[Vm-dev] Strange error while building stack vm

Igor Stasenko siguctua at gmail.com
Thu Dec 30 21:51:05 UTC 2010


On 30 December 2010 22:35, Eliot Miranda <eliot.miranda at gmail.com> wrote:
>
> Hi Igor,
>
> On Thu, Dec 30, 2010 at 1:37 AM, Igor Stasenko <siguctua at gmail.com> wrote:
>>
>> (warnings are supressed)
>>
>> In file included from
>> /Users/sig/projects/cog/sig-cog/build/../platforms/Mac
>> OS/vm/Developer/sqGnu.h:46,
>>                 from /Users/sig/projects/cog/sig-cog/src/vm/gcc3x-interp.c:20:
>> /Users/sig/projects/cog/sig-cog/src/vm/gcc3x-interp.c: At top level:
>> /Users/sig/projects/cog/sig-cog/src/vm/gcc3x-interp.c: At top level:
>> /Users/sig/projects/cog/sig-cog/src/vm/gcc3x-interp.c:15490: note:
>> ‘theIP’ was declared here
>> /Users/sig/projects/cog/sig-cog/src/vm/gcc3x-interp.c:34730: note:
>> ‘hdrTypeBits’ was declared here
>> /Users/sig/projects/cog/sig-cog/src/vm/gcc3x-interp.c:14690: note:
>> ‘hdrTypeBits’ was declared here
>> /Users/sig/projects/cog/sig-cog/src/vm/gcc3x-interp.c:35424: note:
>> ‘freeChunkSize’ was declared here
>> /Users/sig/projects/cog/sig-cog/src/vm/gcc3x-interp.c:35311: note:
>> ‘freeChunkSize’ was declared here
>> /var/folders/kE/kEXZB7BWHXWwxswd6aL4XU+++TM/-Tmp-//ccIxqx9C.s:143250:FATAL:Symbol
>> L0bereaveAllMarriedContexts already defined.
>
> So two things here.  One is what these labels are.  The labels that are added by the code generator simply for performance measurement.  See the define of VM_LABEL in the various sqPlatformSpecific.h files.  So this is not necessary but essential if using the VMProfiler to profile the stack interpreter.  If a method is marked with an <asmLabel: true> pragma the code generator outputs a label for it and one can ten identify the particular code sequence (e.g. a bytecode) in the VMProfiler.
> The other is what causes the error.  The C compiler can choose to inline code, but both gcc and icc appear not to examine the code they're inlining for inserted labels and end up duplicating the labels inserted by VM_LABEL.  Since this duplication depends on the exact source and compiler optimization level in force there's no easy way to predict when a label will be duplicated.  The fix is to add <asmLabel: false> to the relevant method.  A better fix might be to add a <C: #noinline> pragma that would result in turning off inlining for the relevant function (for example I don't think it appropriate at all that the C compiler is inlining bereaveAllMarriedContexts; it doesn't happen all that often - basically at snapshot or voiding machien code - and inlining it is a silly decision on the C compiler's part).  A hack fix is of course to simply define VM_LABEL as 0, but that criples the VMProfiler.

Thanks for explanation and a hint. But as you can see this is a code
generator, which inlining this function,
and then for duplicating label, C compiler inlining
voidVMStateForSnapshot() at least twice.
Yes, since this code is not critical for VM performance, we could
either disable inlining or remove label.

Because my first reaction was disable labels ...

I guess this happens because it compiles it with whole module
optimization flag: -O3 ,
which i took from xcode project 'Release' config.


> HTH
> Eliot
>>
>>
>> i found the only use of it in generated code:
>>
>> voidVMStateForSnapshot(void) {
>> DECL_MAYBE_SQ_GLOBAL_STRUCT
>>    sqInt activeContext;
>>    sqInt header;
>>    sqInt oop;
>>    sqInt sz;
>>
>>        activeContext = divorceAllFrames();
>>        /* begin bereaveAllMarriedContexts */
>>        VM_LABEL(0bereaveAllMarriedContexts);
>>
>>
>> but gcc insists that there are another one???
>>
>> --
>> Best regards,
>> Igor Stasenko AKA sig.
>
>
>



-- 
Best regards,
Igor Stasenko AKA sig.


More information about the Vm-dev mailing list