...
▽

Text file src/cmd/compile/README.md

Documentation: cmd/compile

     1<!---
     2// Copyright 2018 The Go Authors. All rights reserved.
     3// Use of this source code is governed by a BSD-style
     4// license that can be found in the LICENSE file.
     5-->
     6
     7## Introduction to the Go compiler
     8
     9`cmd/compile` contains the main packages that form the Go compiler. The compiler
    10may be logically split in four phases, which we will briefly describe alongside
    11the list of packages that contain their code.
    12
    13You may sometimes hear the terms "front-end" and "back-end" when referring to
    14the compiler. Roughly speaking, these translate to the first two and last two
    15phases we are going to list here. A third term, "middle-end", often refers to
    16much of the work that happens in the second phase.
    17
    18Note that the `go/*` family of packages, such as `go/parser` and
    19`go/types`, are mostly unused by the compiler. Since the compiler was
    20initially written in C, the `go/*` packages were developed to enable
    21writing tools working with Go code, such as `gofmt` and `vet`.
    22However, over time the compiler's internal APIs have slowly evolved to
    23be more familiar to users of the `go/*` packages.
    24
    25It should be clarified that the name "gc" stands for "Go compiler", and has
    26little to do with uppercase "GC", which stands for garbage collection.
    27
    28### 1. Parsing
    29
    30* `cmd/compile/internal/syntax` (lexer, parser, syntax tree)
    31
    32In the first phase of compilation, source code is tokenized (lexical analysis),
    33parsed (syntax analysis), and a syntax tree is constructed for each source
    34file.
    35
    36Each syntax tree is an exact representation of the respective source file, with
    37nodes corresponding to the various elements of the source such as expressions,
    38declarations, and statements. The syntax tree also includes position information
    39which is used for error reporting and the creation of debugging information.
    40
    41### 2. Type checking
    42
    43* `cmd/compile/internal/types2` (type checking)
    44
    45The types2 package is a port of `go/types` to use the syntax package's
    46AST instead of `go/ast`.
    47
    48### 3. IR construction ("noding")
    49
    50* `cmd/compile/internal/types` (compiler types)
    51* `cmd/compile/internal/ir` (compiler AST)
    52* `cmd/compile/internal/noder` (create compiler AST)
    53
    54The compiler middle end uses its own AST definition and representation of Go
    55types carried over from when it was written in C. All of its code is written in
    56terms of these, so the next step after type checking is to convert the syntax
    57and types2 representations to ir and types. This process is referred to as
    58"noding."
    59
    60Noding uses a process called Unified IR, which builds a node representation
    61using a serialized version of the typechecked code from step 2.
    62Unified IR is also involved in import/export of packages and inlining.
    63
    64### 4. Middle end
    65
    66* `cmd/compile/internal/inline` (function call inlining)
    67* `cmd/compile/internal/devirtualize` (devirtualization of known interface method calls)
    68* `cmd/compile/internal/escape` (escape analysis)
    69
    70Several optimization passes are performed on the IR representation:
    71dead code elimination, (early) devirtualization, function call
    72inlining, and escape analysis.
    73
    74The early dead code elimination pass is integrated into the unified IR writer phase.
    75
    76### 5. Walk
    77
    78* `cmd/compile/internal/walk` (order of evaluation, desugaring)
    79
    80The final pass over the IR representation is "walk," which serves two purposes:
    81
    821. It decomposes complex statements into individual, simpler statements,
    83   introducing temporary variables and respecting order of evaluation. This step
    84   is also referred to as "order."
    85
    862. It desugars higher-level Go constructs into more primitive ones. For example,
    87   `switch` statements are turned into binary search or jump tables, and
    88   operations on maps and channels are replaced with runtime calls.
    89
    90### 6. Generic SSA
    91
    92* `cmd/compile/internal/ssa` (SSA passes and rules)
    93* `cmd/compile/internal/ssagen` (converting IR to SSA)
    94
    95In this phase, IR is converted into Static Single Assignment (SSA) form, a
    96lower-level intermediate representation with specific properties that make it
    97easier to implement optimizations and to eventually generate machine code from
    98it.
    99
   100During this conversion, function intrinsics are applied. These are special
   101functions that the compiler has been taught to replace with heavily optimized
   102code on a case-by-case basis.
   103
   104Certain nodes are also lowered into simpler components during the AST to SSA
   105conversion, so that the rest of the compiler can work with them. For instance,
   106the copy builtin is replaced by memory moves, and range loops are rewritten into
   107for loops. Some of these currently happen before the conversion to SSA due to
   108historical reasons, but the long-term plan is to move all of them here.
   109
   110Then, a series of machine-independent passes and rules are applied. These do not
   111concern any single computer architecture, and thus run on all `GOARCH` variants.
   112These passes include dead code elimination, removal of
   113unneeded nil checks, and removal of unused branches. The generic rewrite rules
   114mainly concern expressions, such as replacing some expressions with constant
   115values, and optimizing multiplications and float operations.
   116
   117### 7. Generating machine code
   118
   119* `cmd/compile/internal/ssa` (SSA lowering and arch-specific passes)
   120* `cmd/internal/obj` (machine code generation)
   121
   122The machine-dependent phase of the compiler begins with the "lower" pass, which
   123rewrites generic values into their machine-specific variants. For example, on
   124amd64 memory operands are possible, so many load-store operations may be combined.
   125
   126Note that the lower pass runs all machine-specific rewrite rules, and thus it
   127currently applies lots of optimizations too.
   128
   129Once the SSA has been "lowered" and is more specific to the target architecture,
   130the final code optimization passes are run. This includes yet another dead code
   131elimination pass, moving values closer to their uses, the removal of local
   132variables that are never read from, and register allocation.
   133
   134Other important pieces of work done as part of this step include stack frame
   135layout, which assigns stack offsets to local variables, and pointer liveness
   136analysis, which computes which on-stack pointers are live at each GC safe point.
   137
   138At the end of the SSA generation phase, Go functions have been transformed into
   139a series of obj.Prog instructions. These are passed to the assembler
   140(`cmd/internal/obj`), which turns them into machine code and writes out the
   141final object file. The object file will also contain reflect data, export data,
   142and debugging information.
   143
   144### 7a. Export
   145
   146In addition to writing a file of object code for the linker, the
   147compiler also writes a file of "export data" for downstream
   148compilation units. The export data file holds all the information
   149computed during compilation of package P that may be needed when
   150compiling a package Q that directly imports P. It includes type
   151information for all exported declarations, IR for bodies of functions
   152that are candidates for inlining, IR for bodies of generic functions
   153that may be instantiated in another package, and a summary of the
   154findings of escape analysis on function parameters.
   155
   156The format of the export data file has gone through a number of
   157iterations. Its current form is called "unified", and it is a
   158serialized representation of an object graph, with an index allowing
   159lazy decoding of parts of the whole (since most imports are used to
   160provide only a handful of symbols). See [here](internal/noder/README.md)
   161for details on making changes to unified IR.
   162
   163The GOROOT repository contains a reader and a writer for the unified
   164format; it encodes from/decodes to the compiler's IR.
   165The golang.org/x/tools repository also provides a public API for an export
   166data reader (using the go/types representation) that always supports the
   167compiler's current file format and a small number of historic versions.
   168(It is used by x/tools/go/packages in modes that require type information
   169but not type-annotated syntax.)
   170
   171The x/tools repository also provides public APIs for reading and
   172writing exported type information (but nothing more) using the older
   173"indexed" format. (For example, gopls uses this version for its
   174database of workspace information, which includes types.)
   175
   176Export data usually provides a "deep" summary, so that compilation of
   177package Q can read the export data files only for each direct import,
   178and be assured that these provide all necessary information about
   179declarations in indirect imports, such as the methods and struct
   180fields of types referred to in P's public API. Deep export data is
   181simpler for build systems, since only one file is needed per direct
   182dependency. However, it does have a tendency to grow as one gets
   183higher up the import graph of a big repository: if there is a set of
   184very commonly used types with a large API, nearly every package's
   185export data will include a copy. This problem motivated the "indexed"
   186design, which allowed partial loading on demand.
   187(gopls does less work than the compiler for each import and is thus
   188more sensitive to export data overheads. For this reason, it uses
   189"shallow" export data, in which indirect information is not recorded
   190at all. This demands random access to the export data files of all
   191dependencies, so is not suitable for distributed build systems.)
   192
   193
   194### 8. Tips
   195
   196#### Getting Started
   197
   198* If you have never contributed to the compiler before, a simple way to begin
   199  can be adding a log statement or `panic("here")` to get some
   200  initial insight into whatever you are investigating.
   201
   202* The compiler itself provides logging, debugging and visualization capabilities,
   203  such as:
   204   ```
   205   $ go build -gcflags=-m=2                   # print optimization info, including inlining, escape analysis
   206   $ go build -gcflags=-d=ssa/check_bce/debug # print bounds check info
   207   $ go build -gcflags=-W                     # print internal parse tree after type checking
   208   $ GOSSAFUNC=Foo go build                   # generate ssa.html file for func Foo
   209   $ go build -gcflags=-S                     # print assembly
   210   $ go tool compile -bench=out.txt x.go      # print timing of compiler phases
   211   ```
   212
   213  Some flags alter the compiler behavior, such as:
   214   ```
   215   $ go tool compile -h file.go               # panic on first compile error encountered
   216   $ go build -gcflags=-d=checkptr=2          # enable additional unsafe pointer checking
   217   ```
   218
   219  There are many additional flags. Some descriptions are available via:
   220   ```
   221   $ go tool compile -h              # compiler flags, e.g., go build -gcflags='-m=1 -l'
   222   $ go tool compile -d help         # debug flags, e.g., go build -gcflags=-d=checkptr=2
   223   $ go tool compile -d ssa/help     # ssa flags, e.g., go build -gcflags=-d=ssa/prove/debug=2
   224   ```
   225
   226  There are some additional details about `-gcflags` and the differences between `go build`
   227  vs. `go tool compile` in a [section below](#-gcflags-and-go-build-vs-go-tool-compile).
   228
   229* In general, when investigating a problem in the compiler you usually want to
   230  start with the simplest possible reproduction and understand exactly what is
   231  happening with it.
   232
   233#### Testing your changes
   234
   235* Be sure to read the [Quickly testing your changes](https://go.dev/doc/contribute#quick_test)
   236  section of the Go Contribution Guide.
   237
   238* Some tests live within the cmd/compile packages and can be run by `go test ./...` or similar,
   239  but many cmd/compile tests are in the top-level
   240  [test](https://github.com/golang/go/tree/master/test) directory:
   241
   242  ```
   243  $ go test cmd/internal/testdir                           # all tests in 'test' dir
   244  $ go test cmd/internal/testdir -run='Test/escape.*.go'   # test specific files in 'test' dir
   245  ```
   246  For details, see the [testdir README](https://github.com/golang/go/tree/master/test#readme).
   247  The `errorCheck` method in [testdir_test.go](https://github.com/golang/go/blob/master/src/cmd/internal/testdir/testdir_test.go)
   248  is helpful for a description of the `ERROR` comments used in many of those tests.
   249
   250  In addition, the `go/types` package from the standard library and `cmd/compile/internal/types2`
   251  have shared tests in `src/internal/types/testdata`, and both type checkers
   252  should be checked if anything changes there.
   253
   254* The new [application-based coverage profiling](https://go.dev/testing/coverage/) can be used
   255  with the compiler, such as:
   256
   257  ```
   258  $ go install -cover -coverpkg=cmd/compile/... cmd/compile  # build compiler with coverage instrumentation
   259  $ mkdir /tmp/coverdir                                      # pick location for coverage data
   260  $ GOCOVERDIR=/tmp/coverdir go test [...]                   # use compiler, saving coverage data
   261  $ go tool covdata textfmt -i=/tmp/coverdir -o coverage.out # convert to traditional coverage format
   262  $ go tool cover -html coverage.out                         # view coverage via traditional tools
   263  ```
   264
   265#### Juggling compiler versions
   266
   267* Many of the compiler tests use the version of the `go` command found in your PATH and
   268  its corresponding `compile` binary.
   269
   270* If you are in a branch and your PATH includes `<go-repo>/bin`,
   271  doing `go install cmd/compile` will build the compiler using the code from your
   272  branch and install it to the proper location so that subsequent `go` commands
   273  like `go build` or `go test ./...` will exercise your freshly built compiler.
   274
   275* [toolstash](https://pkg.go.dev/golang.org/x/tools/cmd/toolstash) provides a way
   276  to save, run, and restore a known good copy of the Go toolchain. For example, it can be
   277  a good practice to initially build your branch, save that version of
   278  the toolchain, then restore the known good version of the tools to compile
   279  your work-in-progress version of the compiler.
   280
   281  Sample set up steps:
   282  ```
   283  $ go install golang.org/x/tools/cmd/toolstash@latest
   284  $ git clone https://go.googlesource.com/go
   285  $ export PATH=$PWD/go/bin:$PATH
   286  $ cd go/src
   287  $ git checkout -b mybranch
   288  $ ./all.bash                      # build and confirm good starting point
   289  $ toolstash save                  # save current tools
   290  ```
   291  After that, your edit/compile/test cycle can be similar to:
   292  ```
   293  [... make edits to cmd/compile source ...]
   294  $ toolstash restore && go install cmd/compile   # restore known good tools to build compiler
   295  [... 'go build', 'go test', etc. ...]           # use freshly built compiler
   296  ```
   297
   298* toolstash also allows comparing the installed vs. stashed copy of
   299  the compiler, such as if you expect equivalent behavior after a refactor.
   300  For example, to check that your changed compiler produces identical object files to
   301  the stashed compiler while building the standard library:
   302  ```
   303  $ toolstash restore && go install cmd/compile   # build latest compiler
   304  $ go build -toolexec "toolstash -cmp" -a -v std # compare latest vs. saved compiler
   305  ```
   306
   307* If versions appear to get out of sync (for example, with errors like
   308  `linked object header mismatch` with version strings like
   309  `devel go1.21-db3f952b1f`), you might need to do
   310  `toolstash restore && go install cmd/...` to update all the tools under cmd.
   311
   312#### Additional helpful tools
   313
   314* [compilebench](https://pkg.go.dev/golang.org/x/tools/cmd/compilebench) benchmarks
   315  the speed of the compiler.
   316
   317* [benchstat](https://pkg.go.dev/golang.org/x/perf/cmd/benchstat) is the standard tool
   318  for reporting performance changes resulting from compiler modifications,
   319  including whether any improvements are statistically significant:
   320  ```
   321  $ go test -bench=SomeBenchmarks -count=20 > new.txt   # use new compiler
   322  $ toolstash restore                                   # restore old compiler
   323  $ go test -bench=SomeBenchmarks -count=20 > old.txt   # use old compiler
   324  $ benchstat old.txt new.txt                           # compare old vs. new
   325  ```
   326
   327* [bent](https://pkg.go.dev/golang.org/x/benchmarks/cmd/bent) facilitates running a
   328  large set of benchmarks from various community Go projects inside a Docker container.
   329
   330* [perflock](https://github.com/aclements/perflock) helps obtain more consistent
   331  benchmark results, including by manipulating CPU frequency scaling settings on Linux.
   332
   333* [view-annotated-file](https://github.com/loov/view-annotated-file) (from the community)
   334   overlays inlining, bounds check, and escape info back onto the source code.
   335
   336* [godbolt.org](https://go.godbolt.org) is widely used to examine
   337  and share assembly output from many compilers, including the Go compiler. It can also
   338  [compare](https://go.godbolt.org/z/5Gs1G4bKG) assembly for different versions of
   339  a function or across Go compiler versions, which can be helpful for investigations and
   340  bug reports.
   341
   342#### -gcflags and 'go build' vs. 'go tool compile'
   343
   344* `-gcflags` is a go command [build flag](https://pkg.go.dev/cmd/go#hdr-Compile_packages_and_dependencies).
   345  `go build -gcflags=<args>` passes the supplied `<args>` to the underlying
   346  `compile` invocation(s) while still doing everything that the `go build` command
   347  normally does (e.g., handling the build cache, modules, and so on). In contrast,
   348  `go tool compile <args>` asks the `go` command to invoke `compile <args>` a single time
   349  without involving the standard `go build` machinery. In some cases, it can be helpful to have
   350  fewer moving parts by doing `go tool compile <args>`, such as if you have a
   351  small standalone source file that can be compiled without any assistance from `go build`.
   352  In other cases, it is more convenient to pass `-gcflags` to a build command like
   353  `go build`, `go test`, or `go install`.
   354
   355* `-gcflags` by default applies to the packages named on the command line, but can
   356  use package patterns such as `-gcflags='all=-m=1 -l'`, or multiple package patterns such as
   357  `-gcflags='all=-m=1' -gcflags='fmt=-m=2'`. For details, see the
   358  [cmd/go documentation](https://pkg.go.dev/cmd/go#hdr-Compile_packages_and_dependencies).
   359
   360### Further reading
   361
   362To dig deeper into how the SSA package works, including its passes and rules,
   363head to [cmd/compile/internal/ssa/README.md](internal/ssa/README.md).
   364
   365Finally, if something in this README or the SSA README is unclear
   366or if you have an idea for an improvement, feel free to leave a comment in
   367[issue 30074](https://go.dev/issue/30074).

View as plain text