This was found with fuzzing+UBSan build. Output from UBSan: codec/decoder/core/src/rec_mb.cpp:246:14: runtime error: left shift of negative value -30 Which is referring to: iFullMVx = WELS_CLIP3 (iFullMVx, ((-PADDING_LENGTH + 2) << 2), ((pMCRefMem->iPicWidth + PADDING_LENGTH - 19) << 2));
Hi Tyson, we actually know this but we don't think it is a problem we need to fixed, this left shift won't make a problem since the data range of this input is limited.
(In reply to sijchen from comment #2) > Hi Tyson, we actually know this but we don't think it is a problem we need > to fixed, this left shift won't make a problem since the data range of this > input is limited. This is a runtime error reported by UBSan and the attached test case does actually trigger this issue. I am not exactly sure what the implications are that depends on how data resulting from this operation is used. I will attach a call stack to help debug this issue.
I think it is a common false alarm. The purpose of this implementation is to get the 4 times value of the motion vector or the padding offset. Here, the motion vector is at the range of -512 to 512 and the min value of the offset is -32. So the calculated results will not exceed the expected value. I think we can ignore this type of alarms.
Well it's not a false alarm, things just seem to be working at the moment. The problem is that we are treating a '<< 2' as if it were a '* 4'. The problem isn't the value range, yes it should be a concern, but in this case what is happening in undefined regardless. That means we are at the mercy of the compiler. Different things could happen with different compilers, different build types: x86, arm, 32 or 64-bit, optimized, debug, non-debug, combination of the build types or even different compiler version. So TLDR; a left shift of a negative number by 2 is NOT equal to multiplying it by 4 although it may act that way some or most of the time (even when working within seemingly unaffected value ranges).
Although I think it is not a bug, I have added a fix to remove it at commit af6a9a8 at master branch and commit 9c88070 at openh264v1.5 branch. Please help to verify it.
Thanks Haibo. Verified with commit: fb61733b2.
Talked to Sijia. Close it here since it is fixed and verified.
Status: NEW → RESOLVED
Last Resolved: 3 years ago
Resolution: --- → FIXED
You need to log in before you can comment on or make changes to this bug.