GHC on IA-64

First of all I'd like to note that a registerised build of ghc-6.4.1
"Just Works"tm. I didn't need any special hacks. This is excellent.

This registerised ghc build was able to successfully recompile itself.
It took about 2 hours (dual cpu Itanium 2 @ 900MHz, make -j4 - thanks to
certain Gentoo developers for letting me use their ia64 box!)

I only needed one patch to get parallel make to work. I darcs sent the
patch ("mkDerivedConstants.c depends on ghcplatform.h") to the ghc-cvs
list but it has not come through just yet. It seems that the patch had a
very large context and so the email went over the limit of 40KB and so
got held for moderation. Perhaps it's time to darcs tag/optimise or
increase the size limit of emails that can be sent to the list.

Anyway...

I used this registerised ghc to compile darcs 1.0.6. I ran into one
problem while compiling SHA1.lhs:

ghc  -cpp  -package QuickCheck -package util -package parsec -O
-funbox-strict-fields  -Wall -Werror -I. -DHAVE_CURSES -DHAVE_CURL
-no-auto-all -funfolding-use-threshold20 -c SHA1.lhs
Prologue junk?:         .proc s64t_ret#
s64t_ret:
         .save ar.pfs, r107
         alloc r107 = ar.pfs, 0, 77, 8, 0
         .body

The source code in SHA1.lhs is fairly heavily unrolled for performance
and the darcs build system is using an extra optimisation flag for this
file -funfolding-use-threshold20 (in addition to -O
-funbox-strict-fields).

http://abridgegame.org/cgi-bin/darcs.cgi/darcs/SHA1.lhs?c=annotate

The error looks as though it's coming from the mangler.

It turns out that just omitting the -funfolding-use-threshold20 flag
allows it to compile ok. My suspicion is that with all this heavy
unrolling, the code is using and probably spilling lots of registers and
that this is enough to upset the mangler on ia64.

Now I don't particularly expect anyone to fix this bug since there is a
workaround and there don't seem to be any IA-64/GHC experts about any
more. I just thought I'd mention all this stuff in case anyone finds it
useful in future.

Of course if anyone thinks they can see a fix I'd be happy to test it.
If anyone wants to look into it, I could narrow down the threshold of
the -funfolding-use-threshold flag at which the bug is triggered and
compare the assembly output.

Duncan