Bug 2750 – Optimize slice copy with size known at compile time

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<------------------------------

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

Interesting, since I had already introduced almost the identical patch 3 days ago! Anyhow, this will obviously go out in the next update.

Fixed DMD 1.042 and 2.027