[LAA] Use SCEVUse to add extra NUW flags to pointer bounds.

Use SCEVUse to add a NUW flag to the upper bound of an accessed pointer.
We must already have proved that the pointers do not wrap, as otherwise
we could not use them for runtime check computations.

By adding the use-specific NUW flag, we can detect cases where SCEV can
prove that the compared pointers must overlap, so the runtime checks
will always be false. In that case, there is no point in vectorizing
with runtime checks.

Note that this depends c2895cd27fbf200d1da056bc66d77eeb62690bf0, which
could be submitted separately if desired; without the current change, I
don't think it triggers in practice though.

Author: fhahn

llvm/include/llvm/Analysis/LoopAccessAnalysis.h

@@ -406,10 +406,10 @@ struct RuntimeCheckingPtrGroup {
 
   /// The SCEV expression which represents the upper bound of all the
   /// pointers in this group.
-  const SCEV *High;
+  SCEVUse High;
   /// The SCEV expression which represents the lower bound of all the
   /// pointers in this group.
-  const SCEV *Low;
+  SCEVUse Low;
   /// Indices of all the pointers that constitute this grouping.
   SmallVector<unsigned, 2> Members;
   /// Address space of the involved pointers.
@@ -447,10 +447,10 @@ class RuntimePointerChecking {
     TrackingVH<Value> PointerValue;
     /// Holds the smallest byte address accessed by the pointer throughout all
     /// iterations of the loop.
-    const SCEV *Start;
+    SCEVUse Start;
     /// Holds the largest byte address accessed by the pointer throughout all
     /// iterations of the loop, plus 1.
-    const SCEV *End;
+    SCEVUse End;
     /// Holds the information if this pointer is used for writing to memory.
     bool IsWritePtr;
     /// Holds the id of the set of pointers that could be dependent because of a
@@ -463,7 +463,7 @@ class RuntimePointerChecking {
     /// True if the pointer expressions needs to be frozen after expansion.
     bool NeedsFreeze;
 
-    PointerInfo(Value *PointerValue, const SCEV *Start, const SCEV *End,
+    PointerInfo(Value *PointerValue, SCEVUse Start, SCEVUse End,
                 bool IsWritePtr, unsigned DependencySetId, unsigned AliasSetId,
                 const SCEV *Expr, bool NeedsFreeze)
         : PointerValue(PointerValue), Start(Start), End(End),
@@ -477,8 +477,10 @@ class RuntimePointerChecking {
   /// Reset the state of the pointer runtime information.
   void reset() {
     Need = false;
+    AlwaysFalse = false;
     Pointers.clear();
     Checks.clear();
+    CheckingGroups.clear();
   }
 
   /// Insert a pointer and calculate the start and end SCEVs.
@@ -535,6 +537,8 @@ class RuntimePointerChecking {
   /// This flag indicates if we need to add the runtime check.
   bool Need = false;
 
+  bool AlwaysFalse = false;
+
   /// Information about the pointers that may require checking.
   SmallVector<PointerInfo, 2> Pointers;
 

llvm/lib/Analysis/LoopAccessAnalysis.cpp

@@ -203,13 +203,13 @@ 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 *>
+static std::pair<SCEVUse, SCEVUse>
 getStartAndEndForAccess(const Loop *Lp, const SCEV *PtrExpr, Type *AccessTy,
                         PredicatedScalarEvolution &PSE) {
   ScalarEvolution *SE = PSE.getSE();
 
-  const SCEV *ScStart;
-  const SCEV *ScEnd;
+  SCEVUse ScStart;
+  SCEVUse ScEnd;
 
   if (SE->isLoopInvariant(PtrExpr, Lp)) {
     ScStart = ScEnd = PtrExpr;
@@ -219,6 +219,8 @@ getStartAndEndForAccess(const Loop *Lp, const SCEV *PtrExpr, Type *AccessTy,
     ScStart = AR->getStart();
     ScEnd = AR->evaluateAtIteration(Ex, *SE);
     const SCEV *Step = AR->getStepRecurrence(*SE);
+    if (auto *Comm = dyn_cast<SCEVCommutativeExpr>(ScEnd))
+      ScEnd = SCEVUse(ScEnd, 2);
 
     // For expressions with negative step, the upper bound is ScStart and the
     // lower bound is ScEnd.
@@ -242,7 +244,10 @@ getStartAndEndForAccess(const Loop *Lp, const SCEV *PtrExpr, Type *AccessTy,
   auto &DL = Lp->getHeader()->getModule()->getDataLayout();
   Type *IdxTy = DL.getIndexType(PtrExpr->getType());
   const SCEV *EltSizeSCEV = SE->getStoreSizeOfExpr(IdxTy, AccessTy);
-  ScEnd = SE->getAddExpr(ScEnd, EltSizeSCEV);
+  // TODO: this computes one-past-the-end. ScEnd + EltSizeSCEV - 1 is the last
+  // accessed byte. Not entirely sure if one-past-the-end must also not wrap? If
+  // it does, could compute and use last accessed byte instead.
+  ScEnd = SCEVUse(SE->getAddExpr(ScEnd, EltSizeSCEV), 2);
 
   return {ScStart, ScEnd};
 }
@@ -377,6 +382,11 @@ SmallVector<RuntimePointerCheck, 4> RuntimePointerChecking::generateChecks() {
       if (needsChecking(CGI, CGJ)) {
         CanUseDiffCheck = CanUseDiffCheck && tryToCreateDiffCheck(CGI, CGJ);
         Checks.push_back(std::make_pair(&CGI, &CGJ));
+        if (SE->isKnownPredicate(CmpInst::ICMP_UGT, CGI.High, CGJ.Low) &&
+            SE->isKnownPredicate(CmpInst::ICMP_ULE, CGI.Low, CGJ.High)) {
+          AlwaysFalse = true;
+          return {};
+        }
       }
     }
   }
@@ -631,8 +641,7 @@ void RuntimePointerChecking::print(raw_ostream &OS, unsigned Depth) const {
     const auto &CG = CheckingGroups[I];
 
     OS.indent(Depth + 2) << "Group " << &CG << ":\n";
-    OS.indent(Depth + 4) << "(Low: " << *CG.Low << " High: " << *CG.High
-                         << ")\n";
+    OS.indent(Depth + 4) << "(Low: " << CG.Low << " High: " << CG.High << ")\n";
     for (unsigned J = 0; J < CG.Members.size(); ++J) {
       OS.indent(Depth + 6) << "Member: " << *Pointers[CG.Members[J]].Expr
                            << "\n";
@@ -1268,6 +1277,7 @@ bool AccessAnalysis::canCheckPtrAtRT(RuntimePointerChecking &RtCheck,
   // If we can do run-time checks, but there are no checks, no runtime checks
   // are needed. This can happen when all pointers point to the same underlying
   // object for example.
+  CanDoRT &= !RtCheck.AlwaysFalse;
   RtCheck.Need = CanDoRT ? RtCheck.getNumberOfChecks() != 0 : MayNeedRTCheck;
 
   bool CanDoRTIfNeeded = !RtCheck.Need || CanDoRT;

llvm/lib/Analysis/ScalarEvolution.cpp

@@ -11950,6 +11950,8 @@ bool ScalarEvolution::splitBinaryAdd(SCEVUse Expr, SCEVUse &L, SCEVUse &R,
   L = AE->getOperand(0);
   R = AE->getOperand(1);
   Flags = AE->getNoWrapFlags();
+  Flags = setFlags(AE->getNoWrapFlags(),
+                   static_cast<SCEV::NoWrapFlags>(Expr.getInt()));
   return true;
 }
 

llvm/test/Analysis/LoopAccessAnalysis/different-strides-safe-dep-due-to-backedge-taken-count.ll

@@ -41,10 +41,6 @@ define void @forward_dep_not_known_safe_due_to_backedge_taken_count(ptr %A) {
 ; CHECK-NEXT:    loop:
 ; CHECK-NEXT:      Memory dependences are safe
 ; CHECK-NEXT:      Dependences:
-; CHECK-NEXT:        Forward:
-; CHECK-NEXT:            %l = load i32, ptr %gep.mul.2, align 4 ->
-; CHECK-NEXT:            store i32 %add, ptr %gep, align 4
-; CHECK-EMPTY:
 ; CHECK-NEXT:      Run-time memory checks:
 ; CHECK-NEXT:      Grouped accesses:
 ; CHECK-EMPTY:
@@ -109,13 +105,8 @@ exit:
 define void @unknown_dep_not_known_safe_due_to_backedge_taken_count(ptr %A) {
 ; CHECK-LABEL: 'unknown_dep_not_known_safe_due_to_backedge_taken_count'
 ; CHECK-NEXT:    loop:
-; CHECK-NEXT:      Report: unsafe dependent memory operations in loop. Use #pragma clang loop distribute(enable) to allow loop distribution to attempt to isolate the offending operations into a separate loop
-; CHECK-NEXT:  Unknown data dependence.
+; CHECK-NEXT:      Memory dependences are safe
 ; CHECK-NEXT:      Dependences:
-; CHECK-NEXT:        Unknown:
-; CHECK-NEXT:            %l = load i32, ptr %gep, align 4 ->
-; CHECK-NEXT:            store i32 %add, ptr %gep.mul.2, align 4
-; CHECK-EMPTY:
 ; CHECK-NEXT:      Run-time memory checks:
 ; CHECK-NEXT:      Grouped accesses:
 ; CHECK-EMPTY: