[x86] Add a slight variation on some of the other generic shuffle

lowerings -- one which decomposes into an initial blend followed by
a permute.

Particularly on newer chips, blends are handled independently of
shuffles and so this is much less bottlenecked on the single port that
floating point shuffles are executed with on Intel.

I'll be adding this lowering to a bunch of other code paths in
subsequent commits to handle still more places where we can effectively
leverage blends when they're available in the ISA.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229292 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index cf1583d..3bf4349 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -7543,6 +7543,38 @@ static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
   }
 }
 
+/// \brief Try to lower as a blend of elements from two inputs followed by
+/// a single-input permutation.
+///
+/// This matches the pattern where we can blend elements from two inputs and
+/// then reduce the shuffle to a single-input permutation.
+static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
+                                                   SDValue V2,
+                                                   ArrayRef<int> Mask,
+                                                   SelectionDAG &DAG) {
+  // We build up the blend mask while checking whether a blend is a viable way
+  // to reduce the shuffle.
+  SmallVector<int, 32> BlendMask(Mask.size(), -1);
+  SmallVector<int, 32> PermuteMask(Mask.size(), -1);
+
+  for (int i = 0, Size = Mask.size(); i < Size; ++i) {
+    if (Mask[i] < 0)
+      continue;
+
+    assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
+
+    if (BlendMask[Mask[i] % Size] == -1)
+      BlendMask[Mask[i] % Size] = Mask[i];
+    else if (BlendMask[Mask[i] % Size] != Mask[i])
+      return SDValue(); // Can't blend in the needed input!
+
+    PermuteMask[i] = Mask[i] % Size;
+  }
+
+  SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
+  return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
+}
+
 /// \brief Generic routine to lower a shuffle and blend as a decomposed set of
 /// unblended shuffles followed by an unshuffled blend.
 ///
@@ -8552,6 +8584,24 @@ static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
                      DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
 }
 
+/// \brief Test whether this can be lowered with a single SHUFPS instruction.
+///
+/// This is used to disable more specialized lowerings when the shufps lowering
+/// will happen to be efficient.
+static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
+  // This routine only handles 128-bit shufps.
+  assert(Mask.size() == 4 && "Unsupported mask size!");
+
+  // To lower with a single SHUFPS we need to have the low half and high half
+  // each requiring a single input.
+  if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4 != Mask[1] < 4))
+    return false;
+  if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4 != Mask[3] < 4))
+    return false;
+
+  return true;
+}
+
 /// \brief Lower a vector shuffle using the SHUFPS instruction.
 ///
 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
@@ -8712,6 +8762,11 @@ static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
     // Use INSERTPS if we can complete the shuffle efficiently.
     if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
       return V;
+
+    if (!isSingleSHUFPSMask(Mask))
+      if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
+              DL, MVT::v4f32, V1, V2, Mask, DAG))
+        return BlendPerm;
   }
 
   // Otherwise fall back to a SHUFPS lowering strategy.

diff --git a/test/CodeGen/X86/vector-shuffle-256-v8.ll b/test/CodeGen/X86/vector-shuffle-256-v8.ll
index 45d05a3..7c23765 100644 (file)
--- a/test/CodeGen/X86/vector-shuffle-256-v8.ll
+++ b/test/CodeGen/X86/vector-shuffle-256-v8.ll
@@ -295,11 +295,11 @@ define <8 x float> @shuffle_v8f32_08192a3b(<8 x float> %a, <8 x float> %b) {
 define <8 x float> @shuffle_v8f32_08991abb(<8 x float> %a, <8 x float> %b) {
 ; AVX1-LABEL: shuffle_v8f32_08991abb:
 ; AVX1:       # BB#0:
-; AVX1-NEXT:    vshufps {{.*#+}} xmm2 = xmm0[1,0],xmm1[2,0]
-; AVX1-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[0,2],xmm1[3,3]
-; AVX1-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0]
-; AVX1-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[1,1]
-; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
+; AVX1-NEXT:    vshufps {{.*#+}} xmm2 = xmm0[0,0],xmm1[0,0]
+; AVX1-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[0,2],xmm1[1,1]
+; AVX1-NEXT:    vblendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[1,2,3,3]
+; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: shuffle_v8f32_08991abb:
@@ -1133,11 +1133,11 @@ define <8 x i32> @shuffle_v8i32_08192a3b(<8 x i32> %a, <8 x i32> %b) {
 define <8 x i32> @shuffle_v8i32_08991abb(<8 x i32> %a, <8 x i32> %b) {
 ; AVX1-LABEL: shuffle_v8i32_08991abb:
 ; AVX1:       # BB#0:
-; AVX1-NEXT:    vshufps {{.*#+}} xmm2 = xmm0[1,0],xmm1[2,0]
-; AVX1-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[0,2],xmm1[3,3]
-; AVX1-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[0,0]
-; AVX1-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[1,1]
-; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
+; AVX1-NEXT:    vshufps {{.*#+}} xmm2 = xmm0[0,0],xmm1[0,0]
+; AVX1-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[0,2],xmm1[1,1]
+; AVX1-NEXT:    vblendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
+; AVX1-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[1,2,3,3]
+; AVX1-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: shuffle_v8i32_08991abb:

diff --git a/test/CodeGen/X86/vector-shuffle-combining.ll b/test/CodeGen/X86/vector-shuffle-combining.ll
index 67da5b8..613f7c6 100644 (file)
--- a/test/CodeGen/X86/vector-shuffle-combining.ll
+++ b/test/CodeGen/X86/vector-shuffle-combining.ll
@@ -875,16 +875,28 @@ define <4 x i32> @combine_nested_undef_test15(<4 x i32> %A, <4 x i32> %B) {
 }
 
 define <4 x i32> @combine_nested_undef_test16(<4 x i32> %A, <4 x i32> %B) {
-; SSE-LABEL: combine_nested_undef_test16:
-; SSE:       # BB#0:
-; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[1,3]
-; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2,1,3]
-; SSE-NEXT:    retq
+; SSE2-LABEL: combine_nested_undef_test16:
+; SSE2:       # BB#0:
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[1,3]
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2,1,3]
+; SSE2-NEXT:    retq
+;
+; SSSE3-LABEL: combine_nested_undef_test16:
+; SSSE3:       # BB#0:
+; SSSE3-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[1,3]
+; SSSE3-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2,1,3]
+; SSSE3-NEXT:    retq
+;
+; SSE41-LABEL: combine_nested_undef_test16:
+; SSE41:       # BB#0:
+; SSE41-NEXT:    blendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3]
+; SSE41-NEXT:    shufps {{.*#+}} xmm0 = xmm0[2,1,0,3]
+; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: combine_nested_undef_test16:
 ; AVX:       # BB#0:
-; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[2,0],xmm1[1,3]
-; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,2,1,3]
+; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2],xmm1[3]
+; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[2,1,0,3]
 ; AVX-NEXT:    retq
   %1 = shufflevector <4 x i32> %A, <4 x i32> %B, <4 x i32> <i32 0, i32 5, i32 2, i32 7>
   %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 2, i32 1, i32 0, i32 3>
@@ -892,16 +904,28 @@ define <4 x i32> @combine_nested_undef_test16(<4 x i32> %A, <4 x i32> %B) {
 }
 
 define <4 x i32> @combine_nested_undef_test17(<4 x i32> %A, <4 x i32> %B) {
-; SSE-LABEL: combine_nested_undef_test17:
-; SSE:       # BB#0:
-; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[1,0]
-; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1],xmm1[0,2]
-; SSE-NEXT:    retq
+; SSE2-LABEL: combine_nested_undef_test17:
+; SSE2:       # BB#0:
+; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[1,0]
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1],xmm1[0,2]
+; SSE2-NEXT:    retq
+;
+; SSSE3-LABEL: combine_nested_undef_test17:
+; SSSE3:       # BB#0:
+; SSSE3-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[1,0]
+; SSSE3-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1],xmm1[0,2]
+; SSSE3-NEXT:    retq
+;
+; SSE41-LABEL: combine_nested_undef_test17:
+; SSE41:       # BB#0:
+; SSE41-NEXT:    blendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
+; SSE41-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,0,1]
+; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: combine_nested_undef_test17:
 ; AVX:       # BB#0:
-; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[1,0]
-; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[3,1],xmm1[0,2]
+; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
+; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,0,1]
 ; AVX-NEXT:    retq
   %1 = shufflevector <4 x i32> %A, <4 x i32> %B, <4 x i32> <i32 4, i32 1, i32 3, i32 1>
   %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 2, i32 1, i32 0, i32 3>
@@ -924,17 +948,30 @@ define <4 x i32> @combine_nested_undef_test18(<4 x i32> %A, <4 x i32> %B) {
 }
 
 define <4 x i32> @combine_nested_undef_test19(<4 x i32> %A, <4 x i32> %B) {
-; SSE-LABEL: combine_nested_undef_test19:
-; SSE:       # BB#0:
-; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,0],xmm0[0,0]
-; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2],xmm0[0,0]
-; SSE-NEXT:    movaps %xmm1, %xmm0
-; SSE-NEXT:    retq
+; SSE2-LABEL: combine_nested_undef_test19:
+; SSE2:       # BB#0:
+; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,0],xmm0[0,0]
+; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2],xmm0[0,0]
+; SSE2-NEXT:    movaps %xmm1, %xmm0
+; SSE2-NEXT:    retq
+;
+; SSSE3-LABEL: combine_nested_undef_test19:
+; SSSE3:       # BB#0:
+; SSSE3-NEXT:    shufps {{.*#+}} xmm1 = xmm1[1,0],xmm0[0,0]
+; SSSE3-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2],xmm0[0,0]
+; SSSE3-NEXT:    movaps %xmm1, %xmm0
+; SSSE3-NEXT:    retq
+;
+; SSE41-LABEL: combine_nested_undef_test19:
+; SSE41:       # BB#0:
+; SSE41-NEXT:    blendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3]
+; SSE41-NEXT:    shufps {{.*#+}} xmm0 = xmm0[1,0,0,0]
+; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: combine_nested_undef_test19:
 ; AVX:       # BB#0:
-; AVX-NEXT:    vshufps {{.*#+}} xmm1 = xmm1[1,0],xmm0[0,0]
-; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm1[0,2],xmm0[0,0]
+; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1],xmm0[2,3]
+; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[1,0,0,0]
 ; AVX-NEXT:    retq
   %1 = shufflevector <4 x i32> %A, <4 x i32> %B, <4 x i32> <i32 0, i32 4, i32 5, i32 6>
   %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 2, i32 0, i32 0, i32 0>
@@ -942,17 +979,30 @@ define <4 x i32> @combine_nested_undef_test19(<4 x i32> %A, <4 x i32> %B) {
 }
 
 define <4 x i32> @combine_nested_undef_test20(<4 x i32> %A, <4 x i32> %B) {
-; SSE-LABEL: combine_nested_undef_test20:
-; SSE:       # BB#0:
-; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[2,3]
-; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2,3,1]
-; SSE-NEXT:    movaps %xmm1, %xmm0
-; SSE-NEXT:    retq
+; SSE2-LABEL: combine_nested_undef_test20:
+; SSE2:       # BB#0:
+; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[2,3]
+; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2,3,1]
+; SSE2-NEXT:    movaps %xmm1, %xmm0
+; SSE2-NEXT:    retq
+;
+; SSSE3-LABEL: combine_nested_undef_test20:
+; SSSE3:       # BB#0:
+; SSSE3-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[2,3]
+; SSSE3-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2,3,1]
+; SSSE3-NEXT:    movaps %xmm1, %xmm0
+; SSSE3-NEXT:    retq
+;
+; SSE41-LABEL: combine_nested_undef_test20:
+; SSE41:       # BB#0:
+; SSE41-NEXT:    blendpd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
+; SSE41-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2,3,0]
+; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: combine_nested_undef_test20:
 ; AVX:       # BB#0:
-; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm1[0,0],xmm0[2,3]
-; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,2,3,1]
+; AVX-NEXT:    vblendpd {{.*#+}} xmm0 = xmm1[0],xmm0[1]
+; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,2,3,0]
 ; AVX-NEXT:    retq
   %1 = shufflevector <4 x i32> %A, <4 x i32> %B, <4 x i32> <i32 3, i32 2, i32 4, i32 4>
   %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 2, i32 1, i32 0, i32 3>
@@ -960,17 +1010,30 @@ define <4 x i32> @combine_nested_undef_test20(<4 x i32> %A, <4 x i32> %B) {
 }
 
 define <4 x i32> @combine_nested_undef_test21(<4 x i32> %A, <4 x i32> %B) {
-; SSE-LABEL: combine_nested_undef_test21:
-; SSE:       # BB#0:
-; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[1,1]
-; SSE-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2,1,3]
-; SSE-NEXT:    movaps %xmm1, %xmm0
-; SSE-NEXT:    retq
+; SSE2-LABEL: combine_nested_undef_test21:
+; SSE2:       # BB#0:
+; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[1,1]
+; SSE2-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2,1,3]
+; SSE2-NEXT:    movaps %xmm1, %xmm0
+; SSE2-NEXT:    retq
+;
+; SSSE3-LABEL: combine_nested_undef_test21:
+; SSSE3:       # BB#0:
+; SSSE3-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,0],xmm0[1,1]
+; SSSE3-NEXT:    shufps {{.*#+}} xmm1 = xmm1[0,2,1,3]
+; SSSE3-NEXT:    movaps %xmm1, %xmm0
+; SSSE3-NEXT:    retq
+;
+; SSE41-LABEL: combine_nested_undef_test21:
+; SSE41:       # BB#0:
+; SSE41-NEXT:    blendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
+; SSE41-NEXT:    movddup {{.*#+}} xmm0 = xmm0[0,0]
+; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: combine_nested_undef_test21:
 ; AVX:       # BB#0:
-; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm1[0,0],xmm0[1,1]
-; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,2,1,3]
+; AVX-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1],xmm1[2,3]
+; AVX-NEXT:    vmovddup {{.*#+}} xmm0 = xmm0[0,0]
 ; AVX-NEXT:    retq
   %1 = shufflevector <4 x i32> %A, <4 x i32> %B, <4 x i32> <i32 4, i32 1, i32 3, i32 1>
   %2 = shufflevector <4 x i32> %1, <4 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 3>
@@ -1599,16 +1662,28 @@ define <4 x float> @combine_test2b(<4 x float> %a, <4 x float> %b) {
 }
 
 define <4 x float> @combine_test3b(<4 x float> %a, <4 x float> %b) {
-; SSE-LABEL: combine_test3b:
-; SSE:       # BB#0:
-; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[3,0]
-; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[2,3]
-; SSE-NEXT:    retq
+; SSE2-LABEL: combine_test3b:
+; SSE2:       # BB#0:
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[3,0]
+; SSE2-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[2,3]
+; SSE2-NEXT:    retq
+;
+; SSSE3-LABEL: combine_test3b:
+; SSSE3:       # BB#0:
+; SSSE3-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[3,0]
+; SSSE3-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[2,3]
+; SSSE3-NEXT:    retq
+;
+; SSE41-LABEL: combine_test3b:
+; SSE41:       # BB#0:
+; SSE41-NEXT:    blendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
+; SSE41-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,3,2,3]
+; SSE41-NEXT:    retq
 ;
 ; AVX-LABEL: combine_test3b:
 ; AVX:       # BB#0:
-; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,0],xmm1[3,0]
-; AVX-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[2,3]
+; AVX-NEXT:    vblendpd {{.*#+}} xmm0 = xmm0[0],xmm1[1]
+; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,3,2,3]
 ; AVX-NEXT:    retq
   %1 = shufflevector <4 x float> %a, <4 x float> %b, <4 x i32> <i32 0, i32 0, i32 6, i32 3>
   %2 = shufflevector <4 x float> %1, <4 x float> %b, <4 x i32> <i32 0, i32 7, i32 2, i32 7>