[patch] Check the linkage cache at every recursive step

[John McCall]

On May 28, 2013, at 11:37 AM, Rafael Espíndola <rafael.espindola at gmail.com> wrote:
>> Well, wait, associated question — does this actually cause a speed-up?
> 
> For SemaExpr.ii it is in the noise, which is why I decided to count
> the number of linkage computations.
> 
> The patch has some other advantages:
> * verifyLinkage is gone (inlined to getLVForDecl)
> * When asserts are enabled, we check the linkage even when it is
> computed as part of a recursive getLVForDecl. That is how I found the
> bug fixed in r182773.
> 
> OK?

Sure.  Mostly I wanted to make sure you'd done real performance testing.
As long as it doesn't actually cause a regression (which in theory it could,
due to added bookkeeping checks), I'm fine with the "better in the
abstract" argument.

John.