[X86][SSE] Relax use limits for lowerAddSubToHorizontalOp (PR32433)

Now that we can use HADD/SUB for scalar additions from any pair of extracted elements (D61263), we can relax the one use limit as we will be able to merge multiple uses into using the same HADD/SUB op.

This exposes a couple of missed opportunities in LowerBuildVectorv4x32 which will be committed separately.

Differential Revision: https://reviews.llvm.org/D61782

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@360594 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 58815524b0a41dbeb33e31b8f2525ac3fadb57d7..d4c2b108537938a265dd1b0fcbceb6096269eab8 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -19033,16 +19033,11 @@ static SDValue lowerAddSubToHorizontalOp(SDValue Op, SelectionDAG &DAG,
   if (!IsFP && !Subtarget.hasSSSE3())
     return Op;
 
-  // Defer forming the minimal horizontal op if the vector source has more than
-  // the 2 extract element uses that we're matching here. In that case, we might
-  // form a horizontal op that includes more than 1 add/sub op.
+  // Extract from a common vector.
   if (LHS.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
       RHS.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
       LHS.getOperand(0) != RHS.getOperand(0) ||
-      !LHS.getOperand(0)->hasNUsesOfValue(2, 0))
-    return Op;
-
-  if (!isa<ConstantSDNode>(LHS.getOperand(1)) ||
+      !isa<ConstantSDNode>(LHS.getOperand(1)) ||
       !isa<ConstantSDNode>(RHS.getOperand(1)) ||
       !shouldUseHorizontalOp(true, DAG, Subtarget))
     return Op;

diff --git a/test/CodeGen/X86/haddsub-undef.ll b/test/CodeGen/X86/haddsub-undef.ll
index c8eb9d7d1c302f3d57cf418b9a4ad2f8f4221368..261cce7f595127ac32eabf480d71cbd8b10a41e2 100644 (file)
--- a/test/CodeGen/X86/haddsub-undef.ll
+++ b/test/CodeGen/X86/haddsub-undef.ll
@@ -186,27 +186,39 @@ define <4 x float> @test7_undef(<4 x float> %a, <4 x float> %b) {
 }
 
 define <4 x float> @test8_undef(<4 x float> %a, <4 x float> %b) {
-; SSE-LABEL: test8_undef:
-; SSE:       # %bb.0:
-; SSE-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; SSE-NEXT:    addss %xmm0, %xmm1
-; SSE-NEXT:    movaps %xmm0, %xmm2
-; SSE-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
-; SSE-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; SSE-NEXT:    addss %xmm2, %xmm0
-; SSE-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
-; SSE-NEXT:    movaps %xmm1, %xmm0
-; SSE-NEXT:    retq
+; SSE-SLOW-LABEL: test8_undef:
+; SSE-SLOW:       # %bb.0:
+; SSE-SLOW-NEXT:    movshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; SSE-SLOW-NEXT:    addss %xmm0, %xmm1
+; SSE-SLOW-NEXT:    movaps %xmm0, %xmm2
+; SSE-SLOW-NEXT:    unpckhpd {{.*#+}} xmm2 = xmm2[1],xmm0[1]
+; SSE-SLOW-NEXT:    shufps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; SSE-SLOW-NEXT:    addss %xmm2, %xmm0
+; SSE-SLOW-NEXT:    movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
+; SSE-SLOW-NEXT:    movaps %xmm1, %xmm0
+; SSE-SLOW-NEXT:    retq
 ;
-; AVX-LABEL: test8_undef:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX-NEXT:    vaddss %xmm1, %xmm0, %xmm1
-; AVX-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
-; AVX-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; AVX-NEXT:    vaddss %xmm0, %xmm2, %xmm0
-; AVX-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
-; AVX-NEXT:    retq
+; SSE-FAST-LABEL: test8_undef:
+; SSE-FAST:       # %bb.0:
+; SSE-FAST-NEXT:    haddps %xmm0, %xmm0
+; SSE-FAST-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,1,1,3]
+; SSE-FAST-NEXT:    retq
+;
+; AVX-SLOW-LABEL: test8_undef:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX-SLOW-NEXT:    vaddss %xmm1, %xmm0, %xmm1
+; AVX-SLOW-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
+; AVX-SLOW-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; AVX-SLOW-NEXT:    vaddss %xmm0, %xmm2, %xmm0
+; AVX-SLOW-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1],xmm0[0],xmm1[3]
+; AVX-SLOW-NEXT:    retq
+;
+; AVX-FAST-LABEL: test8_undef:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vhaddps %xmm0, %xmm0, %xmm0
+; AVX-FAST-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,1,3]
+; AVX-FAST-NEXT:    retq
   %vecext = extractelement <4 x float> %a, i32 0
   %vecext1 = extractelement <4 x float> %a, i32 1
   %add = fadd float %vecext, %vecext1
@@ -355,29 +367,29 @@ define <16 x float> @test13_v16f32_undef(<16 x float> %a, <16 x float> %b) {
 ; AVX1-SLOW-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
 ; AVX1-SLOW-NEXT:    retq
 ;
-; AVX1-FAST-LABEL: test13_v16f32_undef:
-; AVX1-FAST:       # %bb.0:
-; AVX1-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
-; AVX1-FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
-; AVX1-FAST-NEXT:    retq
+; AVX-FAST-LABEL: test13_v16f32_undef:
+; AVX-FAST:       # %bb.0:
+; AVX-FAST-NEXT:    vextractf128 $1, %ymm0, %xmm1
+; AVX-FAST-NEXT:    vhaddps %xmm1, %xmm0, %xmm0
+; AVX-FAST-NEXT:    retq
 ;
-; AVX512-LABEL: test13_v16f32_undef:
-; AVX512:       # %bb.0:
-; AVX512-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
-; AVX512-NEXT:    vaddss %xmm1, %xmm0, %xmm1
-; AVX512-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
-; AVX512-NEXT:    vpermilps {{.*#+}} xmm3 = xmm0[3,1,2,3]
-; AVX512-NEXT:    vaddss %xmm3, %xmm2, %xmm2
-; AVX512-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[2,3]
-; AVX512-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX512-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm0[1,1,3,3]
-; AVX512-NEXT:    vaddss %xmm2, %xmm0, %xmm2
-; AVX512-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0,1],xmm2[0],xmm1[3]
-; AVX512-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
-; AVX512-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
-; AVX512-NEXT:    vaddss %xmm0, %xmm2, %xmm0
-; AVX512-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
-; AVX512-NEXT:    retq
+; AVX512-SLOW-LABEL: test13_v16f32_undef:
+; AVX512-SLOW:       # %bb.0:
+; AVX512-SLOW-NEXT:    vmovshdup {{.*#+}} xmm1 = xmm0[1,1,3,3]
+; AVX512-SLOW-NEXT:    vaddss %xmm1, %xmm0, %xmm1
+; AVX512-SLOW-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
+; AVX512-SLOW-NEXT:    vpermilps {{.*#+}} xmm3 = xmm0[3,1,2,3]
+; AVX512-SLOW-NEXT:    vaddss %xmm3, %xmm2, %xmm2
+; AVX512-SLOW-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[2,3]
+; AVX512-SLOW-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; AVX512-SLOW-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm0[1,1,3,3]
+; AVX512-SLOW-NEXT:    vaddss %xmm2, %xmm0, %xmm2
+; AVX512-SLOW-NEXT:    vinsertps {{.*#+}} xmm1 = xmm1[0,1],xmm2[0],xmm1[3]
+; AVX512-SLOW-NEXT:    vpermilpd {{.*#+}} xmm2 = xmm0[1,0]
+; AVX512-SLOW-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[3,1,2,3]
+; AVX512-SLOW-NEXT:    vaddss %xmm0, %xmm2, %xmm0
+; AVX512-SLOW-NEXT:    vinsertps {{.*#+}} xmm0 = xmm1[0,1,2],xmm0[0]
+; AVX512-SLOW-NEXT:    retq
   %vecext = extractelement <16 x float> %a, i32 0
   %vecext1 = extractelement <16 x float> %a, i32 1
   %add1 = fadd float %vecext, %vecext1

diff --git a/test/CodeGen/X86/phaddsub-undef.ll b/test/CodeGen/X86/phaddsub-undef.ll
index b0be5c7eedea9dabd2b1ac47a5ca21c84973b2c1..fed7ec80d81717cab33353de154be23739d45c3e 100644 (file)
--- a/test/CodeGen/X86/phaddsub-undef.ll
+++ b/test/CodeGen/X86/phaddsub-undef.ll
@@ -160,10 +160,26 @@ define <16 x i32> @test16_v16i32_undef(<16 x i32> %a, <16 x i32> %b) {
 ; SSE-NEXT:    phaddd %xmm0, %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX-LABEL: test16_v16i32_undef:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
-; AVX-NEXT:    retq
+; AVX-SLOW-LABEL: test16_v16i32_undef:
+; AVX-SLOW:       # %bb.0:
+; AVX-SLOW-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
+; AVX-SLOW-NEXT:    retq
+;
+; AVX1-FAST-LABEL: test16_v16i32_undef:
+; AVX1-FAST:       # %bb.0:
+; AVX1-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
+; AVX1-FAST-NEXT:    retq
+;
+; AVX2-FAST-LABEL: test16_v16i32_undef:
+; AVX2-FAST:       # %bb.0:
+; AVX2-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
+; AVX2-FAST-NEXT:    retq
+;
+; AVX512-FAST-LABEL: test16_v16i32_undef:
+; AVX512-FAST:       # %bb.0:
+; AVX512-FAST-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
+; AVX512-FAST-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero
+; AVX512-FAST-NEXT:    retq
   %vecext = extractelement <16 x i32> %a, i32 0
   %vecext1 = extractelement <16 x i32> %a, i32 1
   %add = add i32 %vecext, %vecext1