2 * Copyright (c) 2012 Rob Clark <robdclark@gmail.com>
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include "util/bitscan.h"
34 #include "util/ralloc.h"
35 #include "util/u_math.h"
37 #include "instr-a3xx.h"
38 #include "ir3_compiler.h"
40 /* simple allocator to carve allocations out of an up-front allocated heap,
41 * so that we can free everything easily in one shot.
43 void * ir3_alloc(struct ir3
*shader
, int sz
)
45 return rzalloc_size(shader
, sz
); /* TODO: don't use rzalloc */
48 struct ir3
* ir3_create(struct ir3_compiler
*compiler
, gl_shader_stage type
)
50 struct ir3
*shader
= rzalloc(NULL
, struct ir3
);
52 shader
->compiler
= compiler
;
55 list_inithead(&shader
->block_list
);
56 list_inithead(&shader
->array_list
);
61 void ir3_destroy(struct ir3
*shader
)
66 #define iassert(cond) do { \
72 #define iassert_type(reg, full) do { \
74 iassert(!((reg)->flags & IR3_REG_HALF)); \
76 iassert((reg)->flags & IR3_REG_HALF); \
79 static uint32_t reg(struct ir3_register
*reg
, struct ir3_info
*info
,
80 uint32_t repeat
, uint32_t valid_flags
)
82 reg_t val
= { .dummy32
= 0 };
84 if (reg
->flags
& ~valid_flags
) {
85 debug_printf("INVALID FLAGS: %x vs %x\n",
86 reg
->flags
, valid_flags
);
89 if (!(reg
->flags
& IR3_REG_R
))
92 if (reg
->flags
& IR3_REG_IMMED
) {
93 val
.iim_val
= reg
->iim_val
;
98 if (reg
->flags
& IR3_REG_RELATIV
) {
99 components
= reg
->size
;
100 val
.idummy10
= reg
->array
.offset
;
101 max
= (reg
->array
.offset
+ repeat
+ components
- 1);
103 components
= util_last_bit(reg
->wrmask
);
104 val
.comp
= reg
->num
& 0x3;
105 val
.num
= reg
->num
>> 2;
106 max
= (reg
->num
+ repeat
+ components
- 1);
109 if (reg
->flags
& IR3_REG_CONST
) {
110 info
->max_const
= MAX2(info
->max_const
, max
>> 2);
111 } else if (val
.num
== 63) {
112 /* ignore writes to dummy register r63.x */
113 } else if (max
< regid(48, 0)) {
114 if (reg
->flags
& IR3_REG_HALF
) {
115 if (info
->gpu_id
>= 600) {
116 /* starting w/ a6xx, half regs conflict with full regs: */
117 info
->max_reg
= MAX2(info
->max_reg
, max
>> 3);
119 info
->max_half_reg
= MAX2(info
->max_half_reg
, max
>> 2);
122 info
->max_reg
= MAX2(info
->max_reg
, max
>> 2);
130 static int emit_cat0(struct ir3_instruction
*instr
, void *ptr
,
131 struct ir3_info
*info
)
133 instr_cat0_t
*cat0
= ptr
;
135 if (info
->gpu_id
>= 500) {
136 cat0
->a5xx
.immed
= instr
->cat0
.immed
;
137 } else if (info
->gpu_id
>= 400) {
138 cat0
->a4xx
.immed
= instr
->cat0
.immed
;
140 cat0
->a3xx
.immed
= instr
->cat0
.immed
;
142 cat0
->repeat
= instr
->repeat
;
143 cat0
->ss
= !!(instr
->flags
& IR3_INSTR_SS
);
144 cat0
->inv
= instr
->cat0
.inv
;
145 cat0
->comp
= instr
->cat0
.comp
;
146 cat0
->opc
= instr
->opc
;
147 cat0
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
148 cat0
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
151 switch (instr
->opc
) {
164 static int emit_cat1(struct ir3_instruction
*instr
, void *ptr
,
165 struct ir3_info
*info
)
167 struct ir3_register
*dst
= instr
->regs
[0];
168 struct ir3_register
*src
= instr
->regs
[1];
169 instr_cat1_t
*cat1
= ptr
;
171 iassert(instr
->regs_count
== 2);
172 iassert_type(dst
, type_size(instr
->cat1
.dst_type
) == 32);
173 if (!(src
->flags
& IR3_REG_IMMED
))
174 iassert_type(src
, type_size(instr
->cat1
.src_type
) == 32);
176 if (src
->flags
& IR3_REG_IMMED
) {
177 cat1
->iim_val
= src
->iim_val
;
179 } else if (src
->flags
& IR3_REG_RELATIV
) {
180 cat1
->off
= reg(src
, info
, instr
->repeat
,
181 IR3_REG_R
| IR3_REG_CONST
| IR3_REG_HALF
| IR3_REG_RELATIV
);
183 cat1
->src_rel_c
= !!(src
->flags
& IR3_REG_CONST
);
185 cat1
->src
= reg(src
, info
, instr
->repeat
,
186 IR3_REG_R
| IR3_REG_CONST
| IR3_REG_HALF
);
187 cat1
->src_c
= !!(src
->flags
& IR3_REG_CONST
);
190 cat1
->dst
= reg(dst
, info
, instr
->repeat
,
191 IR3_REG_RELATIV
| IR3_REG_EVEN
|
192 IR3_REG_R
| IR3_REG_POS_INF
| IR3_REG_HALF
);
193 cat1
->repeat
= instr
->repeat
;
194 cat1
->src_r
= !!(src
->flags
& IR3_REG_R
);
195 cat1
->ss
= !!(instr
->flags
& IR3_INSTR_SS
);
196 cat1
->ul
= !!(instr
->flags
& IR3_INSTR_UL
);
197 cat1
->dst_type
= instr
->cat1
.dst_type
;
198 cat1
->dst_rel
= !!(dst
->flags
& IR3_REG_RELATIV
);
199 cat1
->src_type
= instr
->cat1
.src_type
;
200 cat1
->even
= !!(dst
->flags
& IR3_REG_EVEN
);
201 cat1
->pos_inf
= !!(dst
->flags
& IR3_REG_POS_INF
);
202 cat1
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
203 cat1
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
209 static int emit_cat2(struct ir3_instruction
*instr
, void *ptr
,
210 struct ir3_info
*info
)
212 struct ir3_register
*dst
= instr
->regs
[0];
213 struct ir3_register
*src1
= instr
->regs
[1];
214 struct ir3_register
*src2
= instr
->regs
[2];
215 instr_cat2_t
*cat2
= ptr
;
216 unsigned absneg
= ir3_cat2_absneg(instr
->opc
);
218 iassert((instr
->regs_count
== 2) || (instr
->regs_count
== 3));
221 iassert(!instr
->repeat
);
222 iassert(instr
->nop
<= 3);
224 cat2
->src1_r
= instr
->nop
& 0x1;
225 cat2
->src2_r
= (instr
->nop
>> 1) & 0x1;
227 cat2
->src1_r
= !!(src1
->flags
& IR3_REG_R
);
229 cat2
->src2_r
= !!(src2
->flags
& IR3_REG_R
);
232 if (src1
->flags
& IR3_REG_RELATIV
) {
233 iassert(src1
->array
.offset
< (1 << 10));
234 cat2
->rel1
.src1
= reg(src1
, info
, instr
->repeat
,
235 IR3_REG_RELATIV
| IR3_REG_CONST
| IR3_REG_R
|
236 IR3_REG_HALF
| absneg
);
237 cat2
->rel1
.src1_c
= !!(src1
->flags
& IR3_REG_CONST
);
238 cat2
->rel1
.src1_rel
= 1;
239 } else if (src1
->flags
& IR3_REG_CONST
) {
240 iassert(src1
->num
< (1 << 12));
241 cat2
->c1
.src1
= reg(src1
, info
, instr
->repeat
,
242 IR3_REG_CONST
| IR3_REG_R
| IR3_REG_HALF
|
246 iassert(src1
->num
< (1 << 11));
247 cat2
->src1
= reg(src1
, info
, instr
->repeat
,
248 IR3_REG_IMMED
| IR3_REG_R
| IR3_REG_HALF
|
251 cat2
->src1_im
= !!(src1
->flags
& IR3_REG_IMMED
);
252 cat2
->src1_neg
= !!(src1
->flags
& (IR3_REG_FNEG
| IR3_REG_SNEG
| IR3_REG_BNOT
));
253 cat2
->src1_abs
= !!(src1
->flags
& (IR3_REG_FABS
| IR3_REG_SABS
));
256 iassert((src2
->flags
& IR3_REG_IMMED
) ||
257 !((src1
->flags
^ src2
->flags
) & IR3_REG_HALF
));
259 if (src2
->flags
& IR3_REG_RELATIV
) {
260 iassert(src2
->array
.offset
< (1 << 10));
261 cat2
->rel2
.src2
= reg(src2
, info
, instr
->repeat
,
262 IR3_REG_RELATIV
| IR3_REG_CONST
| IR3_REG_R
|
263 IR3_REG_HALF
| absneg
);
264 cat2
->rel2
.src2_c
= !!(src2
->flags
& IR3_REG_CONST
);
265 cat2
->rel2
.src2_rel
= 1;
266 } else if (src2
->flags
& IR3_REG_CONST
) {
267 iassert(src2
->num
< (1 << 12));
268 cat2
->c2
.src2
= reg(src2
, info
, instr
->repeat
,
269 IR3_REG_CONST
| IR3_REG_R
| IR3_REG_HALF
|
273 iassert(src2
->num
< (1 << 11));
274 cat2
->src2
= reg(src2
, info
, instr
->repeat
,
275 IR3_REG_IMMED
| IR3_REG_R
| IR3_REG_HALF
|
279 cat2
->src2_im
= !!(src2
->flags
& IR3_REG_IMMED
);
280 cat2
->src2_neg
= !!(src2
->flags
& (IR3_REG_FNEG
| IR3_REG_SNEG
| IR3_REG_BNOT
));
281 cat2
->src2_abs
= !!(src2
->flags
& (IR3_REG_FABS
| IR3_REG_SABS
));
284 cat2
->dst
= reg(dst
, info
, instr
->repeat
,
285 IR3_REG_R
| IR3_REG_EI
| IR3_REG_HALF
);
286 cat2
->repeat
= instr
->repeat
;
287 cat2
->sat
= !!(instr
->flags
& IR3_INSTR_SAT
);
288 cat2
->ss
= !!(instr
->flags
& IR3_INSTR_SS
);
289 cat2
->ul
= !!(instr
->flags
& IR3_INSTR_UL
);
290 cat2
->dst_half
= !!((src1
->flags
^ dst
->flags
) & IR3_REG_HALF
);
291 cat2
->ei
= !!(dst
->flags
& IR3_REG_EI
);
292 cat2
->cond
= instr
->cat2
.condition
;
293 cat2
->full
= ! (src1
->flags
& IR3_REG_HALF
);
294 cat2
->opc
= instr
->opc
;
295 cat2
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
296 cat2
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
302 static int emit_cat3(struct ir3_instruction
*instr
, void *ptr
,
303 struct ir3_info
*info
)
305 struct ir3_register
*dst
= instr
->regs
[0];
306 struct ir3_register
*src1
= instr
->regs
[1];
307 struct ir3_register
*src2
= instr
->regs
[2];
308 struct ir3_register
*src3
= instr
->regs
[3];
309 unsigned absneg
= ir3_cat3_absneg(instr
->opc
);
310 instr_cat3_t
*cat3
= ptr
;
311 uint32_t src_flags
= 0;
313 switch (instr
->opc
) {
321 case OPC_SAD_S32
: // really??
322 src_flags
|= IR3_REG_HALF
;
328 iassert(instr
->regs_count
== 4);
329 iassert(!((src1
->flags
^ src_flags
) & IR3_REG_HALF
));
330 iassert(!((src2
->flags
^ src_flags
) & IR3_REG_HALF
));
331 iassert(!((src3
->flags
^ src_flags
) & IR3_REG_HALF
));
334 iassert(!instr
->repeat
);
335 iassert(instr
->nop
<= 3);
337 cat3
->src1_r
= instr
->nop
& 0x1;
338 cat3
->src2_r
= (instr
->nop
>> 1) & 0x1;
340 cat3
->src1_r
= !!(src1
->flags
& IR3_REG_R
);
341 cat3
->src2_r
= !!(src2
->flags
& IR3_REG_R
);
344 if (src1
->flags
& IR3_REG_RELATIV
) {
345 iassert(src1
->array
.offset
< (1 << 10));
346 cat3
->rel1
.src1
= reg(src1
, info
, instr
->repeat
,
347 IR3_REG_RELATIV
| IR3_REG_CONST
| IR3_REG_R
|
348 IR3_REG_HALF
| absneg
);
349 cat3
->rel1
.src1_c
= !!(src1
->flags
& IR3_REG_CONST
);
350 cat3
->rel1
.src1_rel
= 1;
351 } else if (src1
->flags
& IR3_REG_CONST
) {
352 iassert(src1
->num
< (1 << 12));
353 cat3
->c1
.src1
= reg(src1
, info
, instr
->repeat
,
354 IR3_REG_CONST
| IR3_REG_R
| IR3_REG_HALF
| absneg
);
357 iassert(src1
->num
< (1 << 11));
358 cat3
->src1
= reg(src1
, info
, instr
->repeat
,
359 IR3_REG_R
| IR3_REG_HALF
| absneg
);
362 cat3
->src1_neg
= !!(src1
->flags
& (IR3_REG_FNEG
| IR3_REG_SNEG
| IR3_REG_BNOT
));
364 cat3
->src2
= reg(src2
, info
, instr
->repeat
,
365 IR3_REG_CONST
| IR3_REG_R
| IR3_REG_HALF
| absneg
);
366 cat3
->src2_c
= !!(src2
->flags
& IR3_REG_CONST
);
367 cat3
->src2_neg
= !!(src2
->flags
& (IR3_REG_FNEG
| IR3_REG_SNEG
| IR3_REG_BNOT
));
369 if (src3
->flags
& IR3_REG_RELATIV
) {
370 iassert(src3
->array
.offset
< (1 << 10));
371 cat3
->rel2
.src3
= reg(src3
, info
, instr
->repeat
,
372 IR3_REG_RELATIV
| IR3_REG_CONST
| IR3_REG_R
|
373 IR3_REG_HALF
| absneg
);
374 cat3
->rel2
.src3_c
= !!(src3
->flags
& IR3_REG_CONST
);
375 cat3
->rel2
.src3_rel
= 1;
376 } else if (src3
->flags
& IR3_REG_CONST
) {
377 iassert(src3
->num
< (1 << 12));
378 cat3
->c2
.src3
= reg(src3
, info
, instr
->repeat
,
379 IR3_REG_CONST
| IR3_REG_R
| IR3_REG_HALF
| absneg
);
382 iassert(src3
->num
< (1 << 11));
383 cat3
->src3
= reg(src3
, info
, instr
->repeat
,
384 IR3_REG_R
| IR3_REG_HALF
| absneg
);
387 cat3
->src3_neg
= !!(src3
->flags
& (IR3_REG_FNEG
| IR3_REG_SNEG
| IR3_REG_BNOT
));
388 cat3
->src3_r
= !!(src3
->flags
& IR3_REG_R
);
390 cat3
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
391 cat3
->repeat
= instr
->repeat
;
392 cat3
->sat
= !!(instr
->flags
& IR3_INSTR_SAT
);
393 cat3
->ss
= !!(instr
->flags
& IR3_INSTR_SS
);
394 cat3
->ul
= !!(instr
->flags
& IR3_INSTR_UL
);
395 cat3
->dst_half
= !!((src_flags
^ dst
->flags
) & IR3_REG_HALF
);
396 cat3
->opc
= instr
->opc
;
397 cat3
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
398 cat3
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
404 static int emit_cat4(struct ir3_instruction
*instr
, void *ptr
,
405 struct ir3_info
*info
)
407 struct ir3_register
*dst
= instr
->regs
[0];
408 struct ir3_register
*src
= instr
->regs
[1];
409 instr_cat4_t
*cat4
= ptr
;
411 iassert(instr
->regs_count
== 2);
413 if (src
->flags
& IR3_REG_RELATIV
) {
414 iassert(src
->array
.offset
< (1 << 10));
415 cat4
->rel
.src
= reg(src
, info
, instr
->repeat
,
416 IR3_REG_RELATIV
| IR3_REG_CONST
| IR3_REG_FNEG
|
417 IR3_REG_FABS
| IR3_REG_R
| IR3_REG_HALF
);
418 cat4
->rel
.src_c
= !!(src
->flags
& IR3_REG_CONST
);
419 cat4
->rel
.src_rel
= 1;
420 } else if (src
->flags
& IR3_REG_CONST
) {
421 iassert(src
->num
< (1 << 12));
422 cat4
->c
.src
= reg(src
, info
, instr
->repeat
,
423 IR3_REG_CONST
| IR3_REG_FNEG
| IR3_REG_FABS
|
424 IR3_REG_R
| IR3_REG_HALF
);
427 iassert(src
->num
< (1 << 11));
428 cat4
->src
= reg(src
, info
, instr
->repeat
,
429 IR3_REG_IMMED
| IR3_REG_FNEG
| IR3_REG_FABS
|
430 IR3_REG_R
| IR3_REG_HALF
);
433 cat4
->src_im
= !!(src
->flags
& IR3_REG_IMMED
);
434 cat4
->src_neg
= !!(src
->flags
& IR3_REG_FNEG
);
435 cat4
->src_abs
= !!(src
->flags
& IR3_REG_FABS
);
436 cat4
->src_r
= !!(src
->flags
& IR3_REG_R
);
438 cat4
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
439 cat4
->repeat
= instr
->repeat
;
440 cat4
->sat
= !!(instr
->flags
& IR3_INSTR_SAT
);
441 cat4
->ss
= !!(instr
->flags
& IR3_INSTR_SS
);
442 cat4
->ul
= !!(instr
->flags
& IR3_INSTR_UL
);
443 cat4
->dst_half
= !!((src
->flags
^ dst
->flags
) & IR3_REG_HALF
);
444 cat4
->full
= ! (src
->flags
& IR3_REG_HALF
);
445 cat4
->opc
= instr
->opc
;
446 cat4
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
447 cat4
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
453 static int emit_cat5(struct ir3_instruction
*instr
, void *ptr
,
454 struct ir3_info
*info
)
456 struct ir3_register
*dst
= instr
->regs
[0];
457 /* To simplify things when there could be zero, one, or two args other
458 * than tex/sampler idx, we use the first src reg in the ir to hold
461 struct ir3_register
*src1
;
462 struct ir3_register
*src2
;
463 instr_cat5_t
*cat5
= ptr
;
465 iassert((instr
->regs_count
== 1) ||
466 (instr
->regs_count
== 2) ||
467 (instr
->regs_count
== 3) ||
468 (instr
->regs_count
== 4));
470 if (instr
->flags
& IR3_INSTR_S2EN
) {
471 src1
= instr
->regs
[2];
472 src2
= instr
->regs_count
> 3 ? instr
->regs
[3] : NULL
;
474 src1
= instr
->regs_count
> 1 ? instr
->regs
[1] : NULL
;
475 src2
= instr
->regs_count
> 2 ? instr
->regs
[2] : NULL
;
478 assume(src1
|| !src2
);
481 cat5
->full
= ! (src1
->flags
& IR3_REG_HALF
);
482 cat5
->src1
= reg(src1
, info
, instr
->repeat
, IR3_REG_HALF
);
486 iassert(!((src1
->flags
^ src2
->flags
) & IR3_REG_HALF
));
487 cat5
->src2
= reg(src2
, info
, instr
->repeat
, IR3_REG_HALF
);
490 if (instr
->flags
& IR3_INSTR_B
) {
491 cat5
->s2en_bindless
.base_hi
= instr
->cat5
.tex_base
>> 1;
492 cat5
->base_lo
= instr
->cat5
.tex_base
& 1;
495 if (instr
->flags
& IR3_INSTR_S2EN
) {
496 struct ir3_register
*samp_tex
= instr
->regs
[1];
497 iassert(samp_tex
->flags
& IR3_REG_HALF
);
498 cat5
->s2en_bindless
.src3
= reg(samp_tex
, info
, instr
->repeat
,
499 (instr
->flags
& IR3_INSTR_B
) ? 0 : IR3_REG_HALF
);
500 if (instr
->flags
& IR3_INSTR_B
) {
501 if (instr
->flags
& IR3_INSTR_A1EN
) {
502 cat5
->s2en_bindless
.desc_mode
= CAT5_BINDLESS_A1_UNIFORM
;
504 cat5
->s2en_bindless
.desc_mode
= CAT5_BINDLESS_UNIFORM
;
507 /* TODO: This should probably be CAT5_UNIFORM, at least on a6xx,
508 * as this is what the blob does and it is presumably faster, but
509 * first we should confirm it is actually nonuniform and figure
510 * out when the whole descriptor mode mechanism was introduced.
512 cat5
->s2en_bindless
.desc_mode
= CAT5_NONUNIFORM
;
514 iassert(!(instr
->cat5
.samp
| instr
->cat5
.tex
));
515 } else if (instr
->flags
& IR3_INSTR_B
) {
516 cat5
->s2en_bindless
.src3
= instr
->cat5
.samp
;
517 if (instr
->flags
& IR3_INSTR_A1EN
) {
518 cat5
->s2en_bindless
.desc_mode
= CAT5_BINDLESS_A1_IMM
;
520 cat5
->s2en_bindless
.desc_mode
= CAT5_BINDLESS_IMM
;
523 cat5
->norm
.samp
= instr
->cat5
.samp
;
524 cat5
->norm
.tex
= instr
->cat5
.tex
;
527 cat5
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
528 cat5
->wrmask
= dst
->wrmask
;
529 cat5
->type
= instr
->cat5
.type
;
530 cat5
->is_3d
= !!(instr
->flags
& IR3_INSTR_3D
);
531 cat5
->is_a
= !!(instr
->flags
& IR3_INSTR_A
);
532 cat5
->is_s
= !!(instr
->flags
& IR3_INSTR_S
);
533 cat5
->is_s2en_bindless
= !!(instr
->flags
& (IR3_INSTR_S2EN
| IR3_INSTR_B
));
534 cat5
->is_o
= !!(instr
->flags
& IR3_INSTR_O
);
535 cat5
->is_p
= !!(instr
->flags
& IR3_INSTR_P
);
536 cat5
->opc
= instr
->opc
;
537 cat5
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
538 cat5
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
544 static int emit_cat6_a6xx(struct ir3_instruction
*instr
, void *ptr
,
545 struct ir3_info
*info
)
547 struct ir3_register
*src1
, *src2
, *ssbo
;
548 instr_cat6_a6xx_t
*cat6
= ptr
;
549 bool has_dest
= (instr
->opc
== OPC_LDIB
|| instr
->opc
== OPC_LDC
);
551 ssbo
= instr
->regs
[1];
552 src1
= instr
->regs
[2];
555 /* the src2 field in the instruction is actually the destination
556 * register for load instructions:
558 src2
= instr
->regs
[0];
560 src2
= instr
->regs
[3];
563 cat6
->type
= instr
->cat6
.type
;
564 cat6
->d
= instr
->cat6
.d
- (instr
->opc
== OPC_LDC
? 0 : 1);
565 cat6
->typed
= instr
->cat6
.typed
;
566 cat6
->type_size
= instr
->cat6
.iim_val
- 1;
567 cat6
->opc
= instr
->opc
;
568 cat6
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
569 cat6
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
572 cat6
->src1
= reg(src1
, info
, instr
->repeat
, 0);
573 cat6
->src2
= reg(src2
, info
, instr
->repeat
, 0);
574 cat6
->ssbo
= reg(ssbo
, info
, instr
->repeat
, IR3_REG_IMMED
);
576 if (instr
->flags
& IR3_INSTR_B
) {
577 if (ssbo
->flags
& IR3_REG_IMMED
) {
578 cat6
->desc_mode
= CAT6_BINDLESS_IMM
;
580 cat6
->desc_mode
= CAT6_BINDLESS_UNIFORM
;
582 cat6
->base
= instr
->cat6
.base
;
584 /* TODO figure out mode for indirect SSBO index in !bindless */
585 iassert(ssbo
->flags
& IR3_REG_IMMED
);
586 cat6
->desc_mode
= CAT6_IMM
;
589 switch (instr
->opc
) {
592 case OPC_ATOMIC_XCHG
:
595 case OPC_ATOMIC_CMPXCHG
:
629 static int emit_cat6(struct ir3_instruction
*instr
, void *ptr
,
630 struct ir3_info
*info
)
632 struct ir3_register
*dst
, *src1
, *src2
;
633 instr_cat6_t
*cat6
= ptr
;
635 /* In a6xx we start using a new instruction encoding for some of
636 * these instructions:
638 if (info
->gpu_id
>= 600) {
639 switch (instr
->opc
) {
642 case OPC_ATOMIC_XCHG
:
645 case OPC_ATOMIC_CMPXCHG
:
651 /* The shared variants of these still use the old encoding: */
652 if (!(instr
->flags
& IR3_INSTR_G
))
658 return emit_cat6_a6xx(instr
, ptr
, info
);
664 bool type_full
= type_size(instr
->cat6
.type
) == 32;
666 cat6
->type
= instr
->cat6
.type
;
667 cat6
->opc
= instr
->opc
;
668 cat6
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
669 cat6
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
670 cat6
->g
= !!(instr
->flags
& IR3_INSTR_G
);
673 switch (instr
->opc
) {
676 iassert_type(instr
->regs
[0], type_full
); /* dst */
677 iassert_type(instr
->regs
[1], type_full
); /* src1 */
681 iassert_type(instr
->regs
[0], true); /* dst */
682 iassert_type(instr
->regs
[1], true); /* src1 */
689 /* no dst, so regs[0] is dummy */
690 iassert_type(instr
->regs
[1], true); /* dst */
691 iassert_type(instr
->regs
[2], type_full
); /* src1 */
692 iassert_type(instr
->regs
[3], true); /* src2 */
695 iassert_type(instr
->regs
[0], type_full
); /* dst */
696 iassert_type(instr
->regs
[1], true); /* src1 */
697 if (instr
->regs_count
> 2)
698 iassert_type(instr
->regs
[2], true); /* src1 */
702 /* the "dst" for a store instruction is (from the perspective
703 * of data flow in the shader, ie. register use/def, etc) in
704 * fact a register that is read by the instruction, rather
707 if (is_store(instr
)) {
708 iassert(instr
->regs_count
>= 3);
710 dst
= instr
->regs
[1];
711 src1
= instr
->regs
[2];
712 src2
= (instr
->regs_count
>= 4) ? instr
->regs
[3] : NULL
;
714 iassert(instr
->regs_count
>= 2);
716 dst
= instr
->regs
[0];
717 src1
= instr
->regs
[1];
718 src2
= (instr
->regs_count
>= 3) ? instr
->regs
[2] : NULL
;
721 /* TODO we need a more comprehensive list about which instructions
722 * can be encoded which way. Or possibly use IR3_INSTR_0 flag to
723 * indicate to use the src_off encoding even if offset is zero
724 * (but then what to do about dst_off?)
726 if (is_atomic(instr
->opc
)) {
727 instr_cat6ldgb_t
*ldgb
= ptr
;
729 /* maybe these two bits both determine the instruction encoding? */
730 cat6
->src_off
= false;
732 ldgb
->d
= instr
->cat6
.d
- 1;
733 ldgb
->typed
= instr
->cat6
.typed
;
734 ldgb
->type_size
= instr
->cat6
.iim_val
- 1;
736 ldgb
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
739 struct ir3_register
*src3
= instr
->regs
[3];
740 struct ir3_register
*src4
= instr
->regs
[4];
742 /* first src is src_ssbo: */
743 iassert(src1
->flags
& IR3_REG_IMMED
);
744 ldgb
->src_ssbo
= src1
->uim_val
;
746 ldgb
->src1
= reg(src2
, info
, instr
->repeat
, IR3_REG_IMMED
);
747 ldgb
->src1_im
= !!(src2
->flags
& IR3_REG_IMMED
);
748 ldgb
->src2
= reg(src3
, info
, instr
->repeat
, IR3_REG_IMMED
);
749 ldgb
->src2_im
= !!(src3
->flags
& IR3_REG_IMMED
);
751 ldgb
->src3
= reg(src4
, info
, instr
->repeat
, 0);
755 ldgb
->src1
= reg(src1
, info
, instr
->repeat
, IR3_REG_IMMED
);
756 ldgb
->src1_im
= !!(src1
->flags
& IR3_REG_IMMED
);
757 ldgb
->src2
= reg(src2
, info
, instr
->repeat
, IR3_REG_IMMED
);
758 ldgb
->src2_im
= !!(src2
->flags
& IR3_REG_IMMED
);
764 } else if (instr
->opc
== OPC_LDGB
) {
765 struct ir3_register
*src3
= instr
->regs
[3];
766 instr_cat6ldgb_t
*ldgb
= ptr
;
768 /* maybe these two bits both determine the instruction encoding? */
769 cat6
->src_off
= false;
771 ldgb
->d
= instr
->cat6
.d
- 1;
772 ldgb
->typed
= instr
->cat6
.typed
;
773 ldgb
->type_size
= instr
->cat6
.iim_val
- 1;
775 ldgb
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
777 /* first src is src_ssbo: */
778 iassert(src1
->flags
& IR3_REG_IMMED
);
779 ldgb
->src_ssbo
= src1
->uim_val
;
781 /* then next two are src1/src2: */
782 ldgb
->src1
= reg(src2
, info
, instr
->repeat
, IR3_REG_IMMED
);
783 ldgb
->src1_im
= !!(src2
->flags
& IR3_REG_IMMED
);
784 ldgb
->src2
= reg(src3
, info
, instr
->repeat
, IR3_REG_IMMED
);
785 ldgb
->src2_im
= !!(src3
->flags
& IR3_REG_IMMED
);
791 } else if (instr
->opc
== OPC_RESINFO
) {
792 instr_cat6ldgb_t
*ldgb
= ptr
;
794 ldgb
->d
= instr
->cat6
.d
- 1;
796 ldgb
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
798 /* first src is src_ssbo: */
799 iassert(src1
->flags
& IR3_REG_IMMED
);
800 ldgb
->src_ssbo
= src1
->uim_val
;
803 } else if ((instr
->opc
== OPC_STGB
) || (instr
->opc
== OPC_STIB
)) {
804 struct ir3_register
*src3
= instr
->regs
[4];
805 instr_cat6stgb_t
*stgb
= ptr
;
807 /* maybe these two bits both determine the instruction encoding? */
808 cat6
->src_off
= true;
811 stgb
->d
= instr
->cat6
.d
- 1;
812 stgb
->typed
= instr
->cat6
.typed
;
813 stgb
->type_size
= instr
->cat6
.iim_val
- 1;
815 /* first src is dst_ssbo: */
816 iassert(dst
->flags
& IR3_REG_IMMED
);
817 stgb
->dst_ssbo
= dst
->uim_val
;
819 /* then src1/src2/src3: */
820 stgb
->src1
= reg(src1
, info
, instr
->repeat
, 0);
821 stgb
->src2
= reg(src2
, info
, instr
->repeat
, IR3_REG_IMMED
);
822 stgb
->src2_im
= !!(src2
->flags
& IR3_REG_IMMED
);
823 stgb
->src3
= reg(src3
, info
, instr
->repeat
, IR3_REG_IMMED
);
824 stgb
->src3_im
= !!(src3
->flags
& IR3_REG_IMMED
);
827 } else if (instr
->cat6
.src_offset
|| (instr
->opc
== OPC_LDG
) ||
828 (instr
->opc
== OPC_LDL
) || (instr
->opc
== OPC_LDLW
)) {
829 struct ir3_register
*src3
= instr
->regs
[3];
830 instr_cat6a_t
*cat6a
= ptr
;
832 cat6
->src_off
= true;
834 if (instr
->opc
== OPC_LDG
) {
835 /* For LDG src1 can not be immediate, so src1_imm is redundant and
836 * instead used to signal whether (when true) 'off' is a 32 bit
837 * register or an immediate offset.
839 cat6a
->src1
= reg(src1
, info
, instr
->repeat
, 0);
840 cat6a
->src1_im
= !(src3
->flags
& IR3_REG_IMMED
);
841 cat6a
->off
= reg(src3
, info
, instr
->repeat
, IR3_REG_IMMED
);
843 cat6a
->src1
= reg(src1
, info
, instr
->repeat
, IR3_REG_IMMED
);
844 cat6a
->src1_im
= !!(src1
->flags
& IR3_REG_IMMED
);
845 cat6a
->off
= reg(src3
, info
, instr
->repeat
, IR3_REG_IMMED
);
846 iassert(src3
->flags
& IR3_REG_IMMED
);
850 cat6a
->src2
= reg(src2
, info
, instr
->repeat
, IR3_REG_IMMED
);
851 cat6a
->src2_im
= true;
853 instr_cat6b_t
*cat6b
= ptr
;
855 cat6
->src_off
= false;
857 cat6b
->src1
= reg(src1
, info
, instr
->repeat
, IR3_REG_IMMED
| IR3_REG_HALF
);
858 cat6b
->src1_im
= !!(src1
->flags
& IR3_REG_IMMED
);
860 cat6b
->src2
= reg(src2
, info
, instr
->repeat
, IR3_REG_IMMED
);
861 cat6b
->src2_im
= !!(src2
->flags
& IR3_REG_IMMED
);
865 if (instr
->cat6
.dst_offset
|| (instr
->opc
== OPC_STG
) ||
866 (instr
->opc
== OPC_STL
) || (instr
->opc
== OPC_STLW
)) {
867 instr_cat6c_t
*cat6c
= ptr
;
868 cat6
->dst_off
= true;
869 cat6c
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
871 if (instr
->flags
& IR3_INSTR_G
) {
872 struct ir3_register
*src3
= instr
->regs
[4];
873 cat6c
->off
= reg(src3
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
874 if (src3
->flags
& IR3_REG_IMMED
) {
875 /* Immediate offsets are in bytes... */
880 cat6c
->off
= instr
->cat6
.dst_offset
;
883 instr_cat6d_t
*cat6d
= ptr
;
884 cat6
->dst_off
= false;
885 cat6d
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
891 static int emit_cat7(struct ir3_instruction
*instr
, void *ptr
,
892 struct ir3_info
*info
)
894 instr_cat7_t
*cat7
= ptr
;
896 cat7
->ss
= !!(instr
->flags
& IR3_INSTR_SS
);
897 cat7
->w
= instr
->cat7
.w
;
898 cat7
->r
= instr
->cat7
.r
;
899 cat7
->l
= instr
->cat7
.l
;
900 cat7
->g
= instr
->cat7
.g
;
901 cat7
->opc
= instr
->opc
;
902 cat7
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
903 cat7
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
909 static int (*emit
[])(struct ir3_instruction
*instr
, void *ptr
,
910 struct ir3_info
*info
) = {
911 emit_cat0
, emit_cat1
, emit_cat2
, emit_cat3
, emit_cat4
, emit_cat5
, emit_cat6
,
915 void * ir3_assemble(struct ir3
*shader
, struct ir3_info
*info
,
918 uint32_t *ptr
, *dwords
;
920 memset(info
, 0, sizeof(*info
));
921 info
->gpu_id
= gpu_id
;
923 info
->max_half_reg
= -1;
924 info
->max_const
= -1;
926 foreach_block (block
, &shader
->block_list
) {
927 foreach_instr (instr
, &block
->instr_list
) {
928 info
->sizedwords
+= 2;
932 /* need an integer number of instruction "groups" (sets of 16
933 * instructions on a4xx or sets of 4 instructions on a3xx),
934 * so pad out w/ NOPs if needed: (NOTE each instruction is 64bits)
937 info
->sizedwords
= align(info
->sizedwords
, 16 * 2);
939 info
->sizedwords
= align(info
->sizedwords
, 4 * 2);
942 ptr
= dwords
= calloc(4, info
->sizedwords
);
944 foreach_block (block
, &shader
->block_list
) {
945 unsigned sfu_delay
= 0;
947 foreach_instr (instr
, &block
->instr_list
) {
948 int ret
= emit
[opc_cat(instr
->opc
)](instr
, dwords
, info
);
952 if ((instr
->opc
== OPC_BARY_F
) && (instr
->regs
[0]->flags
& IR3_REG_EI
))
953 info
->last_baryf
= info
->instrs_count
;
955 info
->instrs_count
+= 1 + instr
->repeat
+ instr
->nop
;
956 info
->nops_count
+= instr
->nop
;
957 if (instr
->opc
== OPC_NOP
)
958 info
->nops_count
+= 1 + instr
->repeat
;
959 if (instr
->opc
== OPC_MOV
) {
960 if (instr
->cat1
.src_type
== instr
->cat1
.dst_type
) {
961 info
->mov_count
+= 1 + instr
->repeat
;
963 info
->cov_count
+= 1 + instr
->repeat
;
968 if (instr
->flags
& IR3_INSTR_SS
) {
970 info
->sstall
+= sfu_delay
;
973 if (instr
->flags
& IR3_INSTR_SY
)
978 } else if (sfu_delay
> 0) {
991 static struct ir3_register
* reg_create(struct ir3
*shader
,
994 struct ir3_register
*reg
=
995 ir3_alloc(shader
, sizeof(struct ir3_register
));
999 if (shader
->compiler
->gpu_id
>= 600)
1004 static void insert_instr(struct ir3_block
*block
,
1005 struct ir3_instruction
*instr
)
1007 struct ir3
*shader
= block
->shader
;
1009 instr
->serialno
= ++shader
->instr_count
;
1011 list_addtail(&instr
->node
, &block
->instr_list
);
1013 if (is_input(instr
))
1014 array_insert(shader
, shader
->baryfs
, instr
);
1017 struct ir3_block
* ir3_block_create(struct ir3
*shader
)
1019 struct ir3_block
*block
= ir3_alloc(shader
, sizeof(*block
));
1021 block
->serialno
= ++shader
->block_count
;
1023 block
->shader
= shader
;
1024 list_inithead(&block
->node
);
1025 list_inithead(&block
->instr_list
);
1029 static struct ir3_instruction
*instr_create(struct ir3_block
*block
, int nreg
)
1031 struct ir3_instruction
*instr
;
1032 unsigned sz
= sizeof(*instr
) + (nreg
* sizeof(instr
->regs
[0]));
1033 char *ptr
= ir3_alloc(block
->shader
, sz
);
1035 instr
= (struct ir3_instruction
*)ptr
;
1036 ptr
+= sizeof(*instr
);
1037 instr
->regs
= (struct ir3_register
**)ptr
;
1040 instr
->regs_max
= nreg
;
1046 struct ir3_instruction
* ir3_instr_create2(struct ir3_block
*block
,
1047 opc_t opc
, int nreg
)
1049 struct ir3_instruction
*instr
= instr_create(block
, nreg
);
1050 instr
->block
= block
;
1052 insert_instr(block
, instr
);
1056 struct ir3_instruction
* ir3_instr_create(struct ir3_block
*block
, opc_t opc
)
1058 /* NOTE: we could be slightly more clever, at least for non-meta,
1059 * and choose # of regs based on category.
1061 return ir3_instr_create2(block
, opc
, 4);
1064 struct ir3_instruction
* ir3_instr_clone(struct ir3_instruction
*instr
)
1066 struct ir3_instruction
*new_instr
= instr_create(instr
->block
,
1068 struct ir3_register
**regs
;
1071 regs
= new_instr
->regs
;
1072 *new_instr
= *instr
;
1073 new_instr
->regs
= regs
;
1075 insert_instr(instr
->block
, new_instr
);
1077 /* clone registers: */
1078 new_instr
->regs_count
= 0;
1079 for (i
= 0; i
< instr
->regs_count
; i
++) {
1080 struct ir3_register
*reg
= instr
->regs
[i
];
1081 struct ir3_register
*new_reg
=
1082 ir3_reg_create(new_instr
, reg
->num
, reg
->flags
);
1089 /* Add a false dependency to instruction, to ensure it is scheduled first: */
1090 void ir3_instr_add_dep(struct ir3_instruction
*instr
, struct ir3_instruction
*dep
)
1092 array_insert(instr
, instr
->deps
, dep
);
1095 struct ir3_register
* ir3_reg_create(struct ir3_instruction
*instr
,
1098 struct ir3
*shader
= instr
->block
->shader
;
1099 struct ir3_register
*reg
= reg_create(shader
, num
, flags
);
1101 debug_assert(instr
->regs_count
< instr
->regs_max
);
1103 instr
->regs
[instr
->regs_count
++] = reg
;
1107 struct ir3_register
* ir3_reg_clone(struct ir3
*shader
,
1108 struct ir3_register
*reg
)
1110 struct ir3_register
*new_reg
= reg_create(shader
, 0, 0);
1116 ir3_instr_set_address(struct ir3_instruction
*instr
,
1117 struct ir3_instruction
*addr
)
1119 if (instr
->address
!= addr
) {
1120 struct ir3
*ir
= instr
->block
->shader
;
1122 debug_assert(!instr
->address
);
1123 debug_assert(instr
->block
== addr
->block
);
1125 instr
->address
= addr
;
1126 debug_assert(reg_num(addr
->regs
[0]) == REG_A0
);
1127 unsigned comp
= reg_comp(addr
->regs
[0]);
1129 array_insert(ir
, ir
->a0_users
, instr
);
1131 debug_assert(comp
== 1);
1132 array_insert(ir
, ir
->a1_users
, instr
);
1138 ir3_block_clear_mark(struct ir3_block
*block
)
1140 foreach_instr (instr
, &block
->instr_list
)
1141 instr
->flags
&= ~IR3_INSTR_MARK
;
1145 ir3_clear_mark(struct ir3
*ir
)
1147 foreach_block (block
, &ir
->block_list
) {
1148 ir3_block_clear_mark(block
);
1153 ir3_count_instructions(struct ir3
*ir
)
1156 foreach_block (block
, &ir
->block_list
) {
1157 block
->start_ip
= cnt
;
1158 foreach_instr (instr
, &block
->instr_list
) {
1161 block
->end_ip
= cnt
;
1166 /* When counting instructions for RA, we insert extra fake instructions at the
1167 * beginning of each block, where values become live, and at the end where
1168 * values die. This prevents problems where values live-in at the beginning or
1169 * live-out at the end of a block from being treated as if they were
1170 * live-in/live-out at the first/last instruction, which would be incorrect.
1171 * In ir3_legalize these ip's are assumed to be actual ip's of the final
1172 * program, so it would be incorrect to use this everywhere.
1176 ir3_count_instructions_ra(struct ir3
*ir
)
1179 foreach_block (block
, &ir
->block_list
) {
1180 block
->start_ip
= cnt
++;
1181 foreach_instr (instr
, &block
->instr_list
) {
1184 block
->end_ip
= cnt
++;
1190 ir3_lookup_array(struct ir3
*ir
, unsigned id
)
1192 foreach_array (arr
, &ir
->array_list
)
1199 ir3_find_ssa_uses(struct ir3
*ir
, void *mem_ctx
, bool falsedeps
)
1201 /* We could do this in a single pass if we can assume instructions
1202 * are always sorted. Which currently might not always be true.
1203 * (In particular after ir3_group pass, but maybe other places.)
1205 foreach_block (block
, &ir
->block_list
)
1206 foreach_instr (instr
, &block
->instr_list
)
1209 foreach_block (block
, &ir
->block_list
) {
1210 foreach_instr (instr
, &block
->instr_list
) {
1211 struct ir3_instruction
*src
;
1213 foreach_ssa_src_n (src
, n
, instr
) {
1214 if (__is_false_dep(instr
, n
) && !falsedeps
)
1217 src
->uses
= _mesa_pointer_set_create(mem_ctx
);
1218 _mesa_set_add(src
->uses
, instr
);