[X86][AVX] concat_vectors(scalar_to_vector(x),scalar_to_vector(x)) --> broadcast(x)

RKSimon · RKSimon · commit e08f177ea2cf · 2019-02-23T18:34:05.000Z
For AVX1, limit this to i32/f32/i64/f64 loading cases only.

llvm-svn: 354730
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -41961,6 +41961,13 @@ static SDValue combineInsertSubvector(SDNode *N, SelectionDAG &DAG,
                                 SubVec.getOpcode() == X86ISD::SUBV_BROADCAST))
         return DAG.getNode(SubVec.getOpcode(), dl, OpVT, SubVec.getOperand(0));
 
+      // concat_vectors(scalar_to_vector(x),scalar_to_vector(x)) -> broadcast(x)
+      if (SubVec == SubVec2 && SubVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
+          (Subtarget.hasAVX2() || (OpVT.getScalarSizeInBits() >= 32 &&
+                                   MayFoldLoad(SubVec.getOperand(0)))) &&
+          SubVec.getOperand(0).getValueType() == OpVT.getScalarType())
+        return DAG.getNode(X86ISD::VBROADCAST, dl, OpVT, SubVec.getOperand(0));
+
       // If we're inserting all zeros into the upper half, change this to
       // an insert into an all zeros vector. We will match this to a move
       // with implicit upper bit zeroing during isel.
diff --git a/llvm/test/CodeGen/X86/avx512dqvl-intrinsics-upgrade.ll b/llvm/test/CodeGen/X86/avx512dqvl-intrinsics-upgrade.ll
@@ -2718,10 +2718,9 @@ define <8 x i32>@test_int_x86_avx512_mask_broadcasti32x2_256(<4 x i32> %x0, <8 x
 ; X64-LABEL: test_int_x86_avx512_mask_broadcasti32x2_256:
 ; X64:       # %bb.0:
 ; X64-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
-; X64-NEXT:    vmovq (%rsi), %xmm2 # EVEX TO VEX Compression encoding: [0xc5,0xfa,0x7e,0x16]
-; X64-NEXT:    # xmm2 = mem[0],zero
 ; X64-NEXT:    kmovw %edi, %k1 # encoding: [0xc5,0xf8,0x92,0xcf]
-; X64-NEXT:    vinserti32x4 $1, %xmm2, %ymm2, %ymm1 {%k1} # encoding: [0x62,0xf3,0x6d,0x29,0x38,0xca,0x01]
+; X64-NEXT:    vbroadcasti32x2 (%rsi), %ymm1 {%k1} # encoding: [0x62,0xf2,0x7d,0x29,0x59,0x0e]
+; X64-NEXT:    # ymm1 {%k1} = mem[0,1,0,1,0,1,0,1]
 ; X64-NEXT:    vinserti128 $1, %xmm0, %ymm0, %ymm2 # EVEX TO VEX Compression encoding: [0xc4,0xe3,0x7d,0x38,0xd0,0x01]
 ; X64-NEXT:    vinserti32x4 $1, %xmm0, %ymm0, %ymm0 {%k1} {z} # encoding: [0x62,0xf3,0x7d,0xa9,0x38,0xc0,0x01]
 ; X64-NEXT:    vpaddd %ymm2, %ymm0, %ymm0 # EVEX TO VEX Compression encoding: [0xc5,0xfd,0xfe,0xc2]
diff --git a/llvm/test/CodeGen/X86/subvector-broadcast.ll b/llvm/test/CodeGen/X86/subvector-broadcast.ll
@@ -1657,46 +1657,18 @@ define <4 x double> @broadcast_v4f64_v2f64_4u61(<2 x double>* %vp, <4 x double>
 }
 
 define <8 x float> @broadcast_v8f32_v2f32_u1uu0uEu(<2 x float>* %vp, <8 x float> %default) {
-; X32-AVX1-LABEL: broadcast_v8f32_v2f32_u1uu0uEu:
-; X32-AVX1:       # %bb.0:
-; X32-AVX1-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-AVX1-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; X32-AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm1, %ymm1
-; X32-AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
-; X32-AVX1-NEXT:    retl
-;
-; X32-AVX2-LABEL: broadcast_v8f32_v2f32_u1uu0uEu:
-; X32-AVX2:       # %bb.0:
-; X32-AVX2-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-AVX2-NEXT:    vbroadcastsd (%eax), %ymm1
-; X32-AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
-; X32-AVX2-NEXT:    retl
-;
-; X32-AVX512-LABEL: broadcast_v8f32_v2f32_u1uu0uEu:
-; X32-AVX512:       # %bb.0:
-; X32-AVX512-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X32-AVX512-NEXT:    vbroadcastsd (%eax), %ymm1
-; X32-AVX512-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
-; X32-AVX512-NEXT:    retl
-;
-; X64-AVX1-LABEL: broadcast_v8f32_v2f32_u1uu0uEu:
-; X64-AVX1:       # %bb.0:
-; X64-AVX1-NEXT:    vmovsd {{.*#+}} xmm1 = mem[0],zero
-; X64-AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm1, %ymm1
-; X64-AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
-; X64-AVX1-NEXT:    retq
-;
-; X64-AVX2-LABEL: broadcast_v8f32_v2f32_u1uu0uEu:
-; X64-AVX2:       # %bb.0:
-; X64-AVX2-NEXT:    vbroadcastsd (%rdi), %ymm1
-; X64-AVX2-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
-; X64-AVX2-NEXT:    retq
+; X32-LABEL: broadcast_v8f32_v2f32_u1uu0uEu:
+; X32:       # %bb.0:
+; X32-NEXT:    movl {{[0-9]+}}(%esp), %eax
+; X32-NEXT:    vbroadcastsd (%eax), %ymm1
+; X32-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
+; X32-NEXT:    retl
 ;
-; X64-AVX512-LABEL: broadcast_v8f32_v2f32_u1uu0uEu:
-; X64-AVX512:       # %bb.0:
-; X64-AVX512-NEXT:    vbroadcastsd (%rdi), %ymm1
-; X64-AVX512-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
-; X64-AVX512-NEXT:    retq
+; X64-LABEL: broadcast_v8f32_v2f32_u1uu0uEu:
+; X64:       # %bb.0:
+; X64-NEXT:    vbroadcastsd (%rdi), %ymm1
+; X64-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1,2,3,4,5],ymm0[6,7]
+; X64-NEXT:    retq
   %vec = load <2 x float>, <2 x float>* %vp
   %shuf = shufflevector <2 x float> %vec, <2 x float> undef, <8 x i32> <i32 undef, i32 1, i32 undef, i32 undef, i32 0, i32 2, i32 3, i32 undef>
   %res = select <8 x i1> <i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 0, i1 1>, <8 x float> %shuf, <8 x float> %default
