[InstCombine] a ? (a < 0) : 1 --> (a <= 0)

[k-arrows]

define i32 @src(i32 %0) {
%1:
  %2 = icmp eq i32 %0, 0
  %3 = lshr i32 %0, 31
  %4 = select i1 %2, i32 1, i32 %3
  ret i32 %4
}
=>
define i32 @tgt(i32 %0) {
%1:
  %2 = icmp slt i32 %0, 1
  %3 = zext i1 %2 to i32
  ret i32 %3
}
Transformation seems to be correct!

https://godbolt.org/z/evaE94d7o
https://alive2.llvm.org/ce/z/6ra3nj

[XChy]

Similar problems associated with the fold zext(a < 0) => lshr a, BW-1 have happened many times. For bit-level operations, the latter would be better. Conversely, for logical operations, the former would be better. I don't know which one is better and how to generalize it. But to solve similar problems with current canonicalization, we must consider a << (BW-1) when we fold operation with other (hidden) icmp.
What can we do if we don't want reduntant similar work? My 2c here is, if when we use pattern match, we can match a whole equivalent class like a < 0, a << BW - 1, then we can solve similar problems completely. That's a theoretical solution from my individual view and it's a systematic work (may change architecture) in need of other's approval and efforts.

[elhewaty]

@dtcxzyw I will work on this, too.