[X86] Fix a lowering issue of mask.compress which has undef float passthrough

[yubing]

Summary

Previously, LegaizeDAG didn't check mask.compress's passthrough might be float, and this lead to getConstant crash since it doesn't support fp

[pengfei]

The title is too long. Move it to description.

[pengfei]

llvm/test/CodeGen/X86/compress-undef-float-passthrough.ll

27–30

unreachable seems unnecessary.

[yubing]

address Phoebe's comments

[RKSimon]

LGTM

llvm/lib/Target/X86/X86ISelLowering.cpp

case COMPRESS_EXPAND_IN_REG: {
       SDValue Mask = Op.getOperand(3);
       SDValue DataToCompress = Op.getOperand(1);
       SDValue PassThru = Op.getOperand(2);
       if (ISD::isBuildVectorAllOnes(Mask.getNode())) // return data as is
         return Op.getOperand(1);
 
       // Avoid false dependency.
       if (PassThru.isUndef())
-        PassThru = DAG.getConstant(0, dl, VT);
+        PassThru = getZeroVector(VT, Subtarget, DAG, dl);
 
       return DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress, PassThru,
                          Mask);
     }
     case FIXUPIMM:
     case FIXUPIMM_MASKZ: {
       SDValue Src1 = Op.getOperand(1);
       SDValue Src2 = Op.getOperand(2);

llvm/test/CodeGen/X86/compress-undef-float-passthrough.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-- -mattr=avx512f,avx512vl | FileCheck %s --check-prefix=CHECK
+
+define void @test_compress_undef_float_passthrough() {
+; CHECK-LABEL: test_compress_undef_float_passthrough:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    movb $5, %al
+; CHECK-NEXT:    kmovw %eax, %k1
+; CHECK-NEXT:    vcompresspd %ymm0, %ymm0 {%k1} {z}
+; CHECK-NEXT:    vscatterqpd %ymm0, (,%ymm0) {%k1}
+; CHECK-NEXT:    vzeroupper
+; CHECK-NEXT:    retq
+entry:                                          ; preds = %loop.50
+  %0 = bitcast i4 undef to <4 x i1>
+  %1 = call <4 x double> @llvm.x86.avx512.mask.compress.v4f64(<4 x double> undef, <4 x double> undef, <4 x i1> <i1 1, i1 0, i1 1, i1 0>)
+  call void @llvm.masked.scatter.v4f64.v4p0f64(<4 x double> %1, <4 x double*> undef, i32 0, <4 x i1> %0)
+  ret void
+}
+
+; Function Attrs: nounwind readnone
+declare <4 x double> @llvm.x86.avx512.mask.compress.v4f64(<4 x double>, <4 x double>, <4 x i1>)
+
+; Function Attrs: nocallback nofree nosync nounwind willreturn writeonly
+declare void @llvm.masked.scatter.v4f64.v4p0f64(<4 x double>, <4 x double*>, i32 immarg, <4 x i1>)
+