[llvm] Combine (X ^ Y) and (X == Y) where appropriate (PR #130922)

[Craig Topper via llvm-commits]

topperc wrote:

> I addressed the above comments, but I came across one issue in my tests:
> 
> When working with riscv64 with i32 an i32 input (a0 is i32), the following code generation occurs:
> 
> ```
> entry:
>   %cmp.not = icmp eq i32 %x, 2048
>   br i1 %cmp.not, label %if.end, label %if.then
> if.then:
>   %xor = xor i32 %x, 2048
>   ret i32 %xor
> if.end:
>     tail call void @abort() #2
>   unreachable
> 
> ----->
> 
> # %bb.0:
>    li a1, 1
>    slli a1, a1, 11
>    xor a0, a0, a1
>    sext.w a1, a0
>    beqz a1, .LBB0_2
> # %bb.1: 
>         ret
> .LBB0_2: 
> ```
> 
> Because the result of the sign extension is being used for the comparison, it does not matching the XOR folding pattern.
> 
> One though is that this sign extension is not needed because we are comparing against 0, so we can fold away the sign-extension. However, I'm not sure if that is legal or not (i.e. is an LLVM i32 guaranteed to have the upper bits 0'ed when compiling to a 64 bit architecture).

If %x is a function attribute with the `signext` attribute we can assume the i32 is sign extended to the full register size. Likewise the `zeroext` attribute allows us to assume it is zero extended. If neither attribute is present, we can't make any assumptions.



https://github.com/llvm/llvm-project/pull/130922