Document prior experiences with headers
Several developers have spent time on the problem of "bmcweb makes too
many uses of headers". This document is to attempt to document those
cases, such that others don't duplicate time. If this document is
successful, it will eventually be deleted when we solve this issue.
Tested: Documentation only.
Signed-off-by: Ed Tanous <edtanous@google.com>
Change-Id: I64c84100abbf542d68925060c2f4fe6f6bff1402
diff --git a/HEADERS.md b/HEADERS.md
new file mode 100644
index 0000000..f43cdfd
--- /dev/null
+++ b/HEADERS.md
@@ -0,0 +1,50 @@
+## Why does bmcweb use so many headers? My build times are slow!##
+
+TL; DR, History
+
+bmcweb at one point was a crow-based project. Evidence of this can still be
+seen in the http/.hpp files that still contain references to the crow
+namespaces. Crow makes heavy use of headers and template meta programming, and
+doesn't ship any cpp or implementation files, choosing to put everything in
+include once headers. As bmcweb evolved, it needed more capabilities, so the
+core was ported to Boost Beast, and what remains has very little similarity to
+crow anymore. Boost::beast at the time we ported took the same opinion,
+relying on header files and almost no implementation compile units. A large
+amount of the compile time is taken up in boost::beast template instantiations,
+specifically for boost::beast::http::message (ie Request and Response).
+
+The initial solution that gets proposed is to just move everything as it exists
+to separate compile units, making no other changes. This has been proposed and
+implemented 3-4 times in the project, the latest of which is below. The intent
+of this document is largely to save effort for the next person, so they can at
+least start from the existing prior attempts.
+
+https://gerrit.openbmc.org/c/openbmc/bmcweb/+/49039
+
+Moving to cpp files without handling any architecture has the net result of
+making total compilation slower, not faster, as the slowest-to-compile parts end
+up getting compiled multiple times, then the duplicates deleted at link time.
+This isn't great for the end result.
+
+To actually effect the result that we'd like to see from multiple compile units,
+there have been proposed a few ideas might provide some relief;
+
+- Moving the Request and Response containers to opaque structures, so a majority
+ of code only needs to #include the interface, not any of the template code.
+ https://gerrit.openbmc.org/c/openbmc/bmcweb/+/37445
+ Doing this exposed a number of mediocre practices in the route handlers,
+ where routes made copies of requests/responses, relied on APIs that should've
+ been internal, and other practices that make this migration less
+ straightforward, but is still being pursued by maintainers over time.
+- Moving the internals of Request/Response/Connection to rely on something like
+ [http::proto](https://github.com/CPPAlliance/http_proto) which, written by the
+ same author as boost::beast, claims to have significant reduction in compile
+ time templates, and might not require abstracting the Request/Response
+ objects.
+- Reduce the bmcweb binary size to the point where link time optimization is not
+ required for most usages. About half of the bmcweb build time is spent doing
+ link time optimization, which, as of this time is required to keep bmcweb code
+ small enough to deploy on an actual BMCs (See DEVELOPING.md for details).
+ One could theoretically determine the source of where LTO decreases the binary
+ size the most, and ensure that those were all in the same compile unit, such
+ that they got optimized without requiring LTO.