WAW dependency messages

WAW dependency messages

[Jesse Barnes]

In case anyone's interested, I just found a gcc bugzilla bug for the
dependency violation stuff that always gets printed when we compile the
kernel.  I've added some information to it.  The bug number is 12287,
and you can see the whole thing at
http://gcc.gnu.org/bugzilla/show_bug.cgi?id\x12287.

Jesse

Re: WAW dependency messages

[H. J. Lu]

On Wed, Oct 15, 2003 at 03:24:34PM -0700, Jesse Barnes wrote:
> In case anyone's interested, I just found a gcc bugzilla bug for the
> dependency violation stuff that always gets printed when we compile the
> kernel.  I've added some information to it.  The bug number is 12287,
> and you can see the whole thing at
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id\x12287.
> 

I am working on some DV problems in assembler. With the new assembler,
there are fewer DV warnings when compiling 2.4.21 kernel. If someone
sends me small asm testcases, I will look into them.


H.J.

Re: WAW dependency messages

[Jesse Barnes]

On Wed, Oct 15, 2003 at 03:53:35PM -0700, H. J. Lu wrote:
> On Wed, Oct 15, 2003 at 03:24:34PM -0700, Jesse Barnes wrote:
> > In case anyone's interested, I just found a gcc bugzilla bug for the
> > dependency violation stuff that always gets printed when we compile the
> > kernel.  I've added some information to it.  The bug number is 12287,
> > and you can see the whole thing at
> > http://gcc.gnu.org/bugzilla/show_bug.cgi?id\x12287.
> > 
> 
> I am working on some DV problems in assembler. With the new assembler,
> there are fewer DV warnings when compiling 2.4.21 kernel. If someone
> sends me small asm testcases, I will look into them.

There's a testcase in the bug, but what prompted me to look was all of
the messages that get printed out when compiling a 2.6 kernel.  Does
binutils 2.14.90.0.6 contain the fixes you're talking about?  That's
what I'm using...

Thanks,
Jesse

Re: WAW dependency messages

[H. J. Lu]

On Wed, Oct 15, 2003 at 04:14:16PM -0700, Jesse Barnes wrote:
> On Wed, Oct 15, 2003 at 03:53:35PM -0700, H. J. Lu wrote:
> > On Wed, Oct 15, 2003 at 03:24:34PM -0700, Jesse Barnes wrote:
> > > In case anyone's interested, I just found a gcc bugzilla bug for the
> > > dependency violation stuff that always gets printed when we compile the
> > > kernel.  I've added some information to it.  The bug number is 12287,
> > > and you can see the whole thing at
> > > http://gcc.gnu.org/bugzilla/show_bug.cgi?id\x12287.
> > > 
> > 
> > I am working on some DV problems in assembler. With the new assembler,
> > there are fewer DV warnings when compiling 2.4.21 kernel. If someone
> > sends me small asm testcases, I will look into them.
> 
> There's a testcase in the bug, but what prompted me to look was all of

That testcase should be fixed in the next binutils.

> the messages that get printed out when compiling a 2.6 kernel.  Does

Kernel is a totally different issue. You can send me small asm
testcases. I will see what I can do.

> binutils 2.14.90.0.6 contain the fixes you're talking about?  That's
> what I'm using...
> 

I am working on 2.14.90.0.7 which should have the fix.


H.J.

Re: WAW dependency messages

[Jesse Barnes]

On Wed, Oct 15, 2003 at 04:51:44PM -0700, H. J. Lu wrote:
> That testcase should be fixed in the next binutils.

Great!

> Kernel is a totally different issue. You can send me small asm
> testcases. I will see what I can do.

Ok.

> I am working on 2.14.90.0.7 which should have the fix.

Ok, I'll check that out when you release it and try to put together some
test .s files if I see more failures.

Thanks,
Jesse

RE: WAW dependency messages

[Chen, Kenneth W]

> Kernel is a totally different issue. You can send me small asm
> testcases. I will see what I can do.

Something as simple as:

        itc.i r18
        ;;
        ld8 r19=[r17]

Will generate
aa.S: Assembler messages:
aa.S:3: Warning: Use of 'ld8' violates RAW dependency 'DTC' (data)
aa.S:1: Warning: This is the location of the conflicting usage

Oddly enough, "as test.S" doesn't have warning message, but "as -x
test.S" does.  gcc is execve with "as -x" which triggers the warning
message.

- Ken

Re: WAW dependency messages

[David Mosberger]

>>>>> On Wed, 15 Oct 2003 16:51:44 -0700, "H. J. Lu" <hjl@lucon.org> said:

  HJ> Kernel is a totally different issue. You can send me small asm
  HJ> testcases. I will see what I can do.

Here is one that bugs me (adapted from fsys.S):

$ cat t.S
        mov psr.l = r2;;
        mov r25=ar.rnat
$ as -x t.S
/r/wailua/tmp/t.S: Assembler messages:
/r/wailua/tmp/t.S:3: Warning: Use of 'mov.m' violates RAW dependency 'PSR.si' (data)
/r/wailua/tmp/t.S:1: Warning: This is the location of the conflicting usage
$ as --version |head -1
GNU assembler 2.13.90 20030425

I think the warnings shows up because the assembler doesn't detect
that we're reading ar.rnat, not ar.itc.

	--david

RE: WAW dependency messages

[Jim Wilson]

On Wed, 2003-10-15 at 18:34, Chen, Kenneth W wrote:
> Oddly enough, "as test.S" doesn't have warning message, but "as -x
> test.S" does.  gcc is execve with "as -x" which triggers the warning
> message.

GNU as only does DV checking if the -x option is specified.

The DV checking stuff was never properly finished.  It should be enabled
by default.  And we should emit errors instead of warnings for provable
DVs.  However, if we did that, we would then have the problem that false
positives from bugs would prevent the compiler from working.  So we
would have to add an option to disable DV checking and have the compiler
emit it.  But then we would have the problem that the compiler output
would not get any DV checking from the assembler, which is a worse
problem than what we have now.  We really need to get the bugs fixed
first, and then emit errors by default for DVs.  I'm glad to see that HJ
has time to work on these problems.
-- 
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com