Bug report optimization #4186 and ia64

Bug report optimization #4186 and ia64

[Willi Nüßer]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 854 bytes --]

Hi,

some weeks ago we posted a bug report (optimization/#4186)
on bug-gcc.
For us this bug - although it is "only" in the opt case -
is critical. Without a solution we can not use gcc 3.* for our
development.
 
Now we face a similar optimization problem on ia64 as well
(gcc 3.0, 3.01, snapshots ...).
Again, if and only if we compile with -O2 our app crashes with SIGSEGV.
And again, it seems that some registers get filled with
incorrect values. Unfortunately we are still working
to create a _small_ testcase. 

Does anybody of the gcc gurus know if there is some work
on the register allocator etc going on? Is it possible that
these two gcc bugs are related to the same high level mechanism?

We need a fix really badly. Is it possible to fix this for 3.0.2?

Best regards
Willi Nuesser
SAP Linuxlab

PS:
Please CC me since I´m not on the list.

Re: Bug report optimization #4186 and ia64

[Jim Wilson]

In article < 3BAF3CB0.7CAB5EC9@sap.com > you write:
>some weeks ago we posted a bug report (optimization/#4186)
>on bug-gcc.

>Now we face a similar optimization problem on ia64 as well
>(gcc 3.0, 3.01, snapshots ...).

I took a look at PR 4186.  It is an x86 sibling-call optimization bug.
We accidentally overwrite one of the incoming function arguments before
reading it in the sibling call sequence.  I put this info into the PR.

I think it is unlikely that the IA-64 problem is related, since the
IA-64 does not pass arguments on the stack like the x86 does.

Jim

RE: Bug report optimization #4186 and ia64

[Nuesser, Wilhelm]

Hi,

> -----Original Message-----
> From: Jim Wilson [ mailto:wilson@cygnus.com ]
  
> I took a look at PR 4186.  It is an x86 sibling-call optimization bug.
> We accidentally overwrite one of the incoming function 
> arguments before
> reading it in the sibling call sequence.  I put this info into the PR.

Thanks a lot for your quick response. We will re-build or entire
app asap and check this.

> I think it is unlikely that the IA-64 problem is related, since the
> IA-64 does not pass arguments on the stack like the x86 does.

You're right. We checked this with -fno-optimize-sibling-calls and
- independently - with -fno-strict-aliasing as H.J. Lu suggested.
Neither of them showed any effect.

We'll have to look futher into a testcase.

Thanks
Willi