Question

I am currently working on an implementation of Bitslice DES for x64, and I would like to know how I could avoid 3-byte VEX prefixes as much as possible with the following AVX instructions:

vpor
vpxor
vpand
vpandn

I was told I should use 2-byte VEX prefixes as much as possible instead of 3-byte ones to make the code smaller and faster, but I don't know exactly how to do so. I read somewhere that I should use xmm[0-7] instead of xmm[89] and xmm1[0-5], but I don't know which combinations of operands would give me 2-byte VEX prefixes as opposed to 3-byte ones. These instructions are used with 2 or 3 operands as shown in the following snippet:

%macro sbox1 4
vmovdqa xmm7,  xmm4
vpandn  xmm4,  xmm0
vpor    xmm10, xmm5,  xmm2
vpxor   xmm13, xmm2,  xmm0
vpxor   xmm11, xmm5,  xmm7
vpxor   xmm14, xmm4,  xmm3
vpandn  xmm12, xmm13, xmm11
vpand   xmm13, xmm10
vpxor   xmm15, xmm11, xmm2
vpxor   xmm8,  xmm13, xmm3
vpandn  xmm9,  xmm14, xmm8
vpor    xmm13, xmm5
vpor    xmm5,  xmm0
vpandn  xmm8,  xmm7
vpandn  xmm15, xmm14
vpxor   xmm13, xmm15
vpor    xmm6,  xmm5,  xmm13
vpandn  xmm5,  xmm3
vpandn  xmm15, xmm9,  xmm13
vmovdqa xmm3,  xmm8 
vpxor   xmm8,  xmm6
vpxor   xmm5,  xmm3
vpand   xmm13, xmm10
vpandn  xmm4,  xmm2
vpxor   xmm2,  xmm6,  xmm14
vpxor   xmm6,  xmm10
vpandn  xmm4,  xmm2
vpxor   xmm2,  xmm4,  pnot
vpxor   xmm4,  xmm11
vpxor   xmm13, xmm2
vpor    xmm4,  xmm3
vpandn  xmm2,  xmm1,  xmm8
vpor    xmm14, xmm7
vpxor   xmm4,  xmm10
vpor    xmm9,  xmm1
vpxor   xmm2,  xmm13
vpxor   xmm4,  xmm0
vpxor   xmm13, xmm4
vpxor   xmm9,  xmm13
vpor    xmm5,  xmm12
vpxor   xmm9,  %1

vmovdqa %1,    xmm9
vpor    xmm6,  xmm5
vpor    xmm13, xmm11
vpxor   xmm6,  xmm4
vpor    xmm0,  xmm1,  xmm15
vpxor   xmm13, xmm6
vpxor   xmm2,  %3
vmovdqa %3,    xmm2
vpxor   xmm13, %2
vpand   xmm4,  xmm15
vpandn  xmm6,  xmm14
vpxor   xmm13, xmm0
vmovdqa %2,    xmm13
vpxor   xmm4,  xmm6
vpor    xmm4,  xmm1
vpxor   xmm4,  xmm5
vpxor   xmm4,  %4
vmovdqa %4,    xmm4
%endmacro
Était-ce utile?

La solution 2

It turned out that I just needed to use xmm0...xmm7 as the last operand to have the instruction encoded with a 2-byte VEX prefix.

Autres conseils

You should stick to registers xmm0...xmm7. This also holds true for general purpose registers, the encoding of instructions involving eax,ebx...ebp is most of the time one byte shorter than instructions using r8d...r15d.

Licencié sous: CC-BY-SA avec attribution
Non affilié à StackOverflow
scroll top