LoopVectorizer: A reduction that has multiple uses of the reduction value is not

a reduction.

Really. Under certain circumstances (the use list of an instruction has to be
set up right - hence the extra pass in the test case) we would not recognize
when a value in a potential reduction cycle was used multiple times by the
reduction cycle.

Fixes PR18526.
radar://15851149

llvm-svn: 199570

2 files changed, 53 insertions, 2 deletions

diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index c05288bd070..695ee03ea76 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -4535,13 +4535,22 @@ bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
         continue;
       }
 
-      // Process instructions only once (termination).
+      // Process instructions only once (termination). Each reduction cycle
+      // value must only be used once, except by phi nodes and min/max
+      // reductions which are represented as a cmp followed by a select.
+      ReductionInstDesc IgnoredVal(false, 0);
       if (VisitedInsts.insert(Usr)) {
         if (isa<PHINode>(Usr))
           PHIs.push_back(Usr);
         else
           NonPHIs.push_back(Usr);
-      }
+      } else if (!isa<PHINode>(Usr) &&
+                 ((!isa<FCmpInst>(Usr) &&
+                   !isa<ICmpInst>(Usr) &&
+                   !isa<SelectInst>(Usr)) ||
+                  !isMinMaxSelectCmpPattern(Usr, IgnoredVal).IsReduction))
+        return false;
+
       // Remember that we completed the cycle.
       if (Usr == Phi)
         FoundStartPHI = true;
diff --git a/llvm/test/Transforms/LoopVectorize/multi-use-reduction-bug.ll b/llvm/test/Transforms/LoopVectorize/multi-use-reduction-bug.ll
new file mode 100644
index 00000000000..88a29c50df5
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/multi-use-reduction-bug.ll
@@ -0,0 +1,42 @@
+; RUN: opt -indvars -loop-vectorize -force-vector-width=2 -force-vector-unroll=1 -S < %s | FileCheck %s
+
+target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.9.0"
+
+; We must not vectorize this loop. %add55 is not reduction. Its value is used
+; multiple times.
+
+; PR18526
+
+; CHECK: multiple_use_of_value
+; CHECK-NOT: <2 x i32>
+
+define void @multiple_use_of_value() {
+entry:
+  %n = alloca i32, align 4
+  %k7 = alloca i32, align 4
+  %nf = alloca i32, align 4
+  %0 = load i32* %k7, align 4
+  %.neg1 = sub i32 0, %0
+  %n.promoted = load i32* %n, align 4
+  %nf.promoted = load i32* %nf, align 4
+  br label %for.body
+
+for.body:
+  %inc107 = phi i32 [ undef, %entry ], [ %inc10, %for.body ]
+  %inc6 = phi i32 [ %nf.promoted, %entry ], [ undef, %for.body ]
+  %add55 = phi i32 [ %n.promoted, %entry ], [ %add5, %for.body ]
+  %.neg2 = sub i32 0, %inc6
+  %add.neg = add i32 0, %add55
+  %add4.neg = add i32 %add.neg, %.neg1
+  %sub = add i32 %add4.neg, %.neg2
+  %add5 = add i32 %sub, %add55
+  %inc10 = add i32 %inc107, 1
+  %cmp = icmp ult i32 %inc10, 61
+  br i1 %cmp, label %for.body, label %for.end
+
+for.end:
+  %add5.lcssa = phi i32 [ %add5, %for.body ]
+  store i32 %add5.lcssa, i32* %n, align 4
+  ret void
+}