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

   1 # Backend Agnostic Codegen
   2
   3 <!-- toc -->
   4
   5 As of <!-- date: 2021-01 --> January 2021, `rustc_codegen_ssa` provides an
   6 abstract interface for all backends to implement, to allow other codegen
   7 backends (e.g. [Cranelift]).
   8
   9 [Cranelift]: https://github.com/bytecodealliance/wasmtime/tree/HEAD/cranelift
  10
  11 > The following is a copy/paste of a README from the rust-lang/rust repo.
  12 > Please submit a PR if it needs updating.
  13
  14 # Refactoring of `rustc_codegen_llvm`
  15 by Denis Merigoux, October 23rd 2018
  16
  17 ## State of the code before the refactoring
  18
  19 All the code related to the compilation of MIR into LLVM IR was contained
  20 inside the `rustc_codegen_llvm` crate. Here is the breakdown of the most
  21 important elements:
  22 * the `back` folder (7,800 LOC) implements the mechanisms for creating the
  23   different object files and archive through LLVM, but also the communication
  24   mechanisms for parallel code generation;
  25 * the `debuginfo` (3,200 LOC) folder contains all code that passes debug
  26   information down to LLVM;
  27 * the `llvm` (2,200 LOC) folder defines the FFI necessary to communicate with
  28   LLVM using the C++ API;
  29 * the `mir` (4,300 LOC) folder implements the actual lowering from MIR to LLVM
  30   IR;
  31 * the `base.rs` (1,300 LOC) file contains some helper functions but also the
  32   high-level code that launches the code generation and distributes the work.
  33 * the `builder.rs` (1,200 LOC) file contains all the functions generating
  34   individual LLVM IR instructions inside a basic block;
  35 * the `common.rs` (450 LOC) contains various helper functions and all the
  36   functions generating LLVM static values;
  37 * the `type_.rs` (300 LOC) defines most of the type translations to LLVM IR.
  38
  39 The goal of this refactoring is to separate inside this crate code that is
  40 specific to the LLVM from code that can be reused for other rustc backends. For
  41 instance, the `mir` folder is almost entirely backend-specific but it relies
  42 heavily on other parts of the crate. The separation of the code must not affect
  43 the logic of the code nor its performance.
  44
  45 For these reasons, the separation process involves two transformations that
  46 have to be done at the same time for the resulting code to compile :
  47
  48 1. replace all the LLVM-specific types by generics inside function signatures
  49    and structure definitions;
  50 2. encapsulate all functions calling the LLVM FFI inside a set of traits that
  51    will define the interface between backend-agnostic code and the backend.
  52
  53 While the LLVM-specific code will be left in `rustc_codegen_llvm`, all the new
  54 traits and backend-agnostic code will be moved in `rustc_codegen_ssa` (name
  55 suggestion by @eddyb).
  56
  57 ## Generic types and structures
  58
  59 @irinagpopa started to parametrize the types of `rustc_codegen_llvm` by a
  60 generic `Value` type, implemented in LLVM by a reference `&'ll Value`. This
  61 work has been extended to all structures inside the `mir` folder and elsewhere,
  62 as well as for LLVM's `BasicBlock` and `Type` types.
  63
  64 The two most important structures for the LLVM codegen are `CodegenCx` and
  65 `Builder`. They are parametrized by multiple lifetime parameters and the type
  66 for `Value`.
  67
  68 ```rust,ignore
  69 struct CodegenCx<'ll, 'tcx> {
  70   /* ... */
  71 }
  72
  73 struct Builder<'a, 'll, 'tcx> {
  74   cx: &'a CodegenCx<'ll, 'tcx>,
  75   /* ... */
  76 }
  77 ```
  78
  79 `CodegenCx` is used to compile one codegen-unit that can contain multiple
  80 functions, whereas `Builder` is created to compile one basic block.
  81
  82 The code in `rustc_codegen_llvm` has to deal with multiple explicit lifetime
  83 parameters, that correspond to the following:
  84 * `'tcx` is the longest lifetime, that corresponds to the original `TyCtxt`
  85   containing the program's information;
  86 * `'a` is a short-lived reference of a `CodegenCx` or another object inside a
  87   struct;
  88 * `'ll` is the lifetime of references to LLVM objects such as `Value` or
  89   `Type`.
  90
  91 Although there are already many lifetime parameters in the code, making it
  92 generic uncovered situations where the borrow-checker was passing only due to
  93 the special nature of the LLVM objects manipulated (they are extern pointers).
  94 For instance, an additional lifetime parameter had to be added to
  95 `LocalAnalyser` in `analyse.rs`, leading to the definition:
  96
  97 ```rust,ignore
  98 struct LocalAnalyzer<'mir, 'a, 'tcx> {
  99   /* ... */
 100 }
 101 ```
 102
 103 However, the two most important structures `CodegenCx` and `Builder` are not
 104 defined in the backend-agnostic code. Indeed, their content is highly specific
 105 of the backend and it makes more sense to leave their definition to the backend
 106 implementor than to allow just a narrow spot via a generic field for the
 107 backend's context.
 108
 109 ## Traits and interface
 110
 111 Because they have to be defined by the backend, `CodegenCx` and `Builder` will
 112 be the structures implementing all the traits defining the backend's interface.
 113 These traits are defined in the folder `rustc_codegen_ssa/traits` and all the
 114 backend-agnostic code is parametrized by them. For instance, let us explain how
 115 a function in `base.rs` is parametrized:
 116
 117 ```rust,ignore
 118 pub fn codegen_instance<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
 119     cx: &'a Bx::CodegenCx,
 120     instance: Instance<'tcx>
 121 ) {
 122     /* ... */
 123 }
 124 ```
 125
 126 In this signature, we have the two lifetime parameters explained earlier and
 127 the master type `Bx` which satisfies the trait `BuilderMethods` corresponding
 128 to the interface satisfied by the `Builder` struct. The `BuilderMethods`
 129 defines an associated type `Bx::CodegenCx` that itself satisfies the
 130 `CodegenMethods` traits implemented by the struct `CodegenCx`.
 131
 132 On the trait side, here is an example with part of the definition of
 133 `BuilderMethods` in `traits/builder.rs`:
 134
 135 ```rust,ignore
 136 pub trait BuilderMethods<'a, 'tcx>:
 137     HasCodegen<'tcx>
 138     + DebugInfoBuilderMethods<'tcx>
 139     + ArgTypeMethods<'tcx>
 140     + AbiBuilderMethods<'tcx>
 141     + IntrinsicCallMethods<'tcx>
 142     + AsmBuilderMethods<'tcx>
 143 {
 144     fn new_block<'b>(
 145         cx: &'a Self::CodegenCx,
 146         llfn: Self::Function,
 147         name: &'b str
 148     ) -> Self;
 149     /* ... */
 150     fn cond_br(
 151         &mut self,
 152         cond: Self::Value,
 153         then_llbb: Self::BasicBlock,
 154         else_llbb: Self::BasicBlock,
 155     );
 156     /* ... */
 157 }
 158 ```
 159
 160 Finally, a master structure implementing the `ExtraBackendMethods` trait is
 161 used for high-level codegen-driving functions like `codegen_crate` in
 162 `base.rs`. For LLVM, it is the empty `LlvmCodegenBackend`.
 163 `ExtraBackendMethods` should be implemented by the same structure that
 164 implements the `CodegenBackend` defined in
 165 `rustc_codegen_utils/codegen_backend.rs`.
 166
 167 During the traitification process, certain functions have been converted from
 168 methods of a local structure to methods of `CodegenCx` or `Builder` and a
 169 corresponding `self` parameter has been added. Indeed, LLVM stores information
 170 internally that it can access when called through its API. This information
 171 does not show up in a Rust data structure carried around when these methods are
 172 called. However, when implementing a Rust backend for `rustc`, these methods
 173 will need information from `CodegenCx`, hence the additional parameter (unused
 174 in the LLVM implementation of the trait).
 175
 176 ## State of the code after the refactoring
 177
 178 The traits offer an API which is very similar to the API of LLVM. This is not
 179 the best solution since LLVM has a very special way of doing things: when
 180 adding another backend, the traits definition might be changed in order to
 181 offer more flexibility.
 182
 183 However, the current separation between backend-agnostic and LLVM-specific code
 184 has allowed the reuse of a significant part of the old `rustc_codegen_llvm`.
 185 Here is the new LOC breakdown between backend-agnostic (BA) and LLVM for the
 186 most important elements:
 187
 188 * `back` folder: 3,800 (BA) vs 4,100 (LLVM);
 189 * `mir` folder: 4,400 (BA) vs 0 (LLVM);
 190 * `base.rs`: 1,100 (BA) vs 250 (LLVM);
 191 * `builder.rs`: 1,400 (BA) vs 0 (LLVM);
 192 * `common.rs`: 350 (BA) vs 350 (LLVM);
 193
 194 The `debuginfo` folder has been left almost untouched by the splitting and is
 195 specific to LLVM. Only its high-level features have been traitified.
 196
 197 The new `traits` folder has 1500 LOC only for trait definitions. Overall, the
 198 27,000 LOC-sized old `rustc_codegen_llvm` code has been split into the new
 199 18,500 LOC-sized new `rustc_codegen_llvm` and the 12,000 LOC-sized
 200 `rustc_codegen_ssa`. We can say that this refactoring allowed the reuse of
 201 approximately 10,000 LOC that would otherwise have had to be duplicated between
 202 the multiple backends of `rustc`.
 203
 204 The refactored version of `rustc`'s backend introduced no regression over the
 205 test suite nor in performance benchmark, which is in coherence with the nature
 206 of the refactoring that used only compile-time parametricity (no trait
 207 objects).