2
* Copyright © 2021 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
25
#include "aco_builder.h"
36
constexpr const size_t max_reg_cnt = 512;
39
bool operator==(const Idx& other) const { return block == other.block && instr == other.instr; }
40
bool operator!=(const Idx& other) const { return !operator==(other); }
42
bool found() const { return block != UINT32_MAX; }
48
Idx not_written_in_block{UINT32_MAX, 0};
49
Idx clobbered{UINT32_MAX, 1};
50
Idx const_or_undef{UINT32_MAX, 2};
51
Idx written_by_multiple_instrs{UINT32_MAX, 3};
56
uint32_t current_instr_idx;
57
std::vector<uint16_t> uses;
58
std::vector<std::array<Idx, max_reg_cnt>> instr_idx_by_regs;
60
void reset_block(Block* block)
62
current_block = block;
63
current_instr_idx = 0;
65
if ((block->kind & block_kind_loop_header) || block->linear_preds.empty()) {
66
std::fill(instr_idx_by_regs[block->index].begin(), instr_idx_by_regs[block->index].end(),
67
not_written_in_block);
69
unsigned first_pred = block->linear_preds[0];
70
for (unsigned i = 0; i < max_reg_cnt; i++) {
71
bool all_same = std::all_of(
72
std::next(block->linear_preds.begin()), block->linear_preds.end(),
74
{ return instr_idx_by_regs[pred][i] == instr_idx_by_regs[first_pred][i]; });
77
instr_idx_by_regs[block->index][i] = instr_idx_by_regs[first_pred][i];
79
instr_idx_by_regs[block->index][i] = not_written_in_block;
84
Instruction* get(Idx idx) { return program->blocks[idx.block].instructions[idx.instr].get(); }
88
save_reg_writes(pr_opt_ctx& ctx, aco_ptr<Instruction>& instr)
90
for (const Definition& def : instr->definitions) {
91
assert(def.regClass().type() != RegType::sgpr || def.physReg().reg() <= 255);
92
assert(def.regClass().type() != RegType::vgpr || def.physReg().reg() >= 256);
94
unsigned dw_size = DIV_ROUND_UP(def.bytes(), 4u);
95
unsigned r = def.physReg().reg();
96
Idx idx{ctx.current_block->index, ctx.current_instr_idx};
98
if (def.regClass().is_subdword())
101
assert((r + dw_size) <= max_reg_cnt);
102
assert(def.size() == dw_size || def.regClass().is_subdword());
103
std::fill(ctx.instr_idx_by_regs[ctx.current_block->index].begin() + r,
104
ctx.instr_idx_by_regs[ctx.current_block->index].begin() + r + dw_size, idx);
109
last_writer_idx(pr_opt_ctx& ctx, PhysReg physReg, RegClass rc)
111
/* Verify that all of the operand's registers are written by the same instruction. */
112
assert(physReg.reg() < max_reg_cnt);
113
Idx instr_idx = ctx.instr_idx_by_regs[ctx.current_block->index][physReg.reg()];
114
unsigned dw_size = DIV_ROUND_UP(rc.bytes(), 4u);
115
unsigned r = physReg.reg();
117
std::all_of(ctx.instr_idx_by_regs[ctx.current_block->index].begin() + r,
118
ctx.instr_idx_by_regs[ctx.current_block->index].begin() + r + dw_size,
119
[instr_idx](Idx i) { return i == instr_idx; });
121
return all_same ? instr_idx : written_by_multiple_instrs;
125
last_writer_idx(pr_opt_ctx& ctx, const Operand& op)
127
if (op.isConstant() || op.isUndefined())
128
return const_or_undef;
130
assert(op.physReg().reg() < max_reg_cnt);
131
Idx instr_idx = ctx.instr_idx_by_regs[ctx.current_block->index][op.physReg().reg()];
135
instr_idx = last_writer_idx(ctx, op.physReg(), op.regClass());
136
assert(instr_idx != written_by_multiple_instrs);
143
is_clobbered_since(pr_opt_ctx& ctx, PhysReg reg, RegClass rc, const Idx& idx)
145
/* If we didn't find an instruction, assume that the register is clobbered. */
149
/* TODO: We currently can't keep track of subdword registers. */
150
if (rc.is_subdword())
153
unsigned begin_reg = reg.reg();
154
unsigned end_reg = begin_reg + rc.size();
155
unsigned current_block_idx = ctx.current_block->index;
157
for (unsigned r = begin_reg; r < end_reg; ++r) {
158
Idx& i = ctx.instr_idx_by_regs[current_block_idx][r];
159
if (i == clobbered || i == written_by_multiple_instrs)
161
else if (i == not_written_in_block)
166
if (i.block > idx.block || (i.block == idx.block && i.instr > idx.instr))
173
template <typename T>
175
is_clobbered_since(pr_opt_ctx& ctx, const T& t, const Idx& idx)
177
return is_clobbered_since(ctx, t.physReg(), t.regClass(), idx);
181
try_apply_branch_vcc(pr_opt_ctx& ctx, aco_ptr<Instruction>& instr)
183
/* We are looking for the following pattern:
185
* vcc = ... ; last_vcc_wr
186
* sX, scc = s_and_bXX vcc, exec ; op0_instr
187
* (...vcc and exec must not be clobbered inbetween...)
188
* s_cbranch_XX scc ; instr
190
* If possible, the above is optimized into:
192
* vcc = ... ; last_vcc_wr
193
* s_cbranch_XX vcc ; instr modified to use vcc
196
/* Don't try to optimize this on GFX6-7 because SMEM may corrupt the vccz bit. */
197
if (ctx.program->chip_class < GFX8)
200
if (instr->format != Format::PSEUDO_BRANCH || instr->operands.size() == 0 ||
201
instr->operands[0].physReg() != scc)
204
Idx op0_instr_idx = last_writer_idx(ctx, instr->operands[0]);
205
Idx last_vcc_wr_idx = last_writer_idx(ctx, vcc, ctx.program->lane_mask);
207
/* We need to make sure:
208
* - the instructions that wrote the operand register and VCC are both found
209
* - the operand register used by the branch, and VCC were both written in the current block
210
* - EXEC hasn't been clobbered since the last VCC write
211
* - VCC hasn't been clobbered since the operand register was written
212
* (ie. the last VCC writer precedes the op0 writer)
214
if (!op0_instr_idx.found() || !last_vcc_wr_idx.found() ||
215
op0_instr_idx.block != ctx.current_block->index ||
216
last_vcc_wr_idx.block != ctx.current_block->index ||
217
is_clobbered_since(ctx, exec, ctx.program->lane_mask, last_vcc_wr_idx) ||
218
is_clobbered_since(ctx, vcc, ctx.program->lane_mask, op0_instr_idx))
221
Instruction* op0_instr = ctx.get(op0_instr_idx);
222
Instruction* last_vcc_wr = ctx.get(last_vcc_wr_idx);
224
if ((op0_instr->opcode != aco_opcode::s_and_b64 /* wave64 */ &&
225
op0_instr->opcode != aco_opcode::s_and_b32 /* wave32 */) ||
226
op0_instr->operands[0].physReg() != vcc || op0_instr->operands[1].physReg() != exec ||
227
!last_vcc_wr->isVOPC())
230
assert(last_vcc_wr->definitions[0].tempId() == op0_instr->operands[0].tempId());
232
/* Reduce the uses of the SCC def */
233
ctx.uses[instr->operands[0].tempId()]--;
234
/* Use VCC instead of SCC in the branch */
235
instr->operands[0] = op0_instr->operands[0];
239
try_optimize_scc_nocompare(pr_opt_ctx& ctx, aco_ptr<Instruction>& instr)
241
/* We are looking for the following pattern:
243
* s_bfe_u32 s0, s3, 0x40018 ; outputs SGPR and SCC if the SGPR != 0
244
* s_cmp_eq_i32 s0, 0 ; comparison between the SGPR and 0
245
* s_cbranch_scc0 BB3 ; use the result of the comparison, eg. branch or cselect
247
* If possible, the above is optimized into:
249
* s_bfe_u32 s0, s3, 0x40018 ; original instruction
250
* s_cbranch_scc1 BB3 ; modified to use SCC directly rather than the SGPR with comparison
254
if (!instr->isSALU() && !instr->isBranch())
257
if (instr->isSOPC() &&
258
(instr->opcode == aco_opcode::s_cmp_eq_u32 || instr->opcode == aco_opcode::s_cmp_eq_i32 ||
259
instr->opcode == aco_opcode::s_cmp_lg_u32 || instr->opcode == aco_opcode::s_cmp_lg_i32 ||
260
instr->opcode == aco_opcode::s_cmp_eq_u64 || instr->opcode == aco_opcode::s_cmp_lg_u64) &&
261
(instr->operands[0].constantEquals(0) || instr->operands[1].constantEquals(0)) &&
262
(instr->operands[0].isTemp() || instr->operands[1].isTemp())) {
263
/* Make sure the constant is always in operand 1 */
264
if (instr->operands[0].isConstant())
265
std::swap(instr->operands[0], instr->operands[1]);
267
if (ctx.uses[instr->operands[0].tempId()] > 1)
270
/* Make sure both SCC and Operand 0 are written by the same instruction. */
271
Idx wr_idx = last_writer_idx(ctx, instr->operands[0]);
272
Idx sccwr_idx = last_writer_idx(ctx, scc, s1);
273
if (!wr_idx.found() || wr_idx != sccwr_idx)
276
Instruction* wr_instr = ctx.get(wr_idx);
277
if (!wr_instr->isSALU() || wr_instr->definitions.size() < 2 ||
278
wr_instr->definitions[1].physReg() != scc)
281
/* Look for instructions which set SCC := (D != 0) */
282
switch (wr_instr->opcode) {
283
case aco_opcode::s_bfe_i32:
284
case aco_opcode::s_bfe_i64:
285
case aco_opcode::s_bfe_u32:
286
case aco_opcode::s_bfe_u64:
287
case aco_opcode::s_and_b32:
288
case aco_opcode::s_and_b64:
289
case aco_opcode::s_andn2_b32:
290
case aco_opcode::s_andn2_b64:
291
case aco_opcode::s_or_b32:
292
case aco_opcode::s_or_b64:
293
case aco_opcode::s_orn2_b32:
294
case aco_opcode::s_orn2_b64:
295
case aco_opcode::s_xor_b32:
296
case aco_opcode::s_xor_b64:
297
case aco_opcode::s_not_b32:
298
case aco_opcode::s_not_b64:
299
case aco_opcode::s_nor_b32:
300
case aco_opcode::s_nor_b64:
301
case aco_opcode::s_xnor_b32:
302
case aco_opcode::s_xnor_b64:
303
case aco_opcode::s_nand_b32:
304
case aco_opcode::s_nand_b64:
305
case aco_opcode::s_lshl_b32:
306
case aco_opcode::s_lshl_b64:
307
case aco_opcode::s_lshr_b32:
308
case aco_opcode::s_lshr_b64:
309
case aco_opcode::s_ashr_i32:
310
case aco_opcode::s_ashr_i64:
311
case aco_opcode::s_abs_i32:
312
case aco_opcode::s_absdiff_i32: break;
316
/* Use the SCC def from wr_instr */
317
ctx.uses[instr->operands[0].tempId()]--;
318
instr->operands[0] = Operand(wr_instr->definitions[1].getTemp(), scc);
319
ctx.uses[instr->operands[0].tempId()]++;
321
/* Set the opcode and operand to 32-bit */
322
instr->operands[1] = Operand::zero();
324
(instr->opcode == aco_opcode::s_cmp_eq_u32 || instr->opcode == aco_opcode::s_cmp_eq_i32 ||
325
instr->opcode == aco_opcode::s_cmp_eq_u64)
326
? aco_opcode::s_cmp_eq_u32
327
: aco_opcode::s_cmp_lg_u32;
328
} else if ((instr->format == Format::PSEUDO_BRANCH && instr->operands.size() == 1 &&
329
instr->operands[0].physReg() == scc) ||
330
instr->opcode == aco_opcode::s_cselect_b32) {
332
/* For cselect, operand 2 is the SCC condition */
333
unsigned scc_op_idx = 0;
334
if (instr->opcode == aco_opcode::s_cselect_b32) {
338
Idx wr_idx = last_writer_idx(ctx, instr->operands[scc_op_idx]);
342
Instruction* wr_instr = ctx.get(wr_idx);
344
/* Check if we found the pattern above. */
345
if (wr_instr->opcode != aco_opcode::s_cmp_eq_u32 &&
346
wr_instr->opcode != aco_opcode::s_cmp_lg_u32)
348
if (wr_instr->operands[0].physReg() != scc)
350
if (!wr_instr->operands[1].constantEquals(0))
353
/* The optimization can be unsafe when there are other users. */
354
if (ctx.uses[instr->operands[scc_op_idx].tempId()] > 1)
357
if (wr_instr->opcode == aco_opcode::s_cmp_eq_u32) {
358
/* Flip the meaning of the instruction to correctly use the SCC. */
359
if (instr->format == Format::PSEUDO_BRANCH)
360
instr->opcode = instr->opcode == aco_opcode::p_cbranch_z ? aco_opcode::p_cbranch_nz
361
: aco_opcode::p_cbranch_z;
362
else if (instr->opcode == aco_opcode::s_cselect_b32)
363
std::swap(instr->operands[0], instr->operands[1]);
366
"scc_nocompare optimization is only implemented for p_cbranch and s_cselect");
369
/* Use the SCC def from the original instruction, not the comparison */
370
ctx.uses[instr->operands[scc_op_idx].tempId()]--;
371
instr->operands[scc_op_idx] = wr_instr->operands[0];
376
try_combine_dpp(pr_opt_ctx& ctx, aco_ptr<Instruction>& instr)
378
/* We are looking for the following pattern:
380
* v_mov_dpp vA, vB, ... ; move instruction with DPP
381
* v_xxx vC, vA, ... ; current instr that uses the result from the move
383
* If possible, the above is optimized into:
385
* v_xxx_dpp vC, vB, ... ; current instr modified to use DPP directly
389
if (!instr->isVALU() || instr->isDPP())
392
for (unsigned i = 0; i < MIN2(2, instr->operands.size()); i++) {
393
Idx op_instr_idx = last_writer_idx(ctx, instr->operands[i]);
394
if (!op_instr_idx.found())
397
const Instruction* mov = ctx.get(op_instr_idx);
398
if (mov->opcode != aco_opcode::v_mov_b32 || !mov->isDPP())
400
bool dpp8 = mov->isDPP8();
401
if (!can_use_DPP(instr, false, dpp8))
404
/* If we aren't going to remove the v_mov_b32, we have to ensure that it doesn't overwrite
405
* it's own operand before we use it.
407
if (mov->definitions[0].physReg() == mov->operands[0].physReg() &&
408
(!mov->definitions[0].tempId() || ctx.uses[mov->definitions[0].tempId()] > 1))
411
/* Don't propagate DPP if the source register is overwritten since the move. */
412
if (is_clobbered_since(ctx, mov->operands[0], op_instr_idx))
415
if (i && !can_swap_operands(instr, &instr->opcode))
418
if (!dpp8) /* anything else doesn't make sense in SSA */
419
assert(mov->dpp16().row_mask == 0xf && mov->dpp16().bank_mask == 0xf);
421
if (--ctx.uses[mov->definitions[0].tempId()])
422
ctx.uses[mov->operands[0].tempId()]++;
424
convert_to_DPP(instr, dpp8);
427
DPP8_instruction* dpp = &instr->dpp8();
429
std::swap(dpp->operands[0], dpp->operands[1]);
431
dpp->operands[0] = mov->operands[0];
432
memcpy(dpp->lane_sel, mov->dpp8().lane_sel, sizeof(dpp->lane_sel));
434
DPP16_instruction* dpp = &instr->dpp16();
436
std::swap(dpp->operands[0], dpp->operands[1]);
437
std::swap(dpp->neg[0], dpp->neg[1]);
438
std::swap(dpp->abs[0], dpp->abs[1]);
440
dpp->operands[0] = mov->operands[0];
441
dpp->dpp_ctrl = mov->dpp16().dpp_ctrl;
442
dpp->bound_ctrl = true;
443
dpp->neg[0] ^= mov->dpp16().neg[0] && !dpp->abs[0];
444
dpp->abs[0] |= mov->dpp16().abs[0];
451
process_instruction(pr_opt_ctx& ctx, aco_ptr<Instruction>& instr)
453
try_apply_branch_vcc(ctx, instr);
455
try_optimize_scc_nocompare(ctx, instr);
457
try_combine_dpp(ctx, instr);
460
save_reg_writes(ctx, instr);
462
ctx.current_instr_idx++;
468
optimize_postRA(Program* program)
471
ctx.program = program;
472
ctx.uses = dead_code_analysis(program);
473
ctx.instr_idx_by_regs.resize(program->blocks.size());
476
* Goes through each instruction exactly once, and can transform
477
* instructions or adjust the use counts of temps.
479
for (auto& block : program->blocks) {
480
ctx.reset_block(&block);
482
for (aco_ptr<Instruction>& instr : block.instructions)
483
process_instruction(ctx, instr);
487
* Gets rid of instructions which are manually deleted or
488
* no longer have any uses.
490
for (auto& block : program->blocks) {
491
auto new_end = std::remove_if(block.instructions.begin(), block.instructions.end(),
492
[&ctx](const aco_ptr<Instruction>& instr)
493
{ return !instr || is_dead(ctx.uses, instr.get()); });
494
block.instructions.resize(new_end - block.instructions.begin());