[x86] x >/>=/</<= 0 produce very different assembly

[scottmcm]

Compiling a bunch of icmps against zero

define noundef zeroext i1 @check_slt(i8 %0) unnamed_addr #0 {
start:
  %r = icmp slt i8 %0, 0
  ret i1 %r
}
define noundef zeroext i1 @check_sle(i8 %0) unnamed_addr #0 {
start:
  %r = icmp sle i8 %0, 0
  ret i1 %r
}
define noundef zeroext i1 @check_sgt(i8 %0) unnamed_addr #0 {
start:
  %r = icmp sgt i8 %0, 0
  ret i1 %r
}
define noundef zeroext i1 @check_sge(i8 %0) unnamed_addr #0 {
start:
  %r = icmp sge i8 %0, 0
  ret i1 %r
}

Gives two rather different kinds of checks https://llvm.godbolt.org/z/4sh47s1bn

check_slt:                              # @check_slt
        mov     eax, edi
        shr     al, 7
        ret
check_sle:                              # @check_sle
        test    dil, dil
        setle   al
        ret
check_sgt:                              # @check_sgt
        test    dil, dil
        setg    al
        ret
check_sge:                              # @check_sge
        mov     eax, edi
        not     al
        shr     al, 7
        ret

My instinct is that check_sge should just be test+setge, more like how check_sgt is, rather than needing a shift-by-immediate.

But I'm also really not familiar with the nuances of codegen. I suppose it's also possible that flags are so bad that not using them is better. If that's the case, though, should the others be changed too? Maybe check_sgt could be done with neg and shifting, for example.

(Or perhaps how it is currently is is already best, and I should just go read more architecture manuals to learn what matters.)

---

Inspired by a conversation about the Rust's Ordering::is_lt & friends, and the different ways those could be written https://rust.godbolt.org/z/fq96Yezq7.

[llvmbot]

@llvm/issue-subscribers-backend-x86

[RKSimon]

Interestingly this appears to only be an issue for i8

[scottmcm]

Hmm, for i16, it looks like slt 0 becomes shr 15, but the others are setCC (https://llvm.godbolt.org/z/ThfGbj39v). Maybe that's fine, though? (My μarchitecture knowledge is not nearly good enough to know for sure.)