I ran into a bug with the PPC side of things with both 4.1.1 [came with
FC5] and a freshly built 4.3.2, the snippet is this block
#if 0
#define SQRADDSC(i, j) \
asm( \
" mullw %0,%6,%7 \n\t" \
" mulhwu %1,%6,%7 \n\t" \
" xor %2,%2,%2 \n\t" \
:"=r"(sc0), "=r"(sc1), "=r"(sc2):"0"(sc0), "1"(sc1), "2"(sc2),
"r"(i),"r"(j) : "%cc");
#else
#define SQRADDSC(i, j) \
asm( \
" mullw %0,%3,%4 \n\t" \
" mulhwu %1,%3,%4 \n\t" \
" xor %2,%2,%2 \n\t" \
:"=r"(sc0), "=r"(sc1), "=r"(sc2): "r"(i),"r"(j) : "%cc");
#endif
Where {sc0,sc1,sc2} are not previously initialized. The first block
(that is disabled) is my old code. It would correctly warn about
uninitialized variables. The 2nd block was my fix which to me seems
correct. However, the compiler generates output like this
# 0 "" 2
# 38 "fp_sqr_comba_6.i" 1
mullw 25,25,18
mulhwu 9,25,18
xor 0,0,0
Where r25 is both an input and output it seems. Clearly r25 is not
supposed to be overwritten here.
I put together a tarball of the source, cmdline, uname, gcc version, etc
up at
http://home.libtom.org/gcc/gccbug.tgz
The tarball has all that you need to reproduce the bug with both 4.1.1
and 4.3.2.
Has this been seen before, and/or should I file a gcc bug about it? I
guess more importantly, is my PPC "fix" correct?
Thanks,
Tom
