The old insns_may_alias_p is based checking alias set number. But since 4.5, the new alias oracle doesn't rely the alias set number. may_alias_p() is a new function based on alias oracle. Essentially, it is same as true_dependence function but excluding offset and TBAA based disambiguation, which is not valid for cross-iteration memory dependence analysis.
Cheers, Bingfeng > -----Original Message----- > From: gcc-ow...@gcc.gnu.org [mailto:gcc-ow...@gcc.gnu.org] On Behalf Of > Revital1 Eres > Sent: 15 September 2010 13:28 > To: gcc@gcc.gnu.org > Subject: Question about alias check in ddg.c > > > Hello, > > When trying to compile the following loop with GCC -r164298 with modulo > scheduling pass enabled on PowerPC I get that the inter loop edges > between the memory instructions are created in the DDG although the > following check in ddg.c exists: > > static void > add_inter_loop_mem_dep (ddg_ptr g, ddg_node_ptr from, ddg_node_ptr to) > { > if (!insns_may_alias_p (from->insn, to->insn)) > return; > > With GCC -r146278 the edges are not created in the DDG. > Looking more closely into the code I see the following differences in > ddg.c > between the GCC versions: > > The new version of GCC contains: > > add_inter_loop_mem_dep () -> ...-> may_alias_p() > > while with GCC -r146278: > > add_inter_loop_mem_dep () -> ... -> insn_alias_sets_conflict_p() > > I do not understand this behavior and I appreciate an explanation. > > Thanks, > Revital > > > void foo( int * __restrict__ dst, int * __restrict__ src1, > int * __restrict__ src2) > { > int x, y; > > for( x = 0; x < 100; x+=1 ) > { > dst[x] = ( src1[x] * src2[x] ) ; > } > } >
