This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
RE: [PATCH] check alias sets in add_inter_loop_mem_dep (ddg.c)
Yes, I am aware of this issue. In the true_dependence (alias.c), there are many steps try to determine two memory references that are not aliased. While checking base + offset part (nonoverlapping_memrefs_p) is not applicable to cross-iteration memory dependency analysis, checking alias sets is still fine. The alias set attribute results from restrict keyword (maybe also from incompatiable data types with -fno-strict-aliasing) and guarantees there is no depencency between two memory references of different alias sets regardless intra- or inter- iterations.
Bingfeng
> -----Original Message-----
> From: Revital1 Eres [mailto:ERES@il.ibm.com]
> Sent: 19 January 2009 09:59
> To: Bingfeng Mei
> Cc: Andrey Belevantsev; Ayal Zaks; gcc-patches@gcc.gnu.org
> Subject: RE: [PATCH] check alias sets in
> add_inter_loop_mem_dep (ddg.c)
>
> Hello,
>
> > I asked on the GCC mailing list about the DDG export patch
> a few months
> ago. I
> > tried to merged it recently into the mainline, without much
> success. A
> > immediate issue I have with modulo scheduling is that it
> doesn't handle
> > restrict keyword. I just adapt these the walker functions
> in the above
> patch,
> > which I think is generic and simple.
>
> OK, thanks.
> I am a little confused regarding how alias_sets_conflict_p will handle
> cross-iteration dependencies when constructing DDG for SMS.
>
> From previous talks with Andrey I understood that data dependency
> graph construction for SMS can not use rtl alias analysis
> because of the
> following explanation:
> When adding cross-iteration dependencies, because rtl AA includes
> base+offset disambiguation, which allows to disambiguate a[i]
> and a[i+1]
> on targets with base+offset memory addressing; while a[i] and a[i+1]
> are independent within one iteration, cross-iteration dependency still
> must be added; thus, DDG construction adds dependency edges for all
> pairs of memory references.
>
> Revital
>
> >
> > Cheers,
> > Bingfeng
> >
> > > -----Original Message-----
> > > From: Revital1 Eres [mailto:ERES@il.ibm.com]
> > > Sent: 19 January 2009 06:27
> > > To: Bingfeng Mei
> > > Cc: Andrey Belevantsev; Ayal Zaks; gcc-patches@gcc.gnu.org
> > > Subject: Re: [PATCH] check alias sets in
> > > add_inter_loop_mem_dep (ddg.c)
> > >
> > > Hello,
> > >
> > > It seems that a similar function can be found in the export
> > > DDG patch: http://gcc.gnu.org/ml/gcc/2007-08/msg00342.html
> > > (please see independent_by_exported_info () in
> > > http://gcc.gnu.org/ml/gcc/2007-08/txt00017.txt)
> > > However, alias_sets_conflict_p is not used there.
> > >
> > > Andrey, I appreciate it if you could explain the
> difference between
> > > those two functions.
> > >
> > > Thanks,
> > > Revital
> > >
> > > > I have a following patch to detect if memory references of two
> > > instructions
> > > > belong to non-conflicting alias sets. It is used in
> > > inter-loop dependence
> > >
> > > > analysis (ddg.c) by modulo scheduler. For the following
> > > testcase, GCC
> > > will
> > > > produce better code due to better memory dependency
> info. In fact,
> > > powperpc-
> > > > gcc cannot software pipeline this loop without the patch. OK for
> > > mainline?
> > > > Should I also include this test in the patch?
> > > >
> > > > /* {dg-do run} */
> > > > /* {dg-options "-O2 -fmodulo-sched -w"} */
> > > > void foo( int * __restrict__ a, int * __restrict__ b, int *
> > > __restrict__
> > > c)
> > > > {
> > > > int i;
> > > > for(i = 0; i < 100; i+=4)
> > > > {
> > > > a[i] = b[i] * c[i];
> > > > a[i+1] = b[i+1] * c[i+1];
> > > > a[i+2] = b[i+2] * c[i+2];
> > > > a[i+3] = b[i+3] * c[i+3];
> > > > }
> > > > }
> > > >
> > > >
> > > > int a[100], b[100], c[100];
> > > > int main()
> > > > {
> > > > int i, res;
> > > > for(i = 0; i < 100; i++)
> > > > {
> > > > b[i] = c[i] = i;
> > > > }
> > > > foo(a, b, c);
> > > >
> > > > res = 0;
> > > > for(i = 0; i < 100; i++)
> > > > {
> > > > res += a[i];
> > > > }
> > > > if(res != 328350)
> > > > abort();
> > > >
> > > > return 0;
> > > > }
> > > >
> > > > This is the first time I try to submit a patch. Hope it
> > > conforms to the
> > > coding
> > > > style, patch format, etc. Additionally, my company
> > > (Broadcom) has already
> > >
> > > > signed coporate copyright form with FSF.
> > > >
> > > > Bingfeng Mei
> > > > Broadcom UK
> > > >
> > > >
> > > > 2009-01-15 Bingfeng Mei <bmei@broadcom.com>
> > > >
> > > > *alias.c (walk_mems_1, walk_mems_2,
> > > insn_alias_sets_conflict_p): Check
> > > > whether two instructions have memory references that
> belong to
> > > conflicting
> > > > alias sets.
> > > > walk_mems_1 and walk_mems_2 are helper functions for
> traversing
> > > > *alias.h (insn_alias_sets_confilict_p): new prototypes
> > > > *ddg.c (add_inter_loop_mem_dep): call
> > > insn_alias_sets_conflict_p not
> > > to
> > > > draw dependency edge
> > > > for instructions with non-conflicting alias sets.
> > > >
> > > > Index: gcc/ddg.c
> > > >
> ===================================================================
> > > > --- gcc/ddg.c (revision 143391)
> > > > +++ gcc/ddg.c (working copy)
> > > > @@ -345,6 +345,10 @@
> > > > static void
> > > > add_inter_loop_mem_dep (ddg_ptr g, ddg_node_ptr from,
> > > ddg_node_ptr to)
> > > > {
> > > > + if (!insn_alias_sets_conflict_p (from->insn, to->insn))
> > > > + /* Do not create edge if memory references have
> > > disjoint alias sets.
> > > */
> > > > + return;
> > > > +
> > > > if (mem_write_insn_p (from->insn))
> > > > {
> > > > if (mem_read_insn_p (to->insn))
> > > > Index: gcc/alias.c
> > > >
> ===================================================================
> > > > --- gcc/alias.c (revision 143391)
> > > > +++ gcc/alias.c (working copy)
> > > > @@ -344,6 +344,43 @@
> > > > return 0;
> > > > }
> > > >
> > > > +static int
> > > > +walk_mems_2 (rtx *x, rtx mem)
> > > > +{
> > > > + if (MEM_P (*x))
> > > > + {
> > > > + if (alias_sets_conflict_p (MEM_ALIAS_SET(*x),
> > > MEM_ALIAS_SET(mem)))
> > > > + return 1;
> > > > +
> > > > + return -1;
> > > > + }
> > > > + return 0;
> > > > +}
> > > > +
> > > > +static int
> > > > +walk_mems_1 (rtx *x, rtx *pat)
> > > > +{
> > > > + if (MEM_P (*x))
> > > > + {
> > > > + /* Visit all MEMs in *PAT and check indepedence. */
> > > > + if (for_each_rtx (pat, (rtx_function) walk_mems_2, *x))
> > > > + /* Indicate that dependence was determined and
> > > stop traversal.
> > > */
> > > > + return 1;
> > > > +
> > > > + return -1;
> > > > + }
> > > > + return 0;
> > > > +}
> > > > +
> > > > +/* Return 1 if two specified instructions have mem expr
> > > with conflict
> > > alias sets*/
> > > > +bool
> > > > +insn_alias_sets_conflict_p (rtx insn1, rtx insn2)
> > > > +{
> > > > + /* For each pair of MEMs in INSN1 and INSN2 check their
> > > independence.
> > > */
> > > > + return for_each_rtx (&PATTERN (insn1), (rtx_function)
> > > walk_mems_1,
> > > > + &PATTERN (insn2));
> > > > +}
> > > > +
> > > > /* Return 1 if the two specified alias sets will always
> > > conflict. */
> > > >
> > > > int
> > > > Index: gcc/alias.h
> > > >
> ===================================================================
> > > > --- gcc/alias.h (revision 143391)
> > > > +++ gcc/alias.h (working copy)
> > > > @@ -41,6 +41,7 @@
> > > > extern int alias_sets_must_conflict_p (alias_set_type,
> > > alias_set_type);
> > > > extern int objects_must_conflict_p (tree, tree);
> > > > extern int nonoverlapping_memrefs_p (const_rtx, const_rtx);
> > > > +extern bool insn_alias_sets_conflict_p (rtx, rtx);
> > > >
> > > > /* This alias set can be used to force a memory to
> > > conflict with all
> > > > other memories, creating a barrier across which no
> > > memory reference
> > >
> > >
> > >
>
>
>