Assembly – vmovdqu what are you doing here?

2020-08-21 • Java

I have a Java loop as follows: @ h_ 419_ 7@

public void testMethod() {
    int[] nums = new int[10];
    for (int i = 0; i < nums.length; i++) {
        nums[i] = 0x42;
    }
}

0x00000001296ac845: cmp    %r10d,%ebp
0x00000001296ac848: jae    0x00000001296ac8b4
0x00000001296ac84a: movl   $0x42,0x10(%rbx,%rbp,4)
0x00000001296ac852: inc    %ebp               
0x00000001296ac854: cmp    %r11d,%ebp
0x00000001296ac857: jl     0x00000001296ac845  

0x00000001296ac859: mov    %r10d,%r8d
0x00000001296ac85c: add    $0xfffffffd,%r8d
0x00000001296ac860: mov    $0x80000000,%r9d
0x00000001296ac866: cmp    %r8d,%r10d
0x00000001296ac869: cmovl  %r9d,%r8d
0x00000001296ac86d: cmp    %r8d,%ebp
0x00000001296ac870: jge    0x00000001296ac88e
0x00000001296ac872: vmovq  -0xda(%rip),%xmm0                                                    
0x00000001296ac87a: vpunpcklqdq %xmm0,%xmm0,%xmm0
0x00000001296ac87e: xchg   %ax,%ax

0x00000001296ac880: vmovdqu %xmm0,4)  
0x00000001296ac886: add    $0x4,%ebp          
0x00000001296ac889: cmp    %r8d,%ebp
0x00000001296ac88c: jl     0x00000001296ac880

Solution

The optimizer chooses to vectorize the loop, setting 4 values per iteration (vmovdqu the previous instruction is quite opaque, but it may map 0x42 to all channels of xmm0.) The "misaligned" variable is necessary because there is no guarantee that the array is SIMD aligned in memory (after all, it stores int32s, not int32x4s)

The content of this article comes from the network collection of netizens. It is used as a learning reference. The copyright belongs to the original author.

THE END

Java