-mno-80387 / -mno-fp-ret-in-387 question

[Uros Bizjak]

Hello!

I would just like to ask if -mno-fp-ret-in-387 (and now also -mno-80387) 
violates ABI in some way (just like -mregparm=X). Consider this code:

#include <math.h>

double test(double x) {
    double b;
   
    b = sin(x);
    return b + 1.0;
}

compiled with gcc -O2 -march=pentium4 -mfpmath=sse -mno-80387:

test:
        pushl   %ebp
        movl    %esp, %ebp
        subl    $24, %esp
        movsd   8(%ebp), %xmm0
        movsd   %xmm0, (%esp)
        call    sin
        movl    %eax, -8(%ebp)
        movl    %edx, -4(%ebp)
        movsd   -8(%ebp), %xmm0
        addsd   .LC1, %xmm0
        movsd   %xmm0, -8(%ebp)
        movl    -8(%ebp), %eax
        movl    -4(%ebp), %edx
        leave
        ret

In this code, it is assumed that sin() function returns its result in 
%eax/%edx pair. However, a "normal" libm sin() function returns its 
value in top FP register, so there we have an ABI mismatch. When normal 
libm is linked with this code, wrong results will be produced.

In -mregparm case, there is a big warning in the documentation:

     *Warning:* if you use this switch, and NUM is nonzero, then you
     must build all modules with the same value, including any
     libraries.  This includes the system libraries and startup modules.

I think that something similar should be added to -mno-fp-ret-in-387 / 
-m-no-80387.

Uros.