[FFmpeg-devel] [PATCH] ff_h[yc]scale_fast_mmext always clobbers %rbx

Michael Niedermayer michaelni at gmx.at
Tue May 12 01:18:21 CEST 2015


On Mon, May 11, 2015 at 03:55:40PM -0700, Nick Lewycky wrote:
> On 8 May 2015 at 14:28, Michael Niedermayer <michaelni at gmx.at> wrote:
> 
> > On Fri, May 08, 2015 at 12:43:20PM -0700, Nick Lewycky wrote:
> > > On 8 May 2015 at 12:06, Michael Niedermayer <michaelni at gmx.at> wrote:
> > >
> > > > On Fri, May 08, 2015 at 10:50:49AM -0700, Nick Lewycky wrote:
> > > > > On 6 May 2015 at 18:03, Michael Niedermayer <michaelni at gmx.at>
> > wrote:
> > > > >
> > > > > > On Wed, May 06, 2015 at 04:08:09PM -0700, Nick Lewycky wrote:
> > > > > > > On 6 May 2015 at 15:06, Michael Niedermayer <michaelni at gmx.at>
> > > > wrote:
> > > > > > >
> > > > > > > > Hi
> > > > > > > >
> > > > > > > > On Wed, May 06, 2015 at 11:52:59AM -0700, Nick Lewycky wrote:
> > > > > > > > > Newer versions of clang will allocate %rbx as one of the
> > inline
> > > > asm
> > > > > > > > inputs,
> > > > > > > >
> > > > > > > > Thats great
> > > > > > > >
> > > > > > > >
> > > > > > > > > even in PIC. This occurs when building ffmpeg with clang
> > > > > > > > -fsanitize=address
> > > > > > > > > -O1 -fPIE. Because the asm does clobber %bx whether PIC is
> > on or
> > > > off,
> > > > > > > > just
> > > > > > > > > include %bx in the clobber list regardless of whether PIC is
> > on
> > > > or
> > > > > > off.
> > > > > > > >
> > > > > > > > you cant include *bx in the clobber list with PIC
> > > > > > > > it breaks compilers that are less great, like gcc
> > > > > > > >
> > > > > > >
> > > > > > > Doh!! I did check for this, but only tried x86-64, not x86-32.
> > Sorry!
> > > > > > >
> > > > > > > gcc (Ubuntu/Linaro 4.6.3-1ubuntu5) 4.6.3
> > > > > > > > ../libswscale/x86/hscale_fast_bilinear_simd.c: In function
> > > > > > > > ‘ff_hyscale_fast_mmxext’:
> > > > > > > > ../libswscale/x86/hscale_fast_bilinear_simd.c:205:5: error: PIC
> > > > > > register
> > > > > > > > clobbered by ‘%ebx’ in ‘asm’
> > > > > > > > ../libswscale/x86/hscale_fast_bilinear_simd.c: In function
> > > > > > > > ‘ff_hcscale_fast_mmxext’:
> > > > > > > > ../libswscale/x86/hscale_fast_bilinear_simd.c:276:5: error: PIC
> > > > > > register
> > > > > > > > clobbered by ‘%ebx’ in ‘asm’
> > > > > > > >
> > > > > > > >
> > > > > > > > also what exactly are you trying to fix ?
> > > > > > > > or rather what exactly goes how exactly wrong with the code as
> > it
> > > > is
> > > > > > > > if rbx is used ?
> > > > > > > >
> > > > > > >
> > > > > > > Ok, let's look at ff_hcscale_fast_mmext. Preprocessor directives
> > > > > > evaluated
> > > > > > > in PIC x86-64, the inline constraints work out to:
> > > > > > >
> > > > > > >         :: "m" (src1), "m" (dst1), "m" (filter), "m" (filterPos),
> > > > > > >            "m" (mmxextFilterCode), "m" (src2), "m"(dst2)
> > > > > > >           ,"m" (ebxsave)
> > > > > > >           ,"m"(retsave)
> > > > > > >         : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D
> > > > > > >
> > > > > > > so clang looks at that and decides that it can pick src1 =
> > (%r10),
> > > > dst1 =
> > > > > > > (%r8), filter = (%r11), filterPos = (%r12), mmxextFilterCode =
> > > > (%r15),
> > > > > > src2
> > > > > > > = (%rbx), dst2 = (%r14), ebxsave = (%r13), retsave = (%r9). The
> > > > problem
> > > > > > > there is src2 being (%rbx).
> > > > > > >
> > > > > > > Now let's look at how we use them:
> > > > > > >
> > > > > > >         "mov                 %0, %%"REG_c"  \n\t"
> > > > > > >         "mov                 %1, %%"REG_D"  \n\t"
> > > > > > >         "mov                 %2, %%"REG_d"  \n\t"
> > > > > > >         "mov                 %3, %%"REG_b"  \n\t"  // Clobbers
> > %rbx /
> > > > > > src2
> > > > > > > / %5 here
> > > > > > >         "xor          %%"REG_a", %%"REG_a"  \n\t"
> > > > > > >         PREFETCH"   (%%"REG_c")             \n\t"
> > > > > > >         PREFETCH" 32(%%"REG_c")             \n\t"
> > > > > > >         PREFETCH" 64(%%"REG_c")             \n\t"
> > > > > > >
> > > > > > >         CALL_MMXEXT_FILTER_CODE
> > > > > > >         CALL_MMXEXT_FILTER_CODE
> > > > > > >         CALL_MMXEXT_FILTER_CODE
> > > > > > >         CALL_MMXEXT_FILTER_CODE
> > > > > > >         "xor          %%"REG_a", %%"REG_a"  \n\t"
> > > > > > >         "mov                 %5, %%"REG_c"  \n\t"  // %5 is read
> > too
> > > > > > late,
> > > > > > > we get %3 / filterPos instead
> > > > > > >         "mov                 %6, %%"REG_D"  \n\t"
> > > > > > >         PREFETCH"   (%%"REG_c")             \n\t"
> > > > > > >         PREFETCH" 32(%%"REG_c")             \n\t"
> > > > > > >         PREFETCH" 64(%%"REG_c")             \n\t"
> > > > > > >
> > > > > > >         CALL_MMXEXT_FILTER_CODE  // Crash...
> > > > > > >
> > > > > > > The two marked instructions are intended to refer to different
> > > > contents.
> > > > > > > CALL_MMXEXT_FILTER_CODE moves RBX to ESI and calls
> > mmxextFilterCode:
> > > > > > >
> > > > > > >         "movl            (%%"REG_b"), %%esi     \n\t"\
> > > > > > >         "call                    *%4            \n\t"\  //
> > Crash...
> > > > > > >
> > > > > > > That callee ultimately segfaults:
> > > > > > >
> > > > > > > => 0x7fffefa45000:      movq   (%rdx,%rax,1),%mm3
> > > > > > > => 0x7fffefa45004:      movd   (%rcx,%rsi,1),%mm0
> > > > > > > Program received signal SIGSEGV, Segmentation fault.
> > > > > > > 0x00007fffefa45004 in ?? ()
> > > > > > > => 0x7fffefa45004:      movd   (%rcx,%rsi,1),%mm0
> > > > > > >
> > > > > > > I'm trying to fix this segfault.
> > > > > > >
> > > > > > > also ideally changes to this code should be tested against
> > > > gcc/clang/icc
> > > > > > > > of various versions with and without PIC, 32 and 64 bit
> > > > > > > > this code is more tricky than than the average so theres a good
> > > > > > > > change changes to it will break some compilerAnd if you decide
> > that you don’t want to edit the hunk after all, simply delete everything in
> > the file, save & close it. Git will then abort the editing process and
> > leave the hunk unchanged, returning you to the standard selection.
> > > > > > > >
> > > > > > >
> > > > > > > I understand that, but I may not be able to test as many
> > > > environments as
> > > > > > > you like. I'm going to give testing it my best effort, but I can
> > > > tell you
> > > > > > > up front that I'm only going to test gcc and clang, on an x86
> > Ubuntu
> > > > > > > machine. I don't have ICC, so I can't test that.
> > > > > >
> > > > > > iam not sure it would work but, configure could easily test if
> > > > > > ebx can be put on the clobber list and that then could be used
> > > > > > in place of PIC in the checks
> > > > > >
> > > > >
> > > > > I checked gcc 4.4, 4.6, 4.7, 4.8, and clang 3.3, 3.4, 3.5 in 32 and
> > > > 64-bit
> > > > > builds, and PIC vs. not, cross product thereof. The only one that
> > emits
> > > > > this error is all versions of gcc in 32-bit mode with PIC enabled. On
> > > > > 64-bit, gcc and clang do the same thing, they'll spill BX if they
> > want to
> > > > > preserve it.
> > > > >
> > > > > Based on that, I decided we could sink the manual saving and
> > restoring of
> > > > > ebx down to 32-bit+PIC mode, and just list it as a clobber in 64-bit
> > mode
> > > > > regardless of the PIC setting. That works in every compiler and mode
> > I
> > > > > tried.
> > > > >
> > > > > Tested with make + make check + make fate but not "make fake
> > > > SAMPLES=...".
> > > > > Updated patch attached! Let me know what you think.
> > > > >
> > > > > and thanks for the detailed informaion
> > > > > >
> > > > >
> > > > > No problem, thanks for the review!
> > > >
> > > > please provide a git patch with commit message
> > > >
> > >
> > > Attached!
> > >
> > >
> > > > [...]
> > > >
> > > > >      int i;
> > > > > +#if ARCH_X86_64
> > > > > +    uint64_t retsave;
> > > > > +#else
> > > > >  #if defined(PIC)
> > > > >      uint64_t ebxsave;
> > > > >  #endif
> > > > > -#if ARCH_X86_64
> > > > > -    uint64_t retsave;
> > > > >  #endif
> > > > >
> > > > >      __asm__ volatile(
> > > > > -#if defined(PIC)
> > > > > -        "mov               %%"REG_b", %5        \n\t"
> > > > >  #if ARCH_X86_64
> > > > >          "mov               -8(%%rsp), %%"REG_a" \n\t"
> > > > > -        "mov               %%"REG_a", %6        \n\t"
> > > > > -#endif
> > > > > +        "mov               %%"REG_a", %5        \n\t"  // retsave
> > > > >  #else
> > > > > -#if ARCH_X86_64
> > > > > -        "mov               -8(%%rsp), %%"REG_a" \n\t"
> > > > > -        "mov               %%"REG_a", %5        \n\t"
> > > > > +#if defined(PIC)
> > > > > +        "mov               %%"REG_b", %5        \n\t"  // ebxsave
> > > > >  #endif
> > > > >  #endif
> > > > >          "pxor                  %%mm7, %%mm7     \n\t"
> > > > > @@ -256,28 +253,25 @@ void ff_hyscale_fast_mmxext(SwsContext *c,
> > int16_t
> > > > *dst,
> > > > >          CALL_MMXEXT_FILTER_CODE
> > > > >          CALL_MMXEXT_FILTER_CODE
> > > > >
> > > > > -#if defined(PIC)
> > > > > -        "mov                      %5, %%"REG_b" \n\t"
> > > > > -#if ARCH_X86_64
> > > > > -        "mov                      %6, %%"REG_a" \n\t"
> > > > > -        "mov               %%"REG_a", -8(%%rsp) \n\t"
> > > > > -#endif
> > > > > -#else
> > > > >  #if ARCH_X86_64
> > > > >          "mov                      %5, %%"REG_a" \n\t"
> > > > >          "mov               %%"REG_a", -8(%%rsp) \n\t"
> > > > > +#else
> > > > > +#if defined(PIC)
> > > > > +        "mov                      %5, %%"REG_b" \n\t"
> > > > >  #endif
> > > > >  #endif
> > > > >          :: "m" (src), "m" (dst), "m" (filter), "m" (filterPos),
> > > > >             "m" (mmxextFilterCode)
> > > > > +#if ARCH_X86_64
> > > > > +          ,"m"(retsave)
> > > > > +#else
> > > > >  #if defined(PIC)
> > > > >            ,"m" (ebxsave)
> > > > >  #endif
> > > > > -#if ARCH_X86_64
> > > > > -          ,"m"(retsave)
> > > > >  #endif
> > > > >          : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D
> > > > > -#if !defined(PIC)
> > > > > +#if ARCH_X86_64 || !defined(__PIC__)
> > > >
> > > > this mixes __PIC__ and PIC
> > > >
> > >
> > > Fixed. Thanks!
> > >
> > >
> > > >
> > > >
> > > > [...]
> > > >
> > > > --
> > > > Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
> > > >
> > > > Into a blind darkness they enter who follow after the Ignorance,
> > > > they as if into a greater darkness enter who devote themselves
> > > > to the Knowledge alone. -- Isha Upanishad
> > > >
> > > > _______________________________________________
> > > > ffmpeg-devel mailing list
> > > > ffmpeg-devel at ffmpeg.org
> > > > http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> > > >
> > > >
> >
> > >  hscale_fast_bilinear_simd.c |   65
> > +++++++++++++++++---------------------------
> > >  1 file changed, 26 insertions(+), 39 deletions(-)
> > > 4172e9dc3002b77114f98a067215f0802901b6e2  ebxsave-4.patch
> > > commit a1922839bbc0eaf3fb1451b6a01c0f2b36f67946
> > > Author: Nick Lewycky <nlewycky at google.com>
> > > Date:   Fri May 8 12:40:44 2015 -0700
> > >
> > >     On 64-bit with PIC, the compiler may allocate BX and won't complain
> > about it being in our clobber list. Since we do clobber BX in x86-64+PIC
> > mode, add it to the clobber list. Fixes a miscompile with address sanitizer
> > in clang 3.7-svn.
> >
> > commit messages should start with what is changed like:
> > "swscale/x86/hscale_fast_bilinear_simd:"
> >
> 
> Sorry about that. I should've read prior commit messages and followed the
> style.
> 
> also git am doesnt apply this, this is not a correct git patch
> > its rather the outout of git show i suspect
> > a git patch is created with "git format-patch -1"
> >
> 
> Yes, it was "git show" output. Now with git format-patch -1.
> 
> ideally the commit message should list, what is changed, why it is
> > changed and how it is changed
> >
> 
> Ok, I think the new commit message does?

the first line should be a short description followed by more details
if needed
not really critical but as theres another question below ...


> 
> the actual change is ok but if you like it could be split (i actually
> > think its better if it is split)
> > in one change fixing the bug and another changing the operand
> > numbering and removing ebxsave in the case where its not needed
> > both together make the patch somewhat hard to read
> >
> 
> I don't see any way to fix the bug without changing the operand numbering?
> Fixing the bug requires a change to when ebxsave is in the output vs.
> clobber list, which triggers a renumbering of accesses to retsave.

the renumbering occurs due to removial of ebxsave, i dont think this
removial is needed for the bug fix. It would be unused but that would
then be fixed in the next commit and i think it would make this
alot more readable

[...]
>             "m" (mmxextFilterCode)
> +#if ARCH_X86_64
> +          ,"m"(retsave)
> +#else
>  #if defined(PIC)
>            ,"m" (ebxsave)
>  #endif
> -#if ARCH_X86_64
> -          ,"m"(retsave)
>  #endif
>          : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D

[...]

-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

There will always be a question for which you do not know the correct answer.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 181 bytes
Desc: Digital signature
URL: <https://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20150512/dcfec582/attachment.asc>


More information about the ffmpeg-devel mailing list