[jhc] failure with precompiled libs?

[John Meacham]

On Wed, Feb 07, 2007 at 03:35:50PM -0800, Stefan O'Rear wrote:
> On Wed, Feb 07, 2007 at 03:26:42PM -0500, Isaac Dupree wrote:
> > Stefan O'Rear wrote:
> > > (to change RTS options - turns out I don't have enough RAM to build the
> > > libraries without +RTS -c, and even with it's a tight squeeze)
> > 
> > How much RAM is this amount that worked as a "tight squeeze" for you?
> > That is, I'd like some numbers for how much memory jhc needs (One time I
> > tried this and didn't have enough memory on my current computer; I don't
> > remember if I knew to try +RTS -c)
> 
> 384 MB.
> 
> by which I mean Jhc stayed above 5% cpu the whole time. (Jhc's VIRT stayed
> around 495MB).
> 
> also my limited patience has prevented me from seeing the whole-program
> optimization phase ...
> 
> Jhc is unusably (read: untestably if/when I break something) slow on my
> obsolete computer.  My first hacking target will be some kind of
> -fsymbol-at-a-time use less ram option :)

once the libraries are built, then things go much smoother. I have
basically done very little work on optimizing jhc's memory usage other
than not doing anything blatently silly (I hope). There is most likely a
lot of room for improvement and probably some low-hanging fruit that can
be done easily with a little profiling. 

There are also some obvious things that could be done, like storing the
dependency graph in the binary file, so I need not go through everything
and calculate free variables again on loading, or having a separate
section that doesn't need to be loaded for functions that we know are so
big they will never be inlined.

coming up with a unified annotation to store strictness,cpr, and eta
information and serializing it rather than recalculating it would also
be a big win. (I am working on this actually)

        John

-- 
John Meacham - ⑆repetae.net⑆john⑈