// This file is generated from a similarly-named Perl script in the BoringSSL // source tree. Do not edit by hand. #include #if !defined(OPENSSL_NO_ASM) && defined(OPENSSL_X86_64) && defined(__APPLE__) .section __DATA,__const .p2align 4 L$bswap_mask: .quad 0x08090a0b0c0d0e0f, 0x0001020304050607 L$gfpoly: .quad 1, 0xc200000000000000 L$gfpoly_and_internal_carrybit: .quad 1, 0xc200000000000001 .p2align 5 L$ctr_pattern: .quad 0, 0 .quad 1, 0 L$inc_2blocks: .quad 2, 0 .quad 2, 0 .text .globl _gcm_init_vpclmulqdq_avx2 .private_extern _gcm_init_vpclmulqdq_avx2 .p2align 5 _gcm_init_vpclmulqdq_avx2: _CET_ENDBR vpshufd $0x4e,(%rsi),%xmm3 vpshufd $0xd3,%xmm3,%xmm0 vpsrad $31,%xmm0,%xmm0 vpaddq %xmm3,%xmm3,%xmm3 vpand L$gfpoly_and_internal_carrybit(%rip),%xmm0,%xmm0 vpxor %xmm0,%xmm3,%xmm3 vbroadcasti128 L$gfpoly(%rip),%ymm6 vpclmulqdq $0x00,%xmm3,%xmm3,%xmm0 vpclmulqdq $0x01,%xmm3,%xmm3,%xmm1 vpclmulqdq $0x10,%xmm3,%xmm3,%xmm2 vpxor %xmm2,%xmm1,%xmm1 vpclmulqdq $0x01,%xmm0,%xmm6,%xmm2 vpshufd $0x4e,%xmm0,%xmm0 vpxor %xmm0,%xmm1,%xmm1 vpxor %xmm2,%xmm1,%xmm1 vpclmulqdq $0x11,%xmm3,%xmm3,%xmm5 vpclmulqdq $0x01,%xmm1,%xmm6,%xmm0 vpshufd $0x4e,%xmm1,%xmm1 vpxor %xmm1,%xmm5,%xmm5 vpxor %xmm0,%xmm5,%xmm5 vinserti128 $1,%xmm3,%ymm5,%ymm3 vinserti128 $1,%xmm5,%ymm5,%ymm5 .byte 0xc4,0xe3,0x65,0x44,0xc5,0x00 .byte 0xc4,0xe3,0x65,0x44,0xcd,0x01 .byte 0xc4,0xe3,0x65,0x44,0xd5,0x10 vpxor %ymm2,%ymm1,%ymm1 .byte 0xc4,0xe3,0x4d,0x44,0xd0,0x01 vpshufd $0x4e,%ymm0,%ymm0 vpxor %ymm0,%ymm1,%ymm1 vpxor %ymm2,%ymm1,%ymm1 .byte 0xc4,0xe3,0x65,0x44,0xe5,0x11 .byte 0xc4,0xe3,0x4d,0x44,0xc1,0x01 vpshufd $0x4e,%ymm1,%ymm1 vpxor %ymm1,%ymm4,%ymm4 vpxor %ymm0,%ymm4,%ymm4 vmovdqu %ymm3,96(%rdi) vmovdqu %ymm4,64(%rdi) vpunpcklqdq %ymm3,%ymm4,%ymm0 vpunpckhqdq %ymm3,%ymm4,%ymm1 vpxor %ymm1,%ymm0,%ymm0 vmovdqu %ymm0,128+32(%rdi) .byte 0xc4,0xe3,0x5d,0x44,0xc5,0x00 .byte 0xc4,0xe3,0x5d,0x44,0xcd,0x01 .byte 0xc4,0xe3,0x5d,0x44,0xd5,0x10 vpxor %ymm2,%ymm1,%ymm1 .byte 0xc4,0xe3,0x4d,0x44,0xd0,0x01 vpshufd $0x4e,%ymm0,%ymm0 vpxor %ymm0,%ymm1,%ymm1 vpxor %ymm2,%ymm1,%ymm1 .byte 0xc4,0xe3,0x5d,0x44,0xdd,0x11 .byte 0xc4,0xe3,0x4d,0x44,0xc1,0x01 vpshufd $0x4e,%ymm1,%ymm1 vpxor %ymm1,%ymm3,%ymm3 vpxor %ymm0,%ymm3,%ymm3 .byte 0xc4,0xe3,0x65,0x44,0xc5,0x00 .byte 0xc4,0xe3,0x65,0x44,0xcd,0x01 .byte 0xc4,0xe3,0x65,0x44,0xd5,0x10 vpxor %ymm2,%ymm1,%ymm1 .byte 0xc4,0xe3,0x4d,0x44,0xd0,0x01 vpshufd $0x4e,%ymm0,%ymm0 vpxor %ymm0,%ymm1,%ymm1 vpxor %ymm2,%ymm1,%ymm1 .byte 0xc4,0xe3,0x65,0x44,0xe5,0x11 .byte 0xc4,0xe3,0x4d,0x44,0xc1,0x01 vpshufd $0x4e,%ymm1,%ymm1 vpxor %ymm1,%ymm4,%ymm4 vpxor %ymm0,%ymm4,%ymm4 vmovdqu %ymm3,32(%rdi) vmovdqu %ymm4,0(%rdi) vpunpcklqdq %ymm3,%ymm4,%ymm0 vpunpckhqdq %ymm3,%ymm4,%ymm1 vpxor %ymm1,%ymm0,%ymm0 vmovdqu %ymm0,128(%rdi) vzeroupper ret .globl _gcm_ghash_vpclmulqdq_avx2_1 .private_extern _gcm_ghash_vpclmulqdq_avx2_1 .p2align 5 _gcm_ghash_vpclmulqdq_avx2_1: _CET_ENDBR vmovdqu L$bswap_mask(%rip),%xmm6 vmovdqu L$gfpoly(%rip),%xmm7 vmovdqu (%rdi),%xmm5 vpshufb %xmm6,%xmm5,%xmm5 L$ghash_lastblock: vmovdqu (%rdx),%xmm0 vpshufb %xmm6,%xmm0,%xmm0 vpxor %xmm0,%xmm5,%xmm5 vmovdqu 128-16(%rsi),%xmm0 vpclmulqdq $0x00,%xmm0,%xmm5,%xmm1 vpclmulqdq $0x01,%xmm0,%xmm5,%xmm2 vpclmulqdq $0x10,%xmm0,%xmm5,%xmm3 vpxor %xmm3,%xmm2,%xmm2 vpclmulqdq $0x01,%xmm1,%xmm7,%xmm3 vpshufd $0x4e,%xmm1,%xmm1 vpxor %xmm1,%xmm2,%xmm2 vpxor %xmm3,%xmm2,%xmm2 vpclmulqdq $0x11,%xmm0,%xmm5,%xmm5 vpclmulqdq $0x01,%xmm2,%xmm7,%xmm1 vpshufd $0x4e,%xmm2,%xmm2 vpxor %xmm2,%xmm5,%xmm5 vpxor %xmm1,%xmm5,%xmm5 L$ghash_done: vpshufb %xmm6,%xmm5,%xmm5 vmovdqu %xmm5,(%rdi) vzeroupper ret .globl _aes_gcm_enc_update_vaes_avx2 .private_extern _aes_gcm_enc_update_vaes_avx2 .p2align 5 _aes_gcm_enc_update_vaes_avx2: _CET_ENDBR pushq %r12 movq 16(%rsp),%r12 #ifdef BORINGSSL_DISPATCH_TEST movb $1,_BORINGSSL_function_hit+8(%rip) #endif vbroadcasti128 L$bswap_mask(%rip),%ymm0 vmovdqu (%r12),%xmm1 vpshufb %xmm0,%xmm1,%xmm1 vbroadcasti128 (%r8),%ymm11 vpshufb %ymm0,%ymm11,%ymm11 movl 240(%rcx),%r10d leal -20(,%r10,4),%r10d leaq 96(%rcx,%r10,4),%r11 vbroadcasti128 (%rcx),%ymm9 vbroadcasti128 (%r11),%ymm10 vpaddd L$ctr_pattern(%rip),%ymm11,%ymm11 cmpq $127,%rdx jbe L$crypt_loop_4x_done__func1 vmovdqu 128(%r9),%ymm7 vmovdqu 128+32(%r9),%ymm8 vmovdqu L$inc_2blocks(%rip),%ymm2 vpshufb %ymm0,%ymm11,%ymm12 vpaddd %ymm2,%ymm11,%ymm11 vpshufb %ymm0,%ymm11,%ymm13 vpaddd %ymm2,%ymm11,%ymm11 vpshufb %ymm0,%ymm11,%ymm14 vpaddd %ymm2,%ymm11,%ymm11 vpshufb %ymm0,%ymm11,%ymm15 vpaddd %ymm2,%ymm11,%ymm11 vpxor %ymm9,%ymm12,%ymm12 vpxor %ymm9,%ymm13,%ymm13 vpxor %ymm9,%ymm14,%ymm14 vpxor %ymm9,%ymm15,%ymm15 leaq 16(%rcx),%rax L$vaesenc_loop_first_4_vecs__func1: vbroadcasti128 (%rax),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa addq $16,%rax cmpq %rax,%r11 jne L$vaesenc_loop_first_4_vecs__func1 vpxor 0(%rdi),%ymm10,%ymm2 vpxor 32(%rdi),%ymm10,%ymm3 vpxor 64(%rdi),%ymm10,%ymm5 vpxor 96(%rdi),%ymm10,%ymm6 .byte 0xc4,0x62,0x1d,0xdd,0xe2 .byte 0xc4,0x62,0x15,0xdd,0xeb .byte 0xc4,0x62,0x0d,0xdd,0xf5 .byte 0xc4,0x62,0x05,0xdd,0xfe vmovdqu %ymm12,0(%rsi) vmovdqu %ymm13,32(%rsi) vmovdqu %ymm14,64(%rsi) vmovdqu %ymm15,96(%rsi) subq $-128,%rdi addq $-128,%rdx cmpq $127,%rdx jbe L$ghash_last_ciphertext_4x__func1 .p2align 4 L$crypt_loop_4x__func1: vmovdqu L$inc_2blocks(%rip),%ymm2 vpshufb %ymm0,%ymm11,%ymm12 vpaddd %ymm2,%ymm11,%ymm11 vpshufb %ymm0,%ymm11,%ymm13 vpaddd %ymm2,%ymm11,%ymm11 vpshufb %ymm0,%ymm11,%ymm14 vpaddd %ymm2,%ymm11,%ymm11 vpshufb %ymm0,%ymm11,%ymm15 vpaddd %ymm2,%ymm11,%ymm11 vpxor %ymm9,%ymm12,%ymm12 vpxor %ymm9,%ymm13,%ymm13 vpxor %ymm9,%ymm14,%ymm14 vpxor %ymm9,%ymm15,%ymm15 cmpl $24,%r10d jl L$aes128__func1 je L$aes192__func1 vbroadcasti128 -208(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vbroadcasti128 -192(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa L$aes192__func1: vbroadcasti128 -176(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vbroadcasti128 -160(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa L$aes128__func1: prefetcht0 512(%rdi) prefetcht0 512+64(%rdi) vmovdqu 0(%rsi),%ymm3 vpshufb %ymm0,%ymm3,%ymm3 vmovdqu 0(%r9),%ymm4 vpxor %ymm1,%ymm3,%ymm3 .byte 0xc4,0xe3,0x65,0x44,0xec,0x00 .byte 0xc4,0xe3,0x65,0x44,0xcc,0x11 vpunpckhqdq %ymm3,%ymm3,%ymm2 vpxor %ymm3,%ymm2,%ymm2 .byte 0xc4,0xe3,0x6d,0x44,0xf7,0x00 vbroadcasti128 -144(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vbroadcasti128 -128(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vmovdqu 32(%rsi),%ymm3 vpshufb %ymm0,%ymm3,%ymm3 vmovdqu 32(%r9),%ymm4 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x00 vpxor %ymm2,%ymm5,%ymm5 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x11 vpxor %ymm2,%ymm1,%ymm1 vpunpckhqdq %ymm3,%ymm3,%ymm2 vpxor %ymm3,%ymm2,%ymm2 .byte 0xc4,0xe3,0x6d,0x44,0xd7,0x10 vpxor %ymm2,%ymm6,%ymm6 vbroadcasti128 -112(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vmovdqu 64(%rsi),%ymm3 vpshufb %ymm0,%ymm3,%ymm3 vmovdqu 64(%r9),%ymm4 vbroadcasti128 -96(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa .byte 0xc4,0xe3,0x65,0x44,0xd4,0x00 vpxor %ymm2,%ymm5,%ymm5 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x11 vpxor %ymm2,%ymm1,%ymm1 vbroadcasti128 -80(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vpunpckhqdq %ymm3,%ymm3,%ymm2 vpxor %ymm3,%ymm2,%ymm2 .byte 0xc4,0xc3,0x6d,0x44,0xd0,0x00 vpxor %ymm2,%ymm6,%ymm6 vmovdqu 96(%rsi),%ymm3 vpshufb %ymm0,%ymm3,%ymm3 vbroadcasti128 -64(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vmovdqu 96(%r9),%ymm4 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x00 vpxor %ymm2,%ymm5,%ymm5 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x11 vpxor %ymm2,%ymm1,%ymm1 vpunpckhqdq %ymm3,%ymm3,%ymm2 vpxor %ymm3,%ymm2,%ymm2 .byte 0xc4,0xc3,0x6d,0x44,0xd0,0x10 vpxor %ymm2,%ymm6,%ymm6 vbroadcasti128 -48(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vpxor %ymm5,%ymm6,%ymm6 vpxor %ymm1,%ymm6,%ymm6 vbroadcasti128 L$gfpoly(%rip),%ymm4 .byte 0xc4,0xe3,0x5d,0x44,0xd5,0x01 vpshufd $0x4e,%ymm5,%ymm5 vpxor %ymm5,%ymm6,%ymm6 vpxor %ymm2,%ymm6,%ymm6 vbroadcasti128 -32(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa .byte 0xc4,0xe3,0x5d,0x44,0xd6,0x01 vpshufd $0x4e,%ymm6,%ymm6 vpxor %ymm6,%ymm1,%ymm1 vpxor %ymm2,%ymm1,%ymm1 vbroadcasti128 -16(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vextracti128 $1,%ymm1,%xmm2 vpxor %xmm2,%xmm1,%xmm1 subq $-128,%rsi vpxor 0(%rdi),%ymm10,%ymm2 vpxor 32(%rdi),%ymm10,%ymm3 vpxor 64(%rdi),%ymm10,%ymm5 vpxor 96(%rdi),%ymm10,%ymm6 .byte 0xc4,0x62,0x1d,0xdd,0xe2 .byte 0xc4,0x62,0x15,0xdd,0xeb .byte 0xc4,0x62,0x0d,0xdd,0xf5 .byte 0xc4,0x62,0x05,0xdd,0xfe vmovdqu %ymm12,0(%rsi) vmovdqu %ymm13,32(%rsi) vmovdqu %ymm14,64(%rsi) vmovdqu %ymm15,96(%rsi) subq $-128,%rdi addq $-128,%rdx cmpq $127,%rdx ja L$crypt_loop_4x__func1 L$ghash_last_ciphertext_4x__func1: vmovdqu 0(%rsi),%ymm3 vpshufb %ymm0,%ymm3,%ymm3 vmovdqu 0(%r9),%ymm4 vpxor %ymm1,%ymm3,%ymm3 .byte 0xc4,0xe3,0x65,0x44,0xec,0x00 .byte 0xc4,0xe3,0x65,0x44,0xcc,0x11 vpunpckhqdq %ymm3,%ymm3,%ymm2 vpxor %ymm3,%ymm2,%ymm2 .byte 0xc4,0xe3,0x6d,0x44,0xf7,0x00 vmovdqu 32(%rsi),%ymm3 vpshufb %ymm0,%ymm3,%ymm3 vmovdqu 32(%r9),%ymm4 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x00 vpxor %ymm2,%ymm5,%ymm5 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x11 vpxor %ymm2,%ymm1,%ymm1 vpunpckhqdq %ymm3,%ymm3,%ymm2 vpxor %ymm3,%ymm2,%ymm2 .byte 0xc4,0xe3,0x6d,0x44,0xd7,0x10 vpxor %ymm2,%ymm6,%ymm6 vmovdqu 64(%rsi),%ymm3 vpshufb %ymm0,%ymm3,%ymm3 vmovdqu 64(%r9),%ymm4 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x00 vpxor %ymm2,%ymm5,%ymm5 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x11 vpxor %ymm2,%ymm1,%ymm1 vpunpckhqdq %ymm3,%ymm3,%ymm2 vpxor %ymm3,%ymm2,%ymm2 .byte 0xc4,0xc3,0x6d,0x44,0xd0,0x00 vpxor %ymm2,%ymm6,%ymm6 vmovdqu 96(%rsi),%ymm3 vpshufb %ymm0,%ymm3,%ymm3 vmovdqu 96(%r9),%ymm4 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x00 vpxor %ymm2,%ymm5,%ymm5 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x11 vpxor %ymm2,%ymm1,%ymm1 vpunpckhqdq %ymm3,%ymm3,%ymm2 vpxor %ymm3,%ymm2,%ymm2 .byte 0xc4,0xc3,0x6d,0x44,0xd0,0x10 vpxor %ymm2,%ymm6,%ymm6 vpxor %ymm5,%ymm6,%ymm6 vpxor %ymm1,%ymm6,%ymm6 vbroadcasti128 L$gfpoly(%rip),%ymm4 .byte 0xc4,0xe3,0x5d,0x44,0xd5,0x01 vpshufd $0x4e,%ymm5,%ymm5 vpxor %ymm5,%ymm6,%ymm6 vpxor %ymm2,%ymm6,%ymm6 .byte 0xc4,0xe3,0x5d,0x44,0xd6,0x01 vpshufd $0x4e,%ymm6,%ymm6 vpxor %ymm6,%ymm1,%ymm1 vpxor %ymm2,%ymm1,%ymm1 vextracti128 $1,%ymm1,%xmm2 vpxor %xmm2,%xmm1,%xmm1 subq $-128,%rsi L$crypt_loop_4x_done__func1: testq %rdx,%rdx jz L$done__func1 leaq 128(%r9),%r8 subq %rdx,%r8 vpxor %xmm5,%xmm5,%xmm5 vpxor %xmm6,%xmm6,%xmm6 vpxor %xmm7,%xmm7,%xmm7 cmpq $64,%rdx jb L$lessthan64bytes__func1 vpshufb %ymm0,%ymm11,%ymm12 vpaddd L$inc_2blocks(%rip),%ymm11,%ymm11 vpshufb %ymm0,%ymm11,%ymm13 vpaddd L$inc_2blocks(%rip),%ymm11,%ymm11 vpxor %ymm9,%ymm12,%ymm12 vpxor %ymm9,%ymm13,%ymm13 leaq 16(%rcx),%rax L$vaesenc_loop_tail_1__func1: vbroadcasti128 (%rax),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea addq $16,%rax cmpq %rax,%r11 jne L$vaesenc_loop_tail_1__func1 .byte 0xc4,0x42,0x1d,0xdd,0xe2 .byte 0xc4,0x42,0x15,0xdd,0xea vmovdqu 0(%rdi),%ymm2 vmovdqu 32(%rdi),%ymm3 vpxor %ymm2,%ymm12,%ymm12 vpxor %ymm3,%ymm13,%ymm13 vmovdqu %ymm12,0(%rsi) vmovdqu %ymm13,32(%rsi) vpshufb %ymm0,%ymm12,%ymm12 vpshufb %ymm0,%ymm13,%ymm13 vpxor %ymm1,%ymm12,%ymm12 vmovdqu (%r8),%ymm2 vmovdqu 32(%r8),%ymm3 .byte 0xc4,0xe3,0x1d,0x44,0xea,0x00 .byte 0xc4,0xe3,0x1d,0x44,0xf2,0x01 .byte 0xc4,0xe3,0x1d,0x44,0xe2,0x10 vpxor %ymm4,%ymm6,%ymm6 .byte 0xc4,0xe3,0x1d,0x44,0xfa,0x11 .byte 0xc4,0xe3,0x15,0x44,0xe3,0x00 vpxor %ymm4,%ymm5,%ymm5 .byte 0xc4,0xe3,0x15,0x44,0xe3,0x01 vpxor %ymm4,%ymm6,%ymm6 .byte 0xc4,0xe3,0x15,0x44,0xe3,0x10 vpxor %ymm4,%ymm6,%ymm6 .byte 0xc4,0xe3,0x15,0x44,0xe3,0x11 vpxor %ymm4,%ymm7,%ymm7 addq $64,%r8 addq $64,%rdi addq $64,%rsi subq $64,%rdx jz L$reduce__func1 vpxor %xmm1,%xmm1,%xmm1 L$lessthan64bytes__func1: vpshufb %ymm0,%ymm11,%ymm12 vpaddd L$inc_2blocks(%rip),%ymm11,%ymm11 vpshufb %ymm0,%ymm11,%ymm13 vpxor %ymm9,%ymm12,%ymm12 vpxor %ymm9,%ymm13,%ymm13 leaq 16(%rcx),%rax L$vaesenc_loop_tail_2__func1: vbroadcasti128 (%rax),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea addq $16,%rax cmpq %rax,%r11 jne L$vaesenc_loop_tail_2__func1 .byte 0xc4,0x42,0x1d,0xdd,0xe2 .byte 0xc4,0x42,0x15,0xdd,0xea cmpq $32,%rdx jb L$xor_one_block__func1 je L$xor_two_blocks__func1 L$xor_three_blocks__func1: vmovdqu 0(%rdi),%ymm2 vmovdqu 32(%rdi),%xmm3 vpxor %ymm2,%ymm12,%ymm12 vpxor %xmm3,%xmm13,%xmm13 vmovdqu %ymm12,0(%rsi) vmovdqu %xmm13,32(%rsi) vpshufb %ymm0,%ymm12,%ymm12 vpshufb %xmm0,%xmm13,%xmm13 vpxor %ymm1,%ymm12,%ymm12 vmovdqu (%r8),%ymm2 vmovdqu 32(%r8),%xmm3 vpclmulqdq $0x00,%xmm3,%xmm13,%xmm4 vpxor %ymm4,%ymm5,%ymm5 vpclmulqdq $0x01,%xmm3,%xmm13,%xmm4 vpxor %ymm4,%ymm6,%ymm6 vpclmulqdq $0x10,%xmm3,%xmm13,%xmm4 vpxor %ymm4,%ymm6,%ymm6 vpclmulqdq $0x11,%xmm3,%xmm13,%xmm4 vpxor %ymm4,%ymm7,%ymm7 jmp L$ghash_mul_one_vec_unreduced__func1 L$xor_two_blocks__func1: vmovdqu (%rdi),%ymm2 vpxor %ymm2,%ymm12,%ymm12 vmovdqu %ymm12,(%rsi) vpshufb %ymm0,%ymm12,%ymm12 vpxor %ymm1,%ymm12,%ymm12 vmovdqu (%r8),%ymm2 jmp L$ghash_mul_one_vec_unreduced__func1 L$xor_one_block__func1: vmovdqu (%rdi),%xmm2 vpxor %xmm2,%xmm12,%xmm12 vmovdqu %xmm12,(%rsi) vpshufb %xmm0,%xmm12,%xmm12 vpxor %xmm1,%xmm12,%xmm12 vmovdqu (%r8),%xmm2 L$ghash_mul_one_vec_unreduced__func1: .byte 0xc4,0xe3,0x1d,0x44,0xe2,0x00 vpxor %ymm4,%ymm5,%ymm5 .byte 0xc4,0xe3,0x1d,0x44,0xe2,0x01 vpxor %ymm4,%ymm6,%ymm6 .byte 0xc4,0xe3,0x1d,0x44,0xe2,0x10 vpxor %ymm4,%ymm6,%ymm6 .byte 0xc4,0xe3,0x1d,0x44,0xe2,0x11 vpxor %ymm4,%ymm7,%ymm7 L$reduce__func1: vbroadcasti128 L$gfpoly(%rip),%ymm2 .byte 0xc4,0xe3,0x6d,0x44,0xdd,0x01 vpshufd $0x4e,%ymm5,%ymm5 vpxor %ymm5,%ymm6,%ymm6 vpxor %ymm3,%ymm6,%ymm6 .byte 0xc4,0xe3,0x6d,0x44,0xde,0x01 vpshufd $0x4e,%ymm6,%ymm6 vpxor %ymm6,%ymm7,%ymm7 vpxor %ymm3,%ymm7,%ymm7 vextracti128 $1,%ymm7,%xmm1 vpxor %xmm7,%xmm1,%xmm1 L$done__func1: vpshufb %xmm0,%xmm1,%xmm1 vmovdqu %xmm1,(%r12) vzeroupper popq %r12 ret .globl _aes_gcm_dec_update_vaes_avx2 .private_extern _aes_gcm_dec_update_vaes_avx2 .p2align 5 _aes_gcm_dec_update_vaes_avx2: _CET_ENDBR pushq %r12 movq 16(%rsp),%r12 vbroadcasti128 L$bswap_mask(%rip),%ymm0 vmovdqu (%r12),%xmm1 vpshufb %xmm0,%xmm1,%xmm1 vbroadcasti128 (%r8),%ymm11 vpshufb %ymm0,%ymm11,%ymm11 movl 240(%rcx),%r10d leal -20(,%r10,4),%r10d leaq 96(%rcx,%r10,4),%r11 vbroadcasti128 (%rcx),%ymm9 vbroadcasti128 (%r11),%ymm10 vpaddd L$ctr_pattern(%rip),%ymm11,%ymm11 cmpq $127,%rdx jbe L$crypt_loop_4x_done__func2 vmovdqu 128(%r9),%ymm7 vmovdqu 128+32(%r9),%ymm8 .p2align 4 L$crypt_loop_4x__func2: vmovdqu L$inc_2blocks(%rip),%ymm2 vpshufb %ymm0,%ymm11,%ymm12 vpaddd %ymm2,%ymm11,%ymm11 vpshufb %ymm0,%ymm11,%ymm13 vpaddd %ymm2,%ymm11,%ymm11 vpshufb %ymm0,%ymm11,%ymm14 vpaddd %ymm2,%ymm11,%ymm11 vpshufb %ymm0,%ymm11,%ymm15 vpaddd %ymm2,%ymm11,%ymm11 vpxor %ymm9,%ymm12,%ymm12 vpxor %ymm9,%ymm13,%ymm13 vpxor %ymm9,%ymm14,%ymm14 vpxor %ymm9,%ymm15,%ymm15 cmpl $24,%r10d jl L$aes128__func2 je L$aes192__func2 vbroadcasti128 -208(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vbroadcasti128 -192(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa L$aes192__func2: vbroadcasti128 -176(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vbroadcasti128 -160(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa L$aes128__func2: prefetcht0 512(%rdi) prefetcht0 512+64(%rdi) vmovdqu 0(%rdi),%ymm3 vpshufb %ymm0,%ymm3,%ymm3 vmovdqu 0(%r9),%ymm4 vpxor %ymm1,%ymm3,%ymm3 .byte 0xc4,0xe3,0x65,0x44,0xec,0x00 .byte 0xc4,0xe3,0x65,0x44,0xcc,0x11 vpunpckhqdq %ymm3,%ymm3,%ymm2 vpxor %ymm3,%ymm2,%ymm2 .byte 0xc4,0xe3,0x6d,0x44,0xf7,0x00 vbroadcasti128 -144(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vbroadcasti128 -128(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vmovdqu 32(%rdi),%ymm3 vpshufb %ymm0,%ymm3,%ymm3 vmovdqu 32(%r9),%ymm4 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x00 vpxor %ymm2,%ymm5,%ymm5 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x11 vpxor %ymm2,%ymm1,%ymm1 vpunpckhqdq %ymm3,%ymm3,%ymm2 vpxor %ymm3,%ymm2,%ymm2 .byte 0xc4,0xe3,0x6d,0x44,0xd7,0x10 vpxor %ymm2,%ymm6,%ymm6 vbroadcasti128 -112(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vmovdqu 64(%rdi),%ymm3 vpshufb %ymm0,%ymm3,%ymm3 vmovdqu 64(%r9),%ymm4 vbroadcasti128 -96(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa .byte 0xc4,0xe3,0x65,0x44,0xd4,0x00 vpxor %ymm2,%ymm5,%ymm5 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x11 vpxor %ymm2,%ymm1,%ymm1 vbroadcasti128 -80(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vpunpckhqdq %ymm3,%ymm3,%ymm2 vpxor %ymm3,%ymm2,%ymm2 .byte 0xc4,0xc3,0x6d,0x44,0xd0,0x00 vpxor %ymm2,%ymm6,%ymm6 vmovdqu 96(%rdi),%ymm3 vpshufb %ymm0,%ymm3,%ymm3 vbroadcasti128 -64(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vmovdqu 96(%r9),%ymm4 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x00 vpxor %ymm2,%ymm5,%ymm5 .byte 0xc4,0xe3,0x65,0x44,0xd4,0x11 vpxor %ymm2,%ymm1,%ymm1 vpunpckhqdq %ymm3,%ymm3,%ymm2 vpxor %ymm3,%ymm2,%ymm2 .byte 0xc4,0xc3,0x6d,0x44,0xd0,0x10 vpxor %ymm2,%ymm6,%ymm6 vbroadcasti128 -48(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vpxor %ymm5,%ymm6,%ymm6 vpxor %ymm1,%ymm6,%ymm6 vbroadcasti128 L$gfpoly(%rip),%ymm4 .byte 0xc4,0xe3,0x5d,0x44,0xd5,0x01 vpshufd $0x4e,%ymm5,%ymm5 vpxor %ymm5,%ymm6,%ymm6 vpxor %ymm2,%ymm6,%ymm6 vbroadcasti128 -32(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa .byte 0xc4,0xe3,0x5d,0x44,0xd6,0x01 vpshufd $0x4e,%ymm6,%ymm6 vpxor %ymm6,%ymm1,%ymm1 vpxor %ymm2,%ymm1,%ymm1 vbroadcasti128 -16(%r11),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea .byte 0xc4,0x62,0x0d,0xdc,0xf2 .byte 0xc4,0x62,0x05,0xdc,0xfa vextracti128 $1,%ymm1,%xmm2 vpxor %xmm2,%xmm1,%xmm1 vpxor 0(%rdi),%ymm10,%ymm2 vpxor 32(%rdi),%ymm10,%ymm3 vpxor 64(%rdi),%ymm10,%ymm5 vpxor 96(%rdi),%ymm10,%ymm6 .byte 0xc4,0x62,0x1d,0xdd,0xe2 .byte 0xc4,0x62,0x15,0xdd,0xeb .byte 0xc4,0x62,0x0d,0xdd,0xf5 .byte 0xc4,0x62,0x05,0xdd,0xfe vmovdqu %ymm12,0(%rsi) vmovdqu %ymm13,32(%rsi) vmovdqu %ymm14,64(%rsi) vmovdqu %ymm15,96(%rsi) subq $-128,%rdi subq $-128,%rsi addq $-128,%rdx cmpq $127,%rdx ja L$crypt_loop_4x__func2 L$crypt_loop_4x_done__func2: testq %rdx,%rdx jz L$done__func2 leaq 128(%r9),%r8 subq %rdx,%r8 vpxor %xmm5,%xmm5,%xmm5 vpxor %xmm6,%xmm6,%xmm6 vpxor %xmm7,%xmm7,%xmm7 cmpq $64,%rdx jb L$lessthan64bytes__func2 vpshufb %ymm0,%ymm11,%ymm12 vpaddd L$inc_2blocks(%rip),%ymm11,%ymm11 vpshufb %ymm0,%ymm11,%ymm13 vpaddd L$inc_2blocks(%rip),%ymm11,%ymm11 vpxor %ymm9,%ymm12,%ymm12 vpxor %ymm9,%ymm13,%ymm13 leaq 16(%rcx),%rax L$vaesenc_loop_tail_1__func2: vbroadcasti128 (%rax),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea addq $16,%rax cmpq %rax,%r11 jne L$vaesenc_loop_tail_1__func2 .byte 0xc4,0x42,0x1d,0xdd,0xe2 .byte 0xc4,0x42,0x15,0xdd,0xea vmovdqu 0(%rdi),%ymm2 vmovdqu 32(%rdi),%ymm3 vpxor %ymm2,%ymm12,%ymm12 vpxor %ymm3,%ymm13,%ymm13 vmovdqu %ymm12,0(%rsi) vmovdqu %ymm13,32(%rsi) vpshufb %ymm0,%ymm2,%ymm12 vpshufb %ymm0,%ymm3,%ymm13 vpxor %ymm1,%ymm12,%ymm12 vmovdqu (%r8),%ymm2 vmovdqu 32(%r8),%ymm3 .byte 0xc4,0xe3,0x1d,0x44,0xea,0x00 .byte 0xc4,0xe3,0x1d,0x44,0xf2,0x01 .byte 0xc4,0xe3,0x1d,0x44,0xe2,0x10 vpxor %ymm4,%ymm6,%ymm6 .byte 0xc4,0xe3,0x1d,0x44,0xfa,0x11 .byte 0xc4,0xe3,0x15,0x44,0xe3,0x00 vpxor %ymm4,%ymm5,%ymm5 .byte 0xc4,0xe3,0x15,0x44,0xe3,0x01 vpxor %ymm4,%ymm6,%ymm6 .byte 0xc4,0xe3,0x15,0x44,0xe3,0x10 vpxor %ymm4,%ymm6,%ymm6 .byte 0xc4,0xe3,0x15,0x44,0xe3,0x11 vpxor %ymm4,%ymm7,%ymm7 addq $64,%r8 addq $64,%rdi addq $64,%rsi subq $64,%rdx jz L$reduce__func2 vpxor %xmm1,%xmm1,%xmm1 L$lessthan64bytes__func2: vpshufb %ymm0,%ymm11,%ymm12 vpaddd L$inc_2blocks(%rip),%ymm11,%ymm11 vpshufb %ymm0,%ymm11,%ymm13 vpxor %ymm9,%ymm12,%ymm12 vpxor %ymm9,%ymm13,%ymm13 leaq 16(%rcx),%rax L$vaesenc_loop_tail_2__func2: vbroadcasti128 (%rax),%ymm2 .byte 0xc4,0x62,0x1d,0xdc,0xe2 .byte 0xc4,0x62,0x15,0xdc,0xea addq $16,%rax cmpq %rax,%r11 jne L$vaesenc_loop_tail_2__func2 .byte 0xc4,0x42,0x1d,0xdd,0xe2 .byte 0xc4,0x42,0x15,0xdd,0xea cmpq $32,%rdx jb L$xor_one_block__func2 je L$xor_two_blocks__func2 L$xor_three_blocks__func2: vmovdqu 0(%rdi),%ymm2 vmovdqu 32(%rdi),%xmm3 vpxor %ymm2,%ymm12,%ymm12 vpxor %xmm3,%xmm13,%xmm13 vmovdqu %ymm12,0(%rsi) vmovdqu %xmm13,32(%rsi) vpshufb %ymm0,%ymm2,%ymm12 vpshufb %xmm0,%xmm3,%xmm13 vpxor %ymm1,%ymm12,%ymm12 vmovdqu (%r8),%ymm2 vmovdqu 32(%r8),%xmm3 vpclmulqdq $0x00,%xmm3,%xmm13,%xmm4 vpxor %ymm4,%ymm5,%ymm5 vpclmulqdq $0x01,%xmm3,%xmm13,%xmm4 vpxor %ymm4,%ymm6,%ymm6 vpclmulqdq $0x10,%xmm3,%xmm13,%xmm4 vpxor %ymm4,%ymm6,%ymm6 vpclmulqdq $0x11,%xmm3,%xmm13,%xmm4 vpxor %ymm4,%ymm7,%ymm7 jmp L$ghash_mul_one_vec_unreduced__func2 L$xor_two_blocks__func2: vmovdqu (%rdi),%ymm2 vpxor %ymm2,%ymm12,%ymm12 vmovdqu %ymm12,(%rsi) vpshufb %ymm0,%ymm2,%ymm12 vpxor %ymm1,%ymm12,%ymm12 vmovdqu (%r8),%ymm2 jmp L$ghash_mul_one_vec_unreduced__func2 L$xor_one_block__func2: vmovdqu (%rdi),%xmm2 vpxor %xmm2,%xmm12,%xmm12 vmovdqu %xmm12,(%rsi) vpshufb %xmm0,%xmm2,%xmm12 vpxor %xmm1,%xmm12,%xmm12 vmovdqu (%r8),%xmm2 L$ghash_mul_one_vec_unreduced__func2: .byte 0xc4,0xe3,0x1d,0x44,0xe2,0x00 vpxor %ymm4,%ymm5,%ymm5 .byte 0xc4,0xe3,0x1d,0x44,0xe2,0x01 vpxor %ymm4,%ymm6,%ymm6 .byte 0xc4,0xe3,0x1d,0x44,0xe2,0x10 vpxor %ymm4,%ymm6,%ymm6 .byte 0xc4,0xe3,0x1d,0x44,0xe2,0x11 vpxor %ymm4,%ymm7,%ymm7 L$reduce__func2: vbroadcasti128 L$gfpoly(%rip),%ymm2 .byte 0xc4,0xe3,0x6d,0x44,0xdd,0x01 vpshufd $0x4e,%ymm5,%ymm5 vpxor %ymm5,%ymm6,%ymm6 vpxor %ymm3,%ymm6,%ymm6 .byte 0xc4,0xe3,0x6d,0x44,0xde,0x01 vpshufd $0x4e,%ymm6,%ymm6 vpxor %ymm6,%ymm7,%ymm7 vpxor %ymm3,%ymm7,%ymm7 vextracti128 $1,%ymm7,%xmm1 vpxor %xmm7,%xmm1,%xmm1 L$done__func2: vpshufb %xmm0,%xmm1,%xmm1 vmovdqu %xmm1,(%r12) vzeroupper popq %r12 ret #endif