MVFrame.cpp:HorizontalWiener_sse2() writes to memory protected by VS_FRAME_GUARD

[bryson1234]

HorizontalWiener_sse2() contains this snippet:

for (int x = 2; x < nWidth - 4; x += 8) { ... _mm_storel_epi64((__m128i *)&pDst[x], m0); }

Since _mm_storel_epi64() writes 8 bytes, this overwrites pDst[nWidth], pDst[nWidth + 1], and pDst[nWidth + 2], which may overwrite VS_FRAME_GUARD_PATTERN which is stored before the beginning and after the end of the pixel buffer. This write is caught by VSFrame::verifyGuardPattern() in the src/core/vscore.cpp in the VapourSynth R44 sources and makes VSNode::getFrameInternal() in the same file SIGABRT.

The following is the backtrace when gdb catches (via “watch”) when someone writes into the protected memory:

0 HorizontalWiener_sse2 (pDst=..., pSrc=..., nPitch=736, nWidth=736, nHeight=592, bitsPerSample=...)

at /var/tmp/portage/media-plugins/vapoursynth-mvtools-20/work/vapoursynth-mvtools-20/src/MVFrame.cpp:443

1 0x00001554c1291365 in mvpRefine (mvp=..., sharp=...)

at /var/tmp/portage/media-plugins/vapoursynth-mvtools-20/work/vapoursynth-mvtools-20/src/MVFrame.cpp:1463

2 0x00001554c129297d in mvpRefine (sharp=..., mvp=...)

at /var/tmp/portage/media-plugins/vapoursynth-mvtools-20/work/vapoursynth-mvtools-20/src/MVFrame.cpp:1792

3 mvfRefine (mvf=..., nMode=..., sharp=...)

at /var/tmp/portage/media-plugins/vapoursynth-mvtools-20/work/vapoursynth-mvtools-20/src/MVFrame.cpp:1790

4 0x00001554c1292e71 in mvgofRefine (mvgof=..., nMode=..., sharp=...)

at /var/tmp/portage/media-plugins/vapoursynth-mvtools-20/work/vapoursynth-mvtools-20/src/MVFrame.cpp:1891

5 0x00001554c1296f14 in mvsuperGetFrame (n=..., activationReason=..., instanceData=..., frameData=..., frameCtx=..., core=..., vsapi=...)

at /var/tmp/portage/media-plugins/vapoursynth-mvtools-20/work/vapoursynth-mvtools-20/src/MVSuper.c:110

6 0x0000155551f22a3e in VSNode::getFrameInternal (this=..., n=..., activationReason=..., frameCtx=...) at src/core/vscore.cpp:849

7 0x0000155551f3b7b7 in VSThreadPool::runTasks (owner=..., stop=...) at src/core/vsthreadpool.cpp:186

8 0x0000155554e0026f in std::execute_native_thread_routine (__p=...)

at /var/tmp/portage/sys-devel/gcc-7.3.0-r3/work/gcc-7.3.0/libstdc++-v3/src/c++11/thread.cc:83

9 0x0000155553d01a15 in start_thread (arg=...) at pthread_create.c:465

10 0x000015555484568f in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:95

The above is from a commercial DVD, so I cannot easily post a video source. I’m in fact surprised that I only encounter his bug now since it seems to have been present since at least Oct 2016.

I’m not suggesting a patch because the SSE2 trickery is beyond me (like, six loads are more efficient than one load and five shifts and masks?).

[sekrit-twc]

It looks like the easy fix is to increase the clip width in MVSuper.c by adding 4 to "nSuperWidth".

[dubhater]

:(