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