0], ops[0], pic_reg)); | |
dea01258 | 1575 | } |
1576 | return flag_pic || c == IC_IL2s; | |
1577 | } | |
1578 | break; | |
1579 | case IC_IL1: | |
1580 | case IC_FSMBI: | |
1581 | case IC_CPAT: | |
1582 | break; | |
c8befdb9 | 1583 | } |
dea01258 | 1584 | return 0; |
c8befdb9 | 1585 | } |
1586 | ||
644459d0 | 1587 | /* SAVING is TRUE when we are generating the actual load and store |
1588 | instructions for REGNO. When determining the size of the stack | |
1589 | needed for saving register we must allocate enough space for the | |
1590 | worst case, because we don't always have the information early enough | |
1591 | to not allocate it. But we can at least eliminate the actual loads | |
1592 | and stores during the prologue/epilogue. */ | |
1593 | static int | |
1594 | need_to_save_reg (int regno, int saving) | |
1595 | { | |
3072d30e | 1596 | if (df_regs_ever_live_p (regno) && !call_used_regs[regno]) |
644459d0 | 1597 | return 1; |
1598 | if (flag_pic | |
1599 | && regno == PIC_OFFSET_TABLE_REGNUM | |
5eb28709 | 1600 | && (!saving || cfun->machine->pic_reg == pic_offset_table_rtx)) |
644459d0 | 1601 | return 1; |
1602 | return 0; | |
1603 | } | |
1604 | ||
1605 | /* This function is only correct starting with local register | |
1606 | allocation */ | |
1607 | int | |
1608 | spu_saved_regs_size (void) | |
1609 | { | |
1610 | int reg_save_size = 0; | |
1611 | int regno; | |
1612 | ||
1613 | for (regno = FIRST_PSEUDO_REGISTER - 1; regno >= 0; --regno) | |
1614 | if (need_to_save_reg (regno, 0)) | |
1615 | reg_save_size += 0x10; | |
1616 | return reg_save_size; | |
1617 | } | |
1618 | ||
0af56f80 | 1619 | static rtx_insn * |
644459d0 | 1620 | frame_emit_store (int regno, rtx addr, HOST_WIDE_INT offset) |
1621 | { | |
1622 | rtx reg = gen_rtx_REG (V4SImode, regno); | |
1623 | rtx mem = | |
1624 | gen_frame_mem (V4SImode, gen_rtx_PLUS (Pmode, addr, GEN_INT (offset))); | |
1625 | return emit_insn (gen_movv4si (mem, reg)); | |
1626 | } | |
1627 | ||
0af56f80 | 1628 | static rtx_insn * |
644459d0 | 1629 | frame_emit_load (int regno, rtx addr, HOST_WIDE_INT offset) |
1630 | { | |
1631 | rtx reg = gen_rtx_REG (V4SImode, regno); | |
1632 | rtx mem = | |
1633 | gen_frame_mem (V4SImode, gen_rtx_PLUS (Pmode, addr, GEN_INT (offset))); | |
1634 | return emit_insn (gen_movv4si (reg, mem)); | |
1635 | } | |
1636 | ||
1637 | /* This happens after reload, so we need to expand it. */ | |
0af56f80 | 1638 | static rtx_insn * |
644459d0 | 1639 | frame_emit_add_imm (rtx dst, rtx src, HOST_WIDE_INT imm, rtx scratch) |
1640 | { | |
0af56f80 | 1641 | rtx_insn *insn; |
644459d0 | 1642 | if (satisfies_constraint_K (GEN_INT (imm))) |
1643 | { | |
1644 | insn = emit_insn (gen_addsi3 (dst, src, GEN_INT (imm))); | |
1645 | } | |
1646 | else | |
1647 | { | |
3072d30e | 1648 | emit_insn (gen_movsi (scratch, gen_int_mode (imm, SImode))); |
644459d0 | 1649 | insn = emit_insn (gen_addsi3 (dst, src, scratch)); |
1650 | if (REGNO (src) == REGNO (scratch)) | |
1651 | abort (); | |
1652 | } | |
644459d0 | 1653 | return insn; |
1654 | } | |
1655 | ||
1656 | /* Return nonzero if this function is known to have a null epilogue. */ | |
1657 | ||
1658 | int | |
1659 | direct_return (void) | |
1660 | { | |
1661 | if (reload_completed) | |
1662 | { | |
1663 | if (cfun->static_chain_decl == 0 | |
1664 | && (spu_saved_regs_size () | |
1665 | + get_frame_size () | |
abe32cce | 1666 | + crtl->outgoing_args_size |
1667 | + crtl->args.pretend_args_size == 0) | |
d5bf7b64 | 1668 | && crtl->is_leaf) |
644459d0 | 1669 | return 1; |
1670 | } | |
1671 | return 0; | |
1672 | } | |
1673 | ||
1674 | /* | |
1675 | The stack frame looks like this: | |
1676 | +-------------+ | |
1677 | | incoming | | |
a8e019fa | 1678 | | args | |
1679 | AP -> +-------------+ | |
644459d0 | 1680 | | $lr save | |
1681 | +-------------+ | |
1682 | prev SP | back chain | | |
1683 | +-------------+ | |
1684 | | var args | | |
abe32cce | 1685 | | reg save | crtl->args.pretend_args_size bytes |
644459d0 | 1686 | +-------------+ |
1687 | | ... | | |
1688 | | saved regs | spu_saved_regs_size() bytes | |
a8e019fa | 1689 | FP -> +-------------+ |
644459d0 | 1690 | | ... | |
a8e019fa | 1691 | | vars | get_frame_size() bytes |
1692 | HFP -> +-------------+ | |
644459d0 | 1693 | | ... | |
1694 | | outgoing | | |
abe32cce | 1695 | | args | crtl->outgoing_args_size bytes |
644459d0 | 1696 | +-------------+ |
1697 | | $lr of next | | |
1698 | | frame | | |
1699 | +-------------+ | |
a8e019fa | 1700 | | back chain | |
1701 | SP -> +-------------+ | |
644459d0 | 1702 | |
1703 | */ | |
1704 | void | |
1705 | spu_expand_prologue (void) | |
1706 | { | |
1707 | HOST_WIDE_INT size = get_frame_size (), offset, regno; | |
1708 | HOST_WIDE_INT total_size; | |
1709 | HOST_WIDE_INT saved_regs_size; | |
1710 | rtx sp_reg = gen_rtx_REG (Pmode, STACK_POINTER_REGNUM); | |
1711 | rtx scratch_reg_0, scratch_reg_1; | |
0af56f80 | 1712 | rtx_insn *insn; |
1713 | rtx real; | |
644459d0 | 1714 | |
5eb28709 | 1715 | if (flag_pic && optimize == 0 && !cfun->machine->pic_reg) |
1716 | cfun->machine->pic_reg = pic_offset_table_rtx; | |
644459d0 | 1717 | |
1718 | if (spu_naked_function_p (current_function_decl)) | |
1719 | return; | |
1720 | ||
1721 | scratch_reg_0 = gen_rtx_REG (SImode, LAST_ARG_REGNUM + 1); | |
1722 | scratch_reg_1 = gen_rtx_REG (SImode, LAST_ARG_REGNUM + 2); | |
1723 | ||
1724 | saved_regs_size = spu_saved_regs_size (); | |
1725 | total_size = size + saved_regs_size | |
abe32cce | 1726 | + crtl->outgoing_args_size |
1727 | + crtl->args.pretend_args_size; | |
644459d0 | 1728 | |
d5bf7b64 | 1729 | if (!crtl->is_leaf |
18d50ae6 | 1730 | || cfun->calls_alloca || total_size > 0) |
644459d0 | 1731 | total_size += STACK_POINTER_OFFSET; |
1732 | ||
1733 | /* Save this first because code after this might use the link | |
1734 | register as a scratch register. */ | |
d5bf7b64 | 1735 | if (!crtl->is_leaf) |
644459d0 | 1736 | { |
1737 | insn = frame_emit_store (LINK_REGISTER_REGNUM, sp_reg, 16); | |
1738 | RTX_FRAME_RELATED_P (insn) = 1; | |
1739 | } | |
1740 | ||
1741 | if (total_size > 0) | |
1742 | { | |
abe32cce | 1743 | offset = -crtl->args.pretend_args_size; |
644459d0 | 1744 | for (regno = 0; regno < FIRST_PSEUDO_REGISTER; ++regno) |
1745 | if (need_to_save_reg (regno, 1)) | |
1746 | { | |
1747 | offset -= 16; | |
1748 | insn = frame_emit_store (regno, sp_reg, offset); | |
1749 | RTX_FRAME_RELATED_P (insn) = 1; | |
1750 | } | |
1751 | } | |
1752 | ||
5eb28709 | 1753 | if (flag_pic && cfun->machine->pic_reg) |
644459d0 | 1754 | { |
5eb28709 | 1755 | rtx pic_reg = cfun->machine->pic_reg; |
644459d0 | 1756 | insn = emit_insn (gen_load_pic_offset (pic_reg, scratch_reg_0)); |
644459d0 | 1757 | insn = emit_insn (gen_subsi3 (pic_reg, pic_reg, scratch_reg_0)); |
644459d0 | 1758 | } |
1759 | ||
1760 | if (total_size > 0) | |
1761 | { | |
1762 | if (flag_stack_check) | |
1763 | { | |
d819917f | 1764 | /* We compare against total_size-1 because |
644459d0 | 1765 | ($sp >= total_size) <=> ($sp > total_size-1) */ |
1766 | rtx scratch_v4si = gen_rtx_REG (V4SImode, REGNO (scratch_reg_0)); | |
1767 | rtx sp_v4si = gen_rtx_REG (V4SImode, STACK_POINTER_REGNUM); | |
1768 | rtx size_v4si = spu_const (V4SImode, total_size - 1); | |
1769 | if (!satisfies_constraint_K (GEN_INT (total_size - 1))) | |
1770 | { | |
1771 | emit_move_insn (scratch_v4si, size_v4si); | |
1772 | size_v4si = scratch_v4si; | |
1773 | } | |
1774 | emit_insn (gen_cgt_v4si (scratch_v4si, sp_v4si, size_v4si)); | |
1775 | emit_insn (gen_vec_extractv4si | |
1776 | (scratch_reg_0, scratch_v4si, GEN_INT (1))); | |
1777 | emit_insn (gen_spu_heq (scratch_reg_0, GEN_INT (0))); | |
1778 | } | |
1779 | ||
1780 | /* Adjust the stack pointer, and make sure scratch_reg_0 contains | |
1781 | the value of the previous $sp because we save it as the back | |
1782 | chain. */ | |
1783 | if (total_size <= 2000) | |
1784 | { | |
1785 | /* In this case we save the back chain first. */ | |
1786 | insn = frame_emit_store (STACK_POINTER_REGNUM, sp_reg, -total_size); | |
644459d0 | 1787 | insn = |
1788 | frame_emit_add_imm (sp_reg, sp_reg, -total_size, scratch_reg_0); | |
1789 | } | |
644459d0 | 1790 | else |
1791 | { | |
1792 | insn = emit_move_insn (scratch_reg_0, sp_reg); | |
644459d0 | 1793 | insn = |
1794 | frame_emit_add_imm (sp_reg, sp_reg, -total_size, scratch_reg_1); | |
1795 | } | |
1796 | RTX_FRAME_RELATED_P (insn) = 1; | |
1797 | real = gen_addsi3 (sp_reg, sp_reg, GEN_INT (-total_size)); | |
b9c74b4d | 1798 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, real); |
644459d0 | 1799 | |
1800 | if (total_size > 2000) | |
1801 | { | |
1802 | /* Save the back chain ptr */ | |
1803 | insn = frame_emit_store (REGNO (scratch_reg_0), sp_reg, 0); | |
644459d0 | 1804 | } |
1805 | ||
1806 | if (frame_pointer_needed) | |
1807 | { | |
1808 | rtx fp_reg = gen_rtx_REG (Pmode, HARD_FRAME_POINTER_REGNUM); | |
1809 | HOST_WIDE_INT fp_offset = STACK_POINTER_OFFSET | |
abe32cce | 1810 | + crtl->outgoing_args_size; |
644459d0 | 1811 | /* Set the new frame_pointer */ |
d8dfeb55 | 1812 | insn = frame_emit_add_imm (fp_reg, sp_reg, fp_offset, scratch_reg_0); |
1813 | RTX_FRAME_RELATED_P (insn) = 1; | |
1814 | real = gen_addsi3 (fp_reg, sp_reg, GEN_INT (fp_offset)); | |
b9c74b4d | 1815 | add_reg_note (insn, REG_FRAME_RELATED_EXPR, real); |
5df189be | 1816 | REGNO_POINTER_ALIGN (HARD_FRAME_POINTER_REGNUM) = STACK_BOUNDARY; |
644459d0 | 1817 | } |
1818 | } | |
1819 | ||
8c0dd614 | 1820 | if (flag_stack_usage_info) |
a512540d | 1821 | current_function_static_stack_size = total_size; |
644459d0 | 1822 | } |
1823 | ||
1824 | void | |
1825 | spu_expand_epilogue (bool sibcall_p) | |
1826 | { | |
1827 | int size = get_frame_size (), offset, regno; | |
1828 | HOST_WIDE_INT saved_regs_size, total_size; | |
1829 | rtx sp_reg = gen_rtx_REG (Pmode, STACK_POINTER_REGNUM); | |
431ad7e0 | 1830 | rtx scratch_reg_0; |
644459d0 | 1831 | |
644459d0 | 1832 | if (spu_naked_function_p (current_function_decl)) |
1833 | return; | |
1834 | ||
1835 | scratch_reg_0 = gen_rtx_REG (SImode, LAST_ARG_REGNUM + 1); | |
1836 | ||
1837 | saved_regs_size = spu_saved_regs_size (); | |
1838 | total_size = size + saved_regs_size | |
abe32cce | 1839 | + crtl->outgoing_args_size |
1840 | + crtl->args.pretend_args_size; | |
644459d0 | 1841 | |
d5bf7b64 | 1842 | if (!crtl->is_leaf |
18d50ae6 | 1843 | || cfun->calls_alloca || total_size > 0) |
644459d0 | 1844 | total_size += STACK_POINTER_OFFSET; |
1845 | ||
1846 | if (total_size > 0) | |
1847 | { | |
18d50ae6 | 1848 | if (cfun->calls_alloca) |
644459d0 | 1849 | frame_emit_load (STACK_POINTER_REGNUM, sp_reg, 0); |
1850 | else | |
1851 | frame_emit_add_imm (sp_reg, sp_reg, total_size, scratch_reg_0); | |
1852 | ||
1853 | ||
1854 | if (saved_regs_size > 0) | |
1855 | { | |
abe32cce | 1856 | offset = -crtl->args.pretend_args_size; |
644459d0 | 1857 | for (regno = 0; regno < FIRST_PSEUDO_REGISTER; ++regno) |
1858 | if (need_to_save_reg (regno, 1)) | |
1859 | { | |
1860 | offset -= 0x10; | |
1861 | frame_emit_load (regno, sp_reg, offset); | |
1862 | } | |
1863 | } | |
1864 | } | |
1865 | ||
d5bf7b64 | 1866 | if (!crtl->is_leaf) |
644459d0 | 1867 | frame_emit_load (LINK_REGISTER_REGNUM, sp_reg, 16); |
1868 | ||
1869 | if (!sibcall_p) | |
1870 | { | |
18b42941 | 1871 | emit_use (gen_rtx_REG (SImode, LINK_REGISTER_REGNUM)); |
431ad7e0 | 1872 | emit_jump_insn (gen__return ()); |
644459d0 | 1873 | } |
644459d0 | 1874 | } |
1875 | ||
1876 | rtx | |
1877 | spu_return_addr (int count, rtx frame ATTRIBUTE_UNUSED) | |
1878 | { | |
1879 | if (count != 0) | |
1880 | return 0; | |
1881 | /* This is inefficient because it ends up copying to a save-register | |
1882 | which then gets saved even though $lr has already been saved. But | |
1883 | it does generate better code for leaf functions and we don't need | |
1884 | to use RETURN_ADDRESS_POINTER_REGNUM to get it working. It's only | |
1885 | used for __builtin_return_address anyway, so maybe we don't care if | |
1886 | it's inefficient. */ | |
1887 | return get_hard_reg_initial_val (Pmode, LINK_REGISTER_REGNUM); | |
1888 | } | |
1889 | \f | |
1890 | ||
1891 | /* Given VAL, generate a constant appropriate for MODE. | |
1892 | If MODE is a vector mode, every element will be VAL. | |
1893 | For TImode, VAL will be zero extended to 128 bits. */ | |
1894 | rtx | |
3754d046 | 1895 | spu_const (machine_mode mode, HOST_WIDE_INT val) |
644459d0 | 1896 | { |
1897 | rtx inner; | |
1898 | rtvec v; | |
1899 | int units, i; | |
1900 | ||
1901 | gcc_assert (GET_MODE_CLASS (mode) == MODE_INT | |
1902 | || GET_MODE_CLASS (mode) == MODE_FLOAT | |
1903 | || GET_MODE_CLASS (mode) == MODE_VECTOR_INT | |
1904 | || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT); | |
1905 | ||
1906 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
1907 | return immed_double_const (val, 0, mode); | |
1908 | ||
1909 | /* val is the bit representation of the float */ | |
1910 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
1911 | return hwint_to_const_double (mode, val); | |
1912 | ||
1913 | if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT) | |
1914 | inner = immed_double_const (val, 0, GET_MODE_INNER (mode)); | |
1915 | else | |
1916 | inner = hwint_to_const_double (GET_MODE_INNER (mode), val); | |
1917 | ||
1918 | units = GET_MODE_NUNITS (mode); | |
1919 | ||
1920 | v = rtvec_alloc (units); | |
1921 | ||
1922 | for (i = 0; i < units; ++i) | |
1923 | RTVEC_ELT (v, i) = inner; | |
1924 | ||
1925 | return gen_rtx_CONST_VECTOR (mode, v); | |
1926 | } | |
644459d0 | 1927 | |
5474166e | 1928 | /* Create a MODE vector constant from 4 ints. */ |
1929 | rtx | |
3754d046 | 1930 | spu_const_from_ints(machine_mode mode, int a, int b, int c, int d) |
5474166e | 1931 | { |
1932 | unsigned char arr[16]; | |
1933 | arr[0] = (a >> 24) & 0xff; | |
1934 | arr[1] = (a >> 16) & 0xff; | |
1935 | arr[2] = (a >> 8) & 0xff; | |
1936 | arr[3] = (a >> 0) & 0xff; | |
1937 | arr[4] = (b >> 24) & 0xff; | |
1938 | arr[5] = (b >> 16) & 0xff; | |
1939 | arr[6] = (b >> 8) & 0xff; | |
1940 | arr[7] = (b >> 0) & 0xff; | |
1941 | arr[8] = (c >> 24) & 0xff; | |
1942 | arr[9] = (c >> 16) & 0xff; | |
1943 | arr[10] = (c >> 8) & 0xff; | |
1944 | arr[11] = (c >> 0) & 0xff; | |
1945 | arr[12] = (d >> 24) & 0xff; | |
1946 | arr[13] = (d >> 16) & 0xff; | |
1947 | arr[14] = (d >> 8) & 0xff; | |
1948 | arr[15] = (d >> 0) & 0xff; | |
1949 | return array_to_constant(mode, arr); | |
1950 | } | |
5a976006 | 1951 | \f |
1952 | /* branch hint stuff */ | |
5474166e | 1953 | |
644459d0 | 1954 | /* An array of these is used to propagate hints to predecessor blocks. */ |
1955 | struct spu_bb_info | |
1956 | { | |
0af56f80 | 1957 | rtx_insn *prop_jump; /* propagated from another block */ |
5a976006 | 1958 | int bb_index; /* the original block. */ |
644459d0 | 1959 | }; |
5a976006 | 1960 | static struct spu_bb_info *spu_bb_info; |
644459d0 | 1961 | |
5a976006 | 1962 | #define STOP_HINT_P(INSN) \ |
aa90bb35 | 1963 | (CALL_P(INSN) \ |
5a976006 | 1964 | || INSN_CODE(INSN) == CODE_FOR_divmodsi4 \ |
1965 | || INSN_CODE(INSN) == CODE_FOR_udivmodsi4) | |
1966 | ||
1967 | /* 1 when RTX is a hinted branch or its target. We keep track of | |
1968 | what has been hinted so the safe-hint code can test it easily. */ | |
1969 | #define HINTED_P(RTX) \ | |
1970 | (RTL_FLAG_CHECK3("HINTED_P", (RTX), CODE_LABEL, JUMP_INSN, CALL_INSN)->unchanging) | |
1971 | ||
1972 | /* 1 when RTX is an insn that must be scheduled on an even boundary. */ | |
1973 | #define SCHED_ON_EVEN_P(RTX) \ | |
1974 | (RTL_FLAG_CHECK2("SCHED_ON_EVEN_P", (RTX), JUMP_INSN, CALL_INSN)->in_struct) | |
1975 | ||
1976 | /* Emit a nop for INSN such that the two will dual issue. This assumes | |
1977 | INSN is 8-byte aligned. When INSN is inline asm we emit an lnop. | |
1978 | We check for TImode to handle a MULTI1 insn which has dual issued its | |
b1135d9a | 1979 | first instruction. get_pipe returns -1 for MULTI0 or inline asm. */ |
5a976006 | 1980 | static void |
0af56f80 | 1981 | emit_nop_for_insn (rtx_insn *insn) |
644459d0 | 1982 | { |
5a976006 | 1983 | int p; |
0af56f80 | 1984 | rtx_insn *new_insn; |
b1135d9a | 1985 | |
1986 | /* We need to handle JUMP_TABLE_DATA separately. */ | |
1987 | if (JUMP_TABLE_DATA_P (insn)) | |
1988 | { | |
1989 | new_insn = emit_insn_after (gen_lnop(), insn); | |
1990 | recog_memoized (new_insn); | |
1991 | INSN_LOCATION (new_insn) = UNKNOWN_LOCATION; | |
1992 | return; | |
1993 | } | |
1994 | ||
5a976006 | 1995 | p = get_pipe (insn); |
1996 | if ((CALL_P (insn) || JUMP_P (insn)) && SCHED_ON_EVEN_P (insn)) | |
1997 | new_insn = emit_insn_after (gen_lnop (), insn); | |
1998 | else if (p == 1 && GET_MODE (insn) == TImode) | |
644459d0 | 1999 | { |
5a976006 | 2000 | new_insn = emit_insn_before (gen_nopn (GEN_INT (127)), insn); |
2001 | PUT_MODE (new_insn, TImode); | |
2002 | PUT_MODE (insn, VOIDmode); | |
2003 | } | |
2004 | else | |
2005 | new_insn = emit_insn_after (gen_lnop (), insn); | |
2006 | recog_memoized (new_insn); | |
d53c050c | 2007 | INSN_LOCATION (new_insn) = INSN_LOCATION (insn); |
5a976006 | 2008 | } |
2009 | ||
2010 | /* Insert nops in basic blocks to meet dual issue alignment | |
2011 | requirements. Also make sure hbrp and hint instructions are at least | |
2012 | one cycle apart, possibly inserting a nop. */ | |
2013 | static void | |
2014 | pad_bb(void) | |
2015 | { | |
0af56f80 | 2016 | rtx_insn *insn, *next_insn, *prev_insn, *hbr_insn = 0; |
5a976006 | 2017 | int length; |
2018 | int addr; | |
2019 | ||
2020 | /* This sets up INSN_ADDRESSES. */ | |
2021 | shorten_branches (get_insns ()); | |
2022 | ||
2023 | /* Keep track of length added by nops. */ | |
2024 | length = 0; | |
2025 | ||
2026 | prev_insn = 0; | |
2027 | insn = get_insns (); | |
2028 | if (!active_insn_p (insn)) | |
2029 | insn = next_active_insn (insn); | |
2030 | for (; insn; insn = next_insn) | |
2031 | { | |
2032 | next_insn = next_active_insn (insn); | |
2033 | if (INSN_CODE (insn) == CODE_FOR_iprefetch | |
2034 | || INSN_CODE (insn) == CODE_FOR_hbr) | |
644459d0 | 2035 | { |
5a976006 | 2036 | if (hbr_insn) |
2037 | { | |
2038 | int a0 = INSN_ADDRESSES (INSN_UID (hbr_insn)); | |
2039 | int a1 = INSN_ADDRESSES (INSN_UID (insn)); | |
2040 | if ((a1 - a0 == 8 && GET_MODE (insn) != TImode) | |
2041 | || (a1 - a0 == 4)) | |
2042 | { | |
2043 | prev_insn = emit_insn_before (gen_lnop (), insn); | |
2044 | PUT_MODE (prev_insn, GET_MODE (insn)); | |
2045 | PUT_MODE (insn, TImode); | |
d53c050c | 2046 | INSN_LOCATION (prev_insn) = INSN_LOCATION (insn); |
5a976006 | 2047 | length += 4; |
2048 | } | |
2049 | } | |
2050 | hbr_insn = insn; | |
2051 | } | |
4f8e39e2 | 2052 | if (INSN_CODE (insn) == CODE_FOR_blockage && next_insn) |
5a976006 | 2053 | { |
2054 | if (GET_MODE (insn) == TImode) | |
2055 | PUT_MODE (next_insn, TImode); | |
2056 | insn = next_insn; | |
2057 | next_insn = next_active_insn (insn); | |
2058 | } | |
2059 | addr = INSN_ADDRESSES (INSN_UID (insn)); | |
2060 | if ((CALL_P (insn) || JUMP_P (insn)) && SCHED_ON_EVEN_P (insn)) | |
2061 | { | |
2062 | if (((addr + length) & 7) != 0) | |
2063 | { | |
2064 | emit_nop_for_insn (prev_insn); | |
2065 | length += 4; | |
2066 | } | |
644459d0 | 2067 | } |
5a976006 | 2068 | else if (GET_MODE (insn) == TImode |
2069 | && ((next_insn && GET_MODE (next_insn) != TImode) | |
2070 | || get_attr_type (insn) == TYPE_MULTI0) | |
2071 | && ((addr + length) & 7) != 0) | |
2072 | { | |
2073 | /* prev_insn will always be set because the first insn is | |
2074 | always 8-byte aligned. */ | |
2075 | emit_nop_for_insn (prev_insn); | |
2076 | length += 4; | |
2077 | } | |
2078 | prev_insn = insn; | |
644459d0 | 2079 | } |
644459d0 | 2080 | } |
2081 | ||
5a976006 | 2082 | \f |
2083 | /* Routines for branch hints. */ | |
2084 | ||
644459d0 | 2085 | static void |
0af56f80 | 2086 | spu_emit_branch_hint (rtx_insn *before, rtx_insn *branch, rtx target, |
5a976006 | 2087 | int distance, sbitmap blocks) |
644459d0 | 2088 | { |
5a976006 | 2089 | rtx branch_label = 0; |
0af56f80 | 2090 | rtx_insn *hint; |
2091 | rtx_insn *insn; | |
c86d86ff | 2092 | rtx_jump_table_data *table; |
644459d0 | 2093 | |
2094 | if (before == 0 || branch == 0 || target == 0) | |
2095 | return; | |
2096 | ||
5a976006 | 2097 | /* While scheduling we require hints to be no further than 600, so |
2098 | we need to enforce that here too */ | |
644459d0 | 2099 | if (distance > 600) |
2100 | return; | |
2101 | ||
5a976006 | 2102 | /* If we have a Basic block note, emit it after the basic block note. */ |
37534923 | 2103 | if (NOTE_INSN_BASIC_BLOCK_P (before)) |
5a976006 | 2104 | before = NEXT_INSN (before); |
644459d0 | 2105 | |
2106 | branch_label = gen_label_rtx (); | |
2107 | LABEL_NUSES (branch_label)++; | |
2108 | LABEL_PRESERVE_P (branch_label) = 1; | |
2109 | insn = emit_label_before (branch_label, branch); | |
2110 | branch_label = gen_rtx_LABEL_REF (VOIDmode, branch_label); | |
08b7917c | 2111 | bitmap_set_bit (blocks, BLOCK_FOR_INSN (branch)->index); |
5a976006 | 2112 | |
2113 | hint = emit_insn_before (gen_hbr (branch_label, target), before); | |
2114 | recog_memoized (hint); | |
d53c050c | 2115 | INSN_LOCATION (hint) = INSN_LOCATION (branch); |
5a976006 | 2116 | HINTED_P (branch) = 1; |
644459d0 | 2117 | |
5a976006 | 2118 | if (GET_CODE (target) == LABEL_REF) |
2119 | HINTED_P (XEXP (target, 0)) = 1; | |
2120 | else if (tablejump_p (branch, 0, &table)) | |
644459d0 | 2121 | { |
5a976006 | 2122 | rtvec vec; |
2123 | int j; | |
2124 | if (GET_CODE (PATTERN (table)) == ADDR_VEC) | |
2125 | vec = XVEC (PATTERN (table), 0); | |
2126 | else | |
2127 | vec = XVEC (PATTERN (table), 1); | |
2128 | for (j = GET_NUM_ELEM (vec) - 1; j >= 0; --j) | |
2129 | HINTED_P (XEXP (RTVEC_ELT (vec, j), 0)) = 1; | |
644459d0 | 2130 | } |
5a976006 | 2131 | |
2132 | if (distance >= 588) | |
644459d0 | 2133 | { |
5a976006 | 2134 | /* Make sure the hint isn't scheduled any earlier than this point, |
2135 | which could make it too far for the branch offest to fit */ | |
2fbdf9ef | 2136 | insn = emit_insn_before (gen_blockage (), hint); |
2137 | recog_memoized (insn); | |
d53c050c | 2138 | INSN_LOCATION (insn) = INSN_LOCATION (hint); |
5a976006 | 2139 | } |
2140 | else if (distance <= 8 * 4) | |
2141 | { | |
2142 | /* To guarantee at least 8 insns between the hint and branch we | |
2143 | insert nops. */ | |
2144 | int d; | |
2145 | for (d = distance; d < 8 * 4; d += 4) | |
2146 | { | |
2147 | insn = | |
2148 | emit_insn_after (gen_nopn_nv (gen_rtx_REG (SImode, 127)), hint); | |
2149 | recog_memoized (insn); | |
d53c050c | 2150 | INSN_LOCATION (insn) = INSN_LOCATION (hint); |
5a976006 | 2151 | } |
2152 | ||
2153 | /* Make sure any nops inserted aren't scheduled before the hint. */ | |
2fbdf9ef | 2154 | insn = emit_insn_after (gen_blockage (), hint); |
2155 | recog_memoized (insn); | |
d53c050c | 2156 | INSN_LOCATION (insn) = INSN_LOCATION (hint); |
5a976006 | 2157 | |
2158 | /* Make sure any nops inserted aren't scheduled after the call. */ | |
2159 | if (CALL_P (branch) && distance < 8 * 4) | |
2fbdf9ef | 2160 | { |
2161 | insn = emit_insn_before (gen_blockage (), branch); | |
2162 | recog_memoized (insn); | |
d53c050c | 2163 | INSN_LOCATION (insn) = INSN_LOCATION (branch); |
2fbdf9ef | 2164 | } |
644459d0 | 2165 | } |
644459d0 | 2166 | } |
2167 | ||
2168 | /* Returns 0 if we don't want a hint for this branch. Otherwise return | |
2169 | the rtx for the branch target. */ | |
2170 | static rtx | |
0af56f80 | 2171 | get_branch_target (rtx_insn *branch) |
644459d0 | 2172 | { |
aa90bb35 | 2173 | if (JUMP_P (branch)) |
644459d0 | 2174 | { |
2175 | rtx set, src; | |
2176 | ||
2177 | /* Return statements */ | |
2178 | if (GET_CODE (PATTERN (branch)) == RETURN) | |
2179 | return gen_rtx_REG (SImode, LINK_REGISTER_REGNUM); | |
2180 | ||
fcc31b99 | 2181 | /* ASM GOTOs. */ |
604157f6 | 2182 | if (extract_asm_operands (PATTERN (branch)) != NULL) |
fcc31b99 | 2183 | return NULL; |
2184 | ||
644459d0 | 2185 | set = single_set (branch); |
2186 | src = SET_SRC (set); | |
2187 | if (GET_CODE (SET_DEST (set)) != PC) | |
2188 | abort (); | |
2189 | ||
2190 | if (GET_CODE (src) == IF_THEN_ELSE) | |
2191 | { | |
2192 | rtx lab = 0; | |
2193 | rtx note = find_reg_note (branch, REG_BR_PROB, 0); | |
2194 | if (note) | |
2195 | { | |
2196 | /* If the more probable case is not a fall through, then | |
2197 | try a branch hint. */ | |
9eb946de | 2198 | int prob = XINT (note, 0); |
644459d0 | 2199 | if (prob > (REG_BR_PROB_BASE * 6 / 10) |
2200 | && GET_CODE (XEXP (src, 1)) != PC) | |
2201 | lab = XEXP (src, 1); | |
2202 | else if (prob < (REG_BR_PROB_BASE * 4 / 10) | |
2203 | && GET_CODE (XEXP (src, 2)) != PC) | |
2204 | lab = XEXP (src, 2); | |
2205 | } | |
2206 | if (lab) | |
2207 | { | |
2208 | if (GET_CODE (lab) == RETURN) | |
2209 | return gen_rtx_REG (SImode, LINK_REGISTER_REGNUM); | |
2210 | return lab; | |
2211 | } | |
2212 | return 0; | |
2213 | } | |
2214 | ||
2215 | return src; | |
2216 | } | |
aa90bb35 | 2217 | else if (CALL_P (branch)) |
644459d0 | 2218 | { |
2219 | rtx call; | |
2220 | /* All of our call patterns are in a PARALLEL and the CALL is | |
2221 | the first pattern in the PARALLEL. */ | |
2222 | if (GET_CODE (PATTERN (branch)) != PARALLEL) | |
2223 | abort (); | |
2224 | call = XVECEXP (PATTERN (branch), 0, 0); | |
2225 | if (GET_CODE (call) == SET) | |
2226 | call = SET_SRC (call); | |
2227 | if (GET_CODE (call) != CALL) | |
2228 | abort (); | |
2229 | return XEXP (XEXP (call, 0), 0); | |
2230 | } | |
2231 | return 0; | |
2232 | } | |
2233 | ||
5a976006 | 2234 | /* The special $hbr register is used to prevent the insn scheduler from |
2235 | moving hbr insns across instructions which invalidate them. It | |
2236 | should only be used in a clobber, and this function searches for | |
2237 | insns which clobber it. */ | |
2238 | static bool | |
0af56f80 | 2239 | insn_clobbers_hbr (rtx_insn *insn) |
5a976006 | 2240 | { |
2241 | if (INSN_P (insn) | |
2242 | && GET_CODE (PATTERN (insn)) == PARALLEL) | |
2243 | { | |
2244 | rtx parallel = PATTERN (insn); | |
2245 | rtx clobber; | |
2246 | int j; | |
2247 | for (j = XVECLEN (parallel, 0) - 1; j >= 0; j--) | |
2248 | { | |
2249 | clobber = XVECEXP (parallel, 0, j); | |
2250 | if (GET_CODE (clobber) == CLOBBER | |
2251 | && GET_CODE (XEXP (clobber, 0)) == REG | |
2252 | && REGNO (XEXP (clobber, 0)) == HBR_REGNUM) | |
2253 | return 1; | |
2254 | } | |
2255 | } | |
2256 | return 0; | |
2257 | } | |
2258 | ||
2259 | /* Search up to 32 insns starting at FIRST: | |
2260 | - at any kind of hinted branch, just return | |
2261 | - at any unconditional branch in the first 15 insns, just return | |
2262 | - at a call or indirect branch, after the first 15 insns, force it to | |
2263 | an even address and return | |
2264 | - at any unconditional branch, after the first 15 insns, force it to | |
2265 | an even address. | |
2266 | At then end of the search, insert an hbrp within 4 insns of FIRST, | |
2267 | and an hbrp within 16 instructions of FIRST. | |
2268 | */ | |
644459d0 | 2269 | static void |
0af56f80 | 2270 | insert_hbrp_for_ilb_runout (rtx_insn *first) |
644459d0 | 2271 | { |
0af56f80 | 2272 | rtx_insn *insn, *before_4 = 0, *before_16 = 0; |
5a976006 | 2273 | int addr = 0, length, first_addr = -1; |
2274 | int hbrp_addr0 = 128 * 4, hbrp_addr1 = 128 * 4; | |
2275 | int insert_lnop_after = 0; | |
2276 | for (insn = first; insn; insn = NEXT_INSN (insn)) | |
2277 | if (INSN_P (insn)) | |
2278 | { | |
2279 | if (first_addr == -1) | |
2280 | first_addr = INSN_ADDRESSES (INSN_UID (insn)); | |
2281 | addr = INSN_ADDRESSES (INSN_UID (insn)) - first_addr; | |
2282 | length = get_attr_length (insn); | |
2283 | ||
2284 | if (before_4 == 0 && addr + length >= 4 * 4) | |
2285 | before_4 = insn; | |
2286 | /* We test for 14 instructions because the first hbrp will add | |
2287 | up to 2 instructions. */ | |
2288 | if (before_16 == 0 && addr + length >= 14 * 4) | |
2289 | before_16 = insn; | |
2290 | ||
2291 | if (INSN_CODE (insn) == CODE_FOR_hbr) | |
2292 | { | |
2293 | /* Make sure an hbrp is at least 2 cycles away from a hint. | |
2294 | Insert an lnop after the hbrp when necessary. */ | |
2295 | if (before_4 == 0 && addr > 0) | |
2296 | { | |
2297 | before_4 = insn; | |
2298 | insert_lnop_after |= 1; | |
2299 | } | |
2300 | else if (before_4 && addr <= 4 * 4) | |
2301 | insert_lnop_after |= 1; | |
2302 | if (before_16 == 0 && addr > 10 * 4) | |
2303 | { | |
2304 | before_16 = insn; | |
2305 | insert_lnop_after |= 2; | |
2306 | } | |
2307 | else if (before_16 && addr <= 14 * 4) | |
2308 | insert_lnop_after |= 2; | |
2309 | } | |
644459d0 | 2310 | |
5a976006 | 2311 | if (INSN_CODE (insn) == CODE_FOR_iprefetch) |
2312 | { | |
2313 | if (addr < hbrp_addr0) | |
2314 | hbrp_addr0 = addr; | |
2315 | else if (addr < hbrp_addr1) | |
2316 | hbrp_addr1 = addr; | |
2317 | } | |
644459d0 | 2318 | |
5a976006 | 2319 | if (CALL_P (insn) || JUMP_P (insn)) |
2320 | { | |
2321 | if (HINTED_P (insn)) | |
2322 | return; | |
2323 | ||
2324 | /* Any branch after the first 15 insns should be on an even | |
2325 | address to avoid a special case branch. There might be | |
2326 | some nops and/or hbrps inserted, so we test after 10 | |
2327 | insns. */ | |
2328 | if (addr > 10 * 4) | |
2329 | SCHED_ON_EVEN_P (insn) = 1; | |
2330 | } | |
644459d0 | 2331 | |
5a976006 | 2332 | if (CALL_P (insn) || tablejump_p (insn, 0, 0)) |
2333 | return; | |
2334 | ||
2335 | ||
2336 | if (addr + length >= 32 * 4) | |
644459d0 | 2337 | { |
5a976006 | 2338 | gcc_assert (before_4 && before_16); |
2339 | if (hbrp_addr0 > 4 * 4) | |
644459d0 | 2340 | { |
5a976006 | 2341 | insn = |
2342 | emit_insn_before (gen_iprefetch (GEN_INT (1)), before_4); | |
2343 | recog_memoized (insn); | |
d53c050c | 2344 | INSN_LOCATION (insn) = INSN_LOCATION (before_4); |
5a976006 | 2345 | INSN_ADDRESSES_NEW (insn, |
2346 | INSN_ADDRESSES (INSN_UID (before_4))); | |
2347 | PUT_MODE (insn, GET_MODE (before_4)); | |
2348 | PUT_MODE (before_4, TImode); | |
2349 | if (insert_lnop_after & 1) | |
644459d0 | 2350 | { |
5a976006 | 2351 | insn = emit_insn_before (gen_lnop (), before_4); |
2352 | recog_memoized (insn); | |
d53c050c | 2353 | INSN_LOCATION (insn) = INSN_LOCATION (before_4); |
5a976006 | 2354 | INSN_ADDRESSES_NEW (insn, |
2355 | INSN_ADDRESSES (INSN_UID (before_4))); | |
2356 | PUT_MODE (insn, TImode); | |
644459d0 | 2357 | } |
644459d0 | 2358 | } |
5a976006 | 2359 | if ((hbrp_addr0 <= 4 * 4 || hbrp_addr0 > 16 * 4) |
2360 | && hbrp_addr1 > 16 * 4) | |
644459d0 | 2361 | { |
5a976006 | 2362 | insn = |
2363 | emit_insn_before (gen_iprefetch (GEN_INT (2)), before_16); | |
2364 | recog_memoized (insn); | |
d53c050c | 2365 | INSN_LOCATION (insn) = INSN_LOCATION (before_16); |
5a976006 | 2366 | INSN_ADDRESSES_NEW (insn, |
2367 | INSN_ADDRESSES (INSN_UID (before_16))); | |
2368 | PUT_MODE (insn, GET_MODE (before_16)); | |
2369 | PUT_MODE (before_16, TImode); | |
2370 | if (insert_lnop_after & 2) | |
644459d0 | 2371 | { |
5a976006 | 2372 | insn = emit_insn_before (gen_lnop (), before_16); |
2373 | recog_memoized (insn); | |
d53c050c | 2374 | INSN_LOCATION (insn) = INSN_LOCATION (before_16); |
5a976006 | 2375 | INSN_ADDRESSES_NEW (insn, |
2376 | INSN_ADDRESSES (INSN_UID | |
2377 | (before_16))); | |
2378 | PUT_MODE (insn, TImode); | |
644459d0 | 2379 | } |
2380 | } | |
5a976006 | 2381 | return; |
644459d0 | 2382 | } |
644459d0 | 2383 | } |
5a976006 | 2384 | else if (BARRIER_P (insn)) |
2385 | return; | |
644459d0 | 2386 | |
644459d0 | 2387 | } |
5a976006 | 2388 | |
2389 | /* The SPU might hang when it executes 48 inline instructions after a | |
2390 | hinted branch jumps to its hinted target. The beginning of a | |
851d9296 | 2391 | function and the return from a call might have been hinted, and |
2392 | must be handled as well. To prevent a hang we insert 2 hbrps. The | |
2393 | first should be within 6 insns of the branch target. The second | |
2394 | should be within 22 insns of the branch target. When determining | |
2395 | if hbrps are necessary, we look for only 32 inline instructions, | |
2396 | because up to 12 nops and 4 hbrps could be inserted. Similarily, | |
2397 | when inserting new hbrps, we insert them within 4 and 16 insns of | |
2398 | the target. */ | |
644459d0 | 2399 | static void |
5a976006 | 2400 | insert_hbrp (void) |
644459d0 | 2401 | { |
0af56f80 | 2402 | rtx_insn *insn; |
5a976006 | 2403 | if (TARGET_SAFE_HINTS) |
644459d0 | 2404 | { |
5a976006 | 2405 | shorten_branches (get_insns ()); |
2406 | /* Insert hbrp at beginning of function */ | |
2407 | insn = next_active_insn (get_insns ()); | |
2408 | if (insn) | |
2409 | insert_hbrp_for_ilb_runout (insn); | |
2410 | /* Insert hbrp after hinted targets. */ | |
2411 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
2412 | if ((LABEL_P (insn) && HINTED_P (insn)) || CALL_P (insn)) | |
2413 | insert_hbrp_for_ilb_runout (next_active_insn (insn)); | |
644459d0 | 2414 | } |
644459d0 | 2415 | } |
2416 | ||
5a976006 | 2417 | static int in_spu_reorg; |
2418 | ||
8a42230a | 2419 | static void |
2420 | spu_var_tracking (void) | |
2421 | { | |
2422 | if (flag_var_tracking) | |
2423 | { | |
2424 | df_analyze (); | |
2425 | timevar_push (TV_VAR_TRACKING); | |
2426 | variable_tracking_main (); | |
2427 | timevar_pop (TV_VAR_TRACKING); | |
2428 | df_finish_pass (false); | |
2429 | } | |
2430 | } | |
2431 | ||
5a976006 | 2432 | /* Insert branch hints. There are no branch optimizations after this |
2433 | pass, so it's safe to set our branch hints now. */ | |
644459d0 | 2434 | static void |
5a976006 | 2435 | spu_machine_dependent_reorg (void) |
644459d0 | 2436 | { |
5a976006 | 2437 | sbitmap blocks; |
2438 | basic_block bb; | |
0af56f80 | 2439 | rtx_insn *branch, *insn; |
5a976006 | 2440 | rtx branch_target = 0; |
2441 | int branch_addr = 0, insn_addr, required_dist = 0; | |
2442 | int i; | |
2443 | unsigned int j; | |
644459d0 | 2444 | |
5a976006 | 2445 | if (!TARGET_BRANCH_HINTS || optimize == 0) |
2446 | { | |
2447 | /* We still do it for unoptimized code because an external | |
2448 | function might have hinted a call or return. */ | |
a54ca889 | 2449 | compute_bb_for_insn (); |
5a976006 | 2450 | insert_hbrp (); |
2451 | pad_bb (); | |
8a42230a | 2452 | spu_var_tracking (); |
a54ca889 | 2453 | free_bb_for_insn (); |
5a976006 | 2454 | return; |
2455 | } | |
644459d0 | 2456 | |
fe672ac0 | 2457 | blocks = sbitmap_alloc (last_basic_block_for_fn (cfun)); |
53c5d9d4 | 2458 | bitmap_clear (blocks); |
644459d0 | 2459 | |
5a976006 | 2460 | in_spu_reorg = 1; |
2461 | compute_bb_for_insn (); | |
2462 | ||
a7a0184d | 2463 | /* (Re-)discover loops so that bb->loop_father can be used |
2464 | in the analysis below. */ | |
2465 | loop_optimizer_init (AVOID_CFG_MODIFICATIONS); | |
2466 | ||
5a976006 | 2467 | compact_blocks (); |
2468 | ||
2469 | spu_bb_info = | |
a28770e1 | 2470 | (struct spu_bb_info *) xcalloc (n_basic_blocks_for_fn (cfun), |
5a976006 | 2471 | sizeof (struct spu_bb_info)); |
2472 | ||
2473 | /* We need exact insn addresses and lengths. */ | |
2474 | shorten_branches (get_insns ()); | |
2475 | ||
a28770e1 | 2476 | for (i = n_basic_blocks_for_fn (cfun) - 1; i >= 0; i--) |
644459d0 | 2477 | { |
f5a6b05f | 2478 | bb = BASIC_BLOCK_FOR_FN (cfun, i); |
5a976006 | 2479 | branch = 0; |
2480 | if (spu_bb_info[i].prop_jump) | |
644459d0 | 2481 | { |
5a976006 | 2482 | branch = spu_bb_info[i].prop_jump; |
2483 | branch_target = get_branch_target (branch); | |
2484 | branch_addr = INSN_ADDRESSES (INSN_UID (branch)); | |
2485 | required_dist = spu_hint_dist; | |
2486 | } | |
2487 | /* Search from end of a block to beginning. In this loop, find | |
2488 | jumps which need a branch and emit them only when: | |
2489 | - it's an indirect branch and we're at the insn which sets | |
2490 | the register | |
2491 | - we're at an insn that will invalidate the hint. e.g., a | |
2492 | call, another hint insn, inline asm that clobbers $hbr, and | |
2493 | some inlined operations (divmodsi4). Don't consider jumps | |
2494 | because they are only at the end of a block and are | |
2495 | considered when we are deciding whether to propagate | |
2496 | - we're getting too far away from the branch. The hbr insns | |
2497 | only have a signed 10 bit offset | |
2498 | We go back as far as possible so the branch will be considered | |
2499 | for propagation when we get to the beginning of the block. */ | |
2500 | for (insn = BB_END (bb); insn; insn = PREV_INSN (insn)) | |
2501 | { | |
2502 | if (INSN_P (insn)) | |
2503 | { | |
2504 | insn_addr = INSN_ADDRESSES (INSN_UID (insn)); | |
2505 | if (branch | |
2506 | && ((GET_CODE (branch_target) == REG | |
2507 | && set_of (branch_target, insn) != NULL_RTX) | |
2508 | || insn_clobbers_hbr (insn) | |
2509 | || branch_addr - insn_addr > 600)) | |
2510 | { | |
0af56f80 | 2511 | rtx_insn *next = NEXT_INSN (insn); |
5a976006 | 2512 | int next_addr = INSN_ADDRESSES (INSN_UID (next)); |
2513 | if (insn != BB_END (bb) | |
2514 | && branch_addr - next_addr >= required_dist) | |
2515 | { | |
2516 | if (dump_file) | |
2517 | fprintf (dump_file, | |
2518 | "hint for %i in block %i before %i\n", | |
2519 | INSN_UID (branch), bb->index, | |
2520 | INSN_UID (next)); | |
2521 | spu_emit_branch_hint (next, branch, branch_target, | |
2522 | branch_addr - next_addr, blocks); | |
2523 | } | |
2524 | branch = 0; | |
2525 | } | |
2526 | ||
2527 | /* JUMP_P will only be true at the end of a block. When | |
2528 | branch is already set it means we've previously decided | |
2529 | to propagate a hint for that branch into this block. */ | |
2530 | if (CALL_P (insn) || (JUMP_P (insn) && !branch)) | |
2531 | { | |
2532 | branch = 0; | |
2533 | if ((branch_target = get_branch_target (insn))) | |
2534 | { | |
2535 | branch = insn; | |
2536 | branch_addr = insn_addr; | |
2537 | required_dist = spu_hint_dist; | |
2538 | } | |
2539 | } | |
2540 | } | |
2541 | if (insn == BB_HEAD (bb)) | |
2542 | break; | |
2543 | } | |
2544 | ||
2545 | if (branch) | |
2546 | { | |
2547 | /* If we haven't emitted a hint for this branch yet, it might | |
2548 | be profitable to emit it in one of the predecessor blocks, | |
2549 | especially for loops. */ | |
0af56f80 | 2550 | rtx_insn *bbend; |
5a976006 | 2551 | basic_block prev = 0, prop = 0, prev2 = 0; |
2552 | int loop_exit = 0, simple_loop = 0; | |
2553 | int next_addr = INSN_ADDRESSES (INSN_UID (NEXT_INSN (insn))); | |
2554 | ||
2555 | for (j = 0; j < EDGE_COUNT (bb->preds); j++) | |
2556 | if (EDGE_PRED (bb, j)->flags & EDGE_FALLTHRU) | |
2557 | prev = EDGE_PRED (bb, j)->src; | |
2558 | else | |
2559 | prev2 = EDGE_PRED (bb, j)->src; | |
2560 | ||
2561 | for (j = 0; j < EDGE_COUNT (bb->succs); j++) | |
2562 | if (EDGE_SUCC (bb, j)->flags & EDGE_LOOP_EXIT) | |
2563 | loop_exit = 1; | |
2564 | else if (EDGE_SUCC (bb, j)->dest == bb) | |
2565 | simple_loop = 1; | |
2566 | ||
2567 | /* If this branch is a loop exit then propagate to previous | |
2568 | fallthru block. This catches the cases when it is a simple | |
2569 | loop or when there is an initial branch into the loop. */ | |
2570 | if (prev && (loop_exit || simple_loop) | |
a7a0184d | 2571 | && bb_loop_depth (prev) <= bb_loop_depth (bb)) |
5a976006 | 2572 | prop = prev; |
2573 | ||
2574 | /* If there is only one adjacent predecessor. Don't propagate | |
a7a0184d | 2575 | outside this loop. */ |
5a976006 | 2576 | else if (prev && single_pred_p (bb) |
a7a0184d | 2577 | && prev->loop_father == bb->loop_father) |
5a976006 | 2578 | prop = prev; |
2579 | ||
2580 | /* If this is the JOIN block of a simple IF-THEN then | |
9d75589a | 2581 | propagate the hint to the HEADER block. */ |
5a976006 | 2582 | else if (prev && prev2 |
2583 | && EDGE_COUNT (bb->preds) == 2 | |
2584 | && EDGE_COUNT (prev->preds) == 1 | |
2585 | && EDGE_PRED (prev, 0)->src == prev2 | |
a7a0184d | 2586 | && prev2->loop_father == bb->loop_father |
5a976006 | 2587 | && GET_CODE (branch_target) != REG) |
2588 | prop = prev; | |
2589 | ||
2590 | /* Don't propagate when: | |
2591 | - this is a simple loop and the hint would be too far | |
2592 | - this is not a simple loop and there are 16 insns in | |
2593 | this block already | |
2594 | - the predecessor block ends in a branch that will be | |
2595 | hinted | |
2596 | - the predecessor block ends in an insn that invalidates | |
2597 | the hint */ | |
2598 | if (prop | |
2599 | && prop->index >= 0 | |
2600 | && (bbend = BB_END (prop)) | |
2601 | && branch_addr - INSN_ADDRESSES (INSN_UID (bbend)) < | |
2602 | (simple_loop ? 600 : 16 * 4) && get_branch_target (bbend) == 0 | |
2603 | && (JUMP_P (bbend) || !insn_clobbers_hbr (bbend))) | |
2604 | { | |
2605 | if (dump_file) | |
2606 | fprintf (dump_file, "propagate from %i to %i (loop depth %i) " | |
2607 | "for %i (loop_exit %i simple_loop %i dist %i)\n", | |
a7a0184d | 2608 | bb->index, prop->index, bb_loop_depth (bb), |
5a976006 | 2609 | INSN_UID (branch), loop_exit, simple_loop, |
2610 | branch_addr - INSN_ADDRESSES (INSN_UID (bbend))); | |
2611 | ||
2612 | spu_bb_info[prop->index].prop_jump = branch; | |
2613 | spu_bb_info[prop->index].bb_index = i; | |
2614 | } | |
2615 | else if (branch_addr - next_addr >= required_dist) | |
2616 | { | |
2617 | if (dump_file) | |
2618 | fprintf (dump_file, "hint for %i in block %i before %i\n", | |
2619 | INSN_UID (branch), bb->index, | |
2620 | INSN_UID (NEXT_INSN (insn))); | |
2621 | spu_emit_branch_hint (NEXT_INSN (insn), branch, branch_target, | |
2622 | branch_addr - next_addr, blocks); | |
2623 | } | |
2624 | branch = 0; | |
644459d0 | 2625 | } |
644459d0 | 2626 | } |
5a976006 | 2627 | free (spu_bb_info); |
644459d0 | 2628 | |
53c5d9d4 | 2629 | if (!bitmap_empty_p (blocks)) |
5a976006 | 2630 | find_many_sub_basic_blocks (blocks); |
2631 | ||
2632 | /* We have to schedule to make sure alignment is ok. */ | |
fc00614f | 2633 | FOR_EACH_BB_FN (bb, cfun) bb->flags &= ~BB_DISABLE_SCHEDULE; |
5a976006 | 2634 | |
2635 | /* The hints need to be scheduled, so call it again. */ | |
2636 | schedule_insns (); | |
2fbdf9ef | 2637 | df_finish_pass (true); |
5a976006 | 2638 | |
2639 | insert_hbrp (); | |
2640 | ||
2641 | pad_bb (); | |
2642 | ||
8f1d58ad | 2643 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
2644 | if (NONJUMP_INSN_P (insn) && INSN_CODE (insn) == CODE_FOR_hbr) | |
2645 | { | |
2646 | /* Adjust the LABEL_REF in a hint when we have inserted a nop | |
2647 | between its branch label and the branch . We don't move the | |
2648 | label because GCC expects it at the beginning of the block. */ | |
2649 | rtx unspec = SET_SRC (XVECEXP (PATTERN (insn), 0, 0)); | |
2650 | rtx label_ref = XVECEXP (unspec, 0, 0); | |
4cd001d5 | 2651 | rtx_insn *label = as_a <rtx_insn *> (XEXP (label_ref, 0)); |
2652 | rtx_insn *branch; | |
8f1d58ad | 2653 | int offset = 0; |
2654 | for (branch = NEXT_INSN (label); | |
2655 | !JUMP_P (branch) && !CALL_P (branch); | |
2656 | branch = NEXT_INSN (branch)) | |
2657 | if (NONJUMP_INSN_P (branch)) | |
2658 | offset += get_attr_length (branch); | |
2659 | if (offset > 0) | |
29c05e22 | 2660 | XVECEXP (unspec, 0, 0) = plus_constant (Pmode, label_ref, offset); |
8f1d58ad | 2661 | } |
5a976006 | 2662 | |
8a42230a | 2663 | spu_var_tracking (); |
5a976006 | 2664 | |
a7a0184d | 2665 | loop_optimizer_finalize (); |
2666 | ||
5a976006 | 2667 | free_bb_for_insn (); |
2668 | ||
2669 | in_spu_reorg = 0; | |
644459d0 | 2670 | } |
2671 | \f | |
2672 | ||
2673 | /* Insn scheduling routines, primarily for dual issue. */ | |
2674 | static int | |
2675 | spu_sched_issue_rate (void) | |
2676 | { | |
2677 | return 2; | |
2678 | } | |
2679 | ||
2680 | static int | |
0af56f80 | 2681 | uses_ls_unit(rtx_insn *insn) |
644459d0 | 2682 | { |
5a976006 | 2683 | rtx set = single_set (insn); |
2684 | if (set != 0 | |
2685 | && (GET_CODE (SET_DEST (set)) == MEM | |
2686 | || GET_CODE (SET_SRC (set)) == MEM)) | |
2687 | return 1; | |
2688 | return 0; | |
644459d0 | 2689 | } |
2690 | ||
2691 | static int | |
0af56f80 | 2692 | get_pipe (rtx_insn *insn) |
644459d0 | 2693 | { |
2694 | enum attr_type t; | |
2695 | /* Handle inline asm */ | |
2696 | if (INSN_CODE (insn) == -1) | |
2697 | return -1; | |
2698 | t = get_attr_type (insn); | |
2699 | switch (t) | |
2700 | { | |
2701 | case TYPE_CONVERT: | |
2702 | return -2; | |
2703 | case TYPE_MULTI0: | |
2704 | return -1; | |
2705 | ||
2706 | case TYPE_FX2: | |
2707 | case TYPE_FX3: | |
2708 | case TYPE_SPR: | |
2709 | case TYPE_NOP: | |
2710 | case TYPE_FXB: | |
2711 | case TYPE_FPD: | |
2712 | case TYPE_FP6: | |
2713 | case TYPE_FP7: | |
644459d0 | 2714 | return 0; |
2715 | ||
2716 | case TYPE_LNOP: | |
2717 | case TYPE_SHUF: | |
2718 | case TYPE_LOAD: | |
2719 | case TYPE_STORE: | |
2720 | case TYPE_BR: | |
2721 | case TYPE_MULTI1: | |
2722 | case TYPE_HBR: | |
5a976006 | 2723 | case TYPE_IPREFETCH: |
644459d0 | 2724 | return 1; |
2725 | default: | |
2726 | abort (); | |
2727 | } | |
2728 | } | |
2729 | ||
5a976006 | 2730 | |
2731 | /* haifa-sched.c has a static variable that keeps track of the current | |
2732 | cycle. It is passed to spu_sched_reorder, and we record it here for | |
2733 | use by spu_sched_variable_issue. It won't be accurate if the | |
2734 | scheduler updates it's clock_var between the two calls. */ | |
2735 | static int clock_var; | |
2736 | ||
2737 | /* This is used to keep track of insn alignment. Set to 0 at the | |
2738 | beginning of each block and increased by the "length" attr of each | |
2739 | insn scheduled. */ | |
2740 | static int spu_sched_length; | |
2741 | ||
2742 | /* Record when we've issued pipe0 and pipe1 insns so we can reorder the | |
2743 | ready list appropriately in spu_sched_reorder(). */ | |
2744 | static int pipe0_clock; | |
2745 | static int pipe1_clock; | |
2746 | ||
2747 | static int prev_clock_var; | |
2748 | ||
2749 | static int prev_priority; | |
2750 | ||
2751 | /* The SPU needs to load the next ilb sometime during the execution of | |
2752 | the previous ilb. There is a potential conflict if every cycle has a | |
2753 | load or store. To avoid the conflict we make sure the load/store | |
2754 | unit is free for at least one cycle during the execution of insns in | |
2755 | the previous ilb. */ | |
2756 | static int spu_ls_first; | |
2757 | static int prev_ls_clock; | |
2758 | ||
2759 | static void | |
2760 | spu_sched_init_global (FILE *file ATTRIBUTE_UNUSED, int verbose ATTRIBUTE_UNUSED, | |
2761 | int max_ready ATTRIBUTE_UNUSED) | |
2762 | { | |
2763 | spu_sched_length = 0; | |
2764 | } | |
2765 | ||
2766 | static void | |
2767 | spu_sched_init (FILE *file ATTRIBUTE_UNUSED, int verbose ATTRIBUTE_UNUSED, | |
2768 | int max_ready ATTRIBUTE_UNUSED) | |
2769 | { | |
2770 | if (align_labels > 4 || align_loops > 4 || align_jumps > 4) | |
2771 | { | |
2772 | /* When any block might be at least 8-byte aligned, assume they | |
2773 | will all be at least 8-byte aligned to make sure dual issue | |
2774 | works out correctly. */ | |
2775 | spu_sched_length = 0; | |
2776 | } | |
2777 | spu_ls_first = INT_MAX; | |
2778 | clock_var = -1; | |
2779 | prev_ls_clock = -1; | |
2780 | pipe0_clock = -1; | |
2781 | pipe1_clock = -1; | |
2782 | prev_clock_var = -1; | |
2783 | prev_priority = -1; | |
2784 | } | |
2785 | ||
644459d0 | 2786 | static int |
5a976006 | 2787 | spu_sched_variable_issue (FILE *file ATTRIBUTE_UNUSED, |
0af56f80 | 2788 | int verbose ATTRIBUTE_UNUSED, |
18282db0 | 2789 | rtx_insn *insn, int more) |
644459d0 | 2790 | { |
5a976006 | 2791 | int len; |
2792 | int p; | |
644459d0 | 2793 | if (GET_CODE (PATTERN (insn)) == USE |
2794 | || GET_CODE (PATTERN (insn)) == CLOBBER | |
5a976006 | 2795 | || (len = get_attr_length (insn)) == 0) |
2796 | return more; | |
2797 | ||
2798 | spu_sched_length += len; | |
2799 | ||
2800 | /* Reset on inline asm */ | |
2801 | if (INSN_CODE (insn) == -1) | |
2802 | { | |
2803 | spu_ls_first = INT_MAX; | |
2804 | pipe0_clock = -1; | |
2805 | pipe1_clock = -1; | |
2806 | return 0; | |
2807 | } | |
2808 | p = get_pipe (insn); | |
2809 | if (p == 0) | |
2810 | pipe0_clock = clock_var; | |
2811 | else | |
2812 | pipe1_clock = clock_var; | |
2813 | ||
2814 | if (in_spu_reorg) | |
2815 | { | |
2816 | if (clock_var - prev_ls_clock > 1 | |
2817 | || INSN_CODE (insn) == CODE_FOR_iprefetch) | |
2818 | spu_ls_first = INT_MAX; | |
2819 | if (uses_ls_unit (insn)) | |
2820 | { | |
2821 | if (spu_ls_first == INT_MAX) | |
2822 | spu_ls_first = spu_sched_length; | |
2823 | prev_ls_clock = clock_var; | |
2824 | } | |
2825 | ||
2826 | /* The scheduler hasn't inserted the nop, but we will later on. | |
2827 | Include those nops in spu_sched_length. */ | |
2828 | if (prev_clock_var == clock_var && (spu_sched_length & 7)) | |
2829 | spu_sched_length += 4; | |
2830 | prev_clock_var = clock_var; | |
2831 | ||
2832 | /* more is -1 when called from spu_sched_reorder for new insns | |
2833 | that don't have INSN_PRIORITY */ | |
2834 | if (more >= 0) | |
2835 | prev_priority = INSN_PRIORITY (insn); | |
2836 | } | |
2837 | ||
9d75589a | 2838 | /* Always try issuing more insns. spu_sched_reorder will decide |
5a976006 | 2839 | when the cycle should be advanced. */ |
2840 | return 1; | |
2841 | } | |
2842 | ||
2843 | /* This function is called for both TARGET_SCHED_REORDER and | |
2844 | TARGET_SCHED_REORDER2. */ | |
2845 | static int | |
2846 | spu_sched_reorder (FILE *file ATTRIBUTE_UNUSED, int verbose ATTRIBUTE_UNUSED, | |
b24ef467 | 2847 | rtx_insn **ready, int *nreadyp, int clock) |
5a976006 | 2848 | { |
2849 | int i, nready = *nreadyp; | |
2850 | int pipe_0, pipe_1, pipe_hbrp, pipe_ls, schedule_i; | |
b24ef467 | 2851 | rtx_insn *insn; |
5a976006 | 2852 | |
2853 | clock_var = clock; | |
2854 | ||
2855 | if (nready <= 0 || pipe1_clock >= clock) | |
2856 | return 0; | |
2857 | ||
2858 | /* Find any rtl insns that don't generate assembly insns and schedule | |
2859 | them first. */ | |
2860 | for (i = nready - 1; i >= 0; i--) | |
2861 | { | |
2862 | insn = ready[i]; | |
2863 | if (INSN_CODE (insn) == -1 | |
2864 | || INSN_CODE (insn) == CODE_FOR_blockage | |
9d98604b | 2865 | || (INSN_P (insn) && get_attr_length (insn) == 0)) |
5a976006 | 2866 | { |
2867 | ready[i] = ready[nready - 1]; | |
2868 | ready[nready - 1] = insn; | |
2869 | return 1; | |
2870 | } | |
2871 | } | |
2872 | ||
2873 | pipe_0 = pipe_1 = pipe_hbrp = pipe_ls = schedule_i = -1; | |
2874 | for (i = 0; i < nready; i++) | |
2875 | if (INSN_CODE (ready[i]) != -1) | |
2876 | { | |
2877 | insn = ready[i]; | |
2878 | switch (get_attr_type (insn)) | |
2879 | { | |
2880 | default: | |
2881 | case TYPE_MULTI0: | |
2882 | case TYPE_CONVERT: | |
2883 | case TYPE_FX2: | |
2884 | case TYPE_FX3: | |
2885 | case TYPE_SPR: | |
2886 | case TYPE_NOP: | |
2887 | case TYPE_FXB: | |
2888 | case TYPE_FPD: | |
2889 | case TYPE_FP6: | |
2890 | case TYPE_FP7: | |
2891 | pipe_0 = i; | |
2892 | break; | |
2893 | case TYPE_LOAD: | |
2894 | case TYPE_STORE: | |
2895 | pipe_ls = i; | |
2896 | case TYPE_LNOP: | |
2897 | case TYPE_SHUF: | |
2898 | case TYPE_BR: | |
2899 | case TYPE_MULTI1: | |
2900 | case TYPE_HBR: | |
2901 | pipe_1 = i; | |
2902 | break; | |
2903 | case TYPE_IPREFETCH: | |
2904 | pipe_hbrp = i; | |
2905 | break; | |
2906 | } | |
2907 | } | |
2908 | ||
2909 | /* In the first scheduling phase, schedule loads and stores together | |
2910 | to increase the chance they will get merged during postreload CSE. */ | |
2911 | if (!reload_completed && pipe_ls >= 0) | |
2912 | { | |
2913 | insn = ready[pipe_ls]; | |
2914 | ready[pipe_ls] = ready[nready - 1]; | |
2915 | ready[nready - 1] = insn; | |
2916 | return 1; | |
2917 | } | |
2918 | ||
2919 | /* If there is an hbrp ready, prefer it over other pipe 1 insns. */ | |
2920 | if (pipe_hbrp >= 0) | |
2921 | pipe_1 = pipe_hbrp; | |
2922 | ||
2923 | /* When we have loads/stores in every cycle of the last 15 insns and | |
2924 | we are about to schedule another load/store, emit an hbrp insn | |
2925 | instead. */ | |
2926 | if (in_spu_reorg | |
2927 | && spu_sched_length - spu_ls_first >= 4 * 15 | |
2928 | && !(pipe0_clock < clock && pipe_0 >= 0) && pipe_1 == pipe_ls) | |
2929 | { | |
2930 | insn = sched_emit_insn (gen_iprefetch (GEN_INT (3))); | |
2931 | recog_memoized (insn); | |
2932 | if (pipe0_clock < clock) | |
2933 | PUT_MODE (insn, TImode); | |
2934 | spu_sched_variable_issue (file, verbose, insn, -1); | |
2935 | return 0; | |
2936 | } | |
2937 | ||
2938 | /* In general, we want to emit nops to increase dual issue, but dual | |
2939 | issue isn't faster when one of the insns could be scheduled later | |
2940 | without effecting the critical path. We look at INSN_PRIORITY to | |
2941 | make a good guess, but it isn't perfect so -mdual-nops=n can be | |
2942 | used to effect it. */ | |
2943 | if (in_spu_reorg && spu_dual_nops < 10) | |
2944 | { | |
9d75589a | 2945 | /* When we are at an even address and we are not issuing nops to |
5a976006 | 2946 | improve scheduling then we need to advance the cycle. */ |
2947 | if ((spu_sched_length & 7) == 0 && prev_clock_var == clock | |
2948 | && (spu_dual_nops == 0 | |
2949 | || (pipe_1 != -1 | |
2950 | && prev_priority > | |
2951 | INSN_PRIORITY (ready[pipe_1]) + spu_dual_nops))) | |
2952 | return 0; | |
2953 | ||
2954 | /* When at an odd address, schedule the highest priority insn | |
2955 | without considering pipeline. */ | |
2956 | if ((spu_sched_length & 7) == 4 && prev_clock_var != clock | |
2957 | && (spu_dual_nops == 0 | |
2958 | || (prev_priority > | |
2959 | INSN_PRIORITY (ready[nready - 1]) + spu_dual_nops))) | |
2960 | return 1; | |
2961 | } | |
2962 | ||
2963 | ||
2964 | /* We haven't issued a pipe0 insn yet this cycle, if there is a | |
2965 | pipe0 insn in the ready list, schedule it. */ | |
2966 | if (pipe0_clock < clock && pipe_0 >= 0) | |
2967 | schedule_i = pipe_0; | |
2968 | ||
2969 | /* Either we've scheduled a pipe0 insn already or there is no pipe0 | |
2970 | insn to schedule. Put a pipe1 insn at the front of the ready list. */ | |
2971 | else | |
2972 | schedule_i = pipe_1; | |
2973 | ||
2974 | if (schedule_i > -1) | |
2975 | { | |
2976 | insn = ready[schedule_i]; | |
2977 | ready[schedule_i] = ready[nready - 1]; | |
2978 | ready[nready - 1] = insn; | |
2979 | return 1; | |
2980 | } | |
2981 | return 0; | |
644459d0 | 2982 | } |
2983 | ||
2984 | /* INSN is dependent on DEP_INSN. */ | |
2985 | static int | |
18282db0 | 2986 | spu_sched_adjust_cost (rtx_insn *insn, rtx link, rtx_insn *dep_insn, int cost) |
644459d0 | 2987 | { |
5a976006 | 2988 | rtx set; |
2989 | ||
2990 | /* The blockage pattern is used to prevent instructions from being | |
2991 | moved across it and has no cost. */ | |
2992 | if (INSN_CODE (insn) == CODE_FOR_blockage | |
2993 | || INSN_CODE (dep_insn) == CODE_FOR_blockage) | |
2994 | return 0; | |
2995 | ||
9d98604b | 2996 | if ((INSN_P (insn) && get_attr_length (insn) == 0) |
2997 | || (INSN_P (dep_insn) && get_attr_length (dep_insn) == 0)) | |
5a976006 | 2998 | return 0; |
2999 | ||
3000 | /* Make sure hbrps are spread out. */ | |
3001 | if (INSN_CODE (insn) == CODE_FOR_iprefetch | |
3002 | && INSN_CODE (dep_insn) == CODE_FOR_iprefetch) | |
3003 | return 8; | |
3004 | ||
3005 | /* Make sure hints and hbrps are 2 cycles apart. */ | |
3006 | if ((INSN_CODE (insn) == CODE_FOR_iprefetch | |
3007 | || INSN_CODE (insn) == CODE_FOR_hbr) | |
3008 | && (INSN_CODE (dep_insn) == CODE_FOR_iprefetch | |
3009 | || INSN_CODE (dep_insn) == CODE_FOR_hbr)) | |
3010 | return 2; | |
3011 | ||
3012 | /* An hbrp has no real dependency on other insns. */ | |
3013 | if (INSN_CODE (insn) == CODE_FOR_iprefetch | |
3014 | || INSN_CODE (dep_insn) == CODE_FOR_iprefetch) | |
3015 | return 0; | |
3016 | ||
3017 | /* Assuming that it is unlikely an argument register will be used in | |
3018 | the first cycle of the called function, we reduce the cost for | |
3019 | slightly better scheduling of dep_insn. When not hinted, the | |
3020 | mispredicted branch would hide the cost as well. */ | |
3021 | if (CALL_P (insn)) | |
3022 | { | |
3023 | rtx target = get_branch_target (insn); | |
3024 | if (GET_CODE (target) != REG || !set_of (target, insn)) | |
3025 | return cost - 2; | |
3026 | return cost; | |
3027 | } | |
3028 | ||
3029 | /* And when returning from a function, let's assume the return values | |
3030 | are completed sooner too. */ | |
3031 | if (CALL_P (dep_insn)) | |
644459d0 | 3032 | return cost - 2; |
5a976006 | 3033 | |
3034 | /* Make sure an instruction that loads from the back chain is schedule | |
3035 | away from the return instruction so a hint is more likely to get | |
3036 | issued. */ | |
3037 | if (INSN_CODE (insn) == CODE_FOR__return | |
3038 | && (set = single_set (dep_insn)) | |
3039 | && GET_CODE (SET_DEST (set)) == REG | |
3040 | && REGNO (SET_DEST (set)) == LINK_REGISTER_REGNUM) | |
3041 | return 20; | |
3042 | ||
644459d0 | 3043 | /* The dfa scheduler sets cost to 0 for all anti-dependencies and the |
3044 | scheduler makes every insn in a block anti-dependent on the final | |
3045 | jump_insn. We adjust here so higher cost insns will get scheduled | |
3046 | earlier. */ | |
5a976006 | 3047 | if (JUMP_P (insn) && REG_NOTE_KIND (link) == REG_DEP_ANTI) |
9997bd27 | 3048 | return insn_cost (dep_insn) - 3; |
5a976006 | 3049 | |
644459d0 | 3050 | return cost; |
3051 | } | |
3052 | \f | |
3053 | /* Create a CONST_DOUBLE from a string. */ | |
842ae815 | 3054 | rtx |
3754d046 | 3055 | spu_float_const (const char *string, machine_mode mode) |
644459d0 | 3056 | { |
3057 | REAL_VALUE_TYPE value; | |
3058 | value = REAL_VALUE_ATOF (string, mode); | |
d5f9611d | 3059 | return const_double_from_real_value (value, mode); |
644459d0 | 3060 | } |
3061 | ||
644459d0 | 3062 | int |
3063 | spu_constant_address_p (rtx x) | |
3064 | { | |
3065 | return (GET_CODE (x) == LABEL_REF || GET_CODE (x) == SYMBOL_REF | |
3066 | || GET_CODE (x) == CONST_INT || GET_CODE (x) == CONST | |
3067 | || GET_CODE (x) == HIGH); | |
3068 | } | |
3069 | ||
3070 | static enum spu_immediate | |
3071 | which_immediate_load (HOST_WIDE_INT val) | |
3072 | { | |
3073 | gcc_assert (val == trunc_int_for_mode (val, SImode)); | |
3074 | ||
3075 | if (val >= -0x8000 && val <= 0x7fff) | |
3076 | return SPU_IL; | |
3077 | if (val >= 0 && val <= 0x3ffff) | |
3078 | return SPU_ILA; | |
3079 | if ((val & 0xffff) == ((val >> 16) & 0xffff)) | |
3080 | return SPU_ILH; | |
3081 | if ((val & 0xffff) == 0) | |
3082 | return SPU_ILHU; | |
3083 | ||
3084 | return SPU_NONE; | |
3085 | } | |
3086 | ||
dea01258 | 3087 | /* Return true when OP can be loaded by one of the il instructions, or |
3088 | when flow2 is not completed and OP can be loaded using ilhu and iohl. */ | |
644459d0 | 3089 | int |
3754d046 | 3090 | immediate_load_p (rtx op, machine_mode mode) |
dea01258 | 3091 | { |
3092 | if (CONSTANT_P (op)) | |
3093 | { | |
3094 | enum immediate_class c = classify_immediate (op, mode); | |
5df189be | 3095 | return c == IC_IL1 || c == IC_IL1s |
3072d30e | 3096 | || (!epilogue_completed && (c == IC_IL2 || c == IC_IL2s)); |
dea01258 | 3097 | } |
3098 | return 0; | |
3099 | } | |
3100 | ||
3101 | /* Return true if the first SIZE bytes of arr is a constant that can be | |
3102 | generated with cbd, chd, cwd or cdd. When non-NULL, PRUN and PSTART | |
3103 | represent the size and offset of the instruction to use. */ | |
3104 | static int | |
3105 | cpat_info(unsigned char *arr, int size, int *prun, int *pstart) | |
3106 | { | |
3107 | int cpat, run, i, start; | |
3108 | cpat = 1; | |
3109 | run = 0; | |
3110 | start = -1; | |
3111 | for (i = 0; i < size && cpat; i++) | |
3112 | if (arr[i] != i+16) | |
3113 | { | |
3114 | if (!run) | |
3115 | { | |
3116 | start = i; | |
3117 | if (arr[i] == 3) | |
3118 | run = 1; | |
3119 | else if (arr[i] == 2 && arr[i+1] == 3) | |
3120 | run = 2; | |
3121 | else if (arr[i] == 0) | |
3122 | { | |
3123 | while (arr[i+run] == run && i+run < 16) | |
3124 | run++; | |
3125 | if (run != 4 && run != 8) | |
3126 | cpat = 0; | |
3127 | } | |
3128 | else | |
3129 | cpat = 0; | |
3130 | if ((i & (run-1)) != 0) | |
3131 | cpat = 0; | |
3132 | i += run; | |
3133 | } | |
3134 | else | |
3135 | cpat = 0; | |
3136 | } | |
b01a6dc3 | 3137 | if (cpat && (run || size < 16)) |
dea01258 | 3138 | { |
3139 | if (run == 0) | |
3140 | run = 1; | |
3141 | if (prun) | |
3142 | *prun = run; | |
3143 | if (pstart) | |
3144 | *pstart = start == -1 ? 16-run : start; | |
3145 | return 1; | |
3146 | } | |
3147 | return 0; | |
3148 | } | |
3149 | ||
3150 | /* OP is a CONSTANT_P. Determine what instructions can be used to load | |
d819917f | 3151 | it into a register. MODE is only valid when OP is a CONST_INT. */ |
dea01258 | 3152 | static enum immediate_class |
3754d046 | 3153 | classify_immediate (rtx op, machine_mode mode) |
644459d0 | 3154 | { |
3155 | HOST_WIDE_INT val; | |
3156 | unsigned char arr[16]; | |
5df189be | 3157 | int i, j, repeated, fsmbi, repeat; |
dea01258 | 3158 | |
3159 | gcc_assert (CONSTANT_P (op)); | |
3160 | ||
644459d0 | 3161 | if (GET_MODE (op) != VOIDmode) |
3162 | mode = GET_MODE (op); | |
3163 | ||
dea01258 | 3164 | /* A V4SI const_vector with all identical symbols is ok. */ |
5df189be | 3165 | if (!flag_pic |
3166 | && mode == V4SImode | |
dea01258 | 3167 | && GET_CODE (op) == CONST_VECTOR |
3168 | && GET_CODE (CONST_VECTOR_ELT (op, 0)) != CONST_INT | |
62fdb8e4 | 3169 | && GET_CODE (CONST_VECTOR_ELT (op, 0)) != CONST_DOUBLE) |
3170 | op = unwrap_const_vec_duplicate (op); | |
644459d0 | 3171 | |
dea01258 | 3172 | switch (GET_CODE (op)) |
3173 | { | |
3174 | case SYMBOL_REF: | |
3175 | case LABEL_REF: | |
3176 | return TARGET_LARGE_MEM ? IC_IL2s : IC_IL1s; | |
644459d0 | 3177 | |
dea01258 | 3178 | case CONST: |
0cfc65d4 | 3179 | /* We can never know if the resulting address fits in 18 bits and can be |
3180 | loaded with ila. For now, assume the address will not overflow if | |
3181 | the displacement is "small" (fits 'K' constraint). */ | |
3182 | if (!TARGET_LARGE_MEM && GET_CODE (XEXP (op, 0)) == PLUS) | |
3183 | { | |
3184 | rtx sym = XEXP (XEXP (op, 0), 0); | |
3185 | rtx cst = XEXP (XEXP (op, 0), 1); | |
3186 | ||
3187 | if (GET_CODE (sym) == SYMBOL_REF | |
3188 | && GET_CODE (cst) == CONST_INT | |
3189 | && satisfies_constraint_K (cst)) | |
3190 | return IC_IL1s; | |
3191 | } | |
3192 | return IC_IL2s; | |
644459d0 | 3193 | |
dea01258 | 3194 | case HIGH: |
3195 | return IC_IL1s; | |
3196 | ||
3197 | case CONST_VECTOR: | |
3198 | for (i = 0; i < GET_MODE_NUNITS (mode); i++) | |
3199 | if (GET_CODE (CONST_VECTOR_ELT (op, i)) != CONST_INT | |
3200 | && GET_CODE (CONST_VECTOR_ELT (op, i)) != CONST_DOUBLE) | |
3201 | return IC_POOL; | |
3202 | /* Fall through. */ | |
3203 | ||
3204 | case CONST_INT: | |
3205 | case CONST_DOUBLE: | |
3206 | constant_to_array (mode, op, arr); | |
644459d0 | 3207 | |
dea01258 | 3208 | /* Check that each 4-byte slot is identical. */ |
3209 | repeated = 1; | |
3210 | for (i = 4; i < 16; i += 4) | |
3211 | for (j = 0; j < 4; j++) | |
3212 | if (arr[j] != arr[i + j]) | |
3213 | repeated = 0; | |
3214 | ||
3215 | if (repeated) | |
3216 | { | |
3217 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
3218 | val = trunc_int_for_mode (val, SImode); | |
3219 | ||
3220 | if (which_immediate_load (val) != SPU_NONE) | |
3221 | return IC_IL1; | |
3222 | } | |
3223 | ||
3224 | /* Any mode of 2 bytes or smaller can be loaded with an il | |
3225 | instruction. */ | |
3226 | gcc_assert (GET_MODE_SIZE (mode) > 2); | |
3227 | ||
3228 | fsmbi = 1; | |
5df189be | 3229 | repeat = 0; |
dea01258 | 3230 | for (i = 0; i < 16 && fsmbi; i++) |
5df189be | 3231 | if (arr[i] != 0 && repeat == 0) |
3232 | repeat = arr[i]; | |
3233 | else if (arr[i] != 0 && arr[i] != repeat) | |
dea01258 | 3234 | fsmbi = 0; |
3235 | if (fsmbi) | |
5df189be | 3236 | return repeat == 0xff ? IC_FSMBI : IC_FSMBI2; |
dea01258 | 3237 | |
3238 | if (cpat_info (arr, GET_MODE_SIZE (mode), 0, 0)) | |
3239 | return IC_CPAT; | |
3240 | ||
3241 | if (repeated) | |
3242 | return IC_IL2; | |
3243 | ||
3244 | return IC_POOL; | |
3245 | default: | |
3246 | break; | |
3247 | } | |
3248 | gcc_unreachable (); | |
644459d0 | 3249 | } |
3250 | ||
3251 | static enum spu_immediate | |
3252 | which_logical_immediate (HOST_WIDE_INT val) | |
3253 | { | |
3254 | gcc_assert (val == trunc_int_for_mode (val, SImode)); | |
3255 | ||
3256 | if (val >= -0x200 && val <= 0x1ff) | |
3257 | return SPU_ORI; | |
3258 | if (val >= 0 && val <= 0xffff) | |
3259 | return SPU_IOHL; | |
3260 | if ((val & 0xffff) == ((val >> 16) & 0xffff)) | |
3261 | { | |
3262 | val = trunc_int_for_mode (val, HImode); | |
3263 | if (val >= -0x200 && val <= 0x1ff) | |
3264 | return SPU_ORHI; | |
3265 | if ((val & 0xff) == ((val >> 8) & 0xff)) | |
3266 | { | |
3267 | val = trunc_int_for_mode (val, QImode); | |
3268 | if (val >= -0x200 && val <= 0x1ff) | |
3269 | return SPU_ORBI; | |
3270 | } | |
3271 | } | |
3272 | return SPU_NONE; | |
3273 | } | |
3274 | ||
5df189be | 3275 | /* Return TRUE when X, a CONST_VECTOR, only contains CONST_INTs or |
3276 | CONST_DOUBLEs. */ | |
3277 | static int | |
3278 | const_vector_immediate_p (rtx x) | |
3279 | { | |
3280 | int i; | |
3281 | gcc_assert (GET_CODE (x) == CONST_VECTOR); | |
3282 | for (i = 0; i < GET_MODE_NUNITS (GET_MODE (x)); i++) | |
3283 | if (GET_CODE (CONST_VECTOR_ELT (x, i)) != CONST_INT | |
3284 | && GET_CODE (CONST_VECTOR_ELT (x, i)) != CONST_DOUBLE) | |
3285 | return 0; | |
3286 | return 1; | |
3287 | } | |
3288 | ||
644459d0 | 3289 | int |
3754d046 | 3290 | logical_immediate_p (rtx op, machine_mode mode) |
644459d0 | 3291 | { |
3292 | HOST_WIDE_INT val; | |
3293 | unsigned char arr[16]; | |
3294 | int i, j; | |
3295 | ||
3296 | gcc_assert (GET_CODE (op) == CONST_INT || GET_CODE (op) == CONST_DOUBLE | |
3297 | || GET_CODE (op) == CONST_VECTOR); | |
3298 | ||
5df189be | 3299 | if (GET_CODE (op) == CONST_VECTOR |
3300 | && !const_vector_immediate_p (op)) | |
3301 | return 0; | |
3302 | ||
644459d0 | 3303 | if (GET_MODE (op) != VOIDmode) |
3304 | mode = GET_MODE (op); | |
3305 | ||
3306 | constant_to_array (mode, op, arr); | |
3307 | ||
3308 | /* Check that bytes are repeated. */ | |
3309 | for (i = 4; i < 16; i += 4) | |
3310 | for (j = 0; j < 4; j++) | |
3311 | if (arr[j] != arr[i + j]) | |
3312 | return 0; | |
3313 | ||
3314 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
3315 | val = trunc_int_for_mode (val, SImode); | |
3316 | ||
3317 | i = which_logical_immediate (val); | |
3318 | return i != SPU_NONE && i != SPU_IOHL; | |
3319 | } | |
3320 | ||
3321 | int | |
3754d046 | 3322 | iohl_immediate_p (rtx op, machine_mode mode) |
644459d0 | 3323 | { |
3324 | HOST_WIDE_INT val; | |
3325 | unsigned char arr[16]; | |
3326 | int i, j; | |
3327 | ||
3328 | gcc_assert (GET_CODE (op) == CONST_INT || GET_CODE (op) == CONST_DOUBLE | |
3329 | || GET_CODE (op) == CONST_VECTOR); | |
3330 | ||
5df189be | 3331 | if (GET_CODE (op) == CONST_VECTOR |
3332 | && !const_vector_immediate_p (op)) | |
3333 | return 0; | |
3334 | ||
644459d0 | 3335 | if (GET_MODE (op) != VOIDmode) |
3336 | mode = GET_MODE (op); | |
3337 | ||
3338 | constant_to_array (mode, op, arr); | |
3339 | ||
3340 | /* Check that bytes are repeated. */ | |
3341 | for (i = 4; i < 16; i += 4) | |
3342 | for (j = 0; j < 4; j++) | |
3343 | if (arr[j] != arr[i + j]) | |
3344 | return 0; | |
3345 | ||
3346 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
3347 | val = trunc_int_for_mode (val, SImode); | |
3348 | ||
3349 | return val >= 0 && val <= 0xffff; | |
3350 | } | |
3351 | ||
3352 | int | |
3754d046 | 3353 | arith_immediate_p (rtx op, machine_mode mode, |
644459d0 | 3354 | HOST_WIDE_INT low, HOST_WIDE_INT high) |
3355 | { | |
3356 | HOST_WIDE_INT val; | |
3357 | unsigned char arr[16]; | |
3358 | int bytes, i, j; | |
3359 | ||
3360 | gcc_assert (GET_CODE (op) == CONST_INT || GET_CODE (op) == CONST_DOUBLE | |
3361 | || GET_CODE (op) == CONST_VECTOR); | |
3362 | ||
5df189be | 3363 | if (GET_CODE (op) == CONST_VECTOR |
3364 | && !const_vector_immediate_p (op)) | |
3365 | return 0; | |
3366 | ||
644459d0 | 3367 | if (GET_MODE (op) != VOIDmode) |
3368 | mode = GET_MODE (op); | |
3369 | ||
3370 | constant_to_array (mode, op, arr); | |
3371 | ||
6e256598 | 3372 | bytes = GET_MODE_UNIT_SIZE (mode); |
0436eae5 | 3373 | mode = mode_for_size (GET_MODE_UNIT_BITSIZE (mode), MODE_INT, 0); |
644459d0 | 3374 | |
3375 | /* Check that bytes are repeated. */ | |
3376 | for (i = bytes; i < 16; i += bytes) | |
3377 | for (j = 0; j < bytes; j++) | |
3378 | if (arr[j] != arr[i + j]) | |
3379 | return 0; | |
3380 | ||
3381 | val = arr[0]; | |
3382 | for (j = 1; j < bytes; j++) | |
3383 | val = (val << 8) | arr[j]; | |
3384 | ||
3385 | val = trunc_int_for_mode (val, mode); | |
3386 | ||
3387 | return val >= low && val <= high; | |
3388 | } | |
3389 | ||
56c7bfc2 | 3390 | /* TRUE when op is an immediate and an exact power of 2, and given that |
3391 | OP is 2^scale, scale >= LOW && scale <= HIGH. When OP is a vector, | |
3392 | all entries must be the same. */ | |
3393 | bool | |
3754d046 | 3394 | exp2_immediate_p (rtx op, machine_mode mode, int low, int high) |
56c7bfc2 | 3395 | { |
3754d046 | 3396 | machine_mode int_mode; |
56c7bfc2 | 3397 | HOST_WIDE_INT val; |
3398 | unsigned char arr[16]; | |
3399 | int bytes, i, j; | |
3400 | ||
3401 | gcc_assert (GET_CODE (op) == CONST_INT || GET_CODE (op) == CONST_DOUBLE | |
3402 | || GET_CODE (op) == CONST_VECTOR); | |
3403 | ||
3404 | if (GET_CODE (op) == CONST_VECTOR | |
3405 | && !const_vector_immediate_p (op)) | |
3406 | return 0; | |
3407 | ||
3408 | if (GET_MODE (op) != VOIDmode) | |
3409 | mode = GET_MODE (op); | |
3410 | ||
3411 | constant_to_array (mode, op, arr); | |
3412 | ||
6e256598 | 3413 | mode = GET_MODE_INNER (mode); |
56c7bfc2 | 3414 | |
3415 | bytes = GET_MODE_SIZE (mode); | |
3416 | int_mode = mode_for_size (GET_MODE_BITSIZE (mode), MODE_INT, 0); | |
3417 | ||
3418 | /* Check that bytes are repeated. */ | |
3419 | for (i = bytes; i < 16; i += bytes) | |
3420 | for (j = 0; j < bytes; j++) | |
3421 | if (arr[j] != arr[i + j]) | |
3422 | return 0; | |
3423 | ||
3424 | val = arr[0]; | |
3425 | for (j = 1; j < bytes; j++) | |
3426 | val = (val << 8) | arr[j]; | |
3427 | ||
3428 | val = trunc_int_for_mode (val, int_mode); | |
3429 | ||
3430 | /* Currently, we only handle SFmode */ | |
3431 | gcc_assert (mode == SFmode); | |
3432 | if (mode == SFmode) | |
3433 | { | |
3434 | int exp = (val >> 23) - 127; | |
3435 | return val > 0 && (val & 0x007fffff) == 0 | |
3436 | && exp >= low && exp <= high; | |
3437 | } | |
3438 | return FALSE; | |
3439 | } | |
3440 | ||
6cf5579e | 3441 | /* Return true if X is a SYMBOL_REF to an __ea qualified variable. */ |
3442 | ||
6f4e40cd | 3443 | static bool |
3444 | ea_symbol_ref_p (const_rtx x) | |
6cf5579e | 3445 | { |
6cf5579e | 3446 | tree decl; |
3447 | ||
3448 | if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS) | |
3449 | { | |
3450 | rtx plus = XEXP (x, 0); | |
3451 | rtx op0 = XEXP (plus, 0); | |
3452 | rtx op1 = XEXP (plus, 1); | |
3453 | if (GET_CODE (op1) == CONST_INT) | |
3454 | x = op0; | |
3455 | } | |
3456 | ||
3457 | return (GET_CODE (x) == SYMBOL_REF | |
3458 | && (decl = SYMBOL_REF_DECL (x)) != 0 | |
3459 | && TREE_CODE (decl) == VAR_DECL | |
3460 | && TYPE_ADDR_SPACE (TREE_TYPE (decl))); | |
3461 | } | |
3462 | ||
644459d0 | 3463 | /* We accept: |
5b865faf | 3464 | - any 32-bit constant (SImode, SFmode) |
644459d0 | 3465 | - any constant that can be generated with fsmbi (any mode) |
5b865faf | 3466 | - a 64-bit constant where the high and low bits are identical |
644459d0 | 3467 | (DImode, DFmode) |
5b865faf | 3468 | - a 128-bit constant where the four 32-bit words match. */ |
ca316360 | 3469 | bool |
3754d046 | 3470 | spu_legitimate_constant_p (machine_mode mode, rtx x) |
644459d0 | 3471 | { |
6f4e40cd | 3472 | subrtx_iterator::array_type array; |
5df189be | 3473 | if (GET_CODE (x) == HIGH) |
3474 | x = XEXP (x, 0); | |
6cf5579e | 3475 | |
3476 | /* Reject any __ea qualified reference. These can't appear in | |
3477 | instructions but must be forced to the constant pool. */ | |
6f4e40cd | 3478 | FOR_EACH_SUBRTX (iter, array, x, ALL) |
3479 | if (ea_symbol_ref_p (*iter)) | |
3480 | return 0; | |
6cf5579e | 3481 | |
644459d0 | 3482 | /* V4SI with all identical symbols is valid. */ |
5df189be | 3483 | if (!flag_pic |
ca316360 | 3484 | && mode == V4SImode |
644459d0 | 3485 | && (GET_CODE (CONST_VECTOR_ELT (x, 0)) == SYMBOL_REF |
3486 | || GET_CODE (CONST_VECTOR_ELT (x, 0)) == LABEL_REF | |
5df189be | 3487 | || GET_CODE (CONST_VECTOR_ELT (x, 0)) == CONST)) |
62fdb8e4 | 3488 | return const_vec_duplicate_p (x); |
644459d0 | 3489 | |
5df189be | 3490 | if (GET_CODE (x) == CONST_VECTOR |
3491 | && !const_vector_immediate_p (x)) | |
3492 | return 0; | |
644459d0 | 3493 | return 1; |
3494 | } | |
3495 | ||
3496 | /* Valid address are: | |
3497 | - symbol_ref, label_ref, const | |
3498 | - reg | |
9d98604b | 3499 | - reg + const_int, where const_int is 16 byte aligned |
644459d0 | 3500 | - reg + reg, alignment doesn't matter |
3501 | The alignment matters in the reg+const case because lqd and stqd | |
9d98604b | 3502 | ignore the 4 least significant bits of the const. We only care about |
3503 | 16 byte modes because the expand phase will change all smaller MEM | |
3504 | references to TImode. */ | |
3505 | static bool | |
3754d046 | 3506 | spu_legitimate_address_p (machine_mode mode, |
fd50b071 | 3507 | rtx x, bool reg_ok_strict) |
644459d0 | 3508 | { |
9d98604b | 3509 | int aligned = GET_MODE_SIZE (mode) >= 16; |
3510 | if (aligned | |
3511 | && GET_CODE (x) == AND | |
644459d0 | 3512 | && GET_CODE (XEXP (x, 1)) == CONST_INT |
9d98604b | 3513 | && INTVAL (XEXP (x, 1)) == (HOST_WIDE_INT) - 16) |
644459d0 | 3514 | x = XEXP (x, 0); |
3515 | switch (GET_CODE (x)) | |
3516 | { | |
644459d0 | 3517 | case LABEL_REF: |
6cf5579e | 3518 | return !TARGET_LARGE_MEM; |
3519 | ||
9d98604b | 3520 | case SYMBOL_REF: |
644459d0 | 3521 | case CONST: |
6cf5579e | 3522 | /* Keep __ea references until reload so that spu_expand_mov can see them |
3523 | in MEMs. */ | |
6f4e40cd | 3524 | if (ea_symbol_ref_p (x)) |
6cf5579e | 3525 | return !reload_in_progress && !reload_completed; |
9d98604b | 3526 | return !TARGET_LARGE_MEM; |
644459d0 | 3527 | |
3528 | case CONST_INT: | |
3529 | return INTVAL (x) >= 0 && INTVAL (x) <= 0x3ffff; | |
3530 | ||
3531 | case SUBREG: | |
3532 | x = XEXP (x, 0); | |
9d98604b | 3533 | if (REG_P (x)) |
3534 | return 0; | |
644459d0 | 3535 | |
3536 | case REG: | |
3537 | return INT_REG_OK_FOR_BASE_P (x, reg_ok_strict); | |
3538 | ||
3539 | case PLUS: | |
3540 | case LO_SUM: | |
3541 | { | |
3542 | rtx op0 = XEXP (x, 0); | |
3543 | rtx op1 = XEXP (x, 1); | |
3544 | if (GET_CODE (op0) == SUBREG) | |
3545 | op0 = XEXP (op0, 0); | |
3546 | if (GET_CODE (op1) == SUBREG) | |
3547 | op1 = XEXP (op1, 0); | |
644459d0 | 3548 | if (GET_CODE (op0) == REG |
3549 | && INT_REG_OK_FOR_BASE_P (op0, reg_ok_strict) | |
3550 | && GET_CODE (op1) == CONST_INT | |
fa695424 | 3551 | && ((INTVAL (op1) >= -0x2000 && INTVAL (op1) <= 0x1fff) |
3552 | /* If virtual registers are involved, the displacement will | |
3553 | change later on anyway, so checking would be premature. | |
3554 | Reload will make sure the final displacement after | |
3555 | register elimination is OK. */ | |
3556 | || op0 == arg_pointer_rtx | |
3557 | || op0 == frame_pointer_rtx | |
3558 | || op0 == virtual_stack_vars_rtx) | |
9d98604b | 3559 | && (!aligned || (INTVAL (op1) & 15) == 0)) |
3560 | return TRUE; | |
644459d0 | 3561 | if (GET_CODE (op0) == REG |
3562 | && INT_REG_OK_FOR_BASE_P (op0, reg_ok_strict) | |
3563 | && GET_CODE (op1) == REG | |
3564 | && INT_REG_OK_FOR_INDEX_P (op1, reg_ok_strict)) | |
9d98604b | 3565 | return TRUE; |
644459d0 | 3566 | } |
3567 | break; | |
3568 | ||
3569 | default: | |
3570 | break; | |
3571 | } | |
9d98604b | 3572 | return FALSE; |
644459d0 | 3573 | } |
3574 | ||
6cf5579e | 3575 | /* Like spu_legitimate_address_p, except with named addresses. */ |
3576 | static bool | |
3754d046 | 3577 | spu_addr_space_legitimate_address_p (machine_mode mode, rtx x, |
6cf5579e | 3578 | bool reg_ok_strict, addr_space_t as) |
3579 | { | |
3580 | if (as == ADDR_SPACE_EA) | |
3581 | return (REG_P (x) && (GET_MODE (x) == EAmode)); | |
3582 | ||
3583 | else if (as != ADDR_SPACE_GENERIC) | |
3584 | gcc_unreachable (); | |
3585 | ||
3586 | return spu_legitimate_address_p (mode, x, reg_ok_strict); | |
3587 | } | |
3588 | ||
644459d0 | 3589 | /* When the address is reg + const_int, force the const_int into a |
fa7637bd | 3590 | register. */ |
3defb88e | 3591 | static rtx |
644459d0 | 3592 | spu_legitimize_address (rtx x, rtx oldx ATTRIBUTE_UNUSED, |
3754d046 | 3593 | machine_mode mode ATTRIBUTE_UNUSED) |
644459d0 | 3594 | { |
3595 | rtx op0, op1; | |
3596 | /* Make sure both operands are registers. */ | |
3597 | if (GET_CODE (x) == PLUS) | |
3598 | { | |
3599 | op0 = XEXP (x, 0); | |
3600 | op1 = XEXP (x, 1); | |
3601 | if (ALIGNED_SYMBOL_REF_P (op0)) | |
3602 | { | |
3603 | op0 = force_reg (Pmode, op0); | |
3604 | mark_reg_pointer (op0, 128); | |
3605 | } | |
3606 | else if (GET_CODE (op0) != REG) | |
3607 | op0 = force_reg (Pmode, op0); | |
3608 | if (ALIGNED_SYMBOL_REF_P (op1)) | |
3609 | { | |
3610 | op1 = force_reg (Pmode, op1); | |
3611 | mark_reg_pointer (op1, 128); | |
3612 | } | |
3613 | else if (GET_CODE (op1) != REG) | |
3614 | op1 = force_reg (Pmode, op1); | |
3615 | x = gen_rtx_PLUS (Pmode, op0, op1); | |
644459d0 | 3616 | } |
41e3a0c7 | 3617 | return x; |
644459d0 | 3618 | } |
3619 | ||
6cf5579e | 3620 | /* Like spu_legitimate_address, except with named address support. */ |
3621 | static rtx | |
3754d046 | 3622 | spu_addr_space_legitimize_address (rtx x, rtx oldx, machine_mode mode, |
6cf5579e | 3623 | addr_space_t as) |
3624 | { | |
3625 | if (as != ADDR_SPACE_GENERIC) | |
3626 | return x; | |
3627 | ||
3628 | return spu_legitimize_address (x, oldx, mode); | |
3629 | } | |
3630 | ||
fa695424 | 3631 | /* Reload reg + const_int for out-of-range displacements. */ |
3632 | rtx | |
3754d046 | 3633 | spu_legitimize_reload_address (rtx ad, machine_mode mode ATTRIBUTE_UNUSED, |
fa695424 | 3634 | int opnum, int type) |
3635 | { | |
3636 | bool removed_and = false; | |
3637 | ||
3638 | if (GET_CODE (ad) == AND | |
3639 | && CONST_INT_P (XEXP (ad, 1)) | |
3640 | && INTVAL (XEXP (ad, 1)) == (HOST_WIDE_INT) - 16) | |
3641 | { | |
3642 | ad = XEXP (ad, 0); | |
3643 | removed_and = true; | |
3644 | } | |
3645 | ||
3646 | if (GET_CODE (ad) == PLUS | |
3647 | && REG_P (XEXP (ad, 0)) | |
3648 | && CONST_INT_P (XEXP (ad, 1)) | |
3649 | && !(INTVAL (XEXP (ad, 1)) >= -0x2000 | |
3650 | && INTVAL (XEXP (ad, 1)) <= 0x1fff)) | |
3651 | { | |
3652 | /* Unshare the sum. */ | |
3653 | ad = copy_rtx (ad); | |
3654 | ||
3655 | /* Reload the displacement. */ | |
3656 | push_reload (XEXP (ad, 1), NULL_RTX, &XEXP (ad, 1), NULL, | |
3657 | BASE_REG_CLASS, GET_MODE (ad), VOIDmode, 0, 0, | |
3658 | opnum, (enum reload_type) type); | |
3659 | ||
3660 | /* Add back AND for alignment if we stripped it. */ | |
3661 | if (removed_and) | |
3662 | ad = gen_rtx_AND (GET_MODE (ad), ad, GEN_INT (-16)); | |
3663 | ||
3664 | return ad; | |
3665 | } | |
3666 | ||
3667 | return NULL_RTX; | |
3668 | } | |
3669 | ||
644459d0 | 3670 | /* Handle an attribute requiring a FUNCTION_DECL; arguments as in |
3671 | struct attribute_spec.handler. */ | |
3672 | static tree | |
3673 | spu_handle_fndecl_attribute (tree * node, | |
3674 | tree name, | |
3675 | tree args ATTRIBUTE_UNUSED, | |
3676 | int flags ATTRIBUTE_UNUSED, bool * no_add_attrs) | |
3677 | { | |
3678 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
3679 | { | |
67a779df | 3680 | warning (0, "%qE attribute only applies to functions", |
3681 | name); | |
644459d0 | 3682 | *no_add_attrs = true; |
3683 | } | |
3684 | ||
3685 | return NULL_TREE; | |
3686 | } | |
3687 | ||
3688 | /* Handle the "vector" attribute. */ | |
3689 | static tree | |
3690 | spu_handle_vector_attribute (tree * node, tree name, | |
3691 | tree args ATTRIBUTE_UNUSED, | |
3692 | int flags ATTRIBUTE_UNUSED, bool * no_add_attrs) | |
3693 | { | |
3694 | tree type = *node, result = NULL_TREE; | |
3754d046 | 3695 | machine_mode mode; |
644459d0 | 3696 | int unsigned_p; |
3697 | ||
3698 | while (POINTER_TYPE_P (type) | |
3699 | || TREE_CODE (type) == FUNCTION_TYPE | |
3700 | || TREE_CODE (type) == METHOD_TYPE || TREE_CODE (type) == ARRAY_TYPE) | |
3701 | type = TREE_TYPE (type); | |
3702 | ||
3703 | mode = TYPE_MODE (type); | |
3704 | ||
3705 | unsigned_p = TYPE_UNSIGNED (type); | |
3706 | switch (mode) | |
3707 | { | |
3708 | case DImode: | |
3709 | result = (unsigned_p ? unsigned_V2DI_type_node : V2DI_type_node); | |
3710 | break; | |
3711 | case SImode: | |
3712 | result = (unsigned_p ? unsigned_V4SI_type_node : V4SI_type_node); | |
3713 | break; | |
3714 | case HImode: | |
3715 | result = (unsigned_p ? unsigned_V8HI_type_node : V8HI_type_node); | |
3716 | break; | |
3717 | case QImode: | |
3718 | result = (unsigned_p ? unsigned_V16QI_type_node : V16QI_type_node); | |
3719 | break; | |
3720 | case SFmode: | |
3721 | result = V4SF_type_node; | |
3722 | break; | |
3723 | case DFmode: | |
3724 | result = V2DF_type_node; | |
3725 | break; | |
3726 | default: | |
3727 | break; | |
3728 | } | |
3729 | ||
3730 | /* Propagate qualifiers attached to the element type | |
3731 | onto the vector type. */ | |
3732 | if (result && result != type && TYPE_QUALS (type)) | |
3733 | result = build_qualified_type (result, TYPE_QUALS (type)); | |
3734 | ||
3735 | *no_add_attrs = true; /* No need to hang on to the attribute. */ | |
3736 | ||
3737 | if (!result) | |
67a779df | 3738 | warning (0, "%qE attribute ignored", name); |
644459d0 | 3739 | else |
d991e6e8 | 3740 | *node = lang_hooks.types.reconstruct_complex_type (*node, result); |
644459d0 | 3741 | |
3742 | return NULL_TREE; | |
3743 | } | |
3744 | ||
f2b32076 | 3745 | /* Return nonzero if FUNC is a naked function. */ |
644459d0 | 3746 | static int |
3747 | spu_naked_function_p (tree func) | |
3748 | { | |
3749 | tree a; | |
3750 | ||
3751 | if (TREE_CODE (func) != FUNCTION_DECL) | |
3752 | abort (); | |
3753 | ||
3754 | a = lookup_attribute ("naked", DECL_ATTRIBUTES (func)); | |
3755 | return a != NULL_TREE; | |
3756 | } | |
3757 | ||
3758 | int | |
3759 | spu_initial_elimination_offset (int from, int to) | |
3760 | { | |
3761 | int saved_regs_size = spu_saved_regs_size (); | |
3762 | int sp_offset = 0; | |
d5bf7b64 | 3763 | if (!crtl->is_leaf || crtl->outgoing_args_size |
644459d0 | 3764 | || get_frame_size () || saved_regs_size) |
3765 | sp_offset = STACK_POINTER_OFFSET; | |
3766 | if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM) | |
a8e019fa | 3767 | return get_frame_size () + crtl->outgoing_args_size + sp_offset; |
644459d0 | 3768 | else if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) |
a8e019fa | 3769 | return get_frame_size (); |
644459d0 | 3770 | else if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) |
abe32cce | 3771 | return sp_offset + crtl->outgoing_args_size |
644459d0 | 3772 | + get_frame_size () + saved_regs_size + STACK_POINTER_OFFSET; |
3773 | else if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
3774 | return get_frame_size () + saved_regs_size + sp_offset; | |
a8e019fa | 3775 | else |
3776 | gcc_unreachable (); | |
644459d0 | 3777 | } |
3778 | ||
3779 | rtx | |
fb80456a | 3780 | spu_function_value (const_tree type, const_tree func ATTRIBUTE_UNUSED) |
644459d0 | 3781 | { |
3754d046 | 3782 | machine_mode mode = TYPE_MODE (type); |
644459d0 | 3783 | int byte_size = ((mode == BLKmode) |
3784 | ? int_size_in_bytes (type) : GET_MODE_SIZE (mode)); | |
3785 | ||
3786 | /* Make sure small structs are left justified in a register. */ | |
3787 | if ((mode == BLKmode || (type && AGGREGATE_TYPE_P (type))) | |
3788 | && byte_size <= UNITS_PER_WORD * MAX_REGISTER_RETURN && byte_size > 0) | |
3789 | { | |
3754d046 | 3790 | machine_mode smode; |
644459d0 | 3791 | rtvec v; |
3792 | int i; | |
3793 | int nregs = (byte_size + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
3794 | int n = byte_size / UNITS_PER_WORD; | |
3795 | v = rtvec_alloc (nregs); | |
3796 | for (i = 0; i < n; i++) | |
3797 | { | |
3798 | RTVEC_ELT (v, i) = gen_rtx_EXPR_LIST (VOIDmode, | |
3799 | gen_rtx_REG (TImode, | |
3800 | FIRST_RETURN_REGNUM | |
3801 | + i), | |
3802 | GEN_INT (UNITS_PER_WORD * i)); | |
3803 | byte_size -= UNITS_PER_WORD; | |
3804 | } | |
3805 | ||
3806 | if (n < nregs) | |
3807 | { | |
3808 | if (byte_size < 4) | |
3809 | byte_size = 4; | |
3810 | smode = | |
3811 | smallest_mode_for_size (byte_size * BITS_PER_UNIT, MODE_INT); | |
3812 | RTVEC_ELT (v, n) = | |
3813 | gen_rtx_EXPR_LIST (VOIDmode, | |
3814 | gen_rtx_REG (smode, FIRST_RETURN_REGNUM + n), | |
3815 | GEN_INT (UNITS_PER_WORD * n)); | |
3816 | } | |
3817 | return gen_rtx_PARALLEL (mode, v); | |
3818 | } | |
3819 | return gen_rtx_REG (mode, FIRST_RETURN_REGNUM); | |
3820 | } | |
3821 | ||
ee9034d4 | 3822 | static rtx |
39cba157 | 3823 | spu_function_arg (cumulative_args_t cum_v, |
3754d046 | 3824 | machine_mode mode, |
ee9034d4 | 3825 | const_tree type, bool named ATTRIBUTE_UNUSED) |
644459d0 | 3826 | { |
39cba157 | 3827 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); |
644459d0 | 3828 | int byte_size; |
3829 | ||
a08c5dd0 | 3830 | if (*cum >= MAX_REGISTER_ARGS) |
644459d0 | 3831 | return 0; |
3832 | ||
3833 | byte_size = ((mode == BLKmode) | |
3834 | ? int_size_in_bytes (type) : GET_MODE_SIZE (mode)); | |
3835 | ||
3836 | /* The ABI does not allow parameters to be passed partially in | |
3837 | reg and partially in stack. */ | |
a08c5dd0 | 3838 | if ((*cum + (byte_size + 15) / 16) > MAX_REGISTER_ARGS) |
644459d0 | 3839 | return 0; |
3840 | ||
3841 | /* Make sure small structs are left justified in a register. */ | |
3842 | if ((mode == BLKmode || (type && AGGREGATE_TYPE_P (type))) | |
3843 | && byte_size < UNITS_PER_WORD && byte_size > 0) | |
3844 | { | |
3754d046 | 3845 | machine_mode smode; |
644459d0 | 3846 | rtx gr_reg; |
3847 | if (byte_size < 4) | |
3848 | byte_size = 4; | |
3849 | smode = smallest_mode_for_size (byte_size * BITS_PER_UNIT, MODE_INT); | |
3850 | gr_reg = gen_rtx_EXPR_LIST (VOIDmode, | |
a08c5dd0 | 3851 | gen_rtx_REG (smode, FIRST_ARG_REGNUM + *cum), |
644459d0 | 3852 | const0_rtx); |
3853 | return gen_rtx_PARALLEL (mode, gen_rtvec (1, gr_reg)); | |
3854 | } | |
3855 | else | |
a08c5dd0 | 3856 | return gen_rtx_REG (mode, FIRST_ARG_REGNUM + *cum); |
644459d0 | 3857 | } |
3858 | ||
ee9034d4 | 3859 | static void |
3754d046 | 3860 | spu_function_arg_advance (cumulative_args_t cum_v, machine_mode mode, |
ee9034d4 | 3861 | const_tree type, bool named ATTRIBUTE_UNUSED) |
3862 | { | |
39cba157 | 3863 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); |
3864 | ||
ee9034d4 | 3865 | *cum += (type && TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST |
3866 | ? 1 | |
3867 | : mode == BLKmode | |
3868 | ? ((int_size_in_bytes (type) + 15) / 16) | |
3869 | : mode == VOIDmode | |
3870 | ? 1 | |
3871 | : HARD_REGNO_NREGS (cum, mode)); | |
3872 | } | |
3873 | ||
644459d0 | 3874 | /* Variable sized types are passed by reference. */ |
3875 | static bool | |
39cba157 | 3876 | spu_pass_by_reference (cumulative_args_t cum ATTRIBUTE_UNUSED, |
3754d046 | 3877 | machine_mode mode ATTRIBUTE_UNUSED, |
fb80456a | 3878 | const_tree type, bool named ATTRIBUTE_UNUSED) |
644459d0 | 3879 | { |
3880 | return type && TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST; | |
3881 | } | |
3882 | \f | |
3883 | ||
3884 | /* Var args. */ | |
3885 | ||
3886 | /* Create and return the va_list datatype. | |
3887 | ||
3888 | On SPU, va_list is an array type equivalent to | |
3889 | ||
3890 | typedef struct __va_list_tag | |
3891 | { | |
3892 | void *__args __attribute__((__aligned(16))); | |
3893 | void *__skip __attribute__((__aligned(16))); | |
3894 | ||
3895 | } va_list[1]; | |
3896 | ||
fa7637bd | 3897 | where __args points to the arg that will be returned by the next |
644459d0 | 3898 | va_arg(), and __skip points to the previous stack frame such that |
3899 | when __args == __skip we should advance __args by 32 bytes. */ | |
3900 | static tree | |
3901 | spu_build_builtin_va_list (void) | |
3902 | { | |
3903 | tree f_args, f_skip, record, type_decl; | |
3904 | bool owp; | |
3905 | ||
3906 | record = (*lang_hooks.types.make_type) (RECORD_TYPE); | |
3907 | ||
3908 | type_decl = | |
54e46243 | 3909 | build_decl (BUILTINS_LOCATION, |
3910 | TYPE_DECL, get_identifier ("__va_list_tag"), record); | |
644459d0 | 3911 | |
54e46243 | 3912 | f_args = build_decl (BUILTINS_LOCATION, |
3913 | FIELD_DECL, get_identifier ("__args"), ptr_type_node); | |
3914 | f_skip = build_decl (BUILTINS_LOCATION, | |
3915 | FIELD_DECL, get_identifier ("__skip"), ptr_type_node); | |
644459d0 | 3916 | |
3917 | DECL_FIELD_CONTEXT (f_args) = record; | |
3918 | DECL_ALIGN (f_args) = 128; | |
3919 | DECL_USER_ALIGN (f_args) = 1; | |
3920 | ||
3921 | DECL_FIELD_CONTEXT (f_skip) = record; | |
3922 | DECL_ALIGN (f_skip) = 128; | |
3923 | DECL_USER_ALIGN (f_skip) = 1; | |
3924 | ||
bc907808 | 3925 | TYPE_STUB_DECL (record) = type_decl; |
644459d0 | 3926 | TYPE_NAME (record) = type_decl; |
3927 | TYPE_FIELDS (record) = f_args; | |
1767a056 | 3928 | DECL_CHAIN (f_args) = f_skip; |
644459d0 | 3929 | |
3930 | /* We know this is being padded and we want it too. It is an internal | |
3931 | type so hide the warnings from the user. */ | |
3932 | owp = warn_padded; | |
3933 | warn_padded = false; | |
3934 | ||
3935 | layout_type (record); | |
3936 | ||
3937 | warn_padded = owp; | |
3938 | ||
3939 | /* The correct type is an array type of one element. */ | |
3940 | return build_array_type (record, build_index_type (size_zero_node)); | |
3941 | } | |
3942 | ||
3943 | /* Implement va_start by filling the va_list structure VALIST. | |
3944 | NEXTARG points to the first anonymous stack argument. | |
3945 | ||
3946 | The following global variables are used to initialize | |
3947 | the va_list structure: | |
3948 | ||
abe32cce | 3949 | crtl->args.info; |
644459d0 | 3950 | the CUMULATIVE_ARGS for this function |
3951 | ||
abe32cce | 3952 | crtl->args.arg_offset_rtx: |
644459d0 | 3953 | holds the offset of the first anonymous stack argument |
3954 | (relative to the virtual arg pointer). */ | |
3955 | ||
8a58ed0a | 3956 | static void |
644459d0 | 3957 | spu_va_start (tree valist, rtx nextarg) |
3958 | { | |
3959 | tree f_args, f_skip; | |
3960 | tree args, skip, t; | |
3961 | ||
3962 | f_args = TYPE_FIELDS (TREE_TYPE (va_list_type_node)); | |
1767a056 | 3963 | f_skip = DECL_CHAIN (f_args); |
644459d0 | 3964 | |
170efcd4 | 3965 | valist = build_simple_mem_ref (valist); |
644459d0 | 3966 | args = |
3967 | build3 (COMPONENT_REF, TREE_TYPE (f_args), valist, f_args, NULL_TREE); | |
3968 | skip = | |
3969 | build3 (COMPONENT_REF, TREE_TYPE (f_skip), valist, f_skip, NULL_TREE); | |
3970 | ||
3971 | /* Find the __args area. */ | |
3972 | t = make_tree (TREE_TYPE (args), nextarg); | |
abe32cce | 3973 | if (crtl->args.pretend_args_size > 0) |
2cc66f2a | 3974 | t = fold_build_pointer_plus_hwi (t, -STACK_POINTER_OFFSET); |
75a70cf9 | 3975 | t = build2 (MODIFY_EXPR, TREE_TYPE (args), args, t); |
644459d0 | 3976 | TREE_SIDE_EFFECTS (t) = 1; |
3977 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
3978 | ||
3979 | /* Find the __skip area. */ | |
3980 | t = make_tree (TREE_TYPE (skip), virtual_incoming_args_rtx); | |
2cc66f2a | 3981 | t = fold_build_pointer_plus_hwi (t, (crtl->args.pretend_args_size |
3982 | - STACK_POINTER_OFFSET)); | |
75a70cf9 | 3983 | t = build2 (MODIFY_EXPR, TREE_TYPE (skip), skip, t); |
644459d0 | 3984 | TREE_SIDE_EFFECTS (t) = 1; |
3985 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
3986 | } | |
3987 | ||
3988 | /* Gimplify va_arg by updating the va_list structure | |
3989 | VALIST as required to retrieve an argument of type | |
3990 | TYPE, and returning that argument. | |
3991 | ||
3992 | ret = va_arg(VALIST, TYPE); | |
3993 | ||
3994 | generates code equivalent to: | |
3995 | ||
3996 | paddedsize = (sizeof(TYPE) + 15) & -16; | |
3997 | if (VALIST.__args + paddedsize > VALIST.__skip | |
3998 | && VALIST.__args <= VALIST.__skip) | |
3999 | addr = VALIST.__skip + 32; | |
4000 | else | |
4001 | addr = VALIST.__args; | |
4002 | VALIST.__args = addr + paddedsize; | |
4003 | ret = *(TYPE *)addr; | |
4004 | */ | |
4005 | static tree | |
75a70cf9 | 4006 | spu_gimplify_va_arg_expr (tree valist, tree type, gimple_seq * pre_p, |
4007 | gimple_seq * post_p ATTRIBUTE_UNUSED) | |
644459d0 | 4008 | { |
4009 | tree f_args, f_skip; | |
4010 | tree args, skip; | |
4011 | HOST_WIDE_INT size, rsize; | |
2cc66f2a | 4012 | tree addr, tmp; |
644459d0 | 4013 | bool pass_by_reference_p; |
4014 | ||
4015 | f_args = TYPE_FIELDS (TREE_TYPE (va_list_type_node)); | |
1767a056 | 4016 | f_skip = DECL_CHAIN (f_args); |
644459d0 | 4017 | |
644459d0 | 4018 | args = |
4019 | build3 (COMPONENT_REF, TREE_TYPE (f_args), valist, f_args, NULL_TREE); | |
4020 | skip = | |
4021 | build3 (COMPONENT_REF, TREE_TYPE (f_skip), valist, f_skip, NULL_TREE); | |
4022 | ||
4023 | addr = create_tmp_var (ptr_type_node, "va_arg"); | |
644459d0 | 4024 | |
4025 | /* if an object is dynamically sized, a pointer to it is passed | |
4026 | instead of the object itself. */ | |
27a82950 | 4027 | pass_by_reference_p = pass_by_reference (NULL, TYPE_MODE (type), type, |
4028 | false); | |
644459d0 | 4029 | if (pass_by_reference_p) |
4030 | type = build_pointer_type (type); | |
4031 | size = int_size_in_bytes (type); | |
4032 | rsize = ((size + UNITS_PER_WORD - 1) / UNITS_PER_WORD) * UNITS_PER_WORD; | |
4033 | ||
4034 | /* build conditional expression to calculate addr. The expression | |
4035 | will be gimplified later. */ | |
2cc66f2a | 4036 | tmp = fold_build_pointer_plus_hwi (unshare_expr (args), rsize); |
644459d0 | 4037 | tmp = build2 (TRUTH_AND_EXPR, boolean_type_node, |
75a70cf9 | 4038 | build2 (GT_EXPR, boolean_type_node, tmp, unshare_expr (skip)), |
4039 | build2 (LE_EXPR, boolean_type_node, unshare_expr (args), | |
4040 | unshare_expr (skip))); | |
644459d0 | 4041 | |
4042 | tmp = build3 (COND_EXPR, ptr_type_node, tmp, | |
2cc66f2a | 4043 | fold_build_pointer_plus_hwi (unshare_expr (skip), 32), |
4044 | unshare_expr (args)); | |
644459d0 | 4045 | |
75a70cf9 | 4046 | gimplify_assign (addr, tmp, pre_p); |
644459d0 | 4047 | |
4048 | /* update VALIST.__args */ | |
2cc66f2a | 4049 | tmp = fold_build_pointer_plus_hwi (addr, rsize); |
75a70cf9 | 4050 | gimplify_assign (unshare_expr (args), tmp, pre_p); |
644459d0 | 4051 | |
8115f0af | 4052 | addr = fold_convert (build_pointer_type_for_mode (type, ptr_mode, true), |
4053 | addr); | |
644459d0 | 4054 | |
4055 | if (pass_by_reference_p) | |
4056 | addr = build_va_arg_indirect_ref (addr); | |
4057 | ||
4058 | return build_va_arg_indirect_ref (addr); | |
4059 | } | |
4060 | ||
4061 | /* Save parameter registers starting with the register that corresponds | |
4062 | to the first unnamed parameters. If the first unnamed parameter is | |
4063 | in the stack then save no registers. Set pretend_args_size to the | |
4064 | amount of space needed to save the registers. */ | |
39cba157 | 4065 | static void |
3754d046 | 4066 | spu_setup_incoming_varargs (cumulative_args_t cum, machine_mode mode, |
644459d0 | 4067 | tree type, int *pretend_size, int no_rtl) |
4068 | { | |
4069 | if (!no_rtl) | |
4070 | { | |
4071 | rtx tmp; | |
4072 | int regno; | |
4073 | int offset; | |
39cba157 | 4074 | int ncum = *get_cumulative_args (cum); |
644459d0 | 4075 | |
4076 | /* cum currently points to the last named argument, we want to | |
4077 | start at the next argument. */ | |
39cba157 | 4078 | spu_function_arg_advance (pack_cumulative_args (&ncum), mode, type, true); |
644459d0 | 4079 | |
4080 | offset = -STACK_POINTER_OFFSET; | |
4081 | for (regno = ncum; regno < MAX_REGISTER_ARGS; regno++) | |
4082 | { | |
4083 | tmp = gen_frame_mem (V4SImode, | |
29c05e22 | 4084 | plus_constant (Pmode, virtual_incoming_args_rtx, |
644459d0 | 4085 | offset)); |
4086 | emit_move_insn (tmp, | |
4087 | gen_rtx_REG (V4SImode, FIRST_ARG_REGNUM + regno)); | |
4088 | offset += 16; | |
4089 | } | |
4090 | *pretend_size = offset + STACK_POINTER_OFFSET; | |
4091 | } | |
4092 | } | |
4093 | \f | |
b2d7ede1 | 4094 | static void |
644459d0 | 4095 | spu_conditional_register_usage (void) |
4096 | { | |
4097 | if (flag_pic) | |
4098 | { | |
4099 | fixed_regs[PIC_OFFSET_TABLE_REGNUM] = 1; | |
4100 | call_used_regs[PIC_OFFSET_TABLE_REGNUM] = 1; | |
4101 | } | |
644459d0 | 4102 | } |
4103 | ||
9d98604b | 4104 | /* This is called any time we inspect the alignment of a register for |
4105 | addresses. */ | |
644459d0 | 4106 | static int |
9d98604b | 4107 | reg_aligned_for_addr (rtx x) |
644459d0 | 4108 | { |
9d98604b | 4109 | int regno = |
4110 | REGNO (x) < FIRST_PSEUDO_REGISTER ? ORIGINAL_REGNO (x) : REGNO (x); | |
4111 | return REGNO_POINTER_ALIGN (regno) >= 128; | |
644459d0 | 4112 | } |
4113 | ||
69ced2d6 | 4114 | /* Encode symbol attributes (local vs. global, tls model) of a SYMBOL_REF |
4115 | into its SYMBOL_REF_FLAGS. */ | |
4116 | static void | |
4117 | spu_encode_section_info (tree decl, rtx rtl, int first) | |
4118 | { | |
4119 | default_encode_section_info (decl, rtl, first); | |
4120 | ||
4121 | /* If a variable has a forced alignment to < 16 bytes, mark it with | |
4122 | SYMBOL_FLAG_ALIGN1. */ | |
4123 | if (TREE_CODE (decl) == VAR_DECL | |
4124 | && DECL_USER_ALIGN (decl) && DECL_ALIGN (decl) < 128) | |
4125 | SYMBOL_REF_FLAGS (XEXP (rtl, 0)) |= SYMBOL_FLAG_ALIGN1; | |
4126 | } | |
4127 | ||
644459d0 | 4128 | /* Return TRUE if we are certain the mem refers to a complete object |
4129 | which is both 16-byte aligned and padded to a 16-byte boundary. This | |
4130 | would make it safe to store with a single instruction. | |
4131 | We guarantee the alignment and padding for static objects by aligning | |
4132 | all of them to 16-bytes. (DATA_ALIGNMENT and CONSTANT_ALIGNMENT.) | |
4133 | FIXME: We currently cannot guarantee this for objects on the stack | |
4134 | because assign_parm_setup_stack calls assign_stack_local with the | |
4135 | alignment of the parameter mode and in that case the alignment never | |
4136 | gets adjusted by LOCAL_ALIGNMENT. */ | |
4137 | static int | |
4138 | store_with_one_insn_p (rtx mem) | |
4139 | { | |
3754d046 | 4140 | machine_mode mode = GET_MODE (mem); |
644459d0 | 4141 | rtx addr = XEXP (mem, 0); |
9d98604b | 4142 | if (mode == BLKmode) |
644459d0 | 4143 | return 0; |
9d98604b | 4144 | if (GET_MODE_SIZE (mode) >= 16) |
4145 | return 1; | |
644459d0 | 4146 | /* Only static objects. */ |
4147 | if (GET_CODE (addr) == SYMBOL_REF) | |
4148 | { | |
4149 | /* We use the associated declaration to make sure the access is | |
fa7637bd | 4150 | referring to the whole object. |
851d9296 | 4151 | We check both MEM_EXPR and SYMBOL_REF_DECL. I'm not sure |
644459d0 | 4152 | if it is necessary. Will there be cases where one exists, and |
4153 | the other does not? Will there be cases where both exist, but | |
4154 | have different types? */ | |
4155 | tree decl = MEM_EXPR (mem); | |
4156 | if (decl | |
4157 | && TREE_CODE (decl) == VAR_DECL | |
4158 | && GET_MODE (mem) == TYPE_MODE (TREE_TYPE (decl))) | |
4159 | return 1; | |
4160 | decl = SYMBOL_REF_DECL (addr); | |
4161 | if (decl | |
4162 | && TREE_CODE (decl) == VAR_DECL | |
4163 | && GET_MODE (mem) == TYPE_MODE (TREE_TYPE (decl))) | |
4164 | return 1; | |
4165 | } | |
4166 | return 0; | |
4167 | } | |
4168 | ||
9d98604b | 4169 | /* Return 1 when the address is not valid for a simple load and store as |
4170 | required by the '_mov*' patterns. We could make this less strict | |
4171 | for loads, but we prefer mem's to look the same so they are more | |
4172 | likely to be merged. */ | |
4173 | static int | |
4174 | address_needs_split (rtx mem) | |
4175 | { | |
4176 | if (GET_MODE_SIZE (GET_MODE (mem)) < 16 | |
4177 | && (GET_MODE_SIZE (GET_MODE (mem)) < 4 | |
4178 | || !(store_with_one_insn_p (mem) | |
4179 | || mem_is_padded_component_ref (mem)))) | |
4180 | return 1; | |
4181 | ||
4182 | return 0; | |
4183 | } | |
4184 | ||
6cf5579e | 4185 | static GTY(()) rtx cache_fetch; /* __cache_fetch function */ |
4186 | static GTY(()) rtx cache_fetch_dirty; /* __cache_fetch_dirty function */ | |
4187 | static alias_set_type ea_alias_set = -1; /* alias set for __ea memory */ | |
4188 | ||
4189 | /* MEM is known to be an __ea qualified memory access. Emit a call to | |
4190 | fetch the ppu memory to local store, and return its address in local | |
4191 | store. */ | |
4192 | ||
4193 | static void | |
4194 | ea_load_store (rtx mem, bool is_store, rtx ea_addr, rtx data_addr) | |
4195 | { | |
4196 | if (is_store) | |
4197 | { | |
4198 | rtx ndirty = GEN_INT (GET_MODE_SIZE (GET_MODE (mem))); | |
4199 | if (!cache_fetch_dirty) | |
4200 | cache_fetch_dirty = init_one_libfunc ("__cache_fetch_dirty"); | |
4201 | emit_library_call_value (cache_fetch_dirty, data_addr, LCT_NORMAL, Pmode, | |
4202 | 2, ea_addr, EAmode, ndirty, SImode); | |
4203 | } | |
4204 | else | |
4205 | { | |
4206 | if (!cache_fetch) | |
4207 | cache_fetch = init_one_libfunc ("__cache_fetch"); | |
4208 | emit_library_call_value (cache_fetch, data_addr, LCT_NORMAL, Pmode, | |
4209 | 1, ea_addr, EAmode); | |
4210 | } | |
4211 | } | |
4212 | ||
4213 | /* Like ea_load_store, but do the cache tag comparison and, for stores, | |
4214 | dirty bit marking, inline. | |
4215 | ||
4216 | The cache control data structure is an array of | |
4217 | ||
4218 | struct __cache_tag_array | |
4219 | { | |
4220 | unsigned int tag_lo[4]; | |
4221 | unsigned int tag_hi[4]; | |
4222 | void *data_pointer[4]; | |
4223 | int reserved[4]; | |
4224 | vector unsigned short dirty_bits[4]; | |
4225 | } */ | |
4226 | ||
4227 | static void | |
4228 | ea_load_store_inline (rtx mem, bool is_store, rtx ea_addr, rtx data_addr) | |
4229 | { | |
4230 | rtx ea_addr_si; | |
4231 | HOST_WIDE_INT v; | |
4232 | rtx tag_size_sym = gen_rtx_SYMBOL_REF (Pmode, "__cache_tag_array_size"); | |
4233 | rtx tag_arr_sym = gen_rtx_SYMBOL_REF (Pmode, "__cache_tag_array"); | |
4234 | rtx index_mask = gen_reg_rtx (SImode); | |
4235 | rtx tag_arr = gen_reg_rtx (Pmode); | |
4236 | rtx splat_mask = gen_reg_rtx (TImode); | |
4237 | rtx splat = gen_reg_rtx (V4SImode); | |
4238 | rtx splat_hi = NULL_RTX; | |
4239 | rtx tag_index = gen_reg_rtx (Pmode); | |
4240 | rtx block_off = gen_reg_rtx (SImode); | |
4241 | rtx tag_addr = gen_reg_rtx (Pmode); | |
4242 | rtx tag = gen_reg_rtx (V4SImode); | |
4243 | rtx cache_tag = gen_reg_rtx (V4SImode); | |
4244 | rtx cache_tag_hi = NULL_RTX; | |
4245 | rtx cache_ptrs = gen_reg_rtx (TImode); | |
4246 | rtx cache_ptrs_si = gen_reg_rtx (SImode); | |
4247 | rtx tag_equal = gen_reg_rtx (V4SImode); | |
4248 | rtx tag_equal_hi = NULL_RTX; | |
4249 | rtx tag_eq_pack = gen_reg_rtx (V4SImode); | |
4250 | rtx tag_eq_pack_si = gen_reg_rtx (SImode); | |
4251 | rtx eq_index = gen_reg_rtx (SImode); | |
0af56f80 | 4252 | rtx bcomp, hit_label, hit_ref, cont_label; |
4253 | rtx_insn *insn; | |
6cf5579e | 4254 | |
4255 | if (spu_ea_model != 32) | |
4256 | { | |
4257 | splat_hi = gen_reg_rtx (V4SImode); | |
4258 | cache_tag_hi = gen_reg_rtx (V4SImode); | |
4259 | tag_equal_hi = gen_reg_rtx (V4SImode); | |
4260 | } | |
4261 | ||
29c05e22 | 4262 | emit_move_insn (index_mask, plus_constant (Pmode, tag_size_sym, -128)); |
6cf5579e | 4263 | emit_move_insn (tag_arr, tag_arr_sym); |
4264 | v = 0x0001020300010203LL; | |
4265 | emit_move_insn (splat_mask, immed_double_const (v, v, TImode)); | |
4266 | ea_addr_si = ea_addr; | |
4267 | if (spu_ea_model != 32) | |
4268 | ea_addr_si = convert_to_mode (SImode, ea_addr, 1); | |
4269 | ||
4270 | /* tag_index = ea_addr & (tag_array_size - 128) */ | |
4271 | emit_insn (gen_andsi3 (tag_index, ea_addr_si, index_mask)); | |
4272 | ||
4273 | /* splat ea_addr to all 4 slots. */ | |
4274 | emit_insn (gen_shufb (splat, ea_addr_si, ea_addr_si, splat_mask)); | |
4275 | /* Similarly for high 32 bits of ea_addr. */ | |
4276 | if (spu_ea_model != 32) | |
4277 | emit_insn (gen_shufb (splat_hi, ea_addr, ea_addr, splat_mask)); | |
4278 | ||
4279 | /* block_off = ea_addr & 127 */ | |
4280 | emit_insn (gen_andsi3 (block_off, ea_addr_si, spu_const (SImode, 127))); | |
4281 | ||
4282 | /* tag_addr = tag_arr + tag_index */ | |
4283 | emit_insn (gen_addsi3 (tag_addr, tag_arr, tag_index)); | |
4284 | ||
4285 | /* Read cache tags. */ | |
4286 | emit_move_insn (cache_tag, gen_rtx_MEM (V4SImode, tag_addr)); | |
4287 | if (spu_ea_model != 32) | |
4288 | emit_move_insn (cache_tag_hi, gen_rtx_MEM (V4SImode, | |
29c05e22 | 4289 | plus_constant (Pmode, |
4290 | tag_addr, 16))); | |
6cf5579e | 4291 | |
4292 | /* tag = ea_addr & -128 */ | |
4293 | emit_insn (gen_andv4si3 (tag, splat, spu_const (V4SImode, -128))); | |
4294 | ||
4295 | /* Read all four cache data pointers. */ | |
4296 | emit_move_insn (cache_ptrs, gen_rtx_MEM (TImode, | |
29c05e22 | 4297 | plus_constant (Pmode, |
4298 | tag_addr, 32))); | |
6cf5579e | 4299 | |
4300 | /* Compare tags. */ | |
4301 | emit_insn (gen_ceq_v4si (tag_equal, tag, cache_tag)); | |
4302 | if (spu_ea_model != 32) | |
4303 | { | |
4304 | emit_insn (gen_ceq_v4si (tag_equal_hi, splat_hi, cache_tag_hi)); | |
4305 | emit_insn (gen_andv4si3 (tag_equal, tag_equal, tag_equal_hi)); | |
4306 | } | |
4307 | ||
4308 | /* At most one of the tags compare equal, so tag_equal has one | |
4309 | 32-bit slot set to all 1's, with the other slots all zero. | |
4310 | gbb picks off low bit from each byte in the 128-bit registers, | |
4311 | so tag_eq_pack is one of 0xf000, 0x0f00, 0x00f0, 0x000f, assuming | |
4312 | we have a hit. */ | |
4313 | emit_insn (gen_spu_gbb (tag_eq_pack, spu_gen_subreg (V16QImode, tag_equal))); | |
4314 | emit_insn (gen_spu_convert (tag_eq_pack_si, tag_eq_pack)); | |
4315 | ||
4316 | /* So counting leading zeros will set eq_index to 16, 20, 24 or 28. */ | |
4317 | emit_insn (gen_clzsi2 (eq_index, tag_eq_pack_si)); | |
4318 | ||
4319 | /* Allowing us to rotate the corresponding cache data pointer to slot0. | |
4320 | (rotating eq_index mod 16 bytes). */ | |
4321 | emit_insn (gen_rotqby_ti (cache_ptrs, cache_ptrs, eq_index)); | |
4322 | emit_insn (gen_spu_convert (cache_ptrs_si, cache_ptrs)); | |
4323 | ||
4324 | /* Add block offset to form final data address. */ | |
4325 | emit_insn (gen_addsi3 (data_addr, cache_ptrs_si, block_off)); | |
4326 | ||
4327 | /* Check that we did hit. */ | |
4328 | hit_label = gen_label_rtx (); | |
4329 | hit_ref = gen_rtx_LABEL_REF (VOIDmode, hit_label); | |
4330 | bcomp = gen_rtx_NE (SImode, tag_eq_pack_si, const0_rtx); | |
d1f9b275 | 4331 | insn = emit_jump_insn (gen_rtx_SET (pc_rtx, |
6cf5579e | 4332 | gen_rtx_IF_THEN_ELSE (VOIDmode, bcomp, |
4333 | hit_ref, pc_rtx))); | |
4334 | /* Say that this branch is very likely to happen. */ | |
4335 | v = REG_BR_PROB_BASE - REG_BR_PROB_BASE / 100 - 1; | |
9eb946de | 4336 | add_int_reg_note (insn, REG_BR_PROB, v); |
6cf5579e | 4337 | |
4338 | ea_load_store (mem, is_store, ea_addr, data_addr); | |
4339 | cont_label = gen_label_rtx (); | |
4340 | emit_jump_insn (gen_jump (cont_label)); | |
4341 | emit_barrier (); | |
4342 | ||
4343 | emit_label (hit_label); | |
4344 | ||
4345 | if (is_store) | |
4346 | { | |
4347 | HOST_WIDE_INT v_hi; | |
4348 | rtx dirty_bits = gen_reg_rtx (TImode); | |
4349 | rtx dirty_off = gen_reg_rtx (SImode); | |
4350 | rtx dirty_128 = gen_reg_rtx (TImode); | |
4351 | rtx neg_block_off = gen_reg_rtx (SImode); | |
4352 | ||
4353 | /* Set up mask with one dirty bit per byte of the mem we are | |
4354 | writing, starting from top bit. */ | |
4355 | v_hi = v = -1; | |
4356 | v <<= (128 - GET_MODE_SIZE (GET_MODE (mem))) & 63; | |
4357 | if ((128 - GET_MODE_SIZE (GET_MODE (mem))) >= 64) | |
4358 | { | |
4359 | v_hi = v; | |
4360 | v = 0; | |
4361 | } | |
4362 | emit_move_insn (dirty_bits, immed_double_const (v, v_hi, TImode)); | |
4363 | ||
4364 | /* Form index into cache dirty_bits. eq_index is one of | |
4365 | 0x10, 0x14, 0x18 or 0x1c. Multiplying by 4 gives us | |
4366 | 0x40, 0x50, 0x60 or 0x70 which just happens to be the | |
4367 | offset to each of the four dirty_bits elements. */ | |
4368 | emit_insn (gen_ashlsi3 (dirty_off, eq_index, spu_const (SImode, 2))); | |
4369 | ||
4370 | emit_insn (gen_spu_lqx (dirty_128, tag_addr, dirty_off)); | |
4371 | ||
4372 | /* Rotate bit mask to proper bit. */ | |
4373 | emit_insn (gen_negsi2 (neg_block_off, block_off)); | |
4374 | emit_insn (gen_rotqbybi_ti (dirty_bits, dirty_bits, neg_block_off)); | |
4375 | emit_insn (gen_rotqbi_ti (dirty_bits, dirty_bits, neg_block_off)); | |
4376 | ||
4377 | /* Or in the new dirty bits. */ | |
4378 | emit_insn (gen_iorti3 (dirty_128, dirty_bits, dirty_128)); | |
4379 | ||
4380 | /* Store. */ | |
4381 | emit_insn (gen_spu_stqx (dirty_128, tag_addr, dirty_off)); | |
4382 | } | |
4383 | ||
4384 | emit_label (cont_label); | |
4385 | } | |
4386 | ||
4387 | static rtx | |
4388 | expand_ea_mem (rtx mem, bool is_store) | |
4389 | { | |
4390 | rtx ea_addr; | |
4391 | rtx data_addr = gen_reg_rtx (Pmode); | |
4392 | rtx new_mem; | |
4393 | ||
4394 | ea_addr = force_reg (EAmode, XEXP (mem, 0)); | |
4395 | if (optimize_size || optimize == 0) | |
4396 | ea_load_store (mem, is_store, ea_addr, data_addr); | |
4397 | else | |
4398 | ea_load_store_inline (mem, is_store, ea_addr, data_addr); | |
4399 | ||
4400 | if (ea_alias_set == -1) | |
4401 | ea_alias_set = new_alias_set (); | |
4402 | ||
4403 | /* We generate a new MEM RTX to refer to the copy of the data | |
4404 | in the cache. We do not copy memory attributes (except the | |
4405 | alignment) from the original MEM, as they may no longer apply | |
4406 | to the cache copy. */ | |
4407 | new_mem = gen_rtx_MEM (GET_MODE (mem), data_addr); | |
4408 | set_mem_alias_set (new_mem, ea_alias_set); | |
4409 | set_mem_align (new_mem, MIN (MEM_ALIGN (mem), 128 * 8)); | |
4410 | ||
4411 | return new_mem; | |
4412 | } | |
4413 | ||
644459d0 | 4414 | int |
3754d046 | 4415 | spu_expand_mov (rtx * ops, machine_mode mode) |
644459d0 | 4416 | { |
4417 | if (GET_CODE (ops[0]) == SUBREG && !valid_subreg (ops[0])) | |
abe960bb | 4418 | { |
4419 | /* Perform the move in the destination SUBREG's inner mode. */ | |
4420 | ops[0] = SUBREG_REG (ops[0]); | |
4421 | mode = GET_MODE (ops[0]); | |
4422 | ops[1] = gen_lowpart_common (mode, ops[1]); | |
4423 | gcc_assert (ops[1]); | |
4424 | } | |
644459d0 | 4425 | |
4426 | if (GET_CODE (ops[1]) == SUBREG && !valid_subreg (ops[1])) | |
4427 | { | |
4428 | rtx from = SUBREG_REG (ops[1]); | |
3754d046 | 4429 | machine_mode imode = int_mode_for_mode (GET_MODE (from)); |
644459d0 | 4430 | |
4431 | gcc_assert (GET_MODE_CLASS (mode) == MODE_INT | |
4432 | && GET_MODE_CLASS (imode) == MODE_INT | |
4433 | && subreg_lowpart_p (ops[1])); | |
4434 | ||
4435 | if (GET_MODE_SIZE (imode) < 4) | |
8d72495d | 4436 | imode = SImode; |
4437 | if (imode != GET_MODE (from)) | |
4438 | from = gen_rtx_SUBREG (imode, from, 0); | |
644459d0 | 4439 | |
4440 | if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (imode)) | |
4441 | { | |
d6bf3b14 | 4442 | enum insn_code icode = convert_optab_handler (trunc_optab, |
4443 | mode, imode); | |
644459d0 | 4444 | emit_insn (GEN_FCN (icode) (ops[0], from)); |
4445 | } | |
4446 | else | |
4447 | emit_insn (gen_extend_insn (ops[0], from, mode, imode, 1)); | |
4448 | return 1; | |
4449 | } | |
4450 | ||
4451 | /* At least one of the operands needs to be a register. */ | |
4452 | if ((reload_in_progress | reload_completed) == 0 | |
4453 | && !register_operand (ops[0], mode) && !register_operand (ops[1], mode)) | |
4454 | { | |
4455 | rtx temp = force_reg (mode, ops[1]); | |
4456 | emit_move_insn (ops[0], temp); | |
4457 | return 1; | |
4458 | } | |
4459 | if (reload_in_progress || reload_completed) | |
4460 | { | |
dea01258 | 4461 | if (CONSTANT_P (ops[1])) |
4462 | return spu_split_immediate (ops); | |
644459d0 | 4463 | return 0; |
4464 | } | |
9d98604b | 4465 | |
4466 | /* Catch the SImode immediates greater than 0x7fffffff, and sign | |
4467 | extend them. */ | |
4468 | if (GET_CODE (ops[1]) == CONST_INT) | |
644459d0 | 4469 | { |
9d98604b | 4470 | HOST_WIDE_INT val = trunc_int_for_mode (INTVAL (ops[1]), mode); |
4471 | if (val != INTVAL (ops[1])) | |
644459d0 | 4472 | { |
9d98604b | 4473 | emit_move_insn (ops[0], GEN_INT (val)); |
4474 | return 1; | |
644459d0 | 4475 | } |
4476 | } | |
9d98604b | 4477 | if (MEM_P (ops[0])) |
6cf5579e | 4478 | { |
4479 | if (MEM_ADDR_SPACE (ops[0])) | |
4480 | ops[0] = expand_ea_mem (ops[0], true); | |
4481 | return spu_split_store (ops); | |
4482 | } | |
9d98604b | 4483 | if (MEM_P (ops[1])) |
6cf5579e | 4484 | { |
4485 | if (MEM_ADDR_SPACE (ops[1])) | |
4486 | ops[1] = expand_ea_mem (ops[1], false); | |
4487 | return spu_split_load (ops); | |
4488 | } | |
9d98604b | 4489 | |
644459d0 | 4490 | return 0; |
4491 | } | |
4492 | ||
9d98604b | 4493 | static void |
4494 | spu_convert_move (rtx dst, rtx src) | |
644459d0 | 4495 | { |
3754d046 | 4496 | machine_mode mode = GET_MODE (dst); |
4497 | machine_mode int_mode = mode_for_size (GET_MODE_BITSIZE (mode), MODE_INT, 0); | |
9d98604b | 4498 | rtx reg; |
4499 | gcc_assert (GET_MODE (src) == TImode); | |
4500 | reg = int_mode != mode ? gen_reg_rtx (int_mode) : dst; | |
d1f9b275 | 4501 | emit_insn (gen_rtx_SET (reg, |
9d98604b | 4502 | gen_rtx_TRUNCATE (int_mode, |
4503 | gen_rtx_LSHIFTRT (TImode, src, | |
4504 | GEN_INT (int_mode == DImode ? 64 : 96))))); | |
4505 | if (int_mode != mode) | |
4506 | { | |
4507 | reg = simplify_gen_subreg (mode, reg, int_mode, 0); | |
4508 | emit_move_insn (dst, reg); | |
4509 | } | |
4510 | } | |
644459d0 | 4511 | |
9d98604b | 4512 | /* Load TImode values into DST0 and DST1 (when it is non-NULL) using |
4513 | the address from SRC and SRC+16. Return a REG or CONST_INT that | |
4514 | specifies how many bytes to rotate the loaded registers, plus any | |
4515 | extra from EXTRA_ROTQBY. The address and rotate amounts are | |
4516 | normalized to improve merging of loads and rotate computations. */ | |
4517 | static rtx | |
4518 | spu_expand_load (rtx dst0, rtx dst1, rtx src, int extra_rotby) | |
4519 | { | |
4520 | rtx addr = XEXP (src, 0); | |
4521 | rtx p0, p1, rot, addr0, addr1; | |
4522 | int rot_amt; | |
644459d0 | 4523 | |
4524 | rot = 0; | |
4525 | rot_amt = 0; | |
9d98604b | 4526 | |
4527 | if (MEM_ALIGN (src) >= 128) | |
4528 | /* Address is already aligned; simply perform a TImode load. */ ; | |
4529 | else if (GET_CODE (addr) == PLUS) | |
644459d0 | 4530 | { |
4531 | /* 8 cases: | |
4532 | aligned reg + aligned reg => lqx | |
4533 | aligned reg + unaligned reg => lqx, rotqby | |
4534 | aligned reg + aligned const => lqd | |
4535 | aligned reg + unaligned const => lqd, rotqbyi | |
4536 | unaligned reg + aligned reg => lqx, rotqby | |
4537 | unaligned reg + unaligned reg => lqx, a, rotqby (1 scratch) | |
4538 | unaligned reg + aligned const => lqd, rotqby | |
4539 | unaligned reg + unaligned const -> not allowed by legitimate address | |
4540 | */ | |
4541 | p0 = XEXP (addr, 0); | |
4542 | p1 = XEXP (addr, 1); | |
9d98604b | 4543 | if (!reg_aligned_for_addr (p0)) |
644459d0 | 4544 | { |
9d98604b | 4545 | if (REG_P (p1) && !reg_aligned_for_addr (p1)) |
644459d0 | 4546 | { |
9d98604b | 4547 | rot = gen_reg_rtx (SImode); |
4548 | emit_insn (gen_addsi3 (rot, p0, p1)); | |
4549 | } | |
4550 | else if (GET_CODE (p1) == CONST_INT && (INTVAL (p1) & 15)) | |
4551 | { | |
4552 | if (INTVAL (p1) > 0 | |
4553 | && REG_POINTER (p0) | |
4554 | && INTVAL (p1) * BITS_PER_UNIT | |
4555 | < REGNO_POINTER_ALIGN (REGNO (p0))) | |
4556 | { | |
4557 | rot = gen_reg_rtx (SImode); | |
4558 | emit_insn (gen_addsi3 (rot, p0, p1)); | |
4559 | addr = p0; | |
4560 | } | |
4561 | else | |
4562 | { | |
4563 | rtx x = gen_reg_rtx (SImode); | |
4564 | emit_move_insn (x, p1); | |
4565 | if (!spu_arith_operand (p1, SImode)) | |
4566 | p1 = x; | |
4567 | rot = gen_reg_rtx (SImode); | |
4568 | emit_insn (gen_addsi3 (rot, p0, p1)); | |
4569 | addr = gen_rtx_PLUS (Pmode, p0, x); | |
4570 | } | |
644459d0 | 4571 | } |
4572 | else | |
4573 | rot = p0; | |
4574 | } | |
4575 | else | |
4576 | { | |
4577 | if (GET_CODE (p1) == CONST_INT && (INTVAL (p1) & 15)) | |
4578 | { | |
4579 | rot_amt = INTVAL (p1) & 15; | |
9d98604b | 4580 | if (INTVAL (p1) & -16) |
4581 | { | |
4582 | p1 = GEN_INT (INTVAL (p1) & -16); | |
4583 | addr = gen_rtx_PLUS (SImode, p0, p1); | |
4584 | } | |
4585 | else | |
4586 | addr = p0; | |
644459d0 | 4587 | } |
9d98604b | 4588 | else if (REG_P (p1) && !reg_aligned_for_addr (p1)) |
644459d0 | 4589 | rot = p1; |
4590 | } | |
4591 | } | |
9d98604b | 4592 | else if (REG_P (addr)) |
644459d0 | 4593 | { |
9d98604b | 4594 | if (!reg_aligned_for_addr (addr)) |
644459d0 | 4595 | rot = addr; |
4596 | } | |
4597 | else if (GET_CODE (addr) == CONST) | |
4598 | { | |
4599 | if (GET_CODE (XEXP (addr, 0)) == PLUS | |
4600 | && ALIGNED_SYMBOL_REF_P (XEXP (XEXP (addr, 0), 0)) | |
4601 | && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT) | |
4602 | { | |
4603 | rot_amt = INTVAL (XEXP (XEXP (addr, 0), 1)); | |
4604 | if (rot_amt & -16) | |
4605 | addr = gen_rtx_CONST (Pmode, | |
4606 | gen_rtx_PLUS (Pmode, | |
4607 | XEXP (XEXP (addr, 0), 0), | |
4608 | GEN_INT (rot_amt & -16))); | |
4609 | else | |
4610 | addr = XEXP (XEXP (addr, 0), 0); | |
4611 | } | |
4612 | else | |
9d98604b | 4613 | { |
4614 | rot = gen_reg_rtx (Pmode); | |
4615 | emit_move_insn (rot, addr); | |
4616 | } | |
644459d0 | 4617 | } |
4618 | else if (GET_CODE (addr) == CONST_INT) | |
4619 | { | |
4620 | rot_amt = INTVAL (addr); | |
4621 | addr = GEN_INT (rot_amt & -16); | |
4622 | } | |
4623 | else if (!ALIGNED_SYMBOL_REF_P (addr)) | |
9d98604b | 4624 | { |
4625 | rot = gen_reg_rtx (Pmode); | |
4626 | emit_move_insn (rot, addr); | |
4627 | } | |
644459d0 | 4628 | |
9d98604b | 4629 | rot_amt += extra_rotby; |
644459d0 | 4630 | |
4631 | rot_amt &= 15; | |
4632 | ||
4633 | if (rot && rot_amt) | |
4634 | { | |
9d98604b | 4635 | rtx x = gen_reg_rtx (SImode); |
4636 | emit_insn (gen_addsi3 (x, rot, GEN_INT (rot_amt))); | |
4637 | rot = x; | |
644459d0 | 4638 | rot_amt = 0; |
4639 | } | |
9d98604b | 4640 | if (!rot && rot_amt) |
4641 | rot = GEN_INT (rot_amt); | |
4642 | ||
4643 | addr0 = copy_rtx (addr); | |
4644 | addr0 = gen_rtx_AND (SImode, copy_rtx (addr), GEN_INT (-16)); | |
4645 | emit_insn (gen__movti (dst0, change_address (src, TImode, addr0))); | |
4646 | ||
4647 | if (dst1) | |
4648 | { | |
29c05e22 | 4649 | addr1 = plus_constant (SImode, copy_rtx (addr), 16); |
9d98604b | 4650 | addr1 = gen_rtx_AND (SImode, addr1, GEN_INT (-16)); |
4651 | emit_insn (gen__movti (dst1, change_address (src, TImode, addr1))); | |
4652 | } | |
644459d0 | 4653 | |
9d98604b | 4654 | return rot; |
4655 | } | |
4656 | ||
4657 | int | |
4658 | spu_split_load (rtx * ops) | |
4659 | { | |
3754d046 | 4660 | machine_mode mode = GET_MODE (ops[0]); |
9d98604b | 4661 | rtx addr, load, rot; |
4662 | int rot_amt; | |
644459d0 | 4663 | |
9d98604b | 4664 | if (GET_MODE_SIZE (mode) >= 16) |
4665 | return 0; | |
644459d0 | 4666 | |
9d98604b | 4667 | addr = XEXP (ops[1], 0); |
4668 | gcc_assert (GET_CODE (addr) != AND); | |
4669 | ||
4670 | if (!address_needs_split (ops[1])) | |
4671 | { | |
4672 | ops[1] = change_address (ops[1], TImode, addr); | |
4673 | load = gen_reg_rtx (TImode); | |
4674 | emit_insn (gen__movti (load, ops[1])); | |
4675 | spu_convert_move (ops[0], load); | |
4676 | return 1; | |
4677 | } | |
4678 | ||
4679 | rot_amt = GET_MODE_SIZE (mode) < 4 ? GET_MODE_SIZE (mode) - 4 : 0; | |
4680 | ||
4681 | load = gen_reg_rtx (TImode); | |
4682 | rot = spu_expand_load (load, 0, ops[1], rot_amt); | |
644459d0 | 4683 | |
4684 | if (rot) | |
4685 | emit_insn (gen_rotqby_ti (load, load, rot)); | |
644459d0 | 4686 | |
9d98604b | 4687 | spu_convert_move (ops[0], load); |
4688 | return 1; | |
644459d0 | 4689 | } |
4690 | ||
9d98604b | 4691 | int |
644459d0 | 4692 | spu_split_store (rtx * ops) |
4693 | { | |
3754d046 | 4694 | machine_mode mode = GET_MODE (ops[0]); |
9d98604b | 4695 | rtx reg; |
644459d0 | 4696 | rtx addr, p0, p1, p1_lo, smem; |
4697 | int aform; | |
4698 | int scalar; | |
4699 | ||
9d98604b | 4700 | if (GET_MODE_SIZE (mode) >= 16) |
4701 | return 0; | |
4702 | ||
644459d0 | 4703 | addr = XEXP (ops[0], 0); |
9d98604b | 4704 | gcc_assert (GET_CODE (addr) != AND); |
4705 | ||
4706 | if (!address_needs_split (ops[0])) | |
4707 | { | |
4708 | reg = gen_reg_rtx (TImode); | |
4709 | emit_insn (gen_spu_convert (reg, ops[1])); | |
4710 | ops[0] = change_address (ops[0], TImode, addr); | |
4711 | emit_move_insn (ops[0], reg); | |
4712 | return 1; | |
4713 | } | |
644459d0 | 4714 | |
4715 | if (GET_CODE (addr) == PLUS) | |
4716 | { | |
4717 | /* 8 cases: | |
4718 | aligned reg + aligned reg => lqx, c?x, shuf, stqx | |
4719 | aligned reg + unaligned reg => lqx, c?x, shuf, stqx | |
4720 | aligned reg + aligned const => lqd, c?d, shuf, stqx | |
4721 | aligned reg + unaligned const => lqd, c?d, shuf, stqx | |
4722 | unaligned reg + aligned reg => lqx, c?x, shuf, stqx | |
4723 | unaligned reg + unaligned reg => lqx, c?x, shuf, stqx | |
4724 | unaligned reg + aligned const => lqd, c?d, shuf, stqx | |
9d98604b | 4725 | unaligned reg + unaligned const -> lqx, c?d, shuf, stqx |
644459d0 | 4726 | */ |
4727 | aform = 0; | |
4728 | p0 = XEXP (addr, 0); | |
4729 | p1 = p1_lo = XEXP (addr, 1); | |
9d98604b | 4730 | if (REG_P (p0) && GET_CODE (p1) == CONST_INT) |
644459d0 | 4731 | { |
4732 | p1_lo = GEN_INT (INTVAL (p1) & 15); | |
9d98604b | 4733 | if (reg_aligned_for_addr (p0)) |
4734 | { | |
4735 | p1 = GEN_INT (INTVAL (p1) & -16); | |
4736 | if (p1 == const0_rtx) | |
4737 | addr = p0; | |
4738 | else | |
4739 | addr = gen_rtx_PLUS (SImode, p0, p1); | |
4740 | } | |
4741 | else | |
4742 | { | |
4743 | rtx x = gen_reg_rtx (SImode); | |
4744 | emit_move_insn (x, p1); | |
4745 | addr = gen_rtx_PLUS (SImode, p0, x); | |
4746 | } | |
644459d0 | 4747 | } |
4748 | } | |
9d98604b | 4749 | else if (REG_P (addr)) |
644459d0 | 4750 | { |
4751 | aform = 0; | |
4752 | p0 = addr; | |
4753 | p1 = p1_lo = const0_rtx; | |
4754 | } | |
4755 | else | |
4756 | { | |
4757 | aform = 1; | |
4758 | p0 = gen_rtx_REG (SImode, STACK_POINTER_REGNUM); | |
4759 | p1 = 0; /* aform doesn't use p1 */ | |
4760 | p1_lo = addr; | |
4761 | if (ALIGNED_SYMBOL_REF_P (addr)) | |
4762 | p1_lo = const0_rtx; | |
9d98604b | 4763 | else if (GET_CODE (addr) == CONST |
4764 | && GET_CODE (XEXP (addr, 0)) == PLUS | |
4765 | && ALIGNED_SYMBOL_REF_P (XEXP (XEXP (addr, 0), 0)) | |
4766 | && GET_CODE (XEXP (XEXP (addr, 0), 1)) == CONST_INT) | |
644459d0 | 4767 | { |
9d98604b | 4768 | HOST_WIDE_INT v = INTVAL (XEXP (XEXP (addr, 0), 1)); |
4769 | if ((v & -16) != 0) | |
4770 | addr = gen_rtx_CONST (Pmode, | |
4771 | gen_rtx_PLUS (Pmode, | |
4772 | XEXP (XEXP (addr, 0), 0), | |
4773 | GEN_INT (v & -16))); | |
4774 | else | |
4775 | addr = XEXP (XEXP (addr, 0), 0); | |
4776 | p1_lo = GEN_INT (v & 15); | |
644459d0 | 4777 | } |
4778 | else if (GET_CODE (addr) == CONST_INT) | |
4779 | { | |
4780 | p1_lo = GEN_INT (INTVAL (addr) & 15); | |
4781 | addr = GEN_INT (INTVAL (addr) & -16); | |
4782 | } | |
9d98604b | 4783 | else |
4784 | { | |
4785 | p1_lo = gen_reg_rtx (SImode); | |
4786 | emit_move_insn (p1_lo, addr); | |
4787 | } | |
644459d0 | 4788 | } |
4789 | ||
4cbad5bb | 4790 | gcc_assert (aform == 0 || aform == 1); |
9d98604b | 4791 | reg = gen_reg_rtx (TImode); |
e04cf423 | 4792 | |
644459d0 | 4793 | scalar = store_with_one_insn_p (ops[0]); |
4794 | if (!scalar) | |
4795 | { | |
4796 | /* We could copy the flags from the ops[0] MEM to mem here, | |
4797 | We don't because we want this load to be optimized away if | |
4798 | possible, and copying the flags will prevent that in certain | |
4799 | cases, e.g. consider the volatile flag. */ | |
4800 | ||
9d98604b | 4801 | rtx pat = gen_reg_rtx (TImode); |
e04cf423 | 4802 | rtx lmem = change_address (ops[0], TImode, copy_rtx (addr)); |
4803 | set_mem_alias_set (lmem, 0); | |
4804 | emit_insn (gen_movti (reg, lmem)); | |
644459d0 | 4805 | |
9d98604b | 4806 | if (!p0 || reg_aligned_for_addr (p0)) |
644459d0 | 4807 | p0 = stack_pointer_rtx; |
4808 | if (!p1_lo) | |
4809 | p1_lo = const0_rtx; | |
4810 | ||
4811 | emit_insn (gen_cpat (pat, p0, p1_lo, GEN_INT (GET_MODE_SIZE (mode)))); | |
4812 | emit_insn (gen_shufb (reg, ops[1], reg, pat)); | |
4813 | } | |
644459d0 | 4814 | else |
4815 | { | |
4816 | if (GET_CODE (ops[1]) == REG) | |
4817 | emit_insn (gen_spu_convert (reg, ops[1])); | |
4818 | else if (GET_CODE (ops[1]) == SUBREG) | |
4819 | emit_insn (gen_spu_convert (reg, SUBREG_REG (ops[1]))); | |
4820 | else | |
4821 | abort (); | |
4822 | } | |
4823 | ||
4824 | if (GET_MODE_SIZE (mode) < 4 && scalar) | |
9d98604b | 4825 | emit_insn (gen_ashlti3 |
4826 | (reg, reg, GEN_INT (32 - GET_MODE_BITSIZE (mode)))); | |
644459d0 | 4827 | |
9d98604b | 4828 | smem = change_address (ops[0], TImode, copy_rtx (addr)); |
644459d0 | 4829 | /* We can't use the previous alias set because the memory has changed |
4830 | size and can potentially overlap objects of other types. */ | |
4831 | set_mem_alias_set (smem, 0); | |
4832 | ||
e04cf423 | 4833 | emit_insn (gen_movti (smem, reg)); |
9d98604b | 4834 | return 1; |
644459d0 | 4835 | } |
4836 | ||
4837 | /* Return TRUE if X is MEM which is a struct member reference | |
4838 | and the member can safely be loaded and stored with a single | |
4839 | instruction because it is padded. */ | |
4840 | static int | |
4841 | mem_is_padded_component_ref (rtx x) | |
4842 | { | |
4843 | tree t = MEM_EXPR (x); | |
4844 | tree r; | |
4845 | if (!t || TREE_CODE (t) != COMPONENT_REF) | |
4846 | return 0; | |
4847 | t = TREE_OPERAND (t, 1); | |
4848 | if (!t || TREE_CODE (t) != FIELD_DECL | |
4849 | || DECL_ALIGN (t) < 128 || AGGREGATE_TYPE_P (TREE_TYPE (t))) | |
4850 | return 0; | |
4851 | /* Only do this for RECORD_TYPEs, not UNION_TYPEs. */ | |
4852 | r = DECL_FIELD_CONTEXT (t); | |
4853 | if (!r || TREE_CODE (r) != RECORD_TYPE) | |
4854 | return 0; | |
4855 | /* Make sure they are the same mode */ | |
4856 | if (GET_MODE (x) != TYPE_MODE (TREE_TYPE (t))) | |
4857 | return 0; | |
4858 | /* If there are no following fields then the field alignment assures | |
fa7637bd | 4859 | the structure is padded to the alignment which means this field is |
4860 | padded too. */ | |
644459d0 | 4861 | if (TREE_CHAIN (t) == 0) |
4862 | return 1; | |
4863 | /* If the following field is also aligned then this field will be | |
4864 | padded. */ | |
4865 | t = TREE_CHAIN (t); | |
4866 | if (TREE_CODE (t) == FIELD_DECL && DECL_ALIGN (t) >= 128) | |
4867 | return 1; | |
4868 | return 0; | |
4869 | } | |
4870 | ||
c7b91b14 | 4871 | /* Parse the -mfixed-range= option string. */ |
4872 | static void | |
4873 | fix_range (const char *const_str) | |
4874 | { | |
4875 | int i, first, last; | |
4876 | char *str, *dash, *comma; | |
4877 | ||
4878 | /* str must be of the form REG1'-'REG2{,REG1'-'REG} where REG1 and | |
4879 | REG2 are either register names or register numbers. The effect | |
4880 | of this option is to mark the registers in the range from REG1 to | |
4881 | REG2 as ``fixed'' so they won't be used by the compiler. */ | |
4882 | ||
4883 | i = strlen (const_str); | |
4884 | str = (char *) alloca (i + 1); | |
4885 | memcpy (str, const_str, i + 1); | |
4886 | ||
4887 | while (1) | |
4888 | { | |
4889 | dash = strchr (str, '-'); | |
4890 | if (!dash) | |
4891 | { | |
4892 | warning (0, "value of -mfixed-range must have form REG1-REG2"); | |
4893 | return; | |
4894 | } | |
4895 | *dash = '\0'; | |
4896 | comma = strchr (dash + 1, ','); | |
4897 | if (comma) | |
4898 | *comma = '\0'; | |
4899 | ||
4900 | first = decode_reg_name (str); | |
4901 | if (first < 0) | |
4902 | { | |
4903 | warning (0, "unknown register name: %s", str); | |
4904 | return; | |
4905 | } | |
4906 | ||
4907 | last = decode_reg_name (dash + 1); | |
4908 | if (last < 0) | |
4909 | { | |
4910 | warning (0, "unknown register name: %s", dash + 1); | |
4911 | return; | |
4912 | } | |
4913 | ||
4914 | *dash = '-'; | |
4915 | ||
4916 | if (first > last) | |
4917 | { | |
4918 | warning (0, "%s-%s is an empty range", str, dash + 1); | |
4919 | return; | |
4920 | } | |
4921 | ||
4922 | for (i = first; i <= last; ++i) | |
4923 | fixed_regs[i] = call_used_regs[i] = 1; | |
4924 | ||
4925 | if (!comma) | |
4926 | break; | |
4927 | ||
4928 | *comma = ','; | |
4929 | str = comma + 1; | |
4930 | } | |
4931 | } | |
4932 | ||
644459d0 | 4933 | /* Return TRUE if x is a CONST_INT, CONST_DOUBLE or CONST_VECTOR that |
4934 | can be generated using the fsmbi instruction. */ | |
4935 | int | |
4936 | fsmbi_const_p (rtx x) | |
4937 | { | |
dea01258 | 4938 | if (CONSTANT_P (x)) |
4939 | { | |
5df189be | 4940 | /* We can always choose TImode for CONST_INT because the high bits |
dea01258 | 4941 | of an SImode will always be all 1s, i.e., valid for fsmbi. */ |
5df189be | 4942 | enum immediate_class c = classify_immediate (x, TImode); |
3072d30e | 4943 | return c == IC_FSMBI || (!epilogue_completed && c == IC_FSMBI2); |
dea01258 | 4944 | } |
4945 | return 0; | |
4946 | } | |
4947 | ||
4948 | /* Return TRUE if x is a CONST_INT, CONST_DOUBLE or CONST_VECTOR that | |
4949 | can be generated using the cbd, chd, cwd or cdd instruction. */ | |
4950 | int | |
3754d046 | 4951 | cpat_const_p (rtx x, machine_mode mode) |
dea01258 | 4952 | { |
4953 | if (CONSTANT_P (x)) | |
4954 | { | |
4955 | enum immediate_class c = classify_immediate (x, mode); | |
4956 | return c == IC_CPAT; | |
4957 | } | |
4958 | return 0; | |
4959 | } | |
644459d0 | 4960 | |
dea01258 | 4961 | rtx |
4962 | gen_cpat_const (rtx * ops) | |
4963 | { | |
4964 | unsigned char dst[16]; | |
4965 | int i, offset, shift, isize; | |
4966 | if (GET_CODE (ops[3]) != CONST_INT | |
4967 | || GET_CODE (ops[2]) != CONST_INT | |
4968 | || (GET_CODE (ops[1]) != CONST_INT | |
4969 | && GET_CODE (ops[1]) != REG)) | |
4970 | return 0; | |
4971 | if (GET_CODE (ops[1]) == REG | |
4972 | && (!REG_POINTER (ops[1]) | |
4973 | || REGNO_POINTER_ALIGN (ORIGINAL_REGNO (ops[1])) < 128)) | |
4974 | return 0; | |
644459d0 | 4975 | |
4976 | for (i = 0; i < 16; i++) | |
dea01258 | 4977 | dst[i] = i + 16; |
4978 | isize = INTVAL (ops[3]); | |
4979 | if (isize == 1) | |
4980 | shift = 3; | |
4981 | else if (isize == 2) | |
4982 | shift = 2; | |
4983 | else | |
4984 | shift = 0; | |
4985 | offset = (INTVAL (ops[2]) + | |
4986 | (GET_CODE (ops[1]) == | |
4987 | CONST_INT ? INTVAL (ops[1]) : 0)) & 15; | |
4988 | for (i = 0; i < isize; i++) | |
4989 | dst[offset + i] = i + shift; | |
4990 | return array_to_constant (TImode, dst); | |
644459d0 | 4991 | } |
4992 | ||
4993 | /* Convert a CONST_INT, CONST_DOUBLE, or CONST_VECTOR into a 16 byte | |
4994 | array. Use MODE for CONST_INT's. When the constant's mode is smaller | |
4995 | than 16 bytes, the value is repeated across the rest of the array. */ | |
4996 | void | |
3754d046 | 4997 | constant_to_array (machine_mode mode, rtx x, unsigned char arr[16]) |
644459d0 | 4998 | { |
4999 | HOST_WIDE_INT val; | |
5000 | int i, j, first; | |
5001 | ||
5002 | memset (arr, 0, 16); | |
5003 | mode = GET_MODE (x) != VOIDmode ? GET_MODE (x) : mode; | |
5004 | if (GET_CODE (x) == CONST_INT | |
5005 | || (GET_CODE (x) == CONST_DOUBLE | |
5006 | && (mode == SFmode || mode == DFmode))) | |
5007 | { | |
5008 | gcc_assert (mode != VOIDmode && mode != BLKmode); | |
5009 | ||
5010 | if (GET_CODE (x) == CONST_DOUBLE) | |
5011 | val = const_double_to_hwint (x); | |
5012 | else | |
5013 | val = INTVAL (x); | |
5014 | first = GET_MODE_SIZE (mode) - 1; | |
5015 | for (i = first; i >= 0; i--) | |
5016 | { | |
5017 | arr[i] = val & 0xff; | |
5018 | val >>= 8; | |
5019 | } | |
5020 | /* Splat the constant across the whole array. */ | |
5021 | for (j = 0, i = first + 1; i < 16; i++) | |
5022 | { | |
5023 | arr[i] = arr[j]; | |
5024 | j = (j == first) ? 0 : j + 1; | |
5025 | } | |
5026 | } | |
5027 | else if (GET_CODE (x) == CONST_DOUBLE) | |
5028 | { | |
5029 | val = CONST_DOUBLE_LOW (x); | |
5030 | for (i = 15; i >= 8; i--) | |
5031 | { | |
5032 | arr[i] = val & 0xff; | |
5033 | val >>= 8; | |
5034 | } | |
5035 | val = CONST_DOUBLE_HIGH (x); | |
5036 | for (i = 7; i >= 0; i--) | |
5037 | { | |
5038 | arr[i] = val & 0xff; | |
5039 | val >>= 8; | |
5040 | } | |
5041 | } | |
5042 | else if (GET_CODE (x) == CONST_VECTOR) | |
5043 | { | |
5044 | int units; | |
5045 | rtx elt; | |
5046 | mode = GET_MODE_INNER (mode); | |
5047 | units = CONST_VECTOR_NUNITS (x); | |
5048 | for (i = 0; i < units; i++) | |
5049 | { | |
5050 | elt = CONST_VECTOR_ELT (x, i); | |
5051 | if (GET_CODE (elt) == CONST_INT || GET_CODE (elt) == CONST_DOUBLE) | |
5052 | { | |
5053 | if (GET_CODE (elt) == CONST_DOUBLE) | |
5054 | val = const_double_to_hwint (elt); | |
5055 | else | |
5056 | val = INTVAL (elt); | |
5057 | first = GET_MODE_SIZE (mode) - 1; | |
5058 | if (first + i * GET_MODE_SIZE (mode) > 16) | |
5059 | abort (); | |
5060 | for (j = first; j >= 0; j--) | |
5061 | { | |
5062 | arr[j + i * GET_MODE_SIZE (mode)] = val & 0xff; | |
5063 | val >>= 8; | |
5064 | } | |
5065 | } | |
5066 | } | |
5067 | } | |
5068 | else | |
5069 | gcc_unreachable(); | |
5070 | } | |
5071 | ||
5072 | /* Convert a 16 byte array to a constant of mode MODE. When MODE is | |
5073 | smaller than 16 bytes, use the bytes that would represent that value | |
5074 | in a register, e.g., for QImode return the value of arr[3]. */ | |
5075 | rtx | |
3754d046 | 5076 | array_to_constant (machine_mode mode, const unsigned char arr[16]) |
644459d0 | 5077 | { |
3754d046 | 5078 | machine_mode inner_mode; |
644459d0 | 5079 | rtvec v; |
5080 | int units, size, i, j, k; | |
5081 | HOST_WIDE_INT val; | |
5082 | ||
5083 | if (GET_MODE_CLASS (mode) == MODE_INT | |
5084 | && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT) | |
5085 | { | |
5086 | j = GET_MODE_SIZE (mode); | |
5087 | i = j < 4 ? 4 - j : 0; | |
5088 | for (val = 0; i < j; i++) | |
5089 | val = (val << 8) | arr[i]; | |
5090 | val = trunc_int_for_mode (val, mode); | |
5091 | return GEN_INT (val); | |
5092 | } | |
5093 | ||
5094 | if (mode == TImode) | |
5095 | { | |
5096 | HOST_WIDE_INT high; | |
5097 | for (i = high = 0; i < 8; i++) | |
5098 | high = (high << 8) | arr[i]; | |
5099 | for (i = 8, val = 0; i < 16; i++) | |
5100 | val = (val << 8) | arr[i]; | |
5101 | return immed_double_const (val, high, TImode); | |
5102 | } | |
5103 | if (mode == SFmode) | |
5104 | { | |
5105 | val = (arr[0] << 24) | (arr[1] << 16) | (arr[2] << 8) | arr[3]; | |
5106 | val = trunc_int_for_mode (val, SImode); | |
171b6d22 | 5107 | return hwint_to_const_double (SFmode, val); |
644459d0 | 5108 | } |
5109 | if (mode == DFmode) | |
5110 | { | |
1f915911 | 5111 | for (i = 0, val = 0; i < 8; i++) |
5112 | val = (val << 8) | arr[i]; | |
171b6d22 | 5113 | return hwint_to_const_double (DFmode, val); |
644459d0 | 5114 | } |
5115 | ||
5116 | if (!VECTOR_MODE_P (mode)) | |
5117 | abort (); | |
5118 | ||
5119 | units = GET_MODE_NUNITS (mode); | |
5120 | size = GET_MODE_UNIT_SIZE (mode); | |
5121 | inner_mode = GET_MODE_INNER (mode); | |
5122 | v = rtvec_alloc (units); | |
5123 | ||
5124 | for (k = i = 0; i < units; ++i) | |
5125 | { | |
5126 | val = 0; | |
5127 | for (j = 0; j < size; j++, k++) | |
5128 | val = (val << 8) | arr[k]; | |
5129 | ||
5130 | if (GET_MODE_CLASS (inner_mode) == MODE_FLOAT) | |
5131 | RTVEC_ELT (v, i) = hwint_to_const_double (inner_mode, val); | |
5132 | else | |
5133 | RTVEC_ELT (v, i) = GEN_INT (trunc_int_for_mode (val, inner_mode)); | |
5134 | } | |
5135 | if (k > 16) | |
5136 | abort (); | |
5137 | ||
5138 | return gen_rtx_CONST_VECTOR (mode, v); | |
5139 | } | |
5140 | ||
5141 | static void | |
5142 | reloc_diagnostic (rtx x) | |
5143 | { | |
712d2297 | 5144 | tree decl = 0; |
644459d0 | 5145 | if (!flag_pic || !(TARGET_WARN_RELOC || TARGET_ERROR_RELOC)) |
5146 | return; | |
5147 | ||
5148 | if (GET_CODE (x) == SYMBOL_REF) | |
5149 | decl = SYMBOL_REF_DECL (x); | |
5150 | else if (GET_CODE (x) == CONST | |
5151 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF) | |
5152 | decl = SYMBOL_REF_DECL (XEXP (XEXP (x, 0), 0)); | |
5153 | ||
5154 | /* SYMBOL_REF_DECL is not necessarily a DECL. */ | |
5155 | if (decl && !DECL_P (decl)) | |
5156 | decl = 0; | |
5157 | ||
644459d0 | 5158 | /* The decl could be a string constant. */ |
5159 | if (decl && DECL_P (decl)) | |
712d2297 | 5160 | { |
5161 | location_t loc; | |
5162 | /* We use last_assemble_variable_decl to get line information. It's | |
5163 | not always going to be right and might not even be close, but will | |
5164 | be right for the more common cases. */ | |
5165 | if (!last_assemble_variable_decl || in_section == ctors_section) | |
5166 | loc = DECL_SOURCE_LOCATION (decl); | |
5167 | else | |
5168 | loc = DECL_SOURCE_LOCATION (last_assemble_variable_decl); | |
644459d0 | 5169 | |
712d2297 | 5170 | if (TARGET_WARN_RELOC) |
5171 | warning_at (loc, 0, | |
5172 | "creating run-time relocation for %qD", decl); | |
5173 | else | |
5174 | error_at (loc, | |
5175 | "creating run-time relocation for %qD", decl); | |
5176 | } | |
5177 | else | |
5178 | { | |
5179 | if (TARGET_WARN_RELOC) | |
5180 | warning_at (input_location, 0, "creating run-time relocation"); | |
5181 | else | |
5182 | error_at (input_location, "creating run-time relocation"); | |
5183 | } | |
644459d0 | 5184 | } |
5185 | ||
5186 | /* Hook into assemble_integer so we can generate an error for run-time | |
5187 | relocations. The SPU ABI disallows them. */ | |
5188 | static bool | |
5189 | spu_assemble_integer (rtx x, unsigned int size, int aligned_p) | |
5190 | { | |
5191 | /* By default run-time relocations aren't supported, but we allow them | |
5192 | in case users support it in their own run-time loader. And we provide | |
5193 | a warning for those users that don't. */ | |
5194 | if ((GET_CODE (x) == SYMBOL_REF) | |
5195 | || GET_CODE (x) == LABEL_REF || GET_CODE (x) == CONST) | |
5196 | reloc_diagnostic (x); | |
5197 | ||
5198 | return default_assemble_integer (x, size, aligned_p); | |
5199 | } | |
5200 | ||
5201 | static void | |
5202 | spu_asm_globalize_label (FILE * file, const char *name) | |
5203 | { | |
5204 | fputs ("\t.global\t", file); | |
5205 | assemble_name (file, name); | |
5206 | fputs ("\n", file); | |
5207 | } | |
5208 | ||
5209 | static bool | |
5ae4887d | 5210 | spu_rtx_costs (rtx x, machine_mode mode, int outer_code ATTRIBUTE_UNUSED, |
20d892d1 | 5211 | int opno ATTRIBUTE_UNUSED, int *total, |
f529eb25 | 5212 | bool speed ATTRIBUTE_UNUSED) |
644459d0 | 5213 | { |
5ae4887d | 5214 | int code = GET_CODE (x); |
644459d0 | 5215 | int cost = COSTS_N_INSNS (2); |
5216 | ||
5217 | /* Folding to a CONST_VECTOR will use extra space but there might | |
5218 | be only a small savings in cycles. We'd like to use a CONST_VECTOR | |
9505a73b | 5219 | only if it allows us to fold away multiple insns. Changing the cost |
644459d0 | 5220 | of a CONST_VECTOR here (or in CONST_COSTS) doesn't help though |
5221 | because this cost will only be compared against a single insn. | |
5222 | if (code == CONST_VECTOR) | |
ca316360 | 5223 | return spu_legitimate_constant_p (mode, x) ? cost : COSTS_N_INSNS (6); |
644459d0 | 5224 | */ |
5225 | ||
5226 | /* Use defaults for float operations. Not accurate but good enough. */ | |
5227 | if (mode == DFmode) | |
5228 | { | |
5229 | *total = COSTS_N_INSNS (13); | |
5230 | return true; | |
5231 | } | |
5232 | if (mode == SFmode) | |
5233 | { | |
5234 | *total = COSTS_N_INSNS (6); | |
5235 | return true; | |
5236 | } | |
5237 | switch (code) | |
5238 | { | |
5239 | case CONST_INT: | |
5240 | if (satisfies_constraint_K (x)) | |
5241 | *total = 0; | |
5242 | else if (INTVAL (x) >= -0x80000000ll && INTVAL (x) <= 0xffffffffll) | |
5243 | *total = COSTS_N_INSNS (1); | |
5244 | else | |
5245 | *total = COSTS_N_INSNS (3); | |
5246 | return true; | |
5247 | ||
5248 | case CONST: | |
5249 | *total = COSTS_N_INSNS (3); | |
5250 | return true; | |
5251 | ||
5252 | case LABEL_REF: | |
5253 | case SYMBOL_REF: | |
5254 | *total = COSTS_N_INSNS (0); | |
5255 | return true; | |
5256 | ||
5257 | case CONST_DOUBLE: | |
5258 | *total = COSTS_N_INSNS (5); | |
5259 | return true; | |
5260 | ||
5261 | case FLOAT_EXTEND: | |
5262 | case FLOAT_TRUNCATE: | |
5263 | case FLOAT: | |
5264 | case UNSIGNED_FLOAT: | |
5265 | case FIX: | |
5266 | case UNSIGNED_FIX: | |
5267 | *total = COSTS_N_INSNS (7); | |
5268 | return true; | |
5269 | ||
5270 | case PLUS: | |
5271 | if (mode == TImode) | |
5272 | { | |
5273 | *total = COSTS_N_INSNS (9); | |
5274 | return true; | |
5275 | } | |
5276 | break; | |
5277 | ||
5278 | case MULT: | |
5279 | cost = | |
5280 | GET_CODE (XEXP (x, 0)) == | |
5281 | REG ? COSTS_N_INSNS (12) : COSTS_N_INSNS (7); | |
5282 | if (mode == SImode && GET_CODE (XEXP (x, 0)) == REG) | |
5283 | { | |
5284 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
5285 | { | |
5286 | HOST_WIDE_INT val = INTVAL (XEXP (x, 1)); | |
5287 | cost = COSTS_N_INSNS (14); | |
5288 | if ((val & 0xffff) == 0) | |
5289 | cost = COSTS_N_INSNS (9); | |
5290 | else if (val > 0 && val < 0x10000) | |
5291 | cost = COSTS_N_INSNS (11); | |
5292 | } | |
5293 | } | |
5294 | *total = cost; | |
5295 | return true; | |
5296 | case DIV: | |
5297 | case UDIV: | |
5298 | case MOD: | |
5299 | case UMOD: | |
5300 | *total = COSTS_N_INSNS (20); | |
5301 | return true; | |
5302 | case ROTATE: | |
5303 | case ROTATERT: | |
5304 | case ASHIFT: | |
5305 | case ASHIFTRT: | |
5306 | case LSHIFTRT: | |
5307 | *total = COSTS_N_INSNS (4); | |
5308 | return true; | |
5309 | case UNSPEC: | |
5310 | if (XINT (x, 1) == UNSPEC_CONVERT) | |
5311 | *total = COSTS_N_INSNS (0); | |
5312 | else | |
5313 | *total = COSTS_N_INSNS (4); | |
5314 | return true; | |
5315 | } | |
5316 | /* Scale cost by mode size. Except when initializing (cfun->decl == 0). */ | |
5317 | if (GET_MODE_CLASS (mode) == MODE_INT | |
5318 | && GET_MODE_SIZE (mode) > GET_MODE_SIZE (SImode) && cfun && cfun->decl) | |
5319 | cost = cost * (GET_MODE_SIZE (mode) / GET_MODE_SIZE (SImode)) | |
5320 | * (GET_MODE_SIZE (mode) / GET_MODE_SIZE (SImode)); | |
5321 | *total = cost; | |
5322 | return true; | |
5323 | } | |
5324 | ||
3754d046 | 5325 | static machine_mode |
1bd43494 | 5326 | spu_unwind_word_mode (void) |
644459d0 | 5327 | { |
1bd43494 | 5328 | return SImode; |
644459d0 | 5329 | } |
5330 | ||
5331 | /* Decide whether we can make a sibling call to a function. DECL is the | |
5332 | declaration of the function being targeted by the call and EXP is the | |
5333 | CALL_EXPR representing the call. */ | |
5334 | static bool | |
5335 | spu_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED) | |
5336 | { | |
5337 | return decl && !TARGET_LARGE_MEM; | |
5338 | } | |
5339 | ||
5340 | /* We need to correctly update the back chain pointer and the Available | |
5341 | Stack Size (which is in the second slot of the sp register.) */ | |
5342 | void | |
5343 | spu_allocate_stack (rtx op0, rtx op1) | |
5344 | { | |
5345 | HOST_WIDE_INT v; | |
5346 | rtx chain = gen_reg_rtx (V4SImode); | |
5347 | rtx stack_bot = gen_frame_mem (V4SImode, stack_pointer_rtx); | |
5348 | rtx sp = gen_reg_rtx (V4SImode); | |
5349 | rtx splatted = gen_reg_rtx (V4SImode); | |
5350 | rtx pat = gen_reg_rtx (TImode); | |
5351 | ||
5352 | /* copy the back chain so we can save it back again. */ | |
5353 | emit_move_insn (chain, stack_bot); | |
5354 | ||
5355 | op1 = force_reg (SImode, op1); | |
5356 | ||
5357 | v = 0x1020300010203ll; | |
5358 | emit_move_insn (pat, immed_double_const (v, v, TImode)); | |
5359 | emit_insn (gen_shufb (splatted, op1, op1, pat)); | |
5360 | ||
5361 | emit_insn (gen_spu_convert (sp, stack_pointer_rtx)); | |
5362 | emit_insn (gen_subv4si3 (sp, sp, splatted)); | |
5363 | ||
5364 | if (flag_stack_check) | |
5365 | { | |
5366 | rtx avail = gen_reg_rtx(SImode); | |
5367 | rtx result = gen_reg_rtx(SImode); | |
5368 | emit_insn (gen_vec_extractv4si (avail, sp, GEN_INT (1))); | |
5369 | emit_insn (gen_cgt_si(result, avail, GEN_INT (-1))); | |
5370 | emit_insn (gen_spu_heq (result, GEN_INT(0) )); | |
5371 | } | |
5372 | ||
5373 | emit_insn (gen_spu_convert (stack_pointer_rtx, sp)); | |
5374 | ||
5375 | emit_move_insn (stack_bot, chain); | |
5376 | ||
5377 | emit_move_insn (op0, virtual_stack_dynamic_rtx); | |
5378 | } | |
5379 | ||
5380 | void | |
5381 | spu_restore_stack_nonlocal (rtx op0 ATTRIBUTE_UNUSED, rtx op1) | |
5382 | { | |
5383 | static unsigned char arr[16] = | |
5384 | { 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3 }; | |
5385 | rtx temp = gen_reg_rtx (SImode); | |
5386 | rtx temp2 = gen_reg_rtx (SImode); | |
5387 | rtx temp3 = gen_reg_rtx (V4SImode); | |
5388 | rtx temp4 = gen_reg_rtx (V4SImode); | |
5389 | rtx pat = gen_reg_rtx (TImode); | |
5390 | rtx sp = gen_rtx_REG (V4SImode, STACK_POINTER_REGNUM); | |
5391 | ||
5392 | /* Restore the backchain from the first word, sp from the second. */ | |
5393 | emit_move_insn (temp2, adjust_address_nv (op1, SImode, 0)); | |
5394 | emit_move_insn (temp, adjust_address_nv (op1, SImode, 4)); | |
5395 | ||
5396 | emit_move_insn (pat, array_to_constant (TImode, arr)); | |
5397 | ||
5398 | /* Compute Available Stack Size for sp */ | |
5399 | emit_insn (gen_subsi3 (temp, temp, stack_pointer_rtx)); | |
5400 | emit_insn (gen_shufb (temp3, temp, temp, pat)); | |
5401 | ||
5402 | /* Compute Available Stack Size for back chain */ | |
5403 | emit_insn (gen_subsi3 (temp2, temp2, stack_pointer_rtx)); | |
5404 | emit_insn (gen_shufb (temp4, temp2, temp2, pat)); | |
5405 | emit_insn (gen_addv4si3 (temp4, sp, temp4)); | |
5406 | ||
5407 | emit_insn (gen_addv4si3 (sp, sp, temp3)); | |
5408 | emit_move_insn (gen_frame_mem (V4SImode, stack_pointer_rtx), temp4); | |
5409 | } | |
5410 | ||
5411 | static void | |
5412 | spu_init_libfuncs (void) | |
5413 | { | |
5414 | set_optab_libfunc (smul_optab, DImode, "__muldi3"); | |
5415 | set_optab_libfunc (sdiv_optab, DImode, "__divdi3"); | |
5416 | set_optab_libfunc (smod_optab, DImode, "__moddi3"); | |
5417 | set_optab_libfunc (udiv_optab, DImode, "__udivdi3"); | |
5418 | set_optab_libfunc (umod_optab, DImode, "__umoddi3"); | |
5419 | set_optab_libfunc (udivmod_optab, DImode, "__udivmoddi4"); | |
5420 | set_optab_libfunc (ffs_optab, DImode, "__ffsdi2"); | |
5421 | set_optab_libfunc (clz_optab, DImode, "__clzdi2"); | |
5422 | set_optab_libfunc (ctz_optab, DImode, "__ctzdi2"); | |
4d3aeb29 | 5423 | set_optab_libfunc (clrsb_optab, DImode, "__clrsbdi2"); |
644459d0 | 5424 | set_optab_libfunc (popcount_optab, DImode, "__popcountdi2"); |
5425 | set_optab_libfunc (parity_optab, DImode, "__paritydi2"); | |
5426 | ||
5427 | set_conv_libfunc (ufloat_optab, DFmode, SImode, "__float_unssidf"); | |
5428 | set_conv_libfunc (ufloat_optab, DFmode, DImode, "__float_unsdidf"); | |
19a53068 | 5429 | |
5825ec3f | 5430 | set_optab_libfunc (addv_optab, SImode, "__addvsi3"); |
5431 | set_optab_libfunc (subv_optab, SImode, "__subvsi3"); | |
5432 | set_optab_libfunc (smulv_optab, SImode, "__mulvsi3"); | |
5433 | set_optab_libfunc (sdivv_optab, SImode, "__divvsi3"); | |
5434 | set_optab_libfunc (negv_optab, SImode, "__negvsi2"); | |
5435 | set_optab_libfunc (absv_optab, SImode, "__absvsi2"); | |
5436 | set_optab_libfunc (addv_optab, DImode, "__addvdi3"); | |
5437 | set_optab_libfunc (subv_optab, DImode, "__subvdi3"); | |
5438 | set_optab_libfunc (smulv_optab, DImode, "__mulvdi3"); | |
5439 | set_optab_libfunc (sdivv_optab, DImode, "__divvdi3"); | |
5440 | set_optab_libfunc (negv_optab, DImode, "__negvdi2"); | |
5441 | set_optab_libfunc (absv_optab, DImode, "__absvdi2"); | |
5442 | ||
19a53068 | 5443 | set_optab_libfunc (smul_optab, TImode, "__multi3"); |
5444 | set_optab_libfunc (sdiv_optab, TImode, "__divti3"); | |
5445 | set_optab_libfunc (smod_optab, TImode, "__modti3"); | |
5446 | set_optab_libfunc (udiv_optab, TImode, "__udivti3"); | |
5447 | set_optab_libfunc (umod_optab, TImode, "__umodti3"); | |
5448 | set_optab_libfunc (udivmod_optab, TImode, "__udivmodti4"); | |
644459d0 | 5449 | } |
5450 | ||
5451 | /* Make a subreg, stripping any existing subreg. We could possibly just | |
5452 | call simplify_subreg, but in this case we know what we want. */ | |
5453 | rtx | |
3754d046 | 5454 | spu_gen_subreg (machine_mode mode, rtx x) |
644459d0 | 5455 | { |
5456 | if (GET_CODE (x) == SUBREG) | |
5457 | x = SUBREG_REG (x); | |
5458 | if (GET_MODE (x) == mode) | |
5459 | return x; | |
5460 | return gen_rtx_SUBREG (mode, x, 0); | |
5461 | } | |
5462 | ||
5463 | static bool | |
fb80456a | 5464 | spu_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED) |
644459d0 | 5465 | { |
5466 | return (TYPE_MODE (type) == BLKmode | |
5467 | && ((type) == 0 | |
5468 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST | |
5469 | || int_size_in_bytes (type) > | |
5470 | (MAX_REGISTER_RETURN * UNITS_PER_WORD))); | |
5471 | } | |
5472 | \f | |
5473 | /* Create the built-in types and functions */ | |
5474 | ||
c2233b46 | 5475 | enum spu_function_code |
5476 | { | |
5477 | #define DEF_BUILTIN(fcode, icode, name, type, params) fcode, | |
5478 | #include "spu-builtins.def" | |
5479 | #undef DEF_BUILTIN | |
5480 | NUM_SPU_BUILTINS | |
5481 | }; | |
5482 | ||
5483 | extern GTY(()) struct spu_builtin_description spu_builtins[NUM_SPU_BUILTINS]; | |
5484 | ||
644459d0 | 5485 | struct spu_builtin_description spu_builtins[] = { |
5486 | #define DEF_BUILTIN(fcode, icode, name, type, params) \ | |
0c5c4d59 | 5487 | {fcode, icode, name, type, params}, |
644459d0 | 5488 | #include "spu-builtins.def" |
5489 | #undef DEF_BUILTIN | |
5490 | }; | |
5491 | ||
0c5c4d59 | 5492 | static GTY(()) tree spu_builtin_decls[NUM_SPU_BUILTINS]; |
5493 | ||
5494 | /* Returns the spu builtin decl for CODE. */ | |
e6925042 | 5495 | |
5496 | static tree | |
5497 | spu_builtin_decl (unsigned code, bool initialize_p ATTRIBUTE_UNUSED) | |
5498 | { | |
5499 | if (code >= NUM_SPU_BUILTINS) | |
5500 | return error_mark_node; | |
5501 | ||
0c5c4d59 | 5502 | return spu_builtin_decls[code]; |
e6925042 | 5503 | } |
5504 | ||
5505 | ||
644459d0 | 5506 | static void |
5507 | spu_init_builtins (void) | |
5508 | { | |
5509 | struct spu_builtin_description *d; | |
5510 | unsigned int i; | |
5511 | ||
5512 | V16QI_type_node = build_vector_type (intQI_type_node, 16); | |
5513 | V8HI_type_node = build_vector_type (intHI_type_node, 8); | |
5514 | V4SI_type_node = build_vector_type (intSI_type_node, 4); | |
5515 | V2DI_type_node = build_vector_type (intDI_type_node, 2); | |
5516 | V4SF_type_node = build_vector_type (float_type_node, 4); | |
5517 | V2DF_type_node = build_vector_type (double_type_node, 2); | |
5518 | ||
5519 | unsigned_V16QI_type_node = build_vector_type (unsigned_intQI_type_node, 16); | |
5520 | unsigned_V8HI_type_node = build_vector_type (unsigned_intHI_type_node, 8); | |
5521 | unsigned_V4SI_type_node = build_vector_type (unsigned_intSI_type_node, 4); | |
5522 | unsigned_V2DI_type_node = build_vector_type (unsigned_intDI_type_node, 2); | |
5523 | ||
c4ecce0c | 5524 | spu_builtin_types[SPU_BTI_QUADWORD] = V16QI_type_node; |
644459d0 | 5525 | |
5526 | spu_builtin_types[SPU_BTI_7] = global_trees[TI_INTSI_TYPE]; | |
5527 | spu_builtin_types[SPU_BTI_S7] = global_trees[TI_INTSI_TYPE]; | |
5528 | spu_builtin_types[SPU_BTI_U7] = global_trees[TI_INTSI_TYPE]; | |
5529 | spu_builtin_types[SPU_BTI_S10] = global_trees[TI_INTSI_TYPE]; | |
5530 | spu_builtin_types[SPU_BTI_S10_4] = global_trees[TI_INTSI_TYPE]; | |
5531 | spu_builtin_types[SPU_BTI_U14] = global_trees[TI_INTSI_TYPE]; | |
5532 | spu_builtin_types[SPU_BTI_16] = global_trees[TI_INTSI_TYPE]; | |
5533 | spu_builtin_types[SPU_BTI_S16] = global_trees[TI_INTSI_TYPE]; | |
5534 | spu_builtin_types[SPU_BTI_S16_2] = global_trees[TI_INTSI_TYPE]; | |
5535 | spu_builtin_types[SPU_BTI_U16] = global_trees[TI_INTSI_TYPE]; | |
5536 | spu_builtin_types[SPU_BTI_U16_2] = global_trees[TI_INTSI_TYPE]; | |
5537 | spu_builtin_types[SPU_BTI_U18] = global_trees[TI_INTSI_TYPE]; | |
5538 | ||
5539 | spu_builtin_types[SPU_BTI_INTQI] = global_trees[TI_INTQI_TYPE]; | |
5540 | spu_builtin_types[SPU_BTI_INTHI] = global_trees[TI_INTHI_TYPE]; | |
5541 | spu_builtin_types[SPU_BTI_INTSI] = global_trees[TI_INTSI_TYPE]; | |
5542 | spu_builtin_types[SPU_BTI_INTDI] = global_trees[TI_INTDI_TYPE]; | |
5543 | spu_builtin_types[SPU_BTI_UINTQI] = global_trees[TI_UINTQI_TYPE]; | |
5544 | spu_builtin_types[SPU_BTI_UINTHI] = global_trees[TI_UINTHI_TYPE]; | |
5545 | spu_builtin_types[SPU_BTI_UINTSI] = global_trees[TI_UINTSI_TYPE]; | |
5546 | spu_builtin_types[SPU_BTI_UINTDI] = global_trees[TI_UINTDI_TYPE]; | |
5547 | ||
5548 | spu_builtin_types[SPU_BTI_FLOAT] = global_trees[TI_FLOAT_TYPE]; | |
5549 | spu_builtin_types[SPU_BTI_DOUBLE] = global_trees[TI_DOUBLE_TYPE]; | |
5550 | ||
5551 | spu_builtin_types[SPU_BTI_VOID] = global_trees[TI_VOID_TYPE]; | |
5552 | ||
5553 | spu_builtin_types[SPU_BTI_PTR] = | |
5554 | build_pointer_type (build_qualified_type | |
5555 | (void_type_node, | |
5556 | TYPE_QUAL_CONST | TYPE_QUAL_VOLATILE)); | |
5557 | ||
5558 | /* For each builtin we build a new prototype. The tree code will make | |
5559 | sure nodes are shared. */ | |
5560 | for (i = 0, d = spu_builtins; i < NUM_SPU_BUILTINS; i++, d++) | |
5561 | { | |
5562 | tree p; | |
5563 | char name[64]; /* build_function will make a copy. */ | |
5564 | int parm; | |
5565 | ||
5566 | if (d->name == 0) | |
5567 | continue; | |
5568 | ||
5dfbd18f | 5569 | /* Find last parm. */ |
644459d0 | 5570 | for (parm = 1; d->parm[parm] != SPU_BTI_END_OF_PARAMS; parm++) |
5dfbd18f | 5571 | ; |
644459d0 | 5572 | |
5573 | p = void_list_node; | |
5574 | while (parm > 1) | |
5575 | p = tree_cons (NULL_TREE, spu_builtin_types[d->parm[--parm]], p); | |
5576 | ||
5577 | p = build_function_type (spu_builtin_types[d->parm[0]], p); | |
5578 | ||
5579 | sprintf (name, "__builtin_%s", d->name); | |
0c5c4d59 | 5580 | spu_builtin_decls[i] = |
3726fe5e | 5581 | add_builtin_function (name, p, i, BUILT_IN_MD, NULL, NULL_TREE); |
a76866d3 | 5582 | if (d->fcode == SPU_MASK_FOR_LOAD) |
0c5c4d59 | 5583 | TREE_READONLY (spu_builtin_decls[i]) = 1; |
5dfbd18f | 5584 | |
5585 | /* These builtins don't throw. */ | |
0c5c4d59 | 5586 | TREE_NOTHROW (spu_builtin_decls[i]) = 1; |
644459d0 | 5587 | } |
5588 | } | |
5589 | ||
cf31d486 | 5590 | void |
5591 | spu_restore_stack_block (rtx op0 ATTRIBUTE_UNUSED, rtx op1) | |
5592 | { | |
5593 | static unsigned char arr[16] = | |
5594 | { 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3 }; | |
5595 | ||
5596 | rtx temp = gen_reg_rtx (Pmode); | |
5597 | rtx temp2 = gen_reg_rtx (V4SImode); | |
5598 | rtx temp3 = gen_reg_rtx (V4SImode); | |
5599 | rtx pat = gen_reg_rtx (TImode); | |
5600 | rtx sp = gen_rtx_REG (V4SImode, STACK_POINTER_REGNUM); | |
5601 | ||
5602 | emit_move_insn (pat, array_to_constant (TImode, arr)); | |
5603 | ||
5604 | /* Restore the sp. */ | |
5605 | emit_move_insn (temp, op1); | |
5606 | emit_move_insn (temp2, gen_frame_mem (V4SImode, stack_pointer_rtx)); | |
5607 | ||
5608 | /* Compute available stack size for sp. */ | |
5609 | emit_insn (gen_subsi3 (temp, temp, stack_pointer_rtx)); | |
5610 | emit_insn (gen_shufb (temp3, temp, temp, pat)); | |
5611 | ||
5612 | emit_insn (gen_addv4si3 (sp, sp, temp3)); | |
5613 | emit_move_insn (gen_frame_mem (V4SImode, stack_pointer_rtx), temp2); | |
5614 | } | |
5615 | ||
644459d0 | 5616 | int |
5617 | spu_safe_dma (HOST_WIDE_INT channel) | |
5618 | { | |
006e4b96 | 5619 | return TARGET_SAFE_DMA && channel >= 21 && channel <= 27; |
644459d0 | 5620 | } |
5621 | ||
5622 | void | |
5623 | spu_builtin_splats (rtx ops[]) | |
5624 | { | |
3754d046 | 5625 | machine_mode mode = GET_MODE (ops[0]); |
644459d0 | 5626 | if (GET_CODE (ops[1]) == CONST_INT || GET_CODE (ops[1]) == CONST_DOUBLE) |
5627 | { | |
5628 | unsigned char arr[16]; | |
5629 | constant_to_array (GET_MODE_INNER (mode), ops[1], arr); | |
5630 | emit_move_insn (ops[0], array_to_constant (mode, arr)); | |
5631 | } | |
644459d0 | 5632 | else |
5633 | { | |
5634 | rtx reg = gen_reg_rtx (TImode); | |
5635 | rtx shuf; | |
5636 | if (GET_CODE (ops[1]) != REG | |
5637 | && GET_CODE (ops[1]) != SUBREG) | |
5638 | ops[1] = force_reg (GET_MODE_INNER (mode), ops[1]); | |
5639 | switch (mode) | |
5640 | { | |
5641 | case V2DImode: | |
5642 | case V2DFmode: | |
5643 | shuf = | |
5644 | immed_double_const (0x0001020304050607ll, 0x1011121314151617ll, | |
5645 | TImode); | |
5646 | break; | |
5647 | case V4SImode: | |
5648 | case V4SFmode: | |
5649 | shuf = | |
5650 | immed_double_const (0x0001020300010203ll, 0x0001020300010203ll, | |
5651 | TImode); | |
5652 | break; | |
5653 | case V8HImode: | |
5654 | shuf = | |
5655 | immed_double_const (0x0203020302030203ll, 0x0203020302030203ll, | |
5656 | TImode); | |
5657 | break; | |
5658 | case V16QImode: | |
5659 | shuf = | |
5660 | immed_double_const (0x0303030303030303ll, 0x0303030303030303ll, | |
5661 | TImode); | |
5662 | break; | |
5663 | default: | |
5664 | abort (); | |
5665 | } | |
5666 | emit_move_insn (reg, shuf); | |
5667 | emit_insn (gen_shufb (ops[0], ops[1], ops[1], reg)); | |
5668 | } | |
5669 | } | |
5670 | ||
5671 | void | |
5672 | spu_builtin_extract (rtx ops[]) | |
5673 | { | |
3754d046 | 5674 | machine_mode mode; |
644459d0 | 5675 | rtx rot, from, tmp; |
5676 | ||
5677 | mode = GET_MODE (ops[1]); | |
5678 | ||
5679 | if (GET_CODE (ops[2]) == CONST_INT) | |
5680 | { | |
5681 | switch (mode) | |
5682 | { | |
5683 | case V16QImode: | |
5684 | emit_insn (gen_vec_extractv16qi (ops[0], ops[1], ops[2])); | |
5685 | break; | |
5686 | case V8HImode: | |
5687 | emit_insn (gen_vec_extractv8hi (ops[0], ops[1], ops[2])); | |
5688 | break; | |
5689 | case V4SFmode: | |
5690 | emit_insn (gen_vec_extractv4sf (ops[0], ops[1], ops[2])); | |
5691 | break; | |
5692 | case V4SImode: | |
5693 | emit_insn (gen_vec_extractv4si (ops[0], ops[1], ops[2])); | |
5694 | break; | |
5695 | case V2DImode: | |
5696 | emit_insn (gen_vec_extractv2di (ops[0], ops[1], ops[2])); | |
5697 | break; | |
5698 | case V2DFmode: | |
5699 | emit_insn (gen_vec_extractv2df (ops[0], ops[1], ops[2])); | |
5700 | break; | |
5701 | default: | |
5702 | abort (); | |
5703 | } | |
5704 | return; | |
5705 | } | |
5706 | ||
5707 | from = spu_gen_subreg (TImode, ops[1]); | |
5708 | rot = gen_reg_rtx (TImode); | |
5709 | tmp = gen_reg_rtx (SImode); | |
5710 | ||
5711 | switch (mode) | |
5712 | { | |
5713 | case V16QImode: | |
5714 | emit_insn (gen_addsi3 (tmp, ops[2], GEN_INT (-3))); | |
5715 | break; | |
5716 | case V8HImode: | |
5717 | emit_insn (gen_addsi3 (tmp, ops[2], ops[2])); | |
5718 | emit_insn (gen_addsi3 (tmp, tmp, GEN_INT (-2))); | |
5719 | break; | |
5720 | case V4SFmode: | |
5721 | case V4SImode: | |
5722 | emit_insn (gen_ashlsi3 (tmp, ops[2], GEN_INT (2))); | |
5723 | break; | |
5724 | case V2DImode: | |
5725 | case V2DFmode: | |
5726 | emit_insn (gen_ashlsi3 (tmp, ops[2], GEN_INT (3))); | |
5727 | break; | |
5728 | default: | |
5729 | abort (); | |
5730 | } | |
5731 | emit_insn (gen_rotqby_ti (rot, from, tmp)); | |
5732 | ||
5733 | emit_insn (gen_spu_convert (ops[0], rot)); | |
5734 | } | |
5735 | ||
5736 | void | |
5737 | spu_builtin_insert (rtx ops[]) | |
5738 | { | |
3754d046 | 5739 | machine_mode mode = GET_MODE (ops[0]); |
5740 | machine_mode imode = GET_MODE_INNER (mode); | |
644459d0 | 5741 | rtx mask = gen_reg_rtx (TImode); |
5742 | rtx offset; | |
5743 | ||
5744 | if (GET_CODE (ops[3]) == CONST_INT) | |
5745 | offset = GEN_INT (INTVAL (ops[3]) * GET_MODE_SIZE (imode)); | |
5746 | else | |
5747 | { | |
5748 | offset = gen_reg_rtx (SImode); | |
5749 | emit_insn (gen_mulsi3 | |
5750 | (offset, ops[3], GEN_INT (GET_MODE_SIZE (imode)))); | |
5751 | } | |
5752 | emit_insn (gen_cpat | |
5753 | (mask, stack_pointer_rtx, offset, | |
5754 | GEN_INT (GET_MODE_SIZE (imode)))); | |
5755 | emit_insn (gen_shufb (ops[0], ops[1], ops[2], mask)); | |
5756 | } | |
5757 | ||
5758 | void | |
5759 | spu_builtin_promote (rtx ops[]) | |
5760 | { | |
3754d046 | 5761 | machine_mode mode, imode; |
644459d0 | 5762 | rtx rot, from, offset; |
5763 | HOST_WIDE_INT pos; | |
5764 | ||
5765 | mode = GET_MODE (ops[0]); | |
5766 | imode = GET_MODE_INNER (mode); | |
5767 | ||
5768 | from = gen_reg_rtx (TImode); | |
5769 | rot = spu_gen_subreg (TImode, ops[0]); | |
5770 | ||
5771 | emit_insn (gen_spu_convert (from, ops[1])); | |
5772 | ||
5773 | if (GET_CODE (ops[2]) == CONST_INT) | |
5774 | { | |
5775 | pos = -GET_MODE_SIZE (imode) * INTVAL (ops[2]); | |
5776 | if (GET_MODE_SIZE (imode) < 4) | |
5777 | pos += 4 - GET_MODE_SIZE (imode); | |
5778 | offset = GEN_INT (pos & 15); | |
5779 | } | |
5780 | else | |
5781 | { | |
5782 | offset = gen_reg_rtx (SImode); | |
5783 | switch (mode) | |
5784 | { | |
5785 | case V16QImode: | |
5786 | emit_insn (gen_subsi3 (offset, GEN_INT (3), ops[2])); | |
5787 | break; | |
5788 | case V8HImode: | |
5789 | emit_insn (gen_subsi3 (offset, GEN_INT (1), ops[2])); | |
5790 | emit_insn (gen_addsi3 (offset, offset, offset)); | |
5791 | break; | |
5792 | case V4SFmode: | |
5793 | case V4SImode: | |
5794 | emit_insn (gen_subsi3 (offset, GEN_INT (0), ops[2])); | |
5795 | emit_insn (gen_ashlsi3 (offset, offset, GEN_INT (2))); | |
5796 | break; | |
5797 | case V2DImode: | |
5798 | case V2DFmode: | |
5799 | emit_insn (gen_ashlsi3 (offset, ops[2], GEN_INT (3))); | |
5800 | break; | |
5801 | default: | |
5802 | abort (); | |
5803 | } | |
5804 | } | |
5805 | emit_insn (gen_rotqby_ti (rot, from, offset)); | |
5806 | } | |
5807 | ||
e96f2783 | 5808 | static void |
5809 | spu_trampoline_init (rtx m_tramp, tree fndecl, rtx cxt) | |
644459d0 | 5810 | { |
e96f2783 | 5811 | rtx fnaddr = XEXP (DECL_RTL (fndecl), 0); |
644459d0 | 5812 | rtx shuf = gen_reg_rtx (V4SImode); |
5813 | rtx insn = gen_reg_rtx (V4SImode); | |
5814 | rtx shufc; | |
5815 | rtx insnc; | |
5816 | rtx mem; | |
5817 | ||
5818 | fnaddr = force_reg (SImode, fnaddr); | |
5819 | cxt = force_reg (SImode, cxt); | |
5820 | ||
5821 | if (TARGET_LARGE_MEM) | |
5822 | { | |
5823 | rtx rotl = gen_reg_rtx (V4SImode); | |
5824 | rtx mask = gen_reg_rtx (V4SImode); | |
5825 | rtx bi = gen_reg_rtx (SImode); | |
e96f2783 | 5826 | static unsigned char const shufa[16] = { |
644459d0 | 5827 | 2, 3, 0, 1, 18, 19, 16, 17, |
5828 | 0, 1, 2, 3, 16, 17, 18, 19 | |
5829 | }; | |
e96f2783 | 5830 | static unsigned char const insna[16] = { |
644459d0 | 5831 | 0x41, 0, 0, 79, |
5832 | 0x41, 0, 0, STATIC_CHAIN_REGNUM, | |
5833 | 0x60, 0x80, 0, 79, | |
5834 | 0x60, 0x80, 0, STATIC_CHAIN_REGNUM | |
5835 | }; | |
5836 | ||
5837 | shufc = force_reg (TImode, array_to_constant (TImode, shufa)); | |
5838 | insnc = force_reg (V4SImode, array_to_constant (V4SImode, insna)); | |
5839 | ||
5840 | emit_insn (gen_shufb (shuf, fnaddr, cxt, shufc)); | |
4d54df85 | 5841 | emit_insn (gen_vrotlv4si3 (rotl, shuf, spu_const (V4SImode, 7))); |
644459d0 | 5842 | emit_insn (gen_movv4si (mask, spu_const (V4SImode, 0xffff << 7))); |
5843 | emit_insn (gen_selb (insn, insnc, rotl, mask)); | |
5844 | ||
e96f2783 | 5845 | mem = adjust_address (m_tramp, V4SImode, 0); |
5846 | emit_move_insn (mem, insn); | |
644459d0 | 5847 | |
5848 | emit_move_insn (bi, GEN_INT (0x35000000 + (79 << 7))); | |
e96f2783 | 5849 | mem = adjust_address (m_tramp, Pmode, 16); |
5850 | emit_move_insn (mem, bi); | |
644459d0 | 5851 | } |
5852 | else | |
5853 | { | |
5854 | rtx scxt = gen_reg_rtx (SImode); | |
5855 | rtx sfnaddr = gen_reg_rtx (SImode); | |
e96f2783 | 5856 | static unsigned char const insna[16] = { |
644459d0 | 5857 | 0x42, 0, 0, STATIC_CHAIN_REGNUM, |
5858 | 0x30, 0, 0, 0, | |
5859 | 0, 0, 0, 0, | |
5860 | 0, 0, 0, 0 | |
5861 | }; | |
5862 | ||
5863 | shufc = gen_reg_rtx (TImode); | |
5864 | insnc = force_reg (V4SImode, array_to_constant (V4SImode, insna)); | |
5865 | ||
5866 | /* By or'ing all of cxt with the ila opcode we are assuming cxt | |
5867 | fits 18 bits and the last 4 are zeros. This will be true if | |
5868 | the stack pointer is initialized to 0x3fff0 at program start, | |
5869 | otherwise the ila instruction will be garbage. */ | |
5870 | ||
5871 | emit_insn (gen_ashlsi3 (scxt, cxt, GEN_INT (7))); | |
5872 | emit_insn (gen_ashlsi3 (sfnaddr, fnaddr, GEN_INT (5))); | |
5873 | emit_insn (gen_cpat | |
5874 | (shufc, stack_pointer_rtx, GEN_INT (4), GEN_INT (4))); | |
5875 | emit_insn (gen_shufb (shuf, sfnaddr, scxt, shufc)); | |
5876 | emit_insn (gen_iorv4si3 (insn, insnc, shuf)); | |
5877 | ||
e96f2783 | 5878 | mem = adjust_address (m_tramp, V4SImode, 0); |
5879 | emit_move_insn (mem, insn); | |
644459d0 | 5880 | } |
5881 | emit_insn (gen_sync ()); | |
5882 | } | |
5883 | ||
08c6cbd2 | 5884 | static bool |
5885 | spu_warn_func_return (tree decl) | |
5886 | { | |
5887 | /* Naked functions are implemented entirely in assembly, including the | |
5888 | return sequence, so suppress warnings about this. */ | |
5889 | return !spu_naked_function_p (decl); | |
5890 | } | |
5891 | ||
644459d0 | 5892 | void |
5893 | spu_expand_sign_extend (rtx ops[]) | |
5894 | { | |
5895 | unsigned char arr[16]; | |
5896 | rtx pat = gen_reg_rtx (TImode); | |
5897 | rtx sign, c; | |
5898 | int i, last; | |
5899 | last = GET_MODE (ops[0]) == DImode ? 7 : 15; | |
5900 | if (GET_MODE (ops[1]) == QImode) | |
5901 | { | |
5902 | sign = gen_reg_rtx (HImode); | |
5903 | emit_insn (gen_extendqihi2 (sign, ops[1])); | |
5904 | for (i = 0; i < 16; i++) | |
5905 | arr[i] = 0x12; | |
5906 | arr[last] = 0x13; | |
5907 | } | |
5908 | else | |
5909 | { | |
5910 | for (i = 0; i < 16; i++) | |
5911 | arr[i] = 0x10; | |
5912 | switch (GET_MODE (ops[1])) | |
5913 | { | |
5914 | case HImode: | |
5915 | sign = gen_reg_rtx (SImode); | |
5916 | emit_insn (gen_extendhisi2 (sign, ops[1])); | |
5917 | arr[last] = 0x03; | |
5918 | arr[last - 1] = 0x02; | |
5919 | break; | |
5920 | case SImode: | |
5921 | sign = gen_reg_rtx (SImode); | |
5922 | emit_insn (gen_ashrsi3 (sign, ops[1], GEN_INT (31))); | |
5923 | for (i = 0; i < 4; i++) | |
5924 | arr[last - i] = 3 - i; | |
5925 | break; | |
5926 | case DImode: | |
5927 | sign = gen_reg_rtx (SImode); | |
5928 | c = gen_reg_rtx (SImode); | |
5929 | emit_insn (gen_spu_convert (c, ops[1])); | |
5930 | emit_insn (gen_ashrsi3 (sign, c, GEN_INT (31))); | |
5931 | for (i = 0; i < 8; i++) | |
5932 | arr[last - i] = 7 - i; | |
5933 | break; | |
5934 | default: | |
5935 | abort (); | |
5936 | } | |
5937 | } | |
5938 | emit_move_insn (pat, array_to_constant (TImode, arr)); | |
5939 | emit_insn (gen_shufb (ops[0], ops[1], sign, pat)); | |
5940 | } | |
5941 | ||
5942 | /* expand vector initialization. If there are any constant parts, | |
5943 | load constant parts first. Then load any non-constant parts. */ | |
5944 | void | |
5945 | spu_expand_vector_init (rtx target, rtx vals) | |
5946 | { | |
3754d046 | 5947 | machine_mode mode = GET_MODE (target); |
644459d0 | 5948 | int n_elts = GET_MODE_NUNITS (mode); |
5949 | int n_var = 0; | |
5950 | bool all_same = true; | |
790c536c | 5951 | rtx first, x = NULL_RTX, first_constant = NULL_RTX; |
644459d0 | 5952 | int i; |
5953 | ||
5954 | first = XVECEXP (vals, 0, 0); | |
5955 | for (i = 0; i < n_elts; ++i) | |
5956 | { | |
5957 | x = XVECEXP (vals, 0, i); | |
e442af0b | 5958 | if (!(CONST_INT_P (x) |
5959 | || GET_CODE (x) == CONST_DOUBLE | |
5960 | || GET_CODE (x) == CONST_FIXED)) | |
644459d0 | 5961 | ++n_var; |
5962 | else | |
5963 | { | |
5964 | if (first_constant == NULL_RTX) | |
5965 | first_constant = x; | |
5966 | } | |
5967 | if (i > 0 && !rtx_equal_p (x, first)) | |
5968 | all_same = false; | |
5969 | } | |
5970 | ||
5971 | /* if all elements are the same, use splats to repeat elements */ | |
5972 | if (all_same) | |
5973 | { | |
5974 | if (!CONSTANT_P (first) | |
5975 | && !register_operand (first, GET_MODE (x))) | |
5976 | first = force_reg (GET_MODE (first), first); | |
5977 | emit_insn (gen_spu_splats (target, first)); | |
5978 | return; | |
5979 | } | |
5980 | ||
5981 | /* load constant parts */ | |
5982 | if (n_var != n_elts) | |
5983 | { | |
5984 | if (n_var == 0) | |
5985 | { | |
5986 | emit_move_insn (target, | |
5987 | gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0))); | |
5988 | } | |
5989 | else | |
5990 | { | |
5991 | rtx constant_parts_rtx = copy_rtx (vals); | |
5992 | ||
5993 | gcc_assert (first_constant != NULL_RTX); | |
5994 | /* fill empty slots with the first constant, this increases | |
5995 | our chance of using splats in the recursive call below. */ | |
5996 | for (i = 0; i < n_elts; ++i) | |
e442af0b | 5997 | { |
5998 | x = XVECEXP (constant_parts_rtx, 0, i); | |
5999 | if (!(CONST_INT_P (x) | |
6000 | || GET_CODE (x) == CONST_DOUBLE | |
6001 | || GET_CODE (x) == CONST_FIXED)) | |
6002 | XVECEXP (constant_parts_rtx, 0, i) = first_constant; | |
6003 | } | |
644459d0 | 6004 | |
6005 | spu_expand_vector_init (target, constant_parts_rtx); | |
6006 | } | |
6007 | } | |
6008 | ||
6009 | /* load variable parts */ | |
6010 | if (n_var != 0) | |
6011 | { | |
6012 | rtx insert_operands[4]; | |
6013 | ||
6014 | insert_operands[0] = target; | |
6015 | insert_operands[2] = target; | |
6016 | for (i = 0; i < n_elts; ++i) | |
6017 | { | |
6018 | x = XVECEXP (vals, 0, i); | |
e442af0b | 6019 | if (!(CONST_INT_P (x) |
6020 | || GET_CODE (x) == CONST_DOUBLE | |
6021 | || GET_CODE (x) == CONST_FIXED)) | |
644459d0 | 6022 | { |
6023 | if (!register_operand (x, GET_MODE (x))) | |
6024 | x = force_reg (GET_MODE (x), x); | |
6025 | insert_operands[1] = x; | |
6026 | insert_operands[3] = GEN_INT (i); | |
6027 | spu_builtin_insert (insert_operands); | |
6028 | } | |
6029 | } | |
6030 | } | |
6031 | } | |
6352eedf | 6032 | |
5474166e | 6033 | /* Return insn index for the vector compare instruction for given CODE, |
6034 | and DEST_MODE, OP_MODE. Return -1 if valid insn is not available. */ | |
6035 | ||
6036 | static int | |
6037 | get_vec_cmp_insn (enum rtx_code code, | |
3754d046 | 6038 | machine_mode dest_mode, |
6039 | machine_mode op_mode) | |
5474166e | 6040 | |
6041 | { | |
6042 | switch (code) | |
6043 | { | |
6044 | case EQ: | |
6045 | if (dest_mode == V16QImode && op_mode == V16QImode) | |
6046 | return CODE_FOR_ceq_v16qi; | |
6047 | if (dest_mode == V8HImode && op_mode == V8HImode) | |
6048 | return CODE_FOR_ceq_v8hi; | |
6049 | if (dest_mode == V4SImode && op_mode == V4SImode) | |
6050 | return CODE_FOR_ceq_v4si; | |
6051 | if (dest_mode == V4SImode && op_mode == V4SFmode) | |
6052 | return CODE_FOR_ceq_v4sf; | |
6053 | if (dest_mode == V2DImode && op_mode == V2DFmode) | |
6054 | return CODE_FOR_ceq_v2df; | |
6055 | break; | |
6056 | case GT: | |
6057 | if (dest_mode == V16QImode && op_mode == V16QImode) | |
6058 | return CODE_FOR_cgt_v16qi; | |
6059 | if (dest_mode == V8HImode && op_mode == V8HImode) | |
6060 | return CODE_FOR_cgt_v8hi; | |
6061 | if (dest_mode == V4SImode && op_mode == V4SImode) | |
6062 | return CODE_FOR_cgt_v4si; | |
6063 | if (dest_mode == V4SImode && op_mode == V4SFmode) | |
6064 | return CODE_FOR_cgt_v4sf; | |
6065 | if (dest_mode == V2DImode && op_mode == V2DFmode) | |
6066 | return CODE_FOR_cgt_v2df; | |
6067 | break; | |
6068 | case GTU: | |
6069 | if (dest_mode == V16QImode && op_mode == V16QImode) | |
6070 | return CODE_FOR_clgt_v16qi; | |
6071 | if (dest_mode == V8HImode && op_mode == V8HImode) | |
6072 | return CODE_FOR_clgt_v8hi; | |
6073 | if (dest_mode == V4SImode && op_mode == V4SImode) | |
6074 | return CODE_FOR_clgt_v4si; | |
6075 | break; | |
6076 | default: | |
6077 | break; | |
6078 | } | |
6079 | return -1; | |
6080 | } | |
6081 | ||
6082 | /* Emit vector compare for operands OP0 and OP1 using code RCODE. | |
6083 | DMODE is expected destination mode. This is a recursive function. */ | |
6084 | ||
6085 | static rtx | |
6086 | spu_emit_vector_compare (enum rtx_code rcode, | |
6087 | rtx op0, rtx op1, | |
3754d046 | 6088 | machine_mode dmode) |
5474166e | 6089 | { |
6090 | int vec_cmp_insn; | |
6091 | rtx mask; | |
3754d046 | 6092 | machine_mode dest_mode; |
6093 | machine_mode op_mode = GET_MODE (op1); | |
5474166e | 6094 | |
6095 | gcc_assert (GET_MODE (op0) == GET_MODE (op1)); | |
6096 | ||
6097 | /* Floating point vector compare instructions uses destination V4SImode. | |
6098 | Double floating point vector compare instructions uses destination V2DImode. | |
6099 | Move destination to appropriate mode later. */ | |
6100 | if (dmode == V4SFmode) | |
6101 | dest_mode = V4SImode; | |
6102 | else if (dmode == V2DFmode) | |
6103 | dest_mode = V2DImode; | |
6104 | else | |
6105 | dest_mode = dmode; | |
6106 | ||
6107 | mask = gen_reg_rtx (dest_mode); | |
6108 | vec_cmp_insn = get_vec_cmp_insn (rcode, dest_mode, op_mode); | |
6109 | ||
6110 | if (vec_cmp_insn == -1) | |
6111 | { | |
6112 | bool swap_operands = false; | |
6113 | bool try_again = false; | |
6114 | switch (rcode) | |
6115 | { | |
6116 | case LT: | |
6117 | rcode = GT; | |
6118 | swap_operands = true; | |
6119 | try_again = true; | |
6120 | break; | |
6121 | case LTU: | |
6122 | rcode = GTU; | |
6123 | swap_operands = true; | |
6124 | try_again = true; | |
6125 | break; | |
6126 | case NE: | |
e20943d4 | 6127 | case UNEQ: |
6128 | case UNLE: | |
6129 | case UNLT: | |
6130 | case UNGE: | |
6131 | case UNGT: | |
6132 | case UNORDERED: | |
5474166e | 6133 | /* Treat A != B as ~(A==B). */ |
6134 | { | |
e20943d4 | 6135 | enum rtx_code rev_code; |
5474166e | 6136 | enum insn_code nor_code; |
e20943d4 | 6137 | rtx rev_mask; |
6138 | ||
6139 | rev_code = reverse_condition_maybe_unordered (rcode); | |
6140 | rev_mask = spu_emit_vector_compare (rev_code, op0, op1, dest_mode); | |
6141 | ||
d6bf3b14 | 6142 | nor_code = optab_handler (one_cmpl_optab, dest_mode); |
5474166e | 6143 | gcc_assert (nor_code != CODE_FOR_nothing); |
e20943d4 | 6144 | emit_insn (GEN_FCN (nor_code) (mask, rev_mask)); |
5474166e | 6145 | if (dmode != dest_mode) |
6146 | { | |
6147 | rtx temp = gen_reg_rtx (dest_mode); | |
6148 | convert_move (temp, mask, 0); | |
6149 | return temp; | |
6150 | } | |
6151 | return mask; | |
6152 | } | |
6153 | break; | |
6154 | case GE: | |
6155 | case GEU: | |
6156 | case LE: | |
6157 | case LEU: | |
6158 | /* Try GT/GTU/LT/LTU OR EQ */ | |
6159 | { | |
6160 | rtx c_rtx, eq_rtx; | |
6161 | enum insn_code ior_code; | |
6162 | enum rtx_code new_code; | |
6163 | ||
6164 | switch (rcode) | |
6165 | { | |
6166 | case GE: new_code = GT; break; | |
6167 | case GEU: new_code = GTU; break; | |
6168 | case LE: new_code = LT; break; | |
6169 | case LEU: new_code = LTU; break; | |
6170 | default: | |
6171 | gcc_unreachable (); | |
6172 | } | |
6173 | ||
6174 | c_rtx = spu_emit_vector_compare (new_code, op0, op1, dest_mode); | |
6175 | eq_rtx = spu_emit_vector_compare (EQ, op0, op1, dest_mode); | |
6176 | ||
d6bf3b14 | 6177 | ior_code = optab_handler (ior_optab, dest_mode); |
5474166e | 6178 | gcc_assert (ior_code != CODE_FOR_nothing); |
6179 | emit_insn (GEN_FCN (ior_code) (mask, c_rtx, eq_rtx)); | |
6180 | if (dmode != dest_mode) | |
6181 | { | |
6182 | rtx temp = gen_reg_rtx (dest_mode); | |
6183 | convert_move (temp, mask, 0); | |
6184 | return temp; | |
6185 | } | |
6186 | return mask; | |
6187 | } | |
6188 | break; | |
e20943d4 | 6189 | case LTGT: |
6190 | /* Try LT OR GT */ | |
6191 | { | |
6192 | rtx lt_rtx, gt_rtx; | |
6193 | enum insn_code ior_code; | |
6194 | ||
6195 | lt_rtx = spu_emit_vector_compare (LT, op0, op1, dest_mode); | |
6196 | gt_rtx = spu_emit_vector_compare (GT, op0, op1, dest_mode); | |
6197 | ||
6198 | ior_code = optab_handler (ior_optab, dest_mode); | |
6199 | gcc_assert (ior_code != CODE_FOR_nothing); | |
6200 | emit_insn (GEN_FCN (ior_code) (mask, lt_rtx, gt_rtx)); | |
6201 | if (dmode != dest_mode) | |
6202 | { | |
6203 | rtx temp = gen_reg_rtx (dest_mode); | |
6204 | convert_move (temp, mask, 0); | |
6205 | return temp; | |
6206 | } | |
6207 | return mask; | |
6208 | } | |
6209 | break; | |
6210 | case ORDERED: | |
6211 | /* Implement as (A==A) & (B==B) */ | |
6212 | { | |
6213 | rtx a_rtx, b_rtx; | |
6214 | enum insn_code and_code; | |
6215 | ||
6216 | a_rtx = spu_emit_vector_compare (EQ, op0, op0, dest_mode); | |
6217 | b_rtx = spu_emit_vector_compare (EQ, op1, op1, dest_mode); | |
6218 | ||
6219 | and_code = optab_handler (and_optab, dest_mode); | |
6220 | gcc_assert (and_code != CODE_FOR_nothing); | |
6221 | emit_insn (GEN_FCN (and_code) (mask, a_rtx, b_rtx)); | |
6222 | if (dmode != dest_mode) | |
6223 | { | |
6224 | rtx temp = gen_reg_rtx (dest_mode); | |
6225 | convert_move (temp, mask, 0); | |
6226 | return temp; | |
6227 | } | |
6228 | return mask; | |
6229 | } | |
6230 | break; | |
5474166e | 6231 | default: |
6232 | gcc_unreachable (); | |
6233 | } | |
6234 | ||
6235 | /* You only get two chances. */ | |
6236 | if (try_again) | |
6237 | vec_cmp_insn = get_vec_cmp_insn (rcode, dest_mode, op_mode); | |
6238 | ||
6239 | gcc_assert (vec_cmp_insn != -1); | |
6240 | ||
6241 | if (swap_operands) | |
6242 | { | |
6243 | rtx tmp; | |
6244 | tmp = op0; | |
6245 | op0 = op1; | |
6246 | op1 = tmp; | |
6247 | } | |
6248 | } | |
6249 | ||
6250 | emit_insn (GEN_FCN (vec_cmp_insn) (mask, op0, op1)); | |
6251 | if (dmode != dest_mode) | |
6252 | { | |
6253 | rtx temp = gen_reg_rtx (dest_mode); | |
6254 | convert_move (temp, mask, 0); | |
6255 | return temp; | |
6256 | } | |
6257 | return mask; | |
6258 | } | |
6259 | ||
6260 | ||
6261 | /* Emit vector conditional expression. | |
6262 | DEST is destination. OP1 and OP2 are two VEC_COND_EXPR operands. | |
6263 | CC_OP0 and CC_OP1 are the two operands for the relation operation COND. */ | |
6264 | ||
6265 | int | |
6266 | spu_emit_vector_cond_expr (rtx dest, rtx op1, rtx op2, | |
6267 | rtx cond, rtx cc_op0, rtx cc_op1) | |
6268 | { | |
3754d046 | 6269 | machine_mode dest_mode = GET_MODE (dest); |
5474166e | 6270 | enum rtx_code rcode = GET_CODE (cond); |
6271 | rtx mask; | |
6272 | ||
6273 | /* Get the vector mask for the given relational operations. */ | |
6274 | mask = spu_emit_vector_compare (rcode, cc_op0, cc_op1, dest_mode); | |
6275 | ||
6276 | emit_insn(gen_selb (dest, op2, op1, mask)); | |
6277 | ||
6278 | return 1; | |
6279 | } | |
6280 | ||
6352eedf | 6281 | static rtx |
3754d046 | 6282 | spu_force_reg (machine_mode mode, rtx op) |
6352eedf | 6283 | { |
6284 | rtx x, r; | |
6285 | if (GET_MODE (op) == VOIDmode || GET_MODE (op) == BLKmode) | |
6286 | { | |
6287 | if ((SCALAR_INT_MODE_P (mode) && GET_CODE (op) == CONST_INT) | |
6288 | || GET_MODE (op) == BLKmode) | |
6289 | return force_reg (mode, convert_to_mode (mode, op, 0)); | |
6290 | abort (); | |
6291 | } | |
6292 | ||
6293 | r = force_reg (GET_MODE (op), op); | |
6294 | if (GET_MODE_SIZE (GET_MODE (op)) == GET_MODE_SIZE (mode)) | |
6295 | { | |
6296 | x = simplify_gen_subreg (mode, r, GET_MODE (op), 0); | |
6297 | if (x) | |
6298 | return x; | |
6299 | } | |
6300 | ||
6301 | x = gen_reg_rtx (mode); | |
6302 | emit_insn (gen_spu_convert (x, r)); | |
6303 | return x; | |
6304 | } | |
6305 | ||
6306 | static void | |
6307 | spu_check_builtin_parm (struct spu_builtin_description *d, rtx op, int p) | |
6308 | { | |
6309 | HOST_WIDE_INT v = 0; | |
6310 | int lsbits; | |
6311 | /* Check the range of immediate operands. */ | |
6312 | if (p >= SPU_BTI_7 && p <= SPU_BTI_U18) | |
6313 | { | |
6314 | int range = p - SPU_BTI_7; | |
5df189be | 6315 | |
6316 | if (!CONSTANT_P (op)) | |
bf776685 | 6317 | error ("%s expects an integer literal in the range [%d, %d]", |
6352eedf | 6318 | d->name, |
6319 | spu_builtin_range[range].low, spu_builtin_range[range].high); | |
6320 | ||
6321 | if (GET_CODE (op) == CONST | |
6322 | && (GET_CODE (XEXP (op, 0)) == PLUS | |
6323 | || GET_CODE (XEXP (op, 0)) == MINUS)) | |
6324 | { | |
6325 | v = INTVAL (XEXP (XEXP (op, 0), 1)); | |
6326 | op = XEXP (XEXP (op, 0), 0); | |
6327 | } | |
6328 | else if (GET_CODE (op) == CONST_INT) | |
6329 | v = INTVAL (op); | |
5df189be | 6330 | else if (GET_CODE (op) == CONST_VECTOR |
6331 | && GET_CODE (CONST_VECTOR_ELT (op, 0)) == CONST_INT) | |
6332 | v = INTVAL (CONST_VECTOR_ELT (op, 0)); | |
6333 | ||
6334 | /* The default for v is 0 which is valid in every range. */ | |
6335 | if (v < spu_builtin_range[range].low | |
6336 | || v > spu_builtin_range[range].high) | |
bf776685 | 6337 | error ("%s expects an integer literal in the range [%d, %d]. (%wd)", |
5df189be | 6338 | d->name, |
6339 | spu_builtin_range[range].low, spu_builtin_range[range].high, | |
6340 | v); | |
6352eedf | 6341 | |
6342 | switch (p) | |
6343 | { | |
6344 | case SPU_BTI_S10_4: | |
6345 | lsbits = 4; | |
6346 | break; | |
6347 | case SPU_BTI_U16_2: | |
6348 | /* This is only used in lqa, and stqa. Even though the insns | |
6349 | encode 16 bits of the address (all but the 2 least | |
6350 | significant), only 14 bits are used because it is masked to | |
6351 | be 16 byte aligned. */ | |
6352 | lsbits = 4; | |
6353 | break; | |
6354 | case SPU_BTI_S16_2: | |
6355 | /* This is used for lqr and stqr. */ | |
6356 | lsbits = 2; | |
6357 | break; | |
6358 | default: | |
6359 | lsbits = 0; | |
6360 | } | |
6361 | ||
6362 | if (GET_CODE (op) == LABEL_REF | |
6363 | || (GET_CODE (op) == SYMBOL_REF | |
6364 | && SYMBOL_REF_FUNCTION_P (op)) | |
5df189be | 6365 | || (v & ((1 << lsbits) - 1)) != 0) |
bf776685 | 6366 | warning (0, "%d least significant bits of %s are ignored", lsbits, |
6352eedf | 6367 | d->name); |
6368 | } | |
6369 | } | |
6370 | ||
6371 | ||
70ca06f8 | 6372 | static int |
5df189be | 6373 | expand_builtin_args (struct spu_builtin_description *d, tree exp, |
6352eedf | 6374 | rtx target, rtx ops[]) |
6375 | { | |
bc620c5c | 6376 | enum insn_code icode = (enum insn_code) d->icode; |
5df189be | 6377 | int i = 0, a; |
6352eedf | 6378 | |
6379 | /* Expand the arguments into rtl. */ | |
6380 | ||
6381 | if (d->parm[0] != SPU_BTI_VOID) | |
6382 | ops[i++] = target; | |
6383 | ||
70ca06f8 | 6384 | for (a = 0; d->parm[a+1] != SPU_BTI_END_OF_PARAMS; i++, a++) |
6352eedf | 6385 | { |
5df189be | 6386 | tree arg = CALL_EXPR_ARG (exp, a); |
6352eedf | 6387 | if (arg == 0) |
6388 | abort (); | |
b9c74b4d | 6389 | ops[i] = expand_expr (arg, NULL_RTX, VOIDmode, EXPAND_NORMAL); |
6352eedf | 6390 | } |
70ca06f8 | 6391 | |
32f79657 | 6392 | gcc_assert (i == insn_data[icode].n_generator_args); |
70ca06f8 | 6393 | return i; |
6352eedf | 6394 | } |
6395 | ||
6396 | static rtx | |
6397 | spu_expand_builtin_1 (struct spu_builtin_description *d, | |
5df189be | 6398 | tree exp, rtx target) |
6352eedf | 6399 | { |
6400 | rtx pat; | |
6401 | rtx ops[8]; | |
bc620c5c | 6402 | enum insn_code icode = (enum insn_code) d->icode; |
3754d046 | 6403 | machine_mode mode, tmode; |
6352eedf | 6404 | int i, p; |
70ca06f8 | 6405 | int n_operands; |
6352eedf | 6406 | tree return_type; |
6407 | ||
6408 | /* Set up ops[] with values from arglist. */ | |
70ca06f8 | 6409 | n_operands = expand_builtin_args (d, exp, target, ops); |
6352eedf | 6410 | |
6411 | /* Handle the target operand which must be operand 0. */ | |
6412 | i = 0; | |
6413 | if (d->parm[0] != SPU_BTI_VOID) | |
6414 | { | |
6415 | ||
6416 | /* We prefer the mode specified for the match_operand otherwise | |
6417 | use the mode from the builtin function prototype. */ | |
6418 | tmode = insn_data[d->icode].operand[0].mode; | |
6419 | if (tmode == VOIDmode) | |
6420 | tmode = TYPE_MODE (spu_builtin_types[d->parm[0]]); | |
6421 | ||
6422 | /* Try to use target because not using it can lead to extra copies | |
6423 | and when we are using all of the registers extra copies leads | |
6424 | to extra spills. */ | |
6425 | if (target && GET_CODE (target) == REG && GET_MODE (target) == tmode) | |
6426 | ops[0] = target; | |
6427 | else | |
6428 | target = ops[0] = gen_reg_rtx (tmode); | |
6429 | ||
6430 | if (!(*insn_data[icode].operand[0].predicate) (ops[0], tmode)) | |
6431 | abort (); | |
6432 | ||
6433 | i++; | |
6434 | } | |
6435 | ||
a76866d3 | 6436 | if (d->fcode == SPU_MASK_FOR_LOAD) |
6437 | { | |
3754d046 | 6438 | machine_mode mode = insn_data[icode].operand[1].mode; |
a76866d3 | 6439 | tree arg; |
6440 | rtx addr, op, pat; | |
6441 | ||
6442 | /* get addr */ | |
5df189be | 6443 | arg = CALL_EXPR_ARG (exp, 0); |
4b8ee66a | 6444 | gcc_assert (POINTER_TYPE_P (TREE_TYPE (arg))); |
a76866d3 | 6445 | op = expand_expr (arg, NULL_RTX, Pmode, EXPAND_NORMAL); |
6446 | addr = memory_address (mode, op); | |
6447 | ||
6448 | /* negate addr */ | |
6449 | op = gen_reg_rtx (GET_MODE (addr)); | |
d1f9b275 | 6450 | emit_insn (gen_rtx_SET (op, gen_rtx_NEG (GET_MODE (addr), addr))); |
a76866d3 | 6451 | op = gen_rtx_MEM (mode, op); |
6452 | ||
6453 | pat = GEN_FCN (icode) (target, op); | |
6454 | if (!pat) | |
6455 | return 0; | |
6456 | emit_insn (pat); | |
6457 | return target; | |
6458 | } | |
6459 | ||
6352eedf | 6460 | /* Ignore align_hint, but still expand it's args in case they have |
6461 | side effects. */ | |
6462 | if (icode == CODE_FOR_spu_align_hint) | |
6463 | return 0; | |
6464 | ||
6465 | /* Handle the rest of the operands. */ | |
70ca06f8 | 6466 | for (p = 1; i < n_operands; i++, p++) |
6352eedf | 6467 | { |
6468 | if (insn_data[d->icode].operand[i].mode != VOIDmode) | |
6469 | mode = insn_data[d->icode].operand[i].mode; | |
6470 | else | |
6471 | mode = TYPE_MODE (spu_builtin_types[d->parm[i]]); | |
6472 | ||
6473 | /* mode can be VOIDmode here for labels */ | |
6474 | ||
6475 | /* For specific intrinsics with an immediate operand, e.g., | |
6476 | si_ai(), we sometimes need to convert the scalar argument to a | |
6477 | vector argument by splatting the scalar. */ | |
6478 | if (VECTOR_MODE_P (mode) | |
6479 | && (GET_CODE (ops[i]) == CONST_INT | |
6480 | || GET_MODE_CLASS (GET_MODE (ops[i])) == MODE_INT | |
3b442530 | 6481 | || GET_MODE_CLASS (GET_MODE (ops[i])) == MODE_FLOAT)) |
6352eedf | 6482 | { |
6483 | if (GET_CODE (ops[i]) == CONST_INT) | |
6484 | ops[i] = spu_const (mode, INTVAL (ops[i])); | |
6485 | else | |
6486 | { | |
6487 | rtx reg = gen_reg_rtx (mode); | |
3754d046 | 6488 | machine_mode imode = GET_MODE_INNER (mode); |
6352eedf | 6489 | if (!spu_nonmem_operand (ops[i], GET_MODE (ops[i]))) |
6490 | ops[i] = force_reg (GET_MODE (ops[i]), ops[i]); | |
6491 | if (imode != GET_MODE (ops[i])) | |
6492 | ops[i] = convert_to_mode (imode, ops[i], | |
6493 | TYPE_UNSIGNED (spu_builtin_types | |
6494 | [d->parm[i]])); | |
6495 | emit_insn (gen_spu_splats (reg, ops[i])); | |
6496 | ops[i] = reg; | |
6497 | } | |
6498 | } | |
6499 | ||
5df189be | 6500 | spu_check_builtin_parm (d, ops[i], d->parm[p]); |
6501 | ||
6352eedf | 6502 | if (!(*insn_data[icode].operand[i].predicate) (ops[i], mode)) |
6503 | ops[i] = spu_force_reg (mode, ops[i]); | |
6352eedf | 6504 | } |
6505 | ||
70ca06f8 | 6506 | switch (n_operands) |
6352eedf | 6507 | { |
6508 | case 0: | |
6509 | pat = GEN_FCN (icode) (0); | |
6510 | break; | |
6511 | case 1: | |
6512 | pat = GEN_FCN (icode) (ops[0]); | |
6513 | break; | |
6514 | case 2: | |
6515 | pat = GEN_FCN (icode) (ops[0], ops[1]); | |
6516 | break; | |
6517 | case 3: | |
6518 | pat = GEN_FCN (icode) (ops[0], ops[1], ops[2]); | |
6519 | break; | |
6520 | case 4: | |
6521 | pat = GEN_FCN (icode) (ops[0], ops[1], ops[2], ops[3]); | |
6522 | break; | |
6523 | case 5: | |
6524 | pat = GEN_FCN (icode) (ops[0], ops[1], ops[2], ops[3], ops[4]); | |
6525 | break; | |
6526 | case 6: | |
6527 | pat = GEN_FCN (icode) (ops[0], ops[1], ops[2], ops[3], ops[4], ops[5]); | |
6528 | break; | |
6529 | default: | |
6530 | abort (); | |
6531 | } | |
6532 | ||
6533 | if (!pat) | |
6534 | abort (); | |
6535 | ||
6536 | if (d->type == B_CALL || d->type == B_BISLED) | |
6537 | emit_call_insn (pat); | |
6538 | else if (d->type == B_JUMP) | |
6539 | { | |
6540 | emit_jump_insn (pat); | |
6541 | emit_barrier (); | |
6542 | } | |
6543 | else | |
6544 | emit_insn (pat); | |
6545 | ||
6546 | return_type = spu_builtin_types[d->parm[0]]; | |
6547 | if (d->parm[0] != SPU_BTI_VOID | |
6548 | && GET_MODE (target) != TYPE_MODE (return_type)) | |
6549 | { | |
6550 | /* target is the return value. It should always be the mode of | |
6551 | the builtin function prototype. */ | |
6552 | target = spu_force_reg (TYPE_MODE (return_type), target); | |
6553 | } | |
6554 | ||
6555 | return target; | |
6556 | } | |
6557 | ||
6558 | rtx | |
6559 | spu_expand_builtin (tree exp, | |
6560 | rtx target, | |
6561 | rtx subtarget ATTRIBUTE_UNUSED, | |
3754d046 | 6562 | machine_mode mode ATTRIBUTE_UNUSED, |
6352eedf | 6563 | int ignore ATTRIBUTE_UNUSED) |
6564 | { | |
5df189be | 6565 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); |
3726fe5e | 6566 | unsigned int fcode = DECL_FUNCTION_CODE (fndecl); |
6352eedf | 6567 | struct spu_builtin_description *d; |
6568 | ||
6569 | if (fcode < NUM_SPU_BUILTINS) | |
6570 | { | |
6571 | d = &spu_builtins[fcode]; | |
6572 | ||
5df189be | 6573 | return spu_expand_builtin_1 (d, exp, target); |
6352eedf | 6574 | } |
6575 | abort (); | |
6576 | } | |
6577 | ||
a76866d3 | 6578 | /* Implement targetm.vectorize.builtin_mask_for_load. */ |
6579 | static tree | |
6580 | spu_builtin_mask_for_load (void) | |
6581 | { | |
0c5c4d59 | 6582 | return spu_builtin_decls[SPU_MASK_FOR_LOAD]; |
a76866d3 | 6583 | } |
5df189be | 6584 | |
a28df51d | 6585 | /* Implement targetm.vectorize.builtin_vectorization_cost. */ |
6586 | static int | |
0822b158 | 6587 | spu_builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost, |
d13adc77 | 6588 | tree vectype, |
0822b158 | 6589 | int misalign ATTRIBUTE_UNUSED) |
559093aa | 6590 | { |
d13adc77 | 6591 | unsigned elements; |
6592 | ||
559093aa | 6593 | switch (type_of_cost) |
6594 | { | |
6595 | case scalar_stmt: | |
6596 | case vector_stmt: | |
6597 | case vector_load: | |
6598 | case vector_store: | |
6599 | case vec_to_scalar: | |
6600 | case scalar_to_vec: | |
6601 | case cond_branch_not_taken: | |
6602 | case vec_perm: | |
5df2530b | 6603 | case vec_promote_demote: |
559093aa | 6604 | return 1; |
6605 | ||
6606 | case scalar_store: | |
6607 | return 10; | |
6608 | ||
6609 | case scalar_load: | |
6610 | /* Load + rotate. */ | |
6611 | return 2; | |
6612 | ||
6613 | case unaligned_load: | |
6614 | return 2; | |
6615 | ||
6616 | case cond_branch_taken: | |
6617 | return 6; | |
6618 | ||
d13adc77 | 6619 | case vec_construct: |
6620 | elements = TYPE_VECTOR_SUBPARTS (vectype); | |
6621 | return elements / 2 + 1; | |
6622 | ||
559093aa | 6623 | default: |
6624 | gcc_unreachable (); | |
6625 | } | |
a28df51d | 6626 | } |
6627 | ||
4db2b577 | 6628 | /* Implement targetm.vectorize.init_cost. */ |
6629 | ||
61b33788 | 6630 | static void * |
4db2b577 | 6631 | spu_init_cost (struct loop *loop_info ATTRIBUTE_UNUSED) |
6632 | { | |
f97dec81 | 6633 | unsigned *cost = XNEWVEC (unsigned, 3); |
6634 | cost[vect_prologue] = cost[vect_body] = cost[vect_epilogue] = 0; | |
4db2b577 | 6635 | return cost; |
6636 | } | |
6637 | ||
6638 | /* Implement targetm.vectorize.add_stmt_cost. */ | |
6639 | ||
61b33788 | 6640 | static unsigned |
4db2b577 | 6641 | spu_add_stmt_cost (void *data, int count, enum vect_cost_for_stmt kind, |
f97dec81 | 6642 | struct _stmt_vec_info *stmt_info, int misalign, |
6643 | enum vect_cost_model_location where) | |
4db2b577 | 6644 | { |
6645 | unsigned *cost = (unsigned *) data; | |
6646 | unsigned retval = 0; | |
6647 | ||
6648 | if (flag_vect_cost_model) | |
6649 | { | |
f97dec81 | 6650 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; |
4db2b577 | 6651 | int stmt_cost = spu_builtin_vectorization_cost (kind, vectype, misalign); |
6652 | ||
6653 | /* Statements in an inner loop relative to the loop being | |
6654 | vectorized are weighted more heavily. The value here is | |
6655 | arbitrary and could potentially be improved with analysis. */ | |
f97dec81 | 6656 | if (where == vect_body && stmt_info && stmt_in_inner_loop_p (stmt_info)) |
4db2b577 | 6657 | count *= 50; /* FIXME. */ |
6658 | ||
6659 | retval = (unsigned) (count * stmt_cost); | |
f97dec81 | 6660 | cost[where] += retval; |
4db2b577 | 6661 | } |
6662 | ||
6663 | return retval; | |
6664 | } | |
6665 | ||
6666 | /* Implement targetm.vectorize.finish_cost. */ | |
6667 | ||
f97dec81 | 6668 | static void |
6669 | spu_finish_cost (void *data, unsigned *prologue_cost, | |
6670 | unsigned *body_cost, unsigned *epilogue_cost) | |
4db2b577 | 6671 | { |
f97dec81 | 6672 | unsigned *cost = (unsigned *) data; |
6673 | *prologue_cost = cost[vect_prologue]; | |
6674 | *body_cost = cost[vect_body]; | |
6675 | *epilogue_cost = cost[vect_epilogue]; | |
4db2b577 | 6676 | } |
6677 | ||
6678 | /* Implement targetm.vectorize.destroy_cost_data. */ | |
6679 | ||
61b33788 | 6680 | static void |
4db2b577 | 6681 | spu_destroy_cost_data (void *data) |
6682 | { | |
6683 | free (data); | |
6684 | } | |
6685 | ||
0e87db76 | 6686 | /* Return true iff, data reference of TYPE can reach vector alignment (16) |
6687 | after applying N number of iterations. This routine does not determine | |
6688 | how may iterations are required to reach desired alignment. */ | |
6689 | ||
6690 | static bool | |
a9f1838b | 6691 | spu_vector_alignment_reachable (const_tree type ATTRIBUTE_UNUSED, bool is_packed) |
0e87db76 | 6692 | { |
6693 | if (is_packed) | |
6694 | return false; | |
6695 | ||
6696 | /* All other types are naturally aligned. */ | |
6697 | return true; | |
6698 | } | |
6699 | ||
6cf5579e | 6700 | /* Return the appropriate mode for a named address pointer. */ |
3754d046 | 6701 | static machine_mode |
6cf5579e | 6702 | spu_addr_space_pointer_mode (addr_space_t addrspace) |
6703 | { | |
6704 | switch (addrspace) | |
6705 | { | |
6706 | case ADDR_SPACE_GENERIC: | |
6707 | return ptr_mode; | |
6708 | case ADDR_SPACE_EA: | |
6709 | return EAmode; | |
6710 | default: | |
6711 | gcc_unreachable (); | |
6712 | } | |
6713 | } | |
6714 | ||
6715 | /* Return the appropriate mode for a named address address. */ | |
3754d046 | 6716 | static machine_mode |
6cf5579e | 6717 | spu_addr_space_address_mode (addr_space_t addrspace) |
6718 | { | |
6719 | switch (addrspace) | |
6720 | { | |
6721 | case ADDR_SPACE_GENERIC: | |
6722 | return Pmode; | |
6723 | case ADDR_SPACE_EA: | |
6724 | return EAmode; | |
6725 | default: | |
6726 | gcc_unreachable (); | |
6727 | } | |
6728 | } | |
6729 | ||
6730 | /* Determine if one named address space is a subset of another. */ | |
6731 | ||
6732 | static bool | |
6733 | spu_addr_space_subset_p (addr_space_t subset, addr_space_t superset) | |
6734 | { | |
6735 | gcc_assert (subset == ADDR_SPACE_GENERIC || subset == ADDR_SPACE_EA); | |
6736 | gcc_assert (superset == ADDR_SPACE_GENERIC || superset == ADDR_SPACE_EA); | |
6737 | ||
6738 | if (subset == superset) | |
6739 | return true; | |
6740 | ||
6741 | /* If we have -mno-address-space-conversion, treat __ea and generic as not | |
6742 | being subsets but instead as disjoint address spaces. */ | |
6743 | else if (!TARGET_ADDRESS_SPACE_CONVERSION) | |
6744 | return false; | |
6745 | ||
6746 | else | |
6747 | return (subset == ADDR_SPACE_GENERIC && superset == ADDR_SPACE_EA); | |
6748 | } | |
6749 | ||
6750 | /* Convert from one address space to another. */ | |
6751 | static rtx | |
6752 | spu_addr_space_convert (rtx op, tree from_type, tree to_type) | |
6753 | { | |
6754 | addr_space_t from_as = TYPE_ADDR_SPACE (TREE_TYPE (from_type)); | |
6755 | addr_space_t to_as = TYPE_ADDR_SPACE (TREE_TYPE (to_type)); | |
6756 | ||
6757 | gcc_assert (from_as == ADDR_SPACE_GENERIC || from_as == ADDR_SPACE_EA); | |
6758 | gcc_assert (to_as == ADDR_SPACE_GENERIC || to_as == ADDR_SPACE_EA); | |
6759 | ||
6760 | if (to_as == ADDR_SPACE_GENERIC && from_as == ADDR_SPACE_EA) | |
6761 | { | |
6762 | rtx result, ls; | |
6763 | ||
6764 | ls = gen_const_mem (DImode, | |
6765 | gen_rtx_SYMBOL_REF (Pmode, "__ea_local_store")); | |
6766 | set_mem_align (ls, 128); | |
6767 | ||
6768 | result = gen_reg_rtx (Pmode); | |
6769 | ls = force_reg (Pmode, convert_modes (Pmode, DImode, ls, 1)); | |
6770 | op = force_reg (Pmode, convert_modes (Pmode, EAmode, op, 1)); | |
6771 | ls = emit_conditional_move (ls, NE, op, const0_rtx, Pmode, | |
6772 | ls, const0_rtx, Pmode, 1); | |
6773 | ||
6774 | emit_insn (gen_subsi3 (result, op, ls)); | |
6775 | ||
6776 | return result; | |
6777 | } | |
6778 | ||
6779 | else if (to_as == ADDR_SPACE_EA && from_as == ADDR_SPACE_GENERIC) | |
6780 | { | |
6781 | rtx result, ls; | |
6782 | ||
6783 | ls = gen_const_mem (DImode, | |
6784 | gen_rtx_SYMBOL_REF (Pmode, "__ea_local_store")); | |
6785 | set_mem_align (ls, 128); | |
6786 | ||
6787 | result = gen_reg_rtx (EAmode); | |
6788 | ls = force_reg (EAmode, convert_modes (EAmode, DImode, ls, 1)); | |
6789 | op = force_reg (Pmode, op); | |
6790 | ls = emit_conditional_move (ls, NE, op, const0_rtx, Pmode, | |
6791 | ls, const0_rtx, EAmode, 1); | |
6792 | op = force_reg (EAmode, convert_modes (EAmode, Pmode, op, 1)); | |
6793 | ||
6794 | if (EAmode == SImode) | |
6795 | emit_insn (gen_addsi3 (result, op, ls)); | |
6796 | else | |
6797 | emit_insn (gen_adddi3 (result, op, ls)); | |
6798 | ||
6799 | return result; | |
6800 | } | |
6801 | ||
6802 | else | |
6803 | gcc_unreachable (); | |
6804 | } | |
6805 | ||
6806 | ||
d52fd16a | 6807 | /* Count the total number of instructions in each pipe and return the |
6808 | maximum, which is used as the Minimum Iteration Interval (MII) | |
6809 | in the modulo scheduler. get_pipe() will return -2, -1, 0, or 1. | |
6810 | -2 are instructions that can go in pipe0 or pipe1. */ | |
6811 | static int | |
6812 | spu_sms_res_mii (struct ddg *g) | |
6813 | { | |
6814 | int i; | |
6815 | unsigned t[4] = {0, 0, 0, 0}; | |
6816 | ||
6817 | for (i = 0; i < g->num_nodes; i++) | |
6818 | { | |
0af56f80 | 6819 | rtx_insn *insn = g->nodes[i].insn; |
d52fd16a | 6820 | int p = get_pipe (insn) + 2; |
6821 | ||
1e944a0b | 6822 | gcc_assert (p >= 0); |
6823 | gcc_assert (p < 4); | |
d52fd16a | 6824 | |
6825 | t[p]++; | |
6826 | if (dump_file && INSN_P (insn)) | |
6827 | fprintf (dump_file, "i%d %s %d %d\n", | |
6828 | INSN_UID (insn), | |
6829 | insn_data[INSN_CODE(insn)].name, | |
6830 | p, t[p]); | |
6831 | } | |
6832 | if (dump_file) | |
6833 | fprintf (dump_file, "%d %d %d %d\n", t[0], t[1], t[2], t[3]); | |
6834 | ||
6835 | return MAX ((t[0] + t[2] + t[3] + 1) / 2, MAX (t[2], t[3])); | |
6836 | } | |
6837 | ||
6838 | ||
5df189be | 6839 | void |
6840 | spu_init_expanders (void) | |
9d98604b | 6841 | { |
5df189be | 6842 | if (cfun) |
9d98604b | 6843 | { |
6844 | rtx r0, r1; | |
6845 | /* HARD_FRAME_REGISTER is only 128 bit aligned when | |
6846 | frame_pointer_needed is true. We don't know that until we're | |
6847 | expanding the prologue. */ | |
6848 | REGNO_POINTER_ALIGN (HARD_FRAME_POINTER_REGNUM) = 8; | |
6849 | ||
6850 | /* A number of passes use LAST_VIRTUAL_REGISTER+1 and | |
6851 | LAST_VIRTUAL_REGISTER+2 to test the back-end. We want them | |
6852 | to be treated as aligned, so generate them here. */ | |
6853 | r0 = gen_reg_rtx (SImode); | |
6854 | r1 = gen_reg_rtx (SImode); | |
6855 | mark_reg_pointer (r0, 128); | |
6856 | mark_reg_pointer (r1, 128); | |
6857 | gcc_assert (REGNO (r0) == LAST_VIRTUAL_REGISTER + 1 | |
6858 | && REGNO (r1) == LAST_VIRTUAL_REGISTER + 2); | |
6859 | } | |
ea32e033 | 6860 | } |
6861 | ||
3754d046 | 6862 | static machine_mode |
ea32e033 | 6863 | spu_libgcc_cmp_return_mode (void) |
6864 | { | |
6865 | ||
6866 | /* For SPU word mode is TI mode so it is better to use SImode | |
6867 | for compare returns. */ | |
6868 | return SImode; | |
6869 | } | |
6870 | ||
3754d046 | 6871 | static machine_mode |
ea32e033 | 6872 | spu_libgcc_shift_count_mode (void) |
6873 | { | |
6874 | /* For SPU word mode is TI mode so it is better to use SImode | |
6875 | for shift counts. */ | |
6876 | return SImode; | |
6877 | } | |
5a976006 | 6878 | |
a08dfd55 | 6879 | /* Implement targetm.section_type_flags. */ |
6880 | static unsigned int | |
6881 | spu_section_type_flags (tree decl, const char *name, int reloc) | |
6882 | { | |
6883 | /* .toe needs to have type @nobits. */ | |
6884 | if (strcmp (name, ".toe") == 0) | |
6885 | return SECTION_BSS; | |
6cf5579e | 6886 | /* Don't load _ea into the current address space. */ |
6887 | if (strcmp (name, "._ea") == 0) | |
6888 | return SECTION_WRITE | SECTION_DEBUG; | |
a08dfd55 | 6889 | return default_section_type_flags (decl, name, reloc); |
6890 | } | |
c2233b46 | 6891 | |
6cf5579e | 6892 | /* Implement targetm.select_section. */ |
6893 | static section * | |
6894 | spu_select_section (tree decl, int reloc, unsigned HOST_WIDE_INT align) | |
6895 | { | |
6896 | /* Variables and constants defined in the __ea address space | |
6897 | go into a special section named "._ea". */ | |
6898 | if (TREE_TYPE (decl) != error_mark_node | |
6899 | && TYPE_ADDR_SPACE (TREE_TYPE (decl)) == ADDR_SPACE_EA) | |
6900 | { | |
6901 | /* We might get called with string constants, but get_named_section | |
6902 | doesn't like them as they are not DECLs. Also, we need to set | |
6903 | flags in that case. */ | |
6904 | if (!DECL_P (decl)) | |
6905 | return get_section ("._ea", SECTION_WRITE | SECTION_DEBUG, NULL); | |
6906 | ||
6907 | return get_named_section (decl, "._ea", reloc); | |
6908 | } | |
6909 | ||
6910 | return default_elf_select_section (decl, reloc, align); | |
6911 | } | |
6912 | ||
6913 | /* Implement targetm.unique_section. */ | |
6914 | static void | |
6915 | spu_unique_section (tree decl, int reloc) | |
6916 | { | |
6917 | /* We don't support unique section names in the __ea address | |
6918 | space for now. */ | |
6919 | if (TREE_TYPE (decl) != error_mark_node | |
6920 | && TYPE_ADDR_SPACE (TREE_TYPE (decl)) != 0) | |
6921 | return; | |
6922 | ||
6923 | default_unique_section (decl, reloc); | |
6924 | } | |
6925 | ||
56c7bfc2 | 6926 | /* Generate a constant or register which contains 2^SCALE. We assume |
6927 | the result is valid for MODE. Currently, MODE must be V4SFmode and | |
6928 | SCALE must be SImode. */ | |
6929 | rtx | |
3754d046 | 6930 | spu_gen_exp2 (machine_mode mode, rtx scale) |
56c7bfc2 | 6931 | { |
6932 | gcc_assert (mode == V4SFmode); | |
6933 | gcc_assert (GET_MODE (scale) == SImode || GET_CODE (scale) == CONST_INT); | |
6934 | if (GET_CODE (scale) != CONST_INT) | |
6935 | { | |
6936 | /* unsigned int exp = (127 + scale) << 23; | |
6937 | __vector float m = (__vector float) spu_splats (exp); */ | |
6938 | rtx reg = force_reg (SImode, scale); | |
6939 | rtx exp = gen_reg_rtx (SImode); | |
6940 | rtx mul = gen_reg_rtx (mode); | |
6941 | emit_insn (gen_addsi3 (exp, reg, GEN_INT (127))); | |
6942 | emit_insn (gen_ashlsi3 (exp, exp, GEN_INT (23))); | |
6943 | emit_insn (gen_spu_splats (mul, gen_rtx_SUBREG (GET_MODE_INNER (mode), exp, 0))); | |
6944 | return mul; | |
6945 | } | |
6946 | else | |
6947 | { | |
6948 | HOST_WIDE_INT exp = 127 + INTVAL (scale); | |
6949 | unsigned char arr[16]; | |
6950 | arr[0] = arr[4] = arr[8] = arr[12] = exp >> 1; | |
6951 | arr[1] = arr[5] = arr[9] = arr[13] = exp << 7; | |
6952 | arr[2] = arr[6] = arr[10] = arr[14] = 0; | |
6953 | arr[3] = arr[7] = arr[11] = arr[15] = 0; | |
6954 | return array_to_constant (mode, arr); | |
6955 | } | |
6956 | } | |
6957 | ||
9d98604b | 6958 | /* After reload, just change the convert into a move instruction |
6959 | or a dead instruction. */ | |
6960 | void | |
6961 | spu_split_convert (rtx ops[]) | |
6962 | { | |
6963 | if (REGNO (ops[0]) == REGNO (ops[1])) | |
6964 | emit_note (NOTE_INSN_DELETED); | |
6965 | else | |
6966 | { | |
6967 | /* Use TImode always as this might help hard reg copyprop. */ | |
6968 | rtx op0 = gen_rtx_REG (TImode, REGNO (ops[0])); | |
6969 | rtx op1 = gen_rtx_REG (TImode, REGNO (ops[1])); | |
6970 | emit_insn (gen_move_insn (op0, op1)); | |
6971 | } | |
6972 | } | |
6973 | ||
b3878a6c | 6974 | void |
4cbad5bb | 6975 | spu_function_profiler (FILE * file, int labelno ATTRIBUTE_UNUSED) |
b3878a6c | 6976 | { |
6977 | fprintf (file, "# profile\n"); | |
6978 | fprintf (file, "brsl $75, _mcount\n"); | |
6979 | } | |
6980 | ||
329c1e4e | 6981 | /* Implement targetm.ref_may_alias_errno. */ |
6982 | static bool | |
6983 | spu_ref_may_alias_errno (ao_ref *ref) | |
6984 | { | |
6985 | tree base = ao_ref_base (ref); | |
6986 | ||
6987 | /* With SPU newlib, errno is defined as something like | |
6988 | _impure_data._errno | |
6989 | The default implementation of this target macro does not | |
6990 | recognize such expressions, so special-code for it here. */ | |
6991 | ||
6992 | if (TREE_CODE (base) == VAR_DECL | |
6993 | && !TREE_STATIC (base) | |
6994 | && DECL_EXTERNAL (base) | |
6995 | && TREE_CODE (TREE_TYPE (base)) == RECORD_TYPE | |
6996 | && strcmp (IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (base)), | |
6997 | "_impure_data") == 0 | |
6998 | /* _errno is the first member of _impure_data. */ | |
6999 | && ref->offset == 0) | |
7000 | return true; | |
7001 | ||
7002 | return default_ref_may_alias_errno (ref); | |
7003 | } | |
7004 | ||
f17d2d13 | 7005 | /* Output thunk to FILE that implements a C++ virtual function call (with |
7006 | multiple inheritance) to FUNCTION. The thunk adjusts the this pointer | |
7007 | by DELTA, and unless VCALL_OFFSET is zero, applies an additional adjustment | |
7008 | stored at VCALL_OFFSET in the vtable whose address is located at offset 0 | |
7009 | relative to the resulting this pointer. */ | |
7010 | ||
7011 | static void | |
7012 | spu_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED, | |
7013 | HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset, | |
7014 | tree function) | |
7015 | { | |
7016 | rtx op[8]; | |
7017 | ||
7018 | /* Make sure unwind info is emitted for the thunk if needed. */ | |
7019 | final_start_function (emit_barrier (), file, 1); | |
7020 | ||
7021 | /* Operand 0 is the target function. */ | |
7022 | op[0] = XEXP (DECL_RTL (function), 0); | |
7023 | ||
7024 | /* Operand 1 is the 'this' pointer. */ | |
7025 | if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function)) | |
7026 | op[1] = gen_rtx_REG (Pmode, FIRST_ARG_REGNUM + 1); | |
7027 | else | |
7028 | op[1] = gen_rtx_REG (Pmode, FIRST_ARG_REGNUM); | |
7029 | ||
7030 | /* Operands 2/3 are the low/high halfwords of delta. */ | |
7031 | op[2] = GEN_INT (trunc_int_for_mode (delta, HImode)); | |
7032 | op[3] = GEN_INT (trunc_int_for_mode (delta >> 16, HImode)); | |
7033 | ||
7034 | /* Operands 4/5 are the low/high halfwords of vcall_offset. */ | |
7035 | op[4] = GEN_INT (trunc_int_for_mode (vcall_offset, HImode)); | |
7036 | op[5] = GEN_INT (trunc_int_for_mode (vcall_offset >> 16, HImode)); | |
7037 | ||
7038 | /* Operands 6/7 are temporary registers. */ | |
7039 | op[6] = gen_rtx_REG (Pmode, 79); | |
7040 | op[7] = gen_rtx_REG (Pmode, 78); | |
7041 | ||
7042 | /* Add DELTA to this pointer. */ | |
7043 | if (delta) | |
7044 | { | |
7045 | if (delta >= -0x200 && delta < 0x200) | |
7046 | output_asm_insn ("ai\t%1,%1,%2", op); | |
7047 | else if (delta >= -0x8000 && delta < 0x8000) | |
7048 | { | |
7049 | output_asm_insn ("il\t%6,%2", op); | |
7050 | output_asm_insn ("a\t%1,%1,%6", op); | |
7051 | } | |
7052 | else | |
7053 | { | |
7054 | output_asm_insn ("ilhu\t%6,%3", op); | |
7055 | output_asm_insn ("iohl\t%6,%2", op); | |
7056 | output_asm_insn ("a\t%1,%1,%6", op); | |
7057 | } | |
7058 | } | |
7059 | ||
7060 | /* Perform vcall adjustment. */ | |
7061 | if (vcall_offset) | |
7062 | { | |
7063 | output_asm_insn ("lqd\t%7,0(%1)", op); | |
7064 | output_asm_insn ("rotqby\t%7,%7,%1", op); | |
7065 | ||
7066 | if (vcall_offset >= -0x200 && vcall_offset < 0x200) | |
7067 | output_asm_insn ("ai\t%7,%7,%4", op); | |
7068 | else if (vcall_offset >= -0x8000 && vcall_offset < 0x8000) | |
7069 | { | |
7070 | output_asm_insn ("il\t%6,%4", op); | |
7071 | output_asm_insn ("a\t%7,%7,%6", op); | |
7072 | } | |
7073 | else | |
7074 | { | |
7075 | output_asm_insn ("ilhu\t%6,%5", op); | |
7076 | output_asm_insn ("iohl\t%6,%4", op); | |
7077 | output_asm_insn ("a\t%7,%7,%6", op); | |
7078 | } | |
7079 | ||
7080 | output_asm_insn ("lqd\t%6,0(%7)", op); | |
7081 | output_asm_insn ("rotqby\t%6,%6,%7", op); | |
7082 | output_asm_insn ("a\t%1,%1,%6", op); | |
7083 | } | |
7084 | ||
7085 | /* Jump to target. */ | |
7086 | output_asm_insn ("br\t%0", op); | |
7087 | ||
7088 | final_end_function (); | |
7089 | } | |
7090 | ||
d5065e6e | 7091 | /* Canonicalize a comparison from one we don't have to one we do have. */ |
7092 | static void | |
7093 | spu_canonicalize_comparison (int *code, rtx *op0, rtx *op1, | |
7094 | bool op0_preserve_value) | |
7095 | { | |
7096 | if (!op0_preserve_value | |
7097 | && (*code == LE || *code == LT || *code == LEU || *code == LTU)) | |
7098 | { | |
7099 | rtx tem = *op0; | |
7100 | *op0 = *op1; | |
7101 | *op1 = tem; | |
7102 | *code = (int)swap_condition ((enum rtx_code)*code); | |
7103 | } | |
7104 | } | |
39c0ba8b | 7105 | |
7106 | /* Expand an atomic fetch-and-operate pattern. CODE is the binary operation | |
7107 | to perform. MEM is the memory on which to operate. VAL is the second | |
7108 | operand of the binary operator. BEFORE and AFTER are optional locations to | |
7109 | return the value of MEM either before of after the operation. */ | |
7110 | void | |
7111 | spu_expand_atomic_op (enum rtx_code code, rtx mem, rtx val, | |
7112 | rtx orig_before, rtx orig_after) | |
7113 | { | |
7114 | machine_mode mode = GET_MODE (mem); | |
7115 | rtx before = orig_before, after = orig_after; | |
7116 | ||
7117 | if (before == NULL_RTX) | |
7118 | before = gen_reg_rtx (mode); | |
7119 | ||
7120 | emit_move_insn (before, mem); | |
7121 | ||
7122 | if (code == MULT) /* NAND operation */ | |
7123 | { | |
7124 | rtx x = expand_simple_binop (mode, AND, before, val, | |
7125 | NULL_RTX, 1, OPTAB_LIB_WIDEN); | |
7126 | after = expand_simple_unop (mode, NOT, x, after, 1); | |
7127 | } | |
7128 | else | |
7129 | { | |
7130 | after = expand_simple_binop (mode, code, before, val, | |
7131 | after, 1, OPTAB_LIB_WIDEN); | |
7132 | } | |
7133 | ||
7134 | emit_move_insn (mem, after); | |
7135 | ||
7136 | if (orig_after && after != orig_after) | |
7137 | emit_move_insn (orig_after, after); | |
7138 | } | |
7139 | ||
3defb88e | 7140 | \f |
7141 | /* Table of machine attributes. */ | |
7142 | static const struct attribute_spec spu_attribute_table[] = | |
7143 | { | |
7144 | /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler, | |
7145 | affects_type_identity } */ | |
7146 | { "naked", 0, 0, true, false, false, spu_handle_fndecl_attribute, | |
7147 | false }, | |
7148 | { "spu_vector", 0, 0, false, true, false, spu_handle_vector_attribute, | |
7149 | false }, | |
7150 | { NULL, 0, 0, false, false, false, NULL, false } | |
7151 | }; | |
7152 | ||
7153 | /* TARGET overrides. */ | |
7154 | ||
7155 | #undef TARGET_ADDR_SPACE_POINTER_MODE | |
7156 | #define TARGET_ADDR_SPACE_POINTER_MODE spu_addr_space_pointer_mode | |
7157 | ||
7158 | #undef TARGET_ADDR_SPACE_ADDRESS_MODE | |
7159 | #define TARGET_ADDR_SPACE_ADDRESS_MODE spu_addr_space_address_mode | |
7160 | ||
7161 | #undef TARGET_ADDR_SPACE_LEGITIMATE_ADDRESS_P | |
7162 | #define TARGET_ADDR_SPACE_LEGITIMATE_ADDRESS_P \ | |
7163 | spu_addr_space_legitimate_address_p | |
7164 | ||
7165 | #undef TARGET_ADDR_SPACE_LEGITIMIZE_ADDRESS | |
7166 | #define TARGET_ADDR_SPACE_LEGITIMIZE_ADDRESS spu_addr_space_legitimize_address | |
7167 | ||
7168 | #undef TARGET_ADDR_SPACE_SUBSET_P | |
7169 | #define TARGET_ADDR_SPACE_SUBSET_P spu_addr_space_subset_p | |
7170 | ||
7171 | #undef TARGET_ADDR_SPACE_CONVERT | |
7172 | #define TARGET_ADDR_SPACE_CONVERT spu_addr_space_convert | |
7173 | ||
7174 | #undef TARGET_INIT_BUILTINS | |
7175 | #define TARGET_INIT_BUILTINS spu_init_builtins | |
7176 | #undef TARGET_BUILTIN_DECL | |
7177 | #define TARGET_BUILTIN_DECL spu_builtin_decl | |
7178 | ||
7179 | #undef TARGET_EXPAND_BUILTIN | |
7180 | #define TARGET_EXPAND_BUILTIN spu_expand_builtin | |
7181 | ||
7182 | #undef TARGET_UNWIND_WORD_MODE | |
7183 | #define TARGET_UNWIND_WORD_MODE spu_unwind_word_mode | |
7184 | ||
7185 | #undef TARGET_LEGITIMIZE_ADDRESS | |
7186 | #define TARGET_LEGITIMIZE_ADDRESS spu_legitimize_address | |
7187 | ||
7188 | /* The current assembler doesn't like .4byte foo@ppu, so use the normal .long | |
7189 | and .quad for the debugger. When it is known that the assembler is fixed, | |
7190 | these can be removed. */ | |
7191 | #undef TARGET_ASM_UNALIGNED_SI_OP | |
7192 | #define TARGET_ASM_UNALIGNED_SI_OP "\t.long\t" | |
7193 | ||
7194 | #undef TARGET_ASM_ALIGNED_DI_OP | |
7195 | #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t" | |
7196 | ||
7197 | /* The .8byte directive doesn't seem to work well for a 32 bit | |
7198 | architecture. */ | |
7199 | #undef TARGET_ASM_UNALIGNED_DI_OP | |
7200 | #define TARGET_ASM_UNALIGNED_DI_OP NULL | |
7201 | ||
7202 | #undef TARGET_RTX_COSTS | |
7203 | #define TARGET_RTX_COSTS spu_rtx_costs | |
7204 | ||
7205 | #undef TARGET_ADDRESS_COST | |
d9c5e5f4 | 7206 | #define TARGET_ADDRESS_COST hook_int_rtx_mode_as_bool_0 |
3defb88e | 7207 | |
7208 | #undef TARGET_SCHED_ISSUE_RATE | |
7209 | #define TARGET_SCHED_ISSUE_RATE spu_sched_issue_rate | |
7210 | ||
7211 | #undef TARGET_SCHED_INIT_GLOBAL | |
7212 | #define TARGET_SCHED_INIT_GLOBAL spu_sched_init_global | |
7213 | ||
7214 | #undef TARGET_SCHED_INIT | |
7215 | #define TARGET_SCHED_INIT spu_sched_init | |
7216 | ||
7217 | #undef TARGET_SCHED_VARIABLE_ISSUE | |
7218 | #define TARGET_SCHED_VARIABLE_ISSUE spu_sched_variable_issue | |
7219 | ||
7220 | #undef TARGET_SCHED_REORDER | |
7221 | #define TARGET_SCHED_REORDER spu_sched_reorder | |
7222 | ||
7223 | #undef TARGET_SCHED_REORDER2 | |
7224 | #define TARGET_SCHED_REORDER2 spu_sched_reorder | |
7225 | ||
7226 | #undef TARGET_SCHED_ADJUST_COST | |
7227 | #define TARGET_SCHED_ADJUST_COST spu_sched_adjust_cost | |
7228 | ||
7229 | #undef TARGET_ATTRIBUTE_TABLE | |
7230 | #define TARGET_ATTRIBUTE_TABLE spu_attribute_table | |
7231 | ||
7232 | #undef TARGET_ASM_INTEGER | |
7233 | #define TARGET_ASM_INTEGER spu_assemble_integer | |
7234 | ||
7235 | #undef TARGET_SCALAR_MODE_SUPPORTED_P | |
7236 | #define TARGET_SCALAR_MODE_SUPPORTED_P spu_scalar_mode_supported_p | |
7237 | ||
7238 | #undef TARGET_VECTOR_MODE_SUPPORTED_P | |
7239 | #define TARGET_VECTOR_MODE_SUPPORTED_P spu_vector_mode_supported_p | |
7240 | ||
7241 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL | |
7242 | #define TARGET_FUNCTION_OK_FOR_SIBCALL spu_function_ok_for_sibcall | |
7243 | ||
7244 | #undef TARGET_ASM_GLOBALIZE_LABEL | |
7245 | #define TARGET_ASM_GLOBALIZE_LABEL spu_asm_globalize_label | |
7246 | ||
7247 | #undef TARGET_PASS_BY_REFERENCE | |
7248 | #define TARGET_PASS_BY_REFERENCE spu_pass_by_reference | |
7249 | ||
7250 | #undef TARGET_FUNCTION_ARG | |
7251 | #define TARGET_FUNCTION_ARG spu_function_arg | |
7252 | ||
7253 | #undef TARGET_FUNCTION_ARG_ADVANCE | |
7254 | #define TARGET_FUNCTION_ARG_ADVANCE spu_function_arg_advance | |
7255 | ||
7256 | #undef TARGET_MUST_PASS_IN_STACK | |
7257 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
7258 | ||
7259 | #undef TARGET_BUILD_BUILTIN_VA_LIST | |
7260 | #define TARGET_BUILD_BUILTIN_VA_LIST spu_build_builtin_va_list | |
7261 | ||
7262 | #undef TARGET_EXPAND_BUILTIN_VA_START | |
7263 | #define TARGET_EXPAND_BUILTIN_VA_START spu_va_start | |
7264 | ||
7265 | #undef TARGET_SETUP_INCOMING_VARARGS | |
7266 | #define TARGET_SETUP_INCOMING_VARARGS spu_setup_incoming_varargs | |
7267 | ||
7268 | #undef TARGET_MACHINE_DEPENDENT_REORG | |
7269 | #define TARGET_MACHINE_DEPENDENT_REORG spu_machine_dependent_reorg | |
7270 | ||
7271 | #undef TARGET_GIMPLIFY_VA_ARG_EXPR | |
7272 | #define TARGET_GIMPLIFY_VA_ARG_EXPR spu_gimplify_va_arg_expr | |
7273 | ||
7274 | #undef TARGET_INIT_LIBFUNCS | |
7275 | #define TARGET_INIT_LIBFUNCS spu_init_libfuncs | |
7276 | ||
7277 | #undef TARGET_RETURN_IN_MEMORY | |
7278 | #define TARGET_RETURN_IN_MEMORY spu_return_in_memory | |
7279 | ||
7280 | #undef TARGET_ENCODE_SECTION_INFO | |
7281 | #define TARGET_ENCODE_SECTION_INFO spu_encode_section_info | |
7282 | ||
7283 | #undef TARGET_VECTORIZE_BUILTIN_MASK_FOR_LOAD | |
7284 | #define TARGET_VECTORIZE_BUILTIN_MASK_FOR_LOAD spu_builtin_mask_for_load | |
7285 | ||
7286 | #undef TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST | |
7287 | #define TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST spu_builtin_vectorization_cost | |
7288 | ||
7289 | #undef TARGET_VECTORIZE_INIT_COST | |
7290 | #define TARGET_VECTORIZE_INIT_COST spu_init_cost | |
7291 | ||
7292 | #undef TARGET_VECTORIZE_ADD_STMT_COST | |
7293 | #define TARGET_VECTORIZE_ADD_STMT_COST spu_add_stmt_cost | |
7294 | ||
7295 | #undef TARGET_VECTORIZE_FINISH_COST | |
7296 | #define TARGET_VECTORIZE_FINISH_COST spu_finish_cost | |
7297 | ||
7298 | #undef TARGET_VECTORIZE_DESTROY_COST_DATA | |
7299 | #define TARGET_VECTORIZE_DESTROY_COST_DATA spu_destroy_cost_data | |
7300 | ||
7301 | #undef TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE | |
7302 | #define TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE spu_vector_alignment_reachable | |
7303 | ||
7304 | #undef TARGET_LIBGCC_CMP_RETURN_MODE | |
7305 | #define TARGET_LIBGCC_CMP_RETURN_MODE spu_libgcc_cmp_return_mode | |
7306 | ||
7307 | #undef TARGET_LIBGCC_SHIFT_COUNT_MODE | |
7308 | #define TARGET_LIBGCC_SHIFT_COUNT_MODE spu_libgcc_shift_count_mode | |
7309 | ||
7310 | #undef TARGET_SCHED_SMS_RES_MII | |
7311 | #define TARGET_SCHED_SMS_RES_MII spu_sms_res_mii | |
7312 | ||
7313 | #undef TARGET_SECTION_TYPE_FLAGS | |
7314 | #define TARGET_SECTION_TYPE_FLAGS spu_section_type_flags | |
7315 | ||
7316 | #undef TARGET_ASM_SELECT_SECTION | |
7317 | #define TARGET_ASM_SELECT_SECTION spu_select_section | |
7318 | ||
7319 | #undef TARGET_ASM_UNIQUE_SECTION | |
7320 | #define TARGET_ASM_UNIQUE_SECTION spu_unique_section | |
7321 | ||
7322 | #undef TARGET_LEGITIMATE_ADDRESS_P | |
7323 | #define TARGET_LEGITIMATE_ADDRESS_P spu_legitimate_address_p | |
7324 | ||
7325 | #undef TARGET_LEGITIMATE_CONSTANT_P | |
7326 | #define TARGET_LEGITIMATE_CONSTANT_P spu_legitimate_constant_p | |
7327 | ||
7328 | #undef TARGET_TRAMPOLINE_INIT | |
7329 | #define TARGET_TRAMPOLINE_INIT spu_trampoline_init | |
7330 | ||
08c6cbd2 | 7331 | #undef TARGET_WARN_FUNC_RETURN |
7332 | #define TARGET_WARN_FUNC_RETURN spu_warn_func_return | |
7333 | ||
3defb88e | 7334 | #undef TARGET_OPTION_OVERRIDE |
7335 | #define TARGET_OPTION_OVERRIDE spu_option_override | |
7336 | ||
7337 | #undef TARGET_CONDITIONAL_REGISTER_USAGE | |
7338 | #define TARGET_CONDITIONAL_REGISTER_USAGE spu_conditional_register_usage | |
7339 | ||
7340 | #undef TARGET_REF_MAY_ALIAS_ERRNO | |
7341 | #define TARGET_REF_MAY_ALIAS_ERRNO spu_ref_may_alias_errno | |
7342 | ||
7343 | #undef TARGET_ASM_OUTPUT_MI_THUNK | |
7344 | #define TARGET_ASM_OUTPUT_MI_THUNK spu_output_mi_thunk | |
7345 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
7346 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true | |
7347 | ||
7348 | /* Variable tracking should be run after all optimizations which | |
7349 | change order of insns. It also needs a valid CFG. */ | |
7350 | #undef TARGET_DELAY_VARTRACK | |
7351 | #define TARGET_DELAY_VARTRACK true | |
7352 | ||
d5065e6e | 7353 | #undef TARGET_CANONICALIZE_COMPARISON |
7354 | #define TARGET_CANONICALIZE_COMPARISON spu_canonicalize_comparison | |
7355 | ||
5f35dd0e | 7356 | #undef TARGET_CAN_USE_DOLOOP_P |
7357 | #define TARGET_CAN_USE_DOLOOP_P can_use_doloop_if_innermost | |
7358 | ||
3defb88e | 7359 | struct gcc_target targetm = TARGET_INITIALIZER; |
7360 | ||
c2233b46 | 7361 | #include "gt-spu.h" |