xref: /openbmc/bmcweb/HEADERS.md (revision dfa3fdc3)
1## Why does bmcweb use so many headers? My build times are slow!##
2
3TL; DR, History
4
5bmcweb at one point was a crow-based project. Evidence of this can still be seen
6in the http/.hpp files that still contain references to the crow namespaces.
7Crow makes heavy use of headers and template meta programming, and doesn't ship
8any cpp or implementation files, choosing to put everything in include once
9headers. As bmcweb evolved, it needed more capabilities, so the core was ported
10to Boost Beast, and what remains has very little similarity to crow anymore.
11Boost::beast at the time we ported took the same opinion, relying on header
12files and almost no implementation compile units. A large amount of the compile
13time is taken up in boost::beast template instantiations, specifically for
14boost::beast::http::message (ie Request and Response).
15
16The initial solution that gets proposed is to just move everything as it exists
17to separate compile units, making no other changes. This has been proposed and
18implemented 3-4 times in the project, the latest of which is below. The intent
19of this document is largely to save effort for the next person, so they can at
20least start from the existing prior attempts.
21
22https://gerrit.openbmc.org/c/openbmc/bmcweb/+/49039
23
24Moving to cpp files without handling any architecture has the net result of
25making total compilation slower, not faster, as the slowest-to-compile parts end
26up getting compiled multiple times, then the duplicates deleted at link time.
27This isn't great for the end result.
28
29To actually effect the result that we'd like to see from multiple compile units,
30there have been proposed a few ideas might provide some relief;
31
32- Moving the Request and Response containers to opaque structures, so a majority
33  of code only needs to #include the interface, not any of the template code.
34  https://gerrit.openbmc.org/c/openbmc/bmcweb/+/37445 Doing this exposed a
35  number of mediocre practices in the route handlers, where routes made copies
36  of requests/responses, relied on APIs that should've been internal, and other
37  practices that make this migration less straightforward, but is still being
38  pursued by maintainers over time.
39- Moving the internals of Request/Response/Connection to rely on something like
40  [http::proto](https://github.com/CPPAlliance/http_proto) which, written by the
41  same author as boost::beast, claims to have significant reduction in compile
42  time templates, and might not require abstracting the Request/Response
43  objects.
44- Reduce the bmcweb binary size to the point where link time optimization is not
45  required for most usages. About half of the bmcweb build time is spent doing
46  link time optimization, which, as of this time is required to keep bmcweb code
47  small enough to deploy on an actual BMCs (See DEVELOPING.md for details). One
48  could theoretically determine the source of where LTO decreases the binary
49  size the most, and ensure that those were all in the same compile unit, such
50  that they got optimized without requiring LTO.
51