/src/spirv-tools/source/opt/folding_rules.cpp
Line | Count | Source |
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 <optional> |
18 | | #include <utility> |
19 | | |
20 | | #include "ir_builder.h" |
21 | | #include "source/latest_version_glsl_std_450_header.h" |
22 | | #include "source/opt/ir_context.h" |
23 | | #include "source/spirv_constant.h" |
24 | | #include "source/spirv_target_env.h" |
25 | | |
26 | | namespace spvtools { |
27 | | namespace opt { |
28 | | namespace { |
29 | | |
30 | | constexpr uint32_t kExtractCompositeIdInIdx = 0; |
31 | | constexpr uint32_t kInsertObjectIdInIdx = 0; |
32 | | constexpr uint32_t kInsertCompositeIdInIdx = 1; |
33 | | constexpr uint32_t kExtInstSetIdInIdx = 0; |
34 | | constexpr uint32_t kExtInstInstructionInIdx = 1; |
35 | | constexpr uint32_t kFMixXIdInIdx = 2; |
36 | | constexpr uint32_t kFMixYIdInIdx = 3; |
37 | | constexpr uint32_t kFMixAIdInIdx = 4; |
38 | | constexpr uint32_t kStoreObjectInIdx = 1; |
39 | | |
40 | | // Some image instructions may contain an "image operands" argument. |
41 | | // Returns the operand index for the "image operands". |
42 | | // Returns -1 if the instruction does not have image operands. |
43 | 237k | int32_t ImageOperandsMaskInOperandIndex(Instruction* inst) { |
44 | 237k | const auto opcode = inst->opcode(); |
45 | 237k | switch (opcode) { |
46 | 237k | case spv::Op::OpImageSampleImplicitLod: |
47 | 237k | case spv::Op::OpImageSampleExplicitLod: |
48 | 237k | case spv::Op::OpImageSampleProjImplicitLod: |
49 | 237k | case spv::Op::OpImageSampleProjExplicitLod: |
50 | 237k | case spv::Op::OpImageFetch: |
51 | 237k | case spv::Op::OpImageRead: |
52 | 237k | case spv::Op::OpImageSparseSampleImplicitLod: |
53 | 237k | case spv::Op::OpImageSparseSampleExplicitLod: |
54 | 237k | case spv::Op::OpImageSparseSampleProjImplicitLod: |
55 | 237k | case spv::Op::OpImageSparseSampleProjExplicitLod: |
56 | 237k | case spv::Op::OpImageSparseFetch: |
57 | 237k | case spv::Op::OpImageSparseRead: |
58 | 237k | return inst->NumOperands() > 4 ? 2 : -1; |
59 | 8 | case spv::Op::OpImageSampleDrefImplicitLod: |
60 | 8 | case spv::Op::OpImageSampleDrefExplicitLod: |
61 | 8 | case spv::Op::OpImageSampleProjDrefImplicitLod: |
62 | 8 | case spv::Op::OpImageSampleProjDrefExplicitLod: |
63 | 8 | case spv::Op::OpImageGather: |
64 | 8 | case spv::Op::OpImageDrefGather: |
65 | 8 | case spv::Op::OpImageSparseSampleDrefImplicitLod: |
66 | 8 | case spv::Op::OpImageSparseSampleDrefExplicitLod: |
67 | 8 | case spv::Op::OpImageSparseSampleProjDrefImplicitLod: |
68 | 8 | case spv::Op::OpImageSparseSampleProjDrefExplicitLod: |
69 | 8 | case spv::Op::OpImageSparseGather: |
70 | 8 | case spv::Op::OpImageSparseDrefGather: |
71 | 8 | return inst->NumOperands() > 5 ? 3 : -1; |
72 | 0 | case spv::Op::OpImageWrite: |
73 | 0 | return inst->NumOperands() > 3 ? 3 : -1; |
74 | 0 | default: |
75 | 0 | return -1; |
76 | 237k | } |
77 | 237k | } |
78 | | |
79 | | // Returns the element width of |type|. |
80 | 5.18M | uint32_t ElementWidth(const analysis::Type* type) { |
81 | 5.18M | if (const analysis::CooperativeVectorNV* coopvec_type = |
82 | 5.18M | type->AsCooperativeVectorNV()) { |
83 | 0 | return ElementWidth(coopvec_type->component_type()); |
84 | 5.18M | } else if (const analysis::Vector* vec_type = type->AsVector()) { |
85 | 1.85M | return ElementWidth(vec_type->element_type()); |
86 | 3.32M | } else if (const analysis::Float* float_type = type->AsFloat()) { |
87 | 2.75M | return float_type->width(); |
88 | 2.75M | } else { |
89 | 571k | assert(type->AsInteger()); |
90 | 571k | return type->AsInteger()->width(); |
91 | 571k | } |
92 | 5.18M | } |
93 | | |
94 | | // Returns true if |type| is Float or a vector of Float. |
95 | 5.43M | bool HasFloatingPoint(const analysis::Type* type) { |
96 | 5.43M | if (type->AsFloat()) { |
97 | 1.75M | return true; |
98 | 3.67M | } else if (const analysis::Vector* vec_type = type->AsVector()) { |
99 | 2.98M | return vec_type->element_type()->AsFloat() != nullptr; |
100 | 2.98M | } |
101 | | |
102 | 693k | return false; |
103 | 5.43M | } |
104 | | |
105 | | // Returns false if |val| is NaN, infinite or subnormal. |
106 | | template <typename T> |
107 | 115k | bool IsValidResult(T val) { |
108 | 115k | int classified = std::fpclassify(val); |
109 | 115k | switch (classified) { |
110 | 6.37k | case FP_NAN: |
111 | 25.2k | case FP_INFINITE: |
112 | 27.8k | case FP_SUBNORMAL: |
113 | 27.8k | return false; |
114 | 88.0k | default: |
115 | 88.0k | return true; |
116 | 115k | } |
117 | 115k | } Unexecuted instantiation: folding_rules.cpp:bool spvtools::opt::(anonymous namespace)::IsValidResult<double>(double) folding_rules.cpp:bool spvtools::opt::(anonymous namespace)::IsValidResult<float>(float) Line | Count | Source | 107 | 115k | bool IsValidResult(T val) { | 108 | 115k | int classified = std::fpclassify(val); | 109 | 115k | switch (classified) { | 110 | 6.37k | case FP_NAN: | 111 | 25.2k | case FP_INFINITE: | 112 | 27.8k | case FP_SUBNORMAL: | 113 | 27.8k | return false; | 114 | 88.0k | default: | 115 | 88.0k | return true; | 116 | 115k | } | 117 | 115k | } |
|
118 | | |
119 | | const analysis::Constant* ConstInput( |
120 | 3.80M | const std::vector<const analysis::Constant*>& constants) { |
121 | 3.80M | return constants[0] ? constants[0] : constants[1]; |
122 | 3.80M | } |
123 | | |
124 | | Instruction* NonConstInput(IRContext* context, const analysis::Constant* c, |
125 | 1.79M | Instruction* inst) { |
126 | 1.79M | uint32_t in_op = c ? 1u : 0u; |
127 | 1.79M | return context->get_def_use_mgr()->GetDef( |
128 | 1.79M | inst->GetSingleWordInOperand(in_op)); |
129 | 1.79M | } |
130 | | |
131 | 0 | std::vector<uint32_t> ExtractInts(uint64_t val) { |
132 | 0 | std::vector<uint32_t> words; |
133 | 0 | words.push_back(static_cast<uint32_t>(val)); |
134 | 0 | words.push_back(static_cast<uint32_t>(val >> 32)); |
135 | 0 | return words; |
136 | 0 | } |
137 | | |
138 | | std::vector<uint32_t> GetWordsFromScalarIntConstant( |
139 | 1.58k | const analysis::IntConstant* c) { |
140 | 1.58k | assert(c != nullptr); |
141 | 1.58k | uint32_t width = c->type()->AsInteger()->width(); |
142 | 1.58k | assert(width == 8 || width == 16 || width == 32 || width == 64); |
143 | 1.58k | if (width == 64) { |
144 | 0 | uint64_t uval = static_cast<uint64_t>(c->GetU64()); |
145 | 0 | return ExtractInts(uval); |
146 | 0 | } |
147 | | // Section 2.2.1 of the SPIR-V spec guarantees that all integer types |
148 | | // smaller than 32-bits are automatically zero or sign extended to 32-bits. |
149 | 1.58k | return {c->GetU32BitValue()}; |
150 | 1.58k | } |
151 | | |
152 | | std::vector<uint32_t> GetWordsFromScalarFloatConstant( |
153 | 377 | const analysis::FloatConstant* c) { |
154 | 377 | assert(c != nullptr); |
155 | 377 | uint32_t width = c->type()->AsFloat()->width(); |
156 | 377 | assert(width == 16 || width == 32 || width == 64); |
157 | 377 | if (width == 64) { |
158 | 0 | utils::FloatProxy<double> result(c->GetDouble()); |
159 | 0 | return result.GetWords(); |
160 | 0 | } |
161 | | // Section 2.2.1 of the SPIR-V spec guarantees that all floating-point types |
162 | | // smaller than 32-bits are automatically zero extended to 32-bits. |
163 | 377 | return {c->GetU32BitValue()}; |
164 | 377 | } |
165 | | |
166 | | std::vector<uint32_t> GetWordsFromNumericScalarOrVectorConstant( |
167 | 2.68k | analysis::ConstantManager* const_mgr, const analysis::Constant* c) { |
168 | 2.68k | if (const auto* float_constant = c->AsFloatConstant()) { |
169 | 377 | return GetWordsFromScalarFloatConstant(float_constant); |
170 | 2.31k | } else if (const auto* int_constant = c->AsIntConstant()) { |
171 | 1.58k | return GetWordsFromScalarIntConstant(int_constant); |
172 | 1.58k | } else if (const auto* vec_constant = c->AsVectorConstant()) { |
173 | 117 | std::vector<uint32_t> words; |
174 | | // Retrieve all the components as 32bit words. |
175 | 425 | for (const auto* comp : vec_constant->GetComponents()) { |
176 | 425 | auto comp_in_words = |
177 | 425 | GetWordsFromNumericScalarOrVectorConstant(const_mgr, comp); |
178 | 425 | words.insert(words.end(), comp_in_words.begin(), comp_in_words.end()); |
179 | 425 | } |
180 | | |
181 | 117 | if (ElementWidth(c->type()) >= 32) { |
182 | 117 | return words; |
183 | 117 | } |
184 | | // Check the element width and concactenate if the width is less than 32. |
185 | 0 | if (ElementWidth(c->type()) == 8) { |
186 | 0 | assert(words.size() <= 4); |
187 | | // Each 32-bit word will comprise 4 8-bit integers. |
188 | | // reverse the order when compacting. |
189 | 0 | uint32_t compacted_word = 0; |
190 | 0 | for (int32_t i = static_cast<int32_t>(words.size()) - 1; i >= 0; --i) { |
191 | 0 | compacted_word <<= 8; |
192 | 0 | compacted_word |= (words[i] & 0xFF); |
193 | 0 | } |
194 | 0 | return {compacted_word}; |
195 | 0 | } else if (ElementWidth(c->type()) == 16) { |
196 | 0 | assert(words.size() <= 4); |
197 | 0 | std::vector<uint32_t> compacted_words; |
198 | | // Each 32-bit word will comprise 2 16-bit integers. |
199 | | // reverse the order pair-wise when compacting. |
200 | 0 | for (uint32_t i = 0; i < words.size(); i += 2) { |
201 | 0 | uint32_t word1 = words[i]; |
202 | 0 | uint32_t word2 = (i + 1 < words.size()) ? words[i + 1] : 0; |
203 | 0 | uint32_t compacted_word = (word2 << 16) | (word1 & 0xFFFF); |
204 | 0 | compacted_words.push_back(compacted_word); |
205 | 0 | } |
206 | 0 | return compacted_words; |
207 | 0 | } |
208 | 0 | assert(false && "Unhandled element width"); |
209 | 609 | } else if (c->AsNullConstant()) { |
210 | 609 | uint32_t num_elements = 1; |
211 | | |
212 | 609 | if (const auto* vec_type = c->type()->AsVector()) { |
213 | 0 | num_elements = vec_type->element_count(); |
214 | 0 | } |
215 | | |
216 | | // We need to check the element width to determine how many 32-bit words are |
217 | | // needed. |
218 | 609 | uint32_t element_width = ElementWidth(c->type()); |
219 | 609 | if (element_width < 32) { |
220 | 0 | num_elements = (num_elements + 1) / 2; |
221 | 609 | } else if (element_width == 64) { |
222 | 0 | num_elements = num_elements * 2; |
223 | 0 | } |
224 | 609 | return std::vector<uint32_t>(num_elements, 0); |
225 | 609 | } |
226 | 0 | return {}; |
227 | 2.68k | } |
228 | | |
229 | | const analysis::Constant* ConvertWordsToNumericScalarOrVectorConstant( |
230 | | analysis::ConstantManager* const_mgr, const std::vector<uint32_t>& words, |
231 | 2.26k | const analysis::Type* type) { |
232 | 2.26k | const spvtools::opt::analysis::Integer* int_type = type->AsInteger(); |
233 | | |
234 | 2.26k | if (int_type && int_type->width() <= 32) { |
235 | 1.14k | assert(words.size() == 1); |
236 | 1.14k | return const_mgr->GenerateIntegerConstant(int_type, words[0]); |
237 | 1.14k | } |
238 | | |
239 | 1.11k | if (int_type || type->AsFloat()) return const_mgr->GetConstant(type, words); |
240 | 117 | if (const auto* vec_type = type->AsVector()) |
241 | 117 | return const_mgr->GetNumericVectorConstantWithWords(vec_type, words); |
242 | 0 | return nullptr; |
243 | 117 | } |
244 | | |
245 | | // Returns the negation of |c|. |c| must be a 32 or 64 bit floating point |
246 | | // constant. |
247 | | uint32_t NegateFloatingPointConstant(analysis::ConstantManager* const_mgr, |
248 | 2.52k | const analysis::Constant* c) { |
249 | 2.52k | assert(c); |
250 | 2.52k | assert(c->type()->AsFloat()); |
251 | 2.52k | uint32_t width = c->type()->AsFloat()->width(); |
252 | 2.52k | assert(width == 32 || width == 64); |
253 | 2.52k | std::vector<uint32_t> words; |
254 | 2.52k | if (width == 64) { |
255 | 0 | utils::FloatProxy<double> result(c->GetDouble() * -1.0); |
256 | 0 | words = result.GetWords(); |
257 | 2.52k | } else { |
258 | 2.52k | utils::FloatProxy<float> result(c->GetFloat() * -1.0f); |
259 | 2.52k | words = result.GetWords(); |
260 | 2.52k | } |
261 | | |
262 | 2.52k | const analysis::Constant* negated_const = |
263 | 2.52k | const_mgr->GetConstant(c->type(), std::move(words)); |
264 | 2.52k | return const_mgr->GetDefiningInstruction(negated_const)->result_id(); |
265 | 2.52k | } |
266 | | |
267 | | // Negates the integer constant |c|. Returns the id of the defining instruction. |
268 | | uint32_t NegateIntegerConstant(analysis::ConstantManager* const_mgr, |
269 | 630 | const analysis::Constant* c) { |
270 | 630 | assert(c); |
271 | 630 | assert(c->type()->AsInteger()); |
272 | 630 | uint32_t width = c->type()->AsInteger()->width(); |
273 | 630 | assert(width == 32 || width == 64); |
274 | 630 | std::vector<uint32_t> words; |
275 | 630 | if (width == 64) { |
276 | 0 | uint64_t uval = static_cast<uint64_t>(0 - c->GetU64()); |
277 | 0 | words = ExtractInts(uval); |
278 | 630 | } else { |
279 | 630 | words.push_back(static_cast<uint32_t>(0 - c->GetU32())); |
280 | 630 | } |
281 | | |
282 | 630 | const analysis::Constant* negated_const = |
283 | 630 | const_mgr->GetConstant(c->type(), std::move(words)); |
284 | 630 | return const_mgr->GetDefiningInstruction(negated_const)->result_id(); |
285 | 630 | } |
286 | | |
287 | | // Negates the vector constant |c|. Returns the id of the defining instruction. |
288 | | uint32_t NegateVectorConstant(analysis::ConstantManager* const_mgr, |
289 | 684 | const analysis::Constant* c) { |
290 | 684 | assert(const_mgr && c); |
291 | 684 | assert(c->type()->AsVector()); |
292 | 684 | if (c->AsNullConstant()) { |
293 | | // 0.0 vs -0.0 shouldn't matter. |
294 | 0 | return const_mgr->GetDefiningInstruction(c)->result_id(); |
295 | 684 | } else { |
296 | 684 | const analysis::Type* component_type = |
297 | 684 | c->AsVectorConstant()->component_type(); |
298 | 684 | std::vector<uint32_t> words; |
299 | 1.36k | for (auto& comp : c->AsVectorConstant()->GetComponents()) { |
300 | 1.36k | if (component_type->AsFloat()) { |
301 | 1.36k | words.push_back(NegateFloatingPointConstant(const_mgr, comp)); |
302 | 1.36k | } else { |
303 | 0 | assert(component_type->AsInteger()); |
304 | 0 | words.push_back(NegateIntegerConstant(const_mgr, comp)); |
305 | 0 | } |
306 | 1.36k | } |
307 | | |
308 | 684 | const analysis::Constant* negated_const = |
309 | 684 | const_mgr->GetConstant(c->type(), std::move(words)); |
310 | 684 | return const_mgr->GetDefiningInstruction(negated_const)->result_id(); |
311 | 684 | } |
312 | 684 | } |
313 | | |
314 | | // Negates |c|. Returns the id of the defining instruction. |
315 | | uint32_t NegateConstant(analysis::ConstantManager* const_mgr, |
316 | 2.46k | const analysis::Constant* c) { |
317 | 2.46k | if (c->type()->AsVector()) { |
318 | 684 | return NegateVectorConstant(const_mgr, c); |
319 | 1.78k | } else if (c->type()->AsFloat()) { |
320 | 1.15k | return NegateFloatingPointConstant(const_mgr, c); |
321 | 1.15k | } else { |
322 | 630 | assert(c->type()->AsInteger()); |
323 | 630 | return NegateIntegerConstant(const_mgr, c); |
324 | 630 | } |
325 | 2.46k | } |
326 | | |
327 | | // Takes the reciprocal of |c|. |c|'s type must be Float or a vector of Float. |
328 | | // Returns 0 if the reciprocal is NaN, infinite or subnormal. |
329 | | uint32_t Reciprocal(analysis::ConstantManager* const_mgr, |
330 | 91.6k | const analysis::Constant* c) { |
331 | 91.6k | assert(const_mgr && c); |
332 | 91.6k | assert(c->type()->AsFloat()); |
333 | | |
334 | 91.6k | uint32_t width = c->type()->AsFloat()->width(); |
335 | 91.6k | assert(width == 32 || width == 64); |
336 | 91.6k | std::vector<uint32_t> words; |
337 | | |
338 | 91.6k | if (c->IsZero()) { |
339 | 17.7k | return 0; |
340 | 17.7k | } |
341 | | |
342 | 73.8k | if (width == 64) { |
343 | 0 | spvtools::utils::FloatProxy<double> result(1.0 / c->GetDouble()); |
344 | 0 | if (!IsValidResult(result.getAsFloat())) return 0; |
345 | 0 | words = result.GetWords(); |
346 | 73.8k | } else { |
347 | 73.8k | spvtools::utils::FloatProxy<float> result(1.0f / c->GetFloat()); |
348 | 73.8k | if (!IsValidResult(result.getAsFloat())) return 0; |
349 | 57.7k | words = result.GetWords(); |
350 | 57.7k | } |
351 | | |
352 | 57.7k | const analysis::Constant* negated_const = |
353 | 57.7k | const_mgr->GetConstant(c->type(), std::move(words)); |
354 | 57.7k | return const_mgr->GetDefiningInstruction(negated_const)->result_id(); |
355 | 73.8k | } |
356 | | |
357 | | // Replaces fdiv where second operand is constant with fmul. |
358 | 16.0k | FoldingRule ReciprocalFDiv() { |
359 | 16.0k | return [](IRContext* context, Instruction* inst, |
360 | 103k | const std::vector<const analysis::Constant*>& constants) { |
361 | 103k | assert(inst->opcode() == spv::Op::OpFDiv); |
362 | 103k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
363 | 103k | const analysis::Type* type = |
364 | 103k | context->get_type_mgr()->GetType(inst->type_id()); |
365 | | |
366 | 103k | if (type->IsCooperativeMatrix()) { |
367 | 0 | return false; |
368 | 0 | } |
369 | | |
370 | 103k | if (!inst->IsFloatingPointFoldingAllowed()) return false; |
371 | | |
372 | 102k | uint32_t width = ElementWidth(type); |
373 | 102k | if (width != 32 && width != 64) return false; |
374 | | |
375 | 102k | if (constants[1] != nullptr) { |
376 | 63.4k | uint32_t id = 0; |
377 | 63.4k | if (const analysis::VectorConstant* vector_const = |
378 | 63.4k | constants[1]->AsVectorConstant()) { |
379 | 52.0k | std::vector<uint32_t> neg_ids; |
380 | 80.3k | for (auto& comp : vector_const->GetComponents()) { |
381 | 80.3k | id = Reciprocal(const_mgr, comp); |
382 | 80.3k | if (id == 0) return false; |
383 | 48.8k | neg_ids.push_back(id); |
384 | 48.8k | } |
385 | 20.5k | const analysis::Constant* negated_const = |
386 | 20.5k | const_mgr->GetConstant(constants[1]->type(), std::move(neg_ids)); |
387 | 20.5k | id = const_mgr->GetDefiningInstruction(negated_const)->result_id(); |
388 | 20.5k | } else if (constants[1]->AsFloatConstant()) { |
389 | 11.2k | id = Reciprocal(const_mgr, constants[1]); |
390 | 11.2k | if (id == 0) return false; |
391 | 11.2k | } else { |
392 | | // Don't fold a null constant. |
393 | 263 | return false; |
394 | 263 | } |
395 | 29.3k | inst->SetOpcode(spv::Op::OpFMul); |
396 | 29.3k | inst->SetInOperands( |
397 | 29.3k | {{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(0u)}}, |
398 | 29.3k | {SPV_OPERAND_TYPE_ID, {id}}}); |
399 | 29.3k | return true; |
400 | 63.4k | } |
401 | | |
402 | 39.5k | return false; |
403 | 102k | }; |
404 | 16.0k | } |
405 | | |
406 | | // Elides consecutive negate instructions. |
407 | 32.0k | FoldingRule MergeNegateArithmetic() { |
408 | 32.0k | return [](IRContext* context, Instruction* inst, |
409 | 32.0k | const std::vector<const analysis::Constant*>& constants) { |
410 | 7.30k | assert(inst->opcode() == spv::Op::OpFNegate || |
411 | 7.30k | inst->opcode() == spv::Op::OpSNegate); |
412 | 7.30k | (void)constants; |
413 | 7.30k | const analysis::Type* type = |
414 | 7.30k | context->get_type_mgr()->GetType(inst->type_id()); |
415 | 7.30k | if (HasFloatingPoint(type) && !inst->IsFloatingPointFoldingAllowed()) |
416 | 0 | return false; |
417 | | |
418 | 7.30k | Instruction* op_inst = |
419 | 7.30k | context->get_def_use_mgr()->GetDef(inst->GetSingleWordInOperand(0u)); |
420 | 7.30k | if (HasFloatingPoint(type) && !op_inst->IsFloatingPointFoldingAllowed()) |
421 | 0 | return false; |
422 | | |
423 | 7.30k | if (op_inst->opcode() == inst->opcode()) { |
424 | | // Elide negates. |
425 | 157 | inst->SetOpcode(spv::Op::OpCopyObject); |
426 | 157 | inst->SetInOperands( |
427 | 157 | {{SPV_OPERAND_TYPE_ID, {op_inst->GetSingleWordInOperand(0u)}}}); |
428 | 157 | return true; |
429 | 157 | } |
430 | | |
431 | 7.14k | return false; |
432 | 7.30k | }; |
433 | 32.0k | } |
434 | | |
435 | | // Merges negate into a mul or div operation if that operation contains a |
436 | | // constant operand. |
437 | | // Cases: |
438 | | // -(x * 2) = x * -2 |
439 | | // -(2 * x) = x * -2 |
440 | | // -(x / 2) = x / -2 |
441 | | // -(2 / x) = -2 / x |
442 | 32.0k | FoldingRule MergeNegateMulDivArithmetic() { |
443 | 32.0k | return [](IRContext* context, Instruction* inst, |
444 | 32.0k | const std::vector<const analysis::Constant*>& constants) { |
445 | 6.99k | assert(inst->opcode() == spv::Op::OpFNegate || |
446 | 6.99k | inst->opcode() == spv::Op::OpSNegate); |
447 | 6.99k | (void)constants; |
448 | 6.99k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
449 | 6.99k | const analysis::Type* type = |
450 | 6.99k | context->get_type_mgr()->GetType(inst->type_id()); |
451 | | |
452 | 6.99k | if (type->IsCooperativeMatrix()) { |
453 | 0 | return false; |
454 | 0 | } |
455 | | |
456 | 6.99k | if (HasFloatingPoint(type) && !inst->IsFloatingPointFoldingAllowed()) |
457 | 0 | return false; |
458 | | |
459 | 6.99k | Instruction* op_inst = |
460 | 6.99k | context->get_def_use_mgr()->GetDef(inst->GetSingleWordInOperand(0u)); |
461 | 6.99k | if (HasFloatingPoint(type) && !op_inst->IsFloatingPointFoldingAllowed()) |
462 | 0 | return false; |
463 | | |
464 | 6.99k | uint32_t width = ElementWidth(type); |
465 | 6.99k | if (width != 32 && width != 64) return false; |
466 | | |
467 | 6.99k | spv::Op opcode = op_inst->opcode(); |
468 | 6.99k | if (opcode != spv::Op::OpFMul && opcode != spv::Op::OpFDiv && |
469 | 6.31k | opcode != spv::Op::OpIMul && opcode != spv::Op::OpSDiv) { |
470 | 6.24k | return false; |
471 | 6.24k | } |
472 | | |
473 | 748 | std::vector<const analysis::Constant*> op_constants = |
474 | 748 | const_mgr->GetOperandConstants(op_inst); |
475 | | // Merge negate into mul or div if one operand is constant. |
476 | 748 | if (op_constants[0] == nullptr && op_constants[1] == nullptr) { |
477 | 279 | return false; |
478 | 279 | } |
479 | | |
480 | 469 | bool zero_is_variable = op_constants[0] == nullptr; |
481 | 469 | const analysis::Constant* c = ConstInput(op_constants); |
482 | 469 | uint32_t neg_id = NegateConstant(const_mgr, c); |
483 | 469 | uint32_t non_const_id = zero_is_variable |
484 | 469 | ? op_inst->GetSingleWordInOperand(0u) |
485 | 469 | : op_inst->GetSingleWordInOperand(1u); |
486 | | // Change this instruction to a mul/div. |
487 | 469 | inst->SetOpcode(op_inst->opcode()); |
488 | 469 | if (opcode == spv::Op::OpFDiv || opcode == spv::Op::OpUDiv || |
489 | 441 | opcode == spv::Op::OpSDiv) { |
490 | 34 | uint32_t op0 = zero_is_variable ? non_const_id : neg_id; |
491 | 34 | uint32_t op1 = zero_is_variable ? neg_id : non_const_id; |
492 | 34 | inst->SetInOperands( |
493 | 34 | {{SPV_OPERAND_TYPE_ID, {op0}}, {SPV_OPERAND_TYPE_ID, {op1}}}); |
494 | 435 | } else { |
495 | 435 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {non_const_id}}, |
496 | 435 | {SPV_OPERAND_TYPE_ID, {neg_id}}}); |
497 | 435 | } |
498 | 469 | return true; |
499 | 748 | }; |
500 | 32.0k | } |
501 | | |
502 | | // Merges negate into a add or sub operation if that operation contains a |
503 | | // constant operand. |
504 | | // Cases: |
505 | | // -(x + 2) = -2 - x |
506 | | // -(2 + x) = -2 - x |
507 | | // -(x - 2) = 2 - x |
508 | | // -(2 - x) = x - 2 |
509 | 32.0k | FoldingRule MergeNegateAddSubArithmetic() { |
510 | 32.0k | return [](IRContext* context, Instruction* inst, |
511 | 32.0k | const std::vector<const analysis::Constant*>& constants) { |
512 | 7.12k | assert(inst->opcode() == spv::Op::OpFNegate || |
513 | 7.12k | inst->opcode() == spv::Op::OpSNegate); |
514 | 7.12k | (void)constants; |
515 | 7.12k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
516 | 7.12k | const analysis::Type* type = |
517 | 7.12k | context->get_type_mgr()->GetType(inst->type_id()); |
518 | | |
519 | 7.12k | if (type->IsCooperativeMatrix()) { |
520 | 0 | return false; |
521 | 0 | } |
522 | | |
523 | 7.12k | if (HasFloatingPoint(type) && !inst->IsFloatingPointFoldingAllowed()) |
524 | 0 | return false; |
525 | | |
526 | 7.12k | Instruction* op_inst = |
527 | 7.12k | context->get_def_use_mgr()->GetDef(inst->GetSingleWordInOperand(0u)); |
528 | 7.12k | if (HasFloatingPoint(type) && !op_inst->IsFloatingPointFoldingAllowed()) |
529 | 0 | return false; |
530 | | |
531 | 7.12k | uint32_t width = ElementWidth(type); |
532 | 7.12k | if (width != 32 && width != 64) return false; |
533 | | |
534 | 7.12k | if (op_inst->opcode() == spv::Op::OpFAdd || |
535 | 6.92k | op_inst->opcode() == spv::Op::OpFSub || |
536 | 6.70k | op_inst->opcode() == spv::Op::OpIAdd || |
537 | 6.44k | op_inst->opcode() == spv::Op::OpISub) { |
538 | 683 | std::vector<const analysis::Constant*> op_constants = |
539 | 683 | const_mgr->GetOperandConstants(op_inst); |
540 | 683 | if (op_constants[0] || op_constants[1]) { |
541 | 248 | bool zero_is_variable = op_constants[0] == nullptr; |
542 | 248 | bool is_add = (op_inst->opcode() == spv::Op::OpFAdd) || |
543 | 187 | (op_inst->opcode() == spv::Op::OpIAdd); |
544 | 248 | bool swap_operands = !is_add || zero_is_variable; |
545 | 248 | bool negate_const = is_add; |
546 | 248 | const analysis::Constant* c = ConstInput(op_constants); |
547 | 248 | uint32_t const_id = 0; |
548 | 248 | if (negate_const) { |
549 | 156 | const_id = NegateConstant(const_mgr, c); |
550 | 156 | } else { |
551 | 92 | const_id = zero_is_variable ? op_inst->GetSingleWordInOperand(1u) |
552 | 92 | : op_inst->GetSingleWordInOperand(0u); |
553 | 92 | } |
554 | | |
555 | | // Swap operands if necessary and make the instruction a subtraction. |
556 | 248 | uint32_t op0 = |
557 | 248 | zero_is_variable ? op_inst->GetSingleWordInOperand(0u) : const_id; |
558 | 248 | uint32_t op1 = |
559 | 248 | zero_is_variable ? const_id : op_inst->GetSingleWordInOperand(1u); |
560 | 248 | if (swap_operands) std::swap(op0, op1); |
561 | 248 | inst->SetOpcode(HasFloatingPoint(type) ? spv::Op::OpFSub |
562 | 248 | : spv::Op::OpISub); |
563 | 248 | inst->SetInOperands( |
564 | 248 | {{SPV_OPERAND_TYPE_ID, {op0}}, {SPV_OPERAND_TYPE_ID, {op1}}}); |
565 | 248 | return true; |
566 | 248 | } |
567 | 683 | } |
568 | | |
569 | 6.87k | return false; |
570 | 7.12k | }; |
571 | 32.0k | } |
572 | | |
573 | | // Returns true if |c| has a zero element. |
574 | 202k | bool HasZero(const analysis::Constant* c) { |
575 | 202k | if (c->AsNullConstant()) { |
576 | 534 | return true; |
577 | 534 | } |
578 | 201k | if (const analysis::VectorConstant* vec_const = c->AsVectorConstant()) { |
579 | 70.0k | for (auto& comp : vec_const->GetComponents()) |
580 | 105k | if (HasZero(comp)) return true; |
581 | 131k | } else { |
582 | 131k | assert(c->AsScalarConstant()); |
583 | 131k | return c->AsScalarConstant()->IsZero(); |
584 | 131k | } |
585 | | |
586 | 27.0k | return false; |
587 | 201k | } |
588 | | |
589 | | // Performs |input1| |opcode| |input2| and returns the merged constant result |
590 | | // id. Returns 0 if the result is not a valid value. The input types must be |
591 | | // Float. |
592 | | uint32_t PerformFloatingPointOperation(analysis::ConstantManager* const_mgr, |
593 | | spv::Op opcode, |
594 | | const analysis::Constant* input1, |
595 | 42.0k | const analysis::Constant* input2) { |
596 | 42.0k | const analysis::Type* type = input1->type(); |
597 | 42.0k | assert(type->AsFloat()); |
598 | 42.0k | uint32_t width = type->AsFloat()->width(); |
599 | 42.0k | assert(width == 32 || width == 64); |
600 | 42.0k | std::vector<uint32_t> words; |
601 | 42.0k | #define FOLD_OP(op) \ |
602 | 42.0k | if (width == 64) { \ |
603 | 0 | utils::FloatProxy<double> val = \ |
604 | 0 | input1->GetDouble() op input2->GetDouble(); \ |
605 | 0 | double dval = val.getAsFloat(); \ |
606 | 0 | if (!IsValidResult(dval)) return 0; \ |
607 | 0 | words = val.GetWords(); \ |
608 | 42.0k | } else { \ |
609 | 42.0k | utils::FloatProxy<float> val = input1->GetFloat() op input2->GetFloat(); \ |
610 | 42.0k | float fval = val.getAsFloat(); \ |
611 | 42.0k | if (!IsValidResult(fval)) return 0; \ |
612 | 42.0k | words = val.GetWords(); \ |
613 | 30.3k | } \ |
614 | 42.0k | static_assert(true, "require extra semicolon") |
615 | 42.0k | switch (opcode) { |
616 | 6.84k | case spv::Op::OpFMul: |
617 | 6.84k | FOLD_OP(*); |
618 | 3.69k | break; |
619 | 2.45k | case spv::Op::OpFDiv: |
620 | 2.45k | if (HasZero(input2)) return 0; |
621 | 2.45k | FOLD_OP(/); |
622 | 1.62k | break; |
623 | 25.9k | case spv::Op::OpFAdd: |
624 | 25.9k | FOLD_OP(+); |
625 | 20.2k | break; |
626 | 6.80k | case spv::Op::OpFSub: |
627 | 6.80k | FOLD_OP(-); |
628 | 4.71k | break; |
629 | 0 | default: |
630 | 0 | assert(false && "Unexpected operation"); |
631 | 0 | break; |
632 | 42.0k | } |
633 | 30.3k | #undef FOLD_OP |
634 | 30.3k | const analysis::Constant* merged_const = const_mgr->GetConstant(type, words); |
635 | 30.3k | return const_mgr->GetDefiningInstruction(merged_const)->result_id(); |
636 | 42.0k | } |
637 | | |
638 | | // Performs |input1| |opcode| |input2| and returns the merged constant result |
639 | | // id. Returns 0 if the result is not a valid value. The input types must be |
640 | | // Integers. |
641 | | uint32_t PerformIntegerOperation(analysis::ConstantManager* const_mgr, |
642 | | spv::Op opcode, |
643 | | const analysis::Constant* input1, |
644 | 7.24k | const analysis::Constant* input2) { |
645 | 7.24k | assert(input1->type()->AsInteger()); |
646 | 7.24k | const analysis::Integer* type = input1->type()->AsInteger(); |
647 | 7.24k | uint32_t width = type->AsInteger()->width(); |
648 | 7.24k | assert(width == 32 || width == 64); |
649 | 7.24k | std::vector<uint32_t> words; |
650 | | // Regardless of the sign of the constant, folding is performed on an unsigned |
651 | | // interpretation of the constant data. This avoids signed integer overflow |
652 | | // while folding, and works because sign is irrelevant for the IAdd, ISub and |
653 | | // IMul instructions. |
654 | 7.24k | #define FOLD_OP(op) \ |
655 | 7.24k | if (width == 64) { \ |
656 | 0 | uint64_t val = input1->GetU64() op input2->GetU64(); \ |
657 | 0 | words = ExtractInts(val); \ |
658 | 7.24k | } else { \ |
659 | 7.24k | uint32_t val = input1->GetU32() op input2->GetU32(); \ |
660 | 7.24k | words.push_back(val); \ |
661 | 7.24k | } \ |
662 | 7.24k | static_assert(true, "require extra semicolon") |
663 | 7.24k | switch (opcode) { |
664 | 355 | case spv::Op::OpIMul: |
665 | 355 | FOLD_OP(*); |
666 | 355 | break; |
667 | 0 | case spv::Op::OpSDiv: |
668 | 0 | case spv::Op::OpUDiv: |
669 | 0 | assert(false && "Should not merge integer division"); |
670 | 0 | break; |
671 | 2.75k | case spv::Op::OpIAdd: |
672 | 2.75k | FOLD_OP(+); |
673 | 2.75k | break; |
674 | 1.92k | case spv::Op::OpISub: |
675 | 1.92k | FOLD_OP(-); |
676 | 1.92k | break; |
677 | 11 | case spv::Op::OpBitwiseXor: |
678 | 11 | FOLD_OP(^); |
679 | 11 | break; |
680 | 1.15k | case spv::Op::OpBitwiseOr: |
681 | 1.15k | FOLD_OP(|); |
682 | 1.15k | break; |
683 | 1.04k | case spv::Op::OpBitwiseAnd: |
684 | 1.04k | FOLD_OP(&); |
685 | 1.04k | break; |
686 | 0 | default: |
687 | 0 | assert(false && "Unexpected operation"); |
688 | 0 | break; |
689 | 7.24k | } |
690 | 7.24k | #undef FOLD_OP |
691 | 7.24k | const analysis::Constant* merged_const = const_mgr->GetConstant(type, words); |
692 | 7.24k | return const_mgr->GetDefiningInstruction(merged_const)->result_id(); |
693 | 7.24k | } |
694 | | |
695 | | // Performs |input1| |opcode| |input2| and returns the merged constant result |
696 | | // id. Returns 0 if the result is not a valid value. The input types must be |
697 | | // Integers, Floats or Vectors of such. |
698 | | uint32_t PerformOperation(analysis::ConstantManager* const_mgr, spv::Op opcode, |
699 | | const analysis::Constant* input1, |
700 | 39.5k | const analysis::Constant* input2) { |
701 | 39.5k | assert(input1 && input2); |
702 | 39.5k | const analysis::Type* type = input1->type(); |
703 | 39.5k | std::vector<uint32_t> words; |
704 | 39.5k | if (const analysis::Vector* vector_type = type->AsVector()) { |
705 | 12.4k | const analysis::Type* ele_type = vector_type->element_type(); |
706 | 31.3k | for (uint32_t i = 0; i != vector_type->element_count(); ++i) { |
707 | 22.2k | uint32_t id = 0; |
708 | | |
709 | 22.2k | const analysis::Constant* input1_comp = nullptr; |
710 | 22.2k | if (const analysis::VectorConstant* input1_vector = |
711 | 22.2k | input1->AsVectorConstant()) { |
712 | 22.2k | input1_comp = input1_vector->GetComponents()[i]; |
713 | 22.2k | } else { |
714 | 0 | assert(input1->AsNullConstant()); |
715 | 0 | input1_comp = const_mgr->GetConstant(ele_type, {}); |
716 | 0 | } |
717 | | |
718 | 22.2k | const analysis::Constant* input2_comp = nullptr; |
719 | 22.2k | if (const analysis::VectorConstant* input2_vector = |
720 | 22.2k | input2->AsVectorConstant()) { |
721 | 22.2k | input2_comp = input2_vector->GetComponents()[i]; |
722 | 22.2k | } else { |
723 | 0 | assert(input2->AsNullConstant()); |
724 | 0 | input2_comp = const_mgr->GetConstant(ele_type, {}); |
725 | 0 | } |
726 | | |
727 | 22.2k | if (ele_type->AsFloat()) { |
728 | 22.2k | id = PerformFloatingPointOperation(const_mgr, opcode, input1_comp, |
729 | 22.2k | input2_comp); |
730 | 22.2k | } else { |
731 | 4 | assert(ele_type->AsInteger()); |
732 | 4 | id = PerformIntegerOperation(const_mgr, opcode, input1_comp, |
733 | 4 | input2_comp); |
734 | 4 | } |
735 | 22.2k | if (id == 0) return 0; |
736 | 18.8k | words.push_back(id); |
737 | 18.8k | } |
738 | 9.12k | const analysis::Constant* merged_const = |
739 | 9.12k | const_mgr->GetConstant(type, words); |
740 | 9.12k | return const_mgr->GetDefiningInstruction(merged_const)->result_id(); |
741 | 27.0k | } else if (type->AsFloat()) { |
742 | 19.7k | return PerformFloatingPointOperation(const_mgr, opcode, input1, input2); |
743 | 19.7k | } else { |
744 | 7.24k | assert(type->AsInteger()); |
745 | 7.24k | return PerformIntegerOperation(const_mgr, opcode, input1, input2); |
746 | 7.24k | } |
747 | 39.5k | } |
748 | | |
749 | | // Merges consecutive multiplies where each contains one constant operand. |
750 | | // Cases: |
751 | | // 2 * (x * 2) = x * 4 |
752 | | // 2 * (2 * x) = x * 4 |
753 | | // (x * 2) * 2 = x * 4 |
754 | | // (2 * x) * 2 = x * 4 |
755 | 32.0k | FoldingRule MergeMulMulArithmetic() { |
756 | 32.0k | return [](IRContext* context, Instruction* inst, |
757 | 205k | const std::vector<const analysis::Constant*>& constants) { |
758 | 205k | assert(inst->opcode() == spv::Op::OpFMul || |
759 | 205k | inst->opcode() == spv::Op::OpIMul); |
760 | 205k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
761 | 205k | const analysis::Type* type = |
762 | 205k | context->get_type_mgr()->GetType(inst->type_id()); |
763 | | |
764 | 205k | if (type->IsCooperativeMatrix()) { |
765 | 0 | return false; |
766 | 0 | } |
767 | | |
768 | 205k | if (HasFloatingPoint(type) && !inst->IsFloatingPointFoldingAllowed()) |
769 | 101 | return false; |
770 | | |
771 | 205k | uint32_t width = ElementWidth(type); |
772 | 205k | if (width != 32 && width != 64) return false; |
773 | | |
774 | | // Determine the constant input and the variable input in |inst|. |
775 | 205k | const analysis::Constant* const_input1 = ConstInput(constants); |
776 | 205k | if (!const_input1) return false; |
777 | 135k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
778 | 135k | if (HasFloatingPoint(type) && !other_inst->IsFloatingPointFoldingAllowed()) |
779 | 69 | return false; |
780 | | |
781 | 135k | if (other_inst->opcode() == inst->opcode()) { |
782 | 5.90k | std::vector<const analysis::Constant*> other_constants = |
783 | 5.90k | const_mgr->GetOperandConstants(other_inst); |
784 | 5.90k | const analysis::Constant* const_input2 = ConstInput(other_constants); |
785 | 5.90k | if (!const_input2) return false; |
786 | | |
787 | 4.06k | bool other_first_is_variable = other_constants[0] == nullptr; |
788 | 4.06k | uint32_t merged_id = PerformOperation(const_mgr, inst->opcode(), |
789 | 4.06k | const_input1, const_input2); |
790 | 4.06k | if (merged_id == 0) return false; |
791 | | |
792 | 1.98k | uint32_t non_const_id = other_first_is_variable |
793 | 1.98k | ? other_inst->GetSingleWordInOperand(0u) |
794 | 1.98k | : other_inst->GetSingleWordInOperand(1u); |
795 | 1.98k | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {non_const_id}}, |
796 | 1.98k | {SPV_OPERAND_TYPE_ID, {merged_id}}}); |
797 | 1.98k | return true; |
798 | 4.06k | } |
799 | | |
800 | 129k | return false; |
801 | 135k | }; |
802 | 32.0k | } |
803 | | |
804 | | // Merges divides into subsequent multiplies if each instruction contains one |
805 | | // constant operand. Does not support integer operations. |
806 | | // Cases: |
807 | | // 2 * (x / 2) = x * 1 |
808 | | // 2 * (2 / x) = 4 / x |
809 | | // (x / 2) * 2 = x * 1 |
810 | | // (2 / x) * 2 = 4 / x |
811 | | // (y / x) * x = y |
812 | | // x * (y / x) = y |
813 | 16.0k | FoldingRule MergeMulDivArithmetic() { |
814 | 16.0k | return [](IRContext* context, Instruction* inst, |
815 | 188k | const std::vector<const analysis::Constant*>& constants) { |
816 | 188k | assert(inst->opcode() == spv::Op::OpFMul); |
817 | 188k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
818 | 188k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
819 | | |
820 | 188k | const analysis::Type* type = |
821 | 188k | context->get_type_mgr()->GetType(inst->type_id()); |
822 | | |
823 | 188k | if (type->IsCooperativeMatrix()) { |
824 | 0 | return false; |
825 | 0 | } |
826 | | |
827 | 188k | if (!inst->IsFloatingPointFoldingAllowed()) return false; |
828 | | |
829 | 188k | uint32_t width = ElementWidth(type); |
830 | 188k | if (width != 32 && width != 64) return false; |
831 | | |
832 | 565k | for (uint32_t i = 0; i < 2; i++) { |
833 | 377k | uint32_t op_id = inst->GetSingleWordInOperand(i); |
834 | 377k | Instruction* op_inst = def_use_mgr->GetDef(op_id); |
835 | 377k | if (op_inst->opcode() == spv::Op::OpFDiv) { |
836 | 15.5k | if (op_inst->GetSingleWordInOperand(1) == |
837 | 15.5k | inst->GetSingleWordInOperand(1 - i)) { |
838 | 327 | inst->SetOpcode(spv::Op::OpCopyObject); |
839 | 327 | inst->SetInOperands( |
840 | 327 | {{SPV_OPERAND_TYPE_ID, {op_inst->GetSingleWordInOperand(0)}}}); |
841 | 327 | return true; |
842 | 327 | } |
843 | 15.5k | } |
844 | 377k | } |
845 | | |
846 | 188k | const analysis::Constant* const_input1 = ConstInput(constants); |
847 | 188k | if (!const_input1) return false; |
848 | 121k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
849 | 121k | if (!other_inst->IsFloatingPointFoldingAllowed()) return false; |
850 | | |
851 | 121k | if (other_inst->opcode() == spv::Op::OpFDiv) { |
852 | 2.83k | std::vector<const analysis::Constant*> other_constants = |
853 | 2.83k | const_mgr->GetOperandConstants(other_inst); |
854 | 2.83k | const analysis::Constant* const_input2 = ConstInput(other_constants); |
855 | 2.83k | if (!const_input2 || HasZero(const_input2)) return false; |
856 | | |
857 | 758 | bool other_first_is_variable = other_constants[0] == nullptr; |
858 | | // If the variable value is the second operand of the divide, multiply |
859 | | // the constants together. Otherwise divide the constants. |
860 | 758 | uint32_t merged_id = PerformOperation( |
861 | 758 | const_mgr, |
862 | 758 | other_first_is_variable ? other_inst->opcode() : inst->opcode(), |
863 | 758 | const_input1, const_input2); |
864 | 758 | if (merged_id == 0) return false; |
865 | | |
866 | 320 | uint32_t non_const_id = other_first_is_variable |
867 | 320 | ? other_inst->GetSingleWordInOperand(0u) |
868 | 320 | : other_inst->GetSingleWordInOperand(1u); |
869 | | |
870 | | // If the variable value is on the second operand of the div, then this |
871 | | // operation is a div. Otherwise it should be a multiply. |
872 | 320 | inst->SetOpcode(other_first_is_variable ? inst->opcode() |
873 | 320 | : other_inst->opcode()); |
874 | 320 | if (other_first_is_variable) { |
875 | 28 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {non_const_id}}, |
876 | 28 | {SPV_OPERAND_TYPE_ID, {merged_id}}}); |
877 | 292 | } else { |
878 | 292 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {merged_id}}, |
879 | 292 | {SPV_OPERAND_TYPE_ID, {non_const_id}}}); |
880 | 292 | } |
881 | 320 | return true; |
882 | 758 | } |
883 | | |
884 | 118k | return false; |
885 | 121k | }; |
886 | 16.0k | } |
887 | | |
888 | | // Merges multiply of constant and negation. |
889 | | // Cases: |
890 | | // (-x) * 2 = x * -2 |
891 | | // 2 * (-x) = x * -2 |
892 | 32.0k | FoldingRule MergeMulNegateArithmetic() { |
893 | 32.0k | return [](IRContext* context, Instruction* inst, |
894 | 202k | const std::vector<const analysis::Constant*>& constants) { |
895 | 202k | assert(inst->opcode() == spv::Op::OpFMul || |
896 | 202k | inst->opcode() == spv::Op::OpIMul); |
897 | 202k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
898 | 202k | const analysis::Type* type = |
899 | 202k | context->get_type_mgr()->GetType(inst->type_id()); |
900 | | |
901 | 202k | if (type->IsCooperativeMatrix()) { |
902 | 0 | return false; |
903 | 0 | } |
904 | | |
905 | 202k | bool uses_float = HasFloatingPoint(type); |
906 | 202k | if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
907 | | |
908 | 202k | uint32_t width = ElementWidth(type); |
909 | 202k | if (width != 32 && width != 64) return false; |
910 | | |
911 | 202k | const analysis::Constant* const_input1 = ConstInput(constants); |
912 | 202k | if (!const_input1) return false; |
913 | 133k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
914 | 133k | if (uses_float && !other_inst->IsFloatingPointFoldingAllowed()) |
915 | 69 | return false; |
916 | | |
917 | 133k | if (other_inst->opcode() == spv::Op::OpFNegate || |
918 | 133k | other_inst->opcode() == spv::Op::OpSNegate) { |
919 | 44 | uint32_t neg_id = NegateConstant(const_mgr, const_input1); |
920 | | |
921 | 44 | inst->SetInOperands( |
922 | 44 | {{SPV_OPERAND_TYPE_ID, {other_inst->GetSingleWordInOperand(0u)}}, |
923 | 44 | {SPV_OPERAND_TYPE_ID, {neg_id}}}); |
924 | 44 | return true; |
925 | 44 | } |
926 | | |
927 | 133k | return false; |
928 | 133k | }; |
929 | 32.0k | } |
930 | | |
931 | | // Returns true if |inst| is negation op and is safe to fold. |
932 | 1.76M | static bool IsFoldableNegation(const Instruction* inst) { |
933 | 1.76M | return (inst->opcode() == spv::Op::OpSNegate || |
934 | 1.76M | (inst->opcode() == spv::Op::OpFNegate && |
935 | 888 | inst->IsFloatingPointFoldingAllowed())); |
936 | 1.76M | } |
937 | | |
938 | | // Merges multiplies / divisions of two negations. |
939 | | // Cases: |
940 | | // (-x) * (-y) = x * y |
941 | | // (-x) / (-y) = x / y |
942 | 80.0k | FoldingRule MergeDivMulDoubleNegative() { |
943 | 80.0k | return [](IRContext* context, Instruction* inst, |
944 | 585k | const std::vector<const analysis::Constant*>&) { |
945 | 585k | assert(inst->opcode() == spv::Op::OpFMul || |
946 | 585k | inst->opcode() == spv::Op::OpVectorTimesScalar || |
947 | 585k | inst->opcode() == spv::Op::OpFDiv || |
948 | 585k | inst->opcode() == spv::Op::OpIMul || |
949 | 585k | inst->opcode() == spv::Op::OpSDiv); |
950 | | |
951 | 585k | const analysis::Type* type = |
952 | 585k | context->get_type_mgr()->GetType(inst->type_id()); |
953 | | |
954 | 585k | bool uses_float = HasFloatingPoint(type); |
955 | 585k | if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
956 | | |
957 | 584k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
958 | 584k | Instruction* lhs = def_use_mgr->GetDef(inst->GetSingleWordInOperand(0)); |
959 | 584k | Instruction* rhs = def_use_mgr->GetDef(inst->GetSingleWordInOperand(1)); |
960 | | |
961 | 584k | if (IsFoldableNegation(lhs) && IsFoldableNegation(rhs)) { |
962 | 12 | inst->SetInOperands( |
963 | 12 | {{SPV_OPERAND_TYPE_ID, {lhs->GetSingleWordInOperand(0u)}}, |
964 | 12 | {SPV_OPERAND_TYPE_ID, {rhs->GetSingleWordInOperand(0u)}}}); |
965 | 12 | return true; |
966 | 12 | } |
967 | 584k | return false; |
968 | 584k | }; |
969 | 80.0k | } |
970 | | |
971 | | // Merges consecutive divides if each instruction contains one constant operand. |
972 | | // Does not support integer division. |
973 | | // Cases: |
974 | | // 2 / (x / 2) = 4 / x |
975 | | // 4 / (2 / x) = 2 * x |
976 | | // (4 / x) / 2 = 2 / x |
977 | | // (x / 2) / 2 = x / 4 |
978 | 16.0k | FoldingRule MergeDivDivArithmetic() { |
979 | 16.0k | return [](IRContext* context, Instruction* inst, |
980 | 73.6k | const std::vector<const analysis::Constant*>& constants) { |
981 | 73.6k | assert(inst->opcode() == spv::Op::OpFDiv); |
982 | 73.6k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
983 | 73.6k | const analysis::Type* type = |
984 | 73.6k | context->get_type_mgr()->GetType(inst->type_id()); |
985 | | |
986 | 73.6k | if (type->IsCooperativeMatrix()) { |
987 | 0 | return false; |
988 | 0 | } |
989 | | |
990 | 73.6k | if (!inst->IsFloatingPointFoldingAllowed()) return false; |
991 | | |
992 | 73.6k | uint32_t width = ElementWidth(type); |
993 | 73.6k | if (width != 32 && width != 64) return false; |
994 | | |
995 | 73.6k | const analysis::Constant* const_input1 = ConstInput(constants); |
996 | 73.6k | if (!const_input1 || HasZero(const_input1)) return false; |
997 | 24.0k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
998 | 24.0k | if (!other_inst->IsFloatingPointFoldingAllowed()) return false; |
999 | | |
1000 | 24.0k | bool first_is_variable = constants[0] == nullptr; |
1001 | 24.0k | if (other_inst->opcode() == inst->opcode()) { |
1002 | 2.05k | std::vector<const analysis::Constant*> other_constants = |
1003 | 2.05k | const_mgr->GetOperandConstants(other_inst); |
1004 | 2.05k | const analysis::Constant* const_input2 = ConstInput(other_constants); |
1005 | 2.05k | if (!const_input2 || HasZero(const_input2)) return false; |
1006 | | |
1007 | 1.60k | bool other_first_is_variable = other_constants[0] == nullptr; |
1008 | | |
1009 | 1.60k | spv::Op merge_op = inst->opcode(); |
1010 | 1.60k | if (other_first_is_variable) { |
1011 | | // Constants magnify. |
1012 | 845 | merge_op = spv::Op::OpFMul; |
1013 | 845 | } |
1014 | | |
1015 | | // This is an x / (*) case. Swap the inputs. Doesn't harm multiply |
1016 | | // because it is commutative. |
1017 | 1.60k | if (first_is_variable) std::swap(const_input1, const_input2); |
1018 | 1.60k | uint32_t merged_id = |
1019 | 1.60k | PerformOperation(const_mgr, merge_op, const_input1, const_input2); |
1020 | 1.60k | if (merged_id == 0) return false; |
1021 | | |
1022 | 741 | uint32_t non_const_id = other_first_is_variable |
1023 | 741 | ? other_inst->GetSingleWordInOperand(0u) |
1024 | 741 | : other_inst->GetSingleWordInOperand(1u); |
1025 | | |
1026 | 741 | spv::Op op = inst->opcode(); |
1027 | 741 | if (!first_is_variable && !other_first_is_variable) { |
1028 | | // Effectively div of 1/x, so change to multiply. |
1029 | 499 | op = spv::Op::OpFMul; |
1030 | 499 | } |
1031 | | |
1032 | 741 | uint32_t op1 = merged_id; |
1033 | 741 | uint32_t op2 = non_const_id; |
1034 | 741 | if (first_is_variable && other_first_is_variable) std::swap(op1, op2); |
1035 | 741 | inst->SetOpcode(op); |
1036 | 741 | inst->SetInOperands( |
1037 | 741 | {{SPV_OPERAND_TYPE_ID, {op1}}, {SPV_OPERAND_TYPE_ID, {op2}}}); |
1038 | 741 | return true; |
1039 | 1.60k | } |
1040 | | |
1041 | 22.0k | return false; |
1042 | 24.0k | }; |
1043 | 16.0k | } |
1044 | | |
1045 | | // Fold multiplies succeeded by divides where each instruction contains a |
1046 | | // constant operand. Does not support integer divide. |
1047 | | // Cases: |
1048 | | // 4 / (x * 2) = 2 / x |
1049 | | // 4 / (2 * x) = 2 / x |
1050 | | // (x * 4) / 2 = x * 2 |
1051 | | // (4 * x) / 2 = x * 2 |
1052 | | // (x * y) / x = y |
1053 | | // (y * x) / x = y |
1054 | 16.0k | FoldingRule MergeDivMulArithmetic() { |
1055 | 16.0k | return [](IRContext* context, Instruction* inst, |
1056 | 72.9k | const std::vector<const analysis::Constant*>& constants) { |
1057 | 72.9k | assert(inst->opcode() == spv::Op::OpFDiv); |
1058 | 72.9k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
1059 | 72.9k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
1060 | | |
1061 | 72.9k | const analysis::Type* type = |
1062 | 72.9k | context->get_type_mgr()->GetType(inst->type_id()); |
1063 | | |
1064 | 72.9k | if (type->IsCooperativeMatrix()) { |
1065 | 0 | return false; |
1066 | 0 | } |
1067 | | |
1068 | 72.9k | if (!inst->IsFloatingPointFoldingAllowed()) return false; |
1069 | | |
1070 | 72.8k | uint32_t width = ElementWidth(type); |
1071 | 72.8k | if (width != 32 && width != 64) return false; |
1072 | | |
1073 | 72.8k | uint32_t op_id = inst->GetSingleWordInOperand(0); |
1074 | 72.8k | Instruction* op_inst = def_use_mgr->GetDef(op_id); |
1075 | | |
1076 | 72.8k | if (op_inst->opcode() == spv::Op::OpFMul) { |
1077 | 3.74k | for (uint32_t i = 0; i < 2; i++) { |
1078 | 2.64k | if (op_inst->GetSingleWordInOperand(i) == |
1079 | 2.64k | inst->GetSingleWordInOperand(1)) { |
1080 | 239 | inst->SetOpcode(spv::Op::OpCopyObject); |
1081 | 239 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, |
1082 | 239 | {op_inst->GetSingleWordInOperand(1 - i)}}}); |
1083 | 239 | return true; |
1084 | 239 | } |
1085 | 2.64k | } |
1086 | 1.33k | } |
1087 | | |
1088 | 72.6k | const analysis::Constant* const_input1 = ConstInput(constants); |
1089 | 72.6k | if (!const_input1 || HasZero(const_input1)) return false; |
1090 | 23.1k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
1091 | 23.1k | if (!other_inst->IsFloatingPointFoldingAllowed()) return false; |
1092 | | |
1093 | 23.0k | bool first_is_variable = constants[0] == nullptr; |
1094 | 23.0k | if (other_inst->opcode() == spv::Op::OpFMul) { |
1095 | 1.35k | std::vector<const analysis::Constant*> other_constants = |
1096 | 1.35k | const_mgr->GetOperandConstants(other_inst); |
1097 | 1.35k | const analysis::Constant* const_input2 = ConstInput(other_constants); |
1098 | 1.35k | if (!const_input2) return false; |
1099 | | |
1100 | 991 | bool other_first_is_variable = other_constants[0] == nullptr; |
1101 | | |
1102 | | // This is an x / (*) case. Swap the inputs. |
1103 | 991 | if (first_is_variable) std::swap(const_input1, const_input2); |
1104 | 991 | uint32_t merged_id = PerformOperation(const_mgr, inst->opcode(), |
1105 | 991 | const_input1, const_input2); |
1106 | 991 | if (merged_id == 0) return false; |
1107 | | |
1108 | 658 | uint32_t non_const_id = other_first_is_variable |
1109 | 658 | ? other_inst->GetSingleWordInOperand(0u) |
1110 | 658 | : other_inst->GetSingleWordInOperand(1u); |
1111 | | |
1112 | 658 | uint32_t op1 = merged_id; |
1113 | 658 | uint32_t op2 = non_const_id; |
1114 | 658 | if (first_is_variable) std::swap(op1, op2); |
1115 | | |
1116 | | // Convert to multiply |
1117 | 658 | if (first_is_variable) inst->SetOpcode(other_inst->opcode()); |
1118 | 658 | inst->SetInOperands( |
1119 | 658 | {{SPV_OPERAND_TYPE_ID, {op1}}, {SPV_OPERAND_TYPE_ID, {op2}}}); |
1120 | 658 | return true; |
1121 | 991 | } |
1122 | | |
1123 | 21.7k | return false; |
1124 | 23.0k | }; |
1125 | 16.0k | } |
1126 | | |
1127 | | // Fold divides of a constant and a negation. |
1128 | | // Cases: |
1129 | | // (-x) / 2 = x / -2 |
1130 | | // 2 / (-x) = -2 / x |
1131 | 16.0k | FoldingRule MergeDivNegateArithmetic() { |
1132 | 16.0k | return [](IRContext* context, Instruction* inst, |
1133 | 72.0k | const std::vector<const analysis::Constant*>& constants) { |
1134 | 72.0k | assert(inst->opcode() == spv::Op::OpFDiv); |
1135 | 72.0k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
1136 | 72.0k | if (!inst->IsFloatingPointFoldingAllowed()) return false; |
1137 | | |
1138 | 71.9k | const analysis::Constant* const_input1 = ConstInput(constants); |
1139 | 71.9k | if (!const_input1) return false; |
1140 | 44.4k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
1141 | 44.4k | if (!other_inst->IsFloatingPointFoldingAllowed()) return false; |
1142 | | |
1143 | 44.3k | bool first_is_variable = constants[0] == nullptr; |
1144 | 44.3k | if (other_inst->opcode() == spv::Op::OpFNegate) { |
1145 | 39 | uint32_t neg_id = NegateConstant(const_mgr, const_input1); |
1146 | | |
1147 | 39 | if (first_is_variable) { |
1148 | 24 | inst->SetInOperands( |
1149 | 24 | {{SPV_OPERAND_TYPE_ID, {other_inst->GetSingleWordInOperand(0u)}}, |
1150 | 24 | {SPV_OPERAND_TYPE_ID, {neg_id}}}); |
1151 | 24 | } else { |
1152 | 15 | inst->SetInOperands( |
1153 | 15 | {{SPV_OPERAND_TYPE_ID, {neg_id}}, |
1154 | 15 | {SPV_OPERAND_TYPE_ID, {other_inst->GetSingleWordInOperand(0u)}}}); |
1155 | 15 | } |
1156 | 39 | return true; |
1157 | 39 | } |
1158 | | |
1159 | 44.3k | return false; |
1160 | 44.3k | }; |
1161 | 16.0k | } |
1162 | | |
1163 | | // Folds addition, where one side is a negation. |
1164 | | // (-x) + y = y - x |
1165 | | // y + (-x) = y - x |
1166 | 32.0k | FoldingRule MergeAddNegateArithmetic() { |
1167 | 32.0k | return [](IRContext* context, Instruction* inst, |
1168 | 492k | const std::vector<const analysis::Constant*>&) { |
1169 | 492k | assert(inst->opcode() == spv::Op::OpFAdd || |
1170 | 492k | inst->opcode() == spv::Op::OpIAdd); |
1171 | 492k | const analysis::Type* type = |
1172 | 492k | context->get_type_mgr()->GetType(inst->type_id()); |
1173 | 492k | bool uses_float = HasFloatingPoint(type); |
1174 | 492k | if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
1175 | | |
1176 | 486k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
1177 | 486k | Instruction* lhs = def_use_mgr->GetDef(inst->GetSingleWordInOperand(0)); |
1178 | 486k | Instruction* rhs = def_use_mgr->GetDef(inst->GetSingleWordInOperand(1)); |
1179 | | |
1180 | 486k | auto TrySubstitute = [inst, uses_float](Instruction* first, |
1181 | 972k | Instruction* second) { |
1182 | 972k | if (IsFoldableNegation(first)) { |
1183 | 62 | inst->SetOpcode(uses_float ? spv::Op::OpFSub : spv::Op::OpISub); |
1184 | 62 | inst->SetInOperands( |
1185 | 62 | {{SPV_OPERAND_TYPE_ID, {second->result_id()}}, |
1186 | 62 | {SPV_OPERAND_TYPE_ID, {first->GetSingleWordInOperand(0u)}}}); |
1187 | 62 | return true; |
1188 | 62 | } |
1189 | 971k | return false; |
1190 | 972k | }; |
1191 | | |
1192 | 486k | return TrySubstitute(lhs, rhs) || TrySubstitute(rhs, lhs); |
1193 | 492k | }; |
1194 | 32.0k | } |
1195 | | |
1196 | | // Folds subtraction, where one side is a negation. |
1197 | | // Cases: |
1198 | | // (-x) - 2 = -2 - x |
1199 | | // y - (-x) = x + y |
1200 | 32.0k | FoldingRule MergeSubNegateArithmetic() { |
1201 | 32.0k | return [](IRContext* context, Instruction* inst, |
1202 | 140k | const std::vector<const analysis::Constant*>& constants) { |
1203 | 140k | assert(inst->opcode() == spv::Op::OpFSub || |
1204 | 140k | inst->opcode() == spv::Op::OpISub); |
1205 | 140k | const analysis::Type* type = |
1206 | 140k | context->get_type_mgr()->GetType(inst->type_id()); |
1207 | | |
1208 | 140k | bool uses_float = HasFloatingPoint(type); |
1209 | 140k | if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
1210 | | |
1211 | 135k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
1212 | 135k | Instruction* lhs = def_use_mgr->GetDef(inst->GetSingleWordInOperand(0)); |
1213 | 135k | Instruction* rhs = def_use_mgr->GetDef(inst->GetSingleWordInOperand(1)); |
1214 | | |
1215 | 135k | if (IsFoldableNegation(rhs)) { |
1216 | 650 | inst->SetOpcode(uses_float ? spv::Op::OpFAdd : spv::Op::OpIAdd); |
1217 | 650 | inst->SetInOperands( |
1218 | 650 | {{SPV_OPERAND_TYPE_ID, {lhs->result_id()}}, |
1219 | 650 | {SPV_OPERAND_TYPE_ID, {rhs->GetSingleWordInOperand(0)}}}); |
1220 | 650 | return true; |
1221 | 650 | } |
1222 | | |
1223 | 135k | if (type->IsCooperativeMatrix()) { |
1224 | 0 | return false; |
1225 | 0 | } |
1226 | | |
1227 | 135k | uint32_t width = ElementWidth(type); |
1228 | 135k | if (width != 32 && width != 64) return false; |
1229 | | |
1230 | 135k | if (constants[1] && IsFoldableNegation(lhs)) { |
1231 | 39 | inst->SetInOperands( |
1232 | 39 | {{SPV_OPERAND_TYPE_ID, |
1233 | 39 | {NegateConstant(context->get_constant_mgr(), constants[1])}}, |
1234 | 39 | {SPV_OPERAND_TYPE_ID, {lhs->GetSingleWordInOperand(0)}}}); |
1235 | 39 | return true; |
1236 | 39 | } |
1237 | 135k | return false; |
1238 | 135k | }; |
1239 | 32.0k | } |
1240 | | |
1241 | | // Folds addition of an addition where each operation has a constant operand. |
1242 | | // Cases: |
1243 | | // (x + 2) + 2 = x + 4 |
1244 | | // (2 + x) + 2 = x + 4 |
1245 | | // 2 + (x + 2) = x + 4 |
1246 | | // 2 + (2 + x) = x + 4 |
1247 | 32.0k | FoldingRule MergeAddAddArithmetic() { |
1248 | 32.0k | return [](IRContext* context, Instruction* inst, |
1249 | 492k | const std::vector<const analysis::Constant*>& constants) { |
1250 | 492k | assert(inst->opcode() == spv::Op::OpFAdd || |
1251 | 492k | inst->opcode() == spv::Op::OpIAdd); |
1252 | 492k | const analysis::Type* type = |
1253 | 492k | context->get_type_mgr()->GetType(inst->type_id()); |
1254 | | |
1255 | 492k | if (type->IsCooperativeMatrix()) { |
1256 | 0 | return false; |
1257 | 0 | } |
1258 | | |
1259 | 492k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
1260 | 492k | bool uses_float = HasFloatingPoint(type); |
1261 | 492k | if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
1262 | | |
1263 | 485k | uint32_t width = ElementWidth(type); |
1264 | 485k | if (width != 32 && width != 64) return false; |
1265 | | |
1266 | 485k | const analysis::Constant* const_input1 = ConstInput(constants); |
1267 | 485k | if (!const_input1) return false; |
1268 | 145k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
1269 | 145k | if (uses_float && !other_inst->IsFloatingPointFoldingAllowed()) |
1270 | 373 | return false; |
1271 | | |
1272 | 145k | if (other_inst->opcode() == spv::Op::OpFAdd || |
1273 | 133k | other_inst->opcode() == spv::Op::OpIAdd) { |
1274 | 15.3k | std::vector<const analysis::Constant*> other_constants = |
1275 | 15.3k | const_mgr->GetOperandConstants(other_inst); |
1276 | 15.3k | const analysis::Constant* const_input2 = ConstInput(other_constants); |
1277 | 15.3k | if (!const_input2) return false; |
1278 | | |
1279 | 12.9k | Instruction* non_const_input = |
1280 | 12.9k | NonConstInput(context, other_constants[0], other_inst); |
1281 | 12.9k | uint32_t merged_id = PerformOperation(const_mgr, inst->opcode(), |
1282 | 12.9k | const_input1, const_input2); |
1283 | 12.9k | if (merged_id == 0) return false; |
1284 | | |
1285 | 9.74k | inst->SetInOperands( |
1286 | 9.74k | {{SPV_OPERAND_TYPE_ID, {non_const_input->result_id()}}, |
1287 | 9.74k | {SPV_OPERAND_TYPE_ID, {merged_id}}}); |
1288 | 9.74k | return true; |
1289 | 12.9k | } |
1290 | 130k | return false; |
1291 | 145k | }; |
1292 | 32.0k | } |
1293 | | |
1294 | | // Folds addition of a subtraction where each operation has a constant operand. |
1295 | | // Cases: |
1296 | | // (x - 2) + 2 = x + 0 |
1297 | | // (2 - x) + 2 = 4 - x |
1298 | | // 2 + (x - 2) = x + 0 |
1299 | | // 2 + (2 - x) = 4 - x |
1300 | 32.0k | FoldingRule MergeAddSubArithmetic() { |
1301 | 32.0k | return [](IRContext* context, Instruction* inst, |
1302 | 482k | const std::vector<const analysis::Constant*>& constants) { |
1303 | 482k | assert(inst->opcode() == spv::Op::OpFAdd || |
1304 | 482k | inst->opcode() == spv::Op::OpIAdd); |
1305 | 482k | const analysis::Type* type = |
1306 | 482k | context->get_type_mgr()->GetType(inst->type_id()); |
1307 | | |
1308 | 482k | if (type->IsCooperativeMatrix()) { |
1309 | 0 | return false; |
1310 | 0 | } |
1311 | | |
1312 | 482k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
1313 | 482k | bool uses_float = HasFloatingPoint(type); |
1314 | 482k | if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
1315 | | |
1316 | 476k | uint32_t width = ElementWidth(type); |
1317 | 476k | if (width != 32 && width != 64) return false; |
1318 | | |
1319 | 476k | const analysis::Constant* const_input1 = ConstInput(constants); |
1320 | 476k | if (!const_input1) return false; |
1321 | 136k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
1322 | 136k | if (uses_float && !other_inst->IsFloatingPointFoldingAllowed()) |
1323 | 373 | return false; |
1324 | | |
1325 | 135k | if (other_inst->opcode() == spv::Op::OpFSub || |
1326 | 133k | other_inst->opcode() == spv::Op::OpISub) { |
1327 | 3.17k | std::vector<const analysis::Constant*> other_constants = |
1328 | 3.17k | const_mgr->GetOperandConstants(other_inst); |
1329 | 3.17k | const analysis::Constant* const_input2 = ConstInput(other_constants); |
1330 | 3.17k | if (!const_input2) return false; |
1331 | | |
1332 | 2.99k | bool first_is_variable = other_constants[0] == nullptr; |
1333 | 2.99k | spv::Op op = inst->opcode(); |
1334 | 2.99k | uint32_t op1 = 0; |
1335 | 2.99k | uint32_t op2 = 0; |
1336 | 2.99k | if (first_is_variable) { |
1337 | | // Subtract constants. Non-constant operand is first. |
1338 | 2.70k | op1 = other_inst->GetSingleWordInOperand(0u); |
1339 | 2.70k | op2 = PerformOperation(const_mgr, other_inst->opcode(), const_input1, |
1340 | 2.70k | const_input2); |
1341 | 2.70k | } else { |
1342 | | // Add constants. Constant operand is first. Change the opcode. |
1343 | 291 | op1 = PerformOperation(const_mgr, inst->opcode(), const_input1, |
1344 | 291 | const_input2); |
1345 | 291 | op2 = other_inst->GetSingleWordInOperand(1u); |
1346 | 291 | op = other_inst->opcode(); |
1347 | 291 | } |
1348 | 2.99k | if (op1 == 0 || op2 == 0) return false; |
1349 | | |
1350 | 1.86k | inst->SetOpcode(op); |
1351 | 1.86k | inst->SetInOperands( |
1352 | 1.86k | {{SPV_OPERAND_TYPE_ID, {op1}}, {SPV_OPERAND_TYPE_ID, {op2}}}); |
1353 | 1.86k | return true; |
1354 | 2.99k | } |
1355 | 132k | return false; |
1356 | 135k | }; |
1357 | 32.0k | } |
1358 | | |
1359 | | // Folds subtraction of an addition where each operand has a constant operand. |
1360 | | // Cases: |
1361 | | // (x + 2) - 2 = x + 0 |
1362 | | // (2 + x) - 2 = x + 0 |
1363 | | // 2 - (x + 2) = 0 - x |
1364 | | // 2 - (2 + x) = 0 - x |
1365 | 32.0k | FoldingRule MergeSubAddArithmetic() { |
1366 | 32.0k | return [](IRContext* context, Instruction* inst, |
1367 | 140k | const std::vector<const analysis::Constant*>& constants) { |
1368 | 140k | assert(inst->opcode() == spv::Op::OpFSub || |
1369 | 140k | inst->opcode() == spv::Op::OpISub); |
1370 | 140k | const analysis::Type* type = |
1371 | 140k | context->get_type_mgr()->GetType(inst->type_id()); |
1372 | | |
1373 | 140k | if (type->IsCooperativeMatrix()) { |
1374 | 0 | return false; |
1375 | 0 | } |
1376 | | |
1377 | 140k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
1378 | 140k | bool uses_float = HasFloatingPoint(type); |
1379 | 140k | if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
1380 | | |
1381 | 135k | uint32_t width = ElementWidth(type); |
1382 | 135k | if (width != 32 && width != 64) return false; |
1383 | | |
1384 | 135k | const analysis::Constant* const_input1 = ConstInput(constants); |
1385 | 135k | if (!const_input1) return false; |
1386 | 82.6k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
1387 | 82.6k | if (uses_float && !other_inst->IsFloatingPointFoldingAllowed()) |
1388 | 11 | return false; |
1389 | | |
1390 | 82.6k | if (other_inst->opcode() == spv::Op::OpFAdd || |
1391 | 79.0k | other_inst->opcode() == spv::Op::OpIAdd) { |
1392 | 11.1k | std::vector<const analysis::Constant*> other_constants = |
1393 | 11.1k | const_mgr->GetOperandConstants(other_inst); |
1394 | 11.1k | const analysis::Constant* const_input2 = ConstInput(other_constants); |
1395 | 11.1k | if (!const_input2) return false; |
1396 | | |
1397 | 3.19k | Instruction* non_const_input = |
1398 | 3.19k | NonConstInput(context, other_constants[0], other_inst); |
1399 | | |
1400 | | // If the first operand of the sub is not a constant, swap the constants |
1401 | | // so the subtraction has the correct operands. |
1402 | 3.19k | if (constants[0] == nullptr) std::swap(const_input1, const_input2); |
1403 | | // Subtract the constants. |
1404 | 3.19k | uint32_t merged_id = PerformOperation(const_mgr, inst->opcode(), |
1405 | 3.19k | const_input1, const_input2); |
1406 | 3.19k | spv::Op op = inst->opcode(); |
1407 | 3.19k | uint32_t op1 = 0; |
1408 | 3.19k | uint32_t op2 = 0; |
1409 | 3.19k | if (constants[0] == nullptr) { |
1410 | | // Non-constant operand is first. Change the opcode. |
1411 | 1.88k | op1 = non_const_input->result_id(); |
1412 | 1.88k | op2 = merged_id; |
1413 | 1.88k | op = other_inst->opcode(); |
1414 | 1.88k | } else { |
1415 | | // Constant operand is first. |
1416 | 1.31k | op1 = merged_id; |
1417 | 1.31k | op2 = non_const_input->result_id(); |
1418 | 1.31k | } |
1419 | 3.19k | if (op1 == 0 || op2 == 0) return false; |
1420 | | |
1421 | 2.61k | inst->SetOpcode(op); |
1422 | 2.61k | inst->SetInOperands( |
1423 | 2.61k | {{SPV_OPERAND_TYPE_ID, {op1}}, {SPV_OPERAND_TYPE_ID, {op2}}}); |
1424 | 2.61k | return true; |
1425 | 3.19k | } |
1426 | 71.4k | return false; |
1427 | 82.6k | }; |
1428 | 32.0k | } |
1429 | | |
1430 | | // Folds subtraction of a subtraction where each operand has a constant operand. |
1431 | | // Cases: |
1432 | | // (x - 2) - 2 = x - 4 |
1433 | | // (2 - x) - 2 = 0 - x |
1434 | | // 2 - (x - 2) = 4 - x |
1435 | | // 2 - (2 - x) = x + 0 |
1436 | 32.0k | FoldingRule MergeSubSubArithmetic() { |
1437 | 32.0k | return [](IRContext* context, Instruction* inst, |
1438 | 137k | const std::vector<const analysis::Constant*>& constants) { |
1439 | 137k | assert(inst->opcode() == spv::Op::OpFSub || |
1440 | 137k | inst->opcode() == spv::Op::OpISub); |
1441 | 137k | const analysis::Type* type = |
1442 | 137k | context->get_type_mgr()->GetType(inst->type_id()); |
1443 | | |
1444 | 137k | if (type->IsCooperativeMatrix()) { |
1445 | 0 | return false; |
1446 | 0 | } |
1447 | | |
1448 | 137k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
1449 | 137k | bool uses_float = HasFloatingPoint(type); |
1450 | 137k | if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
1451 | | |
1452 | 132k | uint32_t width = ElementWidth(type); |
1453 | 132k | if (width != 32 && width != 64) return false; |
1454 | | |
1455 | 132k | const analysis::Constant* const_input1 = ConstInput(constants); |
1456 | 132k | if (!const_input1) return false; |
1457 | 80.0k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
1458 | 80.0k | if (uses_float && !other_inst->IsFloatingPointFoldingAllowed()) |
1459 | 11 | return false; |
1460 | | |
1461 | 80.0k | if (other_inst->opcode() == spv::Op::OpFSub || |
1462 | 72.2k | other_inst->opcode() == spv::Op::OpISub) { |
1463 | 8.72k | std::vector<const analysis::Constant*> other_constants = |
1464 | 8.72k | const_mgr->GetOperandConstants(other_inst); |
1465 | 8.72k | const analysis::Constant* const_input2 = ConstInput(other_constants); |
1466 | 8.72k | if (!const_input2) return false; |
1467 | | |
1468 | 8.23k | Instruction* non_const_input = |
1469 | 8.23k | NonConstInput(context, other_constants[0], other_inst); |
1470 | | |
1471 | | // Merge the constants. |
1472 | 8.23k | uint32_t merged_id = 0; |
1473 | 8.23k | spv::Op merge_op = inst->opcode(); |
1474 | 8.23k | if (other_constants[0] == nullptr) { |
1475 | 6.64k | merge_op = uses_float ? spv::Op::OpFAdd : spv::Op::OpIAdd; |
1476 | 6.64k | } else if (constants[0] == nullptr) { |
1477 | 470 | std::swap(const_input1, const_input2); |
1478 | 470 | } |
1479 | 8.23k | merged_id = |
1480 | 8.23k | PerformOperation(const_mgr, merge_op, const_input1, const_input2); |
1481 | 8.23k | if (merged_id == 0) return false; |
1482 | | |
1483 | 5.72k | spv::Op op = inst->opcode(); |
1484 | 5.72k | if (constants[0] != nullptr && other_constants[0] != nullptr) { |
1485 | | // Change the operation. |
1486 | 818 | op = uses_float ? spv::Op::OpFAdd : spv::Op::OpIAdd; |
1487 | 818 | } |
1488 | | |
1489 | 5.72k | uint32_t op1 = 0; |
1490 | 5.72k | uint32_t op2 = 0; |
1491 | 5.72k | if ((constants[0] == nullptr) ^ (other_constants[0] == nullptr)) { |
1492 | 612 | op1 = merged_id; |
1493 | 612 | op2 = non_const_input->result_id(); |
1494 | 5.11k | } else { |
1495 | 5.11k | op1 = non_const_input->result_id(); |
1496 | 5.11k | op2 = merged_id; |
1497 | 5.11k | } |
1498 | | |
1499 | 5.72k | inst->SetOpcode(op); |
1500 | 5.72k | inst->SetInOperands( |
1501 | 5.72k | {{SPV_OPERAND_TYPE_ID, {op1}}, {SPV_OPERAND_TYPE_ID, {op2}}}); |
1502 | 5.72k | return true; |
1503 | 8.23k | } |
1504 | 71.2k | return false; |
1505 | 80.0k | }; |
1506 | 32.0k | } |
1507 | | |
1508 | | // Helper function for MergeGenericAddSubArithmetic. If |addend| and |
1509 | | // subtrahend of |sub| is the same, merge to copy of minuend of |sub|. |
1510 | 948k | bool MergeGenericAddendSub(uint32_t addend, uint32_t sub, Instruction* inst) { |
1511 | 948k | IRContext* context = inst->context(); |
1512 | 948k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
1513 | 948k | Instruction* sub_inst = def_use_mgr->GetDef(sub); |
1514 | 948k | if (sub_inst->opcode() != spv::Op::OpFSub && |
1515 | 940k | sub_inst->opcode() != spv::Op::OpISub) |
1516 | 938k | return false; |
1517 | 9.59k | if (sub_inst->opcode() == spv::Op::OpFSub && |
1518 | 7.79k | !sub_inst->IsFloatingPointFoldingAllowed()) |
1519 | 0 | return false; |
1520 | 9.59k | if (addend != sub_inst->GetSingleWordInOperand(1)) return false; |
1521 | 1.29k | inst->SetOpcode(spv::Op::OpCopyObject); |
1522 | 1.29k | inst->SetInOperands( |
1523 | 1.29k | {{SPV_OPERAND_TYPE_ID, {sub_inst->GetSingleWordInOperand(0)}}}); |
1524 | 1.29k | context->UpdateDefUse(inst); |
1525 | 1.29k | return true; |
1526 | 9.59k | } |
1527 | | |
1528 | | // Folds addition of a subtraction where the subtrahend is equal to the |
1529 | | // other addend. Return a copy of the minuend. Accepts generic (const and |
1530 | | // non-const) operands. |
1531 | | // Cases: |
1532 | | // (a - b) + b = a |
1533 | | // b + (a - b) = a |
1534 | 32.0k | FoldingRule MergeGenericAddSubArithmetic() { |
1535 | 32.0k | return [](IRContext* context, Instruction* inst, |
1536 | 480k | const std::vector<const analysis::Constant*>&) { |
1537 | 480k | assert(inst->opcode() == spv::Op::OpFAdd || |
1538 | 480k | inst->opcode() == spv::Op::OpIAdd); |
1539 | 480k | const analysis::Type* type = |
1540 | 480k | context->get_type_mgr()->GetType(inst->type_id()); |
1541 | | |
1542 | 480k | if (type->IsCooperativeMatrix()) { |
1543 | 0 | return false; |
1544 | 0 | } |
1545 | | |
1546 | 480k | bool uses_float = HasFloatingPoint(type); |
1547 | 480k | if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
1548 | | |
1549 | 474k | uint32_t width = ElementWidth(type); |
1550 | 474k | if (width != 32 && width != 64) return false; |
1551 | | |
1552 | 474k | uint32_t add_op0 = inst->GetSingleWordInOperand(0); |
1553 | 474k | uint32_t add_op1 = inst->GetSingleWordInOperand(1); |
1554 | 474k | if (MergeGenericAddendSub(add_op0, add_op1, inst)) return true; |
1555 | 474k | return MergeGenericAddendSub(add_op1, add_op0, inst); |
1556 | 474k | }; |
1557 | 32.0k | } |
1558 | | |
1559 | | // Helper function for FactorAddSubMuls. |
1560 | | // If |factor0_0| is the same as |factor1_0|, generate: |
1561 | | // |factor0_0| * (|factor0_1| + |factor1_1|) |
1562 | | // |factor0_0| * (|factor0_1| - |factor1_1|) |
1563 | | bool FactorAddSubMulsOpnds(uint32_t factor0_0, uint32_t factor0_1, |
1564 | | uint32_t factor1_0, uint32_t factor1_1, |
1565 | 5.11k | Instruction* inst) { |
1566 | 5.11k | IRContext* context = inst->context(); |
1567 | 5.11k | if (factor0_0 != factor1_0) return false; |
1568 | 238 | InstructionBuilder ir_builder( |
1569 | 238 | context, inst, |
1570 | 238 | IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping); |
1571 | 238 | Instruction* new_add_inst = ir_builder.AddBinaryOp( |
1572 | 238 | inst->type_id(), inst->opcode(), factor0_1, factor1_1); |
1573 | 238 | if (!new_add_inst) { |
1574 | 0 | return false; |
1575 | 0 | } |
1576 | | |
1577 | 238 | bool is_float = |
1578 | 238 | inst->opcode() == spv::Op::OpFAdd || inst->opcode() == spv::Op::OpFSub; |
1579 | 238 | inst->SetOpcode(is_float ? spv::Op::OpFMul : spv::Op::OpIMul); |
1580 | 238 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {factor0_0}}, |
1581 | 238 | {SPV_OPERAND_TYPE_ID, {new_add_inst->result_id()}}}); |
1582 | 238 | context->UpdateDefUse(inst); |
1583 | 238 | return true; |
1584 | 238 | } |
1585 | | |
1586 | | // Perform the following factoring identity, handling all operand order |
1587 | | // combinations: |
1588 | | // (a * b) + (a * c) = a * (b + c) |
1589 | | // (a * b) - (a * c) = a * (b - c) |
1590 | 64.0k | FoldingRule FactorAddSubMuls() { |
1591 | 64.0k | return [](IRContext* context, Instruction* inst, |
1592 | 609k | const std::vector<const analysis::Constant*>&) { |
1593 | 609k | assert(inst->opcode() == spv::Op::OpFAdd || |
1594 | 609k | inst->opcode() == spv::Op::OpFSub || |
1595 | 609k | inst->opcode() == spv::Op::OpIAdd || |
1596 | 609k | inst->opcode() == spv::Op::OpISub); |
1597 | 609k | const analysis::Type* type = |
1598 | 609k | context->get_type_mgr()->GetType(inst->type_id()); |
1599 | 609k | bool uses_float = HasFloatingPoint(type); |
1600 | 609k | if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
1601 | | |
1602 | 598k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
1603 | 598k | uint32_t add_op0 = inst->GetSingleWordInOperand(0); |
1604 | 598k | Instruction* add_op0_inst = def_use_mgr->GetDef(add_op0); |
1605 | 598k | if (add_op0_inst->opcode() != spv::Op::OpFMul && |
1606 | 580k | add_op0_inst->opcode() != spv::Op::OpIMul) |
1607 | 576k | return false; |
1608 | 22.0k | uint32_t add_op1 = inst->GetSingleWordInOperand(1); |
1609 | 22.0k | Instruction* add_op1_inst = def_use_mgr->GetDef(add_op1); |
1610 | 22.0k | if (add_op1_inst->opcode() != spv::Op::OpFMul && |
1611 | 19.9k | add_op1_inst->opcode() != spv::Op::OpIMul) |
1612 | 19.2k | return false; |
1613 | | |
1614 | | // Only perform this optimization if both of the muls only have one use. |
1615 | | // Otherwise this is a deoptimization in size and performance. |
1616 | 2.73k | if (def_use_mgr->NumUses(add_op0_inst) > 1) return false; |
1617 | 1.52k | if (def_use_mgr->NumUses(add_op1_inst) > 1) return false; |
1618 | | |
1619 | 1.40k | if (add_op0_inst->opcode() == spv::Op::OpFMul && |
1620 | 1.32k | (!add_op0_inst->IsFloatingPointFoldingAllowed() || |
1621 | 1.32k | !add_op1_inst->IsFloatingPointFoldingAllowed())) |
1622 | 0 | return false; |
1623 | | |
1624 | 3.76k | for (int i = 0; i < 2; i++) { |
1625 | 7.47k | for (int j = 0; j < 2; j++) { |
1626 | | // Check if operand i in add_op0_inst matches operand j in add_op1_inst. |
1627 | 5.11k | if (FactorAddSubMulsOpnds(add_op0_inst->GetSingleWordInOperand(i), |
1628 | 5.11k | add_op0_inst->GetSingleWordInOperand(1 - i), |
1629 | 5.11k | add_op1_inst->GetSingleWordInOperand(j), |
1630 | 5.11k | add_op1_inst->GetSingleWordInOperand(1 - j), |
1631 | 5.11k | inst)) |
1632 | 238 | return true; |
1633 | 5.11k | } |
1634 | 2.60k | } |
1635 | 1.16k | return false; |
1636 | 1.40k | }; |
1637 | 64.0k | } |
1638 | | |
1639 | | // Reassociate integer instructions where both operands share the same opcode |
1640 | | // and both source instructions contain a constant. |
1641 | | // e.g: |
1642 | | // (a * C0) * (C1 * b) = (C0 * C1) * (a * b) |
1643 | | // (a ^ C0) ^ (b ^ C1) = (C0 ^ C1) ^ (a ^ b) |
1644 | | // (C0 | a) | (b | C1) = (C0 | C1) | (a | b) |
1645 | | // (a & C0) & (b & C1) = (C0 & C1) & (a & b) |
1646 | | static const constexpr spv::Op ReassociateNestedGenericIntOps[] = { |
1647 | | spv::Op::OpIMul, spv::Op::OpBitwiseOr, spv::Op::OpBitwiseXor, |
1648 | | spv::Op::OpBitwiseAnd}; |
1649 | | |
1650 | 64.0k | FoldingRule ReassociateNestedGenericInt(spv::Op opcode) { |
1651 | 64.0k | assert(std::find(std::begin(ReassociateNestedGenericIntOps), |
1652 | 64.0k | std::end(ReassociateNestedGenericIntOps), |
1653 | 64.0k | opcode) != std::end(ReassociateNestedGenericIntOps) && |
1654 | 64.0k | "Wrong opcode."); |
1655 | | |
1656 | 64.0k | return [opcode](IRContext* context, Instruction* inst, |
1657 | 64.0k | const std::vector<const analysis::Constant*>& constants) { |
1658 | | // Handled by other folding rules. |
1659 | 61.7k | if (constants[0] || constants[1]) { |
1660 | 32.4k | return false; |
1661 | 32.4k | } |
1662 | | |
1663 | 29.3k | if (inst->opcode() != opcode) { |
1664 | 0 | return false; |
1665 | 0 | } |
1666 | | |
1667 | 29.3k | const analysis::Type* type = |
1668 | 29.3k | context->get_type_mgr()->GetType(inst->type_id()); |
1669 | | |
1670 | 29.3k | if (type->IsCooperativeMatrix()) { |
1671 | 0 | return false; |
1672 | 0 | } |
1673 | | |
1674 | 29.3k | uint32_t width = ElementWidth(type); |
1675 | 29.3k | if (width != 32 && width != 64) return false; |
1676 | | |
1677 | 29.3k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
1678 | 29.3k | Instruction* lhs = def_use_mgr->GetDef(inst->GetSingleWordInOperand(0)); |
1679 | 29.3k | Instruction* rhs = def_use_mgr->GetDef(inst->GetSingleWordInOperand(1)); |
1680 | | |
1681 | 29.3k | if (lhs->opcode() != opcode || rhs->opcode() != opcode) { |
1682 | 28.1k | return false; |
1683 | 28.1k | } |
1684 | | |
1685 | 1.13k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
1686 | 1.13k | std::vector<const analysis::Constant*> lhs_constants = |
1687 | 1.13k | const_mgr->GetOperandConstants(lhs); |
1688 | 1.13k | const analysis::Constant* lhs_const = ConstInput(lhs_constants); |
1689 | 1.13k | if (!lhs_const) { |
1690 | 806 | return false; |
1691 | 806 | } |
1692 | | |
1693 | 330 | std::vector<const analysis::Constant*> rhs_constants = |
1694 | 330 | const_mgr->GetOperandConstants(rhs); |
1695 | 330 | const analysis::Constant* rhs_const = ConstInput(rhs_constants); |
1696 | 330 | if (!rhs_const) { |
1697 | 129 | return false; |
1698 | 129 | } |
1699 | | |
1700 | 201 | uint32_t merged_constant = |
1701 | 201 | PerformOperation(const_mgr, opcode, lhs_const, rhs_const); |
1702 | 201 | if (!merged_constant) { |
1703 | 0 | return false; |
1704 | 0 | } |
1705 | | |
1706 | 201 | InstructionBuilder ir_builder( |
1707 | 201 | context, inst, |
1708 | 201 | IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping); |
1709 | | |
1710 | 201 | Instruction* new_rhs = ir_builder.AddBinaryOp( |
1711 | 201 | inst->type_id(), opcode, |
1712 | 201 | NonConstInput(context, lhs_constants[0], lhs)->result_id(), |
1713 | 201 | NonConstInput(context, rhs_constants[0], rhs)->result_id()); |
1714 | | |
1715 | 201 | if (!new_rhs) { |
1716 | 0 | return false; |
1717 | 0 | } |
1718 | | |
1719 | 201 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {merged_constant}}, |
1720 | 201 | {SPV_OPERAND_TYPE_ID, {new_rhs->result_id()}}}); |
1721 | 201 | return true; |
1722 | 201 | }; |
1723 | 64.0k | } |
1724 | | |
1725 | | // Reassociate floating point mul/div instructions, which have mul/div inputs, |
1726 | | // both of which contain a constant. |
1727 | | // e.g: |
1728 | | // (a * C0) / (C1 / b) = (C0 / C1) * (a * b) |
1729 | | // (C0 / a) * (b / C1) = (C0 / C1) * (b / a) |
1730 | | // (a / C0) / (b * C1) = (1 / (C0 * C1)) * (a / b) |
1731 | 32.0k | FoldingRule ReassociateNestedMulDivFloat() { |
1732 | 32.0k | return [](IRContext* context, Instruction* inst, |
1733 | 260k | const std::vector<const analysis::Constant*>& constants) { |
1734 | 260k | assert(inst->opcode() == spv::Op::OpFMul || |
1735 | 260k | inst->opcode() == spv::Op::OpFDiv); |
1736 | | |
1737 | | // Handled by other folding rules. |
1738 | 260k | if (constants[0] || constants[1]) { |
1739 | 165k | return false; |
1740 | 165k | } |
1741 | | |
1742 | 94.6k | const analysis::Type* type = |
1743 | 94.6k | context->get_type_mgr()->GetType(inst->type_id()); |
1744 | | |
1745 | 94.6k | if (type->IsCooperativeMatrix()) { |
1746 | 0 | return false; |
1747 | 0 | } |
1748 | | |
1749 | 94.6k | uint32_t width = ElementWidth(type); |
1750 | 94.6k | if (width != 32 && width != 64) return false; |
1751 | | |
1752 | 94.6k | if (!inst->IsFloatingPointFoldingAllowed()) return false; |
1753 | | |
1754 | 94.6k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
1755 | 94.6k | Instruction* lhs = def_use_mgr->GetDef(inst->GetSingleWordInOperand(0)); |
1756 | 94.6k | Instruction* rhs = def_use_mgr->GetDef(inst->GetSingleWordInOperand(1)); |
1757 | | |
1758 | 94.6k | bool lhs_is_mul = lhs->opcode() == spv::Op::OpFMul; |
1759 | 94.6k | bool lhs_is_div = lhs->opcode() == spv::Op::OpFDiv; |
1760 | 94.6k | bool rhs_is_mul = rhs->opcode() == spv::Op::OpFMul; |
1761 | 94.6k | bool rhs_is_div = rhs->opcode() == spv::Op::OpFDiv; |
1762 | 94.6k | if (!(lhs_is_mul || lhs_is_div) || !(rhs_is_mul || rhs_is_div)) { |
1763 | 91.2k | return false; |
1764 | 91.2k | } |
1765 | | |
1766 | 3.46k | if (!lhs->IsFloatingPointFoldingAllowed() || |
1767 | 3.46k | !rhs->IsFloatingPointFoldingAllowed()) { |
1768 | 0 | return false; |
1769 | 0 | } |
1770 | | |
1771 | 3.46k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
1772 | 3.46k | std::vector<const analysis::Constant*> lhs_constants = |
1773 | 3.46k | const_mgr->GetOperandConstants(lhs); |
1774 | 3.46k | if (!lhs_constants[0] && !lhs_constants[1]) { |
1775 | 2.36k | return false; |
1776 | 2.36k | } |
1777 | | |
1778 | 1.09k | std::vector<const analysis::Constant*> rhs_constants = |
1779 | 1.09k | const_mgr->GetOperandConstants(rhs); |
1780 | 1.09k | if (!rhs_constants[0] && !rhs_constants[1]) { |
1781 | 214 | return false; |
1782 | 214 | } |
1783 | | |
1784 | 885 | const analysis::Constant* lhs_const = |
1785 | 885 | lhs_constants[0] ? lhs_constants[0] : lhs_constants[1]; |
1786 | 885 | const analysis::Constant* rhs_const = |
1787 | 885 | rhs_constants[0] ? rhs_constants[0] : rhs_constants[1]; |
1788 | 885 | if (!lhs_const || !rhs_const) return false; |
1789 | | |
1790 | 885 | bool const_lhs_rcp = lhs_constants[0] ? false : lhs_is_div; |
1791 | 885 | bool const_rhs_rcp = rhs_constants[0] ? false : rhs_is_div; |
1792 | | |
1793 | 885 | uint32_t non_const_lhs = lhs_constants[0] ? lhs->GetSingleWordInOperand(1) |
1794 | 885 | : lhs->GetSingleWordInOperand(0); |
1795 | 885 | bool non_const_lhs_rcp = lhs_constants[0] ? lhs_is_div : false; |
1796 | | |
1797 | 885 | uint32_t non_const_rhs = rhs_constants[0] ? rhs->GetSingleWordInOperand(1) |
1798 | 885 | : rhs->GetSingleWordInOperand(0); |
1799 | 885 | bool non_const_rhs_rcp = rhs_constants[0] ? rhs_is_div : false; |
1800 | | |
1801 | | // Rcp the rhs if we're actually dividing it. |
1802 | 885 | if (inst->opcode() == spv::Op::OpFDiv) { |
1803 | 137 | const_rhs_rcp = !const_rhs_rcp; |
1804 | 137 | non_const_rhs_rcp = !non_const_rhs_rcp; |
1805 | 137 | } |
1806 | | |
1807 | 885 | if (const_lhs_rcp) { |
1808 | 20 | lhs_const = |
1809 | 20 | const_mgr->FindDeclaredConstant(Reciprocal(const_mgr, lhs_const)); |
1810 | 20 | if (!lhs_const) { |
1811 | 20 | return false; |
1812 | 20 | } |
1813 | 20 | } |
1814 | 865 | if (const_rhs_rcp) { |
1815 | 133 | rhs_const = |
1816 | 133 | const_mgr->FindDeclaredConstant(Reciprocal(const_mgr, rhs_const)); |
1817 | 133 | if (!rhs_const) { |
1818 | 39 | return false; |
1819 | 39 | } |
1820 | 133 | } |
1821 | | |
1822 | 826 | uint32_t merged_constant = |
1823 | 826 | PerformOperation(const_mgr, spv::Op::OpFMul, lhs_const, rhs_const); |
1824 | | |
1825 | 826 | if (!merged_constant) { |
1826 | 264 | return false; |
1827 | 264 | } |
1828 | | |
1829 | 562 | spv::Op op = spv::Op::OpNop; |
1830 | 562 | Instruction* new_rhs = nullptr; |
1831 | | |
1832 | 562 | InstructionBuilder ir_builder( |
1833 | 562 | context, inst, |
1834 | 562 | IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping); |
1835 | | |
1836 | | // a * b => C * (b * a) |
1837 | 562 | if (!non_const_lhs_rcp && !non_const_rhs_rcp) { |
1838 | 363 | new_rhs = ir_builder.AddBinaryOp(inst->type_id(), spv::Op::OpFMul, |
1839 | 363 | non_const_lhs, non_const_rhs); |
1840 | 363 | op = spv::Op::OpFMul; |
1841 | 363 | } |
1842 | | // 1/a * b => C * (b / a) |
1843 | 199 | else if (non_const_lhs_rcp && !non_const_rhs_rcp) { |
1844 | 7 | new_rhs = ir_builder.AddBinaryOp(inst->type_id(), spv::Op::OpFDiv, |
1845 | 7 | non_const_rhs, non_const_lhs); |
1846 | 7 | op = spv::Op::OpFMul; |
1847 | 7 | } |
1848 | | // a * 1/b => C * (a / b) |
1849 | 192 | else if (!non_const_lhs_rcp && non_const_rhs_rcp) { |
1850 | 67 | new_rhs = ir_builder.AddBinaryOp(inst->type_id(), spv::Op::OpFDiv, |
1851 | 67 | non_const_lhs, non_const_rhs); |
1852 | 67 | op = spv::Op::OpFMul; |
1853 | 67 | } |
1854 | | // 1/a * 1/b => C / (a * b) |
1855 | 125 | else { |
1856 | 125 | new_rhs = ir_builder.AddBinaryOp(inst->type_id(), spv::Op::OpFMul, |
1857 | 125 | non_const_lhs, non_const_rhs); |
1858 | 125 | op = spv::Op::OpFDiv; |
1859 | 125 | } |
1860 | | |
1861 | 562 | if (!new_rhs) { |
1862 | 0 | return false; |
1863 | 0 | } |
1864 | | |
1865 | 562 | inst->SetOpcode(op); |
1866 | 562 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {merged_constant}}, |
1867 | 562 | {SPV_OPERAND_TYPE_ID, {new_rhs->result_id()}}}); |
1868 | 562 | return true; |
1869 | 562 | }; |
1870 | 32.0k | } |
1871 | | |
1872 | | // Reassociate add/sub instructions, which have add/sub inputs, |
1873 | | // both of which contain a constant. |
1874 | | // e.g: |
1875 | | // (a + C0) - (C1 - b) = (C0 - C1) + (a + b) |
1876 | | // (C0 - a) + (b - C1) = (C0 - C1) + (b - a) |
1877 | | // (a - C0) - (b + C1) = (-C0 - C1) + (a - b) |
1878 | 64.0k | FoldingRule ReassociateNestedAddSub() { |
1879 | 64.0k | return [](IRContext* context, Instruction* inst, |
1880 | 610k | const std::vector<const analysis::Constant*>& constants) { |
1881 | 610k | assert(inst->opcode() == spv::Op::OpFAdd || |
1882 | 610k | inst->opcode() == spv::Op::OpIAdd || |
1883 | 610k | inst->opcode() == spv::Op::OpFSub || |
1884 | 610k | inst->opcode() == spv::Op::OpISub); |
1885 | | |
1886 | | // Handled by other folding rules. |
1887 | 610k | if (constants[0] || constants[1]) { |
1888 | 209k | return false; |
1889 | 209k | } |
1890 | | |
1891 | 401k | const analysis::Type* type = |
1892 | 401k | context->get_type_mgr()->GetType(inst->type_id()); |
1893 | | |
1894 | 401k | if (type->IsCooperativeMatrix()) { |
1895 | 0 | return false; |
1896 | 0 | } |
1897 | | |
1898 | 401k | uint32_t width = ElementWidth(type); |
1899 | 401k | if (width != 32 && width != 64) return false; |
1900 | | |
1901 | 401k | bool uses_float = HasFloatingPoint(type); |
1902 | 401k | if (uses_float && !inst->IsFloatingPointFoldingAllowed()) return false; |
1903 | | |
1904 | 392k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
1905 | 392k | Instruction* lhs = def_use_mgr->GetDef(inst->GetSingleWordInOperand(0)); |
1906 | 392k | Instruction* rhs = def_use_mgr->GetDef(inst->GetSingleWordInOperand(1)); |
1907 | | |
1908 | 392k | spv::Op add_op = uses_float ? spv::Op::OpFAdd : spv::Op::OpIAdd; |
1909 | 392k | spv::Op sub_op = uses_float ? spv::Op::OpFSub : spv::Op::OpISub; |
1910 | | |
1911 | 392k | bool lhs_is_add = lhs->opcode() == add_op; |
1912 | 392k | bool lhs_is_sub = lhs->opcode() == sub_op; |
1913 | 392k | bool rhs_is_add = rhs->opcode() == add_op; |
1914 | 392k | bool rhs_is_sub = rhs->opcode() == sub_op; |
1915 | 392k | if (!(lhs_is_add || lhs_is_sub) || !(rhs_is_add || rhs_is_sub)) { |
1916 | 389k | return false; |
1917 | 389k | } |
1918 | | |
1919 | 2.77k | if (uses_float && (!lhs->IsFloatingPointFoldingAllowed() || |
1920 | 2.34k | !rhs->IsFloatingPointFoldingAllowed())) { |
1921 | 0 | return false; |
1922 | 0 | } |
1923 | | |
1924 | 2.77k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
1925 | 2.77k | std::vector<const analysis::Constant*> lhs_constants = |
1926 | 2.77k | const_mgr->GetOperandConstants(lhs); |
1927 | 2.77k | if (!lhs_constants[0] && !lhs_constants[1]) { |
1928 | 1.05k | return false; |
1929 | 1.05k | } |
1930 | | |
1931 | 1.71k | std::vector<const analysis::Constant*> rhs_constants = |
1932 | 1.71k | const_mgr->GetOperandConstants(rhs); |
1933 | 1.71k | if (!rhs_constants[0] && !rhs_constants[1]) { |
1934 | 51 | return false; |
1935 | 51 | } |
1936 | | |
1937 | 1.66k | const analysis::Constant* lhs_const = |
1938 | 1.66k | lhs_constants[0] ? lhs_constants[0] : lhs_constants[1]; |
1939 | 1.66k | const analysis::Constant* rhs_const = |
1940 | 1.66k | rhs_constants[0] ? rhs_constants[0] : rhs_constants[1]; |
1941 | 1.66k | if (!lhs_const || !rhs_const) return false; |
1942 | | |
1943 | 1.66k | bool const_lhs_neg = lhs_constants[0] ? false : lhs_is_sub; |
1944 | 1.66k | bool const_rhs_neg = rhs_constants[0] ? false : rhs_is_sub; |
1945 | | |
1946 | 1.66k | uint32_t non_const_lhs = lhs_constants[0] ? lhs->GetSingleWordInOperand(1) |
1947 | 1.66k | : lhs->GetSingleWordInOperand(0); |
1948 | 1.66k | bool non_const_lhs_neg = lhs_constants[0] ? lhs_is_sub : false; |
1949 | | |
1950 | 1.66k | uint32_t non_const_rhs = rhs_constants[0] ? rhs->GetSingleWordInOperand(1) |
1951 | 1.66k | : rhs->GetSingleWordInOperand(0); |
1952 | 1.66k | bool non_const_rhs_neg = rhs_constants[0] ? rhs_is_sub : false; |
1953 | | |
1954 | | // Negate the rhs if we're actually subtracting it. |
1955 | 1.66k | if (inst->opcode() == spv::Op::OpFSub || |
1956 | 951 | inst->opcode() == spv::Op::OpISub) { |
1957 | 714 | const_rhs_neg = !const_rhs_neg; |
1958 | 714 | non_const_rhs_neg = !non_const_rhs_neg; |
1959 | 714 | } |
1960 | | |
1961 | 1.66k | if (const_lhs_neg) { |
1962 | 495 | lhs_const = |
1963 | 495 | const_mgr->FindDeclaredConstant(NegateConstant(const_mgr, lhs_const)); |
1964 | 495 | if (!lhs_const) { |
1965 | 0 | return false; |
1966 | 0 | } |
1967 | 495 | } |
1968 | 1.66k | if (const_rhs_neg) { |
1969 | 1.22k | rhs_const = |
1970 | 1.22k | const_mgr->FindDeclaredConstant(NegateConstant(const_mgr, rhs_const)); |
1971 | 1.22k | if (!rhs_const) { |
1972 | 0 | return false; |
1973 | 0 | } |
1974 | 1.22k | } |
1975 | | |
1976 | 1.66k | uint32_t merged_constant = |
1977 | 1.66k | PerformOperation(const_mgr, add_op, lhs_const, rhs_const); |
1978 | | |
1979 | 1.66k | if (!merged_constant) { |
1980 | 262 | return false; |
1981 | 262 | } |
1982 | | |
1983 | 1.40k | spv::Op op = spv::Op::OpNop; |
1984 | 1.40k | Instruction* new_rhs = nullptr; |
1985 | | |
1986 | 1.40k | InstructionBuilder ir_builder( |
1987 | 1.40k | context, inst, |
1988 | 1.40k | IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping); |
1989 | | |
1990 | | // a + b => C + (b + a) |
1991 | 1.40k | if (!non_const_lhs_neg && !non_const_rhs_neg) { |
1992 | 734 | new_rhs = ir_builder.AddBinaryOp(inst->type_id(), add_op, non_const_lhs, |
1993 | 734 | non_const_rhs); |
1994 | 734 | op = add_op; |
1995 | 734 | } |
1996 | | // -a + b => C + (b - a) |
1997 | 668 | else if (non_const_lhs_neg && !non_const_rhs_neg) { |
1998 | 33 | new_rhs = ir_builder.AddBinaryOp(inst->type_id(), sub_op, non_const_rhs, |
1999 | 33 | non_const_lhs); |
2000 | 33 | op = add_op; |
2001 | 33 | } |
2002 | | // a + -b => C + (a - b) |
2003 | 635 | else if (!non_const_lhs_neg && non_const_rhs_neg) { |
2004 | 20 | new_rhs = ir_builder.AddBinaryOp(inst->type_id(), sub_op, non_const_lhs, |
2005 | 20 | non_const_rhs); |
2006 | 20 | op = add_op; |
2007 | 20 | } |
2008 | | // -a + -b => C - (a + b) |
2009 | 615 | else { |
2010 | 615 | new_rhs = ir_builder.AddBinaryOp(inst->type_id(), add_op, non_const_lhs, |
2011 | 615 | non_const_rhs); |
2012 | 615 | op = sub_op; |
2013 | 615 | } |
2014 | | |
2015 | 1.40k | if (!new_rhs) { |
2016 | 0 | return false; |
2017 | 0 | } |
2018 | | |
2019 | 1.40k | inst->SetOpcode(op); |
2020 | 1.40k | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {merged_constant}}, |
2021 | 1.40k | {SPV_OPERAND_TYPE_ID, {new_rhs->result_id()}}}); |
2022 | 1.40k | return true; |
2023 | 1.40k | }; |
2024 | 64.0k | } |
2025 | | |
2026 | 16.0k | FoldingRule IntMultipleBy1() { |
2027 | 16.0k | return [](IRContext*, Instruction* inst, |
2028 | 16.0k | const std::vector<const analysis::Constant*>& constants) { |
2029 | 15.1k | assert(inst->opcode() == spv::Op::OpIMul && |
2030 | 15.1k | "Wrong opcode. Should be OpIMul."); |
2031 | 44.9k | for (uint32_t i = 0; i < 2; i++) { |
2032 | 30.3k | if (constants[i] == nullptr) { |
2033 | 17.2k | continue; |
2034 | 17.2k | } |
2035 | 13.0k | const analysis::IntConstant* int_constant = constants[i]->AsIntConstant(); |
2036 | 13.0k | if (int_constant) { |
2037 | 12.8k | uint32_t width = ElementWidth(int_constant->type()); |
2038 | 12.8k | if (width != 32 && width != 64) return false; |
2039 | 12.8k | bool is_one = (width == 32) ? int_constant->GetU32BitValue() == 1u |
2040 | 12.8k | : int_constant->GetU64BitValue() == 1ull; |
2041 | 12.8k | if (is_one) { |
2042 | 487 | inst->SetOpcode(spv::Op::OpCopyObject); |
2043 | 487 | inst->SetInOperands( |
2044 | 487 | {{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(1 - i)}}}); |
2045 | 487 | return true; |
2046 | 487 | } |
2047 | 12.8k | } |
2048 | 13.0k | } |
2049 | 14.6k | return false; |
2050 | 15.1k | }; |
2051 | 16.0k | } |
2052 | | |
2053 | | // Returns the number of elements that the |index|th in operand in |inst| |
2054 | | // contributes to the result of |inst|. |inst| must be an |
2055 | | // OpCompositeConstructInstruction. |
2056 | | uint32_t GetNumOfElementsContributedByOperand(IRContext* context, |
2057 | | const Instruction* inst, |
2058 | 17.7k | uint32_t index) { |
2059 | 17.7k | assert(inst->opcode() == spv::Op::OpCompositeConstruct); |
2060 | 17.7k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
2061 | 17.7k | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
2062 | | |
2063 | 17.7k | analysis::Vector* result_type = |
2064 | 17.7k | type_mgr->GetType(inst->type_id())->AsVector(); |
2065 | 17.7k | if (result_type == nullptr) { |
2066 | | // If the result of the OpCompositeConstruct is not a vector then every |
2067 | | // operands corresponds to a single element in the result. |
2068 | 0 | return 1; |
2069 | 0 | } |
2070 | | |
2071 | | // If the result type is a vector then the operands are either scalars or |
2072 | | // vectors. If it is a scalar, then it corresponds to a single element. If it |
2073 | | // is a vector, then each element in the vector will be an element in the |
2074 | | // result. |
2075 | 17.7k | uint32_t id = inst->GetSingleWordInOperand(index); |
2076 | 17.7k | Instruction* def = def_use_mgr->GetDef(id); |
2077 | 17.7k | analysis::Vector* type = type_mgr->GetType(def->type_id())->AsVector(); |
2078 | 17.7k | if (type == nullptr) { |
2079 | 17.7k | return 1; |
2080 | 17.7k | } |
2081 | 0 | return type->element_count(); |
2082 | 17.7k | } |
2083 | | |
2084 | | // Returns the in-operands for an OpCompositeExtract instruction that are needed |
2085 | | // to extract the |result_index|th element in the result of |inst| without using |
2086 | | // the result of |inst|. Returns the empty vector if |result_index| is |
2087 | | // out-of-bounds. |inst| must be an |OpCompositeConstruct| instruction. |
2088 | | std::vector<Operand> GetExtractOperandsForElementOfCompositeConstruct( |
2089 | 35.6k | IRContext* context, const Instruction* inst, uint32_t result_index) { |
2090 | 35.6k | assert(inst->opcode() == spv::Op::OpCompositeConstruct); |
2091 | 35.6k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
2092 | 35.6k | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
2093 | | |
2094 | 35.6k | analysis::Type* result_type = type_mgr->GetType(inst->type_id()); |
2095 | 35.6k | if (result_type->AsVector() == nullptr) { |
2096 | 24.8k | if (result_index < inst->NumInOperands()) { |
2097 | 24.8k | uint32_t id = inst->GetSingleWordInOperand(result_index); |
2098 | 24.8k | return {Operand(SPV_OPERAND_TYPE_ID, {id})}; |
2099 | 24.8k | } |
2100 | 0 | return {}; |
2101 | 24.8k | } |
2102 | | |
2103 | | // If the result type is a vector, then vector operands are concatenated. |
2104 | 10.7k | uint32_t total_element_count = 0; |
2105 | 17.7k | for (uint32_t idx = 0; idx < inst->NumInOperands(); ++idx) { |
2106 | 17.7k | uint32_t element_count = |
2107 | 17.7k | GetNumOfElementsContributedByOperand(context, inst, idx); |
2108 | 17.7k | total_element_count += element_count; |
2109 | 17.7k | if (result_index < total_element_count) { |
2110 | 10.7k | std::vector<Operand> operands; |
2111 | 10.7k | uint32_t id = inst->GetSingleWordInOperand(idx); |
2112 | 10.7k | Instruction* operand_def = def_use_mgr->GetDef(id); |
2113 | 10.7k | analysis::Type* operand_type = type_mgr->GetType(operand_def->type_id()); |
2114 | | |
2115 | 10.7k | operands.push_back({SPV_OPERAND_TYPE_ID, {id}}); |
2116 | 10.7k | if (operand_type->AsVector()) { |
2117 | 0 | uint32_t start_index_of_id = total_element_count - element_count; |
2118 | 0 | uint32_t index_into_id = result_index - start_index_of_id; |
2119 | 0 | operands.push_back({SPV_OPERAND_TYPE_LITERAL_INTEGER, {index_into_id}}); |
2120 | 0 | } |
2121 | 10.7k | return operands; |
2122 | 10.7k | } |
2123 | 17.7k | } |
2124 | 0 | return {}; |
2125 | 10.7k | } |
2126 | | |
2127 | | // If the OpCompositeConstruct that feeds an OpCopyLogical can be retyped to |
2128 | | // the OpCopyLogical's result type, the layout conversion can be expressed at |
2129 | | // constituent granularity instead of at aggregate granularity. This rewrites |
2130 | | // the OpCopyLogical as an OpCompositeConstruct of the result type, using the |
2131 | | // same constituents where their types already match the corresponding |
2132 | | // field/element of the result type, and inserting per-field OpCopyLogical |
2133 | | // instructions only for the fields that genuinely require a layout |
2134 | | // conversion. |
2135 | | bool CompositeConstructFeedingCopyLogical( |
2136 | | IRContext* context, Instruction* inst, |
2137 | 0 | const std::vector<const analysis::Constant*>&) { |
2138 | 0 | assert(inst->opcode() == spv::Op::OpCopyLogical && |
2139 | 0 | "Wrong opcode. Should be OpCopyLogical."); |
2140 | 0 | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
2141 | |
|
2142 | 0 | uint32_t src_id = inst->GetSingleWordInOperand(0); |
2143 | 0 | Instruction* src_inst = def_use_mgr->GetDef(src_id); |
2144 | 0 | if (src_inst->opcode() != spv::Op::OpCompositeConstruct) { |
2145 | 0 | return false; |
2146 | 0 | } |
2147 | | |
2148 | 0 | Instruction* dst_type_inst = def_use_mgr->GetDef(inst->type_id()); |
2149 | 0 | const uint32_t num_constituents = src_inst->NumInOperands(); |
2150 | | |
2151 | | // Determine the expected type id for each constituent of the destination |
2152 | | // type. |
2153 | 0 | std::vector<uint32_t> expected_type_ids; |
2154 | 0 | expected_type_ids.reserve(num_constituents); |
2155 | 0 | if (dst_type_inst->opcode() == spv::Op::OpTypeStruct) { |
2156 | 0 | if (dst_type_inst->NumInOperands() != num_constituents) { |
2157 | 0 | return false; |
2158 | 0 | } |
2159 | 0 | for (uint32_t i = 0; i < num_constituents; ++i) { |
2160 | 0 | expected_type_ids.push_back(dst_type_inst->GetSingleWordInOperand(i)); |
2161 | 0 | } |
2162 | 0 | } else if (dst_type_inst->opcode() == spv::Op::OpTypeArray) { |
2163 | 0 | const uint32_t elem_type_id = dst_type_inst->GetSingleWordInOperand(0); |
2164 | 0 | for (uint32_t i = 0; i < num_constituents; ++i) { |
2165 | 0 | expected_type_ids.push_back(elem_type_id); |
2166 | 0 | } |
2167 | 0 | } else { |
2168 | 0 | return false; |
2169 | 0 | } |
2170 | | |
2171 | | // Build the new constituent list, inserting OpCopyLogical instructions for |
2172 | | // the fields whose types differ from the result type. |
2173 | 0 | InstructionBuilder ir_builder( |
2174 | 0 | context, inst, |
2175 | 0 | IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping); |
2176 | 0 | std::vector<Operand> operands; |
2177 | 0 | operands.reserve(num_constituents); |
2178 | 0 | for (uint32_t i = 0; i < num_constituents; ++i) { |
2179 | 0 | const uint32_t cid = src_inst->GetSingleWordInOperand(i); |
2180 | 0 | Instruction* cdef = def_use_mgr->GetDef(cid); |
2181 | 0 | if (cdef->type_id() == expected_type_ids[i]) { |
2182 | 0 | operands.push_back({SPV_OPERAND_TYPE_ID, {cid}}); |
2183 | 0 | continue; |
2184 | 0 | } |
2185 | 0 | if (def_use_mgr->GetDef(expected_type_ids[i])->opcode() == |
2186 | 0 | spv::Op::OpTypePointer) { |
2187 | 0 | assert(def_use_mgr->GetDef(expected_type_ids[i])->opcode() != |
2188 | 0 | spv::Op::OpTypePointer && |
2189 | 0 | "Unreachable for valid input"); |
2190 | 0 | } |
2191 | 0 | Instruction* per_field_copy = ir_builder.AddUnaryOp( |
2192 | 0 | expected_type_ids[i], spv::Op::OpCopyLogical, cid); |
2193 | 0 | if (per_field_copy == nullptr) { |
2194 | 0 | return false; |
2195 | 0 | } |
2196 | 0 | operands.push_back({SPV_OPERAND_TYPE_ID, {per_field_copy->result_id()}}); |
2197 | 0 | } |
2198 | | |
2199 | 0 | inst->SetOpcode(spv::Op::OpCompositeConstruct); |
2200 | 0 | inst->SetInOperands(std::move(operands)); |
2201 | 0 | context->UpdateDefUse(inst); |
2202 | 0 | return true; |
2203 | 0 | } |
2204 | | |
2205 | | bool CompositeConstructFeedingExtract( |
2206 | | IRContext* context, Instruction* inst, |
2207 | 278k | const std::vector<const analysis::Constant*>&) { |
2208 | | // If the input to an OpCompositeExtract is an OpCompositeConstruct, |
2209 | | // then we can simply use the appropriate element in the construction. |
2210 | 278k | assert(inst->opcode() == spv::Op::OpCompositeExtract && |
2211 | 278k | "Wrong opcode. Should be OpCompositeExtract."); |
2212 | 278k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
2213 | | |
2214 | | // If there are no index operands, then this rule cannot do anything. |
2215 | 278k | if (inst->NumInOperands() <= 1) { |
2216 | 0 | return false; |
2217 | 0 | } |
2218 | | |
2219 | 278k | uint32_t cid = inst->GetSingleWordInOperand(kExtractCompositeIdInIdx); |
2220 | 278k | Instruction* cinst = def_use_mgr->GetDef(cid); |
2221 | | |
2222 | 278k | if (cinst->opcode() != spv::Op::OpCompositeConstruct) { |
2223 | 242k | return false; |
2224 | 242k | } |
2225 | | |
2226 | 35.6k | uint32_t index_into_result = inst->GetSingleWordInOperand(1); |
2227 | 35.6k | std::vector<Operand> operands = |
2228 | 35.6k | GetExtractOperandsForElementOfCompositeConstruct(context, cinst, |
2229 | 35.6k | index_into_result); |
2230 | | |
2231 | 35.6k | if (operands.empty()) { |
2232 | 0 | return false; |
2233 | 0 | } |
2234 | | |
2235 | | // Add the remaining indices for extraction. |
2236 | 35.6k | for (uint32_t i = 2; i < inst->NumInOperands(); ++i) { |
2237 | 31 | operands.push_back( |
2238 | 31 | {SPV_OPERAND_TYPE_LITERAL_INTEGER, {inst->GetSingleWordInOperand(i)}}); |
2239 | 31 | } |
2240 | | |
2241 | 35.6k | if (operands.size() == 1) { |
2242 | | // If there were no extra indices, then we have the final object. No need |
2243 | | // to extract any more. |
2244 | 35.5k | inst->SetOpcode(spv::Op::OpCopyObject); |
2245 | 35.5k | } |
2246 | | |
2247 | 35.6k | inst->SetInOperands(std::move(operands)); |
2248 | 35.6k | return true; |
2249 | 35.6k | } |
2250 | | |
2251 | | // Walks the indexes chain from |start| to |end| of an OpCompositeInsert or |
2252 | | // OpCompositeExtract instruction, and returns the type id of the final element |
2253 | | // being accessed. Returns 0 if a valid type could not be found. |
2254 | | uint32_t GetElementType(uint32_t type_id, Instruction::iterator start, |
2255 | | Instruction::iterator end, |
2256 | 96.4k | const analysis::DefUseManager* def_use_manager) { |
2257 | 96.4k | for (auto index : make_range(std::move(start), std::move(end))) { |
2258 | 1.08k | const Instruction* type_inst = def_use_manager->GetDef(type_id); |
2259 | 1.08k | assert(index.type == SPV_OPERAND_TYPE_LITERAL_INTEGER && |
2260 | 1.08k | index.words.size() == 1); |
2261 | 1.08k | switch (type_inst->opcode()) { |
2262 | 518 | case spv::Op::OpTypeArray: |
2263 | 518 | case spv::Op::OpTypeMatrix: |
2264 | 518 | case spv::Op::OpTypeVector: |
2265 | 518 | case spv::Op::OpTypeVectorIdEXT: |
2266 | 518 | type_id = type_inst->GetSingleWordInOperand(0); |
2267 | 518 | break; |
2268 | 566 | case spv::Op::OpTypeStruct: |
2269 | 566 | type_id = type_inst->GetSingleWordInOperand(index.words[0]); |
2270 | 566 | break; |
2271 | 0 | default: |
2272 | 0 | return 0; |
2273 | 1.08k | } |
2274 | 1.08k | } |
2275 | 96.4k | return type_id; |
2276 | 96.4k | } |
2277 | | |
2278 | | // If the input to an OpCompositeExtract is an OpCopyLogical, then we can |
2279 | | // hoist the extraction before the copy. |
2280 | | bool CopyLogicalFeedingExtract(IRContext* context, Instruction* inst, |
2281 | 237k | const std::vector<const analysis::Constant*>&) { |
2282 | 237k | assert(inst->opcode() == spv::Op::OpCompositeExtract && |
2283 | 237k | "Wrong opcode. Should be OpCompositeExtract."); |
2284 | | |
2285 | 237k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
2286 | 237k | uint32_t cid = inst->GetSingleWordInOperand(kExtractCompositeIdInIdx); |
2287 | 237k | Instruction* cinst = def_use_mgr->GetDef(cid); |
2288 | | |
2289 | 237k | if (cinst->opcode() != spv::Op::OpCopyLogical) { |
2290 | 237k | return false; |
2291 | 237k | } |
2292 | | |
2293 | 0 | uint32_t original_composite_id = cinst->GetSingleWordInOperand(0); |
2294 | 0 | Instruction* original_composite_inst = |
2295 | 0 | def_use_mgr->GetDef(original_composite_id); |
2296 | |
|
2297 | 0 | std::vector<uint32_t> indices; |
2298 | 0 | for (uint32_t i = 1; i < inst->NumInOperands(); ++i) { |
2299 | 0 | indices.push_back(inst->GetSingleWordInOperand(i)); |
2300 | 0 | } |
2301 | |
|
2302 | 0 | uint32_t original_element_type_id = |
2303 | 0 | GetElementType(original_composite_inst->type_id(), inst->begin() + 3, |
2304 | 0 | inst->end(), def_use_mgr); |
2305 | 0 | assert(original_element_type_id != 0 && |
2306 | 0 | "Could not find the element type. Invalid SPIR-V."); |
2307 | | |
2308 | 0 | InstructionBuilder ir_builder( |
2309 | 0 | context, inst, |
2310 | 0 | IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping); |
2311 | |
|
2312 | 0 | Instruction* new_extract = ir_builder.AddCompositeExtract( |
2313 | 0 | original_element_type_id, original_composite_id, indices); |
2314 | |
|
2315 | 0 | if (original_element_type_id == inst->type_id()) |
2316 | 0 | inst->SetOpcode(spv::Op::OpCopyObject); |
2317 | 0 | else |
2318 | 0 | inst->SetOpcode(spv::Op::OpCopyLogical); |
2319 | 0 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {new_extract->result_id()}}}); |
2320 | 0 | return true; |
2321 | 0 | } |
2322 | | |
2323 | | // If the input to an OpCompositeExtract is an OpLoad, we can change the |
2324 | | // load into a load of an OpAccessChain. |
2325 | | bool LoadFeedingExtract(IRContext* context, Instruction* inst, |
2326 | 237k | const std::vector<const analysis::Constant*>&) { |
2327 | 237k | assert(inst->opcode() == spv::Op::OpCompositeExtract && |
2328 | 237k | "Wrong opcode. Should be OpCompositeExtract."); |
2329 | | |
2330 | 237k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
2331 | 237k | uint32_t cid = inst->GetSingleWordInOperand(kExtractCompositeIdInIdx); |
2332 | 237k | Instruction* cinst = def_use_mgr->GetDef(cid); |
2333 | | |
2334 | 237k | if (cinst->opcode() != spv::Op::OpLoad) { |
2335 | 178k | return false; |
2336 | 178k | } |
2337 | | |
2338 | 59.1k | Instruction* composite_type_inst = def_use_mgr->GetDef(cinst->type_id()); |
2339 | 59.1k | if (composite_type_inst->opcode() != spv::Op::OpTypeStruct && |
2340 | 30.9k | composite_type_inst->opcode() != spv::Op::OpTypeArray) { |
2341 | 12.9k | return false; |
2342 | 12.9k | } |
2343 | | |
2344 | | // Check the memory operands. |
2345 | 46.2k | if (cinst->NumInOperands() > 1) { |
2346 | 597 | uint32_t memory_access_mask = cinst->GetSingleWordInOperand(1); |
2347 | 597 | if (memory_access_mask & uint32_t(spv::MemoryAccessMask::Volatile)) { |
2348 | 22 | return false; |
2349 | 22 | } |
2350 | 597 | } |
2351 | | |
2352 | 46.2k | uint32_t ptr_id = cinst->GetSingleWordInOperand(0); |
2353 | 46.2k | Instruction* ptr_inst = def_use_mgr->GetDef(ptr_id); |
2354 | 46.2k | Instruction* ptr_type_inst = def_use_mgr->GetDef(ptr_inst->type_id()); |
2355 | 46.2k | assert(ptr_type_inst->opcode() == spv::Op::OpTypePointer); |
2356 | 46.2k | spv::StorageClass storage_class = |
2357 | 46.2k | static_cast<spv::StorageClass>(ptr_type_inst->GetSingleWordInOperand(0)); |
2358 | | |
2359 | | // If the storage class is Function or Private, we do not want to fold. |
2360 | | // These are the storage classes that the local-access-chain-convert pass |
2361 | | // works on. |
2362 | 46.2k | if (storage_class == spv::StorageClass::Function || |
2363 | 46.2k | storage_class == spv::StorageClass::Private) { |
2364 | 46.2k | return false; |
2365 | 46.2k | } |
2366 | | |
2367 | 0 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
2368 | 0 | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
2369 | 0 | std::vector<uint32_t> index_ids; |
2370 | 0 | for (uint32_t i = 1; i < inst->NumInOperands(); ++i) { |
2371 | 0 | uint32_t index = inst->GetSingleWordInOperand(i); |
2372 | 0 | const analysis::Constant* index_const = |
2373 | 0 | const_mgr->GetConstant(type_mgr->GetUIntType(), {index}); |
2374 | 0 | index_ids.push_back( |
2375 | 0 | const_mgr->GetDefiningInstruction(index_const)->result_id()); |
2376 | 0 | } |
2377 | |
|
2378 | 0 | InstructionBuilder ir_builder( |
2379 | 0 | context, cinst, |
2380 | 0 | IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping); |
2381 | |
|
2382 | 0 | uint32_t element_ptr_type_id = |
2383 | 0 | type_mgr->FindPointerToType(inst->type_id(), storage_class); |
2384 | 0 | if (element_ptr_type_id == 0) { |
2385 | 0 | return false; |
2386 | 0 | } |
2387 | | |
2388 | 0 | Instruction* access_chain = |
2389 | 0 | ir_builder.AddAccessChain(element_ptr_type_id, ptr_id, index_ids); |
2390 | 0 | std::vector<Operand> load_operands; |
2391 | 0 | load_operands.push_back({SPV_OPERAND_TYPE_ID, {access_chain->result_id()}}); |
2392 | |
|
2393 | 0 | if (cinst->NumInOperands() > 1) { |
2394 | 0 | uint32_t memory_access_mask = cinst->GetSingleWordInOperand(1); |
2395 | 0 | load_operands.push_back( |
2396 | 0 | {SPV_OPERAND_TYPE_MEMORY_ACCESS, {memory_access_mask}}); |
2397 | |
|
2398 | 0 | uint32_t current_operand_index = 2; |
2399 | 0 | if (memory_access_mask & uint32_t(spv::MemoryAccessMask::Aligned)) { |
2400 | 0 | uint32_t original_alignment = |
2401 | 0 | cinst->GetSingleWordInOperand(current_operand_index); |
2402 | |
|
2403 | 0 | std::vector<uint32_t> extract_indices; |
2404 | 0 | for (uint32_t i = 1; i < inst->NumInOperands(); ++i) { |
2405 | 0 | extract_indices.push_back(inst->GetSingleWordInOperand(i)); |
2406 | 0 | } |
2407 | |
|
2408 | 0 | std::optional<uint32_t> offset = |
2409 | 0 | type_mgr->GetType(cinst->type_id())->GetByteOffset(extract_indices); |
2410 | 0 | if (!offset) { |
2411 | 0 | return false; |
2412 | 0 | } |
2413 | | |
2414 | 0 | uint32_t new_alignment = original_alignment; |
2415 | 0 | if (*offset != 0) { |
2416 | 0 | uint32_t offset_alignment = *offset & ~(*offset - 1); |
2417 | 0 | new_alignment = std::min(original_alignment, offset_alignment); |
2418 | 0 | } |
2419 | |
|
2420 | 0 | load_operands.push_back( |
2421 | 0 | {SPV_OPERAND_TYPE_TYPED_LITERAL_NUMBER, {new_alignment}}); |
2422 | 0 | current_operand_index++; |
2423 | 0 | } |
2424 | | |
2425 | | // Copy the remaining operands |
2426 | 0 | for (; current_operand_index < cinst->NumInOperands(); |
2427 | 0 | ++current_operand_index) { |
2428 | 0 | load_operands.push_back(cinst->GetInOperand(current_operand_index)); |
2429 | 0 | } |
2430 | 0 | } |
2431 | | |
2432 | 0 | uint32_t load_result_id = context->TakeNextId(); |
2433 | 0 | if (load_result_id == 0) return false; |
2434 | | |
2435 | 0 | std::unique_ptr<Instruction> new_load_inst( |
2436 | 0 | new Instruction(context, spv::Op::OpLoad, inst->type_id(), load_result_id, |
2437 | 0 | load_operands)); |
2438 | 0 | Instruction* new_load = ir_builder.AddInstruction(std::move(new_load_inst)); |
2439 | |
|
2440 | 0 | inst->SetOpcode(spv::Op::OpCopyObject); |
2441 | 0 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {new_load->result_id()}}}); |
2442 | |
|
2443 | 0 | return true; |
2444 | 0 | } |
2445 | | |
2446 | | // Returns true of |inst_1| and |inst_2| have the same indexes that will be used |
2447 | | // to index into a composite object, excluding the last index. The two |
2448 | | // instructions must have the same opcode, and be either OpCompositeExtract or |
2449 | | // OpCompositeInsert instructions. |
2450 | 1.21M | bool HaveSameIndexesExceptForLast(Instruction* inst_1, Instruction* inst_2) { |
2451 | 1.21M | assert(inst_1->opcode() == inst_2->opcode() && |
2452 | 1.21M | "Expecting the opcodes to be the same."); |
2453 | 1.21M | assert((inst_1->opcode() == spv::Op::OpCompositeInsert || |
2454 | 1.21M | inst_1->opcode() == spv::Op::OpCompositeExtract) && |
2455 | 1.21M | "Instructions must be OpCompositeInsert or OpCompositeExtract."); |
2456 | | |
2457 | 1.21M | if (inst_1->NumInOperands() != inst_2->NumInOperands()) { |
2458 | 1.96k | return false; |
2459 | 1.96k | } |
2460 | | |
2461 | 1.21M | uint32_t first_index_position = |
2462 | 1.21M | (inst_1->opcode() == spv::Op::OpCompositeInsert ? 2 : 1); |
2463 | 1.21M | for (uint32_t i = first_index_position; i < inst_1->NumInOperands() - 1; |
2464 | 1.21M | i++) { |
2465 | 8.02k | if (inst_1->GetSingleWordInOperand(i) != |
2466 | 8.02k | inst_2->GetSingleWordInOperand(i)) { |
2467 | 464 | return false; |
2468 | 464 | } |
2469 | 8.02k | } |
2470 | 1.21M | return true; |
2471 | 1.21M | } |
2472 | | |
2473 | | // If the OpCompositeConstruct is simply putting back together elements that |
2474 | | // where extracted from the same source, we can simply reuse the source. |
2475 | | // |
2476 | | // This is a common code pattern because of the way that scalar replacement |
2477 | | // works. |
2478 | | bool CompositeExtractFeedingConstruct( |
2479 | | IRContext* context, Instruction* inst, |
2480 | 94.4k | const std::vector<const analysis::Constant*>&) { |
2481 | 94.4k | assert(inst->opcode() == spv::Op::OpCompositeConstruct && |
2482 | 94.4k | "Wrong opcode. Should be OpCompositeConstruct."); |
2483 | 94.4k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
2484 | 94.4k | uint32_t original_id = 0; |
2485 | | |
2486 | 94.4k | if (inst->NumInOperands() == 0) { |
2487 | | // The struct being constructed has no members. |
2488 | 0 | return false; |
2489 | 0 | } |
2490 | | |
2491 | | // Check each element to make sure they are: |
2492 | | // - extractions |
2493 | | // - extracting the same position they are inserting |
2494 | | // - all extract from the same id. |
2495 | 94.4k | Instruction* first_element_inst = nullptr; |
2496 | 121k | for (uint32_t i = 0; i < inst->NumInOperands(); ++i) { |
2497 | 119k | const uint32_t element_id = inst->GetSingleWordInOperand(i); |
2498 | 119k | Instruction* element_inst = def_use_mgr->GetDef(element_id); |
2499 | 119k | if (first_element_inst == nullptr) { |
2500 | 94.4k | first_element_inst = element_inst; |
2501 | 94.4k | } |
2502 | | |
2503 | 119k | if (element_inst->opcode() != spv::Op::OpCompositeExtract) { |
2504 | 88.9k | return false; |
2505 | 88.9k | } |
2506 | | |
2507 | 30.8k | if (!HaveSameIndexesExceptForLast(element_inst, first_element_inst)) { |
2508 | 0 | return false; |
2509 | 0 | } |
2510 | | |
2511 | 30.8k | if (element_inst->GetSingleWordInOperand(element_inst->NumInOperands() - |
2512 | 30.8k | 1) != i) { |
2513 | 3.67k | return false; |
2514 | 3.67k | } |
2515 | | |
2516 | 27.1k | if (i == 0) { |
2517 | 9.45k | original_id = |
2518 | 9.45k | element_inst->GetSingleWordInOperand(kExtractCompositeIdInIdx); |
2519 | 17.7k | } else if (original_id != |
2520 | 17.7k | element_inst->GetSingleWordInOperand(kExtractCompositeIdInIdx)) { |
2521 | 424 | return false; |
2522 | 424 | } |
2523 | 27.1k | } |
2524 | 94.4k | assert(first_element_inst != nullptr); |
2525 | | |
2526 | | // The last check it to see that the object being extracted from is the |
2527 | | // correct type. |
2528 | 1.35k | Instruction* original_inst = def_use_mgr->GetDef(original_id); |
2529 | 1.35k | uint32_t original_type_id = |
2530 | 1.35k | GetElementType(original_inst->type_id(), first_element_inst->begin() + 3, |
2531 | 1.35k | first_element_inst->end() - 1, def_use_mgr); |
2532 | | |
2533 | 1.35k | if (inst->type_id() != original_type_id) { |
2534 | 210 | return false; |
2535 | 210 | } |
2536 | | |
2537 | 1.14k | if (first_element_inst->NumInOperands() == 2) { |
2538 | | // Simplify by using the original object. |
2539 | 1.14k | inst->SetOpcode(spv::Op::OpCopyObject); |
2540 | 1.14k | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {original_id}}}); |
2541 | 1.14k | return true; |
2542 | 1.14k | } |
2543 | | |
2544 | | // Copies the original id and all indexes except for the last to the new |
2545 | | // extract instruction. |
2546 | 0 | inst->SetOpcode(spv::Op::OpCompositeExtract); |
2547 | 0 | inst->SetInOperands(std::vector<Operand>(first_element_inst->begin() + 2, |
2548 | 0 | first_element_inst->end() - 1)); |
2549 | 0 | return true; |
2550 | 1.14k | } |
2551 | | |
2552 | 16.0k | FoldingRule InsertFeedingExtract() { |
2553 | 16.0k | return [](IRContext* context, Instruction* inst, |
2554 | 361k | const std::vector<const analysis::Constant*>&) { |
2555 | 361k | assert(inst->opcode() == spv::Op::OpCompositeExtract && |
2556 | 361k | "Wrong opcode. Should be OpCompositeExtract."); |
2557 | 361k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
2558 | 361k | uint32_t cid = inst->GetSingleWordInOperand(kExtractCompositeIdInIdx); |
2559 | 361k | Instruction* cinst = def_use_mgr->GetDef(cid); |
2560 | | |
2561 | 361k | if (cinst->opcode() != spv::Op::OpCompositeInsert) { |
2562 | 277k | return false; |
2563 | 277k | } |
2564 | | |
2565 | | // Find the first position where the list of insert and extract indicies |
2566 | | // differ, if at all. |
2567 | 83.6k | uint32_t i; |
2568 | 104k | for (i = 1; i < inst->NumInOperands(); ++i) { |
2569 | 84.2k | if (i + 1 >= cinst->NumInOperands()) { |
2570 | 0 | break; |
2571 | 0 | } |
2572 | | |
2573 | 84.2k | if (inst->GetSingleWordInOperand(i) != |
2574 | 84.2k | cinst->GetSingleWordInOperand(i + 1)) { |
2575 | 63.0k | break; |
2576 | 63.0k | } |
2577 | 84.2k | } |
2578 | | |
2579 | | // We are extracting the element that was inserted. |
2580 | 83.6k | if (i == inst->NumInOperands() && i + 1 == cinst->NumInOperands()) { |
2581 | 20.3k | inst->SetOpcode(spv::Op::OpCopyObject); |
2582 | 20.3k | inst->SetInOperands( |
2583 | 20.3k | {{SPV_OPERAND_TYPE_ID, |
2584 | 20.3k | {cinst->GetSingleWordInOperand(kInsertObjectIdInIdx)}}}); |
2585 | 20.3k | return true; |
2586 | 20.3k | } |
2587 | | |
2588 | | // Extracting the value that was inserted along with values for the base |
2589 | | // composite. Cannot do anything. |
2590 | 63.3k | if (i == inst->NumInOperands()) { |
2591 | 337 | return false; |
2592 | 337 | } |
2593 | | |
2594 | | // Extracting an element of the value that was inserted. Extract from |
2595 | | // that value directly. |
2596 | 63.0k | if (i + 1 == cinst->NumInOperands()) { |
2597 | 0 | std::vector<Operand> operands; |
2598 | 0 | operands.push_back( |
2599 | 0 | {SPV_OPERAND_TYPE_ID, |
2600 | 0 | {cinst->GetSingleWordInOperand(kInsertObjectIdInIdx)}}); |
2601 | 0 | for (; i < inst->NumInOperands(); ++i) { |
2602 | 0 | operands.push_back({SPV_OPERAND_TYPE_LITERAL_INTEGER, |
2603 | 0 | {inst->GetSingleWordInOperand(i)}}); |
2604 | 0 | } |
2605 | 0 | inst->SetInOperands(std::move(operands)); |
2606 | 0 | return true; |
2607 | 0 | } |
2608 | | |
2609 | | // Extracting a value that is disjoint from the element being inserted. |
2610 | | // Rewrite the extract to use the composite input to the insert. |
2611 | 63.0k | std::vector<Operand> operands; |
2612 | 63.0k | operands.push_back( |
2613 | 63.0k | {SPV_OPERAND_TYPE_ID, |
2614 | 63.0k | {cinst->GetSingleWordInOperand(kInsertCompositeIdInIdx)}}); |
2615 | 126k | for (i = 1; i < inst->NumInOperands(); ++i) { |
2616 | 63.5k | operands.push_back({SPV_OPERAND_TYPE_LITERAL_INTEGER, |
2617 | 63.5k | {inst->GetSingleWordInOperand(i)}}); |
2618 | 63.5k | } |
2619 | 63.0k | inst->SetInOperands(std::move(operands)); |
2620 | 63.0k | return true; |
2621 | 63.0k | }; |
2622 | 16.0k | } |
2623 | | |
2624 | | // When a VectorShuffle is feeding an Extract, we can extract from one of the |
2625 | | // operands of the VectorShuffle. We just need to adjust the index in the |
2626 | | // extract instruction. |
2627 | 16.0k | FoldingRule VectorShuffleFeedingExtract() { |
2628 | 16.0k | return [](IRContext* context, Instruction* inst, |
2629 | 242k | const std::vector<const analysis::Constant*>&) { |
2630 | 242k | assert(inst->opcode() == spv::Op::OpCompositeExtract && |
2631 | 242k | "Wrong opcode. Should be OpCompositeExtract."); |
2632 | 242k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
2633 | 242k | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
2634 | 242k | uint32_t cid = inst->GetSingleWordInOperand(kExtractCompositeIdInIdx); |
2635 | 242k | Instruction* cinst = def_use_mgr->GetDef(cid); |
2636 | | |
2637 | 242k | if (cinst->opcode() != spv::Op::OpVectorShuffle) { |
2638 | 237k | return false; |
2639 | 237k | } |
2640 | | |
2641 | | // Find the size of the first vector operand of the VectorShuffle |
2642 | 4.65k | Instruction* first_input = |
2643 | 4.65k | def_use_mgr->GetDef(cinst->GetSingleWordInOperand(0)); |
2644 | 4.65k | analysis::Type* first_input_type = |
2645 | 4.65k | type_mgr->GetType(first_input->type_id()); |
2646 | 4.65k | assert(first_input_type->AsVector() && |
2647 | 4.65k | "Input to vector shuffle should be vectors."); |
2648 | 4.65k | uint32_t first_input_size = first_input_type->AsVector()->element_count(); |
2649 | | |
2650 | | // Get index of the element the vector shuffle is placing in the position |
2651 | | // being extracted. |
2652 | 4.65k | uint32_t new_index = |
2653 | 4.65k | cinst->GetSingleWordInOperand(2 + inst->GetSingleWordInOperand(1)); |
2654 | | |
2655 | | // Extracting an undefined value so fold this extract into an undef. |
2656 | 4.65k | const uint32_t undef_literal_value = 0xffffffff; |
2657 | 4.65k | if (new_index == undef_literal_value) { |
2658 | 217 | inst->SetOpcode(spv::Op::OpUndef); |
2659 | 217 | inst->SetInOperands({}); |
2660 | 217 | return true; |
2661 | 217 | } |
2662 | | |
2663 | | // Get the id of the of the vector the elemtent comes from, and update the |
2664 | | // index if needed. |
2665 | 4.43k | uint32_t new_vector = 0; |
2666 | 4.43k | if (new_index < first_input_size) { |
2667 | 2.83k | new_vector = cinst->GetSingleWordInOperand(0); |
2668 | 2.83k | } else { |
2669 | 1.59k | new_vector = cinst->GetSingleWordInOperand(1); |
2670 | 1.59k | new_index -= first_input_size; |
2671 | 1.59k | } |
2672 | | |
2673 | | // Update the extract instruction. |
2674 | 4.43k | inst->SetInOperand(kExtractCompositeIdInIdx, {new_vector}); |
2675 | 4.43k | inst->SetInOperand(1, {new_index}); |
2676 | 4.43k | return true; |
2677 | 4.65k | }; |
2678 | 16.0k | } |
2679 | | |
2680 | | // When an FMix with is feeding an Extract that extracts an element whose |
2681 | | // corresponding |a| in the FMix is 0 or 1, we can extract from one of the |
2682 | | // operands of the FMix. |
2683 | 16.0k | FoldingRule FMixFeedingExtract() { |
2684 | 16.0k | return [](IRContext* context, Instruction* inst, |
2685 | 237k | const std::vector<const analysis::Constant*>&) { |
2686 | 237k | assert(inst->opcode() == spv::Op::OpCompositeExtract && |
2687 | 237k | "Wrong opcode. Should be OpCompositeExtract."); |
2688 | 237k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
2689 | 237k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
2690 | | |
2691 | 237k | uint32_t composite_id = |
2692 | 237k | inst->GetSingleWordInOperand(kExtractCompositeIdInIdx); |
2693 | 237k | Instruction* composite_inst = def_use_mgr->GetDef(composite_id); |
2694 | | |
2695 | 237k | if (composite_inst->opcode() != spv::Op::OpExtInst) { |
2696 | 218k | return false; |
2697 | 218k | } |
2698 | | |
2699 | 19.6k | uint32_t inst_set_id = |
2700 | 19.6k | context->get_feature_mgr()->GetExtInstImportId_GLSLstd450(); |
2701 | | |
2702 | 19.6k | if (composite_inst->GetSingleWordInOperand(kExtInstSetIdInIdx) != |
2703 | 19.6k | inst_set_id || |
2704 | 19.6k | composite_inst->GetSingleWordInOperand(kExtInstInstructionInIdx) != |
2705 | 19.6k | GLSLstd450FMix) { |
2706 | 14.5k | return false; |
2707 | 14.5k | } |
2708 | | |
2709 | | // Get the |a| for the FMix instruction. |
2710 | 5.12k | uint32_t a_id = composite_inst->GetSingleWordInOperand(kFMixAIdInIdx); |
2711 | 5.12k | std::unique_ptr<Instruction> a(inst->Clone(context)); |
2712 | 5.12k | a->SetInOperand(kExtractCompositeIdInIdx, {a_id}); |
2713 | 5.12k | context->get_instruction_folder().FoldInstruction(a.get()); |
2714 | | |
2715 | 5.12k | if (a->opcode() != spv::Op::OpCopyObject) { |
2716 | 1.22k | return false; |
2717 | 1.22k | } |
2718 | | |
2719 | 3.90k | const analysis::Constant* a_const = |
2720 | 3.90k | const_mgr->FindDeclaredConstant(a->GetSingleWordInOperand(0)); |
2721 | | |
2722 | 3.90k | if (!a_const) { |
2723 | 1.96k | return false; |
2724 | 1.96k | } |
2725 | | |
2726 | 1.93k | bool use_x = false; |
2727 | | |
2728 | 1.93k | assert(a_const->type()->AsFloat()); |
2729 | | |
2730 | 1.93k | const analysis::Type* type = |
2731 | 1.93k | context->get_type_mgr()->GetType(inst->type_id()); |
2732 | 1.93k | uint32_t width = ElementWidth(type); |
2733 | 1.93k | if (width != 32 && width != 64) { |
2734 | | // We won't support folding half float values. |
2735 | 0 | return false; |
2736 | 0 | } |
2737 | | |
2738 | 1.93k | double element_value = a_const->GetValueAsDouble(); |
2739 | 1.93k | if (element_value == 0.0) { |
2740 | 72 | use_x = true; |
2741 | 1.86k | } else if (element_value == 1.0) { |
2742 | 34 | use_x = false; |
2743 | 1.82k | } else { |
2744 | 1.82k | return false; |
2745 | 1.82k | } |
2746 | | |
2747 | | // Get the id of the of the vector the element comes from. |
2748 | 106 | uint32_t new_vector = 0; |
2749 | 106 | if (use_x) { |
2750 | 72 | new_vector = composite_inst->GetSingleWordInOperand(kFMixXIdInIdx); |
2751 | 72 | } else { |
2752 | 34 | new_vector = composite_inst->GetSingleWordInOperand(kFMixYIdInIdx); |
2753 | 34 | } |
2754 | | |
2755 | | // Update the extract instruction. |
2756 | 106 | inst->SetInOperand(kExtractCompositeIdInIdx, {new_vector}); |
2757 | 106 | return true; |
2758 | 1.93k | }; |
2759 | 16.0k | } |
2760 | | |
2761 | | // Returns the number of elements in the composite type |type|. Returns 0 if |
2762 | | // |type| is a scalar value. Return UINT32_MAX when the size is unknown at |
2763 | | // compile time. |
2764 | 95.1k | uint32_t GetNumberOfElements(const analysis::Type* type) { |
2765 | 95.1k | if (auto* vector_type = type->AsVector()) { |
2766 | 85.9k | return vector_type->element_count(); |
2767 | 85.9k | } |
2768 | 9.20k | if (auto* matrix_type = type->AsMatrix()) { |
2769 | 0 | return matrix_type->element_count(); |
2770 | 0 | } |
2771 | 9.20k | if (auto* struct_type = type->AsStruct()) { |
2772 | 3.56k | return static_cast<uint32_t>(struct_type->element_types().size()); |
2773 | 3.56k | } |
2774 | 5.64k | if (auto* array_type = type->AsArray()) { |
2775 | 5.64k | if (array_type->length_info().words[0] == |
2776 | 5.64k | analysis::Array::LengthInfo::kConstant && |
2777 | 5.64k | array_type->length_info().words.size() == 2) { |
2778 | 5.64k | return array_type->length_info().words[1]; |
2779 | 5.64k | } |
2780 | 0 | return UINT32_MAX; |
2781 | 5.64k | } |
2782 | 0 | return 0; |
2783 | 5.64k | } |
2784 | | |
2785 | | // Returns a map with the set of values that were inserted into an object by |
2786 | | // the chain of OpCompositeInsertInstruction starting with |inst|. |
2787 | | // The map will map the index to the value inserted at that index. An empty map |
2788 | | // will be returned if the map could not be properly generated. |
2789 | 95.1k | std::map<uint32_t, uint32_t> GetInsertedValues(Instruction* inst) { |
2790 | 95.1k | analysis::DefUseManager* def_use_mgr = inst->context()->get_def_use_mgr(); |
2791 | 95.1k | std::map<uint32_t, uint32_t> values_inserted; |
2792 | 95.1k | Instruction* current_inst = inst; |
2793 | 1.27M | while (current_inst->opcode() == spv::Op::OpCompositeInsert) { |
2794 | 1.18M | if (current_inst->NumInOperands() > inst->NumInOperands()) { |
2795 | | // This is to catch the case |
2796 | | // %2 = OpCompositeInsert %m2x2int %v2int_1_0 %m2x2int_undef 0 |
2797 | | // %3 = OpCompositeInsert %m2x2int %int_4 %2 0 0 |
2798 | | // %4 = OpCompositeInsert %m2x2int %v2int_2_3 %3 1 |
2799 | | // In this case we cannot do a single construct to get the matrix. |
2800 | 936 | uint32_t partially_inserted_element_index = |
2801 | 936 | current_inst->GetSingleWordInOperand(inst->NumInOperands() - 1); |
2802 | 936 | if (values_inserted.count(partially_inserted_element_index) == 0) |
2803 | 214 | return {}; |
2804 | 936 | } |
2805 | 1.18M | if (HaveSameIndexesExceptForLast(inst, current_inst)) { |
2806 | 1.18M | values_inserted.insert( |
2807 | 1.18M | {current_inst->GetSingleWordInOperand(current_inst->NumInOperands() - |
2808 | 1.18M | 1), |
2809 | 1.18M | current_inst->GetSingleWordInOperand(kInsertObjectIdInIdx)}); |
2810 | 1.18M | } |
2811 | 1.18M | current_inst = def_use_mgr->GetDef( |
2812 | 1.18M | current_inst->GetSingleWordInOperand(kInsertCompositeIdInIdx)); |
2813 | 1.18M | } |
2814 | 94.9k | return values_inserted; |
2815 | 95.1k | } |
2816 | | |
2817 | | // Returns true of there is an entry in |values_inserted| for every element of |
2818 | | // |Type|. |
2819 | | bool DoInsertedValuesCoverEntireObject( |
2820 | 95.1k | const analysis::Type* type, std::map<uint32_t, uint32_t>& values_inserted) { |
2821 | 95.1k | uint32_t container_size = GetNumberOfElements(type); |
2822 | 95.1k | if (container_size != values_inserted.size()) { |
2823 | 84.2k | return false; |
2824 | 84.2k | } |
2825 | | |
2826 | 10.8k | if (values_inserted.rbegin()->first >= container_size) { |
2827 | 0 | return false; |
2828 | 0 | } |
2829 | 10.8k | return true; |
2830 | 10.8k | } |
2831 | | |
2832 | | // Returns id of the type of the element that immediately contains the element |
2833 | | // being inserted by the OpCompositeInsert instruction |inst|. Returns 0 if it |
2834 | | // could not be found. |
2835 | 95.1k | uint32_t GetContainerTypeId(Instruction* inst) { |
2836 | 95.1k | assert(inst->opcode() == spv::Op::OpCompositeInsert); |
2837 | 95.1k | analysis::DefUseManager* def_use_manager = inst->context()->get_def_use_mgr(); |
2838 | 95.1k | uint32_t container_type_id = GetElementType( |
2839 | 95.1k | inst->type_id(), inst->begin() + 4, inst->end() - 1, def_use_manager); |
2840 | 95.1k | return container_type_id; |
2841 | 95.1k | } |
2842 | | |
2843 | | // Returns an OpCompositeConstruct instruction that build an object with |
2844 | | // |type_id| out of the values in |values_inserted|. Each value will be |
2845 | | // placed at the index corresponding to the value. The new instruction will |
2846 | | // be placed before |insert_before|. |
2847 | | Instruction* BuildCompositeConstruct( |
2848 | | uint32_t type_id, const std::map<uint32_t, uint32_t>& values_inserted, |
2849 | 10.8k | Instruction* insert_before) { |
2850 | 10.8k | InstructionBuilder ir_builder( |
2851 | 10.8k | insert_before->context(), insert_before, |
2852 | 10.8k | IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping); |
2853 | | |
2854 | 10.8k | std::vector<uint32_t> ids_in_order; |
2855 | 22.0k | for (auto it : values_inserted) { |
2856 | 22.0k | ids_in_order.push_back(it.second); |
2857 | 22.0k | } |
2858 | 10.8k | Instruction* construct = |
2859 | 10.8k | ir_builder.AddCompositeConstruct(type_id, ids_in_order); |
2860 | 10.8k | return construct; |
2861 | 10.8k | } |
2862 | | |
2863 | | // Replaces the OpCompositeInsert |inst| that inserts |construct| into the same |
2864 | | // object as |inst| with final index removed. If the resulting |
2865 | | // OpCompositeInsert instruction would have no remaining indexes, the |
2866 | | // instruction is replaced with an OpCopyObject instead. |
2867 | 10.8k | void InsertConstructedObject(Instruction* inst, const Instruction* construct) { |
2868 | 10.8k | if (inst->NumInOperands() == 3) { |
2869 | 10.7k | inst->SetOpcode(spv::Op::OpCopyObject); |
2870 | 10.7k | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {construct->result_id()}}}); |
2871 | 10.7k | } else { |
2872 | 53 | inst->SetInOperand(kInsertObjectIdInIdx, {construct->result_id()}); |
2873 | 53 | inst->RemoveOperand(inst->NumOperands() - 1); |
2874 | 53 | } |
2875 | 10.8k | } |
2876 | | |
2877 | | // Replaces a series of |OpCompositeInsert| instruction that cover the entire |
2878 | | // object with an |OpCompositeConstruct|. |
2879 | | bool CompositeInsertToCompositeConstruct( |
2880 | | IRContext* context, Instruction* inst, |
2881 | 95.1k | const std::vector<const analysis::Constant*>&) { |
2882 | 95.1k | assert(inst->opcode() == spv::Op::OpCompositeInsert && |
2883 | 95.1k | "Wrong opcode. Should be OpCompositeInsert."); |
2884 | 95.1k | if (inst->NumInOperands() < 3) return false; |
2885 | | |
2886 | 95.1k | std::map<uint32_t, uint32_t> values_inserted = GetInsertedValues(inst); |
2887 | 95.1k | uint32_t container_type_id = GetContainerTypeId(inst); |
2888 | 95.1k | if (container_type_id == 0) { |
2889 | 0 | return false; |
2890 | 0 | } |
2891 | | |
2892 | 95.1k | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
2893 | 95.1k | const analysis::Type* container_type = type_mgr->GetType(container_type_id); |
2894 | 95.1k | assert(container_type && "GetContainerTypeId returned a bad id."); |
2895 | 95.1k | if (!DoInsertedValuesCoverEntireObject(container_type, values_inserted)) { |
2896 | 84.2k | return false; |
2897 | 84.2k | } |
2898 | | |
2899 | 10.8k | Instruction* construct = |
2900 | 10.8k | BuildCompositeConstruct(container_type_id, values_inserted, inst); |
2901 | 10.8k | InsertConstructedObject(inst, construct); |
2902 | 10.8k | return true; |
2903 | 95.1k | } |
2904 | | |
2905 | 16.0k | FoldingRule RedundantPhi() { |
2906 | | // An OpPhi instruction where all values are the same or the result of the phi |
2907 | | // itself, can be replaced by the value itself. |
2908 | 16.0k | return [](IRContext*, Instruction* inst, |
2909 | 322k | const std::vector<const analysis::Constant*>&) { |
2910 | 322k | assert(inst->opcode() == spv::Op::OpPhi && |
2911 | 322k | "Wrong opcode. Should be OpPhi."); |
2912 | | |
2913 | 322k | uint32_t incoming_value = 0; |
2914 | | |
2915 | 742k | for (uint32_t i = 0; i < inst->NumInOperands(); i += 2) { |
2916 | 665k | uint32_t op_id = inst->GetSingleWordInOperand(i); |
2917 | 665k | if (op_id == inst->result_id()) { |
2918 | 60.1k | continue; |
2919 | 60.1k | } |
2920 | | |
2921 | 604k | if (incoming_value == 0) { |
2922 | 322k | incoming_value = op_id; |
2923 | 322k | } else if (op_id != incoming_value) { |
2924 | | // Found two possible value. Can't simplify. |
2925 | 244k | return false; |
2926 | 244k | } |
2927 | 604k | } |
2928 | | |
2929 | 77.6k | if (incoming_value == 0) { |
2930 | | // Code looks invalid. Don't do anything. |
2931 | 0 | return false; |
2932 | 0 | } |
2933 | | |
2934 | | // We have a single incoming value. Simplify using that value. |
2935 | 77.6k | inst->SetOpcode(spv::Op::OpCopyObject); |
2936 | 77.6k | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {incoming_value}}}); |
2937 | 77.6k | return true; |
2938 | 77.6k | }; |
2939 | 16.0k | } |
2940 | | |
2941 | 16.0k | FoldingRule BitCastScalarOrVector() { |
2942 | 16.0k | return [](IRContext* context, Instruction* inst, |
2943 | 16.0k | const std::vector<const analysis::Constant*>& constants) { |
2944 | 2.58k | assert(inst->opcode() == spv::Op::OpBitcast && constants.size() == 1); |
2945 | 2.58k | if (constants[0] == nullptr) return false; |
2946 | | |
2947 | 1.78k | const analysis::Type* type = |
2948 | 1.78k | context->get_type_mgr()->GetType(inst->type_id()); |
2949 | 1.78k | if (HasFloatingPoint(type) && !inst->IsFloatingPointFoldingAllowed()) |
2950 | 9 | return false; |
2951 | | |
2952 | 1.78k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
2953 | 1.78k | std::vector<uint32_t> words = |
2954 | 1.78k | GetWordsFromNumericScalarOrVectorConstant(const_mgr, constants[0]); |
2955 | 1.78k | if (words.size() == 0) return false; |
2956 | | |
2957 | 1.78k | const analysis::Constant* bitcasted_constant = |
2958 | 1.78k | ConvertWordsToNumericScalarOrVectorConstant(const_mgr, words, type); |
2959 | 1.78k | if (!bitcasted_constant) return false; |
2960 | | |
2961 | 1.78k | auto new_feeder_id = |
2962 | 1.78k | const_mgr->GetDefiningInstruction(bitcasted_constant, inst->type_id()) |
2963 | 1.78k | ->result_id(); |
2964 | 1.78k | inst->SetOpcode(spv::Op::OpCopyObject); |
2965 | 1.78k | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {new_feeder_id}}}); |
2966 | 1.78k | return true; |
2967 | 1.78k | }; |
2968 | 16.0k | } |
2969 | | |
2970 | | // Remove indirect bitcasts which have no effect. |
2971 | | // uint32 x; asuint32(x) => x |
2972 | | // uint32 x; asuint32(asint32(x)) => x |
2973 | | // float32 x; asuint32(asint32(x)) => asuint32(x) |
2974 | 16.0k | FoldingRule RedundantBitcast() { |
2975 | 16.0k | return [](IRContext* context, Instruction* inst, |
2976 | 16.0k | const std::vector<const analysis::Constant*>&) { |
2977 | 802 | assert(inst->opcode() == spv::Op::OpBitcast); |
2978 | | |
2979 | 802 | analysis::DefUseManager* def_mgr = context->get_def_use_mgr(); |
2980 | 802 | Instruction* child = def_mgr->GetDef(inst->GetSingleWordInOperand(0)); |
2981 | | |
2982 | 802 | if (inst->type_id() == child->type_id()) { |
2983 | 146 | inst->SetOpcode(spv::Op::OpCopyObject); |
2984 | 146 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {child->result_id()}}}); |
2985 | 146 | return true; |
2986 | 146 | } |
2987 | | |
2988 | 656 | if (child->opcode() != spv::Op::OpBitcast) { |
2989 | 656 | return false; |
2990 | 656 | } |
2991 | | |
2992 | 0 | if (def_mgr->GetDef(child->GetSingleWordInOperand(0))->type_id() == |
2993 | 0 | inst->type_id()) { |
2994 | 0 | inst->SetOpcode(spv::Op::OpCopyObject); |
2995 | 0 | } |
2996 | 0 | inst->SetInOperands( |
2997 | 0 | {{SPV_OPERAND_TYPE_ID, {child->GetSingleWordInOperand(0)}}}); |
2998 | |
|
2999 | 0 | return true; |
3000 | 656 | }; |
3001 | 16.0k | } |
3002 | | |
3003 | 16.0k | FoldingRule BitReverseScalarOrVector() { |
3004 | 16.0k | return [](IRContext* context, Instruction* inst, |
3005 | 16.0k | const std::vector<const analysis::Constant*>& constants) { |
3006 | 727 | assert(inst->opcode() == spv::Op::OpBitReverse && constants.size() == 1); |
3007 | 727 | if (constants[0] == nullptr) return false; |
3008 | | |
3009 | 482 | const analysis::Type* type = |
3010 | 482 | context->get_type_mgr()->GetType(inst->type_id()); |
3011 | 482 | assert(!HasFloatingPoint(type) && |
3012 | 482 | "BitReverse cannot be applied to floating point types."); |
3013 | 482 | assert((type->AsInteger() || type->AsVector()) && |
3014 | 482 | "BitReverse can only be applied to integer scalars or vectors."); |
3015 | 482 | assert((ElementWidth(type) == 32) && |
3016 | 482 | "BitReverse can only be applied to integer types of width 32"); |
3017 | | |
3018 | 482 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
3019 | 482 | std::vector<uint32_t> words = |
3020 | 482 | GetWordsFromNumericScalarOrVectorConstant(const_mgr, constants[0]); |
3021 | 482 | if (words.size() == 0) return false; |
3022 | | |
3023 | 486 | for (uint32_t& word : words) { |
3024 | | // Reverse the bits in each word. |
3025 | 486 | word = ((word & 0x55555555) << 1) | ((word >> 1) & 0x55555555); |
3026 | 486 | word = ((word & 0x33333333) << 2) | ((word >> 2) & 0x33333333); |
3027 | 486 | word = ((word & 0x0F0F0F0F) << 4) | ((word >> 4) & 0x0F0F0F0F); |
3028 | 486 | word = ((word & 0x00FF00FF) << 8) | ((word >> 8) & 0x00FF00FF); |
3029 | 486 | word = (word << 16) | (word >> 16); |
3030 | 486 | } |
3031 | | |
3032 | 482 | const analysis::Constant* bitreversed_constant = |
3033 | 482 | ConvertWordsToNumericScalarOrVectorConstant(const_mgr, words, type); |
3034 | 482 | if (!bitreversed_constant) return false; |
3035 | | |
3036 | 482 | auto new_feeder_id = |
3037 | 482 | const_mgr->GetDefiningInstruction(bitreversed_constant, inst->type_id()) |
3038 | 482 | ->result_id(); |
3039 | 482 | inst->SetOpcode(spv::Op::OpCopyObject); |
3040 | 482 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {new_feeder_id}}}); |
3041 | 482 | return true; |
3042 | 482 | }; |
3043 | 16.0k | } |
3044 | | |
3045 | 16.0k | FoldingRule RedundantSelect() { |
3046 | | // An OpSelect instruction where both values are the same or the condition is |
3047 | | // constant can be replaced by one of the values |
3048 | 16.0k | return [](IRContext*, Instruction* inst, |
3049 | 16.6k | const std::vector<const analysis::Constant*>& constants) { |
3050 | 16.6k | assert(inst->opcode() == spv::Op::OpSelect && |
3051 | 16.6k | "Wrong opcode. Should be OpSelect."); |
3052 | 16.6k | assert(inst->NumInOperands() == 3); |
3053 | 16.6k | assert(constants.size() == 3); |
3054 | | |
3055 | 16.6k | uint32_t true_id = inst->GetSingleWordInOperand(1); |
3056 | 16.6k | uint32_t false_id = inst->GetSingleWordInOperand(2); |
3057 | | |
3058 | 16.6k | if (true_id == false_id) { |
3059 | | // Both results are the same, condition doesn't matter |
3060 | 118 | inst->SetOpcode(spv::Op::OpCopyObject); |
3061 | 118 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {true_id}}}); |
3062 | 118 | return true; |
3063 | 16.5k | } else if (constants[0]) { |
3064 | 1.39k | const analysis::Type* type = constants[0]->type(); |
3065 | 1.39k | if (type->AsBool()) { |
3066 | | // Scalar constant value, select the corresponding value. |
3067 | 988 | inst->SetOpcode(spv::Op::OpCopyObject); |
3068 | 988 | if (constants[0]->AsNullConstant() || |
3069 | 988 | !constants[0]->AsBoolConstant()->value()) { |
3070 | 682 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {false_id}}}); |
3071 | 682 | } else { |
3072 | 306 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {true_id}}}); |
3073 | 306 | } |
3074 | 988 | return true; |
3075 | 988 | } else { |
3076 | 403 | assert(type->AsVector()); |
3077 | 403 | if (constants[0]->AsNullConstant()) { |
3078 | | // All values come from false id. |
3079 | 0 | inst->SetOpcode(spv::Op::OpCopyObject); |
3080 | 0 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {false_id}}}); |
3081 | 0 | return true; |
3082 | 403 | } else { |
3083 | | // Convert to a vector shuffle. |
3084 | 403 | std::vector<Operand> ops; |
3085 | 403 | ops.push_back({SPV_OPERAND_TYPE_ID, {true_id}}); |
3086 | 403 | ops.push_back({SPV_OPERAND_TYPE_ID, {false_id}}); |
3087 | 403 | const analysis::VectorConstant* vector_const = |
3088 | 403 | constants[0]->AsVectorConstant(); |
3089 | 403 | uint32_t size = |
3090 | 403 | static_cast<uint32_t>(vector_const->GetComponents().size()); |
3091 | 1.21k | for (uint32_t i = 0; i != size; ++i) { |
3092 | 809 | const analysis::Constant* component = |
3093 | 809 | vector_const->GetComponents()[i]; |
3094 | 809 | if (component->AsNullConstant() || |
3095 | 809 | !component->AsBoolConstant()->value()) { |
3096 | | // Selecting from the false vector which is the second input |
3097 | | // vector to the shuffle. Offset the index by |size|. |
3098 | 31 | ops.push_back({SPV_OPERAND_TYPE_LITERAL_INTEGER, {i + size}}); |
3099 | 778 | } else { |
3100 | | // Selecting from true vector which is the first input vector to |
3101 | | // the shuffle. |
3102 | 778 | ops.push_back({SPV_OPERAND_TYPE_LITERAL_INTEGER, {i}}); |
3103 | 778 | } |
3104 | 809 | } |
3105 | | |
3106 | 403 | inst->SetOpcode(spv::Op::OpVectorShuffle); |
3107 | 403 | inst->SetInOperands(std::move(ops)); |
3108 | 403 | return true; |
3109 | 403 | } |
3110 | 403 | } |
3111 | 1.39k | } |
3112 | | |
3113 | 15.1k | return false; |
3114 | 16.6k | }; |
3115 | 16.0k | } |
3116 | | |
3117 | 11.1k | std::optional<bool> GetBoolConstantKind(const analysis::Constant* c) { |
3118 | 11.1k | if (!c) { |
3119 | 3.85k | return {}; |
3120 | 3.85k | } |
3121 | 7.25k | if (auto composite = c->AsCompositeConstant()) { |
3122 | 0 | auto& components = composite->GetComponents(); |
3123 | 0 | if (components.empty()) { |
3124 | 0 | return {}; |
3125 | 0 | } |
3126 | 0 | auto first = GetBoolConstantKind(components[0]); |
3127 | 0 | if (!first) { |
3128 | 0 | return {}; |
3129 | 0 | } |
3130 | 0 | if (std::all_of(std::begin(components) + 1, std::end(components), |
3131 | 0 | [first](const analysis::Constant* c2) { |
3132 | 0 | return GetBoolConstantKind(c2) == first; |
3133 | 0 | })) { |
3134 | 0 | return first; |
3135 | 0 | } |
3136 | 0 | return {}; |
3137 | 7.25k | } else if (c->AsNullConstant()) { |
3138 | 7 | return false; |
3139 | 7.25k | } else if (c->AsBoolConstant()) { |
3140 | 7.25k | return c->AsBoolConstant()->value(); |
3141 | 7.25k | } |
3142 | 0 | return {}; |
3143 | 7.25k | } |
3144 | | |
3145 | | // Fold OpSelect instructions which have constant booleans as their result. |
3146 | | // x ? true : false = x |
3147 | | // x ? false : true = !x |
3148 | 16.0k | FoldingRule FoldConstantBooleanSelect() { |
3149 | 16.0k | return [](IRContext* context, Instruction* inst, |
3150 | 16.0k | const std::vector<const analysis::Constant*>& constants) { |
3151 | 15.1k | assert(inst->opcode() == spv::Op::OpSelect); |
3152 | 15.1k | assert(inst->NumInOperands() == 3); |
3153 | 15.1k | assert(constants.size() == 3); |
3154 | | |
3155 | 15.1k | if (!constants[1] || !constants[2]) { |
3156 | 10.3k | return false; |
3157 | 10.3k | } |
3158 | | |
3159 | 4.82k | analysis::DefUseManager* def_mgr = context->get_def_use_mgr(); |
3160 | 4.82k | if (inst->type_id() != |
3161 | 4.82k | def_mgr->GetDef(inst->GetSingleWordInOperand(0))->type_id()) { |
3162 | 4.74k | return false; |
3163 | 4.74k | } |
3164 | | |
3165 | 85 | std::optional<bool> uniform_true = GetBoolConstantKind(constants[1]); |
3166 | 85 | std::optional<bool> uniform_false = GetBoolConstantKind(constants[2]); |
3167 | | |
3168 | 85 | if (!uniform_true || !uniform_false) { |
3169 | 0 | return false; |
3170 | 0 | } |
3171 | | |
3172 | 85 | if (uniform_true.value() && !uniform_false.value()) { |
3173 | 29 | inst->SetOpcode(spv::Op::OpCopyObject); |
3174 | 29 | inst->SetInOperands( |
3175 | 29 | {{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(0)}}}); |
3176 | 29 | return true; |
3177 | 56 | } else if (!uniform_true.value() && uniform_false.value()) { |
3178 | 56 | inst->SetOpcode(spv::Op::OpLogicalNot); |
3179 | 56 | inst->SetInOperands( |
3180 | 56 | {{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(0)}}}); |
3181 | 56 | return true; |
3182 | 56 | } |
3183 | 0 | return false; |
3184 | 85 | }; |
3185 | 16.0k | } |
3186 | | |
3187 | | // Fold OpLogicalAnd instructions which have a constant true on one side. |
3188 | | // x && true = x |
3189 | | // true && x = x |
3190 | 16.0k | FoldingRule RedundantLogicalAnd() { |
3191 | 16.0k | return [](IRContext* context, Instruction* inst, |
3192 | 16.0k | const std::vector<const analysis::Constant*>& constants) { |
3193 | 8.51k | assert(inst->opcode() == spv::Op::OpLogicalAnd); |
3194 | | |
3195 | 8.51k | if (GetBoolConstantKind(ConstInput(constants)) == |
3196 | 8.51k | std::optional<bool>(true)) { |
3197 | 5.17k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
3198 | 5.17k | inst->SetOpcode(spv::Op::OpCopyObject); |
3199 | 5.17k | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {other_inst->result_id()}}}); |
3200 | 5.17k | return true; |
3201 | 5.17k | } |
3202 | 3.33k | return false; |
3203 | 8.51k | }; |
3204 | 16.0k | } |
3205 | | |
3206 | | // Fold OpLogicalOr instructions which have a constant false on one side. |
3207 | | // x || false = x |
3208 | | // false || x = x |
3209 | 16.0k | FoldingRule RedundantLogicalOr() { |
3210 | 16.0k | return [](IRContext* context, Instruction* inst, |
3211 | 16.0k | const std::vector<const analysis::Constant*>& constants) { |
3212 | 635 | assert(inst->opcode() == spv::Op::OpLogicalOr); |
3213 | | |
3214 | 635 | if (GetBoolConstantKind(ConstInput(constants)) == |
3215 | 635 | std::optional<bool>(false)) { |
3216 | 108 | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
3217 | 108 | inst->SetOpcode(spv::Op::OpCopyObject); |
3218 | 108 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {other_inst->result_id()}}}); |
3219 | 108 | return true; |
3220 | 108 | } |
3221 | 527 | return false; |
3222 | 635 | }; |
3223 | 16.0k | } |
3224 | | |
3225 | | // Fold concurrent OpLogicalNot instructions: |
3226 | | // !!x = x |
3227 | 16.0k | FoldingRule RedundantLogicalNot() { |
3228 | 16.0k | return [](IRContext* context, Instruction* inst, |
3229 | 16.0k | const std::vector<const analysis::Constant*>&) { |
3230 | 15.6k | assert(inst->opcode() == spv::Op::OpLogicalNot); |
3231 | 15.6k | Instruction* child = |
3232 | 15.6k | context->get_def_use_mgr()->GetDef(inst->GetSingleWordInOperand(0)); |
3233 | 15.6k | if (child->opcode() == spv::Op::OpLogicalNot) { |
3234 | 53 | inst->SetOpcode(spv::Op::OpCopyObject); |
3235 | 53 | inst->SetInOperands( |
3236 | 53 | {{SPV_OPERAND_TYPE_ID, {child->GetSingleWordInOperand(0)}}}); |
3237 | 53 | return true; |
3238 | 53 | } |
3239 | 15.6k | return false; |
3240 | 15.6k | }; |
3241 | 16.0k | } |
3242 | | |
3243 | | // Cases handled: |
3244 | | // ((a ? C0 : C1) == C2) = ((a ? (C0 == C2) : (C1 == C2)) |
3245 | | // ((a ? C0 : C1) != C2) = ((a ? (C0 != C2) : (C1 != C2)) |
3246 | | // ((a ? C0 : C1) < C2) = ((a ? (C0 < C2) : (C1 < C2)) |
3247 | | // ((a ? C0 : C1) <= C2) = ((a ? (C0 <= C2) : (C1 <= C2)) |
3248 | | // ((a ? C0 : C1) > C2) = ((a ? (C0 > C2) : (C1 > C2)) |
3249 | | // ((a ? C0 : C1) >= C2) = ((a ? (C0 >= C2) : (C1 >= C2)) |
3250 | | // ((a ? C0 : C1) || C2) = ((a ? (C0 || C2) : (C1 || C2)) |
3251 | | // ((a ? C0 : C1) && C2) = ((a ? (C0 && C2) : (C1 && C2)) |
3252 | | // ((a ? C0 : C1) + C2) = ((a ? (C0 + C2) : (C1 + C2)) |
3253 | | // ((a ? C0 : C1) - C2) = ((a ? (C0 - C2) : (C1 - C2)) |
3254 | | // ((a ? C0 : C1) * C2) = ((a ? (C0 * C2) : (C1 * C2)) |
3255 | | // ((a ? C0 : C1) / C2) = ((a ? (C0 / C2) : (C1 / C2)) |
3256 | | // ((a ? C0 : C1) >> C2) = ((a ? (C0 >> C2) : (C1 >> C2)) |
3257 | | // ((a ? C0 : C1) << C2) = ((a ? (C0 << C2) : (C1 << C2)) |
3258 | | // ((a ? C0 : C1) ^ C2) = ((a ? (C0 ^ C2) : (C1 ^ C2)) |
3259 | | // ((a ? C0 : C1) | C2) = ((a ? (C0 | C2) : (C1 | C2)) |
3260 | | // ((a ? C0 : C1) & C2) = ((a ? (C0 & C2) : (C1 & C2)) |
3261 | | static const constexpr spv::Op MergeBinaryOpSelectOps[] = { |
3262 | | spv::Op::OpLogicalEqual, |
3263 | | spv::Op::OpLogicalNotEqual, |
3264 | | spv::Op::OpLogicalAnd, |
3265 | | spv::Op::OpLogicalOr, |
3266 | | spv::Op::OpIEqual, |
3267 | | spv::Op::OpINotEqual, |
3268 | | spv::Op::OpUGreaterThan, |
3269 | | spv::Op::OpSGreaterThan, |
3270 | | spv::Op::OpUGreaterThanEqual, |
3271 | | spv::Op::OpSGreaterThanEqual, |
3272 | | spv::Op::OpULessThan, |
3273 | | spv::Op::OpSLessThan, |
3274 | | spv::Op::OpULessThanEqual, |
3275 | | spv::Op::OpSLessThanEqual, |
3276 | | spv::Op::OpFOrdEqual, |
3277 | | spv::Op::OpFUnordEqual, |
3278 | | spv::Op::OpFOrdNotEqual, |
3279 | | spv::Op::OpFUnordNotEqual, |
3280 | | spv::Op::OpFOrdLessThan, |
3281 | | spv::Op::OpFUnordLessThan, |
3282 | | spv::Op::OpFOrdGreaterThan, |
3283 | | spv::Op::OpFUnordGreaterThan, |
3284 | | spv::Op::OpFOrdLessThanEqual, |
3285 | | spv::Op::OpFUnordLessThanEqual, |
3286 | | spv::Op::OpFOrdGreaterThanEqual, |
3287 | | spv::Op::OpFUnordGreaterThanEqual, |
3288 | | spv::Op::OpIAdd, |
3289 | | spv::Op::OpFAdd, |
3290 | | spv::Op::OpISub, |
3291 | | spv::Op::OpFSub, |
3292 | | spv::Op::OpIMul, |
3293 | | spv::Op::OpFMul, |
3294 | | spv::Op::OpUDiv, |
3295 | | spv::Op::OpSDiv, |
3296 | | spv::Op::OpFDiv, |
3297 | | spv::Op::OpVectorTimesScalar, |
3298 | | spv::Op::OpShiftRightLogical, |
3299 | | spv::Op::OpShiftRightArithmetic, |
3300 | | spv::Op::OpShiftLeftLogical, |
3301 | | spv::Op::OpBitwiseXor, |
3302 | | spv::Op::OpBitwiseOr, |
3303 | | spv::Op::OpBitwiseAnd}; |
3304 | | |
3305 | 672k | FoldingRule MergeBinaryOpSelect(spv::Op opcode) { |
3306 | 672k | assert(std::find(std::begin(MergeBinaryOpSelectOps), |
3307 | 672k | std::end(MergeBinaryOpSelectOps), |
3308 | 672k | opcode) != std::end(MergeBinaryOpSelectOps) && |
3309 | 672k | "Wrong opcode."); |
3310 | | |
3311 | 672k | return [opcode](IRContext* context, Instruction* inst, |
3312 | 1.59M | const std::vector<const analysis::Constant*>& constants) { |
3313 | 1.59M | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
3314 | 1.59M | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
3315 | 1.59M | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
3316 | | |
3317 | 1.59M | const analysis::Constant* const_input = ConstInput(constants); |
3318 | 1.59M | if (!const_input) { |
3319 | 816k | return false; |
3320 | 816k | } |
3321 | 780k | Instruction* non_const = NonConstInput(context, constants[0], inst); |
3322 | 780k | if (non_const->opcode() != spv::Op::OpSelect) { |
3323 | 779k | return false; |
3324 | 779k | } |
3325 | 966 | std::vector<const analysis::Constant*> select_constants = |
3326 | 966 | const_mgr->GetOperandConstants(non_const); |
3327 | 966 | if (!select_constants[1] || !select_constants[2]) { |
3328 | 329 | return false; |
3329 | 329 | } |
3330 | | |
3331 | | // The OpSelect that will be created below will use the condition from |
3332 | | // `non_const` and a result type matching `inst`. Before SPIR-V 1.4, |
3333 | | // OpSelect could not have a scalar condition with a vector result. |
3334 | | // We must avoid generating the OpSelect if that would happen. |
3335 | 637 | const analysis::Type* result_type = type_mgr->GetType(inst->type_id()); |
3336 | 637 | if (result_type && result_type->AsVector()) { |
3337 | 0 | Instruction* cond_inst = |
3338 | 0 | def_use_mgr->GetDef(non_const->GetSingleWordInOperand(0)); |
3339 | 0 | const analysis::Type* cond_type = type_mgr->GetType(cond_inst->type_id()); |
3340 | 0 | if (cond_type && !cond_type->AsVector()) { |
3341 | 0 | if (spvVersionForTargetEnv(context->grammar().target_env()) < |
3342 | 0 | SPV_SPIRV_VERSION_WORD(1, 4)) { |
3343 | 0 | return false; |
3344 | 0 | } |
3345 | 0 | } |
3346 | 0 | } |
3347 | | |
3348 | 637 | InstructionBuilder ir_builder( |
3349 | 637 | context, inst, |
3350 | 637 | IRContext::kAnalysisDefUse | IRContext::kAnalysisInstrToBlockMapping); |
3351 | | |
3352 | 637 | Instruction *lhs, *rhs; |
3353 | 637 | if (constants[0]) { |
3354 | 319 | lhs = ir_builder.AddBinaryOp(inst->type_id(), opcode, |
3355 | 319 | inst->GetSingleWordInOperand(0), |
3356 | 319 | non_const->GetSingleWordInOperand(1)); |
3357 | 319 | rhs = ir_builder.AddBinaryOp(inst->type_id(), opcode, |
3358 | 319 | inst->GetSingleWordInOperand(0), |
3359 | 319 | non_const->GetSingleWordInOperand(2)); |
3360 | 319 | } else { |
3361 | 318 | lhs = ir_builder.AddBinaryOp(inst->type_id(), opcode, |
3362 | 318 | non_const->GetSingleWordInOperand(1), |
3363 | 318 | inst->GetSingleWordInOperand(1)); |
3364 | 318 | rhs = ir_builder.AddBinaryOp(inst->type_id(), opcode, |
3365 | 318 | non_const->GetSingleWordInOperand(2), |
3366 | 318 | inst->GetSingleWordInOperand(1)); |
3367 | 318 | } |
3368 | | |
3369 | 637 | if (!lhs || !rhs) { |
3370 | 0 | return false; |
3371 | 0 | } |
3372 | | |
3373 | 637 | if (context->get_instruction_folder().FoldInstruction(lhs)) { |
3374 | 637 | context->AnalyzeDefUse(lhs); |
3375 | 1.27k | while (lhs->opcode() == spv::Op::OpCopyObject) { |
3376 | 637 | lhs = def_use_mgr->GetDef(lhs->GetSingleWordInOperand(0)); |
3377 | 637 | } |
3378 | 637 | } |
3379 | 637 | if (context->get_instruction_folder().FoldInstruction(rhs)) { |
3380 | 637 | context->AnalyzeDefUse(rhs); |
3381 | 1.27k | while (rhs->opcode() == spv::Op::OpCopyObject) { |
3382 | 637 | rhs = def_use_mgr->GetDef(rhs->GetSingleWordInOperand(0)); |
3383 | 637 | } |
3384 | 637 | } |
3385 | 637 | inst->SetOpcode(spv::Op::OpSelect); |
3386 | 637 | inst->SetInOperands( |
3387 | 637 | {{SPV_OPERAND_TYPE_ID, {non_const->GetSingleWordInOperand(0)}}, |
3388 | 637 | {SPV_OPERAND_TYPE_ID, {lhs->result_id()}}, |
3389 | 637 | {SPV_OPERAND_TYPE_ID, {rhs->result_id()}}}); |
3390 | 637 | return true; |
3391 | 637 | }; |
3392 | 672k | } |
3393 | | |
3394 | | // Fold OpLogicalNot instructions that follow a comparison, |
3395 | | // if the comparison is only used by that instruction. |
3396 | | // |
3397 | | // !(a == b) = (a != b) |
3398 | | // !(a != b) = (a == b) |
3399 | | // !(a < b) = (a >= b) |
3400 | | // !(a >= b) = (a < b) |
3401 | | // !(a > b) = (a <= b) |
3402 | | // !(a <= b) = (a > b) |
3403 | 16.0k | FoldingRule FoldLogicalNotComparison() { |
3404 | 16.0k | return [](IRContext* context, Instruction* inst, |
3405 | 16.0k | const std::vector<const analysis::Constant*>&) { |
3406 | 15.6k | assert(inst->opcode() == spv::Op::OpLogicalNot); |
3407 | 15.6k | analysis::DefUseManager* def_mgr = context->get_def_use_mgr(); |
3408 | 15.6k | Instruction* child = |
3409 | 15.6k | context->get_def_use_mgr()->GetDef(inst->GetSingleWordInOperand(0)); |
3410 | | |
3411 | 15.6k | if (def_mgr->NumUses(child) > 1) { |
3412 | 8.23k | return false; |
3413 | 8.23k | } |
3414 | | |
3415 | 7.38k | spv::Op new_opcode = spv::Op::OpNop; |
3416 | 7.38k | switch (child->opcode()) { |
3417 | | // (a == b) <=> (a != b) |
3418 | 2 | case spv::Op::OpIEqual: |
3419 | 2 | new_opcode = spv::Op::OpINotEqual; |
3420 | 2 | break; |
3421 | 29 | case spv::Op::OpINotEqual: |
3422 | 29 | new_opcode = spv::Op::OpIEqual; |
3423 | 29 | break; |
3424 | 134 | case spv::Op::OpFOrdEqual: |
3425 | 134 | new_opcode = spv::Op::OpFUnordNotEqual; |
3426 | 134 | break; |
3427 | 127 | case spv::Op::OpFOrdNotEqual: |
3428 | 127 | new_opcode = spv::Op::OpFUnordEqual; |
3429 | 127 | break; |
3430 | 111 | case spv::Op::OpFUnordEqual: |
3431 | 111 | new_opcode = spv::Op::OpFOrdNotEqual; |
3432 | 111 | break; |
3433 | 55 | case spv::Op::OpFUnordNotEqual: |
3434 | 55 | new_opcode = spv::Op::OpFOrdEqual; |
3435 | 55 | break; |
3436 | 9 | case spv::Op::OpLogicalEqual: |
3437 | 9 | new_opcode = spv::Op::OpLogicalNotEqual; |
3438 | 9 | break; |
3439 | 7 | case spv::Op::OpLogicalNotEqual: |
3440 | 7 | new_opcode = spv::Op::OpLogicalEqual; |
3441 | 7 | break; |
3442 | | |
3443 | | // (a > b) <=> (a <= b) |
3444 | 4 | case spv::Op::OpUGreaterThan: |
3445 | 4 | new_opcode = spv::Op::OpULessThanEqual; |
3446 | 4 | break; |
3447 | 4 | case spv::Op::OpULessThanEqual: |
3448 | 4 | new_opcode = spv::Op::OpUGreaterThan; |
3449 | 4 | break; |
3450 | 38 | case spv::Op::OpSGreaterThan: |
3451 | 38 | new_opcode = spv::Op::OpSLessThanEqual; |
3452 | 38 | break; |
3453 | 2 | case spv::Op::OpSLessThanEqual: |
3454 | 2 | new_opcode = spv::Op::OpSGreaterThan; |
3455 | 2 | break; |
3456 | 2.45k | case spv::Op::OpFOrdGreaterThan: |
3457 | 2.45k | new_opcode = spv::Op::OpFUnordLessThanEqual; |
3458 | 2.45k | break; |
3459 | 139 | case spv::Op::OpFOrdLessThanEqual: |
3460 | 139 | new_opcode = spv::Op::OpFUnordGreaterThan; |
3461 | 139 | break; |
3462 | 117 | case spv::Op::OpFUnordGreaterThan: |
3463 | 117 | new_opcode = spv::Op::OpFOrdLessThanEqual; |
3464 | 117 | break; |
3465 | 138 | case spv::Op::OpFUnordLessThanEqual: |
3466 | 138 | new_opcode = spv::Op::OpFOrdGreaterThan; |
3467 | 138 | break; |
3468 | | |
3469 | | // (a < b) <=> (a >= b) |
3470 | 4 | case spv::Op::OpULessThan: |
3471 | 4 | new_opcode = spv::Op::OpUGreaterThanEqual; |
3472 | 4 | break; |
3473 | 5 | case spv::Op::OpUGreaterThanEqual: |
3474 | 5 | new_opcode = spv::Op::OpULessThan; |
3475 | 5 | break; |
3476 | 2 | case spv::Op::OpSLessThan: |
3477 | 2 | new_opcode = spv::Op::OpSGreaterThanEqual; |
3478 | 2 | break; |
3479 | 4 | case spv::Op::OpSGreaterThanEqual: |
3480 | 4 | new_opcode = spv::Op::OpSLessThan; |
3481 | 4 | break; |
3482 | 2.84k | case spv::Op::OpFOrdLessThan: |
3483 | 2.84k | new_opcode = spv::Op::OpFUnordGreaterThanEqual; |
3484 | 2.84k | break; |
3485 | 70 | case spv::Op::OpFOrdGreaterThanEqual: |
3486 | 70 | new_opcode = spv::Op::OpFUnordLessThan; |
3487 | 70 | break; |
3488 | 30 | case spv::Op::OpFUnordLessThan: |
3489 | 30 | new_opcode = spv::Op::OpFOrdGreaterThanEqual; |
3490 | 30 | break; |
3491 | 52 | case spv::Op::OpFUnordGreaterThanEqual: |
3492 | 52 | new_opcode = spv::Op::OpFOrdLessThan; |
3493 | 52 | break; |
3494 | | |
3495 | 998 | default: |
3496 | 998 | break; |
3497 | 7.38k | } |
3498 | | |
3499 | 7.38k | if (new_opcode == spv::Op::OpNop) { |
3500 | 998 | return false; |
3501 | 998 | } |
3502 | | |
3503 | 6.38k | inst->SetOpcode(new_opcode); |
3504 | 6.38k | inst->SetInOperands( |
3505 | 6.38k | {{SPV_OPERAND_TYPE_ID, {child->GetSingleWordInOperand(0)}}, |
3506 | 6.38k | {SPV_OPERAND_TYPE_ID, {child->GetSingleWordInOperand(1)}}}); |
3507 | | |
3508 | 6.38k | return true; |
3509 | 7.38k | }; |
3510 | 16.0k | } |
3511 | | |
3512 | | // (a == true) = a |
3513 | | // (a == false) = !a |
3514 | | // (a != true) = !a |
3515 | | // (a != false) = a |
3516 | 32.0k | FoldingRule RedundantLogicalEqual() { |
3517 | 32.0k | return [](IRContext* context, Instruction* inst, |
3518 | 32.0k | const std::vector<const analysis::Constant*>& constants) { |
3519 | 2.11k | assert(inst->opcode() == spv::Op::OpLogicalEqual || |
3520 | 2.11k | inst->opcode() == spv::Op::OpLogicalNotEqual); |
3521 | | |
3522 | 2.11k | const analysis::Constant* const_input = ConstInput(constants); |
3523 | 2.11k | if (!const_input) { |
3524 | 316 | return false; |
3525 | 316 | } |
3526 | | |
3527 | 1.79k | analysis::DefUseManager* def_mgr = context->get_def_use_mgr(); |
3528 | 1.79k | if (inst->type_id() != |
3529 | 1.79k | def_mgr->GetDef(inst->GetSingleWordInOperand(0))->type_id()) { |
3530 | 0 | return false; |
3531 | 0 | } |
3532 | | |
3533 | 1.79k | std::optional<bool> uniform_const = GetBoolConstantKind(const_input); |
3534 | 1.79k | if (!uniform_const) { |
3535 | 0 | return false; |
3536 | 0 | } |
3537 | | |
3538 | 1.79k | bool direct_copy = inst->opcode() == spv::Op::OpLogicalEqual |
3539 | 1.79k | ? uniform_const.value() |
3540 | 1.79k | : !uniform_const.value(); |
3541 | | |
3542 | 1.79k | inst->SetOpcode(direct_copy ? spv::Op::OpCopyObject |
3543 | 1.79k | : spv::Op::OpLogicalNot); |
3544 | 1.79k | inst->SetInOperands( |
3545 | 1.79k | {{SPV_OPERAND_TYPE_ID, |
3546 | 1.79k | {NonConstInput(context, constants[0], inst)->result_id()}}}); |
3547 | 1.79k | return true; |
3548 | 1.79k | }; |
3549 | 32.0k | } |
3550 | | |
3551 | | enum class FloatConstantKind { Unknown, Zero, One }; |
3552 | | |
3553 | 2.17M | FloatConstantKind getFloatConstantKind(const analysis::Constant* constant) { |
3554 | 2.17M | if (constant == nullptr) { |
3555 | 1.29M | return FloatConstantKind::Unknown; |
3556 | 1.29M | } |
3557 | | |
3558 | 2.17M | assert(HasFloatingPoint(constant->type()) && "Unexpected constant type"); |
3559 | | |
3560 | 886k | if (constant->AsNullConstant()) { |
3561 | 2.71k | return FloatConstantKind::Zero; |
3562 | 883k | } else if (const analysis::VectorConstant* vc = |
3563 | 883k | constant->AsVectorConstant()) { |
3564 | 221k | const std::vector<const analysis::Constant*>& components = |
3565 | 221k | vc->GetComponents(); |
3566 | 221k | assert(!components.empty()); |
3567 | | |
3568 | 221k | FloatConstantKind kind = getFloatConstantKind(components[0]); |
3569 | | |
3570 | 435k | for (size_t i = 1; i < components.size(); ++i) { |
3571 | 259k | if (getFloatConstantKind(components[i]) != kind) { |
3572 | 45.9k | return FloatConstantKind::Unknown; |
3573 | 45.9k | } |
3574 | 259k | } |
3575 | | |
3576 | 175k | return kind; |
3577 | 662k | } else if (const analysis::FloatConstant* fc = constant->AsFloatConstant()) { |
3578 | 662k | if (fc->IsZero()) return FloatConstantKind::Zero; |
3579 | | |
3580 | 585k | uint32_t width = fc->type()->AsFloat()->width(); |
3581 | 585k | if (width != 32 && width != 64) return FloatConstantKind::Unknown; |
3582 | | |
3583 | 585k | double value = (width == 64) ? fc->GetDoubleValue() : fc->GetFloatValue(); |
3584 | | |
3585 | 585k | if (value == 0.0) { |
3586 | 10.4k | return FloatConstantKind::Zero; |
3587 | 574k | } else if (value == 1.0) { |
3588 | 30.3k | return FloatConstantKind::One; |
3589 | 544k | } else { |
3590 | 544k | return FloatConstantKind::Unknown; |
3591 | 544k | } |
3592 | 585k | } else { |
3593 | 0 | return FloatConstantKind::Unknown; |
3594 | 0 | } |
3595 | 886k | } |
3596 | | |
3597 | 16.0k | FoldingRule RedundantFAdd() { |
3598 | 16.0k | return [](IRContext*, Instruction* inst, |
3599 | 424k | const std::vector<const analysis::Constant*>& constants) { |
3600 | 424k | assert(inst->opcode() == spv::Op::OpFAdd && |
3601 | 424k | "Wrong opcode. Should be OpFAdd."); |
3602 | 424k | assert(constants.size() == 2); |
3603 | | |
3604 | 424k | if (!inst->IsFloatingPointFoldingAllowed()) { |
3605 | 6.06k | return false; |
3606 | 6.06k | } |
3607 | | |
3608 | 418k | FloatConstantKind kind0 = getFloatConstantKind(constants[0]); |
3609 | 418k | FloatConstantKind kind1 = getFloatConstantKind(constants[1]); |
3610 | | |
3611 | 418k | if (kind0 == FloatConstantKind::Zero || kind1 == FloatConstantKind::Zero) { |
3612 | 11.1k | inst->SetOpcode(spv::Op::OpCopyObject); |
3613 | 11.1k | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, |
3614 | 11.1k | {inst->GetSingleWordInOperand( |
3615 | 11.1k | kind0 == FloatConstantKind::Zero ? 1 : 0)}}}); |
3616 | 11.1k | return true; |
3617 | 11.1k | } |
3618 | | |
3619 | 407k | return false; |
3620 | 418k | }; |
3621 | 16.0k | } |
3622 | | |
3623 | 16.0k | FoldingRule RedundantFSub() { |
3624 | 16.0k | return [](IRContext*, Instruction* inst, |
3625 | 96.8k | const std::vector<const analysis::Constant*>& constants) { |
3626 | 96.8k | assert(inst->opcode() == spv::Op::OpFSub && |
3627 | 96.8k | "Wrong opcode. Should be OpFSub."); |
3628 | 96.8k | assert(constants.size() == 2); |
3629 | | |
3630 | 96.8k | if (!inst->IsFloatingPointFoldingAllowed()) { |
3631 | 4.86k | return false; |
3632 | 4.86k | } |
3633 | | |
3634 | 92.0k | FloatConstantKind kind0 = getFloatConstantKind(constants[0]); |
3635 | 92.0k | FloatConstantKind kind1 = getFloatConstantKind(constants[1]); |
3636 | | |
3637 | 92.0k | if (kind0 == FloatConstantKind::Zero) { |
3638 | 1.90k | inst->SetOpcode(spv::Op::OpFNegate); |
3639 | 1.90k | inst->SetInOperands( |
3640 | 1.90k | {{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(1)}}}); |
3641 | 1.90k | return true; |
3642 | 1.90k | } |
3643 | | |
3644 | 90.0k | if (kind1 == FloatConstantKind::Zero) { |
3645 | 2.29k | inst->SetOpcode(spv::Op::OpCopyObject); |
3646 | 2.29k | inst->SetInOperands( |
3647 | 2.29k | {{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(0)}}}); |
3648 | 2.29k | return true; |
3649 | 2.29k | } |
3650 | | |
3651 | 87.8k | return false; |
3652 | 90.0k | }; |
3653 | 16.0k | } |
3654 | | |
3655 | 16.0k | FoldingRule RedundantFMul() { |
3656 | 16.0k | return [](IRContext*, Instruction* inst, |
3657 | 197k | const std::vector<const analysis::Constant*>& constants) { |
3658 | 197k | assert(inst->opcode() == spv::Op::OpFMul && |
3659 | 197k | "Wrong opcode. Should be OpFMul."); |
3660 | 197k | assert(constants.size() == 2); |
3661 | | |
3662 | 197k | if (!inst->IsFloatingPointFoldingAllowed()) { |
3663 | 101 | return false; |
3664 | 101 | } |
3665 | | |
3666 | 197k | FloatConstantKind kind0 = getFloatConstantKind(constants[0]); |
3667 | 197k | FloatConstantKind kind1 = getFloatConstantKind(constants[1]); |
3668 | | |
3669 | 197k | if (kind0 == FloatConstantKind::Zero || kind1 == FloatConstantKind::Zero) { |
3670 | 5.42k | inst->SetOpcode(spv::Op::OpCopyObject); |
3671 | 5.42k | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, |
3672 | 5.42k | {inst->GetSingleWordInOperand( |
3673 | 5.42k | kind0 == FloatConstantKind::Zero ? 0 : 1)}}}); |
3674 | 5.42k | return true; |
3675 | 5.42k | } |
3676 | | |
3677 | 192k | if (kind0 == FloatConstantKind::One || kind1 == FloatConstantKind::One) { |
3678 | 1.63k | inst->SetOpcode(spv::Op::OpCopyObject); |
3679 | 1.63k | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, |
3680 | 1.63k | {inst->GetSingleWordInOperand( |
3681 | 1.63k | kind0 == FloatConstantKind::One ? 1 : 0)}}}); |
3682 | 1.63k | return true; |
3683 | 1.63k | } |
3684 | | |
3685 | 190k | return false; |
3686 | 192k | }; |
3687 | 16.0k | } |
3688 | | |
3689 | 16.0k | FoldingRule RedundantFDiv() { |
3690 | 16.0k | return [](IRContext*, Instruction* inst, |
3691 | 104k | const std::vector<const analysis::Constant*>& constants) { |
3692 | 104k | assert(inst->opcode() == spv::Op::OpFDiv && |
3693 | 104k | "Wrong opcode. Should be OpFDiv."); |
3694 | 104k | assert(constants.size() == 2); |
3695 | | |
3696 | 104k | if (!inst->IsFloatingPointFoldingAllowed()) { |
3697 | 27 | return false; |
3698 | 27 | } |
3699 | | |
3700 | 104k | FloatConstantKind kind0 = getFloatConstantKind(constants[0]); |
3701 | 104k | FloatConstantKind kind1 = getFloatConstantKind(constants[1]); |
3702 | | |
3703 | 104k | if (kind0 == FloatConstantKind::Zero || kind1 == FloatConstantKind::One) { |
3704 | 1.24k | inst->SetOpcode(spv::Op::OpCopyObject); |
3705 | 1.24k | inst->SetInOperands( |
3706 | 1.24k | {{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(0)}}}); |
3707 | 1.24k | return true; |
3708 | 1.24k | } |
3709 | | |
3710 | 102k | return false; |
3711 | 104k | }; |
3712 | 16.0k | } |
3713 | | |
3714 | 16.0k | FoldingRule RedundantFMod() { |
3715 | 16.0k | return [](IRContext*, Instruction* inst, |
3716 | 64.1k | const std::vector<const analysis::Constant*>& constants) { |
3717 | 64.1k | assert(inst->opcode() == spv::Op::OpFMod && |
3718 | 64.1k | "Wrong opcode. Should be OpFMod."); |
3719 | 64.1k | assert(constants.size() == 2); |
3720 | | |
3721 | 64.1k | if (!inst->IsFloatingPointFoldingAllowed()) { |
3722 | 248 | return false; |
3723 | 248 | } |
3724 | | |
3725 | 63.9k | FloatConstantKind kind0 = getFloatConstantKind(constants[0]); |
3726 | | |
3727 | 63.9k | if (kind0 == FloatConstantKind::Zero) { |
3728 | 6.07k | inst->SetOpcode(spv::Op::OpCopyObject); |
3729 | 6.07k | inst->SetInOperands( |
3730 | 6.07k | {{SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(0)}}}); |
3731 | 6.07k | return true; |
3732 | 6.07k | } |
3733 | | |
3734 | 57.8k | return false; |
3735 | 63.9k | }; |
3736 | 16.0k | } |
3737 | | |
3738 | 9.83k | FoldingRule RedundantFMix() { |
3739 | 9.83k | return [](IRContext* context, Instruction* inst, |
3740 | 9.83k | const std::vector<const analysis::Constant*>& constants) { |
3741 | 8.96k | assert(inst->opcode() == spv::Op::OpExtInst && |
3742 | 8.96k | "Wrong opcode. Should be OpExtInst."); |
3743 | | |
3744 | 8.96k | if (!inst->IsFloatingPointFoldingAllowed()) { |
3745 | 0 | return false; |
3746 | 0 | } |
3747 | | |
3748 | 8.96k | uint32_t instSetId = |
3749 | 8.96k | context->get_feature_mgr()->GetExtInstImportId_GLSLstd450(); |
3750 | | |
3751 | 8.96k | if (inst->GetSingleWordInOperand(kExtInstSetIdInIdx) == instSetId && |
3752 | 8.96k | inst->GetSingleWordInOperand(kExtInstInstructionInIdx) == |
3753 | 8.96k | GLSLstd450FMix) { |
3754 | 8.96k | assert(constants.size() == 5); |
3755 | | |
3756 | 8.96k | FloatConstantKind kind4 = getFloatConstantKind(constants[4]); |
3757 | | |
3758 | 8.96k | if (kind4 == FloatConstantKind::Zero || kind4 == FloatConstantKind::One) { |
3759 | 116 | inst->SetOpcode(spv::Op::OpCopyObject); |
3760 | 116 | inst->SetInOperands( |
3761 | 116 | {{SPV_OPERAND_TYPE_ID, |
3762 | 116 | {inst->GetSingleWordInOperand(kind4 == FloatConstantKind::Zero |
3763 | 116 | ? kFMixXIdInIdx |
3764 | 116 | : kFMixYIdInIdx)}}}); |
3765 | 116 | return true; |
3766 | 116 | } |
3767 | 8.96k | } |
3768 | | |
3769 | 8.84k | return false; |
3770 | 8.96k | }; |
3771 | 9.83k | } |
3772 | | |
3773 | | // Returns a folding rule that folds the instruction to operand |foldToArg| |
3774 | | // (0 or 1) if operand |arg| (0 or 1) is a zero constant. |
3775 | 272k | FoldingRule RedundantBinaryOpWithZeroOperand(uint32_t arg, uint32_t foldToArg) { |
3776 | 272k | return [arg, foldToArg]( |
3777 | 272k | IRContext* context, Instruction* inst, |
3778 | 316k | const std::vector<const analysis::Constant*>& constants) { |
3779 | 316k | assert(constants.size() == 2); |
3780 | | |
3781 | 316k | if (constants[arg] && constants[arg]->IsZero()) { |
3782 | 6.89k | auto operand = inst->GetSingleWordInOperand(foldToArg); |
3783 | 6.89k | auto operand_type = constants[arg]->type(); |
3784 | | |
3785 | 6.89k | const analysis::Type* inst_type = |
3786 | 6.89k | context->get_type_mgr()->GetType(inst->type_id()); |
3787 | 6.89k | if (inst_type->IsSame(operand_type)) { |
3788 | 6.80k | inst->SetOpcode(spv::Op::OpCopyObject); |
3789 | 6.80k | } else { |
3790 | 88 | inst->SetOpcode(spv::Op::OpBitcast); |
3791 | 88 | } |
3792 | 6.89k | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {operand}}}); |
3793 | 6.89k | return true; |
3794 | 6.89k | } |
3795 | 309k | return false; |
3796 | 316k | }; |
3797 | 272k | } |
3798 | | |
3799 | | // This rule handles any of RedundantBinaryRhs0Ops with a 0 or vector 0 on the |
3800 | | // right-hand side (a | 0 => a). |
3801 | | static const constexpr spv::Op RedundantBinaryRhs0Ops[] = { |
3802 | | spv::Op::OpBitwiseOr, |
3803 | | spv::Op::OpBitwiseXor, |
3804 | | spv::Op::OpShiftRightLogical, |
3805 | | spv::Op::OpShiftRightArithmetic, |
3806 | | spv::Op::OpShiftLeftLogical, |
3807 | | spv::Op::OpIAdd, |
3808 | | spv::Op::OpISub}; |
3809 | 112k | FoldingRule RedundantBinaryRhs0(spv::Op op) { |
3810 | 112k | assert(std::find(std::begin(RedundantBinaryRhs0Ops), |
3811 | 112k | std::end(RedundantBinaryRhs0Ops), |
3812 | 112k | op) != std::end(RedundantBinaryRhs0Ops) && |
3813 | 112k | "Wrong opcode."); |
3814 | 112k | (void)op; |
3815 | 112k | return RedundantBinaryOpWithZeroOperand(1, 0); |
3816 | 112k | } |
3817 | | |
3818 | | // This rule handles any of RedundantBinaryLhs0Ops with a 0 or vector 0 on the |
3819 | | // left-hand side (0 | a => a). |
3820 | | static const constexpr spv::Op RedundantBinaryLhs0Ops[] = { |
3821 | | spv::Op::OpBitwiseOr, spv::Op::OpBitwiseXor, spv::Op::OpIAdd}; |
3822 | 48.0k | FoldingRule RedundantBinaryLhs0(spv::Op op) { |
3823 | 48.0k | assert(std::find(std::begin(RedundantBinaryLhs0Ops), |
3824 | 48.0k | std::end(RedundantBinaryLhs0Ops), |
3825 | 48.0k | op) != std::end(RedundantBinaryLhs0Ops) && |
3826 | 48.0k | "Wrong opcode."); |
3827 | 48.0k | (void)op; |
3828 | 48.0k | return RedundantBinaryOpWithZeroOperand(0, 1); |
3829 | 48.0k | } |
3830 | | |
3831 | | // This rule handles shifts and divisions of 0 or vector 0 by any amount |
3832 | | // (0 >> a => 0). |
3833 | | static const constexpr spv::Op RedundantBinaryLhs0To0Ops[] = { |
3834 | | spv::Op::OpShiftRightLogical, |
3835 | | spv::Op::OpShiftRightArithmetic, |
3836 | | spv::Op::OpShiftLeftLogical, |
3837 | | spv::Op::OpSDiv, |
3838 | | spv::Op::OpUDiv, |
3839 | | spv::Op::OpSMod, |
3840 | | spv::Op::OpUMod}; |
3841 | 112k | FoldingRule RedundantBinaryLhs0To0(spv::Op op) { |
3842 | 112k | assert(std::find(std::begin(RedundantBinaryLhs0To0Ops), |
3843 | 112k | std::end(RedundantBinaryLhs0To0Ops), |
3844 | 112k | op) != std::end(RedundantBinaryLhs0To0Ops) && |
3845 | 112k | "Wrong opcode."); |
3846 | 112k | (void)op; |
3847 | 112k | return RedundantBinaryOpWithZeroOperand(0, 0); |
3848 | 112k | } |
3849 | | |
3850 | 48.0k | FoldingRule ReassociateCommutiveOp() { |
3851 | 48.0k | return [](IRContext* context, Instruction* inst, |
3852 | 48.5k | const std::vector<const analysis::Constant*>& constants) { |
3853 | 48.5k | const analysis::Type* type = |
3854 | 48.5k | context->get_type_mgr()->GetType(inst->type_id()); |
3855 | 48.5k | uint32_t width = ElementWidth(type); |
3856 | 48.5k | if (width != 32) return false; |
3857 | | |
3858 | 48.5k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
3859 | 48.5k | const analysis::Constant* const_input1 = ConstInput(constants); |
3860 | 48.5k | if (!const_input1) return false; |
3861 | 21.3k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
3862 | | |
3863 | 21.3k | if (other_inst->opcode() == inst->opcode()) { |
3864 | 6.90k | std::vector<const analysis::Constant*> other_constants = |
3865 | 6.90k | const_mgr->GetOperandConstants(other_inst); |
3866 | 6.90k | const analysis::Constant* const_input2 = ConstInput(other_constants); |
3867 | 6.90k | if (!const_input2) return false; |
3868 | | |
3869 | 2.00k | Instruction* non_const_input = |
3870 | 2.00k | NonConstInput(context, other_constants[0], other_inst); |
3871 | 2.00k | uint32_t merged_id = PerformOperation(const_mgr, inst->opcode(), |
3872 | 2.00k | const_input1, const_input2); |
3873 | | |
3874 | 2.00k | if (merged_id == 0) return false; |
3875 | 2.00k | inst->SetInOperands( |
3876 | 2.00k | {{SPV_OPERAND_TYPE_ID, {non_const_input->result_id()}}, |
3877 | 2.00k | {SPV_OPERAND_TYPE_ID, {merged_id}}}); |
3878 | 2.00k | return true; |
3879 | 2.00k | } |
3880 | | |
3881 | 14.4k | return false; |
3882 | 21.3k | }; |
3883 | 48.0k | } |
3884 | | |
3885 | | // A | (b | C) = b | (A | C) |
3886 | | // A ^ (b ^ C) = b ^ (A ^ C) |
3887 | | // A & (b & C) = b & (A & C) |
3888 | | // Where A and C are constants |
3889 | | static const constexpr spv::Op ReassociateCommutiveBitwiseOps[] = { |
3890 | | spv::Op::OpBitwiseOr, spv::Op::OpBitwiseXor, spv::Op::OpBitwiseAnd}; |
3891 | 48.0k | FoldingRule ReassociateCommutiveBitwise(spv::Op op) { |
3892 | 48.0k | assert(std::find(std::begin(ReassociateCommutiveBitwiseOps), |
3893 | 48.0k | std::end(ReassociateCommutiveBitwiseOps), |
3894 | 48.0k | op) != std::end(ReassociateCommutiveBitwiseOps) && |
3895 | 48.0k | "Wrong opcode."); |
3896 | 48.0k | (void)op; |
3897 | 48.0k | return ReassociateCommutiveOp(); |
3898 | 48.0k | } |
3899 | | |
3900 | | // Returns true if all elements in |c| are 1. |
3901 | 17.7k | bool IsAllInt1(const analysis::Constant* c) { |
3902 | 17.7k | if (auto composite = c->AsCompositeConstant()) { |
3903 | 0 | auto& components = composite->GetComponents(); |
3904 | 0 | return std::all_of(std::begin(components), std::end(components), IsAllInt1); |
3905 | 17.7k | } else if (c->AsIntConstant()) { |
3906 | 17.7k | return c->GetSignExtendedValue() == 1; |
3907 | 17.7k | } |
3908 | | |
3909 | 25 | return false; |
3910 | 17.7k | } |
3911 | | |
3912 | | // This rule handles divisions by 1 or vector 1 (a / 1 => a). |
3913 | 32.0k | FoldingRule RedundantSUDiv() { |
3914 | 32.0k | return [](IRContext* context, Instruction* inst, |
3915 | 32.0k | const std::vector<const analysis::Constant*>& constants) { |
3916 | 12.4k | assert(constants.size() == 2); |
3917 | 12.4k | assert((inst->opcode() == spv::Op::OpUDiv || |
3918 | 12.4k | inst->opcode() == spv::Op::OpSDiv) && |
3919 | 12.4k | "Wrong opcode."); |
3920 | | |
3921 | 12.4k | if (constants[1] && IsAllInt1(constants[1])) { |
3922 | 952 | auto operand = inst->GetSingleWordInOperand(0); |
3923 | 952 | auto operand_type = constants[1]->type(); |
3924 | | |
3925 | 952 | const analysis::Type* inst_type = |
3926 | 952 | context->get_type_mgr()->GetType(inst->type_id()); |
3927 | 952 | if (inst_type->IsSame(operand_type)) { |
3928 | 770 | inst->SetOpcode(spv::Op::OpCopyObject); |
3929 | 770 | } else { |
3930 | 182 | inst->SetOpcode(spv::Op::OpBitcast); |
3931 | 182 | } |
3932 | 952 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {operand}}}); |
3933 | 952 | return true; |
3934 | 952 | } |
3935 | 11.5k | return false; |
3936 | 12.4k | }; |
3937 | 32.0k | } |
3938 | | |
3939 | | // This rule handles modulo from division by 1 or vector 1 (a % 1 => 0). |
3940 | 32.0k | FoldingRule RedundantSUMod() { |
3941 | 32.0k | return [](IRContext* context, Instruction* inst, |
3942 | 32.0k | const std::vector<const analysis::Constant*>& constants) { |
3943 | 8.04k | assert(constants.size() == 2); |
3944 | 8.04k | assert((inst->opcode() == spv::Op::OpUMod || |
3945 | 8.04k | inst->opcode() == spv::Op::OpSMod) && |
3946 | 8.04k | "Wrong opcode."); |
3947 | | |
3948 | 8.04k | if (constants[1] && IsAllInt1(constants[1])) { |
3949 | 865 | auto type = context->get_type_mgr()->GetType(inst->type_id()); |
3950 | 865 | auto zero_id = context->get_constant_mgr()->GetNullConstId(type); |
3951 | | |
3952 | 865 | inst->SetOpcode(spv::Op::OpCopyObject); |
3953 | 865 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {zero_id}}}); |
3954 | 865 | return true; |
3955 | 865 | } |
3956 | 7.17k | return false; |
3957 | 8.04k | }; |
3958 | 32.0k | } |
3959 | | |
3960 | | // Utility function for applying |callback| to |input1| and |input2|. |
3961 | | // If they are vectors it applies element wise. |
3962 | | // The constants |input1| and |input2| must be integers or a vector of integers. |
3963 | | template <typename Callback> |
3964 | | void ForEachIntegerConstantPair(analysis::ConstantManager* const_mgr, |
3965 | | const analysis::Constant* input1, |
3966 | | const analysis::Constant* input2, |
3967 | 1.57k | Callback&& callback) { |
3968 | 1.57k | assert(input1 && input2); |
3969 | | |
3970 | 1.57k | auto Dispatch = [&callback](const analysis::Constant* lhs, |
3971 | 1.57k | const analysis::Constant* rhs) { |
3972 | 1.57k | assert(lhs->type()->AsInteger()); |
3973 | 1.57k | const analysis::Integer* type = lhs->type()->AsInteger(); |
3974 | 1.57k | uint32_t width = type->AsInteger()->width(); |
3975 | 1.57k | assert(width == 32 || width == 64); |
3976 | 1.57k | if (width == 32) { |
3977 | 1.57k | callback(lhs->GetU32(), rhs->GetU32()); |
3978 | 1.57k | } else { |
3979 | 0 | callback(lhs->GetU64(), rhs->GetU64()); |
3980 | 0 | } |
3981 | 1.57k | }; folding_rules.cpp:spvtools::opt::(anonymous namespace)::ForEachIntegerConstantPair<spvtools::opt::(anonymous namespace)::RedundantAndOrXor()::$_0::operator()(spvtools::opt::IRContext*, spvtools::opt::Instruction*, std::__1::vector<spvtools::opt::analysis::Constant const*, std::__1::allocator<spvtools::opt::analysis::Constant const*> > const&) const::{lambda(auto:1, auto:2)#1}>(spvtools::opt::analysis::ConstantManager*, spvtools::opt::analysis::Constant const*, spvtools::opt::analysis::Constant const*, spvtools::opt::(anonymous namespace)::RedundantAndOrXor()::$_0::operator()(spvtools::opt::IRContext*, spvtools::opt::Instruction*, std::__1::vector<spvtools::opt::analysis::Constant const*, std::__1::allocator<spvtools::opt::analysis::Constant const*> > const&) const::{lambda(auto:1, auto:2)#1}&&)::{lambda(spvtools::opt::analysis::Constant const*, spvtools::opt::analysis::Constant const*)#1}::operator()(spvtools::opt::analysis::Constant const*, spvtools::opt::analysis::Constant const*) constLine | Count | Source | 3971 | 769 | const analysis::Constant* rhs) { | 3972 | 769 | assert(lhs->type()->AsInteger()); | 3973 | 769 | const analysis::Integer* type = lhs->type()->AsInteger(); | 3974 | 769 | uint32_t width = type->AsInteger()->width(); | 3975 | 769 | assert(width == 32 || width == 64); | 3976 | 769 | if (width == 32) { | 3977 | 769 | callback(lhs->GetU32(), rhs->GetU32()); | 3978 | 769 | } else { | 3979 | 0 | callback(lhs->GetU64(), rhs->GetU64()); | 3980 | 0 | } | 3981 | 769 | }; |
folding_rules.cpp:spvtools::opt::(anonymous namespace)::ForEachIntegerConstantPair<spvtools::opt::(anonymous namespace)::RedundantAndAddSub()::$_0::operator()(spvtools::opt::IRContext*, spvtools::opt::Instruction*, std::__1::vector<spvtools::opt::analysis::Constant const*, std::__1::allocator<spvtools::opt::analysis::Constant const*> > const&) const::{lambda(auto:1, auto:2)#1}>(spvtools::opt::analysis::ConstantManager*, spvtools::opt::analysis::Constant const*, spvtools::opt::analysis::Constant const*, spvtools::opt::(anonymous namespace)::RedundantAndAddSub()::$_0::operator()(spvtools::opt::IRContext*, spvtools::opt::Instruction*, std::__1::vector<spvtools::opt::analysis::Constant const*, std::__1::allocator<spvtools::opt::analysis::Constant const*> > const&) const::{lambda(auto:1, auto:2)#1}&&)::{lambda(spvtools::opt::analysis::Constant const*, spvtools::opt::analysis::Constant const*)#1}::operator()(spvtools::opt::analysis::Constant const*, spvtools::opt::analysis::Constant const*) constLine | Count | Source | 3971 | 801 | const analysis::Constant* rhs) { | 3972 | 801 | assert(lhs->type()->AsInteger()); | 3973 | 801 | const analysis::Integer* type = lhs->type()->AsInteger(); | 3974 | 801 | uint32_t width = type->AsInteger()->width(); | 3975 | 801 | assert(width == 32 || width == 64); | 3976 | 801 | if (width == 32) { | 3977 | 801 | callback(lhs->GetU32(), rhs->GetU32()); | 3978 | 801 | } else { | 3979 | 0 | callback(lhs->GetU64(), rhs->GetU64()); | 3980 | 0 | } | 3981 | 801 | }; |
|
3982 | | |
3983 | 1.57k | const analysis::Type* type = input1->type(); |
3984 | 1.57k | if (const analysis::Vector* vector_type = type->AsVector()) { |
3985 | 0 | const analysis::Type* ele_type = vector_type->element_type(); |
3986 | 0 | assert(ele_type->AsInteger()); |
3987 | 0 | for (uint32_t i = 0; i != vector_type->element_count(); ++i) { |
3988 | 0 | const analysis::Constant* input1_comp = nullptr; |
3989 | 0 | if (const analysis::VectorConstant* input1_vector = |
3990 | 0 | input1->AsVectorConstant()) { |
3991 | 0 | input1_comp = input1_vector->GetComponents()[i]; |
3992 | 0 | } else { |
3993 | 0 | assert(input1->AsNullConstant()); |
3994 | 0 | input1_comp = const_mgr->GetConstant(ele_type, {}); |
3995 | 0 | } |
3996 | | |
3997 | 0 | const analysis::Constant* input2_comp = nullptr; |
3998 | 0 | if (const analysis::VectorConstant* input2_vector = |
3999 | 0 | input2->AsVectorConstant()) { |
4000 | 0 | input2_comp = input2_vector->GetComponents()[i]; |
4001 | 0 | } else { |
4002 | 0 | assert(input2->AsNullConstant()); |
4003 | 0 | input2_comp = const_mgr->GetConstant(ele_type, {}); |
4004 | 0 | } |
4005 | | |
4006 | 0 | assert(ele_type->AsInteger()); |
4007 | 0 | Dispatch(input1_comp, input2_comp); |
4008 | 0 | } |
4009 | |
|
4010 | 1.57k | } else { |
4011 | 1.57k | assert(type->AsInteger()); |
4012 | 1.57k | Dispatch(input1, input2); |
4013 | 1.57k | } |
4014 | 1.57k | } folding_rules.cpp:void spvtools::opt::(anonymous namespace)::ForEachIntegerConstantPair<spvtools::opt::(anonymous namespace)::RedundantAndOrXor()::$_0::operator()(spvtools::opt::IRContext*, spvtools::opt::Instruction*, std::__1::vector<spvtools::opt::analysis::Constant const*, std::__1::allocator<spvtools::opt::analysis::Constant const*> > const&) const::{lambda(auto:1, auto:2)#1}>(spvtools::opt::analysis::ConstantManager*, spvtools::opt::analysis::Constant const*, spvtools::opt::analysis::Constant const*, spvtools::opt::(anonymous namespace)::RedundantAndOrXor()::$_0::operator()(spvtools::opt::IRContext*, spvtools::opt::Instruction*, std::__1::vector<spvtools::opt::analysis::Constant const*, std::__1::allocator<spvtools::opt::analysis::Constant const*> > const&) const::{lambda(auto:1, auto:2)#1}&&)Line | Count | Source | 3967 | 769 | Callback&& callback) { | 3968 | 769 | assert(input1 && input2); | 3969 | | | 3970 | 769 | auto Dispatch = [&callback](const analysis::Constant* lhs, | 3971 | 769 | const analysis::Constant* rhs) { | 3972 | 769 | assert(lhs->type()->AsInteger()); | 3973 | 769 | const analysis::Integer* type = lhs->type()->AsInteger(); | 3974 | 769 | uint32_t width = type->AsInteger()->width(); | 3975 | 769 | assert(width == 32 || width == 64); | 3976 | 769 | if (width == 32) { | 3977 | 769 | callback(lhs->GetU32(), rhs->GetU32()); | 3978 | 769 | } else { | 3979 | 769 | callback(lhs->GetU64(), rhs->GetU64()); | 3980 | 769 | } | 3981 | 769 | }; | 3982 | | | 3983 | 769 | const analysis::Type* type = input1->type(); | 3984 | 769 | if (const analysis::Vector* vector_type = type->AsVector()) { | 3985 | 0 | const analysis::Type* ele_type = vector_type->element_type(); | 3986 | 0 | assert(ele_type->AsInteger()); | 3987 | 0 | for (uint32_t i = 0; i != vector_type->element_count(); ++i) { | 3988 | 0 | const analysis::Constant* input1_comp = nullptr; | 3989 | 0 | if (const analysis::VectorConstant* input1_vector = | 3990 | 0 | input1->AsVectorConstant()) { | 3991 | 0 | input1_comp = input1_vector->GetComponents()[i]; | 3992 | 0 | } else { | 3993 | 0 | assert(input1->AsNullConstant()); | 3994 | 0 | input1_comp = const_mgr->GetConstant(ele_type, {}); | 3995 | 0 | } | 3996 | | | 3997 | 0 | const analysis::Constant* input2_comp = nullptr; | 3998 | 0 | if (const analysis::VectorConstant* input2_vector = | 3999 | 0 | input2->AsVectorConstant()) { | 4000 | 0 | input2_comp = input2_vector->GetComponents()[i]; | 4001 | 0 | } else { | 4002 | 0 | assert(input2->AsNullConstant()); | 4003 | 0 | input2_comp = const_mgr->GetConstant(ele_type, {}); | 4004 | 0 | } | 4005 | | | 4006 | 0 | assert(ele_type->AsInteger()); | 4007 | 0 | Dispatch(input1_comp, input2_comp); | 4008 | 0 | } | 4009 | |
| 4010 | 769 | } else { | 4011 | 769 | assert(type->AsInteger()); | 4012 | 769 | Dispatch(input1, input2); | 4013 | 769 | } | 4014 | 769 | } |
folding_rules.cpp:void spvtools::opt::(anonymous namespace)::ForEachIntegerConstantPair<spvtools::opt::(anonymous namespace)::RedundantAndAddSub()::$_0::operator()(spvtools::opt::IRContext*, spvtools::opt::Instruction*, std::__1::vector<spvtools::opt::analysis::Constant const*, std::__1::allocator<spvtools::opt::analysis::Constant const*> > const&) const::{lambda(auto:1, auto:2)#1}>(spvtools::opt::analysis::ConstantManager*, spvtools::opt::analysis::Constant const*, spvtools::opt::analysis::Constant const*, spvtools::opt::(anonymous namespace)::RedundantAndAddSub()::$_0::operator()(spvtools::opt::IRContext*, spvtools::opt::Instruction*, std::__1::vector<spvtools::opt::analysis::Constant const*, std::__1::allocator<spvtools::opt::analysis::Constant const*> > const&) const::{lambda(auto:1, auto:2)#1}&&)Line | Count | Source | 3967 | 801 | Callback&& callback) { | 3968 | 801 | assert(input1 && input2); | 3969 | | | 3970 | 801 | auto Dispatch = [&callback](const analysis::Constant* lhs, | 3971 | 801 | const analysis::Constant* rhs) { | 3972 | 801 | assert(lhs->type()->AsInteger()); | 3973 | 801 | const analysis::Integer* type = lhs->type()->AsInteger(); | 3974 | 801 | uint32_t width = type->AsInteger()->width(); | 3975 | 801 | assert(width == 32 || width == 64); | 3976 | 801 | if (width == 32) { | 3977 | 801 | callback(lhs->GetU32(), rhs->GetU32()); | 3978 | 801 | } else { | 3979 | 801 | callback(lhs->GetU64(), rhs->GetU64()); | 3980 | 801 | } | 3981 | 801 | }; | 3982 | | | 3983 | 801 | const analysis::Type* type = input1->type(); | 3984 | 801 | if (const analysis::Vector* vector_type = type->AsVector()) { | 3985 | 0 | const analysis::Type* ele_type = vector_type->element_type(); | 3986 | 0 | assert(ele_type->AsInteger()); | 3987 | 0 | for (uint32_t i = 0; i != vector_type->element_count(); ++i) { | 3988 | 0 | const analysis::Constant* input1_comp = nullptr; | 3989 | 0 | if (const analysis::VectorConstant* input1_vector = | 3990 | 0 | input1->AsVectorConstant()) { | 3991 | 0 | input1_comp = input1_vector->GetComponents()[i]; | 3992 | 0 | } else { | 3993 | 0 | assert(input1->AsNullConstant()); | 3994 | 0 | input1_comp = const_mgr->GetConstant(ele_type, {}); | 3995 | 0 | } | 3996 | | | 3997 | 0 | const analysis::Constant* input2_comp = nullptr; | 3998 | 0 | if (const analysis::VectorConstant* input2_vector = | 3999 | 0 | input2->AsVectorConstant()) { | 4000 | 0 | input2_comp = input2_vector->GetComponents()[i]; | 4001 | 0 | } else { | 4002 | 0 | assert(input2->AsNullConstant()); | 4003 | 0 | input2_comp = const_mgr->GetConstant(ele_type, {}); | 4004 | 0 | } | 4005 | | | 4006 | 0 | assert(ele_type->AsInteger()); | 4007 | 0 | Dispatch(input1_comp, input2_comp); | 4008 | 0 | } | 4009 | |
| 4010 | 801 | } else { | 4011 | 801 | assert(type->AsInteger()); | 4012 | 801 | Dispatch(input1, input2); | 4013 | 801 | } | 4014 | 801 | } |
|
4015 | | |
4016 | | // Folds redundant xor and or ops that are part of an and. |
4017 | | // Cases handled: |
4018 | | // 0b1110 & (a | 0b0001) = a & 0b1110 |
4019 | | // 0b1110 & (a ^ 0b0001) = a & 0b1110 |
4020 | | // 0b0110 & (a | 0b1110) = 0b0110 |
4021 | 16.0k | FoldingRule RedundantAndOrXor() { |
4022 | 16.0k | return [](IRContext* context, Instruction* inst, |
4023 | 16.0k | const std::vector<const analysis::Constant*>& constants) { |
4024 | 13.2k | assert(inst->opcode() == spv::Op::OpBitwiseAnd && "Wrong opcode."); |
4025 | 13.2k | const analysis::Type* type = |
4026 | 13.2k | context->get_type_mgr()->GetType(inst->type_id()); |
4027 | 13.2k | uint32_t width = ElementWidth(type); |
4028 | 13.2k | if ((width != 32) && (width != 64)) return false; |
4029 | | |
4030 | 13.2k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
4031 | 13.2k | const analysis::Constant* const_input1 = ConstInput(constants); |
4032 | 13.2k | if (!const_input1) return false; |
4033 | 9.97k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
4034 | | |
4035 | 9.97k | if (other_inst->opcode() == spv::Op::OpBitwiseOr || |
4036 | 8.48k | other_inst->opcode() == spv::Op::OpBitwiseXor) { |
4037 | 1.60k | std::vector<const analysis::Constant*> other_constants = |
4038 | 1.60k | const_mgr->GetOperandConstants(other_inst); |
4039 | 1.60k | const analysis::Constant* const_input2 = ConstInput(other_constants); |
4040 | 1.60k | if (!const_input2) return false; |
4041 | | |
4042 | 769 | bool can_convert_to_const = other_inst->opcode() == spv::Op::OpBitwiseOr; |
4043 | 769 | bool can_remove_inner = true; |
4044 | | |
4045 | 769 | ForEachIntegerConstantPair( |
4046 | 769 | const_mgr, const_input1, const_input2, |
4047 | 769 | [&can_remove_inner, &can_convert_to_const](auto lhs, auto rhs) { |
4048 | | // Only convert to const if 'and' is a subset of 'or' |
4049 | 769 | can_convert_to_const = can_convert_to_const && ((lhs & rhs) == lhs); |
4050 | | // Only remove 'xor'/'or' if no bits intersect with 'and' |
4051 | 769 | can_remove_inner = can_remove_inner && ((lhs & rhs) == 0); |
4052 | 769 | }); folding_rules.cpp:auto spvtools::opt::(anonymous namespace)::RedundantAndOrXor()::$_0::operator()(spvtools::opt::IRContext*, spvtools::opt::Instruction*, std::__1::vector<spvtools::opt::analysis::Constant const*, std::__1::allocator<spvtools::opt::analysis::Constant const*> > const&) const::{lambda(auto:1, auto:2)#1}::operator()<unsigned int, unsigned int>(unsigned int, unsigned int) constLine | Count | Source | 4047 | 769 | [&can_remove_inner, &can_convert_to_const](auto lhs, auto rhs) { | 4048 | | // Only convert to const if 'and' is a subset of 'or' | 4049 | 769 | can_convert_to_const = can_convert_to_const && ((lhs & rhs) == lhs); | 4050 | | // Only remove 'xor'/'or' if no bits intersect with 'and' | 4051 | 769 | can_remove_inner = can_remove_inner && ((lhs & rhs) == 0); | 4052 | 769 | }); |
Unexecuted instantiation: folding_rules.cpp:auto spvtools::opt::(anonymous namespace)::RedundantAndOrXor()::$_0::operator()(spvtools::opt::IRContext*, spvtools::opt::Instruction*, std::__1::vector<spvtools::opt::analysis::Constant const*, std::__1::allocator<spvtools::opt::analysis::Constant const*> > const&) const::{lambda(auto:1, auto:2)#1}::operator()<unsigned long, unsigned long>(unsigned long, unsigned long) const |
4053 | | |
4054 | 769 | if (can_convert_to_const) { |
4055 | 63 | Instruction* const_inst = |
4056 | 63 | const_mgr->GetDefiningInstruction(const_input1); |
4057 | 63 | inst->SetOpcode(spv::Op::OpCopyObject); |
4058 | 63 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {const_inst->result_id()}}}); |
4059 | 63 | return true; |
4060 | 706 | } else if (can_remove_inner) { |
4061 | 59 | Instruction* non_const_input = |
4062 | 59 | NonConstInput(context, other_constants[0], other_inst); |
4063 | 59 | Instruction* const_inst = |
4064 | 59 | const_mgr->GetDefiningInstruction(const_input1); |
4065 | 59 | inst->SetInOperands( |
4066 | 59 | {{SPV_OPERAND_TYPE_ID, {non_const_input->result_id()}}, |
4067 | 59 | {SPV_OPERAND_TYPE_ID, {const_inst->result_id()}}}); |
4068 | 59 | return true; |
4069 | 59 | } |
4070 | 769 | } |
4071 | 9.02k | return false; |
4072 | 9.97k | }; |
4073 | 16.0k | } |
4074 | | |
4075 | | // Folds redundant add and sub ops that are part of an and. |
4076 | | // Cases handled: |
4077 | | // 1 & (b + 2) = b & 1 |
4078 | | // 1 & (b - 2) = b & 1 |
4079 | 16.0k | FoldingRule RedundantAndAddSub() { |
4080 | 16.0k | return [](IRContext* context, Instruction* inst, |
4081 | 16.0k | const std::vector<const analysis::Constant*>& constants) { |
4082 | 13.1k | assert(inst->opcode() == spv::Op::OpBitwiseAnd && "Wrong opcode."); |
4083 | 13.1k | const analysis::Type* type = |
4084 | 13.1k | context->get_type_mgr()->GetType(inst->type_id()); |
4085 | 13.1k | uint32_t width = ElementWidth(type); |
4086 | 13.1k | if ((width != 32) && (width != 64)) return false; |
4087 | | |
4088 | 13.1k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
4089 | 13.1k | const analysis::Constant* const_input1 = ConstInput(constants); |
4090 | 13.1k | if (!const_input1) return false; |
4091 | 9.85k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
4092 | | |
4093 | 9.85k | if (other_inst->opcode() != spv::Op::OpIAdd && |
4094 | 9.05k | other_inst->opcode() != spv::Op::OpISub) { |
4095 | 8.74k | return false; |
4096 | 8.74k | } |
4097 | 1.10k | std::vector<const analysis::Constant*> other_constants = |
4098 | 1.10k | const_mgr->GetOperandConstants(other_inst); |
4099 | 1.10k | const analysis::Constant* const_input2 = ConstInput(other_constants); |
4100 | 1.10k | if (!const_input2) return false; |
4101 | | |
4102 | | // Only valid for subtraction if const is on the right |
4103 | 870 | if ((other_inst->opcode() == spv::Op::OpISub) && other_constants[0]) { |
4104 | 69 | return false; |
4105 | 69 | } |
4106 | | |
4107 | 801 | bool can_remove_inner = true; |
4108 | 801 | ForEachIntegerConstantPair(const_mgr, const_input1, const_input2, |
4109 | 801 | [&can_remove_inner](auto and_op, auto add_op) { |
4110 | 801 | if (can_remove_inner) { |
4111 | | // Only valid if no bits from the +/- could |
4112 | | // affect bits from the & operation. |
4113 | 801 | can_remove_inner = |
4114 | 801 | utils::LSB(add_op) > and_op; |
4115 | 801 | } |
4116 | 801 | }); folding_rules.cpp:auto spvtools::opt::(anonymous namespace)::RedundantAndAddSub()::$_0::operator()(spvtools::opt::IRContext*, spvtools::opt::Instruction*, std::__1::vector<spvtools::opt::analysis::Constant const*, std::__1::allocator<spvtools::opt::analysis::Constant const*> > const&) const::{lambda(auto:1, auto:2)#1}::operator()<unsigned int, unsigned int>(unsigned int, unsigned int) constLine | Count | Source | 4109 | 801 | [&can_remove_inner](auto and_op, auto add_op) { | 4110 | 801 | if (can_remove_inner) { | 4111 | | // Only valid if no bits from the +/- could | 4112 | | // affect bits from the & operation. | 4113 | 801 | can_remove_inner = | 4114 | 801 | utils::LSB(add_op) > and_op; | 4115 | 801 | } | 4116 | 801 | }); |
Unexecuted instantiation: folding_rules.cpp:auto spvtools::opt::(anonymous namespace)::RedundantAndAddSub()::$_0::operator()(spvtools::opt::IRContext*, spvtools::opt::Instruction*, std::__1::vector<spvtools::opt::analysis::Constant const*, std::__1::allocator<spvtools::opt::analysis::Constant const*> > const&) const::{lambda(auto:1, auto:2)#1}::operator()<unsigned long, unsigned long>(unsigned long, unsigned long) const |
4117 | | |
4118 | 801 | if (can_remove_inner) { |
4119 | 71 | Instruction* non_const_input = |
4120 | 71 | NonConstInput(context, other_constants[0], other_inst); |
4121 | 71 | Instruction* const_inst = const_mgr->GetDefiningInstruction(const_input1); |
4122 | 71 | inst->SetInOperands( |
4123 | 71 | {{SPV_OPERAND_TYPE_ID, {non_const_input->result_id()}}, |
4124 | 71 | {SPV_OPERAND_TYPE_ID, {const_inst->result_id()}}}); |
4125 | 71 | return true; |
4126 | 71 | } |
4127 | 730 | return false; |
4128 | 801 | }; |
4129 | 16.0k | } |
4130 | | |
4131 | | // Folds redundant shift ops that are part of an and. |
4132 | | // Cases handled: |
4133 | | // 1 & (b << 1) = 0 |
4134 | | // 0x80000000 & (b >> 1) = 0 |
4135 | 16.0k | FoldingRule RedundantAndShift() { |
4136 | 16.0k | return [](IRContext* context, Instruction* inst, |
4137 | 16.0k | const std::vector<const analysis::Constant*>& constants) { |
4138 | 13.0k | assert(inst->opcode() == spv::Op::OpBitwiseAnd && "Wrong opcode."); |
4139 | 13.0k | const analysis::Type* type = |
4140 | 13.0k | context->get_type_mgr()->GetType(inst->type_id()); |
4141 | 13.0k | uint32_t width = ElementWidth(type); |
4142 | 13.0k | if (width != 8 && width != 16 && width != 32 && width != 64) return false; |
4143 | 13.0k | const uint64_t width_mask = |
4144 | 13.0k | (width == 64) ? ~0ull : ((1ull << width) - 1ull); |
4145 | | |
4146 | 13.0k | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
4147 | 13.0k | const analysis::Constant* const_input1 = ConstInput(constants); |
4148 | 13.0k | if (!const_input1) return false; |
4149 | 9.78k | Instruction* other_inst = NonConstInput(context, constants[0], inst); |
4150 | | |
4151 | 9.78k | spv::Op other_op = other_inst->opcode(); |
4152 | 9.78k | if (other_op != spv::Op::OpShiftLeftLogical && |
4153 | 9.59k | other_op != spv::Op::OpShiftRightLogical) { |
4154 | 9.30k | return false; |
4155 | 9.30k | } |
4156 | | |
4157 | 477 | std::vector<const analysis::Constant*> other_constants = |
4158 | 477 | const_mgr->GetOperandConstants(other_inst); |
4159 | | |
4160 | | // Only valid if const is on the right. |
4161 | 477 | if (other_constants[0]) return false; |
4162 | 375 | const analysis::Constant* const_input2 = other_constants[1]; |
4163 | 375 | if (!const_input2) return false; |
4164 | | |
4165 | 146 | auto get_value_u64 = |
4166 | 292 | [](const analysis::Constant* c) -> std::optional<uint64_t> { |
4167 | 292 | if (!c) return std::nullopt; |
4168 | 292 | const analysis::Integer* int_t = c->type()->AsInteger(); |
4169 | 292 | if (!int_t) return std::nullopt; |
4170 | 292 | return c->GetZeroExtendedValue(); |
4171 | 292 | }; |
4172 | | |
4173 | 146 | auto can_fold_component = |
4174 | 146 | [&](const analysis::Constant* mask_const, |
4175 | 146 | const analysis::Constant* shift_const) -> std::optional<bool> { |
4176 | 146 | auto lhs = get_value_u64(mask_const); |
4177 | 146 | auto rhs = get_value_u64(shift_const); |
4178 | 146 | if (!lhs || !rhs) return std::nullopt; |
4179 | 146 | if (*rhs >= width) return false; |
4180 | 146 | uint64_t lhs_masked = *lhs & width_mask; |
4181 | 146 | if (other_op == spv::Op::OpShiftRightLogical) { |
4182 | 80 | return ((lhs_masked << *rhs) & width_mask) == 0; |
4183 | 80 | } |
4184 | 66 | return ((lhs_masked >> *rhs) & width_mask) == 0; |
4185 | 146 | }; |
4186 | | |
4187 | 146 | if (const analysis::Vector* mask_vec = type->AsVector()) { |
4188 | 0 | const analysis::Vector* shift_vec = const_input2->type()->AsVector(); |
4189 | 0 | if (!shift_vec || |
4190 | 0 | shift_vec->element_count() != mask_vec->element_count()) { |
4191 | 0 | return false; |
4192 | 0 | } |
4193 | 0 | const auto mask_components = const_input1->GetVectorComponents(const_mgr); |
4194 | 0 | const auto shift_components = |
4195 | 0 | const_input2->GetVectorComponents(const_mgr); |
4196 | 0 | for (uint32_t i = 0; i != mask_vec->element_count(); ++i) { |
4197 | 0 | auto result = |
4198 | 0 | can_fold_component(mask_components[i], shift_components[i]); |
4199 | 0 | if (!result || !*result) return false; |
4200 | 0 | } |
4201 | 146 | } else { |
4202 | 146 | if (const_input2->type()->AsVector()) return false; |
4203 | 146 | auto result = can_fold_component(const_input1, const_input2); |
4204 | 146 | if (!result || !*result) return false; |
4205 | 146 | } |
4206 | | |
4207 | 31 | auto zero_id = context->get_constant_mgr()->GetNullConstId(type); |
4208 | 31 | inst->SetOpcode(spv::Op::OpCopyObject); |
4209 | 31 | inst->SetInOperands({{SPV_OPERAND_TYPE_ID, {zero_id}}}); |
4210 | 31 | return true; |
4211 | 146 | }; |
4212 | 16.0k | } |
4213 | | |
4214 | | // This rule look for a dot with a constant vector containing a single 1 and |
4215 | | // the rest 0s. This is the same as doing an extract. |
4216 | 16.0k | FoldingRule DotProductDoingExtract() { |
4217 | 16.0k | return [](IRContext* context, Instruction* inst, |
4218 | 16.0k | const std::vector<const analysis::Constant*>& constants) { |
4219 | 52 | assert(inst->opcode() == spv::Op::OpDot && |
4220 | 52 | "Wrong opcode. Should be OpDot."); |
4221 | | |
4222 | 52 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
4223 | | |
4224 | 52 | if (!inst->IsFloatingPointFoldingAllowed()) { |
4225 | 0 | return false; |
4226 | 0 | } |
4227 | | |
4228 | 154 | for (int i = 0; i < 2; ++i) { |
4229 | 103 | if (!constants[i]) { |
4230 | 58 | continue; |
4231 | 58 | } |
4232 | | |
4233 | 45 | const analysis::Vector* vector_type = constants[i]->type()->AsVector(); |
4234 | 45 | assert(vector_type && "Inputs to OpDot must be vectors."); |
4235 | 45 | const analysis::Float* element_type = |
4236 | 45 | vector_type->element_type()->AsFloat(); |
4237 | 45 | assert(element_type && "Inputs to OpDot must be vectors of floats."); |
4238 | 45 | uint32_t element_width = element_type->width(); |
4239 | 45 | if (element_width != 32 && element_width != 64) { |
4240 | 0 | return false; |
4241 | 0 | } |
4242 | | |
4243 | 45 | std::vector<const analysis::Constant*> components; |
4244 | 45 | components = constants[i]->GetVectorComponents(const_mgr); |
4245 | | |
4246 | 45 | constexpr uint32_t kNotFound = std::numeric_limits<uint32_t>::max(); |
4247 | | |
4248 | 45 | uint32_t component_with_one = kNotFound; |
4249 | 45 | bool all_others_zero = true; |
4250 | 49 | for (uint32_t j = 0; j < components.size(); ++j) { |
4251 | 48 | const analysis::Constant* element = components[j]; |
4252 | 48 | double value = |
4253 | 48 | (element_width == 32 ? element->GetFloat() : element->GetDouble()); |
4254 | 48 | if (value == 0.0) { |
4255 | 2 | continue; |
4256 | 46 | } else if (value == 1.0) { |
4257 | 2 | if (component_with_one == kNotFound) { |
4258 | 2 | component_with_one = j; |
4259 | 2 | } else { |
4260 | 0 | component_with_one = kNotFound; |
4261 | 0 | break; |
4262 | 0 | } |
4263 | 44 | } else { |
4264 | 44 | all_others_zero = false; |
4265 | 44 | break; |
4266 | 44 | } |
4267 | 48 | } |
4268 | | |
4269 | 45 | if (!all_others_zero || component_with_one == kNotFound) { |
4270 | 44 | continue; |
4271 | 44 | } |
4272 | | |
4273 | 1 | std::vector<Operand> operands; |
4274 | 1 | operands.push_back( |
4275 | 1 | {SPV_OPERAND_TYPE_ID, {inst->GetSingleWordInOperand(1u - i)}}); |
4276 | 1 | operands.push_back( |
4277 | 1 | {SPV_OPERAND_TYPE_LITERAL_INTEGER, {component_with_one}}); |
4278 | | |
4279 | 1 | inst->SetOpcode(spv::Op::OpCompositeExtract); |
4280 | 1 | inst->SetInOperands(std::move(operands)); |
4281 | 1 | return true; |
4282 | 45 | } |
4283 | 51 | return false; |
4284 | 52 | }; |
4285 | 16.0k | } |
4286 | | |
4287 | | // If we are storing an undef, then we can remove the store. |
4288 | | // |
4289 | | // TODO: We can do something similar for OpImageWrite, but checking for volatile |
4290 | | // is complicated. Waiting to see if it is needed. |
4291 | 16.0k | FoldingRule StoringUndef() { |
4292 | 16.0k | return [](IRContext* context, Instruction* inst, |
4293 | 1.05M | const std::vector<const analysis::Constant*>&) { |
4294 | 1.05M | assert(inst->opcode() == spv::Op::OpStore && |
4295 | 1.05M | "Wrong opcode. Should be OpStore."); |
4296 | | |
4297 | 1.05M | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
4298 | | |
4299 | | // If this is a volatile store, the store cannot be removed. |
4300 | 1.05M | if (inst->NumInOperands() == 3) { |
4301 | 8.42k | if (inst->GetSingleWordInOperand(2) & |
4302 | 8.42k | uint32_t(spv::MemoryAccessMask::Volatile)) { |
4303 | 5.47k | return false; |
4304 | 5.47k | } |
4305 | 8.42k | } |
4306 | | |
4307 | 1.04M | uint32_t object_id = inst->GetSingleWordInOperand(kStoreObjectInIdx); |
4308 | 1.04M | Instruction* object_inst = def_use_mgr->GetDef(object_id); |
4309 | 1.04M | if (object_inst->opcode() == spv::Op::OpUndef) { |
4310 | 19.1k | inst->ToNop(); |
4311 | 19.1k | return true; |
4312 | 19.1k | } |
4313 | 1.02M | return false; |
4314 | 1.04M | }; |
4315 | 16.0k | } |
4316 | | |
4317 | 16.0k | FoldingRule VectorShuffleFeedingShuffle() { |
4318 | 16.0k | return [](IRContext* context, Instruction* inst, |
4319 | 24.9k | const std::vector<const analysis::Constant*>&) { |
4320 | 24.9k | assert(inst->opcode() == spv::Op::OpVectorShuffle && |
4321 | 24.9k | "Wrong opcode. Should be OpVectorShuffle."); |
4322 | | |
4323 | 24.9k | analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr(); |
4324 | 24.9k | analysis::TypeManager* type_mgr = context->get_type_mgr(); |
4325 | | |
4326 | 24.9k | Instruction* feeding_shuffle_inst = |
4327 | 24.9k | def_use_mgr->GetDef(inst->GetSingleWordInOperand(0)); |
4328 | 24.9k | analysis::Vector* op0_type = |
4329 | 24.9k | type_mgr->GetType(feeding_shuffle_inst->type_id())->AsVector(); |
4330 | 24.9k | uint32_t op0_length = op0_type->element_count(); |
4331 | | |
4332 | 24.9k | bool feeder_is_op0 = true; |
4333 | 24.9k | if (feeding_shuffle_inst->opcode() != spv::Op::OpVectorShuffle) { |
4334 | 24.6k | feeding_shuffle_inst = |
4335 | 24.6k | def_use_mgr->GetDef(inst->GetSingleWordInOperand(1)); |
4336 | 24.6k | feeder_is_op0 = false; |
4337 | 24.6k | } |
4338 | | |
4339 | 24.9k | if (feeding_shuffle_inst->opcode() != spv::Op::OpVectorShuffle) { |
4340 | 24.2k | return false; |
4341 | 24.2k | } |
4342 | | |
4343 | 651 | Instruction* feeder2 = |
4344 | 651 | def_use_mgr->GetDef(feeding_shuffle_inst->GetSingleWordInOperand(0)); |
4345 | 651 | analysis::Vector* feeder_op0_type = |
4346 | 651 | type_mgr->GetType(feeder2->type_id())->AsVector(); |
4347 | 651 | uint32_t feeder_op0_length = feeder_op0_type->element_count(); |
4348 | | |
4349 | 651 | uint32_t new_feeder_id = 0; |
4350 | 651 | std::vector<Operand> new_operands; |
4351 | 651 | new_operands.resize( |
4352 | 651 | 2, {SPV_OPERAND_TYPE_ID, {0}}); // Place holders for vector operands. |
4353 | 651 | const uint32_t undef_literal = 0xffffffff; |
4354 | 2.14k | for (uint32_t op = 2; op < inst->NumInOperands(); ++op) { |
4355 | 1.55k | uint32_t component_index = inst->GetSingleWordInOperand(op); |
4356 | | |
4357 | | // Do not interpret the undefined value literal as coming from operand 1. |
4358 | 1.55k | if (component_index != undef_literal && |
4359 | 1.45k | feeder_is_op0 == (component_index < op0_length)) { |
4360 | | // This component comes from the feeding_shuffle_inst. Update |
4361 | | // |component_index| to be the index into the operand of the feeder. |
4362 | | |
4363 | | // Adjust component_index to get the index into the operands of the |
4364 | | // feeding_shuffle_inst. |
4365 | 802 | if (component_index >= op0_length) { |
4366 | 455 | component_index -= op0_length; |
4367 | 455 | } |
4368 | 802 | component_index = |
4369 | 802 | feeding_shuffle_inst->GetSingleWordInOperand(component_index + 2); |
4370 | | |
4371 | | // Check if we are using a component from the first or second operand of |
4372 | | // the feeding instruction. |
4373 | 802 | if (component_index < feeder_op0_length) { |
4374 | 616 | if (new_feeder_id == 0) { |
4375 | | // First time through, save the id of the operand the element comes |
4376 | | // from. |
4377 | 361 | new_feeder_id = feeding_shuffle_inst->GetSingleWordInOperand(0); |
4378 | 361 | } else if (new_feeder_id != |
4379 | 255 | feeding_shuffle_inst->GetSingleWordInOperand(0)) { |
4380 | | // We need both elements of the feeding_shuffle_inst, so we cannot |
4381 | | // fold. |
4382 | 41 | return false; |
4383 | 41 | } |
4384 | 616 | } else if (component_index != undef_literal) { |
4385 | 121 | if (new_feeder_id == 0) { |
4386 | | // First time through, save the id of the operand the element comes |
4387 | | // from. |
4388 | 82 | new_feeder_id = feeding_shuffle_inst->GetSingleWordInOperand(1); |
4389 | 82 | } else if (new_feeder_id != |
4390 | 39 | feeding_shuffle_inst->GetSingleWordInOperand(1)) { |
4391 | | // We need both elements of the feeding_shuffle_inst, so we cannot |
4392 | | // fold. |
4393 | 22 | return false; |
4394 | 22 | } |
4395 | 99 | component_index -= feeder_op0_length; |
4396 | 99 | } |
4397 | | |
4398 | 739 | if (!feeder_is_op0 && component_index != undef_literal) { |
4399 | 427 | component_index += op0_length; |
4400 | 427 | } |
4401 | 739 | } |
4402 | 1.49k | new_operands.push_back( |
4403 | 1.49k | {SPV_OPERAND_TYPE_LITERAL_INTEGER, {component_index}}); |
4404 | 1.49k | } |
4405 | | |
4406 | 588 | if (new_feeder_id == 0) { |
4407 | 208 | analysis::ConstantManager* const_mgr = context->get_constant_mgr(); |
4408 | 208 | const analysis::Type* type = |
4409 | 208 | type_mgr->GetType(feeding_shuffle_inst->type_id()); |
4410 | 208 | const analysis::Constant* null_const = const_mgr->GetConstant(type, {}); |
4411 | 208 | new_feeder_id = |
4412 | 208 | const_mgr->GetDefiningInstruction(null_const, 0)->result_id(); |
4413 | 208 | } |
4414 | | |
4415 | 588 | if (feeder_is_op0) { |
4416 | | // If the size of the first vector operand changed then the indices |
4417 | | // referring to the second operand need to be adjusted. |
4418 | 170 | Instruction* new_feeder_inst = def_use_mgr->GetDef(new_feeder_id); |
4419 | 170 | analysis::Type* new_feeder_type = |
4420 | 170 | type_mgr->GetType(new_feeder_inst->type_id()); |
4421 | 170 | uint32_t new_op0_size = new_feeder_type->AsVector()->element_count(); |
4422 | 170 | int32_t adjustment = op0_length - new_op0_size; |
4423 | | |
4424 | 170 | if (adjustment != 0) { |
4425 | 254 | for (uint32_t i = 2; i < new_operands.size(); i++) { |
4426 | 171 | uint32_t operand = inst->GetSingleWordInOperand(i); |
4427 | 171 | if (operand >= op0_length && operand != undef_literal) { |
4428 | 49 | new_operands[i].words[0] -= adjustment; |
4429 | 49 | } |
4430 | 171 | } |
4431 | 83 | } |
4432 | | |
4433 | 170 | new_operands[0].words[0] = new_feeder_id; |
4434 | 170 | new_operands[1] = inst->GetInOperand(1); |
4435 | 418 | } else { |
4436 | 418 | new_operands[1].words[0] = new_feeder_id; |
4437 | 418 | new_operands[0] = inst->GetInOperand(0); |
4438 | 418 | } |
4439 | | |
4440 | 588 | inst->SetInOperands(std::move(new_operands)); |
4441 | 588 | return true; |
4442 | 651 | }; |
4443 | 16.0k | } |
4444 | | |
4445 | | // Removes duplicate ids from the interface list of an OpEntryPoint |
4446 | | // instruction. |
4447 | 16.0k | FoldingRule RemoveRedundantOperands() { |
4448 | 16.0k | return [](IRContext*, Instruction* inst, |
4449 | 16.0k | const std::vector<const analysis::Constant*>&) { |
4450 | 0 | assert(inst->opcode() == spv::Op::OpEntryPoint && |
4451 | 0 | "Wrong opcode. Should be OpEntryPoint."); |
4452 | 0 | bool has_redundant_operand = false; |
4453 | 0 | std::unordered_set<uint32_t> seen_operands; |
4454 | 0 | std::vector<Operand> new_operands; |
4455 | |
|
4456 | 0 | new_operands.emplace_back(inst->GetOperand(0)); |
4457 | 0 | new_operands.emplace_back(inst->GetOperand(1)); |
4458 | 0 | new_operands.emplace_back(inst->GetOperand(2)); |
4459 | 0 | for (uint32_t i = 3; i < inst->NumOperands(); ++i) { |
4460 | 0 | if (seen_operands.insert(inst->GetSingleWordOperand(i)).second) { |
4461 | 0 | new_operands.emplace_back(inst->GetOperand(i)); |
4462 | 0 | } else { |
4463 | 0 | has_redundant_operand = true; |
4464 | 0 | } |
4465 | 0 | } |
4466 | |
|
4467 | 0 | if (!has_redundant_operand) { |
4468 | 0 | return false; |
4469 | 0 | } |
4470 | | |
4471 | 0 | inst->SetInOperands(std::move(new_operands)); |
4472 | 0 | return true; |
4473 | 0 | }; |
4474 | 16.0k | } |
4475 | | |
4476 | | // If an image instruction's operand is a constant, updates the image operand |
4477 | | // flag from Offset to ConstOffset. |
4478 | 400k | FoldingRule UpdateImageOperands() { |
4479 | 400k | return [](IRContext*, Instruction* inst, |
4480 | 400k | const std::vector<const analysis::Constant*>& constants) { |
4481 | 237k | const auto opcode = inst->opcode(); |
4482 | 237k | (void)opcode; |
4483 | 237k | assert((opcode == spv::Op::OpImageSampleImplicitLod || |
4484 | 237k | opcode == spv::Op::OpImageSampleExplicitLod || |
4485 | 237k | opcode == spv::Op::OpImageSampleDrefImplicitLod || |
4486 | 237k | opcode == spv::Op::OpImageSampleDrefExplicitLod || |
4487 | 237k | opcode == spv::Op::OpImageSampleProjImplicitLod || |
4488 | 237k | opcode == spv::Op::OpImageSampleProjExplicitLod || |
4489 | 237k | opcode == spv::Op::OpImageSampleProjDrefImplicitLod || |
4490 | 237k | opcode == spv::Op::OpImageSampleProjDrefExplicitLod || |
4491 | 237k | opcode == spv::Op::OpImageFetch || |
4492 | 237k | opcode == spv::Op::OpImageGather || |
4493 | 237k | opcode == spv::Op::OpImageDrefGather || |
4494 | 237k | opcode == spv::Op::OpImageRead || opcode == spv::Op::OpImageWrite || |
4495 | 237k | opcode == spv::Op::OpImageSparseSampleImplicitLod || |
4496 | 237k | opcode == spv::Op::OpImageSparseSampleExplicitLod || |
4497 | 237k | opcode == spv::Op::OpImageSparseSampleDrefImplicitLod || |
4498 | 237k | opcode == spv::Op::OpImageSparseSampleDrefExplicitLod || |
4499 | 237k | opcode == spv::Op::OpImageSparseSampleProjImplicitLod || |
4500 | 237k | opcode == spv::Op::OpImageSparseSampleProjExplicitLod || |
4501 | 237k | opcode == spv::Op::OpImageSparseSampleProjDrefImplicitLod || |
4502 | 237k | opcode == spv::Op::OpImageSparseSampleProjDrefExplicitLod || |
4503 | 237k | opcode == spv::Op::OpImageSparseFetch || |
4504 | 237k | opcode == spv::Op::OpImageSparseGather || |
4505 | 237k | opcode == spv::Op::OpImageSparseDrefGather || |
4506 | 237k | opcode == spv::Op::OpImageSparseRead) && |
4507 | 237k | "Wrong opcode. Should be an image instruction."); |
4508 | | |
4509 | 237k | int32_t operand_index = ImageOperandsMaskInOperandIndex(inst); |
4510 | 237k | if (operand_index >= 0) { |
4511 | 14 | auto image_operands = inst->GetSingleWordInOperand(operand_index); |
4512 | 14 | if (image_operands & uint32_t(spv::ImageOperandsMask::Offset)) { |
4513 | 0 | uint32_t offset_operand_index = operand_index + 1; |
4514 | 0 | if (image_operands & uint32_t(spv::ImageOperandsMask::Bias)) |
4515 | 0 | offset_operand_index++; |
4516 | 0 | if (image_operands & uint32_t(spv::ImageOperandsMask::Lod)) |
4517 | 0 | offset_operand_index++; |
4518 | 0 | if (image_operands & uint32_t(spv::ImageOperandsMask::Grad)) |
4519 | 0 | offset_operand_index += 2; |
4520 | 0 | assert(((image_operands & |
4521 | 0 | uint32_t(spv::ImageOperandsMask::ConstOffset)) == 0) && |
4522 | 0 | "Offset and ConstOffset may not be used together"); |
4523 | 0 | if (offset_operand_index < inst->NumOperands()) { |
4524 | 0 | if (constants[offset_operand_index]) { |
4525 | 0 | if (constants[offset_operand_index]->IsZero()) { |
4526 | 0 | inst->RemoveInOperand(offset_operand_index); |
4527 | 0 | } else { |
4528 | 0 | image_operands = image_operands | |
4529 | 0 | uint32_t(spv::ImageOperandsMask::ConstOffset); |
4530 | 0 | } |
4531 | 0 | image_operands = |
4532 | 0 | image_operands & ~uint32_t(spv::ImageOperandsMask::Offset); |
4533 | 0 | inst->SetInOperand(operand_index, {image_operands}); |
4534 | 0 | return true; |
4535 | 0 | } |
4536 | 0 | } |
4537 | 0 | } |
4538 | 14 | } |
4539 | | |
4540 | 237k | return false; |
4541 | 237k | }; |
4542 | 400k | } |
4543 | | |
4544 | | } // namespace |
4545 | | |
4546 | 16.0k | void FoldingRules::AddFoldingRules() { |
4547 | | // Add all folding rules to the list for the opcodes to which they apply. |
4548 | | // Note that the order in which rules are added to the list matters. If a rule |
4549 | | // applies to the instruction, the rest of the rules will not be attempted. |
4550 | | // Take that into consideration. |
4551 | 16.0k | for (auto op : RedundantBinaryRhs0Ops) |
4552 | 112k | rules_[op].push_back(RedundantBinaryRhs0(op)); |
4553 | 16.0k | for (auto op : RedundantBinaryLhs0Ops) |
4554 | 48.0k | rules_[op].push_back(RedundantBinaryLhs0(op)); |
4555 | 16.0k | for (auto op : RedundantBinaryLhs0To0Ops) |
4556 | 112k | rules_[op].push_back(RedundantBinaryLhs0To0(op)); |
4557 | 16.0k | for (auto op : ReassociateCommutiveBitwiseOps) |
4558 | 48.0k | rules_[op].push_back(ReassociateCommutiveBitwise(op)); |
4559 | 16.0k | for (auto op : ReassociateNestedGenericIntOps) |
4560 | 64.0k | rules_[op].push_back(ReassociateNestedGenericInt(op)); |
4561 | 16.0k | for (auto op : MergeBinaryOpSelectOps) |
4562 | 672k | rules_[op].push_back(MergeBinaryOpSelect(op)); |
4563 | 16.0k | rules_[spv::Op::OpSDiv].push_back(RedundantSUDiv()); |
4564 | 16.0k | rules_[spv::Op::OpUDiv].push_back(RedundantSUDiv()); |
4565 | 16.0k | rules_[spv::Op::OpSMod].push_back(RedundantSUMod()); |
4566 | 16.0k | rules_[spv::Op::OpUMod].push_back(RedundantSUMod()); |
4567 | | |
4568 | 16.0k | rules_[spv::Op::OpBitcast].push_back(BitCastScalarOrVector()); |
4569 | 16.0k | rules_[spv::Op::OpBitcast].push_back(RedundantBitcast()); |
4570 | | |
4571 | 16.0k | rules_[spv::Op::OpBitReverse].push_back(BitReverseScalarOrVector()); |
4572 | | |
4573 | 16.0k | rules_[spv::Op::OpCompositeConstruct].push_back( |
4574 | 16.0k | CompositeExtractFeedingConstruct); |
4575 | | |
4576 | 16.0k | rules_[spv::Op::OpCopyLogical].push_back( |
4577 | 16.0k | CompositeConstructFeedingCopyLogical); |
4578 | | |
4579 | 16.0k | rules_[spv::Op::OpCompositeExtract].push_back(InsertFeedingExtract()); |
4580 | 16.0k | rules_[spv::Op::OpCompositeExtract].push_back( |
4581 | 16.0k | CompositeConstructFeedingExtract); |
4582 | 16.0k | rules_[spv::Op::OpCompositeExtract].push_back(VectorShuffleFeedingExtract()); |
4583 | 16.0k | rules_[spv::Op::OpCompositeExtract].push_back(FMixFeedingExtract()); |
4584 | 16.0k | rules_[spv::Op::OpCompositeExtract].push_back(CopyLogicalFeedingExtract); |
4585 | 16.0k | rules_[spv::Op::OpCompositeExtract].push_back(LoadFeedingExtract); |
4586 | | |
4587 | 16.0k | rules_[spv::Op::OpCompositeInsert].push_back( |
4588 | 16.0k | CompositeInsertToCompositeConstruct); |
4589 | | |
4590 | 16.0k | rules_[spv::Op::OpDot].push_back(DotProductDoingExtract()); |
4591 | | |
4592 | 16.0k | rules_[spv::Op::OpEntryPoint].push_back(RemoveRedundantOperands()); |
4593 | | |
4594 | 16.0k | rules_[spv::Op::OpFAdd].push_back(RedundantFAdd()); |
4595 | 16.0k | rules_[spv::Op::OpFAdd].push_back(MergeAddNegateArithmetic()); |
4596 | 16.0k | rules_[spv::Op::OpFAdd].push_back(MergeAddAddArithmetic()); |
4597 | 16.0k | rules_[spv::Op::OpFAdd].push_back(MergeAddSubArithmetic()); |
4598 | 16.0k | rules_[spv::Op::OpFAdd].push_back(MergeGenericAddSubArithmetic()); |
4599 | 16.0k | rules_[spv::Op::OpFAdd].push_back(ReassociateNestedAddSub()); |
4600 | 16.0k | rules_[spv::Op::OpFAdd].push_back(FactorAddSubMuls()); |
4601 | | |
4602 | 16.0k | rules_[spv::Op::OpFDiv].push_back(RedundantFDiv()); |
4603 | 16.0k | rules_[spv::Op::OpFDiv].push_back(ReciprocalFDiv()); |
4604 | 16.0k | rules_[spv::Op::OpFDiv].push_back(MergeDivDivArithmetic()); |
4605 | 16.0k | rules_[spv::Op::OpFDiv].push_back(MergeDivMulArithmetic()); |
4606 | 16.0k | rules_[spv::Op::OpFDiv].push_back(MergeDivNegateArithmetic()); |
4607 | 16.0k | rules_[spv::Op::OpFDiv].push_back(MergeDivMulDoubleNegative()); |
4608 | 16.0k | rules_[spv::Op::OpFDiv].push_back(ReassociateNestedMulDivFloat()); |
4609 | | |
4610 | 16.0k | rules_[spv::Op::OpFMod].push_back(RedundantFMod()); |
4611 | | |
4612 | 16.0k | rules_[spv::Op::OpFMul].push_back(RedundantFMul()); |
4613 | 16.0k | rules_[spv::Op::OpFMul].push_back(MergeMulMulArithmetic()); |
4614 | 16.0k | rules_[spv::Op::OpFMul].push_back(MergeMulDivArithmetic()); |
4615 | 16.0k | rules_[spv::Op::OpFMul].push_back(MergeMulNegateArithmetic()); |
4616 | 16.0k | rules_[spv::Op::OpFMul].push_back(MergeDivMulDoubleNegative()); |
4617 | 16.0k | rules_[spv::Op::OpFMul].push_back(ReassociateNestedMulDivFloat()); |
4618 | | |
4619 | 16.0k | rules_[spv::Op::OpVectorTimesScalar].push_back(MergeDivMulDoubleNegative()); |
4620 | | |
4621 | 16.0k | rules_[spv::Op::OpFNegate].push_back(MergeNegateArithmetic()); |
4622 | 16.0k | rules_[spv::Op::OpFNegate].push_back(MergeNegateAddSubArithmetic()); |
4623 | 16.0k | rules_[spv::Op::OpFNegate].push_back(MergeNegateMulDivArithmetic()); |
4624 | | |
4625 | 16.0k | rules_[spv::Op::OpFSub].push_back(RedundantFSub()); |
4626 | 16.0k | rules_[spv::Op::OpFSub].push_back(MergeSubNegateArithmetic()); |
4627 | 16.0k | rules_[spv::Op::OpFSub].push_back(MergeSubAddArithmetic()); |
4628 | 16.0k | rules_[spv::Op::OpFSub].push_back(MergeSubSubArithmetic()); |
4629 | 16.0k | rules_[spv::Op::OpFSub].push_back(ReassociateNestedAddSub()); |
4630 | 16.0k | rules_[spv::Op::OpFSub].push_back(FactorAddSubMuls()); |
4631 | | |
4632 | 16.0k | rules_[spv::Op::OpIAdd].push_back(MergeAddNegateArithmetic()); |
4633 | 16.0k | rules_[spv::Op::OpIAdd].push_back(MergeAddAddArithmetic()); |
4634 | 16.0k | rules_[spv::Op::OpIAdd].push_back(MergeAddSubArithmetic()); |
4635 | 16.0k | rules_[spv::Op::OpIAdd].push_back(MergeGenericAddSubArithmetic()); |
4636 | 16.0k | rules_[spv::Op::OpIAdd].push_back(ReassociateNestedAddSub()); |
4637 | 16.0k | rules_[spv::Op::OpIAdd].push_back(FactorAddSubMuls()); |
4638 | | |
4639 | 16.0k | rules_[spv::Op::OpSDiv].push_back(MergeDivMulDoubleNegative()); |
4640 | | |
4641 | 16.0k | rules_[spv::Op::OpIMul].push_back(IntMultipleBy1()); |
4642 | 16.0k | rules_[spv::Op::OpIMul].push_back(MergeMulMulArithmetic()); |
4643 | 16.0k | rules_[spv::Op::OpIMul].push_back(MergeMulNegateArithmetic()); |
4644 | 16.0k | rules_[spv::Op::OpIMul].push_back(MergeDivMulDoubleNegative()); |
4645 | | |
4646 | 16.0k | rules_[spv::Op::OpISub].push_back(MergeSubNegateArithmetic()); |
4647 | 16.0k | rules_[spv::Op::OpISub].push_back(MergeSubAddArithmetic()); |
4648 | 16.0k | rules_[spv::Op::OpISub].push_back(MergeSubSubArithmetic()); |
4649 | 16.0k | rules_[spv::Op::OpISub].push_back(ReassociateNestedAddSub()); |
4650 | 16.0k | rules_[spv::Op::OpISub].push_back(FactorAddSubMuls()); |
4651 | | |
4652 | 16.0k | rules_[spv::Op::OpBitwiseAnd].push_back(RedundantAndOrXor()); |
4653 | 16.0k | rules_[spv::Op::OpBitwiseAnd].push_back(RedundantAndAddSub()); |
4654 | 16.0k | rules_[spv::Op::OpBitwiseAnd].push_back(RedundantAndShift()); |
4655 | | |
4656 | 16.0k | rules_[spv::Op::OpPhi].push_back(RedundantPhi()); |
4657 | | |
4658 | 16.0k | rules_[spv::Op::OpSNegate].push_back(MergeNegateArithmetic()); |
4659 | 16.0k | rules_[spv::Op::OpSNegate].push_back(MergeNegateMulDivArithmetic()); |
4660 | 16.0k | rules_[spv::Op::OpSNegate].push_back(MergeNegateAddSubArithmetic()); |
4661 | | |
4662 | 16.0k | rules_[spv::Op::OpSelect].push_back(RedundantSelect()); |
4663 | 16.0k | rules_[spv::Op::OpSelect].push_back(FoldConstantBooleanSelect()); |
4664 | | |
4665 | 16.0k | rules_[spv::Op::OpLogicalAnd].push_back(RedundantLogicalAnd()); |
4666 | | |
4667 | 16.0k | rules_[spv::Op::OpLogicalOr].push_back(RedundantLogicalOr()); |
4668 | | |
4669 | 16.0k | rules_[spv::Op::OpLogicalNot].push_back(RedundantLogicalNot()); |
4670 | 16.0k | rules_[spv::Op::OpLogicalNot].push_back(FoldLogicalNotComparison()); |
4671 | | |
4672 | 16.0k | rules_[spv::Op::OpLogicalEqual].push_back(RedundantLogicalEqual()); |
4673 | 16.0k | rules_[spv::Op::OpLogicalNotEqual].push_back(RedundantLogicalEqual()); |
4674 | | |
4675 | 16.0k | rules_[spv::Op::OpStore].push_back(StoringUndef()); |
4676 | | |
4677 | 16.0k | rules_[spv::Op::OpVectorShuffle].push_back(VectorShuffleFeedingShuffle()); |
4678 | | |
4679 | 16.0k | rules_[spv::Op::OpImageSampleImplicitLod].push_back(UpdateImageOperands()); |
4680 | 16.0k | rules_[spv::Op::OpImageSampleExplicitLod].push_back(UpdateImageOperands()); |
4681 | 16.0k | rules_[spv::Op::OpImageSampleDrefImplicitLod].push_back( |
4682 | 16.0k | UpdateImageOperands()); |
4683 | 16.0k | rules_[spv::Op::OpImageSampleDrefExplicitLod].push_back( |
4684 | 16.0k | UpdateImageOperands()); |
4685 | 16.0k | rules_[spv::Op::OpImageSampleProjImplicitLod].push_back( |
4686 | 16.0k | UpdateImageOperands()); |
4687 | 16.0k | rules_[spv::Op::OpImageSampleProjExplicitLod].push_back( |
4688 | 16.0k | UpdateImageOperands()); |
4689 | 16.0k | rules_[spv::Op::OpImageSampleProjDrefImplicitLod].push_back( |
4690 | 16.0k | UpdateImageOperands()); |
4691 | 16.0k | rules_[spv::Op::OpImageSampleProjDrefExplicitLod].push_back( |
4692 | 16.0k | UpdateImageOperands()); |
4693 | 16.0k | rules_[spv::Op::OpImageFetch].push_back(UpdateImageOperands()); |
4694 | 16.0k | rules_[spv::Op::OpImageGather].push_back(UpdateImageOperands()); |
4695 | 16.0k | rules_[spv::Op::OpImageDrefGather].push_back(UpdateImageOperands()); |
4696 | 16.0k | rules_[spv::Op::OpImageRead].push_back(UpdateImageOperands()); |
4697 | 16.0k | rules_[spv::Op::OpImageWrite].push_back(UpdateImageOperands()); |
4698 | 16.0k | rules_[spv::Op::OpImageSparseSampleImplicitLod].push_back( |
4699 | 16.0k | UpdateImageOperands()); |
4700 | 16.0k | rules_[spv::Op::OpImageSparseSampleExplicitLod].push_back( |
4701 | 16.0k | UpdateImageOperands()); |
4702 | 16.0k | rules_[spv::Op::OpImageSparseSampleDrefImplicitLod].push_back( |
4703 | 16.0k | UpdateImageOperands()); |
4704 | 16.0k | rules_[spv::Op::OpImageSparseSampleDrefExplicitLod].push_back( |
4705 | 16.0k | UpdateImageOperands()); |
4706 | 16.0k | rules_[spv::Op::OpImageSparseSampleProjImplicitLod].push_back( |
4707 | 16.0k | UpdateImageOperands()); |
4708 | 16.0k | rules_[spv::Op::OpImageSparseSampleProjExplicitLod].push_back( |
4709 | 16.0k | UpdateImageOperands()); |
4710 | 16.0k | rules_[spv::Op::OpImageSparseSampleProjDrefImplicitLod].push_back( |
4711 | 16.0k | UpdateImageOperands()); |
4712 | 16.0k | rules_[spv::Op::OpImageSparseSampleProjDrefExplicitLod].push_back( |
4713 | 16.0k | UpdateImageOperands()); |
4714 | 16.0k | rules_[spv::Op::OpImageSparseFetch].push_back(UpdateImageOperands()); |
4715 | 16.0k | rules_[spv::Op::OpImageSparseGather].push_back(UpdateImageOperands()); |
4716 | 16.0k | rules_[spv::Op::OpImageSparseDrefGather].push_back(UpdateImageOperands()); |
4717 | 16.0k | rules_[spv::Op::OpImageSparseRead].push_back(UpdateImageOperands()); |
4718 | | |
4719 | 16.0k | FeatureManager* feature_manager = context_->get_feature_mgr(); |
4720 | | // Add rules for GLSLstd450 |
4721 | 16.0k | uint32_t ext_inst_glslstd450_id = |
4722 | 16.0k | feature_manager->GetExtInstImportId_GLSLstd450(); |
4723 | 16.0k | if (ext_inst_glslstd450_id != 0) { |
4724 | 9.83k | ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FMix}].push_back( |
4725 | 9.83k | RedundantFMix()); |
4726 | 9.83k | } |
4727 | 16.0k | } |
4728 | | } // namespace opt |
4729 | | } // namespace spvtools |