About mantis 3929 (FPU register stack overflow)

 
http://bugs.squeak.org/view.php?id=3929 has been identified as a
problem of FPU register stack overflow on linux VM.

A single change of line 51 of x86-sysv-asm.S solved this problem:

    fstpl ffiFloatReturnValue

Though I did not write extensive tests, I used modified VM a while
without problems.

What do you think of inclusion in svn trunk ?

Nicolas

 
On Wed, Sep 09, 2009 at 10:50:07PM +0200, Nicolas Cellier wrote:
I recall some extensive discussion of this (some of which is documented
on the Mantis page), but no action since then. Barring any objections,
it sounds like this should be included in the svn trunk.

Ian, you may have missed the earlier discussions.

caveat: I have no expertise with FPUs and the subject is not the sort
of thing that would be improved by a democratic process, so I'll refrain
from casting my "+1" vote.

Dave

 

On 10.09.2009, at 02:32, David T. Lewis wrote:

>
> caveat: I have no expertise with FPUs and the subject is not the sort
> of thing that would be improved by a democratic process, so I'll  
> refrain
> from casting my "+1" vote.


+1 to that.

- Bert -

 
Bert Freudenberg wrote:
> On 10.09.2009, at 02:32, David T. Lewis wrote:
>> caveat: I have no expertise with FPUs and the subject is not the sort
>> of thing that would be improved by a democratic process, so I'll refrain
>> from casting my "+1" vote.
>
> +1 to that.

Ditto. The +1 is well-deserved. The problem is that the FPU stack isn't
popped when returning from a call that has a float return. It needs to
be popped; on x86 the FPU is organized as a stack. The difference
between "fstl" and "fstpl" is the ->p<- which stands for "pop" ;-)

Cheers,
   - Andreas

In looking at this I should point out I made a number of changes to my  
version of x86-sysv-asm.S
Oddly I already had the fstpl

But my other changes might be of interest and or for comment.

(a) usage of REG_TO_EXTERN, PICIFY, and  NON_LAZY_STUB to make it  
compile/link on some variation of gcc in 10.3.x at some point

(b) and     $0xFFFFFFF0,%esp /*align to 128-bits*/

        This was required to properly align the stack for vector logic and to  
make quicktime call FFI calls happy (otherwise at some point a  
quicktime FFI call trashes a value on the stack).

(c)  cld    "Clear Direction"
        I at some point added this based on:
        http://developer.apple.com/mac/library/documentation/DeveloperTools/Conceptual/LowLevelABI/Mac_OS_X_ABI_Function_Calls.pdf
        Contains system flags, such as the direction flag and the carry flag.  
The direction flag must be set to the “forward” direction (that is, 0)  
before entry to and upon exit from a routine.
        Other user flags have no specified role in the standard calling  
sequence and are not preserved.
       
On 2009-09-09, at 5:32 PM, David T. Lewis wrote:

--
=
=
=
========================================================================
John M. McIntosh <[hidden email]>   Twitter:  
squeaker68882
Corporate Smalltalk Consulting Ltd.  http://www.smalltalkconsulting.com
=
=
=
========================================================================



#import <architecture/i386/asm_help.h>
        .text
        .globl _ffiCallAddressOf

_ffiCallAddressOf: /* 8=addr, 12=stack, 16=stackSize */
        pushl %ebp
        movl %esp, %ebp
        movl 16(%ebp), %ecx
        testl %ecx, %ecx
        je 2f
        subl %ecx, %esp
        and     $0xFFFFFFF0,%esp /*align to 128-bits*/
        subl $4, %ecx
        movl 12(%ebp), %edx
1: movl (%edx,%ecx,1), %eax
        movl %eax, (%esp,%ecx,1)
        subl $4, %ecx
        jnc 1b
        jmp 3f
2: and     $0xFFFFFFF0,%esp /*align to 128-bits*/

3:
        cld
        call *8(%ebp)
        REG_TO_EXTERN (%eax, _intReturnValue)
        REG_TO_EXTERN (%edx, _intReturnValue2)
        PICIFY(_floatReturnValue)
        fstpl (%edx)
        NON_LAZY_STUB(_floatReturnValue)
        movl %ebp, %esp
        popl %ebp
        ret