X86: Do splat promotion later, so the optimizer can chew on it first.

This catches many cases where we can emit a more efficient shuffle for a
specific mask or when the mask contains undefs. Once the splat is lowered to
unpacks we can't do that anymore.

There is a possibility of moving the promotion after pshufb matching, but I'm
not sure if pshufb with a mask loaded from memory is faster than 3 shuffles, so
I avoided that for now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@173569 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 17 insertions, 27 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 0c12410..6daa9b6 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -5839,6 +5839,11 @@ LowerVECTOR_SHUFFLEv8i16(SDValue Op, const X86Subtarget *Subtarget,
     }
   }
 
+  // Promote splats to a larger type which usually leads to more efficient code.
+  // FIXME: Is this true if pshufb is available?
+  if (SVOp->isSplat())
+    return PromoteSplat(SVOp, DAG);
+
   // If we have SSSE3, and all words of the result are from 1 input vector,
   // case 2 is generated, otherwise case 3 is generated.  If no SSSE3
   // is present, fall back to case 4.
@@ -5972,6 +5977,11 @@ SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
   DebugLoc dl = SVOp->getDebugLoc();
   ArrayRef<int> MaskVals = SVOp->getMask();
 
+  // Promote splats to a larger type which usually leads to more efficient code.
+  // FIXME: Is this true if pshufb is available?
+  if (SVOp->isSplat())
+    return PromoteSplat(SVOp, DAG);
+
   // If we have SSSE3, case 1 is generated when all result bytes come from
   // one of  the inputs.  Otherwise, case 2 is generated.  If no SSSE3 is
   // present, fall back to case 3.
@@ -6669,20 +6679,10 @@ X86TargetLowering::NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG) const {
 
   // Handle splat operations
   if (SVOp->isSplat()) {
-    unsigned NumElem = VT.getVectorNumElements();
-
     // Use vbroadcast whenever the splat comes from a foldable load
     SDValue Broadcast = LowerVectorBroadcast(Op, DAG);
     if (Broadcast.getNode())
       return Broadcast;
-
-    // Handle splats by matching through known shuffle masks
-    if ((VT.is128BitVector() && NumElem <= 4) ||
-        (VT.is256BitVector() && NumElem <= 8))
-      return SDValue();
-
-    // All remaning splats are promoted to target supported vector shuffles.
-    return PromoteSplat(SVOp, DAG);
   }
 
   // Check integer expanding shuffles.
diff --git a/test/CodeGen/X86/avx-splat.ll b/test/CodeGen/X86/avx-splat.ll
index 67e4b40..5c01c2c 100644
--- a/test/CodeGen/X86/avx-splat.ll
+++ b/test/CodeGen/X86/avx-splat.ll
@@ -3,8 +3,8 @@
 
 ; CHECK: vpunpcklbw %xmm
 ; CHECK-NEXT: vpunpckhbw %xmm
+; CHECK-NEXT: vpshufd $85
 ; CHECK-NEXT: vinsertf128 $1
-; CHECK-NEXT: vpermilps $85
 define <32 x i8> @funcA(<32 x i8> %a) nounwind uwtable readnone ssp {
 entry:
   %shuffle = shufflevector <32 x i8> %a, <32 x i8> undef, <32 x i32> <i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5>
@@ -12,8 +12,8 @@ entry:
 }
 
 ; CHECK: vpunpckhwd %xmm
+; CHECK-NEXT: vpshufd $85
 ; CHECK-NEXT: vinsertf128 $1
-; CHECK-NEXT: vpermilps $85
 define <16 x i16> @funcB(<16 x i16> %a) nounwind uwtable readnone ssp {
 entry:
   %shuffle = shufflevector <16 x i16> %a, <16 x i16> undef, <16 x i32> <i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5, i32 5>
diff --git a/test/CodeGen/X86/vec_splat-3.ll b/test/CodeGen/X86/vec_splat-3.ll
index 293ed48..cf0ecf4 100644
--- a/test/CodeGen/X86/vec_splat-3.ll
+++ b/test/CodeGen/X86/vec_splat-3.ll
@@ -1,14 +1,12 @@
 ; RUN: llc <%s -march=x86 -mcpu=penryn -mattr=sse41 | FileCheck %s
 
 ; Splat test for v8i16
-; Should generate with pshufd with masks $0, $85, $170, $255 (each mask is used twice)
 define <8 x i16> @shuf_8i16_0(<8 x i16> %T0, <8 x i16> %T1) nounwind readnone {
 	%tmp6 = shufflevector <8 x i16> %T0, <8 x i16> %T1, <8 x i32> <i32 0, i32 undef, i32 undef, i32 0, i32 undef, i32 undef, i32 undef, i32 undef>
 	ret <8 x i16> %tmp6
 
 ; CHECK: shuf_8i16_0:
-; CHECK: punpcklwd
-; CHECK-NEXT: pshufd $0
+; CHECK: pshuflw $0
 }
 
 define <8 x i16> @shuf_8i16_1(<8 x i16> %T0, <8 x i16> %T1) nounwind readnone {
@@ -16,8 +14,7 @@ define <8 x i16> @shuf_8i16_1(<8 x i16> %T0, <8 x i16> %T1) nounwind readnone {
 	ret <8 x i16> %tmp6
 
 ; CHECK: shuf_8i16_1:
-; CHECK: punpcklwd
-; CHECK-NEXT: pshufd $85
+; CHECK: pshuflw $5
 }
 
 define <8 x i16> @shuf_8i16_2(<8 x i16> %T0, <8 x i16> %T1) nounwind readnone {
@@ -34,8 +31,7 @@ define <8 x i16> @shuf_8i16_3(<8 x i16> %T0, <8 x i16> %T1) nounwind readnone {
 	ret <8 x i16> %tmp6
 
 ; CHECK: shuf_8i16_3:
-; CHECK: punpcklwd
-; CHECK-NEXT: pshufd $-1
+; CHECK: pshuflw $15
 }
 
 define <8 x i16> @shuf_8i16_4(<8 x i16> %T0, <8 x i16> %T1) nounwind readnone {
@@ -43,8 +39,7 @@ define <8 x i16> @shuf_8i16_4(<8 x i16> %T0, <8 x i16> %T1) nounwind readnone {
 	ret <8 x i16> %tmp6
 
 ; CHECK: shuf_8i16_4:
-; CHECK: punpckhwd
-; CHECK-NEXT: pshufd $0
+; CHECK: movhlps
 }
 
 define <8 x i16> @shuf_8i16_5(<8 x i16> %T0, <8 x i16> %T1) nounwind readnone {
@@ -65,7 +60,6 @@ define <8 x i16> @shuf_8i16_6(<8 x i16> %T0, <8 x i16> %T1) nounwind readnone {
 ; CHECK-NEXT: pshufd $-86
 }
 
-
 define <8 x i16> @shuf_8i16_7(<8 x i16> %T0, <8 x i16> %T1) nounwind readnone {
 	%tmp6 = shufflevector <8 x i16> %T0, <8 x i16> %T1, <8 x i32> <i32 7, i32 undef, i32 undef, i32 7, i32 undef, i32 undef, i32 undef, i32 undef>
 	ret <8 x i16> %tmp6
@@ -75,8 +69,6 @@ define <8 x i16> @shuf_8i16_7(<8 x i16> %T0, <8 x i16> %T1) nounwind readnone {
 ; CHECK-NEXT: pshufd $-1
 }
 
-; Should generate with pshufd with masks $0, $85, $170, $255 (each mask is used 4 times)
-
 ; Splat test for v16i8
 define <16 x i8> @shuf_16i8_8(<16 x i8> %T0, <16 x i8> %T1) nounwind readnone {
 	%tmp6 = shufflevector <16 x i8> %T0, <16 x i8> %T1, <16 x i32> <i32 0, i32 undef, i32 undef, i32 0, i32 undef, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
@@ -124,9 +116,7 @@ define <16 x i8> @shuf_16i8_12(<16 x i8> %T0, <16 x i8> %T1) nounwind readnone {
 	ret <16 x i8> %tmp6
 
 ; CHECK: shuf_16i8_12:
-; CHECK: punpcklbw
-; CHECK-NEXT: punpckhbw
-; CHECK-NEXT: pshufd $0
+; CHECK: pshufd $5
 }
 
 define <16 x i8> @shuf_16i8_13(<16 x i8> %T0, <16 x i8> %T1) nounwind readnone {