This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: patch to tree-profiling branch to do escape type analysis.





> > Tested on darwin.  There are 4 new regressions with this patch.
> All of them
> > are instances of a known problem with the vectorizer that this
> patch happens
> > to tickle.

not exactly

> In a few of the cases, it hits the "PRE creates induction variable" bug
> that i have a patch waiting for 4.1 for

this indeed looks like what happens in vect-[56,60].c.
I guess these can be xfailed until the PRE PR is resolved.

> In the others, it hits the invariant phi blocks vectorization problem
> Dorit says she has a patch for.

the other two are vect-[77,78].c, and it's not the invariant phi problem
that they're hitting as far as I can see, but a different restriction in
the vectorizer:

           /* Supported plus/minus expressions are of the form
             {address_base + offset}, such that address_base is of type
             POINTER/ARRAY, and offset is either an INTEGER_CST of type
POINTER,
             or it's not of type POINTER/ARRAY.
             TODO: swap operands if {offset + address_base}.  */

(this is because in tree-profiling-branch we now get
 '(aint *) (off_17 * 4) + ib_19'
whereas we used to get
 'ib_16 + (aint *) (off_11 * 4)').

It's trivial to add, but probably should wait for 4.1. Will open a
missed-optimization PR for that. Can xfail these in the meantime.


dorit



gcc-patches-owner@gcc.gnu.org wrote on 18/01/2005 22:45:20:

>
>
> On Tue, 18 Jan 2005, Kenneth Zadeck wrote:
>
> > Tested on darwin.  There are 4 new regressions with this patch.
> All of them
> > are instances of a known problem with the vectorizer that this
> patch happens
> > to tickle.
>
> Just to followup, the way we lower the new MEM_REF operation causes PRE
to
> eliminate some redundancies in these tests it didn't before.
>
> In a few of the cases, it hits the "PRE creates induction variable" bug
> that i have a patch waiting for 4.1 for
> In the others, it hits the invariant phi blocks vectorization problem
> Dorit says she has a patch for.
>
> I'm also finishing the documentation stuff for the new MEM_REF operation
> that this patch includes, i'll commit it when it's done.
>
> Then i'll start pushing MEM_REF through till the optimizations that can
> benefit from it, lower it, then rerun the redundancy elimination
optimizers.
>


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]