[X86] combineConcatVectorOps - extend VPERMILPD handling to support 512-bit types

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -58113,11 +58113,14 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
             DAG.getNode(X86ISD::VPERMILPI, DL, FloatVT, Res, Op0.getOperand(1));
         return DAG.getBitcast(VT, Res);
       }
-      // TODO: v8f64 VPERMILPI concatenation.
-      if (!IsSplat && NumOps == 2 && VT == MVT::v4f64) {
-        uint64_t Idx0 = Ops[0].getConstantOperandVal(1);
-        uint64_t Idx1 = Ops[1].getConstantOperandVal(1);
-        uint64_t Idx = ((Idx1 & 3) << 2) | (Idx0 & 3);
+      if (!IsSplat && (VT == MVT::v4f64 || VT == MVT::v8f64)) {
+        unsigned NumSubElts = Op0.getValueType().getVectorNumElements();
+        uint64_t Mask = (1ULL << NumSubElts) - 1;
+        uint64_t Idx = 0;
+        for (unsigned I = 0; I != NumOps; ++I) {
+          uint64_t SubIdx = Ops[I].getConstantOperandVal(1);
+          Idx |= (SubIdx & Mask) << (I * NumSubElts);
+        }
         return DAG.getNode(X86ISD::VPERMILPI, DL, VT,
                            ConcatSubOperand(VT, Ops, 0),
                            DAG.getTargetConstant(Idx, DL, MVT::i8));

llvm/test/CodeGen/X86/vector-shuffle-512-v8.ll

@@ -1678,11 +1678,10 @@ define <8 x double> @concat_shuffle_v8f64_v2f64_10325476(<2 x double> %a0, <2 x
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
 ; AVX512F-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
+; AVX512F-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
 ; AVX512F-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX512F-NEXT:    vshufpd {{.*#+}} ymm0 = ymm0[1,0,3,2]
-; AVX512F-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm1
-; AVX512F-NEXT:    vshufpd {{.*#+}} ymm1 = ymm1[1,0,3,2]
-; AVX512F-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512F-NEXT:    vinsertf64x4 $1, %ymm2, %zmm0, %zmm0
+; AVX512F-NEXT:    vshufpd {{.*#+}} zmm0 = zmm0[1,0,3,2,5,4,7,6]
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512F-32-LABEL: concat_shuffle_v8f64_v2f64_10325476:
@@ -1691,10 +1690,9 @@ define <8 x double> @concat_shuffle_v8f64_v2f64_10325476(<2 x double> %a0, <2 x
 ; AVX512F-32-NEXT:    # kill: def $xmm2 killed $xmm2 def $ymm2
 ; AVX512F-32-NEXT:    # kill: def $xmm0 killed $xmm0 def $ymm0
 ; AVX512F-32-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; AVX512F-32-NEXT:    vshufpd {{.*#+}} ymm0 = ymm0[1,0,3,2]
 ; AVX512F-32-NEXT:    vinsertf128 $1, {{[0-9]+}}(%esp), %ymm2, %ymm1
-; AVX512F-32-NEXT:    vshufpd {{.*#+}} ymm1 = ymm1[1,0,3,2]
 ; AVX512F-32-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
+; AVX512F-32-NEXT:    vshufpd {{.*#+}} zmm0 = zmm0[1,0,3,2,5,4,7,6]
 ; AVX512F-32-NEXT:    addl $12, %esp
 ; AVX512F-32-NEXT:    retl
   %s0 = shufflevector <2 x double> %a0, <2 x double> poison, <2 x i32> <i32 1, i32 0>