[X86][AVX] combineLoopMAddPattern - support 256-bit cases on AVX1 via SplitBinaryOpsAndApply

llvm-svn: 326189

2 files changed, 34 insertions, 36 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 4a9e7e7cb41..5449ef788d8 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -37401,7 +37401,7 @@ static SDValue combineLoopMAddPattern(SDNode *N, SelectionDAG &DAG,
   unsigned RegSize = 128;
   if (Subtarget.useBWIRegs())
     RegSize = 512;
-  else if (Subtarget.hasAVX2())
+  else if (Subtarget.hasAVX())
     RegSize = 256;
   unsigned VectorSize = VT.getVectorNumElements() * 16;
   // If the vector size is less than 128, or greater than the supported RegSize,
@@ -37420,7 +37420,13 @@ static SDValue combineLoopMAddPattern(SDNode *N, SelectionDAG &DAG,
   SDValue N1 = DAG.getNode(ISD::TRUNCATE, DL, ReducedVT, MulOp->getOperand(1));
 
   // Madd vector size is half of the original vector size
-  SDValue Madd = DAG.getNode(X86ISD::VPMADDWD, DL, MAddVT, N0, N1);
+  auto PMADDWDBuilder = [](SelectionDAG &DAG, const SDLoc &DL, SDValue Op0,
+                           SDValue Op1) {
+    MVT VT = MVT::getVectorVT(MVT::i32, Op0.getValueSizeInBits() / 32);
+    return DAG.getNode(X86ISD::VPMADDWD, DL, VT, Op0, Op1);
+  };
+  SDValue Madd = SplitBinaryOpsAndApply(DAG, Subtarget, DL, MAddVT, N0, N1,
+                                        PMADDWDBuilder);
   // Fill the rest of the output with 0
   SDValue Zero = getZeroVector(Madd.getSimpleValueType(), Subtarget, DAG, DL);
   SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Madd, Zero);
diff --git a/llvm/test/CodeGen/X86/madd.ll b/llvm/test/CodeGen/X86/madd.ll
index d21fd4e0ca4..0e472c19cb0 100644
--- a/llvm/test/CodeGen/X86/madd.ll
+++ b/llvm/test/CodeGen/X86/madd.ll
@@ -167,33 +167,25 @@ define i32 @_Z10test_shortPsS_i_512(i16* nocapture readonly, i16* nocapture read
 ; AVX1-NEXT:    .p2align 4, 0x90
 ; AVX1-NEXT:  .LBB1_1: # %vector.body
 ; AVX1-NEXT:    # =>This Inner Loop Header: Depth=1
-; AVX1-NEXT:    vpmovsxwd 8(%rdi,%rcx,2), %xmm2
-; AVX1-NEXT:    vpmovsxwd (%rdi,%rcx,2), %xmm3
-; AVX1-NEXT:    vpackssdw %xmm2, %xmm3, %xmm2
-; AVX1-NEXT:    vpmovsxwd 24(%rdi,%rcx,2), %xmm3
-; AVX1-NEXT:    vpmovsxwd 16(%rdi,%rcx,2), %xmm4
-; AVX1-NEXT:    vpackssdw %xmm3, %xmm4, %xmm3
-; AVX1-NEXT:    vpmovsxwd 8(%rsi,%rcx,2), %xmm4
-; AVX1-NEXT:    vpmovsxwd (%rsi,%rcx,2), %xmm5
-; AVX1-NEXT:    vpackssdw %xmm4, %xmm5, %xmm4
-; AVX1-NEXT:    vpmaddwd %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpmovsxwd 24(%rsi,%rcx,2), %xmm4
-; AVX1-NEXT:    vpmovsxwd 16(%rsi,%rcx,2), %xmm5
-; AVX1-NEXT:    vpackssdw %xmm4, %xmm5, %xmm4
-; AVX1-NEXT:    vpmaddwd %xmm3, %xmm4, %xmm3
-; AVX1-NEXT:    vpaddd %xmm1, %xmm3, %xmm3
-; AVX1-NEXT:    vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4,5,6,7]
-; AVX1-NEXT:    vpaddd %xmm0, %xmm2, %xmm2
-; AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
+; AVX1-NEXT:    vmovdqu (%rdi,%rcx,2), %ymm2
+; AVX1-NEXT:    vmovdqu (%rsi,%rcx,2), %ymm3
+; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm4
+; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm5
+; AVX1-NEXT:    vpmaddwd %xmm4, %xmm5, %xmm4
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm5
+; AVX1-NEXT:    vpaddd %xmm5, %xmm4, %xmm4
+; AVX1-NEXT:    vpmaddwd %xmm2, %xmm3, %xmm2
+; AVX1-NEXT:    vpaddd %xmm1, %xmm2, %xmm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm1, %ymm1
 ; AVX1-NEXT:    addq $16, %rcx
 ; AVX1-NEXT:    cmpq %rcx, %rax
 ; AVX1-NEXT:    jne .LBB1_1
 ; AVX1-NEXT:  # %bb.2: # %middle.block
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm3
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
 ; AVX1-NEXT:    vpaddd %xmm3, %xmm2, %xmm2
-; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
 ; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
@@ -653,25 +645,25 @@ define i32 @_Z9test_charPcS_i(i8* nocapture readonly, i8* nocapture readonly, i3
 ; AVX1-NEXT:    .p2align 4, 0x90
 ; AVX1-NEXT:  .LBB4_1: # %vector.body
 ; AVX1-NEXT:    # =>This Inner Loop Header: Depth=1
-; AVX1-NEXT:    vpmovsxbw (%rdi,%rcx), %xmm2
-; AVX1-NEXT:    vpmovsxbw 8(%rdi,%rcx), %xmm3
+; AVX1-NEXT:    vpmovsxbw 8(%rdi,%rcx), %xmm2
+; AVX1-NEXT:    vpmovsxbw 8(%rsi,%rcx), %xmm3
+; AVX1-NEXT:    vpmaddwd %xmm2, %xmm3, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm3
+; AVX1-NEXT:    vpaddd %xmm3, %xmm2, %xmm2
+; AVX1-NEXT:    vpmovsxbw (%rdi,%rcx), %xmm3
 ; AVX1-NEXT:    vpmovsxbw (%rsi,%rcx), %xmm4
-; AVX1-NEXT:    vpmaddwd %xmm2, %xmm4, %xmm2
-; AVX1-NEXT:    vpmovsxbw 8(%rsi,%rcx), %xmm4
 ; AVX1-NEXT:    vpmaddwd %xmm3, %xmm4, %xmm3
-; AVX1-NEXT:    vpaddd %xmm1, %xmm3, %xmm3
-; AVX1-NEXT:    vblendps {{.*#+}} ymm1 = ymm3[0,1,2,3],ymm1[4,5,6,7]
-; AVX1-NEXT:    vpaddd %xmm0, %xmm2, %xmm2
-; AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
+; AVX1-NEXT:    vpaddd %xmm1, %xmm3, %xmm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
 ; AVX1-NEXT:    addq $16, %rcx
 ; AVX1-NEXT:    cmpq %rcx, %rax
 ; AVX1-NEXT:    jne .LBB4_1
 ; AVX1-NEXT:  # %bb.2: # %middle.block
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm3
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
 ; AVX1-NEXT:    vpaddd %xmm3, %xmm2, %xmm2
-; AVX1-NEXT:    vpaddd %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpaddd %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
 ; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]