On 22-02-19 12:09, Thomas Schwinge wrote: > There is however still a problem here: if you have more than one offload > compiler configured,
Ah, right, that's a configuration I'm not testing, evidently. > all of them will react to > '-foffload=-fdump-rtl-mach', and will overwrite their dump files, or if > different compiler passes configured, that may produce more than one > "mach" dump file, all this resulting in "funny" testing FAILs/ERRORs. > But: we can't just use '-foffload=[nvptx]=-fdump-rtl-mach', because > '[nvptx]' can be 'nvptx', or 'nvptx-none', etc. > > Similar problem for a lot of other test cases that you've recently added > (... but thanks for all these, of course!). > > And, "dg-require-effective-target offload_nvptx" is now obsolete, and > should be removed. > > All this I'm addressing in the commits I've just done, see attached: Thanks for fixing that. - Tom
