| 1 | //===- InferAlignment.cpp -------------------------------------------------===// |
| 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 | // Infer alignment for load, stores and other memory operations based on |
| 10 | // trailing zero known bits information. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "llvm/Transforms/Scalar/InferAlignment.h" |
| 15 | #include "llvm/ADT/APInt.h" |
| 16 | #include "llvm/ADT/STLFunctionalExtras.h" |
| 17 | #include "llvm/Analysis/AssumptionCache.h" |
| 18 | #include "llvm/Analysis/ValueTracking.h" |
| 19 | #include "llvm/IR/Instruction.h" |
| 20 | #include "llvm/IR/Instructions.h" |
| 21 | #include "llvm/IR/IntrinsicInst.h" |
| 22 | #include "llvm/IR/PatternMatch.h" |
| 23 | #include "llvm/Support/KnownBits.h" |
| 24 | #include "llvm/Transforms/Scalar.h" |
| 25 | #include "llvm/Transforms/Utils/Local.h" |
| 26 | |
| 27 | using namespace llvm; |
| 28 | using namespace llvm::PatternMatch; |
| 29 | |
| 30 | static bool tryToImproveAlign( |
| 31 | const DataLayout &DL, Instruction *I, |
| 32 | function_ref<Align(Value *PtrOp, Align OldAlign, Align PrefAlign)> Fn) { |
| 33 | |
| 34 | if (auto *PtrOp = getLoadStorePointerOperand(V: I)) { |
| 35 | Align OldAlign = getLoadStoreAlignment(I); |
| 36 | Align PrefAlign = DL.getPrefTypeAlign(Ty: getLoadStoreType(I)); |
| 37 | |
| 38 | Align NewAlign = Fn(PtrOp, OldAlign, PrefAlign); |
| 39 | if (NewAlign > OldAlign) { |
| 40 | setLoadStoreAlignment(I, NewAlign); |
| 41 | return true; |
| 42 | } |
| 43 | } |
| 44 | |
| 45 | Value *PtrOp; |
| 46 | const APInt *Const; |
| 47 | if (match(V: I, P: m_And(L: m_PtrToIntOrAddr(Op: m_Value(V&: PtrOp)), R: m_APInt(Res&: Const)))) { |
| 48 | Align ActualAlign = Fn(PtrOp, Align(1), Align(1)); |
| 49 | if (Const->ult(RHS: ActualAlign.value())) { |
| 50 | I->replaceAllUsesWith(V: Constant::getNullValue(Ty: I->getType())); |
| 51 | return true; |
| 52 | } |
| 53 | if (Const->uge( |
| 54 | RHS: APInt::getBitsSetFrom(numBits: Const->getBitWidth(), loBit: Log2(A: ActualAlign)))) { |
| 55 | I->replaceAllUsesWith(V: I->getOperand(i: 0)); |
| 56 | return true; |
| 57 | } |
| 58 | } |
| 59 | if (match(V: I, P: m_Trunc(Op: m_PtrToIntOrAddr(Op: m_Value(V&: PtrOp))))) { |
| 60 | Align ActualAlign = Fn(PtrOp, Align(1), Align(1)); |
| 61 | if (Log2(A: ActualAlign) >= I->getType()->getScalarSizeInBits()) { |
| 62 | I->replaceAllUsesWith(V: Constant::getNullValue(Ty: I->getType())); |
| 63 | return true; |
| 64 | } |
| 65 | } |
| 66 | |
| 67 | IntrinsicInst *II = dyn_cast<IntrinsicInst>(Val: I); |
| 68 | if (!II) |
| 69 | return false; |
| 70 | |
| 71 | // TODO: Handle more memory intrinsics. |
| 72 | switch (II->getIntrinsicID()) { |
| 73 | case Intrinsic::masked_load: |
| 74 | case Intrinsic::masked_store: { |
| 75 | unsigned PtrOpIdx = II->getIntrinsicID() == Intrinsic::masked_load ? 0 : 1; |
| 76 | Value *PtrOp = II->getArgOperand(i: PtrOpIdx); |
| 77 | Type *Type = II->getIntrinsicID() == Intrinsic::masked_load |
| 78 | ? II->getType() |
| 79 | : II->getArgOperand(i: 0)->getType(); |
| 80 | |
| 81 | Align OldAlign = II->getParamAlign(ArgNo: PtrOpIdx).valueOrOne(); |
| 82 | Align PrefAlign = DL.getPrefTypeAlign(Ty: Type); |
| 83 | Align NewAlign = Fn(PtrOp, OldAlign, PrefAlign); |
| 84 | if (NewAlign <= OldAlign) |
| 85 | return false; |
| 86 | |
| 87 | II->addParamAttr(ArgNo: PtrOpIdx, |
| 88 | Attr: Attribute::getWithAlignment(Context&: II->getContext(), Alignment: NewAlign)); |
| 89 | return true; |
| 90 | } |
| 91 | default: |
| 92 | return false; |
| 93 | } |
| 94 | } |
| 95 | |
| 96 | bool inferAlignment(Function &F, AssumptionCache &AC, DominatorTree &DT) { |
| 97 | const DataLayout &DL = F.getDataLayout(); |
| 98 | bool Changed = false; |
| 99 | |
| 100 | // Enforce preferred type alignment if possible. We do this as a separate |
| 101 | // pass first, because it may improve the alignments we infer below. |
| 102 | for (BasicBlock &BB : F) { |
| 103 | for (Instruction &I : BB) { |
| 104 | Changed |= tryToImproveAlign( |
| 105 | DL, I: &I, Fn: [&](Value *PtrOp, Align OldAlign, Align PrefAlign) { |
| 106 | if (PrefAlign > OldAlign) |
| 107 | return std::max(a: OldAlign, |
| 108 | b: tryEnforceAlignment(V: PtrOp, PrefAlign, DL)); |
| 109 | return OldAlign; |
| 110 | }); |
| 111 | } |
| 112 | } |
| 113 | |
| 114 | // Compute alignment from known bits. |
| 115 | auto InferFromKnownBits = [&](Instruction &I, Value *PtrOp) { |
| 116 | KnownBits Known = computeKnownBits(V: PtrOp, DL, AC: &AC, CxtI: &I, DT: &DT); |
| 117 | unsigned TrailZ = |
| 118 | std::min(a: Known.countMinTrailingZeros(), b: +Value::MaxAlignmentExponent); |
| 119 | return Align(1ull << std::min(a: Known.getBitWidth() - 1, b: TrailZ)); |
| 120 | }; |
| 121 | |
| 122 | // Propagate alignment between loads and stores that originate from the |
| 123 | // same base pointer. |
| 124 | DenseMap<Value *, Align> BestBasePointerAligns; |
| 125 | auto InferFromBasePointer = [&](Value *PtrOp, Align LoadStoreAlign) { |
| 126 | APInt OffsetFromBase(DL.getIndexTypeSizeInBits(Ty: PtrOp->getType()), 0); |
| 127 | PtrOp = PtrOp->stripAndAccumulateConstantOffsets(DL, Offset&: OffsetFromBase, AllowNonInbounds: true); |
| 128 | // Derive the base pointer alignment from the load/store alignment |
| 129 | // and the offset from the base pointer. |
| 130 | Align BasePointerAlign = |
| 131 | commonAlignment(A: LoadStoreAlign, Offset: OffsetFromBase.getLimitedValue()); |
| 132 | |
| 133 | auto [It, Inserted] = |
| 134 | BestBasePointerAligns.try_emplace(Key: PtrOp, Args&: BasePointerAlign); |
| 135 | if (!Inserted) { |
| 136 | // If the stored base pointer alignment is better than the |
| 137 | // base pointer alignment we derived, we may be able to use it |
| 138 | // to improve the load/store alignment. If not, store the |
| 139 | // improved base pointer alignment for future iterations. |
| 140 | if (It->second > BasePointerAlign) { |
| 141 | Align BetterLoadStoreAlign = |
| 142 | commonAlignment(A: It->second, Offset: OffsetFromBase.getLimitedValue()); |
| 143 | return BetterLoadStoreAlign; |
| 144 | } |
| 145 | It->second = BasePointerAlign; |
| 146 | } |
| 147 | return LoadStoreAlign; |
| 148 | }; |
| 149 | |
| 150 | for (BasicBlock &BB : F) { |
| 151 | // We need to reset the map for each block because alignment information |
| 152 | // can only be propagated from instruction A to B if A dominates B. |
| 153 | // This is because control flow (and exception throwing) could be dependent |
| 154 | // on the address (and its alignment) at runtime. Some sort of dominator |
| 155 | // tree approach could be better, but doing a simple forward pass through a |
| 156 | // single basic block is correct too. |
| 157 | BestBasePointerAligns.clear(); |
| 158 | |
| 159 | for (Instruction &I : BB) { |
| 160 | Changed |= tryToImproveAlign( |
| 161 | DL, I: &I, Fn: [&](Value *PtrOp, Align OldAlign, Align PrefAlign) { |
| 162 | return std::max(a: InferFromKnownBits(I, PtrOp), |
| 163 | b: InferFromBasePointer(PtrOp, OldAlign)); |
| 164 | }); |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | return Changed; |
| 169 | } |
| 170 | |
| 171 | PreservedAnalyses InferAlignmentPass::run(Function &F, |
| 172 | FunctionAnalysisManager &AM) { |
| 173 | AssumptionCache &AC = AM.getResult<AssumptionAnalysis>(IR&: F); |
| 174 | DominatorTree &DT = AM.getResult<DominatorTreeAnalysis>(IR&: F); |
| 175 | inferAlignment(F, AC, DT); |
| 176 | // Changes to alignment shouldn't invalidated analyses. |
| 177 | return PreservedAnalyses::all(); |
| 178 | } |
| 179 | |