[x86] add restriction for lowering to vpermps

This transform was added with rL351346, and we had
an escape for shufps, but we also want one for
unpckps vs. vpermps because vpermps doesn't take
an immediate shuffle index operand.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@352333 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index ffe64f129bb2daba5d08db0c0e63ca72b2d9c652..bca3e74b7d476736feea32ec108af946c0c1c1ec 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -9719,6 +9719,21 @@ static bool isUnpackWdShuffleMask(ArrayRef<int> Mask, MVT VT) {
   return IsUnpackwdMask;
 }
 
+static bool is128BitUnpackShuffleMask(ArrayRef<int> Mask) {
+  // Create 128-bit vector type based on mask size.
+  MVT EltVT = MVT::getIntegerVT(128 / Mask.size());
+  MVT VT = MVT::getVectorVT(EltVT, Mask.size());
+
+  // Match any of unary/binary or low/high.
+  for (unsigned i = 0; i != 4; ++i) {
+    SmallVector<int, 16> UnpackMask;
+    createUnpackShuffleMask(VT, UnpackMask, (i >> 1) % 2, i % 2);
+    if (isTargetShuffleEquivalent(Mask, UnpackMask))
+      return true;
+  }
+  return false;
+}
+
 /// Get a 4-lane 8-bit shuffle immediate for a mask.
 ///
 /// This helper function produces an 8-bit shuffle immediate corresponding to
@@ -11709,8 +11724,10 @@ static SDValue lowerShuffleOfExtractsAsVperm(const SDLoc &DL, SDValue N0,
     return SDValue();
 
   // Final bailout: if the mask is simple, we are better off using an extract
-  // and a simple narrow shuffle.
-  if (NumElts == 4 && isSingleSHUFPSMask(NewMask))
+  // and a simple narrow shuffle. Prefer extract+unpack(h/l)ps to vpermps
+  // because that avoids a constant load from memory.
+  if (NumElts == 4 &&
+      (isSingleSHUFPSMask(NewMask) || is128BitUnpackShuffleMask(NewMask)))
     return SDValue();
 
   // Extend the shuffle mask with undef elements.

diff --git a/test/CodeGen/X86/vector-shuffle-128-unpck.ll b/test/CodeGen/X86/vector-shuffle-128-unpck.ll
index 4bd4a4810692073b66f1e68f3c2c14ad59582a17..47d9c41e01997b2a43abc550cb9cc6cb6c125335 100644 (file)
--- a/test/CodeGen/X86/vector-shuffle-128-unpck.ll
+++ b/test/CodeGen/X86/vector-shuffle-128-unpck.ll
@@ -45,23 +45,15 @@ define <2 x double> @unpckh_unary_extracted_v8f64(<4 x double> %x) {
   ret <2 x double> %r
 }
 
-; FIXME: vpermps requires a constant load for the index op. It's unlikely to be profitable.
+; vpermps requires a constant load for the index op. It's unlikely to be profitable.
 
 define <4 x i32> @unpckh_unary_extracted_v8i32(<8 x i32> %x) {
-; AVX1-LABEL: unpckh_unary_extracted_v8i32:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX1-NEXT:    vzeroupper
-; AVX1-NEXT:    retq
-;
-; AVX2OR512VL-LABEL: unpckh_unary_extracted_v8i32:
-; AVX2OR512VL:       # %bb.0:
-; AVX2OR512VL-NEXT:    vmovaps {{.*#+}} ymm1 = <2,6,3,7,u,u,u,u>
-; AVX2OR512VL-NEXT:    vpermps %ymm0, %ymm1, %ymm0
-; AVX2OR512VL-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
-; AVX2OR512VL-NEXT:    vzeroupper
-; AVX2OR512VL-NEXT:    retq
+; ALL-LABEL: unpckh_unary_extracted_v8i32:
+; ALL:       # %bb.0:
+; ALL-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; ALL-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; ALL-NEXT:    vzeroupper
+; ALL-NEXT:    retq
   %extrl = shufflevector <8 x i32> %x, <8 x i32> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %extrh = shufflevector <8 x i32> %x, <8 x i32> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
   %r = shufflevector <4 x i32> %extrl, <4 x i32> %extrh, <4 x i32> <i32 2, i32 6, i32 3, i32 7>
@@ -69,20 +61,12 @@ define <4 x i32> @unpckh_unary_extracted_v8i32(<8 x i32> %x) {
 }
 
 define <4 x float> @unpckh_unary_extracted_v8f32(<8 x float> %x) {
-; AVX1-LABEL: unpckh_unary_extracted_v8f32:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
-; AVX1-NEXT:    vzeroupper
-; AVX1-NEXT:    retq
-;
-; AVX2OR512VL-LABEL: unpckh_unary_extracted_v8f32:
-; AVX2OR512VL:       # %bb.0:
-; AVX2OR512VL-NEXT:    vmovaps {{.*#+}} ymm1 = <2,6,3,7,u,u,u,u>
-; AVX2OR512VL-NEXT:    vpermps %ymm0, %ymm1, %ymm0
-; AVX2OR512VL-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
-; AVX2OR512VL-NEXT:    vzeroupper
-; AVX2OR512VL-NEXT:    retq
+; ALL-LABEL: unpckh_unary_extracted_v8f32:
+; ALL:       # %bb.0:
+; ALL-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; ALL-NEXT:    vunpckhps {{.*#+}} xmm0 = xmm0[2],xmm1[2],xmm0[3],xmm1[3]
+; ALL-NEXT:    vzeroupper
+; ALL-NEXT:    retq
   %extrl = shufflevector <8 x float> %x, <8 x float> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %extrh = shufflevector <8 x float> %x, <8 x float> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
   %r = shufflevector <4 x float> %extrl, <4 x float> %extrh, <4 x i32> <i32 2, i32 6, i32 3, i32 7>
@@ -169,23 +153,15 @@ define <2 x double> @unpckl_unary_extracted_v8f64(<4 x double> %x) {
   ret <2 x double> %r
 }
 
-; FIXME: vpermps requires a constant load for the index op. It's unlikely to be profitable.
+; vpermps requires a constant load for the index op. It's unlikely to be profitable.
 
 define <4 x i32> @unpckl_unary_extracted_v8i32(<8 x i32> %x) {
-; AVX1-LABEL: unpckl_unary_extracted_v8i32:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
-; AVX1-NEXT:    vzeroupper
-; AVX1-NEXT:    retq
-;
-; AVX2OR512VL-LABEL: unpckl_unary_extracted_v8i32:
-; AVX2OR512VL:       # %bb.0:
-; AVX2OR512VL-NEXT:    vmovaps {{.*#+}} ymm1 = <0,4,1,5,u,u,u,u>
-; AVX2OR512VL-NEXT:    vpermps %ymm0, %ymm1, %ymm0
-; AVX2OR512VL-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
-; AVX2OR512VL-NEXT:    vzeroupper
-; AVX2OR512VL-NEXT:    retq
+; ALL-LABEL: unpckl_unary_extracted_v8i32:
+; ALL:       # %bb.0:
+; ALL-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; ALL-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; ALL-NEXT:    vzeroupper
+; ALL-NEXT:    retq
   %extrl = shufflevector <8 x i32> %x, <8 x i32> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %extrh = shufflevector <8 x i32> %x, <8 x i32> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
   %r = shufflevector <4 x i32> %extrl, <4 x i32> %extrh, <4 x i32> <i32 0, i32 4, i32 1, i32 5>
@@ -193,20 +169,12 @@ define <4 x i32> @unpckl_unary_extracted_v8i32(<8 x i32> %x) {
 }
 
 define <4 x float> @unpckl_unary_extracted_v8f32(<8 x float> %x) {
-; AVX1-LABEL: unpckl_unary_extracted_v8f32:
-; AVX1:       # %bb.0:
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
-; AVX1-NEXT:    vzeroupper
-; AVX1-NEXT:    retq
-;
-; AVX2OR512VL-LABEL: unpckl_unary_extracted_v8f32:
-; AVX2OR512VL:       # %bb.0:
-; AVX2OR512VL-NEXT:    vmovaps {{.*#+}} ymm1 = <0,4,1,5,u,u,u,u>
-; AVX2OR512VL-NEXT:    vpermps %ymm0, %ymm1, %ymm0
-; AVX2OR512VL-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
-; AVX2OR512VL-NEXT:    vzeroupper
-; AVX2OR512VL-NEXT:    retq
+; ALL-LABEL: unpckl_unary_extracted_v8f32:
+; ALL:       # %bb.0:
+; ALL-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; ALL-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
+; ALL-NEXT:    vzeroupper
+; ALL-NEXT:    retq
   %extrl = shufflevector <8 x float> %x, <8 x float> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
   %extrh = shufflevector <8 x float> %x, <8 x float> undef, <4 x i32> <i32 4, i32 5, i32 6, i32 7>
   %r = shufflevector <4 x float> %extrl, <4 x float> %extrh, <4 x i32> <i32 0, i32 4, i32 1, i32 5>