http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60092
--- Comment #13 from Jakub Jelinek <jakub at gcc dot gnu.org> --- (In reply to Richard Biener from comment #12) > (In reply to Andreas Schwab from comment #11) > > If a function is not allowed to change errno this must be explicitly > > documented. > > That means > > Index: gcc/tree-ssa-alias.c > =================================================================== > --- gcc/tree-ssa-alias.c.orig 2014-02-06 15:43:42.138266256 +0100 > +++ gcc/tree-ssa-alias.c 2014-02-06 15:43:33.046266882 +0100 > @@ -1847,7 +1847,9 @@ call_may_clobber_ref_p_1 (gimple call, a > ao_ref dref; > ao_ref_init_from_ptr_and_size (&dref, ptrptr, > TYPE_SIZE_UNIT (ptr_type_node)); > - return refs_may_alias_p_1 (&dref, ref, false); > + return (refs_may_alias_p_1 (&dref, ref, false) > + || (flag_errno_math > + && targetm.ref_may_alias_errno (ref))); > } > /* Freeing memory kills the pointed-to memory. More importantly > the call has to serve as a barrier for moving loads and stores > > is necessary. For posix_memalign? I think errno can contain any value, except that library is not allowed to clear errno. So, IMHO *p = 1; posix_memalign (...); return *p; can be still optimized into return 1;, because if p = &errno; then *p after the call has undefined value (just known not to be zero).