http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60395
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |lto, wrong-code
Status|UNCONFIRMED |NEW
Last reconfirmed| |2014-03-03
CC| |hubicka at gcc dot gnu.org,
| |rguenth at gcc dot gnu.org
Ever confirmed|0 |1
Known to fail| |4.7.2, 4.8.2, 4.9.0
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed on x86_64 with using -mstringop-strategy=libcall. Shorter
single-file testcase:
void *memset(void *dest, int ch, __SIZE_TYPE__ count)
{
char *p = (char *)dest;
while (count--)
*p++ = (char)ch;
return dest;
}
int main ()
{
volatile int dummy[10] = {0};
return dummy[0];
}
> gcc t.c -Os -mstringop-strategy=libcall -flto -fno-builtin
`memset' referenced in section `.text.startup' of
/tmp/ccdDUr5m.ltrans0.ltrans.o: defined in discarded section `.text' of
/tmp/ccApnJzK.o (symbol from plugin)
collect2: error: ld returned 1 exit status
1
t.o 2
177 b0e6a56d4c94b3b5 PREVAILING_DEF_IRONLY memset
180 b0e6a56d4c94b3b5 PREVAILING_DEF main
and then IPA kills the memset definition because it's unused at IPA time
and it isn't a builtin. Then expansion re-introduces a call to memset
an we use the linker resolution and try to bind it locally which obviously
fails (even though a fallback is available in libc).
IMHO even with -fno-builtin IPA needs to preserve certain function bodies
we may introduce calls to late (memset, memcpy, memmove are the obvious
ones), much like we have special_function_p which matches on the function
name and not only the builtin property. Or we need to simply _not_ mark
those definitons as non-builtin as we introduce calls to them late. Or
we need to fixup things at the linker side (we introduce a call to the
_builtin_ memset but we discarded a non-builtin memset).
The special_function_p-like hack is probably easiest but not the robustest
solution. Not sure if we can adjust the linkers idea of where a def resides
at all.
Honza?
