AVX512BW: Fix SRA v64i8 lowering. Use PCMPGTM (cmp result in k register) for 512bit...

Differential Revision: http://reviews.llvm.org/D18204

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@263624 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index bf4cc0f..157c18b 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -19233,6 +19233,11 @@ static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
         // ashr(R, 7)  === cmp_slt(R, 0)
         if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
           SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
+          if (VT.is512BitVector()) {
+            assert(VT == MVT::v64i8 && "Unexpected element type!");
+            SDValue CMP = DAG.getNode(X86ISD::PCMPGTM, dl, MVT::v64i1, Zeros, R);
+            return DAG.getNode(ISD::SIGN_EXTEND, dl, VT, CMP);
+          }
           return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
         }
 

diff --git a/test/CodeGen/X86/vector-shift-ashr-512.ll b/test/CodeGen/X86/vector-shift-ashr-512.ll
index 147e58f..eb074ec 100644 (file)
--- a/test/CodeGen/X86/vector-shift-ashr-512.ll
+++ b/test/CodeGen/X86/vector-shift-ashr-512.ll
@@ -376,3 +376,21 @@ define <64 x i8> @splatconstant_shift_v64i8(<64 x i8> %a) nounwind {
   %shift = ashr <64 x i8> %a, <i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3>
   ret <64 x i8> %shift
 }
+
+define <64 x i8> @ashr_const7_v64i8(<64 x i8> %a) {
+; AVX512DQ-LABEL: ashr_const7_v64i8:
+; AVX512DQ:       ## BB#0:
+; AVX512DQ-NEXT:    vpxor %ymm2, %ymm2, %ymm2
+; AVX512DQ-NEXT:    vpcmpgtb %ymm0, %ymm2, %ymm0
+; AVX512DQ-NEXT:    vpcmpgtb %ymm1, %ymm2, %ymm1
+; AVX512DQ-NEXT:    retq
+;
+; AVX512BW-LABEL: ashr_const7_v64i8:
+; AVX512BW:       ## BB#0:
+; AVX512BW-NEXT:    vpxord %zmm1, %zmm1, %zmm1
+; AVX512BW-NEXT:    vpcmpgtb %zmm0, %zmm1, %k0
+; AVX512BW-NEXT:    vpmovm2b %k0, %zmm0
+; AVX512BW-NEXT:    retq
+  %res = ashr <64 x i8> %a, <i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7>
+  ret <64 x i8> %res
+}