masked_avx_example.jl

# A small example of how to write Julia code that results in masked AVX instructions

"""
    unsafe_max4(C, A, B)

Compute the element-wise maximum of the 4 first elements of vectors `A` and `B`,
and store the result C.

If any of the vectors is too small, the program might crash, or worse.
"""
function unsafe_max4!(C, A, B)
    @inbounds (C[1], C[2], C[3], C[4]) = max.((A[1], A[2], A[3], A[4]), (B[1], B[2], B[3], B[4]))
end

A = rand(Int, 4)
B = rand(Int, 4)
C = zeros(Int, 4)

@code_native unsafe_max4!(C, A, B)

# Results in the following on my machine:

	# .section	__TEXT,__text,regular,pure_instructions
	# decl	%eax
	# movl	(%edx), %eax
	# vmovdqu	(%eax), %ymm0
	# decl	%eax
	# movl	(%ecx), %eax
	# vmovdqu	(%eax), %ymm1
	# vpcmpgtq	%ymm1, %ymm0, %ymm2
	# vblendvpd	%ymm2, %ymm0, %ymm1, %ymm0
	# decl	%eax
	# movl	(%esi), %eax
	# vmovupd	%ymm0, (%eax)
	# vmovupd	%ymm0, (%edi)
	# decl	%eax
	# movl	%edi, %eax
	# vzeroupper
	# retl
	# nopl	(%eax,%eax)

# The vblendvpd instruction in the middle is the masked AVX instruction
# which picks out the element-wise maximum.