[SCEVPatternMatch] Extend m_scev_AffineAddRec with Loop

Author: artagnon

llvm/include/llvm/Analysis/ScalarEvolutionPatternMatch.h

@@ -14,6 +14,7 @@
 #define LLVM_ANALYSIS_SCALAREVOLUTIONPATTERNMATCH_H
 
 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include <type_traits>
 
 namespace llvm {
 namespace SCEVPatternMatch {
@@ -162,13 +163,18 @@ template <typename SCEVTy, typename Op0_t, typename Op1_t>
 struct SCEVBinaryExpr_match {
   Op0_t Op0;
   Op1_t Op1;
+  const Loop *L;
 
-  SCEVBinaryExpr_match(Op0_t Op0, Op1_t Op1) : Op0(Op0), Op1(Op1) {}
+  SCEVBinaryExpr_match(Op0_t Op0, Op1_t Op1, const Loop *L = nullptr)
+      : Op0(Op0), Op1(Op1), L(L) {}
 
   bool match(const SCEV *S) const {
     auto *E = dyn_cast<SCEVTy>(S);
+    bool LoopMatches = true;
+    if constexpr (std::is_same_v<SCEVTy, SCEVAddRecExpr>)
+      LoopMatches = !L || (E && E->getLoop() == L);
     return E && E->getNumOperands() == 2 && Op0.match(E->getOperand(0)) &&
-           Op1.match(E->getOperand(1));
+           Op1.match(E->getOperand(1)) && LoopMatches;
   }
 };
 
@@ -198,7 +204,8 @@ m_scev_UDiv(const Op0_t &Op0, const Op1_t &Op1) {
 
 template <typename Op0_t, typename Op1_t>
 inline SCEVBinaryExpr_match<SCEVAddRecExpr, Op0_t, Op1_t>
-m_scev_AffineAddRec(const Op0_t &Op0, const Op1_t &Op1) {
+m_scev_AffineAddRec(const Op0_t &Op0, const Op1_t &Op1,
+                    const Loop *L = nullptr) {
   return m_scev_Binary<SCEVAddRecExpr>(Op0, Op1);
 }
 } // namespace SCEVPatternMatch

llvm/lib/Transforms/Scalar/IndVarSimplify.cpp

@@ -808,8 +808,7 @@ static bool isLoopCounter(PHINode* Phi, Loop *L,
     return false;
 
   const SCEV *S = SE->getSCEV(Phi);
-  if (!match(S, m_scev_AffineAddRec(m_SCEV(), m_scev_One())) ||
-      cast<SCEVAddRecExpr>(S)->getLoop() != L)
+  if (!match(S, m_scev_AffineAddRec(m_SCEV(), m_scev_One(), L)))
     return false;
 
   int LatchIdx = Phi->getBasicBlockIndex(L->getLoopLatch());

llvm/lib/Transforms/Scalar/LoopIdiomRecognize.cpp

@@ -455,8 +455,8 @@ LoopIdiomRecognize::isLegalStore(StoreInst *SI) {
   // random store we can't handle.
   const SCEV *StoreEv = SE->getSCEV(StorePtr);
   const SCEVConstant *Stride;
-  if (!match(StoreEv, m_scev_AffineAddRec(m_SCEV(), m_SCEVConstant(Stride))) ||
-      cast<SCEVAddRecExpr>(StoreEv)->getLoop() != CurLoop)
+  if (!match(StoreEv,
+             m_scev_AffineAddRec(m_SCEV(), m_SCEVConstant(Stride), CurLoop)))
     return LegalStoreKind::None;
 
   // See if the store can be turned into a memset.
@@ -513,8 +513,7 @@ LoopIdiomRecognize::isLegalStore(StoreInst *SI) {
 
     // The store and load must share the same stride.
     if (!match(LoadEv,
-               m_scev_AffineAddRec(m_SCEV(), m_scev_Specific(Stride))) ||
-        cast<SCEVAddRecExpr>(LoadEv)->getLoop() != CurLoop)
+               m_scev_AffineAddRec(m_SCEV(), m_scev_Specific(Stride), CurLoop)))
       return LegalStoreKind::None;
 
     // Success.  This store can be converted into a memcpy.
@@ -787,11 +786,13 @@ bool LoopIdiomRecognize::processLoopMemCpy(MemCpyInst *MCI,
   // See if the load and store pointer expressions are AddRec like {base,+,1} on
   // the current loop, which indicates a strided load and store.  If we have
   // something else, it's a random load or store we can't handle.
-  const SCEVAddRecExpr *StoreEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Dest));
-  if (!StoreEv || StoreEv->getLoop() != CurLoop || !StoreEv->isAffine())
-    return false;
-  const SCEVAddRecExpr *LoadEv = dyn_cast<SCEVAddRecExpr>(SE->getSCEV(Source));
-  if (!LoadEv || LoadEv->getLoop() != CurLoop || !LoadEv->isAffine())
+  const SCEV *StoreEv = SE->getSCEV(Dest);
+  const SCEV *LoadEv = SE->getSCEV(Source);
+  const APInt *StoreStrideValue, *LoadStrideValue;
+  if (!match(StoreEv, m_scev_AffineAddRec(
+                          m_SCEV(), m_scev_APInt(StoreStrideValue), CurLoop)) ||
+      !match(LoadEv, m_scev_AffineAddRec(
+                         m_SCEV(), m_scev_APInt(LoadStrideValue), CurLoop)))
     return false;
 
   // Reject memcpys that are so large that they overflow an unsigned.
@@ -801,10 +802,6 @@ bool LoopIdiomRecognize::processLoopMemCpy(MemCpyInst *MCI,
 
   // Check if the stride matches the size of the memcpy. If so, then we know
   // that every byte is touched in the loop.
-  const APInt *StoreStrideValue, *LoadStrideValue;
-  if (!match(StoreEv->getOperand(1), m_scev_APInt(StoreStrideValue)) ||
-      !match(LoadEv->getOperand(1), m_scev_APInt(LoadStrideValue)))
-    return false;
 
   // Huge stride value - give up
   if (StoreStrideValue->getBitWidth() > 64 ||
@@ -830,8 +827,8 @@ bool LoopIdiomRecognize::processLoopMemCpy(MemCpyInst *MCI,
 
   return processLoopStoreOfLoopLoad(
       Dest, Source, SE->getConstant(Dest->getType(), SizeInBytes),
-      MCI->getDestAlign(), MCI->getSourceAlign(), MCI, MCI, StoreEv, LoadEv,
-      BECount);
+      MCI->getDestAlign(), MCI->getSourceAlign(), MCI, MCI,
+      cast<SCEVAddRecExpr>(StoreEv), cast<SCEVAddRecExpr>(LoadEv), BECount);
 }
 
 /// processLoopMemSet - See if this memset can be promoted to a large memset.
@@ -852,12 +849,11 @@ bool LoopIdiomRecognize::processLoopMemSet(MemSetInst *MSI,
   // random store we can't handle.
   const SCEV *Ev = SE->getSCEV(Pointer);
   const SCEV *PointerStrideSCEV;
-  if (!match(Ev, m_scev_AffineAddRec(m_SCEV(), m_SCEV(PointerStrideSCEV)))) {
+  if (!match(Ev, m_scev_AffineAddRec(m_SCEV(), m_SCEV(PointerStrideSCEV),
+                                     CurLoop))) {
     LLVM_DEBUG(dbgs() << "  Pointer is not affine, abort\n");
     return false;
   }
-  if (cast<SCEVAddRecExpr>(Ev)->getLoop() != CurLoop)
-    return false;
 
   const SCEV *MemsetSizeSCEV = SE->getSCEV(MSI->getLength());
   if (!PointerStrideSCEV || !MemsetSizeSCEV)