1 |
On Sat, 26 Jul 2008 18:54:20 +0200 |
2 |
Arfrever Frehtes Taifersar Arahesis <arfrever.fta@×××××.com> wrote: |
3 |
> Respecting LDFLAGS provides also some some degree of optimization. |
4 |
|
5 |
It's a *very* small degree, and certainly nowhere near on the scale of |
6 |
the difference made by CFLAGS on some archs. |
7 |
|
8 |
If CFLAGS only made the kind of difference that it made on x86 on other |
9 |
archs, packages ignoring CFLAGS wouldn't be considered a big deal. It's |
10 |
only because ignoring CFLAGS results in either certain types of |
11 |
performance-critical code being *a factor of ten* slower or binaries |
12 |
that won't run that it matters. |
13 |
|
14 |
(As to why you get that factor of ten: gcc on sparc builds v7 code |
15 |
unless you tell it not to. v7 has no hardware integer multiply or |
16 |
divide, and doing it manually takes something like twenty clock cycles. |
17 |
v9 and later, which is what nearly everyone uses, has hardware integer |
18 |
multiply and divide, and can have several scheduled at the same time. |
19 |
For crypto, this matters a hell of a lot.) |
20 |
|
21 |
(And the breakage? Some archs need to be told to use IEEE floating |
22 |
point via a -m CFLAG, or they'll generate code using the wrong ABI.) |
23 |
|
24 |
> Potential benefits of LDFLAGS are sufficient to fix packages which |
25 |
> ignore LDFLAGS. The difference in impact is irrelevant, because even |
26 |
> bugs without any impact can be filed and should be fixed. |
27 |
|
28 |
Of all the things people could be doing, making packages honour LDFLAGS |
29 |
is an extremely minor issue at best. Whilst there's nothing wrong with |
30 |
making the changes, it's not exactly the most productive use of limited |
31 |
resources... |
32 |
|
33 |
-- |
34 |
Ciaran McCreesh |