Hi, folks. I am trying to understand the Intermediate Language in LLVM infrastructure.
But I cannot quite catch the grasp on how the Intermediate Language comes out.
I have some questions.
When developing the new Intermediate Language on Swift Programming Language, Have you started the development from the Parser, which is the beginning of the Compiler?
I searched some good Intermediate Language Projects, but couldn’t find the good specification on IL-like projects. Would you able to give some lists of the good Intermediate Language Projects to get some brief Overview on this?
If you’re interested in swift specifically, it would make sense to ask on the swift dev forum. There is a lot of thought and time that went into the design of SIL.
More generally though, designing an IR like this is all about deciding what level of representational abstraction you need, and this is principally driven by what sorts of optimizations you want to do, and what sort of software abstractions you want to simplify the lowering from one level to another.
In the case of the former, Clang has its “CFG” data structure, but otherwise lacks a great representation for doing language-specific dataflow based analysis and optimization. We should be able to do things like reference count optimizations of std::shared_ptr, should be able to do advanced temporary-elimination optimizations which only appear after inlining, should be able to do STL specific optimizations like inserting “reserve” calls before loops that fill in data structures, etc.
On the lowering side of things, Clang’s CodeGen is monolithic, very complicated, and includes lots of concerns mixed up. The most egregious of which is that the C/C++ ABI lowering stuff all happens in one place, and there is no good way to factor this logic across the LLVM ecosystem, other than generating C++ ASTs (or writing thunks in C++ and compiling them to IR with clang, then inlining them, etc). This makes me really sad, and is something that a nice CIL could fix or at least drastically improve.
What I have in my mind now is making the general Clang Intermediate Language(CIL).
When developing the Graphics card compiler based on LLVM Infrastructure, the optimization is fairly able on LLVM IR. But the problem is that most programmer are more familar to C-style code, not LLVM IR code. Also the fact that LLVM IR is more RISC-like format, there is some limitation to optimize which should be easy if the C/C++/Objective-C Intermediate Language is on clang.
Also what I have in mind is that most software patterns what most programmers want to optimize could be accomplished with just simplified Intermediate Language. With more C-like Intermediate Language, Specified Type-Statement on CIL for making software engineers able to easily optimize.
Though I have some concerns whether this higher-level IR which is more than LLVM IR could help optimize because it feels like duplicating the IR layer. Still doubt on whether this CIL is necessary for LLVM and Clang Infrastructure…
P.S. I wish there are some SPIR-V developers, and share their thought on the CIL.