src/doc/rustc-dev-guide/src/codegen/backend-agnostic.md

   1 # Backend Agnostic Codegen
   2
   3 In the future, it would be nice to allow other codegen backends (e.g.
   4 [Cranelift][cranelift]). To this end, `librustc_codegen_ssa` provides an
   5 abstract interface for all backends to implenent.
   6
   7 [cranelift]: https://github.com/CraneStation/cranelift
   8
   9 > The following is a copy/paste of a README from the rust-lang/rust repo.
  10 > Please submit a PR if it needs updating.
  11
  12 # Refactoring of `rustc_codegen_llvm`
  13 by Denis Merigoux, October 23rd 2018
  14
  15 ## State of the code before the refactoring
  16
  17 All the code related to the compilation of MIR into LLVM IR was contained
  18 inside the `rustc_codegen_llvm` crate. Here is the breakdown of the most
  19 important elements:
  20 * the `back` folder (7,800 LOC) implements the mechanisms for creating the
  21   different object files and archive through LLVM, but also the communication
  22   mechanisms for parallel code generation;
  23 * the `debuginfo` (3,200 LOC) folder contains all code that passes debug
  24   information down to LLVM;
  25 * the `llvm` (2,200 LOC) folder defines the FFI necessary to communicate with
  26   LLVM using the C++ API;
  27 * the `mir` (4,300 LOC) folder implements the actual lowering from MIR to LLVM
  28   IR;
  29 * the `base.rs` (1,300 LOC) file contains some helper functions but also the
  30   high-level code that launches the code generation and distributes the work.
  31 * the `builder.rs` (1,200 LOC) file contains all the functions generating
  32   individual LLVM IR instructions inside a basic block;
  33 * the `common.rs` (450 LOC) contains various helper functions and all the
  34   functions generating LLVM static values;
  35 * the `type_.rs` (300 LOC) defines most of the type translations to LLVM IR.
  36
  37 The goal of this refactoring is to separate inside this crate code that is
  38 specific to the LLVM from code that can be reused for other rustc backends. For
  39 instance, the `mir` folder is almost entirely backend-specific but it relies
  40 heavily on other parts of the crate. The separation of the code must not affect
  41 the logic of the code nor its performance.
  42
  43 For these reasons, the separation process involves two transformations that
  44 have to be done at the same time for the resulting code to compile :
  45
  46 1. replace all the LLVM-specific types by generics inside function signatures
  47    and structure definitions;
  48 2. encapsulate all functions calling the LLVM FFI inside a set of traits that
  49    will define the interface between backend-agnostic code and the backend.
  50
  51 While the LLVM-specific code will be left in `rustc_codegen_llvm`, all the new
  52 traits and backend-agnostic code will be moved in `rustc_codegen_ssa` (name
  53 suggestion by @eddyb).
  54
  55 ## Generic types and structures
  56
  57 @irinagpopa started to parametrize the types of `rustc_codegen_llvm` by a
  58 generic `Value` type, implemented in LLVM by a reference `&'ll Value`. This
  59 work has been extended to all structures inside the `mir` folder and elsewhere,
  60 as well as for LLVM's `BasicBlock` and `Type` types.
  61
  62 The two most important structures for the LLVM codegen are `CodegenCx` and
  63 `Builder`. They are parametrized by multiple lifetime parameters and the type
  64 for `Value`.
  65
  66 ```rust,ignore
  67 struct CodegenCx<'ll, 'tcx> {
  68   /* ... */
  69 }
  70
  71 struct Builder<'a, 'll, 'tcx> {
  72   cx: &'a CodegenCx<'ll, 'tcx>,
  73   /* ... */
  74 }
  75 ```
  76
  77 `CodegenCx` is used to compile one codegen-unit that can contain multiple
  78 functions, whereas `Builder` is created to compile one basic block.
  79
  80 The code in `rustc_codegen_llvm` has to deal with multiple explicit lifetime
  81 parameters, that correspond to the following:
  82 * `'tcx` is the longest lifetime, that corresponds to the original `TyCtxt`
  83   containing the program's information;
  84 * `'a` is a short-lived reference of a `CodegenCx` or another object inside a
  85   struct;
  86 * `'ll` is the lifetime of references to LLVM objects such as `Value` or
  87   `Type`.
  88
  89 Although there are already many lifetime parameters in the code, making it
  90 generic uncovered situations where the borrow-checker was passing only due to
  91 the special nature of the LLVM objects manipulated (they are extern pointers).
  92 For instance, an additional lifetime parameter had to be added to
  93 `LocalAnalyser` in `analyse.rs`, leading to the definition:
  94
  95 ```rust,ignore
  96 struct LocalAnalyzer<'mir, 'a, 'tcx> {
  97   /* ... */
  98 }
  99 ```
 100
 101 However, the two most important structures `CodegenCx` and `Builder` are not
 102 defined in the backend-agnostic code. Indeed, their content is highly specific
 103 of the backend and it makes more sense to leave their definition to the backend
 104 implementor than to allow just a narrow spot via a generic field for the
 105 backend's context.
 106
 107 ## Traits and interface
 108
 109 Because they have to be defined by the backend, `CodegenCx` and `Builder` will
 110 be the structures implementing all the traits defining the backend's interface.
 111 These traits are defined in the folder `rustc_codegen_ssa/traits` and all the
 112 backend-agnostic code is parametrized by them. For instance, let us explain how
 113 a function in `base.rs` is parametrized:
 114
 115 ```rust,ignore
 116 pub fn codegen_instance<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 117     cx: &'a Bx::CodegenCx,
 118     instance: Instance<'tcx>
 119 ) {
 120     /* ... */
 121 }
 122 ```
 123
 124 In this signature, we have the two lifetime parameters explained earlier and
 125 the master type `Bx` which satisfies the trait `BuilderMethods` corresponding
 126 to the interface satisfied by the `Builder` struct. The `BuilderMethods`
 127 defines an associated type `Bx::CodegenCx` that itself satisfies the
 128 `CodegenMethods` traits implemented by the struct `CodegenCx`.
 129
 130 On the trait side, here is an example with part of the definition of
 131 `BuilderMethods` in `traits/builder.rs`:
 132
 133 ```rust,ignore
 134 pub trait BuilderMethods<'a, 'tcx>:
 135     HasCodegen<'tcx>
 136     + DebugInfoBuilderMethods<'tcx>
 137     + ArgTypeMethods<'tcx>
 138     + AbiBuilderMethods<'tcx>
 139     + IntrinsicCallMethods<'tcx>
 140     + AsmBuilderMethods<'tcx>
 141 {
 142     fn new_block<'b>(
 143         cx: &'a Self::CodegenCx,
 144         llfn: Self::Function,
 145         name: &'b str
 146     ) -> Self;
 147     /* ... */
 148     fn cond_br(
 149         &mut self,
 150         cond: Self::Value,
 151         then_llbb: Self::BasicBlock,
 152         else_llbb: Self::BasicBlock,
 153     );
 154     /* ... */
 155 }
 156 ```
 157
 158 Finally, a master structure implementing the `ExtraBackendMethods` trait is
 159 used for high-level codegen-driving functions like `codegen_crate` in
 160 `base.rs`. For LLVM, it is the empty `LlvmCodegenBackend`.
 161 `ExtraBackendMethods` should be implemented by the same structure that
 162 implements the `CodegenBackend` defined in
 163 `rustc_codegen_utils/codegen_backend.rs`.
 164
 165 During the traitification process, certain functions have been converted from
 166 methods of a local structure to methods of `CodegenCx` or `Builder` and a
 167 corresponding `self` parameter has been added. Indeed, LLVM stores information
 168 internally that it can access when called through its API. This information
 169 does not show up in a Rust data structure carried around when these methods are
 170 called. However, when implementing a Rust backend for `rustc`, these methods
 171 will need information from `CodegenCx`, hence the additional parameter (unused
 172 in the LLVM implementation of the trait).
 173
 174 ## State of the code after the refactoring
 175
 176 The traits offer an API which is very similar to the API of LLVM. This is not
 177 the best solution since LLVM has a very special way of doing things: when
 178 addding another backend, the traits definition might be changed in order to
 179 offer more flexibility.
 180
 181 However, the current separation between backend-agnostic and LLVM-specific code
 182 has allowed the reuse of a significant part of the old `rustc_codegen_llvm`.
 183 Here is the new LOC breakdown between backend-agnostic (BA) and LLVM for the
 184 most important elements:
 185
 186 * `back` folder: 3,800 (BA) vs 4,100 (LLVM);
 187 * `mir` folder: 4,400 (BA) vs 0 (LLVM);
 188 * `base.rs`: 1,100 (BA) vs 250 (LLVM);
 189 * `builder.rs`: 1,400 (BA) vs 0 (LLVM);
 190 * `common.rs`: 350 (BA) vs 350 (LLVM);
 191
 192 The `debuginfo` folder has been left almost untouched by the splitting and is
 193 specific to LLVM. Only its high-level features have been traitified.
 194
 195 The new `traits` folder has 1500 LOC only for trait definitions. Overall, the
 196 27,000 LOC-sized old `rustc_codegen_llvm` code has been split into the new
 197 18,500 LOC-sized new `rustc_codegen_llvm` and the 12,000 LOC-sized
 198 `rustc_codegen_ssa`. We can say that this refactoring allowed the reuse of
 199 approximately 10,000 LOC that would otherwise have had to be duplicated between
 200 the multiple backends of `rustc`.
 201
 202 The refactored version of `rustc`'s backend introduced no regression over the
 203 test suite nor in performance benchmark, which is in coherence with the nature
 204 of the refactoring that used only compile-time parametricity (no trait
 205 objects).