Missed optimization: redundant conversion to bool when the value is already in valid bool range

[YanWQ-monad]

For the IR

define i1 @src(ptr %x) {
  %1 = load i8, ptr %x, align 1
  %2 = icmp ult i8 %1, 2
  br i1 %2, label %bb1, label %bb2

bb2:
  %3 = tail call i1 @dynamic()
  br label %bb3

bb1:
  %4 = icmp eq i8 %1, 1
  br label %bb3

bb3:
  %5 = phi i1 [ %4, %bb1 ], [ %3, %bb2 ]
  ret i1 %5
}

declare i1 @dynamic()

which generates the following instructions in x86 (https://godbolt.org/z/hY5EjobTv):

src:                                    # @src
        movzx   eax, byte ptr [rdi]
        cmp     al, 2
        jae     dynamic@PLT                     # TAILCALL
        cmp     al, 1
        sete    al
        ret

The conversion of eax to bool is redundant after jae, we can directly ret (eax) as eax is already in valid bool range [0, 1].

---

This issue is extracted from rust-lang/rust#121673.

I tried to implement this optimization in IR by folding icmp eq %x, 1 to trunc %x to i1 in #83829, but it doesn't work well, since InstCombine would canonicalize trunc to (X & 1) != 0, which results in one more and instruction in backend.

So, I am going to implement this optimization in backend. Please let me know if there is a better way to address this issue.
If possible, please assign this issue to me, thanks.

[andjo403]

InstCombine would canonicalize trunc to (X & 1) != 0

the canonicalizion of trunc is mentioned in the https://discourse.llvm.org/t/rfc-add-nowrap-flags-to-trunc/77453/3 maybe is interesting to you.