[LV]Split store-load forward distance analysis from other checks, NFC (#121156)

The patch splits the store-load forwarding distance analysis from other
dependency analysis in LAA. Currently it supports only power-of-2
distances, required to support non-power-of-2 distances in future.

Part of #100755

Author: alexey-bataev

llvm/include/llvm/Analysis/LoopAccessAnalysis.h

@@ -216,6 +216,21 @@ class MemoryDepChecker {
     return MaxSafeVectorWidthInBits;
   }
 
+  /// Return true if there are no store-load forwarding dependencies.
+  bool isSafeForAnyStoreLoadForwardDistances() const {
+    return MaxStoreLoadForwardSafeDistanceInBits ==
+           std::numeric_limits<uint64_t>::max();
+  }
+
+  /// Return safe power-of-2 number of elements, which do not prevent store-load
+  /// forwarding, multiplied by the size of the elements in bits.
+  uint64_t getStoreLoadForwardSafeDistanceInBits() const {
+    assert(!isSafeForAnyStoreLoadForwardDistances() &&
+           "Expected the distance, that prevent store-load forwarding, to be "
+           "set.");
+    return MaxStoreLoadForwardSafeDistanceInBits;
+  }
+
   /// In same cases when the dependency check fails we can still
   /// vectorize the loop with a dynamic array access check.
   bool shouldRetryWithRuntimeCheck() const {
@@ -304,6 +319,11 @@ class MemoryDepChecker {
   /// restrictive.
   uint64_t MaxSafeVectorWidthInBits = -1U;
 
+  /// Maximum power-of-2 number of elements, which do not prevent store-load
+  /// forwarding, multiplied by the size of the elements in bits.
+  uint64_t MaxStoreLoadForwardSafeDistanceInBits =
+      std::numeric_limits<uint64_t>::max();
+
   /// If we see a non-constant dependence distance we can still try to
   /// vectorize this loop with runtime checks.
   bool FoundNonConstantDistanceDependence = false;
@@ -357,7 +377,8 @@ class MemoryDepChecker {
   ///
   /// \return false if we shouldn't vectorize at all or avoid larger
   /// vectorization factors by limiting MinDepDistBytes.
-  bool couldPreventStoreLoadForward(uint64_t Distance, uint64_t TypeByteSize);
+  bool couldPreventStoreLoadForward(uint64_t Distance, uint64_t TypeByteSize,
+                                    unsigned CommonStride = 0);
 
   /// Updates the current safety status with \p S. We can go from Safe to
   /// either PossiblySafeWithRtChecks or Unsafe and from

llvm/include/llvm/Transforms/Vectorize/LoopVectorizationLegality.h

@@ -382,7 +382,8 @@ class LoopVectorizationLegality {
   const LoopAccessInfo *getLAI() const { return LAI; }
 
   bool isSafeForAnyVectorWidth() const {
-    return LAI->getDepChecker().isSafeForAnyVectorWidth();
+    return LAI->getDepChecker().isSafeForAnyVectorWidth() &&
+           LAI->getDepChecker().isSafeForAnyStoreLoadForwardDistances();
   }
 
   uint64_t getMaxSafeVectorWidthInBits() const {
@@ -406,6 +407,17 @@ class LoopVectorizationLegality {
     return hasUncountableEarlyExit() ? getUncountableEdge()->second : nullptr;
   }
 
+  /// Return true if there is store-load forwarding dependencies.
+  bool isSafeForAnyStoreLoadForwardDistances() const {
+    return LAI->getDepChecker().isSafeForAnyStoreLoadForwardDistances();
+  }
+
+  /// Return safe power-of-2 number of elements, which do not prevent store-load
+  /// forwarding and safe to operate simultaneously.
+  uint64_t getMaxStoreLoadForwardSafeDistanceInBits() const {
+    return LAI->getDepChecker().getStoreLoadForwardSafeDistanceInBits();
+  }
+
   /// Returns true if vector representation of the instruction \p I
   /// requires mask.
   bool isMaskRequired(const Instruction *I) const {

llvm/lib/Analysis/LoopAccessAnalysis.cpp

@@ -1740,7 +1740,8 @@ bool MemoryDepChecker::Dependence::isForward() const {
 }
 
 bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance,
-                                                    uint64_t TypeByteSize) {
+                                                    uint64_t TypeByteSize,
+                                                    unsigned CommonStride) {
   // If loads occur at a distance that is not a multiple of a feasible vector
   // factor store-load forwarding does not take place.
   // Positive dependences might cause troubles because vectorizing them might
@@ -1755,31 +1756,38 @@ bool MemoryDepChecker::couldPreventStoreLoadForward(uint64_t Distance,
   // cause any slowdowns.
   const uint64_t NumItersForStoreLoadThroughMemory = 8 * TypeByteSize;
   // Maximum vector factor.
-  uint64_t MaxVFWithoutSLForwardIssues = std::min(
-      VectorizerParams::MaxVectorWidth * TypeByteSize, MinDepDistBytes);
+  uint64_t MaxVFWithoutSLForwardIssuesPowerOf2 =
+      std::min(VectorizerParams::MaxVectorWidth * TypeByteSize,
+               MaxStoreLoadForwardSafeDistanceInBits);
 
   // Compute the smallest VF at which the store and load would be misaligned.
-  for (uint64_t VF = 2 * TypeByteSize; VF <= MaxVFWithoutSLForwardIssues;
-       VF *= 2) {
+  for (uint64_t VF = 2 * TypeByteSize;
+       VF <= MaxVFWithoutSLForwardIssuesPowerOf2; VF *= 2) {
     // If the number of vector iteration between the store and the load are
     // small we could incur conflicts.
     if (Distance % VF && Distance / VF < NumItersForStoreLoadThroughMemory) {
-      MaxVFWithoutSLForwardIssues = (VF >> 1);
+      MaxVFWithoutSLForwardIssuesPowerOf2 = (VF >> 1);
       break;
     }
   }
 
-  if (MaxVFWithoutSLForwardIssues < 2 * TypeByteSize) {
+  if (MaxVFWithoutSLForwardIssuesPowerOf2 < 2 * TypeByteSize) {
     LLVM_DEBUG(
         dbgs() << "LAA: Distance " << Distance
                << " that could cause a store-load forwarding conflict\n");
     return true;
   }
 
-  if (MaxVFWithoutSLForwardIssues < MinDepDistBytes &&
-      MaxVFWithoutSLForwardIssues !=
-          VectorizerParams::MaxVectorWidth * TypeByteSize)
-    MinDepDistBytes = MaxVFWithoutSLForwardIssues;
+  if (CommonStride &&
+      MaxVFWithoutSLForwardIssuesPowerOf2 <
+          MaxStoreLoadForwardSafeDistanceInBits &&
+      MaxVFWithoutSLForwardIssuesPowerOf2 !=
+          VectorizerParams::MaxVectorWidth * TypeByteSize) {
+    uint64_t MaxVF = MaxVFWithoutSLForwardIssuesPowerOf2 / CommonStride;
+    uint64_t MaxVFInBits = MaxVF * TypeByteSize * 8;
+    MaxStoreLoadForwardSafeDistanceInBits =
+        std::min(MaxStoreLoadForwardSafeDistanceInBits, MaxVFInBits);
+  }
   return false;
 }
 
@@ -2227,20 +2235,10 @@ MemoryDepChecker::isDependent(const MemAccessInfo &A, unsigned AIdx,
       std::min(static_cast<uint64_t>(MinDistance), MinDepDistBytes);
 
   bool IsTrueDataDependence = (!AIsWrite && BIsWrite);
-  uint64_t MinDepDistBytesOld = MinDepDistBytes;
   if (IsTrueDataDependence && EnableForwardingConflictDetection && ConstDist &&
-      couldPreventStoreLoadForward(MinDistance, TypeByteSize)) {
-    // Sanity check that we didn't update MinDepDistBytes when calling
-    // couldPreventStoreLoadForward
-    assert(MinDepDistBytes == MinDepDistBytesOld &&
-           "An update to MinDepDistBytes requires an update to "
-           "MaxSafeVectorWidthInBits");
-    (void)MinDepDistBytesOld;
+      couldPreventStoreLoadForward(MinDistance, TypeByteSize, *CommonStride))
     return Dependence::BackwardVectorizableButPreventsForwarding;
-  }
 
-  // An update to MinDepDistBytes requires an update to MaxSafeVectorWidthInBits
-  // since there is a backwards dependency.
   uint64_t MaxVF = MinDepDistBytes / *CommonStride;
   LLVM_DEBUG(dbgs() << "LAA: Positive min distance " << MinDistance
                     << " with max VF = " << MaxVF << '\n');
@@ -3005,6 +3003,11 @@ void LoopAccessInfo::print(raw_ostream &OS, unsigned Depth) const {
     if (!DC.isSafeForAnyVectorWidth())
       OS << " with a maximum safe vector width of "
          << DC.getMaxSafeVectorWidthInBits() << " bits";
+    if (!DC.isSafeForAnyStoreLoadForwardDistances()) {
+      uint64_t SLDist = DC.getStoreLoadForwardSafeDistanceInBits();
+      OS << ", with a maximum safe store-load forward width of " << SLDist
+         << " bits";
+    }
     if (PtrRtChecking->Need)
       OS << " with run-time checks";
     OS << "\n";

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -3815,13 +3815,18 @@ FixedScalableVFPair LoopVectorizationCostModel::computeFeasibleMaxVF(
   // It is computed by MaxVF * sizeOf(type) * 8, where type is taken from
   // the memory accesses that is most restrictive (involved in the smallest
   // dependence distance).
-  unsigned MaxSafeElements =
-      llvm::bit_floor(Legal->getMaxSafeVectorWidthInBits() / WidestType);
+  unsigned MaxSafeElementsPowerOf2 =
+      bit_floor(Legal->getMaxSafeVectorWidthInBits() / WidestType);
+  if (!Legal->isSafeForAnyStoreLoadForwardDistances()) {
+    unsigned SLDist = Legal->getMaxStoreLoadForwardSafeDistanceInBits();
+    MaxSafeElementsPowerOf2 =
+        std::min(MaxSafeElementsPowerOf2, SLDist / WidestType);
+  }
+  auto MaxSafeFixedVF = ElementCount::getFixed(MaxSafeElementsPowerOf2);
+  auto MaxSafeScalableVF = getMaxLegalScalableVF(MaxSafeElementsPowerOf2);
 
-  auto MaxSafeFixedVF = ElementCount::getFixed(MaxSafeElements);
-  auto MaxSafeScalableVF = getMaxLegalScalableVF(MaxSafeElements);
   if (!Legal->isSafeForAnyVectorWidth())
-    this->MaxSafeElements = MaxSafeElements;
+    this->MaxSafeElements = MaxSafeElementsPowerOf2;
 
   LLVM_DEBUG(dbgs() << "LV: The max safe fixed VF is: " << MaxSafeFixedVF
                     << ".\n");

llvm/test/Analysis/LoopAccessAnalysis/safe-with-dep-distance.ll

@@ -4,7 +4,7 @@
 ;   for (i = 0; i < n; i++)
 ;    A[i + 4] = A[i] * 2;
 
-; CHECK: Memory dependences are safe with a maximum safe vector width of 64 bits
+; CHECK: Memory dependences are safe with a maximum safe vector width of 64 bits, with a maximum safe store-load forward width of 64 bits
 
 target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-apple-macosx10.10.0"

llvm/test/Analysis/LoopAccessAnalysis/stride-access-dependence.ll

@@ -276,7 +276,7 @@ for.body:                                         ; preds = %entry, %for.body
 define void @vectorizable_Read_Write(ptr nocapture %A) {
 ; CHECK-LABEL: 'vectorizable_Read_Write'
 ; CHECK-NEXT:    for.body:
-; CHECK-NEXT:      Memory dependences are safe with a maximum safe vector width of 64 bits
+; CHECK-NEXT:      Memory dependences are safe with a maximum safe vector width of 64 bits, with a maximum safe store-load forward width of 64 bits
 ; CHECK-NEXT:      Dependences:
 ; CHECK-NEXT:        BackwardVectorizable:
 ; CHECK-NEXT:            %0 = load i32, ptr %arrayidx, align 4 ->