x86 missed optimization for __builtin_clz()^31

[Alcaro]

#include <stdint.h>
uint8_t b(uint32_t in)
{
    uint8_t ret = __builtin_clz(in) ^ 31;
    return ret;
}
uint8_t c(uint32_t in)
{
    uint8_t ret = __builtin_clz(in) ^ 31;
    return ret + 1;
}

Expected: Since b optimizes to a single bsr eax,edi, c should optimize to at most one instruction more.
Actual:

b(unsigned int):
        bsr     eax, edi
        ret

c(unsigned int):
        bsr     ecx, edi
        xor     ecx, 31
        mov     al, 32
        sub     al, cl
        ret

GCC gives good output (probably by being less clever about normalization), as does Clang if I add an extra optimization barrier. https://godbolt.org/z/3ansdnMxa

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: None (Alcaro)

```c #include <stdint.h> uint8_t b(uint32_t in) { uint8_t ret = __builtin_clz(in) ^ 31; return ret; } uint8_t c(uint32_t in) { uint8_t ret = __builtin_clz(in) ^ 31; return ret + 1; } ``` Expected: Since b optimizes to a single `bsr eax,edi`, c should optimize to at most one instruction more. Actual: ``` b(unsigned int): bsr eax, edi ret

c(unsigned int):
bsr ecx, edi
xor ecx, 31
mov al, 32
sub al, cl
ret

GCC gives good output (probably by being less clever about normalization), as does Clang if I add an extra optimization barrier. https://godbolt.org/z/3ansdnMxa
</details>

[RKSimon]

define i8 @lz8(i32 noundef %in) {
  %lz = tail call range(i32 0, 33) i32 @llvm.ctlz.i32(i32 %in, i1 true)
  %t = trunc nuw nsw i32 %lz to i8
  %conv2 = sub nuw nsw i8 32, %t
  ret i8 %conv2
}

Interestingly this succeeds as we promote i16 to i32:

define i16 @lz16(i32 noundef %in) {
  %lz = tail call range(i32 0, 33) i32 @llvm.ctlz.i32(i32 %in, i1 true)
  %t = trunc nuw nsw i32 %lz to i16
  %conv2 = sub nuw nsw i16 32, %t
  ret i16 %conv2
}