1 |
With gcc, is there a way to enable all -O3 options but function |
2 |
inlines? Would -fno-inline work or something like that? |
3 |
|
4 |
Mario Domenech Goulart wrote: |
5 |
|
6 |
>Hello, |
7 |
> |
8 |
>There's an interesting discussion in the OpenBSD mailing |
9 |
>list about the use of inline. |
10 |
> |
11 |
>Here's the beginning of the thread about this topic: |
12 |
> |
13 |
>,----[ http://www.sigmasoft.com/cgi-bin/wilma_hiliter/openbsd-tech/200407/msg00175.html ] |
14 |
>| inline considered harmful. |
15 |
>| |
16 |
>| * To: tech@×××××××.org |
17 |
>| * Subject: inline considered harmful. |
18 |
>| * From: Artur Grabowski <art@××××××××.org> |
19 |
>| * Date: 21 Jul 2004 03:54:46 +0200 |
20 |
>| * User-agent: Gnus/5.09 (Gnus v5.9.0) Emacs/21.2 |
21 |
>| |
22 |
>| Today we did a bunch of removal of inline functions in the kernel. |
23 |
>| It all started to make floppies fit, but now it's a quest. |
24 |
>| |
25 |
>| If you think that I'm crazy doing this because it might hurt your |
26 |
>| precious performance, go back to your vax and leave the performance |
27 |
>| tuning to people who have a cache. |
28 |
>| |
29 |
>| Every single inline we removed today (and there are more in the |
30 |
>| pipeline and even more waiting to be fixed) shrunk the code and MADE |
31 |
>| IT FASTER. Yes, modern cpus have something called "cache". The cache |
32 |
>| prefers the code to be smaller, rather than free from function calls. |
33 |
>| Yes, some cpus have expensive function call overhead. Don't use them. |
34 |
>| i386 has quite expensive function calls, on the other hand it doesn't |
35 |
>| have any relevant amount of registers either. So a function call |
36 |
>| instead of the same function inlined can potentially make the job |
37 |
>| easier for the register allocator in the compiler which could eat the |
38 |
>| overhead. At the same time the instruction cache can run the same code |
39 |
>| in the same place, instead of loading it from main memory 4711 times. |
40 |
>| And guess what? The stack on i386 is in the cache too, so the function |
41 |
>| call overhead isn't that bad anyway. |
42 |
>| |
43 |
>| I'm tired of seeing code where everything is made inline just because |
44 |
>| someone acted on a meme that hasn't been true for over a decade. Bloat, |
45 |
>| bloat and more bloat. Since people can't use inline correctly (it does |
46 |
>| have valid and correct uses), from now on inline in the OpenBSD kernel |
47 |
>| is considered to be a bug until proven otherwise. So. Next time I see |
48 |
>| code that adds to the bloat with inlines, I expect performance figures |
49 |
>| and kernel size comparisons that show that the inline actually |
50 |
>| contributes anything. Otherwise the code does not go in. |
51 |
>| |
52 |
>| There's still a lot of work to be done in the kernel (yes, macros can |
53 |
>| be evil too, just see nfs), so send diffs. And there's a whole |
54 |
>| unexplored field in userland too. |
55 |
>| |
56 |
>| //art |
57 |
>`---- |
58 |
> |
59 |
>Mario |
60 |
> |
61 |
> |
62 |
>-- |
63 |
>gentoo-performance@g.o mailing list |
64 |
> |
65 |
> |
66 |
> |
67 |
> |
68 |
|
69 |
-- |
70 |
gentoo-performance@g.o mailing list |