Also I have the theory that the problem may be related to some -fomit-frame-pointer hidden in some place that forces code to use references relative to esp instead of ebp. On Fri, Mar 5, 2010 at 1:20 AM, Javier Pimás <elpochodelagente@gmail.com>wrote:
Wiiiiiii. All tests passing now!!
You were right, this func had so many variables that gcc was reserving some local space to do some intermediate calculations.
To make the story short defining # define STACK_ALIGN_BYTES 16 solved it.
After a lot of debugging I discovered that alloca was aligning the stack, allocating more space than necesary, and placing the args not starting from esp but from esp plus some offset. Adding getsp(argvec) to move argvec to the top of the stack didn't solve the problem because the code that pushes the args also overwrites it (don't know why, it's like alloca augmented the frame in the middle?).
So, that define fixed everything, and
#if __APPLE__ && __MACH__ && __i386__ # define STACK_ALIGN_BYTES 16 #endif
should be changed accordingly (to something I don't know what). In my case I'm using gcc (Ubuntu 4.4.1-4ubuntu9) 4.4.1.
Thanks for all the help!
Regards, Javier.
On Wed, Mar 3, 2010 at 2:36 PM, Javier Pimás <elpochodelagente@gmail.com>wrote:
Exactly, that's what confuses me most! The function doesn't even have locals:
EXPORT(LONGLONG) ffiTestLongLong10a2(char c1, char c2, char c3, char c4, char c5, char c6, char c7, char c8, char c9, char c10, LONGLONG i1, LONGLONG i2) { return c1+c2+c3+c4+c5+c6+c7+c8+c9+c10+i1 + i2; }
I'm quite sure it's substracting because I see the memory contents, and the disassembly is this (notice how it accumulates the result in edx):
0x75d32725 <ffiTestLongLong10a2+308>: movsbl -0x1c(%ebp),%edx 0x75d32729 <ffiTestLongLong10a2+312>: movsbl -0x20(%ebp),%eax 0x75d3272d <ffiTestLongLong10a2+316>: add %eax,%edx 0x75d3272f <ffiTestLongLong10a2+318>: movsbl -0x24(%ebp),%eax 0x75d32733 <ffiTestLongLong10a2+322>: add %eax,%edx 0x75d32735 <ffiTestLongLong10a2+324>: movsbl -0x28(%ebp),%eax 0x75d32739 <ffiTestLongLong10a2+328>: add %eax,%edx 0x75d3273b <ffiTestLongLong10a2+330>: movsbl -0x2c(%ebp),%eax 0x75d3273f <ffiTestLongLong10a2+334>: add %eax,%edx 0x75d32741 <ffiTestLongLong10a2+336>: movsbl -0x30(%ebp),%eax 0x75d32745 <ffiTestLongLong10a2+340>: add %eax,%edx 0x75d32747 <ffiTestLongLong10a2+342>: movsbl -0x34(%ebp),%eax 0x75d3274b <ffiTestLongLong10a2+346>: add %eax,%edx 0x75d3274d <ffiTestLongLong10a2+348>: movsbl -0x38(%ebp),%eax 0x75d32751 <ffiTestLongLong10a2+352>: add %eax,%edx 0x75d32753 <ffiTestLongLong10a2+354>: movsbl -0x3c(%ebp),%eax 0x75d32757 <ffiTestLongLong10a2+358>: add %eax,%edx 0x75d32759 <ffiTestLongLong10a2+360>: movsbl -0x40(%ebp),%eax 0x75d3275d <ffiTestLongLong10a2+364>: lea (%edx,%eax,1),%eax 0x75d32760 <ffiTestLongLong10a2+367>: mov %eax,%edx 0x75d32762 <ffiTestLongLong10a2+369>: sar $0x1f,%edx 0x75d32765 <ffiTestLongLong10a2+372>: add -0x48(%ebp),%eax 0x75d32768 <ffiTestLongLong10a2+375>: adc -0x44(%ebp),%edx 0x75d3276b <ffiTestLongLong10a2+378>: add -0x50(%ebp),%eax 0x75d3276e <ffiTestLongLong10a2+381>: adc -0x4c(%ebp),%edx 0x75d32771 <ffiTestLongLong10a2+384>: add $0xbc,%esp 0x75d32777 <ffiTestLongLong10a2+390>: pop %ebx 0x75d32778 <ffiTestLongLong10a2+391>: pop %esi 0x75d32779 <ffiTestLongLong10a2+392>: pop %edi 0x75d3277a <ffiTestLongLong10a2+393>: pop %ebp 0x75d3277b <ffiTestLongLong10a2+394>: ret
This is really weird! Why could gcc compiled it flipped? I'll continue looking.
Regards, Javier.
2010/3/3 Eliot Miranda <eliot.miranda@gmail.com>
2010/3/3 Javier Pimás <elpochodelagente@gmail.com>
Ok, this is way nicer. The line that was missing and that seems to solve
all this is
self initializeSpecialObjectIndices.
which I think should be placed in Alien>>#initialize just before
self ensureInSpecialObjectsArray.
but then I don't know why it should ensure anything that is done just before.
With that I it's really close to work. After a few hours I understood the way tests are done. In Alien plugin you integrated a couple of exported functions (ffiTest*), which you then call in the tests. These funcs do simple stuff like adding and returning the result, so you can compare, am I right?
Then I realized that (at least in linux) the tests won't work if you compile as internal (there won't be a .so, so you won't be able to load the C test functions!).
Now with the small fix I added, and compiling as external, I can see that it's actually almost almost working, buuuut I get these test results: 36 run, 17 passed, 19 failures, 0 errors (had to remove testCallingSquenceString because it crashed the VM). Notice that before I got 20 errors and 0 failures. This is because the results are wrong.
I debugged the code with ddd to see what's happening, and I can see that everything is going nice before the actual function call, that is done in dabusiness.h.
funcAlien = interpreterProxy->stackValue(funcOffset); f = *(void **)startOfParameterData(funcAlien);
#if STACK_ALIGN_BYTES (in my compiled code it is'nt defined, should it?) /* cut stack back to start of aligned args */ setsp(argstart); #endif r = f();
I took as example test #ffiTestLongLong10a2:, which calls ffiTestLongLong10a2(char c1, char c2, char c3, char c4, char c5, char c6, char c7, char c8, char c9, char c10, LONGLONG i1, LONGLONG i2).
dabusiness.h allocated 56 bytes in argvec (10*4 for chars + 2*8 for longlong), which then fills correctly (ddd says 1 2 3 4 5 6 7 8 9 10 11 0 12 0). I have to admit that I don't fully understand whats happening with the stack then, but ffiTestLongLong10a2 gets its arguments wrong. I'm thinking of a calling convention problem. Here it's why:
In some run I have: &argvec[0] is pos 0aa0 (this is actually direction inside the stack, near it's top). Just before assembly CALL ffiTestLongLong10a2 I have ebp=0b68, and after the call it is pushed and replaced with esp, which is 0a88. Then, we have, args at 0aa0, 0aa4, ..., and current stack frame starting at 0a88 and going lower, 0a84, etc. That may be ok, but when I look into the ffiTestLongLong10a2 assembly it is trying to find the arguments by *substracting* to ebp (arg1 in ebp-1c, 2nd in ebp-20, ...) when it should adding, looking in greater positions (previous stack frame, arg1 is in ebp+18). This has to be some simple calling convention stuff, but I don't know how to solve it, any ideas?.
Subtracting from ebp is for accessing locals, adding to ebp for accessing arguments. Are you sure you're not misinterpreting a local access as an argument access?
Thanks for reading, Javier.
On Mon, Mar 1, 2010 at 9:05 PM, John M McIntosh < johnmci@smalltalkconsulting.com> wrote:
It would be missing, propose a change set.
I would think people usually don't use VMMaker images as their daily work image.
So take image, load vmaker, load vmaker stuff for alien, make alien plugin.
got to my work image that doesn't have vmmaker in it, load alien support stuf, you don't need the alien-vmmaker-support btw. run tests.
On 2010-03-01, at 3:36 PM, Javier Pimás wrote:
Hi, I'm still trying to advance with this. Don't know what caused the ObjectMemory classPools but it seems to be away now, maybe I did some mistake last time.
The thing is that now I have in ObjectMemory classPools all the original ones plus 4 new elements, which I think were added with this Alien-VMMaker-Support override:
ObjectMemory>>#initialize initialize #( #ClassAlien #ClassUnsafeAlien #InvokeCallbackSelector #SelectorAttemptToAssign) do: [:c | [ObjectMemory addClassVarName: c] ifError: []].
the thing is that it never assigns them any value, so my theory is that there's some code missing? Then in Interpreter it happens something quite similar. Interpreter>>#initialize does this:
...
#(#PrimErrBadArgument #PrimErrBadIndex #PrimErrBadNumArgs #PrimErrBadReceiver #PrimErrGenericFailure #PrimErrInappropriate #PrimErrNoCMemory #PrimErrNoMemory #PrimErrNoModification #PrimErrNotFound #PrimErrTableIndex #PrimErrUnsupported #PrimNoErr ) do: [:c | [Interpreter addClassVarName: c] ifError: []].
#(#primFailCode) do: [:i | [Interpreter addInstVarName: i] ifError: []].
...
but then never calls initializePrimitiveErrorCodes, which would set the vars to a meaningfull value.
In conclusion, the question would be, where is this missing initialization code?
Hope you can help, thanks!
Javier.
On Tue, Feb 23, 2010 at 4:39 PM, Javier Pimás < elpochodelagente@gmail.com> wrote:
On Tue, Feb 23, 2010 at 12:19 AM, John M McIntosh < johnmci@smalltalkconsulting.com> wrote:
On 2010-02-22, at 6:48 PM, Javier Pimás wrote:
Nice!!!!!!!! It's compiling now. Now, I loaded tests, and here are the results:
37 run, 17 passed, 0 failures, 20 errors.
TestCallingSequenceChar10Long2 TestCallingSequenceChar2LongLong2 TestCallingSequenceChar8Long2 TestCallingSequenceChar9Long2 TestCallingSequenceCharLongLong2 TestCallingSequenceChars TestCallingSequenceDoubles14 TestCallingSequenceDoubles2 TestCallingSequenceFloats13 TestCallingSequenceFloats14 TestCallingSequenceFloats2 TestCallingSequenceFloats2WithInteger TestCallingSequenceFloats2WithInteger2 TestCallingSequenceFloats7 TestCallingSequenceInt TestCallingSequenceInt8 TestCallingSequenceIntWithFloatArgs TestCallingSequenceLongLong2 TestCallingSequenceShort TestCallingSequenceString
Oh look SUnits, great stuff (*cough* well I wrote most of them).
Cool, this is really really useful (although I sometimes find wrting tests boring ;) ).
It is failing in places where it does primLoadLibrary: 'IA32ABI'. Why should it try to load itself, if it's compiled as an internal plugin? I compiled it as external too but didn't work either.
Well it compiled, but that doesn't mean it works. In fact the error means the plugin code is never loaded, or is callable. Now one thing to consider is that the plugin load fails, because it can't find it. (external usage). Or because (internal and external plugin) the VM Version and the plugin version don't match.
Look for your definition of #define VM_PROXY_MINOR 8
It should be 8 or higher for compiling BOTH the VM and the Plugin. Since the sqVirtualMachine.h is in the IA32ABI folder maybe there is mass confusion about which header is being used?
If for example your VM say it's a VM_PROXY_MINOR of 7, then it won't work with a plugin compiled with VM_PROXY_MINOR = 8. It silently fails... Well actually it gives the primLoadLibrary failure, but good luck in guessing why...
Well, I just overwrote sqVirtualMachine.c/h in Cross/vm, shouldn't have I? It wouldn't compile if I didn't and its sets it to 8. Is it defined in any other place?
Other question, can classic FFI and Alien live nicely together (I mean have x plugin use classic FFI while y uses Alien)?
yes.
One more: should I use IA32ABIPlugin or IA32ABIPluginAttic? You can't have both in, right?
One is a subclass of the other. I use the IA32ABIPluginAttic one.
Ok, what is the difference between them? a performance issue? everything works the same I choose one or the other?
Syntax highligthing is broken for Alien primitive methods like these:
<primitive: 'primUnsignedShortAtPut' error: errorCode module: 'IA32ABI'>
Lastly, as I said when I loaded Alien Core the first time, I got this error while loading it:
Alien class>>#ensureInSpecialObjectsArray: "Index probably wrong".
What should I do about that? ignore it?
Well it seems to be related to
((Smalltalk includesKey: #ObjectMemory) and: [((Smalltalk at: #ObjectMemory) classPool at: #ClassAlien ifAbsent: []) ~~ (index - 1)]) ifTrue: [self error: 'index probably wrong'].
Usually people don't have ObjectMemory loaded in their image, and I"m not sure what it is check for. Why don't you try it in a regular Pharo image versus your VMMaker image.
In Pharo 1.0RC2 without any change, ObjectMemory doesn't exist. When I load VMMaker it's downloaded from monticello, but obviously, #ClassAlien isn't defined inside
(Smalltalk at: #ObjectMemory) classPool
Interestingly, after load alien, #ClassAlien gets added as a key, but all values of (Smalltalk at: #ObjectMemory) classPool are set to nil, like this:
(Smalltalk at: #ObjectMemory) classPool inspect:
- size : 119 [#AllButHashBits] : nil [#AllButMarkBit] : nil [#AllButMarkBitAndTypeMask] : nil [#AllButRootBit] : nil [#AllButTypeMask] : nil ... [#ClassAlien] : nil ...
Any ideas?
-- Javier Pimás Ciudad de Buenos Aires
-- Javier Pimás Ciudad de Buenos Aires
--
=========================================================================== John M. McIntosh <johnmci@smalltalkconsulting.com> Twitter: squeaker68882 Corporate Smalltalk Consulting Ltd. http://www.smalltalkconsulting.com
===========================================================================
-- Javier Pimás Ciudad de Buenos Aires
_______________________________________________ Pharo-project mailing list Pharo-project@lists.gforge.inria.fr http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
_______________________________________________ Pharo-project mailing list Pharo-project@lists.gforge.inria.fr http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project
-- Javier Pimás Ciudad de Buenos Aires
-- Javier Pimás Ciudad de Buenos Aires
-- Javier Pimás Ciudad de Buenos Aires