Reference 10

From PowerMops
Revision as of 07:30, 26 August 2006 by Nao-sacrada (Talk)

Jump to: navigation, search


PowerMops

Introduction

PowerMops is native code Mops for the PowerPC. We've tried to make it as close as possible to 68k Mops ('Mops Classic'?) at the source code level, so that you can compile for either platform without difficulty. However there are inevitably a few differences, which we'll discuss here.

Separate Code and Data

The PowerPC microprocessor uses the idea of code and data being in different areas. Under certain conditions the code area might be read-only. So, in PowerMops, we keep code and data separate. This shouldn't normally affect your source code, but it will if you do certain nonstandard things.

The most common of these will probably be ticking a word, then assuming the resulting execution token (xt) is an address, and making assumptions about what is stored at that address. This is not permitted by the ANSI standard — in fact, you shouldn't assume that an xt is an address at all (even though in Mops, it is).

So, if your code has anything like this:

['] myValue w@

in order to fetch the top 2 bytes of a Value, this will work in 68k Mops, but not in PowerMops. A Value in PowerMops has a header in the code area of the dictionary, but the 4-byte value itself doesn't follow the header, since it has to be in the data area. So, in the code area we put a relocatable address pointing to the value itself which is in the data area. If you tick a Value in PowerMops, you don't even get the address of this pointer; rather, you get this address minus two. The lesson to learn from this is that you shouldn't do this kind of thing at all. Stick to the standard ways of using Values, Variables and objects, and your code should work in both versions of Mops.

No Separate Nucleus and Dictionary

With PowerMops we're abandoning the idea of separate nucleus and dictionary files. This came from a time when everything had to fit on a floppies, and would have added pointless complexity to the PowerMops architecture. So in PowerMops, when you do a save, you get a new application generated, which has everything in it, and you just double-click on it.

Under the Hood

If you want to get an idea of how things work under the hood, at the moment all I can do is refer you to the comments in the code. These, however, are fairly extensive, and should be up-to-date, since I refer to them all the time myself!

The code generator is written entirely in Mops. The vast majority of words are high-level, and for the few assembly words I use the Mops PPC assembler.

The comments in the code, while reasonably adequate, are scattered around, generally near the relevant words, rather than all in one place. In future I may collect them together and include them in an appendix to the manual.

But if you're interested, at the end of the first code generator file 'cg1', there's a full rundown on our basic code generation strategy, register usage, etc. as well as a full description of the PPC version of our class and object formats.

Toolbox Calls

Remember that for Toolbox calls you MUST use the SYSCALL syntax. The old ˜call xxxx' syntax isn't implemented in PowerMops.

As we noted in the Tutorial, if you're passing records to the Toolbox, you have to use the syntax

68k_record{ ... }

This is because the normal record{ ... } uses PowerPC-optimized alignment of fields, which is different to the 68k. The PowerPC prefers information of 4 bytes or more to start at addresses that are multiples of 4, while the 68k is happy with multiples of 2. However, Mac Toolbox calls, for compatibility, all still use 68k-style alignment.

You'll notice that I've used this syntax in files such as Event which pass record addresses to the Toolbox. If your code crashes on a Toolbox call, this is one thing to check.

On the 68k Mops, 68k_record{ is recognized, but treated the same as record{, so you can always use 68k_record{ for records you pass to the Toolbox, whatever version of Mops you're running.

Source Filename Conventions and Loading Order

Files ˜cgx' belong to the code generator proper. During the building of the PowerMops system, these get loaded twice”the first time is so that we can generate PowerPC code, and the second is so that the code generator itself can exist in a native version.

So the basic order of loading is, first the code generator is loaded, then a word CROSS is executed which ˜crosses' us over into PowerPC-land. This begins to lay down a memory image of the native PowerMops application. Then the nucleus is generated, and then the (native version of) the code generator. (This is what is called ˜target compilation' since we're compiling code for a different target machine than the one we're running on”we're running on a 68k-based machine, but compiling code for a PowerPC machine, even though they might actually be the same machine.)

Then finally a word WRITE_PEF is executed which writes out the native application so far generated, as a PEF file (which is the normal Mac OS format for a native PowerPC application).

Subsequent generation of the full PowerMops system consists of running the new native application, and loading all the other necessary files in fully native mode.

Before CROSS-time, we have to redefine some immediate words such as IF, so that although they are still running on the 68k (which they need to be, since we're going to execute them!), they will generate PowerPC code. These are in the files qpCond, qCase, qpCreate, qBase, qpClass and qArgs. Notice that all these filenames start with ˜q'. This is just an arbitrary convention to make things a bit more understandable. Then we execute CROSS and build the nucleus'these files are setup, pnuc1, pnuc2, pnuc3 and pnuc4. (CROSS itself is at the start of setup.)

Then we need to load the immediate words again, but this time they'll be native. This is done in the files pBase, pArgs and qpClass. Note that these files start with a ˜pexcept for qpClass, which gets loaded at ˜q' time as well. I have been able to use conditional compilation of the form PPC? [IF]...[ELSE]...[THEN] to be able to get most of the Class-related code into one file. (The flag PPC? get set true by CROSS).

Then while we're still target compiling, we need to load other Mops files up to where class File is defined, otherwise our new PowerPC application wouldn't be able to save itself to disk. So we have files pStruct, pString, pBytestring and pFiles.

Then after that, we target compile the code generator, by loading the files cgx again. Then we write out the PEF file for the new PowerMops application, and our target compilation is finished.

Now we fire up PowerMops, and continue loading files. These are mainly the normal Mops files, but a few need a special customized version. I had to think of another initial letter for these, and since this is the last time, I used ˜z'. So we have files zBase, zArgs, zClass, zStruct, zModules and zPEF (which is a modified version of the code generator file cg4). A few of the other Mops files have ˜z' versions as well.

Performance

I expect the performance will be quite competitive with C or even assembler. If you have a look at the source, you'll notice that there are very few words written in assembly. I originally rewrote a couple of others in assembly to compare them, but found that the code that had been generated from the high-level version of those words wasn't significantly different! So I ended up only writing those words in assembly that absolutely had to be, because they needed access to special registers or whatever.

Install

This is a little simpler than in the 68k version. You don't get a separate dialog for setting the stack and dictionary sizes, since we only needed that in the 68k version for saving a new nucleus. On the PPC we assume you won't be adding to the dictionary, and you can set the heap size via a SIZE resource or by using ˜Get Info' from the Finder, once you've created your new application.

So when you install, you now go straight to the ˜second' install dialog. When you click ˜OK', you'll get a standard file dialog for saving the new application. If you change your mind about its name, you can do it here if you want to.

If you want to create a ˜FAT' application (with native code for both 68k and PPC), at the moment you'll have to build a 68k and PPC version separately. Then you can use ResEdit to copy all the resources of the 68k version to the PPC version. Here's one way you might do this:

Let's say you want your final fat application to be called ˜MyApp'. Start 68k Mops, load your source files, do Install. Call the application ˜MyApp_68k'. Use ResEdit to copy your resources over to MyApp_68k.

Now run PowerMops, load your source files again, and do Install. This time call the application ˜MyApp'. When you're finished, run ResEdit, and this time copy ALL the resources in MyApp_68k over to MyApp. This will include all your usual resources, since you copied them to MyApp_68k before, and it will also include all the CODE resources from MyApp_68k, which is where all the compiled 68k code lives.

That's all you have to do. MyApp should now be a fat application, which will run in native mode on both the 68k and PowerPC.

In a future Mops version I might provide a simpler way to handle this process, but as you can see it's not very difficult as it is.

Shared Libraries

PowerMops allows you to both call and compile shared libraries. Shared libraries are intended to replace things like extensions or INITs”they are a little like applications, but in another way they are a bit like additions to the Toolbox, in that they contain routines that can be called from any application, which can do all sorts of useful things. And although they may be being used from several applications at once, only one copy of the shared library's code need be in memory. Another feature of shared libraries is that all calls to them use the standard PowerPC calling convention, which means they can be written in any language, and called from any language.

This last point means that since PowerMops can produce a shared library, you can write code in Mops that can be called from some other language.

Calling a Shared Library

You do this in a rather similar way to calling the system. You must first declare your library, like this:

library MySharedLib

(As with SYSCALL, the library name is case sensitive.)

Then you declare all the calls you want to make to that library using the new word LIBCALL (the name is deliberately similar to SYSCALL). You use LIBCALL like this:

libcall myEntry { parm1 parm2 %parm3 -- result }

You'll notice that the syntax is similar to the declaration of named parameters. A floating parameter has ˜%' in front of it. We have to specify all this parameter and result information since Mops doesn't otherwise have it available (unlike SYSCALLs, for which we have the xcalls file). However the parameter syntax is exactly as for named parameters for normal Mops words. If the call returns a floating result, you indicate this the in way you'd expect:

libcall myFloatWord { %parm1 parm2 -- %float_result }

The library that will be used is the one given by the nearest preceding LIBRARY declaration.

The only possibilities for returned results are integer, floating, or nothing. This is determined by the C/Pascal convention, in which functions or procedures can only produce one result at most. We have to observe this convention for shared libraries, because as we mentioned above, they have to be callable from any language.

If you are writing a shared library in Mops, you are not bound by this limitation for your internal words in the code for your library. But for your ˜entry points'”the words that are to be callable from the outside world”you do have to observe this convention.

Generating a Shared Library

You generate a shared library via a new option in the Install dialog. Generating a shared library is a special case of Install, since in both cases we're producing an executable file that runs independently of the Mops development environment.

Now, you need to be able to specify which words in your shared library you want to be callable from the outside world. These are ˜entry points' to your library. You declare them like this:

 :entry  myEntry  { parm1 parm2 %parm3 -- result }
        < your code> 
 ;entry
 

That should look familiar! It's the same as the word we called with a LIBCALL above. You'll see that the syntax is basically the same. And here, the parameters not only have the same syntax as named parameters; they really ARE named parameters. You access them within the definition of myEntry in exactly the way you expect. (And of course you can have local variables in the definition as well, if you want.)


Reference 9 Reference Reference 11
Documentation