[PATCH] WIP: DI: Fold constant arguments into a single MDString

[Duncan P. N. Exon Smith]

> On 2014-Sep-15, at 16:58, David Blaikie <dblaikie at gmail.com> wrote:
> 
>>  1. I think `\0` is cleaner than `,`.  Bike shed?
> 
> OK by me, but depending on how much schema and test case churn you're inclined to (I know you folks at Apple are a bit more concerned about schema churn than I am - because you have static bitcode archives, etc) we might want to discuss the whole thing in a bit more detail. Things like:
> 
> * how's the scheme going to extend to hierarchical data? (eventually we'll hopefully have rolled most of the whole DIE tree into single big strings - separate out the types (which are still huge/non-trivial with all the member function declarations, etc), etc)
> * Do we want to consider a more human readable form (perhaps with an internal clang flag to force it to produce that version of the metadata?) for the purpose of writing test cases? Perhaps self-describing, so we can more easily omit irrelevant fields from tests (when the file/line/col doesn't matter for a particular declaration, etc).
> 
> If you're willing to churn this more later, I don't mind deferring these decisions, if not - we might want to plan a bit more carefully.

I'd prefer to defer these decisions -- IMO, the churn is inevitable.

> HeaderTwine doesn't look like a Twine - it looks like a string builder of sorts. (I'd avoid using the term twine for something that doesn't do the magic delayed string concatenation that llvm's twine does) maybe HeaderString?

Good call.

> Some of the Verify implementations look off - old num operands == new num operands + header fields - 1 (minus one for the header itself being counted in the number of operands), right? Some don't add up (DITemplateTypeParameter::Verify went from 7 to 4 + 3, so isn't it missing a field?

Yup, I found a few myself.  I'll clean those up.

> Any size measurements/test results (does this bootstrap?) on this?

None yet.

Thanks for the early look -- seems like we're on the same page.  I'll
send it out again once I've finished the testcases (somewhat delayed due
to vacation) and gathered some data.