But there's the trickier problem that static variables would be shared across all the versions - avan if you can somehow not get different functions ala sighip_setup() with the same names not to step all over each other (I'm not convinced that's possible).
Somehow the c library (libc) and the math library (libm) manage to exist on only one version - I think that's a modle we should try to follow here.
cheers M
On Tue, Oct 02, 2012 at 06:13:32PM -0400, Hans-Christoph Steiner wrote:
Since Pd manually loads the libraries (.pd_linux), it can also manually map a given function address to the s_thing in the symbol table. There is no need to load the symbols in a .pd_linux in the sense of a public shared library, and therefore no nameclashes. Pd could get the address of the new() function using dlsym() and store that wherever. This is already happening for the setup() function, so we can do the same thing to map the new() function to a symbol..
So for example, pd could map the new() function to the fully qualified name in the symbol table, i.e "vanilla-0.42.5/hip~", then only in the canvas_local namespace would the symbol "hip~" be mapped to the new() function.
.hc
On 10/02/2012 05:09 PM, Miller Puckette wrote:
I'm not sure that any of the Windows, MaacOS, and linux dynamic loading systems will support having multiple versions of a library loaded in the same address space. But here's a simpler way anyhow: libraries such as vanilla could maintain compatibility by querying the version number of the patch at run time.
In the case of hip~ I'm genuinely not sure whether the "correct" behavior would then be to revert to the old behavior for all old patches or only on request. The confusing scenario I worry about is that you have a patch with an old hip~ object in it, save it from 0.44, and then have it switch to the new behavior next time it's loaded.
I think I have found ad hoc ways to fix the other problems without breaking old patches.
cheers Miller
n Tue, Oct 02, 2012 at 11:36:47AM -0400, Hans-Christoph Steiner wrote:
I think having a compatibility version stamp in the patch is a good idea. This ties in well with the experiments I've been doing with splitting out all of the objects from pd itself. If all of the core objects are a standard library, then that means its easy to choose which version of the standard library that a patch is using. In Pd-extended, this is called the 'vanilla' lib, and its been included in some form since 0.42.
Then if a patch has a compatibility version stamp in it, Pd can automatically look to see if it has a copy of that version of the standard library, and load it. Otherwise, it would load the version closest to that, and throw a warning, or optionally that could be considered an error.
To make this work well, the key missing feature is the ability to change which loaded library an object name maps to in the canvas-local namespace. Currently, once an object name is mapped to a loaded .pd_linux, that is a global association. This is needed so that patches using different standard libs can be open at the same time.
Then making the versioned standard libs would be pretty easy, mostly just bundling the right .c files into a lib.
.hc
On 10/02/2012 11:15 AM, Miller Puckette wrote:
This is in my long-range plan but hasn't yet risen to the level of "urgent". However, this migth be a good moment to get started on this because several other backward- and even forward-imcompatible needs are also rising to the fore:
- there's a bug in hip~ - its DC gain is slightly (and possibly considerably)
greater than 1. "fixing" this will change the audio output of older patches, usually much too slightly to matter, but there will have to b a "-pre-0.44-hip" flag or something to allow strict back compatibility;
- There's no place in the pre-0.43 file format to alow specifying individual
box widths and font sizes; I put an "f" (=format) message to the canvas object in 0.43 so that in 0.44 I can make it set font size and box width and perhaps leave an opening for other formatting info.
- the upsampling inlet~ by default zero-pads its input. This is incorrect as
its DC gain is less than one. (Try using that as input to a phasor~ for instance - bad surprise!) I want to change the default so that it acts like a sample-and-hold, which I believe is an option now. To preserve back compatibility I'd keep all the "upsampling methods" in place but only change default behavior for patches with a 0.44 or later version stamped on them.
Each of these presents a different spin on the age-old issue of keeping total back compatibility in place, even when the compatibility is to preserve a big as in (1) and (3) - and arguably in the file searching too; I'm not sure whether to regard that as a bug or just over-hasty design.
cheers Miller
Here's a good sketch of the idea (http://puredata.info/dev/PdSearchPath): Proposed Functionalityfor path in paths do -- the core does this bit for loader in loaders do loader(path, library, object)
Existing Functionalityfor loader in loaders do for path in paths do -- the loader does this bit loader(path, library, object)
.hc
Pd-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list
Pd-list@iem.at mailing list UNSUBSCRIBE and account-management -> http://lists.puredata.info/listinfo/pd-list