www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - StructType.init and StructType() cause a allocation

reply Benjamin Thaut <code benjamin-thaut.de> writes:
If T is a struct both of the following statements will lead to a call of 
the _d_arrayliteral function which will allocate a array of the size of 
the struct just to initialize it!!

void T[] data = ...;

data[0] = T.init;
data[0] = T();

This has proven to be the biggest performance bottleneck in my current 
codebase, replacing this statement with a unneccessary workaround speeds 
my code up by factor 100. This is the cause because all of my container 
classes tend to reinitialize the contents of the data arrays they are 
holding. Also this bug makes it very hard to write non leaking code when 
not using a GC. The replacement code could be

T temp;
data[0] = temp;

Which works correctly as long as T is publicly accessible, if it is a 
private type this will not compile. Then you have to do some more fancy 
stuff like:

void[T.sizeof] temp;
void[] initMem = typeid(T).init();
if(initMem.ptr is null)
   memset(temp.ptr, 0, temp.length);
else
   temp[] = initMem[];
data[0] = *cast(T*)temp.ptr;

How hard would it be to fix this?

It is really a shame that reinitalizing a struct
a) is a major performance bottleneck
b) leaks memory

I created a bug ticket for this 6 months ago, but it has been ignored so 
far: http://d.puremagic.com/issues/show_bug.cgi?id=7271

Kind Regards
Benjamin Thaut
Jun 30 2012
parent reply "bearophile" <bearophileHUGS lycos.com> writes:
Benjamin Thaut:


 I created a bug ticket for this 6 months ago, but it has been 
 ignored so far: 
 http://d.puremagic.com/issues/show_bug.cgi?id=7271


A similar bug was fixed very recently. D used to allocate a 
dynamic array of length 1 when you write "new T" where T is a 
struct. I don't know if that bug fix (in dmd 2.060alpha) has 
fixed your problem too. Such kind of bugs do get fixed, 
eventually :-)

Bye,
bearophile
Jun 30 2012
parent reply "bearophile" <bearophileHUGS lycos.com> writes:
 I don't know if that bug fix (in dmd 2.060alpha) has fixed your 
 problem too.


I have compiled this code from your issue 7271 with the latest 
beta:


struct MemoryBlockInfo {
     size_t size;
     long[10] backtrace;
     int backtraceSize;

     this(size_t size) {
         this.size = size;
     }
}
void main() {
     MemoryBlockInfo info;
     info = MemoryBlockInfo.init; //array allocation here
}


Compiling it with no optimization it gives:


_D4temp15MemoryBlockInfo6__ctorMFkZS4temp15MemoryBlockInfo
         enter   4,0
         mov ECX,8[EBP]
         mov [EAX],ECX
         leave
         ret 4

__Dmain comdat
         enter   060h,0
         push    ESI
         push    EDI
         mov ECX,018h
         xor EAX,EAX
         lea EDI,-060h[EBP]
         rep
         stosd
         mov ESI,offset FLAT:_D4temp15MemoryBlockInfo6__initZ
         lea EDI,-060h[EBP]
         mov CL,018h
         rep
         movsd
         pop EDI
         pop ESI
         leave
         ret


With -O -release -inline it becomes:


_D4temp15MemoryBlockInfo6__ctorMFkZS4temp15MemoryBlockInfo  comdat
         push    EAX
         mov ECX,8[ESP]
         mov [EAX],ECX
         pop ECX
         ret 4

__Dmain comdat
         sub ESP,064h
         mov ECX,018h
         xor EAX,EAX
         push    EDI
         lea EDI,4[ESP]
         rep
         stosd
         pop EDI
         add ESP,064h
         ret


In both cases I don't see a memory allocation.

Bye,
bearophile
Jun 30 2012
parent Benjamin Thaut <code benjamin-thaut.de> writes:
Am 30.06.2012 13:52, schrieb bearophile:

 I don't know if that bug fix (in dmd 2.060alpha) has fixed your
 problem too.


 I have compiled this code from your issue 7271 with the latest beta:


 struct MemoryBlockInfo {
      size_t size;
      long[10] backtrace;
      int backtraceSize;

      this(size_t size) {
          this.size = size;
      }
 }
 void main() {
      MemoryBlockInfo info;
      info = MemoryBlockInfo.init; //array allocation here
 }


 Compiling it with no optimization it gives:


 _D4temp15MemoryBlockInfo6__ctorMFkZS4temp15MemoryBlockInfo
          enter   4,0
          mov ECX,8[EBP]
          mov [EAX],ECX
          leave
          ret 4

 __Dmain comdat
          enter   060h,0
          push    ESI
          push    EDI
          mov ECX,018h
          xor EAX,EAX
          lea EDI,-060h[EBP]
          rep
          stosd
          mov ESI,offset FLAT:_D4temp15MemoryBlockInfo6__initZ
          lea EDI,-060h[EBP]
          mov CL,018h
          rep
          movsd
          pop EDI
          pop ESI
          leave
          ret


 With -O -release -inline it becomes:


 _D4temp15MemoryBlockInfo6__ctorMFkZS4temp15MemoryBlockInfo  comdat
          push    EAX
          mov ECX,8[ESP]
          mov [EAX],ECX
          pop ECX
          ret 4

 __Dmain comdat
          sub ESP,064h
          mov ECX,018h
          xor EAX,EAX
          push    EDI
          lea EDI,4[ESP]
          rep
          stosd
          pop EDI
          add ESP,064h
          ret


 In both cases I don't see a memory allocation.

 Bye,
 bearophile


Thank you, Its very nice to know that this got fixed. Someone might want 
to close the ticket as duplicate then.

Kind Regards
Benjamin Thaut
Jun 30 2012