[X86][AVX512] Suboptimal codegen with llvm.masked.store.v192i8

[Andarwinux]

original C code: https://test.godbolt.org/z/x19oc35f5

reduced IR: https://test.godbolt.org/z/Yfcxcrbzv

define void @UpsampleBgrLinePair_C(<192 x i1> %interleaved.mask474) {
entry:
  br label %vector.body

vector.body:                                      ; preds = %vector.body, %entry
  tail call void @llvm.masked.store.v192i8.p0(<192 x i8> zeroinitializer, ptr null, <192 x i1> %interleaved.mask474)
  br label %vector.body
}

; Function Attrs: nocallback nofree nosync nounwind willreturn memory(argmem: write)
declare void @llvm.masked.store.v192i8.p0(<192 x i8>, ptr captures(none), <192 x i1>) #0

attributes #0 = { nocallback nofree nosync nounwind willreturn memory(argmem: write) }

[Andarwinux]

#198166 did indeed improve the reduced IR, but for some reason the original C code still contains suboptimal sequence, maybe because llvm-reduce excessively reduced it?

[Andarwinux]

I tried reduce again and got a slightly different IR that cannot be improved by #198166.

define void @UpsampleBgrLinePair_C(<192 x i1> %interleaved.mask) {
iter.check:
  br label %vector.body

vector.body:                                      ; preds = %vector.body, %iter.check
  %index = phi i64 [ 0, %iter.check ], [ %index.next, %vector.body ]
  %0 = and i64 %index, 1
  %1 = getelementptr i8, ptr null, i64 %0
  tail call void @llvm.masked.store.v192i8.p0(<192 x i8> zeroinitializer, ptr %1, <192 x i1> %interleaved.mask)
  %index.next = add i64 %index, 1
  br label %vector.body
}

; Function Attrs: nocallback nofree nosync nounwind willreturn memory(argmem: write)
declare void @llvm.masked.store.v192i8.p0(<192 x i8>, ptr captures(none), <192 x i1>) #0

attributes #0 = { nocallback nofree nosync nounwind willreturn memory(argmem: write) }

[llvmorg-github-actions]

@llvm/issue-subscribers-backend-x86

Author: Andarwinux

Details

original C code: https://test.godbolt.org/z/x19oc35f5

reduced IR: https://test.godbolt.org/z/Yfcxcrbzv

define void @<!-- -->UpsampleBgrLinePair_C(&lt;192 x i1&gt; %interleaved.mask474) {
entry:
  br label %vector.body

vector.body:                                      ; preds = %vector.body, %entry
  tail call void @<!-- -->llvm.masked.store.v192i8.p0(&lt;192 x i8&gt; zeroinitializer, ptr null, &lt;192 x i1&gt; %interleaved.mask474)
  br label %vector.body
}

; Function Attrs: nocallback nofree nosync nounwind willreturn memory(argmem: write)
declare void @<!-- -->llvm.masked.store.v192i8.p0(&lt;192 x i8&gt;, ptr captures(none), &lt;192 x i1&gt;) #<!-- -->0

attributes #<!-- -->0 = { nocallback nofree nosync nounwind willreturn memory(argmem: write) }