Certain samples reliably cause oggenc to throw a SIGFPE on an Alpha.  (Only
tested with linux; haven't tried tru64 yet.)  core and sample to be attached.

#0  0x200002119f4 in fit_line (acc=0x11fffdd30, fits=2, y0=0x11ffff24c, 
    y1=0x11ffff250) at ../../lib/floor1.c:511
#1  0x2000021267c in floor1_fit (vb=0x11ffff6a0, look=0x12007ee50, 
    logmdct=0x12007be70, logmask=0x12007bc70) at ../../lib/floor1.c:676
#2  0x20000219ffc in mapping0_forward (vb=0x11ffff6a0)
    at ../../lib/mapping0.c:423
#3  0x20000205054 in vorbis_analysis (vb=0x11ffff6a0, op=0x0)
    at ../../lib/analysis.c:47
#4  0x1200097c8 in oe_encode (opt=0x11ffff9e8) at ../../oggenc/encode.c:283
#5  0x12000316c in main (argc=3, argv=0x11ffffb88) at ../../oggenc/oggenc.c:344

With this sample, starting the encoding anywhere between samples 0-8668 causes
the FPE; if I trim the wav file to start at sample 8669, it encodes fine.  If I
use -q4, it encodes fine.