| 1 |
// Copyright (c) 2018 Google LLC |
| 2 |
// |
| 3 |
// Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 |
// you may not use this file except in compliance with the License. |
| 5 |
// You may obtain a copy of the License at |
| 6 |
// |
| 7 |
// http://www.apache.org/licenses/LICENSE-2.0 |
| 8 |
// |
| 9 |
// Unless required by applicable law or agreed to in writing, software |
| 10 |
// distributed under the License is distributed on an "AS IS" BASIS, |
| 11 |
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 |
// See the License for the specific language governing permissions and |
| 13 |
// limitations under the License. |
| 14 |
|
| 15 |
#include "source/opt/folding_rules.h" |
| 16 |
|
| 17 |
#include <limits> |
| 18 |
#include <memory> |
| 19 |
#include <utility> |
| 20 |
|
| 21 |
#include "ir_builder.h" |
| 22 |
#include "source/latest_version_glsl_std_450_header.h" |
| 23 |
#include "source/opt/ir_context.h" |
| 24 |
|
| 25 |
namespace spvtools { |
| 26 |
namespace opt { |
| 27 |
namespace { |
| 28 |
|
| 29 |
const uint32_t kExtractCompositeIdInIdx = 0; |
| 30 |
const uint32_t kInsertObjectIdInIdx = 0; |
| 31 |
const uint32_t kInsertCompositeIdInIdx = 1; |
| 32 |
const uint32_t kExtInstSetIdInIdx = 0; |
| 33 |
const uint32_t kExtInstInstructionInIdx = 1; |
| 34 |
const uint32_t kFMixXIdInIdx = 2; |
| 35 |
const uint32_t kFMixYIdInIdx = 3; |
| 36 |
const uint32_t kFMixAIdInIdx = 4; |
| 37 |
const uint32_t kStoreObjectInIdx = 1; |
| 38 |
|
| 39 |
// Some image instructions may contain an "image operands" argument. |
| 40 |
// Returns the operand index for the "image operands". |
| 41 |
// Returns -1 if the instruction does not have image operands. |
| 42 |
int32_t ImageOperandsMaskInOperandIndex(Instruction* inst) { |
| 43 |
const auto opcode = inst->opcode(); |
| 44 |
switch (opcode) { |
| 45 |
case SpvOpImageSampleImplicitLod: |
| 46 |
case SpvOpImageSampleExplicitLod: |
| 47 |
case SpvOpImageSampleProjImplicitLod: |
| 48 |
case SpvOpImageSampleProjExplicitLod: |
| 49 |
case SpvOpImageFetch: |
| 50 |
case SpvOpImageRead: |
| 51 |
case SpvOpImageSparseSampleImplicitLod: |
| 52 |
case SpvOpImageSparseSampleExplicitLod: |
| 53 |
case SpvOpImageSparseSampleProjImplicitLod: |
| 54 |
case SpvOpImageSparseSampleProjExplicitLod: |
| 55 |
case SpvOpImageSparseFetch: |
| 56 |
case SpvOpImageSparseRead: |
| 57 |
return inst->NumOperands() > 4 ? 2 : -1; |
| 58 |
case SpvOpImageSampleDrefImplicitLod: |
| 59 |
case SpvOpImageSampleDrefExplicitLod: |
| 60 |
case SpvOpImageSampleProjDrefImplicitLod: |
| 61 |
case SpvOpImageSampleProjDrefExplicitLod: |
| 62 |
case SpvOpImageGather: |
| 63 |
case SpvOpImageDrefGather: |
| 64 |
case SpvOpImageSparseSampleDrefImplicitLod: |
| 65 |
case SpvOpImageSparseSampleDrefExplicitLod: |
| 66 |
case SpvOpImageSparseSampleProjDrefImplicitLod: |
| 67 |
case SpvOpImageSparseSampleProjDrefExplicitLod: |
| 68 |
case SpvOpImageSparseGather: |
| 69 |
case SpvOpImageSparseDrefGather: |
| 70 |
return inst->NumOperands() > 5 ? 3 : -1; |
| 71 |
case SpvOpImageWrite: |
| 72 |
return inst->NumOperands() > 3 ? 3 : -1; |
| 73 |
default: |
| 74 |
return -1; |
| 75 |
} |
| 76 |
} |
| 77 |
|
| 78 |
// Returns the element width of |type|. |
| 79 |
uint32_t ElementWidth(const analysis::Type* type) { |
| 80 |
if (const analysis::Vector* vec_type = type->AsVector()) { |
| 81 |
return ElementWidth(vec_type->element_type()); |
| 82 |
} else if (const analysis::Float* float_type = type->AsFloat()) { |
| 83 |
return float_type->width(); |
| 84 |
} else { |
| 85 |
assert(type->AsInteger()); |
| 86 |
return type->AsInteger()->width(); |
| 87 |
} |
| 88 |
} |
| 89 |
|
| 90 |
// Returns true if |type| is Float or a vector of Float. |
| 91 |
bool HasFloatingPoint(const analysis::Type* type) { |
| 92 |
if (type->AsFloat()) { |
| 93 |
return true; |
| 94 |
} else if (const analysis::Vector* vec_type = type->AsVector()) { |
| 95 |
return vec_type->element_type()->AsFloat() != nullptr; |
| 96 |
} |
| 97 |
|
| 98 |
return false; |
| 99 |
} |
| 100 |
|
| 101 |
// Returns false if |val| is NaN, infinite or subnormal. |
| 102 |
template <typename T> |
| 103 |
bool IsValidResult(T val) { |
| 104 |
int classified = std::fpclassify(val); |
| 105 |
switch (classified) { |
| 106 |
case FP_NAN: |
| 107 |
case FP_INFINITE: |
| 108 |
case FP_SUBNORMAL: |
| 109 |
return false; |
| 110 |
default: |
| 111 |
return true; |
| 112 |
} |
| 113 |
} |
| 114 |
|
| 115 |
const analysis::Constant* ConstInput( |
| 116 |
const std::vector<const analysis::Constant*>& constants) { |
| 117 |
return constants[0] ? constants[0] : constants[1]; |
| 118 |
} |
| 119 |
|
| 120 |
Instruction* NonConstInput(IRContext* context, const analysis::Constant* c, |
| 121 |
Instruction* inst) { |
| 122 |
uint32_t in_op = c ? 1u : 0u; |
| 123 |
return context->get_def_use_mgr()->GetDef( |
| 124 |
inst->GetSingleWordInOperand(in_op)); |
| 125 |
} |
| 126 |
|
| 127 |
// Returns the negation of |c|. |c| must be a 32 or 64 bit floating point |
| 128 |
// constant. |
| 129 |
uint32_t NegateFloatingPointConstant(analysis::ConstantManager* const_mgr, |
| 130 |
const analysis::Constant* c) { |
| 131 |
assert(c); |
| 132 |
assert(c->type()->AsFloat()); |
| 133 |
uint32_t width = c->type()->AsFloat()->width(); |
| 134 |
assert(width == 32 || width == 64); |
| 135 |
std::vector<uint32_t> words; |
| 136 |
if (width == 64) { |
| 137 |
utils::FloatProxy<double> result(c->GetDouble() * -1.0); |
| 138 |
words = result.GetWords(); |
| 139 |
} else { |
| 140 |
utils::FloatProxy<float> result(c->GetFloat() * -1.0f); |
| 141 |
words = result.GetWords(); |
| 142 |
} |
| 143 |
|
| 144 |
const analysis::Constant* negated_const = |
| 145 |
const_mgr->GetConstant(c->type(), std::move(words)); |
| 146 |
return const_mgr->GetDefiningInstruction(negated_const)->result_id(); |
| 147 |
} |
| 148 |
|
| 149 |
std::vector<uint32_t> ExtractInts(uint64_t val) { |
| 150 |
std::vector<uint32_t> words; |
| 151 |
words.push_back(static_cast<uint32_t>(val)); |
| 152 |
words.push_back(static_cast<uint32_t>(val >> 32)); |
| 153 |
return words; |
| 154 |
} |
| 155 |
|
| 156 |
// Negates the integer constant |c|. Returns the id of the defining instruction. |
| 157 |
uint32_t NegateIntegerConstant(analysis::ConstantManager* const_mgr, |
| 158 |
const analysis::Constant* c) { |
| 159 |
assert(c); |
| 160 |
assert(c->type()->AsInteger()); |
| 161 |
uint32_t width = c->type()->AsInteger()->width(); |
| 162 |
assert(width == 32 || width == 64); |
| 163 |
std::vector<uint32_t> words; |
| 164 |
if (width == 64) { |
| 165 |
uint64_t uval = static_cast<uint64_t>(0 - c->GetU64()); |
| 166 |
words = ExtractInts(uval); |
| 167 |
} else { |
| 168 |
words.push_back(static_cast<uint32_t>(0 - c->GetU32())); |
| 169 |
} |
| 170 |
|
| 171 |
const analysis::Constant* negated_const = |
| 172 |
const_mgr->GetConstant(c->type(), std::move(words)); |
| 173 |
return const_mgr->GetDefiningInstruction(negated_const)->result_id(); |
| 174 |
} |
| 175 |
|
| 176 |
// Negates the vector constant |c|. Returns the id of the defining instruction. |
| 177 |
uint32_t NegateVectorConstant(analysis::ConstantManager* const_mgr, |
| 178 |
const analysis::Constant* c) { |
| 179 |
assert(const_mgr && c); |
| 180 |
assert(c->type()->AsVector()); |
| 181 |
if (c->AsNullConstant()) { |
| 182 |
// 0.0 vs -0.0 shouldn't matter. |
| 183 |
return const_mgr->GetDefiningInstruction(c)->result_id(); |
| 184 |
} else { |
| 185 |
const analysis::Type* component_type = |
| 186 |
c->AsVectorConstant()->component_type(); |
| 187 |
std::vector<uint32_t> words; |
| 188 |
for (auto& comp : c->AsVectorConstant()->GetComponents()) { |
| 189 |
if (component_type->AsFloat()) { |
| 190 |
words.push_back(NegateFloatingPointConstant(const_mgr, comp)); |
| 191 |
} else { |
| 192 |
assert(component_type->AsInteger()); |
| 193 |
words.push_back(NegateIntegerConstant(const_mgr, comp)); |
| 194 |
} |
| 195 |
} |
| 196 |
|
| 197 |
const analysis::Constant* negated_const = |
| 198 |
const_mgr->GetConstant(c->type(), std::move(words)); |
| 199 |
return const_mgr->GetDefiningInstruction(negated_const)->result_id(); |
| 200 |
} |
| 201 |
} |
| 202 |
|
| 203 |
// Negates |c|. Returns the id of the defining instruction. |
| 204 |
uint32_t NegateConstant(analysis::ConstantManager* const_mgr, |
| 205 |
const analysis::Constant* c) { |
| 206 |
if (c->type()->AsVector()) { |
| 207 |
return NegateVectorConstant(const_mgr, c); |
| 208 |
} else if (c->type()->AsFloat()) { |
| 209 |
return NegateFloatingPointConstant(const_mgr, c); |
| 210 |
} else { |
| 211 |
assert(c->type()->AsInteger()); |
| 212 |
return NegateIntegerConstant(const_mgr, c); |
| 213 |
} |
| 214 |
} |
| 215 |
|
| 216 |
// Takes the reciprocal of |c|. |c|'s type must be Float or a vector of Float. |
| 217 |
// Returns 0 if the reciprocal is NaN, infinite or subnormal. |
| 218 |
uint32_t Reciprocal(analysis::ConstantManager* const_mgr, |
| 219 |
const analysis::Constant* c) { |
| 220 |
assert(const_mgr && c); |
| 221 |
assert(c->type()->AsFloat()); |
| 222 |
|
| 223 |
uint32_t width = c->type()->AsFloat()->width(); |
| 224 |
assert(width == 32 || width == 64); |
| 225 |
std::vector<uint32_t> words; |
| 226 |
if (width == 64) { |
| 227 |
spvtools::utils::FloatProxy<double> result(1.0 / c->GetDouble()); |
| 228 |
if (!IsValidResult(result.getAsFloat())) return 0; |
| 229 |
words = result.GetWords(); |
| 230 |
} else { |
| 231 |
spvtools::utils::FloatProxy<float> result(1.0f / c->GetFloat()); |
| 232 |
if (!IsValidResult(result.getAsFloat())) return 0; |
| 233 |
words = result.GetWords(); |
| 234 |
} |
| 235 |
|
| 236 |
const analysis::Constant* negated_const = |
| 237 |
const_mgr->GetConstant(c->type(), std::move(words)); |
| 238 |
return const_mgr->GetDefiningInstruction(negated_const)->result_id(); |
| 239 |
} |
| 240 |
|
| 241 |
// Replaces fdiv where second operand is constant with fmul. |
| 242 |
FoldingRule ReciprocalFDiv() { |
| 243 |
return [](IRContext* context, Instruction* inst, |
| 244 |
const std::vector<const analysis::Constant*>& constants) { |
| 245 |
assert(inst->opcode() == SpvOpFDiv); |
| 246 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 247 |
const analysis::Type* type = |
| 248 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 249 |
if (!inst->IsFloatingPointFoldingAllowed()) return false; |
| 250 |
|
| 251 |
uint32_t width = ElementWidth(type); |
| 252 |
if (width != 32 && width != 64) return false; |
| 253 |
|
| 254 |
if (constants[1] != nullptr) { |
| 255 |
uint32_t id = 0; |
| 256 |
if (const analysis::VectorConstant* vector_const = |
| 257 |
constants[1]->AsVectorConstant()) { |
| 258 |
std::vector<uint32_t> neg_ids; |
| 259 |
for (auto& comp : vector_const->GetComponents()) { |
| 260 |
id = Reciprocal(const_mgr, comp); |
| 261 |
if (id == 0) return false; |
| 262 |
neg_ids.push_back(id); |
| 263 |
} |
| 264 |
const analysis::Constant* negated_const = |
| 265 |
const_mgr->GetConstant(constants[1]->type(), std::move(neg_ids)); |
| 266 |
id = const_mgr->GetDefiningInstruction(negated_const)->result_id(); |
| 267 |
} else if (constants[1]->AsFloatConstant()) { |
| 268 |
id = Reciprocal(const_mgr, constants[1]); |
| 269 |
if (id == 0) return false; |
| 270 |
} else { |
| 271 |
// Don't fold a null constant. |
| 272 |
return false; |
| 273 |
} |
| 274 |
inst->SetOpcode(SpvOpFMul); |
| 275 |
inst->SetInOperands( |
| 276 |
{{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(0u)}}, |
| 277 |
{SPV_OPERAND_TYPE_ID, {id}}}); |
| 278 |
return true; |
| 279 |
} |
| 280 |
|
| 281 |
return false; |
| 282 |
}; |
| 283 |
} |
| 284 |
|
| 285 |
// Elides consecutive negate instructions. |
| 286 |
FoldingRule MergeNegateArithmetic() { |
| 287 |
return [](IRContext* context, Instruction* inst, |
| 288 |
const std::vector<const analysis::Constant*>& constants) { |
| 289 |
assert(inst->opcode() == SpvOpFNegate || inst->opcode() == SpvOpSNegate); |
| 290 |
(void)constants; |
| 291 |
const analysis::Type* type = |
| 292 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 293 |
if (HasFloatingPoint(type) && !inst->IsFloatingPointFoldingAllowed()) |
| 294 |
return false; |
| 295 |
|
| 296 |
Instruction* op_inst = |
| 297 |
context->get_def_use_mgr()->GetDef(inst->GetSingleWordInOperand(0u)); |
| 298 |
if (HasFloatingPoint(type) && !op_inst->IsFloatingPointFoldingAllowed()) |
| 299 |
return false; |
| 300 |
|
| 301 |
if (op_inst->opcode() == inst->opcode()) { |
| 302 |
// Elide negates. |
| 303 |
inst->SetOpcode(SpvOpCopyObject); |
| 304 |
inst->SetInOperands( |
| 305 |
{{SPV_OPERAND_TYPE_ID, {op_inst->GetSingleWordInOperand(0u)}}}); |
| 306 |
return true; |
| 307 |
} |
| 308 |
|
| 309 |
return false; |
| 310 |
}; |
| 311 |
} |
| 312 |
|
| 313 |
// Merges negate into a mul or div operation if that operation contains a |
| 314 |
// constant operand. |
| 315 |
// Cases: |
| 316 |
// -(x * 2) = x * -2 |
| 317 |
// -(2 * x) = x * -2 |
| 318 |
// -(x / 2) = x / -2 |
| 319 |
// -(2 / x) = -2 / x |
| 320 |
FoldingRule MergeNegateMulDivArithmetic() { |
| 321 |
return [](IRContext* context, Instruction* inst, |
| 322 |
const std::vector<const analysis::Constant*>& constants) { |
| 323 |
assert(inst->opcode() == SpvOpFNegate || inst->opcode() == SpvOpSNegate); |
| 324 |
(void)constants; |
| 325 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 326 |
const analysis::Type* type = |
| 327 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 328 |
if (HasFloatingPoint(type) && !inst->IsFloatingPointFoldingAllowed()) |
| 329 |
return false; |
| 330 |
|
| 331 |
Instruction* op_inst = |
| 332 |
context->get_def_use_mgr()->GetDef(inst->GetSingleWordInOperand(0u)); |
| 333 |
if (HasFloatingPoint(type) && !op_inst->IsFloatingPointFoldingAllowed()) |
| 334 |
return false; |
| 335 |
|
| 336 |
uint32_t width = ElementWidth(type); |
| 337 |
if (width != 32 && width != 64) return false; |
| 338 |
|
| 339 |
SpvOp opcode = op_inst->opcode(); |
| 340 |
if (opcode == SpvOpFMul || opcode == SpvOpFDiv || opcode == SpvOpIMul || |
| 341 |
opcode == SpvOpSDiv || opcode == SpvOpUDiv) { |
| 342 |
std::vector<const analysis::Constant*> op_constants = |
| 343 |
const_mgr->GetOperandConstants(op_inst); |
| 344 |
// Merge negate into mul or div if one operand is constant. |
| 345 |
if (op_constants[0] || op_constants[1]) { |
| 346 |
bool zero_is_variable = op_constants[0] == nullptr; |
| 347 |
const analysis::Constant* c = ConstInput(op_constants); |
| 348 |
uint32_t neg_id = NegateConstant(const_mgr, c); |
| 349 |
uint32_t non_const_id = zero_is_variable |
| 350 |
? op_inst->GetSingleWordInOperand(0u) |
| 351 |
: op_inst->GetSingleWordInOperand(1u); |
| 352 |
// Change this instruction to a mul/div. |
| 353 |
inst->SetOpcode(op_inst->opcode()); |
| 354 |
if (opcode == SpvOpFDiv || opcode == SpvOpUDiv || opcode == SpvOpSDiv) { |
| 355 |
uint32_t op0 = zero_is_variable ? non_const_id : neg_id; |
| 356 |
uint32_t op1 = zero_is_variable ? neg_id : non_const_id; |
| 357 |
inst->SetInOperands( |
| 358 |
{{SPV_OPERAND_TYPE_ID, {op0}}, {SPV_OPERAND_TYPE_ID, {op1}}}); |
| 359 |
} else { |
| 360 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {non_const_id}}, |
| 361 |
{SPV_OPERAND_TYPE_ID, {neg_id}}}); |
| 362 |
} |
| 363 |
return true; |
| 364 |
} |
| 365 |
} |
| 366 |
|
| 367 |
return false; |
| 368 |
}; |
| 369 |
} |
| 370 |
|
| 371 |
// Merges negate into a add or sub operation if that operation contains a |
| 372 |
// constant operand. |
| 373 |
// Cases: |
| 374 |
// -(x + 2) = -2 - x |
| 375 |
// -(2 + x) = -2 - x |
| 376 |
// -(x - 2) = 2 - x |
| 377 |
// -(2 - x) = x - 2 |
| 378 |
FoldingRule MergeNegateAddSubArithmetic() { |
| 379 |
return [](IRContext* context, Instruction* inst, |
| 380 |
const std::vector<const analysis::Constant*>& constants) { |
| 381 |
assert(inst->opcode() == SpvOpFNegate || inst->opcode() == SpvOpSNegate); |
| 382 |
(void)constants; |
| 383 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 384 |
const analysis::Type* type = |
| 385 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 386 |
if (HasFloatingPoint(type) && !inst->IsFloatingPointFoldingAllowed()) |
| 387 |
return false; |
| 388 |
|
| 389 |
Instruction* op_inst = |
| 390 |
context->get_def_use_mgr()->GetDef(inst->GetSingleWordInOperand(0u)); |
| 391 |
if (HasFloatingPoint(type) && !op_inst->IsFloatingPointFoldingAllowed()) |
| 392 |
return false; |
| 393 |
|
| 394 |
uint32_t width = ElementWidth(type); |
| 395 |
if (width != 32 && width != 64) return false; |
| 396 |
|
| 397 |
if (op_inst->opcode() == SpvOpFAdd || op_inst->opcode() == SpvOpFSub || |
| 398 |
op_inst->opcode() == SpvOpIAdd || op_inst->opcode() == SpvOpISub) { |
| 399 |
std::vector<const analysis::Constant*> op_constants = |
| 400 |
const_mgr->GetOperandConstants(op_inst); |
| 401 |
if (op_constants[0] || op_constants[1]) { |
| 402 |
bool zero_is_variable = op_constants[0] == nullptr; |
| 403 |
bool is_add = (op_inst->opcode() == SpvOpFAdd) || |
| 404 |
(op_inst->opcode() == SpvOpIAdd); |
| 405 |
bool swap_operands = !is_add || zero_is_variable; |
| 406 |
bool negate_const = is_add; |
| 407 |
const analysis::Constant* c = ConstInput(op_constants); |
| 408 |
uint32_t const_id = 0; |
| 409 |
if (negate_const) { |
| 410 |
const_id = NegateConstant(const_mgr, c); |
| 411 |
} else { |
| 412 |
const_id = zero_is_variable ? op_inst->GetSingleWordInOperand(1u) |
| 413 |
: op_inst->GetSingleWordInOperand(0u); |
| 414 |
} |
| 415 |
|
| 416 |
// Swap operands if necessary and make the instruction a subtraction. |
| 417 |
uint32_t op0 = |
| 418 |
zero_is_variable ? op_inst->GetSingleWordInOperand(0u) : const_id; |
| 419 |
uint32_t op1 = |
| 420 |
zero_is_variable ? const_id : op_inst->GetSingleWordInOperand(1u); |
| 421 |
if (swap_operands) std::swap(op0, op1); |
| 422 |
inst->SetOpcode(HasFloatingPoint(type) ? SpvOpFSub : SpvOpISub); |
| 423 |
inst->SetInOperands( |
| 424 |
{{SPV_OPERAND_TYPE_ID, {op0}}, {SPV_OPERAND_TYPE_ID, {op1}}}); |
| 425 |
return true; |
| 426 |
} |
| 427 |
} |
| 428 |
|
| 429 |
return false; |
| 430 |
}; |
| 431 |
} |
| 432 |
|
| 433 |
// Returns true if |c| has a zero element. |
| 434 |
bool HasZero(const analysis::Constant* c) { |
| 435 |
if (c->AsNullConstant()) { |
| 436 |
return true; |
| 437 |
} |
| 438 |
if (const analysis::VectorConstant* vec_const = c->AsVectorConstant()) { |
| 439 |
for (auto& comp : vec_const->GetComponents()) |
| 440 |
if (HasZero(comp)) return true; |
| 441 |
} else { |
| 442 |
assert(c->AsScalarConstant()); |
| 443 |
return c->AsScalarConstant()->IsZero(); |
| 444 |
} |
| 445 |
|
| 446 |
return false; |
| 447 |
} |
| 448 |
|
| 449 |
// Performs |input1| |opcode| |input2| and returns the merged constant result |
| 450 |
// id. Returns 0 if the result is not a valid value. The input types must be |
| 451 |
// Float. |
| 452 |
uint32_t PerformFloatingPointOperation(analysis::ConstantManager* const_mgr, |
| 453 |
SpvOp opcode, |
| 454 |
const analysis::Constant* input1, |
| 455 |
const analysis::Constant* input2) { |
| 456 |
const analysis::Type* type = input1->type(); |
| 457 |
assert(type->AsFloat()); |
| 458 |
uint32_t width = type->AsFloat()->width(); |
| 459 |
assert(width == 32 || width == 64); |
| 460 |
std::vector<uint32_t> words; |
| 461 |
#define FOLD_OP(op) \ |
| 462 |
if (width == 64) { \ |
| 463 |
utils::FloatProxy<double> val = \ |
| 464 |
input1->GetDouble() op input2->GetDouble(); \ |
| 465 |
double dval = val.getAsFloat(); \ |
| 466 |
if (!IsValidResult(dval)) return 0; \ |
| 467 |
words = val.GetWords(); \ |
| 468 |
} else { \ |
| 469 |
utils::FloatProxy<float> val = input1->GetFloat() op input2->GetFloat(); \ |
| 470 |
float fval = val.getAsFloat(); \ |
| 471 |
if (!IsValidResult(fval)) return 0; \ |
| 472 |
words = val.GetWords(); \ |
| 473 |
} |
| 474 |
switch (opcode) { |
| 475 |
case SpvOpFMul: |
| 476 |
FOLD_OP(*); |
| 477 |
break; |
| 478 |
case SpvOpFDiv: |
| 479 |
if (HasZero(input2)) return 0; |
| 480 |
FOLD_OP(/); |
| 481 |
break; |
| 482 |
case SpvOpFAdd: |
| 483 |
FOLD_OP(+); |
| 484 |
break; |
| 485 |
case SpvOpFSub: |
| 486 |
FOLD_OP(-); |
| 487 |
break; |
| 488 |
default: |
| 489 |
assert(false && "Unexpected operation"); |
| 490 |
break; |
| 491 |
} |
| 492 |
#undef FOLD_OP |
| 493 |
const analysis::Constant* merged_const = const_mgr->GetConstant(type, words); |
| 494 |
return const_mgr->GetDefiningInstruction(merged_const)->result_id(); |
| 495 |
} |
| 496 |
|
| 497 |
// Performs |input1| |opcode| |input2| and returns the merged constant result |
| 498 |
// id. Returns 0 if the result is not a valid value. The input types must be |
| 499 |
// Integers. |
| 500 |
uint32_t PerformIntegerOperation(analysis::ConstantManager* const_mgr, |
| 501 |
SpvOp opcode, const analysis::Constant* input1, |
| 502 |
const analysis::Constant* input2) { |
| 503 |
assert(input1->type()->AsInteger()); |
| 504 |
const analysis::Integer* type = input1->type()->AsInteger(); |
| 505 |
uint32_t width = type->AsInteger()->width(); |
| 506 |
assert(width == 32 || width == 64); |
| 507 |
std::vector<uint32_t> words; |
| 508 |
#define FOLD_OP(op) \ |
| 509 |
if (width == 64) { \ |
| 510 |
if (type->IsSigned()) { \ |
| 511 |
int64_t val = input1->GetS64() op input2->GetS64(); \ |
| 512 |
words = ExtractInts(static_cast<uint64_t>(val)); \ |
| 513 |
} else { \ |
| 514 |
uint64_t val = input1->GetU64() op input2->GetU64(); \ |
| 515 |
words = ExtractInts(val); \ |
| 516 |
} \ |
| 517 |
} else { \ |
| 518 |
if (type->IsSigned()) { \ |
| 519 |
int32_t val = input1->GetS32() op input2->GetS32(); \ |
| 520 |
words.push_back(static_cast<uint32_t>(val)); \ |
| 521 |
} else { \ |
| 522 |
uint32_t val = input1->GetU32() op input2->GetU32(); \ |
| 523 |
words.push_back(val); \ |
| 524 |
} \ |
| 525 |
} |
| 526 |
switch (opcode) { |
| 527 |
case SpvOpIMul: |
| 528 |
FOLD_OP(*); |
| 529 |
break; |
| 530 |
case SpvOpSDiv: |
| 531 |
case SpvOpUDiv: |
| 532 |
assert(false && "Should not merge integer division"); |
| 533 |
break; |
| 534 |
case SpvOpIAdd: |
| 535 |
FOLD_OP(+); |
| 536 |
break; |
| 537 |
case SpvOpISub: |
| 538 |
FOLD_OP(-); |
| 539 |
break; |
| 540 |
default: |
| 541 |
assert(false && "Unexpected operation"); |
| 542 |
break; |
| 543 |
} |
| 544 |
#undef FOLD_OP |
| 545 |
const analysis::Constant* merged_const = const_mgr->GetConstant(type, words); |
| 546 |
return const_mgr->GetDefiningInstruction(merged_const)->result_id(); |
| 547 |
} |
| 548 |
|
| 549 |
// Performs |input1| |opcode| |input2| and returns the merged constant result |
| 550 |
// id. Returns 0 if the result is not a valid value. The input types must be |
| 551 |
// Integers, Floats or Vectors of such. |
| 552 |
uint32_t PerformOperation(analysis::ConstantManager* const_mgr, SpvOp opcode, |
| 553 |
const analysis::Constant* input1, |
| 554 |
const analysis::Constant* input2) { |
| 555 |
assert(input1 && input2); |
| 556 |
const analysis::Type* type = input1->type(); |
| 557 |
std::vector<uint32_t> words; |
| 558 |
if (const analysis::Vector* vector_type = type->AsVector()) { |
| 559 |
const analysis::Type* ele_type = vector_type->element_type(); |
| 560 |
for (uint32_t i = 0; i != vector_type->element_count(); ++i) { |
| 561 |
uint32_t id = 0; |
| 562 |
|
| 563 |
const analysis::Constant* input1_comp = nullptr; |
| 564 |
if (const analysis::VectorConstant* input1_vector = |
| 565 |
input1->AsVectorConstant()) { |
| 566 |
input1_comp = input1_vector->GetComponents()[i]; |
| 567 |
} else { |
| 568 |
assert(input1->AsNullConstant()); |
| 569 |
input1_comp = const_mgr->GetConstant(ele_type, {}); |
| 570 |
} |
| 571 |
|
| 572 |
const analysis::Constant* input2_comp = nullptr; |
| 573 |
if (const analysis::VectorConstant* input2_vector = |
| 574 |
input2->AsVectorConstant()) { |
| 575 |
input2_comp = input2_vector->GetComponents()[i]; |
| 576 |
} else { |
| 577 |
assert(input2->AsNullConstant()); |
| 578 |
input2_comp = const_mgr->GetConstant(ele_type, {}); |
| 579 |
} |
| 580 |
|
| 581 |
if (ele_type->AsFloat()) { |
| 582 |
id = PerformFloatingPointOperation(const_mgr, opcode, input1_comp, |
| 583 |
input2_comp); |
| 584 |
} else { |
| 585 |
assert(ele_type->AsInteger()); |
| 586 |
id = PerformIntegerOperation(const_mgr, opcode, input1_comp, |
| 587 |
input2_comp); |
| 588 |
} |
| 589 |
if (id == 0) return 0; |
| 590 |
words.push_back(id); |
| 591 |
} |
| 592 |
const analysis::Constant* merged_const = |
| 593 |
const_mgr->GetConstant(type, words); |
| 594 |
return const_mgr->GetDefiningInstruction(merged_const)->result_id(); |
| 595 |
} else if (type->AsFloat()) { |
| 596 |
return PerformFloatingPointOperation(const_mgr, opcode, input1, input2); |
| 597 |
} else { |
| 598 |
assert(type->AsInteger()); |
| 599 |
return PerformIntegerOperation(const_mgr, opcode, input1, input2); |
| 600 |
} |
| 601 |
} |
| 602 |
|
| 603 |
// Merges consecutive multiplies where each contains one constant operand. |
| 604 |
// Cases: |
| 605 |
// 2 * (x * 2) = x * 4 |
| 606 |
// 2 * (2 * x) = x * 4 |
| 607 |
// (x * 2) * 2 = x * 4 |
| 608 |
// (2 * x) * 2 = x * 4 |
| 609 |
FoldingRule MergeMulMulArithmetic() { |
| 610 |
return [](IRContext* context, Instruction* inst, |
| 611 |
const std::vector<const analysis::Constant*>& constants) { |
| 612 |
assert(inst->opcode() == SpvOpFMul || inst->opcode() == SpvOpIMul); |
| 613 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 614 |
const analysis::Type* type = |
| 615 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 616 |
if (HasFloatingPoint(type) && !inst->IsFloatingPointFoldingAllowed()) |
| 617 |
return false; |
| 618 |
|
| 619 |
uint32_t width = ElementWidth(type); |
| 620 |
if (width != 32 && width != 64) return false; |
| 621 |
|
| 622 |
// Determine the constant input and the variable input in |inst|. |
| 623 |
const analysis::Constant* const_input1 = ConstInput(constants); |
| 624 |
if (!const_input1) return false; |
| 625 |
Instruction* other_inst = NonConstInput(context, constants[0], inst); |
| 626 |
if (HasFloatingPoint(type) && !other_inst->IsFloatingPointFoldingAllowed()) |
| 627 |
return false; |
| 628 |
|
| 629 |
if (other_inst->opcode() == inst->opcode()) { |
| 630 |
std::vector<const analysis::Constant*> other_constants = |
| 631 |
const_mgr->GetOperandConstants(other_inst); |
| 632 |
const analysis::Constant* const_input2 = ConstInput(other_constants); |
| 633 |
if (!const_input2) return false; |
| 634 |
|
| 635 |
bool other_first_is_variable = other_constants[0] == nullptr; |
| 636 |
uint32_t merged_id = PerformOperation(const_mgr, inst->opcode(), |
| 637 |
const_input1, const_input2); |
| 638 |
if (merged_id == 0) return false; |
| 639 |
|
| 640 |
uint32_t non_const_id = other_first_is_variable |
| 641 |
? other_inst->GetSingleWordInOperand(0u) |
| 642 |
: other_inst->GetSingleWordInOperand(1u); |
| 643 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {non_const_id}}, |
| 644 |
{SPV_OPERAND_TYPE_ID, {merged_id}}}); |
| 645 |
return true; |
| 646 |
} |
| 647 |
|
| 648 |
return false; |
| 649 |
}; |
| 650 |
} |
| 651 |
|
| 652 |
// Merges divides into subsequent multiplies if each instruction contains one |
| 653 |
// constant operand. Does not support integer operations. |
| 654 |
// Cases: |
| 655 |
// 2 * (x / 2) = x * 1 |
| 656 |
// 2 * (2 / x) = 4 / x |
| 657 |
// (x / 2) * 2 = x * 1 |
| 658 |
// (2 / x) * 2 = 4 / x |
| 659 |
// (y / x) * x = y |
| 660 |
// x * (y / x) = y |
| 661 |
FoldingRule MergeMulDivArithmetic() { |
| 662 |
return [](IRContext* context, Instruction* inst, |
| 663 |
const std::vector<const analysis::Constant*>& constants) { |
| 664 |
assert(inst->opcode() == SpvOpFMul); |
| 665 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 666 |
analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
| 667 |
|
| 668 |
const analysis::Type* type = |
| 669 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 670 |
if (!inst->IsFloatingPointFoldingAllowed()) return false; |
| 671 |
|
| 672 |
uint32_t width = ElementWidth(type); |
| 673 |
if (width != 32 && width != 64) return false; |
| 674 |
|
| 675 |
for (uint32_t i = 0; i < 2; i++) { |
| 676 |
uint32_t op_id = inst->GetSingleWordInOperand(i); |
| 677 |
Instruction* op_inst = def_use_mgr->GetDef(op_id); |
| 678 |
if (op_inst->opcode() == SpvOpFDiv) { |
| 679 |
if (op_inst->GetSingleWordInOperand(1) == |
| 680 |
inst->GetSingleWordInOperand(1 - i)) { |
| 681 |
inst->SetOpcode(SpvOpCopyObject); |
| 682 |
inst->SetInOperands( |
| 683 |
{{SPV_OPERAND_TYPE_ID, {op_inst->GetSingleWordInOperand(0)}}}); |
| 684 |
return true; |
| 685 |
} |
| 686 |
} |
| 687 |
} |
| 688 |
|
| 689 |
const analysis::Constant* const_input1 = ConstInput(constants); |
| 690 |
if (!const_input1) return false; |
| 691 |
Instruction* other_inst = NonConstInput(context, constants[0], inst); |
| 692 |
if (!other_inst->IsFloatingPointFoldingAllowed()) return false; |
| 693 |
|
| 694 |
if (other_inst->opcode() == SpvOpFDiv) { |
| 695 |
std::vector<const analysis::Constant*> other_constants = |
| 696 |
const_mgr->GetOperandConstants(other_inst); |
| 697 |
const analysis::Constant* const_input2 = ConstInput(other_constants); |
| 698 |
if (!const_input2 || HasZero(const_input2)) return false; |
| 699 |
|
| 700 |
bool other_first_is_variable = other_constants[0] == nullptr; |
| 701 |
// If the variable value is the second operand of the divide, multiply |
| 702 |
// the constants together. Otherwise divide the constants. |
| 703 |
uint32_t merged_id = PerformOperation( |
| 704 |
const_mgr, |
| 705 |
other_first_is_variable ? other_inst->opcode() : inst->opcode(), |
| 706 |
const_input1, const_input2); |
| 707 |
if (merged_id == 0) return false; |
| 708 |
|
| 709 |
uint32_t non_const_id = other_first_is_variable |
| 710 |
? other_inst->GetSingleWordInOperand(0u) |
| 711 |
: other_inst->GetSingleWordInOperand(1u); |
| 712 |
|
| 713 |
// If the variable value is on the second operand of the div, then this |
| 714 |
// operation is a div. Otherwise it should be a multiply. |
| 715 |
inst->SetOpcode(other_first_is_variable ? inst->opcode() |
| 716 |
: other_inst->opcode()); |
| 717 |
if (other_first_is_variable) { |
| 718 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {non_const_id}}, |
| 719 |
{SPV_OPERAND_TYPE_ID, {merged_id}}}); |
| 720 |
} else { |
| 721 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {merged_id}}, |
| 722 |
{SPV_OPERAND_TYPE_ID, {non_const_id}}}); |
| 723 |
} |
| 724 |
return true; |
| 725 |
} |
| 726 |
|
| 727 |
return false; |
| 728 |
}; |
| 729 |
} |
| 730 |
|
| 731 |
// Merges multiply of constant and negation. |
| 732 |
// Cases: |
| 733 |
// (-x) * 2 = x * -2 |
| 734 |
// 2 * (-x) = x * -2 |
| 735 |
FoldingRule MergeMulNegateArithmetic() { |
| 736 |
return [](IRContext* context, Instruction* inst, |
| 737 |
const std::vector<const analysis::Constant*>& constants) { |
| 738 |
assert(inst->opcode() == SpvOpFMul || inst->opcode() == SpvOpIMul); |
| 739 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 740 |
const analysis::Type* type = |
| 741 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 742 |
bool uses_float = HasFloatingPoint(type); |
| 743 |
if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
| 744 |
|
| 745 |
uint32_t width = ElementWidth(type); |
| 746 |
if (width != 32 && width != 64) return false; |
| 747 |
|
| 748 |
const analysis::Constant* const_input1 = ConstInput(constants); |
| 749 |
if (!const_input1) return false; |
| 750 |
Instruction* other_inst = NonConstInput(context, constants[0], inst); |
| 751 |
if (uses_float && !other_inst->IsFloatingPointFoldingAllowed()) |
| 752 |
return false; |
| 753 |
|
| 754 |
if (other_inst->opcode() == SpvOpFNegate || |
| 755 |
other_inst->opcode() == SpvOpSNegate) { |
| 756 |
uint32_t neg_id = NegateConstant(const_mgr, const_input1); |
| 757 |
|
| 758 |
inst->SetInOperands( |
| 759 |
{{SPV_OPERAND_TYPE_ID, {other_inst->GetSingleWordInOperand(0u)}}, |
| 760 |
{SPV_OPERAND_TYPE_ID, {neg_id}}}); |
| 761 |
return true; |
| 762 |
} |
| 763 |
|
| 764 |
return false; |
| 765 |
}; |
| 766 |
} |
| 767 |
|
| 768 |
// Merges consecutive divides if each instruction contains one constant operand. |
| 769 |
// Does not support integer division. |
| 770 |
// Cases: |
| 771 |
// 2 / (x / 2) = 4 / x |
| 772 |
// 4 / (2 / x) = 2 * x |
| 773 |
// (4 / x) / 2 = 2 / x |
| 774 |
// (x / 2) / 2 = x / 4 |
| 775 |
FoldingRule MergeDivDivArithmetic() { |
| 776 |
return [](IRContext* context, Instruction* inst, |
| 777 |
const std::vector<const analysis::Constant*>& constants) { |
| 778 |
assert(inst->opcode() == SpvOpFDiv); |
| 779 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 780 |
const analysis::Type* type = |
| 781 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 782 |
if (!inst->IsFloatingPointFoldingAllowed()) return false; |
| 783 |
|
| 784 |
uint32_t width = ElementWidth(type); |
| 785 |
if (width != 32 && width != 64) return false; |
| 786 |
|
| 787 |
const analysis::Constant* const_input1 = ConstInput(constants); |
| 788 |
if (!const_input1 || HasZero(const_input1)) return false; |
| 789 |
Instruction* other_inst = NonConstInput(context, constants[0], inst); |
| 790 |
if (!other_inst->IsFloatingPointFoldingAllowed()) return false; |
| 791 |
|
| 792 |
bool first_is_variable = constants[0] == nullptr; |
| 793 |
if (other_inst->opcode() == inst->opcode()) { |
| 794 |
std::vector<const analysis::Constant*> other_constants = |
| 795 |
const_mgr->GetOperandConstants(other_inst); |
| 796 |
const analysis::Constant* const_input2 = ConstInput(other_constants); |
| 797 |
if (!const_input2 || HasZero(const_input2)) return false; |
| 798 |
|
| 799 |
bool other_first_is_variable = other_constants[0] == nullptr; |
| 800 |
|
| 801 |
SpvOp merge_op = inst->opcode(); |
| 802 |
if (other_first_is_variable) { |
| 803 |
// Constants magnify. |
| 804 |
merge_op = SpvOpFMul; |
| 805 |
} |
| 806 |
|
| 807 |
// This is an x / (*) case. Swap the inputs. Doesn't harm multiply |
| 808 |
// because it is commutative. |
| 809 |
if (first_is_variable) std::swap(const_input1, const_input2); |
| 810 |
uint32_t merged_id = |
| 811 |
PerformOperation(const_mgr, merge_op, const_input1, const_input2); |
| 812 |
if (merged_id == 0) return false; |
| 813 |
|
| 814 |
uint32_t non_const_id = other_first_is_variable |
| 815 |
? other_inst->GetSingleWordInOperand(0u) |
| 816 |
: other_inst->GetSingleWordInOperand(1u); |
| 817 |
|
| 818 |
SpvOp op = inst->opcode(); |
| 819 |
if (!first_is_variable && !other_first_is_variable) { |
| 820 |
// Effectively div of 1/x, so change to multiply. |
| 821 |
op = SpvOpFMul; |
| 822 |
} |
| 823 |
|
| 824 |
uint32_t op1 = merged_id; |
| 825 |
uint32_t op2 = non_const_id; |
| 826 |
if (first_is_variable && other_first_is_variable) std::swap(op1, op2); |
| 827 |
inst->SetOpcode(op); |
| 828 |
inst->SetInOperands( |
| 829 |
{{SPV_OPERAND_TYPE_ID, {op1}}, {SPV_OPERAND_TYPE_ID, {op2}}}); |
| 830 |
return true; |
| 831 |
} |
| 832 |
|
| 833 |
return false; |
| 834 |
}; |
| 835 |
} |
| 836 |
|
| 837 |
// Fold multiplies succeeded by divides where each instruction contains a |
| 838 |
// constant operand. Does not support integer divide. |
| 839 |
// Cases: |
| 840 |
// 4 / (x * 2) = 2 / x |
| 841 |
// 4 / (2 * x) = 2 / x |
| 842 |
// (x * 4) / 2 = x * 2 |
| 843 |
// (4 * x) / 2 = x * 2 |
| 844 |
// (x * y) / x = y |
| 845 |
// (y * x) / x = y |
| 846 |
FoldingRule MergeDivMulArithmetic() { |
| 847 |
return [](IRContext* context, Instruction* inst, |
| 848 |
const std::vector<const analysis::Constant*>& constants) { |
| 849 |
assert(inst->opcode() == SpvOpFDiv); |
| 850 |
analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
| 851 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 852 |
|
| 853 |
const analysis::Type* type = |
| 854 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 855 |
if (!inst->IsFloatingPointFoldingAllowed()) return false; |
| 856 |
|
| 857 |
uint32_t width = ElementWidth(type); |
| 858 |
if (width != 32 && width != 64) return false; |
| 859 |
|
| 860 |
uint32_t op_id = inst->GetSingleWordInOperand(0); |
| 861 |
Instruction* op_inst = def_use_mgr->GetDef(op_id); |
| 862 |
|
| 863 |
if (op_inst->opcode() == SpvOpFMul) { |
| 864 |
for (uint32_t i = 0; i < 2; i++) { |
| 865 |
if (op_inst->GetSingleWordInOperand(i) == |
| 866 |
inst->GetSingleWordInOperand(1)) { |
| 867 |
inst->SetOpcode(SpvOpCopyObject); |
| 868 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, |
| 869 |
{op_inst->GetSingleWordInOperand(1 - i)}}}); |
| 870 |
return true; |
| 871 |
} |
| 872 |
} |
| 873 |
} |
| 874 |
|
| 875 |
const analysis::Constant* const_input1 = ConstInput(constants); |
| 876 |
if (!const_input1 || HasZero(const_input1)) return false; |
| 877 |
Instruction* other_inst = NonConstInput(context, constants[0], inst); |
| 878 |
if (!other_inst->IsFloatingPointFoldingAllowed()) return false; |
| 879 |
|
| 880 |
bool first_is_variable = constants[0] == nullptr; |
| 881 |
if (other_inst->opcode() == SpvOpFMul) { |
| 882 |
std::vector<const analysis::Constant*> other_constants = |
| 883 |
const_mgr->GetOperandConstants(other_inst); |
| 884 |
const analysis::Constant* const_input2 = ConstInput(other_constants); |
| 885 |
if (!const_input2) return false; |
| 886 |
|
| 887 |
bool other_first_is_variable = other_constants[0] == nullptr; |
| 888 |
|
| 889 |
// This is an x / (*) case. Swap the inputs. |
| 890 |
if (first_is_variable) std::swap(const_input1, const_input2); |
| 891 |
uint32_t merged_id = PerformOperation(const_mgr, inst->opcode(), |
| 892 |
const_input1, const_input2); |
| 893 |
if (merged_id == 0) return false; |
| 894 |
|
| 895 |
uint32_t non_const_id = other_first_is_variable |
| 896 |
? other_inst->GetSingleWordInOperand(0u) |
| 897 |
: other_inst->GetSingleWordInOperand(1u); |
| 898 |
|
| 899 |
uint32_t op1 = merged_id; |
| 900 |
uint32_t op2 = non_const_id; |
| 901 |
if (first_is_variable) std::swap(op1, op2); |
| 902 |
|
| 903 |
// Convert to multiply |
| 904 |
if (first_is_variable) inst->SetOpcode(other_inst->opcode()); |
| 905 |
inst->SetInOperands( |
| 906 |
{{SPV_OPERAND_TYPE_ID, {op1}}, {SPV_OPERAND_TYPE_ID, {op2}}}); |
| 907 |
return true; |
| 908 |
} |
| 909 |
|
| 910 |
return false; |
| 911 |
}; |
| 912 |
} |
| 913 |
|
| 914 |
// Fold divides of a constant and a negation. |
| 915 |
// Cases: |
| 916 |
// (-x) / 2 = x / -2 |
| 917 |
// 2 / (-x) = 2 / -x |
| 918 |
FoldingRule MergeDivNegateArithmetic() { |
| 919 |
return [](IRContext* context, Instruction* inst, |
| 920 |
const std::vector<const analysis::Constant*>& constants) { |
| 921 |
assert(inst->opcode() == SpvOpFDiv || inst->opcode() == SpvOpSDiv || |
| 922 |
inst->opcode() == SpvOpUDiv); |
| 923 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 924 |
const analysis::Type* type = |
| 925 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 926 |
bool uses_float = HasFloatingPoint(type); |
| 927 |
if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
| 928 |
|
| 929 |
uint32_t width = ElementWidth(type); |
| 930 |
if (width != 32 && width != 64) return false; |
| 931 |
|
| 932 |
const analysis::Constant* const_input1 = ConstInput(constants); |
| 933 |
if (!const_input1) return false; |
| 934 |
Instruction* other_inst = NonConstInput(context, constants[0], inst); |
| 935 |
if (uses_float && !other_inst->IsFloatingPointFoldingAllowed()) |
| 936 |
return false; |
| 937 |
|
| 938 |
bool first_is_variable = constants[0] == nullptr; |
| 939 |
if (other_inst->opcode() == SpvOpFNegate || |
| 940 |
other_inst->opcode() == SpvOpSNegate) { |
| 941 |
uint32_t neg_id = NegateConstant(const_mgr, const_input1); |
| 942 |
|
| 943 |
if (first_is_variable) { |
| 944 |
inst->SetInOperands( |
| 945 |
{{SPV_OPERAND_TYPE_ID, {other_inst->GetSingleWordInOperand(0u)}}, |
| 946 |
{SPV_OPERAND_TYPE_ID, {neg_id}}}); |
| 947 |
} else { |
| 948 |
inst->SetInOperands( |
| 949 |
{{SPV_OPERAND_TYPE_ID, {neg_id}}, |
| 950 |
{SPV_OPERAND_TYPE_ID, {other_inst->GetSingleWordInOperand(0u)}}}); |
| 951 |
} |
| 952 |
return true; |
| 953 |
} |
| 954 |
|
| 955 |
return false; |
| 956 |
}; |
| 957 |
} |
| 958 |
|
| 959 |
// Folds addition of a constant and a negation. |
| 960 |
// Cases: |
| 961 |
// (-x) + 2 = 2 - x |
| 962 |
// 2 + (-x) = 2 - x |
| 963 |
FoldingRule MergeAddNegateArithmetic() { |
| 964 |
return [](IRContext* context, Instruction* inst, |
| 965 |
const std::vector<const analysis::Constant*>& constants) { |
| 966 |
assert(inst->opcode() == SpvOpFAdd || inst->opcode() == SpvOpIAdd); |
| 967 |
const analysis::Type* type = |
| 968 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 969 |
bool uses_float = HasFloatingPoint(type); |
| 970 |
if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
| 971 |
|
| 972 |
const analysis::Constant* const_input1 = ConstInput(constants); |
| 973 |
if (!const_input1) return false; |
| 974 |
Instruction* other_inst = NonConstInput(context, constants[0], inst); |
| 975 |
if (uses_float && !other_inst->IsFloatingPointFoldingAllowed()) |
| 976 |
return false; |
| 977 |
|
| 978 |
if (other_inst->opcode() == SpvOpSNegate || |
| 979 |
other_inst->opcode() == SpvOpFNegate) { |
| 980 |
inst->SetOpcode(HasFloatingPoint(type) ? SpvOpFSub : SpvOpISub); |
| 981 |
uint32_t const_id = constants[0] ? inst->GetSingleWordInOperand(0u) |
| 982 |
: inst->GetSingleWordInOperand(1u); |
| 983 |
inst->SetInOperands( |
| 984 |
{{SPV_OPERAND_TYPE_ID, {const_id}}, |
| 985 |
{SPV_OPERAND_TYPE_ID, {other_inst->GetSingleWordInOperand(0u)}}}); |
| 986 |
return true; |
| 987 |
} |
| 988 |
return false; |
| 989 |
}; |
| 990 |
} |
| 991 |
|
| 992 |
// Folds subtraction of a constant and a negation. |
| 993 |
// Cases: |
| 994 |
// (-x) - 2 = -2 - x |
| 995 |
// 2 - (-x) = x + 2 |
| 996 |
FoldingRule MergeSubNegateArithmetic() { |
| 997 |
return [](IRContext* context, Instruction* inst, |
| 998 |
const std::vector<const analysis::Constant*>& constants) { |
| 999 |
assert(inst->opcode() == SpvOpFSub || inst->opcode() == SpvOpISub); |
| 1000 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 1001 |
const analysis::Type* type = |
| 1002 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 1003 |
bool uses_float = HasFloatingPoint(type); |
| 1004 |
if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
| 1005 |
|
| 1006 |
uint32_t width = ElementWidth(type); |
| 1007 |
if (width != 32 && width != 64) return false; |
| 1008 |
|
| 1009 |
const analysis::Constant* const_input1 = ConstInput(constants); |
| 1010 |
if (!const_input1) return false; |
| 1011 |
Instruction* other_inst = NonConstInput(context, constants[0], inst); |
| 1012 |
if (uses_float && !other_inst->IsFloatingPointFoldingAllowed()) |
| 1013 |
return false; |
| 1014 |
|
| 1015 |
if (other_inst->opcode() == SpvOpSNegate || |
| 1016 |
other_inst->opcode() == SpvOpFNegate) { |
| 1017 |
uint32_t op1 = 0; |
| 1018 |
uint32_t op2 = 0; |
| 1019 |
SpvOp opcode = inst->opcode(); |
| 1020 |
if (constants[0] != nullptr) { |
| 1021 |
op1 = other_inst->GetSingleWordInOperand(0u); |
| 1022 |
op2 = inst->GetSingleWordInOperand(0u); |
| 1023 |
opcode = HasFloatingPoint(type) ? SpvOpFAdd : SpvOpIAdd; |
| 1024 |
} else { |
| 1025 |
op1 = NegateConstant(const_mgr, const_input1); |
| 1026 |
op2 = other_inst->GetSingleWordInOperand(0u); |
| 1027 |
} |
| 1028 |
|
| 1029 |
inst->SetOpcode(opcode); |
| 1030 |
inst->SetInOperands( |
| 1031 |
{{SPV_OPERAND_TYPE_ID, {op1}}, {SPV_OPERAND_TYPE_ID, {op2}}}); |
| 1032 |
return true; |
| 1033 |
} |
| 1034 |
return false; |
| 1035 |
}; |
| 1036 |
} |
| 1037 |
|
| 1038 |
// Folds addition of an addition where each operation has a constant operand. |
| 1039 |
// Cases: |
| 1040 |
// (x + 2) + 2 = x + 4 |
| 1041 |
// (2 + x) + 2 = x + 4 |
| 1042 |
// 2 + (x + 2) = x + 4 |
| 1043 |
// 2 + (2 + x) = x + 4 |
| 1044 |
FoldingRule MergeAddAddArithmetic() { |
| 1045 |
return [](IRContext* context, Instruction* inst, |
| 1046 |
const std::vector<const analysis::Constant*>& constants) { |
| 1047 |
assert(inst->opcode() == SpvOpFAdd || inst->opcode() == SpvOpIAdd); |
| 1048 |
const analysis::Type* type = |
| 1049 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 1050 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 1051 |
bool uses_float = HasFloatingPoint(type); |
| 1052 |
if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
| 1053 |
|
| 1054 |
uint32_t width = ElementWidth(type); |
| 1055 |
if (width != 32 && width != 64) return false; |
| 1056 |
|
| 1057 |
const analysis::Constant* const_input1 = ConstInput(constants); |
| 1058 |
if (!const_input1) return false; |
| 1059 |
Instruction* other_inst = NonConstInput(context, constants[0], inst); |
| 1060 |
if (uses_float && !other_inst->IsFloatingPointFoldingAllowed()) |
| 1061 |
return false; |
| 1062 |
|
| 1063 |
if (other_inst->opcode() == SpvOpFAdd || |
| 1064 |
other_inst->opcode() == SpvOpIAdd) { |
| 1065 |
std::vector<const analysis::Constant*> other_constants = |
| 1066 |
const_mgr->GetOperandConstants(other_inst); |
| 1067 |
const analysis::Constant* const_input2 = ConstInput(other_constants); |
| 1068 |
if (!const_input2) return false; |
| 1069 |
|
| 1070 |
Instruction* non_const_input = |
| 1071 |
NonConstInput(context, other_constants[0], other_inst); |
| 1072 |
uint32_t merged_id = PerformOperation(const_mgr, inst->opcode(), |
| 1073 |
const_input1, const_input2); |
| 1074 |
if (merged_id == 0) return false; |
| 1075 |
|
| 1076 |
inst->SetInOperands( |
| 1077 |
{{SPV_OPERAND_TYPE_ID, {non_const_input->result_id()}}, |
| 1078 |
{SPV_OPERAND_TYPE_ID, {merged_id}}}); |
| 1079 |
return true; |
| 1080 |
} |
| 1081 |
return false; |
| 1082 |
}; |
| 1083 |
} |
| 1084 |
|
| 1085 |
// Folds addition of a subtraction where each operation has a constant operand. |
| 1086 |
// Cases: |
| 1087 |
// (x - 2) + 2 = x + 0 |
| 1088 |
// (2 - x) + 2 = 4 - x |
| 1089 |
// 2 + (x - 2) = x + 0 |
| 1090 |
// 2 + (2 - x) = 4 - x |
| 1091 |
FoldingRule MergeAddSubArithmetic() { |
| 1092 |
return [](IRContext* context, Instruction* inst, |
| 1093 |
const std::vector<const analysis::Constant*>& constants) { |
| 1094 |
assert(inst->opcode() == SpvOpFAdd || inst->opcode() == SpvOpIAdd); |
| 1095 |
const analysis::Type* type = |
| 1096 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 1097 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 1098 |
bool uses_float = HasFloatingPoint(type); |
| 1099 |
if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
| 1100 |
|
| 1101 |
uint32_t width = ElementWidth(type); |
| 1102 |
if (width != 32 && width != 64) return false; |
| 1103 |
|
| 1104 |
const analysis::Constant* const_input1 = ConstInput(constants); |
| 1105 |
if (!const_input1) return false; |
| 1106 |
Instruction* other_inst = NonConstInput(context, constants[0], inst); |
| 1107 |
if (uses_float && !other_inst->IsFloatingPointFoldingAllowed()) |
| 1108 |
return false; |
| 1109 |
|
| 1110 |
if (other_inst->opcode() == SpvOpFSub || |
| 1111 |
other_inst->opcode() == SpvOpISub) { |
| 1112 |
std::vector<const analysis::Constant*> other_constants = |
| 1113 |
const_mgr->GetOperandConstants(other_inst); |
| 1114 |
const analysis::Constant* const_input2 = ConstInput(other_constants); |
| 1115 |
if (!const_input2) return false; |
| 1116 |
|
| 1117 |
bool first_is_variable = other_constants[0] == nullptr; |
| 1118 |
SpvOp op = inst->opcode(); |
| 1119 |
uint32_t op1 = 0; |
| 1120 |
uint32_t op2 = 0; |
| 1121 |
if (first_is_variable) { |
| 1122 |
// Subtract constants. Non-constant operand is first. |
| 1123 |
op1 = other_inst->GetSingleWordInOperand(0u); |
| 1124 |
op2 = PerformOperation(const_mgr, other_inst->opcode(), const_input1, |
| 1125 |
const_input2); |
| 1126 |
} else { |
| 1127 |
// Add constants. Constant operand is first. Change the opcode. |
| 1128 |
op1 = PerformOperation(const_mgr, inst->opcode(), const_input1, |
| 1129 |
const_input2); |
| 1130 |
op2 = other_inst->GetSingleWordInOperand(1u); |
| 1131 |
op = other_inst->opcode(); |
| 1132 |
} |
| 1133 |
if (op1 == 0 || op2 == 0) return false; |
| 1134 |
|
| 1135 |
inst->SetOpcode(op); |
| 1136 |
inst->SetInOperands( |
| 1137 |
{{SPV_OPERAND_TYPE_ID, {op1}}, {SPV_OPERAND_TYPE_ID, {op2}}}); |
| 1138 |
return true; |
| 1139 |
} |
| 1140 |
return false; |
| 1141 |
}; |
| 1142 |
} |
| 1143 |
|
| 1144 |
// Folds subtraction of an addition where each operand has a constant operand. |
| 1145 |
// Cases: |
| 1146 |
// (x + 2) - 2 = x + 0 |
| 1147 |
// (2 + x) - 2 = x + 0 |
| 1148 |
// 2 - (x + 2) = 0 - x |
| 1149 |
// 2 - (2 + x) = 0 - x |
| 1150 |
FoldingRule MergeSubAddArithmetic() { |
| 1151 |
return [](IRContext* context, Instruction* inst, |
| 1152 |
const std::vector<const analysis::Constant*>& constants) { |
| 1153 |
assert(inst->opcode() == SpvOpFSub || inst->opcode() == SpvOpISub); |
| 1154 |
const analysis::Type* type = |
| 1155 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 1156 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 1157 |
bool uses_float = HasFloatingPoint(type); |
| 1158 |
if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
| 1159 |
|
| 1160 |
uint32_t width = ElementWidth(type); |
| 1161 |
if (width != 32 && width != 64) return false; |
| 1162 |
|
| 1163 |
const analysis::Constant* const_input1 = ConstInput(constants); |
| 1164 |
if (!const_input1) return false; |
| 1165 |
Instruction* other_inst = NonConstInput(context, constants[0], inst); |
| 1166 |
if (uses_float && !other_inst->IsFloatingPointFoldingAllowed()) |
| 1167 |
return false; |
| 1168 |
|
| 1169 |
if (other_inst->opcode() == SpvOpFAdd || |
| 1170 |
other_inst->opcode() == SpvOpIAdd) { |
| 1171 |
std::vector<const analysis::Constant*> other_constants = |
| 1172 |
const_mgr->GetOperandConstants(other_inst); |
| 1173 |
const analysis::Constant* const_input2 = ConstInput(other_constants); |
| 1174 |
if (!const_input2) return false; |
| 1175 |
|
| 1176 |
Instruction* non_const_input = |
| 1177 |
NonConstInput(context, other_constants[0], other_inst); |
| 1178 |
|
| 1179 |
// If the first operand of the sub is not a constant, swap the constants |
| 1180 |
// so the subtraction has the correct operands. |
| 1181 |
if (constants[0] == nullptr) std::swap(const_input1, const_input2); |
| 1182 |
// Subtract the constants. |
| 1183 |
uint32_t merged_id = PerformOperation(const_mgr, inst->opcode(), |
| 1184 |
const_input1, const_input2); |
| 1185 |
SpvOp op = inst->opcode(); |
| 1186 |
uint32_t op1 = 0; |
| 1187 |
uint32_t op2 = 0; |
| 1188 |
if (constants[0] == nullptr) { |
| 1189 |
// Non-constant operand is first. Change the opcode. |
| 1190 |
op1 = non_const_input->result_id(); |
| 1191 |
op2 = merged_id; |
| 1192 |
op = other_inst->opcode(); |
| 1193 |
} else { |
| 1194 |
// Constant operand is first. |
| 1195 |
op1 = merged_id; |
| 1196 |
op2 = non_const_input->result_id(); |
| 1197 |
} |
| 1198 |
if (op1 == 0 || op2 == 0) return false; |
| 1199 |
|
| 1200 |
inst->SetOpcode(op); |
| 1201 |
inst->SetInOperands( |
| 1202 |
{{SPV_OPERAND_TYPE_ID, {op1}}, {SPV_OPERAND_TYPE_ID, {op2}}}); |
| 1203 |
return true; |
| 1204 |
} |
| 1205 |
return false; |
| 1206 |
}; |
| 1207 |
} |
| 1208 |
|
| 1209 |
// Folds subtraction of a subtraction where each operand has a constant operand. |
| 1210 |
// Cases: |
| 1211 |
// (x - 2) - 2 = x - 4 |
| 1212 |
// (2 - x) - 2 = 0 - x |
| 1213 |
// 2 - (x - 2) = 4 - x |
| 1214 |
// 2 - (2 - x) = x + 0 |
| 1215 |
FoldingRule MergeSubSubArithmetic() { |
| 1216 |
return [](IRContext* context, Instruction* inst, |
| 1217 |
const std::vector<const analysis::Constant*>& constants) { |
| 1218 |
assert(inst->opcode() == SpvOpFSub || inst->opcode() == SpvOpISub); |
| 1219 |
const analysis::Type* type = |
| 1220 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 1221 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 1222 |
bool uses_float = HasFloatingPoint(type); |
| 1223 |
if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
| 1224 |
|
| 1225 |
uint32_t width = ElementWidth(type); |
| 1226 |
if (width != 32 && width != 64) return false; |
| 1227 |
|
| 1228 |
const analysis::Constant* const_input1 = ConstInput(constants); |
| 1229 |
if (!const_input1) return false; |
| 1230 |
Instruction* other_inst = NonConstInput(context, constants[0], inst); |
| 1231 |
if (uses_float && !other_inst->IsFloatingPointFoldingAllowed()) |
| 1232 |
return false; |
| 1233 |
|
| 1234 |
if (other_inst->opcode() == SpvOpFSub || |
| 1235 |
other_inst->opcode() == SpvOpISub) { |
| 1236 |
std::vector<const analysis::Constant*> other_constants = |
| 1237 |
const_mgr->GetOperandConstants(other_inst); |
| 1238 |
const analysis::Constant* const_input2 = ConstInput(other_constants); |
| 1239 |
if (!const_input2) return false; |
| 1240 |
|
| 1241 |
Instruction* non_const_input = |
| 1242 |
NonConstInput(context, other_constants[0], other_inst); |
| 1243 |
|
| 1244 |
// Merge the constants. |
| 1245 |
uint32_t merged_id = 0; |
| 1246 |
SpvOp merge_op = inst->opcode(); |
| 1247 |
if (other_constants[0] == nullptr) { |
| 1248 |
merge_op = uses_float ? SpvOpFAdd : SpvOpIAdd; |
| 1249 |
} else if (constants[0] == nullptr) { |
| 1250 |
std::swap(const_input1, const_input2); |
| 1251 |
} |
| 1252 |
merged_id = |
| 1253 |
PerformOperation(const_mgr, merge_op, const_input1, const_input2); |
| 1254 |
if (merged_id == 0) return false; |
| 1255 |
|
| 1256 |
SpvOp op = inst->opcode(); |
| 1257 |
if (constants[0] != nullptr && other_constants[0] != nullptr) { |
| 1258 |
// Change the operation. |
| 1259 |
op = uses_float ? SpvOpFAdd : SpvOpIAdd; |
| 1260 |
} |
| 1261 |
|
| 1262 |
uint32_t op1 = 0; |
| 1263 |
uint32_t op2 = 0; |
| 1264 |
if ((constants[0] == nullptr) ^ (other_constants[0] == nullptr)) { |
| 1265 |
op1 = merged_id; |
| 1266 |
op2 = non_const_input->result_id(); |
| 1267 |
} else { |
| 1268 |
op1 = non_const_input->result_id(); |
| 1269 |
op2 = merged_id; |
| 1270 |
} |
| 1271 |
|
| 1272 |
inst->SetOpcode(op); |
| 1273 |
inst->SetInOperands( |
| 1274 |
{{SPV_OPERAND_TYPE_ID, {op1}}, {SPV_OPERAND_TYPE_ID, {op2}}}); |
| 1275 |
return true; |
| 1276 |
} |
| 1277 |
return false; |
| 1278 |
}; |
| 1279 |
} |
| 1280 |
|
| 1281 |
// Helper function for MergeGenericAddSubArithmetic. If |addend| and |
| 1282 |
// subtrahend of |sub| is the same, merge to copy of minuend of |sub|. |
| 1283 |
bool MergeGenericAddendSub(uint32_t addend, uint32_t sub, Instruction* inst) { |
| 1284 |
IRContext* context = inst->context(); |
| 1285 |
analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
| 1286 |
Instruction* sub_inst = def_use_mgr->GetDef(sub); |
| 1287 |
if (sub_inst->opcode() != SpvOpFSub && sub_inst->opcode() != SpvOpISub) |
| 1288 |
return false; |
| 1289 |
if (sub_inst->opcode() == SpvOpFSub && |
| 1290 |
!sub_inst->IsFloatingPointFoldingAllowed()) |
| 1291 |
return false; |
| 1292 |
if (addend != sub_inst->GetSingleWordInOperand(1)) return false; |
| 1293 |
inst->SetOpcode(SpvOpCopyObject); |
| 1294 |
inst->SetInOperands( |
| 1295 |
{{SPV_OPERAND_TYPE_ID, {sub_inst->GetSingleWordInOperand(0)}}}); |
| 1296 |
context->UpdateDefUse(inst); |
| 1297 |
return true; |
| 1298 |
} |
| 1299 |
|
| 1300 |
// Folds addition of a subtraction where the subtrahend is equal to the |
| 1301 |
// other addend. Return a copy of the minuend. Accepts generic (const and |
| 1302 |
// non-const) operands. |
| 1303 |
// Cases: |
| 1304 |
// (a - b) + b = a |
| 1305 |
// b + (a - b) = a |
| 1306 |
FoldingRule MergeGenericAddSubArithmetic() { |
| 1307 |
return [](IRContext* context, Instruction* inst, |
| 1308 |
const std::vector<const analysis::Constant*>&) { |
| 1309 |
assert(inst->opcode() == SpvOpFAdd || inst->opcode() == SpvOpIAdd); |
| 1310 |
const analysis::Type* type = |
| 1311 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 1312 |
bool uses_float = HasFloatingPoint(type); |
| 1313 |
if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
| 1314 |
|
| 1315 |
uint32_t width = ElementWidth(type); |
| 1316 |
if (width != 32 && width != 64) return false; |
| 1317 |
|
| 1318 |
uint32_t add_op0 = inst->GetSingleWordInOperand(0); |
| 1319 |
uint32_t add_op1 = inst->GetSingleWordInOperand(1); |
| 1320 |
if (MergeGenericAddendSub(add_op0, add_op1, inst)) return true; |
| 1321 |
return MergeGenericAddendSub(add_op1, add_op0, inst); |
| 1322 |
}; |
| 1323 |
} |
| 1324 |
|
| 1325 |
// Helper function for FactorAddMuls. If |factor0_0| is the same as |factor1_0|, |
| 1326 |
// generate |factor0_0| * (|factor0_1| + |factor1_1|). |
| 1327 |
bool FactorAddMulsOpnds(uint32_t factor0_0, uint32_t factor0_1, |
| 1328 |
uint32_t factor1_0, uint32_t factor1_1, |
| 1329 |
Instruction* inst) { |
| 1330 |
IRContext* context = inst->context(); |
| 1331 |
if (factor0_0 != factor1_0) return false; |
| 1332 |
InstructionBuilder ir_builder( |
| 1333 |
context, inst, |
| 1334 |
IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping); |
| 1335 |
Instruction* new_add_inst = ir_builder.AddBinaryOp( |
| 1336 |
inst->type_id(), inst->opcode(), factor0_1, factor1_1); |
| 1337 |
inst->SetOpcode(inst->opcode() == SpvOpFAdd ? SpvOpFMul : SpvOpIMul); |
| 1338 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {factor0_0}}, |
| 1339 |
{SPV_OPERAND_TYPE_ID, {new_add_inst->result_id()}}}); |
| 1340 |
context->UpdateDefUse(inst); |
| 1341 |
return true; |
| 1342 |
} |
| 1343 |
|
| 1344 |
// Perform the following factoring identity, handling all operand order |
| 1345 |
// combinations: (a * b) + (a * c) = a * (b + c) |
| 1346 |
FoldingRule FactorAddMuls() { |
| 1347 |
return [](IRContext* context, Instruction* inst, |
| 1348 |
const std::vector<const analysis::Constant*>&) { |
| 1349 |
assert(inst->opcode() == SpvOpFAdd || inst->opcode() == SpvOpIAdd); |
| 1350 |
const analysis::Type* type = |
| 1351 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 1352 |
bool uses_float = HasFloatingPoint(type); |
| 1353 |
if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
| 1354 |
|
| 1355 |
analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
| 1356 |
uint32_t add_op0 = inst->GetSingleWordInOperand(0); |
| 1357 |
Instruction* add_op0_inst = def_use_mgr->GetDef(add_op0); |
| 1358 |
if (add_op0_inst->opcode() != SpvOpFMul && |
| 1359 |
add_op0_inst->opcode() != SpvOpIMul) |
| 1360 |
return false; |
| 1361 |
uint32_t add_op1 = inst->GetSingleWordInOperand(1); |
| 1362 |
Instruction* add_op1_inst = def_use_mgr->GetDef(add_op1); |
| 1363 |
if (add_op1_inst->opcode() != SpvOpFMul && |
| 1364 |
add_op1_inst->opcode() != SpvOpIMul) |
| 1365 |
return false; |
| 1366 |
|
| 1367 |
// Only perform this optimization if both of the muls only have one use. |
| 1368 |
// Otherwise this is a deoptimization in size and performance. |
| 1369 |
if (def_use_mgr->NumUses(add_op0_inst) > 1) return false; |
| 1370 |
if (def_use_mgr->NumUses(add_op1_inst) > 1) return false; |
| 1371 |
|
| 1372 |
if (add_op0_inst->opcode() == SpvOpFMul && |
| 1373 |
(!add_op0_inst->IsFloatingPointFoldingAllowed() || |
| 1374 |
!add_op1_inst->IsFloatingPointFoldingAllowed())) |
| 1375 |
return false; |
| 1376 |
|
| 1377 |
for (int i = 0; i < 2; i++) { |
| 1378 |
for (int j = 0; j < 2; j++) { |
| 1379 |
// Check if operand i in add_op0_inst matches operand j in add_op1_inst. |
| 1380 |
if (FactorAddMulsOpnds(add_op0_inst->GetSingleWordInOperand(i), |
| 1381 |
add_op0_inst->GetSingleWordInOperand(1 - i), |
| 1382 |
add_op1_inst->GetSingleWordInOperand(j), |
| 1383 |
add_op1_inst->GetSingleWordInOperand(1 - j), |
| 1384 |
inst)) |
| 1385 |
return true; |
| 1386 |
} |
| 1387 |
} |
| 1388 |
return false; |
| 1389 |
}; |
| 1390 |
} |
| 1391 |
|
| 1392 |
FoldingRule IntMultipleBy1() { |
| 1393 |
return [](IRContext*, Instruction* inst, |
| 1394 |
const std::vector<const analysis::Constant*>& constants) { |
| 1395 |
assert(inst->opcode() == SpvOpIMul && "Wrong opcode. Should be OpIMul."); |
| 1396 |
for (uint32_t i = 0; i < 2; i++) { |
| 1397 |
if (constants[i] == nullptr) { |
| 1398 |
continue; |
| 1399 |
} |
| 1400 |
const analysis::IntConstant* int_constant = constants[i]->AsIntConstant(); |
| 1401 |
if (int_constant) { |
| 1402 |
uint32_t width = ElementWidth(int_constant->type()); |
| 1403 |
if (width != 32 && width != 64) return false; |
| 1404 |
bool is_one = (width == 32) ? int_constant->GetU32BitValue() == 1u |
| 1405 |
: int_constant->GetU64BitValue() == 1ull; |
| 1406 |
if (is_one) { |
| 1407 |
inst->SetOpcode(SpvOpCopyObject); |
| 1408 |
inst->SetInOperands( |
| 1409 |
{{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(1 - i)}}}); |
| 1410 |
return true; |
| 1411 |
} |
| 1412 |
} |
| 1413 |
} |
| 1414 |
return false; |
| 1415 |
}; |
| 1416 |
} |
| 1417 |
|
| 1418 |
FoldingRule CompositeConstructFeedingExtract() { |
| 1419 |
return [](IRContext* context, Instruction* inst, |
| 1420 |
const std::vector<const analysis::Constant*>&) { |
| 1421 |
// If the input to an OpCompositeExtract is an OpCompositeConstruct, |
| 1422 |
// then we can simply use the appropriate element in the construction. |
| 1423 |
assert(inst->opcode() == SpvOpCompositeExtract && |
| 1424 |
"Wrong opcode. Should be OpCompositeExtract."); |
| 1425 |
analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
| 1426 |
analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 1427 |
|
| 1428 |
// If there are no index operands, then this rule cannot do anything. |
| 1429 |
if (inst->NumInOperands() <= 1) { |
| 1430 |
return false; |
| 1431 |
} |
| 1432 |
|
| 1433 |
uint32_t cid = inst->GetSingleWordInOperand(kExtractCompositeIdInIdx); |
| 1434 |
Instruction* cinst = def_use_mgr->GetDef(cid); |
| 1435 |
|
| 1436 |
if (cinst->opcode() != SpvOpCompositeConstruct) { |
| 1437 |
return false; |
| 1438 |
} |
| 1439 |
|
| 1440 |
std::vector<Operand> operands; |
| 1441 |
analysis::Type* composite_type = type_mgr->GetType(cinst->type_id()); |
| 1442 |
if (composite_type->AsVector() == nullptr) { |
| 1443 |
// Get the element being extracted from the OpCompositeConstruct |
| 1444 |
// Since it is not a vector, it is simple to extract the single element. |
| 1445 |
uint32_t element_index = inst->GetSingleWordInOperand(1); |
| 1446 |
uint32_t element_id = cinst->GetSingleWordInOperand(element_index); |
| 1447 |
operands.push_back({SPV_OPERAND_TYPE_ID, {element_id}}); |
| 1448 |
|
| 1449 |
// Add the remaining indices for extraction. |
| 1450 |
for (uint32_t i = 2; i < inst->NumInOperands(); ++i) { |
| 1451 |
operands.push_back({SPV_OPERAND_TYPE_LITERAL_INTEGER, |
| 1452 |
{inst->GetSingleWordInOperand(i)}}); |
| 1453 |
} |
| 1454 |
|
| 1455 |
} else { |
| 1456 |
// With vectors we have to handle the case where it is concatenating |
| 1457 |
// vectors. |
| 1458 |
assert(inst->NumInOperands() == 2 && |
| 1459 |
"Expecting a vector of scalar values."); |
| 1460 |
|
| 1461 |
uint32_t element_index = inst->GetSingleWordInOperand(1); |
| 1462 |
for (uint32_t construct_index = 0; |
| 1463 |
construct_index < cinst->NumInOperands(); ++construct_index) { |
| 1464 |
uint32_t element_id = cinst->GetSingleWordInOperand(construct_index); |
| 1465 |
Instruction* element_def = def_use_mgr->GetDef(element_id); |
| 1466 |
analysis::Vector* element_type = |
| 1467 |
type_mgr->GetType(element_def->type_id())->AsVector(); |
| 1468 |
if (element_type) { |
| 1469 |
uint32_t vector_size = element_type->element_count(); |
| 1470 |
if (vector_size <= element_index) { |
| 1471 |
// The element we want comes after this vector. |
| 1472 |
element_index -= vector_size; |
| 1473 |
} else { |
| 1474 |
// We want an element of this vector. |
| 1475 |
operands.push_back({SPV_OPERAND_TYPE_ID, {element_id}}); |
| 1476 |
operands.push_back( |
| 1477 |
{SPV_OPERAND_TYPE_LITERAL_INTEGER, {element_index}}); |
| 1478 |
break; |
| 1479 |
} |
| 1480 |
} else { |
| 1481 |
if (element_index == 0) { |
| 1482 |
// This is a scalar, and we this is the element we are extracting. |
| 1483 |
operands.push_back({SPV_OPERAND_TYPE_ID, {element_id}}); |
| 1484 |
break; |
| 1485 |
} else { |
| 1486 |
// Skip over this scalar value. |
| 1487 |
--element_index; |
| 1488 |
} |
| 1489 |
} |
| 1490 |
} |
| 1491 |
} |
| 1492 |
|
| 1493 |
// If there were no extra indices, then we have the final object. No need |
| 1494 |
// to extract even more. |
| 1495 |
if (operands.size() == 1) { |
| 1496 |
inst->SetOpcode(SpvOpCopyObject); |
| 1497 |
} |
| 1498 |
|
| 1499 |
inst->SetInOperands(std::move(operands)); |
| 1500 |
return true; |
| 1501 |
}; |
| 1502 |
} |
| 1503 |
|
| 1504 |
// If the OpCompositeConstruct is simply putting back together elements that |
| 1505 |
// where extracted from the same source, we can simply reuse the source. |
| 1506 |
// |
| 1507 |
// This is a common code pattern because of the way that scalar replacement |
| 1508 |
// works. |
| 1509 |
bool CompositeExtractFeedingConstruct( |
| 1510 |
IRContext* context, Instruction* inst, |
| 1511 |
const std::vector<const analysis::Constant*>&) { |
| 1512 |
assert(inst->opcode() == SpvOpCompositeConstruct && |
| 1513 |
"Wrong opcode. Should be OpCompositeConstruct."); |
| 1514 |
analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
| 1515 |
uint32_t original_id = 0; |
| 1516 |
|
| 1517 |
if (inst->NumInOperands() == 0) { |
| 1518 |
// The struct being constructed has no members. |
| 1519 |
return false; |
| 1520 |
} |
| 1521 |
|
| 1522 |
// Check each element to make sure they are: |
| 1523 |
// - extractions |
| 1524 |
// - extracting the same position they are inserting |
| 1525 |
// - all extract from the same id. |
| 1526 |
for (uint32_t i = 0; i < inst->NumInOperands(); ++i) { |
| 1527 |
const uint32_t element_id = inst->GetSingleWordInOperand(i); |
| 1528 |
Instruction* element_inst = def_use_mgr->GetDef(element_id); |
| 1529 |
|
| 1530 |
if (element_inst->opcode() != SpvOpCompositeExtract) { |
| 1531 |
return false; |
| 1532 |
} |
| 1533 |
|
| 1534 |
if (element_inst->NumInOperands() != 2) { |
| 1535 |
return false; |
| 1536 |
} |
| 1537 |
|
| 1538 |
if (element_inst->GetSingleWordInOperand(1) != i) { |
| 1539 |
return false; |
| 1540 |
} |
| 1541 |
|
| 1542 |
if (i == 0) { |
| 1543 |
original_id = |
| 1544 |
element_inst->GetSingleWordInOperand(kExtractCompositeIdInIdx); |
| 1545 |
} else if (original_id != |
| 1546 |
element_inst->GetSingleWordInOperand(kExtractCompositeIdInIdx)) { |
| 1547 |
return false; |
| 1548 |
} |
| 1549 |
} |
| 1550 |
|
| 1551 |
// The last check it to see that the object being extracted from is the |
| 1552 |
// correct type. |
| 1553 |
Instruction* original_inst = def_use_mgr->GetDef(original_id); |
| 1554 |
if (original_inst->type_id() != inst->type_id()) { |
| 1555 |
return false; |
| 1556 |
} |
| 1557 |
|
| 1558 |
// Simplify by using the original object. |
| 1559 |
inst->SetOpcode(SpvOpCopyObject); |
| 1560 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {original_id}}}); |
| 1561 |
return true; |
| 1562 |
} |
| 1563 |
|
| 1564 |
FoldingRule InsertFeedingExtract() { |
| 1565 |
return [](IRContext* context, Instruction* inst, |
| 1566 |
const std::vector<const analysis::Constant*>&) { |
| 1567 |
assert(inst->opcode() == SpvOpCompositeExtract && |
| 1568 |
"Wrong opcode. Should be OpCompositeExtract."); |
| 1569 |
analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
| 1570 |
uint32_t cid = inst->GetSingleWordInOperand(kExtractCompositeIdInIdx); |
| 1571 |
Instruction* cinst = def_use_mgr->GetDef(cid); |
| 1572 |
|
| 1573 |
if (cinst->opcode() != SpvOpCompositeInsert) { |
| 1574 |
return false; |
| 1575 |
} |
| 1576 |
|
| 1577 |
// Find the first position where the list of insert and extract indicies |
| 1578 |
// differ, if at all. |
| 1579 |
uint32_t i; |
| 1580 |
for (i = 1; i < inst->NumInOperands(); ++i) { |
| 1581 |
if (i + 1 >= cinst->NumInOperands()) { |
| 1582 |
break; |
| 1583 |
} |
| 1584 |
|
| 1585 |
if (inst->GetSingleWordInOperand(i) != |
| 1586 |
cinst->GetSingleWordInOperand(i + 1)) { |
| 1587 |
break; |
| 1588 |
} |
| 1589 |
} |
| 1590 |
|
| 1591 |
// We are extracting the element that was inserted. |
| 1592 |
if (i == inst->NumInOperands() && i + 1 == cinst->NumInOperands()) { |
| 1593 |
inst->SetOpcode(SpvOpCopyObject); |
| 1594 |
inst->SetInOperands( |
| 1595 |
{{SPV_OPERAND_TYPE_ID, |
| 1596 |
{cinst->GetSingleWordInOperand(kInsertObjectIdInIdx)}}}); |
| 1597 |
return true; |
| 1598 |
} |
| 1599 |
|
| 1600 |
// Extracting the value that was inserted along with values for the base |
| 1601 |
// composite. Cannot do anything. |
| 1602 |
if (i == inst->NumInOperands()) { |
| 1603 |
return false; |
| 1604 |
} |
| 1605 |
|
| 1606 |
// Extracting an element of the value that was inserted. Extract from |
| 1607 |
// that value directly. |
| 1608 |
if (i + 1 == cinst->NumInOperands()) { |
| 1609 |
std::vector<Operand> operands; |
| 1610 |
operands.push_back( |
| 1611 |
{SPV_OPERAND_TYPE_ID, |
| 1612 |
{cinst->GetSingleWordInOperand(kInsertObjectIdInIdx)}}); |
| 1613 |
for (; i < inst->NumInOperands(); ++i) { |
| 1614 |
operands.push_back({SPV_OPERAND_TYPE_LITERAL_INTEGER, |
| 1615 |
{inst->GetSingleWordInOperand(i)}}); |
| 1616 |
} |
| 1617 |
inst->SetInOperands(std::move(operands)); |
| 1618 |
return true; |
| 1619 |
} |
| 1620 |
|
| 1621 |
// Extracting a value that is disjoint from the element being inserted. |
| 1622 |
// Rewrite the extract to use the composite input to the insert. |
| 1623 |
std::vector<Operand> operands; |
| 1624 |
operands.push_back( |
| 1625 |
{SPV_OPERAND_TYPE_ID, |
| 1626 |
{cinst->GetSingleWordInOperand(kInsertCompositeIdInIdx)}}); |
| 1627 |
for (i = 1; i < inst->NumInOperands(); ++i) { |
| 1628 |
operands.push_back({SPV_OPERAND_TYPE_LITERAL_INTEGER, |
| 1629 |
{inst->GetSingleWordInOperand(i)}}); |
| 1630 |
} |
| 1631 |
inst->SetInOperands(std::move(operands)); |
| 1632 |
return true; |
| 1633 |
}; |
| 1634 |
} |
| 1635 |
|
| 1636 |
// When a VectorShuffle is feeding an Extract, we can extract from one of the |
| 1637 |
// operands of the VectorShuffle. We just need to adjust the index in the |
| 1638 |
// extract instruction. |
| 1639 |
FoldingRule VectorShuffleFeedingExtract() { |
| 1640 |
return [](IRContext* context, Instruction* inst, |
| 1641 |
const std::vector<const analysis::Constant*>&) { |
| 1642 |
assert(inst->opcode() == SpvOpCompositeExtract && |
| 1643 |
"Wrong opcode. Should be OpCompositeExtract."); |
| 1644 |
analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
| 1645 |
analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 1646 |
uint32_t cid = inst->GetSingleWordInOperand(kExtractCompositeIdInIdx); |
| 1647 |
Instruction* cinst = def_use_mgr->GetDef(cid); |
| 1648 |
|
| 1649 |
if (cinst->opcode() != SpvOpVectorShuffle) { |
| 1650 |
return false; |
| 1651 |
} |
| 1652 |
|
| 1653 |
// Find the size of the first vector operand of the VectorShuffle |
| 1654 |
Instruction* first_input = |
| 1655 |
def_use_mgr->GetDef(cinst->GetSingleWordInOperand(0)); |
| 1656 |
analysis::Type* first_input_type = |
| 1657 |
type_mgr->GetType(first_input->type_id()); |
| 1658 |
assert(first_input_type->AsVector() && |
| 1659 |
"Input to vector shuffle should be vectors."); |
| 1660 |
uint32_t first_input_size = first_input_type->AsVector()->element_count(); |
| 1661 |
|
| 1662 |
// Get index of the element the vector shuffle is placing in the position |
| 1663 |
// being extracted. |
| 1664 |
uint32_t new_index = |
| 1665 |
cinst->GetSingleWordInOperand(2 + inst->GetSingleWordInOperand(1)); |
| 1666 |
|
| 1667 |
// Extracting an undefined value so fold this extract into an undef. |
| 1668 |
const uint32_t undef_literal_value = 0xffffffff; |
| 1669 |
if (new_index == undef_literal_value) { |
| 1670 |
inst->SetOpcode(SpvOpUndef); |
| 1671 |
inst->SetInOperands({}); |
| 1672 |
return true; |
| 1673 |
} |
| 1674 |
|
| 1675 |
// Get the id of the of the vector the elemtent comes from, and update the |
| 1676 |
// index if needed. |
| 1677 |
uint32_t new_vector = 0; |
| 1678 |
if (new_index < first_input_size) { |
| 1679 |
new_vector = cinst->GetSingleWordInOperand(0); |
| 1680 |
} else { |
| 1681 |
new_vector = cinst->GetSingleWordInOperand(1); |
| 1682 |
new_index -= first_input_size; |
| 1683 |
} |
| 1684 |
|
| 1685 |
// Update the extract instruction. |
| 1686 |
inst->SetInOperand(kExtractCompositeIdInIdx, {new_vector}); |
| 1687 |
inst->SetInOperand(1, {new_index}); |
| 1688 |
return true; |
| 1689 |
}; |
| 1690 |
} |
| 1691 |
|
| 1692 |
// When an FMix with is feeding an Extract that extracts an element whose |
| 1693 |
// corresponding |a| in the FMix is 0 or 1, we can extract from one of the |
| 1694 |
// operands of the FMix. |
| 1695 |
FoldingRule FMixFeedingExtract() { |
| 1696 |
return [](IRContext* context, Instruction* inst, |
| 1697 |
const std::vector<const analysis::Constant*>&) { |
| 1698 |
assert(inst->opcode() == SpvOpCompositeExtract && |
| 1699 |
"Wrong opcode. Should be OpCompositeExtract."); |
| 1700 |
analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
| 1701 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 1702 |
|
| 1703 |
uint32_t composite_id = |
| 1704 |
inst->GetSingleWordInOperand(kExtractCompositeIdInIdx); |
| 1705 |
Instruction* composite_inst = def_use_mgr->GetDef(composite_id); |
| 1706 |
|
| 1707 |
if (composite_inst->opcode() != SpvOpExtInst) { |
| 1708 |
return false; |
| 1709 |
} |
| 1710 |
|
| 1711 |
uint32_t inst_set_id = |
| 1712 |
context->get_feature_mgr()->GetExtInstImportId_GLSLstd450(); |
| 1713 |
|
| 1714 |
if (composite_inst->GetSingleWordInOperand(kExtInstSetIdInIdx) != |
| 1715 |
inst_set_id || |
| 1716 |
composite_inst->GetSingleWordInOperand(kExtInstInstructionInIdx) != |
| 1717 |
GLSLstd450FMix) { |
| 1718 |
return false; |
| 1719 |
} |
| 1720 |
|
| 1721 |
// Get the |a| for the FMix instruction. |
| 1722 |
uint32_t a_id = composite_inst->GetSingleWordInOperand(kFMixAIdInIdx); |
| 1723 |
std::unique_ptr<Instruction> a(inst->Clone(context)); |
| 1724 |
a->SetInOperand(kExtractCompositeIdInIdx, {a_id}); |
| 1725 |
context->get_instruction_folder().FoldInstruction(a.get()); |
| 1726 |
|
| 1727 |
if (a->opcode() != SpvOpCopyObject) { |
| 1728 |
return false; |
| 1729 |
} |
| 1730 |
|
| 1731 |
const analysis::Constant* a_const = |
| 1732 |
const_mgr->FindDeclaredConstant(a->GetSingleWordInOperand(0)); |
| 1733 |
|
| 1734 |
if (!a_const) { |
| 1735 |
return false; |
| 1736 |
} |
| 1737 |
|
| 1738 |
bool use_x = false; |
| 1739 |
|
| 1740 |
assert(a_const->type()->AsFloat()); |
| 1741 |
double element_value = a_const->GetValueAsDouble(); |
| 1742 |
if (element_value == 0.0) { |
| 1743 |
use_x = true; |
| 1744 |
} else if (element_value == 1.0) { |
| 1745 |
use_x = false; |
| 1746 |
} else { |
| 1747 |
return false; |
| 1748 |
} |
| 1749 |
|
| 1750 |
// Get the id of the of the vector the element comes from. |
| 1751 |
uint32_t new_vector = 0; |
| 1752 |
if (use_x) { |
| 1753 |
new_vector = composite_inst->GetSingleWordInOperand(kFMixXIdInIdx); |
| 1754 |
} else { |
| 1755 |
new_vector = composite_inst->GetSingleWordInOperand(kFMixYIdInIdx); |
| 1756 |
} |
| 1757 |
|
| 1758 |
// Update the extract instruction. |
| 1759 |
inst->SetInOperand(kExtractCompositeIdInIdx, {new_vector}); |
| 1760 |
return true; |
| 1761 |
}; |
| 1762 |
} |
| 1763 |
|
| 1764 |
FoldingRule RedundantPhi() { |
| 1765 |
// An OpPhi instruction where all values are the same or the result of the phi |
| 1766 |
// itself, can be replaced by the value itself. |
| 1767 |
return [](IRContext*, Instruction* inst, |
| 1768 |
const std::vector<const analysis::Constant*>&) { |
| 1769 |
assert(inst->opcode() == SpvOpPhi && "Wrong opcode. Should be OpPhi."); |
| 1770 |
|
| 1771 |
uint32_t incoming_value = 0; |
| 1772 |
|
| 1773 |
for (uint32_t i = 0; i < inst->NumInOperands(); i += 2) { |
| 1774 |
uint32_t op_id = inst->GetSingleWordInOperand(i); |
| 1775 |
if (op_id == inst->result_id()) { |
| 1776 |
continue; |
| 1777 |
} |
| 1778 |
|
| 1779 |
if (incoming_value == 0) { |
| 1780 |
incoming_value = op_id; |
| 1781 |
} else if (op_id != incoming_value) { |
| 1782 |
// Found two possible value. Can't simplify. |
| 1783 |
return false; |
| 1784 |
} |
| 1785 |
} |
| 1786 |
|
| 1787 |
if (incoming_value == 0) { |
| 1788 |
// Code looks invalid. Don't do anything. |
| 1789 |
return false; |
| 1790 |
} |
| 1791 |
|
| 1792 |
// We have a single incoming value. Simplify using that value. |
| 1793 |
inst->SetOpcode(SpvOpCopyObject); |
| 1794 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {incoming_value}}}); |
| 1795 |
return true; |
| 1796 |
}; |
| 1797 |
} |
| 1798 |
|
| 1799 |
FoldingRule RedundantSelect() { |
| 1800 |
// An OpSelect instruction where both values are the same or the condition is |
| 1801 |
// constant can be replaced by one of the values |
| 1802 |
return [](IRContext*, Instruction* inst, |
| 1803 |
const std::vector<const analysis::Constant*>& constants) { |
| 1804 |
assert(inst->opcode() == SpvOpSelect && |
| 1805 |
"Wrong opcode. Should be OpSelect."); |
| 1806 |
assert(inst->NumInOperands() == 3); |
| 1807 |
assert(constants.size() == 3); |
| 1808 |
|
| 1809 |
uint32_t true_id = inst->GetSingleWordInOperand(1); |
| 1810 |
uint32_t false_id = inst->GetSingleWordInOperand(2); |
| 1811 |
|
| 1812 |
if (true_id == false_id) { |
| 1813 |
// Both results are the same, condition doesn't matter |
| 1814 |
inst->SetOpcode(SpvOpCopyObject); |
| 1815 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {true_id}}}); |
| 1816 |
return true; |
| 1817 |
} else if (constants[0]) { |
| 1818 |
const analysis::Type* type = constants[0]->type(); |
| 1819 |
if (type->AsBool()) { |
| 1820 |
// Scalar constant value, select the corresponding value. |
| 1821 |
inst->SetOpcode(SpvOpCopyObject); |
| 1822 |
if (constants[0]->AsNullConstant() || |
| 1823 |
!constants[0]->AsBoolConstant()->value()) { |
| 1824 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {false_id}}}); |
| 1825 |
} else { |
| 1826 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {true_id}}}); |
| 1827 |
} |
| 1828 |
return true; |
| 1829 |
} else { |
| 1830 |
assert(type->AsVector()); |
| 1831 |
if (constants[0]->AsNullConstant()) { |
| 1832 |
// All values come from false id. |
| 1833 |
inst->SetOpcode(SpvOpCopyObject); |
| 1834 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {false_id}}}); |
| 1835 |
return true; |
| 1836 |
} else { |
| 1837 |
// Convert to a vector shuffle. |
| 1838 |
std::vector<Operand> ops; |
| 1839 |
ops.push_back({SPV_OPERAND_TYPE_ID, {true_id}}); |
| 1840 |
ops.push_back({SPV_OPERAND_TYPE_ID, {false_id}}); |
| 1841 |
const analysis::VectorConstant* vector_const = |
| 1842 |
constants[0]->AsVectorConstant(); |
| 1843 |
uint32_t size = |
| 1844 |
static_cast<uint32_t>(vector_const->GetComponents().size()); |
| 1845 |
for (uint32_t i = 0; i != size; ++i) { |
| 1846 |
const analysis::Constant* component = |
| 1847 |
vector_const->GetComponents()[i]; |
| 1848 |
if (component->AsNullConstant() || |
| 1849 |
!component->AsBoolConstant()->value()) { |
| 1850 |
// Selecting from the false vector which is the second input |
| 1851 |
// vector to the shuffle. Offset the index by |size|. |
| 1852 |
ops.push_back({SPV_OPERAND_TYPE_LITERAL_INTEGER, {i + size}}); |
| 1853 |
} else { |
| 1854 |
// Selecting from true vector which is the first input vector to |
| 1855 |
// the shuffle. |
| 1856 |
ops.push_back({SPV_OPERAND_TYPE_LITERAL_INTEGER, {i}}); |
| 1857 |
} |
| 1858 |
} |
| 1859 |
|
| 1860 |
inst->SetOpcode(SpvOpVectorShuffle); |
| 1861 |
inst->SetInOperands(std::move(ops)); |
| 1862 |
return true; |
| 1863 |
} |
| 1864 |
} |
| 1865 |
} |
| 1866 |
|
| 1867 |
return false; |
| 1868 |
}; |
| 1869 |
} |
| 1870 |
|
| 1871 |
enum class FloatConstantKind { Unknown, Zero, One }; |
| 1872 |
|
| 1873 |
FloatConstantKind getFloatConstantKind(const analysis::Constant* constant) { |
| 1874 |
if (constant == nullptr) { |
| 1875 |
return FloatConstantKind::Unknown; |
| 1876 |
} |
| 1877 |
|
| 1878 |
assert(HasFloatingPoint(constant->type()) && "Unexpected constant type"); |
| 1879 |
|
| 1880 |
if (constant->AsNullConstant()) { |
| 1881 |
return FloatConstantKind::Zero; |
| 1882 |
} else if (const analysis::VectorConstant* vc = |
| 1883 |
constant->AsVectorConstant()) { |
| 1884 |
const std::vector<const analysis::Constant*>& components = |
| 1885 |
vc->GetComponents(); |
| 1886 |
assert(!components.empty()); |
| 1887 |
|
| 1888 |
FloatConstantKind kind = getFloatConstantKind(components[0]); |
| 1889 |
|
| 1890 |
for (size_t i = 1; i < components.size(); ++i) { |
| 1891 |
if (getFloatConstantKind(components[i]) != kind) { |
| 1892 |
return FloatConstantKind::Unknown; |
| 1893 |
} |
| 1894 |
} |
| 1895 |
|
| 1896 |
return kind; |
| 1897 |
} else if (const analysis::FloatConstant* fc = constant->AsFloatConstant()) { |
| 1898 |
if (fc->IsZero()) return FloatConstantKind::Zero; |
| 1899 |
|
| 1900 |
uint32_t width = fc->type()->AsFloat()->width(); |
| 1901 |
if (width != 32 && width != 64) return FloatConstantKind::Unknown; |
| 1902 |
|
| 1903 |
double value = (width == 64) ? fc->GetDoubleValue() : fc->GetFloatValue(); |
| 1904 |
|
| 1905 |
if (value == 0.0) { |
| 1906 |
return FloatConstantKind::Zero; |
| 1907 |
} else if (value == 1.0) { |
| 1908 |
return FloatConstantKind::One; |
| 1909 |
} else { |
| 1910 |
return FloatConstantKind::Unknown; |
| 1911 |
} |
| 1912 |
} else { |
| 1913 |
return FloatConstantKind::Unknown; |
| 1914 |
} |
| 1915 |
} |
| 1916 |
|
| 1917 |
FoldingRule RedundantFAdd() { |
| 1918 |
return [](IRContext*, Instruction* inst, |
| 1919 |
const std::vector<const analysis::Constant*>& constants) { |
| 1920 |
assert(inst->opcode() == SpvOpFAdd && "Wrong opcode. Should be OpFAdd."); |
| 1921 |
assert(constants.size() == 2); |
| 1922 |
|
| 1923 |
if (!inst->IsFloatingPointFoldingAllowed()) { |
| 1924 |
return false; |
| 1925 |
} |
| 1926 |
|
| 1927 |
FloatConstantKind kind0 = getFloatConstantKind(constants[0]); |
| 1928 |
FloatConstantKind kind1 = getFloatConstantKind(constants[1]); |
| 1929 |
|
| 1930 |
if (kind0 == FloatConstantKind::Zero || kind1 == FloatConstantKind::Zero) { |
| 1931 |
inst->SetOpcode(SpvOpCopyObject); |
| 1932 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, |
| 1933 |
{inst->GetSingleWordInOperand( |
| 1934 |
kind0 == FloatConstantKind::Zero ? 1 : 0)}}}); |
| 1935 |
return true; |
| 1936 |
} |
| 1937 |
|
| 1938 |
return false; |
| 1939 |
}; |
| 1940 |
} |
| 1941 |
|
| 1942 |
FoldingRule RedundantFSub() { |
| 1943 |
return [](IRContext*, Instruction* inst, |
| 1944 |
const std::vector<const analysis::Constant*>& constants) { |
| 1945 |
assert(inst->opcode() == SpvOpFSub && "Wrong opcode. Should be OpFSub."); |
| 1946 |
assert(constants.size() == 2); |
| 1947 |
|
| 1948 |
if (!inst->IsFloatingPointFoldingAllowed()) { |
| 1949 |
return false; |
| 1950 |
} |
| 1951 |
|
| 1952 |
FloatConstantKind kind0 = getFloatConstantKind(constants[0]); |
| 1953 |
FloatConstantKind kind1 = getFloatConstantKind(constants[1]); |
| 1954 |
|
| 1955 |
if (kind0 == FloatConstantKind::Zero) { |
| 1956 |
inst->SetOpcode(SpvOpFNegate); |
| 1957 |
inst->SetInOperands( |
| 1958 |
{{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(1)}}}); |
| 1959 |
return true; |
| 1960 |
} |
| 1961 |
|
| 1962 |
if (kind1 == FloatConstantKind::Zero) { |
| 1963 |
inst->SetOpcode(SpvOpCopyObject); |
| 1964 |
inst->SetInOperands( |
| 1965 |
{{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(0)}}}); |
| 1966 |
return true; |
| 1967 |
} |
| 1968 |
|
| 1969 |
return false; |
| 1970 |
}; |
| 1971 |
} |
| 1972 |
|
| 1973 |
FoldingRule RedundantFMul() { |
| 1974 |
return [](IRContext*, Instruction* inst, |
| 1975 |
const std::vector<const analysis::Constant*>& constants) { |
| 1976 |
assert(inst->opcode() == SpvOpFMul && "Wrong opcode. Should be OpFMul."); |
| 1977 |
assert(constants.size() == 2); |
| 1978 |
|
| 1979 |
if (!inst->IsFloatingPointFoldingAllowed()) { |
| 1980 |
return false; |
| 1981 |
} |
| 1982 |
|
| 1983 |
FloatConstantKind kind0 = getFloatConstantKind(constants[0]); |
| 1984 |
FloatConstantKind kind1 = getFloatConstantKind(constants[1]); |
| 1985 |
|
| 1986 |
if (kind0 == FloatConstantKind::Zero || kind1 == FloatConstantKind::Zero) { |
| 1987 |
inst->SetOpcode(SpvOpCopyObject); |
| 1988 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, |
| 1989 |
{inst->GetSingleWordInOperand( |
| 1990 |
kind0 == FloatConstantKind::Zero ? 0 : 1)}}}); |
| 1991 |
return true; |
| 1992 |
} |
| 1993 |
|
| 1994 |
if (kind0 == FloatConstantKind::One || kind1 == FloatConstantKind::One) { |
| 1995 |
inst->SetOpcode(SpvOpCopyObject); |
| 1996 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, |
| 1997 |
{inst->GetSingleWordInOperand( |
| 1998 |
kind0 == FloatConstantKind::One ? 1 : 0)}}}); |
| 1999 |
return true; |
| 2000 |
} |
| 2001 |
|
| 2002 |
return false; |
| 2003 |
}; |
| 2004 |
} |
| 2005 |
|
| 2006 |
FoldingRule RedundantFDiv() { |
| 2007 |
return [](IRContext*, Instruction* inst, |
| 2008 |
const std::vector<const analysis::Constant*>& constants) { |
| 2009 |
assert(inst->opcode() == SpvOpFDiv && "Wrong opcode. Should be OpFDiv."); |
| 2010 |
assert(constants.size() == 2); |
| 2011 |
|
| 2012 |
if (!inst->IsFloatingPointFoldingAllowed()) { |
| 2013 |
return false; |
| 2014 |
} |
| 2015 |
|
| 2016 |
FloatConstantKind kind0 = getFloatConstantKind(constants[0]); |
| 2017 |
FloatConstantKind kind1 = getFloatConstantKind(constants[1]); |
| 2018 |
|
| 2019 |
if (kind0 == FloatConstantKind::Zero) { |
| 2020 |
inst->SetOpcode(SpvOpCopyObject); |
| 2021 |
inst->SetInOperands( |
| 2022 |
{{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(0)}}}); |
| 2023 |
return true; |
| 2024 |
} |
| 2025 |
|
| 2026 |
if (kind1 == FloatConstantKind::One) { |
| 2027 |
inst->SetOpcode(SpvOpCopyObject); |
| 2028 |
inst->SetInOperands( |
| 2029 |
{{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(0)}}}); |
| 2030 |
return true; |
| 2031 |
} |
| 2032 |
|
| 2033 |
return false; |
| 2034 |
}; |
| 2035 |
} |
| 2036 |
|
| 2037 |
FoldingRule RedundantFMix() { |
| 2038 |
return [](IRContext* context, Instruction* inst, |
| 2039 |
const std::vector<const analysis::Constant*>& constants) { |
| 2040 |
assert(inst->opcode() == SpvOpExtInst && |
| 2041 |
"Wrong opcode. Should be OpExtInst."); |
| 2042 |
|
| 2043 |
if (!inst->IsFloatingPointFoldingAllowed()) { |
| 2044 |
return false; |
| 2045 |
} |
| 2046 |
|
| 2047 |
uint32_t instSetId = |
| 2048 |
context->get_feature_mgr()->GetExtInstImportId_GLSLstd450(); |
| 2049 |
|
| 2050 |
if (inst->GetSingleWordInOperand(kExtInstSetIdInIdx) == instSetId && |
| 2051 |
inst->GetSingleWordInOperand(kExtInstInstructionInIdx) == |
| 2052 |
GLSLstd450FMix) { |
| 2053 |
assert(constants.size() == 5); |
| 2054 |
|
| 2055 |
FloatConstantKind kind4 = getFloatConstantKind(constants[4]); |
| 2056 |
|
| 2057 |
if (kind4 == FloatConstantKind::Zero || kind4 == FloatConstantKind::One) { |
| 2058 |
inst->SetOpcode(SpvOpCopyObject); |
| 2059 |
inst->SetInOperands( |
| 2060 |
{{SPV_OPERAND_TYPE_ID, |
| 2061 |
{inst->GetSingleWordInOperand(kind4 == FloatConstantKind::Zero |
| 2062 |
? kFMixXIdInIdx |
| 2063 |
: kFMixYIdInIdx)}}}); |
| 2064 |
return true; |
| 2065 |
} |
| 2066 |
} |
| 2067 |
|
| 2068 |
return false; |
| 2069 |
}; |
| 2070 |
} |
| 2071 |
|
| 2072 |
// This rule handles addition of zero for integers. |
| 2073 |
FoldingRule RedundantIAdd() { |
| 2074 |
return [](IRContext* context, Instruction* inst, |
| 2075 |
const std::vector<const analysis::Constant*>& constants) { |
| 2076 |
assert(inst->opcode() == SpvOpIAdd && "Wrong opcode. Should be OpIAdd."); |
| 2077 |
|
| 2078 |
uint32_t operand = std::numeric_limits<uint32_t>::max(); |
| 2079 |
const analysis::Type* operand_type = nullptr; |
| 2080 |
if (constants[0] && constants[0]->IsZero()) { |
| 2081 |
operand = inst->GetSingleWordInOperand(1); |
| 2082 |
operand_type = constants[0]->type(); |
| 2083 |
} else if (constants[1] && constants[1]->IsZero()) { |
| 2084 |
operand = inst->GetSingleWordInOperand(0); |
| 2085 |
operand_type = constants[1]->type(); |
| 2086 |
} |
| 2087 |
|
| 2088 |
if (operand != std::numeric_limits<uint32_t>::max()) { |
| 2089 |
const analysis::Type* inst_type = |
| 2090 |
context->get_type_mgr()->GetType(inst->type_id()); |
| 2091 |
if (inst_type->IsSame(operand_type)) { |
| 2092 |
inst->SetOpcode(SpvOpCopyObject); |
| 2093 |
} else { |
| 2094 |
inst->SetOpcode(SpvOpBitcast); |
| 2095 |
} |
| 2096 |
inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {operand}}}); |
| 2097 |
return true; |
| 2098 |
} |
| 2099 |
return false; |
| 2100 |
}; |
| 2101 |
} |
| 2102 |
|
| 2103 |
// This rule look for a dot with a constant vector containing a single 1 and |
| 2104 |
// the rest 0s. This is the same as doing an extract. |
| 2105 |
FoldingRule DotProductDoingExtract() { |
| 2106 |
return [](IRContext* context, Instruction* inst, |
| 2107 |
const std::vector<const analysis::Constant*>& constants) { |
| 2108 |
assert(inst->opcode() == SpvOpDot && "Wrong opcode. Should be OpDot."); |
| 2109 |
|
| 2110 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 2111 |
|
| 2112 |
if (!inst->IsFloatingPointFoldingAllowed()) { |
| 2113 |
return false; |
| 2114 |
} |
| 2115 |
|
| 2116 |
for (int i = 0; i < 2; ++i) { |
| 2117 |
if (!constants[i]) { |
| 2118 |
continue; |
| 2119 |
} |
| 2120 |
|
| 2121 |
const analysis::Vector* vector_type = constants[i]->type()->AsVector(); |
| 2122 |
assert(vector_type && "Inputs to OpDot must be vectors."); |
| 2123 |
const analysis::Float* element_type = |
| 2124 |
vector_type->element_type()->AsFloat(); |
| 2125 |
assert(element_type && "Inputs to OpDot must be vectors of floats."); |
| 2126 |
uint32_t element_width = element_type->width(); |
| 2127 |
if (element_width != 32 && element_width != 64) { |
| 2128 |
return false; |
| 2129 |
} |
| 2130 |
|
| 2131 |
std::vector<const analysis::Constant*> components; |
| 2132 |
components = constants[i]->GetVectorComponents(const_mgr); |
| 2133 |
|
| 2134 |
const uint32_t kNotFound = std::numeric_limits<uint32_t>::max(); |
| 2135 |
|
| 2136 |
uint32_t component_with_one = kNotFound; |
| 2137 |
bool all_others_zero = true; |
| 2138 |
for (uint32_t j = 0; j < components.size(); ++j) { |
| 2139 |
const analysis::Constant* element = components[j]; |
| 2140 |
double value = |
| 2141 |
(element_width == 32 ? element->GetFloat() : element->GetDouble()); |
| 2142 |
if (value == 0.0) { |
| 2143 |
continue; |
| 2144 |
} else if (value == 1.0) { |
| 2145 |
if (component_with_one == kNotFound) { |
| 2146 |
component_with_one = j; |
| 2147 |
} else { |
| 2148 |
component_with_one = kNotFound; |
| 2149 |
break; |
| 2150 |
} |
| 2151 |
} else { |
| 2152 |
all_others_zero = false; |
| 2153 |
break; |
| 2154 |
} |
| 2155 |
} |
| 2156 |
|
| 2157 |
if (!all_others_zero || component_with_one == kNotFound) { |
| 2158 |
continue; |
| 2159 |
} |
| 2160 |
|
| 2161 |
std::vector<Operand> operands; |
| 2162 |
operands.push_back( |
| 2163 |
{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(1u - i)}}); |
| 2164 |
operands.push_back( |
| 2165 |
{SPV_OPERAND_TYPE_LITERAL_INTEGER, {component_with_one}}); |
| 2166 |
|
| 2167 |
inst->SetOpcode(SpvOpCompositeExtract); |
| 2168 |
inst->SetInOperands(std::move(operands)); |
| 2169 |
return true; |
| 2170 |
} |
| 2171 |
return false; |
| 2172 |
}; |
| 2173 |
} |
| 2174 |
|
| 2175 |
// If we are storing an undef, then we can remove the store. |
| 2176 |
// |
| 2177 |
// TODO: We can do something similar for OpImageWrite, but checking for volatile |
| 2178 |
// is complicated. Waiting to see if it is needed. |
| 2179 |
FoldingRule StoringUndef() { |
| 2180 |
return [](IRContext* context, Instruction* inst, |
| 2181 |
const std::vector<const analysis::Constant*>&) { |
| 2182 |
assert(inst->opcode() == SpvOpStore && "Wrong opcode. Should be OpStore."); |
| 2183 |
|
| 2184 |
analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
| 2185 |
|
| 2186 |
// If this is a volatile store, the store cannot be removed. |
| 2187 |
if (inst->NumInOperands() == 3) { |
| 2188 |
if (inst->GetSingleWordInOperand(2) & SpvMemoryAccessVolatileMask) { |
| 2189 |
return false; |
| 2190 |
} |
| 2191 |
} |
| 2192 |
|
| 2193 |
uint32_t object_id = inst->GetSingleWordInOperand(kStoreObjectInIdx); |
| 2194 |
Instruction* object_inst = def_use_mgr->GetDef(object_id); |
| 2195 |
if (object_inst->opcode() == SpvOpUndef) { |
| 2196 |
inst->ToNop(); |
| 2197 |
return true; |
| 2198 |
} |
| 2199 |
return false; |
| 2200 |
}; |
| 2201 |
} |
| 2202 |
|
| 2203 |
FoldingRule VectorShuffleFeedingShuffle() { |
| 2204 |
return [](IRContext* context, Instruction* inst, |
| 2205 |
const std::vector<const analysis::Constant*>&) { |
| 2206 |
assert(inst->opcode() == SpvOpVectorShuffle && |
| 2207 |
"Wrong opcode. Should be OpVectorShuffle."); |
| 2208 |
|
| 2209 |
analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
| 2210 |
analysis::TypeManager* type_mgr = context->get_type_mgr(); |
| 2211 |
|
| 2212 |
Instruction* feeding_shuffle_inst = |
| 2213 |
def_use_mgr->GetDef(inst->GetSingleWordInOperand(0)); |
| 2214 |
analysis::Vector* op0_type = |
| 2215 |
type_mgr->GetType(feeding_shuffle_inst->type_id())->AsVector(); |
| 2216 |
uint32_t op0_length = op0_type->element_count(); |
| 2217 |
|
| 2218 |
bool feeder_is_op0 = true; |
| 2219 |
if (feeding_shuffle_inst->opcode() != SpvOpVectorShuffle) { |
| 2220 |
feeding_shuffle_inst = |
| 2221 |
def_use_mgr->GetDef(inst->GetSingleWordInOperand(1)); |
| 2222 |
feeder_is_op0 = false; |
| 2223 |
} |
| 2224 |
|
| 2225 |
if (feeding_shuffle_inst->opcode() != SpvOpVectorShuffle) { |
| 2226 |
return false; |
| 2227 |
} |
| 2228 |
|
| 2229 |
Instruction* feeder2 = |
| 2230 |
def_use_mgr->GetDef(feeding_shuffle_inst->GetSingleWordInOperand(0)); |
| 2231 |
analysis::Vector* feeder_op0_type = |
| 2232 |
type_mgr->GetType(feeder2->type_id())->AsVector(); |
| 2233 |
uint32_t feeder_op0_length = feeder_op0_type->element_count(); |
| 2234 |
|
| 2235 |
uint32_t new_feeder_id = 0; |
| 2236 |
std::vector<Operand> new_operands; |
| 2237 |
new_operands.resize( |
| 2238 |
2, {SPV_OPERAND_TYPE_ID, {0}}); // Place holders for vector operands. |
| 2239 |
const uint32_t undef_literal = 0xffffffff; |
| 2240 |
for (uint32_t op = 2; op < inst->NumInOperands(); ++op) { |
| 2241 |
uint32_t component_index = inst->GetSingleWordInOperand(op); |
| 2242 |
|
| 2243 |
// Do not interpret the undefined value literal as coming from operand 1. |
| 2244 |
if (component_index != undef_literal && |
| 2245 |
feeder_is_op0 == (component_index < op0_length)) { |
| 2246 |
// This component comes from the feeding_shuffle_inst. Update |
| 2247 |
// |component_index| to be the index into the operand of the feeder. |
| 2248 |
|
| 2249 |
// Adjust component_index to get the index into the operands of the |
| 2250 |
// feeding_shuffle_inst. |
| 2251 |
if (component_index >= op0_length) { |
| 2252 |
component_index -= op0_length; |
| 2253 |
} |
| 2254 |
component_index = |
| 2255 |
feeding_shuffle_inst->GetSingleWordInOperand(component_index + 2); |
| 2256 |
|
| 2257 |
// Check if we are using a component from the first or second operand of |
| 2258 |
// the feeding instruction. |
| 2259 |
if (component_index < feeder_op0_length) { |
| 2260 |
if (new_feeder_id == 0) { |
| 2261 |
// First time through, save the id of the operand the element comes |
| 2262 |
// from. |
| 2263 |
new_feeder_id = feeding_shuffle_inst->GetSingleWordInOperand(0); |
| 2264 |
} else if (new_feeder_id != |
| 2265 |
feeding_shuffle_inst->GetSingleWordInOperand(0)) { |
| 2266 |
// We need both elements of the feeding_shuffle_inst, so we cannot |
| 2267 |
// fold. |
| 2268 |
return false; |
| 2269 |
} |
| 2270 |
} else { |
| 2271 |
if (new_feeder_id == 0) { |
| 2272 |
// First time through, save the id of the operand the element comes |
| 2273 |
// from. |
| 2274 |
new_feeder_id = feeding_shuffle_inst->GetSingleWordInOperand(1); |
| 2275 |
} else if (new_feeder_id != |
| 2276 |
feeding_shuffle_inst->GetSingleWordInOperand(1)) { |
| 2277 |
// We need both elements of the feeding_shuffle_inst, so we cannot |
| 2278 |
// fold. |
| 2279 |
return false; |
| 2280 |
} |
| 2281 |
component_index -= feeder_op0_length; |
| 2282 |
} |
| 2283 |
|
| 2284 |
if (!feeder_is_op0) { |
| 2285 |
component_index += op0_length; |
| 2286 |
} |
| 2287 |
} |
| 2288 |
new_operands.push_back( |
| 2289 |
{SPV_OPERAND_TYPE_LITERAL_INTEGER, {component_index}}); |
| 2290 |
} |
| 2291 |
|
| 2292 |
if (new_feeder_id == 0) { |
| 2293 |
analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
| 2294 |
const analysis::Type* type = |
| 2295 |
type_mgr->GetType(feeding_shuffle_inst->type_id()); |
| 2296 |
const analysis::Constant* null_const = const_mgr->GetConstant(type, {}); |
| 2297 |
new_feeder_id = |
| 2298 |
const_mgr->GetDefiningInstruction(null_const, 0)->result_id(); |
| 2299 |
} |
| 2300 |
|
| 2301 |
if (feeder_is_op0) { |
| 2302 |
// If the size of the first vector operand changed then the indices |
| 2303 |
// referring to the second operand need to be adjusted. |
| 2304 |
Instruction* new_feeder_inst = def_use_mgr->GetDef(new_feeder_id); |
| 2305 |
analysis::Type* new_feeder_type = |
| 2306 |
type_mgr->GetType(new_feeder_inst->type_id()); |
| 2307 |
uint32_t new_op0_size = new_feeder_type->AsVector()->element_count(); |
| 2308 |
int32_t adjustment = op0_length - new_op0_size; |
| 2309 |
|
| 2310 |
if (adjustment != 0) { |
| 2311 |
for (uint32_t i = 2; i < new_operands.size(); i++) { |
| 2312 |
if (inst->GetSingleWordInOperand(i) >= op0_length) { |
| 2313 |
new_operands[i].words[0] -= adjustment; |
| 2314 |
} |
| 2315 |
} |
| 2316 |
} |
| 2317 |
|
| 2318 |
new_operands[0].words[0] = new_feeder_id; |
| 2319 |
new_operands[1] = inst->GetInOperand(1); |
| 2320 |
} else { |
| 2321 |
new_operands[1].words[0] = new_feeder_id; |
| 2322 |
new_operands[0] = inst->GetInOperand(0); |
| 2323 |
} |
| 2324 |
|
| 2325 |
inst->SetInOperands(std::move(new_operands)); |
| 2326 |
return true; |
| 2327 |
}; |
| 2328 |
} |
| 2329 |
|
| 2330 |
// Removes duplicate ids from the interface list of an OpEntryPoint |
| 2331 |
// instruction. |
| 2332 |
FoldingRule RemoveRedundantOperands() { |
| 2333 |
return [](IRContext*, Instruction* inst, |
| 2334 |
const std::vector<const analysis::Constant*>&) { |
| 2335 |
assert(inst->opcode() == SpvOpEntryPoint && |
| 2336 |
"Wrong opcode. Should be OpEntryPoint."); |
| 2337 |
bool has_redundant_operand = false; |
| 2338 |
std::unordered_set<uint32_t> seen_operands; |
| 2339 |
std::vector<Operand> new_operands; |
| 2340 |
|
| 2341 |
new_operands.emplace_back(inst->GetOperand(0)); |
| 2342 |
new_operands.emplace_back(inst->GetOperand(1)); |
| 2343 |
new_operands.emplace_back(inst->GetOperand(2)); |
| 2344 |
for (uint32_t i = 3; i < inst->NumOperands(); ++i) { |
| 2345 |
if (seen_operands.insert(inst->GetSingleWordOperand(i)).second) { |
| 2346 |
new_operands.emplace_back(inst->GetOperand(i)); |
| 2347 |
} else { |
| 2348 |
has_redundant_operand = true; |
| 2349 |
} |
| 2350 |
} |
| 2351 |
|
| 2352 |
if (!has_redundant_operand) { |
| 2353 |
return false; |
| 2354 |
} |
| 2355 |
|
| 2356 |
inst->SetInOperands(std::move(new_operands)); |
| 2357 |
return true; |
| 2358 |
}; |
| 2359 |
} |
| 2360 |
|
| 2361 |
// If an image instruction's operand is a constant, updates the image operand |
| 2362 |
// flag from Offset to ConstOffset. |
| 2363 |
FoldingRule UpdateImageOperands() { |
| 2364 |
return [](IRContext*, Instruction* inst, |
| 2365 |
const std::vector<const analysis::Constant*>& constants) { |
| 2366 |
const auto opcode = inst->opcode(); |
| 2367 |
(void)opcode; |
| 2368 |
assert((opcode == SpvOpImageSampleImplicitLod || |
| 2369 |
opcode == SpvOpImageSampleExplicitLod || |
| 2370 |
opcode == SpvOpImageSampleDrefImplicitLod || |
| 2371 |
opcode == SpvOpImageSampleDrefExplicitLod || |
| 2372 |
opcode == SpvOpImageSampleProjImplicitLod || |
| 2373 |
opcode == SpvOpImageSampleProjExplicitLod || |
| 2374 |
opcode == SpvOpImageSampleProjDrefImplicitLod || |
| 2375 |
opcode == SpvOpImageSampleProjDrefExplicitLod || |
| 2376 |
opcode == SpvOpImageFetch || opcode == SpvOpImageGather || |
| 2377 |
opcode == SpvOpImageDrefGather || opcode == SpvOpImageRead || |
| 2378 |
opcode == SpvOpImageWrite || |
| 2379 |
opcode == SpvOpImageSparseSampleImplicitLod || |
| 2380 |
opcode == SpvOpImageSparseSampleExplicitLod || |
| 2381 |
opcode == SpvOpImageSparseSampleDrefImplicitLod || |
| 2382 |
opcode == SpvOpImageSparseSampleDrefExplicitLod || |
| 2383 |
opcode == SpvOpImageSparseSampleProjImplicitLod || |
| 2384 |
opcode == SpvOpImageSparseSampleProjExplicitLod || |
| 2385 |
opcode == SpvOpImageSparseSampleProjDrefImplicitLod || |
| 2386 |
opcode == SpvOpImageSparseSampleProjDrefExplicitLod || |
| 2387 |
opcode == SpvOpImageSparseFetch || |
| 2388 |
opcode == SpvOpImageSparseGather || |
| 2389 |
opcode == SpvOpImageSparseDrefGather || |
| 2390 |
opcode == SpvOpImageSparseRead) && |
| 2391 |
"Wrong opcode. Should be an image instruction."); |
| 2392 |
|
| 2393 |
int32_t operand_index = ImageOperandsMaskInOperandIndex(inst); |
| 2394 |
if (operand_index >= 0) { |
| 2395 |
auto image_operands = inst->GetSingleWordInOperand(operand_index); |
| 2396 |
if (image_operands & SpvImageOperandsOffsetMask) { |
| 2397 |
uint32_t offset_operand_index = operand_index + 1; |
| 2398 |
if (image_operands & SpvImageOperandsBiasMask) offset_operand_index++; |
| 2399 |
if (image_operands & SpvImageOperandsLodMask) offset_operand_index++; |
| 2400 |
if (image_operands & SpvImageOperandsGradMask) |
| 2401 |
offset_operand_index += 2; |
| 2402 |
assert(((image_operands & SpvImageOperandsConstOffsetMask) == 0) && |
| 2403 |
"Offset and ConstOffset may not be used together"); |
| 2404 |
if (offset_operand_index < inst->NumOperands()) { |
| 2405 |
if (constants[offset_operand_index]) { |
| 2406 |
image_operands = image_operands | SpvImageOperandsConstOffsetMask; |
| 2407 |
image_operands = image_operands & ~SpvImageOperandsOffsetMask; |
| 2408 |
inst->SetInOperand(operand_index, {image_operands}); |
| 2409 |
return true; |
| 2410 |
} |
| 2411 |
} |
| 2412 |
} |
| 2413 |
} |
| 2414 |
|
| 2415 |
return false; |
| 2416 |
}; |
| 2417 |
} |
| 2418 |
|
| 2419 |
} // namespace |
| 2420 |
|
| 2421 |
void FoldingRules::AddFoldingRules() { |
| 2422 |
// Add all folding rules to the list for the opcodes to which they apply. |
| 2423 |
// Note that the order in which rules are added to the list matters. If a rule |
| 2424 |
// applies to the instruction, the rest of the rules will not be attempted. |
| 2425 |
// Take that into consideration. |
| 2426 |
rules_[SpvOpCompositeConstruct].push_back(CompositeExtractFeedingConstruct); |
| 2427 |
|
| 2428 |
rules_[SpvOpCompositeExtract].push_back(InsertFeedingExtract()); |
| 2429 |
rules_[SpvOpCompositeExtract].push_back(CompositeConstructFeedingExtract()); |
| 2430 |
rules_[SpvOpCompositeExtract].push_back(VectorShuffleFeedingExtract()); |
| 2431 |
rules_[SpvOpCompositeExtract].push_back(FMixFeedingExtract()); |
| 2432 |
|
| 2433 |
rules_[SpvOpDot].push_back(DotProductDoingExtract()); |
| 2434 |
|
| 2435 |
rules_[SpvOpEntryPoint].push_back(RemoveRedundantOperands()); |
| 2436 |
|
| 2437 |
rules_[SpvOpFAdd].push_back(RedundantFAdd()); |
| 2438 |
rules_[SpvOpFAdd].push_back(MergeAddNegateArithmetic()); |
| 2439 |
rules_[SpvOpFAdd].push_back(MergeAddAddArithmetic()); |
| 2440 |
rules_[SpvOpFAdd].push_back(MergeAddSubArithmetic()); |
| 2441 |
rules_[SpvOpFAdd].push_back(MergeGenericAddSubArithmetic()); |
| 2442 |
rules_[SpvOpFAdd].push_back(FactorAddMuls()); |
| 2443 |
|
| 2444 |
rules_[SpvOpFDiv].push_back(RedundantFDiv()); |
| 2445 |
rules_[SpvOpFDiv].push_back(ReciprocalFDiv()); |
| 2446 |
rules_[SpvOpFDiv].push_back(MergeDivDivArithmetic()); |
| 2447 |
rules_[SpvOpFDiv].push_back(MergeDivMulArithmetic()); |
| 2448 |
rules_[SpvOpFDiv].push_back(MergeDivNegateArithmetic()); |
| 2449 |
|
| 2450 |
rules_[SpvOpFMul].push_back(RedundantFMul()); |
| 2451 |
rules_[SpvOpFMul].push_back(MergeMulMulArithmetic()); |
| 2452 |
rules_[SpvOpFMul].push_back(MergeMulDivArithmetic()); |
| 2453 |
rules_[SpvOpFMul].push_back(MergeMulNegateArithmetic()); |
| 2454 |
|
| 2455 |
rules_[SpvOpFNegate].push_back(MergeNegateArithmetic()); |
| 2456 |
rules_[SpvOpFNegate].push_back(MergeNegateAddSubArithmetic()); |
| 2457 |
rules_[SpvOpFNegate].push_back(MergeNegateMulDivArithmetic()); |
| 2458 |
|
| 2459 |
rules_[SpvOpFSub].push_back(RedundantFSub()); |
| 2460 |
rules_[SpvOpFSub].push_back(MergeSubNegateArithmetic()); |
| 2461 |
rules_[SpvOpFSub].push_back(MergeSubAddArithmetic()); |
| 2462 |
rules_[SpvOpFSub].push_back(MergeSubSubArithmetic()); |
| 2463 |
|
| 2464 |
rules_[SpvOpIAdd].push_back(RedundantIAdd()); |
| 2465 |
rules_[SpvOpIAdd].push_back(MergeAddNegateArithmetic()); |
| 2466 |
rules_[SpvOpIAdd].push_back(MergeAddAddArithmetic()); |
| 2467 |
rules_[SpvOpIAdd].push_back(MergeAddSubArithmetic()); |
| 2468 |
rules_[SpvOpIAdd].push_back(MergeGenericAddSubArithmetic()); |
| 2469 |
rules_[SpvOpIAdd].push_back(FactorAddMuls()); |
| 2470 |
|
| 2471 |
rules_[SpvOpIMul].push_back(IntMultipleBy1()); |
| 2472 |
rules_[SpvOpIMul].push_back(MergeMulMulArithmetic()); |
| 2473 |
rules_[SpvOpIMul].push_back(MergeMulNegateArithmetic()); |
| 2474 |
|
| 2475 |
rules_[SpvOpISub].push_back(MergeSubNegateArithmetic()); |
| 2476 |
rules_[SpvOpISub].push_back(MergeSubAddArithmetic()); |
| 2477 |
rules_[SpvOpISub].push_back(MergeSubSubArithmetic()); |
| 2478 |
|
| 2479 |
rules_[SpvOpPhi].push_back(RedundantPhi()); |
| 2480 |
|
| 2481 |
rules_[SpvOpSDiv].push_back(MergeDivNegateArithmetic()); |
| 2482 |
|
| 2483 |
rules_[SpvOpSNegate].push_back(MergeNegateArithmetic()); |
| 2484 |
rules_[SpvOpSNegate].push_back(MergeNegateMulDivArithmetic()); |
| 2485 |
rules_[SpvOpSNegate].push_back(MergeNegateAddSubArithmetic()); |
| 2486 |
|
| 2487 |
rules_[SpvOpSelect].push_back(RedundantSelect()); |
| 2488 |
|
| 2489 |
rules_[SpvOpStore].push_back(StoringUndef()); |
| 2490 |
|
| 2491 |
rules_[SpvOpUDiv].push_back(MergeDivNegateArithmetic()); |
| 2492 |
|
| 2493 |
rules_[SpvOpVectorShuffle].push_back(VectorShuffleFeedingShuffle()); |
| 2494 |
|
| 2495 |
rules_[SpvOpImageSampleImplicitLod].push_back(UpdateImageOperands()); |
| 2496 |
rules_[SpvOpImageSampleExplicitLod].push_back(UpdateImageOperands()); |
| 2497 |
rules_[SpvOpImageSampleDrefImplicitLod].push_back(UpdateImageOperands()); |
| 2498 |
rules_[SpvOpImageSampleDrefExplicitLod].push_back(UpdateImageOperands()); |
| 2499 |
rules_[SpvOpImageSampleProjImplicitLod].push_back(UpdateImageOperands()); |
| 2500 |
rules_[SpvOpImageSampleProjExplicitLod].push_back(UpdateImageOperands()); |
| 2501 |
rules_[SpvOpImageSampleProjDrefImplicitLod].push_back(UpdateImageOperands()); |
| 2502 |
rules_[SpvOpImageSampleProjDrefExplicitLod].push_back(UpdateImageOperands()); |
| 2503 |
rules_[SpvOpImageFetch].push_back(UpdateImageOperands()); |
| 2504 |
rules_[SpvOpImageGather].push_back(UpdateImageOperands()); |
| 2505 |
rules_[SpvOpImageDrefGather].push_back(UpdateImageOperands()); |
| 2506 |
rules_[SpvOpImageRead].push_back(UpdateImageOperands()); |
| 2507 |
rules_[SpvOpImageWrite].push_back(UpdateImageOperands()); |
| 2508 |
rules_[SpvOpImageSparseSampleImplicitLod].push_back(UpdateImageOperands()); |
| 2509 |
rules_[SpvOpImageSparseSampleExplicitLod].push_back(UpdateImageOperands()); |
| 2510 |
rules_[SpvOpImageSparseSampleDrefImplicitLod].push_back( |
| 2511 |
UpdateImageOperands()); |
| 2512 |
rules_[SpvOpImageSparseSampleDrefExplicitLod].push_back( |
| 2513 |
UpdateImageOperands()); |
| 2514 |
rules_[SpvOpImageSparseSampleProjImplicitLod].push_back( |
| 2515 |
UpdateImageOperands()); |
| 2516 |
rules_[SpvOpImageSparseSampleProjExplicitLod].push_back( |
| 2517 |
UpdateImageOperands()); |
| 2518 |
rules_[SpvOpImageSparseSampleProjDrefImplicitLod].push_back( |
| 2519 |
UpdateImageOperands()); |
| 2520 |
rules_[SpvOpImageSparseSampleProjDrefExplicitLod].push_back( |
| 2521 |
UpdateImageOperands()); |
| 2522 |
rules_[SpvOpImageSparseFetch].push_back(UpdateImageOperands()); |
| 2523 |
rules_[SpvOpImageSparseGather].push_back(UpdateImageOperands()); |
| 2524 |
rules_[SpvOpImageSparseDrefGather].push_back(UpdateImageOperands()); |
| 2525 |
rules_[SpvOpImageSparseRead].push_back(UpdateImageOperands()); |
| 2526 |
|
| 2527 |
FeatureManager* feature_manager = context_->get_feature_mgr(); |
| 2528 |
// Add rules for GLSLstd450 |
| 2529 |
uint32_t ext_inst_glslstd450_id = |
| 2530 |
feature_manager->GetExtInstImportId_GLSLstd450(); |
| 2531 |
if (ext_inst_glslstd450_id != 0) { |
| 2532 |
ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FMix}].push_back( |
| 2533 |
RedundantFMix()); |
| 2534 |
} |
| 2535 |
} |
| 2536 |
} // namespace opt |
| 2537 |
} // namespace spvtools |