While working on a framework, I've found that Multiple Return Values (MRV) are clearly an optimization. I'de like to write a small D blog post about this but I don't know If it's clever enough or if it's a well know fact.
My base D material is this:
---
#!runnable-flags: -O -boundscheck=off -release
module runnable;
struct Get
{
static auto all()
{
import std.typecons;
return tuple(0.1f,0.2f,0.3f,0.4f);
}
static float a(){return 0.1f;}
static float b(){return 0.2f;}
static float c(){return 0.3f;}
static float d(){return 0.4f;}
}
void call(float a, float b, float c, float d){}
void tupYes()
{
call(Get.all[0..$]);
}
void tupNo()
{
call(Get.a, Get.b, Get.c, Get.d);
}
void main(string[] args)
{
import disassembler;
import std.stdio;
symbolTable.addModule!runnable;
writeln(prettyDisasm(&tupYes));
writeln;
writeln(prettyDisasm(&tupNo));
}
---
with my d beaengine bindings I get this (bin comes from DMD backend) :
;------- SUB 000000000044C918h -------
; NAMED: tupYes
000000000044C918h push rbp
000000000044C919h mov rbp, rsp
000000000044C91Ch sub rsp, 20h
000000000044C920h call 0044C8A0h
000000000044C925h movsd qword ptr [rbp-20h], xmm0
000000000044C92Ah fld qword ptr [rbp-20h]
000000000044C92Dh movsd qword ptr [rbp-20h], xmm1
000000000044C932h fld qword ptr [rbp-20h]
000000000044C935h fstp qword ptr [rbp-08h]
000000000044C938h fstp qword ptr [rbp-10h]
000000000044C93Bh movss xmm3, dword ptr [rbp-10h]
000000000044C940h movss xmm2, dword ptr [rbp-0Ch]
000000000044C945h movss xmm1, dword ptr [rbp-08h]
000000000044C94Ah movss xmm0, dword ptr [rbp-04h]
000000000044C94Fh call 0044C910h ; (call)
000000000044C954h mov rsp, rbp
000000000044C957h pop rbp
000000000044C958h ret
;-------------------------------------
;------- SUB 000000000044C960h -------
; NAMED: tupNo
000000000044C960h sub rsp, 78h
000000000044C964h call 0044C8D0h
000000000044C969h movss dword ptr [rsp], xmm0
000000000044C96Eh movss xmm3, dword ptr [rsp]
000000000044C973h movapd dqword ptr [rsp+10h], xmm3
000000000044C979h call 0044C8E0h
000000000044C97Eh movss dword ptr [rsp], xmm0
000000000044C983h movss xmm2, dword ptr [rsp]
000000000044C988h movapd xmm3, dqword ptr [rsp+10h]
000000000044C98Eh movapd dqword ptr [rsp+20h], xmm2
000000000044C994h movapd dqword ptr [rsp+30h], xmm3
000000000044C99Ah call 0044C8F0h
000000000044C99Fh movss dword ptr [rsp], xmm0
000000000044C9A4h movss xmm1, dword ptr [rsp]
000000000044C9A9h movapd xmm2, dqword ptr [rsp+20h]
000000000044C9AFh movapd xmm3, dqword ptr [rsp+30h]
000000000044C9B5h movapd dqword ptr [rsp+40h], xmm1
000000000044C9BBh movapd dqword ptr [rsp+50h], xmm2
000000000044C9C1h movapd dqword ptr [rsp+60h], xmm3
000000000044C9C7h call 0044C900h
000000000044C9CCh movapd xmm1, dqword ptr [rsp+40h]
000000000044C9D2h movapd xmm2, dqword ptr [rsp+50h]
000000000044C9D8h movapd xmm3, dqword ptr [rsp+60h]
000000000044C9DEh call 0044C910h ; (call)
000000000044C9E3h add rsp, 78h
000000000044C9E7h ret
;-------------------------------------
Which clearly shows that using the MRV version is faster than grabing each property, since in the second, version there's a call for each parameter.
When I google "MRV optimization tuple", there's nothing, maybe some garbages from the early 2000's...nothing else. I'd like your mind before writing something possibly ridiculous.
| 
Permalink
Reply