Success on DEC OSF/1 4.0 egcs 1.02

Success on DEC OSF/1 4.0 egcs 1.02

[Patrick Schemitz]

Hello,

in ./egcs-1.0.2/INSTALL/finalinstall.html you wrote:

> If you don't mind, please send egcs@cygnus.com a short mail message
> indicating that you successfully built and installed egcs.
> Include the output from running srcdir/config.guess. 

Well, egcs 1.0.2 compiled fine on our DEC Alphastation 500 running OSF/1
Digital UNIX 4.0, and it seems to compile things just fine. 

(Besides the fact that it's core usage grows up to the hilarious amount
of 127 MB when compiling through some --very nasty-- template instantiation
and then produces assembler output that induces an assembler  warning saying
"line too long", followed by a nearly 18 MB core dump of the DEC Assembler.
But looking at the source file, I guess it's not actually egcs' fault. The
file at hand is really obfuscated. Here's some example:

template class map<int, multiset<ZTrackStub,less<ZTrackStub> >, less<int> >;
template class pair<int const, multiset<ZTrackStub,less<ZTrackStub> > >;
template class rb_tree<int,
                       pair<int const,multiset<ZTrackStub,less<ZTrackStub> > >,
                       select1st<pair<int
                       const,multiset<ZTrackStub,less<ZTrackStub> > >, int>,
                       less<int> >;

I just want to note that this code is *not* written by me :-)



Here's the config.guess output:

alphaev56-dec-osf4.0b

Cheers,	patrick


--
Patrick Schemitz, schemitz@ekp.physik.uni-karlsruhe.de
University of Karlsruhe (TH), Germany
Faculty for Physics, EKP, Ka'CDF Tracking Group

Re: Success on DEC OSF/1 4.0 egcs 1.02

[Joe Buck]

> Well, egcs 1.0.2 compiled fine on our DEC Alphastation 500 running OSF/1
> Digital UNIX 4.0, and it seems to compile things just fine. 
> 
> (Besides the fact that it's core usage grows up to the hilarious amount
> of 127 MB when compiling through some --very nasty-- template instantiation
> and then produces assembler output that induces an assembler  warning saying
> "line too long", followed by a nearly 18 MB core dump of the DEC Assembler.
> But looking at the source file, I guess it's not actually egcs' fault.

Well, it is in a way: the name mangling scheme wasn't really designed
for templates.  And that memory use is still too high, even for complex
templates.

There is a new, experimental mangling scheme called "squangle" (a
combination of "squish" and "mangle") that is much more compact for
STL-style template names; in the snapshots it is turned on by -fsquangle.
Problem: you'd need to build libstdc++ with this switch as well, and
the binutils would need to be updated to understand it.

(but -fsquangle is only in the snapshots, not 1.0.2)

Re: Success on DEC OSF/1 4.0 egcs 1.02

[Jeffrey A Law]

  In message < 199804011631.IAA29108@atrus.synopsys.com >you write:
  > Well, it is in a way: the name mangling scheme wasn't really designed
  > for templates.  And that memory use is still too high, even for complex
  > templates.
I'd bet the memory problem is the haifa memory exposion problem we
recently fixed in the mainline sources.


Patrick -- can you try to build with

-fno-schedule-insns -fno-schedule-insns2

And see if that reduces the memory consumption by an order of magnitude?

jeff