www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - Array allocation on struct initialization

reply Benjamin Thaut <code benjamin-thaut.de> writes:
I'm currently trying to make a nongc safe version of the d-runtime and 
stumbeled on my favorite WTF moment so far:

The following programm:

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

   this(size_t size)
   {
     this.size = size;
   }
}

int main(string[] argv)
{
   MemoryBlockInfo info;
   info = MemoryBlockInfo.init;
}

This produces the following disassembly:
0040385A  mov         dword ptr [info],eax
0040385D  push        0Ah
0040385F  mov         ecx,offset TypeInfo_G10l __init (41E260h)
00403864  push        ecx
00403865  call        rt lifetime _d_arrayliteralTX (404664h)
0040386A  mov         dword ptr [_TMP0],eax
0040386D  mov         dword ptr [eax],0
00403873  mov         dword ptr [eax+4],0
0040387A  mov         dword ptr [eax+8],0
00403881  mov         dword ptr [eax+0Ch],0
00403888  mov         dword ptr [eax+10h],0
0040388F  mov         dword ptr [eax+14h],0
00403896  mov         dword ptr [eax+18h],0
0040389D  mov         dword ptr [eax+1Ch],0
004038A4  mov         dword ptr [eax+20h],0
004038AB  mov         dword ptr [eax+24h],0
004038B2  mov         dword ptr [eax+28h],0

Why in hell is there a call to rt.lieftime._d_arrayliteralTX ???
I always thought that the init data is some global immutable data block 
that never changes thorughout the program execution, so why does it need 
to be copied? With behaviour like this it is pretty much impossible to 
make a nongc safe version of D at all.

Is this a compiler bug or intended behaviour?
-- 
Kind Regards
Benjamin Thaut
Dec 29 2011
next sibling parent reply Benjamin Thaut <code benjamin-thaut.de> writes:
Anyone else caring about this?
Even with a GC it is a unneccessary allocation and performance impact.

Kind Regards
Benjamin Thaut
Dec 31 2011
parent Timon Gehr <timon.gehr gmx.ch> writes:
On 01/01/2012 04:30 AM, Benjamin Thaut wrote:

 Anyone else caring about this?
 Even with a GC it is a unneccessary allocation and performance impact.

 Kind Regards
 Benjamin Thaut


I think you should file a bug report.
Dec 31 2011
prev sibling parent Peter Alexander <peter.alexander.au gmail.com> writes:
On 29/12/11 5:13 PM, Benjamin Thaut wrote:

 I'm currently trying to make a nongc safe version of the d-runtime and
 stumbeled on my favorite WTF moment so far:

 The following programm:

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

 this(size_t size)
 {
 this.size = size;
 }
 }

 int main(string[] argv)
 {
 MemoryBlockInfo info;
 info = MemoryBlockInfo.init;
 }

 This produces the following disassembly:
 0040385A mov dword ptr [info],eax
 0040385D push 0Ah
 0040385F mov ecx,offset TypeInfo_G10l __init (41E260h)
 00403864 push ecx
 00403865 call rt lifetime _d_arrayliteralTX (404664h)
 0040386A mov dword ptr [_TMP0],eax
 0040386D mov dword ptr [eax],0
 00403873 mov dword ptr [eax+4],0
 0040387A mov dword ptr [eax+8],0
 00403881 mov dword ptr [eax+0Ch],0
 00403888 mov dword ptr [eax+10h],0
 0040388F mov dword ptr [eax+14h],0
 00403896 mov dword ptr [eax+18h],0
 0040389D mov dword ptr [eax+1Ch],0
 004038A4 mov dword ptr [eax+20h],0
 004038AB mov dword ptr [eax+24h],0
 004038B2 mov dword ptr [eax+28h],0

 Why in hell is there a call to rt.lieftime._d_arrayliteralTX ???
 I always thought that the init data is some global immutable data block
 that never changes thorughout the program execution, so why does it need
 to be copied? With behaviour like this it is pretty much impossible to
 make a nongc safe version of D at all.

 Is this a compiler bug or intended behaviour?


Seems like a bug (or at least a misfeature).

If you do this instead:

MemoryBlockInfo info;
MemoryBlockInfo infoInit;
info = infoInit;

Then you don't get the array literal memory allocation i.e. it only 
happens when assigning to T.init explicitly.
Jan 01 2012