[llvm] r320298 - [SCEV] Fix wrong Equal predicate created in getAddRecForPhiWithCasts

[Dorit Nuzman via llvm-commits]

Author: dorit
Date: Sun Dec 10 03:13:35 2017
New Revision: 320298

URL: http://llvm.org/viewvc/llvm-project?rev=320298&view=rev
Log:
[SCEV] Fix wrong Equal predicate created in getAddRecForPhiWithCasts

CreateAddRecFromPHIWithCastsImpl() adds an IncrementNUSW overflow predicate
which allows the PSCEV rewriter to rewrite this scev expression:
 (zext i8 {0, + , (trunc i32 step to i8)} to i32)
into
 {0, +, (sext i8 (trunc i32 step to i8) to i32)}

But then it adds the wrong Equal predicate:
 %step == (zext i8 (trunc i32 %step to i8) to i32).
instead of:
 %step == (sext i8 (trunc i32 %step to i8) to i32)

This is fixed here.

Differential Revision: https://reviews.llvm.org/D40641


Modified:
    llvm/trunk/lib/Analysis/ScalarEvolution.cpp
    llvm/trunk/test/Transforms/LoopVectorize/pr30654-phiscev-sext-trunc.ll

Modified: llvm/trunk/lib/Analysis/ScalarEvolution.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/ScalarEvolution.cpp?rev=320298&r1=320297&r2=320298&view=diff
==============================================================================
--- llvm/trunk/lib/Analysis/ScalarEvolution.cpp (original)
+++ llvm/trunk/lib/Analysis/ScalarEvolution.cpp Sun Dec 10 03:13:35 2017
@@ -4636,18 +4636,19 @@ ScalarEvolution::createAddRecFromPHIWith
 
   // Construct the extended SCEV: (Ext ix (Trunc iy (Expr) to ix) to iy)
   // for each of StartVal and Accum
-  auto GetExtendedExpr = [&](const SCEV *Expr) -> const SCEV * {
+  auto getExtendedExpr = [&](const SCEV *Expr, 
+                             bool CreateSignExtend) -> const SCEV * {
     assert(isLoopInvariant(Expr, L) && "Expr is expected to be invariant");
     const SCEV *TruncatedExpr = getTruncateExpr(Expr, TruncTy);
     const SCEV *ExtendedExpr =
-        Signed ? getSignExtendExpr(TruncatedExpr, Expr->getType())
-               : getZeroExtendExpr(TruncatedExpr, Expr->getType());
+        CreateSignExtend ? getSignExtendExpr(TruncatedExpr, Expr->getType())
+                         : getZeroExtendExpr(TruncatedExpr, Expr->getType());
     return ExtendedExpr;
   };
 
   // Given:
   //  ExtendedExpr = (Ext ix (Trunc iy (Expr) to ix) to iy
-  //               = GetExtendedExpr(Expr)
+  //               = getExtendedExpr(Expr)
   // Determine whether the predicate P: Expr == ExtendedExpr
   // is known to be false at compile time
   auto PredIsKnownFalse = [&](const SCEV *Expr,
@@ -4656,13 +4657,15 @@ ScalarEvolution::createAddRecFromPHIWith
            isKnownPredicate(ICmpInst::ICMP_NE, Expr, ExtendedExpr);
   };
 
-  const SCEV *StartExtended = GetExtendedExpr(StartVal);
+  const SCEV *StartExtended = getExtendedExpr(StartVal, Signed);
   if (PredIsKnownFalse(StartVal, StartExtended)) {
     DEBUG(dbgs() << "P2 is compile-time false\n";);
     return None;
   }
 
-  const SCEV *AccumExtended = GetExtendedExpr(Accum);
+  // The Step is always Signed (because the overflow checks are either
+  // NSSW or NUSW)
+  const SCEV *AccumExtended = getExtendedExpr(Accum, /*CreateSignExtend=*/true);
   if (PredIsKnownFalse(Accum, AccumExtended)) {
     DEBUG(dbgs() << "P3 is compile-time false\n";);
     return None;

Modified: llvm/trunk/test/Transforms/LoopVectorize/pr30654-phiscev-sext-trunc.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/LoopVectorize/pr30654-phiscev-sext-trunc.ll?rev=320298&r1=320297&r2=320298&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/LoopVectorize/pr30654-phiscev-sext-trunc.ll (original)
+++ llvm/trunk/test/Transforms/LoopVectorize/pr30654-phiscev-sext-trunc.ll Sun Dec 10 03:13:35 2017
@@ -74,7 +74,7 @@ for.end:
 ; Same as above, but for checking the SCEV "zext(trunc(%p.09)) + %step".
 ; Here we expect the following two predicates to be added for runtime checking:
 ; 1) {0,+,(trunc i32 %step to i8)}<%for.body> Added Flags: <nusw>
-; 2) Equal predicate: %step == (zext i8 (trunc i32 %step to i8) to i32)
+; 2) Equal predicate: %step == (sext i8 (trunc i32 %step to i8) to i32)
 ;
 ; int a[N];
 ; void doit2(int n, int step) {
@@ -93,7 +93,8 @@ for.end:
 ; CHECK-NOT: %mul = call { i8, i1 } @llvm.umul.with.overflow.i8(i8 {{.*}}, i8 {{.*}})
 ; CHECK: %[[TEST:[0-9]+]] = or i1 {{.*}}, %mul.overflow
 ; CHECK: %[[NTEST:[0-9]+]] = or i1 false, %[[TEST]]
-; CHECK: %ident.check = icmp ne i32 {{.*}}, %{{.*}}
+; CHECK: %[[EXT:[0-9]+]] = sext i8 {{.*}} to i32
+; CHECK: %ident.check = icmp ne i32 {{.*}}, %[[EXT]]
 ; CHECK: %{{.*}} = or i1 %[[NTEST]], %ident.check
 ; CHECK-NOT: %mul = call { i8, i1 } @llvm.umul.with.overflow.i8(i8 {{.*}}, i8 {{.*}})
 ; CHECK: vector.body: