[Vm-dev] 64bits Pharo VM for windows

Tue Mar 21 21:18:20 UTC 2017

2017-03-21 7:48 GMT+01:00 Nicolas Cellier <
nicolas.cellier.aka.nice at gmail.com>:

> Hi Eliot,
> I'll try the assert if I can find a time slot today, otherwise this
> evening.
> I could also have printed the jump_buf address in gdb, but it was too late
> yesterday ;)
>
>
> 2017-03-21 2:19 GMT+01:00 Eliot Miranda <eliot.miranda at gmail.com>:
>
>>
>>
>>
>> On Mon, Mar 20, 2017 at 6:01 PM, Eliot Miranda <eliot.miranda at gmail.com>
>> wrote:
>>
>>> Hi Nicolas,
>>>
>>> On Mon, Mar 20, 2017 at 5:56 PM, Eliot Miranda <eliot.miranda at gmail.com>
>>> wrote:
>>>
>>>> Hi Nicolas,
>>>>
>>>> On Mon, Mar 20, 2017 at 3:45 PM, Nicolas Cellier <
>>>> nicolas.cellier.aka.nice at gmail.com> wrote:
>>>>
>>>>>
>>>>> Thanks Eliot for pushing WIN64 ABI further!
>>>>>
>>>>> So the failure is this one:
>>>>>
>>>>> gdb: unknown target exception 0xc0000028 at 0x774c8078
>>>>>
>>>>> Program received signal ?, Unknown signal.
>>>>> 0x00000000774c8078 in ntdll!RtlRaiseStatus () from
>>>>> /cygdrive/c/Windows/SYSTEM32/ntdll.dll
>>>>> (gdb) where
>>>>> #0  0x00000000774c8078 in ntdll!RtlRaiseStatus () from
>>>>> /cygdrive/c/Windows/SYSTEM32/ntdll.dll
>>>>> #1  0x0000000077467eb6 in ntdll!TpAlpcRegisterCompletionList () from
>>>>> /cygdrive/c/Windows/SYSTEM32/ntdll.dll
>>>>> #2  0x000007fefe08e5a3 in msvcrt!longjmp () from
>>>>> /cygdrive/c/Windows/system32/msvcrt.dll
>>>>> #3  0x0000000000419502 in ceReturnToInterpreter (anOop=176164968) at
>>>>> ../../spur64src/vm/gcc3x-cointerp.c:16504
>>>>> #4  0x000000000a801086 in ?? ()
>>>>> Backtrace stopped: previous frame inner to this frame (corrupt stack?)
>>>>>
>>>>> I now suspect the jmp_buf alignment problem that I had to workaround
>>>>> in jpeg plugin:
>>>>> It must be aligned on 16 bytes boundary in Win64, but sometimes the
>>>>> compiler fails to honour this requirement
>>>>> See https://github.com/OpenSmalltalk/opensmalltalk-vm/pull/120
>>>>>
>>>>
>>>> Hmph.  So the stack /should/ be aligned on a 16-byte boundary and hence
>>>> the compiler /should/ be able to maintain the invariant (see
>>>> platforms/Cross/vm/sqCogStackAlignment.h; in fact on Mac OS X the
>>>> alignment is 32 bytes).
>>>>
>>>> Let me suggest that you add the following to the preamble:
>>>>
>>>> #if WIN64
>>>> # define sigsetjmp(jb,ssmf) (assert(((int)jb & 15) == 0, setjmp(jb))
>>>> # define siglongjmp(jb,v) (assert(((int)jb & 15) == 0, longjmp(jb,v))
>>>> #elif WIN32
>>>> ...
>>>>
>>>> and make sure there's a self assertCStackWellAligned send in
>>>> ceReturnToInterpreter.
>>>>
>>>
>>> Hmmm.  I expect we need code in the ceReturnToInterpreterTrampoline
>>> that establishes the stack alignment requirement.  ceReturnToInterpreter:
>>> would be called from machine code where there is only 8 byte alignment on
>>> x64 (& 4 byte alignment on 32-bit VMs).  If you like I can try and
>>> implement this tomorrow.
>>>
>>
>>  Looking at the code, that's not necessary.  The trampoline still
>> switches to the C stack, which should be correctly aligned.  So yes,
>> definitely add the "self assertCStackWellAligned" to both
>> ceReturnToInterpreter: and ceBaseFrameReturn:.  It looks like the issue is
>> whether the returnToInterpreter jmpbuf is correctly aligned.
>>
>>

The STACK_ALIGN_BYTES is 16 in good agreement with WIN64 ABI and
assertCStackWellAligned() do succeed in ceReturnToInterpreter...

So, I added the assertion on setjmp and longjmp, but they do not fail.
The jmp_buf reenterInterpreter is correctly aligned on 16 bytes boundary.
There is a declaration like this in setjmp.h
 typedef _CRT_ALIGN(16) struct _SETJMP_FLOAT128
So local and global jmp_buf variables are allways well aligned by the C
compiler(s)

The jpeg problem was caused by putting the jmp_buf into a structure.
Apparently, gcc and clang fail to correctly handle that case
(they should first align the whole struct on 16 bytes then align the offset
of the jmp_buf member).

There are other cases that will be problematic in Win64 exactly like the
jpeg case, for example I see:
siglongjmp((vmCallbackContext->trampoline)... in
returnAsThroughCallbackContext
longjmp(GIV(jmpBuf)[GIV(jmpDepth)]...in callbackLeave (only #if
SQ_USE_GLOBAL_STRUCT)

We should memcpy these to a local jmp_buf for WIN64 compatibility.

But the VM failed before ever reaching one of these callbacks (we should
have none in Squeak by default)

So, wrong guess from my side for the moment, seeing longjmp on the call
stack raised that false alarm.

>
>>> 2017-03-19 21:03 GMT+01:00 Nicolas Cellier <
>>>>> nicolas.cellier.aka.nice at gmail.com>:
>>>>>
>>>>>> And currently https://github.com/OpenSmallta
>>>>>> lk/opensmalltalk-vm/blob/Cog/spur64src/vm/cogitX64.c is generated
>>>>>> for SysV only.
>>>>>> It's necessary to hack the CogX64Compiler SysV class var
>>>>>> initialization and generate a win64 specific cogitX64.c.
>>>>>>
>>>>>>
>>>>>> 2017-03-19 20:34 GMT+01:00 Nicolas Cellier <
>>>>>> nicolas.cellier.aka.nice at gmail.com>:
>>>>>>
>>>>>>> Hi Clement,
>>>>>>> it's been a while since I last tested, but in a few words:
>>>>>>> - win64 use it's own ABI
>>>>>>> - we have to assign the registers differently than sysV
>>>>>>> - the experiments I did resulted in VM crashing early (before
>>>>>>> opening window)
>>>>>>>
>>>>>>> Nicolas
>>>>>>>
>>>>>>> 2017-03-19 20:29 GMT+01:00 Clément Bera <bera.clement at gmail.com>:
>>>>>>>
>>>>>>>>
>>>>>>>> Thank you very much for doing Nicolas. It is very important for
>>>>>>>> many Pharo users to use Pharo 64 bits on Windows.
>>>>>>>>
>>>>>>>> What are the problems you have when trying to build the VM with the
>>>>>>>> JIT that you don't have when building the stack VM ? Is it about API to
>>>>>>>> make the memory executable, is it about calling conventions ?
>>>>>>>>
>>>>>>>> On Sun, Mar 19, 2017 at 12:14 PM, Nicolas Cellier <
>>>>>>>> nicolas.cellier.aka.nice at gmail.com> wrote:
>>>>>>>>
>>>>>>>>>
>>>>>>>>> And the appveyor builds are green
>>>>>>>>> https://ci.appveyor.com/project/OpenSmalltalk/vm/build/1.0.579
>>>>>>>>>
>>>>>>>>> 2017-03-19 17:31 GMT+01:00 Nicolas Cellier <
>>>>>>>>> nicolas.cellier.aka.nice at gmail.com>:
>>>>>>>>>
>>>>>>>>>> Hi,
>>>>>>>>>> I've built a 64bits pharo.stack.spur VM for windows on my machine,
>>>>>>>>>> and I'm uploading the changes to opensmalltalk-vm in branch
>>>>>>>>>> build_pharo_win32_with_cygwin
>>>>>>>>>>
>>>>>>>>>> If the appveyor job correctly succeed, I will emit a pull request.
>>>>>>>>>>
>>>>>>>>>> The VM does not have the SqueakSSL plugin yet.
>>>>>>>>>>
>>>>>>>>>> The 64bits squeak/pharo.cog.spur JIT for windows is still to come,
>>>>>>>>>> but I did not work on it for a few months...
>>>>>>>>>> One thing at a time.
>>>>>>>>>>
>>>>>>>>>> Let's cross finger
>>>>>>>>>>
>>>>>>>>>> Nicolas
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>> --
>>>> _,,,^..^,,,_
>>>> best, Eliot
>>>>
>>>
>>>
>>>
>>> --
>>> _,,,^..^,,,_
>>> best, Eliot
>>>
>>
>>
>>
>> --
>> _,,,^..^,,,_
>> best, Eliot
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.squeakfoundation.org/pipermail/vm-dev/attachments/20170321/d225d551/attachment-0001.html>