Skip to contentSkip to conversation

huge gcc 4.1.{0,1} __weak problem

DISCUSSION Chris Knadle 7 participants 10 messages May 2, 2008

lore
Share & tools
lore
#0 lore raw
Quoted 10 lines · Linus Torvalds
Quoted 8 lines · Andrew Morton
Quoted 6 lines

I see only the following choices:
- remove __weak and replace all current usages
- move all __weak functions into own files, and ensure that also happens for future usages
- #error for gcc 4.1.{0,1}

Can we detect the {0,1}? __GNUC_EVEN_MORE_MINOR__?

Quoted 8 lines

It's __GNUC_PATCHLEVEL__, I believe.

So yes, we can distinguish 4.1.2 (good, and very common) from 4.1.{0,1} (bad, and rather uncommon).
And yes, considering that 4.1.1 (and even more so 4.1.0) should be rare to begin with, I think it's better to just not support it.

Linus

Unfortunately Debian Stable (i.e. Etch), which is relatively popular for server
use, is still using 4.1.1 :-( (The current gcc package is gcc-4.1.1-21)

I have not looked to see if Debian Stable's gcc-4.1.1-21 has been patched for
the currently discussed __weak bug.

-- Chris

Chris Knadle
[email protected]

#1 lore raw
Quoted 13 lines · Chris Knadle
Quoted 12 lines · Linus Torvalds
Quoted 8 lines · Andrew Morton
Quoted 6 lines

I see only the following choices:
- remove __weak and replace all current usages
- move all __weak functions into own files, and ensure that also happens for future usages
- #error for gcc 4.1.{0,1}

Can we detect the {0,1}? __GNUC_EVEN_MORE_MINOR__?

It's __GNUC_PATCHLEVEL__, I believe.

Quoted 7 lines
Quoted 4 lines

So yes, we can distinguish 4.1.2 (good, and very common) from 4.1.{0,1} (bad, and rather uncommon).
And yes, considering that 4.1.1 (and even more so 4.1.0) should be rare to begin with, I think it's better to just not support it.

Unfortunately Debian Stable (i.e. Etch), which is relatively popular for server
use, is still using 4.1.1 :-( (The current gcc package is gcc-4.1.1-21)

Well the package version string is 4.1.1, but the compiler thinks it's 4.1.2, at least on i386, amd64 and ia64:

$ cc -v
gcc version 4.1.2 20061115 (prerelease) (Debian 4.1.1-21)

A small test program agrees:

        printf("%d\n", __GNUC_PATCHLEVEL__);
$ ./a.out 
2

Mike.

#2 lore raw
In reply to Chris Knadle

Hi Chris,

(I fixed the corrupted CC and Reply-to: address from your email.)

Quoted 17 lines · On Friday 02 May 2008 00:55:58 Chris Knadle wrote:
Quoted 16 lines · Linus Torvalds
Quoted 7 lines · Andrew Morton
Quoted 5 lines

I see only the following choices:
- remove __weak and replace all current usages
- move all __weak functions into own files, and ensure that also happens for future usages
- #error for gcc 4.1.{0,1}

Can we detect the {0,1}? __GNUC_EVEN_MORE_MINOR__?

It's __GNUC_PATCHLEVEL__, I believe.

So yes, we can distinguish 4.1.2 (good, and very common) from 4.1.{0,1} (bad, and rather uncommon).
And yes, considering that 4.1.1 (and even more so 4.1.0) should be rare to begin with, I think it's better to just not support it.

Quoted 8 lines
Quoted 1 line

Linus

Unfortunately Debian Stable (i.e. Etch), which is relatively popular for
server use, is still using 4.1.1 :-( (The current gcc package is gcc-4.1.1-21)

I have not looked to see if Debian Stable's gcc-4.1.1-21 has been
patched for the currently discussed __weak bug.

I checked and it has been patched in 4.1.1-21. This would make checking for 4.1.1 via __GNUC_PATCHLEVEL__ potentially invalid, as patched distro compilers may (and in this case do) have this fixed.

Signature · 4 lines
Cheers,
Alistair.

137/1 Warrender Park Road, Edinburgh, UK.
#3 lore raw
Quoted 28 lines · Alistair John Strachan

Hi Chris,

(I fixed the corrupted CC and Reply-to: address from your email.)

Quoted 23 lines · On Friday 02 May 2008 00:55:58 Chris Knadle wrote:
Quoted 9 lines · Linus Torvalds
Quoted 7 lines · Andrew Morton
Quoted 5 lines

I see only the following choices:
- remove __weak and replace all current usages
- move all __weak functions into own files, and ensure that also happens for future usages
- #error for gcc 4.1.{0,1}

Can we detect the {0,1}? __GNUC_EVEN_MORE_MINOR__?

Quoted 8 lines

It's __GNUC_PATCHLEVEL__, I believe.

So yes, we can distinguish 4.1.2 (good, and very common) from 4.1.{0,1} (bad, and rather uncommon).
And yes, considering that 4.1.1 (and even more so 4.1.0) should be rare to begin with, I think it's better to just not support it.

Linus

Unfortunately Debian Stable (i.e. Etch), which is relatively popular for
server use, is still using 4.1.1 :-( (The current gcc package is gcc-4.1.1-21)

Quoted 6 lines
Quoted 2 lines

I have not looked to see if Debian Stable's gcc-4.1.1-21 has been
patched for the currently discussed __weak bug.

I checked and it has been patched in 4.1.1-21. This would make checking for 4.1.1 via __GNUC_PATCHLEVEL__ potentially invalid, as patched distro compilers may (and in this case do) have this fixed.

Is it possible to cook up a small sample file we could build as part of the kernel build. If it fails => error out.
If someone comes up with the code I shall try to integrate it in the build system.

Sam

#4 lore raw

On Friday 02 May 2008 11:43:48 Sam Ravnborg wrote:
[snip]

Quoted 16 lines
Quoted 16 lines
Quoted 15 lines
Quoted 8 lines

It's __GNUC_PATCHLEVEL__, I believe.

So yes, we can distinguish 4.1.2 (good, and very common) from 4.1.{0,1} (bad, and rather uncommon).
And yes, considering that 4.1.1 (and even more so 4.1.0) should be rare to begin with, I think it's better to just not support it.

Linus

Unfortunately Debian Stable (i.e. Etch), which is relatively popular
for server use, is still using 4.1.1 :-( (The current gcc package is gcc-4.1.1-21)

I have not looked to see if Debian Stable's gcc-4.1.1-21 has been
patched for the currently discussed __weak bug.

Quoted 8 lines
Quoted 3 lines

I checked and it has been patched in 4.1.1-21. This would make checking for 4.1.1 via __GNUC_PATCHLEVEL__ potentially invalid, as patched distro compilers may (and in this case do) have this fixed.

Is it possible to cook up a small sample file we could build as part of the kernel build. If it fails => error out.
If someone comes up with the code I shall try to integrate it in the build system.

The GCC PR has a test case for this regression which might be usable.

http://gcc.gnu.org/viewcvs/branches/gcc-4_1-branch/gcc/testsuite/gcc.dg/tree-ssa/pr27781.c?view=markup&pathrev=114852

Signature · 4 lines
Cheers,
Alistair.

137/1 Warrender Park Road, Edinburgh, UK.
#5 lore raw
Quoted 22 lines · Alistair John Strachan

On Friday 02 May 2008 11:43:48 Sam Ravnborg wrote:
[snip]

Quoted 20 lines
Quoted 19 lines
Quoted 7 lines
Quoted 7 lines

It's __GNUC_PATCHLEVEL__, I believe.

So yes, we can distinguish 4.1.2 (good, and very common) from 4.1.{0,1} (bad, and rather uncommon).
And yes, considering that 4.1.1 (and even more so 4.1.0) should be rare to begin with, I think it's better to just not support it.

Quoted 8 lines
Quoted 1 line

Linus

Unfortunately Debian Stable (i.e. Etch), which is relatively popular
for server use, is still using 4.1.1 :-( (The current gcc package is gcc-4.1.1-21)

I have not looked to see if Debian Stable's gcc-4.1.1-21 has been
patched for the currently discussed __weak bug.

I checked and it has been patched in 4.1.1-21. This would make checking for 4.1.1 via __GNUC_PATCHLEVEL__ potentially invalid, as patched distro compilers may (and in this case do) have this fixed.

Quoted 8 lines
Quoted 4 lines

Is it possible to cook up a small sample file we could build as part of the kernel build. If it fails => error out.
If someone comes up with the code I shall try to integrate it in the build system.

The GCC PR has a test case for this regression which might be usable.

http://gcc.gnu.org/viewcvs/branches/gcc-4_1-branch/gcc/testsuite/gcc.dg/tree-ssa/pr27781.c?view=markup&pathrev=114852

OK, can anyone confirm that this fails to build which a buggy gcc:

void __attribute__((weak)) func(void)

{
    /* no code */
}

int main()

{
    func();
    return 0;
}

Sam

#6 lore raw
Quoted 9 lines · Sam Ravnborg

OK, can anyone confirm that this fails to build which a buggy gcc:

void __attribute__((weak)) func(void)

{
    /* no code */
}
Quoted 5 lines

int main()

{
    func();
    return 0;
}

Of course it doesn't fail to build. With buggy gcc it will be optimized to
void __attribute__((weak)) func (void)
{
}

int main ()
{
return 0;
}

(similarly how all recent gccs optimize this without the weak attribute) while non-buggy gcc keeps the func call.
So, you either need to grep the assembly (that's what e.g. the GCC testcase does), or you can e.g. use a runtime testcase:
extern void abort (void);
void __attribute__((weak)) func (void) { }
int main () { func (); abort (); }
in one compilation unit and
extern void exit (int);
void func (void) { exit (0); }
in another one. I doubt a runtime testcase is acceptable though for the kernel, as the cross compiler used to build the kernel might not be able to create userland executables (missing C library, etc.).

Jakub

#7 lore raw
Quoted 17 lines · On Friday 02 May 2008 15:11:32 Jakub Jelinek wrote:
Quoted 9 lines · Sam Ravnborg

OK, can anyone confirm that this fails to build which a buggy gcc:

void __attribute__((weak)) func(void)

{
    /* no code */
}
Quoted 5 lines

int main()

{
    func();
    return 0;
}

Of course it doesn't fail to build.

[snip]

Quoted 3 lines

 I doubt a runtime testcase is acceptable though for the kernel, as the cross compiler used to build the kernel might not be able to create userland executables (missing C library, etc.).

I assume the GCC testsuite has the same generic problem, which is probably why it uses the -fdump-tree-optimized parameter to gcc. For example:

alistair@just:~$ cat test.c
void __attribute__((weak)) func(void)
{
    /* no code */
}

int main()

{
    func();
    return 0;
}
alistair@just:~$ gcc -O2 -fdump-tree-optimized test.c
alistair@just:~$ rm -f a.out
alistair@just:~$ cat test.c*.optimized | egrep "func \\(\\);"
  func ();

Ergo, my compiler isn't buggy. As this doesn't require a runtime test I think it would be OK for the kernel.

Of course, whether or not it's worth it is now debatable, given the information Sven-Haegar Koch provided (Debian's GCC version number is actually 4.1.2).

Signature · 4 lines
Cheers,
Alistair.

137/1 Warrender Park Road, Edinburgh, UK.
#8 lore raw
In reply to Sam Ravnborg
Quoted 9 lines · Sam Ravnborg

OK, can anyone confirm that this fails to build which a buggy gcc:

void __attribute__((weak)) func(void)

{
    /* no code */
}
Quoted 6 lines

int main()

{
    func();
    return 0;
}

I think the problem is that main() would have no call to func(), not that it wouldn't build.

Does making func noinline fix it? I wonder if we should make __weak implicitly make the function noinline too.

J

#9 lore raw
In reply to Alistair John Strachan
Quoted 24 lines · Alistair John Strachan
Quoted 23 lines · On Friday 02 May 2008 00:55:58 Chris Knadle wrote:
Quoted 9 lines · Linus Torvalds
Quoted 7 lines · Andrew Morton
Quoted 5 lines

I see only the following choices:
- remove __weak and replace all current usages
- move all __weak functions into own files, and ensure that also happens for future usages
- #error for gcc 4.1.{0,1}

Can we detect the {0,1}? __GNUC_EVEN_MORE_MINOR__?

Quoted 8 lines

It's __GNUC_PATCHLEVEL__, I believe.

So yes, we can distinguish 4.1.2 (good, and very common) from 4.1.{0,1} (bad, and rather uncommon).
And yes, considering that 4.1.1 (and even more so 4.1.0) should be rare to begin with, I think it's better to just not support it.

Linus

Unfortunately Debian Stable (i.e. Etch), which is relatively popular for
server use, is still using 4.1.1 :-( (The current gcc package is gcc-4.1.1-21)

Quoted 6 lines
Quoted 2 lines

I have not looked to see if Debian Stable's gcc-4.1.1-21 has been
patched for the currently discussed __weak bug.

I checked and it has been patched in 4.1.1-21. This would make checking for 4.1.1 via __GNUC_PATCHLEVEL__ potentially invalid, as patched distro compilers may (and in this case do) have this fixed.

Not a problem for the Debian Stable Version, as someone already wrote, it calls itself 4.1.2, even if the package version number is 4.1.1-xx

Quoted 1 line

gcc --version

gcc (GCC) 4.1.2 20061115 (prerelease) (Debian 4.1.1-21) Copyright (C) 2006 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

c'ya
sven

Signature · 2 lines
The Internet treats censorship as a routing problem, and routes around
it. (John Gilmore on http://www.cygnus.com/~gnu/)