split OPpair into two separate loads

[WalterBright]

Making it two separate operations instead of one works better with the code gen designed for register operands.

[WalterBright]

For an idea of the effect of this, consider the program:

void foo() {
    __gshared uint[2]  a1 = [0x1001_1100, 0x0220_0012];
    uint[] a = a1;
    foreach (e; a)
        bar(e);
}
void bar(uint);

Before:

_D3foo3fooFZv:
                push    EAX
                mov     ECX,offset FLAT:_D3foo3fooFZ2a1G2k
                push    EAX
                mov     EAX,2
                push    EBX
                xor     EBX,EBX
                mov     4[ESP],EAX
                mov     8[ESP],ECX
                cmp     4[ESP],EBX
                je      L30
L1D:            mov     EDX,8[ESP]
                mov     EAX,[EBX*4][EDX]
                call    near ptr _D3foo3barFkZv
                inc     EBX
                cmp     EBX,4[ESP]
                jb      L1D
L30:            pop     EBX
                add     ESP,8
                ret

After:

_D3foo3fooFZv:
                push    EBX
                xor     EBX,EBX
L3:             mov     EAX,_D3foo3fooFZ2a1G2k[EBX*4]
                call    near ptr _D3foo3barFkZv
                inc     EBX
                cmp     EBX,2
                jb      L3
                pop     EBX
                ret

[CyberShadow]

This pull request caused a ~30% increase in compilation real time for the "Vibe.d diet templates" test on http://digger.k3.1azy.net/trend/

[MartinNowak]

Seems like this introduced a bug.
Issue 15272 – [2.069-rc2,inline] nothing written to output when -inline is set

[MartinNowak]

This pull request caused a ~30% increase in compilation real time for the "Vibe.d diet templates" test on http://digger.k3.1azy.net/trend/

Have you filed that as regression? Is it resolved by now?
Github comments have the tendency to get ignored.

[CyberShadow]

Have you filed that as regression?

From http://forum.dlang.org/post/55F0868F.7070108@dawg.eu I understood that filing such regressions was not a desired course of action.

[WalterBright]

This pull request caused a ~30% increase in compilation real time for the "Vibe.d diet templates" test on http://digger.k3.1azy.net/trend/

1. do you mean dmd time or compiled application time?
2. the link doesn't work

[CyberShadow]

do you mean dmd time or compiled application time?

DMD time.

the link doesn't work

Try another web browser, some web browsers choke on big JSON blobs with nonsensical errors for some reason.

[MartinNowak]

From http://forum.dlang.org/post/55F0868F.7070108@dawg.eu I understood that filing such regressions was not a desired course of action.

Well 30% slowdown is a huge change, not just a small sideeffect.

[CyberShadow]

Well 30% slowdown is a huge change, not just a small sideeffect.

An even bigger one happened here:
#4924 (comment)