On 07/16/2012 08:04 PM, Eli Zaretskii wrote:
> we need a facility for doing this
> only with a few functions that affect performance.

It turns out that when compiling with -O0, always_inline
functions are often slower than macros, as the inlined code
also contains unnecessary instructions to copy arguments and
results.  So instead of using always_inline, it's better to
do this performance-critical inlining by hand.  I did that
(patch relative to trunk bzr 109195 attached), inlining
enough so that CPU performance improved by 8.7% compared to
the current trunk, when compiled with gcc -O0.  (This is the
same benchmark as before, on x86-64 with GCC 4.7.1.)  The
performance win is because I inlined a bit more cleverly
than the current code does.

Like the earlier version, this patch should improve
performance slightly in the default-optimization case too,
since this patch is identical to the earlier one when
default optimization is used.

In short, it should take only a relatively small amount of
hand-inlining to address the -O0 performance issue.