[FFmpeg-devel] fate : clang x86

[Eli Friedman]

On Mon, Aug 30, 2010 at 11:56 AM, Ronald S. Bultje <rsbultje at gmail.com> wrote:
> Hi,
>
> On Mon, Aug 30, 2010 at 2:20 PM, Eli Friedman <eli.friedman at gmail.com> wrote:
>> On Mon, Aug 30, 2010 at 9:43 AM, Ronald S. Bultje <rsbultje at gmail.com> wrote:
>>> Hi,
>>>
>>> On Tue, Aug 24, 2010 at 11:56 AM, Michael Niedermayer <michaelni at gmx.at> wrote:
>>>> On Tue, Aug 24, 2010 at 11:48:45AM -0400, Jason Garrett-Glaser wrote:
>>>>> On Tue, Aug 24, 2010 at 11:19 AM, Michael Niedermayer <michaelni at gmx.at> wrote:
>>>>> > On Mon, Aug 23, 2010 at 06:57:20PM -0700, Eli Friedman wrote:
>>>>> >> 2010/8/23 M?ns Rullg?rd <mans at mansr.com>:
>>>>> >> > Eli Friedman <eli.friedman at gmail.com> writes:
>>>>> >> >
>>>>> >> >> 2010/8/21 M?ns Rullg?rd <mans at mansr.com>:
>>>>> >> >>> Eli Friedman <eli.friedman at gmail.com> writes:
>>>>> >> >>>
>>>>> >> >>>> 2010/8/21 M?ns Rullg?rd <mans at mansr.com>:
>>>>> >> >>>>> castet.matthieu at free.fr writes:
>>>>> >> >>>>>
>>>>> >> >>>>>> Hi,
>>>>> >> >>>>>>
>>>>> >> >>>>>> on freebsd "-mllvm -regalloc=fast" cflags are used to make clang/llvm accept
>>>>> >> >>>>>> some inline asm.
>>>>> >> >>>>>>
>>>>> >> >>>>>> May be we should do the same on linux ?
>>>>> >> >>>>>
>>>>> >> >>>>> I tried and failed to figure out what that flag does. ?I assume it
>>>>> >> >>>>> does something with the register allocator, but I'd like to know what.
>>>>> >> >>>>
>>>>> >> >>>> It's a workaround of sorts for
>>>>> >> >>>> http://llvm.org/bugs/show_bug.cgi?id=4668 . ?LLVM essentially has two
>>>>> >> >>>> register allocator implementations: one is the "fast" allocator, which
>>>>> >> >>>> is a local register allocator used for -O0, and the other is the
>>>>> >> >>>> "linear scan" allocator, which is the slower global register allocator
>>>>> >> >>>> used for -O1+. ?"-mllvm -regalloc=fast" forces the use of the "fast"
>>>>> >> >>>> allocator, which leads to slower generated code, but isn't affected by
>>>>> >> >>>> the bug in question.
>>>>> >> >>>
>>>>> >> >>> Sounds like it's not suitable for production use. ?Any chance they'll
>>>>> >> >>> fix the bug?
>>>>> >> >>
>>>>> >> >> In the near future? ?Not very likely... from what I understand, it's a
>>>>> >> >> relatively difficult issue to solve, and bugs rejecting valid inline
>>>>> >> >> asm are generally considered low priority for the people who know the
>>>>> >> >> register allocator well enough to fix this.
>>>>> >> >
>>>>> >> > That leaves two options:
>>>>> >> >
>>>>> >
>>>>> >> > 1. Declare clang officially unsupported for x86_32.
>>>>> >
>>>>> > ok with me
>>>>> >
>>>>> >
>>>>> >> > 2. Fix our code.
>>>>> >>
>>>>> >> Attached attempts option 2. ?It's essentially a straight port of the
>>>>> >> inline asm to yasm, with a couple minor changes to reduce the number
>>>>> >> of arguments. ?This removes all the inline asm blocks from
>>>>> >> mpegvideo_mmx_template.c, which allows clang to successfully build
>>>>> >> ffmpeg. ?Passes regression tests.
>>>>> >
>>>>> > and is slower due to additional call overhead
>>>>>
>>>>> Why would it have additional overhead? ?Both are called by the same
>>>>> function pointer.
>>>>
>>>> - ? ? ? ? ? ?: "+a" (last_non_zero_p1)
>>>> - ? ? ? ? ? ?: "r" (block+64), "r" (qmat), "r" (bias),
>>>> - ? ? ? ? ? ? ?"r" (inv_zigzag_direct16+64), "r" (temp_block+64)
>>>> - ? ? ? ?);
>>>> + ? ? ? ?last_non_zero_p1 =
>>>> + ? ? ? ? ? ?RENAMEcore(ff_dct_quantize_core_h263)(block+64, qmat, bias,
>>>> + ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? inv_zigzag_direct16+64,
>>>> + ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? temp_block+64, overflow);
>>>> ? ? }else{ // FMT_H263
>>>
>>> Eli, are you still working on this?
>>
>> Do you have any suggestions for putting the patch into an acceptable form?
>
> Make the called function a macro and inline it manually in yasm?

Writing dct_quantize in its current form completely in yasm is a
non-starter; it makes function calls, and depends on offsets into
structs defined in C.

-Eli