Re: [Qemu-devel] [RISU PATCH v3 00/18] Support for generating x86 SIMD test images

["Alex Bennée" <alex.bennee@linaro.org>]

Jan Bobek <jan.bobek@gmail.com> writes:

> On 7/12/19 9:34 AM, Alex Bennée wrote:
>>
>> Jan Bobek <jan.bobek@gmail.com> writes:
>>
>>> This is v3 of the patch series posted in [1] and [2]. Note that this
>>> is the first fully-featured patch series implementing all desired
>>> functionality, including (V)LDMXCSR and VSIB-based instructions like
>>> VGATHER*.
>>>
>>> While implementing the last bits required in order to support VGATHERx
>>> instructions, I ran into problems which required a larger redesign;
>>> namely, there are no more !emit blocks as their functionality is now
>>> implemented in regular !constraints blocks. Also, memory constraints
>>> are specified in !memory blocks, similarly to other architectures.
>>>
>>> I tested these changes on my machine; both master and slave modes work
>>> in both 32-bit and 64-bit modes.
>>
>> Two things I've noticed:
>>
>>   ./contrib/generate_all.sh -n 1 x86.risu testcases.x86
>>
>> takes a very long time. I wonder if this is a consequence of constantly
>> needing to re-query the random number generator?
>
> I believe so. While other architectures can be as cheap as a single rand()
> call per instruction, x86 does more like 5-10.

OK

> Even worse, there are some instructions which cannot be generated in
> 32-bit mode (those requiring REX.W prefix, e.g. MMX MOVQ). When I let
> the script run for a little bit, risugen would get stuck in an
> infinite loop, because it could only choose from a single instruction
> which wasn't valid for 32-bit....

The first instruction I see hang is:

  Running: /home/alex/lsrc/tests/risu.git/risugen --xfeatures avx  --pattern CVTSD2SI_64 x86.risu testcases.x86/insn_CVTSD2SI_64__INC.risu.bin
  Generating code using patterns: CVTSD2SI_64 SSE2...
  [                                                                            ]

I wonder if this means we should split the x86.risu by mode? Or some
other way of filtering out patterns that are invalid for a mode?

We do have the concept of classes, see the @ annotations in
aarch64.risu. I guess the generate_all script doesn't handle that nicely
yet though. For now I lumped them all together with:

  ./risugen --pattern "MMX" --xfeatures=sse x86.risu all_mmx.risu.bin
  ./risugen --pattern "SSE" --xfeatures=sse x86.risu all_sse.risu.bin
  ./risugen --pattern "SSE2" --xfeatures=sse x86.risu all_sse2.risu.bin
  ./risugen --pattern "SSE3" --xfeatures=sse x86.risu all_sse3.risu.bin
  ./risugen --pattern "AVX" --xfeatures=avx x86.risu all_avx.risu.bin
  ./risugen --pattern "AVX2" --xfeatures=avx x86.risu all_avx2.risu.bin

>
>> The other is:
>>
>>   set -x RISU ./build/i686-linux-gnu/risu
>>   ./contrib/record_traces.sh testcases.x86/*.risu.bin
>>
>> fails on the first trace when validating the playback. Might want to
>> check why that is.
>
> The SIMD registers aren't getting initialized; both master and
> apprentice need an --xfeatures=XXX parameter for that. Right now the
> default is 'none'; unless the instructions are filtered, you'd need
> --xfeatures=avx (or --xfeatures=sse, and that only works because on my
> laptop, the upper part of ymm registers seems to be always zeroed when
> risu starts).

Ahh OK, I did a lot better with:

  ./contrib/generate_all.sh -n 1 x86.risu testcases.x86 -- --xfeatures avx
  set -x RISU ./build/i686-linux-gnu/risu --xfeatures=avx
  ./contrib/record_traces.sh testcases.x86-avx/*.risu.bin

There are enough failures when you run those against QEMU for now that
we don't need to worry too much about coverage yet ;-)

>>>
>>> Cheers,
>>>  -Jan
>>>
>>> Changes since v2:
>>>   Too many to be listed individually; this patch series might be
>>>   better reviewed on its own.
>>>
>>> References:
>>>   1. https://lists.nongnu.org/archive/html/qemu-devel/2019-06/msg04123.html
>>>   2. https://lists.nongnu.org/archive/html/qemu-devel/2019-07/msg00001.html
>>>
>>> Jan Bobek (18):
>>>   risugen_common: add helper functions insnv, randint
>>>   risugen_common: split eval_with_fields into extract_fields and
>>>     eval_block
>>>   risugen_x86_asm: add module
>>>   risugen_x86_constraints: add module
>>>   risugen_x86_memory: add module
>>>   risugen_x86: add module
>>>   risugen: allow all byte-aligned instructions
>>>   risugen: add command-line flag --x86_64
>>>   risugen: add --xfeatures option for x86
>>>   x86.risu: add MMX instructions
>>>   x86.risu: add SSE instructions
>>>   x86.risu: add SSE2 instructions
>>>   x86.risu: add SSE3 instructions
>>>   x86.risu: add SSSE3 instructions
>>>   x86.risu: add SSE4.1 and SSE4.2 instructions
>>>   x86.risu: add AES and PCLMULQDQ instructions
>>>   x86.risu: add AVX instructions
>>>   x86.risu: add AVX2 instructions
>>>
>>>  risugen                    |   27 +-
>>>  risugen_arm.pm             |    6 +-
>>>  risugen_common.pm          |  117 +-
>>>  risugen_m68k.pm            |    3 +-
>>>  risugen_ppc64.pm           |    6 +-
>>>  risugen_x86.pm             |  518 +++++
>>>  risugen_x86_asm.pm         |  918 ++++++++
>>>  risugen_x86_constraints.pm |  154 ++
>>>  risugen_x86_memory.pm      |   87 +
>>>  x86.risu                   | 4499 ++++++++++++++++++++++++++++++++++++
>>>  10 files changed, 6293 insertions(+), 42 deletions(-)
>>>  create mode 100644 risugen_x86.pm
>>>  create mode 100644 risugen_x86_asm.pm
>>>  create mode 100644 risugen_x86_constraints.pm
>>>  create mode 100644 risugen_x86_memory.pm
>>>  create mode 100644 x86.risu
>>
>>
>> --
>> Alex Bennée
>>


--
Alex Bennée