[PATCH] D131947: [X86] Fix a lowering issue of mask.compress which has undef float passthrough Previously, LegaizeDAG didn't check mask.compress's passthrough might be float , and this lead to getConstant crash since it doesn't support fp.

[Bing Yu via Phabricator via llvm-commits]

yubing created this revision.
Herald added subscribers: pengfei, hiraditya.
Herald added a project: All.
yubing requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D131947

Files:
  llvm/lib/Target/X86/X86ISelLowering.cpp
  llvm/test/CodeGen/X86/compress-undef-float-passthrough.ll


Index: llvm/test/CodeGen/X86/compress-undef-float-passthrough.ll
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/X86/compress-undef-float-passthrough.ll
@@ -0,0 +1,37 @@
+; 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 dso_local void @test_compress_undef_float_passthrough() {
+; CHECK-LABEL: test_compress_undef_float_passthrough:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    xorl %eax, %eax
+; CHECK-NEXT:    testb %al, %al
+; CHECK-NEXT:    je .LBB0_1
+; CHECK-NEXT:  # %bb.2: # %ifmerge.80
+; CHECK-NEXT:  .LBB0_1: # %then.80
+; CHECK-NEXT:    movb $5, %al
+; CHECK-NEXT:    kmovw %eax, %k1
+; CHECK-NEXT:    vcompresspd %ymm0, %ymm0 {%k1} {z}
+; CHECK-NEXT:    vscatterqpd %ymm0, (,%ymm0) {%k1}
+entry:
+  br label %loop.50
+
+loop.50:                                          ; preds = %entry
+  br i1 undef, label %ifmerge.80, label %then.80
+
+then.80:                                          ; 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)
+  unreachable
+
+ifmerge.80:                                       ; preds = %loop.50
+  unreachable
+}
+
+; 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>)
+
Index: llvm/lib/Target/X86/X86ISelLowering.cpp
===================================================================
--- llvm/lib/Target/X86/X86ISelLowering.cpp
+++ llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -27038,7 +27038,7 @@
 
       // 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);


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D131947.452911.patch
Type: text/x-patch
Size: 2331 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220816/e6bcda01/attachment.bin>