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_THRSW
] = { "thrsw", 0, 0, true },
87 [QOP_LOAD_IMM
] = { "load_imm", 0, 1 },
88 [QOP_LOAD_IMM_U2
] = { "load_imm_u2", 0, 1 },
89 [QOP_LOAD_IMM_I2
] = { "load_imm_i2", 0, 1 },
91 [QOP_ROT_MUL
] = { "rot_mul", 0, 2 },
93 [QOP_BRANCH
] = { "branch", 0, 0, true },
94 [QOP_UNIFORMS_RESET
] = { "uniforms_reset", 0, 2, true },
98 qir_get_op_name(enum qop qop
)
100 if (qop
< ARRAY_SIZE(qir_op_info
) && qir_op_info
[qop
].name
)
101 return qir_op_info
[qop
].name
;
107 qir_get_op_nsrc(enum qop qop
)
109 if (qop
< ARRAY_SIZE(qir_op_info
) && qir_op_info
[qop
].name
)
110 return qir_op_info
[qop
].nsrc
;
116 * Returns whether the instruction has any side effects that must be
120 qir_has_side_effects(struct vc4_compile
*c
, struct qinst
*inst
)
122 switch (inst
->dst
.file
) {
123 case QFILE_TLB_Z_WRITE
:
124 case QFILE_TLB_COLOR_WRITE
:
125 case QFILE_TLB_COLOR_WRITE_MS
:
126 case QFILE_TLB_STENCIL_SETUP
:
132 return qir_op_info
[inst
->op
].has_side_effects
;
136 qir_has_side_effect_reads(struct vc4_compile
*c
, struct qinst
*inst
)
138 /* We can dead-code eliminate varyings, because we only tell the VS
139 * about the live ones at the end. But we have to preserve the
140 * point/line coordinates reads, because they're generated by
141 * fixed-function hardware.
143 for (int i
= 0; i
< qir_get_op_nsrc(inst
->op
); i
++) {
144 if (inst
->src
[i
].file
== QFILE_VARY
&&
145 c
->input_slots
[inst
->src
[i
].index
].slot
== 0xff) {
149 if (inst
->src
[i
].file
== QFILE_VPM
)
153 if (inst
->dst
.file
== QFILE_VPM
)
160 qir_is_mul(struct qinst
*inst
)
179 qir_is_float_input(struct qinst
*inst
)
198 qir_is_raw_mov(struct qinst
*inst
)
200 return ((inst
->op
== QOP_MOV
||
201 inst
->op
== QOP_FMOV
||
202 inst
->op
== QOP_MMOV
) &&
203 inst
->cond
== QPU_COND_ALWAYS
&&
209 qir_is_tex(struct qinst
*inst
)
211 return inst
->op
>= QOP_TEX_S
&& inst
->op
<= QOP_TEX_DIRECT
;
215 qir_depends_on_flags(struct qinst
*inst
)
217 if (inst
->op
== QOP_BRANCH
) {
218 return inst
->cond
!= QPU_COND_BRANCH_ALWAYS
;
220 return (inst
->cond
!= QPU_COND_ALWAYS
&&
221 inst
->cond
!= QPU_COND_NEVER
);
226 qir_writes_r4(struct qinst
*inst
)
230 case QOP_TLB_COLOR_READ
:
242 qir_channels_written(struct qinst
*inst
)
244 if (qir_is_mul(inst
)) {
245 switch (inst
->dst
.pack
) {
246 case QPU_PACK_MUL_NOP
:
247 case QPU_PACK_MUL_8888
:
249 case QPU_PACK_MUL_8A
:
251 case QPU_PACK_MUL_8B
:
253 case QPU_PACK_MUL_8C
:
255 case QPU_PACK_MUL_8D
:
259 switch (inst
->dst
.pack
) {
261 case QPU_PACK_A_8888
:
262 case QPU_PACK_A_8888_SAT
:
263 case QPU_PACK_A_32_SAT
:
266 case QPU_PACK_A_8A_SAT
:
269 case QPU_PACK_A_8B_SAT
:
272 case QPU_PACK_A_8C_SAT
:
275 case QPU_PACK_A_8D_SAT
:
278 case QPU_PACK_A_16A_SAT
:
281 case QPU_PACK_A_16B_SAT
:
285 unreachable("Bad pack field");
289 qir_print_reg(struct vc4_compile
*c
, struct qreg reg
, bool write
)
291 static const char *files
[] = {
295 [QFILE_TLB_COLOR_WRITE
] = "tlb_c",
296 [QFILE_TLB_COLOR_WRITE_MS
] = "tlb_c_ms",
297 [QFILE_TLB_Z_WRITE
] = "tlb_z",
298 [QFILE_TLB_STENCIL_SETUP
] = "tlb_stencil",
299 [QFILE_FRAG_X
] = "frag_x",
300 [QFILE_FRAG_Y
] = "frag_y",
301 [QFILE_FRAG_REV_FLAG
] = "frag_rev_flag",
302 [QFILE_QPU_ELEMENT
] = "elem",
308 fprintf(stderr
, "null");
312 fprintf(stderr
, "0x%08x (%f)", reg
.index
, uif(reg
.index
));
315 case QFILE_SMALL_IMM
:
316 if ((int)reg
.index
>= -16 && (int)reg
.index
<= 15)
317 fprintf(stderr
, "%d", reg
.index
);
319 fprintf(stderr
, "%f", uif(reg
.index
));
324 fprintf(stderr
, "vpm");
326 fprintf(stderr
, "vpm%d.%d",
327 reg
.index
/ 4, reg
.index
% 4);
331 case QFILE_TLB_COLOR_WRITE
:
332 case QFILE_TLB_COLOR_WRITE_MS
:
333 case QFILE_TLB_Z_WRITE
:
334 case QFILE_TLB_STENCIL_SETUP
:
335 fprintf(stderr
, "%s", files
[reg
.file
]);
339 fprintf(stderr
, "%s%d", files
[reg
.file
], reg
.index
);
343 if (reg
.file
== QFILE_UNIF
&&
344 c
->uniform_contents
[reg
.index
] == QUNIFORM_CONSTANT
) {
345 fprintf(stderr
, " (0x%08x / %f)",
346 c
->uniform_data
[reg
.index
],
347 uif(c
->uniform_data
[reg
.index
]));
352 qir_dump_inst(struct vc4_compile
*c
, struct qinst
*inst
)
354 fprintf(stderr
, "%s", qir_get_op_name(inst
->op
));
355 if (inst
->op
== QOP_BRANCH
)
356 vc4_qpu_disasm_cond_branch(stderr
, inst
->cond
);
358 vc4_qpu_disasm_cond(stderr
, inst
->cond
);
360 fprintf(stderr
, ".sf");
361 fprintf(stderr
, " ");
363 if (inst
->op
!= QOP_BRANCH
) {
364 qir_print_reg(c
, inst
->dst
, true);
365 if (inst
->dst
.pack
) {
366 if (inst
->dst
.pack
) {
367 if (qir_is_mul(inst
))
368 vc4_qpu_disasm_pack_mul(stderr
, inst
->dst
.pack
);
370 vc4_qpu_disasm_pack_a(stderr
, inst
->dst
.pack
);
375 for (int i
= 0; i
< qir_get_op_nsrc(inst
->op
); i
++) {
376 fprintf(stderr
, ", ");
377 qir_print_reg(c
, inst
->src
[i
], false);
378 vc4_qpu_disasm_unpack(stderr
, inst
->src
[i
].pack
);
383 qir_dump(struct vc4_compile
*c
)
388 qir_for_each_block(block
, c
) {
389 fprintf(stderr
, "BLOCK %d:\n", block
->index
);
390 qir_for_each_inst(inst
, block
) {
394 fprintf(stderr
, "%3d ", pressure
);
396 for (int i
= 0; i
< c
->num_temps
; i
++) {
397 if (c
->temp_start
[i
] != ip
)
403 fprintf(stderr
, ", ");
405 fprintf(stderr
, "S%4d", i
);
410 fprintf(stderr
, " ");
412 fprintf(stderr
, " ");
418 for (int i
= 0; i
< c
->num_temps
; i
++) {
419 if (c
->temp_end
[i
] != ip
)
425 fprintf(stderr
, ", ");
427 fprintf(stderr
, "E%4d", i
);
432 fprintf(stderr
, " ");
434 fprintf(stderr
, " ");
437 qir_dump_inst(c
, inst
);
438 fprintf(stderr
, "\n");
441 if (block
->successors
[1]) {
442 fprintf(stderr
, "-> BLOCK %d, %d\n",
443 block
->successors
[0]->index
,
444 block
->successors
[1]->index
);
445 } else if (block
->successors
[0]) {
446 fprintf(stderr
, "-> BLOCK %d\n",
447 block
->successors
[0]->index
);
453 qir_get_temp(struct vc4_compile
*c
)
457 reg
.file
= QFILE_TEMP
;
458 reg
.index
= c
->num_temps
++;
461 if (c
->num_temps
> c
->defs_array_size
) {
462 uint32_t old_size
= c
->defs_array_size
;
463 c
->defs_array_size
= MAX2(old_size
* 2, 16);
464 c
->defs
= reralloc(c
, c
->defs
, struct qinst
*,
466 memset(&c
->defs
[old_size
], 0,
467 sizeof(c
->defs
[0]) * (c
->defs_array_size
- old_size
));
474 qir_inst(enum qop op
, struct qreg dst
, struct qreg src0
, struct qreg src1
)
476 struct qinst
*inst
= CALLOC_STRUCT(qinst
);
482 inst
->cond
= QPU_COND_ALWAYS
;
488 qir_emit(struct vc4_compile
*c
, struct qinst
*inst
)
490 list_addtail(&inst
->link
, &c
->cur_block
->instructions
);
493 /* Updates inst to write to a new temporary, emits it, and notes the def. */
495 qir_emit_def(struct vc4_compile
*c
, struct qinst
*inst
)
497 assert(inst
->dst
.file
== QFILE_NULL
);
499 inst
->dst
= qir_get_temp(c
);
501 if (inst
->dst
.file
== QFILE_TEMP
)
502 c
->defs
[inst
->dst
.index
] = inst
;
510 qir_emit_nondef(struct vc4_compile
*c
, struct qinst
*inst
)
512 if (inst
->dst
.file
== QFILE_TEMP
)
513 c
->defs
[inst
->dst
.index
] = NULL
;
521 qir_reg_equals(struct qreg a
, struct qreg b
)
523 return a
.file
== b
.file
&& a
.index
== b
.index
&& a
.pack
== b
.pack
;
527 qir_new_block(struct vc4_compile
*c
)
529 struct qblock
*block
= rzalloc(c
, struct qblock
);
531 list_inithead(&block
->instructions
);
532 list_inithead(&block
->qpu_inst_list
);
534 block
->predecessors
= _mesa_set_create(block
,
536 _mesa_key_pointer_equal
);
538 block
->index
= c
->next_block_index
++;
544 qir_set_emit_block(struct vc4_compile
*c
, struct qblock
*block
)
546 c
->cur_block
= block
;
547 list_addtail(&block
->link
, &c
->blocks
);
551 qir_entry_block(struct vc4_compile
*c
)
553 return list_first_entry(&c
->blocks
, struct qblock
, link
);
557 qir_exit_block(struct vc4_compile
*c
)
559 return list_last_entry(&c
->blocks
, struct qblock
, link
);
563 qir_link_blocks(struct qblock
*predecessor
, struct qblock
*successor
)
565 _mesa_set_add(successor
->predecessors
, predecessor
);
566 if (predecessor
->successors
[0]) {
567 assert(!predecessor
->successors
[1]);
568 predecessor
->successors
[1] = successor
;
570 predecessor
->successors
[0] = successor
;
575 qir_compile_init(void)
577 struct vc4_compile
*c
= rzalloc(NULL
, struct vc4_compile
);
579 list_inithead(&c
->blocks
);
580 qir_set_emit_block(c
, qir_new_block(c
));
582 c
->output_position_index
= -1;
583 c
->output_color_index
= -1;
584 c
->output_point_size_index
= -1;
585 c
->output_sample_mask_index
= -1;
587 c
->def_ht
= _mesa_hash_table_create(c
, _mesa_hash_pointer
,
588 _mesa_key_pointer_equal
);
594 qir_remove_instruction(struct vc4_compile
*c
, struct qinst
*qinst
)
596 if (qinst
->dst
.file
== QFILE_TEMP
)
597 c
->defs
[qinst
->dst
.index
] = NULL
;
599 list_del(&qinst
->link
);
604 qir_follow_movs(struct vc4_compile
*c
, struct qreg reg
)
608 while (reg
.file
== QFILE_TEMP
&&
609 c
->defs
[reg
.index
] &&
610 (c
->defs
[reg
.index
]->op
== QOP_MOV
||
611 c
->defs
[reg
.index
]->op
== QOP_FMOV
||
612 c
->defs
[reg
.index
]->op
== QOP_MMOV
)&&
613 !c
->defs
[reg
.index
]->dst
.pack
&&
614 !c
->defs
[reg
.index
]->src
[0].pack
) {
615 reg
= c
->defs
[reg
.index
]->src
[0];
623 qir_compile_destroy(struct vc4_compile
*c
)
625 qir_for_each_block(block
, c
) {
626 while (!list_empty(&block
->instructions
)) {
627 struct qinst
*qinst
=
628 list_first_entry(&block
->instructions
,
630 qir_remove_instruction(c
, qinst
);
638 qir_get_stage_name(enum qstage stage
)
640 static const char *names
[] = {
641 [QSTAGE_FRAG
] = "FS",
642 [QSTAGE_VERT
] = "VS",
643 [QSTAGE_COORD
] = "CS",
650 qir_uniform(struct vc4_compile
*c
,
651 enum quniform_contents contents
,
654 for (int i
= 0; i
< c
->num_uniforms
; i
++) {
655 if (c
->uniform_contents
[i
] == contents
&&
656 c
->uniform_data
[i
] == data
) {
657 return qir_reg(QFILE_UNIF
, i
);
661 uint32_t uniform
= c
->num_uniforms
++;
663 if (uniform
>= c
->uniform_array_size
) {
664 c
->uniform_array_size
= MAX2(MAX2(16, uniform
+ 1),
665 c
->uniform_array_size
* 2);
667 c
->uniform_data
= reralloc(c
, c
->uniform_data
,
669 c
->uniform_array_size
);
670 c
->uniform_contents
= reralloc(c
, c
->uniform_contents
,
671 enum quniform_contents
,
672 c
->uniform_array_size
);
675 c
->uniform_contents
[uniform
] = contents
;
676 c
->uniform_data
[uniform
] = data
;
678 return qir_reg(QFILE_UNIF
, uniform
);
682 qir_SF(struct vc4_compile
*c
, struct qreg src
)
684 struct qinst
*last_inst
= NULL
;
686 if (!list_empty(&c
->cur_block
->instructions
))
687 last_inst
= (struct qinst
*)c
->cur_block
->instructions
.prev
;
689 /* We don't have any way to guess which kind of MOV is implied. */
692 if (src
.file
!= QFILE_TEMP
||
693 !c
->defs
[src
.index
] ||
694 last_inst
!= c
->defs
[src
.index
]) {
695 last_inst
= qir_MOV_dest(c
, qir_reg(QFILE_NULL
, 0), src
);
696 last_inst
= (struct qinst
*)c
->cur_block
->instructions
.prev
;
698 last_inst
->sf
= true;
701 #define OPTPASS(func) \
703 bool stage_progress = func(c); \
704 if (stage_progress) { \
706 if (print_opt_debug) { \
708 "QIR opt pass %2d: %s progress\n", \
716 qir_optimize(struct vc4_compile
*c
)
718 bool print_opt_debug
= false;
722 bool progress
= false;
724 OPTPASS(qir_opt_algebraic
);
725 OPTPASS(qir_opt_constant_folding
);
726 OPTPASS(qir_opt_copy_propagation
);
727 OPTPASS(qir_opt_peephole_sf
);
728 OPTPASS(qir_opt_dead_code
);
729 OPTPASS(qir_opt_small_immediates
);
730 OPTPASS(qir_opt_vpm
);