PATCH - [tree-ssa] fix for PRs 14036 and 14065

Fariborz Jahanian fjahanian@apple.com
Tue Feb 10 00:27:00 GMT 2004


On Feb 9, 2004, at 2:12 PM, Zack Weinberg wrote:

> Fariborz Jahanian <fjahanian@apple.com> writes:
>
> [Please fix your mailer so it word wraps before 80 columns.]
>
>> These PRs were filed when tree-ssa was used to build gap of the SPEC
>> benchmark with IMI.  But bugs also show up without IMI in a single
>> file compile mode. Two tests are attached.
>
> Can you reproduce these failures with mainline?  I do not see what
> they have to do with tree-ssa.

tree-ssa has new routines: expand_used_vars, expand_var which
postpone rtl generation of variable declarations, thus exposing this 
bug.

>
>> Both bugs are related and have to do with declaration or definition
>> of a locally declared 'extern' variable in global scope. Bug results
>> in ICE in the first test and duplicate definition in the second
>> test.  Fix is to preserve the flag in call to merge_decls
>
> ... this would be fine, but the whole concept is slated to disappear;

Can you please give us a timeframe of when this will be available.
We are very interested in tree-ssa compiling gap in IMI, which is
impacted by this.

>
>> and to keep locally declared extern of the old declaration 'extern'
>> after call to merge_decls in pushdecl.
>
> ... and this is flat wrong, as far as I can tell, and would cause
> standards violations once the existing (thoroughly broken) data
> structure is replaced.

I am keeping the tree entry for a previous, not-in-scope declaration
to remain an extern decl.
The new declaration is in no way impacted by this patch. Why is
this violation of the standard rules? Do you have an example?

>
> Please do the following:
>
> 1) Attempt to duplicate these problems on mainline.  Report results.

Can't be done. Please see above.

> 2) Assign PRs 14036 and 14065 to me.

Done.

> 3) Do NOT check anything in at this time.

OK. I will keep my changes local since we need it to build gap.

- Thanks, fariborz

>
> i) if you feel like being helpful, please test the patch in
>    http://gcc.gnu.org/ml/gcc-patches/2004-02/msg00768.html
>    on a system that uses stabs, SDB, or XCOFF debug format by default.
>
> zw



More information about the Gcc-patches mailing list