On Fri, May 10, 2013 at 05:18:34PM -0600, Jeff Law wrote:
> On 05/10/2013 08:53 AM, Jakub Jelinek wrote:
> >Hi!
> >
> >Our rotate expansion if rotate optab isn't present for the respective
> >mode looks unsafe for rotations by variable count if that count could
> >be 0, because then it invokes righ
Jan Hubicka wrote:
>>
>> The existing check should work ok with lto. If not then we should
>figure out why we do not merge the main variants properly.
>Hmm, adding:
>Index: tree.c
>===
>--- tree.c (revision 198796)
>+++ tree.c
On Sat, May 11, 2013 at 09:05:52AM +0200, Jakub Jelinek wrote:
> > Seems that we ought to have a testcase, even though it probably
> > means scanning the tree dumps to pick up the undefined behaviour.
> > Approved with a testcase.
>
> I have added lots of testcases recently, for rotation by zero p
Second try.
I removed the fold_single_bit_test thing (I thought I'd handle it, so I
started by the easy part, and never did the rest).
Adapting invert_truthvalue_loc for vectors, I thought: calling
fold_truth_not_expr and build1 if it fails is just the same as
fold_build1. Except that it was
> >
> > BTW2: We badly need to figure out a way to create test cases for FDO... :-(
>
> Yes. I had tried testing awhile back with the gcc regression tests and
> enabling -freorder-blocks-and-partition, but none of the issues I was
> having with larger benchmarks fired. I think there just aren't en
> >But
> >glancing over the the dumps, I see many of them just have different
> >name
> >spaces. Do we even attempt to merge namespace_decl? How types from
> >same
> >namespaces in different units are supposed to match?
> >
> We do not merge namespace decls, which is likely the issue here. My
> i
On Sat, May 11, 2013 at 5:21 AM, Teresa Johnson wrote:
> Here there was a block that happened to be laid out at the very start
> of the cold section (it was jumped to from elsewhere, not reached via
> fall through from its layout predecessor). Thus it was preceded by a
> switch section note, which
On Sat, May 11, 2013 at 1:19 PM, Jan Hubicka wrote:
> Once -freorder-blocks-and-partition actually works, we should enable it by
> default with -fprofile-generate (I recall I was trying to do that once, but
> I am not sure what was outcome back then and why it did not happen).
> That should get it
> On Sat, May 11, 2013 at 1:19 PM, Jan Hubicka wrote:
> > Once -freorder-blocks-and-partition actually works, we should enable it by
> > default with -fprofile-generate (I recall I was trying to do that once, but
> > I am not sure what was outcome back then and why it did not happen).
> > That shou
On Sat, May 11, 2013 at 4:44 AM, Steven Bosscher wrote:
> On Sat, May 11, 2013 at 5:21 AM, Teresa Johnson wrote:
>> Here there was a block that happened to be laid out at the very start
>> of the cold section (it was jumped to from elsewhere, not reached via
>> fall through from its layout predece
On Sat, May 11, 2013 at 4:19 AM, Jan Hubicka wrote:
>> >
>> > BTW2: We badly need to figure out a way to create test cases for FDO... :-(
>>
>> Yes. I had tried testing awhile back with the gcc regression tests and
>> enabling -freorder-blocks-and-partition, but none of the issues I was
>> having
On Sat, May 11, 2013 at 4:38 PM, Teresa Johnson wrote:
> /* If we are partitioning hot/cold basic blocks, we don't want to
> mess up unconditional or indirect jumps that cross between hot
> and cold sections.
>
> Basic block partitioning may result in some jumps that appear to
>
>>> However, I'm not sure if the place where you do it is the most
>>> preferable: I think it should rather go inside of
>>> check_dummy_characteristics (since any check for the dummy
>>> characteristics should include the 'strict' type check, and not only a
>>> check of type compatibility, cf. F08
>> in principle I think your idea to tighten up the type check by
>> symmetrizing it is ok.
>
> I actually kind of copied it from the proc-pointer assignment check, which
> does the same.
Btw, I think you refer to this piece of code in expr.c
(gfc_check_pointer_assign), right?
if (!gfc_com
Hello,
This just removes one unused function, and moves two functions from
emit-rtl.c to reorg.c which is the only place where they're used.
Will commit in a few days, barring objections.
Ciao!
Steven
* rtl.h (next_label, skip_consecutive_labels, link_cc0_insns):
Remove prototy
In AutoFDO, we early-inline callsites that was inlined in profiling
runs regardless of the size limit. With this change, the existing
ipa-inline tunings for AutoFDO is unnecessary: it's fine to just use
the traditional FDO based heuristic. This patch cleans up the original
tunings and make it easie
Looks good.
David
On Sat, May 11, 2013 at 3:53 PM, Dehao Chen wrote:
> In AutoFDO, we early-inline callsites that was inlined in profiling
> runs regardless of the size limit. With this change, the existing
> ipa-inline tunings for AutoFDO is unnecessary: it's fine to just use
> the traditional
17 matches
Mail list logo
