This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH][RFC][OpenMP] Forbid target* pragmas in target regions
- From: Jakub Jelinek <jakub at redhat dot com>
- To: Ilya Verbin <iverbin at gmail dot com>
- Cc: gcc-patches at gcc dot gnu dot org, Kirill Yukhin <kirill dot yukhin at gmail dot com>, Thomas Schwinge <thomas at codesourcery dot com>
- Date: Mon, 2 Feb 2015 13:15:34 +0100
- Subject: Re: [PATCH][RFC][OpenMP] Forbid target* pragmas in target regions
- Authentication-results: sourceware.org; auth=none
- References: <20150111212244 dot GD30445 at msticlxl57 dot ims dot intel dot com>
- Reply-to: Jakub Jelinek <jakub at redhat dot com>
On Mon, Jan 12, 2015 at 12:22:44AM +0300, Ilya Verbin wrote:
> Currently if a target* pragma appears within a target region, GCC successfully
> compiles such code (with a warning). But the binary fails at run-time, since it
> tries to call GOMP_target* functions on target.
>
> The spec says: "If a target, target update, or target data construct appears
> within a target region then the behavior is unspecified."
>
> I see 2 options to make the behavior more user-friendly:
> 1. To return an error at compile-time.
> 2. To check at run-time in libgomp whether GOMP_target* is called on target, and
> perform target-fallback if so.
>
> If we will select option #1, the patch is ready.
Option #1 is just wrong. There is nothing wrong with such constructs
appearing in #pragma omp declare target functions etc., the problem is
if you hit them at runtime. You can very well have say #pragma omp declare
target function, that optionally invokes #pragma omp target region e.g.
based on its parameters, state of global variables, what other functions
return etc. - and the program can be written so that that condition just
never happens if the function is already offloaded.
Jakub