Floating point comparison miscompilation

[aykevl]

I have the following C code:

void sideeffect(void);

int identify(float v) {
    if (v+v == v && v > 0) {
        return 2;
    }
    if (v < 0) {
        sideeffect();
        return 5;
    }
    return 4;
}

When I call this with identify(20) (where 20 is the floating point value with bits 0x41a00000, see here) the function will return 5. This is clearly incorrect, as v is 20 which is bigger than 0. Note that sideeffect doesn't do anything, but without some external call the misbehavior appears to be optimized away.
I have seen the same issue with other positive integers, I don't think the actual number matters.

I found the issue originally in TinyGo (runtime.printfloat32) but managed to reduce it and convert it to C while still maintaining the incorrect behavior.

This issue might be a duplicate of #20.

[andreisfr]

Hi @aykevl ! This problem corrected in xtensa_release_11.0.0 branch, please check it

[aykevl]

Thank you! I have tested this and it is indeed fixed with the LLVM 11 branch.