[LoopVectorize] Add support for reverse loops in isDereferenceableAndAlignedInLoop

Currently when we encounter a negative step in the induction
variable isDereferenceableAndAlignedInLoop bails out because
the element size is signed greater than the step. This patch
adds support for negative steps in cases where we detect the
start address for the load is of the form base + offset. In
this case the address decrements in each iteration so we need
to calculate the access size differently. I have done this by
caling getStartAndEndForAccess from LoopAccessAnalysis.cpp.

The changed test in LoopVectorize/X86/load-deref-pred.ll now
passes because previously we were calculating the total access
size incorrectly, whereas now it is 412 bytes and fits
perfectly into the alloca.

Author: david-arm

llvm/include/llvm/Analysis/LoopAccessAnalysis.h

@@ -853,6 +853,15 @@ bool sortPtrAccesses(ArrayRef<Value *> VL, Type *ElemTy, const DataLayout &DL,
 bool isConsecutiveAccess(Value *A, Value *B, const DataLayout &DL,
                          ScalarEvolution &SE, bool CheckType = true);
 
+/// For a given Loop \p Lp and pointer \p PtrExpr return a pair of SCEV values
+/// representing the maximum range of addresses accessed in the loop, i.e.
+/// [min,max).
+std::pair<const SCEV *, const SCEV *> getStartAndEndForAccess(
+    const Loop *Lp, const SCEV *PtrExpr, Type *AccessTy, const SCEV *MaxBECount,
+    ScalarEvolution *SE,
+    DenseMap<std::pair<const SCEV *, Type *>,
+             std::pair<const SCEV *, const SCEV *>> *PointerBounds);
+
 class LoopAccessInfoManager {
   /// The cache.
   DenseMap<Loop *, std::unique_ptr<LoopAccessInfo>> LoopAccessInfoMap;

llvm/lib/Analysis/Loads.cpp

@@ -13,6 +13,7 @@
 #include "llvm/Analysis/Loads.h"
 #include "llvm/Analysis/AliasAnalysis.h"
 #include "llvm/Analysis/AssumeBundleQueries.h"
+#include "llvm/Analysis/LoopAccessAnalysis.h"
 #include "llvm/Analysis/LoopInfo.h"
 #include "llvm/Analysis/MemoryBuiltins.h"
 #include "llvm/Analysis/MemoryLocation.h"
@@ -275,84 +276,85 @@ static bool AreEquivalentAddressValues(const Value *A, const Value *B) {
 bool llvm::isDereferenceableAndAlignedInLoop(
     LoadInst *LI, Loop *L, ScalarEvolution &SE, DominatorTree &DT,
     AssumptionCache *AC, SmallVectorImpl<const SCEVPredicate *> *Predicates) {
-  auto &DL = LI->getDataLayout();
-  Value *Ptr = LI->getPointerOperand();
-
-  APInt EltSize(DL.getIndexTypeSizeInBits(Ptr->getType()),
-                DL.getTypeStoreSize(LI->getType()).getFixedValue());
-  const Align Alignment = LI->getAlign();
+  const SCEV *Ptr = SE.getSCEV(LI->getPointerOperand());
+  auto *AddRec = dyn_cast<SCEVAddRecExpr>(Ptr);
 
-  Instruction *HeaderFirstNonPHI = L->getHeader()->getFirstNonPHI();
-
-  // If given a uniform (i.e. non-varying) address, see if we can prove the
-  // access is safe within the loop w/o needing predication.
-  if (L->isLoopInvariant(Ptr))
-    return isDereferenceableAndAlignedPointer(Ptr, Alignment, EltSize, DL,
-                                              HeaderFirstNonPHI, AC, &DT);
-
-  // Otherwise, check to see if we have a repeating access pattern where we can
-  // prove that all accesses are well aligned and dereferenceable.
-  auto *AddRec = dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Ptr));
+  // Check to see if we have a repeating access pattern and it's possible
+  // to prove all accesses are well aligned.
   if (!AddRec || AddRec->getLoop() != L || !AddRec->isAffine())
     return false;
+
   auto* Step = dyn_cast<SCEVConstant>(AddRec->getStepRecurrence(SE));
   if (!Step)
     return false;
 
-  auto TC = SE.getSmallConstantMaxTripCount(L, Predicates);
-  if (!TC)
+  // For the moment, restrict ourselves to the case where the access size is a
+  // multiple of the requested alignment and the base is aligned.
+  // TODO: generalize if a case found which warrants
+  const Align Alignment = LI->getAlign();
+  auto &DL = LI->getDataLayout();
+  APInt EltSize(DL.getIndexTypeSizeInBits(Ptr->getType()),
+                DL.getTypeStoreSize(LI->getType()).getFixedValue());
+  if (EltSize.urem(Alignment.value()) != 0)
     return false;
 
   // TODO: Handle overlapping accesses.
-  // We should be computing AccessSize as (TC - 1) * Step + EltSize.
-  if (EltSize.sgt(Step->getAPInt()))
+  if (EltSize.ugt(Step->getAPInt().abs()))
     return false;
 
-  // Compute the total access size for access patterns with unit stride and
-  // patterns with gaps. For patterns with unit stride, Step and EltSize are the
-  // same.
-  // For patterns with gaps (i.e. non unit stride), we are
-  // accessing EltSize bytes at every Step.
-  APInt AccessSize = TC * Step->getAPInt();
+  const SCEV *MaxBECount =
+      SE.getPredicatedSymbolicMaxBackedgeTakenCount(L, *Predicates);
+  if (isa<SCEVCouldNotCompute>(MaxBECount))
+    return false;
 
-  assert(SE.isLoopInvariant(AddRec->getStart(), L) &&
-         "implied by addrec definition");
-  Value *Base = nullptr;
-  if (auto *StartS = dyn_cast<SCEVUnknown>(AddRec->getStart())) {
-    Base = StartS->getValue();
-  } else if (auto *StartS = dyn_cast<SCEVAddExpr>(AddRec->getStart())) {
-    // Handle (NewBase + offset) as start value.
-    const auto *Offset = dyn_cast<SCEVConstant>(StartS->getOperand(0));
-    const auto *NewBase = dyn_cast<SCEVUnknown>(StartS->getOperand(1));
-    if (StartS->getNumOperands() == 2 && Offset && NewBase) {
-      // The following code below assumes the offset is unsigned, but GEP
-      // offsets are treated as signed so we can end up with a signed value
-      // here too. For example, suppose the initial PHI value is (i8 255),
-      // the offset will be treated as (i8 -1) and sign-extended to (i64 -1).
-      if (Offset->getAPInt().isNegative())
-        return false;
+  const auto &[AccessStart, AccessEnd] =
+      getStartAndEndForAccess(L, Ptr, LI->getType(), MaxBECount, &SE, nullptr);
+  if (isa<SCEVCouldNotCompute>(AccessStart) ||
+      isa<SCEVCouldNotCompute>(AccessEnd))
+    return false;
 
-      // For the moment, restrict ourselves to the case where the offset is a
-      // multiple of the requested alignment and the base is aligned.
-      // TODO: generalize if a case found which warrants
-      if (Offset->getAPInt().urem(Alignment.value()) != 0)
-        return false;
-      Base = NewBase->getValue();
-      bool Overflow = false;
-      AccessSize = AccessSize.uadd_ov(Offset->getAPInt(), Overflow);
-      if (Overflow)
-        return false;
-    }
-  }
+  // Try to get the access size.
+  const SCEV *PtrDiff = SE.getMinusSCEV(AccessEnd, AccessStart);
+  APInt MaxPtrDiff = SE.getUnsignedRangeMax(PtrDiff);
 
-  if (!Base)
+  // If the (max) pointer difference is > 32 bits then it's unlikely to be
+  // dereferenceable.
+  if (MaxPtrDiff.getActiveBits() > 32)
     return false;
 
-  // For the moment, restrict ourselves to the case where the access size is a
-  // multiple of the requested alignment and the base is aligned.
-  // TODO: generalize if a case found which warrants
-  if (EltSize.urem(Alignment.value()) != 0)
+  Value *Base = nullptr;
+  APInt AccessSize;
+  if (const SCEVUnknown *NewBase = dyn_cast<SCEVUnknown>(AccessStart)) {
+    Base = NewBase->getValue();
+    AccessSize = MaxPtrDiff;
+  } else if (auto *MinAdd = dyn_cast<SCEVAddExpr>(AccessStart)) {
+    if (MinAdd->getNumOperands() != 2)
+      return false;
+
+    const auto *Offset = dyn_cast<SCEVConstant>(MinAdd->getOperand(0));
+    const auto *NewBase = dyn_cast<SCEVUnknown>(MinAdd->getOperand(1));
+    if (!Offset || !NewBase)
+      return false;
+
+    // The following code below assumes the offset is unsigned, but GEP
+    // offsets are treated as signed so we can end up with a signed value
+    // here too. For example, suppose the initial PHI value is (i8 255),
+    // the offset will be treated as (i8 -1) and sign-extended to (i64 -1).
+    if (Offset->getAPInt().isNegative())
+      return false;
+
+    // For the moment, restrict ourselves to the case where the offset is a
+    // multiple of the requested alignment and the base is aligned.
+    // TODO: generalize if a case found which warrants
+    if (Offset->getAPInt().urem(Alignment.value()) != 0)
+      return false;
+
+    AccessSize = MaxPtrDiff + Offset->getAPInt();
+    Base = NewBase->getValue();
+  } else
     return false;
+
+  Instruction *HeaderFirstNonPHI = L->getHeader()->getFirstNonPHI();
   return isDereferenceableAndAlignedPointer(Base, Alignment, AccessSize, DL,
                                             HeaderFirstNonPHI, AC, &DT);
 }

llvm/lib/Analysis/LoopAccessAnalysis.cpp

@@ -203,29 +203,29 @@ RuntimeCheckingPtrGroup::RuntimeCheckingPtrGroup(
 ///
 /// There is no conflict when the intervals are disjoint:
 /// NoConflict = (P2.Start >= P1.End) || (P1.Start >= P2.End)
-static std::pair<const SCEV *, const SCEV *> getStartAndEndForAccess(
-    const Loop *Lp, const SCEV *PtrExpr, Type *AccessTy,
-    PredicatedScalarEvolution &PSE,
+std::pair<const SCEV *, const SCEV *> llvm::getStartAndEndForAccess(
+    const Loop *Lp, const SCEV *PtrExpr, Type *AccessTy, const SCEV *MaxBECount,
+    ScalarEvolution *SE,
     DenseMap<std::pair<const SCEV *, Type *>,
-             std::pair<const SCEV *, const SCEV *>> &PointerBounds) {
-  ScalarEvolution *SE = PSE.getSE();
-
-  auto [Iter, Ins] = PointerBounds.insert(
-      {{PtrExpr, AccessTy},
-       {SE->getCouldNotCompute(), SE->getCouldNotCompute()}});
-  if (!Ins)
-    return Iter->second;
+             std::pair<const SCEV *, const SCEV *>> *PointerBounds) {
+  std::pair<const SCEV *, const SCEV *> *PtrBoundsPair;
+  if (PointerBounds) {
+    auto [Iter, Ins] = PointerBounds->insert(
+        {{PtrExpr, AccessTy},
+         {SE->getCouldNotCompute(), SE->getCouldNotCompute()}});
+    if (!Ins)
+      return Iter->second;
+    PtrBoundsPair = &Iter->second;
+  }
 
   const SCEV *ScStart;
   const SCEV *ScEnd;
 
   if (SE->isLoopInvariant(PtrExpr, Lp)) {
     ScStart = ScEnd = PtrExpr;
   } else if (auto *AR = dyn_cast<SCEVAddRecExpr>(PtrExpr)) {
-    const SCEV *Ex = PSE.getSymbolicMaxBackedgeTakenCount();
-
     ScStart = AR->getStart();
-    ScEnd = AR->evaluateAtIteration(Ex, *SE);
+    ScEnd = AR->evaluateAtIteration(MaxBECount, *SE);
     const SCEV *Step = AR->getStepRecurrence(*SE);
 
     // For expressions with negative step, the upper bound is ScStart and the
@@ -244,16 +244,18 @@ static std::pair<const SCEV *, const SCEV *> getStartAndEndForAccess(
     return {SE->getCouldNotCompute(), SE->getCouldNotCompute()};
 
   assert(SE->isLoopInvariant(ScStart, Lp) && "ScStart needs to be invariant");
-  assert(SE->isLoopInvariant(ScEnd, Lp)&& "ScEnd needs to be invariant");
+  assert(SE->isLoopInvariant(ScEnd, Lp) && "ScEnd needs to be invariant");
 
   // Add the size of the pointed element to ScEnd.
   auto &DL = Lp->getHeader()->getDataLayout();
   Type *IdxTy = DL.getIndexType(PtrExpr->getType());
   const SCEV *EltSizeSCEV = SE->getStoreSizeOfExpr(IdxTy, AccessTy);
   ScEnd = SE->getAddExpr(ScEnd, EltSizeSCEV);
 
-  Iter->second = {ScStart, ScEnd};
-  return Iter->second;
+  std::pair<const SCEV *, const SCEV *> Res = {ScStart, ScEnd};
+  if (PointerBounds)
+    *PtrBoundsPair = Res;
+  return Res;
 }
 
 /// Calculate Start and End points of memory access using
@@ -263,8 +265,9 @@ void RuntimePointerChecking::insert(Loop *Lp, Value *Ptr, const SCEV *PtrExpr,
                                     unsigned DepSetId, unsigned ASId,
                                     PredicatedScalarEvolution &PSE,
                                     bool NeedsFreeze) {
+  const SCEV *MaxBECount = PSE.getSymbolicMaxBackedgeTakenCount();
   const auto &[ScStart, ScEnd] = getStartAndEndForAccess(
-      Lp, PtrExpr, AccessTy, PSE, DC.getPointerBounds());
+      Lp, PtrExpr, AccessTy, MaxBECount, PSE.getSE(), &DC.getPointerBounds());
   assert(!isa<SCEVCouldNotCompute>(ScStart) &&
          !isa<SCEVCouldNotCompute>(ScEnd) &&
          "must be able to compute both start and end expressions");
@@ -1938,10 +1941,11 @@ MemoryDepChecker::getDependenceDistanceStrideAndSize(
   // required for correctness.
   if (SE.isLoopInvariant(Src, InnermostLoop) ||
       SE.isLoopInvariant(Sink, InnermostLoop)) {
-    const auto &[SrcStart_, SrcEnd_] =
-        getStartAndEndForAccess(InnermostLoop, Src, ATy, PSE, PointerBounds);
-    const auto &[SinkStart_, SinkEnd_] =
-        getStartAndEndForAccess(InnermostLoop, Sink, BTy, PSE, PointerBounds);
+    const SCEV *MaxBECount = PSE.getSymbolicMaxBackedgeTakenCount();
+    const auto &[SrcStart_, SrcEnd_] = getStartAndEndForAccess(
+        InnermostLoop, Src, ATy, MaxBECount, PSE.getSE(), &PointerBounds);
+    const auto &[SinkStart_, SinkEnd_] = getStartAndEndForAccess(
+        InnermostLoop, Sink, BTy, MaxBECount, PSE.getSE(), &PointerBounds);
     if (!isa<SCEVCouldNotCompute>(SrcStart_) &&
         !isa<SCEVCouldNotCompute>(SrcEnd_) &&
         !isa<SCEVCouldNotCompute>(SinkStart_) &&