• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/const_folding_rules.h"
16 
17 #include "source/opt/ir_context.h"
18 
19 namespace spvtools {
20 namespace opt {
21 namespace {
22 constexpr uint32_t kExtractCompositeIdInIdx = 0;
23 
24 // Returns the value obtained by extracting the |number_of_bits| least
25 // significant bits from |value|, and sign-extending it to 64-bits.
SignExtendValue(uint64_t value,uint32_t number_of_bits)26 uint64_t SignExtendValue(uint64_t value, uint32_t number_of_bits) {
27   if (number_of_bits == 64) return value;
28 
29   uint64_t mask_for_sign_bit = 1ull << (number_of_bits - 1);
30   uint64_t mask_for_significant_bits = (mask_for_sign_bit << 1) - 1ull;
31   if (value & mask_for_sign_bit) {
32     // Set upper bits to 1
33     value |= ~mask_for_significant_bits;
34   } else {
35     // Clear the upper bits
36     value &= mask_for_significant_bits;
37   }
38   return value;
39 }
40 
41 // Returns the value obtained by extracting the |number_of_bits| least
42 // significant bits from |value|, and zero-extending it to 64-bits.
ZeroExtendValue(uint64_t value,uint32_t number_of_bits)43 uint64_t ZeroExtendValue(uint64_t value, uint32_t number_of_bits) {
44   if (number_of_bits == 64) return value;
45 
46   uint64_t mask_for_first_bit_to_clear = 1ull << (number_of_bits);
47   uint64_t mask_for_bits_to_keep = mask_for_first_bit_to_clear - 1;
48   value &= mask_for_bits_to_keep;
49   return value;
50 }
51 
52 // Returns a constant whose value is `value` and type is `type`. This constant
53 // will be generated by `const_mgr`. The type must be a scalar integer type.
GenerateIntegerConstant(const analysis::Integer * integer_type,uint64_t result,analysis::ConstantManager * const_mgr)54 const analysis::Constant* GenerateIntegerConstant(
55     const analysis::Integer* integer_type, uint64_t result,
56     analysis::ConstantManager* const_mgr) {
57   assert(integer_type != nullptr);
58 
59   std::vector<uint32_t> words;
60   if (integer_type->width() == 64) {
61     // In the 64-bit case, two words are needed to represent the value.
62     words = {static_cast<uint32_t>(result),
63              static_cast<uint32_t>(result >> 32)};
64   } else {
65     // In all other cases, only a single word is needed.
66     assert(integer_type->width() <= 32);
67     if (integer_type->IsSigned()) {
68       result = SignExtendValue(result, integer_type->width());
69     } else {
70       result = ZeroExtendValue(result, integer_type->width());
71     }
72     words = {static_cast<uint32_t>(result)};
73   }
74   return const_mgr->GetConstant(integer_type, words);
75 }
76 
77 // Returns a constants with the value NaN of the given type.  Only works for
78 // 32-bit and 64-bit float point types.  Returns |nullptr| if an error occurs.
GetNan(const analysis::Type * type,analysis::ConstantManager * const_mgr)79 const analysis::Constant* GetNan(const analysis::Type* type,
80                                  analysis::ConstantManager* const_mgr) {
81   const analysis::Float* float_type = type->AsFloat();
82   if (float_type == nullptr) {
83     return nullptr;
84   }
85 
86   switch (float_type->width()) {
87     case 32:
88       return const_mgr->GetFloatConst(std::numeric_limits<float>::quiet_NaN());
89     case 64:
90       return const_mgr->GetDoubleConst(
91           std::numeric_limits<double>::quiet_NaN());
92     default:
93       return nullptr;
94   }
95 }
96 
97 // Returns a constants with the value INF of the given type.  Only works for
98 // 32-bit and 64-bit float point types.  Returns |nullptr| if an error occurs.
GetInf(const analysis::Type * type,analysis::ConstantManager * const_mgr)99 const analysis::Constant* GetInf(const analysis::Type* type,
100                                  analysis::ConstantManager* const_mgr) {
101   const analysis::Float* float_type = type->AsFloat();
102   if (float_type == nullptr) {
103     return nullptr;
104   }
105 
106   switch (float_type->width()) {
107     case 32:
108       return const_mgr->GetFloatConst(std::numeric_limits<float>::infinity());
109     case 64:
110       return const_mgr->GetDoubleConst(std::numeric_limits<double>::infinity());
111     default:
112       return nullptr;
113   }
114 }
115 
116 // Returns true if |type| is Float or a vector of Float.
HasFloatingPoint(const analysis::Type * type)117 bool HasFloatingPoint(const analysis::Type* type) {
118   if (type->AsFloat()) {
119     return true;
120   } else if (const analysis::Vector* vec_type = type->AsVector()) {
121     return vec_type->element_type()->AsFloat() != nullptr;
122   }
123 
124   return false;
125 }
126 
127 // Returns a constants with the value |-val| of the given type.  Only works for
128 // 32-bit and 64-bit float point types.  Returns |nullptr| if an error occurs.
NegateFPConst(const analysis::Type * result_type,const analysis::Constant * val,analysis::ConstantManager * const_mgr)129 const analysis::Constant* NegateFPConst(const analysis::Type* result_type,
130                                         const analysis::Constant* val,
131                                         analysis::ConstantManager* const_mgr) {
132   const analysis::Float* float_type = result_type->AsFloat();
133   assert(float_type != nullptr);
134   if (float_type->width() == 32) {
135     float fa = val->GetFloat();
136     return const_mgr->GetFloatConst(-fa);
137   } else if (float_type->width() == 64) {
138     double da = val->GetDouble();
139     return const_mgr->GetDoubleConst(-da);
140   }
141   return nullptr;
142 }
143 
144 // Returns a constants with the value |-val| of the given type.
NegateIntConst(const analysis::Type * result_type,const analysis::Constant * val,analysis::ConstantManager * const_mgr)145 const analysis::Constant* NegateIntConst(const analysis::Type* result_type,
146                                          const analysis::Constant* val,
147                                          analysis::ConstantManager* const_mgr) {
148   const analysis::Integer* int_type = result_type->AsInteger();
149   assert(int_type != nullptr);
150 
151   if (val->AsNullConstant()) {
152     return val;
153   }
154 
155   uint64_t new_value = static_cast<uint64_t>(-val->GetSignExtendedValue());
156   return const_mgr->GetIntConst(new_value, int_type->width(),
157                                 int_type->IsSigned());
158 }
159 
160 // Folds an OpcompositeExtract where input is a composite constant.
FoldExtractWithConstants()161 ConstantFoldingRule FoldExtractWithConstants() {
162   return [](IRContext* context, Instruction* inst,
163             const std::vector<const analysis::Constant*>& constants)
164              -> const analysis::Constant* {
165     const analysis::Constant* c = constants[kExtractCompositeIdInIdx];
166     if (c == nullptr) {
167       return nullptr;
168     }
169 
170     for (uint32_t i = 1; i < inst->NumInOperands(); ++i) {
171       uint32_t element_index = inst->GetSingleWordInOperand(i);
172       if (c->AsNullConstant()) {
173         // Return Null for the return type.
174         analysis::ConstantManager* const_mgr = context->get_constant_mgr();
175         analysis::TypeManager* type_mgr = context->get_type_mgr();
176         return const_mgr->GetConstant(type_mgr->GetType(inst->type_id()), {});
177       }
178 
179       auto cc = c->AsCompositeConstant();
180       assert(cc != nullptr);
181       auto components = cc->GetComponents();
182       // Protect against invalid IR.  Refuse to fold if the index is out
183       // of bounds.
184       if (element_index >= components.size()) return nullptr;
185       c = components[element_index];
186     }
187     return c;
188   };
189 }
190 
191 // Folds an OpcompositeInsert where input is a composite constant.
FoldInsertWithConstants()192 ConstantFoldingRule FoldInsertWithConstants() {
193   return [](IRContext* context, Instruction* inst,
194             const std::vector<const analysis::Constant*>& constants)
195              -> const analysis::Constant* {
196     analysis::ConstantManager* const_mgr = context->get_constant_mgr();
197     const analysis::Constant* object = constants[0];
198     const analysis::Constant* composite = constants[1];
199     if (object == nullptr || composite == nullptr) {
200       return nullptr;
201     }
202 
203     // If there is more than 1 index, then each additional constant used by the
204     // index will need to be recreated to use the inserted object.
205     std::vector<const analysis::Constant*> chain;
206     std::vector<const analysis::Constant*> components;
207     const analysis::Type* type = nullptr;
208     const uint32_t final_index = (inst->NumInOperands() - 1);
209 
210     // Work down hierarchy of all indexes
211     for (uint32_t i = 2; i < inst->NumInOperands(); ++i) {
212       type = composite->type();
213 
214       if (composite->AsNullConstant()) {
215         // Make new composite so it can be inserted in the index with the
216         // non-null value
217         if (const auto new_composite =
218                 const_mgr->GetNullCompositeConstant(type)) {
219           // Keep track of any indexes along the way to last index
220           if (i != final_index) {
221             chain.push_back(new_composite);
222           }
223           components = new_composite->AsCompositeConstant()->GetComponents();
224         } else {
225           // Unsupported input type (such as structs)
226           return nullptr;
227         }
228       } else {
229         // Keep track of any indexes along the way to last index
230         if (i != final_index) {
231           chain.push_back(composite);
232         }
233         components = composite->AsCompositeConstant()->GetComponents();
234       }
235       const uint32_t index = inst->GetSingleWordInOperand(i);
236       composite = components[index];
237     }
238 
239     // Final index in hierarchy is inserted with new object.
240     const uint32_t final_operand = inst->GetSingleWordInOperand(final_index);
241     std::vector<uint32_t> ids;
242     for (size_t i = 0; i < components.size(); i++) {
243       const analysis::Constant* constant =
244           (i == final_operand) ? object : components[i];
245       Instruction* member_inst = const_mgr->GetDefiningInstruction(constant);
246       ids.push_back(member_inst->result_id());
247     }
248     const analysis::Constant* new_constant = const_mgr->GetConstant(type, ids);
249 
250     // Work backwards up the chain and replace each index with new constant.
251     for (size_t i = chain.size(); i > 0; i--) {
252       // Need to insert any previous instruction into the module first.
253       // Can't just insert in types_values_begin() because it will move above
254       // where the types are declared.
255       // Can't compare with location of inst because not all new added
256       // instructions are added to types_values_
257       auto iter = context->types_values_end();
258       Module::inst_iterator* pos = &iter;
259       const_mgr->BuildInstructionAndAddToModule(new_constant, pos);
260 
261       composite = chain[i - 1];
262       components = composite->AsCompositeConstant()->GetComponents();
263       type = composite->type();
264       ids.clear();
265       for (size_t k = 0; k < components.size(); k++) {
266         const uint32_t index =
267             inst->GetSingleWordInOperand(1 + static_cast<uint32_t>(i));
268         const analysis::Constant* constant =
269             (k == index) ? new_constant : components[k];
270         const uint32_t constant_id =
271             const_mgr->FindDeclaredConstant(constant, 0);
272         ids.push_back(constant_id);
273       }
274       new_constant = const_mgr->GetConstant(type, ids);
275     }
276 
277     // If multiple constants were created, only need to return the top index.
278     return new_constant;
279   };
280 }
281 
FoldVectorShuffleWithConstants()282 ConstantFoldingRule FoldVectorShuffleWithConstants() {
283   return [](IRContext* context, Instruction* inst,
284             const std::vector<const analysis::Constant*>& constants)
285              -> const analysis::Constant* {
286     assert(inst->opcode() == spv::Op::OpVectorShuffle);
287     const analysis::Constant* c1 = constants[0];
288     const analysis::Constant* c2 = constants[1];
289     if (c1 == nullptr || c2 == nullptr) {
290       return nullptr;
291     }
292 
293     analysis::ConstantManager* const_mgr = context->get_constant_mgr();
294     const analysis::Type* element_type = c1->type()->AsVector()->element_type();
295 
296     std::vector<const analysis::Constant*> c1_components;
297     if (const analysis::VectorConstant* vec_const = c1->AsVectorConstant()) {
298       c1_components = vec_const->GetComponents();
299     } else {
300       assert(c1->AsNullConstant());
301       const analysis::Constant* element =
302           const_mgr->GetConstant(element_type, {});
303       c1_components.resize(c1->type()->AsVector()->element_count(), element);
304     }
305     std::vector<const analysis::Constant*> c2_components;
306     if (const analysis::VectorConstant* vec_const = c2->AsVectorConstant()) {
307       c2_components = vec_const->GetComponents();
308     } else {
309       assert(c2->AsNullConstant());
310       const analysis::Constant* element =
311           const_mgr->GetConstant(element_type, {});
312       c2_components.resize(c2->type()->AsVector()->element_count(), element);
313     }
314 
315     std::vector<uint32_t> ids;
316     const uint32_t undef_literal_value = 0xffffffff;
317     for (uint32_t i = 2; i < inst->NumInOperands(); ++i) {
318       uint32_t index = inst->GetSingleWordInOperand(i);
319       if (index == undef_literal_value) {
320         // Don't fold shuffle with undef literal value.
321         return nullptr;
322       } else if (index < c1_components.size()) {
323         Instruction* member_inst =
324             const_mgr->GetDefiningInstruction(c1_components[index]);
325         ids.push_back(member_inst->result_id());
326       } else {
327         Instruction* member_inst = const_mgr->GetDefiningInstruction(
328             c2_components[index - c1_components.size()]);
329         ids.push_back(member_inst->result_id());
330       }
331     }
332 
333     analysis::TypeManager* type_mgr = context->get_type_mgr();
334     return const_mgr->GetConstant(type_mgr->GetType(inst->type_id()), ids);
335   };
336 }
337 
FoldVectorTimesScalar()338 ConstantFoldingRule FoldVectorTimesScalar() {
339   return [](IRContext* context, Instruction* inst,
340             const std::vector<const analysis::Constant*>& constants)
341              -> const analysis::Constant* {
342     assert(inst->opcode() == spv::Op::OpVectorTimesScalar);
343     analysis::ConstantManager* const_mgr = context->get_constant_mgr();
344     analysis::TypeManager* type_mgr = context->get_type_mgr();
345 
346     if (!inst->IsFloatingPointFoldingAllowed()) {
347       if (HasFloatingPoint(type_mgr->GetType(inst->type_id()))) {
348         return nullptr;
349       }
350     }
351 
352     const analysis::Constant* c1 = constants[0];
353     const analysis::Constant* c2 = constants[1];
354 
355     if (c1 && c1->IsZero()) {
356       return c1;
357     }
358 
359     if (c2 && c2->IsZero()) {
360       // Get or create the NullConstant for this type.
361       std::vector<uint32_t> ids;
362       return const_mgr->GetConstant(type_mgr->GetType(inst->type_id()), ids);
363     }
364 
365     if (c1 == nullptr || c2 == nullptr) {
366       return nullptr;
367     }
368 
369     // Check result type.
370     const analysis::Type* result_type = type_mgr->GetType(inst->type_id());
371     const analysis::Vector* vector_type = result_type->AsVector();
372     assert(vector_type != nullptr);
373     const analysis::Type* element_type = vector_type->element_type();
374     assert(element_type != nullptr);
375     const analysis::Float* float_type = element_type->AsFloat();
376     assert(float_type != nullptr);
377 
378     // Check types of c1 and c2.
379     assert(c1->type()->AsVector() == vector_type);
380     assert(c1->type()->AsVector()->element_type() == element_type &&
381            c2->type() == element_type);
382 
383     // Get a float vector that is the result of vector-times-scalar.
384     std::vector<const analysis::Constant*> c1_components =
385         c1->GetVectorComponents(const_mgr);
386     std::vector<uint32_t> ids;
387     if (float_type->width() == 32) {
388       float scalar = c2->GetFloat();
389       for (uint32_t i = 0; i < c1_components.size(); ++i) {
390         utils::FloatProxy<float> result(c1_components[i]->GetFloat() * scalar);
391         std::vector<uint32_t> words = result.GetWords();
392         const analysis::Constant* new_elem =
393             const_mgr->GetConstant(float_type, words);
394         ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id());
395       }
396       return const_mgr->GetConstant(vector_type, ids);
397     } else if (float_type->width() == 64) {
398       double scalar = c2->GetDouble();
399       for (uint32_t i = 0; i < c1_components.size(); ++i) {
400         utils::FloatProxy<double> result(c1_components[i]->GetDouble() *
401                                          scalar);
402         std::vector<uint32_t> words = result.GetWords();
403         const analysis::Constant* new_elem =
404             const_mgr->GetConstant(float_type, words);
405         ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id());
406       }
407       return const_mgr->GetConstant(vector_type, ids);
408     }
409     return nullptr;
410   };
411 }
412 
413 // Returns to the constant that results from tranposing |matrix|. The result
414 // will have type |result_type|, and |matrix| must exist in |context|. The
415 // result constant will also exist in |context|.
TransposeMatrix(const analysis::Constant * matrix,analysis::Matrix * result_type,IRContext * context)416 const analysis::Constant* TransposeMatrix(const analysis::Constant* matrix,
417                                           analysis::Matrix* result_type,
418                                           IRContext* context) {
419   analysis::ConstantManager* const_mgr = context->get_constant_mgr();
420   if (matrix->AsNullConstant() != nullptr) {
421     return const_mgr->GetNullCompositeConstant(result_type);
422   }
423 
424   const auto& columns = matrix->AsMatrixConstant()->GetComponents();
425   uint32_t number_of_rows = columns[0]->type()->AsVector()->element_count();
426 
427   // Collect the ids of the elements in their new positions.
428   std::vector<std::vector<uint32_t>> result_elements(number_of_rows);
429   for (const analysis::Constant* column : columns) {
430     if (column->AsNullConstant()) {
431       column = const_mgr->GetNullCompositeConstant(column->type());
432     }
433     const auto& column_components = column->AsVectorConstant()->GetComponents();
434 
435     for (uint32_t row = 0; row < number_of_rows; ++row) {
436       result_elements[row].push_back(
437           const_mgr->GetDefiningInstruction(column_components[row])
438               ->result_id());
439     }
440   }
441 
442   // Create the constant for each row in the result, and collect the ids.
443   std::vector<uint32_t> result_columns(number_of_rows);
444   for (uint32_t col = 0; col < number_of_rows; ++col) {
445     auto* element = const_mgr->GetConstant(result_type->element_type(),
446                                            result_elements[col]);
447     result_columns[col] =
448         const_mgr->GetDefiningInstruction(element)->result_id();
449   }
450 
451   // Create the matrix constant from the row ids, and return it.
452   return const_mgr->GetConstant(result_type, result_columns);
453 }
454 
FoldTranspose(IRContext * context,Instruction * inst,const std::vector<const analysis::Constant * > & constants)455 const analysis::Constant* FoldTranspose(
456     IRContext* context, Instruction* inst,
457     const std::vector<const analysis::Constant*>& constants) {
458   assert(inst->opcode() == spv::Op::OpTranspose);
459 
460   analysis::TypeManager* type_mgr = context->get_type_mgr();
461   if (!inst->IsFloatingPointFoldingAllowed()) {
462     if (HasFloatingPoint(type_mgr->GetType(inst->type_id()))) {
463       return nullptr;
464     }
465   }
466 
467   const analysis::Constant* matrix = constants[0];
468   if (matrix == nullptr) {
469     return nullptr;
470   }
471 
472   auto* result_type = type_mgr->GetType(inst->type_id());
473   return TransposeMatrix(matrix, result_type->AsMatrix(), context);
474 }
475 
FoldVectorTimesMatrix()476 ConstantFoldingRule FoldVectorTimesMatrix() {
477   return [](IRContext* context, Instruction* inst,
478             const std::vector<const analysis::Constant*>& constants)
479              -> const analysis::Constant* {
480     assert(inst->opcode() == spv::Op::OpVectorTimesMatrix);
481     analysis::ConstantManager* const_mgr = context->get_constant_mgr();
482     analysis::TypeManager* type_mgr = context->get_type_mgr();
483 
484     if (!inst->IsFloatingPointFoldingAllowed()) {
485       if (HasFloatingPoint(type_mgr->GetType(inst->type_id()))) {
486         return nullptr;
487       }
488     }
489 
490     const analysis::Constant* c1 = constants[0];
491     const analysis::Constant* c2 = constants[1];
492 
493     if (c1 == nullptr || c2 == nullptr) {
494       return nullptr;
495     }
496 
497     // Check result type.
498     const analysis::Type* result_type = type_mgr->GetType(inst->type_id());
499     const analysis::Vector* vector_type = result_type->AsVector();
500     assert(vector_type != nullptr);
501     const analysis::Type* element_type = vector_type->element_type();
502     assert(element_type != nullptr);
503     const analysis::Float* float_type = element_type->AsFloat();
504     assert(float_type != nullptr);
505 
506     // Check types of c1 and c2.
507     assert(c1->type()->AsVector() == vector_type);
508     assert(c1->type()->AsVector()->element_type() == element_type &&
509            c2->type()->AsMatrix()->element_type() == vector_type);
510 
511     uint32_t resultVectorSize = result_type->AsVector()->element_count();
512     std::vector<uint32_t> ids;
513 
514     if ((c1 && c1->IsZero()) || (c2 && c2->IsZero())) {
515       std::vector<uint32_t> words(float_type->width() / 32, 0);
516       for (uint32_t i = 0; i < resultVectorSize; ++i) {
517         const analysis::Constant* new_elem =
518             const_mgr->GetConstant(float_type, words);
519         ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id());
520       }
521       return const_mgr->GetConstant(vector_type, ids);
522     }
523 
524     // Get a float vector that is the result of vector-times-matrix.
525     std::vector<const analysis::Constant*> c1_components =
526         c1->GetVectorComponents(const_mgr);
527     std::vector<const analysis::Constant*> c2_components =
528         c2->AsMatrixConstant()->GetComponents();
529 
530     if (float_type->width() == 32) {
531       for (uint32_t i = 0; i < resultVectorSize; ++i) {
532         float result_scalar = 0.0f;
533         if (!c2_components[i]->AsNullConstant()) {
534           const analysis::VectorConstant* c2_vec =
535               c2_components[i]->AsVectorConstant();
536           for (uint32_t j = 0; j < c2_vec->GetComponents().size(); ++j) {
537             float c1_scalar = c1_components[j]->GetFloat();
538             float c2_scalar = c2_vec->GetComponents()[j]->GetFloat();
539             result_scalar += c1_scalar * c2_scalar;
540           }
541         }
542         utils::FloatProxy<float> result(result_scalar);
543         std::vector<uint32_t> words = result.GetWords();
544         const analysis::Constant* new_elem =
545             const_mgr->GetConstant(float_type, words);
546         ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id());
547       }
548       return const_mgr->GetConstant(vector_type, ids);
549     } else if (float_type->width() == 64) {
550       for (uint32_t i = 0; i < c2_components.size(); ++i) {
551         double result_scalar = 0.0;
552         if (!c2_components[i]->AsNullConstant()) {
553           const analysis::VectorConstant* c2_vec =
554               c2_components[i]->AsVectorConstant();
555           for (uint32_t j = 0; j < c2_vec->GetComponents().size(); ++j) {
556             double c1_scalar = c1_components[j]->GetDouble();
557             double c2_scalar = c2_vec->GetComponents()[j]->GetDouble();
558             result_scalar += c1_scalar * c2_scalar;
559           }
560         }
561         utils::FloatProxy<double> result(result_scalar);
562         std::vector<uint32_t> words = result.GetWords();
563         const analysis::Constant* new_elem =
564             const_mgr->GetConstant(float_type, words);
565         ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id());
566       }
567       return const_mgr->GetConstant(vector_type, ids);
568     }
569     return nullptr;
570   };
571 }
572 
FoldMatrixTimesVector()573 ConstantFoldingRule FoldMatrixTimesVector() {
574   return [](IRContext* context, Instruction* inst,
575             const std::vector<const analysis::Constant*>& constants)
576              -> const analysis::Constant* {
577     assert(inst->opcode() == spv::Op::OpMatrixTimesVector);
578     analysis::ConstantManager* const_mgr = context->get_constant_mgr();
579     analysis::TypeManager* type_mgr = context->get_type_mgr();
580 
581     if (!inst->IsFloatingPointFoldingAllowed()) {
582       if (HasFloatingPoint(type_mgr->GetType(inst->type_id()))) {
583         return nullptr;
584       }
585     }
586 
587     const analysis::Constant* c1 = constants[0];
588     const analysis::Constant* c2 = constants[1];
589 
590     if (c1 == nullptr || c2 == nullptr) {
591       return nullptr;
592     }
593 
594     // Check result type.
595     const analysis::Type* result_type = type_mgr->GetType(inst->type_id());
596     const analysis::Vector* vector_type = result_type->AsVector();
597     assert(vector_type != nullptr);
598     const analysis::Type* element_type = vector_type->element_type();
599     assert(element_type != nullptr);
600     const analysis::Float* float_type = element_type->AsFloat();
601     assert(float_type != nullptr);
602 
603     // Check types of c1 and c2.
604     assert(c1->type()->AsMatrix()->element_type() == vector_type);
605     assert(c2->type()->AsVector()->element_type() == element_type);
606 
607     uint32_t resultVectorSize = result_type->AsVector()->element_count();
608     std::vector<uint32_t> ids;
609 
610     if ((c1 && c1->IsZero()) || (c2 && c2->IsZero())) {
611       std::vector<uint32_t> words(float_type->width() / 32, 0);
612       for (uint32_t i = 0; i < resultVectorSize; ++i) {
613         const analysis::Constant* new_elem =
614             const_mgr->GetConstant(float_type, words);
615         ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id());
616       }
617       return const_mgr->GetConstant(vector_type, ids);
618     }
619 
620     // Get a float vector that is the result of matrix-times-vector.
621     std::vector<const analysis::Constant*> c1_components =
622         c1->AsMatrixConstant()->GetComponents();
623     std::vector<const analysis::Constant*> c2_components =
624         c2->GetVectorComponents(const_mgr);
625 
626     if (float_type->width() == 32) {
627       for (uint32_t i = 0; i < resultVectorSize; ++i) {
628         float result_scalar = 0.0f;
629         for (uint32_t j = 0; j < c1_components.size(); ++j) {
630           if (!c1_components[j]->AsNullConstant()) {
631             float c1_scalar = c1_components[j]
632                                   ->AsVectorConstant()
633                                   ->GetComponents()[i]
634                                   ->GetFloat();
635             float c2_scalar = c2_components[j]->GetFloat();
636             result_scalar += c1_scalar * c2_scalar;
637           }
638         }
639         utils::FloatProxy<float> result(result_scalar);
640         std::vector<uint32_t> words = result.GetWords();
641         const analysis::Constant* new_elem =
642             const_mgr->GetConstant(float_type, words);
643         ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id());
644       }
645       return const_mgr->GetConstant(vector_type, ids);
646     } else if (float_type->width() == 64) {
647       for (uint32_t i = 0; i < resultVectorSize; ++i) {
648         double result_scalar = 0.0;
649         for (uint32_t j = 0; j < c1_components.size(); ++j) {
650           if (!c1_components[j]->AsNullConstant()) {
651             double c1_scalar = c1_components[j]
652                                    ->AsVectorConstant()
653                                    ->GetComponents()[i]
654                                    ->GetDouble();
655             double c2_scalar = c2_components[j]->GetDouble();
656             result_scalar += c1_scalar * c2_scalar;
657           }
658         }
659         utils::FloatProxy<double> result(result_scalar);
660         std::vector<uint32_t> words = result.GetWords();
661         const analysis::Constant* new_elem =
662             const_mgr->GetConstant(float_type, words);
663         ids.push_back(const_mgr->GetDefiningInstruction(new_elem)->result_id());
664       }
665       return const_mgr->GetConstant(vector_type, ids);
666     }
667     return nullptr;
668   };
669 }
670 
FoldCompositeWithConstants()671 ConstantFoldingRule FoldCompositeWithConstants() {
672   // Folds an OpCompositeConstruct where all of the inputs are constants to a
673   // constant.  A new constant is created if necessary.
674   return [](IRContext* context, Instruction* inst,
675             const std::vector<const analysis::Constant*>& constants)
676              -> const analysis::Constant* {
677     analysis::ConstantManager* const_mgr = context->get_constant_mgr();
678     analysis::TypeManager* type_mgr = context->get_type_mgr();
679     const analysis::Type* new_type = type_mgr->GetType(inst->type_id());
680     Instruction* type_inst =
681         context->get_def_use_mgr()->GetDef(inst->type_id());
682 
683     std::vector<uint32_t> ids;
684     for (uint32_t i = 0; i < constants.size(); ++i) {
685       const analysis::Constant* element_const = constants[i];
686       if (element_const == nullptr) {
687         return nullptr;
688       }
689 
690       uint32_t component_type_id = 0;
691       if (type_inst->opcode() == spv::Op::OpTypeStruct) {
692         component_type_id = type_inst->GetSingleWordInOperand(i);
693       } else if (type_inst->opcode() == spv::Op::OpTypeArray) {
694         component_type_id = type_inst->GetSingleWordInOperand(0);
695       }
696 
697       uint32_t element_id =
698           const_mgr->FindDeclaredConstant(element_const, component_type_id);
699       if (element_id == 0) {
700         return nullptr;
701       }
702       ids.push_back(element_id);
703     }
704     return const_mgr->GetConstant(new_type, ids);
705   };
706 }
707 
708 // The interface for a function that returns the result of applying a scalar
709 // floating-point binary operation on |a| and |b|.  The type of the return value
710 // will be |type|.  The input constants must also be of type |type|.
711 using UnaryScalarFoldingRule = std::function<const analysis::Constant*(
712     const analysis::Type* result_type, const analysis::Constant* a,
713     analysis::ConstantManager*)>;
714 
715 // The interface for a function that returns the result of applying a scalar
716 // floating-point binary operation on |a| and |b|.  The type of the return value
717 // will be |type|.  The input constants must also be of type |type|.
718 using BinaryScalarFoldingRule = std::function<const analysis::Constant*(
719     const analysis::Type* result_type, const analysis::Constant* a,
720     const analysis::Constant* b, analysis::ConstantManager*)>;
721 
722 // Returns a |ConstantFoldingRule| that folds unary scalar ops
723 // using |scalar_rule| and unary vectors ops by applying
724 // |scalar_rule| to the elements of the vector.  The |ConstantFoldingRule|
725 // that is returned assumes that |constants| contains 1 entry.  If they are
726 // not |nullptr|, then their type is either |Float| or |Integer| or a |Vector|
727 // whose element type is |Float| or |Integer|.
FoldUnaryOp(UnaryScalarFoldingRule scalar_rule)728 ConstantFoldingRule FoldUnaryOp(UnaryScalarFoldingRule scalar_rule) {
729   return [scalar_rule](IRContext* context, Instruction* inst,
730                        const std::vector<const analysis::Constant*>& constants)
731              -> const analysis::Constant* {
732     analysis::ConstantManager* const_mgr = context->get_constant_mgr();
733     analysis::TypeManager* type_mgr = context->get_type_mgr();
734     const analysis::Type* result_type = type_mgr->GetType(inst->type_id());
735     const analysis::Vector* vector_type = result_type->AsVector();
736 
737     const analysis::Constant* arg =
738         (inst->opcode() == spv::Op::OpExtInst) ? constants[1] : constants[0];
739 
740     if (arg == nullptr) {
741       return nullptr;
742     }
743 
744     if (vector_type != nullptr) {
745       std::vector<const analysis::Constant*> a_components;
746       std::vector<const analysis::Constant*> results_components;
747 
748       a_components = arg->GetVectorComponents(const_mgr);
749 
750       // Fold each component of the vector.
751       for (uint32_t i = 0; i < a_components.size(); ++i) {
752         results_components.push_back(scalar_rule(vector_type->element_type(),
753                                                  a_components[i], const_mgr));
754         if (results_components[i] == nullptr) {
755           return nullptr;
756         }
757       }
758 
759       // Build the constant object and return it.
760       std::vector<uint32_t> ids;
761       for (const analysis::Constant* member : results_components) {
762         ids.push_back(const_mgr->GetDefiningInstruction(member)->result_id());
763       }
764       return const_mgr->GetConstant(vector_type, ids);
765     } else {
766       return scalar_rule(result_type, arg, const_mgr);
767     }
768   };
769 }
770 
771 // Returns a |ConstantFoldingRule| that folds binary scalar ops
772 // using |scalar_rule| and binary vectors ops by applying
773 // |scalar_rule| to the elements of the vector. The folding rule assumes that op
774 // has two inputs. For regular instruction, those are in operands 0 and 1. For
775 // extended instruction, they are in operands 1 and 2. If an element in
776 // |constants| is not nullprt, then the constant's type is |Float|, |Integer|,
777 // or |Vector| whose element type is |Float| or |Integer|.
FoldBinaryOp(BinaryScalarFoldingRule scalar_rule)778 ConstantFoldingRule FoldBinaryOp(BinaryScalarFoldingRule scalar_rule) {
779   return [scalar_rule](IRContext* context, Instruction* inst,
780                        const std::vector<const analysis::Constant*>& constants)
781              -> const analysis::Constant* {
782     assert(constants.size() == inst->NumInOperands());
783     assert(constants.size() == (inst->opcode() == spv::Op::OpExtInst ? 3 : 2));
784     analysis::ConstantManager* const_mgr = context->get_constant_mgr();
785     analysis::TypeManager* type_mgr = context->get_type_mgr();
786     const analysis::Type* result_type = type_mgr->GetType(inst->type_id());
787     const analysis::Vector* vector_type = result_type->AsVector();
788 
789     const analysis::Constant* arg1 =
790         (inst->opcode() == spv::Op::OpExtInst) ? constants[1] : constants[0];
791     const analysis::Constant* arg2 =
792         (inst->opcode() == spv::Op::OpExtInst) ? constants[2] : constants[1];
793 
794     if (arg1 == nullptr || arg2 == nullptr) {
795       return nullptr;
796     }
797 
798     if (vector_type == nullptr) {
799       return scalar_rule(result_type, arg1, arg2, const_mgr);
800     }
801 
802     std::vector<const analysis::Constant*> a_components;
803     std::vector<const analysis::Constant*> b_components;
804     std::vector<const analysis::Constant*> results_components;
805 
806     a_components = arg1->GetVectorComponents(const_mgr);
807     b_components = arg2->GetVectorComponents(const_mgr);
808     assert(a_components.size() == b_components.size());
809 
810     // Fold each component of the vector.
811     for (uint32_t i = 0; i < a_components.size(); ++i) {
812       results_components.push_back(scalar_rule(vector_type->element_type(),
813                                                a_components[i], b_components[i],
814                                                const_mgr));
815       if (results_components[i] == nullptr) {
816         return nullptr;
817       }
818     }
819 
820     // Build the constant object and return it.
821     std::vector<uint32_t> ids;
822     for (const analysis::Constant* member : results_components) {
823       ids.push_back(const_mgr->GetDefiningInstruction(member)->result_id());
824     }
825     return const_mgr->GetConstant(vector_type, ids);
826   };
827 }
828 
829 // Returns a |ConstantFoldingRule| that folds unary floating point scalar ops
830 // using |scalar_rule| and unary float point vectors ops by applying
831 // |scalar_rule| to the elements of the vector.  The |ConstantFoldingRule|
832 // that is returned assumes that |constants| contains 1 entry.  If they are
833 // not |nullptr|, then their type is either |Float| or |Integer| or a |Vector|
834 // whose element type is |Float| or |Integer|.
FoldFPUnaryOp(UnaryScalarFoldingRule scalar_rule)835 ConstantFoldingRule FoldFPUnaryOp(UnaryScalarFoldingRule scalar_rule) {
836   auto folding_rule = FoldUnaryOp(scalar_rule);
837   return [folding_rule](IRContext* context, Instruction* inst,
838                         const std::vector<const analysis::Constant*>& constants)
839              -> const analysis::Constant* {
840     if (!inst->IsFloatingPointFoldingAllowed()) {
841       return nullptr;
842     }
843 
844     return folding_rule(context, inst, constants);
845   };
846 }
847 
848 // Returns the result of folding the constants in |constants| according the
849 // |scalar_rule|.  If |result_type| is a vector, then |scalar_rule| is applied
850 // per component.
FoldFPBinaryOp(BinaryScalarFoldingRule scalar_rule,uint32_t result_type_id,const std::vector<const analysis::Constant * > & constants,IRContext * context)851 const analysis::Constant* FoldFPBinaryOp(
852     BinaryScalarFoldingRule scalar_rule, uint32_t result_type_id,
853     const std::vector<const analysis::Constant*>& constants,
854     IRContext* context) {
855   analysis::ConstantManager* const_mgr = context->get_constant_mgr();
856   analysis::TypeManager* type_mgr = context->get_type_mgr();
857   const analysis::Type* result_type = type_mgr->GetType(result_type_id);
858   const analysis::Vector* vector_type = result_type->AsVector();
859 
860   if (constants[0] == nullptr || constants[1] == nullptr) {
861     return nullptr;
862   }
863 
864   if (vector_type != nullptr) {
865     std::vector<const analysis::Constant*> a_components;
866     std::vector<const analysis::Constant*> b_components;
867     std::vector<const analysis::Constant*> results_components;
868 
869     a_components = constants[0]->GetVectorComponents(const_mgr);
870     b_components = constants[1]->GetVectorComponents(const_mgr);
871 
872     // Fold each component of the vector.
873     for (uint32_t i = 0; i < a_components.size(); ++i) {
874       results_components.push_back(scalar_rule(vector_type->element_type(),
875                                                a_components[i], b_components[i],
876                                                const_mgr));
877       if (results_components[i] == nullptr) {
878         return nullptr;
879       }
880     }
881 
882     // Build the constant object and return it.
883     std::vector<uint32_t> ids;
884     for (const analysis::Constant* member : results_components) {
885       ids.push_back(const_mgr->GetDefiningInstruction(member)->result_id());
886     }
887     return const_mgr->GetConstant(vector_type, ids);
888   } else {
889     return scalar_rule(result_type, constants[0], constants[1], const_mgr);
890   }
891 }
892 
893 // Returns a |ConstantFoldingRule| that folds floating point scalars using
894 // |scalar_rule| and vectors of floating point by applying |scalar_rule| to the
895 // elements of the vector.  The |ConstantFoldingRule| that is returned assumes
896 // that |constants| contains 2 entries.  If they are not |nullptr|, then their
897 // type is either |Float| or a |Vector| whose element type is |Float|.
FoldFPBinaryOp(BinaryScalarFoldingRule scalar_rule)898 ConstantFoldingRule FoldFPBinaryOp(BinaryScalarFoldingRule scalar_rule) {
899   return [scalar_rule](IRContext* context, Instruction* inst,
900                        const std::vector<const analysis::Constant*>& constants)
901              -> const analysis::Constant* {
902     if (!inst->IsFloatingPointFoldingAllowed()) {
903       return nullptr;
904     }
905     if (inst->opcode() == spv::Op::OpExtInst) {
906       return FoldFPBinaryOp(scalar_rule, inst->type_id(),
907                             {constants[1], constants[2]}, context);
908     }
909     return FoldFPBinaryOp(scalar_rule, inst->type_id(), constants, context);
910   };
911 }
912 
913 // This macro defines a |UnaryScalarFoldingRule| that performs float to
914 // integer conversion.
915 // TODO(greg-lunarg): Support for 64-bit integer types.
FoldFToIOp()916 UnaryScalarFoldingRule FoldFToIOp() {
917   return [](const analysis::Type* result_type, const analysis::Constant* a,
918             analysis::ConstantManager* const_mgr) -> const analysis::Constant* {
919     assert(result_type != nullptr && a != nullptr);
920     const analysis::Integer* integer_type = result_type->AsInteger();
921     const analysis::Float* float_type = a->type()->AsFloat();
922     assert(float_type != nullptr);
923     assert(integer_type != nullptr);
924     if (integer_type->width() != 32) return nullptr;
925     if (float_type->width() == 32) {
926       float fa = a->GetFloat();
927       uint32_t result = integer_type->IsSigned()
928                             ? static_cast<uint32_t>(static_cast<int32_t>(fa))
929                             : static_cast<uint32_t>(fa);
930       std::vector<uint32_t> words = {result};
931       return const_mgr->GetConstant(result_type, words);
932     } else if (float_type->width() == 64) {
933       double fa = a->GetDouble();
934       uint32_t result = integer_type->IsSigned()
935                             ? static_cast<uint32_t>(static_cast<int32_t>(fa))
936                             : static_cast<uint32_t>(fa);
937       std::vector<uint32_t> words = {result};
938       return const_mgr->GetConstant(result_type, words);
939     }
940     return nullptr;
941   };
942 }
943 
944 // This function defines a |UnaryScalarFoldingRule| that performs integer to
945 // float conversion.
946 // TODO(greg-lunarg): Support for 64-bit integer types.
FoldIToFOp()947 UnaryScalarFoldingRule FoldIToFOp() {
948   return [](const analysis::Type* result_type, const analysis::Constant* a,
949             analysis::ConstantManager* const_mgr) -> const analysis::Constant* {
950     assert(result_type != nullptr && a != nullptr);
951     const analysis::Integer* integer_type = a->type()->AsInteger();
952     const analysis::Float* float_type = result_type->AsFloat();
953     assert(float_type != nullptr);
954     assert(integer_type != nullptr);
955     if (integer_type->width() != 32) return nullptr;
956     uint32_t ua = a->GetU32();
957     if (float_type->width() == 32) {
958       float result_val = integer_type->IsSigned()
959                              ? static_cast<float>(static_cast<int32_t>(ua))
960                              : static_cast<float>(ua);
961       utils::FloatProxy<float> result(result_val);
962       std::vector<uint32_t> words = {result.data()};
963       return const_mgr->GetConstant(result_type, words);
964     } else if (float_type->width() == 64) {
965       double result_val = integer_type->IsSigned()
966                               ? static_cast<double>(static_cast<int32_t>(ua))
967                               : static_cast<double>(ua);
968       utils::FloatProxy<double> result(result_val);
969       std::vector<uint32_t> words = result.GetWords();
970       return const_mgr->GetConstant(result_type, words);
971     }
972     return nullptr;
973   };
974 }
975 
976 // This defines a |UnaryScalarFoldingRule| that performs |OpQuantizeToF16|.
FoldQuantizeToF16Scalar()977 UnaryScalarFoldingRule FoldQuantizeToF16Scalar() {
978   return [](const analysis::Type* result_type, const analysis::Constant* a,
979             analysis::ConstantManager* const_mgr) -> const analysis::Constant* {
980     assert(result_type != nullptr && a != nullptr);
981     const analysis::Float* float_type = a->type()->AsFloat();
982     assert(float_type != nullptr);
983     if (float_type->width() != 32) {
984       return nullptr;
985     }
986 
987     float fa = a->GetFloat();
988     utils::HexFloat<utils::FloatProxy<float>> orignal(fa);
989     utils::HexFloat<utils::FloatProxy<utils::Float16>> quantized(0);
990     utils::HexFloat<utils::FloatProxy<float>> result(0.0f);
991     orignal.castTo(quantized, utils::round_direction::kToZero);
992     quantized.castTo(result, utils::round_direction::kToZero);
993     std::vector<uint32_t> words = {result.getBits()};
994     return const_mgr->GetConstant(result_type, words);
995   };
996 }
997 
998 // This macro defines a |BinaryScalarFoldingRule| that applies |op|.  The
999 // operator |op| must work for both float and double, and use syntax "f1 op f2".
1000 #define FOLD_FPARITH_OP(op)                                                   \
1001   [](const analysis::Type* result_type_in_macro, const analysis::Constant* a, \
1002      const analysis::Constant* b,                                             \
1003      analysis::ConstantManager* const_mgr_in_macro)                           \
1004       -> const analysis::Constant* {                                          \
1005     assert(result_type_in_macro != nullptr && a != nullptr && b != nullptr);  \
1006     assert(result_type_in_macro == a->type() &&                               \
1007            result_type_in_macro == b->type());                                \
1008     const analysis::Float* float_type_in_macro =                              \
1009         result_type_in_macro->AsFloat();                                      \
1010     assert(float_type_in_macro != nullptr);                                   \
1011     if (float_type_in_macro->width() == 32) {                                 \
1012       float fa = a->GetFloat();                                               \
1013       float fb = b->GetFloat();                                               \
1014       utils::FloatProxy<float> result_in_macro(fa op fb);                     \
1015       std::vector<uint32_t> words_in_macro = result_in_macro.GetWords();      \
1016       return const_mgr_in_macro->GetConstant(result_type_in_macro,            \
1017                                              words_in_macro);                 \
1018     } else if (float_type_in_macro->width() == 64) {                          \
1019       double fa = a->GetDouble();                                             \
1020       double fb = b->GetDouble();                                             \
1021       utils::FloatProxy<double> result_in_macro(fa op fb);                    \
1022       std::vector<uint32_t> words_in_macro = result_in_macro.GetWords();      \
1023       return const_mgr_in_macro->GetConstant(result_type_in_macro,            \
1024                                              words_in_macro);                 \
1025     }                                                                         \
1026     return nullptr;                                                           \
1027   }
1028 
1029 // Define the folding rule for conversion between floating point and integer
FoldFToI()1030 ConstantFoldingRule FoldFToI() { return FoldFPUnaryOp(FoldFToIOp()); }
FoldIToF()1031 ConstantFoldingRule FoldIToF() { return FoldFPUnaryOp(FoldIToFOp()); }
FoldQuantizeToF16()1032 ConstantFoldingRule FoldQuantizeToF16() {
1033   return FoldFPUnaryOp(FoldQuantizeToF16Scalar());
1034 }
1035 
1036 // Define the folding rules for subtraction, addition, multiplication, and
1037 // division for floating point values.
FoldFSub()1038 ConstantFoldingRule FoldFSub() { return FoldFPBinaryOp(FOLD_FPARITH_OP(-)); }
FoldFAdd()1039 ConstantFoldingRule FoldFAdd() { return FoldFPBinaryOp(FOLD_FPARITH_OP(+)); }
FoldFMul()1040 ConstantFoldingRule FoldFMul() { return FoldFPBinaryOp(FOLD_FPARITH_OP(*)); }
1041 
1042 // Returns the constant that results from evaluating |numerator| / 0.0.  Returns
1043 // |nullptr| if the result could not be evaluated.
FoldFPScalarDivideByZero(const analysis::Type * result_type,const analysis::Constant * numerator,analysis::ConstantManager * const_mgr)1044 const analysis::Constant* FoldFPScalarDivideByZero(
1045     const analysis::Type* result_type, const analysis::Constant* numerator,
1046     analysis::ConstantManager* const_mgr) {
1047   if (numerator == nullptr) {
1048     return nullptr;
1049   }
1050 
1051   if (numerator->IsZero()) {
1052     return GetNan(result_type, const_mgr);
1053   }
1054 
1055   const analysis::Constant* result = GetInf(result_type, const_mgr);
1056   if (result == nullptr) {
1057     return nullptr;
1058   }
1059 
1060   if (numerator->AsFloatConstant()->GetValueAsDouble() < 0.0) {
1061     result = NegateFPConst(result_type, result, const_mgr);
1062   }
1063   return result;
1064 }
1065 
1066 // Returns the result of folding |numerator| / |denominator|.  Returns |nullptr|
1067 // if it cannot be folded.
FoldScalarFPDivide(const analysis::Type * result_type,const analysis::Constant * numerator,const analysis::Constant * denominator,analysis::ConstantManager * const_mgr)1068 const analysis::Constant* FoldScalarFPDivide(
1069     const analysis::Type* result_type, const analysis::Constant* numerator,
1070     const analysis::Constant* denominator,
1071     analysis::ConstantManager* const_mgr) {
1072   if (denominator == nullptr) {
1073     return nullptr;
1074   }
1075 
1076   if (denominator->IsZero()) {
1077     return FoldFPScalarDivideByZero(result_type, numerator, const_mgr);
1078   }
1079 
1080   uint32_t width = denominator->type()->AsFloat()->width();
1081   if (width != 32 && width != 64) {
1082     return nullptr;
1083   }
1084 
1085   const analysis::FloatConstant* denominator_float =
1086       denominator->AsFloatConstant();
1087   if (denominator_float && denominator->GetValueAsDouble() == -0.0) {
1088     const analysis::Constant* result =
1089         FoldFPScalarDivideByZero(result_type, numerator, const_mgr);
1090     if (result != nullptr)
1091       result = NegateFPConst(result_type, result, const_mgr);
1092     return result;
1093   } else {
1094     return FOLD_FPARITH_OP(/)(result_type, numerator, denominator, const_mgr);
1095   }
1096 }
1097 
1098 // Returns the constant folding rule to fold |OpFDiv| with two constants.
FoldFDiv()1099 ConstantFoldingRule FoldFDiv() { return FoldFPBinaryOp(FoldScalarFPDivide); }
1100 
CompareFloatingPoint(bool op_result,bool op_unordered,bool need_ordered)1101 bool CompareFloatingPoint(bool op_result, bool op_unordered,
1102                           bool need_ordered) {
1103   if (need_ordered) {
1104     // operands are ordered and Operand 1 is |op| Operand 2
1105     return !op_unordered && op_result;
1106   } else {
1107     // operands are unordered or Operand 1 is |op| Operand 2
1108     return op_unordered || op_result;
1109   }
1110 }
1111 
1112 // This macro defines a |BinaryScalarFoldingRule| that applies |op|.  The
1113 // operator |op| must work for both float and double, and use syntax "f1 op f2".
1114 #define FOLD_FPCMP_OP(op, ord)                                            \
1115   [](const analysis::Type* result_type, const analysis::Constant* a,      \
1116      const analysis::Constant* b,                                         \
1117      analysis::ConstantManager* const_mgr) -> const analysis::Constant* { \
1118     assert(result_type != nullptr && a != nullptr && b != nullptr);       \
1119     assert(result_type->AsBool());                                        \
1120     assert(a->type() == b->type());                                       \
1121     const analysis::Float* float_type = a->type()->AsFloat();             \
1122     assert(float_type != nullptr);                                        \
1123     if (float_type->width() == 32) {                                      \
1124       float fa = a->GetFloat();                                           \
1125       float fb = b->GetFloat();                                           \
1126       bool result = CompareFloatingPoint(                                 \
1127           fa op fb, std::isnan(fa) || std::isnan(fb), ord);               \
1128       std::vector<uint32_t> words = {uint32_t(result)};                   \
1129       return const_mgr->GetConstant(result_type, words);                  \
1130     } else if (float_type->width() == 64) {                               \
1131       double fa = a->GetDouble();                                         \
1132       double fb = b->GetDouble();                                         \
1133       bool result = CompareFloatingPoint(                                 \
1134           fa op fb, std::isnan(fa) || std::isnan(fb), ord);               \
1135       std::vector<uint32_t> words = {uint32_t(result)};                   \
1136       return const_mgr->GetConstant(result_type, words);                  \
1137     }                                                                     \
1138     return nullptr;                                                       \
1139   }
1140 
1141 // Define the folding rules for ordered and unordered comparison for floating
1142 // point values.
FoldFOrdEqual()1143 ConstantFoldingRule FoldFOrdEqual() {
1144   return FoldFPBinaryOp(FOLD_FPCMP_OP(==, true));
1145 }
FoldFUnordEqual()1146 ConstantFoldingRule FoldFUnordEqual() {
1147   return FoldFPBinaryOp(FOLD_FPCMP_OP(==, false));
1148 }
FoldFOrdNotEqual()1149 ConstantFoldingRule FoldFOrdNotEqual() {
1150   return FoldFPBinaryOp(FOLD_FPCMP_OP(!=, true));
1151 }
FoldFUnordNotEqual()1152 ConstantFoldingRule FoldFUnordNotEqual() {
1153   return FoldFPBinaryOp(FOLD_FPCMP_OP(!=, false));
1154 }
FoldFOrdLessThan()1155 ConstantFoldingRule FoldFOrdLessThan() {
1156   return FoldFPBinaryOp(FOLD_FPCMP_OP(<, true));
1157 }
FoldFUnordLessThan()1158 ConstantFoldingRule FoldFUnordLessThan() {
1159   return FoldFPBinaryOp(FOLD_FPCMP_OP(<, false));
1160 }
FoldFOrdGreaterThan()1161 ConstantFoldingRule FoldFOrdGreaterThan() {
1162   return FoldFPBinaryOp(FOLD_FPCMP_OP(>, true));
1163 }
FoldFUnordGreaterThan()1164 ConstantFoldingRule FoldFUnordGreaterThan() {
1165   return FoldFPBinaryOp(FOLD_FPCMP_OP(>, false));
1166 }
FoldFOrdLessThanEqual()1167 ConstantFoldingRule FoldFOrdLessThanEqual() {
1168   return FoldFPBinaryOp(FOLD_FPCMP_OP(<=, true));
1169 }
FoldFUnordLessThanEqual()1170 ConstantFoldingRule FoldFUnordLessThanEqual() {
1171   return FoldFPBinaryOp(FOLD_FPCMP_OP(<=, false));
1172 }
FoldFOrdGreaterThanEqual()1173 ConstantFoldingRule FoldFOrdGreaterThanEqual() {
1174   return FoldFPBinaryOp(FOLD_FPCMP_OP(>=, true));
1175 }
FoldFUnordGreaterThanEqual()1176 ConstantFoldingRule FoldFUnordGreaterThanEqual() {
1177   return FoldFPBinaryOp(FOLD_FPCMP_OP(>=, false));
1178 }
1179 
1180 // Folds an OpDot where all of the inputs are constants to a
1181 // constant.  A new constant is created if necessary.
FoldOpDotWithConstants()1182 ConstantFoldingRule FoldOpDotWithConstants() {
1183   return [](IRContext* context, Instruction* inst,
1184             const std::vector<const analysis::Constant*>& constants)
1185              -> const analysis::Constant* {
1186     analysis::ConstantManager* const_mgr = context->get_constant_mgr();
1187     analysis::TypeManager* type_mgr = context->get_type_mgr();
1188     const analysis::Type* new_type = type_mgr->GetType(inst->type_id());
1189     assert(new_type->AsFloat() && "OpDot should have a float return type.");
1190     const analysis::Float* float_type = new_type->AsFloat();
1191 
1192     if (!inst->IsFloatingPointFoldingAllowed()) {
1193       return nullptr;
1194     }
1195 
1196     // If one of the operands is 0, then the result is 0.
1197     bool has_zero_operand = false;
1198 
1199     for (int i = 0; i < 2; ++i) {
1200       if (constants[i]) {
1201         if (constants[i]->AsNullConstant() ||
1202             constants[i]->AsVectorConstant()->IsZero()) {
1203           has_zero_operand = true;
1204           break;
1205         }
1206       }
1207     }
1208 
1209     if (has_zero_operand) {
1210       if (float_type->width() == 32) {
1211         utils::FloatProxy<float> result(0.0f);
1212         std::vector<uint32_t> words = result.GetWords();
1213         return const_mgr->GetConstant(float_type, words);
1214       }
1215       if (float_type->width() == 64) {
1216         utils::FloatProxy<double> result(0.0);
1217         std::vector<uint32_t> words = result.GetWords();
1218         return const_mgr->GetConstant(float_type, words);
1219       }
1220       return nullptr;
1221     }
1222 
1223     if (constants[0] == nullptr || constants[1] == nullptr) {
1224       return nullptr;
1225     }
1226 
1227     std::vector<const analysis::Constant*> a_components;
1228     std::vector<const analysis::Constant*> b_components;
1229 
1230     a_components = constants[0]->GetVectorComponents(const_mgr);
1231     b_components = constants[1]->GetVectorComponents(const_mgr);
1232 
1233     utils::FloatProxy<double> result(0.0);
1234     std::vector<uint32_t> words = result.GetWords();
1235     const analysis::Constant* result_const =
1236         const_mgr->GetConstant(float_type, words);
1237     for (uint32_t i = 0; i < a_components.size() && result_const != nullptr;
1238          ++i) {
1239       if (a_components[i] == nullptr || b_components[i] == nullptr) {
1240         return nullptr;
1241       }
1242 
1243       const analysis::Constant* component = FOLD_FPARITH_OP(*)(
1244           new_type, a_components[i], b_components[i], const_mgr);
1245       if (component == nullptr) {
1246         return nullptr;
1247       }
1248       result_const =
1249           FOLD_FPARITH_OP(+)(new_type, result_const, component, const_mgr);
1250     }
1251     return result_const;
1252   };
1253 }
1254 
FoldFNegate()1255 ConstantFoldingRule FoldFNegate() { return FoldFPUnaryOp(NegateFPConst); }
FoldSNegate()1256 ConstantFoldingRule FoldSNegate() { return FoldUnaryOp(NegateIntConst); }
1257 
FoldFClampFeedingCompare(spv::Op cmp_opcode)1258 ConstantFoldingRule FoldFClampFeedingCompare(spv::Op cmp_opcode) {
1259   return [cmp_opcode](IRContext* context, Instruction* inst,
1260                       const std::vector<const analysis::Constant*>& constants)
1261              -> const analysis::Constant* {
1262     analysis::ConstantManager* const_mgr = context->get_constant_mgr();
1263     analysis::DefUseManager* def_use_mgr = context->get_def_use_mgr();
1264 
1265     if (!inst->IsFloatingPointFoldingAllowed()) {
1266       return nullptr;
1267     }
1268 
1269     uint32_t non_const_idx = (constants[0] ? 1 : 0);
1270     uint32_t operand_id = inst->GetSingleWordInOperand(non_const_idx);
1271     Instruction* operand_inst = def_use_mgr->GetDef(operand_id);
1272 
1273     analysis::TypeManager* type_mgr = context->get_type_mgr();
1274     const analysis::Type* operand_type =
1275         type_mgr->GetType(operand_inst->type_id());
1276 
1277     if (!operand_type->AsFloat()) {
1278       return nullptr;
1279     }
1280 
1281     if (operand_type->AsFloat()->width() != 32 &&
1282         operand_type->AsFloat()->width() != 64) {
1283       return nullptr;
1284     }
1285 
1286     if (operand_inst->opcode() != spv::Op::OpExtInst) {
1287       return nullptr;
1288     }
1289 
1290     if (operand_inst->GetSingleWordInOperand(1) != GLSLstd450FClamp) {
1291       return nullptr;
1292     }
1293 
1294     if (constants[1] == nullptr && constants[0] == nullptr) {
1295       return nullptr;
1296     }
1297 
1298     uint32_t max_id = operand_inst->GetSingleWordInOperand(4);
1299     const analysis::Constant* max_const =
1300         const_mgr->FindDeclaredConstant(max_id);
1301 
1302     uint32_t min_id = operand_inst->GetSingleWordInOperand(3);
1303     const analysis::Constant* min_const =
1304         const_mgr->FindDeclaredConstant(min_id);
1305 
1306     bool found_result = false;
1307     bool result = false;
1308 
1309     switch (cmp_opcode) {
1310       case spv::Op::OpFOrdLessThan:
1311       case spv::Op::OpFUnordLessThan:
1312       case spv::Op::OpFOrdGreaterThanEqual:
1313       case spv::Op::OpFUnordGreaterThanEqual:
1314         if (constants[0]) {
1315           if (min_const) {
1316             if (constants[0]->GetValueAsDouble() <
1317                 min_const->GetValueAsDouble()) {
1318               found_result = true;
1319               result = (cmp_opcode == spv::Op::OpFOrdLessThan ||
1320                         cmp_opcode == spv::Op::OpFUnordLessThan);
1321             }
1322           }
1323           if (max_const) {
1324             if (constants[0]->GetValueAsDouble() >=
1325                 max_const->GetValueAsDouble()) {
1326               found_result = true;
1327               result = !(cmp_opcode == spv::Op::OpFOrdLessThan ||
1328                          cmp_opcode == spv::Op::OpFUnordLessThan);
1329             }
1330           }
1331         }
1332 
1333         if (constants[1]) {
1334           if (max_const) {
1335             if (max_const->GetValueAsDouble() <
1336                 constants[1]->GetValueAsDouble()) {
1337               found_result = true;
1338               result = (cmp_opcode == spv::Op::OpFOrdLessThan ||
1339                         cmp_opcode == spv::Op::OpFUnordLessThan);
1340             }
1341           }
1342 
1343           if (min_const) {
1344             if (min_const->GetValueAsDouble() >=
1345                 constants[1]->GetValueAsDouble()) {
1346               found_result = true;
1347               result = !(cmp_opcode == spv::Op::OpFOrdLessThan ||
1348                          cmp_opcode == spv::Op::OpFUnordLessThan);
1349             }
1350           }
1351         }
1352         break;
1353       case spv::Op::OpFOrdGreaterThan:
1354       case spv::Op::OpFUnordGreaterThan:
1355       case spv::Op::OpFOrdLessThanEqual:
1356       case spv::Op::OpFUnordLessThanEqual:
1357         if (constants[0]) {
1358           if (min_const) {
1359             if (constants[0]->GetValueAsDouble() <=
1360                 min_const->GetValueAsDouble()) {
1361               found_result = true;
1362               result = (cmp_opcode == spv::Op::OpFOrdLessThanEqual ||
1363                         cmp_opcode == spv::Op::OpFUnordLessThanEqual);
1364             }
1365           }
1366           if (max_const) {
1367             if (constants[0]->GetValueAsDouble() >
1368                 max_const->GetValueAsDouble()) {
1369               found_result = true;
1370               result = !(cmp_opcode == spv::Op::OpFOrdLessThanEqual ||
1371                          cmp_opcode == spv::Op::OpFUnordLessThanEqual);
1372             }
1373           }
1374         }
1375 
1376         if (constants[1]) {
1377           if (max_const) {
1378             if (max_const->GetValueAsDouble() <=
1379                 constants[1]->GetValueAsDouble()) {
1380               found_result = true;
1381               result = (cmp_opcode == spv::Op::OpFOrdLessThanEqual ||
1382                         cmp_opcode == spv::Op::OpFUnordLessThanEqual);
1383             }
1384           }
1385 
1386           if (min_const) {
1387             if (min_const->GetValueAsDouble() >
1388                 constants[1]->GetValueAsDouble()) {
1389               found_result = true;
1390               result = !(cmp_opcode == spv::Op::OpFOrdLessThanEqual ||
1391                          cmp_opcode == spv::Op::OpFUnordLessThanEqual);
1392             }
1393           }
1394         }
1395         break;
1396       default:
1397         return nullptr;
1398     }
1399 
1400     if (!found_result) {
1401       return nullptr;
1402     }
1403 
1404     const analysis::Type* bool_type =
1405         context->get_type_mgr()->GetType(inst->type_id());
1406     const analysis::Constant* result_const =
1407         const_mgr->GetConstant(bool_type, {static_cast<uint32_t>(result)});
1408     assert(result_const);
1409     return result_const;
1410   };
1411 }
1412 
FoldFMix()1413 ConstantFoldingRule FoldFMix() {
1414   return [](IRContext* context, Instruction* inst,
1415             const std::vector<const analysis::Constant*>& constants)
1416              -> const analysis::Constant* {
1417     analysis::ConstantManager* const_mgr = context->get_constant_mgr();
1418     assert(inst->opcode() == spv::Op::OpExtInst &&
1419            "Expecting an extended instruction.");
1420     assert(inst->GetSingleWordInOperand(0) ==
1421                context->get_feature_mgr()->GetExtInstImportId_GLSLstd450() &&
1422            "Expecting a GLSLstd450 extended instruction.");
1423     assert(inst->GetSingleWordInOperand(1) == GLSLstd450FMix &&
1424            "Expecting and FMix instruction.");
1425 
1426     if (!inst->IsFloatingPointFoldingAllowed()) {
1427       return nullptr;
1428     }
1429 
1430     // Make sure all FMix operands are constants.
1431     for (uint32_t i = 1; i < 4; i++) {
1432       if (constants[i] == nullptr) {
1433         return nullptr;
1434       }
1435     }
1436 
1437     const analysis::Constant* one;
1438     bool is_vector = false;
1439     const analysis::Type* result_type = constants[1]->type();
1440     const analysis::Type* base_type = result_type;
1441     if (base_type->AsVector()) {
1442       is_vector = true;
1443       base_type = base_type->AsVector()->element_type();
1444     }
1445     assert(base_type->AsFloat() != nullptr &&
1446            "FMix is suppose to act on floats or vectors of floats.");
1447 
1448     if (base_type->AsFloat()->width() == 32) {
1449       one = const_mgr->GetConstant(base_type,
1450                                    utils::FloatProxy<float>(1.0f).GetWords());
1451     } else {
1452       one = const_mgr->GetConstant(base_type,
1453                                    utils::FloatProxy<double>(1.0).GetWords());
1454     }
1455 
1456     if (is_vector) {
1457       uint32_t one_id = const_mgr->GetDefiningInstruction(one)->result_id();
1458       one =
1459           const_mgr->GetConstant(result_type, std::vector<uint32_t>(4, one_id));
1460     }
1461 
1462     const analysis::Constant* temp1 = FoldFPBinaryOp(
1463         FOLD_FPARITH_OP(-), inst->type_id(), {one, constants[3]}, context);
1464     if (temp1 == nullptr) {
1465       return nullptr;
1466     }
1467 
1468     const analysis::Constant* temp2 = FoldFPBinaryOp(
1469         FOLD_FPARITH_OP(*), inst->type_id(), {constants[1], temp1}, context);
1470     if (temp2 == nullptr) {
1471       return nullptr;
1472     }
1473     const analysis::Constant* temp3 =
1474         FoldFPBinaryOp(FOLD_FPARITH_OP(*), inst->type_id(),
1475                        {constants[2], constants[3]}, context);
1476     if (temp3 == nullptr) {
1477       return nullptr;
1478     }
1479     return FoldFPBinaryOp(FOLD_FPARITH_OP(+), inst->type_id(), {temp2, temp3},
1480                           context);
1481   };
1482 }
1483 
FoldMin(const analysis::Type * result_type,const analysis::Constant * a,const analysis::Constant * b,analysis::ConstantManager *)1484 const analysis::Constant* FoldMin(const analysis::Type* result_type,
1485                                   const analysis::Constant* a,
1486                                   const analysis::Constant* b,
1487                                   analysis::ConstantManager*) {
1488   if (const analysis::Integer* int_type = result_type->AsInteger()) {
1489     if (int_type->width() == 32) {
1490       if (int_type->IsSigned()) {
1491         int32_t va = a->GetS32();
1492         int32_t vb = b->GetS32();
1493         return (va < vb ? a : b);
1494       } else {
1495         uint32_t va = a->GetU32();
1496         uint32_t vb = b->GetU32();
1497         return (va < vb ? a : b);
1498       }
1499     } else if (int_type->width() == 64) {
1500       if (int_type->IsSigned()) {
1501         int64_t va = a->GetS64();
1502         int64_t vb = b->GetS64();
1503         return (va < vb ? a : b);
1504       } else {
1505         uint64_t va = a->GetU64();
1506         uint64_t vb = b->GetU64();
1507         return (va < vb ? a : b);
1508       }
1509     }
1510   } else if (const analysis::Float* float_type = result_type->AsFloat()) {
1511     if (float_type->width() == 32) {
1512       float va = a->GetFloat();
1513       float vb = b->GetFloat();
1514       return (va < vb ? a : b);
1515     } else if (float_type->width() == 64) {
1516       double va = a->GetDouble();
1517       double vb = b->GetDouble();
1518       return (va < vb ? a : b);
1519     }
1520   }
1521   return nullptr;
1522 }
1523 
FoldMax(const analysis::Type * result_type,const analysis::Constant * a,const analysis::Constant * b,analysis::ConstantManager *)1524 const analysis::Constant* FoldMax(const analysis::Type* result_type,
1525                                   const analysis::Constant* a,
1526                                   const analysis::Constant* b,
1527                                   analysis::ConstantManager*) {
1528   if (const analysis::Integer* int_type = result_type->AsInteger()) {
1529     if (int_type->width() == 32) {
1530       if (int_type->IsSigned()) {
1531         int32_t va = a->GetS32();
1532         int32_t vb = b->GetS32();
1533         return (va > vb ? a : b);
1534       } else {
1535         uint32_t va = a->GetU32();
1536         uint32_t vb = b->GetU32();
1537         return (va > vb ? a : b);
1538       }
1539     } else if (int_type->width() == 64) {
1540       if (int_type->IsSigned()) {
1541         int64_t va = a->GetS64();
1542         int64_t vb = b->GetS64();
1543         return (va > vb ? a : b);
1544       } else {
1545         uint64_t va = a->GetU64();
1546         uint64_t vb = b->GetU64();
1547         return (va > vb ? a : b);
1548       }
1549     }
1550   } else if (const analysis::Float* float_type = result_type->AsFloat()) {
1551     if (float_type->width() == 32) {
1552       float va = a->GetFloat();
1553       float vb = b->GetFloat();
1554       return (va > vb ? a : b);
1555     } else if (float_type->width() == 64) {
1556       double va = a->GetDouble();
1557       double vb = b->GetDouble();
1558       return (va > vb ? a : b);
1559     }
1560   }
1561   return nullptr;
1562 }
1563 
1564 // Fold an clamp instruction when all three operands are constant.
FoldClamp1(IRContext * context,Instruction * inst,const std::vector<const analysis::Constant * > & constants)1565 const analysis::Constant* FoldClamp1(
1566     IRContext* context, Instruction* inst,
1567     const std::vector<const analysis::Constant*>& constants) {
1568   assert(inst->opcode() == spv::Op::OpExtInst &&
1569          "Expecting an extended instruction.");
1570   assert(inst->GetSingleWordInOperand(0) ==
1571              context->get_feature_mgr()->GetExtInstImportId_GLSLstd450() &&
1572          "Expecting a GLSLstd450 extended instruction.");
1573 
1574   // Make sure all Clamp operands are constants.
1575   for (uint32_t i = 1; i < 4; i++) {
1576     if (constants[i] == nullptr) {
1577       return nullptr;
1578     }
1579   }
1580 
1581   const analysis::Constant* temp = FoldFPBinaryOp(
1582       FoldMax, inst->type_id(), {constants[1], constants[2]}, context);
1583   if (temp == nullptr) {
1584     return nullptr;
1585   }
1586   return FoldFPBinaryOp(FoldMin, inst->type_id(), {temp, constants[3]},
1587                         context);
1588 }
1589 
1590 // Fold a clamp instruction when |x <= min_val|.
FoldClamp2(IRContext * context,Instruction * inst,const std::vector<const analysis::Constant * > & constants)1591 const analysis::Constant* FoldClamp2(
1592     IRContext* context, Instruction* inst,
1593     const std::vector<const analysis::Constant*>& constants) {
1594   assert(inst->opcode() == spv::Op::OpExtInst &&
1595          "Expecting an extended instruction.");
1596   assert(inst->GetSingleWordInOperand(0) ==
1597              context->get_feature_mgr()->GetExtInstImportId_GLSLstd450() &&
1598          "Expecting a GLSLstd450 extended instruction.");
1599 
1600   const analysis::Constant* x = constants[1];
1601   const analysis::Constant* min_val = constants[2];
1602 
1603   if (x == nullptr || min_val == nullptr) {
1604     return nullptr;
1605   }
1606 
1607   const analysis::Constant* temp =
1608       FoldFPBinaryOp(FoldMax, inst->type_id(), {x, min_val}, context);
1609   if (temp == min_val) {
1610     // We can assume that |min_val| is less than |max_val|.  Therefore, if the
1611     // result of the max operation is |min_val|, we know the result of the min
1612     // operation, even if |max_val| is not a constant.
1613     return min_val;
1614   }
1615   return nullptr;
1616 }
1617 
1618 // Fold a clamp instruction when |x >= max_val|.
FoldClamp3(IRContext * context,Instruction * inst,const std::vector<const analysis::Constant * > & constants)1619 const analysis::Constant* FoldClamp3(
1620     IRContext* context, Instruction* inst,
1621     const std::vector<const analysis::Constant*>& constants) {
1622   assert(inst->opcode() == spv::Op::OpExtInst &&
1623          "Expecting an extended instruction.");
1624   assert(inst->GetSingleWordInOperand(0) ==
1625              context->get_feature_mgr()->GetExtInstImportId_GLSLstd450() &&
1626          "Expecting a GLSLstd450 extended instruction.");
1627 
1628   const analysis::Constant* x = constants[1];
1629   const analysis::Constant* max_val = constants[3];
1630 
1631   if (x == nullptr || max_val == nullptr) {
1632     return nullptr;
1633   }
1634 
1635   const analysis::Constant* temp =
1636       FoldFPBinaryOp(FoldMin, inst->type_id(), {x, max_val}, context);
1637   if (temp == max_val) {
1638     // We can assume that |min_val| is less than |max_val|.  Therefore, if the
1639     // result of the max operation is |min_val|, we know the result of the min
1640     // operation, even if |max_val| is not a constant.
1641     return max_val;
1642   }
1643   return nullptr;
1644 }
1645 
FoldFTranscendentalUnary(double (* fp)(double))1646 UnaryScalarFoldingRule FoldFTranscendentalUnary(double (*fp)(double)) {
1647   return
1648       [fp](const analysis::Type* result_type, const analysis::Constant* a,
1649            analysis::ConstantManager* const_mgr) -> const analysis::Constant* {
1650         assert(result_type != nullptr && a != nullptr);
1651         const analysis::Float* float_type = a->type()->AsFloat();
1652         assert(float_type != nullptr);
1653         assert(float_type == result_type->AsFloat());
1654         if (float_type->width() == 32) {
1655           float fa = a->GetFloat();
1656           float res = static_cast<float>(fp(fa));
1657           utils::FloatProxy<float> result(res);
1658           std::vector<uint32_t> words = result.GetWords();
1659           return const_mgr->GetConstant(result_type, words);
1660         } else if (float_type->width() == 64) {
1661           double fa = a->GetDouble();
1662           double res = fp(fa);
1663           utils::FloatProxy<double> result(res);
1664           std::vector<uint32_t> words = result.GetWords();
1665           return const_mgr->GetConstant(result_type, words);
1666         }
1667         return nullptr;
1668       };
1669 }
1670 
FoldFTranscendentalBinary(double (* fp)(double,double))1671 BinaryScalarFoldingRule FoldFTranscendentalBinary(double (*fp)(double,
1672                                                                double)) {
1673   return
1674       [fp](const analysis::Type* result_type, const analysis::Constant* a,
1675            const analysis::Constant* b,
1676            analysis::ConstantManager* const_mgr) -> const analysis::Constant* {
1677         assert(result_type != nullptr && a != nullptr);
1678         const analysis::Float* float_type = a->type()->AsFloat();
1679         assert(float_type != nullptr);
1680         assert(float_type == result_type->AsFloat());
1681         assert(float_type == b->type()->AsFloat());
1682         if (float_type->width() == 32) {
1683           float fa = a->GetFloat();
1684           float fb = b->GetFloat();
1685           float res = static_cast<float>(fp(fa, fb));
1686           utils::FloatProxy<float> result(res);
1687           std::vector<uint32_t> words = result.GetWords();
1688           return const_mgr->GetConstant(result_type, words);
1689         } else if (float_type->width() == 64) {
1690           double fa = a->GetDouble();
1691           double fb = b->GetDouble();
1692           double res = fp(fa, fb);
1693           utils::FloatProxy<double> result(res);
1694           std::vector<uint32_t> words = result.GetWords();
1695           return const_mgr->GetConstant(result_type, words);
1696         }
1697         return nullptr;
1698       };
1699 }
1700 
1701 enum Sign { Signed, Unsigned };
1702 
1703 // Returns a BinaryScalarFoldingRule that applies `op` to the scalars.
1704 // The `signedness` is used to determine if the operands should be interpreted
1705 // as signed or unsigned. If the operands are signed, the value will be sign
1706 // extended before the value is passed to `op`. Otherwise the values will be
1707 // zero extended.
1708 template <Sign signedness>
FoldBinaryIntegerOperation(uint64_t (* op)(uint64_t,uint64_t))1709 BinaryScalarFoldingRule FoldBinaryIntegerOperation(uint64_t (*op)(uint64_t,
1710                                                                   uint64_t)) {
1711   return
1712       [op](const analysis::Type* result_type, const analysis::Constant* a,
1713            const analysis::Constant* b,
1714            analysis::ConstantManager* const_mgr) -> const analysis::Constant* {
1715         assert(result_type != nullptr && a != nullptr && b != nullptr);
1716         const analysis::Integer* integer_type = result_type->AsInteger();
1717         assert(integer_type != nullptr);
1718         assert(a->type()->kind() == analysis::Type::kInteger);
1719         assert(b->type()->kind() == analysis::Type::kInteger);
1720         assert(integer_type->width() == a->type()->AsInteger()->width());
1721         assert(integer_type->width() == b->type()->AsInteger()->width());
1722 
1723         // In SPIR-V, all operations support unsigned types, but the way they
1724         // are interpreted depends on the opcode. This is why we use the
1725         // template argument to determine how to interpret the operands.
1726         uint64_t ia = (signedness == Signed ? a->GetSignExtendedValue()
1727                                             : a->GetZeroExtendedValue());
1728         uint64_t ib = (signedness == Signed ? b->GetSignExtendedValue()
1729                                             : b->GetZeroExtendedValue());
1730         uint64_t result = op(ia, ib);
1731 
1732         const analysis::Constant* result_constant =
1733             GenerateIntegerConstant(integer_type, result, const_mgr);
1734         return result_constant;
1735       };
1736 }
1737 
1738 // A scalar folding rule that folds OpSConvert.
FoldScalarSConvert(const analysis::Type * result_type,const analysis::Constant * a,analysis::ConstantManager * const_mgr)1739 const analysis::Constant* FoldScalarSConvert(
1740     const analysis::Type* result_type, const analysis::Constant* a,
1741     analysis::ConstantManager* const_mgr) {
1742   assert(result_type != nullptr);
1743   assert(a != nullptr);
1744   assert(const_mgr != nullptr);
1745   const analysis::Integer* integer_type = result_type->AsInteger();
1746   assert(integer_type && "The result type of an SConvert");
1747   int64_t value = a->GetSignExtendedValue();
1748   return GenerateIntegerConstant(integer_type, value, const_mgr);
1749 }
1750 
1751 // A scalar folding rule that folds OpUConvert.
FoldScalarUConvert(const analysis::Type * result_type,const analysis::Constant * a,analysis::ConstantManager * const_mgr)1752 const analysis::Constant* FoldScalarUConvert(
1753     const analysis::Type* result_type, const analysis::Constant* a,
1754     analysis::ConstantManager* const_mgr) {
1755   assert(result_type != nullptr);
1756   assert(a != nullptr);
1757   assert(const_mgr != nullptr);
1758   const analysis::Integer* integer_type = result_type->AsInteger();
1759   assert(integer_type && "The result type of an UConvert");
1760   uint64_t value = a->GetZeroExtendedValue();
1761 
1762   // If the operand was an unsigned value with less than 32-bit, it would have
1763   // been sign extended earlier, and we need to clear those bits.
1764   auto* operand_type = a->type()->AsInteger();
1765   value = ZeroExtendValue(value, operand_type->width());
1766   return GenerateIntegerConstant(integer_type, value, const_mgr);
1767 }
1768 }  // namespace
1769 
AddFoldingRules()1770 void ConstantFoldingRules::AddFoldingRules() {
1771   // Add all folding rules to the list for the opcodes to which they apply.
1772   // Note that the order in which rules are added to the list matters. If a rule
1773   // applies to the instruction, the rest of the rules will not be attempted.
1774   // Take that into consideration.
1775 
1776   rules_[spv::Op::OpCompositeConstruct].push_back(FoldCompositeWithConstants());
1777 
1778   rules_[spv::Op::OpCompositeExtract].push_back(FoldExtractWithConstants());
1779   rules_[spv::Op::OpCompositeInsert].push_back(FoldInsertWithConstants());
1780 
1781   rules_[spv::Op::OpConvertFToS].push_back(FoldFToI());
1782   rules_[spv::Op::OpConvertFToU].push_back(FoldFToI());
1783   rules_[spv::Op::OpConvertSToF].push_back(FoldIToF());
1784   rules_[spv::Op::OpConvertUToF].push_back(FoldIToF());
1785   rules_[spv::Op::OpSConvert].push_back(FoldUnaryOp(FoldScalarSConvert));
1786   rules_[spv::Op::OpUConvert].push_back(FoldUnaryOp(FoldScalarUConvert));
1787 
1788   rules_[spv::Op::OpDot].push_back(FoldOpDotWithConstants());
1789   rules_[spv::Op::OpFAdd].push_back(FoldFAdd());
1790   rules_[spv::Op::OpFDiv].push_back(FoldFDiv());
1791   rules_[spv::Op::OpFMul].push_back(FoldFMul());
1792   rules_[spv::Op::OpFSub].push_back(FoldFSub());
1793 
1794   rules_[spv::Op::OpFOrdEqual].push_back(FoldFOrdEqual());
1795 
1796   rules_[spv::Op::OpFUnordEqual].push_back(FoldFUnordEqual());
1797 
1798   rules_[spv::Op::OpFOrdNotEqual].push_back(FoldFOrdNotEqual());
1799 
1800   rules_[spv::Op::OpFUnordNotEqual].push_back(FoldFUnordNotEqual());
1801 
1802   rules_[spv::Op::OpFOrdLessThan].push_back(FoldFOrdLessThan());
1803   rules_[spv::Op::OpFOrdLessThan].push_back(
1804       FoldFClampFeedingCompare(spv::Op::OpFOrdLessThan));
1805 
1806   rules_[spv::Op::OpFUnordLessThan].push_back(FoldFUnordLessThan());
1807   rules_[spv::Op::OpFUnordLessThan].push_back(
1808       FoldFClampFeedingCompare(spv::Op::OpFUnordLessThan));
1809 
1810   rules_[spv::Op::OpFOrdGreaterThan].push_back(FoldFOrdGreaterThan());
1811   rules_[spv::Op::OpFOrdGreaterThan].push_back(
1812       FoldFClampFeedingCompare(spv::Op::OpFOrdGreaterThan));
1813 
1814   rules_[spv::Op::OpFUnordGreaterThan].push_back(FoldFUnordGreaterThan());
1815   rules_[spv::Op::OpFUnordGreaterThan].push_back(
1816       FoldFClampFeedingCompare(spv::Op::OpFUnordGreaterThan));
1817 
1818   rules_[spv::Op::OpFOrdLessThanEqual].push_back(FoldFOrdLessThanEqual());
1819   rules_[spv::Op::OpFOrdLessThanEqual].push_back(
1820       FoldFClampFeedingCompare(spv::Op::OpFOrdLessThanEqual));
1821 
1822   rules_[spv::Op::OpFUnordLessThanEqual].push_back(FoldFUnordLessThanEqual());
1823   rules_[spv::Op::OpFUnordLessThanEqual].push_back(
1824       FoldFClampFeedingCompare(spv::Op::OpFUnordLessThanEqual));
1825 
1826   rules_[spv::Op::OpFOrdGreaterThanEqual].push_back(FoldFOrdGreaterThanEqual());
1827   rules_[spv::Op::OpFOrdGreaterThanEqual].push_back(
1828       FoldFClampFeedingCompare(spv::Op::OpFOrdGreaterThanEqual));
1829 
1830   rules_[spv::Op::OpFUnordGreaterThanEqual].push_back(
1831       FoldFUnordGreaterThanEqual());
1832   rules_[spv::Op::OpFUnordGreaterThanEqual].push_back(
1833       FoldFClampFeedingCompare(spv::Op::OpFUnordGreaterThanEqual));
1834 
1835   rules_[spv::Op::OpVectorShuffle].push_back(FoldVectorShuffleWithConstants());
1836   rules_[spv::Op::OpVectorTimesScalar].push_back(FoldVectorTimesScalar());
1837   rules_[spv::Op::OpVectorTimesMatrix].push_back(FoldVectorTimesMatrix());
1838   rules_[spv::Op::OpMatrixTimesVector].push_back(FoldMatrixTimesVector());
1839   rules_[spv::Op::OpTranspose].push_back(FoldTranspose);
1840 
1841   rules_[spv::Op::OpFNegate].push_back(FoldFNegate());
1842   rules_[spv::Op::OpSNegate].push_back(FoldSNegate());
1843   rules_[spv::Op::OpQuantizeToF16].push_back(FoldQuantizeToF16());
1844 
1845   rules_[spv::Op::OpIAdd].push_back(
1846       FoldBinaryOp(FoldBinaryIntegerOperation<Unsigned>(
1847           [](uint64_t a, uint64_t b) { return a + b; })));
1848   rules_[spv::Op::OpISub].push_back(
1849       FoldBinaryOp(FoldBinaryIntegerOperation<Unsigned>(
1850           [](uint64_t a, uint64_t b) { return a - b; })));
1851   rules_[spv::Op::OpIMul].push_back(
1852       FoldBinaryOp(FoldBinaryIntegerOperation<Unsigned>(
1853           [](uint64_t a, uint64_t b) { return a * b; })));
1854   rules_[spv::Op::OpUDiv].push_back(
1855       FoldBinaryOp(FoldBinaryIntegerOperation<Unsigned>(
1856           [](uint64_t a, uint64_t b) { return (b != 0 ? a / b : 0); })));
1857   rules_[spv::Op::OpSDiv].push_back(FoldBinaryOp(
1858       FoldBinaryIntegerOperation<Signed>([](uint64_t a, uint64_t b) {
1859         return (b != 0 ? static_cast<uint64_t>(static_cast<int64_t>(a) /
1860                                                static_cast<int64_t>(b))
1861                        : 0);
1862       })));
1863   rules_[spv::Op::OpUMod].push_back(
1864       FoldBinaryOp(FoldBinaryIntegerOperation<Unsigned>(
1865           [](uint64_t a, uint64_t b) { return (b != 0 ? a % b : 0); })));
1866 
1867   rules_[spv::Op::OpSRem].push_back(FoldBinaryOp(
1868       FoldBinaryIntegerOperation<Signed>([](uint64_t a, uint64_t b) {
1869         return (b != 0 ? static_cast<uint64_t>(static_cast<int64_t>(a) %
1870                                                static_cast<int64_t>(b))
1871                        : 0);
1872       })));
1873 
1874   rules_[spv::Op::OpSMod].push_back(FoldBinaryOp(
1875       FoldBinaryIntegerOperation<Signed>([](uint64_t a, uint64_t b) {
1876         if (b == 0) return static_cast<uint64_t>(0ull);
1877 
1878         int64_t signed_a = static_cast<int64_t>(a);
1879         int64_t signed_b = static_cast<int64_t>(b);
1880         int64_t result = signed_a % signed_b;
1881         if ((signed_b < 0) != (result < 0)) result += signed_b;
1882         return static_cast<uint64_t>(result);
1883       })));
1884 
1885   // Add rules for GLSLstd450
1886   FeatureManager* feature_manager = context_->get_feature_mgr();
1887   uint32_t ext_inst_glslstd450_id =
1888       feature_manager->GetExtInstImportId_GLSLstd450();
1889   if (ext_inst_glslstd450_id != 0) {
1890     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FMix}].push_back(FoldFMix());
1891     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SMin}].push_back(
1892         FoldFPBinaryOp(FoldMin));
1893     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UMin}].push_back(
1894         FoldFPBinaryOp(FoldMin));
1895     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FMin}].push_back(
1896         FoldFPBinaryOp(FoldMin));
1897     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SMax}].push_back(
1898         FoldFPBinaryOp(FoldMax));
1899     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UMax}].push_back(
1900         FoldFPBinaryOp(FoldMax));
1901     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FMax}].push_back(
1902         FoldFPBinaryOp(FoldMax));
1903     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UClamp}].push_back(
1904         FoldClamp1);
1905     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UClamp}].push_back(
1906         FoldClamp2);
1907     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450UClamp}].push_back(
1908         FoldClamp3);
1909     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SClamp}].push_back(
1910         FoldClamp1);
1911     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SClamp}].push_back(
1912         FoldClamp2);
1913     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450SClamp}].push_back(
1914         FoldClamp3);
1915     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FClamp}].push_back(
1916         FoldClamp1);
1917     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FClamp}].push_back(
1918         FoldClamp2);
1919     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450FClamp}].push_back(
1920         FoldClamp3);
1921     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Sin}].push_back(
1922         FoldFPUnaryOp(FoldFTranscendentalUnary(std::sin)));
1923     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Cos}].push_back(
1924         FoldFPUnaryOp(FoldFTranscendentalUnary(std::cos)));
1925     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Tan}].push_back(
1926         FoldFPUnaryOp(FoldFTranscendentalUnary(std::tan)));
1927     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Asin}].push_back(
1928         FoldFPUnaryOp(FoldFTranscendentalUnary(std::asin)));
1929     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Acos}].push_back(
1930         FoldFPUnaryOp(FoldFTranscendentalUnary(std::acos)));
1931     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Atan}].push_back(
1932         FoldFPUnaryOp(FoldFTranscendentalUnary(std::atan)));
1933     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Exp}].push_back(
1934         FoldFPUnaryOp(FoldFTranscendentalUnary(std::exp)));
1935     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Log}].push_back(
1936         FoldFPUnaryOp(FoldFTranscendentalUnary(std::log)));
1937 
1938 #ifdef __ANDROID__
1939     // Android NDK r15c targeting ABI 15 doesn't have full support for C++11
1940     // (no std::exp2/log2). ::exp2 is available from C99 but ::log2 isn't
1941     // available up until ABI 18 so we use a shim
1942     auto log2_shim = [](double v) -> double { return log(v) / log(2.0); };
1943     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Exp2}].push_back(
1944         FoldFPUnaryOp(FoldFTranscendentalUnary(::exp2)));
1945     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Log2}].push_back(
1946         FoldFPUnaryOp(FoldFTranscendentalUnary(log2_shim)));
1947 #else
1948     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Exp2}].push_back(
1949         FoldFPUnaryOp(FoldFTranscendentalUnary(std::exp2)));
1950     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Log2}].push_back(
1951         FoldFPUnaryOp(FoldFTranscendentalUnary(std::log2)));
1952 #endif
1953 
1954     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Sqrt}].push_back(
1955         FoldFPUnaryOp(FoldFTranscendentalUnary(std::sqrt)));
1956     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Atan2}].push_back(
1957         FoldFPBinaryOp(FoldFTranscendentalBinary(std::atan2)));
1958     ext_rules_[{ext_inst_glslstd450_id, GLSLstd450Pow}].push_back(
1959         FoldFPBinaryOp(FoldFTranscendentalBinary(std::pow)));
1960   }
1961 }
1962 }  // namespace opt
1963 }  // namespace spvtools
1964