Producing abstract types can be more expensive than leaftypes

[mauro3]

Making a abstract type seems more expensive than a concrete one (although I didn't benchmark it):

julia> abstract A{T}; a() = A{Int}
a (generic function with 1 method)

julia> immutable B{T} end; b() = B{Int}
b (generic function with 1 method)

julia> @code_native a()
        .text
Filename: none
Source line: 1
        push    RBP
        mov     RBP, RSP
        push    RBX
        sub     RSP, 40
        mov     QWORD PTR [RBP - 40], 4
Source line: 1
        movabs  RBX, 140026519774896
        mov     RAX, QWORD PTR [RBX]
        mov     QWORD PTR [RBP - 32], RAX
        lea     RAX, QWORD PTR [RBP - 40]
        mov     QWORD PTR [RBX], RAX
        vxorps  XMM0, XMM0, XMM0
        vmovups XMMWORD PTR [RBP - 24], XMM0
Source line: 1
        mov     RDI, QWORD PTR [65703328]
        mov     RAX, QWORD PTR [RDI + 8]
Source line: 1
        lea     RSI, QWORD PTR [RBP - 24]
Source line: 1
        mov     RCX, QWORD PTR [169843264]
        mov     QWORD PTR [RBP - 24], RCX
        mov     RCX, QWORD PTR [65704528]
        mov     QWORD PTR [RBP - 16], RCX
        mov     EDX, 2
        call    RAX
        mov     RCX, QWORD PTR [RBP - 32]
        mov     QWORD PTR [RBX], RCX
        add     RSP, 40
        pop     RBX
        pop     RBP
        ret

julia> @code_native b()
        .text
Filename: none
Source line: 1
        push    RBP
        mov     RBP, RSP
        mov     EAX, 173166928
Source line: 1
        pop     RBP
        ret

Note that @code_native f() = A is just as short as for b.

[carnaval]

I think it's just about weakening

julia/base/inference.jl

Line 2146 in 042c4d5

isleaftype(e.typ.parameters[1])

this should not need a leaftype, just something that is not a typevar should be ok

[JeffBezanson]

+1 to that

[mauro3]

Thanks for looking into this. I hope fixing this will make trait methods in Traits.jl just as efficient as normal methods. At the moment, they usually produce more native code.

[tkelman]

@JeffBezanson backport this or no?

[JeffBezanson]

Yes should be safe.

[tkelman]

thanks, though there are some nasty conflicts so I think I'll skip this for now? it might depend on a verdict on #13491 and #13553 edit: depends on #13434 so will leave off for now.

[JeffBezanson]

This is not crucial.