http://gcc.gnu.org/bugzilla/show_bug.cgi?id=61054
Bug ID: 61054
Summary: Clearly note effect of memcpy optimization
improvements in 4.9 changes
Product: gcc
Version: 4.9.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: web
Assignee: unassigned at gcc dot gnu.org
Reporter: harald at gigawatt dot nl
Consider this reduced program:
void *volatile null;
volatile long zerolen;
int main() {
void *dst = null, *src = null;
__builtin_memcpy (dst, src, zerolen);
if (dst) __builtin_abort ();
}
This program is invalid, but until 4.8 it would compile and run without an
issue. As of 4.9, it aborts.
I'm not asking for GCC to change (the programs should be fixed instead), but it
is not clear from the GCC 4.9 Changes page that such code no longer works. This
affects real code using custom allocators and realloc-like functions, where
some of those realloc-like functions don't bother to check for null pointers
and assume zero-length memcpy is a no-op. When the old memory is then freed,
and their free-like function performs a null pointer check, that check is
optimised away if the function gets inlined.
Could the GCC 4.9 Changes page get a mention of this?
I do see "Better inlining of memcpy and memset that is aware of value ranges
and produces shorter alignment prologues." in the IA-32/x86-64 section, but I
cannot tell if it is related to that, or if it is actually arch-independent.
