Files and the Firefox Build System

February 28, 2013 at 07:45 PM | categories: Mozilla, Firefox, build system

The next time you update mozilla-central you may notice some significant changes with the build system. That's because this morning we finally landed the start of a massive overhaul of the build system! There are many end goals to this effort. The important ones for most will be faster build times and a build system that is friendlier to make changes to.

Introducing Files

If you look in the tree, you'll notice that nearly every directory now has a file. files are what we are using to define the build system. Think of them each as a descriptor that describes how to build its own part of the tree. An individual file will contain the C++ sources to compile, the headers to export, the tests to run, etc. Eventually. Currently, they are limited to directory traversal information. files essentially add a level of indirection between the build system definition and how the tree is actually built. Before files, the same metadata we are now capturing in files (or plan to capture) was captured in files. We performed simple variable substitution on these files to produce Makefile files in the object directory. These Makefile files were used by GNU Make (or Pymake on Windows) to build the tree.

As I outlined in Improving Mozilla's Build System, are suboptimal for a number of reasons. The important bit is they essentially tie us to the use of make (recursive or otherwise). We are very interested in supporting modern build systems such as Tup (the theory being they will build the tree faster).

Enter files. Storing our build configuration in files allows us to decouple the definition of the build system from the tool used to build it.

How Files Work

At the tail end of configure, the build system invokes the config.status script in the object directory. The role of config.status is to combine the information collected during configure with the build configuration obtained from files and take the necessary actions to ensure the build backend (make) can build the tree.

Before today, config.status essentially iterated over the source tree and converted files to Makefile in the object directory. Things are now slightly more complicated with files.

When config.status runs, it starts with the root from the source tree. It feeds this file into a Python interpreter. It then looks for special variables like DIRS and PARALLEL_DIRS to determine which directories contain additional files. It then descends into all the referenced directories, reading their files. While this is happening, we are converting the results of each file execution into files that make knows how to build. It also performs the to Makefile conversion like it always has. When the whole process has finished, the object directory contains a bunch of Makefile and files. make runs like it always has. The only real difference is some variables are coming from the files instead of Makefile.

This is just a brief overview, of course. If you want to know more, see the code in /python/mozbuild/mozbuild/frontend and /python/mozbuild/mozbuild/backend.

Build System Future

With the introduction of files, the intent is to eventually completely eliminate and have all the build definition live in files.

Doing this all at once would have been next to impossible. So, we decided to eliminate gradually. The first step is what landed today: essentially moving DIRS variables out of and into files. Next, we will be eliminating empty (bug 844635) and will be moving more parts of the build definition from to files. The next part to move over will likely be IDLs (bug 818246). After that, it may be exported files (EXPORTS in parlance). And repeat until we have no more in the tree.

Each migration of build definition data to files will likely occur in two phases:

  1. A largely mechanical move of the variables from to
  2. Better build backend integration resulting from the move.

In phase #1, we will essentially cut and paste variable assignments to files. make will read the same variables it does today and perform the same actions. The only difference is the values in these variables will be defined in files.

In phase #2, we will leverage the fact that our build definition now has an API. We will change our existing make backend to be more efficient. For example, we should soon be able to compile IDLs and copy exported headers without make traversing through the directory tree at build time. We will be able to do this because the traversal at pre-build time sees data from all files and with this complete world view is able to produce more optimal make files than what you would get if you recursed into multiple directories. In short: it will make the build faster.

Once we have a sufficient portion of the build definition moved to files we will be able to experiment with new build backends (like Tup), look into automatic Visual Studio project generation, and more easily experiment with different ways of building (such as precompiled headers, fewer compiler process invocations, etc). These should all contribute to faster build times.

Frequently Asked Questions

What impact will I see from this change?

If you never touched files in the tree, you should not notice anything different about how the tree builds or how the build system works. You should have nothing to fear.

The most obvious changes to the source tree are:

  1. There is a file in almost every directory now.
  2. The variables related to directory traversal (those containing DIRS in their name) are now defined in files instead of
  3. If your contains a variable that has been moved to files, make will spew an error when processing that file and the build will fail.

Will this change how I build?

It shouldn't. You should build the tree just like you always have. Most of you won't notice any differences.

If you see something weird, speak up in #build or file a bug if you are really confident it is wrong.

What are the risks to this change?

The migration of variables from to files is largely mechanical and significant portions are done manually. This can be a mind-numbing and tedious process. Not helping things is the fact that Splinter's review interface for these kinds of patches is hard to read.

This all means that there is a non-trivial risk for transcription errors. All it takes is an inverted conditional block and all of a sudden large parts of the tree are no longer built, for example.

We have established bug 846825 to investigate any oddities from the initial transfer. Developers are encouraged to help with the effort. Please spot check that your directories are still being built, tests run, etc. Pay special attention to changes made in the last 4 months as these parts of would have been bit rotted and more prone to data loss.

Parts of the tree not enabled in standard configurations are more prone to breakage due to less testing. i.e. build configurations not captured by TBPL have a higher chance of breaking.

Will this make the tree build faster?

Not yet. But eventually it will. This initial landing paves the groundwork to making the tree build faster (see the Future section above).

I see a lot of empty files!

Yes. Sorry about that. The good news is they shouldn't be empty for long. As things move from to we'll see fewer and fewer empty files. We'll also see fewer and fewer files once we start deleting empty

If you want to know why we must have empty files, it's mainly for validation. If we allowed files to be optional, how would you detect a typo in a directory name? Directory exists? What if that directory exists but isn't supposed to have a file?

You bitrotted my patches!

Yes. I'm sorry. The transition period to files could be a little messy. There will be lots of changes to and files and lots of chances for bit rot. Uplifts could be especially nasty. (Although I don't believe many uplifts involve significant changes to the build configuration.)

This all means there is a strong incentive for us to complete the transition as quickly as possible.

Can I help with the transition to files?


The transition is largely mechanical (at least phase #1). If you are interested in moving a variable or set of variables, hop in #build on IRC and speak up!

You said files are actually Python files?!

Yes they are! However, they are executed in a very tightly controlled sandbox. You can't import modules, open files, etc. UPPERCASE variable names are reserved and only a few functions are exposed. If you attempt to assign to an unknown UPPERCASE variable or assign an invalid value, an error will occur. This is already much better than Makefile because we can now detect errors earlier in the build process (rather than 15 minutes into a build).

What variables and functions are available in files?

If you run |./mach mozbuild-reference| you will see a print-out of all the variables, functions, and symbols that are exposed to the Python sandbox that files execute in. There are even tests that will fail the build if the sandbox contains symbols not in this output!

The output should be valid reSTructuredText (in case you want to convert to HTML for reading in your browser).

What if a file contains an error?

The build will break.

A lot of work has gone into making the output of errors human friendly and actionable. If you do something wrong, it won't just complain: it will tell you how to fix it!

Besides build times, how else will this improve the build system?

There are several ways!

As mentioned above, are more strict about what data is allowed to be defined. If you assign to an UPPERCASE variable, that variable must be known to the sandbox or else the assignment will error. This means that if you assign to an UPPERCASE variable, you know it has a side-effect. No more cargo culting of old, meaningless variables!

To change the behavior of files (add new variables or functions, change how makefile generation works, etc) will require changes to the code in /python/mozbuild. This code belongs squarely to the build module and requires appropriate review. A problem with Makefiles is that they have lots of foot guns by default and its easy for self-inflicted wounds to land in the tree without explicit build peer review. This problem largely goes away with files because the sandbox takes away all of make's foot guns.

The output of a execution is essentially a static data structure. It's easy to validate them for conformance. If we discover bad practices in our build definition, we can centrally add tests for them and enforce best practices.

We will also see user experience wins by moving data to files. Take mochitests for an example. We currently have several flavors (plain, browser, chrome, etc). Sometimes you cannot distinguish the flavor by the filename alone. With files, it will be easier to answer questions like "what mochitest flavor is this file?" mach could hook into this so you can type |mach mochitest path/to/file.html| instead of |mach mochitest-plain path/to/file.html|. Even better, you should just be able to type |mach path/to/test.html| and mach knows from the build definition that path/to/test.html is a plain mochitest file and assumes you want to run it. There are dozens of small development workflow wins to be gained here!

If I change a file, what happens?

If you change a file, then make should detect that it has changed and it will update the dynamically generated file and reinvoke the requested build action. This should all happen automatically (just like to Makefile conversion works automatically).

My build seems to pause for a few seconds before starting!

A change to any file will cause a full traversal of the entire tree. On modern machines, this should only take 1-3 seconds. If your source tree is not in the page cache (and you need to load files from disk) or if you are on older hardware, this could be a little slower.

This is an unfortunate side-effect of having a whole world view of the build definition. The build delay incurred by these full scans should eventually be cancelled out by build backend optimizations resulting from having this whole world view, however.

The good news is this full scan should only occur if a mozbuild file changes. And, if you are performing make recursion, it should only happen once (not in every directory). If you notice multiple scanning-related pauses during builds, please file a bug in Core :: Build Config!

Finally, we are performing the reads on a single thread currently. We can throw more cores at the task if someone codes up a patch.

What happened to

It has been sacked. was an optimization to perform all the to Makefile conversion in one go. The directory traversal performed by reading effectively replaces the role of Not only that, but the build definition is always up to date! was typically out of sync with reality and was a burden to maintain.

Did we just invent our own build system?

Kinda. We invented a generic Python sandboxing infrastructure. Then we hooked up code to populate it with variables from our build system and told it how to perform file traversal by reading specific variables set during file execution. Then we hooked up code for taking the evaluated results of all those sandboxes and convert them into make files.

Conceptually, what we invented is like GYP but with a different config file format. We have dabbled with the idea of converting the parsed build definition into GYP classes and then leveraging GYP to produce Makefiles, Ninja files, Visual Studio Projects, etc. This would an interesting experiment!

If you want to call it a build system, call it a build system. However, it is currently tightly coupled to Mozilla's needs, so you can't just use it anywhere. The concept might be worth exploring, however.

Is there anything else you'd like to share?

I think we set the record for most parts in a bug: 61. Although, they are numbered 1-17, 19-20. Part 18 has 30+ sub-parts using letters from the English and Greek alphabet for identifiers. Part 61 uses the infinity symbol as its number. See the pushlog.

Finally, I'd like to thank everyone who helped with this effort. The bug itself was only 6 months old and had active development off and on for a lot of it. Ted Mielczarek and Mike Hommey provided invaluable feedback on the core build system patches. A number of module owners stepped in to lend eyes to the mechanical conversion of their files. Last but not least, Ms2ger provided invaluable review aid on many of the patches. The work was so good that we all agreed that an Ms2ger f+ was good enough for a build peer rs! If reviewing the patches wasn't enough, Ms2ger also oversaw the tree closure and merging of the landing. I don't know how I will repay this debt.

Any more questions?

If you have more questions, drop in #build on and ask away.