[Issue 2750] New: Optimize slice copy with size known at compile time

Mar 19, 2009

d-bugmail

Mar 19, 2009

d-bugmail

Mar 20, 2009

d-bugmail

Apr 01, 2009

d-bugmail

March 19, 2009

[Issue 2750] New: Optimize slice copy with size known at compile time

Posted by d-bugmail

Permalink

d-bugmail

Permalink

http://d.puremagic.com/issues/show_bug.cgi?id=2750

           Summary: Optimize slice copy with size known at compile time
           Product: D
           Version: 1.041
          Platform: PC
        OS/Version: Windows
            Status: NEW
          Keywords: patch, wrong-code
          Severity: normal
          Priority: P2
         Component: DMD
        AssignedTo: bugzilla@digitalmars.com
        ReportedBy: snake.scaly@gmail.com


It was discussed recently that the compiler intrinsic for slice copying was
slower than CRT memcpy():

http://tinyurl.com/cfxmva

In that particular case it was generating rep movsb despite the fact that the slice size was known at compile time.

I'm proposing a patch which fixes this problem.  Here is an example.  This code:

void main() {
    auto h = "hello\n";
    char buf[16];
    buf[0 .. h.length] = h;
}

compiled with -O -release -inline by the current 1.041:

__Dmain comdat
        assume  CS:__Dmain
                sub     ESP,020h
                mov     EDX,FLAT:_DATA[0Ch]
                mov     EAX,FLAT:_DATA[08h]
                push    EBX
                push    ESI
                mov     ESI,EDX
                push    EDI
                lea     EDI,0Ch[ESP]
                movsd
                movsb
                movsb
                lea     ECX,01Ch[ESP]
                mov     EBX,0FFFFFFFFh
                mov     [ECX],EBX
                mov     EAX,6
                lea     ESI,0Ch[ESP]
                mov     4[ECX],EBX
                lea     EDI,01Ch[ESP]
                mov     8[ECX],EBX
                mov     0Ch[ECX],EBX
                mov     ECX,EAX
                rep
                movsb
                xor     EAX,EAX
                pop     EDI
                pop     ESI
                pop     EBX
                add     ESP,020h
                ret
__Dmain ends

and by a patched compiler:

__Dmain comdat
        assume  CS:__Dmain
                sub     ESP,020h
                mov     EDX,FLAT:_DATA[0Ch]
                mov     EAX,FLAT:_DATA[08h]
                push    EBX
                push    ESI
                mov     ESI,EDX
                push    EDI
                lea     EDI,0Ch[ESP]
                movsd
                movsb
                movsb
                lea     ECX,01Ch[ESP]
                mov     EBX,0FFFFFFFFh
                mov     [ECX],EBX
                xor     EAX,EAX
                mov     4[ECX],EBX
                mov     8[ECX],EBX
                mov     0Ch[ECX],EBX
                pop     EDI
                pop     ESI
                pop     EBX
                add     ESP,020h
                ret
__Dmain ends

Here is the patch:

-------8<------------------------------
diff --git a/dmd/backend/cgelem.c b/dmd/backend/cgelem.c index a2a4a1f..a80eefb 100644
--- a/dmd/backend/cgelem.c
+++ b/dmd/backend/cgelem.c
@@ -3773,6 +3773,16 @@ STATIC elem * el64_32(elem *e)
            e->E1 = el_selecte1(e->E1);
        }
        break;
+
+    case OPpair:
+       e = el_selecte1(el_selecte1(e));
+       goto L1;
+    case OPrpair:
+       e = el_selecte2(el_selecte1(e));
+       goto L1;
+    L1:
+       e->Ety = ty;
+       break;
   }
   return e;
 }
-------8<------------------------------


--

http://d.puremagic.com/issues/show_bug.cgi?id=2750 ------- Comment #1 from snake.scaly@gmail.com 2009-03-19 15:40 ------- Sorry, bad example. The patched compiler simply optimized the copy away, which wasn't bad in itself, but neither was what I wanted to demonstrate. Here's a better example: void main() { auto h = "hello\n"; auto buf = new char[16]; buf[0 .. h.length] = h; } Original 1.041: __Dmain comdat assume CS:__Dmain L0: sub ESP,018h mov EDX,FLAT:_DATA[0Ch] mov EAX,FLAT:_DATA[08h] push EBX push ESI mov ESI,EDX push EDI lea EDI,0Ch[ESP] movsd movsb movsb lea ESI,0Ch[ESP] mov ECX,offset FLAT:_D11TypeInfo_Aa6__initZ push 010h push ECX call near ptr __d_newarrayiT mov EBX,6 mov ECX,EBX mov EDI,EDX rep movsb add ESP,8 xor EAX,EAX pop EDI pop ESI pop EBX add ESP,018h ret __Dmain ends Assembly produced by a patched compiler: __Dmain comdat assume CS:__Dmain L0: sub ESP,01Ch mov EDX,FLAT:_DATA[0Ch] mov EAX,FLAT:_DATA[08h] push ESI mov ESI,EDX push EDI lea EDI,0Ch[ESP] movsd movsb movsb mov ECX,offset FLAT:_D11TypeInfo_Aa6__initZ push 010h push ECX call near ptr __d_newarrayiT lea ESI,014h[ESP] mov EDI,EDX movsd movsb movsb add ESP,8 xor EAX,EAX pop EDI pop ESI add ESP,01Ch ret __Dmain ends --

http://d.puremagic.com/issues/show_bug.cgi?id=2750 ------- Comment #2 from bugzilla@digitalmars.com 2009-03-20 00:12 ------- Interesting, since I had already introduced almost the identical patch 3 days ago! Anyhow, this will obviously go out in the next update. --

http://d.puremagic.com/issues/show_bug.cgi?id=2750 bugzilla@digitalmars.com changed: What |Removed |Added ---------------------------------------------------------------------------- Status|NEW |RESOLVED Resolution| |FIXED ------- Comment #3 from bugzilla@digitalmars.com 2009-04-01 13:49 ------- Fixed DMD 1.042 and 2.027 --

Forums