Re: Speed up _cpp_clean_line and _cpp_skip_block_comment

["Zack Weinberg" <zack at codesourcery dot com>]

Neil Booth <neil@daikokuya.co.uk> writes:

> I see.  It looked like 3 sets of 3 columns; Mike made clear that the
> spaces were spurious.

Ah.  The spaces are kcachegrind's convention for thousands separators;
it uses thinner spaces, though, so it's clearer.  (Can't do that with
fixed width fonts.)

I did some more tests, cooking up an input file that has lots of phase
4 tokens but no macro expansion.  In this case _cpp_lex_direct itself
is now the top of the flat profile with 25% of runtime. _cpp_clean_line
is still responsible for most of the L2 cache misses, but then it would
be; I suppose we could play with __builtin_prefetch.  However, it only
takes 10% or so of runtime.  It is also interesting to observe that in
the call chain cpp_get_token -> _cpp_lex_token -> _cpp_lex_direct, the
higher layers have something like 14% overhead.

I have some tweaks in mind for _cpp_clean_line still, but no ideas for
_cpp_lex_direct and its callers.  Maybe you can think of something?
Or maybe that's as tuned as it gets.

zw