Re: [PATCH] New attribute to create target clones

[Evgeny Stupachenko <evstupac at gmail dot com>]

Bootstrap and make check for x86 passed. No new fails.
Please ignore an empty line added to omp-low.c in the patch, the
misprint will be removed prior to a commit.

Thanks,
Evgeny

On Tue, Oct 13, 2015 at 2:35 AM, Evgeny Stupachenko <evstupac@gmail.com> wrote:
> Hi All,
>
> Here is a new version of patch (attached).
> Bootstrap and make check are in progress (all new tests passed).
>
> New test case g++.dg/ext/mvc4.C fails with ICE, when options lower
> than "-mavx" are passed.
> However it has the same behavior if "target_clones" attribute is
> replaced by 2 corresponding "target" attributes.
> I've filed PR67946 on this:
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=67946
>
> Thanks,
> Evgeny
>
> ChangeLog:
>
> 2015-10-13  Evgeny Stupachenko  <evstupac@gmail.com>
> gcc/
>         * Makefile.in (OBJS): Add multiple_target.o.
>         * attrib.c (make_attribute): Moved from config/i386/i386.c
>         * config/i386/i386.c (make_attribute): Deleted.
>         * multiple_target.c (make_attribute): New.
>         (create_dispatcher_calls): Ditto.
>         (get_attr_len): Ditto.
>         (get_attr_str): Ditto.
>         (is_valid_asm_symbol): Ditto.
>         (create_new_asm_name): Ditto.
>         (create_target_clone): Ditto.
>         (expand_target_clones): Ditto.
>         (ipa_target_clone): Ditto.
>         (ipa_dispatcher_calls): Ditto.
>         * passes.def (pass_target_clone): Two new ipa passes.
>         * tree-pass.h (make_pass_target_clone): Ditto.
>
> gcc/c-family
>         * c-common.c (handle_target_clones_attribute): New.
>         * (c_common_attribute_table): Add handle_target_clones_attribute.
>         * (handle_always_inline_attribute): Add check on target_clones
>         attribute.
>         * (handle_target_attribute): Ditto.
>
> gcc/testsuite
>         * gcc.dg/mvc1.c: New test for multiple targets cloning.
>         * gcc.dg/mvc2.c: Ditto.
>         * gcc.dg/mvc3.c: Ditto.
>         * gcc.dg/mvc4.c: Ditto.
>         * gcc.dg/mvc5.c: Ditto.
>         * gcc.dg/mvc6.c: Ditto.
>         * gcc.dg/mvc7.c: Ditto.
>         * g++.dg/ext/mvc1.C: Ditto.
>         * g++.dg/ext/mvc2.C: Ditto.
>         * g++.dg/ext/mvc3.C: Ditto.
>         * g++.dg/ext/mvc4.C: Ditto.
>
> gcc/doc
>         * doc/extend.texi (target_clones): New attribute description.
>
> On Sat, Oct 10, 2015 at 12:44 AM, Evgeny Stupachenko <evstupac@gmail.com> wrote:
>> On Fri, Oct 9, 2015 at 11:04 PM, Jan Hubicka <hubicka@ucw.cz> wrote:
>>>> On Fri, Oct 9, 2015 at 9:27 PM, Jan Hubicka <hubicka@ucw.cz> wrote:
>>>> >> >Of course it also depends what you inline into function. You can have
>>>> >> >
>>>> >> >bar() target(-mavx) {fancy avx code}
>>>> >> >foobar() { ...... if (avx) bar();}
>>>> >> >foo() ctarget(-mavx,-mno-avx) {....foobar();....}
>>>>
>>>> "no-" targets are not supported
>>>
>>> Why not? I suppose I can use -march=x86_64 in a file compiled with -march=core-avx2 or something like that, too.
>> Sure, you can. target(arch=x86-64) is ok. I mean exactly target(no-avx) returns:
>>
>> aaa.cpp: In function '<built-in>':
>> aaa.cpp:7:5: error: No dispatcher found for no-avx
>>  int bar()
>>      ^
>>
>>>>
>>>> >> >
>>>> >> >Now if you compile with -mavx and because ctarget takes effect only after inlining,
>>>> >> >at inlining time the target attributes will match and we can edn up inline bar->foobar->foo.
>>>> >> >After that we multiversion foo and drop AVX flag we will likely get ICE at expansion
>>>> >> >time.
>>>> >> But isn't that avoided by fixing up the call graph so that all calls
>>>> >> to the affected function are going through the dispatcher?  Or is
>>>> >> that happening too late?
>>>> >
>>>> > There is dispatcher only for foo that is the root of the callgarph tree.
>>>> > When inlining we compare target attributes for match (in can_inline_edge_p).
>>>> > We do not compare ctarget attributes.  Expanding ctarget to target early would
>>>> > avoid need for ctarget handling.
>>>> Currently inlining is disabled for functions with target_clone attribute:
>>>
>>> Do you also disable inlining into functions with target_clone?
>>> What I am concerned about is early inliner inlining (say) AVX code into ctarget
>>> function because at early inlining time the target is not applied, yet.
>> Right. Now I've got your point and ICE on the test.
>> Yes the solution is to disable inline into target_clones function.
>> Or to move the pass creating clones before inline (as you suggested)
>> and leave dispatcher creator after inline.
>>
>> I like you suggestion. It fixes the ICE.
>> I'll fix the patch and retest.
>>
>> Thank you for the review,
>> Evgeny.
>>
>>
>>>
>>> Honza