Eli Zaretskii wrote:
> Here are my results, with today's emacs-25 and master:
>
>              emacs-25     master    emacs-25.0.92    emacs-23.3
> foo          33.375       33.89       2.859            1.375
> bar         213.250      213.578     15.547            8.625

I get somewhat different results: when comparing Emacs 24.2 to Emacs master, 
24.2 is 8% faster with -O2, 25% faster with -Og, and 3x faster with -O0. So, 
although there's a bit of slowdown otherwise (which should get looked at, at 
some point), the main culprit is -O0, and it probably is due to -O0's not 
inlining functions.

A quick build with --enable-profiling and CFLAGS=-O0 suggests that the primary 
bottlenecks with -O0 are CHAR_TABLE_REF_ASCII (18%), make_lisp_symbol (15%), 
builtin_lisp_symbol (11%), and PSEUDOVECTORP (6%).  I suppose we could spend 
some time trying to make these go faster when compiling with -O0, but I think 
our time is better spent asking developers to use -Og, or to use -O0 more 
selectively.

I used master commit 181e92c4e060a7ce4740b561375f9ec9f473f144 src/xdisp.c, with 
Emacs built on Fedora 23 x86-64 (GCC 5.3.1):

	      foo bar
   24.2   -O2  2.4  13
   24.2   -Og  3.0  17
   24.2   -O0  5.2  32
   master -O2  2.6  14
   master -Og  3.8  21
   master -O0 15    96

Fedora 23 can't build Emacs 23.3 due to some minor portability problems in 23.3, 
so I didn't benchmark that. I used the attached Lisp code to benchmark.