This reverts commit 526416b1e4.
Figured out a fix for the deadlocks; resubmitting the patch.
The first version was deadlocking because the switch statement in
zipWriter.write would choose a specific zip entry to start writing,
but the individual chunks may not have all necessarily been compressed
yet. When each individual chunk was made to require to request its own
allocations, the compression of the chunks of the file being currently
written could be blocked waiting for memory to be freed by chunks from
other files that hadn't yet started being written.
This patch is much like the original except it preallocates the memory
for the entire file upfront (and happens to use the total file size
rather than the compressed size, but I didn't observe that to cause any
performance differences).
Bug: 64536066
Test: m -j dist showcommands # which runs soong_zip to package everything
Change-Id: Id1d7ff415e54d3a6be71188abbdbbbab5a719fcf
This wasn't intended to change the behavior, but it does slightly.
Previously any requests to acquire memory wouldn't block; only
(subsequent) requests for execution would block (if another
caller had already consumed the memory quota). Now the requests
for memory can also also block.
It turns out in a brief test on my computer that soong_zip
runs about 10 to 20% faster with this change than without it.
The final step involving soong_zip decreased from about
3.6 sec to about 3.3 sec in tests on my computer.
When testing the process of re-zipping the contents of
angler-target_files*.zip , the time decreased from about
6.3 sec to about 5.3 sec in tests on my computer, and the
peak memory usage reported by `top` decreased from about
1.5g to 1g
Bug: 64536066
Test: m -j dist showcommands # which runs soong_zip to package everything
Change-Id: I0422e4c363c675bb7a93309fac4616c768bfbe8f
