[llvm] [InstCombine] avoid extra instructions in foldSelectICmpAnd (PR #127398)

[Yingwei Zheng via llvm-commits]

https://github.com/dtcxzyw approved this pull request.

LGTM. Thank you!

Can you pre-commit this test? https://alive2.llvm.org/ce/z/BWBK9_
```
define i32 @src(i32 %x) {
  %cmp = icmp sgt i32 %x, -1
  %ext = zext i1 %cmp to i32
  call void @use(i32 %ext)
  %and = and i32 %x, 2147483647
  %masksel = select i1 %cmp, i32 -2147483648, i32 0
  %res = or disjoint i32 %masksel, %and
  ret i32 %res
}

declare void @use(i32)
```
A possible solution: https://godbolt.org/z/TfdMKKb4E
convert
```
%and = and i32 %x, 2147483647
%masksel = select i1 %cmp, i32 -2147483648, i32 0
%res = or disjoint i32 %masksel, %and
```
into
```
  %and = and i32 %x, 2147483647 ; bset
  %or = or i32 %x, -2147483648 ; bclr
  %masksel = select i1 %.not9, i32 %or, i32 %and ; cmov
```


https://github.com/llvm/llvm-project/pull/127398