website/slides/speaker-notes.md

how rust compiles

• welcome to my talk about how rust compiles
  • we'll take a look at rusts compilation model
• before we get started
  • example of confusing behavior arising from the model
• me
• behind cargo build
  • we use cargo
  • one cargo many rustc per crate
  • use -v
• what does rustc do
• source example
  • add function that adds
• frontend and backend
• frontend in a single slide
  • parse into AST
  • resolve, typecheck
• MIR
  • intermediate representation
  • CFG
  • borrow checker
• LLVM IR
  • LLVM backend
  • clang, swift, julia
  • IR to assembly
  • optimizes
• Assembly
  • machine code
• it's not that simple
• backend orchestration
  • how, what, and when invoking LLVM
• codegen unit
  • LLVM module/object file
  • group of functions
  • linked together
• more codegen units
  • three functions into two codegen units
  • functions grouped into codegen units is complicated
• codegen units cross crate
  • a great add function
  • add into cgu in math crate
  • object file into rlibs
• generics
  • mono
• instantiating generics
  • MIR to LLVM IR
• generics cross crate
  • math crate only MIR
  • instantiated in our own crate
  • read MIR from rlib
• inlining
  • a function call
  • inefficient call
  • maybe just turn it into 5
• inlining
  • very good
  • everywhere, or very slow
• cross crate inlining catch
  • access to the body
• #[inline]
  • why?
• #[inline]
  • like a generic function
  • instantiated downstream
• #[inline]
  • non-generic
  • automatically for small
  • not for others
  • cross-crate-inline-threshold=always
  • dont always use it because its slow
  • dont forget about it
  • benchmark
• if you dont want to worry about this as an application
• LTO
  • no cross-crate-inline-threshold
  • LTO
  • applies optimizations once for everything at the end
  • breaks crate boundaries
  • slow
• fat LTO
  • LLVM IR instead of machine code in libs
  • combines all LLVM IR
  • this CGU is huge, not parallel
  • see everything
  • not all optimizations for larger programs
• thin LTO
  • confusing name
  • try both
  • LLVM IR for libs
  • for generics, just in instantiating crate
  • combine for summary
  • summary of function (size and where to find it)
  • many different units in parallel
  • look at the summary for inlining
• linker plugin LTO
  • fat LTO or thin LTO
  • not by rustc, but by the linker
  • across languages
  • LLVM IR from Rust, LLVM IR from clang
  • annoying to set up