tree: 3723b3e6cef088bb7ad53ecdbb88e441358775f6 [path history] [tgz]
  1. 7z/

Automatic updates of the Windows toolchain

On the consumer side, e.g. in Chromium src.git:

  • update is called early during DEPS. Update() asks depot_tools to put in place a particular version of the toolchain (whatever src will currently build with). src provides an output .json file, where Update() saves relevant information about the toolchain, the paths, version numbers, etc.
  • Later in DEPS, build/gyp_chromium uses vs_toolchain:SetEnvironmentAndGetRuntimeDllDirs(), which loads the .json file, and uses it to set a few GYP_ variables and update the PATH to include CRT runtime directories (see below).
  • Then, gyp_chromium runs gyp generation.
  • Finally, it uses vs_toolchain again to copy runtime dlls to the output directories.

The reason the logic was split between depot_tools and src was because at some point, the bots had insufficient hard drive space and if there were > 1 build directories (say, if a build machine handled the Release and Debug builds for a given configuration) then the duplication of the toolchain in both trees would cause the bot to run out of disk space.

On the depot_tools side: takes an output .json file (per above) and an input SHA1. It tries to confirm that the user is probably a Google employee (or a bot) to encourage them to use the automatic toolchain rather than using a system installed one. It then uses gsutil to download the zip corresponding to the hash. This requires authentication with credentials, so it walks the user through that process if necessary.

(Previously in the VS2010 and early VS2013 timeframe, we also supported building with Express editions of VS. Along with this script dealt with all the complexity of acquiring the Express ISO, SDK bits, patches, etc. and applying them all in the correct sequence. However, Express no longer works, and Community is not too hard to install properly, so we just let the user do that. The primary benefit of having an automatically updated toolchain is that it works for bots, allows changes to the toolchain to be tryjob'd, reduces Infra/Labs work, and ensures that devs match bots.)

For the above convoluted reason get_toolchain_if_necessary uses to extract the zip file, but the majority of the code in there is no longer used and what remains should be inlined into get_toolchain_if_necessary in the future.

When the zip file is extracted, the mtimes of all the files, and the sha1 of the entire tree are saved to a local file. This allows future updates to compare whether the bits of the toolchain currently on disk are different than expected (the passed in SHA1), and if so, replace it with a toolchain with the correct SHA1. This is probably a bit more complicated than necessary, and again dates back to when the toolchain was assembled from many pieces. It could probably just write a stamp file with the SHA1, or just a version number, and trust that on future runs.

Finally, it copies the json file to the location that the caller requested (the json file is generated during the unzip/acquire process in

Building a .zip

Ignoring the steps to acquire a toolchain automatically from bits for Express, the procedure is roughly:

  • Get a clean Windows VM,
  • Install Visual Studio 2013 with updates as you want it,
  • Install Windows 8.1 SDK,
  • Run,
  • Upload the resulting zip file to the chrome-wintoolchain GS bucket.

That script first builds a zip file of the required pieces, including generating a batch file corresponding to SetEnv.cmd or vcvarsall.bat. It then extracts that zip to a temporary location and calculates the SHA1 in the same way that the depot_tools update procedure would do, so that it knows what to rename the zip file to.