[AVX512] Fix intrinsic cmp{sd|ss} lowering.

[igorb]

Summary

[AVX512] Fix intrinsic cmp{sd|ss} lowering. Intrinsic should return 1/0 and not FF/0.

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

case CMP_MASK_SCALAR_CC: {
       if(!Cmp.getNode())
         Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
 
       SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
                                              DAG.getTargetConstant(0, dl,
                                                                    MVT::i1),
                                              Subtarget, DAG);
 
-      return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
-                         DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
-                         DAG.getValueType(MVT::i1));
+      return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, CmpMask);
     }
     case COMI: { // Comparison intrinsics
       ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
       SDValue LHS = Op.getOperand(1);
       SDValue RHS = Op.getOperand(2);
       SDValue Comi = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
       SDValue InvComi = DAG.getNode(IntrData->Opc0, dl, MVT::i32, RHS, LHS);
       SDValue SetCC;

llvm/trunk/test/CodeGen/X86/avx512-intrinsics.ll

 
 define i8@test_int_x86_avx512_mask_cmp_sd(<2 x double> %x0, <2 x double> %x1, i8 %x3, i32 %x4) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_cmp_sd:
 ; CHECK:       ## BB#0:
 ; CHECK-NEXT:    andl $1, %edi
 ; CHECK-NEXT:    kmovw %edi, %k1
 ; CHECK-NEXT:    vcmpnltsd {sae}, %xmm1, %xmm0, %k0 {%k1}
 ; CHECK-NEXT:    kmovw %k0, %eax
-; CHECK-NEXT:    shlb $7, %al
-; CHECK-NEXT:    sarb $7, %al
+; CHECK-NEXT:    andl $1, %eax
 ; CHECK-NEXT:    retq
 
   %res4 = call i8 @llvm.x86.avx512.mask.cmp.sd(<2 x double> %x0, <2 x double> %x1, i32 5, i8 %x3, i32 8)
   ret i8 %res4
 }
 
 define i8@test_int_x86_avx512_mask_cmp_sd_all(<2 x double> %x0, <2 x double> %x1, i8 %x3, i32 %x4) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_cmp_sd_all:
 ; CHECK:       ## BB#0:
 ; CHECK-NEXT:    vcmpunordsd {sae}, %xmm1, %xmm0, %k0
 ; CHECK-NEXT:    vcmplesd %xmm1, %xmm0, %k1
 ; CHECK-NEXT:    korw %k0, %k1, %k0
 ; CHECK-NEXT:    vcmpnltsd {sae}, %xmm1, %xmm0, %k1
 ; CHECK-NEXT:    vcmpneqsd %xmm1, %xmm0, %k2
 ; CHECK-NEXT:    korw %k1, %k2, %k1
 ; CHECK-NEXT:    andl $1, %edi
 ; CHECK-NEXT:    kmovw %edi, %k2
 ; CHECK-NEXT:    kandw %k2, %k1, %k1
 ; CHECK-NEXT:    korw %k1, %k0, %k0
 ; CHECK-NEXT:    kmovw %k0, %eax
-; CHECK-NEXT:    shlb $7, %al
-; CHECK-NEXT:    sarb $7, %al
+; CHECK-NEXT:    andl $1, %eax
 ; CHECK-NEXT:    retq
 
   %res1 = call i8 @llvm.x86.avx512.mask.cmp.sd(<2 x double> %x0, <2 x double> %x1, i32 2, i8 -1, i32 4)
   %res2 = call i8 @llvm.x86.avx512.mask.cmp.sd(<2 x double> %x0, <2 x double> %x1, i32 3, i8 -1, i32 8)
   %res3 = call i8 @llvm.x86.avx512.mask.cmp.sd(<2 x double> %x0, <2 x double> %x1, i32 4, i8 %x3, i32 4)
   %res4 = call i8 @llvm.x86.avx512.mask.cmp.sd(<2 x double> %x0, <2 x double> %x1, i32 5, i8 %x3, i32 8)
 
   %res11 = or i8 %res1, %res2
   %res12 = or i8 %res3, %res4
   %res13 = or i8 %res11, %res12
   ret i8 %res13
 }
 
 declare i8 @llvm.x86.avx512.mask.cmp.ss(<4 x float>, <4 x float>, i32, i8, i32)
 
 define i8@test_int_x86_avx512_mask_cmp_ss(<4 x float> %x0, <4 x float> %x1, i8 %x3, i32 %x4) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_cmp_ss:
 ; CHECK:       ## BB#0:
 ; CHECK-NEXT:    andl $1, %edi
 ; CHECK-NEXT:    kmovw %edi, %k1
 ; CHECK-NEXT:    vcmpunordss %xmm1, %xmm0, %k0 {%k1}
 ; CHECK-NEXT:    kmovw %k0, %eax
-; CHECK-NEXT:    shlb $7, %al
-; CHECK-NEXT:    sarb $7, %al
+; CHECK-NEXT:    andl $1, %eax
 ; CHECK-NEXT:    retq
 
   %res2 = call i8 @llvm.x86.avx512.mask.cmp.ss(<4 x float> %x0, <4 x float> %x1, i32 3, i8 %x3, i32 4)
   ret i8 %res2
 }
 
 
 define i8@test_int_x86_avx512_mask_cmp_ss_all(<4 x float> %x0, <4 x float> %x1, i8 %x3, i32 %x4) {
 ; CHECK-LABEL: test_int_x86_avx512_mask_cmp_ss_all:
 ; CHECK:       ## BB#0:
 ; CHECK-NEXT:    vcmpless %xmm1, %xmm0, %k1
 ; CHECK-NEXT:    vcmpunordss {sae}, %xmm1, %xmm0, %k0 {%k1}
-; CHECK-NEXT:    vcmpneqss %xmm1, %xmm0, %k1
-; CHECK-NEXT:    vcmpnltss {sae}, %xmm1, %xmm0, %k1 {%k1}
 ; CHECK-NEXT:    andl $1, %edi
-; CHECK-NEXT:    kmovw %edi, %k2
-; CHECK-NEXT:    kandw %k2, %k1, %k1
-; CHECK-NEXT:    kandw %k1, %k0, %k0
+; CHECK-NEXT:    kmovw %edi, %k1
+; CHECK-NEXT:    vcmpneqss %xmm1, %xmm0, %k2 {%k1}
+; CHECK-NEXT:    kmovw %k2, %ecx
+; CHECK-NEXT:    vcmpnltss {sae}, %xmm1, %xmm0, %k1 {%k1}
+; CHECK-NEXT:    kmovw %k1, %edx
+; CHECK-NEXT:    andl $1, %edx
 ; CHECK-NEXT:    kmovw %k0, %eax
-; CHECK-NEXT:    shlb $7, %al
-; CHECK-NEXT:    sarb $7, %al
+; CHECK-NEXT:    andb %cl, %al
+; CHECK-NEXT:    andb %dl, %al
 ; CHECK-NEXT:    retq
   %res1 = call i8 @llvm.x86.avx512.mask.cmp.ss(<4 x float> %x0, <4 x float> %x1, i32 2, i8 -1, i32 4)
   %res2 = call i8 @llvm.x86.avx512.mask.cmp.ss(<4 x float> %x0, <4 x float> %x1, i32 3, i8 -1, i32 8)
   %res3 = call i8 @llvm.x86.avx512.mask.cmp.ss(<4 x float> %x0, <4 x float> %x1, i32 4, i8 %x3, i32 4)
   %res4 = call i8 @llvm.x86.avx512.mask.cmp.ss(<4 x float> %x0, <4 x float> %x1, i32 5, i8 %x3, i32 8)
 
   %res11 = and i8 %res1, %res2
   %res12 = and i8 %res3, %res4