[X86] getFauxShuffleMask - Handle undef + sentinel values in subvector insertion

llvm-svn: 343926

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 6ea5ee1..08a9294 100644 (file)
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -6409,7 +6409,9 @@ static bool getFauxShuffleMask(SDValue N, SmallVectorImpl<int> &Mask,
       Mask.push_back(i);
     for (int i = 0; i != (int)NumSubElts; ++i) {
       int M = SubMask[i];
-      if (0 <= M) {
+      if (M < 0) {
+        Mask[i + InsertIdx] = M;
+      } else {
         int InputIdx = M / NumSubElts;
         int ExtractIdx = SubInputs[InputIdx].getConstantOperandVal(1);
         Mask[i + InsertIdx] = (NumElts * (1 + InputIdx)) + ExtractIdx + M;

diff --git a/llvm/test/CodeGen/X86/vector-shuffle-avx512.ll b/llvm/test/CodeGen/X86/vector-shuffle-avx512.ll
index 0da2bd8..e00e376 100644 (file)
--- a/llvm/test/CodeGen/X86/vector-shuffle-avx512.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-avx512.ll
@@ -541,34 +541,30 @@ define <8 x float> @expand14(<4 x float> %a) {
 define <8 x float> @expand15(<4 x float> %a) {
 ; SKX64-LABEL: expand15:
 ; SKX64:       # %bb.0:
-; SKX64-NEXT:    vpermilps {{.*#+}} xmm1 = xmm0[0,1,1,3]
-; SKX64-NEXT:    vmovaps {{.*#+}} ymm2 = [0,1,8,3,10,5,6,7]
-; SKX64-NEXT:    vbroadcastss {{.*#+}} ymm0 = [2,2,2,2,2,2,2,2]
-; SKX64-NEXT:    vpermt2ps %ymm1, %ymm2, %ymm0
+; SKX64-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,1,3]
+; SKX64-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,0,1,3]
+; SKX64-NEXT:    vblendps {{.*#+}} ymm0 = mem[0,1],ymm0[2],mem[3],ymm0[4],mem[5,6,7]
 ; SKX64-NEXT:    retq
 ;
 ; KNL64-LABEL: expand15:
 ; KNL64:       # %bb.0:
 ; KNL64-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,1,3]
 ; KNL64-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,0,1,3]
-; KNL64-NEXT:    vbroadcastss {{.*#+}} ymm1 = [2,2,2,2,2,2,2,2]
-; KNL64-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3],ymm0[4],ymm1[5,6,7]
+; KNL64-NEXT:    vblendps {{.*#+}} ymm0 = mem[0,1],ymm0[2],mem[3],ymm0[4],mem[5,6,7]
 ; KNL64-NEXT:    retq
 ;
 ; SKX32-LABEL: expand15:
 ; SKX32:       # %bb.0:
-; SKX32-NEXT:    vpermilps {{.*#+}} xmm1 = xmm0[0,1,1,3]
-; SKX32-NEXT:    vmovaps {{.*#+}} ymm2 = [0,1,8,3,10,5,6,7]
-; SKX32-NEXT:    vbroadcastss {{.*#+}} ymm0 = [2,2,2,2,2,2,2,2]
-; SKX32-NEXT:    vpermt2ps %ymm1, %ymm2, %ymm0
+; SKX32-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,1,3]
+; SKX32-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,0,1,3]
+; SKX32-NEXT:    vblendps {{.*#+}} ymm0 = mem[0,1],ymm0[2],mem[3],ymm0[4],mem[5,6,7]
 ; SKX32-NEXT:    retl
 ;
 ; KNL32-LABEL: expand15:
 ; KNL32:       # %bb.0:
 ; KNL32-NEXT:    vpermilps {{.*#+}} xmm0 = xmm0[0,1,1,3]
 ; KNL32-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,0,1,3]
-; KNL32-NEXT:    vbroadcastss {{.*#+}} ymm1 = [2,2,2,2,2,2,2,2]
-; KNL32-NEXT:    vblendps {{.*#+}} ymm0 = ymm1[0,1],ymm0[2],ymm1[3],ymm0[4],ymm1[5,6,7]
+; KNL32-NEXT:    vblendps {{.*#+}} ymm0 = mem[0,1],ymm0[2],mem[3],ymm0[4],mem[5,6,7]
 ; KNL32-NEXT:    retl
    %addV = fadd <4 x float> <float 0.0,float 1.0,float 2.0,float 0.0> , <float 0.0,float 1.0,float 2.0,float 0.0>
    %res = shufflevector <4 x float> %addV, <4 x float> %a, <8 x i32> <i32 0, i32 1, i32 4, i32 0, i32 5, i32 0, i32 0, i32 0>