Array fixes for non-power-of-2 sized elements

[JeffBezanson]

Before:

julia> primitive type TypeWith24Bits 24 end

julia> x = Core.Intrinsics.trunc_int(TypeWith24Bits, 0x112233); a = [x,x];

julia> Core.arrayset(true, a, x, 2)
2-element Array{TypeWith24Bits,1}:
 TypeWith24Bits(0x112233)
 TypeWith24Bits(0x111122)

After: works.

Fixes JuliaData/JuliaDB.jl#287 (comment)

[vtjnash]

a[2] = x, or are you intentionally trying to avoid this being compiled (could use a comment)?

Can you add a test for a = [(x, x), (x, x)] also?

Might be good to use something other than x here also, so that the mutation here is observable.

[JeffBezanson]

OK. Yes, I was trying to avoid the compiled form since I believe it does the right thing.

[chethega]

Can we add the backport label?

Reason is that @code_native Base.elsize([]) is a real mess before (afaiu because Core.sizeof(Ptr) and Ptr is non-concrete, which messes up constant folding). Now we use Core.sizeof(Ptr{Cvoid}) which constant folds properly.

[JeffBezanson]

In julia 1.x:

julia> @code_native Base.elsize([])
	.text
; Function elsize {
; Location: abstractarray.jl:106
	movl	$8, %eax
	retq
	nopw	%cs:(%rax,%rax)
;}

How to reproduce what you're seeing?

[chethega]

julia> versioninfo()
Julia Version 1.2.0
Commit c6da87ff4b (2019-08-20 00:03 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i5-5###U CPU @ 2.00GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-6.0.1 (ORCJIT, broadwell)

julia> @code_typed Base.elsize([])
CodeInfo(
1 ─ %1 = Base.Ptr::Type{Ptr}
│   %2 = Core.sizeof::typeof(Core.sizeof)
│   %3 = (%2)(%1)::Int64
└──      return %3
) => Int64

julia> @code_native Base.elsize([])
	.text
; ┌ @ abstractarray.jl:153 within `elsize' @ array.jl:201
; │┌ @ abstractarray.jl:153 within `sizeof'
	pushq	%rax
	movabsq	$140487341557840, %rax  # imm = 0x7FC5C216C850
	movq	%rax, (%rsp)
	movabsq	$jl_f_sizeof, %rax
	movq	%rsp, %rsi
	xorl	%edi, %edi
	movl	$1, %edx
	callq	*%rax
; │└
; │ @ abstractarray.jl:153 within `elsize'
	movq	(%rax), %rax
	popq	%rcx
	retq
	nopw	(%rax,%rax)

That is fixed on master. I saw the spurious jl_f_sizeof call in some pointer(a, i) and tracked the fix back to this PR. I have only tested master and 1.2.0, but git blame suggests that it is not a recent regression.

Core.sizeof(Ptr) is not constant-folded on master either, but your PR changed to Core.sizeof(Ptr{Cvoid}) (and also fixed the original issue, and also added @pure for good measure). A minimal backport would just replace Core.sizeof(Ptr) with Core.sizeof(Ptr{Cvoid}).

[KristofferC]

julia> @code_native Base.elsize([])
        .text
; ┌ @ abstractarray.jl:153 within `elsize'
        movl    $8, %eax
        retq
        nopw    %cs:(%rax,%rax)
; └

on 1.3 at least.