[Mesa-dev] [PATCH 8/7] i965: Accurately bail on SIMD16 compiles.

[Matt Turner]

On Wed, Mar 12, 2014 at 10:36 AM, Ian Romanick <idr at freedesktop.org> wrote:
> On 03/12/2014 02:51 AM, Ilia Mirkin wrote:
>> On Wed, Mar 12, 2014 at 4:28 AM, Kenneth Graunke <kenneth at whitecape.org> wrote:
>>> Ideally, we'd like to never even attempt the SIMD16 compile if we could
>>> know ahead of time that it won't succeed---it's purely a waste of time.
>>> This is especially important for state-based recompiles, which happen at
>>> draw time.
>>>
>>> The fragment shader compiler has a number of checks like:
>>>
>>>    if (dispatch_width == 16)
>>>       fail("...some reason...");
>>>
>>> This patch introduces a new no16() function which replaces the above
>>> pattern.  In the SIMD8 compile, it sets a "SIMD16 will never work" flag.
>>> Then, brw_wm_fs_emit can check that flag, skip the SIMD16 compile, and
>>> issue a helpful performance warning if INTEL_DEBUG=perf is set.  (In
>>> SIMD16 mode, no16() calls fail(), for safety's sake.)
>>>
>>> The great part is that this is not a heuristic---if the flag is set, we
>>> know with 100% certainty that the SIMD16 compile would fail.  (It might
>>> fail anyway if we run out of registers, but it's always worth trying.)
>>>
>>> Signed-off-by: Kenneth Graunke <kenneth at whitecape.org>
>>> ---
>>>  src/mesa/drivers/dri/i965/brw_fs.cpp         | 69 +++++++++++++++++++++++-----
>>>  src/mesa/drivers/dri/i965/brw_fs.h           |  4 ++
>>>  src/mesa/drivers/dri/i965/brw_fs_visitor.cpp | 44 +++++++++---------
>>>  3 files changed, 83 insertions(+), 34 deletions(-)
>>>
>>> I forgot to send this one out...it applies on top of the previous 7 patches.
>>>
>>> diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp
>>> index 62848be..9ad80c5 100644
>>> --- a/src/mesa/drivers/dri/i965/brw_fs.cpp
>>> +++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
>>> @@ -647,9 +647,8 @@ fs_visitor::emit_shader_time_write(enum shader_time_shader_type type,
>>>  }
>>>
>>>  void
>>> -fs_visitor::fail(const char *format, ...)
>>> +fs_visitor::vfail(const char *format, va_list va)
>>>  {
>>> -   va_list va;
>>>     char *msg;
>>>
>>>     if (failed)
>>> @@ -657,9 +656,7 @@ fs_visitor::fail(const char *format, ...)
>>>
>>>     failed = true;
>>>
>>> -   va_start(va, format);
>>>     msg = ralloc_vasprintf(mem_ctx, format, va);
>>> -   va_end(va);
>>>     msg = ralloc_asprintf(mem_ctx, "FS compile failed: %s\n", msg);
>>>
>>>     this->fail_msg = msg;
>>> @@ -669,6 +666,49 @@ fs_visitor::fail(const char *format, ...)
>>>     }
>>>  }
>>>
>>> +void
>>> +fs_visitor::fail(const char *format, ...)
>>> +{
>>> +   va_list va;
>>> +
>>> +   va_start(va, format);
>>> +   vfail(format, va);
>>> +   va_end(va);
>>> +}
>>> +
>>> +/**
>>> + * Mark this program as impossible to compile in SIMD16 mode.
>>> + *
>>> + * During the SIMD8 compile (which happens first), we can detect and flag
>>> + * things that are unsupported in SIMD16 mode, so the compiler can skip
>>> + * the SIMD16 compile altogether.
>>> + *
>>> + * During a SIMD16 compile (if one happens anyway), this just calls fail().
>>> + */
>>> +void
>>> +fs_visitor::no16(const char *format, ...)
>>> +{
>>> +   va_list va;
>>> +
>>> +   va_start(va, format);
>>> +
>>> +   if (dispatch_width == 16) {
>>> +      vfail(format, va);
>>
>> I think there's a va_end() missing in this path. Not sure what the
>> end-effect of that is, but I'm pretty sure that the recommendation is
>> to always end the list before returning.
>
> I don't know about that.  On some implementations, va_start and va_end
> are macros that introduce a new block.  You can imagine an
> implementation like:
>
> typedef intptr_t va_list;
>
> #define va_start(va, last) \
>     do { \
>         void *__va_temp; \
>         va = ((intptr_t)(void *)&last) + sizeof(last)
>
> #define va_arg(va, type) \
>     (__va_temp = (void *) va, va += sizeof(type), *(type *) __va_temp)
>
> #define va_end(va) \
>     } while (0)
>
> This exact implementation doesn't work because it doesn't align the
> pointers, but the general idea is valid.
>
> I know that some implementations of varargs.h  used to work that way.
> ANSI C may require that stdarg.h be more civilized.

My va_start man page says "On some systems, va_end contains a closing
'}' matching a '{' in va_start" in the NOTES section about the old
varargs.h macros, so I don't think that's the case anymore.

In fact, gcc defines them as:

#define va_start(v,l)   __builtin_va_start(v,l)
#define va_end(v)       __builtin_va_end(v)