Re: g++ versions

[Tim Prince <TimothyPrince@sbcglobal.net>]

Robert Jones wrote:

> I'm unclear what you mean by the Cygwin install menu - is that the setup
> program?
Yes, cygwin setup menu has included g++-4 and its cousins for several months.
> 

> My aim is simply to arrive at fairly recent (ie version 4) g++ compiler for
> Windows, either a binary download or one I build myself, but the build
> reports seem to suggest that all the v4 windows build attempts have many
> failures.

Well, let's be more specific.  Windows targets are at best secondary for
gcc.  This means the gcc developers have latitude about the degree of
Windows support in testsuite for each front end.
Out of 13 failures in the g++ section of gcc-testsuite for cygwin in last
Friday's gcc-4.4 snapshot, 9 are due to minor formatting discrepancies
between the linux and Windows versions of .s file.  Let's give the g++
people some credit for major progress.
Most of the libstdc++ testsuite failures are due to the broken scheme in
testsuite of attempting to link against libiconv.  Many people don't care
about libiconv, which probably bears on why it isn't fixed.
The gcc people don't care to make testsuite entirely meaningful on
Windows.  In addition to testing functionality which will never be
supported on Windows, without marking those tests as XFAIL, and not
allowing the bsd variations of tests rather than linux ones to be used for
Windows, there are major categories of Windows failures, e.g. pch,
vectorization (even though vectorization basically works).  The greatest
gross number of failures I have seen in repetitions of gcc testsuite is
177, out of more than 20000.  If you prefer a version of gcc which hasn't
been run through testsuite publically, be my guest.
The gfortran tests which fail are all OS-specific, and have been discussed
on the gfortran list.
objc testsuite gives the same results as on most primary targets.
gomp testsuite, if enabled, doesn't give many errors.  That doesn't
necessarily translate into a reliable option.  After all, it's built on
pthreads without much specific Windows stuff, aside from the proprietary
mingw offering with an alternate library of unknown origin or licensing
compliance.
mingw versions of gcc often have major gaps in exception handling,
fast-math, profiling, and thread support, so there are many expected
testsuite failures.  That doesn't make them unusable, within their limits.
You certainly can't expect the testsuite cases which test for presence of
glibc specific directories to pass on Windows.
I'll post this week's gcc-testsuite for cygwin soon.  Have to decide
whether to let it complete the extra rounds it chooses to make when there
are failures.
Do you care about work-arounds for building gcc on cygwin?