bitcoin/depends/packages.md
Carl Dong 683b7d7a3f
depends: Purge libtool archives
We use pkg-config where we can, which generally replaces libtool at a
higher level and does not have the same downsides as libtool. These
archives sit in our depends tree with no purpose and pollute the final
bitcoin build with massive overlinking.
2019-05-28 11:02:42 -04:00

6.7 KiB

Each recipe consists of 3 main parts: defining identifiers, setting build variables, and defining build commands.

The package "mylib" will be used here as an example

General tips:

  • mylib_foo is written as $(package)_foo in order to make recipes more similar.
  • Secondary dependency packages relative to the bitcoin binaries/libraries (i.e. those not in ALLOWED_LIBRARIES in contrib/devtools/symbol-check.py) don't need to be shared and should be built statically whenever possible. See below for more details.

Identifiers

Each package is required to define at least these variables:

$(package)_version:
Version of the upstream library or program. If there is no version, a
placeholder such as 1.0 can be used.

$(package)_download_path:
Location of the upstream source, without the file-name. Usually http, https
or ftp. Secure transmission options like https should be preferred if
available.

$(package)_file_name:
The upstream source filename available at the download path.

$(package)_sha256_hash:
The sha256 hash of the upstream file

These variables are optional:

$(package)_build_subdir:
cd to this dir before running configure/build/stage commands.

$(package)_download_file:
The file-name of the upstream source if it differs from how it should be
stored locally. This can be used to avoid storing file-names with strange
characters.

$(package)_dependencies:
Names of any other packages that this one depends on.

$(package)_patches:
Filenames of any patches needed to build the package

$(package)_extra_sources:
Any extra files that will be fetched via $(package)_fetch_cmds. These are
specified so that they can be fetched and verified via 'make download'.

Build Variables:

After defining the main identifiers, build variables may be added or customized before running the build commands. They should be added to a function called $(package)_set_vars. For example:

define $(package)_set_vars
...
endef

Most variables can be prefixed with the host, architecture, or both, to make the modifications specific to that case. For example:

Universal:     $(package)_cc=gcc
Linux only:    $(package)_linux_cc=gcc
x86_64 only:       $(package)_x86_64_cc = gcc
x86_64 linux only: $(package)_x86_64_linux_cc = gcc

These variables may be set to override or append their default values.

$(package)_cc
$(package)_cxx
$(package)_objc
$(package)_objcxx
$(package)_ar
$(package)_ranlib
$(package)_libtool
$(package)_nm
$(package)_cflags
$(package)_cxxflags
$(package)_ldflags
$(package)_cppflags
$(package)_config_env
$(package)_build_env
$(package)_stage_env
$(package)_build_opts
$(package)_config_opts

The *_env variables are used to add environment variables to the respective commands.

Many variables respect a debug/release suffix as well, in order to use them for only the appropriate build config. For example:

$(package)_cflags_release = -O3
$(package)_cflags_i686_debug = -g
$(package)_config_opts_release = --disable-debug

These will be used in addition to the options that do not specify debug/release. All builds are considered to be release unless DEBUG=1 is set by the user. Other variables may be defined as needed.

Build commands:

For each build, a unique build dir and staging dir are created. For example, work/build/mylib/1.0-1adac830f6e and work/staging/mylib/1.0-1adac830f6e.

The following build commands are available for each recipe:

$(package)_fetch_cmds:
Runs from: build dir
Fetch the source file. If undefined, it will be fetched and verified
against its hash.

$(package)_extract_cmds:
Runs from: build dir
Verify the source file against its hash and extract it. If undefined, the
source is assumed to be a tarball.

$(package)_preprocess_cmds:
Runs from: build dir/$(package)_build_subdir
Preprocess the source as necessary. If undefined, does nothing.

$(package)_config_cmds:
Runs from: build dir/$(package)_build_subdir
Configure the source. If undefined, does nothing.

$(package)_build_cmds:
Runs from: build dir/$(package)_build_subdir
Build the source. If undefined, does nothing.

$(package)_stage_cmds:
Runs from: build dir/$(package)_build_subdir
Stage the build results. If undefined, does nothing.

The following variables are available for each recipe:

$(1)_staging_dir: package's destination sysroot path
$(1)_staging_prefix_dir: prefix path inside of the package's staging dir
$(1)_extract_dir: path to the package's extracted sources
$(1)_build_dir: path where configure/build/stage commands will be run
$(1)_patch_dir: path where the package's patches (if any) are found

Notes on build commands:

For packages built with autotools, $($(package)_autoconf) can be used in the configure step to (usually) correctly configure automatically. Any $($(package)_config_opts) will be appended.

Most autotools projects can be properly staged using:

$(MAKE) DESTDIR=$($(package)_staging_dir) install

Build outputs:

In general, the output of a depends package should not contain any libtool archives. Instead, the package should output .pc (pkg-config) files where possible.

From the Gentoo Wiki entry:

Libtool pulls in all direct and indirect dependencies into the .la files it creates. This leads to massive overlinking, which is toxic to the Gentoo ecosystem, as it leads to a massive number of unnecessary rebuilds.

Secondary dependencies:

Secondary dependency packages relative to the bitcoin binaries/libraries (i.e. those not in ALLOWED_LIBRARIES in contrib/devtools/symbol-check.py) don't need to be shared and should be built statically whenever possible. This improves general build reliability as illustrated by the following example:

When linking an executable against a shared library libprimary that has its own shared dependency libsecondary, we may need to specify the path to libsecondary on the link command using the -rpath/-rpath-link options, it is not sufficient to just say libprimary.

For us, it's much easier to just link a static libsecondary into a shared libprimary. Especially because in our case, we are linking against a dummy libprimary anyway that we'll throw away. We don't care if the end-user has a static or dynamic libseconday, that's not our concern. With a static libseconday, when we need to link libprimary into our executable, there's no dependency chain to worry about as libprimary has all the symbols.