sinfl can read beyond the end of input buffer

[cdmaczane]

This bug was detected using Clang address sanitizer on Windows (version shipped with Visual Studio 2026).

The stack where the invalid read was detected was:
sinfl_read64
sinfl_refill
sinfl_decompress
sinflate

The fix was to over-allocate the buffer size to be a multiple of 8. I'm guessing this isn't normally a problem because most allocators will return blocks of multiples of 8 on 64 bit systems. In my case I was using a custom scratch allocator that poisons all leftover space that wasn't part of the initially requested size. From what I can tell, the leftover bytes do not need to be zeroed. Could you confirm this?

I'm not sure that the fix is to change reading 8 bytes at a time, which would affect performance. Perhaps just a note at the top of the file?

[cdmaczane]

After some testing, it doesn't seem to be related to being a multiple of 8. The memory was already aligned to 32 bytes. To prevent reading over the end of the buffer, I have to over-allocate by 12 bytes. This suggests that the 8 byte reads are not aligned. The safe thing seems to be to always over-allocate by 16 bytes. I think this is an actual bug.