Fix Scalarizer insertion point when replacing PHIs with insertelements

If the Scalarizer scalarized a vector PHI but could not scalarize
all uses of it, it would insert a series of insertelements to reconstruct
the vector PHI value from the scalar ones.  The problem was that it would
emit these insertelements immediately after the PHI, even if there were
other PHIs after it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@197909 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/Scalar/Scalarizer.cpp b/lib/Transforms/Scalar/Scalarizer.cpp
index 33ccb7350d85d7f0674fb6beb0c6b7492a05dde5..e614ecee53a0f9f78d6c461a65e4b476c1d7c904 100644 (file)
--- a/lib/Transforms/Scalar/Scalarizer.cpp
+++ b/lib/Transforms/Scalar/Scalarizer.cpp
@@ -638,8 +638,11 @@ bool Scalarizer::finish() {
       // InsertElements.
       Type *Ty = Op->getType();
       Value *Res = UndefValue::get(Ty);
+      BasicBlock *BB = Op->getParent();
       unsigned Count = Ty->getVectorNumElements();
-      IRBuilder<> Builder(Op->getParent(), Op);
+      IRBuilder<> Builder(BB, Op);
+      if (isa<PHINode>(Op))
+        Builder.SetInsertPoint(BB, BB->getFirstInsertionPt());
       for (unsigned I = 0; I < Count; ++I)
         Res = Builder.CreateInsertElement(Res, CV[I], Builder.getInt32(I),
                                           Op->getName() + ".upto" + Twine(I));

diff --git a/test/Transforms/Scalarizer/basic.ll b/test/Transforms/Scalarizer/basic.ll
index 67e61057bc5e5c052f9ff67f8bb16d04e006e423..1cfc0dd290172854275901f006eec2d75b1dc41c 100644 (file)
--- a/test/Transforms/Scalarizer/basic.ll
+++ b/test/Transforms/Scalarizer/basic.ll
@@ -414,6 +414,35 @@ define void @f13(<4 x float *> *%dest, <4 x [4 x float] *> %ptr, <4 x i32> %i,
   ret void
 }
 
+; Test combinations of vector and non-vector PHIs.
+define <4 x float> @f14(<4 x float> %acc, i32 %count) {
+; CHECK-LABEL: @f14(
+; CHECK: %this_acc.i0 = phi float [ %acc.i0, %entry ], [ %next_acc.i0, %loop ]
+; CHECK: %this_acc.i1 = phi float [ %acc.i1, %entry ], [ %next_acc.i1, %loop ]
+; CHECK: %this_acc.i2 = phi float [ %acc.i2, %entry ], [ %next_acc.i2, %loop ]
+; CHECK: %this_acc.i3 = phi float [ %acc.i3, %entry ], [ %next_acc.i3, %loop ]
+; CHECK: %this_count = phi i32 [ %count, %entry ], [ %next_count, %loop ]
+; CHECK: %this_acc.upto0 = insertelement <4 x float> undef, float %this_acc.i0, i32 0
+; CHECK: %this_acc.upto1 = insertelement <4 x float> %this_acc.upto0, float %this_acc.i1, i32 1
+; CHECK: %this_acc.upto2 = insertelement <4 x float> %this_acc.upto1, float %this_acc.i2, i32 2
+; CHECK: %this_acc = insertelement <4 x float> %this_acc.upto2, float %this_acc.i3, i32 3
+; CHECK: ret <4 x float> %next_acc
+entry:
+  br label %loop
+
+loop:
+  %this_acc = phi <4 x float> [ %acc, %entry ], [ %next_acc, %loop ]
+  %this_count = phi i32 [ %count, %entry ], [ %next_count, %loop ]
+  %foo = call <4 x float> @ext(<4 x float> %this_acc)
+  %next_acc = fadd <4 x float> %this_acc, %foo
+  %next_count = sub i32 %this_count, 1
+  %cmp = icmp eq i32 %next_count, 0
+  br i1 %cmp, label %loop, label %exit
+
+exit:
+  ret <4 x float> %next_acc
+}
+
 !0 = metadata !{ metadata !"root" }
 !1 = metadata !{ metadata !"set1", metadata !0 }
 !2 = metadata !{ metadata !"set2", metadata !0 }