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_SEL_X_0_NS
] = { "fsel_x_0_ns", 1, 1, false, true },
69 [QOP_SEL_X_0_NC
] = { "fsel_x_0_nc", 1, 1, false, true },
70 [QOP_SEL_X_0_ZS
] = { "fsel_x_0_zs", 1, 1, false, true },
71 [QOP_SEL_X_0_ZC
] = { "fsel_x_0_zc", 1, 1, false, true },
72 [QOP_SEL_X_0_CS
] = { "fsel_x_0_cs", 1, 1, false, true },
73 [QOP_SEL_X_0_CC
] = { "fsel_x_0_cc", 1, 1, false, true },
74 [QOP_SEL_X_Y_NS
] = { "fsel_x_y_ns", 1, 2, false, true },
75 [QOP_SEL_X_Y_NC
] = { "fsel_x_y_nc", 1, 2, false, true },
76 [QOP_SEL_X_Y_ZS
] = { "fsel_x_y_zs", 1, 2, false, true },
77 [QOP_SEL_X_Y_ZC
] = { "fsel_x_y_zc", 1, 2, false, true },
78 [QOP_SEL_X_Y_CS
] = { "fsel_x_y_cs", 1, 2, false, true },
79 [QOP_SEL_X_Y_CC
] = { "fsel_x_y_cc", 1, 2, false, true },
81 [QOP_RCP
] = { "rcp", 1, 1, false, true },
82 [QOP_RSQ
] = { "rsq", 1, 1, false, true },
83 [QOP_EXP2
] = { "exp2", 1, 2, false, true },
84 [QOP_LOG2
] = { "log2", 1, 2, false, true },
85 [QOP_TLB_DISCARD_SETUP
] = { "discard", 0, 1, true },
86 [QOP_TLB_STENCIL_SETUP
] = { "tlb_stencil_setup", 0, 1, true },
87 [QOP_TLB_Z_WRITE
] = { "tlb_z", 0, 1, true },
88 [QOP_TLB_COLOR_WRITE
] = { "tlb_color", 0, 1, true },
89 [QOP_TLB_COLOR_WRITE_MS
] = { "tlb_color_ms", 0, 1, true },
90 [QOP_TLB_COLOR_READ
] = { "tlb_color_read", 1, 0 },
91 [QOP_MS_MASK
] = { "ms_mask", 0, 1, true },
92 [QOP_VARY_ADD_C
] = { "vary_add_c", 1, 1 },
94 [QOP_FRAG_X
] = { "frag_x", 1, 0 },
95 [QOP_FRAG_Y
] = { "frag_y", 1, 0 },
96 [QOP_FRAG_Z
] = { "frag_z", 1, 0 },
97 [QOP_FRAG_W
] = { "frag_w", 1, 0 },
98 [QOP_FRAG_REV_FLAG
] = { "frag_rev_flag", 1, 0 },
100 [QOP_TEX_S
] = { "tex_s", 0, 2 },
101 [QOP_TEX_T
] = { "tex_t", 0, 2 },
102 [QOP_TEX_R
] = { "tex_r", 0, 2 },
103 [QOP_TEX_B
] = { "tex_b", 0, 2 },
104 [QOP_TEX_DIRECT
] = { "tex_direct", 0, 2 },
105 [QOP_TEX_RESULT
] = { "tex_result", 1, 0, true },
109 qir_get_op_name(enum qop qop
)
111 if (qop
< ARRAY_SIZE(qir_op_info
) && qir_op_info
[qop
].name
)
112 return qir_op_info
[qop
].name
;
118 qir_get_op_nsrc(enum qop qop
)
120 if (qop
< ARRAY_SIZE(qir_op_info
) && qir_op_info
[qop
].name
)
121 return qir_op_info
[qop
].nsrc
;
127 * Returns whether the instruction has any side effects that must be
131 qir_has_side_effects(struct vc4_compile
*c
, struct qinst
*inst
)
133 return qir_op_info
[inst
->op
].has_side_effects
;
137 qir_has_side_effect_reads(struct vc4_compile
*c
, struct qinst
*inst
)
139 /* We can dead-code eliminate varyings, because we only tell the VS
140 * about the live ones at the end. But we have to preserve the
141 * point/line coordinates reads, because they're generated by
142 * fixed-function hardware.
144 for (int i
= 0; i
< qir_get_op_nsrc(inst
->op
); i
++) {
145 if (inst
->src
[i
].file
== QFILE_VARY
&&
146 c
->input_slots
[inst
->src
[i
].index
].slot
== 0xff) {
150 if (inst
->src
[i
].file
== QFILE_VPM
)
154 if (inst
->dst
.file
== QFILE_VPM
)
161 qir_is_multi_instruction(struct qinst
*inst
)
163 return qir_op_info
[inst
->op
].multi_instruction
;
167 qir_is_mul(struct qinst
*inst
)
185 qir_is_float_input(struct qinst
*inst
)
204 qir_is_raw_mov(struct qinst
*inst
)
206 return ((inst
->op
== QOP_MOV
||
207 inst
->op
== QOP_FMOV
||
208 inst
->op
== QOP_MMOV
) &&
214 qir_is_tex(struct qinst
*inst
)
216 return inst
->op
>= QOP_TEX_S
&& inst
->op
<= QOP_TEX_DIRECT
;
220 qir_depends_on_flags(struct qinst
*inst
)
242 qir_writes_r4(struct qinst
*inst
)
246 case QOP_TLB_COLOR_READ
:
258 qir_print_reg(struct vc4_compile
*c
, struct qreg reg
, bool write
)
260 static const char *files
[] = {
266 if (reg
.file
== QFILE_NULL
) {
267 fprintf(stderr
, "null");
268 } else if (reg
.file
== QFILE_SMALL_IMM
) {
269 if ((int)reg
.index
>= -16 && (int)reg
.index
<= 15)
270 fprintf(stderr
, "%d", reg
.index
);
272 fprintf(stderr
, "%f", uif(reg
.index
));
273 } else if (reg
.file
== QFILE_VPM
) {
275 fprintf(stderr
, "vpm");
277 fprintf(stderr
, "vpm%d.%d",
278 reg
.index
/ 4, reg
.index
% 4);
281 fprintf(stderr
, "%s%d", files
[reg
.file
], reg
.index
);
284 if (reg
.file
== QFILE_UNIF
&&
285 c
->uniform_contents
[reg
.index
] == QUNIFORM_CONSTANT
) {
286 fprintf(stderr
, " (0x%08x / %f)",
287 c
->uniform_data
[reg
.index
],
288 uif(c
->uniform_data
[reg
.index
]));
293 qir_dump_inst(struct vc4_compile
*c
, struct qinst
*inst
)
295 fprintf(stderr
, "%s%s ",
296 qir_get_op_name(inst
->op
),
297 inst
->sf
? ".sf" : "");
299 qir_print_reg(c
, inst
->dst
, true);
300 if (inst
->dst
.pack
) {
301 if (inst
->dst
.pack
) {
302 if (qir_is_mul(inst
))
303 vc4_qpu_disasm_pack_mul(stderr
, inst
->dst
.pack
);
305 vc4_qpu_disasm_pack_a(stderr
, inst
->dst
.pack
);
308 for (int i
= 0; i
< qir_get_op_nsrc(inst
->op
); i
++) {
309 fprintf(stderr
, ", ");
310 qir_print_reg(c
, inst
->src
[i
], false);
311 vc4_qpu_disasm_unpack(stderr
, inst
->src
[i
].pack
);
316 qir_dump(struct vc4_compile
*c
)
318 list_for_each_entry(struct qinst
, inst
, &c
->instructions
, link
) {
319 qir_dump_inst(c
, inst
);
320 fprintf(stderr
, "\n");
325 qir_get_temp(struct vc4_compile
*c
)
329 reg
.file
= QFILE_TEMP
;
330 reg
.index
= c
->num_temps
++;
333 if (c
->num_temps
> c
->defs_array_size
) {
334 uint32_t old_size
= c
->defs_array_size
;
335 c
->defs_array_size
= MAX2(old_size
* 2, 16);
336 c
->defs
= reralloc(c
, c
->defs
, struct qinst
*,
338 memset(&c
->defs
[old_size
], 0,
339 sizeof(c
->defs
[0]) * (c
->defs_array_size
- old_size
));
346 qir_inst(enum qop op
, struct qreg dst
, struct qreg src0
, struct qreg src1
)
348 struct qinst
*inst
= CALLOC_STRUCT(qinst
);
352 inst
->src
= calloc(2, sizeof(inst
->src
[0]));
360 qir_inst4(enum qop op
, struct qreg dst
,
366 struct qinst
*inst
= CALLOC_STRUCT(qinst
);
370 inst
->src
= calloc(4, sizeof(*inst
->src
));
380 qir_emit(struct vc4_compile
*c
, struct qinst
*inst
)
382 if (inst
->dst
.file
== QFILE_TEMP
)
383 c
->defs
[inst
->dst
.index
] = inst
;
385 qir_emit_nodef(c
, inst
);
389 qir_reg_equals(struct qreg a
, struct qreg b
)
391 return a
.file
== b
.file
&& a
.index
== b
.index
;
395 qir_compile_init(void)
397 struct vc4_compile
*c
= rzalloc(NULL
, struct vc4_compile
);
399 list_inithead(&c
->instructions
);
401 c
->output_position_index
= -1;
402 c
->output_color_index
= -1;
403 c
->output_point_size_index
= -1;
404 c
->output_sample_mask_index
= -1;
406 c
->def_ht
= _mesa_hash_table_create(c
, _mesa_hash_pointer
,
407 _mesa_key_pointer_equal
);
413 qir_remove_instruction(struct vc4_compile
*c
, struct qinst
*qinst
)
415 if (qinst
->dst
.file
== QFILE_TEMP
)
416 c
->defs
[qinst
->dst
.index
] = NULL
;
418 list_del(&qinst
->link
);
424 qir_follow_movs(struct vc4_compile
*c
, struct qreg reg
)
428 while (reg
.file
== QFILE_TEMP
&&
429 c
->defs
[reg
.index
] &&
430 (c
->defs
[reg
.index
]->op
== QOP_MOV
||
431 c
->defs
[reg
.index
]->op
== QOP_FMOV
||
432 c
->defs
[reg
.index
]->op
== QOP_MMOV
)&&
433 !c
->defs
[reg
.index
]->dst
.pack
&&
434 !c
->defs
[reg
.index
]->src
[0].pack
) {
435 reg
= c
->defs
[reg
.index
]->src
[0];
443 qir_compile_destroy(struct vc4_compile
*c
)
445 while (!list_empty(&c
->instructions
)) {
446 struct qinst
*qinst
=
447 (struct qinst
*)c
->instructions
.next
;
448 qir_remove_instruction(c
, qinst
);
455 qir_get_stage_name(enum qstage stage
)
457 static const char *names
[] = {
458 [QSTAGE_FRAG
] = "FS",
459 [QSTAGE_VERT
] = "VS",
460 [QSTAGE_COORD
] = "CS",
467 qir_uniform(struct vc4_compile
*c
,
468 enum quniform_contents contents
,
471 for (int i
= 0; i
< c
->num_uniforms
; i
++) {
472 if (c
->uniform_contents
[i
] == contents
&&
473 c
->uniform_data
[i
] == data
) {
474 return (struct qreg
) { QFILE_UNIF
, i
};
478 uint32_t uniform
= c
->num_uniforms
++;
479 struct qreg u
= { QFILE_UNIF
, uniform
};
481 if (uniform
>= c
->uniform_array_size
) {
482 c
->uniform_array_size
= MAX2(MAX2(16, uniform
+ 1),
483 c
->uniform_array_size
* 2);
485 c
->uniform_data
= reralloc(c
, c
->uniform_data
,
487 c
->uniform_array_size
);
488 c
->uniform_contents
= reralloc(c
, c
->uniform_contents
,
489 enum quniform_contents
,
490 c
->uniform_array_size
);
493 c
->uniform_contents
[uniform
] = contents
;
494 c
->uniform_data
[uniform
] = data
;
500 qir_SF(struct vc4_compile
*c
, struct qreg src
)
502 struct qinst
*last_inst
= NULL
;
503 if (!list_empty(&c
->instructions
))
504 last_inst
= (struct qinst
*)c
->instructions
.prev
;
507 last_inst
->dst
.file
!= src
.file
||
508 last_inst
->dst
.index
!= src
.index
||
509 qir_is_multi_instruction(last_inst
)) {
510 src
= qir_MOV(c
, src
);
511 last_inst
= (struct qinst
*)c
->instructions
.prev
;
513 last_inst
->sf
= true;
516 #define OPTPASS(func) \
518 bool stage_progress = func(c); \
519 if (stage_progress) { \
521 if (print_opt_debug) { \
523 "QIR opt pass %2d: %s progress\n", \
530 qir_optimize(struct vc4_compile
*c
)
532 bool print_opt_debug
= false;
536 bool progress
= false;
538 OPTPASS(qir_opt_algebraic
);
539 OPTPASS(qir_opt_cse
);
540 OPTPASS(qir_opt_constant_folding
);
541 OPTPASS(qir_opt_copy_propagation
);
542 OPTPASS(qir_opt_dead_code
);
543 OPTPASS(qir_opt_small_immediates
);
544 OPTPASS(qir_opt_vpm_writes
);