ed14c343faa48b616a2fef3051738868d029cee2
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"
39 /* simple allocator to carve allocations out of an up-front allocated heap,
40 * so that we can free everything easily in one shot.
42 void * ir3_alloc(struct ir3
*shader
, int sz
)
44 return rzalloc_size(shader
, sz
); /* TODO: don't use rzalloc */
47 struct ir3
* ir3_create(struct ir3_compiler
*compiler
,
48 unsigned nin
, unsigned nout
)
50 struct ir3
*shader
= rzalloc(compiler
, struct ir3
);
52 shader
->compiler
= compiler
;
53 shader
->ninputs
= nin
;
54 shader
->inputs
= ir3_alloc(shader
, sizeof(shader
->inputs
[0]) * nin
);
56 shader
->noutputs
= nout
;
57 shader
->outputs
= ir3_alloc(shader
, sizeof(shader
->outputs
[0]) * nout
);
59 list_inithead(&shader
->block_list
);
60 list_inithead(&shader
->array_list
);
65 void ir3_destroy(struct ir3
*shader
)
70 #define iassert(cond) do { \
76 #define iassert_type(reg, full) do { \
78 iassert(!((reg)->flags & IR3_REG_HALF)); \
80 iassert((reg)->flags & IR3_REG_HALF); \
83 static uint32_t reg(struct ir3_register
*reg
, struct ir3_info
*info
,
84 uint32_t repeat
, uint32_t valid_flags
)
86 reg_t val
= { .dummy32
= 0 };
88 if (reg
->flags
& ~valid_flags
) {
89 debug_printf("INVALID FLAGS: %x vs %x\n",
90 reg
->flags
, valid_flags
);
93 if (!(reg
->flags
& IR3_REG_R
))
96 if (reg
->flags
& IR3_REG_IMMED
) {
97 val
.iim_val
= reg
->iim_val
;
102 if (reg
->flags
& IR3_REG_RELATIV
) {
103 components
= reg
->size
;
104 val
.idummy10
= reg
->array
.offset
;
105 max
= (reg
->array
.offset
+ repeat
+ components
- 1) >> 2;
107 components
= util_last_bit(reg
->wrmask
);
108 val
.comp
= reg
->num
& 0x3;
109 val
.num
= reg
->num
>> 2;
110 max
= (reg
->num
+ repeat
+ components
- 1) >> 2;
113 if (reg
->flags
& IR3_REG_CONST
) {
114 info
->max_const
= MAX2(info
->max_const
, max
);
115 } else if (val
.num
== 63) {
116 /* ignore writes to dummy register r63.x */
117 } else if (max
< 48) {
118 if (reg
->flags
& IR3_REG_HALF
) {
119 if (info
->gpu_id
>= 600) {
120 /* starting w/ a6xx, half regs conflict with full regs: */
121 info
->max_reg
= MAX2(info
->max_reg
, (max
+1)/2);
123 info
->max_half_reg
= MAX2(info
->max_half_reg
, max
);
126 info
->max_reg
= MAX2(info
->max_reg
, max
);
134 static int emit_cat0(struct ir3_instruction
*instr
, void *ptr
,
135 struct ir3_info
*info
)
137 instr_cat0_t
*cat0
= ptr
;
139 if (info
->gpu_id
>= 500) {
140 cat0
->a5xx
.immed
= instr
->cat0
.immed
;
141 } else if (info
->gpu_id
>= 400) {
142 cat0
->a4xx
.immed
= instr
->cat0
.immed
;
144 cat0
->a3xx
.immed
= instr
->cat0
.immed
;
146 cat0
->repeat
= instr
->repeat
;
147 cat0
->ss
= !!(instr
->flags
& IR3_INSTR_SS
);
148 cat0
->inv
= instr
->cat0
.inv
;
149 cat0
->comp
= instr
->cat0
.comp
;
150 cat0
->opc
= instr
->opc
;
151 cat0
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
152 cat0
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
158 static int emit_cat1(struct ir3_instruction
*instr
, void *ptr
,
159 struct ir3_info
*info
)
161 struct ir3_register
*dst
= instr
->regs
[0];
162 struct ir3_register
*src
= instr
->regs
[1];
163 instr_cat1_t
*cat1
= ptr
;
165 iassert(instr
->regs_count
== 2);
166 iassert_type(dst
, type_size(instr
->cat1
.dst_type
) == 32);
167 if (!(src
->flags
& IR3_REG_IMMED
))
168 iassert_type(src
, type_size(instr
->cat1
.src_type
) == 32);
170 if (src
->flags
& IR3_REG_IMMED
) {
171 cat1
->iim_val
= src
->iim_val
;
173 } else if (src
->flags
& IR3_REG_RELATIV
) {
174 cat1
->off
= reg(src
, info
, instr
->repeat
,
175 IR3_REG_R
| IR3_REG_CONST
| IR3_REG_HALF
| IR3_REG_RELATIV
);
177 cat1
->src_rel_c
= !!(src
->flags
& IR3_REG_CONST
);
179 cat1
->src
= reg(src
, info
, instr
->repeat
,
180 IR3_REG_R
| IR3_REG_CONST
| IR3_REG_HALF
);
181 cat1
->src_c
= !!(src
->flags
& IR3_REG_CONST
);
184 cat1
->dst
= reg(dst
, info
, instr
->repeat
,
185 IR3_REG_RELATIV
| IR3_REG_EVEN
|
186 IR3_REG_R
| IR3_REG_POS_INF
| IR3_REG_HALF
);
187 cat1
->repeat
= instr
->repeat
;
188 cat1
->src_r
= !!(src
->flags
& IR3_REG_R
);
189 cat1
->ss
= !!(instr
->flags
& IR3_INSTR_SS
);
190 cat1
->ul
= !!(instr
->flags
& IR3_INSTR_UL
);
191 cat1
->dst_type
= instr
->cat1
.dst_type
;
192 cat1
->dst_rel
= !!(dst
->flags
& IR3_REG_RELATIV
);
193 cat1
->src_type
= instr
->cat1
.src_type
;
194 cat1
->even
= !!(dst
->flags
& IR3_REG_EVEN
);
195 cat1
->pos_inf
= !!(dst
->flags
& IR3_REG_POS_INF
);
196 cat1
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
197 cat1
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
203 static int emit_cat2(struct ir3_instruction
*instr
, void *ptr
,
204 struct ir3_info
*info
)
206 struct ir3_register
*dst
= instr
->regs
[0];
207 struct ir3_register
*src1
= instr
->regs
[1];
208 struct ir3_register
*src2
= instr
->regs
[2];
209 instr_cat2_t
*cat2
= ptr
;
210 unsigned absneg
= ir3_cat2_absneg(instr
->opc
);
212 iassert((instr
->regs_count
== 2) || (instr
->regs_count
== 3));
215 iassert(!instr
->repeat
);
216 iassert(instr
->nop
<= 3);
218 cat2
->src1_r
= instr
->nop
& 0x1;
219 cat2
->src2_r
= (instr
->nop
>> 1) & 0x1;
221 cat2
->src1_r
= !!(src1
->flags
& IR3_REG_R
);
223 cat2
->src2_r
= !!(src2
->flags
& IR3_REG_R
);
226 if (src1
->flags
& IR3_REG_RELATIV
) {
227 iassert(src1
->array
.offset
< (1 << 10));
228 cat2
->rel1
.src1
= reg(src1
, info
, instr
->repeat
,
229 IR3_REG_RELATIV
| IR3_REG_CONST
| IR3_REG_R
|
230 IR3_REG_HALF
| absneg
);
231 cat2
->rel1
.src1_c
= !!(src1
->flags
& IR3_REG_CONST
);
232 cat2
->rel1
.src1_rel
= 1;
233 } else if (src1
->flags
& IR3_REG_CONST
) {
234 iassert(src1
->num
< (1 << 12));
235 cat2
->c1
.src1
= reg(src1
, info
, instr
->repeat
,
236 IR3_REG_CONST
| IR3_REG_R
| IR3_REG_HALF
);
239 iassert(src1
->num
< (1 << 11));
240 cat2
->src1
= reg(src1
, info
, instr
->repeat
,
241 IR3_REG_IMMED
| IR3_REG_R
| IR3_REG_HALF
|
244 cat2
->src1_im
= !!(src1
->flags
& IR3_REG_IMMED
);
245 cat2
->src1_neg
= !!(src1
->flags
& (IR3_REG_FNEG
| IR3_REG_SNEG
| IR3_REG_BNOT
));
246 cat2
->src1_abs
= !!(src1
->flags
& (IR3_REG_FABS
| IR3_REG_SABS
));
249 iassert((src2
->flags
& IR3_REG_IMMED
) ||
250 !((src1
->flags
^ src2
->flags
) & IR3_REG_HALF
));
252 if (src2
->flags
& IR3_REG_RELATIV
) {
253 iassert(src2
->array
.offset
< (1 << 10));
254 cat2
->rel2
.src2
= reg(src2
, info
, instr
->repeat
,
255 IR3_REG_RELATIV
| IR3_REG_CONST
| IR3_REG_R
|
256 IR3_REG_HALF
| absneg
);
257 cat2
->rel2
.src2_c
= !!(src2
->flags
& IR3_REG_CONST
);
258 cat2
->rel2
.src2_rel
= 1;
259 } else if (src2
->flags
& IR3_REG_CONST
) {
260 iassert(src2
->num
< (1 << 12));
261 cat2
->c2
.src2
= reg(src2
, info
, instr
->repeat
,
262 IR3_REG_CONST
| IR3_REG_R
| IR3_REG_HALF
);
265 iassert(src2
->num
< (1 << 11));
266 cat2
->src2
= reg(src2
, info
, instr
->repeat
,
267 IR3_REG_IMMED
| IR3_REG_R
| IR3_REG_HALF
|
271 cat2
->src2_im
= !!(src2
->flags
& IR3_REG_IMMED
);
272 cat2
->src2_neg
= !!(src2
->flags
& (IR3_REG_FNEG
| IR3_REG_SNEG
| IR3_REG_BNOT
));
273 cat2
->src2_abs
= !!(src2
->flags
& (IR3_REG_FABS
| IR3_REG_SABS
));
276 cat2
->dst
= reg(dst
, info
, instr
->repeat
,
277 IR3_REG_R
| IR3_REG_EI
| IR3_REG_HALF
);
278 cat2
->repeat
= instr
->repeat
;
279 cat2
->sat
= !!(instr
->flags
& IR3_INSTR_SAT
);
280 cat2
->ss
= !!(instr
->flags
& IR3_INSTR_SS
);
281 cat2
->ul
= !!(instr
->flags
& IR3_INSTR_UL
);
282 cat2
->dst_half
= !!((src1
->flags
^ dst
->flags
) & IR3_REG_HALF
);
283 cat2
->ei
= !!(dst
->flags
& IR3_REG_EI
);
284 cat2
->cond
= instr
->cat2
.condition
;
285 cat2
->full
= ! (src1
->flags
& IR3_REG_HALF
);
286 cat2
->opc
= instr
->opc
;
287 cat2
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
288 cat2
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
294 static int emit_cat3(struct ir3_instruction
*instr
, void *ptr
,
295 struct ir3_info
*info
)
297 struct ir3_register
*dst
= instr
->regs
[0];
298 struct ir3_register
*src1
= instr
->regs
[1];
299 struct ir3_register
*src2
= instr
->regs
[2];
300 struct ir3_register
*src3
= instr
->regs
[3];
301 unsigned absneg
= ir3_cat3_absneg(instr
->opc
);
302 instr_cat3_t
*cat3
= ptr
;
303 uint32_t src_flags
= 0;
305 switch (instr
->opc
) {
313 case OPC_SAD_S32
: // really??
314 src_flags
|= IR3_REG_HALF
;
320 iassert(instr
->regs_count
== 4);
321 iassert(!((src1
->flags
^ src_flags
) & IR3_REG_HALF
));
322 iassert(!((src2
->flags
^ src_flags
) & IR3_REG_HALF
));
323 iassert(!((src3
->flags
^ src_flags
) & IR3_REG_HALF
));
326 iassert(!instr
->repeat
);
327 iassert(instr
->nop
<= 3);
329 cat3
->src1_r
= instr
->nop
& 0x1;
330 cat3
->src2_r
= (instr
->nop
>> 1) & 0x1;
332 cat3
->src1_r
= !!(src1
->flags
& IR3_REG_R
);
333 cat3
->src2_r
= !!(src2
->flags
& IR3_REG_R
);
336 if (src1
->flags
& IR3_REG_RELATIV
) {
337 iassert(src1
->array
.offset
< (1 << 10));
338 cat3
->rel1
.src1
= reg(src1
, info
, instr
->repeat
,
339 IR3_REG_RELATIV
| IR3_REG_CONST
| IR3_REG_R
|
340 IR3_REG_HALF
| absneg
);
341 cat3
->rel1
.src1_c
= !!(src1
->flags
& IR3_REG_CONST
);
342 cat3
->rel1
.src1_rel
= 1;
343 } else if (src1
->flags
& IR3_REG_CONST
) {
344 iassert(src1
->num
< (1 << 12));
345 cat3
->c1
.src1
= reg(src1
, info
, instr
->repeat
,
346 IR3_REG_CONST
| IR3_REG_R
| IR3_REG_HALF
);
349 iassert(src1
->num
< (1 << 11));
350 cat3
->src1
= reg(src1
, info
, instr
->repeat
,
351 IR3_REG_R
| IR3_REG_HALF
| absneg
);
354 cat3
->src1_neg
= !!(src1
->flags
& (IR3_REG_FNEG
| IR3_REG_SNEG
| IR3_REG_BNOT
));
356 cat3
->src2
= reg(src2
, info
, instr
->repeat
,
357 IR3_REG_CONST
| IR3_REG_R
| IR3_REG_HALF
| absneg
);
358 cat3
->src2_c
= !!(src2
->flags
& IR3_REG_CONST
);
359 cat3
->src2_neg
= !!(src2
->flags
& (IR3_REG_FNEG
| IR3_REG_SNEG
| IR3_REG_BNOT
));
361 if (src3
->flags
& IR3_REG_RELATIV
) {
362 iassert(src3
->array
.offset
< (1 << 10));
363 cat3
->rel2
.src3
= reg(src3
, info
, instr
->repeat
,
364 IR3_REG_RELATIV
| IR3_REG_CONST
| IR3_REG_R
|
365 IR3_REG_HALF
| absneg
);
366 cat3
->rel2
.src3_c
= !!(src3
->flags
& IR3_REG_CONST
);
367 cat3
->rel2
.src3_rel
= 1;
368 } else if (src3
->flags
& IR3_REG_CONST
) {
369 iassert(src3
->num
< (1 << 12));
370 cat3
->c2
.src3
= reg(src3
, info
, instr
->repeat
,
371 IR3_REG_CONST
| IR3_REG_R
| IR3_REG_HALF
);
374 iassert(src3
->num
< (1 << 11));
375 cat3
->src3
= reg(src3
, info
, instr
->repeat
,
376 IR3_REG_R
| IR3_REG_HALF
| absneg
);
379 cat3
->src3_neg
= !!(src3
->flags
& (IR3_REG_FNEG
| IR3_REG_SNEG
| IR3_REG_BNOT
));
380 cat3
->src3_r
= !!(src3
->flags
& IR3_REG_R
);
382 cat3
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
383 cat3
->repeat
= instr
->repeat
;
384 cat3
->sat
= !!(instr
->flags
& IR3_INSTR_SAT
);
385 cat3
->ss
= !!(instr
->flags
& IR3_INSTR_SS
);
386 cat3
->ul
= !!(instr
->flags
& IR3_INSTR_UL
);
387 cat3
->dst_half
= !!((src_flags
^ dst
->flags
) & IR3_REG_HALF
);
388 cat3
->opc
= instr
->opc
;
389 cat3
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
390 cat3
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
396 static int emit_cat4(struct ir3_instruction
*instr
, void *ptr
,
397 struct ir3_info
*info
)
399 struct ir3_register
*dst
= instr
->regs
[0];
400 struct ir3_register
*src
= instr
->regs
[1];
401 instr_cat4_t
*cat4
= ptr
;
403 iassert(instr
->regs_count
== 2);
405 if (src
->flags
& IR3_REG_RELATIV
) {
406 iassert(src
->array
.offset
< (1 << 10));
407 cat4
->rel
.src
= reg(src
, info
, instr
->repeat
,
408 IR3_REG_RELATIV
| IR3_REG_CONST
| IR3_REG_FNEG
|
409 IR3_REG_FABS
| IR3_REG_R
| IR3_REG_HALF
);
410 cat4
->rel
.src_c
= !!(src
->flags
& IR3_REG_CONST
);
411 cat4
->rel
.src_rel
= 1;
412 } else if (src
->flags
& IR3_REG_CONST
) {
413 iassert(src
->num
< (1 << 12));
414 cat4
->c
.src
= reg(src
, info
, instr
->repeat
,
415 IR3_REG_CONST
| IR3_REG_FNEG
| IR3_REG_FABS
|
416 IR3_REG_R
| IR3_REG_HALF
);
419 iassert(src
->num
< (1 << 11));
420 cat4
->src
= reg(src
, info
, instr
->repeat
,
421 IR3_REG_IMMED
| IR3_REG_FNEG
| IR3_REG_FABS
|
422 IR3_REG_R
| IR3_REG_HALF
);
425 cat4
->src_im
= !!(src
->flags
& IR3_REG_IMMED
);
426 cat4
->src_neg
= !!(src
->flags
& IR3_REG_FNEG
);
427 cat4
->src_abs
= !!(src
->flags
& IR3_REG_FABS
);
428 cat4
->src_r
= !!(src
->flags
& IR3_REG_R
);
430 cat4
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
431 cat4
->repeat
= instr
->repeat
;
432 cat4
->sat
= !!(instr
->flags
& IR3_INSTR_SAT
);
433 cat4
->ss
= !!(instr
->flags
& IR3_INSTR_SS
);
434 cat4
->ul
= !!(instr
->flags
& IR3_INSTR_UL
);
435 cat4
->dst_half
= !!((src
->flags
^ dst
->flags
) & IR3_REG_HALF
);
436 cat4
->full
= ! (src
->flags
& IR3_REG_HALF
);
437 cat4
->opc
= instr
->opc
;
438 cat4
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
439 cat4
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
445 static int emit_cat5(struct ir3_instruction
*instr
, void *ptr
,
446 struct ir3_info
*info
)
448 struct ir3_register
*dst
= instr
->regs
[0];
449 struct ir3_register
*src1
= instr
->regs
[1];
450 struct ir3_register
*src2
= instr
->regs
[2];
451 struct ir3_register
*src3
= instr
->regs
[3];
452 instr_cat5_t
*cat5
= ptr
;
454 iassert_type(dst
, type_size(instr
->cat5
.type
) == 32)
456 assume(src1
|| !src2
);
457 assume(src2
|| !src3
);
460 cat5
->full
= ! (src1
->flags
& IR3_REG_HALF
);
461 cat5
->src1
= reg(src1
, info
, instr
->repeat
, IR3_REG_HALF
);
464 if (instr
->flags
& IR3_INSTR_S2EN
) {
466 iassert(!((src1
->flags
^ src2
->flags
) & IR3_REG_HALF
));
467 cat5
->s2en
.src2
= reg(src2
, info
, instr
->repeat
, IR3_REG_HALF
);
470 iassert(src3
->flags
& IR3_REG_HALF
);
471 cat5
->s2en
.src3
= reg(src3
, info
, instr
->repeat
, IR3_REG_HALF
);
473 iassert(!(instr
->cat5
.samp
| instr
->cat5
.tex
));
477 iassert(!((src1
->flags
^ src2
->flags
) & IR3_REG_HALF
));
478 cat5
->norm
.src2
= reg(src2
, info
, instr
->repeat
, IR3_REG_HALF
);
480 cat5
->norm
.samp
= instr
->cat5
.samp
;
481 cat5
->norm
.tex
= instr
->cat5
.tex
;
484 cat5
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
485 cat5
->wrmask
= dst
->wrmask
;
486 cat5
->type
= instr
->cat5
.type
;
487 cat5
->is_3d
= !!(instr
->flags
& IR3_INSTR_3D
);
488 cat5
->is_a
= !!(instr
->flags
& IR3_INSTR_A
);
489 cat5
->is_s
= !!(instr
->flags
& IR3_INSTR_S
);
490 cat5
->is_s2en
= !!(instr
->flags
& IR3_INSTR_S2EN
);
491 cat5
->is_o
= !!(instr
->flags
& IR3_INSTR_O
);
492 cat5
->is_p
= !!(instr
->flags
& IR3_INSTR_P
);
493 cat5
->opc
= instr
->opc
;
494 cat5
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
495 cat5
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
501 static int emit_cat6_a6xx(struct ir3_instruction
*instr
, void *ptr
,
502 struct ir3_info
*info
)
504 struct ir3_register
*src1
, *src2
;
505 instr_cat6_a6xx_t
*cat6
= ptr
;
506 bool has_dest
= (instr
->opc
== OPC_LDIB
);
508 /* first reg should be SSBO binding point: */
509 iassert(instr
->regs
[1]->flags
& IR3_REG_IMMED
);
511 src1
= instr
->regs
[2];
514 /* the src2 field in the instruction is actually the destination
515 * register for load instructions:
517 src2
= instr
->regs
[0];
519 src2
= instr
->regs
[3];
522 cat6
->type
= instr
->cat6
.type
;
523 cat6
->d
= instr
->cat6
.d
- 1;
524 cat6
->typed
= instr
->cat6
.typed
;
525 cat6
->type_size
= instr
->cat6
.iim_val
- 1;
526 cat6
->opc
= instr
->opc
;
527 cat6
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
528 cat6
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
531 cat6
->src1
= reg(src1
, info
, instr
->repeat
, 0);
532 cat6
->src2
= reg(src2
, info
, instr
->repeat
, 0);
533 cat6
->ssbo
= instr
->regs
[1]->iim_val
;
535 switch (instr
->opc
) {
538 case OPC_ATOMIC_XCHG
:
541 case OPC_ATOMIC_CMPXCHG
:
577 static int emit_cat6(struct ir3_instruction
*instr
, void *ptr
,
578 struct ir3_info
*info
)
580 struct ir3_register
*dst
, *src1
, *src2
;
581 instr_cat6_t
*cat6
= ptr
;
583 /* In a6xx we start using a new instruction encoding for some of
584 * these instructions:
586 if (info
->gpu_id
>= 600) {
587 switch (instr
->opc
) {
590 case OPC_ATOMIC_XCHG
:
593 case OPC_ATOMIC_CMPXCHG
:
599 /* The shared variants of these still use the old encoding: */
600 if (!(instr
->flags
& IR3_INSTR_G
))
606 return emit_cat6_a6xx(instr
, ptr
, info
);
612 bool type_full
= type_size(instr
->cat6
.type
) == 32;
614 cat6
->type
= instr
->cat6
.type
;
615 cat6
->opc
= instr
->opc
;
616 cat6
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
617 cat6
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
618 cat6
->g
= !!(instr
->flags
& IR3_INSTR_G
);
621 switch (instr
->opc
) {
624 iassert_type(instr
->regs
[0], type_full
); /* dst */
625 iassert_type(instr
->regs
[1], type_full
); /* src1 */
629 iassert_type(instr
->regs
[0], true); /* dst */
630 iassert_type(instr
->regs
[1], true); /* src1 */
637 /* no dst, so regs[0] is dummy */
638 iassert_type(instr
->regs
[1], true); /* dst */
639 iassert_type(instr
->regs
[2], type_full
); /* src1 */
640 iassert_type(instr
->regs
[3], true); /* src2 */
643 iassert_type(instr
->regs
[0], type_full
); /* dst */
644 iassert_type(instr
->regs
[1], true); /* src1 */
645 if (instr
->regs_count
> 2)
646 iassert_type(instr
->regs
[2], true); /* src1 */
650 /* the "dst" for a store instruction is (from the perspective
651 * of data flow in the shader, ie. register use/def, etc) in
652 * fact a register that is read by the instruction, rather
655 if (is_store(instr
)) {
656 iassert(instr
->regs_count
>= 3);
658 dst
= instr
->regs
[1];
659 src1
= instr
->regs
[2];
660 src2
= (instr
->regs_count
>= 4) ? instr
->regs
[3] : NULL
;
662 iassert(instr
->regs_count
>= 2);
664 dst
= instr
->regs
[0];
665 src1
= instr
->regs
[1];
666 src2
= (instr
->regs_count
>= 3) ? instr
->regs
[2] : NULL
;
669 /* TODO we need a more comprehensive list about which instructions
670 * can be encoded which way. Or possibly use IR3_INSTR_0 flag to
671 * indicate to use the src_off encoding even if offset is zero
672 * (but then what to do about dst_off?)
674 if (is_atomic(instr
->opc
)) {
675 instr_cat6ldgb_t
*ldgb
= ptr
;
677 /* maybe these two bits both determine the instruction encoding? */
678 cat6
->src_off
= false;
680 ldgb
->d
= instr
->cat6
.d
- 1;
681 ldgb
->typed
= instr
->cat6
.typed
;
682 ldgb
->type_size
= instr
->cat6
.iim_val
- 1;
684 ldgb
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
687 struct ir3_register
*src3
= instr
->regs
[3];
688 struct ir3_register
*src4
= instr
->regs
[4];
690 /* first src is src_ssbo: */
691 iassert(src1
->flags
& IR3_REG_IMMED
);
692 ldgb
->src_ssbo
= src1
->uim_val
;
694 ldgb
->src1
= reg(src2
, info
, instr
->repeat
, IR3_REG_IMMED
);
695 ldgb
->src1_im
= !!(src2
->flags
& IR3_REG_IMMED
);
696 ldgb
->src2
= reg(src3
, info
, instr
->repeat
, IR3_REG_IMMED
);
697 ldgb
->src2_im
= !!(src3
->flags
& IR3_REG_IMMED
);
699 ldgb
->src3
= reg(src4
, info
, instr
->repeat
, 0);
703 ldgb
->src1
= reg(src1
, info
, instr
->repeat
, IR3_REG_IMMED
);
704 ldgb
->src1_im
= !!(src1
->flags
& IR3_REG_IMMED
);
705 ldgb
->src2
= reg(src2
, info
, instr
->repeat
, IR3_REG_IMMED
);
706 ldgb
->src2_im
= !!(src2
->flags
& IR3_REG_IMMED
);
712 } else if (instr
->opc
== OPC_LDGB
) {
713 struct ir3_register
*src3
= instr
->regs
[3];
714 instr_cat6ldgb_t
*ldgb
= ptr
;
716 /* maybe these two bits both determine the instruction encoding? */
717 cat6
->src_off
= false;
719 ldgb
->d
= instr
->cat6
.d
- 1;
720 ldgb
->typed
= instr
->cat6
.typed
;
721 ldgb
->type_size
= instr
->cat6
.iim_val
- 1;
723 ldgb
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
725 /* first src is src_ssbo: */
726 iassert(src1
->flags
& IR3_REG_IMMED
);
727 ldgb
->src_ssbo
= src1
->uim_val
;
729 /* then next two are src1/src2: */
730 ldgb
->src1
= reg(src2
, info
, instr
->repeat
, IR3_REG_IMMED
);
731 ldgb
->src1_im
= !!(src2
->flags
& IR3_REG_IMMED
);
732 ldgb
->src2
= reg(src3
, info
, instr
->repeat
, IR3_REG_IMMED
);
733 ldgb
->src2_im
= !!(src3
->flags
& IR3_REG_IMMED
);
739 } else if (instr
->opc
== OPC_RESINFO
) {
740 instr_cat6ldgb_t
*ldgb
= ptr
;
742 ldgb
->d
= instr
->cat6
.d
- 1;
744 ldgb
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
746 /* first src is src_ssbo: */
747 iassert(src1
->flags
& IR3_REG_IMMED
);
748 ldgb
->src_ssbo
= src1
->uim_val
;
751 } else if ((instr
->opc
== OPC_STGB
) || (instr
->opc
== OPC_STIB
)) {
752 struct ir3_register
*src3
= instr
->regs
[4];
753 instr_cat6stgb_t
*stgb
= ptr
;
755 /* maybe these two bits both determine the instruction encoding? */
756 cat6
->src_off
= true;
759 stgb
->d
= instr
->cat6
.d
- 1;
760 stgb
->typed
= instr
->cat6
.typed
;
761 stgb
->type_size
= instr
->cat6
.iim_val
- 1;
763 /* first src is dst_ssbo: */
764 iassert(dst
->flags
& IR3_REG_IMMED
);
765 stgb
->dst_ssbo
= dst
->uim_val
;
767 /* then src1/src2/src3: */
768 stgb
->src1
= reg(src1
, info
, instr
->repeat
, 0);
769 stgb
->src2
= reg(src2
, info
, instr
->repeat
, IR3_REG_IMMED
);
770 stgb
->src2_im
= !!(src2
->flags
& IR3_REG_IMMED
);
771 stgb
->src3
= reg(src3
, info
, instr
->repeat
, IR3_REG_IMMED
);
772 stgb
->src3_im
= !!(src3
->flags
& IR3_REG_IMMED
);
775 } else if (instr
->cat6
.src_offset
|| (instr
->opc
== OPC_LDG
) ||
776 (instr
->opc
== OPC_LDL
)) {
777 instr_cat6a_t
*cat6a
= ptr
;
779 cat6
->src_off
= true;
781 cat6a
->src1
= reg(src1
, info
, instr
->repeat
, IR3_REG_IMMED
);
782 cat6a
->src1_im
= !!(src1
->flags
& IR3_REG_IMMED
);
784 cat6a
->src2
= reg(src2
, info
, instr
->repeat
, IR3_REG_IMMED
);
785 cat6a
->src2_im
= !!(src2
->flags
& IR3_REG_IMMED
);
787 cat6a
->off
= instr
->cat6
.src_offset
;
789 instr_cat6b_t
*cat6b
= ptr
;
791 cat6
->src_off
= false;
793 cat6b
->src1
= reg(src1
, info
, instr
->repeat
, IR3_REG_IMMED
| IR3_REG_HALF
);
794 cat6b
->src1_im
= !!(src1
->flags
& IR3_REG_IMMED
);
796 cat6b
->src2
= reg(src2
, info
, instr
->repeat
, IR3_REG_IMMED
);
797 cat6b
->src2_im
= !!(src2
->flags
& IR3_REG_IMMED
);
801 if (instr
->cat6
.dst_offset
|| (instr
->opc
== OPC_STG
) ||
802 (instr
->opc
== OPC_STL
)) {
803 instr_cat6c_t
*cat6c
= ptr
;
804 cat6
->dst_off
= true;
805 cat6c
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
806 cat6c
->off
= instr
->cat6
.dst_offset
;
808 instr_cat6d_t
*cat6d
= ptr
;
809 cat6
->dst_off
= false;
810 cat6d
->dst
= reg(dst
, info
, instr
->repeat
, IR3_REG_R
| IR3_REG_HALF
);
816 static int emit_cat7(struct ir3_instruction
*instr
, void *ptr
,
817 struct ir3_info
*info
)
819 instr_cat7_t
*cat7
= ptr
;
821 cat7
->ss
= !!(instr
->flags
& IR3_INSTR_SS
);
822 cat7
->w
= instr
->cat7
.w
;
823 cat7
->r
= instr
->cat7
.r
;
824 cat7
->l
= instr
->cat7
.l
;
825 cat7
->g
= instr
->cat7
.g
;
826 cat7
->opc
= instr
->opc
;
827 cat7
->jmp_tgt
= !!(instr
->flags
& IR3_INSTR_JP
);
828 cat7
->sync
= !!(instr
->flags
& IR3_INSTR_SY
);
834 static int (*emit
[])(struct ir3_instruction
*instr
, void *ptr
,
835 struct ir3_info
*info
) = {
836 emit_cat0
, emit_cat1
, emit_cat2
, emit_cat3
, emit_cat4
, emit_cat5
, emit_cat6
,
840 void * ir3_assemble(struct ir3
*shader
, struct ir3_info
*info
,
843 uint32_t *ptr
, *dwords
;
845 info
->gpu_id
= gpu_id
;
847 info
->max_half_reg
= -1;
848 info
->max_const
= -1;
849 info
->instrs_count
= 0;
850 info
->sizedwords
= 0;
851 info
->ss
= info
->sy
= 0;
853 list_for_each_entry (struct ir3_block
, block
, &shader
->block_list
, node
) {
854 list_for_each_entry (struct ir3_instruction
, instr
, &block
->instr_list
, node
) {
855 info
->sizedwords
+= 2;
859 /* need an integer number of instruction "groups" (sets of 16
860 * instructions on a4xx or sets of 4 instructions on a3xx),
861 * so pad out w/ NOPs if needed: (NOTE each instruction is 64bits)
864 info
->sizedwords
= align(info
->sizedwords
, 16 * 2);
866 info
->sizedwords
= align(info
->sizedwords
, 4 * 2);
869 ptr
= dwords
= calloc(4, info
->sizedwords
);
871 list_for_each_entry (struct ir3_block
, block
, &shader
->block_list
, node
) {
872 list_for_each_entry (struct ir3_instruction
, instr
, &block
->instr_list
, node
) {
873 int ret
= emit
[opc_cat(instr
->opc
)](instr
, dwords
, info
);
876 info
->instrs_count
+= 1 + instr
->repeat
;
879 if (instr
->flags
& IR3_INSTR_SS
)
882 if (instr
->flags
& IR3_INSTR_SY
)
894 static struct ir3_register
* reg_create(struct ir3
*shader
,
897 struct ir3_register
*reg
=
898 ir3_alloc(shader
, sizeof(struct ir3_register
));
905 static void insert_instr(struct ir3_block
*block
,
906 struct ir3_instruction
*instr
)
908 struct ir3
*shader
= block
->shader
;
910 instr
->serialno
= ++shader
->instr_count
;
912 list_addtail(&instr
->node
, &block
->instr_list
);
915 array_insert(shader
, shader
->baryfs
, instr
);
918 struct ir3_block
* ir3_block_create(struct ir3
*shader
)
920 struct ir3_block
*block
= ir3_alloc(shader
, sizeof(*block
));
922 block
->serialno
= ++shader
->block_count
;
924 block
->shader
= shader
;
925 list_inithead(&block
->node
);
926 list_inithead(&block
->instr_list
);
930 static struct ir3_instruction
*instr_create(struct ir3_block
*block
, int nreg
)
932 struct ir3_instruction
*instr
;
933 unsigned sz
= sizeof(*instr
) + (nreg
* sizeof(instr
->regs
[0]));
934 char *ptr
= ir3_alloc(block
->shader
, sz
);
936 instr
= (struct ir3_instruction
*)ptr
;
937 ptr
+= sizeof(*instr
);
938 instr
->regs
= (struct ir3_register
**)ptr
;
941 instr
->regs_max
= nreg
;
947 struct ir3_instruction
* ir3_instr_create2(struct ir3_block
*block
,
950 struct ir3_instruction
*instr
= instr_create(block
, nreg
);
951 instr
->block
= block
;
953 insert_instr(block
, instr
);
957 struct ir3_instruction
* ir3_instr_create(struct ir3_block
*block
, opc_t opc
)
959 /* NOTE: we could be slightly more clever, at least for non-meta,
960 * and choose # of regs based on category.
962 return ir3_instr_create2(block
, opc
, 4);
965 struct ir3_instruction
* ir3_instr_clone(struct ir3_instruction
*instr
)
967 struct ir3_instruction
*new_instr
= instr_create(instr
->block
,
969 struct ir3_register
**regs
;
972 regs
= new_instr
->regs
;
974 new_instr
->regs
= regs
;
976 insert_instr(instr
->block
, new_instr
);
978 /* clone registers: */
979 new_instr
->regs_count
= 0;
980 for (i
= 0; i
< instr
->regs_count
; i
++) {
981 struct ir3_register
*reg
= instr
->regs
[i
];
982 struct ir3_register
*new_reg
=
983 ir3_reg_create(new_instr
, reg
->num
, reg
->flags
);
990 /* Add a false dependency to instruction, to ensure it is scheduled first: */
991 void ir3_instr_add_dep(struct ir3_instruction
*instr
, struct ir3_instruction
*dep
)
993 array_insert(instr
, instr
->deps
, dep
);
996 struct ir3_register
* ir3_reg_create(struct ir3_instruction
*instr
,
999 struct ir3
*shader
= instr
->block
->shader
;
1000 struct ir3_register
*reg
= reg_create(shader
, num
, flags
);
1002 debug_assert(instr
->regs_count
< instr
->regs_max
);
1004 instr
->regs
[instr
->regs_count
++] = reg
;
1008 struct ir3_register
* ir3_reg_clone(struct ir3
*shader
,
1009 struct ir3_register
*reg
)
1011 struct ir3_register
*new_reg
= reg_create(shader
, 0, 0);
1017 ir3_instr_set_address(struct ir3_instruction
*instr
,
1018 struct ir3_instruction
*addr
)
1020 if (instr
->address
!= addr
) {
1021 struct ir3
*ir
= instr
->block
->shader
;
1022 instr
->address
= addr
;
1023 array_insert(ir
, ir
->indirects
, instr
);
1028 ir3_block_clear_mark(struct ir3_block
*block
)
1030 list_for_each_entry (struct ir3_instruction
, instr
, &block
->instr_list
, node
)
1031 instr
->flags
&= ~IR3_INSTR_MARK
;
1035 ir3_clear_mark(struct ir3
*ir
)
1037 list_for_each_entry (struct ir3_block
, block
, &ir
->block_list
, node
) {
1038 ir3_block_clear_mark(block
);
1042 /* note: this will destroy instr->depth, don't do it until after sched! */
1044 ir3_count_instructions(struct ir3
*ir
)
1047 list_for_each_entry (struct ir3_block
, block
, &ir
->block_list
, node
) {
1048 list_for_each_entry (struct ir3_instruction
, instr
, &block
->instr_list
, node
) {
1051 block
->start_ip
= list_first_entry(&block
->instr_list
, struct ir3_instruction
, node
)->ip
;
1052 block
->end_ip
= list_last_entry(&block
->instr_list
, struct ir3_instruction
, node
)->ip
;
1058 ir3_lookup_array(struct ir3
*ir
, unsigned id
)
1060 list_for_each_entry (struct ir3_array
, arr
, &ir
->array_list
, node
)