On 09/10/2017 03:22 AM, Luc Van Oostenryck wrote:
> On Sat, Sep 9, 2017 at 11:02 PM, Uwe Kleine-König <uwe@kleine-koenig.org> wrote:
>>
>> I tried this on ppc64le and it fixes 2 tests, so were at
>>
>>         Out of 287 tests, 273 passed, 14 failed (10 of them are known to fail)
>>
>> The repaired tests are:
>>
>>         backend/hello.c
>>         backend/sum.c
>>
>> unexpected failures are:
>>
>>         backend/arithmetic-ops.c
>>         backend/cmp-ops.c
>>         backend/int-cond.c
>>         backend/logical-ops.c
>>
>> These are not about missing preprocessor tokens as there are no system
>> includes used, but the error there is
>>
>>         Error: unrecognized opcode: `...`
>>
>> . I didn't look into what the problem is there, but attached the test
>> log.
> 
> It clearly looks as the code generated by LLVM  (the machine code/assembly
> not LLVM's bytecode) is not understood by the assembler (or at least some
> instructions). Probably a mismatch with the architecture version or something
> like that.
> 
>> I did a build test on a few other Debian machines, arm64 was fine, mips
>> and mipx64el had 15 failures, ppc64 (i.e. big endian) had 12. I didn't
>> look in more detail and suggest to tackle one after the other :-)
> 
> I fully test on x86, x86-64, arm & ARM64 (with LLVM 3.9 or 4.0).
> I also test on ppc64 but not the LLVM part because the machines I have
> access to have not LLVM installed and I never bothered to install it myself.
> 
> Would it be possible to have access to a machine with the architectures
> you care about?

Debian provides access to porter boxes for such problems. See
https://dsa.debian.org/doc/guest-account/.

> Meanwhile, is it possible to have the build logs but with 'make V=1 ...' ?
> It would also be useful to have:
> - the output of 'uname -a'
> - the details about the version of LLVM you're using

Sure, can do. Attached is a build from the ppc64el machine with Chris'
patch applied. Tell me if it contains everything you need.

> On the other hand, you/us should disable the sparse-llvm part since:
> - it's something that is bundled and build by default but absolutely not
>   needed (or even useful) to use sparse.
> - it hasn't been written for anything else than x86/x86-64 (no 'layout'
>   for anything else than those architectures.

With your patch applied I get (independent of having Chris' patch
applied or not):

	Out of 265 tests, 255 passed, 10 failed (10 of them are known to fail)

Best regards
Uwe