Disable szclass 12 for i686 systems.

[vchuravy]

The interplay between platform ABI and alignment for atomic ops is
tricky. This disables szclass 12 on i686 until we can use libatomic on all 32bit platforms (Windows support is currently missing).

[vchuravy]

I tested this on a cross-compile with wine and LLVM 3.9

using Base.Threads

function f(a)
    a[] = 128
end

for T in (UInt32, UInt64)
  @show T
  open("atomic_$(T)_$(Sys.ARCH).ll", "w") do io
    code_llvm(io, f, (Atomic{T},))
  end
  open("atomic_$(T)_$(Sys.ARCH).S", "w") do io
    code_native(io, f, (Atomic{T},))
  end
end

Prior:

; Function Attrs: alignstack(16)
define i32 @julia_f_60451(%jl_value_t*) #0 !dbg !5 {
top:
  %1 = bitcast %jl_value_t* %0 to i64*
  store atomic i64 128, i64* %1 release, align 4
  ret i32 128
}

After:

; Function Attrs: alignstack(16)
define i32 @julia_f_60448(%jl_value_t*) #0 !dbg !5 {
top:
  %1 = bitcast %jl_value_t* %0 to i64*
  store atomic i64 128, i64* %1 release, align 16
  ret i32 128
}

code_native only works after the change:

	.text
Filename: atomic.jl
	pushl	%ebp
	movl	%esp, %ebp
	pushl	%ebx
	pushl	%esi
	andl	$-16, %esp
	movl	8(%ebp), %esi
Source line: 4
	movl	(%esi), %eax
	movl	4(%esi), %edx
L16:
	xorl	%ecx, %ecx
	movl	$128, %ebx
	lock
	cmpxchg8b	(%esi)
	jne	L16
	movl	$128, %eax
	leal	-8(%ebp), %esp
	popl	%esi
	popl	%ebx
	popl	%ebp
	retl
	nopl	(%eax)

[tkelman]

The patch looks fine. I'll test locally - need to remember to rebase (when I test locally) to pick up the arm patches just added, just in case