[patch][pr22217] Use the most recent decl for mangling

[John McCall]

> On Jan 29, 2015, at 4:16 PM, Richard Smith <richard at metafoo.co.uk> wrote:
> On Thu, Jan 29, 2015 at 4:04 PM, John McCall <rjmccall at apple.com <mailto:rjmccall at apple.com>> wrote:
>> On Jan 29, 2015, at 3:59 PM, Richard Smith <richard at metafoo.co.uk <mailto:richard at metafoo.co.uk>> wrote:
>> On Thu, Jan 29, 2015 at 3:57 PM, Richard Smith <richard at metafoo.co.uk <mailto:richard at metafoo.co.uk>> wrote:
>> On Thu, Jan 29, 2015 at 3:47 PM, John McCall <rjmccall at apple.com <mailto:rjmccall at apple.com>> wrote:
>> 
>>> On Jan 29, 2015, at 2:47 PM, Richard Smith <richard at metafoo.co.uk <mailto:richard at metafoo.co.uk>> wrote:
>>> 
>>> On Fri, Jan 23, 2015 at 12:25 AM, John McCall <rjmccall at apple.com <mailto:rjmccall at apple.com>> wrote:
>>> > On Jan 22, 2015, at 4:52 PM, Rafael Espíndola <rafael.espindola at gmail.com <mailto:rafael.espindola at gmail.com>> wrote:
>>> >
>>> > Sent the email a bit early.
>>> >
>>> >
>>> >>> That is not what I am seeing with gcc. Given
>>> >>>
>>> >>> int pr22217_foo;
>>> >>> int *b = &pr22217_foo;
>>> >>> extern int pr22217_foo __attribute__((section("zed")));
>>> 
>>> This should be an error in both C and C++.  I see absolutely no reason to allow a declaration following a definition (even a tentative definition) to add a section attribute.  We should not be afraid to reject stupidly-written code when it abuses language extensions, even when they’re not “our” extensions.
>>> 
>>> I completely agree with the principle here. It is not reasonable to write attributes that affect a definition after the definition. It is not reasonable to write attributes that affect how a symbol is referenced (such as an asm label) after the first use (and perhaps we should simply require them on the first declaration).
>>> 
>>> (Segue away from attributes and towards tentative definitons follows...)
>>> 
>>> I don't agree with what you said about tentative definitions. The C standard is very clear on the model for tentative definitions: they act exactly like non-defining declarations until you get to the end of the translation unit; if you've not seen a non-tentative definition by that point "then the behavior is exactly as if the translation unit contains a file scope declaration of that identifier, with the composite type as of the end of the translation unit, with an initializer equal to 0.”
>> 
>> So, this is interesting.  Unix C compilers have traditionally defaulted to -fcommon, i.e. to treating uninitialized variables as common definitions that are overridable not just within a translation unit, but within the entire program.  (I’m not sure whether ELF platforms implement this as “program” or “linkage unit”.  Darwin uses “linkage unit”.)  Whether that’s actually compliant is arguable, but regardless, it’s the semantics we use, and so we really do have to maintain the tri-state, because tentative definitions are semantically quite different from non-tentative definitions.
>> 
>> But in the sense that non-tentative definitions fully replace tentative definitions, I agree that the correct behavior is probably to allow a non-tentative definition with a section attribute to “override” a tentative definition which lacks the attribute.
>> 
>> That's reasonable as long as section attributes never affect the code-generation of accesses to an object.  I think we can agree that section attributes that do affect code-generation of references (in an incompatible way) would clearly need to be on all declarations.  But that’s more like an address-space attribute than a section attribute.
>> 
>>> Based on that simple semantic model, it is not reasonable for us to reject this:
>>> 
>>> int pr22217_foo;
>>> int *b = &pr22217_foo;
>>> extern int pr22217_foo __attribute__((section("zed")));
>>> int pr22217_foo = 123;
>>> 
>>> See also PR20688, which is a rejects-valid for standard C11 code due to our being confused about how tentative definitions work.
>>> 
>>> And here's another case we get wrong:
>>> 
>>>   int a[];
>>>   extern int a[5];
>>> 
>>> We're required to emit a definition of 'a' with type 'int[5]', but we emit it with type 'int[1]'. We get the corresponding case with an incomplete struct correct:
>>> 
>>>   struct foo x; // ok, tentative definition
>>>   struct foo { int n, m; };
>>>   // definition emitted now and has complete type; initializer is {0}.
>>> 
>>> There are lots of ways we can fix this; perhaps the easiest one would be to literally follow what the C standard says: synthesize a definition for each tentatively-defined variable at the end of the translation unit. Then we can change isThisDeclarationADefinition to simply return 'bool' instead of an enum, and have it return 'false' for tentative definitions. Sema would track the tentative definitions it's seen, and consider converting each one to a definition at end-of-TU.
>> 
>> Like I mentioned above, this isn’t actually allowed under -fcommon.
>> 
>> I don't see why not. We just need to make sure that the definition we create at the end of TU is emitted as a common definition.
> 
> Ah.  I see what you’re saying.  This does become a channel of out-of-bound information that other clients have to be aware of, but yes, it’s possible.
> 
> Yeah, we were intending on tracking this on VarDecl so that consumers who care can find out that a particular implicit definition was synthesized to represent the definition of a tentatively-defined variable. I contend that very few consumers of the AST will need to be aware of this; fewer than need to be aware of the isThisDeclarationADefinition special case today.

Yeah.  I think this is a reasonable design, and Doug concurs.

>> (There are also a small number of checks we need to turn off for the synthesized definition; in particular, for weird cases like:
>> 
>> int a[];
>> void f() { extern int a[5]; }
>>  
>> ... we should not reject even though the synthesized definition has type 'int a[1];' which is not compatible with the declaration in 'f'. I suppose we could even detect this case and not emit a definition of 'a' at all here, because we know that another TU must be providing one with the right size.)
> 
> Odds that that wouldn’t cause a failure somewhere? :)
> 
> If it does, the code is doing scary things today -- we'll emit 'a' as an array of 1 element. But just because we could, doesn't mean we should... =) *jurassic park flashback*

Well, it’s pretty easy to imagine f() not really caring about how long a[] is but putting a useless bound on it anyway.  I assume that’s technically a violation, but it’s a harmless one.

That said, the rule about defaulting incomplete array types to length 1 is just bad language design, especially under the official model that lacks common definitions.  We should consider warning about it, at least in translation units that actually use the array.

John.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20150129/3ab3aed0/attachment.html>