Is there an inherent reason Value Propagation is so memory hungry? It might
be an independently useful improvement to improve the memory usage of Value
Propagation so that it isn't the high water mark.
-- Sean Silva
On Fri, Dec 12, 2014 at 5:16 AM, Joerg Sonnenberger <joerg at
britannica.bec.de> wrote:
>
> Hi all,
> for various reasons, I am limiting clang during builds in pkgsrc to 2GB
> address space. Given that e.g. ARM boards typically have only 1 or 2GB,
> that's a very practical limit. There are a number of builds that fail
> because clang hits that limit and for the majority, they all fail in the
> Value Propagation pass. I'd like to have an option for more selectively
> dealing with this -- -O0 is a very blunt hammer.
>
> (a) Is it possible to restrict the memory use and the associated CPU
> time use at least for -O2? The problematic cases also tend to be very
> slow, 5min+ builds are seen on a E3 Xeon.
>
> (b) Add an internal flag and go with -mllvm -disable-value-propagation.
> This is a relatively small patch, but I dislike the dependency on the
> implementation details.
>
> (c) Add a proper flag for clang and associated PM option.
>
> Joerg
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20141212/284044db/attachment.html>