[clang] [llvm] [clang][Modules][Serialization] Avoid serializing duplicate source location entries for shared input files (PR #209795)

[Ayokunle Amodu via cfe-commits]

ayokunle321 wrote:

> @qiongsiwu has offered to test this PR against something that resembles our workflows against Apple SDKs. We've been keen on reducing the size overhead that pcms accumulate across a developer's iteration cycle.

Thanks! Though I just realized that workload I tested on (ROOT) builds with -fmodules-embed-all-files, which means that when mod A imports mod B, mod B's textual content is also serialized into mod A. So this patch is deduping both the SLoc entries and the textual content.

Across ROOT's 88 dictionaries that's a 4.1% PCM size reduction, but without embedded files it's only around 0.4%, since the SLoc entries themselves are pretty small. So the ROOT numbers may not be that representative, but I'd still like to see what Qiongsi reports.

https://github.com/llvm/llvm-project/pull/209795