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
;
34 bool multi_instruction
;
37 static const struct qir_op_info qir_op_info
[] = {
38 [QOP_MOV
] = { "mov", 1, 1 },
39 [QOP_FMOV
] = { "fmov", 1, 1 },
40 [QOP_MMOV
] = { "mmov", 1, 1 },
41 [QOP_FADD
] = { "fadd", 1, 2 },
42 [QOP_FSUB
] = { "fsub", 1, 2 },
43 [QOP_FMUL
] = { "fmul", 1, 2 },
44 [QOP_MUL24
] = { "mul24", 1, 2 },
45 [QOP_V8MULD
] = {"v8muld", 1, 2 },
46 [QOP_V8MIN
] = {"v8min", 1, 2 },
47 [QOP_V8MAX
] = {"v8max", 1, 2 },
48 [QOP_V8ADDS
] = {"v8adds", 1, 2 },
49 [QOP_V8SUBS
] = {"v8subs", 1, 2 },
50 [QOP_FMIN
] = { "fmin", 1, 2 },
51 [QOP_FMAX
] = { "fmax", 1, 2 },
52 [QOP_FMINABS
] = { "fminabs", 1, 2 },
53 [QOP_FMAXABS
] = { "fmaxabs", 1, 2 },
54 [QOP_FTOI
] = { "ftoi", 1, 1 },
55 [QOP_ITOF
] = { "itof", 1, 1 },
56 [QOP_ADD
] = { "add", 1, 2 },
57 [QOP_SUB
] = { "sub", 1, 2 },
58 [QOP_SHR
] = { "shr", 1, 2 },
59 [QOP_ASR
] = { "asr", 1, 2 },
60 [QOP_SHL
] = { "shl", 1, 2 },
61 [QOP_MIN
] = { "min", 1, 2 },
62 [QOP_MAX
] = { "max", 1, 2 },
63 [QOP_AND
] = { "and", 1, 2 },
64 [QOP_OR
] = { "or", 1, 2 },
65 [QOP_XOR
] = { "xor", 1, 2 },
66 [QOP_NOT
] = { "not", 1, 1 },
68 [QOP_RCP
] = { "rcp", 1, 1, false, true },
69 [QOP_RSQ
] = { "rsq", 1, 1, false, true },
70 [QOP_EXP2
] = { "exp2", 1, 2, false, true },
71 [QOP_LOG2
] = { "log2", 1, 2, false, true },
72 [QOP_TLB_DISCARD_SETUP
] = { "discard", 0, 1, true },
73 [QOP_TLB_STENCIL_SETUP
] = { "tlb_stencil_setup", 0, 1, true },
74 [QOP_TLB_Z_WRITE
] = { "tlb_z", 0, 1, true },
75 [QOP_TLB_COLOR_WRITE
] = { "tlb_color", 0, 1, true },
76 [QOP_TLB_COLOR_WRITE_MS
] = { "tlb_color_ms", 0, 1, true },
77 [QOP_TLB_COLOR_READ
] = { "tlb_color_read", 1, 0 },
78 [QOP_MS_MASK
] = { "ms_mask", 0, 1, true },
79 [QOP_VARY_ADD_C
] = { "vary_add_c", 1, 1 },
81 [QOP_FRAG_X
] = { "frag_x", 1, 0 },
82 [QOP_FRAG_Y
] = { "frag_y", 1, 0 },
83 [QOP_FRAG_Z
] = { "frag_z", 1, 0 },
84 [QOP_FRAG_W
] = { "frag_w", 1, 0 },
85 [QOP_FRAG_REV_FLAG
] = { "frag_rev_flag", 1, 0 },
87 [QOP_TEX_S
] = { "tex_s", 0, 2 },
88 [QOP_TEX_T
] = { "tex_t", 0, 2 },
89 [QOP_TEX_R
] = { "tex_r", 0, 2 },
90 [QOP_TEX_B
] = { "tex_b", 0, 2 },
91 [QOP_TEX_DIRECT
] = { "tex_direct", 0, 2 },
92 [QOP_TEX_RESULT
] = { "tex_result", 1, 0, true },
96 qir_get_op_name(enum qop qop
)
98 if (qop
< ARRAY_SIZE(qir_op_info
) && qir_op_info
[qop
].name
)
99 return qir_op_info
[qop
].name
;
105 qir_get_op_nsrc(enum qop qop
)
107 if (qop
< ARRAY_SIZE(qir_op_info
) && qir_op_info
[qop
].name
)
108 return qir_op_info
[qop
].nsrc
;
114 * Returns whether the instruction has any side effects that must be
118 qir_has_side_effects(struct vc4_compile
*c
, struct qinst
*inst
)
120 return qir_op_info
[inst
->op
].has_side_effects
;
124 qir_has_side_effect_reads(struct vc4_compile
*c
, struct qinst
*inst
)
126 /* We can dead-code eliminate varyings, because we only tell the VS
127 * about the live ones at the end. But we have to preserve the
128 * point/line coordinates reads, because they're generated by
129 * fixed-function hardware.
131 for (int i
= 0; i
< qir_get_op_nsrc(inst
->op
); i
++) {
132 if (inst
->src
[i
].file
== QFILE_VARY
&&
133 c
->input_slots
[inst
->src
[i
].index
].slot
== 0xff) {
137 if (inst
->src
[i
].file
== QFILE_VPM
)
141 if (inst
->dst
.file
== QFILE_VPM
)
148 qir_is_multi_instruction(struct qinst
*inst
)
150 return qir_op_info
[inst
->op
].multi_instruction
;
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
) &&
201 qir_is_tex(struct qinst
*inst
)
203 return inst
->op
>= QOP_TEX_S
&& inst
->op
<= QOP_TEX_DIRECT
;
207 qir_depends_on_flags(struct qinst
*inst
)
209 return (inst
->cond
!= QPU_COND_ALWAYS
&&
210 inst
->cond
!= QPU_COND_NEVER
);
214 qir_writes_r4(struct qinst
*inst
)
218 case QOP_TLB_COLOR_READ
:
230 qir_print_reg(struct vc4_compile
*c
, struct qreg reg
, bool write
)
232 static const char *files
[] = {
238 if (reg
.file
== QFILE_NULL
) {
239 fprintf(stderr
, "null");
240 } else if (reg
.file
== QFILE_SMALL_IMM
) {
241 if ((int)reg
.index
>= -16 && (int)reg
.index
<= 15)
242 fprintf(stderr
, "%d", reg
.index
);
244 fprintf(stderr
, "%f", uif(reg
.index
));
245 } else if (reg
.file
== QFILE_VPM
) {
247 fprintf(stderr
, "vpm");
249 fprintf(stderr
, "vpm%d.%d",
250 reg
.index
/ 4, reg
.index
% 4);
253 fprintf(stderr
, "%s%d", files
[reg
.file
], reg
.index
);
256 if (reg
.file
== QFILE_UNIF
&&
257 c
->uniform_contents
[reg
.index
] == QUNIFORM_CONSTANT
) {
258 fprintf(stderr
, " (0x%08x / %f)",
259 c
->uniform_data
[reg
.index
],
260 uif(c
->uniform_data
[reg
.index
]));
265 qir_dump_inst(struct vc4_compile
*c
, struct qinst
*inst
)
267 static const char *conditions
[] = {
268 [QPU_COND_ALWAYS
] = "",
269 [QPU_COND_NEVER
] = ".never",
270 [QPU_COND_ZS
] = ".zs",
271 [QPU_COND_ZC
] = ".zc",
272 [QPU_COND_NS
] = ".ns",
273 [QPU_COND_NC
] = ".nc",
274 [QPU_COND_CS
] = ".cs",
275 [QPU_COND_CC
] = ".cc",
277 fprintf(stderr
, "%s%s%s ",
278 qir_get_op_name(inst
->op
),
279 conditions
[inst
->cond
],
280 inst
->sf
? ".sf" : "");
282 qir_print_reg(c
, inst
->dst
, true);
283 if (inst
->dst
.pack
) {
284 if (inst
->dst
.pack
) {
285 if (qir_is_mul(inst
))
286 vc4_qpu_disasm_pack_mul(stderr
, inst
->dst
.pack
);
288 vc4_qpu_disasm_pack_a(stderr
, inst
->dst
.pack
);
291 for (int i
= 0; i
< qir_get_op_nsrc(inst
->op
); i
++) {
292 fprintf(stderr
, ", ");
293 qir_print_reg(c
, inst
->src
[i
], false);
294 vc4_qpu_disasm_unpack(stderr
, inst
->src
[i
].pack
);
299 qir_dump(struct vc4_compile
*c
)
301 list_for_each_entry(struct qinst
, inst
, &c
->instructions
, link
) {
302 qir_dump_inst(c
, inst
);
303 fprintf(stderr
, "\n");
308 qir_get_temp(struct vc4_compile
*c
)
312 reg
.file
= QFILE_TEMP
;
313 reg
.index
= c
->num_temps
++;
316 if (c
->num_temps
> c
->defs_array_size
) {
317 uint32_t old_size
= c
->defs_array_size
;
318 c
->defs_array_size
= MAX2(old_size
* 2, 16);
319 c
->defs
= reralloc(c
, c
->defs
, struct qinst
*,
321 memset(&c
->defs
[old_size
], 0,
322 sizeof(c
->defs
[0]) * (c
->defs_array_size
- old_size
));
329 qir_inst(enum qop op
, struct qreg dst
, struct qreg src0
, struct qreg src1
)
331 struct qinst
*inst
= CALLOC_STRUCT(qinst
);
335 inst
->src
= calloc(2, sizeof(inst
->src
[0]));
338 inst
->cond
= QPU_COND_ALWAYS
;
344 qir_inst4(enum qop op
, struct qreg dst
,
350 struct qinst
*inst
= CALLOC_STRUCT(qinst
);
354 inst
->src
= calloc(4, sizeof(*inst
->src
));
364 qir_emit(struct vc4_compile
*c
, struct qinst
*inst
)
366 if (inst
->dst
.file
== QFILE_TEMP
)
367 c
->defs
[inst
->dst
.index
] = inst
;
369 qir_emit_nodef(c
, inst
);
373 qir_reg_equals(struct qreg a
, struct qreg b
)
375 return a
.file
== b
.file
&& a
.index
== b
.index
;
379 qir_compile_init(void)
381 struct vc4_compile
*c
= rzalloc(NULL
, struct vc4_compile
);
383 list_inithead(&c
->instructions
);
385 c
->output_position_index
= -1;
386 c
->output_color_index
= -1;
387 c
->output_point_size_index
= -1;
388 c
->output_sample_mask_index
= -1;
390 c
->def_ht
= _mesa_hash_table_create(c
, _mesa_hash_pointer
,
391 _mesa_key_pointer_equal
);
397 qir_remove_instruction(struct vc4_compile
*c
, struct qinst
*qinst
)
399 if (qinst
->dst
.file
== QFILE_TEMP
)
400 c
->defs
[qinst
->dst
.index
] = NULL
;
402 list_del(&qinst
->link
);
408 qir_follow_movs(struct vc4_compile
*c
, struct qreg reg
)
412 while (reg
.file
== QFILE_TEMP
&&
413 c
->defs
[reg
.index
] &&
414 (c
->defs
[reg
.index
]->op
== QOP_MOV
||
415 c
->defs
[reg
.index
]->op
== QOP_FMOV
||
416 c
->defs
[reg
.index
]->op
== QOP_MMOV
)&&
417 !c
->defs
[reg
.index
]->dst
.pack
&&
418 !c
->defs
[reg
.index
]->src
[0].pack
) {
419 reg
= c
->defs
[reg
.index
]->src
[0];
427 qir_compile_destroy(struct vc4_compile
*c
)
429 while (!list_empty(&c
->instructions
)) {
430 struct qinst
*qinst
=
431 (struct qinst
*)c
->instructions
.next
;
432 qir_remove_instruction(c
, qinst
);
439 qir_get_stage_name(enum qstage stage
)
441 static const char *names
[] = {
442 [QSTAGE_FRAG
] = "FS",
443 [QSTAGE_VERT
] = "VS",
444 [QSTAGE_COORD
] = "CS",
451 qir_uniform(struct vc4_compile
*c
,
452 enum quniform_contents contents
,
455 for (int i
= 0; i
< c
->num_uniforms
; i
++) {
456 if (c
->uniform_contents
[i
] == contents
&&
457 c
->uniform_data
[i
] == data
) {
458 return (struct qreg
) { QFILE_UNIF
, i
};
462 uint32_t uniform
= c
->num_uniforms
++;
463 struct qreg u
= { QFILE_UNIF
, uniform
};
465 if (uniform
>= c
->uniform_array_size
) {
466 c
->uniform_array_size
= MAX2(MAX2(16, uniform
+ 1),
467 c
->uniform_array_size
* 2);
469 c
->uniform_data
= reralloc(c
, c
->uniform_data
,
471 c
->uniform_array_size
);
472 c
->uniform_contents
= reralloc(c
, c
->uniform_contents
,
473 enum quniform_contents
,
474 c
->uniform_array_size
);
477 c
->uniform_contents
[uniform
] = contents
;
478 c
->uniform_data
[uniform
] = data
;
484 qir_SF(struct vc4_compile
*c
, struct qreg src
)
486 struct qinst
*last_inst
= NULL
;
487 if (!list_empty(&c
->instructions
))
488 last_inst
= (struct qinst
*)c
->instructions
.prev
;
490 if (src
.file
!= QFILE_TEMP
||
491 !c
->defs
[src
.index
] ||
492 last_inst
!= c
->defs
[src
.index
] ||
493 qir_is_multi_instruction(last_inst
)) {
494 src
= qir_MOV(c
, src
);
495 last_inst
= (struct qinst
*)c
->instructions
.prev
;
497 last_inst
->sf
= true;
500 #define OPTPASS(func) \
502 bool stage_progress = func(c); \
503 if (stage_progress) { \
505 if (print_opt_debug) { \
507 "QIR opt pass %2d: %s progress\n", \
514 qir_optimize(struct vc4_compile
*c
)
516 bool print_opt_debug
= false;
520 bool progress
= false;
522 OPTPASS(qir_opt_algebraic
);
523 OPTPASS(qir_opt_cse
);
524 OPTPASS(qir_opt_constant_folding
);
525 OPTPASS(qir_opt_copy_propagation
);
526 OPTPASS(qir_opt_dead_code
);
527 OPTPASS(qir_opt_small_immediates
);
528 OPTPASS(qir_opt_vpm_writes
);