1 //===- AggressiveInstCombine.cpp ------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the aggressive expression pattern combiner classes.
11 // Currently, it handles expression patterns for:
12 // * Truncate instruction
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Transforms/AggressiveInstCombine/AggressiveInstCombine.h"
17 #include "AggressiveInstCombineInternal.h"
18 #include "llvm-c/Initialization.h"
19 #include "llvm-c/Transforms/AggressiveInstCombine.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/BasicAliasAnalysis.h"
22 #include "llvm/Analysis/GlobalsModRef.h"
23 #include "llvm/Analysis/TargetLibraryInfo.h"
24 #include "llvm/IR/DataLayout.h"
25 #include "llvm/IR/Dominators.h"
26 #include "llvm/IR/IRBuilder.h"
27 #include "llvm/IR/LegacyPassManager.h"
28 #include "llvm/IR/PatternMatch.h"
29 #include "llvm/Pass.h"
30 #include "llvm/Transforms/Utils/Local.h"
32 using namespace PatternMatch;
34 #define DEBUG_TYPE "aggressive-instcombine"
37 /// Contains expression pattern combiner logic.
38 /// This class provides both the logic to combine expression patterns and
39 /// combine them. It differs from InstCombiner class in that each pattern
40 /// combiner runs only once as opposed to InstCombine's multi-iteration,
41 /// which allows pattern combiner to have higher complexity than the O(1)
42 /// required by the instruction combiner.
43 class AggressiveInstCombinerLegacyPass : public FunctionPass {
45 static char ID; // Pass identification, replacement for typeid
47 AggressiveInstCombinerLegacyPass() : FunctionPass(ID) {
48 initializeAggressiveInstCombinerLegacyPassPass(
49 *PassRegistry::getPassRegistry());
52 void getAnalysisUsage(AnalysisUsage &AU) const override;
54 /// Run all expression pattern optimizations on the given /p F function.
56 /// \param F function to optimize.
57 /// \returns true if the IR is changed.
58 bool runOnFunction(Function &F) override;
62 /// Match a pattern for a bitwise rotate operation that partially guards
63 /// against undefined behavior by branching around the rotation when the shift
65 static bool foldGuardedRotateToFunnelShift(Instruction &I) {
66 if (I.getOpcode() != Instruction::PHI || I.getNumOperands() != 2)
69 // As with the one-use checks below, this is not strictly necessary, but we
70 // are being cautious to avoid potential perf regressions on targets that
71 // do not actually have a rotate instruction (where the funnel shift would be
72 // expanded back into math/shift/logic ops).
73 if (!isPowerOf2_32(I.getType()->getScalarSizeInBits()))
76 // Match V to funnel shift left/right and capture the source operand and
77 // shift amount in X and Y.
78 auto matchRotate = [](Value *V, Value *&X, Value *&Y) {
79 Value *L0, *L1, *R0, *R1;
80 unsigned Width = V->getType()->getScalarSizeInBits();
81 auto Sub = m_Sub(m_SpecificInt(Width), m_Value(R1));
83 // rotate_left(X, Y) == (X << Y) | (X >> (Width - Y))
85 m_c_Or(m_Shl(m_Value(L0), m_Value(L1)), m_LShr(m_Value(R0), Sub)));
86 if (RotL.match(V) && L0 == R0 && L1 == R1) {
89 return Intrinsic::fshl;
92 // rotate_right(X, Y) == (X >> Y) | (X << (Width - Y))
94 m_c_Or(m_LShr(m_Value(L0), m_Value(L1)), m_Shl(m_Value(R0), Sub)));
95 if (RotR.match(V) && L0 == R0 && L1 == R1) {
98 return Intrinsic::fshr;
101 return Intrinsic::not_intrinsic;
104 // One phi operand must be a rotate operation, and the other phi operand must
105 // be the source value of that rotate operation:
106 // phi [ rotate(RotSrc, RotAmt), RotBB ], [ RotSrc, GuardBB ]
107 PHINode &Phi = cast<PHINode>(I);
108 Value *P0 = Phi.getOperand(0), *P1 = Phi.getOperand(1);
109 Value *RotSrc, *RotAmt;
110 Intrinsic::ID IID = matchRotate(P0, RotSrc, RotAmt);
111 if (IID == Intrinsic::not_intrinsic || RotSrc != P1) {
112 IID = matchRotate(P1, RotSrc, RotAmt);
113 if (IID == Intrinsic::not_intrinsic || RotSrc != P0)
115 assert((IID == Intrinsic::fshl || IID == Intrinsic::fshr) &&
116 "Pattern must match funnel shift left or right");
119 // The incoming block with our source operand must be the "guard" block.
120 // That must contain a cmp+branch to avoid the rotate when the shift amount
121 // is equal to 0. The other incoming block is the block with the rotate.
122 BasicBlock *GuardBB = Phi.getIncomingBlock(RotSrc == P1);
123 BasicBlock *RotBB = Phi.getIncomingBlock(RotSrc != P1);
124 Instruction *TermI = GuardBB->getTerminator();
125 BasicBlock *TrueBB, *FalseBB;
126 ICmpInst::Predicate Pred;
127 if (!match(TermI, m_Br(m_ICmp(Pred, m_Specific(RotAmt), m_ZeroInt()), TrueBB,
131 BasicBlock *PhiBB = Phi.getParent();
132 if (Pred != CmpInst::ICMP_EQ || TrueBB != PhiBB || FalseBB != RotBB)
135 // We matched a variation of this IR pattern:
137 // %cmp = icmp eq i32 %RotAmt, 0
138 // br i1 %cmp, label %PhiBB, label %RotBB
140 // %sub = sub i32 32, %RotAmt
141 // %shr = lshr i32 %X, %sub
142 // %shl = shl i32 %X, %RotAmt
143 // %rot = or i32 %shr, %shl
146 // %cond = phi i32 [ %rot, %RotBB ], [ %X, %GuardBB ]
148 // llvm.fshl.i32(i32 %X, i32 %RotAmt)
149 IRBuilder<> Builder(PhiBB, PhiBB->getFirstInsertionPt());
150 Function *F = Intrinsic::getDeclaration(Phi.getModule(), IID, Phi.getType());
151 Phi.replaceAllUsesWith(Builder.CreateCall(F, {RotSrc, RotSrc, RotAmt}));
155 /// This is used by foldAnyOrAllBitsSet() to capture a source value (Root) and
156 /// the bit indexes (Mask) needed by a masked compare. If we're matching a chain
157 /// of 'and' ops, then we also need to capture the fact that we saw an
158 /// "and X, 1", so that's an extra return value for that case.
165 MaskOps(unsigned BitWidth, bool MatchAnds)
166 : Root(nullptr), Mask(APInt::getNullValue(BitWidth)),
167 MatchAndChain(MatchAnds), FoundAnd1(false) {}
170 /// This is a recursive helper for foldAnyOrAllBitsSet() that walks through a
171 /// chain of 'and' or 'or' instructions looking for shift ops of a common source
173 /// or (or (or X, (X >> 3)), (X >> 5)), (X >> 8)
174 /// returns { X, 0x129 }
175 /// and (and (X >> 1), 1), (X >> 4)
176 /// returns { X, 0x12 }
177 static bool matchAndOrChain(Value *V, MaskOps &MOps) {
179 if (MOps.MatchAndChain) {
180 // Recurse through a chain of 'and' operands. This requires an extra check
181 // vs. the 'or' matcher: we must find an "and X, 1" instruction somewhere
182 // in the chain to know that all of the high bits are cleared.
183 if (match(V, m_And(m_Value(Op0), m_One()))) {
184 MOps.FoundAnd1 = true;
185 return matchAndOrChain(Op0, MOps);
187 if (match(V, m_And(m_Value(Op0), m_Value(Op1))))
188 return matchAndOrChain(Op0, MOps) && matchAndOrChain(Op1, MOps);
190 // Recurse through a chain of 'or' operands.
191 if (match(V, m_Or(m_Value(Op0), m_Value(Op1))))
192 return matchAndOrChain(Op0, MOps) && matchAndOrChain(Op1, MOps);
195 // We need a shift-right or a bare value representing a compare of bit 0 of
196 // the original source operand.
198 uint64_t BitIndex = 0;
199 if (!match(V, m_LShr(m_Value(Candidate), m_ConstantInt(BitIndex))))
202 // Initialize result source operand.
204 MOps.Root = Candidate;
206 // The shift constant is out-of-range? This code hasn't been simplified.
207 if (BitIndex >= MOps.Mask.getBitWidth())
210 // Fill in the mask bit derived from the shift constant.
211 MOps.Mask.setBit(BitIndex);
212 return MOps.Root == Candidate;
215 /// Match patterns that correspond to "any-bits-set" and "all-bits-set".
216 /// These will include a chain of 'or' or 'and'-shifted bits from a
217 /// common source value:
218 /// and (or (lshr X, C), ...), 1 --> (X & CMask) != 0
219 /// and (and (lshr X, C), ...), 1 --> (X & CMask) == CMask
220 /// Note: "any-bits-clear" and "all-bits-clear" are variations of these patterns
221 /// that differ only with a final 'not' of the result. We expect that final
222 /// 'not' to be folded with the compare that we create here (invert predicate).
223 static bool foldAnyOrAllBitsSet(Instruction &I) {
224 // The 'any-bits-set' ('or' chain) pattern is simpler to match because the
225 // final "and X, 1" instruction must be the final op in the sequence.
226 bool MatchAllBitsSet;
227 if (match(&I, m_c_And(m_OneUse(m_And(m_Value(), m_Value())), m_Value())))
228 MatchAllBitsSet = true;
229 else if (match(&I, m_And(m_OneUse(m_Or(m_Value(), m_Value())), m_One())))
230 MatchAllBitsSet = false;
234 MaskOps MOps(I.getType()->getScalarSizeInBits(), MatchAllBitsSet);
235 if (MatchAllBitsSet) {
236 if (!matchAndOrChain(cast<BinaryOperator>(&I), MOps) || !MOps.FoundAnd1)
239 if (!matchAndOrChain(cast<BinaryOperator>(&I)->getOperand(0), MOps))
243 // The pattern was found. Create a masked compare that replaces all of the
244 // shift and logic ops.
245 IRBuilder<> Builder(&I);
246 Constant *Mask = ConstantInt::get(I.getType(), MOps.Mask);
247 Value *And = Builder.CreateAnd(MOps.Root, Mask);
248 Value *Cmp = MatchAllBitsSet ? Builder.CreateICmpEQ(And, Mask)
249 : Builder.CreateIsNotNull(And);
250 Value *Zext = Builder.CreateZExt(Cmp, I.getType());
251 I.replaceAllUsesWith(Zext);
255 /// This is the entry point for folds that could be implemented in regular
256 /// InstCombine, but they are separated because they are not expected to
257 /// occur frequently and/or have more than a constant-length pattern match.
258 static bool foldUnusualPatterns(Function &F, DominatorTree &DT) {
259 bool MadeChange = false;
260 for (BasicBlock &BB : F) {
261 // Ignore unreachable basic blocks.
262 if (!DT.isReachableFromEntry(&BB))
264 // Do not delete instructions under here and invalidate the iterator.
265 // Walk the block backwards for efficiency. We're matching a chain of
266 // use->defs, so we're more likely to succeed by starting from the bottom.
267 // Also, we want to avoid matching partial patterns.
268 // TODO: It would be more efficient if we removed dead instructions
269 // iteratively in this loop rather than waiting until the end.
270 for (Instruction &I : make_range(BB.rbegin(), BB.rend())) {
271 MadeChange |= foldAnyOrAllBitsSet(I);
272 MadeChange |= foldGuardedRotateToFunnelShift(I);
276 // We're done with transforms, so remove dead instructions.
278 for (BasicBlock &BB : F)
279 SimplifyInstructionsInBlock(&BB);
284 /// This is the entry point for all transforms. Pass manager differences are
285 /// handled in the callers of this function.
286 static bool runImpl(Function &F, TargetLibraryInfo &TLI, DominatorTree &DT) {
287 bool MadeChange = false;
288 const DataLayout &DL = F.getParent()->getDataLayout();
289 TruncInstCombine TIC(TLI, DL, DT);
290 MadeChange |= TIC.run(F);
291 MadeChange |= foldUnusualPatterns(F, DT);
295 void AggressiveInstCombinerLegacyPass::getAnalysisUsage(
296 AnalysisUsage &AU) const {
297 AU.setPreservesCFG();
298 AU.addRequired<DominatorTreeWrapperPass>();
299 AU.addRequired<TargetLibraryInfoWrapperPass>();
300 AU.addPreserved<AAResultsWrapperPass>();
301 AU.addPreserved<BasicAAWrapperPass>();
302 AU.addPreserved<DominatorTreeWrapperPass>();
303 AU.addPreserved<GlobalsAAWrapperPass>();
306 bool AggressiveInstCombinerLegacyPass::runOnFunction(Function &F) {
307 auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
308 auto &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
309 return runImpl(F, TLI, DT);
312 PreservedAnalyses AggressiveInstCombinePass::run(Function &F,
313 FunctionAnalysisManager &AM) {
314 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
315 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
316 if (!runImpl(F, TLI, DT)) {
317 // No changes, all analyses are preserved.
318 return PreservedAnalyses::all();
320 // Mark all the analyses that instcombine updates as preserved.
321 PreservedAnalyses PA;
322 PA.preserveSet<CFGAnalyses>();
323 PA.preserve<AAManager>();
324 PA.preserve<GlobalsAA>();
328 char AggressiveInstCombinerLegacyPass::ID = 0;
329 INITIALIZE_PASS_BEGIN(AggressiveInstCombinerLegacyPass,
330 "aggressive-instcombine",
331 "Combine pattern based expressions", false, false)
332 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
333 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
334 INITIALIZE_PASS_END(AggressiveInstCombinerLegacyPass, "aggressive-instcombine",
335 "Combine pattern based expressions", false, false)
337 // Initialization Routines
338 void llvm::initializeAggressiveInstCombine(PassRegistry &Registry) {
339 initializeAggressiveInstCombinerLegacyPassPass(Registry);
342 void LLVMInitializeAggressiveInstCombiner(LLVMPassRegistryRef R) {
343 initializeAggressiveInstCombinerLegacyPassPass(*unwrap(R));
346 FunctionPass *llvm::createAggressiveInstCombinerPass() {
347 return new AggressiveInstCombinerLegacyPass();
350 void LLVMAddAggressiveInstCombinerPass(LLVMPassManagerRef PM) {
351 unwrap(PM)->add(createAggressiveInstCombinerPass());