[X86][AVX] createVariablePermute - fix v2i64/v2f64 VPERMILPD index cr…

…eation. The input indices vector will put the index in bit0, but VPERMILPD actually selects off bit1 - so we need to scale accordingly. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@327159 91177308-0d34-0410-b5e6-96231b3b80d8
RPCS3 · Mar 9, 2018 · c805ab0 · c805ab0
1 parent 9ae21b3
commit c805ab0
Show file tree

Hide file tree

Showing 2 changed files with 4 additions and 0 deletions.
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
@@ -7963,6 +7963,8 @@ SDValue createVariablePermute(MVT VT, SDValue SrcVec, SDValue IndicesVec,
   case MVT::v2f64:
   case MVT::v2i64:
     if (Subtarget.hasAVX()) {
+      // VPERMILPD selects using bit#1 of the index vector, so scale IndicesVec.
+      IndicesVec = DAG.getNode(ISD::ADD, DL, IndicesVT, IndicesVec, IndicesVec);
       Opcode = X86ISD::VPERMILPV;
       ShuffleVT = MVT::v2f64;
     }

diff --git a/test/CodeGen/X86/var-permute-128.ll b/test/CodeGen/X86/var-permute-128.ll
@@ -37,6 +37,7 @@ define <2 x i64> @var_shuffle_v2i64(<2 x i64> %v, <2 x i64> %indices) nounwind {
 ;
 ; AVX-LABEL: var_shuffle_v2i64:
 ; AVX:       # %bb.0:
+; AVX-NEXT:    vpaddq %xmm1, %xmm1, %xmm1
 ; AVX-NEXT:    vpermilpd %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %index0 = extractelement <2 x i64> %indices, i32 0
@@ -221,6 +222,7 @@ define <2 x double> @var_shuffle_v2f64(<2 x double> %v, <2 x i64> %indices) noun
 ;
 ; AVX-LABEL: var_shuffle_v2f64:
 ; AVX:       # %bb.0:
+; AVX-NEXT:    vpaddq %xmm1, %xmm1, %xmm1
 ; AVX-NEXT:    vpermilpd %xmm1, %xmm0, %xmm0
 ; AVX-NEXT:    retq
   %index0 = extractelement <2 x i64> %indices, i32 0