diff options
author | Chris Robinson <[email protected]> | 2023-09-16 06:19:02 -0700 |
---|---|---|
committer | Chris Robinson <[email protected]> | 2023-09-16 06:19:02 -0700 |
commit | 2e293ee7cbc6f0dbd069ba7c678d20c1af16872f (patch) | |
tree | 56558220d4ffab395eef821d7aef12b4cd8c5cae /XCompile.txt | |
parent | 10f2acdabb307df25c11b8ee11f770e4e62bb5fd (diff) |
Don't inline some big functions
This is very dumb. Template functions are implicitly marked inline according to
C++, and contrary to popular belief, "inline" *does* influence the compiler's
decision to inline a function. A function the compiler may not have decided to
inline normally may be inlined anyway, or issue a warning if it still decides
not to inline, if explicitly or implicitly marked as such (or does inline it as
requested, but then decides to not inline different functions it normally would
because of a cumulative code size increase or something).
Furthermore, once a function becomes inline due to being a template function,
there's no way to undo it. Marking an inline function "noinline" pushes the
problem the other way, causing the compiler to not inline a function it may
have decided was beneficial to inline. There's no way to declare a template
function to be inlined based solely on compiler heuristics, it will always be
influenced by the implicit "inline" or explicit "noinline".
That's what's happening here. A number of functions had been combined into a
smaller number of large-ish template functions to reduce code duplication and
ease maintanence, causing them to be implicitly inline as a side-effect. GCC
then manages to inline these larger functions as implicitly requested, but in
doing so prevents other smaller functions (which are explicitly marked inline)
from being inlined due to excessive code increase and issue a warning. The
"noinline" is a heavy-handed method of un-pessimizing the optimization pass, on
the assumption the compiler apparently doesn't actually want to inline the
template functions, but does so because they're technically marked inline.
There's no good option here until it gets acknowledged that inline does mean
something beyond allowing multiple definitions, and that template (and other
types of) function definitions sometimes (if not most often) want to allow
multiple definitions but don't want an artificial/detrimental boost in inline
prioritization.
/rant
Diffstat (limited to 'XCompile.txt')
0 files changed, 0 insertions, 0 deletions