RFC: Avoid unwanted signed/unsigned conversion check in pointer + and -

[toivoh]

After #8420, each use of +(::Ptr,::Int) and -(::Ptr,::Int) includes a (probably unintentional) check for illegal signed/unsigned conversion: (maybe also for other Integer subtypes)

julia> code_native(+, (Ptr{Int}, Int))
    .text
Filename: pointer.jl
Source line: 65
    push    RBP
    mov RBP, RSP
    test    RSI, RSI
Source line: 65
    js  8
    add RDI, RSI
    mov RAX, RDI
    pop RBP
    ret
    movabs  RAX, 140346252604120
    mov RDI, QWORD PTR [RAX]
    movabs  RAX, 140346237622048
    mov ESI, 65
    call    RAX

I can only guess that this was not intended. Pointer arithmetic is inherently unsafe and impossible to check automatically, but it can be quite useful in tight hand-optimized loops, where the above check adds overhead and precludes further optimizations.

This patch brings back the previous, simple behavior:

julia> code_native(+, (Ptr{Int}, Int))
    .text
Filename: pointer.jl
Source line: 65
    push    RBP
    mov RBP, RSP
Source line: 65
    lea RAX, QWORD PTR [RDI + RSI]
    pop RBP
    ret

I put RFC in the title to make sure that this is actually desirable to fix, and to get feedback on the way that I have fixed it:

• The new code assumes implicitly that a UInt has the same size as a Ptr. Afaik this is always the case in Julia, but should I rely on it?
• A new fix will be needed if we get checked arithmetic by default.

Btw, I noticed that -(::Ptr,::Ptr) gives an unsigned result. Might also be worth fixing?

[JeffBezanson]

Yes I agree -(::Ptr,::Ptr) should be Int.