From: Richard Biener <rguenther@suse.de>
To: Tom de Vries <Tom_deVries@mentor.com>
Cc: GCC Patches <gcc-patches@gcc.gnu.org>,
Thomas Schwinge <thomas@codesourcery.com>
Subject: Re: [PATCH, PR81430] Use finalize_options in lto1
Date: Tue, 25 Jul 2017 11:24:00 -0000 [thread overview]
Message-ID: <alpine.LSU.2.20.1707251324070.10808@zhemvz.fhfr.qr> (raw)
In-Reply-To: <7b5c0846-6a77-cc4a-5ad2-475552468ead@mentor.com>
On Fri, 21 Jul 2017, Tom de Vries wrote:
> On 07/21/2017 11:41 AM, Richard Biener wrote:
> > On Thu, 20 Jul 2017, Tom de Vries wrote:
> >
> > > On 07/20/2017 12:10 PM, Richard Biener wrote:
> > > > On Thu, 20 Jul 2017, Tom de Vries wrote:
> > > >
> > > > > Hi,
> > > > >
> > > > > this patch fixes PR81430, an ICE in the libgomp testsuite for both
> > > > > openmp
> > > > > and
> > > > > openacc test-cases for x86_64 with nvptx accelerator.
> > > > >
> > > > > The scenario how we hit the ICE is as follows:
> > > > > - a testcase is compiled with -O2
> > > > > - ix86_option_optimization_table enables
> > > > > OPT_freorder_blocks_and_partition at -O2
> > > > > - cc1 writes out the flag as part of
> > > > > DECL_FUNCTION_SPECIFIC_OPTIMIZATION
> > > > > - lto1 reads in the flag as part of
> > > > > DECL_FUNCTION_SPECIFIC_OPTIMIZATION
> > > > > - lto1 uses the flag, and runs pass_partition_blocks
> > > > > - pass_partition_blocks ICEs, because it generates code that is not
> > > > > supported by the nvptx target.
> > > > >
> > > > > Note that for standalone compilation for single-thread ptx execution,
> > > > > we
> > > > > don't
> > > > > attempt to run pass_partition_blocks. This is because for nvptx,
> > > > > TARGET_HAVE_NAMED_SECTIONS is set to false, and this bit in
> > > > > finish_options
> > > > > switches off pass_partition_blocks:
> > > > > ...
> > > > > /* If the target requested unwind info, then turn off the
> > > > > partitioning optimization with a different message. Likewise,
> > > > > if
> > > > > the target does not support named sections. */
> > > > >
> > > > > if (opts->x_flag_reorder_blocks_and_partition
> > > > > && (!targetm_common.have_named_sections
> > > > > || (opts->x_flag_unwind_tables
> > > > > && targetm_common.unwind_tables_default
> > > > > && (ui_except == UI_SJLJ || ui_except >= UI_TARGET))))
> > > > > {
> > > > > if (opts_set->x_flag_reorder_blocks_and_partition)
> > > > > inform (loc,
> > > > > "-freorder-blocks-and-partition does not work "
> > > > > "on this architecture");
> > > > > opts->x_flag_reorder_blocks_and_partition = 0;
> > > > > opts->x_flag_reorder_blocks = 1;
> > > > > }
> > > > > ...
> > > > >
> > > > > The patch fixes this by calling finish_options in lto1 after
> > > > > cl_optimization_restore.
> > > > >
> > > > > Points for review:
> > > > > 1. I'm uncertain though about the placement of the call. Perhaps it
> > > > > should
> > > > > be
> > > > > in cl_optimization_restore, before
> > > > > targetm.override_options_after_change?
> > > > >
> > > > > 2. I think that this is offloading specific, so perhaps this should be
> > > > > guarded
> > > > > with lto_stream_offload_p or #ifdef ACCEL_COMPILER or some such.
> > > >
> > > > Hmm, I agree with #2. I think it conceptually is a LTO stream
> > > > adjustment
> > > > and thus we should do this at the time we stream in the
> > > > optimization/target nodes (like we remap modes for example). Not
> > > > sure if it's possible to do this at that point, but it looks like
> > > > finish_options takes two option structs and thus we should be able to
> > > > call it.
> > > >
> > >
> > > With what parameters? Patch below tries with same option struct, but ...
> > >
> > > > Do you get the inform note? I suppose we don't really want that, no?
> > > >
> > >
> > > ... I think that way we'll get the inform note (while the previous
> > > solution
> > > did not).
> > >
> > > I've also tried with a tmp2 memset to 0, but that ran into problems when
> > > doing
> > > a maybe_set_param_value.
> >
> > Use global_options_set? That should do what the other patch did.
> >
>
> I managed to get it working now. The variable tmp was only partly
> initialized, which caused the problems when calling maybe_set_param_value. I'm
> now using init_options_struct.
>
> There's no note when using -O2 or "-O2 -freorder-blocks-and-partition".
>
> But when I do "-O2 -foffload=-freorder-blocks-and-partition" I get:
> ...
> lto1: note: '-freorder-blocks-and-partition' does not work on this
> architecture
> lto1: note: '-freorder-blocks-and-partition' does not support unwind info on
> this architecture
> ...
>
> And for "-O0 -foffload=-freorder-blocks-and-partition" I just get:
> ...
> lto1: note: '-freorder-blocks-and-partition' does not work on this
> architecture
> ...
Ok.
Thanks,
Richard.
> Thanks,
> - Tom
>
--
Richard Biener <rguenther@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)
next prev parent reply other threads:[~2017-07-25 11:24 UTC|newest]
Thread overview: 10+ messages / expand[flat|nested] mbox.gz Atom feed top
2017-07-20 9:22 Tom de Vries
2017-07-20 10:10 ` Richard Biener
2017-07-20 15:35 ` Tom de Vries
2017-07-21 9:41 ` Richard Biener
2017-07-21 13:40 ` Tom de Vries
2017-07-25 11:24 ` Richard Biener [this message]
2017-07-21 9:49 ` [committed, nvptx] Add nvptx_override_options_after_change Tom de Vries
2017-08-08 8:46 ` Thomas Schwinge
2017-08-08 8:58 ` Tom de Vries
2017-08-11 16:46 ` Thomas Schwinge
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=alpine.LSU.2.20.1707251324070.10808@zhemvz.fhfr.qr \
--to=rguenther@suse.de \
--cc=Tom_deVries@mentor.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=thomas@codesourcery.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).