1 |
-----BEGIN PGP SIGNED MESSAGE----- |
2 |
Hash: SHA1 |
3 |
|
4 |
|
5 |
|
6 |
Ciaran McCreesh wrote: |
7 |
| On Wed, 22 Sep 2004 21:40:39 -0400 John Richard Moser |
8 |
| <nigelenki@×××××××.net> wrote: |
9 |
| | | Personally, I would be *very* wary about giving our x86 users a 5% |
10 |
| | | performance hit |
11 |
| | |
12 |
| | 1. Where are you getting 5% from? |
13 |
| |
14 |
| I did some quick measurements on vim's test suite with and without |
15 |
| -fstack-protector. The value varies per app, of course, but vim's regex |
16 |
| stuff was taking a ~10% hit (not surprising) and the file ops were |
17 |
| affected considerably less. *shrug* depends upon the app, of course. |
18 |
| |
19 |
|
20 |
Mmm, interesting. |
21 |
|
22 |
Because SSP's protections are not generally applied (-fstack-protector |
23 |
only affects functions with strings), it's difficult to measure I guess. |
24 |
|
25 |
| | 2. What context is this "Performance" hit in? gcc would take a |
26 |
| | "performance" hit because it eats 100% CPU; most nothing else would |
27 |
| | take a "performance" hit unless the *overhead* pulled CPU usage up to |
28 |
| | 100% for a time. |
29 |
| |
30 |
| Well, performance when a box is at low load is of no interest... |
31 |
| |
32 |
|
33 |
Which is 99% of the time of course. :) Hell if your box is running at |
34 |
80% load performance hit is of no interest (although adding enough for |
35 |
it to run at 80% load instead of 1% is bad). |
36 |
|
37 |
| | I guess this is the point where I have to ebuild unpack nbyte and |
38 |
| | generate SSP benchmarks. Why oh why didn't I bench ssp when I was |
39 |
| | doing PIC? |
40 |
| |
41 |
| Note that some kinds of benchmarks, such as integer op tests, won't |
42 |
| be affected at all by SSP. Certain kinds of string handling, on the |
43 |
| other hand, will be hit really badly. Better to pick an app and |
44 |
| benchmark it than running dedicated benchmarkers. |
45 |
|
46 |
|
47 |
Heh. True, it's too specific to quantify I guess. |
48 |
|
49 |
The impact is also less with large functions. If you have a 3 |
50 |
instruction function, the extra SSP instructions are bigger than the |
51 |
damn function; if you have a function made of 150 lines of C code, |
52 |
obviously that's going to have little to no impact. |
53 |
| |
54 |
|
55 |
- -- |
56 |
All content of all messages exchanged herein are left in the |
57 |
Public Domain, unless otherwise explicitly stated. |
58 |
|
59 |
-----BEGIN PGP SIGNATURE----- |
60 |
Version: GnuPG v1.2.6 (GNU/Linux) |
61 |
Comment: Using GnuPG with Thunderbird - http://enigmail.mozdev.org |
62 |
|
63 |
iD8DBQFBUjPWhDd4aOud5P8RAp38AJ9UA1auQhRZ+efdMOUL8BE9Z3SZwACePwr2 |
64 |
Y/wD9+B75+MNoUEd/4C39Tk= |
65 |
=NxBt |
66 |
-----END PGP SIGNATURE----- |
67 |
|
68 |
-- |
69 |
gentoo-dev@g.o mailing list |