Re: Bogus 'not declared' warning

[Josh Triplett <josh@freedesktop.org>]

[-- Attachment #1: Type: text/plain, Size: 2360 bytes --]

Josh Triplett wrote:
> Matthew Wilcox wrote:
>> Seems to me that sparse ignores 'static' forward declarations, leading
>> to false warnings like this:
>>
>> /home/willy/kernel/linux-2.6/drivers/scsi/sym53c8xx_2/sym_hipd.c:3786:5: warning: symbol 'sym_compute_residual' was not declared. Should it be static?
>>
>> $ grep -n sym_compute_residual drivers/scsi/sym53c8xx_2/*
>> drivers/scsi/sym53c8xx_2/sym_hipd.c:61:static int sym_compute_residual(struct sym_hcb *np, struct sym_ccb *cp);
>> drivers/scsi/sym53c8xx_2/sym_hipd.c:3033:               cp->sv_resid = sym_compute_residual(np, cp);
>> drivers/scsi/sym53c8xx_2/sym_hipd.c:3786:int sym_compute_residual(struct sym_hcb *np, struct sym_ccb *cp)
> 
> Interesting; yes, it looks like the routine emitting that warning
> doesn't check for a symbol marked static by having a previous static
> declaration.  That warning comes from check_duplicates in evaluate.c.
> Seeing it means Sparse didn't see any previous declaration of the
> symbol by checking the same_symbol linked list, which seems wrong.
> check_declaration in symbol.c hooks symbols into those lists.  That
> implies 1) the scopes don't match and 2) one or the other symbol
> doesn't have extern.  The latter should clearly hold true (that case,
> IIRC, handles letting you put extern declarations inside inner
> scopes).  I think the former occurs because one declaration has file
> scope and the other one global scope, but that has a chicken-and-egg
> issue in the case of a static forward declaration: the static forward
> declaration applies, and makes the later declaration have file scope.
> 
> I *think* the right fix involves changing check_declaration to check
> for this case specifically:
> if (next->scope == file_scope && sym->scope == global_scope) {
> 	sym->scope = file_scope;
> 	sym->same_symbol = next;
> 	return;
> }

Turns out this will take a bit more work to fix.  check_declaration
seems like the wrong place for this check, and in any case the code
above won't properly hook the symbol into its scope since it doesn't
call bind_symbol.  bind_symbol in turn has some dead code and other
problems.  I don't plan to address this for the 0.4.1 release,
other than by adding a test case for this, because I don't want to
further delay the other bugfixes in 0.4.1.

- Josh Triplett


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 252 bytes --]