Turn shuffles into unpacks for VT == MVT::v2i64 and MVT::v2f64

too. Patch by Jeff Muizelaar.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@135789 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 472eb647a0eb646f5c8fbfd7bd32f5421dbc1fa9..c3fa9d1fa2deeb572f09d3bd7e63e64dcac7f705 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -5837,11 +5837,9 @@ X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
   // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
   // unpckh_undef). Only use pshufd if speed is more important than size.
   if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
-    if (VT != MVT::v2i64 && VT != MVT::v2f64)
-      return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()), dl, VT, V1, V1, DAG);
+    return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()), dl, VT, V1, V1, DAG);
   if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
-    if (VT != MVT::v2i64 && VT != MVT::v2f64)
-      return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
+    return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
 
   if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
       RelaxedMayFoldVectorLoad(V1))
@@ -6038,12 +6036,10 @@ X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
   }
 
   if (X86::isUNPCKL_v_undef_Mask(SVOp))
-    if (VT != MVT::v2i64 && VT != MVT::v2f64)
-      return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
-                                  dl, VT, V1, V1, DAG);
+    return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
+                               dl, VT, V1, V1, DAG);
   if (X86::isUNPCKH_v_undef_Mask(SVOp))
-    if (VT != MVT::v2i64 && VT != MVT::v2f64)
-      return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
+    return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
 
   // Handle v8i16 specifically since SSE can do byte extraction and insertion.
   if (VT == MVT::v8i16) {

diff --git a/test/CodeGen/X86/vec_shuffle-38.ll b/test/CodeGen/X86/vec_shuffle-38.ll
new file mode 100644 (file)
index 0000000..fe83cd4
--- /dev/null
+++ b/test/CodeGen/X86/vec_shuffle-38.ll
@@ -0,0 +1,26 @@
+; RUN: llc < %s -march=x86-64 | FileCheck %s
+
+define <2 x double> @ld(<2 x double> %p) nounwind optsize ssp {
+; CHECK: unpcklpd
+  %shuffle = shufflevector <2 x double> %p, <2 x double> undef, <2 x i32> zeroinitializer
+  ret <2 x double> %shuffle
+}
+
+define <2 x double> @hd(<2 x double> %p) nounwind optsize ssp {
+; CHECK: unpckhpd
+  %shuffle = shufflevector <2 x double> %p, <2 x double> undef, <2 x i32> <i32 1, i32 1>
+  ret <2 x double> %shuffle
+}
+
+define <2 x i64> @ldi(<2 x i64> %p) nounwind optsize ssp {
+; CHECK: punpcklqdq
+  %shuffle = shufflevector <2 x i64> %p, <2 x i64> undef, <2 x i32> zeroinitializer
+  ret <2 x i64> %shuffle
+}
+
+define <2 x i64> @hdi(<2 x i64> %p) nounwind optsize ssp {
+; CHECK: punpckhqdq
+  %shuffle = shufflevector <2 x i64> %p, <2 x i64> undef, <2 x i32> <i32 1, i32 1>
+  ret <2 x i64> %shuffle
+}
+