https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61001
Emmanuel Blot <eblot.ml at gmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|WAITING |RESOLVED
Resolution|--- |INVALID
--- Comment #4 from Emmanuel Blot <eblot.ml at gmail dot com> ---
(In reply to Hans-Peter Nilsson from comment #3)
> (In reply to Marc Glisse from comment #2)
> > -fno-builtin-memcmp is not a hack, it is necessary, even without -flto.
> > Without it, linking is not the only thing that may fail. gcc may look at
> > your implementation, recognize that it is doing the same thing as memcmp,
> > and replace it with a call to memcmp...
>
> True in part. Each code implementing the standard NAME function should
> themselves be compiled with the corresponding -fno-builtin-NAME. But other
> than that, there should be no restrictions or required options. There are
> likely bugs hiding (like, -flto combined with compiling the implementation
> of such a function) so PLEASE:
>
> Provide a test-case or you'll at most receive cryptic comments like this one.
Sorry about the delay, I've been out of the office for 2 weeks.
I spent some time reproducing the issue with a small code base (2 source files
+ 1 linker file), then I realised I was misusing the -fno-builtin-NAME option
switch.
With -fno-builtin-memcpy -fno-builtin-memset properly set, both the sample test
case and the original code work as expected.
Thanks for the support, and very sorry for the noise, I did not know about the
-fno-builtin-NAME flag.
