optimization/2415: ICE (regression 2.95.2->gcc3.0) when compiling fdlibm

[Jan Vroonhof <jan dot vroonhof at insignia dot com>]

>Number:         2415
>Category:       optimization
>Synopsis:       ICE (regression 2.95.2->gcc3.0) when compiling pow from fdlibm
>Confidential:   no
>Severity:       critical
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          ice-on-legal-code
>Submitter-Id:   net
>Arrival-Date:   Wed Mar 28 08:46:01 PST 2001
>Closed-Date:
>Last-Modified:
>Originator:     Jan Vroonhof <jan.vroonhof@insignia.com>
>Release:        3.0 20010328 (prerelease)
>Organization:
Insignia Solutions
>Environment:
System: Linux bart 2.2.17 #2 Tue Mar 13 22:44:15 GMT 2001 i686 unknown
Architecture: i686

	
host: i686-pc-linux-gnu
build: i686-pc-linux-gnu
target: i686-pc-linux-gnu
configured with: ../configure --prefix /local/cross/i386/gcc-3.0
>Description:
	When compiling __ieee754_pow gcc 3.0 (compiled from CVS checkout,
    verified with online test compiler for 20010327 snapshot)
    the compiler ICE's at -O2 (previous versions produced incorrect code).
>How-To-Repeat:
	I have tried stripping down the function as far as possible. The result
    is inserted below. Compile with
    gcc -O2 -o printpow printpow.i
    and you will get
printpow.i: In function `__ieee754_pow':
printpow.i:75: Unable to find a register to spill in class `AREG'.
printpow.i:75: This is the insn:
(insn 209 79 82 (set (subreg:SI (reg/v:DF 15 st(7) [76]) 0)
        (const_int 0 [0x0])) 35 {*movsi_1} (nil)
    (nil))
printpow.i:75: confused by earlier errors, bailing out

The file [Note: This has been stripped down aggressively by deleting
blocks of code so might no longer make sense.]:

double __ieee754_pow(double x,
                     double y)
{
    double z,
        ax,
        z_h,
        z_l,
        p_h,
        p_l;
    double y1,
        t1,
        t2,
        r,
        s,
        t,
        u,
        v,
        w;
    int i,
        j,
        k,
        yisint,
        n;
    int hx,
        hy,
        ix,
        iy;
    unsigned lx,
        ly;

    if (0)
    {
    }
    else
    {
        double s2,
            s_h,
            s_l,
            t_h,
            t_l;

        if (ix < 0x00100000)
        {
            ax *= 0;
            n -= 53;
            ix = (*(1+(int*)&ax));
        }
        n += ((ix) >> 20) - 0x3ff;
        (*(int*)&s_h) = 0;
        r += s_l * (s_h + s);
        s2 = s_h * s_h;
        t_h = 3.0 + s2 + r;
        (*(int*)&t_h) = 0;
        t_l = r - ((t_h - 3.0) - s2);
        u = s_h * t_h;
        v = s_l * t_h + t_l * s;
        p_h = u + v;
        (*(int*)&p_h) = 0;
        p_l = v - (p_h - u);
        z_h = 0;
        z_l = p_h + p_l;
        t = (double) n;
        t1 = (((z_h + z_l)) + t);
        (*(int*)&t1) = 0;
        t2 = z_l - (((t1 - t)) - z_h);
    }

    y1 = y;
    (*(int*)&y1) = 0;
    p_l = (y - y1) * t1 + y * t2;
    p_h = y1 * t1;
    z = p_l + p_h;
    return s * z;
}



>Fix:
	No known fix or workaround
>Release-Note:
>Audit-Trail:
>Unformatted: