[llvm-branch-commits] [llvm] cbbfc82 - [X86][SSE] canonicalizeShuffleMaskWithHorizOp - simplify shuffle(HOP(HOP(X, Y), HOP(Z, W))) style chains.

Simon Pilgrim via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Wed Jan 13 09:24:27 PST 2021


Author: Simon Pilgrim
Date: 2021-01-13T17:19:40Z
New Revision: cbbfc8258615bc971a54c6287abe33c4215d2eac

URL: https://github.com/llvm/llvm-project/commit/cbbfc8258615bc971a54c6287abe33c4215d2eac
DIFF: https://github.com/llvm/llvm-project/commit/cbbfc8258615bc971a54c6287abe33c4215d2eac.diff

LOG: [X86][SSE] canonicalizeShuffleMaskWithHorizOp - simplify shuffle(HOP(HOP(X,Y),HOP(Z,W))) style chains.

See if we can remove the shuffle by resorting a HOP chain so that the HOP args are pre-shuffled.

This initial version just handles (the most common) v4i32/v4f32 hadd/hsub reduction patterns - future work can extend this to v8i16 types plus PACK chains (2f64 HADD/HSUB should already be handled in the half-lane combine code later on).

Added: 
    

Modified: 
    llvm/lib/Target/X86/X86ISelLowering.cpp
    llvm/test/CodeGen/X86/horizontal-sum.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 821cfc5f0c27..d45eb5366bfe 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -36115,6 +36115,38 @@ static SDValue canonicalizeShuffleMaskWithHorizOp(
   int NumEltsPerLane = NumElts / NumLanes;
   int NumHalfEltsPerLane = NumEltsPerLane / 2;
 
+  // See if we can remove the shuffle by resorting the HOP chain so that
+  // the HOP args are pre-shuffled.
+  // TODO: Generalize to any sized/depth chain.
+  // TODO: Add support for PACKSS/PACKUS.
+  if (isHoriz && NumEltsPerLane == 4 && VT0.is128BitVector() &&
+      shouldUseHorizontalOp(Ops.size() == 1, DAG, Subtarget)) {
+    SmallVector<int> ScaledMask;
+    if (scaleShuffleElements(Mask, 4, ScaledMask)) {
+      // Attempt to find a HOP(HOP(X,Y),HOP(Z,W)) source operand.
+      auto GetHOpSrc = [&](int M) {
+        if (M == SM_SentinelUndef)
+          return DAG.getUNDEF(VT0);
+        if (M == SM_SentinelZero)
+          return getZeroVector(VT0.getSimpleVT(), Subtarget, DAG, DL);
+        SDValue Src0 = BC[M / NumElts];
+        SDValue Src1 = Src0.getOperand((M % 4) >= 2);
+        if (Src1.getOpcode() == Opcode0 && Src0->isOnlyUserOf(Src1.getNode()))
+          return Src1.getOperand(M % 2);
+        return SDValue();
+      };
+      SDValue M0 = GetHOpSrc(ScaledMask[0]);
+      SDValue M1 = GetHOpSrc(ScaledMask[1]);
+      SDValue M2 = GetHOpSrc(ScaledMask[2]);
+      SDValue M3 = GetHOpSrc(ScaledMask[3]);
+      if (M0 && M1 && M2 && M3) {
+        SDValue LHS = DAG.getNode(Opcode0, DL, VT0, M0, M1);
+        SDValue RHS = DAG.getNode(Opcode0, DL, VT0, M2, M3);
+        return DAG.getNode(Opcode0, DL, VT0, LHS, RHS);
+      }
+    }
+  }
+
   if (2 < Ops.size())
     return SDValue();
 

diff  --git a/llvm/test/CodeGen/X86/horizontal-sum.ll b/llvm/test/CodeGen/X86/horizontal-sum.ll
index 47d44171d99a..315e795d7a37 100644
--- a/llvm/test/CodeGen/X86/horizontal-sum.ll
+++ b/llvm/test/CodeGen/X86/horizontal-sum.ll
@@ -38,13 +38,9 @@ define <4 x float> @pair_sum_v4f32_v4f32(<4 x float> %0, <4 x float> %1, <4 x fl
 ;
 ; SSSE3-FAST-LABEL: pair_sum_v4f32_v4f32:
 ; SSSE3-FAST:       # %bb.0:
-; SSSE3-FAST-NEXT:    haddps %xmm0, %xmm0
-; SSSE3-FAST-NEXT:    haddps %xmm1, %xmm1
 ; SSSE3-FAST-NEXT:    haddps %xmm1, %xmm0
-; SSSE3-FAST-NEXT:    haddps %xmm2, %xmm2
-; SSSE3-FAST-NEXT:    haddps %xmm3, %xmm3
-; SSSE3-FAST-NEXT:    haddps %xmm2, %xmm3
-; SSSE3-FAST-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm3[2,0]
+; SSSE3-FAST-NEXT:    haddps %xmm3, %xmm2
+; SSSE3-FAST-NEXT:    haddps %xmm2, %xmm0
 ; SSSE3-FAST-NEXT:    retq
 ;
 ; AVX1-SLOW-LABEL: pair_sum_v4f32_v4f32:
@@ -66,18 +62,12 @@ define <4 x float> @pair_sum_v4f32_v4f32(<4 x float> %0, <4 x float> %1, <4 x fl
 ; AVX1-SLOW-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[0]
 ; AVX1-SLOW-NEXT:    retq
 ;
-; AVX1-FAST-LABEL: pair_sum_v4f32_v4f32:
-; AVX1-FAST:       # %bb.0:
-; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
-; AVX1-FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vhaddps %xmm2, %xmm2, %xmm1
-; AVX1-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
-; AVX1-FAST-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,1]
-; AVX1-FAST-NEXT:    vhaddps %xmm3, %xmm3, %xmm1
-; AVX1-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
-; AVX1-FAST-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[0]
-; AVX1-FAST-NEXT:    retq
+; AVX-FAST-LABEL: pair_sum_v4f32_v4f32:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vhaddps %xmm3, %xmm2, %xmm1
+; AVX-FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
+; AVX-FAST-NEXT:    retq
 ;
 ; AVX2-SLOW-LABEL: pair_sum_v4f32_v4f32:
 ; AVX2-SLOW:       # %bb.0:
@@ -97,19 +87,6 @@ define <4 x float> @pair_sum_v4f32_v4f32(<4 x float> %0, <4 x float> %1, <4 x fl
 ; AVX2-SLOW-NEXT:    vaddps %xmm2, %xmm1, %xmm1
 ; AVX2-SLOW-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[0]
 ; AVX2-SLOW-NEXT:    retq
-;
-; AVX2-FAST-LABEL: pair_sum_v4f32_v4f32:
-; AVX2-FAST:       # %bb.0:
-; AVX2-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
-; AVX2-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
-; AVX2-FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
-; AVX2-FAST-NEXT:    vhaddps %xmm2, %xmm2, %xmm1
-; AVX2-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
-; AVX2-FAST-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,3]
-; AVX2-FAST-NEXT:    vhaddps %xmm3, %xmm3, %xmm1
-; AVX2-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
-; AVX2-FAST-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[0]
-; AVX2-FAST-NEXT:    retq
   %5 = shufflevector <4 x float> %0, <4 x float> poison, <2 x i32> <i32 0, i32 2>
   %6 = shufflevector <4 x float> %0, <4 x float> poison, <2 x i32> <i32 1, i32 3>
   %7 = fadd <2 x float> %5, %6
@@ -160,13 +137,9 @@ define <4 x i32> @pair_sum_v4i32_v4i32(<4 x i32> %0, <4 x i32> %1, <4 x i32> %2,
 ;
 ; SSSE3-FAST-LABEL: pair_sum_v4i32_v4i32:
 ; SSSE3-FAST:       # %bb.0:
-; SSSE3-FAST-NEXT:    phaddd %xmm0, %xmm0
-; SSSE3-FAST-NEXT:    phaddd %xmm1, %xmm1
+; SSSE3-FAST-NEXT:    phaddd %xmm3, %xmm2
 ; SSSE3-FAST-NEXT:    phaddd %xmm1, %xmm0
-; SSSE3-FAST-NEXT:    phaddd %xmm2, %xmm2
-; SSSE3-FAST-NEXT:    phaddd %xmm3, %xmm3
-; SSSE3-FAST-NEXT:    phaddd %xmm2, %xmm3
-; SSSE3-FAST-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm3[2,0]
+; SSSE3-FAST-NEXT:    phaddd %xmm2, %xmm0
 ; SSSE3-FAST-NEXT:    retq
 ;
 ; AVX1-SLOW-LABEL: pair_sum_v4i32_v4i32:
@@ -189,20 +162,12 @@ define <4 x i32> @pair_sum_v4i32_v4i32(<4 x i32> %0, <4 x i32> %1, <4 x i32> %2,
 ; AVX1-SLOW-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5],xmm1[6,7]
 ; AVX1-SLOW-NEXT:    retq
 ;
-; AVX1-FAST-LABEL: pair_sum_v4i32_v4i32:
-; AVX1-FAST:       # %bb.0:
-; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vphaddd %xmm1, %xmm1, %xmm1
-; AVX1-FAST-NEXT:    vphaddd %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
-; AVX1-FAST-NEXT:    vphaddd %xmm2, %xmm2, %xmm1
-; AVX1-FAST-NEXT:    vphaddd %xmm1, %xmm1, %xmm1
-; AVX1-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX1-FAST-NEXT:    vphaddd %xmm3, %xmm3, %xmm1
-; AVX1-FAST-NEXT:    vphaddd %xmm1, %xmm1, %xmm1
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,0,0,0]
-; AVX1-FAST-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3,4,5],xmm1[6,7]
-; AVX1-FAST-NEXT:    retq
+; AVX-FAST-LABEL: pair_sum_v4i32_v4i32:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vphaddd %xmm3, %xmm2, %xmm2
+; AVX-FAST-NEXT:    vphaddd %xmm1, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vphaddd %xmm2, %xmm0, %xmm0
+; AVX-FAST-NEXT:    retq
 ;
 ; AVX2-SLOW-LABEL: pair_sum_v4i32_v4i32:
 ; AVX2-SLOW:       # %bb.0:
@@ -222,19 +187,6 @@ define <4 x i32> @pair_sum_v4i32_v4i32(<4 x i32> %0, <4 x i32> %1, <4 x i32> %2,
 ; AVX2-SLOW-NEXT:    vpaddd %xmm2, %xmm1, %xmm1
 ; AVX2-SLOW-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[0]
 ; AVX2-SLOW-NEXT:    retq
-;
-; AVX2-FAST-LABEL: pair_sum_v4i32_v4i32:
-; AVX2-FAST:       # %bb.0:
-; AVX2-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
-; AVX2-FAST-NEXT:    vphaddd %xmm1, %xmm1, %xmm1
-; AVX2-FAST-NEXT:    vphaddd %xmm1, %xmm0, %xmm0
-; AVX2-FAST-NEXT:    vphaddd %xmm2, %xmm2, %xmm1
-; AVX2-FAST-NEXT:    vphaddd %xmm1, %xmm1, %xmm1
-; AVX2-FAST-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,3]
-; AVX2-FAST-NEXT:    vphaddd %xmm3, %xmm3, %xmm1
-; AVX2-FAST-NEXT:    vphaddd %xmm1, %xmm1, %xmm1
-; AVX2-FAST-NEXT:    vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[0]
-; AVX2-FAST-NEXT:    retq
   %5 = shufflevector <4 x i32> %0, <4 x i32> poison, <2 x i32> <i32 0, i32 2>
   %6 = shufflevector <4 x i32> %0, <4 x i32> poison, <2 x i32> <i32 1, i32 3>
   %7 = add <2 x i32> %5, %6
@@ -295,10 +247,8 @@ define <8 x float> @pair_sum_v8f32_v4f32(<4 x float> %0, <4 x float> %1, <4 x fl
 ;
 ; SSSE3-FAST-LABEL: pair_sum_v8f32_v4f32:
 ; SSSE3-FAST:       # %bb.0:
-; SSSE3-FAST-NEXT:    haddps %xmm0, %xmm0
-; SSSE3-FAST-NEXT:    haddps %xmm1, %xmm1
 ; SSSE3-FAST-NEXT:    haddps %xmm1, %xmm0
-; SSSE3-FAST-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2,1,3]
+; SSSE3-FAST-NEXT:    haddps %xmm0, %xmm0
 ; SSSE3-FAST-NEXT:    movaps %xmm2, %xmm1
 ; SSSE3-FAST-NEXT:    haddps %xmm3, %xmm1
 ; SSSE3-FAST-NEXT:    haddps %xmm2, %xmm2
@@ -352,9 +302,8 @@ define <8 x float> @pair_sum_v8f32_v4f32(<4 x float> %0, <4 x float> %1, <4 x fl
 ;
 ; AVX1-FAST-LABEL: pair_sum_v8f32_v4f32:
 ; AVX1-FAST:       # %bb.0:
-; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
 ; AVX1-FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX1-FAST-NEXT:    vhaddps %xmm3, %xmm2, %xmm1
 ; AVX1-FAST-NEXT:    vhaddps %xmm2, %xmm2, %xmm2
 ; AVX1-FAST-NEXT:    vhaddps %xmm3, %xmm3, %xmm3
@@ -366,13 +315,11 @@ define <8 x float> @pair_sum_v8f32_v4f32(<4 x float> %0, <4 x float> %1, <4 x fl
 ; AVX1-FAST-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[0,1],xmm3[1,3]
 ; AVX1-FAST-NEXT:    vinsertps {{.*#+}} xmm2 = xmm2[0,1,2],xmm4[1]
 ; AVX1-FAST-NEXT:    vaddps %xmm2, %xmm1, %xmm1
-; AVX1-FAST-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,1]
+; AVX1-FAST-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 ; AVX1-FAST-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm1[1,0]
 ; AVX1-FAST-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
-; AVX1-FAST-NEXT:    vhaddps %xmm6, %xmm6, %xmm2
-; AVX1-FAST-NEXT:    vhaddps %xmm7, %xmm7, %xmm3
-; AVX1-FAST-NEXT:    vhaddps %xmm3, %xmm2, %xmm2
-; AVX1-FAST-NEXT:    vpermilps {{.*#+}} xmm2 = xmm2[0,2,2,3]
+; AVX1-FAST-NEXT:    vhaddps %xmm7, %xmm6, %xmm2
+; AVX1-FAST-NEXT:    vhaddps %xmm0, %xmm2, %xmm2
 ; AVX1-FAST-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
 ; AVX1-FAST-NEXT:    vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[2]
 ; AVX1-FAST-NEXT:    retq
@@ -407,9 +354,8 @@ define <8 x float> @pair_sum_v8f32_v4f32(<4 x float> %0, <4 x float> %1, <4 x fl
 ;
 ; AVX2-FAST-LABEL: pair_sum_v8f32_v4f32:
 ; AVX2-FAST:       # %bb.0:
-; AVX2-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
-; AVX2-FAST-NEXT:    vhaddps %xmm1, %xmm1, %xmm1
 ; AVX2-FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
+; AVX2-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
 ; AVX2-FAST-NEXT:    vhaddps %xmm4, %xmm4, %xmm1
 ; AVX2-FAST-NEXT:    vhaddps %xmm5, %xmm5, %xmm4
 ; AVX2-FAST-NEXT:    vhaddps %xmm3, %xmm2, %xmm2
@@ -418,13 +364,11 @@ define <8 x float> @pair_sum_v8f32_v4f32(<4 x float> %0, <4 x float> %1, <4 x fl
 ; AVX2-FAST-NEXT:    vshufps {{.*#+}} xmm1 = xmm2[1,3],xmm1[1,3]
 ; AVX2-FAST-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0,1,2],xmm4[1]
 ; AVX2-FAST-NEXT:    vaddps %xmm1, %xmm3, %xmm1
-; AVX2-FAST-NEXT:    vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,1]
+; AVX2-FAST-NEXT:    vmovlhps {{.*#+}} xmm0 = xmm0[0],xmm1[0]
 ; AVX2-FAST-NEXT:    vpermilpd {{.*#+}} xmm1 = xmm1[1,0]
 ; AVX2-FAST-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm1
-; AVX2-FAST-NEXT:    vhaddps %xmm6, %xmm6, %xmm2
-; AVX2-FAST-NEXT:    vhaddps %xmm7, %xmm7, %xmm3
-; AVX2-FAST-NEXT:    vhaddps %xmm3, %xmm2, %xmm2
-; AVX2-FAST-NEXT:    vpermilps {{.*#+}} xmm2 = xmm2[0,2,2,3]
+; AVX2-FAST-NEXT:    vhaddps %xmm7, %xmm6, %xmm2
+; AVX2-FAST-NEXT:    vhaddps %xmm0, %xmm2, %xmm2
 ; AVX2-FAST-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
 ; AVX2-FAST-NEXT:    vshufpd {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[2]
 ; AVX2-FAST-NEXT:    retq
@@ -507,10 +451,8 @@ define <8 x i32> @pair_sum_v8i32_v4i32(<4 x i32> %0, <4 x i32> %1, <4 x i32> %2,
 ;
 ; SSSE3-FAST-LABEL: pair_sum_v8i32_v4i32:
 ; SSSE3-FAST:       # %bb.0:
-; SSSE3-FAST-NEXT:    phaddd %xmm0, %xmm0
-; SSSE3-FAST-NEXT:    phaddd %xmm1, %xmm1
 ; SSSE3-FAST-NEXT:    phaddd %xmm1, %xmm0
-; SSSE3-FAST-NEXT:    pshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
+; SSSE3-FAST-NEXT:    phaddd %xmm0, %xmm0
 ; SSSE3-FAST-NEXT:    phaddd %xmm3, %xmm2
 ; SSSE3-FAST-NEXT:    movdqa %xmm5, %xmm1
 ; SSSE3-FAST-NEXT:    phaddd %xmm4, %xmm5
@@ -565,10 +507,8 @@ define <8 x i32> @pair_sum_v8i32_v4i32(<4 x i32> %0, <4 x i32> %1, <4 x i32> %2,
 ;
 ; AVX1-FAST-LABEL: pair_sum_v8i32_v4i32:
 ; AVX1-FAST:       # %bb.0:
-; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vphaddd %xmm1, %xmm1, %xmm1
 ; AVX1-FAST-NEXT:    vphaddd %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm8 = xmm0[0,2,1,3]
+; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm8
 ; AVX1-FAST-NEXT:    vphaddd %xmm3, %xmm2, %xmm1
 ; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,2,1,3]
 ; AVX1-FAST-NEXT:    vphaddd %xmm2, %xmm2, %xmm2
@@ -586,10 +526,8 @@ define <8 x i32> @pair_sum_v8i32_v4i32(<4 x i32> %0, <4 x i32> %1, <4 x i32> %2,
 ; AVX1-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm8[0],xmm0[0]
 ; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[2,3,2,3]
 ; AVX1-FAST-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; AVX1-FAST-NEXT:    vphaddd %xmm6, %xmm6, %xmm2
-; AVX1-FAST-NEXT:    vphaddd %xmm7, %xmm7, %xmm3
-; AVX1-FAST-NEXT:    vphaddd %xmm3, %xmm2, %xmm2
-; AVX1-FAST-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
+; AVX1-FAST-NEXT:    vphaddd %xmm7, %xmm6, %xmm2
+; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm2, %xmm2
 ; AVX1-FAST-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
 ; AVX1-FAST-NEXT:    vshufpd {{.*#+}} ymm0 = ymm0[0],ymm1[1],ymm0[2],ymm1[2]
 ; AVX1-FAST-NEXT:    retq
@@ -624,25 +562,22 @@ define <8 x i32> @pair_sum_v8i32_v4i32(<4 x i32> %0, <4 x i32> %1, <4 x i32> %2,
 ;
 ; AVX2-FAST-LABEL: pair_sum_v8i32_v4i32:
 ; AVX2-FAST:       # %bb.0:
-; AVX2-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
-; AVX2-FAST-NEXT:    vphaddd %xmm1, %xmm1, %xmm1
 ; AVX2-FAST-NEXT:    vphaddd %xmm1, %xmm0, %xmm0
-; AVX2-FAST-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
-; AVX2-FAST-NEXT:    vphaddd %xmm4, %xmm4, %xmm1
-; AVX2-FAST-NEXT:    vphaddd %xmm5, %xmm5, %xmm4
+; AVX2-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm1
+; AVX2-FAST-NEXT:    vphaddd %xmm1, %xmm0, %xmm0
+; AVX2-FAST-NEXT:    vphaddd %xmm4, %xmm4, %xmm4
+; AVX2-FAST-NEXT:    vphaddd %xmm5, %xmm5, %xmm5
 ; AVX2-FAST-NEXT:    vphaddd %xmm3, %xmm2, %xmm2
-; AVX2-FAST-NEXT:    vshufps {{.*#+}} xmm3 = xmm2[0,2],xmm1[0,3]
-; AVX2-FAST-NEXT:    vinsertps {{.*#+}} xmm3 = xmm3[0,1,2],xmm4[0]
-; AVX2-FAST-NEXT:    vshufps {{.*#+}} xmm1 = xmm2[1,3],xmm1[1,3]
-; AVX2-FAST-NEXT:    vblendps {{.*#+}} xmm1 = xmm1[0,1,2],xmm4[3]
-; AVX2-FAST-NEXT:    vpaddd %xmm1, %xmm3, %xmm1
-; AVX2-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
-; AVX2-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[2,3,2,3]
-; AVX2-FAST-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
-; AVX2-FAST-NEXT:    vphaddd %xmm6, %xmm6, %xmm1
-; AVX2-FAST-NEXT:    vphaddd %xmm7, %xmm7, %xmm2
-; AVX2-FAST-NEXT:    vphaddd %xmm2, %xmm1, %xmm1
-; AVX2-FAST-NEXT:    vpshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
+; AVX2-FAST-NEXT:    vshufps {{.*#+}} xmm3 = xmm2[0,2],xmm4[0,3]
+; AVX2-FAST-NEXT:    vinsertps {{.*#+}} xmm3 = xmm3[0,1,2],xmm5[0]
+; AVX2-FAST-NEXT:    vshufps {{.*#+}} xmm2 = xmm2[1,3],xmm4[1,3]
+; AVX2-FAST-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0,1,2],xmm5[3]
+; AVX2-FAST-NEXT:    vpaddd %xmm2, %xmm3, %xmm2
+; AVX2-FAST-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm2[0]
+; AVX2-FAST-NEXT:    vpshufd {{.*#+}} xmm2 = xmm2[2,3,2,3]
+; AVX2-FAST-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
+; AVX2-FAST-NEXT:    vphaddd %xmm7, %xmm6, %xmm2
+; AVX2-FAST-NEXT:    vphaddd %xmm1, %xmm2, %xmm1
 ; AVX2-FAST-NEXT:    vpbroadcastq %xmm1, %ymm1
 ; AVX2-FAST-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
 ; AVX2-FAST-NEXT:    retq


        


More information about the llvm-branch-commits mailing list