[ConstraintElimination] Missed optimization: fold `max(SExt(b), SExt(a) + 1)` to `SExt(b)` if `a s< b`

[XChy]

Alive2 proof: https://alive2.llvm.org/ce/z/A8dtGp

Motivating example

Similar to #75155

define i64 @src(i32 %a, i32 %b){
entry:
  %cmp = icmp slt i32 %a, %b
  br i1 %cmp, label %then, label %else

then:
  %sa = sext i32 %a to i64
  %sb = sext i32 %b to i64
  %add = add i64 %sa, 1
  %smax = call i64 @llvm.smax.i64(i64 %sb, i64 %add)
  call void @dummy()
  ret i64 %smax

else:
  ret i64 0
}

can be folded to:

define i64 @tgt(i32 %a, i32 %b){
entry:
  %cmp = icmp slt i32 %a, %b
  br i1 %cmp, label %then, label %else

then:
  %sb = sext i32 %b to i64
  call void @dummy()
  ret i64 %sb

else:
  ret i64 0
}

Real-world motivation

This snippet of IR is derived from redis/src/db.c@xreadGetKeys (after O3 pipeline).
The example above is a reduced version. If you're interested in the original suboptimal IR and optimal IR, see also:https://godbolt.org/z/Kz4M6WPxn

Let me know if you can confirm that it's an optimization opportunity, thanks.

[fhahn]

Initial support for simplifying min/max has been added in #75306, not sure exactly what is missing. cc @zjaffal

[XChy]

It seemed that zext-like insts had been decomposed before, but opt on godbolt didn't optimize it away when I opened this issue...
Correct this issue now.

[fhahn]

Thanks for clarifying re unsigned, I was surprised that this wasn’t handled yet