[PATCH] D51317: Handle BumpPtrAllocator::Allocate(0) without undefined behavior

[Brent Royal-Gordon via llvm-commits]

> On Sep 4, 2018, at 2:23 AM, Hans Wennborg via Phabricator <reviews at reviews.llvm.org> wrote:
> 
> I'd like to see some kind of benchmark to verify this doesn't have negative performance impact.

This is a little rough because it uses time(1), but here's twenty compiles of oggenc.c (1.7 MB) at -O3 on macOS with as much extraneous stuff closed as possible:

Without change:
      103.81 real       101.95 user         1.48 sys

With change:
      101.31 real        99.55 user         1.46 sys

Additional runs under less controlled conditions (e.g. a web browser open in the background) didn't necessarily show the changed version being faster, but they were always quite close. Test runs the other day with bzip2.c (cited in the original commit) also looked pretty close, but it compiled so fast that I had a tough time getting solid numbers. I had trouble getting gcc.c to compile on macOS.

-- 
Brent Royal-Gordon
Architechies

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180905/e8aab90b/attachment.html>