1 |
On Fri, 2006-06-09 at 20:50 +0200, Hemmann, Volker Armin wrote: |
2 |
|
3 |
> > If -ffast-math is filtered or stripped out, there is no harm in leaving |
4 |
> > it in CFLAGS, right? |
5 |
> |
6 |
> no, because a lot of other flags are filtered too! |
7 |
|
8 |
What??? Whether only -ffast-math is filtered or all flags are filtered, |
9 |
-ffast-math still filtered. I don't see where the harm lies. |
10 |
|
11 |
> Some ebuilds filter ALL flags, only because some ricers had to use ffast-math |
12 |
> and other 'controversial' flags. |
13 |
> |
14 |
> And some day, you will hit a package, that will built a damaged binary doing |
15 |
> the wrong stuff, that does not filter out, because nobody knew about it. And |
16 |
> then you are f*. |
17 |
|
18 |
> |
19 |
> > |
20 |
> > But, on my system, only 14 packages filter out -ffast-math: gnubg |
21 |
> > postgresql pgcluster libpq zoom mpfr gmp octave openoffice gsl goffice |
22 |
> > rrdtool xv gimp. None strip it out. So, the huge majority of the |
23 |
> > packages on my system are compiled with -ffast-math, unless I've made a |
24 |
> > mistake in grepping for "fast-math" in ebuilds that contain |
25 |
> > "filter-flags". |
26 |
> |
27 |
> and look for 'strip-flags' |
28 |
|
29 |
There are only small number of ebuilds that strip out all flags. |
30 |
Certainly less than 30. |
31 |
|
32 |
> |
33 |
> |
34 |
> > http://gcc.gnu.org/ml/gcc/2004-03/msg01511.html |
35 |
> > For me, it's very simple: If -ffast-math leads to answers that are |
36 |
> > less accurate (in the verification-against-observation-sense) or unphysical |
37 |
> > (against theoretical limit analysis), then I'll inspect my Fortran code and |
38 |
> > repair the formulation (either a single expression or the layout of loop |
39 |
> > code) that is responsible for the mayhem. |
40 |
> > |
41 |
> > Under no circumstances I will give up using -ffast-math. |
42 |
> > |
43 |
> > |
44 |
> > http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/320 |
45 |
> >35.pdf -ffast-math is recommended by AMD. |
46 |
> |
47 |
> FOR SPECIAL APPS. AND THIS DOCUMENT IS MEANT FOR PROGRAMMERS! PEOPLE WHO KNOW |
48 |
> THEIR CODE! |
49 |
|
50 |
I didn't see anywhere that the AMD document is for special applications |
51 |
only. Where does it say that? |
52 |
|
53 |
What it does say is that applications the rely on exact implementation |
54 |
of IEEE rules or specifications for floating-point behavior shouldn't |
55 |
use -ffast-math. |
56 |
|
57 |
> |
58 |
> > |
59 |
> > These opinions seem to contradict your advice. Could you be more |
60 |
> > specific about why -ffast-math should not be used? |
61 |
> |
62 |
> no, they don't contradict me. You did not understand them. |
63 |
> |
64 |
> Nowhere does AMD tells endusers to build their complete system with that flag. |
65 |
|
66 |
Chapter 1 Introduction, Section 1.1 Audience says: |
67 |
|
68 |
This document is intended for ISVs and end-users of the AMD |
69 |
Athlon 64 processor-based platforms... |
70 |
|
71 |
It's clear that you think using -ffast-math is dangerous. I don't, so I |
72 |
plan on continuing to use it. So far, the evidence from my system is in |
73 |
my favor. (I did run across two more instances, just today, where, in |
74 |
addition, to busybox, ebuilds fail with -ffast-math: cdparanoia and |
75 |
pcmciautils.) |
76 |
|
77 |
What would make me change my mind is some evidence, or some authority |
78 |
who could convincingly explain why the experiences of those who do use |
79 |
-ffast-math is anomalous. SHOUTING OR MERELY ASSERTING THAT IT'S A |
80 |
PROBLEM is not really convincing. |
81 |
|
82 |
How about continuing this discussion offline? |
83 |
|
84 |
--- Vladimir |
85 |
|
86 |
-- |
87 |
Vladimir G. Ivanovic <vgivanovic@×××××××.net> |
88 |
-- |
89 |
gentoo-amd64@g.o mailing list |