[PATCH] Fix alignment of data in TypeLocs (PR16144)

[John McCall]

On Jun 5, 2013, at 8:34 PM, Richard Smith <richard at metafoo.co.uk> wrote:
> On Wed, Jun 5, 2013 at 8:06 PM, Eli Friedman <eli.friedman at gmail.com> wrote:
> Patch attached; fixes PR16144.
> 
> Currently, the data in TypeLocs consists of a bunch of tightly packed structures, and the structures can become misaligned because there isn't any padding.
> 
> There are basically three possible approaches to fixing the alignment issues in TypeLocs:
> 
> 1) Force every piece of the TypeLoc's data to have alignment 8.
> 2) Perform dynamic alignment adjustments.
> 3) Use #pragma pack to let the compiler know the data is intentionally misaligned.
> 
> (1) has a substantial impact on memory usage (something like 1% on Cocoa.h), so I'd like to avoid it if possible.  (2) is the attached patch; it avoids both misaligned loads and unnecessary memory usage.  The primary downside is that TypeLocBuilder becomes a lot more complicated, because it doesn't know in advance where it needs to insert padding.  (3) keeps around to misaligned data: there's a potential performance penalty, it requires being careful not to introduce incorrect accesses to the data, and it's just plain ugly.
> 
> 
> Two questions to focus on for review: would (3) be a better approach?
> 
> I've tried this. We expose pointers into the type source info block in a couple of places (for instance, the array of ParmVarDecl*s on a FunctionTypeLoc) and the misalignment is then exposed to quite a large body of code. Maybe a MisalignedArrayRef<...> would help, but the damage is still not very contained.

Hmm.  We could have an API that potentially copies out to a temporary buffer if the data is misaligned.  Dynamic alignment is probably cleaner, though;  ugh.

> And is there any way to make the TypeLocBuilder implementation a bit less ugly?
> 
> Perhaps we could remove the guarantee that the child locations of a TypeLoc produced by push<T> are valid, or require some explicit action to fix them?

I *think* that should be fine.

John.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20130606/3a667501/attachment.html>