[llvm-dev] [RFC] Removing Waymarking from Use.
Ehud Katz via llvm-dev
llvm-dev at lists.llvm.org
Fri Apr 3 09:56:54 PDT 2020
I'm in favor for this change.
I've also done some testing myself and noticed only about 1-2% increase of
memory, in exchange for about 1-3% increase of speed.
I can't say that a speedup of 3% (at most, and usually 1%), worth working
for, this is a simple change, that give a lot more than this; especially
simpler code path (also easier debugging).
As part of my measurements, while compiling "deal.II", the maximum number
of simultaneous `Use` instances was 2,728,963 (and in most cases, about
half of it).
In a 64-bit process (Clang for that matter), we will have roughly 21 MB of
memory spent on the extra pointer (of 8 bytes size).
On a 16-core machine, for which, in many cases, a build will trigger that
many simultaneous Clang processes, we will reach, in the worst case, 336 MB
more memory spent on compilation.
Is it worth it? I think it is. But I am not sure I see the whole picture -
are there low-memory systems that need to run LLVM on?
I am not sure what needs to be done to approve such a fundamental change;
especially when we can't prove the Waymarking was needed at all.
On Thu, Apr 2, 2020 at 8:10 PM Chris Lattner via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> On Apr 1, 2020, at 5:51 AM, Tyker1 at outlook.com via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
> Hi llvm-dev,
> I have a patch open for review that removes waymarking
> This patch removes waymarking and replaces it with storing a pointer to
> the User in the Use.
> when compiling the CTMark tests of the test suite, this give an average
> of +1.8% max memory use and -1.1% compile time.
> Removing Waymarking also simplifies the code of Use and User.
> Waymarking is very complicated, and I’d be generally +1 on removing it if
> it is not paying for itself. It is also nice to see compiler speedups!
> I’m curious about kimwitu++ though - when waymarking initially landed, it
> was a significant memory size win for that. Do you have any idea why
> you’re not seeing the same difference? Is it possible that different
> things were being measured back then?
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-dev