[gentoo-amd64] Re: Unexpected side effect of GCC 4

[Duncan]

Peter Humphrey <prh@××××××××××.uk> posted
200611051210.25023.prh@××××××××××.uk, excerpted below, on  Sun, 05 Nov
2006 12:10:24 +0000:

[rewrapped]

> $ emerge --info | grep FLAGS

> CFLAGS="-march=k8 -Os -pipe -frename-registers -fweb -freorder-blocks
> -freorder-blocks-and-partition -funit-at-a-time -fgcse-sm -fgcse-las
> -fgcse-after-reload -fmerge-all-constants"

> CXXFLAGS="-march=k8 -Os -pipe -frename-registers -fweb -freorder-blocks
> -funit-at-a-time -fgcse-sm -fgcse-las -fgcse-after-reload
> -fmerge-all-constants"

> Unset:  CTARGET, INSTALL_MASK, LDFLAGS, PORTAGE_RSYNC_EXTRA_OPTS
 
> So I have a pointer away from glibc and linux-headers. CFLAGS on the new
> system are as on the old but with -combine added.The old and new systems
> have the same versions of kernel, linux-headers and glibc, but different
> versions of gcc. Rsync is crippled on the old system but fine on the new
> one.

Well, I /did/ say gcc 4.1.1 seemed way (as in, vastly) better than 3.4.x,
for me, so at least that's being borne out.

As for CFLAGS, the only change I've made recently is adding
-ftree-vectorize, experimentally, given the discussion here.  I'm not sure
it's the problem, particularly as where I /am/ experiencing issues the
first thing I tried was compiling stuff without it, so if it's the
problem, it's in some obscure dependency somewhere, but I certainly had no
issues before trying it and now I do, so for anyone else thinking of
trying it, I'd recommend staying well away from -ftree-vectorize for the
time being.  I'd really like to be able to say for sure it is the problem,
and find it very frustrating not to be able to nail it down, but suffice
it to say /something's/ causing me problems right now and that's one of my
recent changes, so I'd recommend staying well away from it.

I mentioned that I need to do an emerge -emptytree again, as it has been
awhile.  I'm still debating whether to try it with or without
-ftree-vectorize.  If that's the problem, it may well go away if
everything is compiled with it.  OTOH, it may not.  That's an awful lot of
compiling to do and still have a problem if -ftree-vectorize is causing it
and it doesn't go away with everything compiled with it, but OTOH, if
that's /not/ the problem, and I decide not to try it, I'll never know
whether that was the problem or not.

I'm about to decide to simply play it safe and pretend nobody here ever
called -ftree-vectorize to my attention, at least until the flag has had a
bit more time to mature (say gcc 4.2 or 4.3).  Only just because I'm the
type of person I am, that would bother me as I will have never nailed it
for sure.  In any case, my best guess right now are that the issues I am
having are related to a nasty interplay between -ftree-vectorize and
glibc-2.5.  In looking around at Gentoo glibc bugs, I found at least one
equally strange one, on x86.  The bug required glibc 2.5, and it required
that a certain package be compiled with a certain cflag.  With that cflag
on glibc 2.4, it worked fine, as it did without the cflag on 2.5, but put
them together and things blew up.  It wasn't -ftree-vectorize, and it was
on x86, but I've a rather strong suspicion my case is similar, only on
amd64 and with a different cflag, -ftree-vectorize most likely.  I've just
not figured out which particular package is doing it, and not being able
to figure it out as I said is VERY annoying to me, WAY more so than the
bit of inconvenience the actual bug is causing.  (Yes, I know I said all
that already, but it's still annoying and I'm still griping about it! =8^(

-- 
Duncan - List replies preferred.   No HTML msgs.
"Every nonfree program has a lord, a master --
and if you use the program, he is your master."  Richard Stallman

-- 
gentoo-amd64@g.o mailing list