Jump to content

  • Log In with Google      Sign In   
  • Create Account

VC2012: cost of recursive capturing lambda function


Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.

  • You cannot reply to this topic
No replies to this topic

#1 tanzanite7   Members   -  Reputation: 1291

Like
0Likes
Like

Posted 12 November 2012 - 10:18 AM

Not a burning problem - just curiosity:

Using "static std::function<bool(int64u)> fnAdd = [&amp;](int64u tag)->bool { ... fnAdd(...) ... }" produces the suspected/expected, but unnecessary, inline initialization / shutdown code + other pointless overhead:
0000000140021BA2  mov		 eax,dword ptr [$S2 (0141678770h)] // static initialization check overhead ...
0000000140021BA8  and		 eax,1 
0000000140021BAB  test	    eax,eax 
0000000140021BAD  jne		 ResGlShader::update+85h (0140021BF5h) // ... end of check.
0000000140021BAF  mov		 eax,dword ptr [$S2 (0141678770h)] // one time overhead ...
0000000140021BB5  or		  eax,1
0000000140021BB8  mov		 dword ptr [$S2 (0141678770h)],eax // this whole flag check/update looks completely bonkers - i suspect some crazy cache optimization (bit-packing multiple flags into one 32bit memory location?).
// capture code ... all i need is "this" and one local ("valid"), but it stores also what i am using via this - probably for performance reasons.
0000000140021BBE  lea		 r9,[valid] 
0000000140021BC3  mov		 r8,qword ptr [this] 
0000000140021BCB  lea		 rdx,[sources] 
0000000140021BD0  lea		 rcx,[rsp+60h] 
// the storing part is moved elsewhere (over a dozen loads and stores) - probably for instruction cache reasons (it should have moved most of the leading garbage there too though)
0000000140021BD5  call	    <lambda_ccd461a2e723c170260c9697d9a65f88>::<lambda_ccd461a2e723c170260c9697d9a65f88> (0140021EA0h) 
0000000140021BDA  mov		 rdx,rax 
0000000140021BDD  lea		 rcx,[fnAdd (0141678780h)]
// this is terrible ... introduces "malloc"/exceptions ... ffs. None of that is needed ... argh!
0000000140021BE4  call	    std::function<bool __cdecl(unsigned __int64)>::function<bool __cdecl(unsigned __int64)><<lambda_ccd461a2e723c170260c9697d9a65f88> > (014003A2E0h) 
// destruction at program exit ... there is nothing in lambda function to destroy - dealing with std::function ("free") is probably all it does.
0000000140021BE9  lea		 rcx,[`ResGlShader::update'::`2'::`dynamic atexit destructor for 'fnAdd'' (01400C6380h)] 
0000000140021BF0  call	    atexit (014003E7F0h)
0000000140021BF5 // ... end of skip

// The "fnAdd" function call has some extra std::function overhead too (argh*2):
// /.../ input: int64u param + "this". Stack pointer for local variable does not need passing. Clean and neat.
0000000140021C0C  call	    std::_Func_class<bool,unsigned __int64,std::_Nil,std::_Nil,std::_Nil,std::_Nil,std::_Nil,std::_Nil>::operator() (0140036E00h)
// ... call will land in (std::function check for valid contained function ... completely redundant):
0000000140036E00  mov		 qword ptr [rsp+10h],rdx 
0000000140036E05  sub		 rsp,28h 
0000000140036E09  mov		 rcx,qword ptr [rcx+18h] 
0000000140036E0D  test	    rcx,rcx 
0000000140036E10  je		  std::_Func_class<bool,unsigned __int64,std::_Nil,std::_Nil,std::_Nil,std::_Nil,std::_Nil,std::_Nil>::operator()+22h (0140036E22h) 
0000000140036E12  mov		 rax,qword ptr [rcx] 
0000000140036E15  lea		 rdx,[_Vx0] 
0000000140036E1A  call	    qword ptr [rax+10h] // call for contained function object ... used as variable and not a fixed constant which it actually is x_x. It will land ...
0000000140036E1D  add		 rsp,28h 
0000000140036E21  ret
// ... here:
000000014003D430  mov		 rdx,qword ptr [rdx] 
000000014003D433  add		 rcx,8 
000000014003D437  jmp		 <lambda_ccd461a2e723c170260c9697d9a65f88>::operator() (0140021EF0h) // finally ...

VC does the smart thing on certain conditions (stuff that does not have side effects mainly) and moves such static initializations out of the function body - but it can not do it here. Can i help it?

The "std::function" has many side effects and hence can not be optimized out, but i can not do recursion without it (or did i miss some other workaround that works with VC2012?) !

This is not performance related question (i do not see any situation where it would matter ... if it would matter then i would make a real member function that would have none of the considerable overhead) - it is just for my own educational purposes in regards of lambda functions. So far very pleased with them - just that doing recursion with them is a f* disaster (works fine, just looks like a disaster).

edit:
Well, one can not use "static" often with lambdas, so overhead there is mostly unavoidable - except getting rid of std::function. Giving up for now (aka. using a real function) - but leaving the question open in case a better solution shows up.

Edited by tanzanite7, 12 November 2012 - 10:47 AM.


Sponsor:



Old topic!
Guest, the last post of this topic is over 60 days old and at this point you may not reply in this topic. If you wish to continue this conversation start a new topic.



PARTNERS