Hi Tobias!
On 2020-06-12T16:12:44+0200, Tobias Burnus <tob...@codesourcery.com> wrote:
> For real offloading compilers, the configure-time ENABLE_OFFLOAD
> macro is set to true.
I once toyed with the idea of getting rid of that configure-time
'ENABLE_OFFLOAD' flag, moving it to GCC run time, implicitly set by
'-fopenacc' etc. But it's not trivial, in particular inside libgomp, if
I now remember correctly.
And, that wouldn't help with this issue here, anyway. ;-)
> If it is set, some additional code paths
> are enabled – which can affect code generation with -fopenmp/-fopenacc.
>
> As the PR shows, this can make differences for passing/failing a
> testcase. – In principle, it is supposed to *pass* with ENABLE_OFFLOAD
> but until the PR is fixed, XFAIL those test cases.
>
> OK for the trunk and GCC 10?
> (The patches causing the to-be-xfailed issue will be backported.)
OK, thanks.
Reviewed-by: Thomas Schwinge <tho...@codesourcery.com>
> --- a/gcc/testsuite/lib/target-supports.exp
> +++ b/gcc/testsuite/lib/target-supports.exp
> @@ -997,6 +997,12 @@ proc check_effective_target_fgraphite {} {
> } "-O1 -fgraphite"]
> }
>
> +# Return 1 if compiled with --enable-offload-targets=
> +# This affects host compilation as ENABLE_OFFLOAD then evaluates to true.
> +proc check_effective_target_offloading_enabled {} {
> + return [check_configured_with "--enable-offload-targets"]
Ah, nice and easy! :-)
Eh, actually too easy, strictly speaking: we'd need to 'return 0' if
(only!) configured with '--enable-offload-targets=hsa'. But, as Martin
Jambor recently has gotten approved his "Request to deprecate offloading
to HSAIL in GCC", we probably don't have to worry about that now.
> +}
> +
> # Return 1 if compilation with -fopenacc is error-free for trivial
> # code, 0 otherwise.
>
Grüße
Thomas
-----------------
Mentor Graphics (Deutschland) GmbH, Arnulfstraße 201, 80634 München / Germany
Registergericht München HRB 106955, Geschäftsführer: Thomas Heurung, Alexander
Walter
