Skip to content

Latest commit

 

History

History
273 lines (220 loc) · 11.6 KB

build-overview.md

File metadata and controls

273 lines (220 loc) · 11.6 KB

Build Overview

NOTE This document should not be considered accurate until this note is removed.

This is a work-in-progress document covering the build process used by Stack. It was started following the Pantry rewrite work in Stack (likely to land as Stack 2.0), and contains some significant changes/simplifications from how things used to work. This document will likely not fully be reflected in the behavior of Stack itself until late in the Stack 2.0 development cycle.

Terminology

  • Project package: anything listed in packages in stack.yaml
  • Dependency: anything listed in extra-deps or a snapshot
  • Target: package and/or component listed on the command line to be built. Can be either project package or dependency. If none specified, automatically targets all project packages
  • Immutable package: a package which comes from Hackage, an archive, or a repository. In contrast to...
  • Mutable package: a package which comes from a local file path. The contents of such a package are assumed to mutate over time.
  • Write only database: a package database and set of executables for a given set of immutable packages. Only packages from immutable sources and which depend exclusively on other immutable packages can be in this database. NOTE formerly this was the snapshot database.
  • Mutable database: a package database and set of executables for packages which are either mutable or depend on such mutable packages. Importantly, packages in this database can be unregister, replaced, etc, depending on what happens with the source packages. NOTE formerly this was the local database.

Outdated terminology to be purged:

  • Wanted
  • Local
  • Snapshot package

Inputs

Stack pays attention to the following inputs:

  • Current working directory, used for finding the default stack.yaml file and resolving relative paths
  • The STACK_YAML environment variable
  • Command line arguments (CLI args), as will be referenced below

Given these inputs, Stack attempts the following process when performing a build.

Find the stack.yaml file

  • Check for a --stack-yaml CLI arg, and use that
  • Check for a STACK_YAML env var
  • Look for a stack.yaml in this directory or ancestor directories
  • Fall back to the default global project

This file is parsed to provide the following config values:

  • resolver (required field)
  • compiler (optional field)
  • packages (optional field, defaults to ["."])
  • extra-deps (optional field, defaults to [])
  • flags (optional field, defaults to {})
  • ghc-options (optional field, defaults to {})

flags and ghc-options break down into both by name (applied to a specific package) and general (general option * for flags is only available in CLI).

Wanted compiler, dependencies, and project packages

  • If the --resolver CLI is present, ignore the resolver and compiler config values
  • Load up the snapshot indicated by the resolver (either config value or CLI arg). This will provide:
    • A map from package name to package location, flags, GHC options, and if a package should be hidden. All package locations here are immutable.
    • A wanted compiler version, e.g. ghc-8.4.3
  • If the --compiler CLI arg is set, or the compiler config value is set (and --resolver CLI arg is not set), ignore the wanted compiler from the snapshot and use the specified wanted compiler
  • Parse extra-deps into a Map PackageName PackageLocation, containing both mutable and immutable package locations. Parse packages into a Map PackageName ProjectPackage.
  • Ensure there are no duplicates between these two sets of packages
  • Delete any packages from the snapshot packages that appear in packages or extra-deps
  • Perform a left biased union between the immutable extra-deps values and the snapshot packages. Ignore any settings in the snapshot packages that have been replaced.
  • Apply the flags and ghc-options by name to these packages overwriting any previous values coming from a snapshot. If any values are specified but no matching package is found, it's an error. If a flag is not defined in the corresponding package cabal file, it's an error.
  • We are now left with the following:
    • A wanted compiler version
    • A map from package name to immutable packages with package config (flags, GHC options, hidden)
    • A map from package name to mutable packages as dependencies with package config
    • A map from package name to mutable packages as project packages with package config

Get actual compiler

Use the wanted compiler and various other Stack config values (not all listed here) to find the actual compiler, potentially installing it in the process.

Global package sources

With the actual compiler discovered, list out the packages available in its database and create a map from package name to version/GhcPkgId. Remove any packages from this map which are present in one of the other three maps mentioned above.

Resolve targets

Take the CLI args for targets as raw text values and turn them into actual targets.

  • Do a basic parse of the values into one of the following:
    • Package name
    • Package identifier
    • Package name + component
    • Directory
  • An empty target list is equivalent to listing the package names of all project packages
  • For any directories specified, find all project packages in that directory or subdirectories therefore and convert to those package names
  • For all package identifiers, ensure that either the package name does not exist in any of the three parsed maps from the "wanted compiler" step above, or that the package is present as an immutable dependency from Hackage. If so, create an immutable dependency entry with default flags, GHC options, and hidden status, and add this package to the set of immutable package dependencies.
  • For all package names, ensure the package is in one of the four maps we have, and if so add to either the dependency or project package target set.
  • For all package name + component, ensure that the package is a project package, and add that package + component to the set of project targets.
  • Ensure that no target has been specified multiple times. (FIXME Mihai states: I think we will need an extra consistency step for internal libraries. Sometimes stack needs to use the mangled name (z-package-internallibname-z..), sometimes the package:internallibname one. But I think this will become obvious when doing the code changes.)

We now have an update four package maps, a new set of dependency targets, and a new set of project package targets (potentially with specific components).

Apply named CLI flags

Named CLI flags are applied to specific packages by updating the config in one of the four maps. If a flag is specified and no package is found, it's an error. Note that flag settings are added on top of previous settings in this case, and does not replace them. That is, if previously we have singleton (FlagName "foo") True and now add singleton (FlagName "bar") True, both foo and bar will now be true. If any flags are specified but no matching package is found, it's an error. If a flag is not defined in the corresponding package cabal file, it's an error.

Apply CLI GHC options

CLI GHC options are applied as general GHC options according to apply-ghc-options setting.

Apply general flags from CLI

--flag *:flagname[:bool] specified on the CLI are applied to any project package which uses that flag name.

Apply general GHC options

General options are divided into the following categories:

  • $locals is deprecated, it's now a synonym for $project
  • $project applies to all project packages, not to any dependencies
  • $targets applies to all project packages that are targets, not to any dependencies or non-target project packages. This is the default option for apply-ghc-options
  • $everything applies to all packages in the source map excluding global packages

These options get applied to any corresponding packages in the source map. If some GHC options already exist for such a package then they get prepended otherwise they get used as is.

Determine snapshot hash

Use some deterministic binary serialization and SHA256 thereof to get a hash of the following information:

  • Actual compiler (GHC version, path, FIXME probably some other unique info from GHC, I've heard that ghc --info gives you something)
  • Global database map
  • Immutable dependency map

Motivation: Any package built from the immutable dependency map and installed in this database will never need to be rebuilt.

FIXME Caveat: do we need to take profiling settings into account here? How about Haddock status?

Determine actual target components

  • Dependencies: "default" components (all libraries and executables)
  • Project packages:
    • If specific components named: only those, plus any libraries present
    • If no specific components, include the following:
      • All libraries, always
      • All executables, always
      • All test suites, if --test specified on command line
      • All benchmarks, if --bench specified on command line

Construct build plan

  • Applied to every target (project package or dependency)
  • Apply flags, platform, and actual GHC version to resolve dependencies in any package analyzed
  • Include all library dependencies for all enabled components
  • Include all build tool dependencies for all enabled components (using the fun backwards compat logic for build-tools)
  • Apply the logic recursively to come up with a full build plan
  • If a task depends exclusively on immutable packages, mark it as immutable. Otherwise, it's mutable. The former go into the snapshot database, the latter into the local database.

We now have a set of tasks of packages/components to build, with full config information for each package, and dependencies that must be built first.

FIXME There's some logic to deal with cyclic dependencies between test suites and benchmarks, where a task can be broken up into individual components versus be kept as a single task. Need to document this better. Currently it's the "all in one" logic.

Unregister local modified packages

  • For all mutable packages in the set of tasks, see if any files have changed since last successful build and, if so, unregister + delete their executables
  • For anything which depends on them directly or transitively, unregister + delete their executables

Perform the tasks

  • Topological sort, find things which have no dependencies remaining
  • Check if already installed in the relevant database
    • Check package database
    • Check Stack specific "is installed" flags, necessary for non-library packages
    • For project packages, need to also check which components were built, if tests were run, if we need to rerun tests, etc
  • If all good: do nothing
  • Otherwise, for immutable tasks: check the precompiled cache for an identical package installation (same GHC, dependencies, etc). If present: copy that over, and we're done.
  • Otherwise, perform the build, register, write to the Stack specific "is installed" stuff, and (for immutable tasks) register to the precompiled cache

"Perform the build" consists of:

  • Do a cabal configure, if needed
  • Build the desired components
  • For all test suites built, unless "no rerun tests" logic is on and we already ran the test, or "no run tests" is on, run the test
  • For all benchmarks built, unless "no run benchmarks" is on, run the benchmark