2 * Copyright © 2014 Broadcom
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
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 #include "util/u_memory.h"
25 #include "util/ralloc.h"
33 bool has_side_effects
;
36 static const struct qir_op_info qir_op_info
[] = {
37 [QOP_MOV
] = { "mov", 1, 1 },
38 [QOP_FMOV
] = { "fmov", 1, 1 },
39 [QOP_MMOV
] = { "mmov", 1, 1 },
40 [QOP_FADD
] = { "fadd", 1, 2 },
41 [QOP_FSUB
] = { "fsub", 1, 2 },
42 [QOP_FMUL
] = { "fmul", 1, 2 },
43 [QOP_MUL24
] = { "mul24", 1, 2 },
44 [QOP_V8MULD
] = {"v8muld", 1, 2 },
45 [QOP_V8MIN
] = {"v8min", 1, 2 },
46 [QOP_V8MAX
] = {"v8max", 1, 2 },
47 [QOP_V8ADDS
] = {"v8adds", 1, 2 },
48 [QOP_V8SUBS
] = {"v8subs", 1, 2 },
49 [QOP_FMIN
] = { "fmin", 1, 2 },
50 [QOP_FMAX
] = { "fmax", 1, 2 },
51 [QOP_FMINABS
] = { "fminabs", 1, 2 },
52 [QOP_FMAXABS
] = { "fmaxabs", 1, 2 },
53 [QOP_FTOI
] = { "ftoi", 1, 1 },
54 [QOP_ITOF
] = { "itof", 1, 1 },
55 [QOP_ADD
] = { "add", 1, 2 },
56 [QOP_SUB
] = { "sub", 1, 2 },
57 [QOP_SHR
] = { "shr", 1, 2 },
58 [QOP_ASR
] = { "asr", 1, 2 },
59 [QOP_SHL
] = { "shl", 1, 2 },
60 [QOP_MIN
] = { "min", 1, 2 },
61 [QOP_MAX
] = { "max", 1, 2 },
62 [QOP_AND
] = { "and", 1, 2 },
63 [QOP_OR
] = { "or", 1, 2 },
64 [QOP_XOR
] = { "xor", 1, 2 },
65 [QOP_NOT
] = { "not", 1, 1 },
67 [QOP_RCP
] = { "rcp", 1, 1 },
68 [QOP_RSQ
] = { "rsq", 1, 1 },
69 [QOP_EXP2
] = { "exp2", 1, 1 },
70 [QOP_LOG2
] = { "log2", 1, 1 },
71 [QOP_TLB_COLOR_READ
] = { "tlb_color_read", 1, 0 },
72 [QOP_MS_MASK
] = { "ms_mask", 0, 1, true },
73 [QOP_VARY_ADD_C
] = { "vary_add_c", 1, 1 },
75 [QOP_FRAG_Z
] = { "frag_z", 1, 0 },
76 [QOP_FRAG_W
] = { "frag_w", 1, 0 },
78 [QOP_TEX_S
] = { "tex_s", 0, 2, true },
79 [QOP_TEX_T
] = { "tex_t", 0, 2, true },
80 [QOP_TEX_R
] = { "tex_r", 0, 2, true },
81 [QOP_TEX_B
] = { "tex_b", 0, 2, true },
82 [QOP_TEX_DIRECT
] = { "tex_direct", 0, 2, true },
83 [QOP_TEX_RESULT
] = { "tex_result", 1, 0, true },
85 [QOP_LOAD_IMM
] = { "load_imm", 0, 1 },
87 [QOP_BRANCH
] = { "branch", 0, 0, true },
88 [QOP_UNIFORMS_RESET
] = { "uniforms_reset", 0, 2, true },
92 qir_get_op_name(enum qop qop
)
94 if (qop
< ARRAY_SIZE(qir_op_info
) && qir_op_info
[qop
].name
)
95 return qir_op_info
[qop
].name
;
101 qir_get_op_nsrc(enum qop qop
)
103 if (qop
< ARRAY_SIZE(qir_op_info
) && qir_op_info
[qop
].name
)
104 return qir_op_info
[qop
].nsrc
;
110 * Returns whether the instruction has any side effects that must be
114 qir_has_side_effects(struct vc4_compile
*c
, struct qinst
*inst
)
116 switch (inst
->dst
.file
) {
117 case QFILE_TLB_Z_WRITE
:
118 case QFILE_TLB_COLOR_WRITE
:
119 case QFILE_TLB_COLOR_WRITE_MS
:
120 case QFILE_TLB_STENCIL_SETUP
:
126 return qir_op_info
[inst
->op
].has_side_effects
;
130 qir_has_side_effect_reads(struct vc4_compile
*c
, struct qinst
*inst
)
132 /* We can dead-code eliminate varyings, because we only tell the VS
133 * about the live ones at the end. But we have to preserve the
134 * point/line coordinates reads, because they're generated by
135 * fixed-function hardware.
137 for (int i
= 0; i
< qir_get_op_nsrc(inst
->op
); i
++) {
138 if (inst
->src
[i
].file
== QFILE_VARY
&&
139 c
->input_slots
[inst
->src
[i
].index
].slot
== 0xff) {
143 if (inst
->src
[i
].file
== QFILE_VPM
)
147 if (inst
->dst
.file
== QFILE_VPM
)
154 qir_is_mul(struct qinst
*inst
)
172 qir_is_float_input(struct qinst
*inst
)
191 qir_is_raw_mov(struct qinst
*inst
)
193 return ((inst
->op
== QOP_MOV
||
194 inst
->op
== QOP_FMOV
||
195 inst
->op
== QOP_MMOV
) &&
196 inst
->cond
== QPU_COND_ALWAYS
&&
202 qir_is_tex(struct qinst
*inst
)
204 return inst
->op
>= QOP_TEX_S
&& inst
->op
<= QOP_TEX_DIRECT
;
208 qir_depends_on_flags(struct qinst
*inst
)
210 if (inst
->op
== QOP_BRANCH
) {
211 return inst
->cond
!= QPU_COND_BRANCH_ALWAYS
;
213 return (inst
->cond
!= QPU_COND_ALWAYS
&&
214 inst
->cond
!= QPU_COND_NEVER
);
219 qir_writes_r4(struct qinst
*inst
)
223 case QOP_TLB_COLOR_READ
:
235 qir_channels_written(struct qinst
*inst
)
237 if (qir_is_mul(inst
)) {
238 switch (inst
->dst
.pack
) {
239 case QPU_PACK_MUL_NOP
:
240 case QPU_PACK_MUL_8888
:
242 case QPU_PACK_MUL_8A
:
244 case QPU_PACK_MUL_8B
:
246 case QPU_PACK_MUL_8C
:
248 case QPU_PACK_MUL_8D
:
252 switch (inst
->dst
.pack
) {
254 case QPU_PACK_A_8888
:
255 case QPU_PACK_A_8888_SAT
:
256 case QPU_PACK_A_32_SAT
:
259 case QPU_PACK_A_8A_SAT
:
262 case QPU_PACK_A_8B_SAT
:
265 case QPU_PACK_A_8C_SAT
:
268 case QPU_PACK_A_8D_SAT
:
271 case QPU_PACK_A_16A_SAT
:
274 case QPU_PACK_A_16B_SAT
:
278 unreachable("Bad pack field");
282 qir_print_reg(struct vc4_compile
*c
, struct qreg reg
, bool write
)
284 static const char *files
[] = {
288 [QFILE_TLB_COLOR_WRITE
] = "tlb_c",
289 [QFILE_TLB_COLOR_WRITE_MS
] = "tlb_c_ms",
290 [QFILE_TLB_Z_WRITE
] = "tlb_z",
291 [QFILE_TLB_STENCIL_SETUP
] = "tlb_stencil",
292 [QFILE_FRAG_X
] = "frag_x",
293 [QFILE_FRAG_Y
] = "frag_y",
294 [QFILE_FRAG_REV_FLAG
] = "frag_rev_flag",
295 [QFILE_QPU_ELEMENT
] = "elem",
301 fprintf(stderr
, "null");
305 fprintf(stderr
, "0x%08x (%f)", reg
.index
, uif(reg
.index
));
308 case QFILE_SMALL_IMM
:
309 if ((int)reg
.index
>= -16 && (int)reg
.index
<= 15)
310 fprintf(stderr
, "%d", reg
.index
);
312 fprintf(stderr
, "%f", uif(reg
.index
));
317 fprintf(stderr
, "vpm");
319 fprintf(stderr
, "vpm%d.%d",
320 reg
.index
/ 4, reg
.index
% 4);
324 case QFILE_TLB_COLOR_WRITE
:
325 case QFILE_TLB_COLOR_WRITE_MS
:
326 case QFILE_TLB_Z_WRITE
:
327 case QFILE_TLB_STENCIL_SETUP
:
328 fprintf(stderr
, "%s", files
[reg
.file
]);
332 fprintf(stderr
, "%s%d", files
[reg
.file
], reg
.index
);
336 if (reg
.file
== QFILE_UNIF
&&
337 c
->uniform_contents
[reg
.index
] == QUNIFORM_CONSTANT
) {
338 fprintf(stderr
, " (0x%08x / %f)",
339 c
->uniform_data
[reg
.index
],
340 uif(c
->uniform_data
[reg
.index
]));
345 qir_dump_inst(struct vc4_compile
*c
, struct qinst
*inst
)
347 fprintf(stderr
, "%s", qir_get_op_name(inst
->op
));
348 if (inst
->op
== QOP_BRANCH
)
349 vc4_qpu_disasm_cond_branch(stderr
, inst
->cond
);
351 vc4_qpu_disasm_cond(stderr
, inst
->cond
);
353 fprintf(stderr
, ".sf");
354 fprintf(stderr
, " ");
356 if (inst
->op
!= QOP_BRANCH
) {
357 qir_print_reg(c
, inst
->dst
, true);
358 if (inst
->dst
.pack
) {
359 if (inst
->dst
.pack
) {
360 if (qir_is_mul(inst
))
361 vc4_qpu_disasm_pack_mul(stderr
, inst
->dst
.pack
);
363 vc4_qpu_disasm_pack_a(stderr
, inst
->dst
.pack
);
368 for (int i
= 0; i
< qir_get_op_nsrc(inst
->op
); i
++) {
369 fprintf(stderr
, ", ");
370 qir_print_reg(c
, inst
->src
[i
], false);
371 vc4_qpu_disasm_unpack(stderr
, inst
->src
[i
].pack
);
376 qir_dump(struct vc4_compile
*c
)
380 qir_for_each_block(block
, c
) {
381 fprintf(stderr
, "BLOCK %d:\n", block
->index
);
382 qir_for_each_inst(inst
, block
) {
386 for (int i
= 0; i
< c
->num_temps
; i
++) {
387 if (c
->temp_start
[i
] != ip
)
393 fprintf(stderr
, ", ");
395 fprintf(stderr
, "S%4d", i
);
399 fprintf(stderr
, " ");
401 fprintf(stderr
, " ");
407 for (int i
= 0; i
< c
->num_temps
; i
++) {
408 if (c
->temp_end
[i
] != ip
)
414 fprintf(stderr
, ", ");
416 fprintf(stderr
, "E%4d", i
);
420 fprintf(stderr
, " ");
422 fprintf(stderr
, " ");
425 qir_dump_inst(c
, inst
);
426 fprintf(stderr
, "\n");
429 if (block
->successors
[1]) {
430 fprintf(stderr
, "-> BLOCK %d, %d\n",
431 block
->successors
[0]->index
,
432 block
->successors
[1]->index
);
433 } else if (block
->successors
[0]) {
434 fprintf(stderr
, "-> BLOCK %d\n",
435 block
->successors
[0]->index
);
441 qir_get_temp(struct vc4_compile
*c
)
445 reg
.file
= QFILE_TEMP
;
446 reg
.index
= c
->num_temps
++;
449 if (c
->num_temps
> c
->defs_array_size
) {
450 uint32_t old_size
= c
->defs_array_size
;
451 c
->defs_array_size
= MAX2(old_size
* 2, 16);
452 c
->defs
= reralloc(c
, c
->defs
, struct qinst
*,
454 memset(&c
->defs
[old_size
], 0,
455 sizeof(c
->defs
[0]) * (c
->defs_array_size
- old_size
));
462 qir_inst(enum qop op
, struct qreg dst
, struct qreg src0
, struct qreg src1
)
464 struct qinst
*inst
= CALLOC_STRUCT(qinst
);
468 inst
->src
= calloc(2, sizeof(inst
->src
[0]));
471 inst
->cond
= QPU_COND_ALWAYS
;
477 qir_inst4(enum qop op
, struct qreg dst
,
483 struct qinst
*inst
= CALLOC_STRUCT(qinst
);
487 inst
->src
= calloc(4, sizeof(*inst
->src
));
497 qir_emit(struct vc4_compile
*c
, struct qinst
*inst
)
499 list_addtail(&inst
->link
, &c
->cur_block
->instructions
);
502 /* Updates inst to write to a new temporary, emits it, and notes the def. */
504 qir_emit_def(struct vc4_compile
*c
, struct qinst
*inst
)
506 assert(inst
->dst
.file
== QFILE_NULL
);
508 inst
->dst
= qir_get_temp(c
);
510 if (inst
->dst
.file
== QFILE_TEMP
)
511 c
->defs
[inst
->dst
.index
] = inst
;
519 qir_emit_nondef(struct vc4_compile
*c
, struct qinst
*inst
)
521 if (inst
->dst
.file
== QFILE_TEMP
)
522 c
->defs
[inst
->dst
.index
] = NULL
;
530 qir_reg_equals(struct qreg a
, struct qreg b
)
532 return a
.file
== b
.file
&& a
.index
== b
.index
&& a
.pack
== b
.pack
;
536 qir_new_block(struct vc4_compile
*c
)
538 struct qblock
*block
= rzalloc(c
, struct qblock
);
540 list_inithead(&block
->instructions
);
541 list_inithead(&block
->qpu_inst_list
);
543 block
->predecessors
= _mesa_set_create(block
,
545 _mesa_key_pointer_equal
);
547 block
->index
= c
->next_block_index
++;
553 qir_set_emit_block(struct vc4_compile
*c
, struct qblock
*block
)
555 c
->cur_block
= block
;
556 list_addtail(&block
->link
, &c
->blocks
);
560 qir_entry_block(struct vc4_compile
*c
)
562 return list_first_entry(&c
->blocks
, struct qblock
, link
);
566 qir_exit_block(struct vc4_compile
*c
)
568 return list_last_entry(&c
->blocks
, struct qblock
, link
);
572 qir_link_blocks(struct qblock
*predecessor
, struct qblock
*successor
)
574 _mesa_set_add(successor
->predecessors
, predecessor
);
575 if (predecessor
->successors
[0]) {
576 assert(!predecessor
->successors
[1]);
577 predecessor
->successors
[1] = successor
;
579 predecessor
->successors
[0] = successor
;
584 qir_compile_init(void)
586 struct vc4_compile
*c
= rzalloc(NULL
, struct vc4_compile
);
588 list_inithead(&c
->blocks
);
589 qir_set_emit_block(c
, qir_new_block(c
));
591 c
->output_position_index
= -1;
592 c
->output_color_index
= -1;
593 c
->output_point_size_index
= -1;
594 c
->output_sample_mask_index
= -1;
596 c
->def_ht
= _mesa_hash_table_create(c
, _mesa_hash_pointer
,
597 _mesa_key_pointer_equal
);
603 qir_remove_instruction(struct vc4_compile
*c
, struct qinst
*qinst
)
605 if (qinst
->dst
.file
== QFILE_TEMP
)
606 c
->defs
[qinst
->dst
.index
] = NULL
;
608 list_del(&qinst
->link
);
614 qir_follow_movs(struct vc4_compile
*c
, struct qreg reg
)
618 while (reg
.file
== QFILE_TEMP
&&
619 c
->defs
[reg
.index
] &&
620 (c
->defs
[reg
.index
]->op
== QOP_MOV
||
621 c
->defs
[reg
.index
]->op
== QOP_FMOV
||
622 c
->defs
[reg
.index
]->op
== QOP_MMOV
)&&
623 !c
->defs
[reg
.index
]->dst
.pack
&&
624 !c
->defs
[reg
.index
]->src
[0].pack
) {
625 reg
= c
->defs
[reg
.index
]->src
[0];
633 qir_compile_destroy(struct vc4_compile
*c
)
635 qir_for_each_block(block
, c
) {
636 while (!list_empty(&block
->instructions
)) {
637 struct qinst
*qinst
=
638 list_first_entry(&block
->instructions
,
640 qir_remove_instruction(c
, qinst
);
648 qir_get_stage_name(enum qstage stage
)
650 static const char *names
[] = {
651 [QSTAGE_FRAG
] = "FS",
652 [QSTAGE_VERT
] = "VS",
653 [QSTAGE_COORD
] = "CS",
660 qir_uniform(struct vc4_compile
*c
,
661 enum quniform_contents contents
,
664 for (int i
= 0; i
< c
->num_uniforms
; i
++) {
665 if (c
->uniform_contents
[i
] == contents
&&
666 c
->uniform_data
[i
] == data
) {
667 return qir_reg(QFILE_UNIF
, i
);
671 uint32_t uniform
= c
->num_uniforms
++;
673 if (uniform
>= c
->uniform_array_size
) {
674 c
->uniform_array_size
= MAX2(MAX2(16, uniform
+ 1),
675 c
->uniform_array_size
* 2);
677 c
->uniform_data
= reralloc(c
, c
->uniform_data
,
679 c
->uniform_array_size
);
680 c
->uniform_contents
= reralloc(c
, c
->uniform_contents
,
681 enum quniform_contents
,
682 c
->uniform_array_size
);
685 c
->uniform_contents
[uniform
] = contents
;
686 c
->uniform_data
[uniform
] = data
;
688 return qir_reg(QFILE_UNIF
, uniform
);
692 qir_SF(struct vc4_compile
*c
, struct qreg src
)
694 struct qinst
*last_inst
= NULL
;
696 if (!list_empty(&c
->cur_block
->instructions
))
697 last_inst
= (struct qinst
*)c
->cur_block
->instructions
.prev
;
699 /* We don't have any way to guess which kind of MOV is implied. */
702 if (src
.file
!= QFILE_TEMP
||
703 !c
->defs
[src
.index
] ||
704 last_inst
!= c
->defs
[src
.index
]) {
705 last_inst
= qir_MOV_dest(c
, qir_reg(QFILE_NULL
, 0), src
);
706 last_inst
= (struct qinst
*)c
->cur_block
->instructions
.prev
;
708 last_inst
->sf
= true;
711 #define OPTPASS(func) \
713 bool stage_progress = func(c); \
714 if (stage_progress) { \
716 if (print_opt_debug) { \
718 "QIR opt pass %2d: %s progress\n", \
726 qir_optimize(struct vc4_compile
*c
)
728 bool print_opt_debug
= false;
732 bool progress
= false;
734 OPTPASS(qir_opt_algebraic
);
735 OPTPASS(qir_opt_constant_folding
);
736 OPTPASS(qir_opt_copy_propagation
);
737 OPTPASS(qir_opt_peephole_sf
);
738 OPTPASS(qir_opt_dead_code
);
739 OPTPASS(qir_opt_small_immediates
);
740 OPTPASS(qir_opt_vpm
);