Fix: SLPVectorizer tried to move an instruction which was replaced by a vector instruction.

For a detailed description of the problem see the comment in the test file.
The problematic moveBefore() calls are not required anymore because the new
scheduling algorithm ensures a correct ordering anyway.

llvm-svn: 216656

2 files changed, 41 insertions, 4 deletions

diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index c0d44f179bc..bbf3e03854b 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -3112,11 +3112,9 @@ bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
     BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
     BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
     if (tryToVectorizePair(A, B0, R)) {
-      B->moveBefore(V);
       return true;
     }
     if (tryToVectorizePair(A, B1, R)) {
-      B->moveBefore(V);
       return true;
     }
   }
@@ -3126,11 +3124,9 @@ bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
     BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
     BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
     if (tryToVectorizePair(A0, B, R)) {
-      A->moveBefore(V);
       return true;
     }
     if (tryToVectorizePair(A1, B, R)) {
-      A->moveBefore(V);
       return true;
     }
   }
diff --git a/llvm/test/Transforms/SLPVectorizer/X86/crash_binaryop.ll b/llvm/test/Transforms/SLPVectorizer/X86/crash_binaryop.ll
new file mode 100644
index 00000000000..dc99366e1da
--- /dev/null
+++ b/llvm/test/Transforms/SLPVectorizer/X86/crash_binaryop.ll
@@ -0,0 +1,41 @@
+; RUN: opt < %s -basicaa -slp-vectorizer -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7
+
+target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-darwin13.3.0"
+
+@a = common global double 0.000000e+00, align 8
+
+define i32 @fn1() {
+entry:
+  %init = load double* @a, align 8
+  br label %loop
+
+loop:
+  %phi = phi double [ %add2, %loop ], [ %init, %entry ]
+  %postadd1_phi = phi double [ %postadd1, %loop ], [ %init, %entry ]
+  %postadd2_phi = phi double [ %postadd2, %loop ], [ %init, %entry ]
+  %add1 = fadd double %postadd1_phi, undef
+  %add2 = fadd double %postadd2_phi, %phi
+  %mul2 = fmul double %add2, 0.000000e+00
+  %binaryop_B = fadd double %postadd1_phi, %mul2
+  %mul1 = fmul double %add1, 0.000000e+00
+  %tmp = fadd double %postadd2_phi, 0.000000e+00
+
+  ; tryToVectorize() starts with this binary instruction.
+  ; At the same time vectorization wraps around the loop, vectorizes
+  ; postadd1/2 and eventually binary_V and tmp. So binary_V itself is replaced
+  ; with a vector instruction.
+  ; The SLPVectorizer crashed because it tried to use binary_V
+  ; after vectorization to re-arrange instructions.
+  %binary_V = fadd double %mul1, %binaryop_B
+
+  %postadd1 = fadd double %binary_V, 0.000000e+00
+  %postadd2 = fadd double %tmp, 1.000000e+00
+  %tobool = fcmp une double %postadd1, 0.000000e+00
+  br i1 %tobool, label %exit, label %loop
+
+exit:
+  ret i32 1
+}
+
+