.code
ALIGN 16
KeyExpansionStdcall proc
  movdqu xmm1, xmmword ptr [rcx + 0]
  movdqu xmmword ptr [rdx + 0], xmm1
  aeskeygenassist xmm2, xmm1, 1
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 16], xmm1
  aeskeygenassist xmm2, xmm1, 2
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 32], xmm1
  aeskeygenassist xmm2, xmm1, 4
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 48], xmm1
  aeskeygenassist xmm2, xmm1, 8
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 64], xmm1
  aeskeygenassist xmm2, xmm1, 16
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 80], xmm1
  aeskeygenassist xmm2, xmm1, 32
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 96], xmm1
  aeskeygenassist xmm2, xmm1, 64
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 112], xmm1
  aeskeygenassist xmm2, xmm1, 128
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 128], xmm1
  aeskeygenassist xmm2, xmm1, 27
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 144], xmm1
  aeskeygenassist xmm2, xmm1, 54
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 160], xmm1
  pxor xmm1, xmm1
  pxor xmm2, xmm2
  pxor xmm3, xmm3
  ret 
KeyExpansionStdcall endp
ALIGN 16
KeyExpansionAndInversionStdcall proc
  movdqu xmm1, xmmword ptr [rcx + 0]
  movdqu xmmword ptr [rdx + 0], xmm1
  aeskeygenassist xmm2, xmm1, 1
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 16], xmm1
  aeskeygenassist xmm2, xmm1, 2
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 32], xmm1
  aeskeygenassist xmm2, xmm1, 4
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 48], xmm1
  aeskeygenassist xmm2, xmm1, 8
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 64], xmm1
  aeskeygenassist xmm2, xmm1, 16
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 80], xmm1
  aeskeygenassist xmm2, xmm1, 32
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 96], xmm1
  aeskeygenassist xmm2, xmm1, 64
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 112], xmm1
  aeskeygenassist xmm2, xmm1, 128
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 128], xmm1
  aeskeygenassist xmm2, xmm1, 27
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 144], xmm1
  aeskeygenassist xmm2, xmm1, 54
  pshufd xmm2, xmm2, 255
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  vpslldq xmm3, xmm1, 4
  pxor xmm1, xmm3
  pxor xmm1, xmm2
  movdqu xmmword ptr [rdx + 160], xmm1
  movdqu xmm1, xmmword ptr [rdx + 16]
  aesimc xmm1, xmm1
  movdqu xmmword ptr [rdx + 16], xmm1
  movdqu xmm1, xmmword ptr [rdx + 32]
  aesimc xmm1, xmm1
  movdqu xmmword ptr [rdx + 32], xmm1
  movdqu xmm1, xmmword ptr [rdx + 48]
  aesimc xmm1, xmm1
  movdqu xmmword ptr [rdx + 48], xmm1
  movdqu xmm1, xmmword ptr [rdx + 64]
  aesimc xmm1, xmm1
  movdqu xmmword ptr [rdx + 64], xmm1
  movdqu xmm1, xmmword ptr [rdx + 80]
  aesimc xmm1, xmm1
  movdqu xmmword ptr [rdx + 80], xmm1
  movdqu xmm1, xmmword ptr [rdx + 96]
  aesimc xmm1, xmm1
  movdqu xmmword ptr [rdx + 96], xmm1
  movdqu xmm1, xmmword ptr [rdx + 112]
  aesimc xmm1, xmm1
  movdqu xmmword ptr [rdx + 112], xmm1
  movdqu xmm1, xmmword ptr [rdx + 128]
  aesimc xmm1, xmm1
  movdqu xmmword ptr [rdx + 128], xmm1
  movdqu xmm1, xmmword ptr [rdx + 144]
  aesimc xmm1, xmm1
  movdqu xmmword ptr [rdx + 144], xmm1
  pxor xmm1, xmm1
  pxor xmm2, xmm2
  pxor xmm3, xmm3
  ret 
KeyExpansionAndInversionStdcall endp
ALIGN 16
AES128EncryptOneBlockStdcall proc
  movdqu xmm0, xmmword ptr [rdx + 0]
  movdqu xmm2, xmmword ptr [r8 + 0]
  pxor xmm0, xmm2
  movdqu xmm2, xmmword ptr [r8 + 16]
  aesenc xmm0, xmm2
  movdqu xmm2, xmmword ptr [r8 + 32]
  aesenc xmm0, xmm2
  movdqu xmm2, xmmword ptr [r8 + 48]
  aesenc xmm0, xmm2
  movdqu xmm2, xmmword ptr [r8 + 64]
  aesenc xmm0, xmm2
  movdqu xmm2, xmmword ptr [r8 + 80]
  aesenc xmm0, xmm2
  movdqu xmm2, xmmword ptr [r8 + 96]
  aesenc xmm0, xmm2
  movdqu xmm2, xmmword ptr [r8 + 112]
  aesenc xmm0, xmm2
  movdqu xmm2, xmmword ptr [r8 + 128]
  aesenc xmm0, xmm2
  movdqu xmm2, xmmword ptr [r8 + 144]
  aesenc xmm0, xmm2
  movdqu xmm2, xmmword ptr [r8 + 160]
  aesenclast xmm0, xmm2
  pxor xmm2, xmm2
  movdqu xmmword ptr [rcx + 0], xmm0
  ret 
AES128EncryptOneBlockStdcall endp
end