2 * Copyright © 2018 Valve Corporation
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
33 void perfwarn(bool cond
, const char *msg
, Instruction
*instr
)
36 fprintf(stderr
, "ACO performance warning: %s\n", msg
);
38 fprintf(stderr
, "instruction: ");
39 aco_print_instr(instr
, stderr
);
40 fprintf(stderr
, "\n");
43 if (debug_flags
& DEBUG_PERFWARN
)
49 void validate(Program
* program
, FILE * output
)
51 if (!(debug_flags
& DEBUG_VALIDATE
))
55 auto check
= [&output
, &is_valid
](bool check
, const char * msg
, aco::Instruction
* instr
) -> void {
57 fprintf(output
, "%s: ", msg
);
58 aco_print_instr(instr
, output
);
59 fprintf(output
, "\n");
63 auto check_block
= [&output
, &is_valid
](bool check
, const char * msg
, aco::Block
* block
) -> void {
65 fprintf(output
, "%s: BB%u\n", msg
, block
->index
);
70 for (Block
& block
: program
->blocks
) {
71 for (aco_ptr
<Instruction
>& instr
: block
.instructions
) {
73 /* check base format */
74 Format base_format
= instr
->format
;
75 base_format
= (Format
)((uint32_t)base_format
& ~(uint32_t)Format::SDWA
);
76 base_format
= (Format
)((uint32_t)base_format
& ~(uint32_t)Format::DPP
);
77 if ((uint32_t)base_format
& (uint32_t)Format::VOP1
)
78 base_format
= Format::VOP1
;
79 else if ((uint32_t)base_format
& (uint32_t)Format::VOP2
)
80 base_format
= Format::VOP2
;
81 else if ((uint32_t)base_format
& (uint32_t)Format::VOPC
)
82 base_format
= Format::VOPC
;
83 else if ((uint32_t)base_format
& (uint32_t)Format::VINTRP
)
84 base_format
= Format::VINTRP
;
85 check(base_format
== instr_info
.format
[(int)instr
->opcode
], "Wrong base format for instruction", instr
.get());
87 /* check VOP3 modifiers */
88 if (((uint32_t)instr
->format
& (uint32_t)Format::VOP3
) && instr
->format
!= Format::VOP3
) {
89 check(base_format
== Format::VOP2
||
90 base_format
== Format::VOP1
||
91 base_format
== Format::VOPC
||
92 base_format
== Format::VINTRP
,
93 "Format cannot have VOP3A/VOP3B applied", instr
.get());
96 /* check for undefs */
97 for (unsigned i
= 0; i
< instr
->operands
.size(); i
++) {
98 if (instr
->operands
[i
].isUndefined()) {
99 bool flat
= instr
->format
== Format::FLAT
|| instr
->format
== Format::SCRATCH
|| instr
->format
== Format::GLOBAL
;
100 bool can_be_undef
= is_phi(instr
) || instr
->format
== Format::EXP
||
101 instr
->format
== Format::PSEUDO_REDUCTION
||
102 (flat
&& i
== 1) || (instr
->format
== Format::MIMG
&& i
== 2) ||
103 ((instr
->format
== Format::MUBUF
|| instr
->format
== Format::MTBUF
) && i
== 0);
104 check(can_be_undef
, "Undefs can only be used in certain operands", instr
.get());
108 /* check num literals */
109 if (instr
->isSALU() || instr
->isVALU()) {
110 unsigned num_literals
= 0;
111 for (unsigned i
= 0; i
< instr
->operands
.size(); i
++)
113 if (instr
->operands
[i
].isLiteral() && instr
->isVOP3() && program
->chip_class
>= GFX10
) {
115 } else if (instr
->operands
[i
].isLiteral()) {
116 check(instr
->format
== Format::SOP1
||
117 instr
->format
== Format::SOP2
||
118 instr
->format
== Format::SOPC
||
119 instr
->format
== Format::VOP1
||
120 instr
->format
== Format::VOP2
||
121 instr
->format
== Format::VOPC
,
122 "Literal applied on wrong instruction format", instr
.get());
125 check(!instr
->isVALU() || i
== 0 || i
== 2, "Wrong source position for Literal argument", instr
.get());
128 check(num_literals
<= 1, "Only 1 Literal allowed", instr
.get());
130 /* check num sgprs for VALU */
131 if (instr
->isVALU()) {
132 check(instr
->definitions
[0].getTemp().type() == RegType::vgpr
||
133 (int) instr
->format
& (int) Format::VOPC
||
134 instr
->opcode
== aco_opcode::v_readfirstlane_b32
||
135 instr
->opcode
== aco_opcode::v_readlane_b32
||
136 instr
->opcode
== aco_opcode::v_readlane_b32_e64
,
137 "Wrong Definition type for VALU instruction", instr
.get());
138 unsigned num_sgpr
= 0;
139 unsigned sgpr_idx
= instr
->operands
.size();
140 for (unsigned i
= 0; i
< instr
->operands
.size(); i
++) {
141 if (instr
->opcode
== aco_opcode::v_readfirstlane_b32
||
142 instr
->opcode
== aco_opcode::v_readlane_b32
||
143 instr
->opcode
== aco_opcode::v_readlane_b32_e64
||
144 instr
->opcode
== aco_opcode::v_writelane_b32
||
145 instr
->opcode
== aco_opcode::v_writelane_b32_e64
) {
146 check(!instr
->operands
[i
].isLiteral(), "No literal allowed on VALU instruction", instr
.get());
147 check(i
== 1 || (instr
->operands
[i
].isTemp() && instr
->operands
[i
].regClass() == v1
), "Wrong Operand type for VALU instruction", instr
.get());
150 if (instr
->operands
[i
].isTemp() && instr
->operands
[i
].regClass().type() == RegType::sgpr
) {
151 check(i
!= 1 || (int) instr
->format
& (int) Format::VOP3A
, "Wrong source position for SGPR argument", instr
.get());
153 if (sgpr_idx
== instr
->operands
.size() || instr
->operands
[sgpr_idx
].tempId() != instr
->operands
[i
].tempId())
158 if (instr
->operands
[i
].isConstant() && !instr
->operands
[i
].isLiteral())
159 check(i
== 0 || (int) instr
->format
& (int) Format::VOP3A
, "Wrong source position for constant argument", instr
.get());
161 check(num_sgpr
+ num_literals
<= 1, "Only 1 Literal OR 1 SGPR allowed", instr
.get());
164 if (instr
->format
== Format::SOP1
|| instr
->format
== Format::SOP2
) {
165 check(instr
->definitions
[0].getTemp().type() == RegType::sgpr
, "Wrong Definition type for SALU instruction", instr
.get());
166 for (const Operand
& op
: instr
->operands
) {
167 check(op
.isConstant() || op
.regClass().type() <= RegType::sgpr
,
168 "Wrong Operand type for SALU instruction", instr
.get());
173 switch (instr
->format
) {
174 case Format::PSEUDO
: {
175 if (instr
->opcode
== aco_opcode::p_create_vector
) {
177 for (const Operand
& op
: instr
->operands
) {
180 check(size
== instr
->definitions
[0].size(), "Definition size does not match operand sizes", instr
.get());
181 if (instr
->definitions
[0].getTemp().type() == RegType::sgpr
) {
182 for (const Operand
& op
: instr
->operands
) {
183 check(op
.isConstant() || op
.regClass().type() == RegType::sgpr
,
184 "Wrong Operand type for scalar vector", instr
.get());
187 } else if (instr
->opcode
== aco_opcode::p_extract_vector
) {
188 check((instr
->operands
[0].isTemp()) && instr
->operands
[1].isConstant(), "Wrong Operand types", instr
.get());
189 check(instr
->operands
[1].constantValue() < instr
->operands
[0].size(), "Index out of range", instr
.get());
190 check(instr
->definitions
[0].getTemp().type() == RegType::vgpr
|| instr
->operands
[0].regClass().type() == RegType::sgpr
,
191 "Cannot extract SGPR value from VGPR vector", instr
.get());
192 } else if (instr
->opcode
== aco_opcode::p_parallelcopy
) {
193 check(instr
->definitions
.size() == instr
->operands
.size(), "Number of Operands does not match number of Definitions", instr
.get());
194 for (unsigned i
= 0; i
< instr
->operands
.size(); i
++) {
195 if (instr
->operands
[i
].isTemp())
196 check((instr
->definitions
[i
].getTemp().type() == instr
->operands
[i
].regClass().type()) ||
197 (instr
->definitions
[i
].getTemp().type() == RegType::vgpr
&& instr
->operands
[i
].regClass().type() == RegType::sgpr
),
198 "Operand and Definition types do not match", instr
.get());
200 } else if (instr
->opcode
== aco_opcode::p_phi
) {
201 check(instr
->operands
.size() == block
.logical_preds
.size(), "Number of Operands does not match number of predecessors", instr
.get());
202 check(instr
->definitions
[0].getTemp().type() == RegType::vgpr
|| instr
->definitions
[0].getTemp().regClass() == program
->lane_mask
, "Logical Phi Definition must be vgpr or divergent boolean", instr
.get());
203 } else if (instr
->opcode
== aco_opcode::p_linear_phi
) {
204 for (const Operand
& op
: instr
->operands
)
205 check(!op
.isTemp() || op
.getTemp().is_linear(), "Wrong Operand type", instr
.get());
206 check(instr
->operands
.size() == block
.linear_preds
.size(), "Number of Operands does not match number of predecessors", instr
.get());
211 if (instr
->operands
.size() >= 1)
212 check(instr
->operands
[0].isTemp() && instr
->operands
[0].regClass().type() == RegType::sgpr
, "SMEM operands must be sgpr", instr
.get());
213 if (instr
->operands
.size() >= 2)
214 check(instr
->operands
[1].isConstant() || (instr
->operands
[1].isTemp() && instr
->operands
[1].regClass().type() == RegType::sgpr
),
215 "SMEM offset must be constant or sgpr", instr
.get());
216 if (!instr
->definitions
.empty())
217 check(instr
->definitions
[0].getTemp().type() == RegType::sgpr
, "SMEM result must be sgpr", instr
.get());
223 check(instr
->operands
.size() > 1, "VMEM instructions must have at least one operand", instr
.get());
224 check(instr
->operands
[0].hasRegClass() && instr
->operands
[0].regClass().type() == RegType::vgpr
,
225 "VADDR must be in vgpr for VMEM instructions", instr
.get());
226 check(instr
->operands
[1].isTemp() && instr
->operands
[1].regClass().type() == RegType::sgpr
, "VMEM resource constant must be sgpr", instr
.get());
227 check(instr
->operands
.size() < 4 || (instr
->operands
[3].isTemp() && instr
->operands
[3].regClass().type() == RegType::vgpr
), "VMEM write data must be vgpr", instr
.get());
231 for (const Operand
& op
: instr
->operands
) {
232 check((op
.isTemp() && op
.regClass().type() == RegType::vgpr
) || op
.physReg() == m0
,
233 "Only VGPRs are valid DS instruction operands", instr
.get());
235 if (!instr
->definitions
.empty())
236 check(instr
->definitions
[0].getTemp().type() == RegType::vgpr
, "DS instruction must return VGPR", instr
.get());
240 for (unsigned i
= 0; i
< 4; i
++)
241 check(instr
->operands
[i
].hasRegClass() && instr
->operands
[i
].regClass().type() == RegType::vgpr
,
242 "Only VGPRs are valid Export arguments", instr
.get());
246 check(instr
->operands
[1].isUndefined(), "Flat instructions don't support SADDR", instr
.get());
249 case Format::SCRATCH
: {
250 check(instr
->operands
[0].isTemp() && instr
->operands
[0].regClass().type() == RegType::vgpr
, "FLAT/GLOBAL/SCRATCH address must be vgpr", instr
.get());
251 check(instr
->operands
[1].hasRegClass() && instr
->operands
[1].regClass().type() == RegType::sgpr
,
252 "FLAT/GLOBAL/SCRATCH sgpr address must be undefined or sgpr", instr
.get());
253 if (!instr
->definitions
.empty())
254 check(instr
->definitions
[0].getTemp().type() == RegType::vgpr
, "FLAT/GLOBAL/SCRATCH result must be vgpr", instr
.get());
256 check(instr
->operands
[2].regClass().type() == RegType::vgpr
, "FLAT/GLOBAL/SCRATCH data must be vgpr", instr
.get());
266 for (unsigned i
= 0; i
< program
->blocks
.size(); i
++) {
267 Block
& block
= program
->blocks
[i
];
268 check_block(block
.index
== i
, "block.index must match actual index", &block
);
270 /* predecessors/successors should be sorted */
271 for (unsigned j
= 0; j
+ 1 < block
.linear_preds
.size(); j
++)
272 check_block(block
.linear_preds
[j
] < block
.linear_preds
[j
+ 1], "linear predecessors must be sorted", &block
);
273 for (unsigned j
= 0; j
+ 1 < block
.logical_preds
.size(); j
++)
274 check_block(block
.logical_preds
[j
] < block
.logical_preds
[j
+ 1], "logical predecessors must be sorted", &block
);
275 for (unsigned j
= 0; j
+ 1 < block
.linear_succs
.size(); j
++)
276 check_block(block
.linear_succs
[j
] < block
.linear_succs
[j
+ 1], "linear successors must be sorted", &block
);
277 for (unsigned j
= 0; j
+ 1 < block
.logical_succs
.size(); j
++)
278 check_block(block
.logical_succs
[j
] < block
.logical_succs
[j
+ 1], "logical successors must be sorted", &block
);
280 /* critical edges are not allowed */
281 if (block
.linear_preds
.size() > 1) {
282 for (unsigned pred
: block
.linear_preds
)
283 check_block(program
->blocks
[pred
].linear_succs
.size() == 1, "linear critical edges are not allowed", &program
->blocks
[pred
]);
284 for (unsigned pred
: block
.logical_preds
)
285 check_block(program
->blocks
[pred
].logical_succs
.size() == 1, "logical critical edges are not allowed", &program
->blocks
[pred
]);
296 Location() : block(NULL
), instr(NULL
) {}
299 Instruction
*instr
; //NULL if it's the block's live-in
308 bool ra_fail(FILE *output
, Location loc
, Location loc2
, const char *fmt
, ...) {
312 vsprintf(msg
, fmt
, args
);
315 fprintf(stderr
, "RA error found at instruction in BB%d:\n", loc
.block
->index
);
317 aco_print_instr(loc
.instr
, stderr
);
318 fprintf(stderr
, "\n%s", msg
);
320 fprintf(stderr
, "%s", msg
);
323 fprintf(stderr
, " in BB%d:\n", loc2
.block
->index
);
324 aco_print_instr(loc2
.instr
, stderr
);
326 fprintf(stderr
, "\n\n");
331 } /* end namespace */
333 bool validate_ra(Program
*program
, const struct radv_nir_compiler_options
*options
, FILE *output
) {
334 if (!(debug_flags
& DEBUG_VALIDATE_RA
))
338 aco::live live_vars
= aco::live_var_analysis(program
, options
);
339 std::vector
<std::vector
<Temp
>> phi_sgpr_ops(program
->blocks
.size());
341 std::map
<unsigned, Assignment
> assignments
;
342 for (Block
& block
: program
->blocks
) {
345 for (aco_ptr
<Instruction
>& instr
: block
.instructions
) {
346 if (instr
->opcode
== aco_opcode::p_phi
) {
347 for (unsigned i
= 0; i
< instr
->operands
.size(); i
++) {
348 if (instr
->operands
[i
].isTemp() &&
349 instr
->operands
[i
].getTemp().type() == RegType::sgpr
&&
350 instr
->operands
[i
].isFirstKill())
351 phi_sgpr_ops
[block
.logical_preds
[i
]].emplace_back(instr
->operands
[i
].getTemp());
355 loc
.instr
= instr
.get();
356 for (unsigned i
= 0; i
< instr
->operands
.size(); i
++) {
357 Operand
& op
= instr
->operands
[i
];
361 err
|= ra_fail(output
, loc
, Location(), "Operand %d is not assigned a register", i
);
362 if (assignments
.count(op
.tempId()) && assignments
[op
.tempId()].reg
!= op
.physReg())
363 err
|= ra_fail(output
, loc
, assignments
.at(op
.tempId()).firstloc
, "Operand %d has an inconsistent register assignment with instruction", i
);
364 if ((op
.getTemp().type() == RegType::vgpr
&& op
.physReg() + op
.size() > 256 + program
->config
->num_vgprs
) ||
365 (op
.getTemp().type() == RegType::sgpr
&& op
.physReg() + op
.size() > program
->config
->num_sgprs
&& op
.physReg() < program
->sgpr_limit
))
366 err
|= ra_fail(output
, loc
, assignments
.at(op
.tempId()).firstloc
, "Operand %d has an out-of-bounds register assignment", i
);
367 if (!assignments
[op
.tempId()].firstloc
.block
)
368 assignments
[op
.tempId()].firstloc
= loc
;
369 if (!assignments
[op
.tempId()].defloc
.block
)
370 assignments
[op
.tempId()].reg
= op
.physReg();
373 for (unsigned i
= 0; i
< instr
->definitions
.size(); i
++) {
374 Definition
& def
= instr
->definitions
[i
];
378 err
|= ra_fail(output
, loc
, Location(), "Definition %d is not assigned a register", i
);
379 if (assignments
[def
.tempId()].defloc
.block
)
380 err
|= ra_fail(output
, loc
, assignments
.at(def
.tempId()).defloc
, "Temporary %%%d also defined by instruction", def
.tempId());
381 if ((def
.getTemp().type() == RegType::vgpr
&& def
.physReg() + def
.size() > 256 + program
->config
->num_vgprs
) ||
382 (def
.getTemp().type() == RegType::sgpr
&& def
.physReg() + def
.size() > program
->config
->num_sgprs
&& def
.physReg() < program
->sgpr_limit
))
383 err
|= ra_fail(output
, loc
, assignments
.at(def
.tempId()).firstloc
, "Definition %d has an out-of-bounds register assignment", i
);
384 if (!assignments
[def
.tempId()].firstloc
.block
)
385 assignments
[def
.tempId()].firstloc
= loc
;
386 assignments
[def
.tempId()].defloc
= loc
;
387 assignments
[def
.tempId()].reg
= def
.physReg();
392 for (Block
& block
: program
->blocks
) {
396 std::array
<unsigned, 512> regs
;
400 live
.insert(live_vars
.live_out
[block
.index
].begin(), live_vars
.live_out
[block
.index
].end());
401 /* remove killed p_phi sgpr operands */
402 for (Temp tmp
: phi_sgpr_ops
[block
.index
])
406 for (Temp tmp
: live
) {
407 PhysReg reg
= assignments
.at(tmp
.id()).reg
;
408 for (unsigned i
= 0; i
< tmp
.size(); i
++) {
410 err
|= ra_fail(output
, loc
, Location(), "Assignment of element %d of %%%d already taken by %%%d in live-out", i
, tmp
.id(), regs
[reg
+ i
]);
412 regs
[reg
+ i
] = tmp
.id();
417 for (auto it
= block
.instructions
.rbegin(); it
!= block
.instructions
.rend(); ++it
) {
418 aco_ptr
<Instruction
>& instr
= *it
;
420 /* check killed p_phi sgpr operands */
421 if (instr
->opcode
== aco_opcode::p_logical_end
) {
422 for (Temp tmp
: phi_sgpr_ops
[block
.index
]) {
423 PhysReg reg
= assignments
.at(tmp
.id()).reg
;
424 for (unsigned i
= 0; i
< tmp
.size(); i
++) {
426 err
|= ra_fail(output
, loc
, Location(), "Assignment of element %d of %%%d already taken by %%%d in live-out", i
, tmp
.id(), regs
[reg
+ i
]);
432 for (const Definition
& def
: instr
->definitions
) {
435 live
.erase(def
.getTemp());
438 /* don't count phi operands as live-in, since they are actually
439 * killed when they are copied at the predecessor */
440 if (instr
->opcode
!= aco_opcode::p_phi
&& instr
->opcode
!= aco_opcode::p_linear_phi
) {
441 for (const Operand
& op
: instr
->operands
) {
444 live
.insert(op
.getTemp());
449 for (Temp tmp
: live
) {
450 PhysReg reg
= assignments
.at(tmp
.id()).reg
;
451 for (unsigned i
= 0; i
< tmp
.size(); i
++)
452 regs
[reg
+ i
] = tmp
.id();
455 for (aco_ptr
<Instruction
>& instr
: block
.instructions
) {
456 loc
.instr
= instr
.get();
458 /* remove killed p_phi operands from regs */
459 if (instr
->opcode
== aco_opcode::p_logical_end
) {
460 for (Temp tmp
: phi_sgpr_ops
[block
.index
]) {
461 PhysReg reg
= assignments
.at(tmp
.id()).reg
;
466 if (instr
->opcode
!= aco_opcode::p_phi
&& instr
->opcode
!= aco_opcode::p_linear_phi
) {
467 for (const Operand
& op
: instr
->operands
) {
470 if (op
.isFirstKill()) {
471 for (unsigned j
= 0; j
< op
.getTemp().size(); j
++)
472 regs
[op
.physReg() + j
] = 0;
477 for (unsigned i
= 0; i
< instr
->definitions
.size(); i
++) {
478 Definition
& def
= instr
->definitions
[i
];
481 Temp tmp
= def
.getTemp();
482 PhysReg reg
= assignments
.at(tmp
.id()).reg
;
483 for (unsigned j
= 0; j
< tmp
.size(); j
++) {
485 err
|= ra_fail(output
, loc
, assignments
.at(regs
[reg
+ i
]).defloc
, "Assignment of element %d of %%%d already taken by %%%d from instruction", i
, tmp
.id(), regs
[reg
+ j
]);
486 regs
[reg
+ j
] = tmp
.id();
490 for (const Definition
& def
: instr
->definitions
) {
494 for (unsigned j
= 0; j
< def
.getTemp().size(); j
++)
495 regs
[def
.physReg() + j
] = 0;