Reapply "[LAA,Loads] Use loop guards and max BTC if needed when checking deref. (llvm#155672)"

fhahn · fhahn · commit 57bd0c8e0b80 · 2025-11-04T11:13:41.000Z
This reverts commit f0df1e3. Recommit with extra check for SCEVCouldNotCompute. Test has been added in b169302. Original message: Remove the fall-back to constant max BTC if the backedge-taken-count cannot be computed. The constant max backedge-taken count is computed considering loop guards, so to avoid regressions we need to apply loop guards as needed. Also remove the special handling for Mul in willNotOverflow, as this should not longer be needed after 9143746 (llvm#155300). PR: llvm#155672 (cherry picked from commit a434a7a)
diff --git a/llvm/include/llvm/Analysis/LoopAccessAnalysis.h b/llvm/include/llvm/Analysis/LoopAccessAnalysis.h
@@ -183,10 +183,12 @@ class MemoryDepChecker {
   MemoryDepChecker(PredicatedScalarEvolution &PSE, AssumptionCache *AC,
                    DominatorTree *DT, const Loop *L,
                    const DenseMap<Value *, const SCEV *> &SymbolicStrides,
-                   unsigned MaxTargetVectorWidthInBits)
+                   unsigned MaxTargetVectorWidthInBits,
+                   std::optional<ScalarEvolution::LoopGuards> &LoopGuards)
       : PSE(PSE), AC(AC), DT(DT), InnermostLoop(L),
         SymbolicStrides(SymbolicStrides),
-        MaxTargetVectorWidthInBits(MaxTargetVectorWidthInBits) {}
+        MaxTargetVectorWidthInBits(MaxTargetVectorWidthInBits),
+        LoopGuards(LoopGuards) {}
 
   /// Register the location (instructions are given increasing numbers)
   /// of a write access.
@@ -373,7 +375,7 @@ class MemoryDepChecker {
       PointerBounds;
 
   /// Cache for the loop guards of InnermostLoop.
-  std::optional<ScalarEvolution::LoopGuards> LoopGuards;
+  std::optional<ScalarEvolution::LoopGuards> &LoopGuards;
 
   /// Check whether there is a plausible dependence between the two
   /// accesses.
@@ -531,8 +533,9 @@ class RuntimePointerChecking {
           AliasSetId(AliasSetId), Expr(Expr), NeedsFreeze(NeedsFreeze) {}
   };
 
-  RuntimePointerChecking(MemoryDepChecker &DC, ScalarEvolution *SE)
-      : DC(DC), SE(SE) {}
+  RuntimePointerChecking(MemoryDepChecker &DC, ScalarEvolution *SE,
+                         std::optional<ScalarEvolution::LoopGuards> &LoopGuards)
+      : DC(DC), SE(SE), LoopGuards(LoopGuards) {}
 
   /// Reset the state of the pointer runtime information.
   void reset() {
@@ -646,6 +649,9 @@ class RuntimePointerChecking {
   /// Holds a pointer to the ScalarEvolution analysis.
   ScalarEvolution *SE;
 
+  /// Cache for the loop guards of the loop.
+  std::optional<ScalarEvolution::LoopGuards> &LoopGuards;
+
   /// Set of run-time checks required to establish independence of
   /// otherwise may-aliasing pointers in the loop.
   SmallVector<RuntimePointerCheck, 4> Checks;
@@ -821,6 +827,9 @@ class LoopAccessInfo {
 
   Loop *TheLoop;
 
+  /// Cache for the loop guards of TheLoop.
+  std::optional<ScalarEvolution::LoopGuards> LoopGuards;
+
   /// Determines whether we should generate partial runtime checks when not all
   /// memory accesses could be analyzed.
   bool AllowPartial;
@@ -938,7 +947,8 @@ LLVM_ABI std::pair<const SCEV *, const SCEV *> getStartAndEndForAccess(
     const SCEV *MaxBTC, ScalarEvolution *SE,
     DenseMap<std::pair<const SCEV *, Type *>,
              std::pair<const SCEV *, const SCEV *>> *PointerBounds,
-    DominatorTree *DT, AssumptionCache *AC);
+    DominatorTree *DT, AssumptionCache *AC,
+    std::optional<ScalarEvolution::LoopGuards> &LoopGuards);
 
 class LoopAccessInfoManager {
   /// The cache.
diff --git a/llvm/lib/Analysis/Loads.cpp b/llvm/lib/Analysis/Loads.cpp
@@ -26,10 +26,6 @@
 
 using namespace llvm;
 
-static cl::opt<bool>
-    UseSymbolicMaxBTCForDerefInLoop("use-symbolic-maxbtc-deref-loop",
-                                    cl::init(false));
-
 static bool isAligned(const Value *Base, Align Alignment,
                       const DataLayout &DL) {
   return Base->getPointerAlignment(DL) >= Alignment;
@@ -336,18 +332,10 @@ bool llvm::isDereferenceableAndAlignedInLoop(
                             : SE.getBackedgeTakenCount(L);
   if (isa<SCEVCouldNotCompute>(MaxBECount))
     return false;
-
-  if (isa<SCEVCouldNotCompute>(BECount) && !UseSymbolicMaxBTCForDerefInLoop) {
-    // TODO: Support symbolic max backedge taken counts for loops without
-    // computable backedge taken counts.
-    MaxBECount =
-        Predicates
-            ? SE.getPredicatedConstantMaxBackedgeTakenCount(L, *Predicates)
-            : SE.getConstantMaxBackedgeTakenCount(L);
-  }
-
-  const auto &[AccessStart, AccessEnd] = getStartAndEndForAccess(
-      L, PtrScev, LI->getType(), BECount, MaxBECount, &SE, nullptr, &DT, AC);
+  std::optional<ScalarEvolution::LoopGuards> LoopGuards;
+  const auto &[AccessStart, AccessEnd] =
+      getStartAndEndForAccess(L, PtrScev, LI->getType(), BECount, MaxBECount,
+                              &SE, nullptr, &DT, AC, LoopGuards);
   if (isa<SCEVCouldNotCompute>(AccessStart) ||
       isa<SCEVCouldNotCompute>(AccessEnd))
     return false;
@@ -356,7 +344,13 @@ bool llvm::isDereferenceableAndAlignedInLoop(
   const SCEV *PtrDiff = SE.getMinusSCEV(AccessEnd, AccessStart);
   if (isa<SCEVCouldNotCompute>(PtrDiff))
     return false;
-  APInt MaxPtrDiff = SE.getUnsignedRangeMax(PtrDiff);
+
+  if (!LoopGuards)
+    LoopGuards.emplace(
+        ScalarEvolution::LoopGuards::collect(AddRec->getLoop(), SE));
+
+  APInt MaxPtrDiff =
+      SE.getUnsignedRangeMax(SE.applyLoopGuards(PtrDiff, *LoopGuards));
 
   Value *Base = nullptr;
   APInt AccessSize;
@@ -399,9 +393,10 @@ bool llvm::isDereferenceableAndAlignedInLoop(
   Instruction *HeaderFirstNonPHI = &*L->getHeader()->getFirstNonPHIIt();
   return isDereferenceableAndAlignedPointerViaAssumption(
              Base, Alignment,
-             [&SE, AccessSizeSCEV](const RetainedKnowledge &RK) {
-               return SE.isKnownPredicate(CmpInst::ICMP_ULE, AccessSizeSCEV,
-                                          SE.getSCEV(RK.IRArgValue));
+             [&SE, AccessSizeSCEV, &LoopGuards](const RetainedKnowledge &RK) {
+               return SE.isKnownPredicate(
+                   CmpInst::ICMP_ULE, AccessSizeSCEV,
+                   SE.applyLoopGuards(SE.getSCEV(RK.IRArgValue), *LoopGuards));
              },
              DL, HeaderFirstNonPHI, AC, &DT) ||
          isDereferenceableAndAlignedPointer(Base, Alignment, AccessSize, DL,
diff --git a/llvm/lib/Analysis/LoopAccessAnalysis.cpp b/llvm/lib/Analysis/LoopAccessAnalysis.cpp
@@ -193,30 +193,28 @@ RuntimeCheckingPtrGroup::RuntimeCheckingPtrGroup(
 /// Returns \p A + \p B, if it is guaranteed not to unsigned wrap. Otherwise
 /// return nullptr. \p A and \p B must have the same type.
 static const SCEV *addSCEVNoOverflow(const SCEV *A, const SCEV *B,
-                                     ScalarEvolution &SE,
-                                     const Instruction *CtxI) {
-  if (!SE.willNotOverflow(Instruction::Add, /*IsSigned=*/false, A, B, CtxI))
+                                     ScalarEvolution &SE) {
+  if (!SE.willNotOverflow(Instruction::Add, /*IsSigned=*/false, A, B))
     return nullptr;
   return SE.getAddExpr(A, B);
 }
 
 /// Returns \p A * \p B, if it is guaranteed not to unsigned wrap. Otherwise
 /// return nullptr. \p A and \p B must have the same type.
 static const SCEV *mulSCEVOverflow(const SCEV *A, const SCEV *B,
-                                   ScalarEvolution &SE,
-                                   const Instruction *CtxI) {
-  if (!SE.willNotOverflow(Instruction::Mul, /*IsSigned=*/false, A, B, CtxI))
+                                   ScalarEvolution &SE) {
+  if (!SE.willNotOverflow(Instruction::Mul, /*IsSigned=*/false, A, B))
     return nullptr;
   return SE.getMulExpr(A, B);
 }
 
 /// Return true, if evaluating \p AR at \p MaxBTC cannot wrap, because \p AR at
 /// \p MaxBTC is guaranteed inbounds of the accessed object.
-static bool
-evaluatePtrAddRecAtMaxBTCWillNotWrap(const SCEVAddRecExpr *AR,
-                                     const SCEV *MaxBTC, const SCEV *EltSize,
-                                     ScalarEvolution &SE, const DataLayout &DL,
-                                     DominatorTree *DT, AssumptionCache *AC) {
+static bool evaluatePtrAddRecAtMaxBTCWillNotWrap(
+    const SCEVAddRecExpr *AR, const SCEV *MaxBTC, const SCEV *EltSize,
+    ScalarEvolution &SE, const DataLayout &DL, DominatorTree *DT,
+    AssumptionCache *AC,
+    std::optional<ScalarEvolution::LoopGuards> &LoopGuards) {
   auto *PointerBase = SE.getPointerBase(AR->getStart());
   auto *StartPtr = dyn_cast<SCEVUnknown>(PointerBase);
   if (!StartPtr)
@@ -234,12 +232,11 @@ evaluatePtrAddRecAtMaxBTCWillNotWrap(const SCEVAddRecExpr *AR,
   Type *WiderTy = SE.getWiderType(MaxBTC->getType(), Step->getType());
   const SCEV *DerefBytesSCEV = SE.getConstant(WiderTy, DerefBytes);
 
-  // Context which dominates the entire loop.
-  auto *CtxI = L->getLoopPredecessor()->getTerminator();
   // Check if we have a suitable dereferencable assumption we can use.
   if (!StartPtrV->canBeFreed()) {
     RetainedKnowledge DerefRK = getKnowledgeValidInContext(
-        StartPtrV, {Attribute::Dereferenceable}, *AC, CtxI, DT);
+        StartPtrV, {Attribute::Dereferenceable}, *AC,
+        L->getLoopPredecessor()->getTerminator(), DT);
     if (DerefRK) {
       DerefBytesSCEV = SE.getUMaxExpr(
           DerefBytesSCEV, SE.getConstant(WiderTy, DerefRK.ArgValue));
@@ -263,23 +260,39 @@ evaluatePtrAddRecAtMaxBTCWillNotWrap(const SCEVAddRecExpr *AR,
       SE.getMinusSCEV(AR->getStart(), StartPtr), WiderTy);
 
   const SCEV *OffsetAtLastIter =
-      mulSCEVOverflow(MaxBTC, SE.getAbsExpr(Step, /*IsNSW=*/false), SE, CtxI);
-  if (!OffsetAtLastIter)
-    return false;
+      mulSCEVOverflow(MaxBTC, SE.getAbsExpr(Step, /*IsNSW=*/false), SE);
+  if (!OffsetAtLastIter) {
+    // Re-try with constant max backedge-taken count if using the symbolic one
+    // failed.
+    MaxBTC = SE.getConstantMaxBackedgeTakenCount(AR->getLoop());
+    if (isa<SCEVCouldNotCompute>(MaxBTC))
+      return false;
+    MaxBTC = SE.getNoopOrZeroExtend(
+        MaxBTC, WiderTy);
+    OffsetAtLastIter =
+        mulSCEVOverflow(MaxBTC, SE.getAbsExpr(Step, /*IsNSW=*/false), SE);
+    if (!OffsetAtLastIter)
+      return false;
+  }
 
   const SCEV *OffsetEndBytes = addSCEVNoOverflow(
-      OffsetAtLastIter, SE.getNoopOrZeroExtend(EltSize, WiderTy), SE, CtxI);
+      OffsetAtLastIter, SE.getNoopOrZeroExtend(EltSize, WiderTy), SE);
   if (!OffsetEndBytes)
     return false;
 
   if (IsKnownNonNegative) {
     // For positive steps, check if
     //  (AR->getStart() - StartPtr) + (MaxBTC  * Step) + EltSize <= DerefBytes,
     // while making sure none of the computations unsigned wrap themselves.
-    const SCEV *EndBytes =
-        addSCEVNoOverflow(StartOffset, OffsetEndBytes, SE, CtxI);
+    const SCEV *EndBytes = addSCEVNoOverflow(StartOffset, OffsetEndBytes, SE);
     if (!EndBytes)
       return false;
+
+    if (!LoopGuards)
+      LoopGuards.emplace(
+          ScalarEvolution::LoopGuards::collect(AR->getLoop(), SE));
+
+    EndBytes = SE.applyLoopGuards(EndBytes, *LoopGuards);
     return SE.isKnownPredicate(CmpInst::ICMP_ULE, EndBytes, DerefBytesSCEV);
   }
 
@@ -296,7 +309,8 @@ std::pair<const SCEV *, const SCEV *> llvm::getStartAndEndForAccess(
     const SCEV *MaxBTC, ScalarEvolution *SE,
     DenseMap<std::pair<const SCEV *, Type *>,
              std::pair<const SCEV *, const SCEV *>> *PointerBounds,
-    DominatorTree *DT, AssumptionCache *AC) {
+    DominatorTree *DT, AssumptionCache *AC,
+    std::optional<ScalarEvolution::LoopGuards> &LoopGuards) {
   std::pair<const SCEV *, const SCEV *> *PtrBoundsPair;
   if (PointerBounds) {
     auto [Iter, Ins] = PointerBounds->insert(
@@ -332,7 +346,7 @@ std::pair<const SCEV *, const SCEV *> llvm::getStartAndEndForAccess(
       // separately checks that accesses cannot not wrap, so unsigned max
       // represents an upper bound.
       if (evaluatePtrAddRecAtMaxBTCWillNotWrap(AR, MaxBTC, EltSizeSCEV, *SE, DL,
-                                               DT, AC)) {
+                                               DT, AC, LoopGuards)) {
         ScEnd = AR->evaluateAtIteration(MaxBTC, *SE);
       } else {
         ScEnd = SE->getAddExpr(
@@ -381,7 +395,7 @@ void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, const SCEV *PtrExpr,
   const SCEV *BTC = PSE.getBackedgeTakenCount();
   const auto &[ScStart, ScEnd] = getStartAndEndForAccess(
       Lp, PtrExpr, AccessTy, BTC, SymbolicMaxBTC, PSE.getSE(),
-      &DC.getPointerBounds(), DC.getDT(), DC.getAC());
+      &DC.getPointerBounds(), DC.getDT(), DC.getAC(), LoopGuards);
   assert(!isa<SCEVCouldNotCompute>(ScStart) &&
          !isa<SCEVCouldNotCompute>(ScEnd) &&
          "must be able to compute both start and end expressions");
@@ -1986,13 +2000,13 @@ bool MemoryDepChecker::areAccessesCompletelyBeforeOrAfter(const SCEV *Src,
   ScalarEvolution &SE = *PSE.getSE();
   const auto &[SrcStart_, SrcEnd_] =
       getStartAndEndForAccess(InnermostLoop, Src, SrcTy, BTC, SymbolicMaxBTC,
-                              &SE, &PointerBounds, DT, AC);
+                              &SE, &PointerBounds, DT, AC, LoopGuards);
   if (isa<SCEVCouldNotCompute>(SrcStart_) || isa<SCEVCouldNotCompute>(SrcEnd_))
     return false;
 
   const auto &[SinkStart_, SinkEnd_] =
       getStartAndEndForAccess(InnermostLoop, Sink, SinkTy, BTC, SymbolicMaxBTC,
-                              &SE, &PointerBounds, DT, AC);
+                              &SE, &PointerBounds, DT, AC, LoopGuards);
   if (isa<SCEVCouldNotCompute>(SinkStart_) ||
       isa<SCEVCouldNotCompute>(SinkEnd_))
     return false;
@@ -3039,8 +3053,9 @@ LoopAccessInfo::LoopAccessInfo(Loop *L, ScalarEvolution *SE,
         TTI->getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector) * 2;
 
   DepChecker = std::make_unique<MemoryDepChecker>(
-      *PSE, AC, DT, L, SymbolicStrides, MaxTargetVectorWidthInBits);
-  PtrRtChecking = std::make_unique<RuntimePointerChecking>(*DepChecker, SE);
+      *PSE, AC, DT, L, SymbolicStrides, MaxTargetVectorWidthInBits, LoopGuards);
+  PtrRtChecking =
+      std::make_unique<RuntimePointerChecking>(*DepChecker, SE, LoopGuards);
   if (canAnalyzeLoop())
     CanVecMem = analyzeLoop(AA, LI, TLI, DT);
 }
diff --git a/llvm/lib/Analysis/ScalarEvolution.cpp b/llvm/lib/Analysis/ScalarEvolution.cpp
@@ -2338,23 +2338,15 @@ bool ScalarEvolution::willNotOverflow(Instruction::BinaryOps BinOp, bool Signed,
   // Can we use context to prove the fact we need?
   if (!CtxI)
     return false;
+  // TODO: Support mul.
+  if (BinOp == Instruction::Mul)
+    return false;
   auto *RHSC = dyn_cast<SCEVConstant>(RHS);
   // TODO: Lift this limitation.
   if (!RHSC)
     return false;
   APInt C = RHSC->getAPInt();
   unsigned NumBits = C.getBitWidth();
-  if (BinOp == Instruction::Mul) {
-    // Multiplying by 0 or 1 never overflows
-    if (C.isZero() || C.isOne())
-      return true;
-    if (Signed)
-      return false;
-    APInt Limit = APInt::getMaxValue(NumBits).udiv(C);
-    // To avoid overflow, we need to make sure that LHS <= MAX / C.
-    return isKnownPredicateAt(ICmpInst::ICMP_ULE, LHS, getConstant(Limit),
-                              CtxI);
-  }
   bool IsSub = (BinOp == Instruction::Sub);
   bool IsNegativeConst = (Signed && C.isNegative());
   // Compute the direction and magnitude by which we need to check overflow.
diff --git a/llvm/test/Transforms/LoopVectorize/vect.stats.ll b/llvm/test/Transforms/LoopVectorize/vect.stats.ll
@@ -1,4 +1,4 @@
-; RUN: opt < %s -passes=loop-vectorize -force-vector-interleave=4 -force-vector-width=4 -debug-only=loop-vectorize -enable-early-exit-vectorization -use-symbolic-maxbtc-deref-loop --disable-output -stats -S 2>&1 | FileCheck %s
+; RUN: opt < %s -passes=loop-vectorize -force-vector-interleave=4 -force-vector-width=4 -debug-only=loop-vectorize -enable-early-exit-vectorization --disable-output -stats -S 2>&1 | FileCheck %s
 ; REQUIRES: asserts
 
 ; We have 3 loops, two of them are vectorizable (with one being early-exit

Original file line number	Diff line number	Diff line change
`@@ -1,4 +1,4 @@`
`1`		`-; RUN: opt < %s -passes=loop-vectorize -force-vector-interleave=4 -force-vector-width=4 -debug-only=loop-vectorize -enable-early-exit-vectorization -use-symbolic-maxbtc-deref-loop --disable-output -stats -S 2>&1 \| FileCheck %s`
	`1`	`+; RUN: opt < %s -passes=loop-vectorize -force-vector-interleave=4 -force-vector-width=4 -debug-only=loop-vectorize -enable-early-exit-vectorization --disable-output -stats -S 2>&1 \| FileCheck %s`
`2`	`2`	`; REQUIRES: asserts`
`3`	`3`
`4`	`4`	`; We have 3 loops, two of them are vectorizable (with one being early-exit`