Hi Jan,
The way I understand it is that the static libs, the .a files, aren't
actually linked yet - they are just an archive of .o object files collected
together.
When you link the app with the static lib it has to do a full read of the
symbols in the .a file and write the relevant ones into the executable.
There are 1.3 million symbols in the .a file (from nm libKit_X86_64_debug.a
| wc -l )
Adding -print_statistics to the linker flags gives this:
ld total time: 894.4 seconds ( 100.0%)
option parsing time: 11.0 milliseconds ( 0.0%)
object file processing: 0.1 milliseconds ( 0.0%)
resolve symbols: 612.3 seconds ( 68.4%)
build atom list: 0.0 milliseconds ( 0.0%)
passess: 486.3 milliseconds ( 0.0%)
write output: 281.6 seconds ( 31.4%)
pageins=447127, pageouts=2622, faults=3099715
processed 7 object files, totaling 476,732 bytes
processed 2 archive files, totaling 456,642,248 bytes
processed 32 dylib files
wrote output file totaling 324,771,884 bytes
Eg most of the time is spent reading symbols from the .a file, and then
writing them into the executable.
I also played around of using ld independantly of XCode and just trying to
link the .a file with no app and no swift involved - times were about the
same.
So to improve the linking of the .a file I think the main game would be
reducing the number of symbols in the .a. Probably not that practical.
So what I think you will get with a Framework project is actually linking
the .a file, seperately to the application executable. That means you won't
have that full link every time you build the app, but only when the
framework needs to build. That's the theory, anyway.
iOS has allowed dylibs since iOS 8 and the introduction of Swift. In fact
you can't produce a static lib with Swift, only a dylib.
For our Pdfium wrapper we produce a static lib out of the Pdfium code
itself, and link that in a framework project with the swift wrappers, to
produce a dylib which is a swift module that can be imported into an app.
Beside the linking benefits this then gives you a nice encapsulated library
that is easy to consume from swift.
If I have time over the next couple of days I'll have a go at creating a
framework to see if it does actually behave as I expect.
Cheers
Jon
On Sat, Dec 30, 2017 at 2:43 AM, jan iversen <jani@apache.org> wrote:
Please do not misunderstand my comments, I am open to any enhancements especially on the linking process of the app...but I would like to see a benefit of the changes. rgds jan I. On 29 December 2017 at 09:36, Jon Nermut <jon.nermut@gmail.com> wrote:Thanks for the reply Jan. My main point about BridgeLOKit was that you don't really need to add another FFI on top of the existing LibreOfficeKit.h FFI. Swift can import and use the existing _LibreOfficeKit / _LibreOfficeKitClass and _LibreOfficeKitDocument / _LibreOfficeKitDocumentClass structs and their functions just fine. Swift actually has excellent C integration (and non-existent C++). To do so I just added: #define LOK_USE_UNSTABLE_API 1 #import "../../../include/LibreOfficeKit/LibreOfficeKit.h" Into lokit-Bridging-Header.h, and the struct types mentioned above are usable directly within Swift without anymore C needed - LibreOfficeKitWrapper.swift is an example of using the functions from these structs directly, and wrapping the C struct pointers in swift classes, making them encapsulated and easier to use. The exception being the lok_init functions, which need to be included and called from within a C file.The LIBRARY_SEARCH_PATH should be overwritten by the xcconfig file,but I will need to check that. Oh, I couldn't find it... where is it supposed to be generated to? I changed the the path settings in LibreOfficeLight.xcodeproj like this: - LIBRARY_SEARCH_PATHS = /Users/jani/LO/core/ios/generated/; + LIBRARY_SEARCH_PATHS = $PROJECT_DIR/../generated/; I tried a few settings on the linking, couldn't make it better. Need to find a way to work out what its doing... I'll have more of a play. My suggestion is to split the C integration, and it's swift wrappers, into a separate Framework project, and let that produce a dylib. That should link pretty much instantly to the app, and should only rebuild and link when the libreoffice lib changes, or the code in the Framework Cheers Jon On Fri, Dec 29, 2017 at 6:38 PM, jan iversen <jani@apache.org> wrote:Hi adding dev list to reply, so that others might benefit from the info. Hope you don't mind the unsolicited email, I figured you were the guy totalk to about this from the git commits.I am working actively on creating a version of LO for the iPad. So I got it compiling via lode, with just a couple of hitches (had toinstall libassuan, had to make sure to use the make out of lode, and there is a hard coded LIBRARY_SEARCH_PATH to /Users/jani/... in the ios project file)I do not understand why you had to install extra libraries. I work on high sierra with xcode 9 and have not installed that library. The LIBRARY_SEARCH_PATH should be overwritten by the xcconfig file, but I will need to check that. There are 2 projects, but I assume you talk about the kit project?1. The app doesnt actually attempt to render yet? Were you planning on using CATiledLayer for that? I've used it a couple of times (for PDFs)... it's funNo it doesn’t. As you probably have seen the render function is near empty, I am strugling to find out what the tiled calls returns and how to use that in the swift app. 2. The static lib, and the compiled app, are pretty fat. (At least indebug for the simulator - ~400mb, I havent tried the release build yet). Too fat to embed in my app, it would have to be a separate app. Any insight as to whether this could ever be cut down to a reasonable size?Well is it actually quite reduced. LO is simply big. 3. The link time on the app is outrageously slow at the moment - atleast on my macbook pro - I guess this is related to the size and number of symbols in the static lib. That's what the dummy.c file is all about? Needs to be quarantined from the app somehow. Perhaps by keeping it in a Framework project? Or cutting down its size. I was too scared to turn on LTO...The link time is my biggest problem, linking the kit is a fraction of linking the app, and It seems to be the swift interface that is the problem. dummy.c is to link without the kit, and it is automatically quarantined, look in build phases, where you will see it is not being compiled. 4. Just wondering the reasoning for starting a new C interface intoLibreOfficeKit (eg BridgeLOkit_* )?How else would you make a C/C++ interface for swift ? I had success in talking to the main LibreOfficeKit.h file directly fromswift by including it in the bridging file. Using it directly would take away a lot of duplication needed to flesh out BridgeLOkit. Granted the main C api isnt that friendly to use, but IMHO it would be better to do the wrapping and making the API friendly on the Swift side, rather than another layer of C, which then still needs swift friendly classes around it. The main problem is with the way LibreOfficeKitInit works (which seems weird...), for which I reused BridgeLOkit_Init and added a func to get the pointer to kit out. See the attached LibreOfficeKitWrapper.swift file - it has just a couple wrapped functions done but you can see what I mean. Needs the rest filled in and memory handling done.Functions not declared in the bridge are unlikely to work in swift (according to the swift documentation).I've done this before for Pdfium - which also has a C based FFI. We created a framework called PdfiumSwift which had swift classes like PDFDocument, PDFPage etc which wrapped the C interface and made consuming it easy in Swift. We hooked the memory management off the swift deinit() etc. It used an internal private module to consume the C API so it was just the Swift API exposed outside of the framework / module.this is basically the same the kit project does, except it does not use classes. rgds jan iOnce the basic wrapping is done, then these classes provide a good place to add stuff like converting the raw tiles into iOS friendly bitmaps etc. Anyway, good job on getting it this far, and happy new year. Cheers Jon Nermut-- Sent from My iPad, sorry for any misspellings.