| 1 | //===- CorrelatedValuePropagation.cpp - Propagate CFG-derived info --------===// |
| 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 Correlated Value Propagation pass. |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "llvm/Transforms/Scalar/CorrelatedValuePropagation.h" |
| 14 | #include "llvm/ADT/DepthFirstIterator.h" |
| 15 | #include "llvm/ADT/SmallVector.h" |
| 16 | #include "llvm/ADT/Statistic.h" |
| 17 | #include "llvm/Analysis/DomTreeUpdater.h" |
| 18 | #include "llvm/Analysis/GlobalsModRef.h" |
| 19 | #include "llvm/Analysis/InstructionSimplify.h" |
| 20 | #include "llvm/Analysis/LazyValueInfo.h" |
| 21 | #include "llvm/Analysis/ValueTracking.h" |
| 22 | #include "llvm/IR/Attributes.h" |
| 23 | #include "llvm/IR/BasicBlock.h" |
| 24 | #include "llvm/IR/CFG.h" |
| 25 | #include "llvm/IR/Constant.h" |
| 26 | #include "llvm/IR/ConstantRange.h" |
| 27 | #include "llvm/IR/Constants.h" |
| 28 | #include "llvm/IR/DebugInfo.h" |
| 29 | #include "llvm/IR/DerivedTypes.h" |
| 30 | #include "llvm/IR/Function.h" |
| 31 | #include "llvm/IR/IRBuilder.h" |
| 32 | #include "llvm/IR/InstrTypes.h" |
| 33 | #include "llvm/IR/Instruction.h" |
| 34 | #include "llvm/IR/Instructions.h" |
| 35 | #include "llvm/IR/IntrinsicInst.h" |
| 36 | #include "llvm/IR/MDBuilder.h" |
| 37 | #include "llvm/IR/Operator.h" |
| 38 | #include "llvm/IR/PassManager.h" |
| 39 | #include "llvm/IR/PatternMatch.h" |
| 40 | #include "llvm/IR/Type.h" |
| 41 | #include "llvm/IR/Value.h" |
| 42 | #include "llvm/Support/Casting.h" |
| 43 | #include "llvm/Transforms/Utils/Local.h" |
| 44 | #include <cassert> |
| 45 | #include <optional> |
| 46 | #include <utility> |
| 47 | |
| 48 | using namespace llvm; |
| 49 | |
| 50 | #define DEBUG_TYPE "correlated-value-propagation" |
| 51 | |
| 52 | STATISTIC(NumPhis, "Number of phis propagated" ); |
| 53 | STATISTIC(NumPhiCommon, "Number of phis deleted via common incoming value" ); |
| 54 | STATISTIC(NumSelects, "Number of selects propagated" ); |
| 55 | STATISTIC(NumCmps, "Number of comparisons propagated" ); |
| 56 | STATISTIC(NumReturns, "Number of return values propagated" ); |
| 57 | STATISTIC(NumDeadCases, "Number of switch cases removed" ); |
| 58 | STATISTIC(NumSDivSRemsNarrowed, |
| 59 | "Number of sdivs/srems whose width was decreased" ); |
| 60 | STATISTIC(NumSDivs, "Number of sdiv converted to udiv" ); |
| 61 | STATISTIC(NumUDivURemsNarrowed, |
| 62 | "Number of udivs/urems whose width was decreased" ); |
| 63 | STATISTIC(NumAShrsConverted, "Number of ashr converted to lshr" ); |
| 64 | STATISTIC(NumAShrsRemoved, "Number of ashr removed" ); |
| 65 | STATISTIC(NumSRems, "Number of srem converted to urem" ); |
| 66 | STATISTIC(NumSExt, "Number of sext converted to zext" ); |
| 67 | STATISTIC(NumSIToFP, "Number of sitofp converted to uitofp" ); |
| 68 | STATISTIC(NumSICmps, "Number of signed icmp preds simplified to unsigned" ); |
| 69 | STATISTIC(NumAnd, "Number of ands removed" ); |
| 70 | STATISTIC(NumNW, "Number of no-wrap deductions" ); |
| 71 | STATISTIC(NumNSW, "Number of no-signed-wrap deductions" ); |
| 72 | STATISTIC(NumNUW, "Number of no-unsigned-wrap deductions" ); |
| 73 | STATISTIC(NumAddNW, "Number of no-wrap deductions for add" ); |
| 74 | STATISTIC(NumAddNSW, "Number of no-signed-wrap deductions for add" ); |
| 75 | STATISTIC(NumAddNUW, "Number of no-unsigned-wrap deductions for add" ); |
| 76 | STATISTIC(NumSubNW, "Number of no-wrap deductions for sub" ); |
| 77 | STATISTIC(NumSubNSW, "Number of no-signed-wrap deductions for sub" ); |
| 78 | STATISTIC(NumSubNUW, "Number of no-unsigned-wrap deductions for sub" ); |
| 79 | STATISTIC(NumMulNW, "Number of no-wrap deductions for mul" ); |
| 80 | STATISTIC(NumMulNSW, "Number of no-signed-wrap deductions for mul" ); |
| 81 | STATISTIC(NumMulNUW, "Number of no-unsigned-wrap deductions for mul" ); |
| 82 | STATISTIC(NumShlNW, "Number of no-wrap deductions for shl" ); |
| 83 | STATISTIC(NumShlNSW, "Number of no-signed-wrap deductions for shl" ); |
| 84 | STATISTIC(NumShlNUW, "Number of no-unsigned-wrap deductions for shl" ); |
| 85 | STATISTIC(NumAbs, "Number of llvm.abs intrinsics removed" ); |
| 86 | STATISTIC(NumOverflows, "Number of overflow checks removed" ); |
| 87 | STATISTIC(NumSaturating, |
| 88 | "Number of saturating arithmetics converted to normal arithmetics" ); |
| 89 | STATISTIC(NumNonNull, "Number of function pointer arguments marked non-null" ); |
| 90 | STATISTIC(NumCmpIntr, "Number of llvm.[us]cmp intrinsics removed" ); |
| 91 | STATISTIC(NumMinMax, "Number of llvm.[us]{min,max} intrinsics removed" ); |
| 92 | STATISTIC(NumSMinMax, |
| 93 | "Number of llvm.s{min,max} intrinsics simplified to unsigned" ); |
| 94 | STATISTIC(NumUDivURemsNarrowedExpanded, |
| 95 | "Number of bound udiv's/urem's expanded" ); |
| 96 | STATISTIC(NumNNeg, "Number of zext/uitofp non-negative deductions" ); |
| 97 | |
| 98 | namespace { |
| 99 | struct NoWrapFlags; |
| 100 | |
| 101 | class CVPImpl { |
| 102 | Function &F; |
| 103 | LazyValueInfo *LVI; |
| 104 | DominatorTree *DT; |
| 105 | const SimplifyQuery SQ; |
| 106 | std::optional<ConstantRange> RetRange; |
| 107 | |
| 108 | public: |
| 109 | CVPImpl(Function &F, LazyValueInfo *LVI, DominatorTree *DT, |
| 110 | const SimplifyQuery &SQ) |
| 111 | : F(F), LVI(LVI), DT(DT), SQ(SQ) {} |
| 112 | |
| 113 | bool run(); |
| 114 | |
| 115 | private: |
| 116 | // Driver methods. |
| 117 | bool processInstruction(Instruction &I); |
| 118 | bool inferReturnRangeAttr(); |
| 119 | |
| 120 | // Process instructions. |
| 121 | bool processSelect(SelectInst *S); |
| 122 | bool processPHI(PHINode *P); |
| 123 | bool processICmp(ICmpInst *Cmp); |
| 124 | bool processCmp(CmpInst *Cmp); |
| 125 | bool processSwitch(SwitchInst *I); |
| 126 | bool processBinOp(BinaryOperator *BinOp); |
| 127 | bool processAbsIntrinsic(IntrinsicInst *II); |
| 128 | bool processCmpIntrinsic(CmpIntrinsic *CI); |
| 129 | bool processMinMaxIntrinsic(MinMaxIntrinsic *MM); |
| 130 | bool processOverflowIntrinsic(WithOverflowInst *WO); |
| 131 | bool processSaturatingInst(SaturatingInst *SI); |
| 132 | bool processCallSite(CallBase &CB); |
| 133 | bool processUDivOrURem(BinaryOperator *Instr); |
| 134 | bool processSRem(BinaryOperator *SDI, const ConstantRange &LCR, |
| 135 | const ConstantRange &RCR); |
| 136 | bool processSDiv(BinaryOperator *SDI, const ConstantRange &LCR, |
| 137 | const ConstantRange &RCR); |
| 138 | bool processSDivOrSRem(BinaryOperator *Instr); |
| 139 | bool processAShr(BinaryOperator *SDI); |
| 140 | bool processSExt(SExtInst *SDI); |
| 141 | bool processPossibleNonNeg(PossiblyNonNegInst *I); |
| 142 | bool processSIToFP(SIToFPInst *SIToFP); |
| 143 | bool processAnd(BinaryOperator *BinOp); |
| 144 | bool processTrunc(TruncInst *TI); |
| 145 | bool processReturn(ReturnInst *RI); |
| 146 | |
| 147 | // Helpers. |
| 148 | Constant *getConstantAt(Value *V, Instruction *At); |
| 149 | bool simplifyCommonValuePhi(PHINode *P); |
| 150 | bool constantFoldCmp(CmpInst *Cmp); |
| 151 | bool willNotOverflow(BinaryOpIntrinsic *BO); |
| 152 | Value *getValueOnEdge(Value *Incoming, BasicBlock *From, BasicBlock *To, |
| 153 | Instruction *CtxI); |
| 154 | NoWrapFlags inferNoWrapFromUses(BinaryOperator *BinOp, bool WantNSW, |
| 155 | bool WantNUW); |
| 156 | }; |
| 157 | } // end anonymous namespace |
| 158 | |
| 159 | Constant *CVPImpl::getConstantAt(Value *V, Instruction *At) { |
| 160 | if (Constant *C = LVI->getConstant(V, CtxI: At)) |
| 161 | return C; |
| 162 | |
| 163 | // TODO: The following really should be sunk inside LVI's core algorithm, or |
| 164 | // at least the outer shims around such. |
| 165 | auto *C = dyn_cast<CmpInst>(Val: V); |
| 166 | if (!C) |
| 167 | return nullptr; |
| 168 | |
| 169 | Value *Op0 = C->getOperand(i_nocapture: 0); |
| 170 | Constant *Op1 = dyn_cast<Constant>(Val: C->getOperand(i_nocapture: 1)); |
| 171 | if (!Op1) |
| 172 | return nullptr; |
| 173 | |
| 174 | return LVI->getPredicateAt(Pred: C->getPredicate(), V: Op0, C: Op1, CtxI: At, |
| 175 | /*UseBlockValue=*/false); |
| 176 | } |
| 177 | |
| 178 | bool CVPImpl::processSelect(SelectInst *S) { |
| 179 | if (S->getType()->isVectorTy() || isa<Constant>(Val: S->getCondition())) |
| 180 | return false; |
| 181 | |
| 182 | bool Changed = false; |
| 183 | for (Use &U : make_early_inc_range(Range: S->uses())) { |
| 184 | auto *I = cast<Instruction>(Val: U.getUser()); |
| 185 | Constant *C; |
| 186 | if (auto *PN = dyn_cast<PHINode>(Val: I)) |
| 187 | C = LVI->getConstantOnEdge(V: S->getCondition(), FromBB: PN->getIncomingBlock(U), |
| 188 | ToBB: I->getParent(), CtxI: I); |
| 189 | else |
| 190 | C = getConstantAt(V: S->getCondition(), At: I); |
| 191 | |
| 192 | auto *CI = dyn_cast_or_null<ConstantInt>(Val: C); |
| 193 | if (!CI) |
| 194 | continue; |
| 195 | |
| 196 | U.set(CI->isOne() ? S->getTrueValue() : S->getFalseValue()); |
| 197 | Changed = true; |
| 198 | ++NumSelects; |
| 199 | } |
| 200 | |
| 201 | if (Changed && S->use_empty()) |
| 202 | S->eraseFromParent(); |
| 203 | |
| 204 | return Changed; |
| 205 | } |
| 206 | |
| 207 | /// Try to simplify a phi with constant incoming values that match the edge |
| 208 | /// values of a non-constant value on all other edges: |
| 209 | /// bb0: |
| 210 | /// %isnull = icmp eq i8* %x, null |
| 211 | /// br i1 %isnull, label %bb2, label %bb1 |
| 212 | /// bb1: |
| 213 | /// br label %bb2 |
| 214 | /// bb2: |
| 215 | /// %r = phi i8* [ %x, %bb1 ], [ null, %bb0 ] |
| 216 | /// --> |
| 217 | /// %r = %x |
| 218 | bool CVPImpl::simplifyCommonValuePhi(PHINode *P) { |
| 219 | // Collect incoming constants and initialize possible common value. |
| 220 | SmallVector<std::pair<Constant *, unsigned>, 4> IncomingConstants; |
| 221 | Value *CommonValue = nullptr; |
| 222 | for (unsigned I = 0, E = P->getNumIncomingValues(); I != E; ++I) { |
| 223 | Value *Incoming = P->getIncomingValue(i: I); |
| 224 | if (auto *IncomingConstant = dyn_cast<Constant>(Val: Incoming)) { |
| 225 | IncomingConstants.push_back(Elt: std::make_pair(x&: IncomingConstant, y&: I)); |
| 226 | } else if (!CommonValue) { |
| 227 | // The potential common value is initialized to the first non-constant. |
| 228 | CommonValue = Incoming; |
| 229 | } else if (Incoming != CommonValue) { |
| 230 | // There can be only one non-constant common value. |
| 231 | return false; |
| 232 | } |
| 233 | } |
| 234 | |
| 235 | if (!CommonValue || IncomingConstants.empty()) |
| 236 | return false; |
| 237 | |
| 238 | // The common value must be valid in all incoming blocks. |
| 239 | BasicBlock *ToBB = P->getParent(); |
| 240 | if (auto *CommonInst = dyn_cast<Instruction>(Val: CommonValue)) |
| 241 | if (!DT->dominates(Def: CommonInst, BB: ToBB)) |
| 242 | return false; |
| 243 | |
| 244 | // We have a phi with exactly 1 variable incoming value and 1 or more constant |
| 245 | // incoming values. See if all constant incoming values can be mapped back to |
| 246 | // the same incoming variable value. |
| 247 | for (auto &IncomingConstant : IncomingConstants) { |
| 248 | Constant *C = IncomingConstant.first; |
| 249 | BasicBlock *IncomingBB = P->getIncomingBlock(i: IncomingConstant.second); |
| 250 | if (C != LVI->getConstantOnEdge(V: CommonValue, FromBB: IncomingBB, ToBB, CtxI: P)) |
| 251 | return false; |
| 252 | } |
| 253 | |
| 254 | // LVI only guarantees that the value matches a certain constant if the value |
| 255 | // is not poison. Make sure we don't replace a well-defined value with poison. |
| 256 | // This is usually satisfied due to a prior branch on the value. |
| 257 | if (!isGuaranteedNotToBePoison(V: CommonValue, AC: nullptr, CtxI: P, DT)) |
| 258 | return false; |
| 259 | |
| 260 | // All constant incoming values map to the same variable along the incoming |
| 261 | // edges of the phi. The phi is unnecessary. |
| 262 | P->replaceAllUsesWith(V: CommonValue); |
| 263 | P->eraseFromParent(); |
| 264 | ++NumPhiCommon; |
| 265 | return true; |
| 266 | } |
| 267 | |
| 268 | Value *CVPImpl::getValueOnEdge(Value *Incoming, BasicBlock *From, |
| 269 | BasicBlock *To, Instruction *CtxI) { |
| 270 | if (Constant *C = LVI->getConstantOnEdge(V: Incoming, FromBB: From, ToBB: To, CtxI)) |
| 271 | return C; |
| 272 | |
| 273 | // Look if the incoming value is a select with a scalar condition for which |
| 274 | // LVI can tells us the value. In that case replace the incoming value with |
| 275 | // the appropriate value of the select. This often allows us to remove the |
| 276 | // select later. |
| 277 | auto *SI = dyn_cast<SelectInst>(Val: Incoming); |
| 278 | if (!SI) |
| 279 | return nullptr; |
| 280 | |
| 281 | // Once LVI learns to handle vector types, we could also add support |
| 282 | // for vector type constants that are not all zeroes or all ones. |
| 283 | Value *Condition = SI->getCondition(); |
| 284 | if (!Condition->getType()->isVectorTy()) { |
| 285 | if (Constant *C = LVI->getConstantOnEdge(V: Condition, FromBB: From, ToBB: To, CtxI)) { |
| 286 | if (C->isOneValue()) |
| 287 | return SI->getTrueValue(); |
| 288 | if (C->isNullValue()) |
| 289 | return SI->getFalseValue(); |
| 290 | } |
| 291 | } |
| 292 | |
| 293 | // Look if the select has a constant but LVI tells us that the incoming |
| 294 | // value can never be that constant. In that case replace the incoming |
| 295 | // value with the other value of the select. This often allows us to |
| 296 | // remove the select later. |
| 297 | |
| 298 | // The "false" case |
| 299 | if (auto *C = dyn_cast<Constant>(Val: SI->getFalseValue())) |
| 300 | if (auto *Res = dyn_cast_or_null<ConstantInt>( |
| 301 | Val: LVI->getPredicateOnEdge(Pred: ICmpInst::ICMP_EQ, V: SI, C, FromBB: From, ToBB: To, CtxI)); |
| 302 | Res && Res->isZero()) |
| 303 | return SI->getTrueValue(); |
| 304 | |
| 305 | // The "true" case, |
| 306 | // similar to the select "false" case, but try the select "true" value |
| 307 | if (auto *C = dyn_cast<Constant>(Val: SI->getTrueValue())) |
| 308 | if (auto *Res = dyn_cast_or_null<ConstantInt>( |
| 309 | Val: LVI->getPredicateOnEdge(Pred: ICmpInst::ICMP_EQ, V: SI, C, FromBB: From, ToBB: To, CtxI)); |
| 310 | Res && Res->isZero()) |
| 311 | return SI->getFalseValue(); |
| 312 | |
| 313 | return nullptr; |
| 314 | } |
| 315 | |
| 316 | bool CVPImpl::processPHI(PHINode *P) { |
| 317 | bool Changed = false; |
| 318 | |
| 319 | BasicBlock *BB = P->getParent(); |
| 320 | for (unsigned I = 0, E = P->getNumIncomingValues(); I < E; ++I) { |
| 321 | Value *Incoming = P->getIncomingValue(i: I); |
| 322 | if (isa<Constant>(Val: Incoming)) |
| 323 | continue; |
| 324 | |
| 325 | Value *V = getValueOnEdge(Incoming, From: P->getIncomingBlock(i: I), To: BB, CtxI: P); |
| 326 | if (V) { |
| 327 | P->setIncomingValue(i: I, V); |
| 328 | Changed = true; |
| 329 | } |
| 330 | } |
| 331 | |
| 332 | if (Value *V = simplifyInstruction(I: P, Q: SQ)) { |
| 333 | P->replaceAllUsesWith(V); |
| 334 | P->eraseFromParent(); |
| 335 | Changed = true; |
| 336 | } |
| 337 | |
| 338 | if (!Changed) |
| 339 | Changed = simplifyCommonValuePhi(P); |
| 340 | |
| 341 | if (Changed) |
| 342 | ++NumPhis; |
| 343 | |
| 344 | return Changed; |
| 345 | } |
| 346 | |
| 347 | bool CVPImpl::processICmp(ICmpInst *Cmp) { |
| 348 | // Only for signed relational comparisons of integers. |
| 349 | if (!Cmp->getOperand(i_nocapture: 0)->getType()->isIntOrIntVectorTy()) |
| 350 | return false; |
| 351 | |
| 352 | if (!Cmp->isSigned() && (!Cmp->isUnsigned() || Cmp->hasSameSign())) |
| 353 | return false; |
| 354 | |
| 355 | bool Changed = false; |
| 356 | SmallVector<DbgVariableRecord *, 2> DebugUsers; |
| 357 | ConstantRange CR1 = LVI->getConstantRangeAtUse(U: Cmp->getOperandUse(i: 0), |
| 358 | /*UndefAllowed=*/false, |
| 359 | DebugUsers: &DebugUsers), |
| 360 | CR2 = LVI->getConstantRangeAtUse(U: Cmp->getOperandUse(i: 1), |
| 361 | /*UndefAllowed=*/false, |
| 362 | DebugUsers: &DebugUsers); |
| 363 | if (Cmp->isSigned()) { |
| 364 | ICmpInst::Predicate UnsignedPred = |
| 365 | ConstantRange::getEquivalentPredWithFlippedSignedness( |
| 366 | Pred: Cmp->getPredicate(), CR1, CR2); |
| 367 | |
| 368 | if (UnsignedPred == ICmpInst::Predicate::BAD_ICMP_PREDICATE) |
| 369 | return false; |
| 370 | |
| 371 | // Since the instruction changed, all debug users of |
| 372 | // this value must be poisoned |
| 373 | for (auto *Record : DebugUsers) |
| 374 | Record->setKillLocation(); |
| 375 | |
| 376 | ++NumSICmps; |
| 377 | Cmp->setPredicate(UnsignedPred); |
| 378 | Changed = true; |
| 379 | } |
| 380 | |
| 381 | if (ConstantRange::areInsensitiveToSignednessOfICmpPredicate(CR1, CR2)) { |
| 382 | Cmp->setSameSign(); |
| 383 | Changed = true; |
| 384 | } |
| 385 | |
| 386 | return Changed; |
| 387 | } |
| 388 | |
| 389 | /// See if LazyValueInfo's ability to exploit edge conditions or range |
| 390 | /// information is sufficient to prove this comparison. Even for local |
| 391 | /// conditions, this can sometimes prove conditions instcombine can't by |
| 392 | /// exploiting range information. |
| 393 | bool CVPImpl::constantFoldCmp(CmpInst *Cmp) { |
| 394 | Value *Op0 = Cmp->getOperand(i_nocapture: 0); |
| 395 | Value *Op1 = Cmp->getOperand(i_nocapture: 1); |
| 396 | Constant *Res = LVI->getPredicateAt(Pred: Cmp->getPredicate(), LHS: Op0, RHS: Op1, CtxI: Cmp, |
| 397 | /*UseBlockValue=*/true); |
| 398 | if (!Res) |
| 399 | return false; |
| 400 | |
| 401 | bool Changed = Cmp->replaceUsesWithIf( |
| 402 | New: Res, ShouldReplace: [](Use &U) { return !isa<AssumeInst>(Val: U.getUser()); }); |
| 403 | if (Cmp->use_empty()) { |
| 404 | Cmp->eraseFromParent(); |
| 405 | Changed = true; |
| 406 | } |
| 407 | |
| 408 | if (Changed) |
| 409 | ++NumCmps; |
| 410 | |
| 411 | return Changed; |
| 412 | } |
| 413 | |
| 414 | bool CVPImpl::processCmp(CmpInst *Cmp) { |
| 415 | if (constantFoldCmp(Cmp)) |
| 416 | return true; |
| 417 | |
| 418 | if (auto *ICmp = dyn_cast<ICmpInst>(Val: Cmp)) |
| 419 | if (processICmp(Cmp: ICmp)) |
| 420 | return true; |
| 421 | |
| 422 | return false; |
| 423 | } |
| 424 | |
| 425 | /// Simplify a switch instruction by removing cases which can never fire. If the |
| 426 | /// uselessness of a case could be determined locally then constant propagation |
| 427 | /// would already have figured it out. Instead, walk the predecessors and |
| 428 | /// statically evaluate cases based on information available on that edge. Cases |
| 429 | /// that cannot fire no matter what the incoming edge can safely be removed. If |
| 430 | /// a case fires on every incoming edge then the entire switch can be removed |
| 431 | /// and replaced with a branch to the case destination. |
| 432 | bool CVPImpl::processSwitch(SwitchInst *I) { |
| 433 | DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy); |
| 434 | Value *Cond = I->getCondition(); |
| 435 | BasicBlock *BB = I->getParent(); |
| 436 | |
| 437 | // Analyse each switch case in turn. |
| 438 | bool Changed = false; |
| 439 | DenseMap<BasicBlock*, int> SuccessorsCount; |
| 440 | for (auto *Succ : successors(BB)) |
| 441 | SuccessorsCount[Succ]++; |
| 442 | |
| 443 | { // Scope for SwitchInstProfUpdateWrapper. It must not live during |
| 444 | // ConstantFoldTerminator() as the underlying SwitchInst can be changed. |
| 445 | SwitchInstProfUpdateWrapper SI(*I); |
| 446 | ConstantRange CR = |
| 447 | LVI->getConstantRangeAtUse(U: I->getOperandUse(i: 0), /*UndefAllowed=*/false); |
| 448 | unsigned ReachableCaseCount = 0; |
| 449 | |
| 450 | for (auto CI = SI->case_begin(), CE = SI->case_end(); CI != CE;) { |
| 451 | ConstantInt *Case = CI->getCaseValue(); |
| 452 | std::optional<bool> Predicate = std::nullopt; |
| 453 | if (!CR.contains(Val: Case->getValue())) |
| 454 | Predicate = false; |
| 455 | else if (CR.isSingleElement() && |
| 456 | *CR.getSingleElement() == Case->getValue()) |
| 457 | Predicate = true; |
| 458 | if (!Predicate) { |
| 459 | // Handle missing cases, e.g., the range has a hole. |
| 460 | auto *Res = dyn_cast_or_null<ConstantInt>( |
| 461 | Val: LVI->getPredicateAt(Pred: CmpInst::ICMP_EQ, V: Cond, C: Case, CtxI: I, |
| 462 | /* UseBlockValue=*/true)); |
| 463 | if (Res && Res->isZero()) |
| 464 | Predicate = false; |
| 465 | else if (Res && Res->isOne()) |
| 466 | Predicate = true; |
| 467 | } |
| 468 | |
| 469 | if (Predicate && !*Predicate) { |
| 470 | // This case never fires - remove it. |
| 471 | BasicBlock *Succ = CI->getCaseSuccessor(); |
| 472 | Succ->removePredecessor(Pred: BB); |
| 473 | CI = SI.removeCase(I: CI); |
| 474 | CE = SI->case_end(); |
| 475 | |
| 476 | // The condition can be modified by removePredecessor's PHI simplification |
| 477 | // logic. |
| 478 | Cond = SI->getCondition(); |
| 479 | |
| 480 | ++NumDeadCases; |
| 481 | Changed = true; |
| 482 | if (--SuccessorsCount[Succ] == 0) |
| 483 | DTU.applyUpdatesPermissive(Updates: {{DominatorTree::Delete, BB, Succ}}); |
| 484 | continue; |
| 485 | } |
| 486 | if (Predicate && *Predicate) { |
| 487 | // This case always fires. Arrange for the switch to be turned into an |
| 488 | // unconditional branch by replacing the switch condition with the case |
| 489 | // value. |
| 490 | SI->setCondition(Case); |
| 491 | NumDeadCases += SI->getNumCases(); |
| 492 | Changed = true; |
| 493 | break; |
| 494 | } |
| 495 | |
| 496 | // Increment the case iterator since we didn't delete it. |
| 497 | ++CI; |
| 498 | ++ReachableCaseCount; |
| 499 | } |
| 500 | |
| 501 | // The default dest is unreachable if all cases are covered. |
| 502 | if (!SI->defaultDestUnreachable() && |
| 503 | !CR.isSizeLargerThan(MaxSize: ReachableCaseCount)) { |
| 504 | BasicBlock *DefaultDest = SI->getDefaultDest(); |
| 505 | BasicBlock *NewUnreachableBB = |
| 506 | BasicBlock::Create(Context&: BB->getContext(), Name: "default.unreachable" , |
| 507 | Parent: BB->getParent(), InsertBefore: DefaultDest); |
| 508 | auto *UI = new UnreachableInst(BB->getContext(), NewUnreachableBB); |
| 509 | UI->setDebugLoc(DebugLoc::getTemporary()); |
| 510 | |
| 511 | DefaultDest->removePredecessor(Pred: BB); |
| 512 | SI->setDefaultDest(NewUnreachableBB); |
| 513 | |
| 514 | if (SuccessorsCount[DefaultDest] == 1) |
| 515 | DTU.applyUpdates(Updates: {{DominatorTree::Delete, BB, DefaultDest}}); |
| 516 | DTU.applyUpdates(Updates: {{DominatorTree::Insert, BB, NewUnreachableBB}}); |
| 517 | |
| 518 | ++NumDeadCases; |
| 519 | Changed = true; |
| 520 | } |
| 521 | } |
| 522 | |
| 523 | if (Changed) |
| 524 | // If the switch has been simplified to the point where it can be replaced |
| 525 | // by a branch then do so now. |
| 526 | ConstantFoldTerminator(BB, /*DeleteDeadConditions = */ false, |
| 527 | /*TLI = */ nullptr, DTU: &DTU); |
| 528 | return Changed; |
| 529 | } |
| 530 | |
| 531 | // See if we can prove that the given binary op intrinsic will not overflow. |
| 532 | bool CVPImpl::willNotOverflow(BinaryOpIntrinsic *BO) { |
| 533 | ConstantRange LRange = |
| 534 | LVI->getConstantRangeAtUse(U: BO->getOperandUse(i: 0), /*UndefAllowed*/ false); |
| 535 | ConstantRange RRange = |
| 536 | LVI->getConstantRangeAtUse(U: BO->getOperandUse(i: 1), /*UndefAllowed*/ false); |
| 537 | ConstantRange NWRegion = ConstantRange::makeGuaranteedNoWrapRegion( |
| 538 | BinOp: BO->getBinaryOp(), Other: RRange, NoWrapKind: BO->getNoWrapKind()); |
| 539 | return NWRegion.contains(CR: LRange); |
| 540 | } |
| 541 | |
| 542 | static void setDeducedOverflowingFlags(Value *V, Instruction::BinaryOps Opcode, |
| 543 | bool NewNSW, bool NewNUW) { |
| 544 | Statistic *OpcNW, *OpcNSW, *OpcNUW; |
| 545 | switch (Opcode) { |
| 546 | case Instruction::Add: |
| 547 | OpcNW = &NumAddNW; |
| 548 | OpcNSW = &NumAddNSW; |
| 549 | OpcNUW = &NumAddNUW; |
| 550 | break; |
| 551 | case Instruction::Sub: |
| 552 | OpcNW = &NumSubNW; |
| 553 | OpcNSW = &NumSubNSW; |
| 554 | OpcNUW = &NumSubNUW; |
| 555 | break; |
| 556 | case Instruction::Mul: |
| 557 | OpcNW = &NumMulNW; |
| 558 | OpcNSW = &NumMulNSW; |
| 559 | OpcNUW = &NumMulNUW; |
| 560 | break; |
| 561 | case Instruction::Shl: |
| 562 | OpcNW = &NumShlNW; |
| 563 | OpcNSW = &NumShlNSW; |
| 564 | OpcNUW = &NumShlNUW; |
| 565 | break; |
| 566 | default: |
| 567 | llvm_unreachable("Will not be called with other binops" ); |
| 568 | } |
| 569 | |
| 570 | auto *Inst = dyn_cast<Instruction>(Val: V); |
| 571 | if (NewNSW) { |
| 572 | ++NumNW; |
| 573 | ++*OpcNW; |
| 574 | ++NumNSW; |
| 575 | ++*OpcNSW; |
| 576 | if (Inst) |
| 577 | Inst->setHasNoSignedWrap(); |
| 578 | } |
| 579 | if (NewNUW) { |
| 580 | ++NumNW; |
| 581 | ++*OpcNW; |
| 582 | ++NumNUW; |
| 583 | ++*OpcNUW; |
| 584 | if (Inst) |
| 585 | Inst->setHasNoUnsignedWrap(); |
| 586 | } |
| 587 | } |
| 588 | |
| 589 | // See if @llvm.abs argument is alays positive/negative, and simplify. |
| 590 | // Notably, INT_MIN can belong to either range, regardless of the NSW, |
| 591 | // because it is negation-invariant. |
| 592 | bool CVPImpl::processAbsIntrinsic(IntrinsicInst *II) { |
| 593 | Value *X = II->getArgOperand(i: 0); |
| 594 | bool IsIntMinPoison = cast<ConstantInt>(Val: II->getArgOperand(i: 1))->isOne(); |
| 595 | APInt IntMin = APInt::getSignedMinValue(numBits: X->getType()->getScalarSizeInBits()); |
| 596 | ConstantRange Range = LVI->getConstantRangeAtUse( |
| 597 | U: II->getOperandUse(i: 0), /*UndefAllowed*/ IsIntMinPoison); |
| 598 | |
| 599 | // Is X in [0, IntMin]? NOTE: INT_MIN is fine! |
| 600 | if (Range.icmp(Pred: CmpInst::ICMP_ULE, Other: IntMin)) { |
| 601 | ++NumAbs; |
| 602 | II->replaceAllUsesWith(V: X); |
| 603 | II->eraseFromParent(); |
| 604 | return true; |
| 605 | } |
| 606 | |
| 607 | // Is X in [IntMin, 0]? NOTE: INT_MIN is fine! |
| 608 | if (Range.getSignedMax().isNonPositive()) { |
| 609 | IRBuilder<> B(II); |
| 610 | Value *NegX = B.CreateNeg(V: X, Name: II->getName(), |
| 611 | /*HasNSW=*/IsIntMinPoison); |
| 612 | ++NumAbs; |
| 613 | II->replaceAllUsesWith(V: NegX); |
| 614 | II->eraseFromParent(); |
| 615 | |
| 616 | // See if we can infer some no-wrap flags. |
| 617 | if (auto *BO = dyn_cast<BinaryOperator>(Val: NegX)) |
| 618 | processBinOp(BinOp: BO); |
| 619 | |
| 620 | return true; |
| 621 | } |
| 622 | |
| 623 | // Argument's range crosses zero. |
| 624 | // Can we at least tell that the argument is never INT_MIN? |
| 625 | if (!IsIntMinPoison && !Range.contains(Val: IntMin)) { |
| 626 | ++NumNSW; |
| 627 | ++NumSubNSW; |
| 628 | II->setArgOperand(i: 1, v: ConstantInt::getTrue(Context&: II->getContext())); |
| 629 | return true; |
| 630 | } |
| 631 | return false; |
| 632 | } |
| 633 | |
| 634 | bool CVPImpl::processCmpIntrinsic(CmpIntrinsic *CI) { |
| 635 | ConstantRange LRange = |
| 636 | LVI->getConstantRangeAtUse(U: CI->getOperandUse(i: 0), /*UndefAllowed*/ false); |
| 637 | ConstantRange RRange = |
| 638 | LVI->getConstantRangeAtUse(U: CI->getOperandUse(i: 1), /*UndefAllowed*/ false); |
| 639 | |
| 640 | if (LRange.icmp(Pred: CI->getGTPredicate(), Other: RRange)) { |
| 641 | ++NumCmpIntr; |
| 642 | CI->replaceAllUsesWith(V: ConstantInt::get(Ty: CI->getType(), V: 1)); |
| 643 | CI->eraseFromParent(); |
| 644 | return true; |
| 645 | } |
| 646 | if (LRange.icmp(Pred: CI->getLTPredicate(), Other: RRange)) { |
| 647 | ++NumCmpIntr; |
| 648 | CI->replaceAllUsesWith(V: ConstantInt::getSigned(Ty: CI->getType(), V: -1)); |
| 649 | CI->eraseFromParent(); |
| 650 | return true; |
| 651 | } |
| 652 | if (LRange.icmp(Pred: ICmpInst::ICMP_EQ, Other: RRange)) { |
| 653 | ++NumCmpIntr; |
| 654 | CI->replaceAllUsesWith(V: ConstantInt::get(Ty: CI->getType(), V: 0)); |
| 655 | CI->eraseFromParent(); |
| 656 | return true; |
| 657 | } |
| 658 | |
| 659 | return false; |
| 660 | } |
| 661 | |
| 662 | // See if this min/max intrinsic always picks it's one specific operand. |
| 663 | // If not, check whether we can canonicalize signed minmax into unsigned version |
| 664 | bool CVPImpl::processMinMaxIntrinsic(MinMaxIntrinsic *MM) { |
| 665 | CmpInst::Predicate Pred = CmpInst::getNonStrictPredicate(pred: MM->getPredicate()); |
| 666 | ConstantRange LRange = LVI->getConstantRangeAtUse(U: MM->getOperandUse(i: 0), |
| 667 | /*UndefAllowed*/ false); |
| 668 | ConstantRange RRange = LVI->getConstantRangeAtUse(U: MM->getOperandUse(i: 1), |
| 669 | /*UndefAllowed*/ false); |
| 670 | if (LRange.icmp(Pred, Other: RRange)) { |
| 671 | ++NumMinMax; |
| 672 | MM->replaceAllUsesWith(V: MM->getLHS()); |
| 673 | MM->eraseFromParent(); |
| 674 | return true; |
| 675 | } |
| 676 | if (RRange.icmp(Pred, Other: LRange)) { |
| 677 | ++NumMinMax; |
| 678 | MM->replaceAllUsesWith(V: MM->getRHS()); |
| 679 | MM->eraseFromParent(); |
| 680 | return true; |
| 681 | } |
| 682 | |
| 683 | if (MM->isSigned() && |
| 684 | ConstantRange::areInsensitiveToSignednessOfICmpPredicate(CR1: LRange, |
| 685 | CR2: RRange)) { |
| 686 | ++NumSMinMax; |
| 687 | IRBuilder<> B(MM); |
| 688 | salvageDebugInfo(I&: *MM); |
| 689 | MM->replaceAllUsesWith(V: B.CreateBinaryIntrinsic( |
| 690 | ID: MM->getIntrinsicID() == Intrinsic::smin ? Intrinsic::umin |
| 691 | : Intrinsic::umax, |
| 692 | LHS: MM->getLHS(), RHS: MM->getRHS())); |
| 693 | MM->eraseFromParent(); |
| 694 | return true; |
| 695 | } |
| 696 | |
| 697 | return false; |
| 698 | } |
| 699 | |
| 700 | // Rewrite this with.overflow intrinsic as non-overflowing. |
| 701 | bool CVPImpl::processOverflowIntrinsic(WithOverflowInst *WO) { |
| 702 | IRBuilder<> B(WO); |
| 703 | Instruction::BinaryOps Opcode = WO->getBinaryOp(); |
| 704 | bool NSW = WO->isSigned(); |
| 705 | bool NUW = !WO->isSigned(); |
| 706 | |
| 707 | Value *NewOp = |
| 708 | B.CreateBinOp(Opc: Opcode, LHS: WO->getLHS(), RHS: WO->getRHS(), Name: WO->getName()); |
| 709 | setDeducedOverflowingFlags(V: NewOp, Opcode, NewNSW: NSW, NewNUW: NUW); |
| 710 | |
| 711 | StructType *ST = cast<StructType>(Val: WO->getType()); |
| 712 | Constant *Struct = ConstantStruct::get(T: ST, |
| 713 | V: { PoisonValue::get(T: ST->getElementType(N: 0)), |
| 714 | ConstantInt::getFalse(Ty: ST->getElementType(N: 1)) }); |
| 715 | Value *NewI = B.CreateInsertValue(Agg: Struct, Val: NewOp, Idxs: 0); |
| 716 | WO->replaceAllUsesWith(V: NewI); |
| 717 | WO->eraseFromParent(); |
| 718 | ++NumOverflows; |
| 719 | |
| 720 | // See if we can infer the other no-wrap too. |
| 721 | if (auto *BO = dyn_cast<BinaryOperator>(Val: NewOp)) |
| 722 | processBinOp(BinOp: BO); |
| 723 | |
| 724 | return true; |
| 725 | } |
| 726 | |
| 727 | bool CVPImpl::processSaturatingInst(SaturatingInst *SI) { |
| 728 | Instruction::BinaryOps Opcode = SI->getBinaryOp(); |
| 729 | bool NSW = SI->isSigned(); |
| 730 | bool NUW = !SI->isSigned(); |
| 731 | BinaryOperator *BinOp = BinaryOperator::Create( |
| 732 | Op: Opcode, S1: SI->getLHS(), S2: SI->getRHS(), Name: SI->getName(), InsertBefore: SI->getIterator()); |
| 733 | BinOp->setDebugLoc(SI->getDebugLoc()); |
| 734 | setDeducedOverflowingFlags(V: BinOp, Opcode, NewNSW: NSW, NewNUW: NUW); |
| 735 | |
| 736 | salvageDebugInfo(I&: *SI); |
| 737 | SI->replaceAllUsesWith(V: BinOp); |
| 738 | SI->eraseFromParent(); |
| 739 | ++NumSaturating; |
| 740 | |
| 741 | // See if we can infer the other no-wrap too. |
| 742 | processBinOp(BinOp); |
| 743 | |
| 744 | return true; |
| 745 | } |
| 746 | |
| 747 | /// Infer nonnull attributes for the arguments at the specified callsite. |
| 748 | bool CVPImpl::processCallSite(CallBase &CB) { |
| 749 | |
| 750 | if (CB.getIntrinsicID() == Intrinsic::abs) { |
| 751 | return processAbsIntrinsic(II: &cast<IntrinsicInst>(Val&: CB)); |
| 752 | } |
| 753 | |
| 754 | if (auto *CI = dyn_cast<CmpIntrinsic>(Val: &CB)) { |
| 755 | return processCmpIntrinsic(CI); |
| 756 | } |
| 757 | |
| 758 | if (auto *MM = dyn_cast<MinMaxIntrinsic>(Val: &CB)) { |
| 759 | return processMinMaxIntrinsic(MM); |
| 760 | } |
| 761 | |
| 762 | if (auto *WO = dyn_cast<WithOverflowInst>(Val: &CB)) { |
| 763 | if (willNotOverflow(BO: WO)) |
| 764 | return processOverflowIntrinsic(WO); |
| 765 | } |
| 766 | |
| 767 | if (auto *SI = dyn_cast<SaturatingInst>(Val: &CB)) { |
| 768 | if (willNotOverflow(BO: SI)) |
| 769 | return processSaturatingInst(SI); |
| 770 | } |
| 771 | |
| 772 | bool Changed = false; |
| 773 | |
| 774 | // Deopt bundle operands are intended to capture state with minimal |
| 775 | // perturbance of the code otherwise. If we can find a constant value for |
| 776 | // any such operand and remove a use of the original value, that's |
| 777 | // desireable since it may allow further optimization of that value (e.g. via |
| 778 | // single use rules in instcombine). Since deopt uses tend to, |
| 779 | // idiomatically, appear along rare conditional paths, it's reasonable likely |
| 780 | // we may have a conditional fact with which LVI can fold. |
| 781 | if (auto DeoptBundle = CB.getOperandBundle(ID: LLVMContext::OB_deopt)) { |
| 782 | for (const Use &ConstU : DeoptBundle->Inputs) { |
| 783 | Use &U = const_cast<Use&>(ConstU); |
| 784 | Value *V = U.get(); |
| 785 | if (V->getType()->isVectorTy()) continue; |
| 786 | if (isa<Constant>(Val: V)) continue; |
| 787 | |
| 788 | Constant *C = LVI->getConstant(V, CtxI: &CB); |
| 789 | if (!C) continue; |
| 790 | U.set(C); |
| 791 | Changed = true; |
| 792 | } |
| 793 | } |
| 794 | |
| 795 | SmallVector<unsigned, 4> ArgNos; |
| 796 | unsigned ArgNo = 0; |
| 797 | |
| 798 | for (Value *V : CB.args()) { |
| 799 | PointerType *Type = dyn_cast<PointerType>(Val: V->getType()); |
| 800 | // Try to mark pointer typed parameters as non-null. We skip the |
| 801 | // relatively expensive analysis for constants which are obviously either |
| 802 | // null or non-null to start with. |
| 803 | if (Type && !CB.paramHasAttr(ArgNo, Kind: Attribute::NonNull) && |
| 804 | !isa<Constant>(Val: V)) |
| 805 | if (auto *Res = dyn_cast_or_null<ConstantInt>(Val: LVI->getPredicateAt( |
| 806 | Pred: ICmpInst::ICMP_EQ, V, C: ConstantPointerNull::get(T: Type), CtxI: &CB, |
| 807 | /*UseBlockValue=*/false)); |
| 808 | Res && Res->isZero()) |
| 809 | ArgNos.push_back(Elt: ArgNo); |
| 810 | ArgNo++; |
| 811 | } |
| 812 | |
| 813 | assert(ArgNo == CB.arg_size() && "Call arguments not processed correctly." ); |
| 814 | |
| 815 | if (ArgNos.empty()) |
| 816 | return Changed; |
| 817 | |
| 818 | NumNonNull += ArgNos.size(); |
| 819 | AttributeList AS = CB.getAttributes(); |
| 820 | LLVMContext &Ctx = CB.getContext(); |
| 821 | AS = AS.addParamAttribute(C&: Ctx, ArgNos, |
| 822 | A: Attribute::get(Context&: Ctx, Kind: Attribute::NonNull)); |
| 823 | CB.setAttributes(AS); |
| 824 | |
| 825 | return true; |
| 826 | } |
| 827 | |
| 828 | enum class Domain { NonNegative, NonPositive, Unknown }; |
| 829 | |
| 830 | static Domain getDomain(const ConstantRange &CR) { |
| 831 | if (CR.isAllNonNegative()) |
| 832 | return Domain::NonNegative; |
| 833 | if (CR.icmp(Pred: ICmpInst::ICMP_SLE, Other: APInt::getZero(numBits: CR.getBitWidth()))) |
| 834 | return Domain::NonPositive; |
| 835 | return Domain::Unknown; |
| 836 | } |
| 837 | |
| 838 | /// Try to shrink a sdiv/srem's width down to the smallest power of two that's |
| 839 | /// sufficient to contain its operands. |
| 840 | static bool narrowSDivOrSRem(BinaryOperator *Instr, const ConstantRange &LCR, |
| 841 | const ConstantRange &RCR) { |
| 842 | assert(Instr->getOpcode() == Instruction::SDiv || |
| 843 | Instr->getOpcode() == Instruction::SRem); |
| 844 | |
| 845 | // Find the smallest power of two bitwidth that's sufficient to hold Instr's |
| 846 | // operands. |
| 847 | unsigned OrigWidth = Instr->getType()->getScalarSizeInBits(); |
| 848 | |
| 849 | // What is the smallest bit width that can accommodate the entire value ranges |
| 850 | // of both of the operands? |
| 851 | unsigned MinSignedBits = |
| 852 | std::max(a: LCR.getMinSignedBits(), b: RCR.getMinSignedBits()); |
| 853 | |
| 854 | // sdiv/srem is UB if divisor is -1 and divident is INT_MIN, so unless we can |
| 855 | // prove that such a combination is impossible, we need to bump the bitwidth. |
| 856 | if (RCR.contains(Val: APInt::getAllOnes(numBits: OrigWidth)) && |
| 857 | LCR.contains(Val: APInt::getSignedMinValue(numBits: MinSignedBits).sext(width: OrigWidth))) |
| 858 | ++MinSignedBits; |
| 859 | |
| 860 | // Don't shrink below 8 bits wide. |
| 861 | unsigned NewWidth = std::max<unsigned>(a: PowerOf2Ceil(A: MinSignedBits), b: 8); |
| 862 | |
| 863 | // NewWidth might be greater than OrigWidth if OrigWidth is not a power of |
| 864 | // two. |
| 865 | if (NewWidth >= OrigWidth) |
| 866 | return false; |
| 867 | |
| 868 | ++NumSDivSRemsNarrowed; |
| 869 | IRBuilder<> B{Instr}; |
| 870 | auto *TruncTy = Instr->getType()->getWithNewBitWidth(NewBitWidth: NewWidth); |
| 871 | auto *LHS = B.CreateTruncOrBitCast(V: Instr->getOperand(i_nocapture: 0), DestTy: TruncTy, |
| 872 | Name: Instr->getName() + ".lhs.trunc" ); |
| 873 | auto *RHS = B.CreateTruncOrBitCast(V: Instr->getOperand(i_nocapture: 1), DestTy: TruncTy, |
| 874 | Name: Instr->getName() + ".rhs.trunc" ); |
| 875 | auto *BO = B.CreateBinOp(Opc: Instr->getOpcode(), LHS, RHS, Name: Instr->getName()); |
| 876 | auto *Sext = B.CreateSExt(V: BO, DestTy: Instr->getType(), Name: Instr->getName() + ".sext" ); |
| 877 | if (auto *BinOp = dyn_cast<BinaryOperator>(Val: BO)) |
| 878 | if (BinOp->getOpcode() == Instruction::SDiv) |
| 879 | BinOp->setIsExact(Instr->isExact()); |
| 880 | |
| 881 | Instr->replaceAllUsesWith(V: Sext); |
| 882 | Instr->eraseFromParent(); |
| 883 | return true; |
| 884 | } |
| 885 | |
| 886 | static bool expandUDivOrURem(BinaryOperator *Instr, const ConstantRange &XCR, |
| 887 | const ConstantRange &YCR) { |
| 888 | Type *Ty = Instr->getType(); |
| 889 | assert(Instr->getOpcode() == Instruction::UDiv || |
| 890 | Instr->getOpcode() == Instruction::URem); |
| 891 | bool IsRem = Instr->getOpcode() == Instruction::URem; |
| 892 | |
| 893 | Value *X = Instr->getOperand(i_nocapture: 0); |
| 894 | Value *Y = Instr->getOperand(i_nocapture: 1); |
| 895 | |
| 896 | // X u/ Y -> 0 iff X u< Y |
| 897 | // X u% Y -> X iff X u< Y |
| 898 | if (XCR.icmp(Pred: ICmpInst::ICMP_ULT, Other: YCR)) { |
| 899 | salvageDebugInfo(I&: *Instr); |
| 900 | Instr->replaceAllUsesWith(V: IsRem ? X : Constant::getNullValue(Ty)); |
| 901 | Instr->eraseFromParent(); |
| 902 | ++NumUDivURemsNarrowedExpanded; |
| 903 | return true; |
| 904 | } |
| 905 | |
| 906 | // Given |
| 907 | // R = X u% Y |
| 908 | // We can represent the modulo operation as a loop/self-recursion: |
| 909 | // urem_rec(X, Y): |
| 910 | // Z = X - Y |
| 911 | // if X u< Y |
| 912 | // ret X |
| 913 | // else |
| 914 | // ret urem_rec(Z, Y) |
| 915 | // which isn't better, but if we only need a single iteration |
| 916 | // to compute the answer, this becomes quite good: |
| 917 | // R = X < Y ? X : X - Y iff X u< 2*Y (w/ unsigned saturation) |
| 918 | // Now, we do not care about all full multiples of Y in X, they do not change |
| 919 | // the answer, thus we could rewrite the expression as: |
| 920 | // X* = X - (Y * |_ X / Y _|) |
| 921 | // R = X* % Y |
| 922 | // so we don't need the *first* iteration to return, we just need to |
| 923 | // know *which* iteration will always return, so we could also rewrite it as: |
| 924 | // X* = X - (Y * |_ X / Y _|) |
| 925 | // R = X* % Y iff X* u< 2*Y (w/ unsigned saturation) |
| 926 | // but that does not seem profitable here. |
| 927 | |
| 928 | // Even if we don't know X's range, the divisor may be so large, X can't ever |
| 929 | // be 2x larger than that. I.e. if divisor is always negative. |
| 930 | if (!XCR.icmp(Pred: ICmpInst::ICMP_ULT, Other: YCR.uadd_sat(Other: YCR)) && !YCR.isAllNegative()) |
| 931 | return false; |
| 932 | |
| 933 | IRBuilder<> B(Instr); |
| 934 | Value *ExpandedOp; |
| 935 | if (XCR.icmp(Pred: ICmpInst::ICMP_UGE, Other: YCR)) { |
| 936 | // If X is between Y and 2*Y the result is known. |
| 937 | if (IsRem) |
| 938 | ExpandedOp = B.CreateNUWSub(LHS: X, RHS: Y); |
| 939 | else |
| 940 | ExpandedOp = ConstantInt::get(Ty: Instr->getType(), V: 1); |
| 941 | } else if (IsRem) { |
| 942 | // NOTE: this transformation introduces two uses of X, |
| 943 | // but it may be undef so we must freeze it first. |
| 944 | Value *FrozenX = X; |
| 945 | if (!isGuaranteedNotToBeUndef(V: X)) |
| 946 | FrozenX = B.CreateFreeze(V: X, Name: X->getName() + ".frozen" ); |
| 947 | Value *FrozenY = Y; |
| 948 | if (!isGuaranteedNotToBeUndef(V: Y)) |
| 949 | FrozenY = B.CreateFreeze(V: Y, Name: Y->getName() + ".frozen" ); |
| 950 | auto *AdjX = B.CreateNUWSub(LHS: FrozenX, RHS: FrozenY, Name: Instr->getName() + ".urem" ); |
| 951 | auto *Cmp = B.CreateICmp(P: ICmpInst::ICMP_ULT, LHS: FrozenX, RHS: FrozenY, |
| 952 | Name: Instr->getName() + ".cmp" ); |
| 953 | ExpandedOp = |
| 954 | B.CreateSelectWithUnknownProfile(C: Cmp, True: FrozenX, False: AdjX, DEBUG_TYPE); |
| 955 | } else { |
| 956 | auto *Cmp = |
| 957 | B.CreateICmp(P: ICmpInst::ICMP_UGE, LHS: X, RHS: Y, Name: Instr->getName() + ".cmp" ); |
| 958 | ExpandedOp = B.CreateZExt(V: Cmp, DestTy: Ty, Name: Instr->getName() + ".udiv" ); |
| 959 | } |
| 960 | ExpandedOp->takeName(V: Instr); |
| 961 | Instr->replaceAllUsesWith(V: ExpandedOp); |
| 962 | Instr->eraseFromParent(); |
| 963 | ++NumUDivURemsNarrowedExpanded; |
| 964 | return true; |
| 965 | } |
| 966 | |
| 967 | /// Try to shrink a udiv/urem's width down to the smallest power of two that's |
| 968 | /// sufficient to contain its operands. |
| 969 | static bool narrowUDivOrURem(BinaryOperator *Instr, const ConstantRange &XCR, |
| 970 | const ConstantRange &YCR) { |
| 971 | assert(Instr->getOpcode() == Instruction::UDiv || |
| 972 | Instr->getOpcode() == Instruction::URem); |
| 973 | |
| 974 | // Find the smallest power of two bitwidth that's sufficient to hold Instr's |
| 975 | // operands. |
| 976 | |
| 977 | // What is the smallest bit width that can accommodate the entire value ranges |
| 978 | // of both of the operands? |
| 979 | unsigned MaxActiveBits = std::max(a: XCR.getActiveBits(), b: YCR.getActiveBits()); |
| 980 | // Don't shrink below 8 bits wide. |
| 981 | unsigned NewWidth = std::max<unsigned>(a: PowerOf2Ceil(A: MaxActiveBits), b: 8); |
| 982 | |
| 983 | // NewWidth might be greater than OrigWidth if OrigWidth is not a power of |
| 984 | // two. |
| 985 | if (NewWidth >= Instr->getType()->getScalarSizeInBits()) |
| 986 | return false; |
| 987 | |
| 988 | ++NumUDivURemsNarrowed; |
| 989 | IRBuilder<> B{Instr}; |
| 990 | auto *TruncTy = Instr->getType()->getWithNewBitWidth(NewBitWidth: NewWidth); |
| 991 | auto *LHS = B.CreateTruncOrBitCast(V: Instr->getOperand(i_nocapture: 0), DestTy: TruncTy, |
| 992 | Name: Instr->getName() + ".lhs.trunc" ); |
| 993 | auto *RHS = B.CreateTruncOrBitCast(V: Instr->getOperand(i_nocapture: 1), DestTy: TruncTy, |
| 994 | Name: Instr->getName() + ".rhs.trunc" ); |
| 995 | auto *BO = B.CreateBinOp(Opc: Instr->getOpcode(), LHS, RHS, Name: Instr->getName()); |
| 996 | auto *Zext = B.CreateZExt(V: BO, DestTy: Instr->getType(), Name: Instr->getName() + ".zext" ); |
| 997 | if (auto *BinOp = dyn_cast<BinaryOperator>(Val: BO)) |
| 998 | if (BinOp->getOpcode() == Instruction::UDiv) |
| 999 | BinOp->setIsExact(Instr->isExact()); |
| 1000 | |
| 1001 | Instr->replaceAllUsesWith(V: Zext); |
| 1002 | Instr->eraseFromParent(); |
| 1003 | return true; |
| 1004 | } |
| 1005 | |
| 1006 | bool CVPImpl::processUDivOrURem(BinaryOperator *Instr) { |
| 1007 | assert(Instr->getOpcode() == Instruction::UDiv || |
| 1008 | Instr->getOpcode() == Instruction::URem); |
| 1009 | ConstantRange XCR = LVI->getConstantRangeAtUse(U: Instr->getOperandUse(i: 0), |
| 1010 | /*UndefAllowed*/ false); |
| 1011 | // Allow undef for RHS, as we can assume it is division by zero UB. |
| 1012 | ConstantRange YCR = LVI->getConstantRangeAtUse(U: Instr->getOperandUse(i: 1), |
| 1013 | /*UndefAllowed*/ true); |
| 1014 | if (expandUDivOrURem(Instr, XCR, YCR)) |
| 1015 | return true; |
| 1016 | |
| 1017 | return narrowUDivOrURem(Instr, XCR, YCR); |
| 1018 | } |
| 1019 | |
| 1020 | bool CVPImpl::processSRem(BinaryOperator *SDI, const ConstantRange &LCR, |
| 1021 | const ConstantRange &RCR) { |
| 1022 | assert(SDI->getOpcode() == Instruction::SRem); |
| 1023 | |
| 1024 | if (LCR.abs().icmp(Pred: CmpInst::ICMP_ULT, Other: RCR.abs())) { |
| 1025 | salvageDebugInfo(I&: *SDI); |
| 1026 | SDI->replaceAllUsesWith(V: SDI->getOperand(i_nocapture: 0)); |
| 1027 | SDI->eraseFromParent(); |
| 1028 | return true; |
| 1029 | } |
| 1030 | |
| 1031 | struct Operand { |
| 1032 | Value *V; |
| 1033 | Domain D; |
| 1034 | }; |
| 1035 | std::array<Operand, 2> Ops = {._M_elems: {{.V: SDI->getOperand(i_nocapture: 0), .D: getDomain(CR: LCR)}, |
| 1036 | {.V: SDI->getOperand(i_nocapture: 1), .D: getDomain(CR: RCR)}}}; |
| 1037 | if (Ops[0].D == Domain::Unknown || Ops[1].D == Domain::Unknown) |
| 1038 | return false; |
| 1039 | |
| 1040 | // We know domains of both of the operands! |
| 1041 | ++NumSRems; |
| 1042 | |
| 1043 | // We need operands to be non-negative, so negate each one that isn't. |
| 1044 | for (Operand &Op : Ops) { |
| 1045 | if (Op.D == Domain::NonNegative) |
| 1046 | continue; |
| 1047 | auto *BO = BinaryOperator::CreateNeg(Op: Op.V, Name: Op.V->getName() + ".nonneg" , |
| 1048 | InsertBefore: SDI->getIterator()); |
| 1049 | BO->setDebugLoc(SDI->getDebugLoc()); |
| 1050 | Op.V = BO; |
| 1051 | } |
| 1052 | |
| 1053 | auto *URem = BinaryOperator::CreateURem(V1: Ops[0].V, V2: Ops[1].V, Name: SDI->getName(), |
| 1054 | InsertBefore: SDI->getIterator()); |
| 1055 | URem->setDebugLoc(SDI->getDebugLoc()); |
| 1056 | |
| 1057 | auto *Res = URem; |
| 1058 | |
| 1059 | // If the divident was non-positive, we need to negate the result. |
| 1060 | if (Ops[0].D == Domain::NonPositive) { |
| 1061 | Res = BinaryOperator::CreateNeg(Op: Res, Name: Res->getName() + ".neg" , |
| 1062 | InsertBefore: SDI->getIterator()); |
| 1063 | Res->setDebugLoc(SDI->getDebugLoc()); |
| 1064 | } |
| 1065 | |
| 1066 | salvageDebugInfo(I&: *SDI); |
| 1067 | SDI->replaceAllUsesWith(V: Res); |
| 1068 | SDI->eraseFromParent(); |
| 1069 | |
| 1070 | // Try to simplify our new urem. |
| 1071 | processUDivOrURem(Instr: URem); |
| 1072 | |
| 1073 | return true; |
| 1074 | } |
| 1075 | |
| 1076 | /// See if LazyValueInfo's ability to exploit edge conditions or range |
| 1077 | /// information is sufficient to prove the signs of both operands of this SDiv. |
| 1078 | /// If this is the case, replace the SDiv with a UDiv. Even for local |
| 1079 | /// conditions, this can sometimes prove conditions instcombine can't by |
| 1080 | /// exploiting range information. |
| 1081 | bool CVPImpl::processSDiv(BinaryOperator *SDI, const ConstantRange &LCR, |
| 1082 | const ConstantRange &RCR) { |
| 1083 | assert(SDI->getOpcode() == Instruction::SDiv); |
| 1084 | |
| 1085 | // Check whether the division folds to a constant. |
| 1086 | ConstantRange DivCR = LCR.sdiv(Other: RCR); |
| 1087 | if (const APInt *Elem = DivCR.getSingleElement()) { |
| 1088 | salvageDebugInfo(I&: *SDI); |
| 1089 | SDI->replaceAllUsesWith(V: ConstantInt::get(Ty: SDI->getType(), V: *Elem)); |
| 1090 | SDI->eraseFromParent(); |
| 1091 | return true; |
| 1092 | } |
| 1093 | |
| 1094 | struct Operand { |
| 1095 | Value *V; |
| 1096 | Domain D; |
| 1097 | }; |
| 1098 | std::array<Operand, 2> Ops = {._M_elems: {{.V: SDI->getOperand(i_nocapture: 0), .D: getDomain(CR: LCR)}, |
| 1099 | {.V: SDI->getOperand(i_nocapture: 1), .D: getDomain(CR: RCR)}}}; |
| 1100 | if (Ops[0].D == Domain::Unknown || Ops[1].D == Domain::Unknown) |
| 1101 | return false; |
| 1102 | |
| 1103 | // We know domains of both of the operands! |
| 1104 | ++NumSDivs; |
| 1105 | |
| 1106 | // We need operands to be non-negative, so negate each one that isn't. |
| 1107 | for (Operand &Op : Ops) { |
| 1108 | if (Op.D == Domain::NonNegative) |
| 1109 | continue; |
| 1110 | auto *BO = BinaryOperator::CreateNeg(Op: Op.V, Name: Op.V->getName() + ".nonneg" , |
| 1111 | InsertBefore: SDI->getIterator()); |
| 1112 | BO->setDebugLoc(SDI->getDebugLoc()); |
| 1113 | Op.V = BO; |
| 1114 | } |
| 1115 | |
| 1116 | auto *UDiv = BinaryOperator::CreateUDiv(V1: Ops[0].V, V2: Ops[1].V, Name: SDI->getName(), |
| 1117 | InsertBefore: SDI->getIterator()); |
| 1118 | UDiv->setDebugLoc(SDI->getDebugLoc()); |
| 1119 | UDiv->setIsExact(SDI->isExact()); |
| 1120 | |
| 1121 | auto *Res = UDiv; |
| 1122 | |
| 1123 | // If the operands had two different domains, we need to negate the result. |
| 1124 | if (Ops[0].D != Ops[1].D) { |
| 1125 | Res = BinaryOperator::CreateNeg(Op: Res, Name: Res->getName() + ".neg" , |
| 1126 | InsertBefore: SDI->getIterator()); |
| 1127 | Res->setDebugLoc(SDI->getDebugLoc()); |
| 1128 | } |
| 1129 | |
| 1130 | salvageDebugInfo(I&: *SDI); |
| 1131 | SDI->replaceAllUsesWith(V: Res); |
| 1132 | SDI->eraseFromParent(); |
| 1133 | |
| 1134 | // Try to simplify our new udiv. |
| 1135 | processUDivOrURem(Instr: UDiv); |
| 1136 | |
| 1137 | return true; |
| 1138 | } |
| 1139 | |
| 1140 | bool CVPImpl::processSDivOrSRem(BinaryOperator *Instr) { |
| 1141 | assert(Instr->getOpcode() == Instruction::SDiv || |
| 1142 | Instr->getOpcode() == Instruction::SRem); |
| 1143 | ConstantRange LCR = |
| 1144 | LVI->getConstantRangeAtUse(U: Instr->getOperandUse(i: 0), /*AllowUndef*/ UndefAllowed: false); |
| 1145 | // Allow undef for RHS, as we can assume it is division by zero UB. |
| 1146 | ConstantRange RCR = |
| 1147 | LVI->getConstantRangeAtUse(U: Instr->getOperandUse(i: 1), /*AlloweUndef*/ UndefAllowed: true); |
| 1148 | if (Instr->getOpcode() == Instruction::SDiv) |
| 1149 | if (processSDiv(SDI: Instr, LCR, RCR)) |
| 1150 | return true; |
| 1151 | |
| 1152 | if (Instr->getOpcode() == Instruction::SRem) { |
| 1153 | if (processSRem(SDI: Instr, LCR, RCR)) |
| 1154 | return true; |
| 1155 | } |
| 1156 | |
| 1157 | return narrowSDivOrSRem(Instr, LCR, RCR); |
| 1158 | } |
| 1159 | |
| 1160 | bool CVPImpl::processAShr(BinaryOperator *SDI) { |
| 1161 | ConstantRange LRange = |
| 1162 | LVI->getConstantRangeAtUse(U: SDI->getOperandUse(i: 0), /*UndefAllowed*/ false); |
| 1163 | unsigned OrigWidth = SDI->getType()->getScalarSizeInBits(); |
| 1164 | ConstantRange NegOneOrZero = |
| 1165 | ConstantRange(APInt(OrigWidth, (uint64_t)-1, true), APInt(OrigWidth, 1)); |
| 1166 | if (NegOneOrZero.contains(CR: LRange)) { |
| 1167 | // ashr of -1 or 0 never changes the value, so drop the whole instruction |
| 1168 | ++NumAShrsRemoved; |
| 1169 | salvageDebugInfo(I&: *SDI); |
| 1170 | SDI->replaceAllUsesWith(V: SDI->getOperand(i_nocapture: 0)); |
| 1171 | SDI->eraseFromParent(); |
| 1172 | return true; |
| 1173 | } |
| 1174 | |
| 1175 | if (!LRange.isAllNonNegative()) |
| 1176 | return false; |
| 1177 | |
| 1178 | ++NumAShrsConverted; |
| 1179 | auto *BO = BinaryOperator::CreateLShr(V1: SDI->getOperand(i_nocapture: 0), V2: SDI->getOperand(i_nocapture: 1), |
| 1180 | Name: "" , InsertBefore: SDI->getIterator()); |
| 1181 | BO->takeName(V: SDI); |
| 1182 | BO->setDebugLoc(SDI->getDebugLoc()); |
| 1183 | BO->setIsExact(SDI->isExact()); |
| 1184 | salvageDebugInfo(I&: *SDI); |
| 1185 | SDI->replaceAllUsesWith(V: BO); |
| 1186 | SDI->eraseFromParent(); |
| 1187 | |
| 1188 | return true; |
| 1189 | } |
| 1190 | |
| 1191 | bool CVPImpl::processSExt(SExtInst *SDI) { |
| 1192 | const Use &Base = SDI->getOperandUse(i: 0); |
| 1193 | if (!LVI->getConstantRangeAtUse(U: Base, /*UndefAllowed*/ false) |
| 1194 | .isAllNonNegative()) |
| 1195 | return false; |
| 1196 | |
| 1197 | ++NumSExt; |
| 1198 | auto *ZExt = CastInst::CreateZExtOrBitCast(S: Base, Ty: SDI->getType(), Name: "" , |
| 1199 | InsertBefore: SDI->getIterator()); |
| 1200 | ZExt->takeName(V: SDI); |
| 1201 | ZExt->setDebugLoc(SDI->getDebugLoc()); |
| 1202 | ZExt->setNonNeg(); |
| 1203 | salvageDebugInfo(I&: *SDI); |
| 1204 | SDI->replaceAllUsesWith(V: ZExt); |
| 1205 | SDI->eraseFromParent(); |
| 1206 | |
| 1207 | return true; |
| 1208 | } |
| 1209 | |
| 1210 | bool CVPImpl::processPossibleNonNeg(PossiblyNonNegInst *I) { |
| 1211 | if (I->hasNonNeg()) |
| 1212 | return false; |
| 1213 | |
| 1214 | const Use &Base = I->getOperandUse(i: 0); |
| 1215 | if (!LVI->getConstantRangeAtUse(U: Base, /*UndefAllowed*/ false) |
| 1216 | .isAllNonNegative()) |
| 1217 | return false; |
| 1218 | |
| 1219 | ++NumNNeg; |
| 1220 | I->setNonNeg(); |
| 1221 | |
| 1222 | return true; |
| 1223 | } |
| 1224 | |
| 1225 | bool CVPImpl::processSIToFP(SIToFPInst *SIToFP) { |
| 1226 | const Use &Base = SIToFP->getOperandUse(i: 0); |
| 1227 | if (!LVI->getConstantRangeAtUse(U: Base, /*UndefAllowed*/ false) |
| 1228 | .isAllNonNegative()) |
| 1229 | return false; |
| 1230 | |
| 1231 | ++NumSIToFP; |
| 1232 | auto *UIToFP = CastInst::Create(Instruction::UIToFP, S: Base, Ty: SIToFP->getType(), |
| 1233 | Name: "" , InsertBefore: SIToFP->getIterator()); |
| 1234 | UIToFP->takeName(V: SIToFP); |
| 1235 | UIToFP->setDebugLoc(SIToFP->getDebugLoc()); |
| 1236 | UIToFP->setNonNeg(); |
| 1237 | salvageDebugInfo(I&: *SIToFP); |
| 1238 | SIToFP->replaceAllUsesWith(V: UIToFP); |
| 1239 | SIToFP->eraseFromParent(); |
| 1240 | |
| 1241 | return true; |
| 1242 | } |
| 1243 | |
| 1244 | namespace { |
| 1245 | struct NoWrapFlags { |
| 1246 | bool NSW = false; |
| 1247 | bool NUW = false; |
| 1248 | }; |
| 1249 | } // namespace |
| 1250 | |
| 1251 | // Check if the requested no-wrap flags are valid for \p Opcode on \p LRange and |
| 1252 | // \p RRange. |
| 1253 | static NoWrapFlags computeNoWrapFlags(Instruction::BinaryOps Opcode, |
| 1254 | const ConstantRange &LRange, |
| 1255 | const ConstantRange &RRange, |
| 1256 | bool CheckNSW, bool CheckNUW) { |
| 1257 | using OBO = OverflowingBinaryOperator; |
| 1258 | NoWrapFlags Flags; |
| 1259 | if (CheckNUW) |
| 1260 | Flags.NUW = ConstantRange::makeGuaranteedNoWrapRegion(BinOp: Opcode, Other: RRange, |
| 1261 | NoWrapKind: OBO::NoUnsignedWrap) |
| 1262 | .contains(CR: LRange); |
| 1263 | if (CheckNSW) |
| 1264 | Flags.NSW = ConstantRange::makeGuaranteedNoWrapRegion(BinOp: Opcode, Other: RRange, |
| 1265 | NoWrapKind: OBO::NoSignedWrap) |
| 1266 | .contains(CR: LRange); |
| 1267 | return Flags; |
| 1268 | } |
| 1269 | |
| 1270 | // Try to prove that \p BinOp does not wrap by looking at the operand ranges |
| 1271 | // constrained at each of its use sites, rather than at the definition. This |
| 1272 | // improves results, e.g. when all uses are constrained by a runtime check. |
| 1273 | NoWrapFlags CVPImpl::inferNoWrapFromUses(BinaryOperator *BinOp, bool WantNSW, |
| 1274 | bool WantNUW) { |
| 1275 | // Skip analysis, when there are too many uses to check or any use is in the |
| 1276 | // same block. |
| 1277 | const unsigned MaxUsesToInspect = 4; |
| 1278 | BasicBlock *DefBB = BinOp->getParent(); |
| 1279 | unsigned NumUses = 0; |
| 1280 | for (Use &U : BinOp->uses()) { |
| 1281 | if (++NumUses > MaxUsesToInspect) |
| 1282 | return {}; |
| 1283 | auto *UserI = cast<Instruction>(Val: U.getUser()); |
| 1284 | if (isa<PHINode>(Val: UserI) || UserI->getParent() == DefBB) |
| 1285 | return {}; |
| 1286 | } |
| 1287 | if (NumUses == 0) |
| 1288 | return {}; |
| 1289 | |
| 1290 | Instruction::BinaryOps Opcode = BinOp->getOpcode(); |
| 1291 | NoWrapFlags Flags; |
| 1292 | Flags.NSW = WantNSW; |
| 1293 | Flags.NUW = WantNUW; |
| 1294 | for (Use &U : BinOp->uses()) { |
| 1295 | auto *UserI = cast<Instruction>(Val: U.getUser()); |
| 1296 | // Constrain both operands at this use site and see which flags still hold. |
| 1297 | ConstantRange LRange = LVI->getConstantRange(V: BinOp->getOperand(i_nocapture: 0), CtxI: UserI, |
| 1298 | /*UndefAllowed=*/false); |
| 1299 | ConstantRange RRange = LVI->getConstantRange(V: BinOp->getOperand(i_nocapture: 1), CtxI: UserI, |
| 1300 | /*UndefAllowed=*/false); |
| 1301 | Flags = computeNoWrapFlags(Opcode, LRange, RRange, CheckNSW: Flags.NSW, CheckNUW: Flags.NUW); |
| 1302 | if (!Flags.NSW && !Flags.NUW) |
| 1303 | return {}; |
| 1304 | } |
| 1305 | |
| 1306 | return Flags; |
| 1307 | } |
| 1308 | |
| 1309 | bool CVPImpl::processBinOp(BinaryOperator *BinOp) { |
| 1310 | bool NSW = BinOp->hasNoSignedWrap(); |
| 1311 | bool NUW = BinOp->hasNoUnsignedWrap(); |
| 1312 | if (NSW && NUW) |
| 1313 | return false; |
| 1314 | |
| 1315 | Instruction::BinaryOps Opcode = BinOp->getOpcode(); |
| 1316 | ConstantRange LRange = LVI->getConstantRangeAtUse(U: BinOp->getOperandUse(i: 0), |
| 1317 | /*UndefAllowed=*/false); |
| 1318 | ConstantRange RRange = LVI->getConstantRangeAtUse(U: BinOp->getOperandUse(i: 1), |
| 1319 | /*UndefAllowed=*/false); |
| 1320 | |
| 1321 | NoWrapFlags New = |
| 1322 | computeNoWrapFlags(Opcode, LRange, RRange, /*CheckNSW=*/!NSW, |
| 1323 | /*CheckNUW=*/!NUW); |
| 1324 | |
| 1325 | // If a still-wanted flag could not be proven at the definition, retry using |
| 1326 | // the operand ranges constrained at the use sites. This is the more |
| 1327 | // expensive path, so it only runs when the cheap query above came up short. |
| 1328 | bool WantNSW = !NSW && !New.NSW; |
| 1329 | bool WantNUW = !NUW && !New.NUW; |
| 1330 | if (WantNSW || WantNUW) { |
| 1331 | NoWrapFlags FromUses = inferNoWrapFromUses(BinOp, WantNSW, WantNUW); |
| 1332 | New.NSW |= FromUses.NSW; |
| 1333 | New.NUW |= FromUses.NUW; |
| 1334 | } |
| 1335 | |
| 1336 | setDeducedOverflowingFlags(V: BinOp, Opcode, NewNSW: New.NSW, NewNUW: New.NUW); |
| 1337 | |
| 1338 | return New.NSW || New.NUW; |
| 1339 | } |
| 1340 | |
| 1341 | bool CVPImpl::processAnd(BinaryOperator *BinOp) { |
| 1342 | using namespace llvm::PatternMatch; |
| 1343 | |
| 1344 | // Pattern match (and lhs, C) where C includes a superset of bits which might |
| 1345 | // be set in lhs. This is a common truncation idiom created by instcombine. |
| 1346 | const Use &LHS = BinOp->getOperandUse(i: 0); |
| 1347 | const APInt *RHS; |
| 1348 | if (!match(V: BinOp->getOperand(i_nocapture: 1), P: m_LowBitMask(V&: RHS))) |
| 1349 | return false; |
| 1350 | |
| 1351 | // We can only replace the AND with LHS based on range info if the range does |
| 1352 | // not include undef. |
| 1353 | ConstantRange LRange = |
| 1354 | LVI->getConstantRangeAtUse(U: LHS, /*UndefAllowed=*/false); |
| 1355 | if (!LRange.getUnsignedMax().ule(RHS: *RHS)) |
| 1356 | return false; |
| 1357 | |
| 1358 | salvageDebugInfo(I&: *BinOp); |
| 1359 | BinOp->replaceAllUsesWith(V: LHS); |
| 1360 | BinOp->eraseFromParent(); |
| 1361 | NumAnd++; |
| 1362 | return true; |
| 1363 | } |
| 1364 | |
| 1365 | bool CVPImpl::processTrunc(TruncInst *TI) { |
| 1366 | if (TI->hasNoSignedWrap() && TI->hasNoUnsignedWrap()) |
| 1367 | return false; |
| 1368 | |
| 1369 | ConstantRange Range = |
| 1370 | LVI->getConstantRangeAtUse(U: TI->getOperandUse(i: 0), /*UndefAllowed=*/false); |
| 1371 | uint64_t DestWidth = TI->getDestTy()->getScalarSizeInBits(); |
| 1372 | bool Changed = false; |
| 1373 | |
| 1374 | if (!TI->hasNoUnsignedWrap()) { |
| 1375 | if (Range.getActiveBits() <= DestWidth) { |
| 1376 | TI->setHasNoUnsignedWrap(true); |
| 1377 | ++NumNUW; |
| 1378 | Changed = true; |
| 1379 | } |
| 1380 | } |
| 1381 | |
| 1382 | if (!TI->hasNoSignedWrap()) { |
| 1383 | if (Range.getMinSignedBits() <= DestWidth) { |
| 1384 | TI->setHasNoSignedWrap(true); |
| 1385 | ++NumNSW; |
| 1386 | Changed = true; |
| 1387 | } |
| 1388 | } |
| 1389 | |
| 1390 | return Changed; |
| 1391 | } |
| 1392 | |
| 1393 | bool CVPImpl::processReturn(ReturnInst *RI) { |
| 1394 | // Try to determine the return value if we can. This is mainly here to |
| 1395 | // simplify the writing of unit tests, but also helps to enable IPO by |
| 1396 | // constant folding the return values of callees. |
| 1397 | Value *RetVal = RI->getReturnValue(); |
| 1398 | if (!RetVal) |
| 1399 | return false; // handle "ret void" |
| 1400 | if (RetRange && !RetRange->isFullSet()) |
| 1401 | RetRange = RetRange->unionWith( |
| 1402 | CR: LVI->getConstantRange(V: RetVal, CtxI: RI, /*UndefAllowed=*/false)); |
| 1403 | |
| 1404 | if (isa<Constant>(Val: RetVal)) |
| 1405 | return false; // nothing to do |
| 1406 | if (Constant *C = getConstantAt(V: RetVal, At: RI)) { |
| 1407 | ++NumReturns; |
| 1408 | RI->replaceUsesOfWith(From: RetVal, To: C); |
| 1409 | return true; |
| 1410 | } |
| 1411 | return false; |
| 1412 | } |
| 1413 | |
| 1414 | bool CVPImpl::processInstruction(Instruction &I) { |
| 1415 | switch (I.getOpcode()) { |
| 1416 | case Instruction::Select: |
| 1417 | return processSelect(S: cast<SelectInst>(Val: &I)); |
| 1418 | case Instruction::PHI: |
| 1419 | return processPHI(P: cast<PHINode>(Val: &I)); |
| 1420 | case Instruction::ICmp: |
| 1421 | case Instruction::FCmp: |
| 1422 | return processCmp(Cmp: cast<CmpInst>(Val: &I)); |
| 1423 | case Instruction::Call: |
| 1424 | case Instruction::Invoke: |
| 1425 | return processCallSite(CB&: cast<CallBase>(Val&: I)); |
| 1426 | case Instruction::SRem: |
| 1427 | case Instruction::SDiv: |
| 1428 | return processSDivOrSRem(Instr: cast<BinaryOperator>(Val: &I)); |
| 1429 | case Instruction::UDiv: |
| 1430 | case Instruction::URem: |
| 1431 | return processUDivOrURem(Instr: cast<BinaryOperator>(Val: &I)); |
| 1432 | case Instruction::AShr: |
| 1433 | return processAShr(SDI: cast<BinaryOperator>(Val: &I)); |
| 1434 | case Instruction::SExt: |
| 1435 | return processSExt(SDI: cast<SExtInst>(Val: &I)); |
| 1436 | case Instruction::ZExt: |
| 1437 | case Instruction::UIToFP: |
| 1438 | return processPossibleNonNeg(I: cast<PossiblyNonNegInst>(Val: &I)); |
| 1439 | case Instruction::SIToFP: |
| 1440 | return processSIToFP(SIToFP: cast<SIToFPInst>(Val: &I)); |
| 1441 | case Instruction::Add: |
| 1442 | case Instruction::Sub: |
| 1443 | case Instruction::Mul: |
| 1444 | case Instruction::Shl: |
| 1445 | return processBinOp(BinOp: cast<BinaryOperator>(Val: &I)); |
| 1446 | case Instruction::And: |
| 1447 | return processAnd(BinOp: cast<BinaryOperator>(Val: &I)); |
| 1448 | case Instruction::Trunc: |
| 1449 | return processTrunc(TI: cast<TruncInst>(Val: &I)); |
| 1450 | case Instruction::Switch: |
| 1451 | return processSwitch(I: cast<SwitchInst>(Val: &I)); |
| 1452 | case Instruction::Ret: |
| 1453 | return processReturn(RI: cast<ReturnInst>(Val: &I)); |
| 1454 | default: |
| 1455 | return false; |
| 1456 | } |
| 1457 | } |
| 1458 | |
| 1459 | bool CVPImpl::inferReturnRangeAttr() { |
| 1460 | // Infer range attribute on return value. |
| 1461 | if (!RetRange || RetRange->isFullSet()) |
| 1462 | return false; |
| 1463 | Attribute RangeAttr = F.getRetAttribute(Kind: Attribute::Range); |
| 1464 | if (RangeAttr.isValid()) |
| 1465 | RetRange = RetRange->intersectWith(CR: RangeAttr.getRange()); |
| 1466 | // Don't add attribute for constant integer returns to reduce noise. These |
| 1467 | // are propagated across functions by IPSCCP. |
| 1468 | if (RetRange->isEmptySet() || RetRange->isSingleElement()) |
| 1469 | return false; |
| 1470 | F.addRangeRetAttr(CR: *RetRange); |
| 1471 | return true; |
| 1472 | } |
| 1473 | |
| 1474 | bool CVPImpl::run() { |
| 1475 | if (F.hasExactDefinition() && F.getReturnType()->isIntOrIntVectorTy()) |
| 1476 | RetRange = |
| 1477 | ConstantRange::getEmpty(BitWidth: F.getReturnType()->getScalarSizeInBits()); |
| 1478 | |
| 1479 | bool Changed = false; |
| 1480 | // Visiting in a pre-order depth-first traversal causes us to simplify early |
| 1481 | // blocks before querying later blocks (which require us to analyze early |
| 1482 | // blocks). Eagerly simplifying shallow blocks means there is strictly less |
| 1483 | // work to do for deep blocks. This also means we don't visit unreachable |
| 1484 | // blocks. |
| 1485 | for (BasicBlock *BB : depth_first(G: &F.getEntryBlock())) |
| 1486 | for (Instruction &I : llvm::make_early_inc_range(Range&: *BB)) |
| 1487 | Changed |= processInstruction(I); |
| 1488 | |
| 1489 | Changed |= inferReturnRangeAttr(); |
| 1490 | return Changed; |
| 1491 | } |
| 1492 | |
| 1493 | PreservedAnalyses |
| 1494 | CorrelatedValuePropagationPass::run(Function &F, FunctionAnalysisManager &AM) { |
| 1495 | LazyValueInfo *LVI = &AM.getResult<LazyValueAnalysis>(IR&: F); |
| 1496 | DominatorTree *DT = &AM.getResult<DominatorTreeAnalysis>(IR&: F); |
| 1497 | |
| 1498 | CVPImpl Impl(F, LVI, DT, getBestSimplifyQuery(AM, F)); |
| 1499 | bool Changed = Impl.run(); |
| 1500 | |
| 1501 | PreservedAnalyses PA; |
| 1502 | if (!Changed) { |
| 1503 | PA = PreservedAnalyses::all(); |
| 1504 | } else { |
| 1505 | #if defined(EXPENSIVE_CHECKS) |
| 1506 | assert(DT->verify(DominatorTree::VerificationLevel::Full)); |
| 1507 | #else |
| 1508 | assert(DT->verify(DominatorTree::VerificationLevel::Fast)); |
| 1509 | #endif // EXPENSIVE_CHECKS |
| 1510 | |
| 1511 | PA.preserve<DominatorTreeAnalysis>(); |
| 1512 | PA.preserve<LazyValueAnalysis>(); |
| 1513 | } |
| 1514 | |
| 1515 | // Keeping LVI alive is expensive, both because it uses a lot of memory, and |
| 1516 | // because invalidating values in LVI is expensive. While CVP does preserve |
| 1517 | // LVI, we know that passes after JumpThreading+CVP will not need the result |
| 1518 | // of this analysis, so we forcefully discard it early. |
| 1519 | PA.abandon<LazyValueAnalysis>(); |
| 1520 | return PA; |
| 1521 | } |
| 1522 | |