gcc 3.2 - float-related errors with -O -march=athlon

[Jakub Bogusz]

On Thu, Sep 05, 2002 at 11:54:07PM +0200, Jakub Bogusz wrote:
> I get errors when trying to compile with -march=athlon and at least -O
> (-O1, -O2, -O3).
> Errors stop appearing after changing optimization to -O0, -Os or
> -march=i686.
> 
> $ gcc -v
> Reading specs from /usr/lib/gcc-lib/athlon-pld-linux/3.2/specs
> Configured with: ../configure --prefix=/usr --infodir=/usr/share/info --mandir=/usr/share/man --enable-shared --enable-threads=posix --enable-__cxa_atexit --enable-languages=c,c++,f77,gcov,java,objc,ksi,ada --enable-c99 --enable-long-long --enable-multilib --enable-nls --with-gnu-as --with-gnu-ld --with-system-zlib --with-slibdir=/lib --without-x athlon-pld-linux
> Thread model: posix
> gcc version 3.2
> 
> 
> This testcase is taken from lbreakout2, but file is cut down to less
> than 1kB (and it still triggers bug).
> It isn't rare case - I got very similar errors when trying to compile
> XFree86 4.2.1 with "-O2 -march=athlon" optimization.
> I'm using lbreakout2 code as example because it's simpler.
> 
> Example command line is:
> 
> $ gcc -c misc-b.i -O -march=athlon

I've done more tests, maybe this would help.
Errors occur with (shellglob-like notation):
-O{2,3} -march=athlon{,-tbird,-4,-xp,-mp}
-O{2,3} -march=pentium{2,3,4}
-O{2,3} -march=pentiumpro -mmx
-O{2,3} -march=k6{,-2,-3}

And -march=pentium-mmx does _not_ trigger bug.

> Then errors appear:
> 
> misc-b.c: In function `circle_intersect':
> misc-b.c:58: unable to find a register to spill in class `FLOAT_REGS'
> misc-b.c:58: this is the insn:
> (insn 23 21 24 (set (subreg:SF (reg/v:DI 29 rmm0 [65]) 0)
>         (minus:SF (subreg:SF (reg/v:DI 61) 0)
>             (subreg:SF (reg/v:DI 1 rdx [59]) 0))) 533 {*fop_sf_1_nosse} (insn_list 8 (insn_list 4 (nil)))
>     (nil))
> misc-b.c:58: confused by earlier errors, bailing out

And using "-O -march=athlon-xp -mfpmath=sse" changes error message to:

misc-b.c: In function `circle_intersect':
misc-b.c:58: unable to find a register to spill in class `SSE_REGS'
misc-b.c:58: this is the insn:
(insn 23 21 24 (set (subreg:SF (reg/v:DI 29 rmm0 [65]) 0)
        (minus:SF (subreg:SF (reg/v:DI 61) 0)
            (subreg:SF (reg/v:DI 1 rdx [59]) 0))) 535 {*fop_sf_1_sse} (insn_list 8 (insn_list 4 (nil)))
    (nil))
misc-b.c:58: confused by earlier errors, bailing out

> char circle_msg[256];
> 
> typedef struct {
>     float x, y;
> } Vector;
> # 36 "misc-b.c"
> int circle_intersect( Vector m, int r, Vector pos, Vector v, Vector *t1, Vector *t2 )
> {
>     Vector delta = { pos.x - m.x, pos.y - m.y };
>     float delta_v = delta.x * v.x + delta.y * v.y;
>     float dis = delta_v * delta_v + r * r - ( delta.x * delta.x + delta.y * delta.y );
>     float t;
> 
>     if ( dis < 0 ) {
> 
> 
> 
>                 return 0;
>         }
> 
>     t = -delta_v + dis;
>     t1->x = pos.x + t * v.x; t1->y = pos.y + t * v.y;
>     t = -delta_v - dis;
>     t2->x = pos.x + t * v.x; t2->y = pos.y + t * v.y;
> 
> 
> 
>     return 1;
> }

Shorter version that still triggers bug:

typedef struct {
    float x, y;
} Vector;
float circle_intersect( Vector m, Vector pos)
{
    Vector delta = { pos.x - m.x, pos.y - m.y};
    return delta.x+delta.y;
}

The key seems to be that struct is initialized by "= { , }" construct,
and then both fields are used in some calculations.

After changing to "Vector delta; delta.x=pos.x-m.x; delta.y=pos.y-m.y"
errors doesn't occur.


-- 
Jakub Bogusz    http://prioris.mini.pw.edu.pl/~qboosh/
PLD Linux       http://www.pld.org.pl/