[X86][AVX] Support shuffle combining for VBROADCAST with smaller vector sources

getTargetShuffleMask can only do this safely if we're extracting the lowest subvector from a vector of the same result type.

llvm-svn: 352999

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -6773,6 +6773,26 @@ static bool getFauxShuffleMask(SDValue N, SmallVectorImpl<int> &Mask,
     }
     return true;
   }
+  case X86ISD::VBROADCAST: {
+    SDValue Src = N.getOperand(0);
+    MVT SrcVT = Src.getSimpleValueType();
+    if (!SrcVT.isVector())
+      return false;
+
+    if (NumSizeInBits != SrcVT.getSizeInBits()) {
+      assert((NumSizeInBits % SrcVT.getSizeInBits()) == 0 &&
+             "Illegal broadcast type");
+      SrcVT = MVT::getVectorVT(SrcVT.getScalarType(),
+                               NumSizeInBits / SrcVT.getScalarSizeInBits());
+      Src = DAG.getNode(ISD::INSERT_SUBVECTOR, SDLoc(N), SrcVT,
+                        DAG.getUNDEF(SrcVT), Src,
+                        DAG.getIntPtrConstant(0, SDLoc(N)));
+    }
+
+    Ops.push_back(Src);
+    Mask.append(NumElts, 0);
+    return true;
+  }
   case ISD::ZERO_EXTEND_VECTOR_INREG:
   case ISD::ZERO_EXTEND: {
     SDValue Src = N.getOperand(0);

llvm/test/CodeGen/X86/avx512-shuffles/partial_permute.ll

@@ -2174,8 +2174,7 @@ define <4 x i64> @test_masked_8xi64_to_4xi64_perm_mask7(<8 x i64> %vec, <4 x i64
 ; CHECK-LABEL: test_masked_8xi64_to_4xi64_perm_mask7:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vextracti32x4 $2, %zmm0, %xmm3
-; CHECK-NEXT:    vpbroadcastq %xmm3, %ymm3
-; CHECK-NEXT:    vmovdqa {{.*#+}} ymm4 = [2,0,3,7]
+; CHECK-NEXT:    vmovdqa {{.*#+}} ymm4 = [2,0,3,4]
 ; CHECK-NEXT:    vpermi2q %ymm3, %ymm0, %ymm4
 ; CHECK-NEXT:    vptestnmq %ymm2, %ymm2, %k1
 ; CHECK-NEXT:    vpblendmq %ymm4, %ymm1, %ymm0 {%k1}
@@ -2189,9 +2188,8 @@ define <4 x i64> @test_masked_8xi64_to_4xi64_perm_mask7(<8 x i64> %vec, <4 x i64
 define <4 x i64> @test_masked_z_8xi64_to_4xi64_perm_mask7(<8 x i64> %vec, <4 x i64> %mask) {
 ; CHECK-LABEL: test_masked_z_8xi64_to_4xi64_perm_mask7:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextracti32x4 $2, %zmm0, %xmm2
-; CHECK-NEXT:    vpbroadcastq %xmm2, %ymm3
-; CHECK-NEXT:    vmovdqa {{.*#+}} ymm2 = [2,0,3,7]
+; CHECK-NEXT:    vextracti32x4 $2, %zmm0, %xmm3
+; CHECK-NEXT:    vmovdqa {{.*#+}} ymm2 = [2,0,3,4]
 ; CHECK-NEXT:    vptestnmq %ymm1, %ymm1, %k1
 ; CHECK-NEXT:    vpermi2q %ymm3, %ymm0, %ymm2 {%k1} {z}
 ; CHECK-NEXT:    vmovdqa %ymm2, %ymm0
@@ -3806,9 +3804,8 @@ define <4 x double> @test_masked_z_8xdouble_to_4xdouble_perm_mask2(<8 x double>
 define <4 x double> @test_8xdouble_to_4xdouble_perm_mask3(<8 x double> %vec) {
 ; CHECK-LABEL: test_8xdouble_to_4xdouble_perm_mask3:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf32x4 $2, %zmm0, %xmm1
-; CHECK-NEXT:    vbroadcastsd %xmm1, %ymm2
-; CHECK-NEXT:    vmovapd {{.*#+}} ymm1 = [0,2,1,7]
+; CHECK-NEXT:    vextractf32x4 $2, %zmm0, %xmm2
+; CHECK-NEXT:    vmovapd {{.*#+}} ymm1 = [0,2,1,4]
 ; CHECK-NEXT:    vpermi2pd %ymm2, %ymm0, %ymm1
 ; CHECK-NEXT:    vmovapd %ymm1, %ymm0
 ; CHECK-NEXT:    retq
@@ -3819,8 +3816,7 @@ define <4 x double> @test_masked_8xdouble_to_4xdouble_perm_mask3(<8 x double> %v
 ; CHECK-LABEL: test_masked_8xdouble_to_4xdouble_perm_mask3:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vextractf32x4 $2, %zmm0, %xmm3
-; CHECK-NEXT:    vbroadcastsd %xmm3, %ymm3
-; CHECK-NEXT:    vmovapd {{.*#+}} ymm4 = [0,2,1,7]
+; CHECK-NEXT:    vmovapd {{.*#+}} ymm4 = [0,2,1,4]
 ; CHECK-NEXT:    vpermi2pd %ymm3, %ymm0, %ymm4
 ; CHECK-NEXT:    vxorpd %xmm0, %xmm0, %xmm0
 ; CHECK-NEXT:    vcmpeqpd %ymm0, %ymm2, %k1
@@ -3835,9 +3831,8 @@ define <4 x double> @test_masked_8xdouble_to_4xdouble_perm_mask3(<8 x double> %v
 define <4 x double> @test_masked_z_8xdouble_to_4xdouble_perm_mask3(<8 x double> %vec, <4 x double> %mask) {
 ; CHECK-LABEL: test_masked_z_8xdouble_to_4xdouble_perm_mask3:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vextractf32x4 $2, %zmm0, %xmm2
-; CHECK-NEXT:    vbroadcastsd %xmm2, %ymm3
-; CHECK-NEXT:    vmovapd {{.*#+}} ymm2 = [0,2,1,7]
+; CHECK-NEXT:    vextractf32x4 $2, %zmm0, %xmm3
+; CHECK-NEXT:    vmovapd {{.*#+}} ymm2 = [0,2,1,4]
 ; CHECK-NEXT:    vxorpd %xmm4, %xmm4, %xmm4
 ; CHECK-NEXT:    vcmpeqpd %ymm4, %ymm1, %k1
 ; CHECK-NEXT:    vpermi2pd %ymm3, %ymm0, %ymm2 {%k1} {z}

llvm/test/CodeGen/X86/vector-shuffle-avx512.ll

@@ -190,9 +190,10 @@ define <4 x i64> @expand4(<2 x i64> %a ) {
 define <8 x float> @expand5(<4 x float> %a ) {
 ; SKX64-LABEL: expand5:
 ; SKX64:       # %bb.0:
-; SKX64-NEXT:    vbroadcastss %xmm0, %ymm0
+; SKX64-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; SKX64-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; SKX64-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
+; SKX64-NEXT:    vmovaps {{.*#+}} ymm2 = [8,0,10,0,12,0,14,0]
+; SKX64-NEXT:    vpermt2ps %ymm1, %ymm2, %ymm0
 ; SKX64-NEXT:    retq
 ;
 ; KNL64-LABEL: expand5:
@@ -204,9 +205,10 @@ define <8 x float> @expand5(<4 x float> %a ) {
 ;
 ; SKX32-LABEL: expand5:
 ; SKX32:       # %bb.0:
-; SKX32-NEXT:    vbroadcastss %xmm0, %ymm0
+; SKX32-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; SKX32-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; SKX32-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0],ymm0[1],ymm1[2],ymm0[3],ymm1[4],ymm0[5],ymm1[6],ymm0[7]
+; SKX32-NEXT:    vmovaps {{.*#+}} ymm2 = [8,0,10,0,12,0,14,0]
+; SKX32-NEXT:    vpermt2ps %ymm1, %ymm2, %ymm0
 ; SKX32-NEXT:    retl
 ;
 ; KNL32-LABEL: expand5: