[cfe-commits] r146959 - in /cfe/trunk: include/clang/AST/ include/clang/Analysis/Analyses/ include/clang/Basic/ include/clang/CodeGen/ include/clang/Driver/ include/clang/Frontend/ include/clang/Lex/ include/clang/Sema/ include/clang/StaticAnalyzer/Core/ include/clang/StaticAnalyzer/Core/BugReporter/ include/clang/StaticAnalyzer/Core/PathSensitive/ lib/AST/ lib/Analysis/ lib/Basic/ lib/CodeGen/ lib/Driver/ lib/Frontend/ lib/Lex/ lib/Parse/ lib/Sema/ lib/StaticAnalyzer/Core/ utils/TableGen/

[Richard Smith]

On Tue, December 20, 2011 08:54, Chris Lattner wrote:
> On Dec 19, 2011, at 9:57 PM, Ted Kremenek wrote:
>> On Dec 19, 2011, at 9:48 PM, Rafael Ávila de Espíndola wrote:
>>> On 19/12/11 11:33 PM, Ted Kremenek wrote:
>>>> Hi David,
>>>>
>>> When building with gcc 4.6, the sum of the size of every .o goes from
>>> 81926520 to 82165328 (so the patch saved a bit) and the clang binary
>>> goes from 36017976 to 35894528.
>>
>> Beyond the file size, my main concern is why do we need an anchor() methods
>> at all?  Can't we just move more methods to have out-of-line definitions in
>> .cpp files and accomplish the same thing?  I'm completely willing to admit
>> that I'm missing some key idea, but if so can someone please point it out
>> to me?  No offense David (as I really appreciate you spear heading this
>> effort), but they really look like turds in the codebase to me.
>
> If there is a simple virtual method that can be moved out of line without
> impacting performance, then that is clearly preferable to adding an explicit
> anchor method.  The main advantage of this patch IMO is that we could turn on
> the warning flag in the build to catch future cases that should have an
> anchor.

This issue seems indicative of a missing C++ language feature for explicit
suppression / generation of class support data -- akin to gcc's inline
template feature, but for any class. We are considering presenting a paper at
Kona suggesting such a feature.

In the absence of such language support, might it be better to use macros to
provide the anchor function declaration and definition rather than explicitly
duplicating this hack all over the code?

- Richard