[RISCV] Fix mgather -> riscv.masked.strided.load combine not extending indices (...
[llvm-project.git] / llvm / lib / IR / Operator.cpp
blob16a89534b4b3ec62ffd9fcaf2fc8d537371d3708
1 //===-- Operator.cpp - Implement the LLVM operators -----------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the non-inline methods for the LLVM Operator classes.
11 //===----------------------------------------------------------------------===//
13 #include "llvm/IR/Operator.h"
14 #include "llvm/IR/DataLayout.h"
15 #include "llvm/IR/GetElementPtrTypeIterator.h"
16 #include "llvm/IR/Instructions.h"
18 #include "ConstantsContext.h"
20 namespace llvm {
21 bool Operator::hasPoisonGeneratingFlags() const {
22 switch (getOpcode()) {
23 case Instruction::Add:
24 case Instruction::Sub:
25 case Instruction::Mul:
26 case Instruction::Shl: {
27 auto *OBO = cast<OverflowingBinaryOperator>(this);
28 return OBO->hasNoUnsignedWrap() || OBO->hasNoSignedWrap();
30 case Instruction::UDiv:
31 case Instruction::SDiv:
32 case Instruction::AShr:
33 case Instruction::LShr:
34 return cast<PossiblyExactOperator>(this)->isExact();
35 case Instruction::Or:
36 return cast<PossiblyDisjointInst>(this)->isDisjoint();
37 case Instruction::GetElementPtr: {
38 auto *GEP = cast<GEPOperator>(this);
39 // Note: inrange exists on constexpr only
40 return GEP->isInBounds() || GEP->getInRangeIndex() != std::nullopt;
42 case Instruction::ZExt:
43 if (auto *NNI = dyn_cast<PossiblyNonNegInst>(this))
44 return NNI->hasNonNeg();
45 return false;
46 default:
47 if (const auto *FP = dyn_cast<FPMathOperator>(this))
48 return FP->hasNoNaNs() || FP->hasNoInfs();
49 return false;
53 bool Operator::hasPoisonGeneratingFlagsOrMetadata() const {
54 if (hasPoisonGeneratingFlags())
55 return true;
56 auto *I = dyn_cast<Instruction>(this);
57 return I && I->hasPoisonGeneratingMetadata();
60 Type *GEPOperator::getSourceElementType() const {
61 if (auto *I = dyn_cast<GetElementPtrInst>(this))
62 return I->getSourceElementType();
63 return cast<GetElementPtrConstantExpr>(this)->getSourceElementType();
66 Type *GEPOperator::getResultElementType() const {
67 if (auto *I = dyn_cast<GetElementPtrInst>(this))
68 return I->getResultElementType();
69 return cast<GetElementPtrConstantExpr>(this)->getResultElementType();
72 Align GEPOperator::getMaxPreservedAlignment(const DataLayout &DL) const {
73 /// compute the worse possible offset for every level of the GEP et accumulate
74 /// the minimum alignment into Result.
76 Align Result = Align(llvm::Value::MaximumAlignment);
77 for (gep_type_iterator GTI = gep_type_begin(this), GTE = gep_type_end(this);
78 GTI != GTE; ++GTI) {
79 uint64_t Offset;
80 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
82 if (StructType *STy = GTI.getStructTypeOrNull()) {
83 const StructLayout *SL = DL.getStructLayout(STy);
84 Offset = SL->getElementOffset(OpC->getZExtValue());
85 } else {
86 assert(GTI.isSequential() && "should be sequencial");
87 /// If the index isn't known, we take 1 because it is the index that will
88 /// give the worse alignment of the offset.
89 const uint64_t ElemCount = OpC ? OpC->getZExtValue() : 1;
90 Offset = GTI.getSequentialElementStride(DL) * ElemCount;
92 Result = Align(MinAlign(Offset, Result.value()));
94 return Result;
97 bool GEPOperator::accumulateConstantOffset(
98 const DataLayout &DL, APInt &Offset,
99 function_ref<bool(Value &, APInt &)> ExternalAnalysis) const {
100 assert(Offset.getBitWidth() ==
101 DL.getIndexSizeInBits(getPointerAddressSpace()) &&
102 "The offset bit width does not match DL specification.");
103 SmallVector<const Value *> Index(llvm::drop_begin(operand_values()));
104 return GEPOperator::accumulateConstantOffset(getSourceElementType(), Index,
105 DL, Offset, ExternalAnalysis);
108 bool GEPOperator::accumulateConstantOffset(
109 Type *SourceType, ArrayRef<const Value *> Index, const DataLayout &DL,
110 APInt &Offset, function_ref<bool(Value &, APInt &)> ExternalAnalysis) {
111 bool UsedExternalAnalysis = false;
112 auto AccumulateOffset = [&](APInt Index, uint64_t Size) -> bool {
113 Index = Index.sextOrTrunc(Offset.getBitWidth());
114 APInt IndexedSize = APInt(Offset.getBitWidth(), Size);
115 // For array or vector indices, scale the index by the size of the type.
116 if (!UsedExternalAnalysis) {
117 Offset += Index * IndexedSize;
118 } else {
119 // External Analysis can return a result higher/lower than the value
120 // represents. We need to detect overflow/underflow.
121 bool Overflow = false;
122 APInt OffsetPlus = Index.smul_ov(IndexedSize, Overflow);
123 if (Overflow)
124 return false;
125 Offset = Offset.sadd_ov(OffsetPlus, Overflow);
126 if (Overflow)
127 return false;
129 return true;
131 auto begin = generic_gep_type_iterator<decltype(Index.begin())>::begin(
132 SourceType, Index.begin());
133 auto end = generic_gep_type_iterator<decltype(Index.end())>::end(Index.end());
134 for (auto GTI = begin, GTE = end; GTI != GTE; ++GTI) {
135 // Scalable vectors are multiplied by a runtime constant.
136 bool ScalableType = GTI.getIndexedType()->isScalableTy();
138 Value *V = GTI.getOperand();
139 StructType *STy = GTI.getStructTypeOrNull();
140 // Handle ConstantInt if possible.
141 if (auto ConstOffset = dyn_cast<ConstantInt>(V)) {
142 if (ConstOffset->isZero())
143 continue;
144 // if the type is scalable and the constant is not zero (vscale * n * 0 =
145 // 0) bailout.
146 if (ScalableType)
147 return false;
148 // Handle a struct index, which adds its field offset to the pointer.
149 if (STy) {
150 unsigned ElementIdx = ConstOffset->getZExtValue();
151 const StructLayout *SL = DL.getStructLayout(STy);
152 // Element offset is in bytes.
153 if (!AccumulateOffset(
154 APInt(Offset.getBitWidth(), SL->getElementOffset(ElementIdx)),
156 return false;
157 continue;
159 if (!AccumulateOffset(ConstOffset->getValue(),
160 GTI.getSequentialElementStride(DL)))
161 return false;
162 continue;
165 // The operand is not constant, check if an external analysis was provided.
166 // External analsis is not applicable to a struct type.
167 if (!ExternalAnalysis || STy || ScalableType)
168 return false;
169 APInt AnalysisIndex;
170 if (!ExternalAnalysis(*V, AnalysisIndex))
171 return false;
172 UsedExternalAnalysis = true;
173 if (!AccumulateOffset(AnalysisIndex, GTI.getSequentialElementStride(DL)))
174 return false;
176 return true;
179 bool GEPOperator::collectOffset(
180 const DataLayout &DL, unsigned BitWidth,
181 MapVector<Value *, APInt> &VariableOffsets,
182 APInt &ConstantOffset) const {
183 assert(BitWidth == DL.getIndexSizeInBits(getPointerAddressSpace()) &&
184 "The offset bit width does not match DL specification.");
186 auto CollectConstantOffset = [&](APInt Index, uint64_t Size) {
187 Index = Index.sextOrTrunc(BitWidth);
188 APInt IndexedSize = APInt(BitWidth, Size);
189 ConstantOffset += Index * IndexedSize;
192 for (gep_type_iterator GTI = gep_type_begin(this), GTE = gep_type_end(this);
193 GTI != GTE; ++GTI) {
194 // Scalable vectors are multiplied by a runtime constant.
195 bool ScalableType = GTI.getIndexedType()->isScalableTy();
197 Value *V = GTI.getOperand();
198 StructType *STy = GTI.getStructTypeOrNull();
199 // Handle ConstantInt if possible.
200 if (auto ConstOffset = dyn_cast<ConstantInt>(V)) {
201 if (ConstOffset->isZero())
202 continue;
203 // If the type is scalable and the constant is not zero (vscale * n * 0 =
204 // 0) bailout.
205 // TODO: If the runtime value is accessible at any point before DWARF
206 // emission, then we could potentially keep a forward reference to it
207 // in the debug value to be filled in later.
208 if (ScalableType)
209 return false;
210 // Handle a struct index, which adds its field offset to the pointer.
211 if (STy) {
212 unsigned ElementIdx = ConstOffset->getZExtValue();
213 const StructLayout *SL = DL.getStructLayout(STy);
214 // Element offset is in bytes.
215 CollectConstantOffset(APInt(BitWidth, SL->getElementOffset(ElementIdx)),
217 continue;
219 CollectConstantOffset(ConstOffset->getValue(),
220 GTI.getSequentialElementStride(DL));
221 continue;
224 if (STy || ScalableType)
225 return false;
226 APInt IndexedSize = APInt(BitWidth, GTI.getSequentialElementStride(DL));
227 // Insert an initial offset of 0 for V iff none exists already, then
228 // increment the offset by IndexedSize.
229 if (!IndexedSize.isZero()) {
230 auto *It = VariableOffsets.insert({V, APInt(BitWidth, 0)}).first;
231 It->second += IndexedSize;
234 return true;
237 void FastMathFlags::print(raw_ostream &O) const {
238 if (all())
239 O << " fast";
240 else {
241 if (allowReassoc())
242 O << " reassoc";
243 if (noNaNs())
244 O << " nnan";
245 if (noInfs())
246 O << " ninf";
247 if (noSignedZeros())
248 O << " nsz";
249 if (allowReciprocal())
250 O << " arcp";
251 if (allowContract())
252 O << " contract";
253 if (approxFunc())
254 O << " afn";
257 } // namespace llvm