RE: Compiling ghc 6.4 with ghc 5.04.3

On 06 July 2005 10:50, Axel Simon wrote:

> On Wed, 2005-07-06 at 10:06 +0100, Simon Peyton-Jones wrote:
>> That's weird.  We compile 6.4 with 5.04.3 every night.
>> 
>> Maybe you are using different flags than we are?
>> 
>> - can you send your build.mk
> 
> I haven't changed anything, i.e. no build.mk.
> 
>> - plus a full transcript so we can see what flags are being passed to
>> ghc (e.g. -O2)
> 
> below is the full last invocation of the link stage that fails (stage1
> compiler), attached is that invocation redone with -v
> 
>> Is your source tree for 6.4
>>      - the STABLE branch, or
>>      - the exact bits for 6.4 downloaded from the download site?
> 
> It's the tar ball from the web-site.
> 
> Under /lib there is libc-2.2.5.so and libc.so.6. Is there any other
> prerequisite that this Debian machine might not fulfil?

I just built both a clean 6.4 tree and the current STABLE branch with
the 5.04.3 we have installed here, no problems at all.

It's probably a subtle bug in 5.04.3, but there's not much we can do
without a way to reproduce, and since it's in an old version of GHC the
value of tracking down the bug is reduced (it may or may not still be
present).

You can work around it by compiling stage1 without the native code
generator: add GhcWithNativeCodeGen=NO to mk/build.mk, and then remove
it before compiling stage 2.  Other workarounds may work; eg. compiling
AsmCodeGen without optimisation (you have to hack ghc/compiler/Makefile
to do this).  Or simply compile the whole of stage1 without
optimisation: add GhcStage1HcOpts+=-Onot to mk/build.mk.

Cheers,
        Simon