How to define package ABI compatibility

Introduction

Each conan package recipe can generate N binary packages from it, depending on three things: the package settings, the package options and the package requires.

When any of the settings of a package recipe changes, it will reference a different binary:

class MyLibConanPackage(ConanFile):
    name = "MyLib"
    version = "1.0"
    settings = "os", "arch", "compiler", "build_type"

When this package is installed, by a conanfile.txt, another package conanfile.py, or directly:

conan install MyLib/1.0@user/channel -s arch=x86_64 -s ...

The process will be:

  1. Conan will get the user input settings and options, which can come from the command line, be default values defined in ~/.conan/conan.conf, defined in profile files, or cached from the latest conan install execution.
  2. Conan will retrieve the MyLib/1.0@user/channel recipe, read the settings attribute, and assign the necessary values.
  3. With the current package values for settings (also options and requires), it will compute a SHA1 hash, that will be the binary package ID, e.g. c6d75a933080ca17eb7f076813e7fb21aaa740f2.
  4. conan will try to find the c6d75... package binary. If it’s present conan will retrieve it, if not, it can be built from sources with conan install --build.

If the package is installed again with different settings, for example, for 32bits architecture:

conan install MyLib/1.0@user/channel -s arch=x86 -s ...

The process will be repeated, but now generating a different package ID, because the arch setting will have a different value.

Also, the same process can be repeated for different compilers, compiler versions, build type, etc., generating multiple binaries, one for each configuration. Later, those binaries can be uploaded to a conan server.

When users of the package define the same settings as one of those binaries that have been uploaded, the computed package ID will be the same, such binary will be retrieved, and they will be able to reuse the binary without building it from sources.

The use case for options is very similar, the main difference is that options can be more easily defined at the package level and they can be defaulted. Check the options, default_options reference.

Note the simple scenario of a header-only library. Such package does not need to be built, and it will not have any ABI issues at all. The recipe of such package will have to generate exactly 1 binary package, no more. This is easily achieved, just by no declaring settings nor options in the recipe:

class MyLibConanPackage(ConanFile):
    name = "MyLib"
    version = "1.0"
    # no settings defined!

Then, no matter what are the settings defined by the users, which compiler or version, the package settings and options will always be the same (empty) and they will hash to the same package binary ID, that will typically contain just the header files.

What happens if we have a library that we know we can build with gcc 4.8 and we know it will remain ABI compatibility with gcc 4.9? This kind of compatibility is easier to achieve for example for pure C libraries. Although it could be argued that it is worth rebuilding with 4.9 too, to get fixes and performance improvements, lets suppose that we don’t want to create 2 different binaries, just one, built with gcc 4.8 and be able to use it from gcc 4.9 installs.

Defining a custom package_id()

The default package_id() uses the settings and options directly as defined, and assumes semver behavior for dependencies requires.

This package_id() recipe method can be overriden to control the package ID generation. Within the package_id() method we have access to the self.info object, which is the actual object being hashed for computing the binary ID:

  • self.info.settings: Contains all the declared settings, always as string values. We can access/alter the settings. E.g: self.info.settings.compiler.version
  • self.info.options: Contains all the declared options, always as string values. E.g: self.info.options.shared

Initially, this info object will contain the original settings and options, stored as strings. They can be changed without constraints, to any other string value.

For example, if you are sure your package ABI compatibility is fine for GCC versions > 4.5 and < 5.0, (just an example, not a real case) you could do this:

from conans import ConanFile, CMake, tools
from conans.model.version import Version

class PkgConan(ConanFile):
    name = "Pkg"
    version = "0.1"
    settings = "compiler", "build_type"

    def package_id(self):
        v = Version(str(self.settings.compiler.version))
        if self.settings.compiler == "gcc" and (v >= "4.5" and v < "5.0"):
            self.info.settings.compiler.version = "GCC 4 between 4.5 and 5.0"

We have set the self.info.settings.compiler.version with an arbitrary string, it’s not really important, could be any string. The only important thing is that won’t change for any GCC[4.5-5.0], for those gcc versions, it will be always the same string, and then it will be always hashed to the same ID.

Let’s check that it works properly, lets install the package for gcc 4.5:

$ conan export myuser/mychannel
$ conan install Pkg/1.0@myuser/mychannel -s compiler=gcc -s compiler.version=4.5 ...

Requirements
    Pkg/1.0@myuser/mychannel from local
Packages
    Pkg/1.0@myuser/mychannel:mychannel:af044f9619574eceb8e1cca737a64bdad88246ad
...

We can see that the computed package ID is af04...46ad (not real). What would happen if we specify GCC 4.6?

$ conan install Pkg/1.0@myuser/mychannel -s compiler=gcc -s compiler.version=4.6 ...

Requirements
    Pkg/1.0@myuser/mychannel from local
Packages
    Pkg/1.0@myuser/mychannel:mychannel:af044f9619574eceb8e1cca737a64bdad88246ad

Same result, the required package is again af04...46ad. Now we can try with GCC 4.4 (<4.5).

$ conan install Pkg/1.0@myuser/mychannel -s compiler=gcc -s compiler.version=4.6 ...

Requirements
    Pkg/1.0@myuser/mychannel from local
Packages
    Pkg/1.0@myuser/mychannel:mychannel:7d02dc01581029782b59dcc8c9783a73ab3c22dd

Now the computed package ID is different, that means that we need a different binary package for GCC 4.4

The same way we have adjusted the self.info.settings we could set the self.info.options values if necessary.

The problem of dependencies

Let’s define a simple scenario in which there are two packages, one for MyLib/1.0 which depends on (requires) MyOtherLib/2.0. The recipes and binaries for them have been created and uploaded to a conan server.

A new release for MyOtherLib/2.1 comes out, with improved recipe and new binaries. The MyLib/1.0 is modified to upgrade the requires to MyOtherLib/2.1. (Note that this is not strictly necessary, we would face the same problem if the downstream, consuming project defines a dependency to MyOtherLib/2.1, which would have precedence over the existing one in MyLib)

The question is: Is it necessary to build new MyLib/1.0 packages binaries? Or the existing packages are still valid?

The answer: it depends.

Let’s suppose that both are being compiled as static libraries, and that the API exposed by MyOtherLib to MyLib/1.0 through the public headers has not changed at all. Then, it is not necessary to build new binaries for MyLib/1.0, because the final consumer will link against both Mylib/1.0 and MyOtherLib/2.1.

It could happen that the API exposed by MyOtherLib``in public headers has changed, but without affecting the ``MyLib/1.0 binary, for whatever reasons, like changes consisting on new functions, not used by MyLib. The same reasoning would still be valid if MyOtherLib was header only.

But what if one header file of MyOtherLib, named myadd.h has changed from 2.0:

int addition (int a, int b) { return a - b; }

to the``myadd.h`` file in 2.1:

int addition (int a, int b) { return a + b; }

And the addition function is being called from compiled .cpp files of MyLib/1.0?

Then, in this case, MyLib/0.1 has to build a new binary for the new dependency version. Otherwise, it will maintain the old, buggy addition version. Even if MyLib/0.1 hasn’t change a line, not the code, neither the recipe, still the resulting binary would be different.

Using package_id() for package dependencies

The self.info object also have a requires object. It is a dictionary with the necessary information for each requirement, all direct and transitive dependencies. E.g. self.info.requires["MyOtherLib"] is a RequirementInfo object

  • Each RequirementInfo has the following read only reference fields:

    • full_name: Full require’s name. E.g MyOtherLib
    • full_version: Full require’s version. E.g 1.2
    • full_user: Full require’s user. E.g my_user
    • full_channel: Full require’s channel. E.g stable
    • full_package_id: Full require’s package ID. E.g c6d75a...
  • The following fields are the ones used in the package_id() evaluation:

    • name: By default same value as full_name. E.g MyOtherLib
    • version: By default the major version representation of the full_version. E.g 1.Y for a 1.2 full_version field and 1.Y.Z for a 1.2.3 full_version field.
    • user: By default None (doesn’t affect the package ID)
    • channel: By default None (doesn’t affect the package ID)
    • package_id: By default None (doesn’t affect the package ID)

When defining a package ID to model dependencies, it is necessary to take into account two factors:

  • The versioning schema followed by our requirements (semver?, custom?)
  • Type of library being built and type of library being reused (shared: so, dll, dylib, static).

Versioning schema

By default conan assumes semver compatibility, i.e, if a version changes from minor 2.0 to 2.1 conan will assume that the API is compatible (headers not changing), and that it is not necessary to build a new binary for it. Exactly the same for patches, changing from 2.1.10 to 2.1.11 doesn’t require a re-build. Those rules are defined by semver.

If it is necessary to change the default behavior, the applied versioning schema can be customized within the package_id() method:

from conans import ConanFile, CMake, tools
from conans.model.version import Version

class PkgConan(ConanFile):
    name = "Mylib"
    version = "1.0"
    settings = "os", "compiler", "build_type", "arch"
    requires = "MyOtherLib/2.0@lasote/stable"

    def package_id(self):
        myotherlib = self.info.requires["MyOtherLib"]

        # Any change in the MyOtherLib version will change current Package ID
        myotherlib.version = myotherlib.full_version

        # Changes in major and stable versions will change the Package ID but
        # only a MyOtherLib revision won't. E.j: From 1.2.3 to 1.2.89 won't change.
        myotherlib.version = myotherlib.full_version.minor()

Besides the version, there are some other helpers that can be used, to decide whether the channel and user of one dependency also affects the package binary, or even the required package ID can change your own package ID:

def package_id(self):
    # Default behavior, only major release changes the package ID
    self.info.requires["MyOtherLib"].semver_mode()

    # Any change in the require version will change the package ID
    self.info.requires["MyOtherLib"].full_version_mode()

    # Any change in the MyOtherLib version, user or channel will affect our package ID
    self.info.requires["MyOtherLib"].full_recipe_mode()

    # Any change in the MyOtherLib version, user or channel or Package ID will affect our package ID
    self.info.requires["MyOtherLib"].full_package_mode()

    # The requires won't affect at all to the package ID
    self.info.requires["MyOtherLib"].unrelated_mode()

You can also adjust the individual properties manually:

def package_id(self):
    myotherlib = self.info.requires["MyOtherLib"]

    # Same as myotherlib.semver_mode()
    myotherlib.name = myotherlib.full_name
    myotherlib.version = myotherlib.full_version.stable()
    myotherlib.user = myotherlib.channel = myotherlib.package_id = None

    # Only the channel (and the name) matters
    myotherlib.name = myotherlib.full_name
    myotherlib.user = myotherlib.package_id = myotherlib.version = None
    myotherlib.channel = myotherlib.full_channel

The result of the package_id() is the package ID hash, but the details can be checked in the generated conaninfo.txt file. The [requires], [options] and [settings] are those taken into account to generate the SHA1 hash for the package ID, while the [full_xxxx] fields show the complete reference information.

The default behavior produces a conaninfo.txt that looks like:

[requires]
   MyOtherLib/2.Y.Z

[full_requires]
   MyOtherLib/2.2@demo/testing:73bce3fd7eb82b2eabc19fe11317d37da81afa56

Library types: Shared, static, header only

Let’s see some examples, corresponding to common scenarios:

  • MyLib/1.0 is a shared library, linking with a static library MyOtherLib/2.0 package. When a new MyOtherLib/2.1 version is released: Do I need to create a new binary for MyLib/1.0 to link with it?

    Yes, always, because the implementation is embedded in the MyLib/1.0 shared library. If we always want to rebuild our library, even if the channel changes (we assume a channel change could mean a source code change):

def package_id(self):
    # Any change in the MyOtherLib version, user or
    # channel or Package ID will affect our package ID
    self.info.requires["MyOtherLib"].full_package_mode()
  • MyLib/1.0 is a shared library, requiring another shared library MyOtherLib/2.0 package. When a new MyOtherLib/2.1 version is released: Do I need to create a new binary for MyLib/1.0 to link with it? It depends, if the public headers have not changed at all, it is not necessary. Actually it might be necessary to consider transitive dependencies that are shared among the public headers, how they are linked and if they cross the frontiers of the API, it might also lead to incompatibilities. If public headers have changed, it would depend on what changes and how are they used in MyLib/1.0. Adding new methods to the public headers will have no impact, but changing the implementation of some functions that will be inlined when compiled from MyLib/1.0 will definitely require re-building. For this case, it could make sense:
def package_id(self):
    # Any change in the MyOtherLib version, user or channel
    # or Package ID will affect our package ID
    self.info.requires["MyOtherLib"].full_package_mode()

    # Or any change in the MyOtherLib version, user or
    # channel will affect our package ID
    self.info.requires["MyOtherLib"].full_recipe_mode()
  • MyLib/1.0 is a header-only library, linking with any kind (header, static, shared) of library in MyOtherLib/2.0 package. When a new MyOtherLib/2.1 version is released: Do I need to create a new binary for MyLib/1.0 to link with it? Never, the package should always be the same, there are no settings, no options, and in any way a dependency can affect a binary, because there is no such binary. The default behavior should be changed to:
def package_id(self):
    self.info.requires.clear()
  • MyLib/1.0 is a static library, linking with a header only library in MyOtherLib/2.0 package. When a new MyOtherLib/2.1 version is released: Do I need to create a new binary for MyLib/1.0 to link with it? It could happen that the MyOtherLib headers are strictly used in some MyLib headers, which are not compiled, but transitively #included. But in the general case it is likely that MyOtherLib headers are used in MyLib implementation files, so every change in them should imply a new binary to be built. If we know that changes in the channel never imply a source code change, because it is the way we have defined our workflow/lifecycle, we could write:
def package_id(self):

    self.info.requires["MyOtherLib"].full_package()
    self.info.requires["MyOtherLib"].channel = None # Channel doesn't change out package ID