From: Andrew Stubbs <ams@codesourcery.com>
To: Richard Biener <richard.guenther@gmail.com>,
Thomas Schwinge <thomas@codesourcery.com>
Cc: Jakub Jelinek <jakub@redhat.com>,
Julian Brown <julian@codesourcery.com>,
GCC Patches <gcc-patches@gcc.gnu.org>
Subject: Re: Host and offload targets have no common meaning of address spaces (was: [ping] Re-unify 'omp_build_component_ref' and 'oacc_build_component_ref')
Date: Fri, 3 Sep 2021 12:42:50 +0100 [thread overview]
Message-ID: <75e539b9-ce8a-c98b-78e3-af999cb2a473@codesourcery.com> (raw)
In-Reply-To: <CAFiYyc1XenNypFzSKv3rYhTzR-y89TvZCvLTzC8fGQmF1sv0hw@mail.gmail.com>
On 24/08/2021 12:43, Richard Biener via Gcc-patches wrote:
> On Tue, Aug 24, 2021 at 12:23 PM Thomas Schwinge
> <thomas@codesourcery.com> wrote:
>>
>> Hi!
>>
>> On 2021-08-19T22:13:56+0200, I wrote:
>>> On 2021-08-16T10:21:04+0200, Jakub Jelinek <jakub@redhat.com> wrote:
>>>> On Mon, Aug 16, 2021 at 10:08:42AM +0200, Thomas Schwinge wrote:
>>> |> Concerning the current 'gcc/omp-low.c:omp_build_component_ref', for the
>>> |> current set of offloading testcases, we never see a
>>> |> '!ADDR_SPACE_GENERIC_P' there, so the address space handling doesn't seem
>>> |> to be necessary there (but also won't do any harm: no-op).
>>>>
>>>> Are you sure this can't trigger?
>>>> Say
>>>> extern int __seg_fs a;
>>>>
>>>> void
>>>> foo (void)
>>>> {
>>>> #pragma omp parallel private (a)
>>>> a = 2;
>>>> }
>>>
>>> That test case doesn't run into 'omp_build_component_ref' at all,
>>> but [I've pushed an altered and extended variant that does],
>>> "Add 'libgomp.c/address-space-1.c'".
>>>
>>> In this case, 'omp_build_component_ref' called via host compilation
>>> 'pass_lower_omp', it's the 'field_type' that has 'address-space-1', not
>>> 'obj_type', so indeed Kwok's new code is a no-op:
>>>
>>> (gdb) call debug_tree(field_type)
>>> <pointer_type 0x7ffff7686b28
>>> type <integer_type 0x7ffff7686498 int address-space-1 SI
>>
>>>> I think keeping the qual addr space here is the wrong thing to do,
>>>> it should keep the other quals and clear the address space instead,
>>>> the whole struct is going to be in generic addres space, isn't it?
>>>
>>> Correct for 'omp_build_component_ref' called via host compilation
>>> 'pass_lower_omp'
>>
>>> However, regarding the former comment -- shouldn't we force generic
>>> address space for all 'tree' types read in via LTO streaming for
>>> offloading compilation? I assume that (in the general case) address
>>> spaces are never compatible between host and offloading compilation?
>>> For [...] "Add 'libgomp.c/address-space-1.c'", propagating the
>>> '__seg_fs' address space across the offloading boundary (assuming I did
>>> interpret the dumps correctly) doesn't seem to cause any problems
>>
>> As I found later, actually the 'address-space-1' per host '__seg_fs' does
>> cause the "Intel MIC (emulated) offloading execution failure"
>> mentioned/XFAILed for 'libgomp.c/address-space-1.c': SIGSEGV, like
>> (expected) for host execution. For GCN offloading target, it maps to
>> GCN 'ADDR_SPACE_FLAT' which apparently doesn't cause any ill effects (for
>> that simple test case). The nvptx offloading target doesn't consider
>> address spaces at all.
>>
>> Is the attached "Host and offload targets have no common meaning of
>> address spaces" OK to push?
>>
>>
>> Then, is that the way to do this, or should we add in
>> 'gcc/tree-streamer-out.c:pack_ts_base_value_fields':
>>
>> if (lto_stream_offload_p)
>> gcc_assert (ADDR_SPACE_GENERIC_P (TYPE_ADDR_SPACE (expr)));
>>
>> ..., and elsewhere sanitize this for offloading compilation? Jakub's
>> suggestion above, regarding 'gcc/omp-low.c:omp_build_component_ref':
>>
>> | I think keeping the qual addr space here is the wrong thing to do,
>> | it should keep the other quals and clear the address space instead
>>
>> But it's not obvious to me that indeed this is the one place where this
>> would need to be done? (It ought to work for
>> 'libgomp.c/address-space-1.c', and any other occurrences would run into
>> the 'assert', so that ought to be "fine", though?)
>>
>>
>> And, should we have a new hook
>> 'void targetm.addr_space.validate (addr_space_t as)' (better name?),
>> called via 'gcc/emit-rtl.c:set_mem_attrs' (only? -- assuming this is the
>> appropriate canonic function where address space use is observed?), to
>> make sure that the requested 'as' is valid for the target?
>> 'default_addr_space_validate' would refuse everything but
>> 'ADDR_SPACE_GENERIC_P (as)'; this hook would need implementing for all
>> handful of targets making use of address spaces (supposedly matching the
>> logic how they call 'c_register_addr_space'?). (The closest existing
>> hook seems to be 'targetm.addr_space.diagnose_usage', only defined for
>> AVR, and called from "the front ends" (C only).)
>
> Are address-spaces to be used in any way for OpenMP offload code? That is,
> does the OpenMP standard talk about them and how to remap things? I'd
> say I agree that any host address-space should go away when the corresponding
> data is offloaded and in case OpenMP allows to specify a target address-space
> that would need to be instantiated in a way so the LTO streaming knows about
> a mapping from the host to the target representation.
The new OpenMP 5 allocator features will permit allocations to different
memories (we're planning an implementation soon). Whether that means a
different address space may be target specific, but I would certainly
expect that it could be. For AMD GCN there is a "flat" address space
that covers most memories, but if you know what memory an address refers
to then there's often a more efficient instruction you can use.
Certainly the numeric address space codes for the host system
architecture have no meaning on the accelerator architecture.
Andrew
next prev parent reply other threads:[~2021-09-03 11:42 UTC|newest]
Thread overview: 15+ messages / expand[flat|nested] mbox.gz Atom feed top
[not found] <992c7c29-5773-45b6-6fb7-ffb71299a98f@mentor.com>
[not found] ` <f2129af9-8314-6eb6-d342-8fd725d0a028@mentor.com>
[not found] ` <ea33e1e0-71ac-1561-01e9-67d875292904@mentor.com>
[not found] ` <fe7bc7ae-6807-ff00-1b3a-e3c7ac41b723@mentor.com>
2021-08-09 14:16 ` Re-unify 'omp_build_component_ref' and 'oacc_build_component_ref' Thomas Schwinge
2021-08-16 8:08 ` [ping] " Thomas Schwinge
2021-08-16 8:21 ` Jakub Jelinek
2021-08-19 20:13 ` Thomas Schwinge
2021-08-20 7:51 ` Richard Biener
2021-08-23 14:30 ` Thomas Schwinge
2021-08-24 7:43 ` Richard Biener
2021-08-20 14:49 ` Jakub Jelinek
2021-08-23 15:55 ` Add 'libgomp.c/address-space-1.c' (was: [ping] Re-unify 'omp_build_component_ref' and 'oacc_build_component_ref') Thomas Schwinge
2021-08-24 10:23 ` Host and offload targets have no common meaning of address spaces " Thomas Schwinge
2021-08-24 11:43 ` Richard Biener
2021-09-03 11:42 ` Andrew Stubbs [this message]
2022-01-13 10:24 ` Host and offload targets have no common meaning of address spaces Thomas Schwinge
2021-09-10 8:03 ` Thomas Schwinge
2022-02-22 17:00 ` Get rid of 'gcc/omp-oacc-neuter-broadcast.cc:oacc_build_component_ref' (was: Re-unify 'omp_build_component_ref' and 'oacc_build_component_ref') 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=75e539b9-ce8a-c98b-78e3-af999cb2a473@codesourcery.com \
--to=ams@codesourcery.com \
--cc=gcc-patches@gcc.gnu.org \
--cc=jakub@redhat.com \
--cc=julian@codesourcery.com \
--cc=richard.guenther@gmail.com \
--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).