1. Precompiled include files
- Posted by jaygade Jul 11, 2013
- 1664 views
As I'm working with the source, building and running tests, and waiting for the same files to be translated over and over again, a thought occurred to me:
Would it be a useful feature if Euphoria included precompiled .e files from include/std in object format? The files could either be built with the language itself or else they could be cached after other builds (including building the language itself). The standard library files don't change much at all either during Euphoria development or during user program development. It might save a bit of time.
Just a thought for a possible future feature.
2. Re: Precompiled include files
- Posted by mattlewis (admin) Jul 11, 2013
- 1599 views
As I'm working with the source, building and running tests, and waiting for the same files to be translated over and over again, a thought occurred to me:
Would it be a useful feature if Euphoria included precompiled .e files from include/std in object format? The files could either be built with the language itself or else they could be cached after other builds (including building the language itself). The standard library files don't change much at all either during Euphoria development or during user program development. It might save a bit of time.
Just a thought for a possible future feature.
Yes, I've thought about this some. See RoadmapBeyond:
- Integration of compiled libraries with interpreted code
- Compile a euphoria include as a dll.
- Translator adds a function that returns data about the public / exported routines, which the interpreter adds to the symbol table as special euphoria routines
- User can call compiled routines like normal euphoria routines, and the interpreter handles the call transparently
- Possibly use ifdefs to determine whether regular interpreted or compiled library to be used (if available)
- Add a lib/ dir in the euphoria directory structure as a place for the compiled libs to live, or possibly specify other locations through config file
Matt
3. Re: Precompiled include files
- Posted by useless_ Jul 11, 2013
- 1594 views
As I'm working with the source, building and running tests, and waiting for the same files to be translated over and over again, a thought occurred to me:
Would it be a useful feature if Euphoria included precompiled .e files from include/std in object format? The files could either be built with the language itself or else they could be cached after other builds (including building the language itself). The standard library files don't change much at all either during Euphoria development or during user program development. It might save a bit of time.
Just a thought for a possible future feature.
Yes, I've thought about this some. See RoadmapBeyond:
- Integration of compiled libraries with interpreted code
- Compile a euphoria include as a dll.
- Translator adds a function that returns data about the public / exported routines, which the interpreter adds to the symbol table as special euphoria routines
- User can call compiled routines like normal euphoria routines, and the interpreter handles the call transparently
- Possibly use ifdefs to determine whether regular interpreted or compiled library to be used (if available)
- Add a lib/ dir in the euphoria directory structure as a place for the compiled libs to live, or possibly specify other locations through config file
Matt
. . . . because as a dll we can unload it and reload it during runtime as much as we like to change it . . . .
useless
4. Re: Precompiled include files
- Posted by mattlewis (admin) Jul 11, 2013
- 1574 views
. . . . because as a dll we can unload it and reload it during runtime as much as we like to change it . . . .
It depends on the implementation, but probably not. My idea makes it all work seamlessly from the user's perspective, which means that it's integrated with how the symbol table is built and how code is parsed.
Matt
5. Re: Precompiled include files
- Posted by BRyan Jul 11, 2013
- 1583 views
It depends on the implementation, but probably not. My idea makes it all work seamlessly from the user's perspective, which means that it's integrated with how the symbol table is built and how code is parsed.
Matt
Isn't that the same thing as the Rob's original shrouded code ? Wasn't that originally rejected by the developers for some reason when it was requested ?
6. Re: Precompiled include files
- Posted by jimcbrown (admin) Jul 11, 2013
- 1612 views
It depends on the implementation, but probably not. My idea makes it all work seamlessly from the user's perspective, which means that it's integrated with how the symbol table is built and how code is parsed.
Matt
Isn't that the same thing as the Rob's original shrouded code ?
No. The shrouded and scrambled code was just a form of obsfucation. The interpreter (rather, the parser) would essentially unobsfucate it back into something like its original form before executing it. Rather than being more efficient like precompiled code, it instead added a step to the execution process.
Wasn't that originally rejected by the developers for some reason when it was requested ?
Shrouded code uses the same bytes that are used in other file encodings that go beyond ASCII (e.g. UTF-8). Getting the old shrouded code to work now that we're moving to support Unicode is not a trivial task.
7. Re: Precompiled include files
- Posted by mattlewis (admin) Jul 11, 2013
- 1611 views
It depends on the implementation, but probably not. My idea makes it all work seamlessly from the user's perspective, which means that it's integrated with how the symbol table is built and how code is parsed.
Isn't that the same thing as the Rob's original shrouded code ? Wasn't that originally rejected by the developers for some reason when it was requested ?
No, it's actually using a compiled library as though it was included as a normal source file. I imagine that it would probably speed up parsing time for the application and would definitely speed up run time (which shrouded code would not do).
I don't believe that Rob ever open sourced his shrouding code. I think this was due to people possibly relying on released code staying encrypted. I'm not positive about this, though, and I remember him talking a bit about how stuff was encrypted at one point. Either way, the code was never released, and I don't think there was enough interest in reviving it.
Also, changes that started around 2.5 made shrouding more difficult from an implementation standpoint. I believe that was when the front end was rewritten in euphoria and no execution was done until everything was parsed.
Matt
8. Re: Precompiled include files
- Posted by jaygade Jul 11, 2013
- 1594 views
My main point was more for translating than for other uses. Files like std/convert.e could be compiled into convert.o and then linked to the rest of the .o files when doing a translate.
Currently when doing "make" on the source, the std/include files get compiled, linked, and then discarded several times in the process.
9. Re: Precompiled include files
- Posted by _tom (admin) Jul 11, 2013
- 1596 views
Trying to understand how things work now:
- I write a program, say test.ex
- eushroud test.ex produces test.il
- eub test.il executes the "shrouded" file
The advantage of an .il file is that no parsing of source-code is needed and unused code is excluded; the result is faster startup compared to using eui test.ex .
The "current shrouding" is not the same as "historical shrouding." For example, a code number was used to make extra lines workable in the "old free" Euphoria interpreter.
It looks like Python lets you "compile" .py scripts into .pyc (which may be analogous to Euphoria .il code). (Python does not provide a real compiler like euc.) Python .pyc files skip parsing but do not improve program speed.
How practical would it be for eui to read .il files (say from a library) with the idea that some parsing time is saved?
10. Re: Precompiled include files
- Posted by mattlewis (admin) Jul 11, 2013
- 1545 views
How practical would it be for eui to read .il files (say from a library) with the idea that some parsing time is saved?
Not very. The .il files are dumps of the symbol tree (which includes all of the code). So all of the references in the .il file only work for that particular file. You can't combine them.
There has been talk in the past about coming up with a more flexible way that would be similar to compiling a file to IL, then linking various IL files together. I think that would take quite a bit of work, and I'm not sure how much of a benefit it would be.
Matt
11. Re: Precompiled include files
- Posted by SDPringle Aug 11, 2013
- 1448 views
For me it wouldn't be .il files it would be .dll files. The Euphoria functions would be compiled C code and the dlls loaded when a EUPHORIA include is included.
-- really loads compiled code from a DLL say std/convert.dll: translator doesn't really use .e file if older include std/convert.e
As the other devs know, symbol names get mangled according to Symtab values each time they are parsed and they are not the same on subsequent translations. We would want a way that always produces the same name and a way to load the symbols from the DLL while parsing.
12. Re: Precompiled include files
- Posted by jaygade Aug 11, 2013
- 1438 views
I'm not talking about .il files OR .dll (or .so or .dylib) files. I'm talking about translating and compiling std/*.e files to .c and .o files during a project build.
I'm mainly talking about building Euphoria, although the same could apply to other larger projects. The same std/*.e file gets translated into a .c file and then compiled into a .o file and then finally linked in to the project. This seems wasteful when it happens several times during a build, when these files change extremely rarely.
13. Re: Precompiled include files
- Posted by SDPringle Aug 12, 2013
- 1444 views
I'm not talking about .il files OR .dll (or .so or .dylib) files. I'm talking about translating and compiling std/*.e files to .c and .o files during a project build.
I'm mainly talking about building Euphoria, although the same could apply to other larger projects. The same std/*.e file gets translated into a .c file and then compiled into a .o file and then finally linked in to the project. This seems wasteful when it happens several times during a build, when these files change extremely rarely.
I ran seven tests looking for a bug in the interpreter and so I had compiled several build directories. I compared the function name in the C from the EUPHORIA routine screen_output and each of them had the same prefix of _56. Arguably this file didn't need to be compiled each time unless the EUPHORIA code they used had changed. Since their code and even their function names remain the same across builds. It is not only the contents of the translation but also the defines that are used that affect the object code and of course which translator you are using.
What about inlining? If routines are inlined from other dot - e files than if any included file is changed you have to recompile it. So already the library files are interdependent in this way.
14. Re: Precompiled include files
- Posted by jaygade Aug 12, 2013
- 1402 views
I don't know anything about inlining.
It does seem to me, though, that the Translator can be at least as smart as Make when it comes to which files have been modified.
It may not be a simple change, but it should be a doable change.
Scenario:
- Make is building the interpreter.
- One of the interpreter files includes std/convert.e.
- The translator accepts a command flag that says "reuse compiled library files".
- But that the library file hasn't been translated yet, or the .e file is newer than the .c or .o file.
- The translator translates and compiles the file, and saves it to a known location. It does this for all std/*.e files included in the project. (I don't whether it's worthwhile for project-specific or all .e files in a project?)
- The translator finishes its build of the interpreter and deletes all of its old files EXCEPT the compiled library files.
- Make is building the translator.
- One of the translator files includes std/convert.e
- The translator accepts a command flag that says "reuse compiled library files".
- The translator finds the precompiled .c or .o file and compares it to the .e file, and sees that it's newer.
- The translator uses the precompiled file in its build of the translator, saving cycles and time.
- The translator finishes its build of the interpreter and deletes all of its old files EXCEPT the compiled library files.
And finally, the developer makes some changes to some project file or other and does another build; but he saves a lot cycles and time because a third or so of the project is already built. If there's a conflict, there can be a 'make clean' target which removes the precompiled files from the build directory.
Again, this is just brainstorming. If this ONLY affects building of Euphoria and not other projects then it might not be worth doing.
15. Re: Precompiled include files
- Posted by mattlewis (admin) Aug 12, 2013
- 1386 views
It does seem to me, though, that the Translator can be at least as smart as Make when it comes to which files have been modified.
This was tried, and I think there are still some remnants in the translator source from this effort. If we were doing a simple translation to C, it would probably be fairly easy. However, the translator itself does a fair amount of optimization, which includes getting rid of unused code. This feeds into other optimizations. So we can't just look at when that file has changed. We also have to know if anything has changed about the way it's being used.
So, for instance, if we know that a certain function is only called with a particular value, we can replace the variable with the value constant (assuming it's an integer) and we reduce run time tests and entire code branches. The C compiler does its own optimizations, too.
This all means that we may speed up translation (done once) at the expense of slower run times and larger executables. Ultimately, we dropped the idea.
Matt
16. Re: Precompiled include files
- Posted by jaygade Aug 12, 2013
- 1430 views
That's interesting, and it doesn't seem to match my recent experiment. I created these two files and ran "euc -nobuild" on them to examine the .c files.
-- convert1.ex include std/console.e include std/convert.e object cmd cmd = command_line() if length(cmd) < 3 then display("[] requires at least 1 argument", {cmd[2]}) abort(1) end if display(to_string(cmd[3]))
-- convert2.ex include std/console.e include std/convert.e object cmd cmd = command_line() if length(cmd) < 3 then display("[] requires at least 1 argument", {cmd[2]}) abort(1) end if display(to_number(cmd[3]))
convert.c was identical in both cases, and seemed to contain all of the exported functions of convert.e, as were all of the .c files created from other std/*.e files. I don't see any stripping or optimization being performed on these files.
I may be misunderstanding, but doesn't the translator treat files on its command line differently from files "included" within another file? I'm only guessing that from experiments not from actually studying the code.
Admittedly I'm testing with 4.05 on Windows right now. Maybe my example is too small or it's different on 4.1.0 -- I'll have to test it later.
Edit: Nevermind. I just did another experiment and I see that it creates a much different file.
17. Re: Precompiled include files
- Posted by jaygade Aug 12, 2013
- 1389 views
Nevermind. I just did another experiment with very different results.
18. Re: Precompiled include files
- Posted by mattlewis (admin) Aug 12, 2013
- 1378 views
Admittedly I'm testing with 4.05 on Windows right now. Maybe my example is too small or it's different on 4.1.0 -- I'll have to test it later.
It looks like 4.1 is better at removing things. I think this is probably due to better routine_id analysis. When I tried your examples using a 4.1 translator, I checked for the respective to_string/to_number that wasn't used, and they weren't there where you'd expect them to not be there.
Matt
19. Re: Precompiled include files
- Posted by jaygade Aug 12, 2013
- 1411 views
Yeah, I just translated int.ex (with 4.05) and it created a totally different convert.c with different identifiers and everything.
The funny thing is that it was also a much smaller file -- 37K vs. 60K.
20. Re: Precompiled include files
- Posted by SDPringle Aug 13, 2013
- 1363 views
Yeah, I just translated int.ex (with 4.05) and it created a totally different convert.c with different identifiers and everything.
The funny thing is that it was also a much smaller file -- 37K vs. 60K.
Yes, the version of the translator the defined words and the defined words will effect the resulting C code. Even if you compare the convert.c in intobj and transobj using the same interpreter/translator you will find different symbols inside the routines and different names for the routines themselves. They must be named this way to avoid conflicts among routines that have the same name.
These routines are not even necessarily functionally equivalent if I understand what Matt is saying.
21. Re: Precompiled include files
- Posted by mattlewis (admin) Aug 13, 2013
- 1405 views
Yes, the version of the translator the defined words and the defined words will effect the resulting C code. Even if you compare the convert.c in intobj and transobj using the same interpreter/translator you will find different symbols inside the routines and different names for the routines themselves. They must be named this way to avoid conflicts among routines that have the same name.
These routines are not even necessarily functionally equivalent if I understand what Matt is saying.
At the C level, this can be true. Consider the following simple example:
-- foo.ex export procedure foo( object o ) if sequence( o ) then ? 1 elsif integer( o ) then ? 2 elsif atom( o ) then ? 3 end if end procedure foo( 1 ) foo( "bar" ) foo( 1.5 )
Try commenting out different calls of foo, and look at the translated result. The translator isn't smart enough to omit the if blocks, but you'll note that you get things like:
/** foo.ex:5 elsif integer( o ) then*/ _9 = 1; if (_9 == 0) { _9 = NOVALUE; goto L3; // [22] 33 } else{ _9 = NOVALUE; }
...and so the C compiler will be able to optimize this away.
Matt