<html><head><meta http-equiv="Content-Type" content="text/html charset=utf-8"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" class=""><br class=""><div><blockquote type="cite" class=""><div class="">On Jan 29, 2015, at 2:47 PM, Richard Smith <<a href="mailto:richard@metafoo.co.uk" class="">richard@metafoo.co.uk</a>> wrote:</div><br class="Apple-interchange-newline"><div class=""><div dir="ltr" class=""><div class="gmail_extra"><div class="gmail_quote">On Fri, Jan 23, 2015 at 12:25 AM, John McCall <span dir="ltr" class=""><<a href="mailto:rjmccall@apple.com" target="_blank" class="">rjmccall@apple.com</a>></span> wrote:<br class=""><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><span class="">> On Jan 22, 2015, at 4:52 PM, Rafael Espíndola <<a href="mailto:rafael.espindola@gmail.com" class="">rafael.espindola@gmail.com</a>> wrote:<br class="">
><br class="">
> Sent the email a bit early.<br class="">
><br class="">
><br class="">
>>> That is not what I am seeing with gcc. Given<br class="">
>>><br class="">
>>> int pr22217_foo;<br class="">
>>> int *b = &pr22217_foo;<br class="">
>>> extern int pr22217_foo __attribute__((section("zed")));<br class="">
<br class="">
</span>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.<br class=""></blockquote><div class=""><br class=""></div><div class="">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).</div><div class=""><br class=""></div><div class="">(Segue away from attributes and towards tentative definitons follows...)</div><div class=""><br class=""></div><div class="">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.”</div></div></div></div></div></blockquote><div><br class=""></div>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.</div><div><br class=""></div><div>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.</div><div><br class=""></div><div>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.</div><div><br class=""></div><div><blockquote type="cite" class=""><div class=""><div dir="ltr" class=""><div class="gmail_extra"><div class="gmail_quote"><div class="">Based on that simple semantic model, it is not reasonable for us to reject this:</div><div class=""><div class=""><br class=""></div><div class="">int pr22217_foo;</div><div class="">int *b = &pr22217_foo;</div><div class="">extern int pr22217_foo __attribute__((section("zed")));</div></div><div class="">int pr22217_foo = 123;</div><div class=""><br class=""></div><div class="">See also PR20688, which is a rejects-valid for standard C11 code due to our being confused about how tentative definitions work.</div><div class=""><br class=""></div><div class="">And here's another case we get wrong:</div><div class=""><br class=""></div><div class=""><div class=""> int a[];</div><div class=""> extern int a[5];</div></div><div class=""><br class=""></div><div class="">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:</div><div class=""><br class=""></div><div class=""><div class=""> struct foo x; // ok, tentative definition</div><div class=""> struct foo { int n, m; };</div></div><div class=""> // definition emitted now and has complete type; initializer is {0}.</div><div class=""><br class=""></div><div class="">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.</div></div></div></div></div></blockquote><div><br class=""></div>Like I mentioned above, this isn’t actually allowed under -fcommon.</div><div><br class=""><blockquote type="cite" class=""><div class=""><div dir="ltr" class=""><div class="gmail_extra"><div class="gmail_quote"><div class="">Or we can try to keep our current model with a tristate for whether a declaration is a definition, but that requires both Sema and IRGen to get a lot smarter with regard to handling of tentative definitions.<br class=""></div></div></div></div>
</div></blockquote></div><br class=""><div class="">I think this is reasonable. IRGen should be able to just completely replace an existing tentative definition. As I mentioned up-thread, IRGen needs to hold persistent references to global variables with handles anyway just because types can change.</div><div class=""><br class=""></div><div class="">John.</div></body></html>