Reclaim size regression for aligning fallthrough-only blocks #303
Conversation
abhinay-anubola
force-pushed
the
sanubola.align.fallthrough.only.blocks
branch
from
8c4d73a to
3d412e5
Compare
| if (AlignOffset == 0) | ||
| return; | ||
| unsigned PadBytes = 16 - AlignOffset; | ||
| while (PadBytes) { |
There was a problem hiding this comment.
Ah, I understand again. Could you please add a comment here? We have bundles formats that can only grow by four bytes, and we don't want to elongate these if they would overshoot the requested padding and others could supply an exact fit. Therefore we run in two stages, one which doesn't allow overshoot, and a second that does allow it. We may be forced to overshoot, in which case we will iterate the entire process. We will always terminate because we will saturate on full 16 byte bundles, which don't need padding.
|
Looks good. I think the most important case is a block ending in a JNZ that is 2 bytes bigger than a multiple of 16 followed by a small block, e.g. one 32 bit instruction followed by a fallthrough. In the old situation you would pad the first block with 14 bytes and the second one with 12. In the new situation, you would only pad the second with 10 bytes. Perhaps there are existing tests to check that the end of the function is aligned in the presence of fallthrough blocks. Otherwise one should be added. |
Yet to add tests