since Clang 17 `-ffast-math` prevents tail call optimization for math.h functions

[namniav]

https://godbolt.org/z/Ea5ccfebr

#include <math.h>

float foo(float x) {
    return expf(x);
}

Clang 17.0.1 -O3 -ffast-math:

foo:
        push    rax
        call    expf@PLT
        pop     rax
        ret

Clang 17.0.1 -O3:

foo:
        jmp     expf@PLT

Clang 16.0.0 -O3 -ffast-math:

foo:
        jmp     expf@PLT

GCC 14.2 -O3 -ffast-math:

foo:
        jmp     expf

[namniav]

They're not redundant for call due to alignment requirement, but a single jmp should be sufficient.

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: None (namniav)

https://godbolt.org/z/Ea5ccfebr

#include <math.h>

float foo(float x) {
    return expf(x);
}

Clang 17.0.1 -O3 -ffast-math:

foo:
        push    rax
        call    expf@<!-- -->PLT
        pop     rax
        ret

Clang 17.0.1 -O3:

foo:
        jmp     expf@<!-- -->PLT

Clang 16.0.0 -O3 -ffast-math:

foo:
        jmp     expf@<!-- -->PLT

GCC 14.2 -O3 -ffast-math:

foo:
        jmp     expf

[dtcxzyw]

With -ffast-math we convert expf into llvm.exp.f32. I don't know why tail call optimization doesn't work here...