[Mesa-dev] [PATCH] llvmpipe: disable vsx in ppc due to LLVM PPC bug

[Oded Gabbay]

On Tue, Nov 17, 2015 at 4:42 PM, Roland Scheidegger <sroland at vmware.com> wrote:
> Am 17.11.2015 um 15:19 schrieb Oded Gabbay:
>> This patch makes sure that if we use altivec (VMX) instructions, we don't
>> use VSX instructions as well, as this cause piglit tests to fail
>>
>> For more details, see: https://llvm.org/bugs/show_bug.cgi?id=25503#c7
>>
>> With this patch, ppc64le reaches parity with x86-64 as far as piglit test
>> suite is concerned.
>>
>> Signed-off-by: Oded Gabbay <oded.gabbay at gmail.com>
>> Cc: "11.0" <mesa-stable at lists.freedesktop.org>
>> ---
>>  src/gallium/auxiliary/gallivm/lp_bld_misc.cpp | 1 +
>>  1 file changed, 1 insertion(+)
>>
>> diff --git a/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp b/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp
>> index 7bda118..8c74cb8 100644
>> --- a/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp
>> +++ b/src/gallium/auxiliary/gallivm/lp_bld_misc.cpp
>> @@ -536,6 +536,7 @@ lp_build_create_jit_compiler_for_module(LLVMExecutionEngineRef *OutJIT,
>>
>>  #if defined(PIPE_ARCH_PPC)
>>     MAttrs.push_back(util_cpu_caps.has_altivec ? "+altivec" : "-altivec");
>> +   MAttrs.push_back("-vsx");
>>  #endif
>>
>>     builder.setMAttrs(MAttrs);
>>
>
> Doesn't that need some llvm version check, otherwise the attribute might
> be unsupported (not entirely sure what happens in this case)? From a
> very quick look, vsx seems to be a fairly recent addition (that is,
> might not be in llvm 3.3 which is the minimum, at least for x86).
> Otherwise looks ok to me. Albeit if llvm indeed miscompiles this ought
> to be fixed as that's quite nasty (we had some bugs in the past where we
> thought it might be miscompilation and it was due to us relying on
> undefined behavior too).
>
> Roland
>

Hi Roland,
vsx was added in release 3.4, according to git log. I believe there is
no need for version check because I don't think ppc64le distributions
uses llvm lower than that (I know that in Red Hat/Fedora we have at
least 3.4). But if you insist I can add that check.

I'm in contact with LLVM devs from IBM to solve this bug and from my
debugging (and I did a LOT of it), I don't believe its our misuse of
the code. Based on different experiments I did (remove optimizations,
replace vmx with vsx instructions and more), I truly think this is a
bug in LLVM backend. It is not an easy bug - you need to go over each
vsx instruction and make sure if its lane-sensitive or not. If so, you
need to add a special treatment inside LLVM code for it.

Once I get a resolution from IBM's devs, I will of course remove this
workaround. Moreover, once that happens, I will probably move some of
the vmx calls to vsx, as vsx have a larger register file.

    Oded