[X86][SSE2] Fix zero/any extension shuffles that don't start from the first element

Fix for D12561 - we weren't correctly ensuring that the base element for extension was moved to start on a boundary suitable for UNPCKL/H

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@248536 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index f273404..0a089d9 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -7485,13 +7485,15 @@ static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
                                                   MVT::v16i8, PSHUFBMask)));
   }
 
-  // If we are extending from an (odd)offset, shuffle them by 1 element.
-  if (Offset & 1) {
+  // If we are extending from an offset, ensure we start on a boundary that
+  // we can unpack from.
+  int AlignToUnpack = Offset % (NumElements / Scale);
+  if (AlignToUnpack) {
     SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
-    for (int i = 1; i < NumElements; ++i)
-      ShMask[i - 1] = i;
+    for (int i = AlignToUnpack; i < NumElements; ++i)
+      ShMask[i - AlignToUnpack] = i;
     InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
-    Offset--;
+    Offset -= AlignToUnpack;
   }
 
   // Otherwise emit a sequence of unpacks.

diff --git a/test/CodeGen/X86/vector-zext.ll b/test/CodeGen/X86/vector-zext.ll
index eb8fd16..9feac05 100644 (file)
--- a/test/CodeGen/X86/vector-zext.ll
+++ b/test/CodeGen/X86/vector-zext.ll
@@ -1283,7 +1283,7 @@ define <8 x i32> @shuf_zext_8i16_to_8i32_offset3(<8 x i16> %A) nounwind uwtable
 ; SSE2-LABEL: shuf_zext_8i16_to_8i32_offset3:
 ; SSE2:       # BB#0: # %entry
 ; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrldq {{.*#+}} xmm0 = xmm0[2,3,4,5,6,7,8,9,10,11,12,13,14,15],zero,zero
+; SSE2-NEXT:    psrldq {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
 ; SSE2-NEXT:    pxor %xmm2, %xmm2
 ; SSE2-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
 ; SSE2-NEXT:    psrldq {{.*#+}} xmm1 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
@@ -1292,7 +1292,7 @@ define <8 x i32> @shuf_zext_8i16_to_8i32_offset3(<8 x i16> %A) nounwind uwtable
 ; SSSE3-LABEL: shuf_zext_8i16_to_8i32_offset3:
 ; SSSE3:       # BB#0: # %entry
 ; SSSE3-NEXT:    movdqa %xmm0, %xmm1
-; SSSE3-NEXT:    psrldq {{.*#+}} xmm0 = xmm0[2,3,4,5,6,7,8,9,10,11,12,13,14,15],zero,zero
+; SSSE3-NEXT:    psrldq {{.*#+}} xmm0 = xmm0[6,7,8,9,10,11,12,13,14,15],zero,zero,zero,zero,zero,zero
 ; SSSE3-NEXT:    pxor %xmm2, %xmm2
 ; SSSE3-NEXT:    punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
 ; SSSE3-NEXT:    psrldq {{.*#+}} xmm1 = xmm1[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero