[LLD][COFF] Early dependency detection

[aganea]

Summary

This patch adds more fine grained information in ObjFile about PDB type server dependencies and precomp OBJ dependencies.
Dependencies are required to be loaded & merged first, before any dependent OBJ that relies on it.

Before, dependencies were discovered and loaded along the way, when the type stream for a given OBJ was being merged (in PDB.cpp).

This patch is part of ground work towards parallelizing the type merging. The objective is to split PDBLinker::addObjFile into several smaller parallelizable pieces. See D59226 for the overall intention.

[rnk]

COFF/InputFiles.h

168–182

I'm starting to feel like mergeDebugT should really be behind a virtual interface, something like this:

class TPISource { // TpiSource?
public:
  virtual CVIndexMap *mergeTPI(ObjFile *File, CVIndexMap *ObjectIndexMap) = 0;
};
...
class SectionTPISource { ... }; // ObjectTPISource? RegularTPISource? Z7TPISource? TpiSource?
class TypeServerTPISource { ... };
class PrecompTPISource { ... };
class UsePrecompTPISource { ... };

The class names I came up with are kind of crummy, though.

I feel like 4 public, optional members of ObjFile leaks a lot of PDB details into the rest of the linker. Instead, we could have a pointer to some interface. Do you think this, or something like it, is worth doing now, ever, or later?

The "regular" or .debug$T based type merger could just be some stateless global variable that all the regular objs point to, since we pass the object file in through the interface, and that's all the info it needs. Every file using a type server would point to the same TypeServerTPISource. I'm not sure how the PCH stuff would work, since I'm less familiar.

[aganea]

COFF/InputFiles.h

168–182

Yes you're completly right - somehow I started doing that at first, but there were too many changes, so I meant to do this incrementally. I'll lay out this new class hierarchy in this patch, and then I'll implement mergeTPI() in a subsequent patch.

[aganea]

Introduce TpiSource and derived types as suggested by @rnk.
Please see D59226 for more information as for where this is going.

[ruiu]

COFF/DebugTypes.cpp

47

What is this? You are not using this. (If you need it in another patch, move this to the another patch.)

66

nit: Please add a blank line between function definitions.

COFF/InputFiles.cpp

688

If you do know the first argument for all function calls of makeTpiSource, you can define one function for each value of the first argument, instead of calling the same function and then dispatch according to the first argument.

693

Avoid auto.

COFF/PDB.cpp

406

Please do not use auto unless its type is obvious from the right-hand side (e.g. cast or new).

[aganea]

Updated with changes as suggested by @ruiu

COFF/DebugTypes.cpp

47

Normally we should have std::unique_ptr<TpiSource> DebugTypesObj in COFF/InputFiles.h, L166, but I wanted to avoid #including DebugTypes.h just so we can use std::unique_ptr. I'm keeping hard references to TpiSource objects here instead.

COFF/InputFiles.cpp

688

I did like you said - however now there are details leaking out of DebugTypes.h, like TypeServer2Record parameter to makeUseTypeServerSource(), which now requires having #include "llvm/DebugInfo/CodeView/TypeRecord.h" directly in DebugTypes.h, which can slow down compilation - if DebugTypes.h is included elsewhere. It's a tradeoff between terseness and compilation speed, please let me know which you prefer.

693

The right side already specifies that it deserializes as a TypeServer2Record, but yes I made that explicit.

[aganea]

Remove reference to llvm/DebugInfo/CodeView/TypeRecord.h from DebugTypes.h

COFF/InputFiles.cpp

688

Forget what I said, it was late last night. Changed the parameters to pointers and forward declarations.

[ruiu]

COFF/InputFiles.cpp

662

Could you write a function comment as to what this function is supposed to do so that it can be easily understood for a first-time reader?

688

I was about to say that you probably could make use of forward declaration, but you already did. :)

[aganea]

Document ObjFile::initializeDependencies().

[aganea]

Ping! Any further comments?

[rnk]

Not from me. :)