Hi,

I am using DMD 2.062 on Windows 7 64-bit.

I am writing performance critical functions that need switch statements to use an indirect jump table... current I'm analysing the assembly dump, and the code is compiled to nested ifs instead. This happens with switch and final switch. Is there any way to force the compiler to use a jump table?

Steve

On Thursday, April 04, 2013 03:06:44 Steve Kucera wrote:
> Hi,
> 
> I am using DMD 2.062 on Windows 7 64-bit.
> 
> I am writing performance critical functions that need switch statements to use an indirect jump table... current I'm analysing the assembly dump, and the code is compiled to nested ifs instead. This happens with switch and final switch. Is there any way to force the compiler to use a jump table?

If you have performance-critical code, then use gdc or ldc (probably ldc, since you're on Windows). dmd's optimizer is nowhere near as good as theirs are. dmd is very well optimized for compiling quickly, but it fails in optimizing the code that it generates. It's okay at it, but it really doesn't compare to gdc or ldc (which use gcc and LLVM as their backends respectively). And dmd doesn't have much in the way of switches which affect optimization, so if neither -O, -inline, or -release gets you jump tables, then I expect that there's no way that dmd is going to give you jump tables.

- Jonathan M Davis

Steve Kucera:

> I am using DMD 2.062 on Windows 7 64-bit.
>
> I am writing performance critical functions that need switch statements to use an indirect jump table... current I'm analysing the assembly dump, and the code is compiled to nested ifs instead. This happens with switch and final switch. Is there any way to force the compiler to use a jump table?

What kind of switch do you have? Are you switching on strings?

Here I am not seeing nested ifs, this is efficient enough:


import core.stdc.stdio: puts;
void main(string[] args) {
    switch (args.length) {
        case 0: puts("0"); break;
        case 1: puts("1"); break;
        case 2: puts("2"); break;
        case 3: puts("3"); break;
        default: puts("default"); break;
    }
}




main:
L0:     push    EAX
        push    EBX
        mov EBX,0Ch[ESP]
        cmp EBX,3
        push    ESI
        ja  L53
        jmp dword ptr FLAT:_DATA[018h][EBX*4]

        mov ESI,offset FLAT:_DATA
        push    ESI
        call    near ptr _puts
        add ESP,4
        jmp short   L61

        mov EDX,offset FLAT:_DATA[4]
        push    EDX
        call    near ptr _puts
        add ESP,4
        jmp short   L61

        mov ECX,offset FLAT:_DATA[8]
        push    ECX
        call    near ptr _puts
        add ESP,4
        jmp short   L61

        mov EAX,offset FLAT:_DATA[0Ch]
        push    EAX
        call    near ptr _puts
        add ESP,4
        jmp short   L61

L53:        mov EBX,offset FLAT:_DATA[010h]
        push    EBX
        call    near ptr _puts
        add ESP,4

L61:        pop ESI
        xor EAX,EAX
        pop EBX
        pop ECX
        ret


Bye,
bearophile

On Thursday, 4 April 2013 at 01:06:45 UTC, Steve Kucera wrote:
> Hi,
>
> I am using DMD 2.062 on Windows 7 64-bit.
>
> I am writing performance critical functions that need switch statements to use an indirect jump table... current I'm analysing the assembly dump, and the code is compiled to nested ifs instead. This happens with switch and final switch. Is there any way to force the compiler to use a jump table?
>
> Steve

Ldc or gdc may be able to do this for you if dmd cannot.

Other than that, the inline asm in dmd is easy enough to use.

On Thursday, 4 April 2013 at 09:51:15 UTC, John Colvin wrote:
> On Thursday, 4 April 2013 at 01:06:45 UTC, Steve Kucera wrote:
>> Hi,
>>
>> I am using DMD 2.062 on Windows 7 64-bit.
>>
>> I am writing performance critical functions that need switch statements to use an indirect jump table... current I'm analysing the assembly dump, and the code is compiled to nested ifs instead. This happens with switch and final switch. Is there any way to force the compiler to use a jump table?
>>
>> Steve
>
> Ldc or gdc may be able to do this for you if dmd cannot.
>
> Other than that, the inline asm in dmd is easy enough to use.

Sorry, inline asm in D, not just dmd.

Hi,

OK thanks guys. LDC compiled a jump table. The code runs ~15% faster than dmd2, even though checking the ASM it did not inline functions it could have, and core.bitop.bsf compiled to a function call instead of an ASM instruction.

Can you force a function to be inline, and/or make core.bitop.bsf an instruction with ldc? (in the latter case, maybe the last resort is to dip into assembly?)

Steve

On Thursday, 4 April 2013 at 09:51:15 UTC, John Colvin wrote:
> On Thursday, 4 April 2013 at 01:06:45 UTC, Steve Kucera wrote:
>> Hi,
>>
>> I am using DMD 2.062 on Windows 7 64-bit.
>>
>> I am writing performance critical functions that need switch statements to use an indirect jump table... current I'm analysing the assembly dump, and the code is compiled to nested ifs instead. This happens with switch and final switch. Is there any way to force the compiler to use a jump table?
>>
>> Steve
>
> Ldc or gdc may be able to do this for you if dmd cannot.
>
> Other than that, the inline asm in dmd is easy enough to use.

Steven Kucera:

> and core.bitop.bsf compiled to a function call instead of an ASM instruction.

It used to be an intrinsics. I don't know if in the meantime
things have changed for LDC2.


> Can you force a function to be inline, and/or make core.bitop.bsf an instruction with ldc?

In LDC there are two different ways to inline asm, see the LDC
docs.

Bye,
bearophile