[ was: Re: x86_64 specific tests in dwz vs. other architectures ]

On 11-12-2019 23:12, Jakub Jelinek wrote:
> On Wed, Dec 11, 2019 at 10:33:33PM +0100, Tom de Vries wrote:
>>> /builddir/build/BUILD/dwz/testsuite/dwz.tests/dw2-skip-prologue.S:309: Error: cannot represent relocation type BFD_RELOC_64
>>> as: /tmp/ccg3sw27.o: unsupported relocation type: 0x1
>>> on i686 etc.  Can't we ensure that the tests are assembled only on the
>>> corresponding architecture and nowhere else if they are architecture
>>> specific?
>>> It is fine if some tests are UNSUPPORTED, just we should completely skip
>>> them if they can't be assembled.
>>
>> Indeed, the messages are ugly, but it's not a correctness problem, just
>> an annoyance.
>>
>> The test execs are build directly from the makefile, not by dejagnu, so
>> we can't use the target testing available there. I suppose I could test
>> for uname -p:
>> ...
>> $ uname -p
>> x86_64
>> ...
> 
> Yeah, and if we e.g. require GNU make to build dwz,

I thought we already required that, given the use of '+=', which is
listed here (
https://www.gnu.org/software/make/manual/html_node/Features.html ) as
something to avoid for portable Makefiles.

Anyway, we're now using the addsuffix function, which I think is a
GNUism. Also, I suspect the use of the shell function is non-portable,
though I didn't find it mentioned at the link above.

> we could use all kinds
> of GNU make features to conditionalize it easily.
> Though, at least for the tests where gdb isn't run on it, perhaps it might
> be good enough to replace actual instructions in the assembly sources with
> .skip or similar; of course it depends also if the location info doesn't
> contain something horribly wrong on other architecture, probably can't mix
> 32-bit and 64-bit and some arches use different characters in assembly, so
> maybe some portability macros and preprocessing would be needed (e.g.
> arm and a few others require %progbits rather than @progbits).

I've fixed things for now in the most restrictive way: only compile
hardcoded assembly test-cases on x86_64, if they were generated on
x86_64, which should fix all those warnings.

I hope to eventually replace all hardcoded assembly as well as
compressed execs with dwarf assembly, which should work for all
architectures.

Thanks,
- Tom