On Mon, Sep 19, 2016 at 11:22:27AM +0200, Jan Hubicka wrote:
> > On Mon, Sep 19, 2016 at 2:48 AM, Jan Hubicka wrote:
> > > Hi,
> > > this is the patch compensating testsuite I commited after re-testing
> > > on x86_64-linux.
> > >
> > > Other placements of early_thread_jumps does not work veyr wel
On 19/09/16 08:02, Christophe Lyon wrote:
Index: testsuite/gcc.dg/tree-ssa/pr68198.c
===
--- testsuite/gcc.dg/tree-ssa/pr68198.c (revision 240109)
+++ testsuite/gcc.dg/tree-ssa/pr68198.c (working copy)
@@ -1,5 +1,5 @@
/* { dg-do com
ind_control_statement_thread_paths,
> > find_jump_threads_backwards): Add speed_p parameter.
> > (pass_data_early_thread_jumps): New pass.
> > (make_pass_early_thread_jumps): New function.
> >
> > * g++.dg/predict-loop-exit-1.C: Disa
,
> fsm_find_control_statement_thread_paths,
> find_jump_threads_backwards): Add speed_p parameter.
> (pass_data_early_thread_jumps): New pass.
> (make_pass_early_thread_jumps): New function.
>
> * g++.dg/predict-loop-exit-1.C: Disable early jump th
>
> * g++.dg/predict-loop-exit-1.C: Disable early jump threading.
> * g++.dg/predict-loop-exit-2.C: Disable early jump threading.
> * g++.dg/predict-loop-exit-3.C: Disable early jump threading.
> * gcc.dg/tree-ssa/pr69196-1.c: Disable early jum
): New pass.
(make_pass_early_thread_jumps): New function.
* g++.dg/predict-loop-exit-1.C: Disable early jump threading.
* g++.dg/predict-loop-exit-2.C: Disable early jump threading.
* g++.dg/predict-loop-exit-3.C: Disable early jump threading.
* gcc.dg/tree-ssa
On 08/11/2016 08:02 AM, Jan Hubicka wrote:
Hi,
this patch adds early jump threading pass. Jump threading is one of most
common cases where estimated profile becomes corrupted, because the branches
are predicted independently beforehand. This patch simply adds early mode to
jump threader that
On 08/12/2016 02:02 AM, Richard Biener wrote:
On Thu, 11 Aug 2016, Jan Hubicka wrote:
Hi,
this patch adds early jump threading pass. Jump threading is one of most
common cases where estimated profile becomes corrupted, because the branches
are predicted independently beforehand. This patch
On 08/16/2016 05:02 AM, Jan Hubicka wrote:
I don't think the backwards/FSM threader tries to update the profile
data at all right now.
This seems to be cause of the regression andrew is speaking about. I wrote
updating
of profile after threading few times and there was also Theresa's patch.
I
> I don't think the backwards/FSM threader tries to update the profile
> data at all right now.
This seems to be cause of the regression andrew is speaking about. I wrote
updating
of profile after threading few times and there was also Theresa's patch.
I tought all the threaders use common dupli
On 08/12/2016 05:27 AM, Jan Hubicka wrote:
* passes.def (pass_early_thread_jumps): Schedule after forwprop.
* tree-pass.h (make_pass_early_thread_jumps): Declare.
* tree-ssa-threadbackward.c (fsm_find_thread_path,
fsm_find_thread_path, profitable_jump_thread_path,
On 08/12/2016 12:02 AM, Richard Biener wrote:
Hmm, isn't walking backwards from uses doing a lot of redundant stmt
walking compared to walking stmts once in forward direction? To me
it sounds like a 'local' patterns matching like optimization rather
than a global one with proper data flow or a
> > * passes.def (pass_early_thread_jumps): Schedule after forwprop.
> > * tree-pass.h (make_pass_early_thread_jumps): Declare.
> > * tree-ssa-threadbackward.c (fsm_find_thread_path,
> > fsm_find_thread_path, profitable_jump_thread_path,
> > fsm_find_control_statement_thread_pat
On Thu, 11 Aug 2016, Jan Hubicka wrote:
> Hi,
> this patch adds early jump threading pass. Jump threading is one of most
> common cases where estimated profile becomes corrupted, because the branches
> are predicted independently beforehand. This patch simply adds early mode to
>
On August 11, 2016 8:41:53 PM GMT+02:00, Jeff Law wrote:
>On 08/11/2016 08:27 AM, Jan Hubicka wrote:
>>
>> On tramp3d all VRP passes threads together 730 branches, all DOM
>passes 393, so
>> FSM threading (with 1957 branches) is the most effective one. Perhaps
>eventually
>> early VRP can also do
On 08/11/2016 09:50 AM, Richard Biener wrote:
Ah, I thought it was exclusively dealing with threading through back
edges which is sth I'd avoid doing early?
No, that's one of the fundamental changes we made for gcc-6, namely
using the FSM threader for general purpose threading rather than just
On 08/11/2016 08:27 AM, Jan Hubicka wrote:
On tramp3d all VRP passes threads together 730 branches, all DOM passes 393, so
FSM threading (with 1957 branches) is the most effective one. Perhaps eventually
early VRP can also do bit of work.
That's roughly consistent with what I've seen. I have so
On 08/11/2016 08:06 AM, Richard Biener wrote:
On Thu, 11 Aug 2016, Jan Hubicka wrote:
Hi,
this patch adds early jump threading pass. Jump threading is one of most
common cases where estimated profile becomes corrupted, because the branches
are predicted independently beforehand. This patch
On 08/11/2016 08:02 AM, Jan Hubicka wrote:
Hi,
this patch adds early jump threading pass. Jump threading is one of most
common cases where estimated profile becomes corrupted, because the branches
are predicted independently beforehand. This patch simply adds early mode to
jump threader that
On August 11, 2016 4:27:00 PM GMT+02:00, Jan Hubicka wrote:
>> On Thu, 11 Aug 2016, Jan Hubicka wrote:
>>
>> > Hi,
>> > this patch adds early jump threading pass. Jump threading is one
>of most
>> > common cases where estimated profile becomes corrupted
> On Thu, 11 Aug 2016, Jan Hubicka wrote:
>
> > Hi,
> > this patch adds early jump threading pass. Jump threading is one of most
> > common cases where estimated profile becomes corrupted, because the branches
> > are predicted independently beforehand. This p
On Thu, 11 Aug 2016, Jan Hubicka wrote:
> Hi,
> this patch adds early jump threading pass. Jump threading is one of most
> common cases where estimated profile becomes corrupted, because the branches
> are predicted independently beforehand. This patch simply adds early mode to
>
Hi,
this patch adds early jump threading pass. Jump threading is one of most
common cases where estimated profile becomes corrupted, because the branches
are predicted independently beforehand. This patch simply adds early mode to
jump threader that does not permit code growth and thus only win
23 matches
Mail list logo
