[cfe-dev] __builtin_constant_p and floating-point constants

[Eli Friedman]

On Mon, Jun 20, 2011 at 1:34 PM, Jonathan Sauer <jonathan.sauer at gmx.de> wrote:
> Hello,
>
> please consider the following (not very useful) program:
>
> inline int foo(float a, int b)
> {
>        if (__builtin_constant_p(a))
>                return int(a) * b;
>        else
>                return a * b;
> }
>
>
> int bar(int b)
> {
>        return foo(23.0f, b);
> }
>
>
> int main(int, char**)
> {
> }
>
>
> If <a> is a constant when <foo> is called, only its integer part should be multiplied with <b>, resulting
> in an integer multiplication. If <a> is not constant, <b> should be converted to a float, multiplied with
> <a> and truncated back to an <int> instead.
>
> Apple's GCC 4.2.1 handles this as expected ("gcc RandomTest.cpp -S -O3 -o -", small cleanups of the output):
>
>        pushq   %rbp
>        movq    %rsp, %rbp
>        imull   $23, %edi, %eax
>        leave
>        ret
>
>
> Clang r132676 on the other hand ("clang RandomTest.cpp -S -O3 -o -", again slightly cleaned output):
>
>        pushq   %rbp
>        movq    %rsp, %rbp
>        cvtsi2ss        %edi, %xmm0
>        mulss   LCPI0_0(%rip), %xmm0
>        cvttss2si       %xmm0, %eax
>        popq    %rbp
>        ret
>
>
> It seems that clang does not evaluate __builtin_constant_p in the same way as GCC does. Is this a bug
> or intentional? (personally, I would think the latter, as there does not seem to exist a reason why
> __builtin_constant_p should not work with floating-point constants)

http://llvm.org/bugs/show_bug.cgi?id=4898

Essentially, the issue here is that we fold __builtin_constant_p
before the inliner runs.

-Eli