diff options
| -rw-r--r-- | llvm/lib/Target/X86/X86ISelLowering.cpp | 15 | ||||
| -rw-r--r-- | llvm/test/CodeGen/X86/haddsub-shuf.ll | 25 |
2 files changed, 34 insertions, 6 deletions
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index 1f352ae4772..8a63163cade 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -33266,12 +33266,15 @@ static SDValue foldShuffleOfHorizOp(SDNode *N) { // the result is the same as the high half. If a target shuffle is also // replicating low and high halves, we don't need the shuffle. if (Opcode == X86ISD::MOVDDUP || Opcode == X86ISD::VBROADCAST) { - // movddup (hadd X, X) --> hadd X, X - // broadcast (extract_vec_elt (hadd X, X), 0) --> hadd X, X - assert((HOp.getValueType() == MVT::v2f64 || - HOp.getValueType() == MVT::v4f64) && HOp.getValueType() == VT && - "Unexpected type for h-op"); - return HOp; + if (HOp.getScalarValueSizeInBits() == 64) { + // movddup (hadd X, X) --> hadd X, X + // broadcast (extract_vec_elt (hadd X, X), 0) --> hadd X, X + assert((HOp.getValueType() == MVT::v2f64 || + HOp.getValueType() == MVT::v4f64) && HOp.getValueType() == VT && + "Unexpected type for h-op"); + return HOp; + } + return SDValue(); } // shuffle (hadd X, X), undef, [low half...high half] --> hadd X, X diff --git a/llvm/test/CodeGen/X86/haddsub-shuf.ll b/llvm/test/CodeGen/X86/haddsub-shuf.ll index b5b7cfde8ae..0787269072b 100644 --- a/llvm/test/CodeGen/X86/haddsub-shuf.ll +++ b/llvm/test/CodeGen/X86/haddsub-shuf.ll @@ -700,3 +700,28 @@ define <16 x i16> @hsub_v16i16b(<16 x i16> %a) { %shuf = shufflevector <16 x i16> %hop, <16 x i16> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 8, i32 9, i32 10, i32 11, i32 8, i32 9, i32 10, i32 11> ret <16 x i16> %shuf } + +define <4 x float> @broadcast_haddps_v4f32(<4 x float> %a0) { +; SSSE3-LABEL: broadcast_haddps_v4f32: +; SSSE3: # %bb.0: +; SSSE3-NEXT: haddps %xmm0, %xmm0 +; SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,0,0,0] +; SSSE3-NEXT: retq +; +; AVX1-LABEL: broadcast_haddps_v4f32: +; AVX1: # %bb.0: +; AVX1-NEXT: vhaddps %xmm0, %xmm0, %xmm0 +; AVX1-NEXT: vpermilps {{.*#+}} xmm0 = xmm0[0,0,0,0] +; AVX1-NEXT: retq +; +; AVX2-LABEL: broadcast_haddps_v4f32: +; AVX2: # %bb.0: +; AVX2-NEXT: vhaddps %xmm0, %xmm0, %xmm0 +; AVX2-NEXT: vbroadcastss %xmm0, %xmm0 +; AVX2-NEXT: retq + %1 = tail call <4 x float> @llvm.x86.sse3.hadd.ps(<4 x float> %a0, <4 x float> %a0) + %2 = shufflevector <4 x float> %1, <4 x float> undef, <4 x i32> zeroinitializer + ret <4 x float> %2 +} + +declare <4 x float> @llvm.x86.sse3.hadd.ps(<4 x float>, <4 x float>) |
