[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Aicas again
From: |
Casey Marshall |
Subject: |
Re: Aicas again |
Date: |
Mon, 23 Jan 2006 20:57:57 -0800 |
On Jan 23, 2006, at 7:02 PM, Dalibor Topic wrote:
If the new native layer code turns out to be something that does
more harm then good, then it should be either autoconfiscated, to
save what makes sense saving, or reverted and presumably gradually
rewritten to remove the macros for good. I am not sure if it's too
early to tell, I have not looked at the rewrite in past week myself.
One salient point, I think, is that this goes against the standard
way GNU packages are written, which seems to me to basically be to
target GNU, but try to be as portable as possible, and use autoconf
to sort out the messy bits.
And, a lot of the code I've looked at breaks GNU's style rules
(besides using macros absolutely everywhere), which I don't like.
It does make debugging much harder, so that right there is a negative.
I doubt that will change much for Darwin, either way, since Darwin
is different enough from Linux to trigger portability issues. The
same would be the case if GNU Classpath had more developers using
Cygwin, which also gets regularly broken, in my experience, and
will get regularly broken in the future no matter what approach to
portability layers we take.
I do also think there is value in just writing UNIX-portable code; it
seems, to me, to encourage better thinking about what one is writing,
without relying on some obscure OS-specific crutch. It is like when I
hear C# or Java programmers say they can't write anything lower
level, because they depend too much on features (which are often
overly generic and inefficient) tied to that environment.
Runtimes that care about niche platforms like Darwin, Cygwin,
Solaris, or eCos are at any time able to write and maintain their
own code for the native layer anyway thanks to the VM interface
layer, and several runtimes seem to be working fine that way.
And that's what Aicas should do! This target-native stuff seems to
work really great for them, since they push it so hard. But in the
end it isn't terribly portable or usable by the community, and for
the life of me I don't see what value they get out of the JNI code
that wraps all of these macros. It seems to me like writing a POSIX
layer for whatever they are targeting would be easier -- why is
`TARGET_NATIVE_ERROR_TRY_AGAIN' better than just writing your own
`#define EAGAIN?' Why can't you just call `TARGET_NATIVE_LAST_ERROR'
`errno' for platforms that don't have a real errno?
By the way, I don't want to sound like I'm just picking on Aicas
here; I should have introduced this as a discussion on just the
native layer, not this company in particular. If they had a good
suggestion on how to write the native interface, then I'd be fine
with it. I just think this one is extraordinarily bad.
I think the interesting question is where to draw the line what
needs to be included in GNU Classpath (and at what cost) and what
doesn't.
I think a nice, simple, almost reference-implementation style native
library makes sense for Classpath (for the things that make sense to
write as straight JNI code; truly VM-specific stuff needn't be
included). As far as targets go, popular platforms that free software
users use is best -- and that may not include Darwin, but GNU/Linux
and the sane BSD's should be supported.
I believe we'll be able to draw a clearer line once the last native
layer experiment is over, but I haven't been following the
discussion closely enough to be able to tell if it is over by now,
or if the current state of afairs is some transition state to
something else.
Is this really mostly an experimental thing? If it is, then it has
even less business being the only way of compiling native code right
now. It should have been done on a branch, especially since it is
continuously breaking things.