[chrisaycock]

The beauty of separate IRs is that each can represent a separate stage of compilation. For example, Rust goes from abstract syntax tree (AST) to high-level IR (HIR) to typed high-level IR (THIR):

https://rustc-dev-guide.rust-lang.org/part-3-intro.html

Additionally, Many language implementations use a mid-level IR to sit above LLVM: Rust has MIR, Swift has SIL, Julia has an SSA-form IR, etc. The article mentions GHC's Core, which is also an example of an IR that comes after type checking but before LLVM.

[retrac]

Multiple IRs can get verbose, but it also helps compartmentalize complexity.

In a toy C compiler I wrote, I had a pass that did a transform of structs and arrays, turning everything into pointer arithmetic -- only thing that pass did. After that phase, the IR doesn't need to know about arrays or structs.

Though I found a downside with that method. If i encountered an error after parsing, since I had repeatedly transformed the program, it was impossible to give a meaningful error message. Which line in the source or which source-level feature corresponds to the broken AST node? That kind of info is just as tricky to carry around in the AST as type tagging.

[lionkor]

https://rustc-dev-guide.rust-lang.org/rustc-driver-interacti...

Love seeing that even Rust compiler devs are forced to be verbose and dont bother breaking some of this code out into functions, because specifying the constraints is too much work