Update 2022-08-02 The -Wmax-tokens experiment was retired as the downsides (annoyance to developers) were determined to outweigh the benefits (preventing include bloat), especially in the face of libc++ rolls. We now have better include graph analysis and tracking of total build size over time. The goal is to have a Gerrit warning for flagging include bloat growth in the future.
This is an experiment that uses the compiler to limit the size of certain header files as a way of tackling #include bloat.
The main idea is to insert “back stops” for header sizes to ensure that header cleanups don't regress without anyone noticing.
A back stop is implementing using this pragma:
#pragma clang max_tokens_here <num>
It makes the compiler issue a -Wmax-tokens warning (which we treat as an error) if the number of tokens after preprocessing at that point in the translation unit exceeds num.
For more background and data, see clang -Wmax-tokens: A Backstop for #include Bloat.
The purpose of the pragma is to avoid accidentally growing the size of widely included headers. That is, they are there to make developers aware of an increase, not to block it.
There are two common scenarios for hitting the pragma:
An incremental increase in header size. If the increase is small, don't worry about it (unless you want to), and just bump the limit.
If the increase is significant, say a doubling of the size or more, consider using techniques such as forward declarations to avoid increasing the header size. Even complex classes may have forward declarations available, see for example https://source.chromium.org/chromium/chromium/src/+/HEAD:base/functional/callback_forward.h. Many types defined in .mojom.h files have forward declarations in a corresponding .mojom-forward.h file. If the size increase is unavoidable, raise the limit.
To limit the size of a header foo.h, insert the pragma directly after including the header into its corresponding .cc file:
// in foo.cc: #include "foo.h" #pragma clang max_tokens_here 123456 #include "blah.h"
By inserting the pragma at this point, it captures the token count from including the header in isolation.
In order to not create unnecessary friction for developers, token limits should only be imposed on headers with significant impact on build times: headers that are (or were) large and widely included.
System headers on Chrome OS differ between boards and are not covered by the commit queue. This means the token limits were not tailored to those builds, causing build problems downstream. To avoid this, the -Wmax-tokens warning was disabled for Chrome OS (see crbug.com/1079053).
Good examples of -Wmax-tokens preventing regressions: