diff --git a/src/cmd/compile/internal/types2/stdlib_test.go b/src/cmd/compile/internal/types2/stdlib_test.go index 0454e5b93aa..d91d60d1241 100644 --- a/src/cmd/compile/internal/types2/stdlib_test.go +++ b/src/cmd/compile/internal/types2/stdlib_test.go @@ -359,6 +359,7 @@ var excluded = map[string]bool{ "crypto/internal/edwards25519/field/_asm": true, "crypto/sha1/_asm": true, "crypto/sha256/_asm": true, + "crypto/sha512/_asm": true, } // printPackageMu synchronizes the printing of type-checked package files in diff --git a/src/crypto/sha512/_asm/go.mod b/src/crypto/sha512/_asm/go.mod new file mode 100644 index 00000000000..63f129f7c6b --- /dev/null +++ b/src/crypto/sha512/_asm/go.mod @@ -0,0 +1,11 @@ +module std/crypto/sha512/_asm + +go 1.24 + +require github.com/mmcloughlin/avo v0.6.0 + +require ( + golang.org/x/mod v0.20.0 // indirect + golang.org/x/sync v0.8.0 // indirect + golang.org/x/tools v0.24.0 // indirect +) diff --git a/src/crypto/sha512/_asm/go.sum b/src/crypto/sha512/_asm/go.sum new file mode 100644 index 00000000000..76af484b2eb --- /dev/null +++ b/src/crypto/sha512/_asm/go.sum @@ -0,0 +1,8 @@ +github.com/mmcloughlin/avo v0.6.0 h1:QH6FU8SKoTLaVs80GA8TJuLNkUYl4VokHKlPhVDg4YY= +github.com/mmcloughlin/avo v0.6.0/go.mod h1:8CoAGaCSYXtCPR+8y18Y9aB/kxb8JSS6FRI7mSkvD+8= +golang.org/x/mod v0.20.0 h1:utOm6MM3R3dnawAiJgn0y+xvuYRsm1RKM/4giyfDgV0= +golang.org/x/mod v0.20.0/go.mod h1:hTbmBsO62+eylJbnUtE2MGJUyE7QWk4xUqPFrRgJ+7c= +golang.org/x/sync v0.8.0 h1:3NFvSEYkUoMifnESzZl15y791HH1qU2xm6eCJU5ZPXQ= +golang.org/x/sync v0.8.0/go.mod h1:Czt+wKu1gCyEFDUtn0jG5QVvpJ6rzVqr5aXyt9drQfk= +golang.org/x/tools v0.24.0 h1:J1shsA93PJUEVaUSaay7UXAyE8aimq3GW0pjlolpa24= +golang.org/x/tools v0.24.0/go.mod h1:YhNqVBIfWHdzvTLs0d8LCuMhkKUgSUKldakyV7W/WDQ= diff --git a/src/crypto/sha512/_asm/sha512block_amd64_asm.go b/src/crypto/sha512/_asm/sha512block_amd64_asm.go new file mode 100644 index 00000000000..c0959714d01 --- /dev/null +++ b/src/crypto/sha512/_asm/sha512block_amd64_asm.go @@ -0,0 +1,1604 @@ +// Copyright 2024 The Go Authors. All rights reserved. +// Use of this source code is governed by a BSD-style +// license that can be found in the LICENSE file. + +package main + +import ( + . "github.com/mmcloughlin/avo/build" + . "github.com/mmcloughlin/avo/operand" + . "github.com/mmcloughlin/avo/reg" +) + +//go:generate go run . -out ../sha512block_amd64.s -pkg sha512 + +// SHA512 block routine. See sha512block.go for Go equivalent. +// +// The algorithm is detailed in FIPS 180-4: +// +// https://csrc.nist.gov/publications/fips/fips180-4/fips-180-4.pdf +// +// Wt = Mt; for 0 <= t <= 15 +// Wt = SIGMA1(Wt-2) + SIGMA0(Wt-15) + Wt-16; for 16 <= t <= 79 +// +// a = H0 +// b = H1 +// c = H2 +// d = H3 +// e = H4 +// f = H5 +// g = H6 +// h = H7 +// +// for t = 0 to 79 { +// T1 = h + BIGSIGMA1(e) + Ch(e,f,g) + Kt + Wt +// T2 = BIGSIGMA0(a) + Maj(a,b,c) +// h = g +// g = f +// f = e +// e = d + T1 +// d = c +// c = b +// b = a +// a = T1 + T2 +// } +// +// H0 = a + H0 +// H1 = b + H1 +// H2 = c + H2 +// H3 = d + H3 +// H4 = e + H4 +// H5 = f + H5 +// H6 = g + H6 +// H7 = h + H7 + +const ThatPeskyUnicodeDot = "\u00b7" + +var _K = []uint64{ + 0x428a2f98d728ae22, + 0x7137449123ef65cd, + 0xb5c0fbcfec4d3b2f, + 0xe9b5dba58189dbbc, + 0x3956c25bf348b538, + 0x59f111f1b605d019, + 0x923f82a4af194f9b, + 0xab1c5ed5da6d8118, + 0xd807aa98a3030242, + 0x12835b0145706fbe, + 0x243185be4ee4b28c, + 0x550c7dc3d5ffb4e2, + 0x72be5d74f27b896f, + 0x80deb1fe3b1696b1, + 0x9bdc06a725c71235, + 0xc19bf174cf692694, + 0xe49b69c19ef14ad2, + 0xefbe4786384f25e3, + 0x0fc19dc68b8cd5b5, + 0x240ca1cc77ac9c65, + 0x2de92c6f592b0275, + 0x4a7484aa6ea6e483, + 0x5cb0a9dcbd41fbd4, + 0x76f988da831153b5, + 0x983e5152ee66dfab, + 0xa831c66d2db43210, + 0xb00327c898fb213f, + 0xbf597fc7beef0ee4, + 0xc6e00bf33da88fc2, + 0xd5a79147930aa725, + 0x06ca6351e003826f, + 0x142929670a0e6e70, + 0x27b70a8546d22ffc, + 0x2e1b21385c26c926, + 0x4d2c6dfc5ac42aed, + 0x53380d139d95b3df, + 0x650a73548baf63de, + 0x766a0abb3c77b2a8, + 0x81c2c92e47edaee6, + 0x92722c851482353b, + 0xa2bfe8a14cf10364, + 0xa81a664bbc423001, + 0xc24b8b70d0f89791, + 0xc76c51a30654be30, + 0xd192e819d6ef5218, + 0xd69906245565a910, + 0xf40e35855771202a, + 0x106aa07032bbd1b8, + 0x19a4c116b8d2d0c8, + 0x1e376c085141ab53, + 0x2748774cdf8eeb99, + 0x34b0bcb5e19b48a8, + 0x391c0cb3c5c95a63, + 0x4ed8aa4ae3418acb, + 0x5b9cca4f7763e373, + 0x682e6ff3d6b2b8a3, + 0x748f82ee5defb2fc, + 0x78a5636f43172f60, + 0x84c87814a1f0ab72, + 0x8cc702081a6439ec, + 0x90befffa23631e28, + 0xa4506cebde82bde9, + 0xbef9a3f7b2c67915, + 0xc67178f2e372532b, + 0xca273eceea26619c, + 0xd186b8c721c0c207, + 0xeada7dd6cde0eb1e, + 0xf57d4f7fee6ed178, + 0x06f067aa72176fba, + 0x0a637dc5a2c898a6, + 0x113f9804bef90dae, + 0x1b710b35131c471b, + 0x28db77f523047d84, + 0x32caab7b40c72493, + 0x3c9ebe0a15c9bebc, + 0x431d67c49c100d4c, + 0x4cc5d4becb3e42b6, + 0x597f299cfc657e2a, + 0x5fcb6fab3ad6faec, + 0x6c44198c4a475817, +} + +func main() { + Package("crypto/sha512") + ConstraintExpr("!purego") + blockAMD64() + blockAVX2() + Generate() +} + +// Wt = Mt; for 0 <= t <= 15 +// +// Line 50 +func MSGSCHEDULE0(index int) { + MOVQ(Mem{Base: SI}.Offset(index*8), RAX) + BSWAPQ(RAX) + MOVQ(RAX, Mem{Base: BP}.Offset(index*8)) +} + +// Wt = SIGMA1(Wt-2) + Wt-7 + SIGMA0(Wt-15) + Wt-16; for 16 <= t <= 79 +// +// SIGMA0(x) = ROTR(1,x) XOR ROTR(8,x) XOR SHR(7,x) +// SIGMA1(x) = ROTR(19,x) XOR ROTR(61,x) XOR SHR(6,x) +// +// Line 58 +func MSGSCHEDULE1(index int) { + MOVQ(Mem{Base: BP}.Offset((index-2)*8), RAX) + MOVQ(RAX, RCX) + RORQ(Imm(19), RAX) + MOVQ(RCX, RDX) + RORQ(Imm(61), RCX) + SHRQ(Imm(6), RDX) + MOVQ(Mem{Base: BP}.Offset((index-15)*8), RBX) + XORQ(RCX, RAX) + MOVQ(RBX, RCX) + XORQ(RDX, RAX) + RORQ(Imm(1), RBX) + MOVQ(RCX, RDX) + SHRQ(Imm(7), RDX) + RORQ(Imm(8), RCX) + ADDQ(Mem{Base: BP}.Offset((index-7)*8), RAX) + XORQ(RCX, RBX) + XORQ(RDX, RBX) + ADDQ(Mem{Base: BP}.Offset((index-16)*8), RBX) + ADDQ(RBX, RAX) + MOVQ(RAX, Mem{Base: BP}.Offset((index)*8)) +} + +// Calculate T1 in AX - uses AX, CX and DX registers. +// h is also used as an accumulator. Wt is passed in AX. +// +// T1 = h + BIGSIGMA1(e) + Ch(e, f, g) + Kt + Wt +// BIGSIGMA1(x) = ROTR(14,x) XOR ROTR(18,x) XOR ROTR(41,x) +// Ch(x, y, z) = (x AND y) XOR (NOT x AND z) +// +// Line 85 +func SHA512T1(konst uint64, e, f, g, h GPPhysical) { + MOVQ(U64(konst), RDX) + ADDQ(RAX, h) + MOVQ(e, RAX) + ADDQ(RDX, h) + MOVQ(e, RCX) + RORQ(U8(14), RAX) + MOVQ(e, RDX) + RORQ(U8(18), RCX) + XORQ(RCX, RAX) + MOVQ(e, RCX) + RORQ(U8(41), RDX) + ANDQ(f, RCX) + XORQ(RAX, RDX) + MOVQ(e, RAX) + NOTQ(RAX) + ADDQ(RDX, h) + ANDQ(g, RAX) + XORQ(RCX, RAX) + ADDQ(h, RAX) +} + +// Calculate T2 in BX - uses BX, CX, DX and DI registers. +// +// T2 = BIGSIGMA0(a) + Maj(a, b, c) +// BIGSIGMA0(x) = ROTR(28,x) XOR ROTR(34,x) XOR ROTR(39,x) +// Maj(x, y, z) = (x AND y) XOR (x AND z) XOR (y AND z) +// +// Line 110 +func SHA512T2(a, b, c GPPhysical) { + MOVQ(a, RDI) + MOVQ(c, RBX) + RORQ(Imm(28), RDI) + MOVQ(a, RDX) + ANDQ(b, RBX) + RORQ(Imm(34), RDX) + MOVQ(a, RCX) + ANDQ(c, RCX) + XORQ(RDX, RDI) + XORQ(RCX, RBX) + MOVQ(a, RDX) + MOVQ(b, RCX) + RORQ(Imm(39), RDX) + ANDQ(a, RCX) + XORQ(RCX, RBX) + XORQ(RDX, RDI) + ADDQ(RDI, RBX) +} + +// Calculate T1 and T2, then e = d + T1 and a = T1 + T2. +// The values for e and a are stored in d and h, ready for rotation. +// +// Line 131 +func SHA512ROUND(index int, konst uint64, a, b, c, d, e, f, g, h GPPhysical) { + SHA512T1(konst, e, f, g, h) + SHA512T2(a, b, c) + MOVQ(RBX, h) + ADDQ(RAX, d) + ADDQ(RAX, h) +} + +// Line 169 +func SHA512ROUND0(index int, konst uint64, a, b, c, d, e, f, g, h GPPhysical) { + MSGSCHEDULE0(index) + SHA512ROUND(index, konst, a, b, c, d, e, f, g, h) +} + +// Line 142 +func SHA512ROUND1(index int, konst uint64, a, b, c, d, e, f, g, h GPPhysical) { + MSGSCHEDULE1(index) + SHA512ROUND(index, konst, a, b, c, d, e, f, g, h) +} + +// Line 146 +func blockAMD64() { + Implement("blockAMD64") + AllocLocal(648) + + Load(Param("p").Base(), RSI) + Load(Param("p").Len(), RDX) + SHRQ(Imm(7), RDX) + SHLQ(Imm(7), RDX) + + LEAQ(Mem{Base: SI, Index: DX, Scale: 1}, RDI) + MOVQ(RDI, Mem{Base: SP}.Offset(640)) + CMPQ(RSI, RDI) + JEQ(LabelRef("end")) + + Load(Param("dig"), RBP) + MOVQ(Mem{Base: BP}.Offset(0*8), R8) // a = H0 + MOVQ(Mem{Base: BP}.Offset(1*8), R9) // b = H1 + MOVQ(Mem{Base: BP}.Offset(2*8), R10) // c = H2 + MOVQ(Mem{Base: BP}.Offset(3*8), R11) // d = H3 + MOVQ(Mem{Base: BP}.Offset(4*8), R12) // e = H4 + MOVQ(Mem{Base: BP}.Offset(5*8), R13) // f = H5 + MOVQ(Mem{Base: BP}.Offset(6*8), R14) // g = H6 + MOVQ(Mem{Base: BP}.Offset(7*8), R15) // h = H7 + PSHUFFLE_BYTE_FLIP_MASK_DATA() + loop() + end() +} + +func rotateRight(slice *[]GPPhysical) []GPPhysical { + n := len(*slice) + new := make([]GPPhysical, n) + for i, reg := range *slice { + new[(i+1)%n] = reg + } + return new +} + +// Line 167 +func loop() { + Label("loop") + MOVQ(RSP, RBP) // message schedule + + n := len(_K) + regs := []GPPhysical{R8, R9, R10, R11, R12, R13, R14, R15} + + for i := 0; i < 16; i++ { + SHA512ROUND0(i, _K[i], regs[0], regs[1], regs[2], regs[3], regs[4], regs[5], regs[6], regs[7]) + regs = rotateRight(®s) + } + + for i := 16; i < n; i++ { + SHA512ROUND1(i, _K[i], regs[0], regs[1], regs[2], regs[3], regs[4], regs[5], regs[6], regs[7]) + regs = rotateRight(®s) + } + + Load(Param("dig"), RBP) + + BP_Mem := Mem{Base: BP} + ADDQ(BP_Mem.Offset(0*8), R8) // H0 = a + H0 + MOVQ(R8, BP_Mem.Offset(0*8)) + ADDQ(BP_Mem.Offset(1*8), R9) // H1 = b + H1 + MOVQ(R9, BP_Mem.Offset(1*8)) + ADDQ(BP_Mem.Offset(2*8), R10) // H2 = c + H2 + MOVQ(R10, BP_Mem.Offset(2*8)) + ADDQ(BP_Mem.Offset(3*8), R11) // H3 = d + H3 + MOVQ(R11, BP_Mem.Offset(3*8)) + ADDQ(BP_Mem.Offset(4*8), R12) // H4 = e + H4 + MOVQ(R12, BP_Mem.Offset(4*8)) + ADDQ(BP_Mem.Offset(5*8), R13) // H5 = f + H5 + MOVQ(R13, BP_Mem.Offset(5*8)) + ADDQ(BP_Mem.Offset(6*8), R14) // H6 = g + H6 + MOVQ(R14, BP_Mem.Offset(6*8)) + ADDQ(BP_Mem.Offset(7*8), R15) // H7 = h + H7 + MOVQ(R15, BP_Mem.Offset(7*8)) + + ADDQ(Imm(128), RSI) + CMPQ(RSI, Mem{Base: SP}.Offset(640)) + JB(LabelRef("loop")) +} + +// Line 274 +func end() { + Label("end") + RET() +} + +// Version below is based on "Fast SHA512 Implementations on Intel +// Architecture Processors" White-paper +// https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/fast-sha512-implementations-ia-processors-paper.pdf +// AVX2 version by Intel, same algorithm in Linux kernel: +// https://github.com/torvalds/linux/blob/master/arch/x86/crypto/sha512-avx2-asm.S + +// James Guilford +// Kirk Yap +// Tim Chen +// David Cote +// Aleksey Sidorov + +// Line 289 +var ( + YFER_SIZE int = (4 * 8) + SRND_SIZE = (1 * 8) + INP_SIZE = (1 * 8) + + frame_YFER = (0) + frame_SRND = (frame_YFER + YFER_SIZE) + frame_INP = (frame_SRND + SRND_SIZE) + frame_INPEND = (frame_INP + INP_SIZE) +) + +// Line 298 +func addm(p1 Mem, p2 GPPhysical) { + ADDQ(p1, p2) + MOVQ(p2, p1) +} + +// Line 302 +func COPY_YMM_AND_BSWAP(p1 VecPhysical, p2 Mem, p3 VecPhysical) { + VMOVDQU(p2, p1) + VPSHUFB(p3, p1, p1) +} + +// Line 306 +func MY_VPALIGNR(YDST, YSRC1, YSRC2 VecPhysical, RVAL int) { + VPERM2F128(U8(0x3), YSRC2, YSRC1, YDST) + VPALIGNR(U8(RVAL), YSRC2, YDST, YDST) +} + +// Line 324 +func blockAVX2() { + Implement("blockAVX2") + Attributes(NOSPLIT) + AllocLocal(56) + + Load(Param("dig"), RSI) + Load(Param("p").Base(), RDI) + Load(Param("p").Len(), RDX) + + SHRQ(Imm(7), RDX) + SHLQ(Imm(7), RDX) + + JZ(LabelRef("done_hash")) + ADDQ(RDI, RDX) + MOVQ(RDX, Mem{Base: SP}.Offset(frame_INPEND)) + + MOVQ(Mem{Base: SI}.Offset(0*8), RAX) + MOVQ(Mem{Base: SI}.Offset(1*8), RBX) + MOVQ(Mem{Base: SI}.Offset(2*8), RCX) + MOVQ(Mem{Base: SI}.Offset(3*8), R8) + MOVQ(Mem{Base: SI}.Offset(4*8), RDX) + MOVQ(Mem{Base: SI}.Offset(5*8), R9) + MOVQ(Mem{Base: SI}.Offset(6*8), R10) + MOVQ(Mem{Base: SI}.Offset(7*8), R11) + + PSHUFFLE_BYTE_FLIP_MASK := PSHUFFLE_BYTE_FLIP_MASK_DATA() + VMOVDQU(PSHUFFLE_BYTE_FLIP_MASK, Y9) + + loop0() + loop1() + loop2() + done_hash() +} + +// Line 347 +func loop0() { + Label("loop0") + + _K := NewDataAddr(Symbol{Name: ThatPeskyUnicodeDot + "_K"}, 0) + MOVQ(_K, RBP) + + // byte swap first 16 dwords + COPY_YMM_AND_BSWAP(Y4, Mem{Base: DI}.Offset(0*32), Y9) + COPY_YMM_AND_BSWAP(Y5, Mem{Base: DI}.Offset(1*32), Y9) + COPY_YMM_AND_BSWAP(Y6, Mem{Base: DI}.Offset(2*32), Y9) + COPY_YMM_AND_BSWAP(Y7, Mem{Base: DI}.Offset(3*32), Y9) + + MOVQ(RDI, Mem{Base: SP}.Offset(frame_INP)) + + // schedule 64 input dwords, by doing 12 rounds of 4 each + MOVQ(U32(4), Mem{Base: SP}.Offset(frame_SRND)) +} + +// Line 361 +func loop1() { + Label("loop1") + VPADDQ(Mem{Base: BP}, Y4, Y0) + VMOVDQU(Y0, Mem{Base: SP}.Offset(frame_YFER)) + + MY_VPALIGNR(Y0, Y7, Y6, 8) + + VPADDQ(Y4, Y0, Y0) + + MY_VPALIGNR(Y1, Y5, Y4, 8) + + VPSRLQ(Imm(1), Y1, Y2) + VPSLLQ(Imm(64-1), Y1, Y3) + VPOR(Y2, Y3, Y3) + + VPSRLQ(Imm(7), Y1, Y8) + + MOVQ(RAX, RDI) + RORXQ(Imm(41), RDX, R13) + RORXQ(Imm(18), RDX, R14) + ADDQ(Mem{Base: SP}.Offset(frame_YFER), R11) + ORQ(RCX, RDI) + MOVQ(R9, R15) + RORXQ(Imm(34), RAX, R12) + + XORQ(R14, R13) + XORQ(R10, R15) + RORXQ(Imm(14), RDX, R14) + + ANDQ(RDX, R15) + XORQ(R14, R13) + RORXQ(Imm(39), RAX, R14) + ADDQ(R11, R8) + + ANDQ(RBX, RDI) + XORQ(R12, R14) + RORXQ(Imm(28), RAX, R12) + + XORQ(R10, R15) + XORQ(R12, R14) + MOVQ(RAX, R12) + ANDQ(RCX, R12) + + ADDQ(R13, R15) + ORQ(R12, RDI) + ADDQ(R14, R11) + + ADDQ(R15, R8) + + ADDQ(R15, R11) + ADDQ(RDI, R11) + + VPSRLQ(Imm(8), Y1, Y2) + VPSLLQ(Imm(64-8), Y1, Y1) + VPOR(Y2, Y1, Y1) + + VPXOR(Y8, Y3, Y3) + VPXOR(Y1, Y3, Y1) + + VPADDQ(Y1, Y0, Y0) + + VPERM2F128(Imm(0x0), Y0, Y0, Y4) + + MASK_YMM_LO := MASK_YMM_LO_DATA() + VPAND(MASK_YMM_LO, Y0, Y0) + + VPERM2F128(Imm(0x11), Y7, Y7, Y2) + VPSRLQ(Imm(6), Y2, Y8) + + MOVQ(R11, RDI) + RORXQ(Imm(41), R8, R13) + RORXQ(Imm(18), R8, R14) + ADDQ(Mem{Base: SP}.Offset(1*8+frame_YFER), R10) + ORQ(RBX, RDI) + + MOVQ(RDX, R15) + RORXQ(Imm(34), R11, R12) + XORQ(R14, R13) + XORQ(R9, R15) + + RORXQ(Imm(14), R8, R14) + XORQ(R14, R13) + RORXQ(Imm(39), R11, R14) + ANDQ(R8, R15) + ADDQ(R10, RCX) + + ANDQ(RAX, RDI) + XORQ(R12, R14) + + RORXQ(Imm(28), R11, R12) + XORQ(R9, R15) + + XORQ(R12, R14) + MOVQ(R11, R12) + ANDQ(RBX, R12) + ADDQ(R13, R15) + + ORQ(R12, RDI) + ADDQ(R14, R10) + + ADDQ(R15, RCX) + ADDQ(R15, R10) + ADDQ(RDI, R10) + + VPSRLQ(Imm(19), Y2, Y3) + VPSLLQ(Imm(64-19), Y2, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + VPSRLQ(Imm(61), Y2, Y3) + VPSLLQ(Imm(64-61), Y2, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + + VPADDQ(Y8, Y4, Y4) + + VPSRLQ(Imm(6), Y4, Y8) + + MOVQ(R10, RDI) + RORXQ(Imm(41), RCX, R13) + ADDQ(Mem{Base: SP}.Offset(2*8+frame_YFER), R9) + + RORXQ(Imm(18), RCX, R14) + ORQ(RAX, RDI) + MOVQ(R8, R15) + XORQ(RDX, R15) + + RORXQ(Imm(34), R10, R12) + XORQ(R14, R13) + ANDQ(RCX, R15) + + RORXQ(Imm(14), RCX, R14) + ADDQ(R9, RBX) + ANDQ(R11, RDI) + + XORQ(R14, R13) + RORXQ(Imm(39), R10, R14) + XORQ(RDX, R15) + + XORQ(R12, R14) + RORXQ(Imm(28), R10, R12) + + XORQ(R12, R14) + MOVQ(R10, R12) + ANDQ(RAX, R12) + ADDQ(R13, R15) + + ORQ(R12, RDI) + ADDQ(R14, R9) + ADDQ(R15, RBX) + ADDQ(R15, R9) + + ADDQ(RDI, R9) + + VPSRLQ(Imm(19), Y4, Y3) + VPSLLQ(Imm(64-19), Y4, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + VPSRLQ(Imm(61), Y4, Y3) + VPSLLQ(Imm(64-61), Y4, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + + VPADDQ(Y8, Y0, Y2) + + VPBLENDD(Imm(0xF0), Y2, Y4, Y4) + + MOVQ(R9, RDI) + RORXQ(Imm(41), RBX, R13) + RORXQ(Imm(18), RBX, R14) + ADDQ(Mem{Base: SP}.Offset(3*8+frame_YFER), RDX) + ORQ(R11, RDI) + + MOVQ(RCX, R15) + RORXQ(Imm(34), R9, R12) + XORQ(R14, R13) + XORQ(R8, R15) + + RORXQ(Imm(14), RBX, R14) + ANDQ(RBX, R15) + ADDQ(RDX, RAX) + ANDQ(R10, RDI) + + XORQ(R14, R13) + XORQ(R8, R15) + + RORXQ(Imm(39), R9, R14) + ADDQ(R13, R15) + + XORQ(R12, R14) + ADDQ(R15, RAX) + + RORXQ(Imm(28), R9, R12) + + XORQ(R12, R14) + MOVQ(R9, R12) + ANDQ(R11, R12) + ORQ(R12, RDI) + + ADDQ(R14, RDX) + ADDQ(R15, RDX) + ADDQ(RDI, RDX) + + VPADDQ(Mem{Base: BP}.Offset(1*32), Y5, Y0) + VMOVDQU(Y0, Mem{Base: SP}.Offset(frame_YFER)) + + MY_VPALIGNR(Y0, Y4, Y7, 8) + + VPADDQ(Y5, Y0, Y0) + + MY_VPALIGNR(Y1, Y6, Y5, 8) + + VPSRLQ(Imm(1), Y1, Y2) + VPSLLQ(Imm(64-1), Y1, Y3) + VPOR(Y2, Y3, Y3) + + VPSRLQ(Imm(7), Y1, Y8) + + MOVQ(RDX, RDI) + RORXQ(Imm(41), RAX, R13) + RORXQ(Imm(18), RAX, R14) + ADDQ(Mem{Base: SP}.Offset(frame_YFER), R8) + ORQ(R10, RDI) + MOVQ(RBX, R15) + RORXQ(Imm(34), RDX, R12) + + XORQ(R14, R13) + XORQ(RCX, R15) + RORXQ(Imm(14), RAX, R14) + + ANDQ(RAX, R15) + XORQ(R14, R13) + RORXQ(Imm(39), RDX, R14) + ADDQ(R8, R11) + + ANDQ(R9, RDI) + XORQ(R12, R14) + RORXQ(Imm(28), RDX, R12) + + XORQ(RCX, R15) + XORQ(R12, R14) + MOVQ(RDX, R12) + ANDQ(R10, R12) + + ADDQ(R13, R15) + ORQ(R12, RDI) + ADDQ(R14, R8) + + ADDQ(R15, R11) + + ADDQ(R15, R8) + ADDQ(RDI, R8) + + VPSRLQ(Imm(8), Y1, Y2) + VPSLLQ(Imm(64-8), Y1, Y1) + VPOR(Y2, Y1, Y1) + + VPXOR(Y8, Y3, Y3) + VPXOR(Y1, Y3, Y1) + + VPADDQ(Y1, Y0, Y0) + + VPERM2F128(Imm(0x0), Y0, Y0, Y5) + + VPAND(MASK_YMM_LO, Y0, Y0) + + VPERM2F128(Imm(0x11), Y4, Y4, Y2) + VPSRLQ(Imm(6), Y2, Y8) + + MOVQ(R8, RDI) + RORXQ(Imm(41), R11, R13) + RORXQ(Imm(18), R11, R14) + ADDQ(Mem{Base: SP}.Offset(1*8+frame_YFER), RCX) + ORQ(R9, RDI) + + MOVQ(RAX, R15) + RORXQ(Imm(34), R8, R12) + XORQ(R14, R13) + XORQ(RBX, R15) + + RORXQ(Imm(14), R11, R14) + XORQ(R14, R13) + RORXQ(Imm(39), R8, R14) + ANDQ(R11, R15) + ADDQ(RCX, R10) + + ANDQ(RDX, RDI) + XORQ(R12, R14) + + RORXQ(Imm(28), R8, R12) + XORQ(RBX, R15) + + XORQ(R12, R14) + MOVQ(R8, R12) + ANDQ(R9, R12) + ADDQ(R13, R15) + + ORQ(R12, RDI) + ADDQ(R14, RCX) + + ADDQ(R15, R10) + ADDQ(R15, RCX) + ADDQ(RDI, RCX) + + VPSRLQ(Imm(19), Y2, Y3) + VPSLLQ(Imm(64-19), Y2, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + VPSRLQ(Imm(61), Y2, Y3) + VPSLLQ(Imm(64-61), Y2, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + + VPADDQ(Y8, Y5, Y5) + + VPSRLQ(Imm(6), Y5, Y8) + + MOVQ(RCX, RDI) + RORXQ(Imm(41), R10, R13) + ADDQ(Mem{Base: SP}.Offset(2*8+frame_YFER), RBX) + + RORXQ(Imm(18), R10, R14) + ORQ(RDX, RDI) + MOVQ(R11, R15) + XORQ(RAX, R15) + + RORXQ(Imm(34), RCX, R12) + XORQ(R14, R13) + ANDQ(R10, R15) + + RORXQ(Imm(14), R10, R14) + ADDQ(RBX, R9) + ANDQ(R8, RDI) + + XORQ(R14, R13) + RORXQ(Imm(39), RCX, R14) + XORQ(RAX, R15) + + XORQ(R12, R14) + RORXQ(Imm(28), RCX, R12) + + XORQ(R12, R14) + MOVQ(RCX, R12) + ANDQ(RDX, R12) + ADDQ(R13, R15) + + ORQ(R12, RDI) + ADDQ(R14, RBX) + ADDQ(R15, R9) + ADDQ(R15, RBX) + + ADDQ(RDI, RBX) + + VPSRLQ(Imm(19), Y5, Y3) + VPSLLQ(Imm(64-19), Y5, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + VPSRLQ(Imm(61), Y5, Y3) + VPSLLQ(Imm(64-61), Y5, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + + VPADDQ(Y8, Y0, Y2) + + VPBLENDD(Imm(0xF0), Y2, Y5, Y5) + + MOVQ(RBX, RDI) + RORXQ(Imm(41), R9, R13) + RORXQ(Imm(18), R9, R14) + ADDQ(Mem{Base: SP}.Offset(3*8+frame_YFER), RAX) + ORQ(R8, RDI) + + MOVQ(R10, R15) + RORXQ(Imm(34), RBX, R12) + XORQ(R14, R13) + XORQ(R11, R15) + + RORXQ(Imm(14), R9, R14) + ANDQ(R9, R15) + ADDQ(RAX, RDX) + ANDQ(RCX, RDI) + + XORQ(R14, R13) + XORQ(R11, R15) + + RORXQ(Imm(39), RBX, R14) + ADDQ(R13, R15) + + XORQ(R12, R14) + ADDQ(R15, RDX) + + RORXQ(Imm(28), RBX, R12) + + XORQ(R12, R14) + MOVQ(RBX, R12) + ANDQ(R8, R12) + ORQ(R12, RDI) + + ADDQ(R14, RAX) + ADDQ(R15, RAX) + ADDQ(RDI, RAX) + + VPADDQ(Mem{Base: BP}.Offset(2*32), Y6, Y0) + VMOVDQU(Y0, Mem{Base: SP}.Offset(frame_YFER)) + + MY_VPALIGNR(Y0, Y5, Y4, 8) + + VPADDQ(Y6, Y0, Y0) + + MY_VPALIGNR(Y1, Y7, Y6, 8) + + VPSRLQ(Imm(1), Y1, Y2) + VPSLLQ(Imm(64-1), Y1, Y3) + VPOR(Y2, Y3, Y3) + + VPSRLQ(Imm(7), Y1, Y8) + + MOVQ(RAX, RDI) + RORXQ(Imm(41), RDX, R13) + RORXQ(Imm(18), RDX, R14) + ADDQ(Mem{Base: SP}.Offset(frame_YFER), R11) + ORQ(RCX, RDI) + MOVQ(R9, R15) + RORXQ(Imm(34), RAX, R12) + + XORQ(R14, R13) + XORQ(R10, R15) + RORXQ(Imm(14), RDX, R14) + + ANDQ(RDX, R15) + XORQ(R14, R13) + RORXQ(Imm(39), RAX, R14) + ADDQ(R11, R8) + + ANDQ(RBX, RDI) + XORQ(R12, R14) + RORXQ(Imm(28), RAX, R12) + + XORQ(R10, R15) + XORQ(R12, R14) + MOVQ(RAX, R12) + ANDQ(RCX, R12) + + ADDQ(R13, R15) + ORQ(R12, RDI) + ADDQ(R14, R11) + + ADDQ(R15, R8) + + ADDQ(R15, R11) + ADDQ(RDI, R11) + + VPSRLQ(Imm(8), Y1, Y2) + VPSLLQ(Imm(64-8), Y1, Y1) + VPOR(Y2, Y1, Y1) + + VPXOR(Y8, Y3, Y3) + VPXOR(Y1, Y3, Y1) + + VPADDQ(Y1, Y0, Y0) + + VPERM2F128(Imm(0x0), Y0, Y0, Y6) + + VPAND(MASK_YMM_LO, Y0, Y0) + + VPERM2F128(Imm(0x11), Y5, Y5, Y2) + VPSRLQ(Imm(6), Y2, Y8) + + MOVQ(R11, RDI) + RORXQ(Imm(41), R8, R13) + RORXQ(Imm(18), R8, R14) + ADDQ(Mem{Base: SP}.Offset(1*8+frame_YFER), R10) + ORQ(RBX, RDI) + + MOVQ(RDX, R15) + RORXQ(Imm(34), R11, R12) + XORQ(R14, R13) + XORQ(R9, R15) + + RORXQ(Imm(14), R8, R14) + XORQ(R14, R13) + RORXQ(Imm(39), R11, R14) + ANDQ(R8, R15) + ADDQ(R10, RCX) + + ANDQ(RAX, RDI) + XORQ(R12, R14) + + RORXQ(Imm(28), R11, R12) + XORQ(R9, R15) + + XORQ(R12, R14) + MOVQ(R11, R12) + ANDQ(RBX, R12) + ADDQ(R13, R15) + + ORQ(R12, RDI) + ADDQ(R14, R10) + + ADDQ(R15, RCX) + ADDQ(R15, R10) + ADDQ(RDI, R10) + + VPSRLQ(Imm(19), Y2, Y3) + VPSLLQ(Imm(64-19), Y2, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + VPSRLQ(Imm(61), Y2, Y3) + VPSLLQ(Imm(64-61), Y2, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + + VPADDQ(Y8, Y6, Y6) + + VPSRLQ(Imm(6), Y6, Y8) + + MOVQ(R10, RDI) + RORXQ(Imm(41), RCX, R13) + ADDQ(Mem{Base: SP}.Offset(2*8+frame_YFER), R9) + + RORXQ(Imm(18), RCX, R14) + ORQ(RAX, RDI) + MOVQ(R8, R15) + XORQ(RDX, R15) + + RORXQ(Imm(34), R10, R12) + XORQ(R14, R13) + ANDQ(RCX, R15) + + RORXQ(Imm(14), RCX, R14) + ADDQ(R9, RBX) + ANDQ(R11, RDI) + + XORQ(R14, R13) + RORXQ(Imm(39), R10, R14) + XORQ(RDX, R15) + + XORQ(R12, R14) + RORXQ(Imm(28), R10, R12) + + XORQ(R12, R14) + MOVQ(R10, R12) + ANDQ(RAX, R12) + ADDQ(R13, R15) + + ORQ(R12, RDI) + ADDQ(R14, R9) + ADDQ(R15, RBX) + ADDQ(R15, R9) + + ADDQ(RDI, R9) + + VPSRLQ(Imm(19), Y6, Y3) + VPSLLQ(Imm(64-19), Y6, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + VPSRLQ(Imm(61), Y6, Y3) + VPSLLQ(Imm(64-61), Y6, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + + VPADDQ(Y8, Y0, Y2) + + VPBLENDD(Imm(0xF0), Y2, Y6, Y6) + + MOVQ(R9, RDI) + RORXQ(Imm(41), RBX, R13) + RORXQ(Imm(18), RBX, R14) + ADDQ(Mem{Base: SP}.Offset(3*8+frame_YFER), RDX) + ORQ(R11, RDI) + + MOVQ(RCX, R15) + RORXQ(Imm(34), R9, R12) + XORQ(R14, R13) + XORQ(R8, R15) + + RORXQ(Imm(14), RBX, R14) + ANDQ(RBX, R15) + ADDQ(RDX, RAX) + ANDQ(R10, RDI) + + XORQ(R14, R13) + XORQ(R8, R15) + + RORXQ(Imm(39), R9, R14) + ADDQ(R13, R15) + + XORQ(R12, R14) + ADDQ(R15, RAX) + + RORXQ(Imm(28), R9, R12) + + XORQ(R12, R14) + MOVQ(R9, R12) + ANDQ(R11, R12) + ORQ(R12, RDI) + + ADDQ(R14, RDX) + ADDQ(R15, RDX) + ADDQ(RDI, RDX) + + VPADDQ(Mem{Base: BP}.Offset(3*32), Y7, Y0) + VMOVDQU(Y0, Mem{Base: SP}.Offset(frame_YFER)) + ADDQ(U8(4*32), RBP) + + MY_VPALIGNR(Y0, Y6, Y5, 8) + + VPADDQ(Y7, Y0, Y0) + + MY_VPALIGNR(Y1, Y4, Y7, 8) + + VPSRLQ(Imm(1), Y1, Y2) + VPSLLQ(Imm(64-1), Y1, Y3) + VPOR(Y2, Y3, Y3) + + VPSRLQ(Imm(7), Y1, Y8) + + MOVQ(RDX, RDI) + RORXQ(Imm(41), RAX, R13) + RORXQ(Imm(18), RAX, R14) + ADDQ(Mem{Base: SP}.Offset(frame_YFER), R8) + ORQ(R10, RDI) + MOVQ(RBX, R15) + RORXQ(Imm(34), RDX, R12) + + XORQ(R14, R13) + XORQ(RCX, R15) + RORXQ(Imm(14), RAX, R14) + + ANDQ(RAX, R15) + XORQ(R14, R13) + RORXQ(Imm(39), RDX, R14) + ADDQ(R8, R11) + + ANDQ(R9, RDI) + XORQ(R12, R14) + RORXQ(Imm(28), RDX, R12) + + XORQ(RCX, R15) + XORQ(R12, R14) + MOVQ(RDX, R12) + ANDQ(R10, R12) + + ADDQ(R13, R15) + ORQ(R12, RDI) + ADDQ(R14, R8) + + ADDQ(R15, R11) + + ADDQ(R15, R8) + ADDQ(RDI, R8) + + VPSRLQ(Imm(8), Y1, Y2) + VPSLLQ(Imm(64-8), Y1, Y1) + VPOR(Y2, Y1, Y1) + + VPXOR(Y8, Y3, Y3) + VPXOR(Y1, Y3, Y1) + + VPADDQ(Y1, Y0, Y0) + + VPERM2F128(Imm(0x0), Y0, Y0, Y7) + + VPAND(MASK_YMM_LO, Y0, Y0) + + VPERM2F128(Imm(0x11), Y6, Y6, Y2) + VPSRLQ(Imm(6), Y2, Y8) + + MOVQ(R8, RDI) + RORXQ(Imm(41), R11, R13) + RORXQ(Imm(18), R11, R14) + ADDQ(Mem{Base: SP}.Offset(1*8+frame_YFER), RCX) + ORQ(R9, RDI) + + MOVQ(RAX, R15) + RORXQ(Imm(34), R8, R12) + XORQ(R14, R13) + XORQ(RBX, R15) + + RORXQ(Imm(14), R11, R14) + XORQ(R14, R13) + RORXQ(Imm(39), R8, R14) + ANDQ(R11, R15) + ADDQ(RCX, R10) + + ANDQ(RDX, RDI) + XORQ(R12, R14) + + RORXQ(Imm(28), R8, R12) + XORQ(RBX, R15) + + XORQ(R12, R14) + MOVQ(R8, R12) + ANDQ(R9, R12) + ADDQ(R13, R15) + + ORQ(R12, RDI) + ADDQ(R14, RCX) + + ADDQ(R15, R10) + ADDQ(R15, RCX) + ADDQ(RDI, RCX) + + VPSRLQ(Imm(19), Y2, Y3) + VPSLLQ(Imm(64-19), Y2, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + VPSRLQ(Imm(61), Y2, Y3) + VPSLLQ(Imm(64-61), Y2, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + + VPADDQ(Y8, Y7, Y7) + + VPSRLQ(Imm(6), Y7, Y8) + + MOVQ(RCX, RDI) + RORXQ(Imm(41), R10, R13) + ADDQ(Mem{Base: SP}.Offset(2*8+frame_YFER), RBX) + + RORXQ(Imm(18), R10, R14) + ORQ(RDX, RDI) + MOVQ(R11, R15) + XORQ(RAX, R15) + + RORXQ(Imm(34), RCX, R12) + XORQ(R14, R13) + ANDQ(R10, R15) + + RORXQ(Imm(14), R10, R14) + ADDQ(RBX, R9) + ANDQ(R8, RDI) + + XORQ(R14, R13) + RORXQ(Imm(39), RCX, R14) + XORQ(RAX, R15) + + XORQ(R12, R14) + RORXQ(Imm(28), RCX, R12) + + XORQ(R12, R14) + MOVQ(RCX, R12) + ANDQ(RDX, R12) + ADDQ(R13, R15) + + ORQ(R12, RDI) + ADDQ(R14, RBX) + ADDQ(R15, R9) + ADDQ(R15, RBX) + + ADDQ(RDI, RBX) + + VPSRLQ(Imm(19), Y7, Y3) + VPSLLQ(Imm(64-19), Y7, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + VPSRLQ(Imm(61), Y7, Y3) + VPSLLQ(Imm(64-61), Y7, Y1) + VPOR(Y1, Y3, Y3) + VPXOR(Y3, Y8, Y8) + + VPADDQ(Y8, Y0, Y2) + + VPBLENDD(Imm(0xF0), Y2, Y7, Y7) + + MOVQ(RBX, RDI) + RORXQ(Imm(41), R9, R13) + RORXQ(Imm(18), R9, R14) + ADDQ(Mem{Base: SP}.Offset(3*8+frame_YFER), RAX) + ORQ(R8, RDI) + + MOVQ(R10, R15) + RORXQ(Imm(34), RBX, R12) + XORQ(R14, R13) + XORQ(R11, R15) + + RORXQ(Imm(14), R9, R14) + ANDQ(R9, R15) + ADDQ(RAX, RDX) + ANDQ(RCX, RDI) + + XORQ(R14, R13) + XORQ(R11, R15) + + RORXQ(Imm(39), RBX, R14) + ADDQ(R13, R15) + + XORQ(R12, R14) + ADDQ(R15, RDX) + + RORXQ(Imm(28), RBX, R12) + + XORQ(R12, R14) + MOVQ(RBX, R12) + ANDQ(R8, R12) + ORQ(R12, RDI) + + ADDQ(R14, RAX) + ADDQ(R15, RAX) + ADDQ(RDI, RAX) + + SUBQ(Imm(1), Mem{Base: SP}.Offset(frame_SRND)) + JNE(LabelRef("loop1")) + + MOVQ(U32(2), Mem{Base: SP}.Offset(frame_SRND)) +} + +// Line 1164 +func loop2() { + Label("loop2") + VPADDQ(Mem{Base: BP}, Y4, Y0) + VMOVDQU(Y0, Mem{Base: SP}.Offset(frame_YFER)) + + MOVQ(R9, R15) + RORXQ(Imm(41), RDX, R13) + RORXQ(Imm(18), RDX, R14) + XORQ(R10, R15) + + XORQ(R14, R13) + RORXQ(Imm(14), RDX, R14) + ANDQ(RDX, R15) + + XORQ(R14, R13) + RORXQ(Imm(34), RAX, R12) + XORQ(R10, R15) + RORXQ(Imm(39), RAX, R14) + MOVQ(RAX, RDI) + + XORQ(R12, R14) + RORXQ(Imm(28), RAX, R12) + ADDQ(Mem{Base: SP}.Offset(frame_YFER), R11) + ORQ(RCX, RDI) + + XORQ(R12, R14) + MOVQ(RAX, R12) + ANDQ(RBX, RDI) + ANDQ(RCX, R12) + ADDQ(R13, R15) + + ADDQ(R11, R8) + ORQ(R12, RDI) + ADDQ(R14, R11) + + ADDQ(R15, R8) + + ADDQ(R15, R11) + MOVQ(RDX, R15) + RORXQ(Imm(41), R8, R13) + RORXQ(Imm(18), R8, R14) + XORQ(R9, R15) + + XORQ(R14, R13) + RORXQ(Imm(14), R8, R14) + ANDQ(R8, R15) + ADDQ(RDI, R11) + + XORQ(R14, R13) + RORXQ(Imm(34), R11, R12) + XORQ(R9, R15) + RORXQ(Imm(39), R11, R14) + MOVQ(R11, RDI) + + XORQ(R12, R14) + RORXQ(Imm(28), R11, R12) + ADDQ(Mem{Base: SP}.Offset(8*1+frame_YFER), R10) + ORQ(RBX, RDI) + + XORQ(R12, R14) + MOVQ(R11, R12) + ANDQ(RAX, RDI) + ANDQ(RBX, R12) + ADDQ(R13, R15) + + ADDQ(R10, RCX) + ORQ(R12, RDI) + ADDQ(R14, R10) + + ADDQ(R15, RCX) + + ADDQ(R15, R10) + MOVQ(R8, R15) + RORXQ(Imm(41), RCX, R13) + RORXQ(Imm(18), RCX, R14) + XORQ(RDX, R15) + + XORQ(R14, R13) + RORXQ(Imm(14), RCX, R14) + ANDQ(RCX, R15) + ADDQ(RDI, R10) + + XORQ(R14, R13) + RORXQ(Imm(34), R10, R12) + XORQ(RDX, R15) + RORXQ(Imm(39), R10, R14) + MOVQ(R10, RDI) + + XORQ(R12, R14) + RORXQ(Imm(28), R10, R12) + ADDQ(Mem{Base: SP}.Offset(8*2+frame_YFER), R9) + ORQ(RAX, RDI) + + XORQ(R12, R14) + MOVQ(R10, R12) + ANDQ(R11, RDI) + ANDQ(RAX, R12) + ADDQ(R13, R15) + + ADDQ(R9, RBX) + ORQ(R12, RDI) + ADDQ(R14, R9) + + ADDQ(R15, RBX) + + ADDQ(R15, R9) + MOVQ(RCX, R15) + RORXQ(Imm(41), RBX, R13) + RORXQ(Imm(18), RBX, R14) + XORQ(R8, R15) + + XORQ(R14, R13) + RORXQ(Imm(14), RBX, R14) + ANDQ(RBX, R15) + ADDQ(RDI, R9) + + XORQ(R14, R13) + RORXQ(Imm(34), R9, R12) + XORQ(R8, R15) + RORXQ(Imm(39), R9, R14) + MOVQ(R9, RDI) + + XORQ(R12, R14) + RORXQ(Imm(28), R9, R12) + ADDQ(Mem{Base: SP}.Offset(8*3+frame_YFER), RDX) + ORQ(R11, RDI) + + XORQ(R12, R14) + MOVQ(R9, R12) + ANDQ(R10, RDI) + ANDQ(R11, R12) + ADDQ(R13, R15) + + ADDQ(RDX, RAX) + ORQ(R12, RDI) + ADDQ(R14, RDX) + + ADDQ(R15, RAX) + + ADDQ(R15, RDX) + + ADDQ(RDI, RDX) + + VPADDQ(Mem{Base: BP}.Offset(1*32), Y5, Y0) + VMOVDQU(Y0, Mem{Base: SP}.Offset(frame_YFER)) + ADDQ(U8(2*32), RBP) + + MOVQ(RBX, R15) + RORXQ(Imm(41), RAX, R13) + RORXQ(Imm(18), RAX, R14) + XORQ(RCX, R15) + + XORQ(R14, R13) + RORXQ(Imm(14), RAX, R14) + ANDQ(RAX, R15) + + XORQ(R14, R13) + RORXQ(Imm(34), RDX, R12) + XORQ(RCX, R15) + RORXQ(Imm(39), RDX, R14) + MOVQ(RDX, RDI) + + XORQ(R12, R14) + RORXQ(Imm(28), RDX, R12) + ADDQ(Mem{Base: SP}.Offset(frame_YFER), R8) + ORQ(R10, RDI) + + XORQ(R12, R14) + MOVQ(RDX, R12) + ANDQ(R9, RDI) + ANDQ(R10, R12) + ADDQ(R13, R15) + + ADDQ(R8, R11) + ORQ(R12, RDI) + ADDQ(R14, R8) + + ADDQ(R15, R11) + + ADDQ(R15, R8) + MOVQ(RAX, R15) + RORXQ(Imm(41), R11, R13) + RORXQ(Imm(18), R11, R14) + XORQ(RBX, R15) + + XORQ(R14, R13) + RORXQ(Imm(14), R11, R14) + ANDQ(R11, R15) + ADDQ(RDI, R8) + + XORQ(R14, R13) + RORXQ(Imm(34), R8, R12) + XORQ(RBX, R15) + RORXQ(Imm(39), R8, R14) + MOVQ(R8, RDI) + + XORQ(R12, R14) + RORXQ(Imm(28), R8, R12) + ADDQ(Mem{Base: SP}.Offset(8*1+frame_YFER), RCX) + ORQ(R9, RDI) + + XORQ(R12, R14) + MOVQ(R8, R12) + ANDQ(RDX, RDI) + ANDQ(R9, R12) + ADDQ(R13, R15) + + ADDQ(RCX, R10) + ORQ(R12, RDI) + ADDQ(R14, RCX) + + ADDQ(R15, R10) + + ADDQ(R15, RCX) + MOVQ(R11, R15) + RORXQ(Imm(41), R10, R13) + RORXQ(Imm(18), R10, R14) + XORQ(RAX, R15) + + XORQ(R14, R13) + RORXQ(Imm(14), R10, R14) + ANDQ(R10, R15) + ADDQ(RDI, RCX) + + XORQ(R14, R13) + RORXQ(Imm(34), RCX, R12) + XORQ(RAX, R15) + RORXQ(Imm(39), RCX, R14) + MOVQ(RCX, RDI) + + XORQ(R12, R14) + RORXQ(Imm(28), RCX, R12) + ADDQ(Mem{Base: SP}.Offset(8*2+frame_YFER), RBX) + ORQ(RDX, RDI) + + XORQ(R12, R14) + MOVQ(RCX, R12) + ANDQ(R8, RDI) + ANDQ(RDX, R12) + ADDQ(R13, R15) + + ADDQ(RBX, R9) + ORQ(R12, RDI) + ADDQ(R14, RBX) + + ADDQ(R15, R9) + + ADDQ(R15, RBX) + MOVQ(R10, R15) + RORXQ(Imm(41), R9, R13) + RORXQ(Imm(18), R9, R14) + XORQ(R11, R15) + + XORQ(R14, R13) + RORXQ(Imm(14), R9, R14) + ANDQ(R9, R15) + ADDQ(RDI, RBX) + + XORQ(R14, R13) + RORXQ(Imm(34), RBX, R12) + XORQ(R11, R15) + RORXQ(Imm(39), RBX, R14) + MOVQ(RBX, RDI) + + XORQ(R12, R14) + RORXQ(Imm(28), RBX, R12) + ADDQ(Mem{Base: SP}.Offset(8*3+frame_YFER), RAX) + ORQ(R8, RDI) + + XORQ(R12, R14) + MOVQ(RBX, R12) + ANDQ(RCX, RDI) + ANDQ(R8, R12) + ADDQ(R13, R15) + + ADDQ(RAX, RDX) + ORQ(R12, RDI) + ADDQ(R14, RAX) + + ADDQ(R15, RDX) + + ADDQ(R15, RAX) + + ADDQ(RDI, RAX) + + VMOVDQU(Y6, Y4) + VMOVDQU(Y7, Y5) + + SUBQ(Imm(1), Mem{Base: SP}.Offset(frame_SRND)) + JNE(LabelRef("loop2")) + + addm(Mem{Base: SI}.Offset(8*0), RAX) + addm(Mem{Base: SI}.Offset(8*1), RBX) + addm(Mem{Base: SI}.Offset(8*2), RCX) + addm(Mem{Base: SI}.Offset(8*3), R8) + addm(Mem{Base: SI}.Offset(8*4), RDX) + addm(Mem{Base: SI}.Offset(8*5), R9) + addm(Mem{Base: SI}.Offset(8*6), R10) + addm(Mem{Base: SI}.Offset(8*7), R11) + + MOVQ(Mem{Base: SP}.Offset(frame_INP), RDI) + ADDQ(Imm(128), RDI) + CMPQ(RDI, Mem{Base: SP}.Offset(frame_INPEND)) + JNE(LabelRef("loop0")) +} + +// Line 1468 +func done_hash() { + Label("done_hash") + VZEROUPPER() + RET() +} + +// ##~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~DATA SECTION~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~## + +// Pointers for memoizing Data section symbols +var PSHUFFLE_BYTE_FLIP_MASK_DATA_ptr, MASK_YMM_LO_ptr *Mem + +// Line 310 +func PSHUFFLE_BYTE_FLIP_MASK_DATA() Mem { + if PSHUFFLE_BYTE_FLIP_MASK_DATA_ptr != nil { + return *PSHUFFLE_BYTE_FLIP_MASK_DATA_ptr + } + + PSHUFFLE_BYTE_FLIP_MASK_DATA := GLOBL("PSHUFFLE_BYTE_FLIP_MASK", NOPTR|RODATA) + PSHUFFLE_BYTE_FLIP_MASK_DATA_ptr = &PSHUFFLE_BYTE_FLIP_MASK_DATA + DATA(0x00, U64(0x0001020304050607)) + DATA(0x08, U64(0x08090a0b0c0d0e0f)) + DATA(0x10, U64(0x1011121314151617)) + DATA(0x18, U64(0x18191a1b1c1d1e1f)) + return PSHUFFLE_BYTE_FLIP_MASK_DATA +} + +// Line 317 +func MASK_YMM_LO_DATA() Mem { + if MASK_YMM_LO_ptr != nil { + return *MASK_YMM_LO_ptr + } + + MASK_YMM_LO := GLOBL("MASK_YMM_LO", NOPTR|RODATA) + MASK_YMM_LO_ptr = &MASK_YMM_LO + DATA(0x00, U64(0x0000000000000000)) + DATA(0x08, U64(0x0000000000000000)) + DATA(0x10, U64(0xFFFFFFFFFFFFFFFF)) + DATA(0x18, U64(0xFFFFFFFFFFFFFFFF)) + return MASK_YMM_LO +} diff --git a/src/crypto/sha512/sha512block_amd64.s b/src/crypto/sha512/sha512block_amd64.s index bd4e3e6669f..fdcef222150 100644 --- a/src/crypto/sha512/sha512block_amd64.s +++ b/src/crypto/sha512/sha512block_amd64.s @@ -1,1470 +1,5398 @@ -// Copyright 2013 The Go Authors. All rights reserved. -// Use of this source code is governed by a BSD-style -// license that can be found in the LICENSE file. +// Code generated by command: go run sha512block_amd64_asm.go -out ../sha512block_amd64.s -pkg sha512. DO NOT EDIT. //go:build !purego #include "textflag.h" -// SHA512 block routine. See sha512block.go for Go equivalent. -// -// The algorithm is detailed in FIPS 180-4: -// -// https://csrc.nist.gov/publications/fips/fips180-4/fips-180-4.pdf -// -// Wt = Mt; for 0 <= t <= 15 -// Wt = SIGMA1(Wt-2) + SIGMA0(Wt-15) + Wt-16; for 16 <= t <= 79 -// -// a = H0 -// b = H1 -// c = H2 -// d = H3 -// e = H4 -// f = H5 -// g = H6 -// h = H7 -// -// for t = 0 to 79 { -// T1 = h + BIGSIGMA1(e) + Ch(e,f,g) + Kt + Wt -// T2 = BIGSIGMA0(a) + Maj(a,b,c) -// h = g -// g = f -// f = e -// e = d + T1 -// d = c -// c = b -// b = a -// a = T1 + T2 -// } -// -// H0 = a + H0 -// H1 = b + H1 -// H2 = c + H2 -// H3 = d + H3 -// H4 = e + H4 -// H5 = f + H5 -// H6 = g + H6 -// H7 = h + H7 - -// Wt = Mt; for 0 <= t <= 15 -#define MSGSCHEDULE0(index) \ - MOVQ (index*8)(SI), AX; \ - BSWAPQ AX; \ - MOVQ AX, (index*8)(BP) - -// Wt = SIGMA1(Wt-2) + Wt-7 + SIGMA0(Wt-15) + Wt-16; for 16 <= t <= 79 -// SIGMA0(x) = ROTR(1,x) XOR ROTR(8,x) XOR SHR(7,x) -// SIGMA1(x) = ROTR(19,x) XOR ROTR(61,x) XOR SHR(6,x) -#define MSGSCHEDULE1(index) \ - MOVQ ((index-2)*8)(BP), AX; \ - MOVQ AX, CX; \ - RORQ $19, AX; \ - MOVQ CX, DX; \ - RORQ $61, CX; \ - SHRQ $6, DX; \ - MOVQ ((index-15)*8)(BP), BX; \ - XORQ CX, AX; \ - MOVQ BX, CX; \ - XORQ DX, AX; \ - RORQ $1, BX; \ - MOVQ CX, DX; \ - SHRQ $7, DX; \ - RORQ $8, CX; \ - ADDQ ((index-7)*8)(BP), AX; \ - XORQ CX, BX; \ - XORQ DX, BX; \ - ADDQ ((index-16)*8)(BP), BX; \ - ADDQ BX, AX; \ - MOVQ AX, ((index)*8)(BP) - -// Calculate T1 in AX - uses AX, CX and DX registers. -// h is also used as an accumulator. Wt is passed in AX. -// T1 = h + BIGSIGMA1(e) + Ch(e, f, g) + Kt + Wt -// BIGSIGMA1(x) = ROTR(14,x) XOR ROTR(18,x) XOR ROTR(41,x) -// Ch(x, y, z) = (x AND y) XOR (NOT x AND z) -#define SHA512T1(const, e, f, g, h) \ - MOVQ $const, DX; \ - ADDQ AX, h; \ - MOVQ e, AX; \ - ADDQ DX, h; \ - MOVQ e, CX; \ - RORQ $14, AX; \ - MOVQ e, DX; \ - RORQ $18, CX; \ - XORQ CX, AX; \ - MOVQ e, CX; \ - RORQ $41, DX; \ - ANDQ f, CX; \ - XORQ AX, DX; \ - MOVQ e, AX; \ - NOTQ AX; \ - ADDQ DX, h; \ - ANDQ g, AX; \ - XORQ CX, AX; \ - ADDQ h, AX - -// Calculate T2 in BX - uses BX, CX, DX and DI registers. -// T2 = BIGSIGMA0(a) + Maj(a, b, c) -// BIGSIGMA0(x) = ROTR(28,x) XOR ROTR(34,x) XOR ROTR(39,x) -// Maj(x, y, z) = (x AND y) XOR (x AND z) XOR (y AND z) -#define SHA512T2(a, b, c) \ - MOVQ a, DI; \ - MOVQ c, BX; \ - RORQ $28, DI; \ - MOVQ a, DX; \ - ANDQ b, BX; \ - RORQ $34, DX; \ - MOVQ a, CX; \ - ANDQ c, CX; \ - XORQ DX, DI; \ - XORQ CX, BX; \ - MOVQ a, DX; \ - MOVQ b, CX; \ - RORQ $39, DX; \ - ANDQ a, CX; \ - XORQ CX, BX; \ - XORQ DX, DI; \ - ADDQ DI, BX - -// Calculate T1 and T2, then e = d + T1 and a = T1 + T2. -// The values for e and a are stored in d and h, ready for rotation. -#define SHA512ROUND(index, const, a, b, c, d, e, f, g, h) \ - SHA512T1(const, e, f, g, h); \ - SHA512T2(a, b, c); \ - MOVQ BX, h; \ - ADDQ AX, d; \ - ADDQ AX, h - -#define SHA512ROUND0(index, const, a, b, c, d, e, f, g, h) \ - MSGSCHEDULE0(index); \ - SHA512ROUND(index, const, a, b, c, d, e, f, g, h) - -#define SHA512ROUND1(index, const, a, b, c, d, e, f, g, h) \ - MSGSCHEDULE1(index); \ - SHA512ROUND(index, const, a, b, c, d, e, f, g, h) - -TEXT ·blockAMD64(SB),0,$648-32 - MOVQ p_base+8(FP), SI - MOVQ p_len+16(FP), DX - SHRQ $7, DX - SHLQ $7, DX - - LEAQ (SI)(DX*1), DI - MOVQ DI, 640(SP) - CMPQ SI, DI - JEQ end - - MOVQ dig+0(FP), BP - MOVQ (0*8)(BP), R8 // a = H0 - MOVQ (1*8)(BP), R9 // b = H1 - MOVQ (2*8)(BP), R10 // c = H2 - MOVQ (3*8)(BP), R11 // d = H3 - MOVQ (4*8)(BP), R12 // e = H4 - MOVQ (5*8)(BP), R13 // f = H5 - MOVQ (6*8)(BP), R14 // g = H6 - MOVQ (7*8)(BP), R15 // h = H7 +// func blockAMD64(dig *digest, p []byte) +TEXT ·blockAMD64(SB), $648-32 + MOVQ p_base+8(FP), SI + MOVQ p_len+16(FP), DX + SHRQ $0x07, DX + SHLQ $0x07, DX + LEAQ (SI)(DX*1), DI + MOVQ DI, 640(SP) + CMPQ SI, DI + JEQ end + MOVQ dig+0(FP), BP + MOVQ (BP), R8 + MOVQ 8(BP), R9 + MOVQ 16(BP), R10 + MOVQ 24(BP), R11 + MOVQ 32(BP), R12 + MOVQ 40(BP), R13 + MOVQ 48(BP), R14 + MOVQ 56(BP), R15 loop: - MOVQ SP, BP // message schedule - - SHA512ROUND0(0, 0x428a2f98d728ae22, R8, R9, R10, R11, R12, R13, R14, R15) - SHA512ROUND0(1, 0x7137449123ef65cd, R15, R8, R9, R10, R11, R12, R13, R14) - SHA512ROUND0(2, 0xb5c0fbcfec4d3b2f, R14, R15, R8, R9, R10, R11, R12, R13) - SHA512ROUND0(3, 0xe9b5dba58189dbbc, R13, R14, R15, R8, R9, R10, R11, R12) - SHA512ROUND0(4, 0x3956c25bf348b538, R12, R13, R14, R15, R8, R9, R10, R11) - SHA512ROUND0(5, 0x59f111f1b605d019, R11, R12, R13, R14, R15, R8, R9, R10) - SHA512ROUND0(6, 0x923f82a4af194f9b, R10, R11, R12, R13, R14, R15, R8, R9) - SHA512ROUND0(7, 0xab1c5ed5da6d8118, R9, R10, R11, R12, R13, R14, R15, R8) - SHA512ROUND0(8, 0xd807aa98a3030242, R8, R9, R10, R11, R12, R13, R14, R15) - SHA512ROUND0(9, 0x12835b0145706fbe, R15, R8, R9, R10, R11, R12, R13, R14) - SHA512ROUND0(10, 0x243185be4ee4b28c, R14, R15, R8, R9, R10, R11, R12, R13) - SHA512ROUND0(11, 0x550c7dc3d5ffb4e2, R13, R14, R15, R8, R9, R10, R11, R12) - SHA512ROUND0(12, 0x72be5d74f27b896f, R12, R13, R14, R15, R8, R9, R10, R11) - SHA512ROUND0(13, 0x80deb1fe3b1696b1, R11, R12, R13, R14, R15, R8, R9, R10) - SHA512ROUND0(14, 0x9bdc06a725c71235, R10, R11, R12, R13, R14, R15, R8, R9) - SHA512ROUND0(15, 0xc19bf174cf692694, R9, R10, R11, R12, R13, R14, R15, R8) - - SHA512ROUND1(16, 0xe49b69c19ef14ad2, R8, R9, R10, R11, R12, R13, R14, R15) - SHA512ROUND1(17, 0xefbe4786384f25e3, R15, R8, R9, R10, R11, R12, R13, R14) - SHA512ROUND1(18, 0x0fc19dc68b8cd5b5, R14, R15, R8, R9, R10, R11, R12, R13) - SHA512ROUND1(19, 0x240ca1cc77ac9c65, R13, R14, R15, R8, R9, R10, R11, R12) - SHA512ROUND1(20, 0x2de92c6f592b0275, R12, R13, R14, R15, R8, R9, R10, R11) - SHA512ROUND1(21, 0x4a7484aa6ea6e483, R11, R12, R13, R14, R15, R8, R9, R10) - SHA512ROUND1(22, 0x5cb0a9dcbd41fbd4, R10, R11, R12, R13, R14, R15, R8, R9) - SHA512ROUND1(23, 0x76f988da831153b5, R9, R10, R11, R12, R13, R14, R15, R8) - SHA512ROUND1(24, 0x983e5152ee66dfab, R8, R9, R10, R11, R12, R13, R14, R15) - SHA512ROUND1(25, 0xa831c66d2db43210, R15, R8, R9, R10, R11, R12, R13, R14) - SHA512ROUND1(26, 0xb00327c898fb213f, R14, R15, R8, R9, R10, R11, R12, R13) - SHA512ROUND1(27, 0xbf597fc7beef0ee4, R13, R14, R15, R8, R9, R10, R11, R12) - SHA512ROUND1(28, 0xc6e00bf33da88fc2, R12, R13, R14, R15, R8, R9, R10, R11) - SHA512ROUND1(29, 0xd5a79147930aa725, R11, R12, R13, R14, R15, R8, R9, R10) - SHA512ROUND1(30, 0x06ca6351e003826f, R10, R11, R12, R13, R14, R15, R8, R9) - SHA512ROUND1(31, 0x142929670a0e6e70, R9, R10, R11, R12, R13, R14, R15, R8) - SHA512ROUND1(32, 0x27b70a8546d22ffc, R8, R9, R10, R11, R12, R13, R14, R15) - SHA512ROUND1(33, 0x2e1b21385c26c926, R15, R8, R9, R10, R11, R12, R13, R14) - SHA512ROUND1(34, 0x4d2c6dfc5ac42aed, R14, R15, R8, R9, R10, R11, R12, R13) - SHA512ROUND1(35, 0x53380d139d95b3df, R13, R14, R15, R8, R9, R10, R11, R12) - SHA512ROUND1(36, 0x650a73548baf63de, R12, R13, R14, R15, R8, R9, R10, R11) - SHA512ROUND1(37, 0x766a0abb3c77b2a8, R11, R12, R13, R14, R15, R8, R9, R10) - SHA512ROUND1(38, 0x81c2c92e47edaee6, R10, R11, R12, R13, R14, R15, R8, R9) - SHA512ROUND1(39, 0x92722c851482353b, R9, R10, R11, R12, R13, R14, R15, R8) - SHA512ROUND1(40, 0xa2bfe8a14cf10364, R8, R9, R10, R11, R12, R13, R14, R15) - SHA512ROUND1(41, 0xa81a664bbc423001, R15, R8, R9, R10, R11, R12, R13, R14) - SHA512ROUND1(42, 0xc24b8b70d0f89791, R14, R15, R8, R9, R10, R11, R12, R13) - SHA512ROUND1(43, 0xc76c51a30654be30, R13, R14, R15, R8, R9, R10, R11, R12) - SHA512ROUND1(44, 0xd192e819d6ef5218, R12, R13, R14, R15, R8, R9, R10, R11) - SHA512ROUND1(45, 0xd69906245565a910, R11, R12, R13, R14, R15, R8, R9, R10) - SHA512ROUND1(46, 0xf40e35855771202a, R10, R11, R12, R13, R14, R15, R8, R9) - SHA512ROUND1(47, 0x106aa07032bbd1b8, R9, R10, R11, R12, R13, R14, R15, R8) - SHA512ROUND1(48, 0x19a4c116b8d2d0c8, R8, R9, R10, R11, R12, R13, R14, R15) - SHA512ROUND1(49, 0x1e376c085141ab53, R15, R8, R9, R10, R11, R12, R13, R14) - SHA512ROUND1(50, 0x2748774cdf8eeb99, R14, R15, R8, R9, R10, R11, R12, R13) - SHA512ROUND1(51, 0x34b0bcb5e19b48a8, R13, R14, R15, R8, R9, R10, R11, R12) - SHA512ROUND1(52, 0x391c0cb3c5c95a63, R12, R13, R14, R15, R8, R9, R10, R11) - SHA512ROUND1(53, 0x4ed8aa4ae3418acb, R11, R12, R13, R14, R15, R8, R9, R10) - SHA512ROUND1(54, 0x5b9cca4f7763e373, R10, R11, R12, R13, R14, R15, R8, R9) - SHA512ROUND1(55, 0x682e6ff3d6b2b8a3, R9, R10, R11, R12, R13, R14, R15, R8) - SHA512ROUND1(56, 0x748f82ee5defb2fc, R8, R9, R10, R11, R12, R13, R14, R15) - SHA512ROUND1(57, 0x78a5636f43172f60, R15, R8, R9, R10, R11, R12, R13, R14) - SHA512ROUND1(58, 0x84c87814a1f0ab72, R14, R15, R8, R9, R10, R11, R12, R13) - SHA512ROUND1(59, 0x8cc702081a6439ec, R13, R14, R15, R8, R9, R10, R11, R12) - SHA512ROUND1(60, 0x90befffa23631e28, R12, R13, R14, R15, R8, R9, R10, R11) - SHA512ROUND1(61, 0xa4506cebde82bde9, R11, R12, R13, R14, R15, R8, R9, R10) - SHA512ROUND1(62, 0xbef9a3f7b2c67915, R10, R11, R12, R13, R14, R15, R8, R9) - SHA512ROUND1(63, 0xc67178f2e372532b, R9, R10, R11, R12, R13, R14, R15, R8) - SHA512ROUND1(64, 0xca273eceea26619c, R8, R9, R10, R11, R12, R13, R14, R15) - SHA512ROUND1(65, 0xd186b8c721c0c207, R15, R8, R9, R10, R11, R12, R13, R14) - SHA512ROUND1(66, 0xeada7dd6cde0eb1e, R14, R15, R8, R9, R10, R11, R12, R13) - SHA512ROUND1(67, 0xf57d4f7fee6ed178, R13, R14, R15, R8, R9, R10, R11, R12) - SHA512ROUND1(68, 0x06f067aa72176fba, R12, R13, R14, R15, R8, R9, R10, R11) - SHA512ROUND1(69, 0x0a637dc5a2c898a6, R11, R12, R13, R14, R15, R8, R9, R10) - SHA512ROUND1(70, 0x113f9804bef90dae, R10, R11, R12, R13, R14, R15, R8, R9) - SHA512ROUND1(71, 0x1b710b35131c471b, R9, R10, R11, R12, R13, R14, R15, R8) - SHA512ROUND1(72, 0x28db77f523047d84, R8, R9, R10, R11, R12, R13, R14, R15) - SHA512ROUND1(73, 0x32caab7b40c72493, R15, R8, R9, R10, R11, R12, R13, R14) - SHA512ROUND1(74, 0x3c9ebe0a15c9bebc, R14, R15, R8, R9, R10, R11, R12, R13) - SHA512ROUND1(75, 0x431d67c49c100d4c, R13, R14, R15, R8, R9, R10, R11, R12) - SHA512ROUND1(76, 0x4cc5d4becb3e42b6, R12, R13, R14, R15, R8, R9, R10, R11) - SHA512ROUND1(77, 0x597f299cfc657e2a, R11, R12, R13, R14, R15, R8, R9, R10) - SHA512ROUND1(78, 0x5fcb6fab3ad6faec, R10, R11, R12, R13, R14, R15, R8, R9) - SHA512ROUND1(79, 0x6c44198c4a475817, R9, R10, R11, R12, R13, R14, R15, R8) - - MOVQ dig+0(FP), BP - ADDQ (0*8)(BP), R8 // H0 = a + H0 - MOVQ R8, (0*8)(BP) - ADDQ (1*8)(BP), R9 // H1 = b + H1 - MOVQ R9, (1*8)(BP) - ADDQ (2*8)(BP), R10 // H2 = c + H2 - MOVQ R10, (2*8)(BP) - ADDQ (3*8)(BP), R11 // H3 = d + H3 - MOVQ R11, (3*8)(BP) - ADDQ (4*8)(BP), R12 // H4 = e + H4 - MOVQ R12, (4*8)(BP) - ADDQ (5*8)(BP), R13 // H5 = f + H5 - MOVQ R13, (5*8)(BP) - ADDQ (6*8)(BP), R14 // H6 = g + H6 - MOVQ R14, (6*8)(BP) - ADDQ (7*8)(BP), R15 // H7 = h + H7 - MOVQ R15, (7*8)(BP) - - ADDQ $128, SI - CMPQ SI, 640(SP) - JB loop + MOVQ SP, BP + MOVQ (SI), AX + BSWAPQ AX + MOVQ AX, (BP) + MOVQ $0x428a2f98d728ae22, DX + ADDQ AX, R15 + MOVQ R12, AX + ADDQ DX, R15 + MOVQ R12, CX + RORQ $0x0e, AX + MOVQ R12, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R12, CX + RORQ $0x29, DX + ANDQ R13, CX + XORQ AX, DX + MOVQ R12, AX + NOTQ AX + ADDQ DX, R15 + ANDQ R14, AX + XORQ CX, AX + ADDQ R15, AX + MOVQ R8, DI + MOVQ R10, BX + RORQ $0x1c, DI + MOVQ R8, DX + ANDQ R9, BX + RORQ $0x22, DX + MOVQ R8, CX + ANDQ R10, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R8, DX + MOVQ R9, CX + RORQ $0x27, DX + ANDQ R8, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R15 + ADDQ AX, R11 + ADDQ AX, R15 + MOVQ 8(SI), AX + BSWAPQ AX + MOVQ AX, 8(BP) + MOVQ $0x7137449123ef65cd, DX + ADDQ AX, R14 + MOVQ R11, AX + ADDQ DX, R14 + MOVQ R11, CX + RORQ $0x0e, AX + MOVQ R11, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R11, CX + RORQ $0x29, DX + ANDQ R12, CX + XORQ AX, DX + MOVQ R11, AX + NOTQ AX + ADDQ DX, R14 + ANDQ R13, AX + XORQ CX, AX + ADDQ R14, AX + MOVQ R15, DI + MOVQ R9, BX + RORQ $0x1c, DI + MOVQ R15, DX + ANDQ R8, BX + RORQ $0x22, DX + MOVQ R15, CX + ANDQ R9, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R15, DX + MOVQ R8, CX + RORQ $0x27, DX + ANDQ R15, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R14 + ADDQ AX, R10 + ADDQ AX, R14 + MOVQ 16(SI), AX + BSWAPQ AX + MOVQ AX, 16(BP) + MOVQ $0xb5c0fbcfec4d3b2f, DX + ADDQ AX, R13 + MOVQ R10, AX + ADDQ DX, R13 + MOVQ R10, CX + RORQ $0x0e, AX + MOVQ R10, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R10, CX + RORQ $0x29, DX + ANDQ R11, CX + XORQ AX, DX + MOVQ R10, AX + NOTQ AX + ADDQ DX, R13 + ANDQ R12, AX + XORQ CX, AX + ADDQ R13, AX + MOVQ R14, DI + MOVQ R8, BX + RORQ $0x1c, DI + MOVQ R14, DX + ANDQ R15, BX + RORQ $0x22, DX + MOVQ R14, CX + ANDQ R8, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R14, DX + MOVQ R15, CX + RORQ $0x27, DX + ANDQ R14, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R13 + ADDQ AX, R9 + ADDQ AX, R13 + MOVQ 24(SI), AX + BSWAPQ AX + MOVQ AX, 24(BP) + MOVQ $0xe9b5dba58189dbbc, DX + ADDQ AX, R12 + MOVQ R9, AX + ADDQ DX, R12 + MOVQ R9, CX + RORQ $0x0e, AX + MOVQ R9, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R9, CX + RORQ $0x29, DX + ANDQ R10, CX + XORQ AX, DX + MOVQ R9, AX + NOTQ AX + ADDQ DX, R12 + ANDQ R11, AX + XORQ CX, AX + ADDQ R12, AX + MOVQ R13, DI + MOVQ R15, BX + RORQ $0x1c, DI + MOVQ R13, DX + ANDQ R14, BX + RORQ $0x22, DX + MOVQ R13, CX + ANDQ R15, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R13, DX + MOVQ R14, CX + RORQ $0x27, DX + ANDQ R13, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R12 + ADDQ AX, R8 + ADDQ AX, R12 + MOVQ 32(SI), AX + BSWAPQ AX + MOVQ AX, 32(BP) + MOVQ $0x3956c25bf348b538, DX + ADDQ AX, R11 + MOVQ R8, AX + ADDQ DX, R11 + MOVQ R8, CX + RORQ $0x0e, AX + MOVQ R8, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R8, CX + RORQ $0x29, DX + ANDQ R9, CX + XORQ AX, DX + MOVQ R8, AX + NOTQ AX + ADDQ DX, R11 + ANDQ R10, AX + XORQ CX, AX + ADDQ R11, AX + MOVQ R12, DI + MOVQ R14, BX + RORQ $0x1c, DI + MOVQ R12, DX + ANDQ R13, BX + RORQ $0x22, DX + MOVQ R12, CX + ANDQ R14, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R12, DX + MOVQ R13, CX + RORQ $0x27, DX + ANDQ R12, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R11 + ADDQ AX, R15 + ADDQ AX, R11 + MOVQ 40(SI), AX + BSWAPQ AX + MOVQ AX, 40(BP) + MOVQ $0x59f111f1b605d019, DX + ADDQ AX, R10 + MOVQ R15, AX + ADDQ DX, R10 + MOVQ R15, CX + RORQ $0x0e, AX + MOVQ R15, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R15, CX + RORQ $0x29, DX + ANDQ R8, CX + XORQ AX, DX + MOVQ R15, AX + NOTQ AX + ADDQ DX, R10 + ANDQ R9, AX + XORQ CX, AX + ADDQ R10, AX + MOVQ R11, DI + MOVQ R13, BX + RORQ $0x1c, DI + MOVQ R11, DX + ANDQ R12, BX + RORQ $0x22, DX + MOVQ R11, CX + ANDQ R13, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R11, DX + MOVQ R12, CX + RORQ $0x27, DX + ANDQ R11, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R10 + ADDQ AX, R14 + ADDQ AX, R10 + MOVQ 48(SI), AX + BSWAPQ AX + MOVQ AX, 48(BP) + MOVQ $0x923f82a4af194f9b, DX + ADDQ AX, R9 + MOVQ R14, AX + ADDQ DX, R9 + MOVQ R14, CX + RORQ $0x0e, AX + MOVQ R14, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R14, CX + RORQ $0x29, DX + ANDQ R15, CX + XORQ AX, DX + MOVQ R14, AX + NOTQ AX + ADDQ DX, R9 + ANDQ R8, AX + XORQ CX, AX + ADDQ R9, AX + MOVQ R10, DI + MOVQ R12, BX + RORQ $0x1c, DI + MOVQ R10, DX + ANDQ R11, BX + RORQ $0x22, DX + MOVQ R10, CX + ANDQ R12, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R10, DX + MOVQ R11, CX + RORQ $0x27, DX + ANDQ R10, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R9 + ADDQ AX, R13 + ADDQ AX, R9 + MOVQ 56(SI), AX + BSWAPQ AX + MOVQ AX, 56(BP) + MOVQ $0xab1c5ed5da6d8118, DX + ADDQ AX, R8 + MOVQ R13, AX + ADDQ DX, R8 + MOVQ R13, CX + RORQ $0x0e, AX + MOVQ R13, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R13, CX + RORQ $0x29, DX + ANDQ R14, CX + XORQ AX, DX + MOVQ R13, AX + NOTQ AX + ADDQ DX, R8 + ANDQ R15, AX + XORQ CX, AX + ADDQ R8, AX + MOVQ R9, DI + MOVQ R11, BX + RORQ $0x1c, DI + MOVQ R9, DX + ANDQ R10, BX + RORQ $0x22, DX + MOVQ R9, CX + ANDQ R11, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R9, DX + MOVQ R10, CX + RORQ $0x27, DX + ANDQ R9, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R8 + ADDQ AX, R12 + ADDQ AX, R8 + MOVQ 64(SI), AX + BSWAPQ AX + MOVQ AX, 64(BP) + MOVQ $0xd807aa98a3030242, DX + ADDQ AX, R15 + MOVQ R12, AX + ADDQ DX, R15 + MOVQ R12, CX + RORQ $0x0e, AX + MOVQ R12, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R12, CX + RORQ $0x29, DX + ANDQ R13, CX + XORQ AX, DX + MOVQ R12, AX + NOTQ AX + ADDQ DX, R15 + ANDQ R14, AX + XORQ CX, AX + ADDQ R15, AX + MOVQ R8, DI + MOVQ R10, BX + RORQ $0x1c, DI + MOVQ R8, DX + ANDQ R9, BX + RORQ $0x22, DX + MOVQ R8, CX + ANDQ R10, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R8, DX + MOVQ R9, CX + RORQ $0x27, DX + ANDQ R8, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R15 + ADDQ AX, R11 + ADDQ AX, R15 + MOVQ 72(SI), AX + BSWAPQ AX + MOVQ AX, 72(BP) + MOVQ $0x12835b0145706fbe, DX + ADDQ AX, R14 + MOVQ R11, AX + ADDQ DX, R14 + MOVQ R11, CX + RORQ $0x0e, AX + MOVQ R11, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R11, CX + RORQ $0x29, DX + ANDQ R12, CX + XORQ AX, DX + MOVQ R11, AX + NOTQ AX + ADDQ DX, R14 + ANDQ R13, AX + XORQ CX, AX + ADDQ R14, AX + MOVQ R15, DI + MOVQ R9, BX + RORQ $0x1c, DI + MOVQ R15, DX + ANDQ R8, BX + RORQ $0x22, DX + MOVQ R15, CX + ANDQ R9, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R15, DX + MOVQ R8, CX + RORQ $0x27, DX + ANDQ R15, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R14 + ADDQ AX, R10 + ADDQ AX, R14 + MOVQ 80(SI), AX + BSWAPQ AX + MOVQ AX, 80(BP) + MOVQ $0x243185be4ee4b28c, DX + ADDQ AX, R13 + MOVQ R10, AX + ADDQ DX, R13 + MOVQ R10, CX + RORQ $0x0e, AX + MOVQ R10, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R10, CX + RORQ $0x29, DX + ANDQ R11, CX + XORQ AX, DX + MOVQ R10, AX + NOTQ AX + ADDQ DX, R13 + ANDQ R12, AX + XORQ CX, AX + ADDQ R13, AX + MOVQ R14, DI + MOVQ R8, BX + RORQ $0x1c, DI + MOVQ R14, DX + ANDQ R15, BX + RORQ $0x22, DX + MOVQ R14, CX + ANDQ R8, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R14, DX + MOVQ R15, CX + RORQ $0x27, DX + ANDQ R14, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R13 + ADDQ AX, R9 + ADDQ AX, R13 + MOVQ 88(SI), AX + BSWAPQ AX + MOVQ AX, 88(BP) + MOVQ $0x550c7dc3d5ffb4e2, DX + ADDQ AX, R12 + MOVQ R9, AX + ADDQ DX, R12 + MOVQ R9, CX + RORQ $0x0e, AX + MOVQ R9, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R9, CX + RORQ $0x29, DX + ANDQ R10, CX + XORQ AX, DX + MOVQ R9, AX + NOTQ AX + ADDQ DX, R12 + ANDQ R11, AX + XORQ CX, AX + ADDQ R12, AX + MOVQ R13, DI + MOVQ R15, BX + RORQ $0x1c, DI + MOVQ R13, DX + ANDQ R14, BX + RORQ $0x22, DX + MOVQ R13, CX + ANDQ R15, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R13, DX + MOVQ R14, CX + RORQ $0x27, DX + ANDQ R13, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R12 + ADDQ AX, R8 + ADDQ AX, R12 + MOVQ 96(SI), AX + BSWAPQ AX + MOVQ AX, 96(BP) + MOVQ $0x72be5d74f27b896f, DX + ADDQ AX, R11 + MOVQ R8, AX + ADDQ DX, R11 + MOVQ R8, CX + RORQ $0x0e, AX + MOVQ R8, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R8, CX + RORQ $0x29, DX + ANDQ R9, CX + XORQ AX, DX + MOVQ R8, AX + NOTQ AX + ADDQ DX, R11 + ANDQ R10, AX + XORQ CX, AX + ADDQ R11, AX + MOVQ R12, DI + MOVQ R14, BX + RORQ $0x1c, DI + MOVQ R12, DX + ANDQ R13, BX + RORQ $0x22, DX + MOVQ R12, CX + ANDQ R14, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R12, DX + MOVQ R13, CX + RORQ $0x27, DX + ANDQ R12, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R11 + ADDQ AX, R15 + ADDQ AX, R11 + MOVQ 104(SI), AX + BSWAPQ AX + MOVQ AX, 104(BP) + MOVQ $0x80deb1fe3b1696b1, DX + ADDQ AX, R10 + MOVQ R15, AX + ADDQ DX, R10 + MOVQ R15, CX + RORQ $0x0e, AX + MOVQ R15, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R15, CX + RORQ $0x29, DX + ANDQ R8, CX + XORQ AX, DX + MOVQ R15, AX + NOTQ AX + ADDQ DX, R10 + ANDQ R9, AX + XORQ CX, AX + ADDQ R10, AX + MOVQ R11, DI + MOVQ R13, BX + RORQ $0x1c, DI + MOVQ R11, DX + ANDQ R12, BX + RORQ $0x22, DX + MOVQ R11, CX + ANDQ R13, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R11, DX + MOVQ R12, CX + RORQ $0x27, DX + ANDQ R11, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R10 + ADDQ AX, R14 + ADDQ AX, R10 + MOVQ 112(SI), AX + BSWAPQ AX + MOVQ AX, 112(BP) + MOVQ $0x9bdc06a725c71235, DX + ADDQ AX, R9 + MOVQ R14, AX + ADDQ DX, R9 + MOVQ R14, CX + RORQ $0x0e, AX + MOVQ R14, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R14, CX + RORQ $0x29, DX + ANDQ R15, CX + XORQ AX, DX + MOVQ R14, AX + NOTQ AX + ADDQ DX, R9 + ANDQ R8, AX + XORQ CX, AX + ADDQ R9, AX + MOVQ R10, DI + MOVQ R12, BX + RORQ $0x1c, DI + MOVQ R10, DX + ANDQ R11, BX + RORQ $0x22, DX + MOVQ R10, CX + ANDQ R12, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R10, DX + MOVQ R11, CX + RORQ $0x27, DX + ANDQ R10, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R9 + ADDQ AX, R13 + ADDQ AX, R9 + MOVQ 120(SI), AX + BSWAPQ AX + MOVQ AX, 120(BP) + MOVQ $0xc19bf174cf692694, DX + ADDQ AX, R8 + MOVQ R13, AX + ADDQ DX, R8 + MOVQ R13, CX + RORQ $0x0e, AX + MOVQ R13, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R13, CX + RORQ $0x29, DX + ANDQ R14, CX + XORQ AX, DX + MOVQ R13, AX + NOTQ AX + ADDQ DX, R8 + ANDQ R15, AX + XORQ CX, AX + ADDQ R8, AX + MOVQ R9, DI + MOVQ R11, BX + RORQ $0x1c, DI + MOVQ R9, DX + ANDQ R10, BX + RORQ $0x22, DX + MOVQ R9, CX + ANDQ R11, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R9, DX + MOVQ R10, CX + RORQ $0x27, DX + ANDQ R9, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R8 + ADDQ AX, R12 + ADDQ AX, R8 + MOVQ 112(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 8(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 72(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ (BP), BX + ADDQ BX, AX + MOVQ AX, 128(BP) + MOVQ $0xe49b69c19ef14ad2, DX + ADDQ AX, R15 + MOVQ R12, AX + ADDQ DX, R15 + MOVQ R12, CX + RORQ $0x0e, AX + MOVQ R12, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R12, CX + RORQ $0x29, DX + ANDQ R13, CX + XORQ AX, DX + MOVQ R12, AX + NOTQ AX + ADDQ DX, R15 + ANDQ R14, AX + XORQ CX, AX + ADDQ R15, AX + MOVQ R8, DI + MOVQ R10, BX + RORQ $0x1c, DI + MOVQ R8, DX + ANDQ R9, BX + RORQ $0x22, DX + MOVQ R8, CX + ANDQ R10, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R8, DX + MOVQ R9, CX + RORQ $0x27, DX + ANDQ R8, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R15 + ADDQ AX, R11 + ADDQ AX, R15 + MOVQ 120(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 16(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 80(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 8(BP), BX + ADDQ BX, AX + MOVQ AX, 136(BP) + MOVQ $0xefbe4786384f25e3, DX + ADDQ AX, R14 + MOVQ R11, AX + ADDQ DX, R14 + MOVQ R11, CX + RORQ $0x0e, AX + MOVQ R11, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R11, CX + RORQ $0x29, DX + ANDQ R12, CX + XORQ AX, DX + MOVQ R11, AX + NOTQ AX + ADDQ DX, R14 + ANDQ R13, AX + XORQ CX, AX + ADDQ R14, AX + MOVQ R15, DI + MOVQ R9, BX + RORQ $0x1c, DI + MOVQ R15, DX + ANDQ R8, BX + RORQ $0x22, DX + MOVQ R15, CX + ANDQ R9, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R15, DX + MOVQ R8, CX + RORQ $0x27, DX + ANDQ R15, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R14 + ADDQ AX, R10 + ADDQ AX, R14 + MOVQ 128(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 24(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 88(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 16(BP), BX + ADDQ BX, AX + MOVQ AX, 144(BP) + MOVQ $0x0fc19dc68b8cd5b5, DX + ADDQ AX, R13 + MOVQ R10, AX + ADDQ DX, R13 + MOVQ R10, CX + RORQ $0x0e, AX + MOVQ R10, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R10, CX + RORQ $0x29, DX + ANDQ R11, CX + XORQ AX, DX + MOVQ R10, AX + NOTQ AX + ADDQ DX, R13 + ANDQ R12, AX + XORQ CX, AX + ADDQ R13, AX + MOVQ R14, DI + MOVQ R8, BX + RORQ $0x1c, DI + MOVQ R14, DX + ANDQ R15, BX + RORQ $0x22, DX + MOVQ R14, CX + ANDQ R8, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R14, DX + MOVQ R15, CX + RORQ $0x27, DX + ANDQ R14, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R13 + ADDQ AX, R9 + ADDQ AX, R13 + MOVQ 136(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 32(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 96(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 24(BP), BX + ADDQ BX, AX + MOVQ AX, 152(BP) + MOVQ $0x240ca1cc77ac9c65, DX + ADDQ AX, R12 + MOVQ R9, AX + ADDQ DX, R12 + MOVQ R9, CX + RORQ $0x0e, AX + MOVQ R9, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R9, CX + RORQ $0x29, DX + ANDQ R10, CX + XORQ AX, DX + MOVQ R9, AX + NOTQ AX + ADDQ DX, R12 + ANDQ R11, AX + XORQ CX, AX + ADDQ R12, AX + MOVQ R13, DI + MOVQ R15, BX + RORQ $0x1c, DI + MOVQ R13, DX + ANDQ R14, BX + RORQ $0x22, DX + MOVQ R13, CX + ANDQ R15, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R13, DX + MOVQ R14, CX + RORQ $0x27, DX + ANDQ R13, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R12 + ADDQ AX, R8 + ADDQ AX, R12 + MOVQ 144(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 40(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 104(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 32(BP), BX + ADDQ BX, AX + MOVQ AX, 160(BP) + MOVQ $0x2de92c6f592b0275, DX + ADDQ AX, R11 + MOVQ R8, AX + ADDQ DX, R11 + MOVQ R8, CX + RORQ $0x0e, AX + MOVQ R8, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R8, CX + RORQ $0x29, DX + ANDQ R9, CX + XORQ AX, DX + MOVQ R8, AX + NOTQ AX + ADDQ DX, R11 + ANDQ R10, AX + XORQ CX, AX + ADDQ R11, AX + MOVQ R12, DI + MOVQ R14, BX + RORQ $0x1c, DI + MOVQ R12, DX + ANDQ R13, BX + RORQ $0x22, DX + MOVQ R12, CX + ANDQ R14, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R12, DX + MOVQ R13, CX + RORQ $0x27, DX + ANDQ R12, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R11 + ADDQ AX, R15 + ADDQ AX, R11 + MOVQ 152(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 48(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 112(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 40(BP), BX + ADDQ BX, AX + MOVQ AX, 168(BP) + MOVQ $0x4a7484aa6ea6e483, DX + ADDQ AX, R10 + MOVQ R15, AX + ADDQ DX, R10 + MOVQ R15, CX + RORQ $0x0e, AX + MOVQ R15, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R15, CX + RORQ $0x29, DX + ANDQ R8, CX + XORQ AX, DX + MOVQ R15, AX + NOTQ AX + ADDQ DX, R10 + ANDQ R9, AX + XORQ CX, AX + ADDQ R10, AX + MOVQ R11, DI + MOVQ R13, BX + RORQ $0x1c, DI + MOVQ R11, DX + ANDQ R12, BX + RORQ $0x22, DX + MOVQ R11, CX + ANDQ R13, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R11, DX + MOVQ R12, CX + RORQ $0x27, DX + ANDQ R11, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R10 + ADDQ AX, R14 + ADDQ AX, R10 + MOVQ 160(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 56(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 120(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 48(BP), BX + ADDQ BX, AX + MOVQ AX, 176(BP) + MOVQ $0x5cb0a9dcbd41fbd4, DX + ADDQ AX, R9 + MOVQ R14, AX + ADDQ DX, R9 + MOVQ R14, CX + RORQ $0x0e, AX + MOVQ R14, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R14, CX + RORQ $0x29, DX + ANDQ R15, CX + XORQ AX, DX + MOVQ R14, AX + NOTQ AX + ADDQ DX, R9 + ANDQ R8, AX + XORQ CX, AX + ADDQ R9, AX + MOVQ R10, DI + MOVQ R12, BX + RORQ $0x1c, DI + MOVQ R10, DX + ANDQ R11, BX + RORQ $0x22, DX + MOVQ R10, CX + ANDQ R12, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R10, DX + MOVQ R11, CX + RORQ $0x27, DX + ANDQ R10, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R9 + ADDQ AX, R13 + ADDQ AX, R9 + MOVQ 168(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 64(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 128(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 56(BP), BX + ADDQ BX, AX + MOVQ AX, 184(BP) + MOVQ $0x76f988da831153b5, DX + ADDQ AX, R8 + MOVQ R13, AX + ADDQ DX, R8 + MOVQ R13, CX + RORQ $0x0e, AX + MOVQ R13, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R13, CX + RORQ $0x29, DX + ANDQ R14, CX + XORQ AX, DX + MOVQ R13, AX + NOTQ AX + ADDQ DX, R8 + ANDQ R15, AX + XORQ CX, AX + ADDQ R8, AX + MOVQ R9, DI + MOVQ R11, BX + RORQ $0x1c, DI + MOVQ R9, DX + ANDQ R10, BX + RORQ $0x22, DX + MOVQ R9, CX + ANDQ R11, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R9, DX + MOVQ R10, CX + RORQ $0x27, DX + ANDQ R9, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R8 + ADDQ AX, R12 + ADDQ AX, R8 + MOVQ 176(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 72(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 136(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 64(BP), BX + ADDQ BX, AX + MOVQ AX, 192(BP) + MOVQ $0x983e5152ee66dfab, DX + ADDQ AX, R15 + MOVQ R12, AX + ADDQ DX, R15 + MOVQ R12, CX + RORQ $0x0e, AX + MOVQ R12, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R12, CX + RORQ $0x29, DX + ANDQ R13, CX + XORQ AX, DX + MOVQ R12, AX + NOTQ AX + ADDQ DX, R15 + ANDQ R14, AX + XORQ CX, AX + ADDQ R15, AX + MOVQ R8, DI + MOVQ R10, BX + RORQ $0x1c, DI + MOVQ R8, DX + ANDQ R9, BX + RORQ $0x22, DX + MOVQ R8, CX + ANDQ R10, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R8, DX + MOVQ R9, CX + RORQ $0x27, DX + ANDQ R8, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R15 + ADDQ AX, R11 + ADDQ AX, R15 + MOVQ 184(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 80(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 144(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 72(BP), BX + ADDQ BX, AX + MOVQ AX, 200(BP) + MOVQ $0xa831c66d2db43210, DX + ADDQ AX, R14 + MOVQ R11, AX + ADDQ DX, R14 + MOVQ R11, CX + RORQ $0x0e, AX + MOVQ R11, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R11, CX + RORQ $0x29, DX + ANDQ R12, CX + XORQ AX, DX + MOVQ R11, AX + NOTQ AX + ADDQ DX, R14 + ANDQ R13, AX + XORQ CX, AX + ADDQ R14, AX + MOVQ R15, DI + MOVQ R9, BX + RORQ $0x1c, DI + MOVQ R15, DX + ANDQ R8, BX + RORQ $0x22, DX + MOVQ R15, CX + ANDQ R9, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R15, DX + MOVQ R8, CX + RORQ $0x27, DX + ANDQ R15, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R14 + ADDQ AX, R10 + ADDQ AX, R14 + MOVQ 192(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 88(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 152(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 80(BP), BX + ADDQ BX, AX + MOVQ AX, 208(BP) + MOVQ $0xb00327c898fb213f, DX + ADDQ AX, R13 + MOVQ R10, AX + ADDQ DX, R13 + MOVQ R10, CX + RORQ $0x0e, AX + MOVQ R10, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R10, CX + RORQ $0x29, DX + ANDQ R11, CX + XORQ AX, DX + MOVQ R10, AX + NOTQ AX + ADDQ DX, R13 + ANDQ R12, AX + XORQ CX, AX + ADDQ R13, AX + MOVQ R14, DI + MOVQ R8, BX + RORQ $0x1c, DI + MOVQ R14, DX + ANDQ R15, BX + RORQ $0x22, DX + MOVQ R14, CX + ANDQ R8, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R14, DX + MOVQ R15, CX + RORQ $0x27, DX + ANDQ R14, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R13 + ADDQ AX, R9 + ADDQ AX, R13 + MOVQ 200(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 96(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 160(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 88(BP), BX + ADDQ BX, AX + MOVQ AX, 216(BP) + MOVQ $0xbf597fc7beef0ee4, DX + ADDQ AX, R12 + MOVQ R9, AX + ADDQ DX, R12 + MOVQ R9, CX + RORQ $0x0e, AX + MOVQ R9, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R9, CX + RORQ $0x29, DX + ANDQ R10, CX + XORQ AX, DX + MOVQ R9, AX + NOTQ AX + ADDQ DX, R12 + ANDQ R11, AX + XORQ CX, AX + ADDQ R12, AX + MOVQ R13, DI + MOVQ R15, BX + RORQ $0x1c, DI + MOVQ R13, DX + ANDQ R14, BX + RORQ $0x22, DX + MOVQ R13, CX + ANDQ R15, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R13, DX + MOVQ R14, CX + RORQ $0x27, DX + ANDQ R13, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R12 + ADDQ AX, R8 + ADDQ AX, R12 + MOVQ 208(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 104(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 168(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 96(BP), BX + ADDQ BX, AX + MOVQ AX, 224(BP) + MOVQ $0xc6e00bf33da88fc2, DX + ADDQ AX, R11 + MOVQ R8, AX + ADDQ DX, R11 + MOVQ R8, CX + RORQ $0x0e, AX + MOVQ R8, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R8, CX + RORQ $0x29, DX + ANDQ R9, CX + XORQ AX, DX + MOVQ R8, AX + NOTQ AX + ADDQ DX, R11 + ANDQ R10, AX + XORQ CX, AX + ADDQ R11, AX + MOVQ R12, DI + MOVQ R14, BX + RORQ $0x1c, DI + MOVQ R12, DX + ANDQ R13, BX + RORQ $0x22, DX + MOVQ R12, CX + ANDQ R14, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R12, DX + MOVQ R13, CX + RORQ $0x27, DX + ANDQ R12, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R11 + ADDQ AX, R15 + ADDQ AX, R11 + MOVQ 216(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 112(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 176(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 104(BP), BX + ADDQ BX, AX + MOVQ AX, 232(BP) + MOVQ $0xd5a79147930aa725, DX + ADDQ AX, R10 + MOVQ R15, AX + ADDQ DX, R10 + MOVQ R15, CX + RORQ $0x0e, AX + MOVQ R15, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R15, CX + RORQ $0x29, DX + ANDQ R8, CX + XORQ AX, DX + MOVQ R15, AX + NOTQ AX + ADDQ DX, R10 + ANDQ R9, AX + XORQ CX, AX + ADDQ R10, AX + MOVQ R11, DI + MOVQ R13, BX + RORQ $0x1c, DI + MOVQ R11, DX + ANDQ R12, BX + RORQ $0x22, DX + MOVQ R11, CX + ANDQ R13, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R11, DX + MOVQ R12, CX + RORQ $0x27, DX + ANDQ R11, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R10 + ADDQ AX, R14 + ADDQ AX, R10 + MOVQ 224(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 120(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 184(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 112(BP), BX + ADDQ BX, AX + MOVQ AX, 240(BP) + MOVQ $0x06ca6351e003826f, DX + ADDQ AX, R9 + MOVQ R14, AX + ADDQ DX, R9 + MOVQ R14, CX + RORQ $0x0e, AX + MOVQ R14, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R14, CX + RORQ $0x29, DX + ANDQ R15, CX + XORQ AX, DX + MOVQ R14, AX + NOTQ AX + ADDQ DX, R9 + ANDQ R8, AX + XORQ CX, AX + ADDQ R9, AX + MOVQ R10, DI + MOVQ R12, BX + RORQ $0x1c, DI + MOVQ R10, DX + ANDQ R11, BX + RORQ $0x22, DX + MOVQ R10, CX + ANDQ R12, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R10, DX + MOVQ R11, CX + RORQ $0x27, DX + ANDQ R10, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R9 + ADDQ AX, R13 + ADDQ AX, R9 + MOVQ 232(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 128(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 192(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 120(BP), BX + ADDQ BX, AX + MOVQ AX, 248(BP) + MOVQ $0x142929670a0e6e70, DX + ADDQ AX, R8 + MOVQ R13, AX + ADDQ DX, R8 + MOVQ R13, CX + RORQ $0x0e, AX + MOVQ R13, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R13, CX + RORQ $0x29, DX + ANDQ R14, CX + XORQ AX, DX + MOVQ R13, AX + NOTQ AX + ADDQ DX, R8 + ANDQ R15, AX + XORQ CX, AX + ADDQ R8, AX + MOVQ R9, DI + MOVQ R11, BX + RORQ $0x1c, DI + MOVQ R9, DX + ANDQ R10, BX + RORQ $0x22, DX + MOVQ R9, CX + ANDQ R11, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R9, DX + MOVQ R10, CX + RORQ $0x27, DX + ANDQ R9, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R8 + ADDQ AX, R12 + ADDQ AX, R8 + MOVQ 240(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 136(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 200(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 128(BP), BX + ADDQ BX, AX + MOVQ AX, 256(BP) + MOVQ $0x27b70a8546d22ffc, DX + ADDQ AX, R15 + MOVQ R12, AX + ADDQ DX, R15 + MOVQ R12, CX + RORQ $0x0e, AX + MOVQ R12, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R12, CX + RORQ $0x29, DX + ANDQ R13, CX + XORQ AX, DX + MOVQ R12, AX + NOTQ AX + ADDQ DX, R15 + ANDQ R14, AX + XORQ CX, AX + ADDQ R15, AX + MOVQ R8, DI + MOVQ R10, BX + RORQ $0x1c, DI + MOVQ R8, DX + ANDQ R9, BX + RORQ $0x22, DX + MOVQ R8, CX + ANDQ R10, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R8, DX + MOVQ R9, CX + RORQ $0x27, DX + ANDQ R8, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R15 + ADDQ AX, R11 + ADDQ AX, R15 + MOVQ 248(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 144(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 208(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 136(BP), BX + ADDQ BX, AX + MOVQ AX, 264(BP) + MOVQ $0x2e1b21385c26c926, DX + ADDQ AX, R14 + MOVQ R11, AX + ADDQ DX, R14 + MOVQ R11, CX + RORQ $0x0e, AX + MOVQ R11, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R11, CX + RORQ $0x29, DX + ANDQ R12, CX + XORQ AX, DX + MOVQ R11, AX + NOTQ AX + ADDQ DX, R14 + ANDQ R13, AX + XORQ CX, AX + ADDQ R14, AX + MOVQ R15, DI + MOVQ R9, BX + RORQ $0x1c, DI + MOVQ R15, DX + ANDQ R8, BX + RORQ $0x22, DX + MOVQ R15, CX + ANDQ R9, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R15, DX + MOVQ R8, CX + RORQ $0x27, DX + ANDQ R15, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R14 + ADDQ AX, R10 + ADDQ AX, R14 + MOVQ 256(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 152(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 216(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 144(BP), BX + ADDQ BX, AX + MOVQ AX, 272(BP) + MOVQ $0x4d2c6dfc5ac42aed, DX + ADDQ AX, R13 + MOVQ R10, AX + ADDQ DX, R13 + MOVQ R10, CX + RORQ $0x0e, AX + MOVQ R10, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R10, CX + RORQ $0x29, DX + ANDQ R11, CX + XORQ AX, DX + MOVQ R10, AX + NOTQ AX + ADDQ DX, R13 + ANDQ R12, AX + XORQ CX, AX + ADDQ R13, AX + MOVQ R14, DI + MOVQ R8, BX + RORQ $0x1c, DI + MOVQ R14, DX + ANDQ R15, BX + RORQ $0x22, DX + MOVQ R14, CX + ANDQ R8, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R14, DX + MOVQ R15, CX + RORQ $0x27, DX + ANDQ R14, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R13 + ADDQ AX, R9 + ADDQ AX, R13 + MOVQ 264(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 160(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 224(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 152(BP), BX + ADDQ BX, AX + MOVQ AX, 280(BP) + MOVQ $0x53380d139d95b3df, DX + ADDQ AX, R12 + MOVQ R9, AX + ADDQ DX, R12 + MOVQ R9, CX + RORQ $0x0e, AX + MOVQ R9, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R9, CX + RORQ $0x29, DX + ANDQ R10, CX + XORQ AX, DX + MOVQ R9, AX + NOTQ AX + ADDQ DX, R12 + ANDQ R11, AX + XORQ CX, AX + ADDQ R12, AX + MOVQ R13, DI + MOVQ R15, BX + RORQ $0x1c, DI + MOVQ R13, DX + ANDQ R14, BX + RORQ $0x22, DX + MOVQ R13, CX + ANDQ R15, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R13, DX + MOVQ R14, CX + RORQ $0x27, DX + ANDQ R13, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R12 + ADDQ AX, R8 + ADDQ AX, R12 + MOVQ 272(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 168(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 232(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 160(BP), BX + ADDQ BX, AX + MOVQ AX, 288(BP) + MOVQ $0x650a73548baf63de, DX + ADDQ AX, R11 + MOVQ R8, AX + ADDQ DX, R11 + MOVQ R8, CX + RORQ $0x0e, AX + MOVQ R8, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R8, CX + RORQ $0x29, DX + ANDQ R9, CX + XORQ AX, DX + MOVQ R8, AX + NOTQ AX + ADDQ DX, R11 + ANDQ R10, AX + XORQ CX, AX + ADDQ R11, AX + MOVQ R12, DI + MOVQ R14, BX + RORQ $0x1c, DI + MOVQ R12, DX + ANDQ R13, BX + RORQ $0x22, DX + MOVQ R12, CX + ANDQ R14, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R12, DX + MOVQ R13, CX + RORQ $0x27, DX + ANDQ R12, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R11 + ADDQ AX, R15 + ADDQ AX, R11 + MOVQ 280(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 176(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 240(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 168(BP), BX + ADDQ BX, AX + MOVQ AX, 296(BP) + MOVQ $0x766a0abb3c77b2a8, DX + ADDQ AX, R10 + MOVQ R15, AX + ADDQ DX, R10 + MOVQ R15, CX + RORQ $0x0e, AX + MOVQ R15, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R15, CX + RORQ $0x29, DX + ANDQ R8, CX + XORQ AX, DX + MOVQ R15, AX + NOTQ AX + ADDQ DX, R10 + ANDQ R9, AX + XORQ CX, AX + ADDQ R10, AX + MOVQ R11, DI + MOVQ R13, BX + RORQ $0x1c, DI + MOVQ R11, DX + ANDQ R12, BX + RORQ $0x22, DX + MOVQ R11, CX + ANDQ R13, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R11, DX + MOVQ R12, CX + RORQ $0x27, DX + ANDQ R11, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R10 + ADDQ AX, R14 + ADDQ AX, R10 + MOVQ 288(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 184(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 248(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 176(BP), BX + ADDQ BX, AX + MOVQ AX, 304(BP) + MOVQ $0x81c2c92e47edaee6, DX + ADDQ AX, R9 + MOVQ R14, AX + ADDQ DX, R9 + MOVQ R14, CX + RORQ $0x0e, AX + MOVQ R14, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R14, CX + RORQ $0x29, DX + ANDQ R15, CX + XORQ AX, DX + MOVQ R14, AX + NOTQ AX + ADDQ DX, R9 + ANDQ R8, AX + XORQ CX, AX + ADDQ R9, AX + MOVQ R10, DI + MOVQ R12, BX + RORQ $0x1c, DI + MOVQ R10, DX + ANDQ R11, BX + RORQ $0x22, DX + MOVQ R10, CX + ANDQ R12, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R10, DX + MOVQ R11, CX + RORQ $0x27, DX + ANDQ R10, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R9 + ADDQ AX, R13 + ADDQ AX, R9 + MOVQ 296(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 192(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 256(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 184(BP), BX + ADDQ BX, AX + MOVQ AX, 312(BP) + MOVQ $0x92722c851482353b, DX + ADDQ AX, R8 + MOVQ R13, AX + ADDQ DX, R8 + MOVQ R13, CX + RORQ $0x0e, AX + MOVQ R13, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R13, CX + RORQ $0x29, DX + ANDQ R14, CX + XORQ AX, DX + MOVQ R13, AX + NOTQ AX + ADDQ DX, R8 + ANDQ R15, AX + XORQ CX, AX + ADDQ R8, AX + MOVQ R9, DI + MOVQ R11, BX + RORQ $0x1c, DI + MOVQ R9, DX + ANDQ R10, BX + RORQ $0x22, DX + MOVQ R9, CX + ANDQ R11, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R9, DX + MOVQ R10, CX + RORQ $0x27, DX + ANDQ R9, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R8 + ADDQ AX, R12 + ADDQ AX, R8 + MOVQ 304(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 200(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 264(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 192(BP), BX + ADDQ BX, AX + MOVQ AX, 320(BP) + MOVQ $0xa2bfe8a14cf10364, DX + ADDQ AX, R15 + MOVQ R12, AX + ADDQ DX, R15 + MOVQ R12, CX + RORQ $0x0e, AX + MOVQ R12, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R12, CX + RORQ $0x29, DX + ANDQ R13, CX + XORQ AX, DX + MOVQ R12, AX + NOTQ AX + ADDQ DX, R15 + ANDQ R14, AX + XORQ CX, AX + ADDQ R15, AX + MOVQ R8, DI + MOVQ R10, BX + RORQ $0x1c, DI + MOVQ R8, DX + ANDQ R9, BX + RORQ $0x22, DX + MOVQ R8, CX + ANDQ R10, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R8, DX + MOVQ R9, CX + RORQ $0x27, DX + ANDQ R8, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R15 + ADDQ AX, R11 + ADDQ AX, R15 + MOVQ 312(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 208(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 272(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 200(BP), BX + ADDQ BX, AX + MOVQ AX, 328(BP) + MOVQ $0xa81a664bbc423001, DX + ADDQ AX, R14 + MOVQ R11, AX + ADDQ DX, R14 + MOVQ R11, CX + RORQ $0x0e, AX + MOVQ R11, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R11, CX + RORQ $0x29, DX + ANDQ R12, CX + XORQ AX, DX + MOVQ R11, AX + NOTQ AX + ADDQ DX, R14 + ANDQ R13, AX + XORQ CX, AX + ADDQ R14, AX + MOVQ R15, DI + MOVQ R9, BX + RORQ $0x1c, DI + MOVQ R15, DX + ANDQ R8, BX + RORQ $0x22, DX + MOVQ R15, CX + ANDQ R9, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R15, DX + MOVQ R8, CX + RORQ $0x27, DX + ANDQ R15, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R14 + ADDQ AX, R10 + ADDQ AX, R14 + MOVQ 320(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 216(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 280(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 208(BP), BX + ADDQ BX, AX + MOVQ AX, 336(BP) + MOVQ $0xc24b8b70d0f89791, DX + ADDQ AX, R13 + MOVQ R10, AX + ADDQ DX, R13 + MOVQ R10, CX + RORQ $0x0e, AX + MOVQ R10, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R10, CX + RORQ $0x29, DX + ANDQ R11, CX + XORQ AX, DX + MOVQ R10, AX + NOTQ AX + ADDQ DX, R13 + ANDQ R12, AX + XORQ CX, AX + ADDQ R13, AX + MOVQ R14, DI + MOVQ R8, BX + RORQ $0x1c, DI + MOVQ R14, DX + ANDQ R15, BX + RORQ $0x22, DX + MOVQ R14, CX + ANDQ R8, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R14, DX + MOVQ R15, CX + RORQ $0x27, DX + ANDQ R14, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R13 + ADDQ AX, R9 + ADDQ AX, R13 + MOVQ 328(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 224(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 288(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 216(BP), BX + ADDQ BX, AX + MOVQ AX, 344(BP) + MOVQ $0xc76c51a30654be30, DX + ADDQ AX, R12 + MOVQ R9, AX + ADDQ DX, R12 + MOVQ R9, CX + RORQ $0x0e, AX + MOVQ R9, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R9, CX + RORQ $0x29, DX + ANDQ R10, CX + XORQ AX, DX + MOVQ R9, AX + NOTQ AX + ADDQ DX, R12 + ANDQ R11, AX + XORQ CX, AX + ADDQ R12, AX + MOVQ R13, DI + MOVQ R15, BX + RORQ $0x1c, DI + MOVQ R13, DX + ANDQ R14, BX + RORQ $0x22, DX + MOVQ R13, CX + ANDQ R15, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R13, DX + MOVQ R14, CX + RORQ $0x27, DX + ANDQ R13, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R12 + ADDQ AX, R8 + ADDQ AX, R12 + MOVQ 336(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 232(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 296(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 224(BP), BX + ADDQ BX, AX + MOVQ AX, 352(BP) + MOVQ $0xd192e819d6ef5218, DX + ADDQ AX, R11 + MOVQ R8, AX + ADDQ DX, R11 + MOVQ R8, CX + RORQ $0x0e, AX + MOVQ R8, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R8, CX + RORQ $0x29, DX + ANDQ R9, CX + XORQ AX, DX + MOVQ R8, AX + NOTQ AX + ADDQ DX, R11 + ANDQ R10, AX + XORQ CX, AX + ADDQ R11, AX + MOVQ R12, DI + MOVQ R14, BX + RORQ $0x1c, DI + MOVQ R12, DX + ANDQ R13, BX + RORQ $0x22, DX + MOVQ R12, CX + ANDQ R14, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R12, DX + MOVQ R13, CX + RORQ $0x27, DX + ANDQ R12, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R11 + ADDQ AX, R15 + ADDQ AX, R11 + MOVQ 344(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 240(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 304(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 232(BP), BX + ADDQ BX, AX + MOVQ AX, 360(BP) + MOVQ $0xd69906245565a910, DX + ADDQ AX, R10 + MOVQ R15, AX + ADDQ DX, R10 + MOVQ R15, CX + RORQ $0x0e, AX + MOVQ R15, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R15, CX + RORQ $0x29, DX + ANDQ R8, CX + XORQ AX, DX + MOVQ R15, AX + NOTQ AX + ADDQ DX, R10 + ANDQ R9, AX + XORQ CX, AX + ADDQ R10, AX + MOVQ R11, DI + MOVQ R13, BX + RORQ $0x1c, DI + MOVQ R11, DX + ANDQ R12, BX + RORQ $0x22, DX + MOVQ R11, CX + ANDQ R13, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R11, DX + MOVQ R12, CX + RORQ $0x27, DX + ANDQ R11, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R10 + ADDQ AX, R14 + ADDQ AX, R10 + MOVQ 352(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 248(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 312(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 240(BP), BX + ADDQ BX, AX + MOVQ AX, 368(BP) + MOVQ $0xf40e35855771202a, DX + ADDQ AX, R9 + MOVQ R14, AX + ADDQ DX, R9 + MOVQ R14, CX + RORQ $0x0e, AX + MOVQ R14, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R14, CX + RORQ $0x29, DX + ANDQ R15, CX + XORQ AX, DX + MOVQ R14, AX + NOTQ AX + ADDQ DX, R9 + ANDQ R8, AX + XORQ CX, AX + ADDQ R9, AX + MOVQ R10, DI + MOVQ R12, BX + RORQ $0x1c, DI + MOVQ R10, DX + ANDQ R11, BX + RORQ $0x22, DX + MOVQ R10, CX + ANDQ R12, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R10, DX + MOVQ R11, CX + RORQ $0x27, DX + ANDQ R10, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R9 + ADDQ AX, R13 + ADDQ AX, R9 + MOVQ 360(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 256(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 320(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 248(BP), BX + ADDQ BX, AX + MOVQ AX, 376(BP) + MOVQ $0x106aa07032bbd1b8, DX + ADDQ AX, R8 + MOVQ R13, AX + ADDQ DX, R8 + MOVQ R13, CX + RORQ $0x0e, AX + MOVQ R13, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R13, CX + RORQ $0x29, DX + ANDQ R14, CX + XORQ AX, DX + MOVQ R13, AX + NOTQ AX + ADDQ DX, R8 + ANDQ R15, AX + XORQ CX, AX + ADDQ R8, AX + MOVQ R9, DI + MOVQ R11, BX + RORQ $0x1c, DI + MOVQ R9, DX + ANDQ R10, BX + RORQ $0x22, DX + MOVQ R9, CX + ANDQ R11, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R9, DX + MOVQ R10, CX + RORQ $0x27, DX + ANDQ R9, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R8 + ADDQ AX, R12 + ADDQ AX, R8 + MOVQ 368(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 264(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 328(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 256(BP), BX + ADDQ BX, AX + MOVQ AX, 384(BP) + MOVQ $0x19a4c116b8d2d0c8, DX + ADDQ AX, R15 + MOVQ R12, AX + ADDQ DX, R15 + MOVQ R12, CX + RORQ $0x0e, AX + MOVQ R12, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R12, CX + RORQ $0x29, DX + ANDQ R13, CX + XORQ AX, DX + MOVQ R12, AX + NOTQ AX + ADDQ DX, R15 + ANDQ R14, AX + XORQ CX, AX + ADDQ R15, AX + MOVQ R8, DI + MOVQ R10, BX + RORQ $0x1c, DI + MOVQ R8, DX + ANDQ R9, BX + RORQ $0x22, DX + MOVQ R8, CX + ANDQ R10, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R8, DX + MOVQ R9, CX + RORQ $0x27, DX + ANDQ R8, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R15 + ADDQ AX, R11 + ADDQ AX, R15 + MOVQ 376(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 272(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 336(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 264(BP), BX + ADDQ BX, AX + MOVQ AX, 392(BP) + MOVQ $0x1e376c085141ab53, DX + ADDQ AX, R14 + MOVQ R11, AX + ADDQ DX, R14 + MOVQ R11, CX + RORQ $0x0e, AX + MOVQ R11, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R11, CX + RORQ $0x29, DX + ANDQ R12, CX + XORQ AX, DX + MOVQ R11, AX + NOTQ AX + ADDQ DX, R14 + ANDQ R13, AX + XORQ CX, AX + ADDQ R14, AX + MOVQ R15, DI + MOVQ R9, BX + RORQ $0x1c, DI + MOVQ R15, DX + ANDQ R8, BX + RORQ $0x22, DX + MOVQ R15, CX + ANDQ R9, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R15, DX + MOVQ R8, CX + RORQ $0x27, DX + ANDQ R15, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R14 + ADDQ AX, R10 + ADDQ AX, R14 + MOVQ 384(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 280(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 344(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 272(BP), BX + ADDQ BX, AX + MOVQ AX, 400(BP) + MOVQ $0x2748774cdf8eeb99, DX + ADDQ AX, R13 + MOVQ R10, AX + ADDQ DX, R13 + MOVQ R10, CX + RORQ $0x0e, AX + MOVQ R10, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R10, CX + RORQ $0x29, DX + ANDQ R11, CX + XORQ AX, DX + MOVQ R10, AX + NOTQ AX + ADDQ DX, R13 + ANDQ R12, AX + XORQ CX, AX + ADDQ R13, AX + MOVQ R14, DI + MOVQ R8, BX + RORQ $0x1c, DI + MOVQ R14, DX + ANDQ R15, BX + RORQ $0x22, DX + MOVQ R14, CX + ANDQ R8, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R14, DX + MOVQ R15, CX + RORQ $0x27, DX + ANDQ R14, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R13 + ADDQ AX, R9 + ADDQ AX, R13 + MOVQ 392(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 288(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 352(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 280(BP), BX + ADDQ BX, AX + MOVQ AX, 408(BP) + MOVQ $0x34b0bcb5e19b48a8, DX + ADDQ AX, R12 + MOVQ R9, AX + ADDQ DX, R12 + MOVQ R9, CX + RORQ $0x0e, AX + MOVQ R9, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R9, CX + RORQ $0x29, DX + ANDQ R10, CX + XORQ AX, DX + MOVQ R9, AX + NOTQ AX + ADDQ DX, R12 + ANDQ R11, AX + XORQ CX, AX + ADDQ R12, AX + MOVQ R13, DI + MOVQ R15, BX + RORQ $0x1c, DI + MOVQ R13, DX + ANDQ R14, BX + RORQ $0x22, DX + MOVQ R13, CX + ANDQ R15, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R13, DX + MOVQ R14, CX + RORQ $0x27, DX + ANDQ R13, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R12 + ADDQ AX, R8 + ADDQ AX, R12 + MOVQ 400(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 296(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 360(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 288(BP), BX + ADDQ BX, AX + MOVQ AX, 416(BP) + MOVQ $0x391c0cb3c5c95a63, DX + ADDQ AX, R11 + MOVQ R8, AX + ADDQ DX, R11 + MOVQ R8, CX + RORQ $0x0e, AX + MOVQ R8, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R8, CX + RORQ $0x29, DX + ANDQ R9, CX + XORQ AX, DX + MOVQ R8, AX + NOTQ AX + ADDQ DX, R11 + ANDQ R10, AX + XORQ CX, AX + ADDQ R11, AX + MOVQ R12, DI + MOVQ R14, BX + RORQ $0x1c, DI + MOVQ R12, DX + ANDQ R13, BX + RORQ $0x22, DX + MOVQ R12, CX + ANDQ R14, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R12, DX + MOVQ R13, CX + RORQ $0x27, DX + ANDQ R12, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R11 + ADDQ AX, R15 + ADDQ AX, R11 + MOVQ 408(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 304(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 368(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 296(BP), BX + ADDQ BX, AX + MOVQ AX, 424(BP) + MOVQ $0x4ed8aa4ae3418acb, DX + ADDQ AX, R10 + MOVQ R15, AX + ADDQ DX, R10 + MOVQ R15, CX + RORQ $0x0e, AX + MOVQ R15, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R15, CX + RORQ $0x29, DX + ANDQ R8, CX + XORQ AX, DX + MOVQ R15, AX + NOTQ AX + ADDQ DX, R10 + ANDQ R9, AX + XORQ CX, AX + ADDQ R10, AX + MOVQ R11, DI + MOVQ R13, BX + RORQ $0x1c, DI + MOVQ R11, DX + ANDQ R12, BX + RORQ $0x22, DX + MOVQ R11, CX + ANDQ R13, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R11, DX + MOVQ R12, CX + RORQ $0x27, DX + ANDQ R11, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R10 + ADDQ AX, R14 + ADDQ AX, R10 + MOVQ 416(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 312(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 376(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 304(BP), BX + ADDQ BX, AX + MOVQ AX, 432(BP) + MOVQ $0x5b9cca4f7763e373, DX + ADDQ AX, R9 + MOVQ R14, AX + ADDQ DX, R9 + MOVQ R14, CX + RORQ $0x0e, AX + MOVQ R14, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R14, CX + RORQ $0x29, DX + ANDQ R15, CX + XORQ AX, DX + MOVQ R14, AX + NOTQ AX + ADDQ DX, R9 + ANDQ R8, AX + XORQ CX, AX + ADDQ R9, AX + MOVQ R10, DI + MOVQ R12, BX + RORQ $0x1c, DI + MOVQ R10, DX + ANDQ R11, BX + RORQ $0x22, DX + MOVQ R10, CX + ANDQ R12, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R10, DX + MOVQ R11, CX + RORQ $0x27, DX + ANDQ R10, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R9 + ADDQ AX, R13 + ADDQ AX, R9 + MOVQ 424(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 320(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 384(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 312(BP), BX + ADDQ BX, AX + MOVQ AX, 440(BP) + MOVQ $0x682e6ff3d6b2b8a3, DX + ADDQ AX, R8 + MOVQ R13, AX + ADDQ DX, R8 + MOVQ R13, CX + RORQ $0x0e, AX + MOVQ R13, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R13, CX + RORQ $0x29, DX + ANDQ R14, CX + XORQ AX, DX + MOVQ R13, AX + NOTQ AX + ADDQ DX, R8 + ANDQ R15, AX + XORQ CX, AX + ADDQ R8, AX + MOVQ R9, DI + MOVQ R11, BX + RORQ $0x1c, DI + MOVQ R9, DX + ANDQ R10, BX + RORQ $0x22, DX + MOVQ R9, CX + ANDQ R11, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R9, DX + MOVQ R10, CX + RORQ $0x27, DX + ANDQ R9, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R8 + ADDQ AX, R12 + ADDQ AX, R8 + MOVQ 432(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 328(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 392(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 320(BP), BX + ADDQ BX, AX + MOVQ AX, 448(BP) + MOVQ $0x748f82ee5defb2fc, DX + ADDQ AX, R15 + MOVQ R12, AX + ADDQ DX, R15 + MOVQ R12, CX + RORQ $0x0e, AX + MOVQ R12, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R12, CX + RORQ $0x29, DX + ANDQ R13, CX + XORQ AX, DX + MOVQ R12, AX + NOTQ AX + ADDQ DX, R15 + ANDQ R14, AX + XORQ CX, AX + ADDQ R15, AX + MOVQ R8, DI + MOVQ R10, BX + RORQ $0x1c, DI + MOVQ R8, DX + ANDQ R9, BX + RORQ $0x22, DX + MOVQ R8, CX + ANDQ R10, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R8, DX + MOVQ R9, CX + RORQ $0x27, DX + ANDQ R8, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R15 + ADDQ AX, R11 + ADDQ AX, R15 + MOVQ 440(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 336(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 400(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 328(BP), BX + ADDQ BX, AX + MOVQ AX, 456(BP) + MOVQ $0x78a5636f43172f60, DX + ADDQ AX, R14 + MOVQ R11, AX + ADDQ DX, R14 + MOVQ R11, CX + RORQ $0x0e, AX + MOVQ R11, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R11, CX + RORQ $0x29, DX + ANDQ R12, CX + XORQ AX, DX + MOVQ R11, AX + NOTQ AX + ADDQ DX, R14 + ANDQ R13, AX + XORQ CX, AX + ADDQ R14, AX + MOVQ R15, DI + MOVQ R9, BX + RORQ $0x1c, DI + MOVQ R15, DX + ANDQ R8, BX + RORQ $0x22, DX + MOVQ R15, CX + ANDQ R9, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R15, DX + MOVQ R8, CX + RORQ $0x27, DX + ANDQ R15, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R14 + ADDQ AX, R10 + ADDQ AX, R14 + MOVQ 448(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 344(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 408(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 336(BP), BX + ADDQ BX, AX + MOVQ AX, 464(BP) + MOVQ $0x84c87814a1f0ab72, DX + ADDQ AX, R13 + MOVQ R10, AX + ADDQ DX, R13 + MOVQ R10, CX + RORQ $0x0e, AX + MOVQ R10, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R10, CX + RORQ $0x29, DX + ANDQ R11, CX + XORQ AX, DX + MOVQ R10, AX + NOTQ AX + ADDQ DX, R13 + ANDQ R12, AX + XORQ CX, AX + ADDQ R13, AX + MOVQ R14, DI + MOVQ R8, BX + RORQ $0x1c, DI + MOVQ R14, DX + ANDQ R15, BX + RORQ $0x22, DX + MOVQ R14, CX + ANDQ R8, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R14, DX + MOVQ R15, CX + RORQ $0x27, DX + ANDQ R14, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R13 + ADDQ AX, R9 + ADDQ AX, R13 + MOVQ 456(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 352(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 416(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 344(BP), BX + ADDQ BX, AX + MOVQ AX, 472(BP) + MOVQ $0x8cc702081a6439ec, DX + ADDQ AX, R12 + MOVQ R9, AX + ADDQ DX, R12 + MOVQ R9, CX + RORQ $0x0e, AX + MOVQ R9, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R9, CX + RORQ $0x29, DX + ANDQ R10, CX + XORQ AX, DX + MOVQ R9, AX + NOTQ AX + ADDQ DX, R12 + ANDQ R11, AX + XORQ CX, AX + ADDQ R12, AX + MOVQ R13, DI + MOVQ R15, BX + RORQ $0x1c, DI + MOVQ R13, DX + ANDQ R14, BX + RORQ $0x22, DX + MOVQ R13, CX + ANDQ R15, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R13, DX + MOVQ R14, CX + RORQ $0x27, DX + ANDQ R13, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R12 + ADDQ AX, R8 + ADDQ AX, R12 + MOVQ 464(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 360(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 424(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 352(BP), BX + ADDQ BX, AX + MOVQ AX, 480(BP) + MOVQ $0x90befffa23631e28, DX + ADDQ AX, R11 + MOVQ R8, AX + ADDQ DX, R11 + MOVQ R8, CX + RORQ $0x0e, AX + MOVQ R8, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R8, CX + RORQ $0x29, DX + ANDQ R9, CX + XORQ AX, DX + MOVQ R8, AX + NOTQ AX + ADDQ DX, R11 + ANDQ R10, AX + XORQ CX, AX + ADDQ R11, AX + MOVQ R12, DI + MOVQ R14, BX + RORQ $0x1c, DI + MOVQ R12, DX + ANDQ R13, BX + RORQ $0x22, DX + MOVQ R12, CX + ANDQ R14, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R12, DX + MOVQ R13, CX + RORQ $0x27, DX + ANDQ R12, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R11 + ADDQ AX, R15 + ADDQ AX, R11 + MOVQ 472(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 368(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 432(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 360(BP), BX + ADDQ BX, AX + MOVQ AX, 488(BP) + MOVQ $0xa4506cebde82bde9, DX + ADDQ AX, R10 + MOVQ R15, AX + ADDQ DX, R10 + MOVQ R15, CX + RORQ $0x0e, AX + MOVQ R15, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R15, CX + RORQ $0x29, DX + ANDQ R8, CX + XORQ AX, DX + MOVQ R15, AX + NOTQ AX + ADDQ DX, R10 + ANDQ R9, AX + XORQ CX, AX + ADDQ R10, AX + MOVQ R11, DI + MOVQ R13, BX + RORQ $0x1c, DI + MOVQ R11, DX + ANDQ R12, BX + RORQ $0x22, DX + MOVQ R11, CX + ANDQ R13, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R11, DX + MOVQ R12, CX + RORQ $0x27, DX + ANDQ R11, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R10 + ADDQ AX, R14 + ADDQ AX, R10 + MOVQ 480(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 376(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 440(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 368(BP), BX + ADDQ BX, AX + MOVQ AX, 496(BP) + MOVQ $0xbef9a3f7b2c67915, DX + ADDQ AX, R9 + MOVQ R14, AX + ADDQ DX, R9 + MOVQ R14, CX + RORQ $0x0e, AX + MOVQ R14, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R14, CX + RORQ $0x29, DX + ANDQ R15, CX + XORQ AX, DX + MOVQ R14, AX + NOTQ AX + ADDQ DX, R9 + ANDQ R8, AX + XORQ CX, AX + ADDQ R9, AX + MOVQ R10, DI + MOVQ R12, BX + RORQ $0x1c, DI + MOVQ R10, DX + ANDQ R11, BX + RORQ $0x22, DX + MOVQ R10, CX + ANDQ R12, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R10, DX + MOVQ R11, CX + RORQ $0x27, DX + ANDQ R10, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R9 + ADDQ AX, R13 + ADDQ AX, R9 + MOVQ 488(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 384(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 448(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 376(BP), BX + ADDQ BX, AX + MOVQ AX, 504(BP) + MOVQ $0xc67178f2e372532b, DX + ADDQ AX, R8 + MOVQ R13, AX + ADDQ DX, R8 + MOVQ R13, CX + RORQ $0x0e, AX + MOVQ R13, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R13, CX + RORQ $0x29, DX + ANDQ R14, CX + XORQ AX, DX + MOVQ R13, AX + NOTQ AX + ADDQ DX, R8 + ANDQ R15, AX + XORQ CX, AX + ADDQ R8, AX + MOVQ R9, DI + MOVQ R11, BX + RORQ $0x1c, DI + MOVQ R9, DX + ANDQ R10, BX + RORQ $0x22, DX + MOVQ R9, CX + ANDQ R11, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R9, DX + MOVQ R10, CX + RORQ $0x27, DX + ANDQ R9, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R8 + ADDQ AX, R12 + ADDQ AX, R8 + MOVQ 496(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 392(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 456(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 384(BP), BX + ADDQ BX, AX + MOVQ AX, 512(BP) + MOVQ $0xca273eceea26619c, DX + ADDQ AX, R15 + MOVQ R12, AX + ADDQ DX, R15 + MOVQ R12, CX + RORQ $0x0e, AX + MOVQ R12, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R12, CX + RORQ $0x29, DX + ANDQ R13, CX + XORQ AX, DX + MOVQ R12, AX + NOTQ AX + ADDQ DX, R15 + ANDQ R14, AX + XORQ CX, AX + ADDQ R15, AX + MOVQ R8, DI + MOVQ R10, BX + RORQ $0x1c, DI + MOVQ R8, DX + ANDQ R9, BX + RORQ $0x22, DX + MOVQ R8, CX + ANDQ R10, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R8, DX + MOVQ R9, CX + RORQ $0x27, DX + ANDQ R8, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R15 + ADDQ AX, R11 + ADDQ AX, R15 + MOVQ 504(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 400(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 464(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 392(BP), BX + ADDQ BX, AX + MOVQ AX, 520(BP) + MOVQ $0xd186b8c721c0c207, DX + ADDQ AX, R14 + MOVQ R11, AX + ADDQ DX, R14 + MOVQ R11, CX + RORQ $0x0e, AX + MOVQ R11, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R11, CX + RORQ $0x29, DX + ANDQ R12, CX + XORQ AX, DX + MOVQ R11, AX + NOTQ AX + ADDQ DX, R14 + ANDQ R13, AX + XORQ CX, AX + ADDQ R14, AX + MOVQ R15, DI + MOVQ R9, BX + RORQ $0x1c, DI + MOVQ R15, DX + ANDQ R8, BX + RORQ $0x22, DX + MOVQ R15, CX + ANDQ R9, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R15, DX + MOVQ R8, CX + RORQ $0x27, DX + ANDQ R15, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R14 + ADDQ AX, R10 + ADDQ AX, R14 + MOVQ 512(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 408(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 472(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 400(BP), BX + ADDQ BX, AX + MOVQ AX, 528(BP) + MOVQ $0xeada7dd6cde0eb1e, DX + ADDQ AX, R13 + MOVQ R10, AX + ADDQ DX, R13 + MOVQ R10, CX + RORQ $0x0e, AX + MOVQ R10, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R10, CX + RORQ $0x29, DX + ANDQ R11, CX + XORQ AX, DX + MOVQ R10, AX + NOTQ AX + ADDQ DX, R13 + ANDQ R12, AX + XORQ CX, AX + ADDQ R13, AX + MOVQ R14, DI + MOVQ R8, BX + RORQ $0x1c, DI + MOVQ R14, DX + ANDQ R15, BX + RORQ $0x22, DX + MOVQ R14, CX + ANDQ R8, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R14, DX + MOVQ R15, CX + RORQ $0x27, DX + ANDQ R14, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R13 + ADDQ AX, R9 + ADDQ AX, R13 + MOVQ 520(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 416(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 480(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 408(BP), BX + ADDQ BX, AX + MOVQ AX, 536(BP) + MOVQ $0xf57d4f7fee6ed178, DX + ADDQ AX, R12 + MOVQ R9, AX + ADDQ DX, R12 + MOVQ R9, CX + RORQ $0x0e, AX + MOVQ R9, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R9, CX + RORQ $0x29, DX + ANDQ R10, CX + XORQ AX, DX + MOVQ R9, AX + NOTQ AX + ADDQ DX, R12 + ANDQ R11, AX + XORQ CX, AX + ADDQ R12, AX + MOVQ R13, DI + MOVQ R15, BX + RORQ $0x1c, DI + MOVQ R13, DX + ANDQ R14, BX + RORQ $0x22, DX + MOVQ R13, CX + ANDQ R15, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R13, DX + MOVQ R14, CX + RORQ $0x27, DX + ANDQ R13, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R12 + ADDQ AX, R8 + ADDQ AX, R12 + MOVQ 528(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 424(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 488(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 416(BP), BX + ADDQ BX, AX + MOVQ AX, 544(BP) + MOVQ $0x06f067aa72176fba, DX + ADDQ AX, R11 + MOVQ R8, AX + ADDQ DX, R11 + MOVQ R8, CX + RORQ $0x0e, AX + MOVQ R8, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R8, CX + RORQ $0x29, DX + ANDQ R9, CX + XORQ AX, DX + MOVQ R8, AX + NOTQ AX + ADDQ DX, R11 + ANDQ R10, AX + XORQ CX, AX + ADDQ R11, AX + MOVQ R12, DI + MOVQ R14, BX + RORQ $0x1c, DI + MOVQ R12, DX + ANDQ R13, BX + RORQ $0x22, DX + MOVQ R12, CX + ANDQ R14, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R12, DX + MOVQ R13, CX + RORQ $0x27, DX + ANDQ R12, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R11 + ADDQ AX, R15 + ADDQ AX, R11 + MOVQ 536(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 432(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 496(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 424(BP), BX + ADDQ BX, AX + MOVQ AX, 552(BP) + MOVQ $0x0a637dc5a2c898a6, DX + ADDQ AX, R10 + MOVQ R15, AX + ADDQ DX, R10 + MOVQ R15, CX + RORQ $0x0e, AX + MOVQ R15, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R15, CX + RORQ $0x29, DX + ANDQ R8, CX + XORQ AX, DX + MOVQ R15, AX + NOTQ AX + ADDQ DX, R10 + ANDQ R9, AX + XORQ CX, AX + ADDQ R10, AX + MOVQ R11, DI + MOVQ R13, BX + RORQ $0x1c, DI + MOVQ R11, DX + ANDQ R12, BX + RORQ $0x22, DX + MOVQ R11, CX + ANDQ R13, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R11, DX + MOVQ R12, CX + RORQ $0x27, DX + ANDQ R11, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R10 + ADDQ AX, R14 + ADDQ AX, R10 + MOVQ 544(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 440(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 504(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 432(BP), BX + ADDQ BX, AX + MOVQ AX, 560(BP) + MOVQ $0x113f9804bef90dae, DX + ADDQ AX, R9 + MOVQ R14, AX + ADDQ DX, R9 + MOVQ R14, CX + RORQ $0x0e, AX + MOVQ R14, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R14, CX + RORQ $0x29, DX + ANDQ R15, CX + XORQ AX, DX + MOVQ R14, AX + NOTQ AX + ADDQ DX, R9 + ANDQ R8, AX + XORQ CX, AX + ADDQ R9, AX + MOVQ R10, DI + MOVQ R12, BX + RORQ $0x1c, DI + MOVQ R10, DX + ANDQ R11, BX + RORQ $0x22, DX + MOVQ R10, CX + ANDQ R12, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R10, DX + MOVQ R11, CX + RORQ $0x27, DX + ANDQ R10, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R9 + ADDQ AX, R13 + ADDQ AX, R9 + MOVQ 552(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 448(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 512(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 440(BP), BX + ADDQ BX, AX + MOVQ AX, 568(BP) + MOVQ $0x1b710b35131c471b, DX + ADDQ AX, R8 + MOVQ R13, AX + ADDQ DX, R8 + MOVQ R13, CX + RORQ $0x0e, AX + MOVQ R13, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R13, CX + RORQ $0x29, DX + ANDQ R14, CX + XORQ AX, DX + MOVQ R13, AX + NOTQ AX + ADDQ DX, R8 + ANDQ R15, AX + XORQ CX, AX + ADDQ R8, AX + MOVQ R9, DI + MOVQ R11, BX + RORQ $0x1c, DI + MOVQ R9, DX + ANDQ R10, BX + RORQ $0x22, DX + MOVQ R9, CX + ANDQ R11, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R9, DX + MOVQ R10, CX + RORQ $0x27, DX + ANDQ R9, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R8 + ADDQ AX, R12 + ADDQ AX, R8 + MOVQ 560(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 456(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 520(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 448(BP), BX + ADDQ BX, AX + MOVQ AX, 576(BP) + MOVQ $0x28db77f523047d84, DX + ADDQ AX, R15 + MOVQ R12, AX + ADDQ DX, R15 + MOVQ R12, CX + RORQ $0x0e, AX + MOVQ R12, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R12, CX + RORQ $0x29, DX + ANDQ R13, CX + XORQ AX, DX + MOVQ R12, AX + NOTQ AX + ADDQ DX, R15 + ANDQ R14, AX + XORQ CX, AX + ADDQ R15, AX + MOVQ R8, DI + MOVQ R10, BX + RORQ $0x1c, DI + MOVQ R8, DX + ANDQ R9, BX + RORQ $0x22, DX + MOVQ R8, CX + ANDQ R10, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R8, DX + MOVQ R9, CX + RORQ $0x27, DX + ANDQ R8, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R15 + ADDQ AX, R11 + ADDQ AX, R15 + MOVQ 568(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 464(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 528(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 456(BP), BX + ADDQ BX, AX + MOVQ AX, 584(BP) + MOVQ $0x32caab7b40c72493, DX + ADDQ AX, R14 + MOVQ R11, AX + ADDQ DX, R14 + MOVQ R11, CX + RORQ $0x0e, AX + MOVQ R11, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R11, CX + RORQ $0x29, DX + ANDQ R12, CX + XORQ AX, DX + MOVQ R11, AX + NOTQ AX + ADDQ DX, R14 + ANDQ R13, AX + XORQ CX, AX + ADDQ R14, AX + MOVQ R15, DI + MOVQ R9, BX + RORQ $0x1c, DI + MOVQ R15, DX + ANDQ R8, BX + RORQ $0x22, DX + MOVQ R15, CX + ANDQ R9, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R15, DX + MOVQ R8, CX + RORQ $0x27, DX + ANDQ R15, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R14 + ADDQ AX, R10 + ADDQ AX, R14 + MOVQ 576(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 472(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 536(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 464(BP), BX + ADDQ BX, AX + MOVQ AX, 592(BP) + MOVQ $0x3c9ebe0a15c9bebc, DX + ADDQ AX, R13 + MOVQ R10, AX + ADDQ DX, R13 + MOVQ R10, CX + RORQ $0x0e, AX + MOVQ R10, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R10, CX + RORQ $0x29, DX + ANDQ R11, CX + XORQ AX, DX + MOVQ R10, AX + NOTQ AX + ADDQ DX, R13 + ANDQ R12, AX + XORQ CX, AX + ADDQ R13, AX + MOVQ R14, DI + MOVQ R8, BX + RORQ $0x1c, DI + MOVQ R14, DX + ANDQ R15, BX + RORQ $0x22, DX + MOVQ R14, CX + ANDQ R8, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R14, DX + MOVQ R15, CX + RORQ $0x27, DX + ANDQ R14, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R13 + ADDQ AX, R9 + ADDQ AX, R13 + MOVQ 584(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 480(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 544(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 472(BP), BX + ADDQ BX, AX + MOVQ AX, 600(BP) + MOVQ $0x431d67c49c100d4c, DX + ADDQ AX, R12 + MOVQ R9, AX + ADDQ DX, R12 + MOVQ R9, CX + RORQ $0x0e, AX + MOVQ R9, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R9, CX + RORQ $0x29, DX + ANDQ R10, CX + XORQ AX, DX + MOVQ R9, AX + NOTQ AX + ADDQ DX, R12 + ANDQ R11, AX + XORQ CX, AX + ADDQ R12, AX + MOVQ R13, DI + MOVQ R15, BX + RORQ $0x1c, DI + MOVQ R13, DX + ANDQ R14, BX + RORQ $0x22, DX + MOVQ R13, CX + ANDQ R15, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R13, DX + MOVQ R14, CX + RORQ $0x27, DX + ANDQ R13, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R12 + ADDQ AX, R8 + ADDQ AX, R12 + MOVQ 592(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 488(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 552(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 480(BP), BX + ADDQ BX, AX + MOVQ AX, 608(BP) + MOVQ $0x4cc5d4becb3e42b6, DX + ADDQ AX, R11 + MOVQ R8, AX + ADDQ DX, R11 + MOVQ R8, CX + RORQ $0x0e, AX + MOVQ R8, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R8, CX + RORQ $0x29, DX + ANDQ R9, CX + XORQ AX, DX + MOVQ R8, AX + NOTQ AX + ADDQ DX, R11 + ANDQ R10, AX + XORQ CX, AX + ADDQ R11, AX + MOVQ R12, DI + MOVQ R14, BX + RORQ $0x1c, DI + MOVQ R12, DX + ANDQ R13, BX + RORQ $0x22, DX + MOVQ R12, CX + ANDQ R14, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R12, DX + MOVQ R13, CX + RORQ $0x27, DX + ANDQ R12, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R11 + ADDQ AX, R15 + ADDQ AX, R11 + MOVQ 600(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 496(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 560(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 488(BP), BX + ADDQ BX, AX + MOVQ AX, 616(BP) + MOVQ $0x597f299cfc657e2a, DX + ADDQ AX, R10 + MOVQ R15, AX + ADDQ DX, R10 + MOVQ R15, CX + RORQ $0x0e, AX + MOVQ R15, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R15, CX + RORQ $0x29, DX + ANDQ R8, CX + XORQ AX, DX + MOVQ R15, AX + NOTQ AX + ADDQ DX, R10 + ANDQ R9, AX + XORQ CX, AX + ADDQ R10, AX + MOVQ R11, DI + MOVQ R13, BX + RORQ $0x1c, DI + MOVQ R11, DX + ANDQ R12, BX + RORQ $0x22, DX + MOVQ R11, CX + ANDQ R13, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R11, DX + MOVQ R12, CX + RORQ $0x27, DX + ANDQ R11, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R10 + ADDQ AX, R14 + ADDQ AX, R10 + MOVQ 608(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 504(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 568(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 496(BP), BX + ADDQ BX, AX + MOVQ AX, 624(BP) + MOVQ $0x5fcb6fab3ad6faec, DX + ADDQ AX, R9 + MOVQ R14, AX + ADDQ DX, R9 + MOVQ R14, CX + RORQ $0x0e, AX + MOVQ R14, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R14, CX + RORQ $0x29, DX + ANDQ R15, CX + XORQ AX, DX + MOVQ R14, AX + NOTQ AX + ADDQ DX, R9 + ANDQ R8, AX + XORQ CX, AX + ADDQ R9, AX + MOVQ R10, DI + MOVQ R12, BX + RORQ $0x1c, DI + MOVQ R10, DX + ANDQ R11, BX + RORQ $0x22, DX + MOVQ R10, CX + ANDQ R12, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R10, DX + MOVQ R11, CX + RORQ $0x27, DX + ANDQ R10, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R9 + ADDQ AX, R13 + ADDQ AX, R9 + MOVQ 616(BP), AX + MOVQ AX, CX + RORQ $0x13, AX + MOVQ CX, DX + RORQ $0x3d, CX + SHRQ $0x06, DX + MOVQ 512(BP), BX + XORQ CX, AX + MOVQ BX, CX + XORQ DX, AX + RORQ $0x01, BX + MOVQ CX, DX + SHRQ $0x07, DX + RORQ $0x08, CX + ADDQ 576(BP), AX + XORQ CX, BX + XORQ DX, BX + ADDQ 504(BP), BX + ADDQ BX, AX + MOVQ AX, 632(BP) + MOVQ $0x6c44198c4a475817, DX + ADDQ AX, R8 + MOVQ R13, AX + ADDQ DX, R8 + MOVQ R13, CX + RORQ $0x0e, AX + MOVQ R13, DX + RORQ $0x12, CX + XORQ CX, AX + MOVQ R13, CX + RORQ $0x29, DX + ANDQ R14, CX + XORQ AX, DX + MOVQ R13, AX + NOTQ AX + ADDQ DX, R8 + ANDQ R15, AX + XORQ CX, AX + ADDQ R8, AX + MOVQ R9, DI + MOVQ R11, BX + RORQ $0x1c, DI + MOVQ R9, DX + ANDQ R10, BX + RORQ $0x22, DX + MOVQ R9, CX + ANDQ R11, CX + XORQ DX, DI + XORQ CX, BX + MOVQ R9, DX + MOVQ R10, CX + RORQ $0x27, DX + ANDQ R9, CX + XORQ CX, BX + XORQ DX, DI + ADDQ DI, BX + MOVQ BX, R8 + ADDQ AX, R12 + ADDQ AX, R8 + MOVQ dig+0(FP), BP + ADDQ (BP), R8 + MOVQ R8, (BP) + ADDQ 8(BP), R9 + MOVQ R9, 8(BP) + ADDQ 16(BP), R10 + MOVQ R10, 16(BP) + ADDQ 24(BP), R11 + MOVQ R11, 24(BP) + ADDQ 32(BP), R12 + MOVQ R12, 32(BP) + ADDQ 40(BP), R13 + MOVQ R13, 40(BP) + ADDQ 48(BP), R14 + MOVQ R14, 48(BP) + ADDQ 56(BP), R15 + MOVQ R15, 56(BP) + ADDQ $0x80, SI + CMPQ SI, 640(SP) + JB loop end: RET -// Version below is based on "Fast SHA512 Implementations on Intel -// Architecture Processors" White-paper -// https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/fast-sha512-implementations-ia-processors-paper.pdf -// AVX2 version by Intel, same algorithm in Linux kernel: -// https://github.com/torvalds/linux/blob/master/arch/x86/crypto/sha512-avx2-asm.S - -// James Guilford -// Kirk Yap -// Tim Chen -// David Cote -// Aleksey Sidorov - -#define YFER_SIZE (4*8) -#define SRND_SIZE (1*8) -#define INP_SIZE (1*8) - -#define frame_YFER (0) -#define frame_SRND (frame_YFER + YFER_SIZE) -#define frame_INP (frame_SRND + SRND_SIZE) -#define frame_INPEND (frame_INP + INP_SIZE) - -#define addm(p1, p2) \ - ADDQ p1, p2; \ - MOVQ p2, p1 - -#define COPY_YMM_AND_BSWAP(p1, p2, p3) \ - VMOVDQU p2, p1; \ - VPSHUFB p3, p1, p1 - -#define MY_VPALIGNR(YDST, YSRC1, YSRC2, RVAL) \ - VPERM2F128 $0x3, YSRC2, YSRC1, YDST; \ - VPALIGNR $RVAL, YSRC2, YDST, YDST - -DATA PSHUFFLE_BYTE_FLIP_MASK<>+0x00(SB)/8, $0x0001020304050607 -DATA PSHUFFLE_BYTE_FLIP_MASK<>+0x08(SB)/8, $0x08090a0b0c0d0e0f -DATA PSHUFFLE_BYTE_FLIP_MASK<>+0x10(SB)/8, $0x1011121314151617 -DATA PSHUFFLE_BYTE_FLIP_MASK<>+0x18(SB)/8, $0x18191a1b1c1d1e1f - -GLOBL PSHUFFLE_BYTE_FLIP_MASK<>(SB), (NOPTR+RODATA), $32 - -DATA MASK_YMM_LO<>+0x00(SB)/8, $0x0000000000000000 -DATA MASK_YMM_LO<>+0x08(SB)/8, $0x0000000000000000 -DATA MASK_YMM_LO<>+0x10(SB)/8, $0xFFFFFFFFFFFFFFFF -DATA MASK_YMM_LO<>+0x18(SB)/8, $0xFFFFFFFFFFFFFFFF - -GLOBL MASK_YMM_LO<>(SB), (NOPTR+RODATA), $32 +DATA PSHUFFLE_BYTE_FLIP_MASK<>+0(SB)/8, $0x0001020304050607 +DATA PSHUFFLE_BYTE_FLIP_MASK<>+8(SB)/8, $0x08090a0b0c0d0e0f +DATA PSHUFFLE_BYTE_FLIP_MASK<>+16(SB)/8, $0x1011121314151617 +DATA PSHUFFLE_BYTE_FLIP_MASK<>+24(SB)/8, $0x18191a1b1c1d1e1f +GLOBL PSHUFFLE_BYTE_FLIP_MASK<>(SB), RODATA|NOPTR, $32 +// func blockAVX2(dig *digest, p []byte) +// Requires: AVX, AVX2, BMI2 TEXT ·blockAVX2(SB), NOSPLIT, $56-32 - MOVQ dig+0(FP), SI - MOVQ p_base+8(FP), DI - MOVQ p_len+16(FP), DX - - SHRQ $7, DX - SHLQ $7, DX - - JZ done_hash - ADDQ DI, DX - MOVQ DX, frame_INPEND(SP) - - MOVQ (0*8)(SI), AX - MOVQ (1*8)(SI), BX - MOVQ (2*8)(SI), CX - MOVQ (3*8)(SI), R8 - MOVQ (4*8)(SI), DX - MOVQ (5*8)(SI), R9 - MOVQ (6*8)(SI), R10 - MOVQ (7*8)(SI), R11 - - VMOVDQU PSHUFFLE_BYTE_FLIP_MASK<>(SB), Y9 + MOVQ dig+0(FP), SI + MOVQ p_base+8(FP), DI + MOVQ p_len+16(FP), DX + SHRQ $0x07, DX + SHLQ $0x07, DX + JZ done_hash + ADDQ DI, DX + MOVQ DX, 48(SP) + MOVQ (SI), AX + MOVQ 8(SI), BX + MOVQ 16(SI), CX + MOVQ 24(SI), R8 + MOVQ 32(SI), DX + MOVQ 40(SI), R9 + MOVQ 48(SI), R10 + MOVQ 56(SI), R11 + VMOVDQU PSHUFFLE_BYTE_FLIP_MASK<>+0(SB), Y9 loop0: - MOVQ ·_K+0(SB), BP - - // byte swap first 16 dwords - COPY_YMM_AND_BSWAP(Y4, (0*32)(DI), Y9) - COPY_YMM_AND_BSWAP(Y5, (1*32)(DI), Y9) - COPY_YMM_AND_BSWAP(Y6, (2*32)(DI), Y9) - COPY_YMM_AND_BSWAP(Y7, (3*32)(DI), Y9) - - MOVQ DI, frame_INP(SP) - - // schedule 64 input dwords, by doing 12 rounds of 4 each - MOVQ $4, frame_SRND(SP) + MOVQ ·_K+0(SB), BP + VMOVDQU (DI), Y4 + VPSHUFB Y9, Y4, Y4 + VMOVDQU 32(DI), Y5 + VPSHUFB Y9, Y5, Y5 + VMOVDQU 64(DI), Y6 + VPSHUFB Y9, Y6, Y6 + VMOVDQU 96(DI), Y7 + VPSHUFB Y9, Y7, Y7 + MOVQ DI, 40(SP) + MOVQ $0x00000004, 32(SP) loop1: - VPADDQ (BP), Y4, Y0 - VMOVDQU Y0, frame_YFER(SP) - - MY_VPALIGNR(Y0, Y7, Y6, 8) - - VPADDQ Y4, Y0, Y0 - - MY_VPALIGNR(Y1, Y5, Y4, 8) - - VPSRLQ $1, Y1, Y2 - VPSLLQ $(64-1), Y1, Y3 - VPOR Y2, Y3, Y3 - - VPSRLQ $7, Y1, Y8 - - MOVQ AX, DI - RORXQ $41, DX, R13 - RORXQ $18, DX, R14 - ADDQ frame_YFER(SP), R11 - ORQ CX, DI - MOVQ R9, R15 - RORXQ $34, AX, R12 - - XORQ R14, R13 - XORQ R10, R15 - RORXQ $14, DX, R14 - - ANDQ DX, R15 - XORQ R14, R13 - RORXQ $39, AX, R14 - ADDQ R11, R8 - - ANDQ BX, DI - XORQ R12, R14 - RORXQ $28, AX, R12 - - XORQ R10, R15 - XORQ R12, R14 - MOVQ AX, R12 - ANDQ CX, R12 - - ADDQ R13, R15 - ORQ R12, DI - ADDQ R14, R11 - - ADDQ R15, R8 - - ADDQ R15, R11 - ADDQ DI, R11 - - VPSRLQ $8, Y1, Y2 - VPSLLQ $(64-8), Y1, Y1 - VPOR Y2, Y1, Y1 - - VPXOR Y8, Y3, Y3 - VPXOR Y1, Y3, Y1 - - VPADDQ Y1, Y0, Y0 - - VPERM2F128 $0x0, Y0, Y0, Y4 - - VPAND MASK_YMM_LO<>(SB), Y0, Y0 - + VPADDQ (BP), Y4, Y0 + VMOVDQU Y0, (SP) + VPERM2F128 $0x03, Y6, Y7, Y0 + VPALIGNR $0x08, Y6, Y0, Y0 + VPADDQ Y4, Y0, Y0 + VPERM2F128 $0x03, Y4, Y5, Y1 + VPALIGNR $0x08, Y4, Y1, Y1 + VPSRLQ $0x01, Y1, Y2 + VPSLLQ $0x3f, Y1, Y3 + VPOR Y2, Y3, Y3 + VPSRLQ $0x07, Y1, Y8 + MOVQ AX, DI + RORXQ $0x29, DX, R13 + RORXQ $0x12, DX, R14 + ADDQ (SP), R11 + ORQ CX, DI + MOVQ R9, R15 + RORXQ $0x22, AX, R12 + XORQ R14, R13 + XORQ R10, R15 + RORXQ $0x0e, DX, R14 + ANDQ DX, R15 + XORQ R14, R13 + RORXQ $0x27, AX, R14 + ADDQ R11, R8 + ANDQ BX, DI + XORQ R12, R14 + RORXQ $0x1c, AX, R12 + XORQ R10, R15 + XORQ R12, R14 + MOVQ AX, R12 + ANDQ CX, R12 + ADDQ R13, R15 + ORQ R12, DI + ADDQ R14, R11 + ADDQ R15, R8 + ADDQ R15, R11 + ADDQ DI, R11 + VPSRLQ $0x08, Y1, Y2 + VPSLLQ $0x38, Y1, Y1 + VPOR Y2, Y1, Y1 + VPXOR Y8, Y3, Y3 + VPXOR Y1, Y3, Y1 + VPADDQ Y1, Y0, Y0 + VPERM2F128 $0x00, Y0, Y0, Y4 + VPAND MASK_YMM_LO<>+0(SB), Y0, Y0 VPERM2F128 $0x11, Y7, Y7, Y2 - VPSRLQ $6, Y2, Y8 - - MOVQ R11, DI - RORXQ $41, R8, R13 - RORXQ $18, R8, R14 - ADDQ 1*8+frame_YFER(SP), R10 - ORQ BX, DI - - MOVQ DX, R15 - RORXQ $34, R11, R12 - XORQ R14, R13 - XORQ R9, R15 - - RORXQ $14, R8, R14 - XORQ R14, R13 - RORXQ $39, R11, R14 - ANDQ R8, R15 - ADDQ R10, CX - - ANDQ AX, DI - XORQ R12, R14 - - RORXQ $28, R11, R12 - XORQ R9, R15 - - XORQ R12, R14 - MOVQ R11, R12 - ANDQ BX, R12 - ADDQ R13, R15 - - ORQ R12, DI - ADDQ R14, R10 - - ADDQ R15, CX - ADDQ R15, R10 - ADDQ DI, R10 - - VPSRLQ $19, Y2, Y3 - VPSLLQ $(64-19), Y2, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - VPSRLQ $61, Y2, Y3 - VPSLLQ $(64-61), Y2, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - - VPADDQ Y8, Y4, Y4 - - VPSRLQ $6, Y4, Y8 - - MOVQ R10, DI - RORXQ $41, CX, R13 - ADDQ 2*8+frame_YFER(SP), R9 - - RORXQ $18, CX, R14 - ORQ AX, DI - MOVQ R8, R15 - XORQ DX, R15 - - RORXQ $34, R10, R12 - XORQ R14, R13 - ANDQ CX, R15 - - RORXQ $14, CX, R14 - ADDQ R9, BX - ANDQ R11, DI - - XORQ R14, R13 - RORXQ $39, R10, R14 - XORQ DX, R15 - - XORQ R12, R14 - RORXQ $28, R10, R12 - - XORQ R12, R14 - MOVQ R10, R12 - ANDQ AX, R12 - ADDQ R13, R15 - - ORQ R12, DI - ADDQ R14, R9 - ADDQ R15, BX - ADDQ R15, R9 - - ADDQ DI, R9 - - VPSRLQ $19, Y4, Y3 - VPSLLQ $(64-19), Y4, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - VPSRLQ $61, Y4, Y3 - VPSLLQ $(64-61), Y4, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - - VPADDQ Y8, Y0, Y2 - - VPBLENDD $0xF0, Y2, Y4, Y4 - - MOVQ R9, DI - RORXQ $41, BX, R13 - RORXQ $18, BX, R14 - ADDQ 3*8+frame_YFER(SP), DX - ORQ R11, DI - - MOVQ CX, R15 - RORXQ $34, R9, R12 - XORQ R14, R13 - XORQ R8, R15 - - RORXQ $14, BX, R14 - ANDQ BX, R15 - ADDQ DX, AX - ANDQ R10, DI - - XORQ R14, R13 - XORQ R8, R15 - - RORXQ $39, R9, R14 - ADDQ R13, R15 - - XORQ R12, R14 - ADDQ R15, AX - - RORXQ $28, R9, R12 - - XORQ R12, R14 - MOVQ R9, R12 - ANDQ R11, R12 - ORQ R12, DI - - ADDQ R14, DX - ADDQ R15, DX - ADDQ DI, DX - - VPADDQ 1*32(BP), Y5, Y0 - VMOVDQU Y0, frame_YFER(SP) - - MY_VPALIGNR(Y0, Y4, Y7, 8) - - VPADDQ Y5, Y0, Y0 - - MY_VPALIGNR(Y1, Y6, Y5, 8) - - VPSRLQ $1, Y1, Y2 - VPSLLQ $(64-1), Y1, Y3 - VPOR Y2, Y3, Y3 - - VPSRLQ $7, Y1, Y8 - - MOVQ DX, DI - RORXQ $41, AX, R13 - RORXQ $18, AX, R14 - ADDQ frame_YFER(SP), R8 - ORQ R10, DI - MOVQ BX, R15 - RORXQ $34, DX, R12 - - XORQ R14, R13 - XORQ CX, R15 - RORXQ $14, AX, R14 - - ANDQ AX, R15 - XORQ R14, R13 - RORXQ $39, DX, R14 - ADDQ R8, R11 - - ANDQ R9, DI - XORQ R12, R14 - RORXQ $28, DX, R12 - - XORQ CX, R15 - XORQ R12, R14 - MOVQ DX, R12 - ANDQ R10, R12 - - ADDQ R13, R15 - ORQ R12, DI - ADDQ R14, R8 - - ADDQ R15, R11 - - ADDQ R15, R8 - ADDQ DI, R8 - - VPSRLQ $8, Y1, Y2 - VPSLLQ $(64-8), Y1, Y1 - VPOR Y2, Y1, Y1 - - VPXOR Y8, Y3, Y3 - VPXOR Y1, Y3, Y1 - - VPADDQ Y1, Y0, Y0 - - VPERM2F128 $0x0, Y0, Y0, Y5 - - VPAND MASK_YMM_LO<>(SB), Y0, Y0 - + VPSRLQ $0x06, Y2, Y8 + MOVQ R11, DI + RORXQ $0x29, R8, R13 + RORXQ $0x12, R8, R14 + ADDQ 8(SP), R10 + ORQ BX, DI + MOVQ DX, R15 + RORXQ $0x22, R11, R12 + XORQ R14, R13 + XORQ R9, R15 + RORXQ $0x0e, R8, R14 + XORQ R14, R13 + RORXQ $0x27, R11, R14 + ANDQ R8, R15 + ADDQ R10, CX + ANDQ AX, DI + XORQ R12, R14 + RORXQ $0x1c, R11, R12 + XORQ R9, R15 + XORQ R12, R14 + MOVQ R11, R12 + ANDQ BX, R12 + ADDQ R13, R15 + ORQ R12, DI + ADDQ R14, R10 + ADDQ R15, CX + ADDQ R15, R10 + ADDQ DI, R10 + VPSRLQ $0x13, Y2, Y3 + VPSLLQ $0x2d, Y2, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPSRLQ $0x3d, Y2, Y3 + VPSLLQ $0x03, Y2, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPADDQ Y8, Y4, Y4 + VPSRLQ $0x06, Y4, Y8 + MOVQ R10, DI + RORXQ $0x29, CX, R13 + ADDQ 16(SP), R9 + RORXQ $0x12, CX, R14 + ORQ AX, DI + MOVQ R8, R15 + XORQ DX, R15 + RORXQ $0x22, R10, R12 + XORQ R14, R13 + ANDQ CX, R15 + RORXQ $0x0e, CX, R14 + ADDQ R9, BX + ANDQ R11, DI + XORQ R14, R13 + RORXQ $0x27, R10, R14 + XORQ DX, R15 + XORQ R12, R14 + RORXQ $0x1c, R10, R12 + XORQ R12, R14 + MOVQ R10, R12 + ANDQ AX, R12 + ADDQ R13, R15 + ORQ R12, DI + ADDQ R14, R9 + ADDQ R15, BX + ADDQ R15, R9 + ADDQ DI, R9 + VPSRLQ $0x13, Y4, Y3 + VPSLLQ $0x2d, Y4, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPSRLQ $0x3d, Y4, Y3 + VPSLLQ $0x03, Y4, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPADDQ Y8, Y0, Y2 + VPBLENDD $0xf0, Y2, Y4, Y4 + MOVQ R9, DI + RORXQ $0x29, BX, R13 + RORXQ $0x12, BX, R14 + ADDQ 24(SP), DX + ORQ R11, DI + MOVQ CX, R15 + RORXQ $0x22, R9, R12 + XORQ R14, R13 + XORQ R8, R15 + RORXQ $0x0e, BX, R14 + ANDQ BX, R15 + ADDQ DX, AX + ANDQ R10, DI + XORQ R14, R13 + XORQ R8, R15 + RORXQ $0x27, R9, R14 + ADDQ R13, R15 + XORQ R12, R14 + ADDQ R15, AX + RORXQ $0x1c, R9, R12 + XORQ R12, R14 + MOVQ R9, R12 + ANDQ R11, R12 + ORQ R12, DI + ADDQ R14, DX + ADDQ R15, DX + ADDQ DI, DX + VPADDQ 32(BP), Y5, Y0 + VMOVDQU Y0, (SP) + VPERM2F128 $0x03, Y7, Y4, Y0 + VPALIGNR $0x08, Y7, Y0, Y0 + VPADDQ Y5, Y0, Y0 + VPERM2F128 $0x03, Y5, Y6, Y1 + VPALIGNR $0x08, Y5, Y1, Y1 + VPSRLQ $0x01, Y1, Y2 + VPSLLQ $0x3f, Y1, Y3 + VPOR Y2, Y3, Y3 + VPSRLQ $0x07, Y1, Y8 + MOVQ DX, DI + RORXQ $0x29, AX, R13 + RORXQ $0x12, AX, R14 + ADDQ (SP), R8 + ORQ R10, DI + MOVQ BX, R15 + RORXQ $0x22, DX, R12 + XORQ R14, R13 + XORQ CX, R15 + RORXQ $0x0e, AX, R14 + ANDQ AX, R15 + XORQ R14, R13 + RORXQ $0x27, DX, R14 + ADDQ R8, R11 + ANDQ R9, DI + XORQ R12, R14 + RORXQ $0x1c, DX, R12 + XORQ CX, R15 + XORQ R12, R14 + MOVQ DX, R12 + ANDQ R10, R12 + ADDQ R13, R15 + ORQ R12, DI + ADDQ R14, R8 + ADDQ R15, R11 + ADDQ R15, R8 + ADDQ DI, R8 + VPSRLQ $0x08, Y1, Y2 + VPSLLQ $0x38, Y1, Y1 + VPOR Y2, Y1, Y1 + VPXOR Y8, Y3, Y3 + VPXOR Y1, Y3, Y1 + VPADDQ Y1, Y0, Y0 + VPERM2F128 $0x00, Y0, Y0, Y5 + VPAND MASK_YMM_LO<>+0(SB), Y0, Y0 VPERM2F128 $0x11, Y4, Y4, Y2 - VPSRLQ $6, Y2, Y8 - - MOVQ R8, DI - RORXQ $41, R11, R13 - RORXQ $18, R11, R14 - ADDQ 1*8+frame_YFER(SP), CX - ORQ R9, DI - - MOVQ AX, R15 - RORXQ $34, R8, R12 - XORQ R14, R13 - XORQ BX, R15 - - RORXQ $14, R11, R14 - XORQ R14, R13 - RORXQ $39, R8, R14 - ANDQ R11, R15 - ADDQ CX, R10 - - ANDQ DX, DI - XORQ R12, R14 - - RORXQ $28, R8, R12 - XORQ BX, R15 - - XORQ R12, R14 - MOVQ R8, R12 - ANDQ R9, R12 - ADDQ R13, R15 - - ORQ R12, DI - ADDQ R14, CX - - ADDQ R15, R10 - ADDQ R15, CX - ADDQ DI, CX - - VPSRLQ $19, Y2, Y3 - VPSLLQ $(64-19), Y2, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - VPSRLQ $61, Y2, Y3 - VPSLLQ $(64-61), Y2, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - - VPADDQ Y8, Y5, Y5 - - VPSRLQ $6, Y5, Y8 - - MOVQ CX, DI - RORXQ $41, R10, R13 - ADDQ 2*8+frame_YFER(SP), BX - - RORXQ $18, R10, R14 - ORQ DX, DI - MOVQ R11, R15 - XORQ AX, R15 - - RORXQ $34, CX, R12 - XORQ R14, R13 - ANDQ R10, R15 - - RORXQ $14, R10, R14 - ADDQ BX, R9 - ANDQ R8, DI - - XORQ R14, R13 - RORXQ $39, CX, R14 - XORQ AX, R15 - - XORQ R12, R14 - RORXQ $28, CX, R12 - - XORQ R12, R14 - MOVQ CX, R12 - ANDQ DX, R12 - ADDQ R13, R15 - - ORQ R12, DI - ADDQ R14, BX - ADDQ R15, R9 - ADDQ R15, BX - - ADDQ DI, BX - - VPSRLQ $19, Y5, Y3 - VPSLLQ $(64-19), Y5, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - VPSRLQ $61, Y5, Y3 - VPSLLQ $(64-61), Y5, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - - VPADDQ Y8, Y0, Y2 - - VPBLENDD $0xF0, Y2, Y5, Y5 - - MOVQ BX, DI - RORXQ $41, R9, R13 - RORXQ $18, R9, R14 - ADDQ 3*8+frame_YFER(SP), AX - ORQ R8, DI - - MOVQ R10, R15 - RORXQ $34, BX, R12 - XORQ R14, R13 - XORQ R11, R15 - - RORXQ $14, R9, R14 - ANDQ R9, R15 - ADDQ AX, DX - ANDQ CX, DI - - XORQ R14, R13 - XORQ R11, R15 - - RORXQ $39, BX, R14 - ADDQ R13, R15 - - XORQ R12, R14 - ADDQ R15, DX - - RORXQ $28, BX, R12 - - XORQ R12, R14 - MOVQ BX, R12 - ANDQ R8, R12 - ORQ R12, DI - - ADDQ R14, AX - ADDQ R15, AX - ADDQ DI, AX - - VPADDQ 2*32(BP), Y6, Y0 - VMOVDQU Y0, frame_YFER(SP) - - MY_VPALIGNR(Y0, Y5, Y4, 8) - - VPADDQ Y6, Y0, Y0 - - MY_VPALIGNR(Y1, Y7, Y6, 8) - - VPSRLQ $1, Y1, Y2 - VPSLLQ $(64-1), Y1, Y3 - VPOR Y2, Y3, Y3 - - VPSRLQ $7, Y1, Y8 - - MOVQ AX, DI - RORXQ $41, DX, R13 - RORXQ $18, DX, R14 - ADDQ frame_YFER(SP), R11 - ORQ CX, DI - MOVQ R9, R15 - RORXQ $34, AX, R12 - - XORQ R14, R13 - XORQ R10, R15 - RORXQ $14, DX, R14 - - ANDQ DX, R15 - XORQ R14, R13 - RORXQ $39, AX, R14 - ADDQ R11, R8 - - ANDQ BX, DI - XORQ R12, R14 - RORXQ $28, AX, R12 - - XORQ R10, R15 - XORQ R12, R14 - MOVQ AX, R12 - ANDQ CX, R12 - - ADDQ R13, R15 - ORQ R12, DI - ADDQ R14, R11 - - ADDQ R15, R8 - - ADDQ R15, R11 - ADDQ DI, R11 - - VPSRLQ $8, Y1, Y2 - VPSLLQ $(64-8), Y1, Y1 - VPOR Y2, Y1, Y1 - - VPXOR Y8, Y3, Y3 - VPXOR Y1, Y3, Y1 - - VPADDQ Y1, Y0, Y0 - - VPERM2F128 $0x0, Y0, Y0, Y6 - - VPAND MASK_YMM_LO<>(SB), Y0, Y0 - + VPSRLQ $0x06, Y2, Y8 + MOVQ R8, DI + RORXQ $0x29, R11, R13 + RORXQ $0x12, R11, R14 + ADDQ 8(SP), CX + ORQ R9, DI + MOVQ AX, R15 + RORXQ $0x22, R8, R12 + XORQ R14, R13 + XORQ BX, R15 + RORXQ $0x0e, R11, R14 + XORQ R14, R13 + RORXQ $0x27, R8, R14 + ANDQ R11, R15 + ADDQ CX, R10 + ANDQ DX, DI + XORQ R12, R14 + RORXQ $0x1c, R8, R12 + XORQ BX, R15 + XORQ R12, R14 + MOVQ R8, R12 + ANDQ R9, R12 + ADDQ R13, R15 + ORQ R12, DI + ADDQ R14, CX + ADDQ R15, R10 + ADDQ R15, CX + ADDQ DI, CX + VPSRLQ $0x13, Y2, Y3 + VPSLLQ $0x2d, Y2, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPSRLQ $0x3d, Y2, Y3 + VPSLLQ $0x03, Y2, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPADDQ Y8, Y5, Y5 + VPSRLQ $0x06, Y5, Y8 + MOVQ CX, DI + RORXQ $0x29, R10, R13 + ADDQ 16(SP), BX + RORXQ $0x12, R10, R14 + ORQ DX, DI + MOVQ R11, R15 + XORQ AX, R15 + RORXQ $0x22, CX, R12 + XORQ R14, R13 + ANDQ R10, R15 + RORXQ $0x0e, R10, R14 + ADDQ BX, R9 + ANDQ R8, DI + XORQ R14, R13 + RORXQ $0x27, CX, R14 + XORQ AX, R15 + XORQ R12, R14 + RORXQ $0x1c, CX, R12 + XORQ R12, R14 + MOVQ CX, R12 + ANDQ DX, R12 + ADDQ R13, R15 + ORQ R12, DI + ADDQ R14, BX + ADDQ R15, R9 + ADDQ R15, BX + ADDQ DI, BX + VPSRLQ $0x13, Y5, Y3 + VPSLLQ $0x2d, Y5, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPSRLQ $0x3d, Y5, Y3 + VPSLLQ $0x03, Y5, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPADDQ Y8, Y0, Y2 + VPBLENDD $0xf0, Y2, Y5, Y5 + MOVQ BX, DI + RORXQ $0x29, R9, R13 + RORXQ $0x12, R9, R14 + ADDQ 24(SP), AX + ORQ R8, DI + MOVQ R10, R15 + RORXQ $0x22, BX, R12 + XORQ R14, R13 + XORQ R11, R15 + RORXQ $0x0e, R9, R14 + ANDQ R9, R15 + ADDQ AX, DX + ANDQ CX, DI + XORQ R14, R13 + XORQ R11, R15 + RORXQ $0x27, BX, R14 + ADDQ R13, R15 + XORQ R12, R14 + ADDQ R15, DX + RORXQ $0x1c, BX, R12 + XORQ R12, R14 + MOVQ BX, R12 + ANDQ R8, R12 + ORQ R12, DI + ADDQ R14, AX + ADDQ R15, AX + ADDQ DI, AX + VPADDQ 64(BP), Y6, Y0 + VMOVDQU Y0, (SP) + VPERM2F128 $0x03, Y4, Y5, Y0 + VPALIGNR $0x08, Y4, Y0, Y0 + VPADDQ Y6, Y0, Y0 + VPERM2F128 $0x03, Y6, Y7, Y1 + VPALIGNR $0x08, Y6, Y1, Y1 + VPSRLQ $0x01, Y1, Y2 + VPSLLQ $0x3f, Y1, Y3 + VPOR Y2, Y3, Y3 + VPSRLQ $0x07, Y1, Y8 + MOVQ AX, DI + RORXQ $0x29, DX, R13 + RORXQ $0x12, DX, R14 + ADDQ (SP), R11 + ORQ CX, DI + MOVQ R9, R15 + RORXQ $0x22, AX, R12 + XORQ R14, R13 + XORQ R10, R15 + RORXQ $0x0e, DX, R14 + ANDQ DX, R15 + XORQ R14, R13 + RORXQ $0x27, AX, R14 + ADDQ R11, R8 + ANDQ BX, DI + XORQ R12, R14 + RORXQ $0x1c, AX, R12 + XORQ R10, R15 + XORQ R12, R14 + MOVQ AX, R12 + ANDQ CX, R12 + ADDQ R13, R15 + ORQ R12, DI + ADDQ R14, R11 + ADDQ R15, R8 + ADDQ R15, R11 + ADDQ DI, R11 + VPSRLQ $0x08, Y1, Y2 + VPSLLQ $0x38, Y1, Y1 + VPOR Y2, Y1, Y1 + VPXOR Y8, Y3, Y3 + VPXOR Y1, Y3, Y1 + VPADDQ Y1, Y0, Y0 + VPERM2F128 $0x00, Y0, Y0, Y6 + VPAND MASK_YMM_LO<>+0(SB), Y0, Y0 VPERM2F128 $0x11, Y5, Y5, Y2 - VPSRLQ $6, Y2, Y8 - - MOVQ R11, DI - RORXQ $41, R8, R13 - RORXQ $18, R8, R14 - ADDQ 1*8+frame_YFER(SP), R10 - ORQ BX, DI - - MOVQ DX, R15 - RORXQ $34, R11, R12 - XORQ R14, R13 - XORQ R9, R15 - - RORXQ $14, R8, R14 - XORQ R14, R13 - RORXQ $39, R11, R14 - ANDQ R8, R15 - ADDQ R10, CX - - ANDQ AX, DI - XORQ R12, R14 - - RORXQ $28, R11, R12 - XORQ R9, R15 - - XORQ R12, R14 - MOVQ R11, R12 - ANDQ BX, R12 - ADDQ R13, R15 - - ORQ R12, DI - ADDQ R14, R10 - - ADDQ R15, CX - ADDQ R15, R10 - ADDQ DI, R10 - - VPSRLQ $19, Y2, Y3 - VPSLLQ $(64-19), Y2, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - VPSRLQ $61, Y2, Y3 - VPSLLQ $(64-61), Y2, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - - VPADDQ Y8, Y6, Y6 - - VPSRLQ $6, Y6, Y8 - - MOVQ R10, DI - RORXQ $41, CX, R13 - ADDQ 2*8+frame_YFER(SP), R9 - - RORXQ $18, CX, R14 - ORQ AX, DI - MOVQ R8, R15 - XORQ DX, R15 - - RORXQ $34, R10, R12 - XORQ R14, R13 - ANDQ CX, R15 - - RORXQ $14, CX, R14 - ADDQ R9, BX - ANDQ R11, DI - - XORQ R14, R13 - RORXQ $39, R10, R14 - XORQ DX, R15 - - XORQ R12, R14 - RORXQ $28, R10, R12 - - XORQ R12, R14 - MOVQ R10, R12 - ANDQ AX, R12 - ADDQ R13, R15 - - ORQ R12, DI - ADDQ R14, R9 - ADDQ R15, BX - ADDQ R15, R9 - - ADDQ DI, R9 - - VPSRLQ $19, Y6, Y3 - VPSLLQ $(64-19), Y6, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - VPSRLQ $61, Y6, Y3 - VPSLLQ $(64-61), Y6, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - - VPADDQ Y8, Y0, Y2 - - VPBLENDD $0xF0, Y2, Y6, Y6 - - MOVQ R9, DI - RORXQ $41, BX, R13 - RORXQ $18, BX, R14 - ADDQ 3*8+frame_YFER(SP), DX - ORQ R11, DI - - MOVQ CX, R15 - RORXQ $34, R9, R12 - XORQ R14, R13 - XORQ R8, R15 - - RORXQ $14, BX, R14 - ANDQ BX, R15 - ADDQ DX, AX - ANDQ R10, DI - - XORQ R14, R13 - XORQ R8, R15 - - RORXQ $39, R9, R14 - ADDQ R13, R15 - - XORQ R12, R14 - ADDQ R15, AX - - RORXQ $28, R9, R12 - - XORQ R12, R14 - MOVQ R9, R12 - ANDQ R11, R12 - ORQ R12, DI - - ADDQ R14, DX - ADDQ R15, DX - ADDQ DI, DX - - VPADDQ 3*32(BP), Y7, Y0 - VMOVDQU Y0, frame_YFER(SP) - ADDQ $(4*32), BP - - MY_VPALIGNR(Y0, Y6, Y5, 8) - - VPADDQ Y7, Y0, Y0 - - MY_VPALIGNR(Y1, Y4, Y7, 8) - - VPSRLQ $1, Y1, Y2 - VPSLLQ $(64-1), Y1, Y3 - VPOR Y2, Y3, Y3 - - VPSRLQ $7, Y1, Y8 - - MOVQ DX, DI - RORXQ $41, AX, R13 - RORXQ $18, AX, R14 - ADDQ frame_YFER(SP), R8 - ORQ R10, DI - MOVQ BX, R15 - RORXQ $34, DX, R12 - - XORQ R14, R13 - XORQ CX, R15 - RORXQ $14, AX, R14 - - ANDQ AX, R15 - XORQ R14, R13 - RORXQ $39, DX, R14 - ADDQ R8, R11 - - ANDQ R9, DI - XORQ R12, R14 - RORXQ $28, DX, R12 - - XORQ CX, R15 - XORQ R12, R14 - MOVQ DX, R12 - ANDQ R10, R12 - - ADDQ R13, R15 - ORQ R12, DI - ADDQ R14, R8 - - ADDQ R15, R11 - - ADDQ R15, R8 - ADDQ DI, R8 - - VPSRLQ $8, Y1, Y2 - VPSLLQ $(64-8), Y1, Y1 - VPOR Y2, Y1, Y1 - - VPXOR Y8, Y3, Y3 - VPXOR Y1, Y3, Y1 - - VPADDQ Y1, Y0, Y0 - - VPERM2F128 $0x0, Y0, Y0, Y7 - - VPAND MASK_YMM_LO<>(SB), Y0, Y0 - + VPSRLQ $0x06, Y2, Y8 + MOVQ R11, DI + RORXQ $0x29, R8, R13 + RORXQ $0x12, R8, R14 + ADDQ 8(SP), R10 + ORQ BX, DI + MOVQ DX, R15 + RORXQ $0x22, R11, R12 + XORQ R14, R13 + XORQ R9, R15 + RORXQ $0x0e, R8, R14 + XORQ R14, R13 + RORXQ $0x27, R11, R14 + ANDQ R8, R15 + ADDQ R10, CX + ANDQ AX, DI + XORQ R12, R14 + RORXQ $0x1c, R11, R12 + XORQ R9, R15 + XORQ R12, R14 + MOVQ R11, R12 + ANDQ BX, R12 + ADDQ R13, R15 + ORQ R12, DI + ADDQ R14, R10 + ADDQ R15, CX + ADDQ R15, R10 + ADDQ DI, R10 + VPSRLQ $0x13, Y2, Y3 + VPSLLQ $0x2d, Y2, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPSRLQ $0x3d, Y2, Y3 + VPSLLQ $0x03, Y2, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPADDQ Y8, Y6, Y6 + VPSRLQ $0x06, Y6, Y8 + MOVQ R10, DI + RORXQ $0x29, CX, R13 + ADDQ 16(SP), R9 + RORXQ $0x12, CX, R14 + ORQ AX, DI + MOVQ R8, R15 + XORQ DX, R15 + RORXQ $0x22, R10, R12 + XORQ R14, R13 + ANDQ CX, R15 + RORXQ $0x0e, CX, R14 + ADDQ R9, BX + ANDQ R11, DI + XORQ R14, R13 + RORXQ $0x27, R10, R14 + XORQ DX, R15 + XORQ R12, R14 + RORXQ $0x1c, R10, R12 + XORQ R12, R14 + MOVQ R10, R12 + ANDQ AX, R12 + ADDQ R13, R15 + ORQ R12, DI + ADDQ R14, R9 + ADDQ R15, BX + ADDQ R15, R9 + ADDQ DI, R9 + VPSRLQ $0x13, Y6, Y3 + VPSLLQ $0x2d, Y6, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPSRLQ $0x3d, Y6, Y3 + VPSLLQ $0x03, Y6, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPADDQ Y8, Y0, Y2 + VPBLENDD $0xf0, Y2, Y6, Y6 + MOVQ R9, DI + RORXQ $0x29, BX, R13 + RORXQ $0x12, BX, R14 + ADDQ 24(SP), DX + ORQ R11, DI + MOVQ CX, R15 + RORXQ $0x22, R9, R12 + XORQ R14, R13 + XORQ R8, R15 + RORXQ $0x0e, BX, R14 + ANDQ BX, R15 + ADDQ DX, AX + ANDQ R10, DI + XORQ R14, R13 + XORQ R8, R15 + RORXQ $0x27, R9, R14 + ADDQ R13, R15 + XORQ R12, R14 + ADDQ R15, AX + RORXQ $0x1c, R9, R12 + XORQ R12, R14 + MOVQ R9, R12 + ANDQ R11, R12 + ORQ R12, DI + ADDQ R14, DX + ADDQ R15, DX + ADDQ DI, DX + VPADDQ 96(BP), Y7, Y0 + VMOVDQU Y0, (SP) + ADDQ $0x80, BP + VPERM2F128 $0x03, Y5, Y6, Y0 + VPALIGNR $0x08, Y5, Y0, Y0 + VPADDQ Y7, Y0, Y0 + VPERM2F128 $0x03, Y7, Y4, Y1 + VPALIGNR $0x08, Y7, Y1, Y1 + VPSRLQ $0x01, Y1, Y2 + VPSLLQ $0x3f, Y1, Y3 + VPOR Y2, Y3, Y3 + VPSRLQ $0x07, Y1, Y8 + MOVQ DX, DI + RORXQ $0x29, AX, R13 + RORXQ $0x12, AX, R14 + ADDQ (SP), R8 + ORQ R10, DI + MOVQ BX, R15 + RORXQ $0x22, DX, R12 + XORQ R14, R13 + XORQ CX, R15 + RORXQ $0x0e, AX, R14 + ANDQ AX, R15 + XORQ R14, R13 + RORXQ $0x27, DX, R14 + ADDQ R8, R11 + ANDQ R9, DI + XORQ R12, R14 + RORXQ $0x1c, DX, R12 + XORQ CX, R15 + XORQ R12, R14 + MOVQ DX, R12 + ANDQ R10, R12 + ADDQ R13, R15 + ORQ R12, DI + ADDQ R14, R8 + ADDQ R15, R11 + ADDQ R15, R8 + ADDQ DI, R8 + VPSRLQ $0x08, Y1, Y2 + VPSLLQ $0x38, Y1, Y1 + VPOR Y2, Y1, Y1 + VPXOR Y8, Y3, Y3 + VPXOR Y1, Y3, Y1 + VPADDQ Y1, Y0, Y0 + VPERM2F128 $0x00, Y0, Y0, Y7 + VPAND MASK_YMM_LO<>+0(SB), Y0, Y0 VPERM2F128 $0x11, Y6, Y6, Y2 - VPSRLQ $6, Y2, Y8 - - MOVQ R8, DI - RORXQ $41, R11, R13 - RORXQ $18, R11, R14 - ADDQ 1*8+frame_YFER(SP), CX - ORQ R9, DI - - MOVQ AX, R15 - RORXQ $34, R8, R12 - XORQ R14, R13 - XORQ BX, R15 - - RORXQ $14, R11, R14 - XORQ R14, R13 - RORXQ $39, R8, R14 - ANDQ R11, R15 - ADDQ CX, R10 - - ANDQ DX, DI - XORQ R12, R14 - - RORXQ $28, R8, R12 - XORQ BX, R15 - - XORQ R12, R14 - MOVQ R8, R12 - ANDQ R9, R12 - ADDQ R13, R15 - - ORQ R12, DI - ADDQ R14, CX - - ADDQ R15, R10 - ADDQ R15, CX - ADDQ DI, CX - - VPSRLQ $19, Y2, Y3 - VPSLLQ $(64-19), Y2, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - VPSRLQ $61, Y2, Y3 - VPSLLQ $(64-61), Y2, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - - VPADDQ Y8, Y7, Y7 - - VPSRLQ $6, Y7, Y8 - - MOVQ CX, DI - RORXQ $41, R10, R13 - ADDQ 2*8+frame_YFER(SP), BX - - RORXQ $18, R10, R14 - ORQ DX, DI - MOVQ R11, R15 - XORQ AX, R15 - - RORXQ $34, CX, R12 - XORQ R14, R13 - ANDQ R10, R15 - - RORXQ $14, R10, R14 - ADDQ BX, R9 - ANDQ R8, DI - - XORQ R14, R13 - RORXQ $39, CX, R14 - XORQ AX, R15 - - XORQ R12, R14 - RORXQ $28, CX, R12 - - XORQ R12, R14 - MOVQ CX, R12 - ANDQ DX, R12 - ADDQ R13, R15 - - ORQ R12, DI - ADDQ R14, BX - ADDQ R15, R9 - ADDQ R15, BX - - ADDQ DI, BX - - VPSRLQ $19, Y7, Y3 - VPSLLQ $(64-19), Y7, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - VPSRLQ $61, Y7, Y3 - VPSLLQ $(64-61), Y7, Y1 - VPOR Y1, Y3, Y3 - VPXOR Y3, Y8, Y8 - - VPADDQ Y8, Y0, Y2 - - VPBLENDD $0xF0, Y2, Y7, Y7 - - MOVQ BX, DI - RORXQ $41, R9, R13 - RORXQ $18, R9, R14 - ADDQ 3*8+frame_YFER(SP), AX - ORQ R8, DI - - MOVQ R10, R15 - RORXQ $34, BX, R12 - XORQ R14, R13 - XORQ R11, R15 - - RORXQ $14, R9, R14 - ANDQ R9, R15 - ADDQ AX, DX - ANDQ CX, DI - - XORQ R14, R13 - XORQ R11, R15 - - RORXQ $39, BX, R14 - ADDQ R13, R15 - - XORQ R12, R14 - ADDQ R15, DX - - RORXQ $28, BX, R12 - - XORQ R12, R14 - MOVQ BX, R12 - ANDQ R8, R12 - ORQ R12, DI - - ADDQ R14, AX - ADDQ R15, AX - ADDQ DI, AX - - SUBQ $1, frame_SRND(SP) - JNE loop1 - - MOVQ $2, frame_SRND(SP) + VPSRLQ $0x06, Y2, Y8 + MOVQ R8, DI + RORXQ $0x29, R11, R13 + RORXQ $0x12, R11, R14 + ADDQ 8(SP), CX + ORQ R9, DI + MOVQ AX, R15 + RORXQ $0x22, R8, R12 + XORQ R14, R13 + XORQ BX, R15 + RORXQ $0x0e, R11, R14 + XORQ R14, R13 + RORXQ $0x27, R8, R14 + ANDQ R11, R15 + ADDQ CX, R10 + ANDQ DX, DI + XORQ R12, R14 + RORXQ $0x1c, R8, R12 + XORQ BX, R15 + XORQ R12, R14 + MOVQ R8, R12 + ANDQ R9, R12 + ADDQ R13, R15 + ORQ R12, DI + ADDQ R14, CX + ADDQ R15, R10 + ADDQ R15, CX + ADDQ DI, CX + VPSRLQ $0x13, Y2, Y3 + VPSLLQ $0x2d, Y2, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPSRLQ $0x3d, Y2, Y3 + VPSLLQ $0x03, Y2, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPADDQ Y8, Y7, Y7 + VPSRLQ $0x06, Y7, Y8 + MOVQ CX, DI + RORXQ $0x29, R10, R13 + ADDQ 16(SP), BX + RORXQ $0x12, R10, R14 + ORQ DX, DI + MOVQ R11, R15 + XORQ AX, R15 + RORXQ $0x22, CX, R12 + XORQ R14, R13 + ANDQ R10, R15 + RORXQ $0x0e, R10, R14 + ADDQ BX, R9 + ANDQ R8, DI + XORQ R14, R13 + RORXQ $0x27, CX, R14 + XORQ AX, R15 + XORQ R12, R14 + RORXQ $0x1c, CX, R12 + XORQ R12, R14 + MOVQ CX, R12 + ANDQ DX, R12 + ADDQ R13, R15 + ORQ R12, DI + ADDQ R14, BX + ADDQ R15, R9 + ADDQ R15, BX + ADDQ DI, BX + VPSRLQ $0x13, Y7, Y3 + VPSLLQ $0x2d, Y7, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPSRLQ $0x3d, Y7, Y3 + VPSLLQ $0x03, Y7, Y1 + VPOR Y1, Y3, Y3 + VPXOR Y3, Y8, Y8 + VPADDQ Y8, Y0, Y2 + VPBLENDD $0xf0, Y2, Y7, Y7 + MOVQ BX, DI + RORXQ $0x29, R9, R13 + RORXQ $0x12, R9, R14 + ADDQ 24(SP), AX + ORQ R8, DI + MOVQ R10, R15 + RORXQ $0x22, BX, R12 + XORQ R14, R13 + XORQ R11, R15 + RORXQ $0x0e, R9, R14 + ANDQ R9, R15 + ADDQ AX, DX + ANDQ CX, DI + XORQ R14, R13 + XORQ R11, R15 + RORXQ $0x27, BX, R14 + ADDQ R13, R15 + XORQ R12, R14 + ADDQ R15, DX + RORXQ $0x1c, BX, R12 + XORQ R12, R14 + MOVQ BX, R12 + ANDQ R8, R12 + ORQ R12, DI + ADDQ R14, AX + ADDQ R15, AX + ADDQ DI, AX + SUBQ $0x01, 32(SP) + JNE loop1 + MOVQ $0x00000002, 32(SP) loop2: VPADDQ (BP), Y4, Y0 - VMOVDQU Y0, frame_YFER(SP) - - MOVQ R9, R15 - RORXQ $41, DX, R13 - RORXQ $18, DX, R14 - XORQ R10, R15 - - XORQ R14, R13 - RORXQ $14, DX, R14 - ANDQ DX, R15 - - XORQ R14, R13 - RORXQ $34, AX, R12 - XORQ R10, R15 - RORXQ $39, AX, R14 - MOVQ AX, DI - - XORQ R12, R14 - RORXQ $28, AX, R12 - ADDQ frame_YFER(SP), R11 - ORQ CX, DI - - XORQ R12, R14 - MOVQ AX, R12 - ANDQ BX, DI - ANDQ CX, R12 - ADDQ R13, R15 - - ADDQ R11, R8 - ORQ R12, DI - ADDQ R14, R11 - - ADDQ R15, R8 - - ADDQ R15, R11 - MOVQ DX, R15 - RORXQ $41, R8, R13 - RORXQ $18, R8, R14 - XORQ R9, R15 - - XORQ R14, R13 - RORXQ $14, R8, R14 - ANDQ R8, R15 - ADDQ DI, R11 - - XORQ R14, R13 - RORXQ $34, R11, R12 - XORQ R9, R15 - RORXQ $39, R11, R14 - MOVQ R11, DI - - XORQ R12, R14 - RORXQ $28, R11, R12 - ADDQ 8*1+frame_YFER(SP), R10 - ORQ BX, DI - - XORQ R12, R14 - MOVQ R11, R12 - ANDQ AX, DI - ANDQ BX, R12 - ADDQ R13, R15 - - ADDQ R10, CX - ORQ R12, DI - ADDQ R14, R10 - - ADDQ R15, CX - - ADDQ R15, R10 - MOVQ R8, R15 - RORXQ $41, CX, R13 - RORXQ $18, CX, R14 - XORQ DX, R15 - - XORQ R14, R13 - RORXQ $14, CX, R14 - ANDQ CX, R15 - ADDQ DI, R10 - - XORQ R14, R13 - RORXQ $34, R10, R12 - XORQ DX, R15 - RORXQ $39, R10, R14 - MOVQ R10, DI - - XORQ R12, R14 - RORXQ $28, R10, R12 - ADDQ 8*2+frame_YFER(SP), R9 - ORQ AX, DI - - XORQ R12, R14 - MOVQ R10, R12 - ANDQ R11, DI - ANDQ AX, R12 - ADDQ R13, R15 - - ADDQ R9, BX - ORQ R12, DI - ADDQ R14, R9 - - ADDQ R15, BX - - ADDQ R15, R9 - MOVQ CX, R15 - RORXQ $41, BX, R13 - RORXQ $18, BX, R14 - XORQ R8, R15 - - XORQ R14, R13 - RORXQ $14, BX, R14 - ANDQ BX, R15 - ADDQ DI, R9 - - XORQ R14, R13 - RORXQ $34, R9, R12 - XORQ R8, R15 - RORXQ $39, R9, R14 - MOVQ R9, DI - - XORQ R12, R14 - RORXQ $28, R9, R12 - ADDQ 8*3+frame_YFER(SP), DX - ORQ R11, DI - - XORQ R12, R14 - MOVQ R9, R12 - ANDQ R10, DI - ANDQ R11, R12 - ADDQ R13, R15 - - ADDQ DX, AX - ORQ R12, DI - ADDQ R14, DX - - ADDQ R15, AX - - ADDQ R15, DX - - ADDQ DI, DX - - VPADDQ 1*32(BP), Y5, Y0 - VMOVDQU Y0, frame_YFER(SP) - ADDQ $(2*32), BP - - MOVQ BX, R15 - RORXQ $41, AX, R13 - RORXQ $18, AX, R14 - XORQ CX, R15 - - XORQ R14, R13 - RORXQ $14, AX, R14 - ANDQ AX, R15 - - XORQ R14, R13 - RORXQ $34, DX, R12 - XORQ CX, R15 - RORXQ $39, DX, R14 - MOVQ DX, DI - - XORQ R12, R14 - RORXQ $28, DX, R12 - ADDQ frame_YFER(SP), R8 - ORQ R10, DI - - XORQ R12, R14 - MOVQ DX, R12 - ANDQ R9, DI - ANDQ R10, R12 - ADDQ R13, R15 - - ADDQ R8, R11 - ORQ R12, DI - ADDQ R14, R8 - - ADDQ R15, R11 - - ADDQ R15, R8 - MOVQ AX, R15 - RORXQ $41, R11, R13 - RORXQ $18, R11, R14 - XORQ BX, R15 - - XORQ R14, R13 - RORXQ $14, R11, R14 - ANDQ R11, R15 - ADDQ DI, R8 - - XORQ R14, R13 - RORXQ $34, R8, R12 - XORQ BX, R15 - RORXQ $39, R8, R14 - MOVQ R8, DI - - XORQ R12, R14 - RORXQ $28, R8, R12 - ADDQ 8*1+frame_YFER(SP), CX - ORQ R9, DI - - XORQ R12, R14 - MOVQ R8, R12 - ANDQ DX, DI - ANDQ R9, R12 - ADDQ R13, R15 - - ADDQ CX, R10 - ORQ R12, DI - ADDQ R14, CX - - ADDQ R15, R10 - - ADDQ R15, CX - MOVQ R11, R15 - RORXQ $41, R10, R13 - RORXQ $18, R10, R14 - XORQ AX, R15 - - XORQ R14, R13 - RORXQ $14, R10, R14 - ANDQ R10, R15 - ADDQ DI, CX - - XORQ R14, R13 - RORXQ $34, CX, R12 - XORQ AX, R15 - RORXQ $39, CX, R14 - MOVQ CX, DI - - XORQ R12, R14 - RORXQ $28, CX, R12 - ADDQ 8*2+frame_YFER(SP), BX - ORQ DX, DI - - XORQ R12, R14 - MOVQ CX, R12 - ANDQ R8, DI - ANDQ DX, R12 - ADDQ R13, R15 - - ADDQ BX, R9 - ORQ R12, DI - ADDQ R14, BX - - ADDQ R15, R9 - - ADDQ R15, BX - MOVQ R10, R15 - RORXQ $41, R9, R13 - RORXQ $18, R9, R14 - XORQ R11, R15 - - XORQ R14, R13 - RORXQ $14, R9, R14 - ANDQ R9, R15 - ADDQ DI, BX - - XORQ R14, R13 - RORXQ $34, BX, R12 - XORQ R11, R15 - RORXQ $39, BX, R14 - MOVQ BX, DI - - XORQ R12, R14 - RORXQ $28, BX, R12 - ADDQ 8*3+frame_YFER(SP), AX - ORQ R8, DI - - XORQ R12, R14 - MOVQ BX, R12 - ANDQ CX, DI - ANDQ R8, R12 - ADDQ R13, R15 - - ADDQ AX, DX - ORQ R12, DI - ADDQ R14, AX - - ADDQ R15, DX - - ADDQ R15, AX - - ADDQ DI, AX - + VMOVDQU Y0, (SP) + MOVQ R9, R15 + RORXQ $0x29, DX, R13 + RORXQ $0x12, DX, R14 + XORQ R10, R15 + XORQ R14, R13 + RORXQ $0x0e, DX, R14 + ANDQ DX, R15 + XORQ R14, R13 + RORXQ $0x22, AX, R12 + XORQ R10, R15 + RORXQ $0x27, AX, R14 + MOVQ AX, DI + XORQ R12, R14 + RORXQ $0x1c, AX, R12 + ADDQ (SP), R11 + ORQ CX, DI + XORQ R12, R14 + MOVQ AX, R12 + ANDQ BX, DI + ANDQ CX, R12 + ADDQ R13, R15 + ADDQ R11, R8 + ORQ R12, DI + ADDQ R14, R11 + ADDQ R15, R8 + ADDQ R15, R11 + MOVQ DX, R15 + RORXQ $0x29, R8, R13 + RORXQ $0x12, R8, R14 + XORQ R9, R15 + XORQ R14, R13 + RORXQ $0x0e, R8, R14 + ANDQ R8, R15 + ADDQ DI, R11 + XORQ R14, R13 + RORXQ $0x22, R11, R12 + XORQ R9, R15 + RORXQ $0x27, R11, R14 + MOVQ R11, DI + XORQ R12, R14 + RORXQ $0x1c, R11, R12 + ADDQ 8(SP), R10 + ORQ BX, DI + XORQ R12, R14 + MOVQ R11, R12 + ANDQ AX, DI + ANDQ BX, R12 + ADDQ R13, R15 + ADDQ R10, CX + ORQ R12, DI + ADDQ R14, R10 + ADDQ R15, CX + ADDQ R15, R10 + MOVQ R8, R15 + RORXQ $0x29, CX, R13 + RORXQ $0x12, CX, R14 + XORQ DX, R15 + XORQ R14, R13 + RORXQ $0x0e, CX, R14 + ANDQ CX, R15 + ADDQ DI, R10 + XORQ R14, R13 + RORXQ $0x22, R10, R12 + XORQ DX, R15 + RORXQ $0x27, R10, R14 + MOVQ R10, DI + XORQ R12, R14 + RORXQ $0x1c, R10, R12 + ADDQ 16(SP), R9 + ORQ AX, DI + XORQ R12, R14 + MOVQ R10, R12 + ANDQ R11, DI + ANDQ AX, R12 + ADDQ R13, R15 + ADDQ R9, BX + ORQ R12, DI + ADDQ R14, R9 + ADDQ R15, BX + ADDQ R15, R9 + MOVQ CX, R15 + RORXQ $0x29, BX, R13 + RORXQ $0x12, BX, R14 + XORQ R8, R15 + XORQ R14, R13 + RORXQ $0x0e, BX, R14 + ANDQ BX, R15 + ADDQ DI, R9 + XORQ R14, R13 + RORXQ $0x22, R9, R12 + XORQ R8, R15 + RORXQ $0x27, R9, R14 + MOVQ R9, DI + XORQ R12, R14 + RORXQ $0x1c, R9, R12 + ADDQ 24(SP), DX + ORQ R11, DI + XORQ R12, R14 + MOVQ R9, R12 + ANDQ R10, DI + ANDQ R11, R12 + ADDQ R13, R15 + ADDQ DX, AX + ORQ R12, DI + ADDQ R14, DX + ADDQ R15, AX + ADDQ R15, DX + ADDQ DI, DX + VPADDQ 32(BP), Y5, Y0 + VMOVDQU Y0, (SP) + ADDQ $0x40, BP + MOVQ BX, R15 + RORXQ $0x29, AX, R13 + RORXQ $0x12, AX, R14 + XORQ CX, R15 + XORQ R14, R13 + RORXQ $0x0e, AX, R14 + ANDQ AX, R15 + XORQ R14, R13 + RORXQ $0x22, DX, R12 + XORQ CX, R15 + RORXQ $0x27, DX, R14 + MOVQ DX, DI + XORQ R12, R14 + RORXQ $0x1c, DX, R12 + ADDQ (SP), R8 + ORQ R10, DI + XORQ R12, R14 + MOVQ DX, R12 + ANDQ R9, DI + ANDQ R10, R12 + ADDQ R13, R15 + ADDQ R8, R11 + ORQ R12, DI + ADDQ R14, R8 + ADDQ R15, R11 + ADDQ R15, R8 + MOVQ AX, R15 + RORXQ $0x29, R11, R13 + RORXQ $0x12, R11, R14 + XORQ BX, R15 + XORQ R14, R13 + RORXQ $0x0e, R11, R14 + ANDQ R11, R15 + ADDQ DI, R8 + XORQ R14, R13 + RORXQ $0x22, R8, R12 + XORQ BX, R15 + RORXQ $0x27, R8, R14 + MOVQ R8, DI + XORQ R12, R14 + RORXQ $0x1c, R8, R12 + ADDQ 8(SP), CX + ORQ R9, DI + XORQ R12, R14 + MOVQ R8, R12 + ANDQ DX, DI + ANDQ R9, R12 + ADDQ R13, R15 + ADDQ CX, R10 + ORQ R12, DI + ADDQ R14, CX + ADDQ R15, R10 + ADDQ R15, CX + MOVQ R11, R15 + RORXQ $0x29, R10, R13 + RORXQ $0x12, R10, R14 + XORQ AX, R15 + XORQ R14, R13 + RORXQ $0x0e, R10, R14 + ANDQ R10, R15 + ADDQ DI, CX + XORQ R14, R13 + RORXQ $0x22, CX, R12 + XORQ AX, R15 + RORXQ $0x27, CX, R14 + MOVQ CX, DI + XORQ R12, R14 + RORXQ $0x1c, CX, R12 + ADDQ 16(SP), BX + ORQ DX, DI + XORQ R12, R14 + MOVQ CX, R12 + ANDQ R8, DI + ANDQ DX, R12 + ADDQ R13, R15 + ADDQ BX, R9 + ORQ R12, DI + ADDQ R14, BX + ADDQ R15, R9 + ADDQ R15, BX + MOVQ R10, R15 + RORXQ $0x29, R9, R13 + RORXQ $0x12, R9, R14 + XORQ R11, R15 + XORQ R14, R13 + RORXQ $0x0e, R9, R14 + ANDQ R9, R15 + ADDQ DI, BX + XORQ R14, R13 + RORXQ $0x22, BX, R12 + XORQ R11, R15 + RORXQ $0x27, BX, R14 + MOVQ BX, DI + XORQ R12, R14 + RORXQ $0x1c, BX, R12 + ADDQ 24(SP), AX + ORQ R8, DI + XORQ R12, R14 + MOVQ BX, R12 + ANDQ CX, DI + ANDQ R8, R12 + ADDQ R13, R15 + ADDQ AX, DX + ORQ R12, DI + ADDQ R14, AX + ADDQ R15, DX + ADDQ R15, AX + ADDQ DI, AX VMOVDQU Y6, Y4 VMOVDQU Y7, Y5 - - SUBQ $1, frame_SRND(SP) - JNE loop2 - - addm(8*0(SI),AX) - addm(8*1(SI),BX) - addm(8*2(SI),CX) - addm(8*3(SI),R8) - addm(8*4(SI),DX) - addm(8*5(SI),R9) - addm(8*6(SI),R10) - addm(8*7(SI),R11) - - MOVQ frame_INP(SP), DI - ADDQ $128, DI - CMPQ DI, frame_INPEND(SP) - JNE loop0 + SUBQ $0x01, 32(SP) + JNE loop2 + ADDQ (SI), AX + MOVQ AX, (SI) + ADDQ 8(SI), BX + MOVQ BX, 8(SI) + ADDQ 16(SI), CX + MOVQ CX, 16(SI) + ADDQ 24(SI), R8 + MOVQ R8, 24(SI) + ADDQ 32(SI), DX + MOVQ DX, 32(SI) + ADDQ 40(SI), R9 + MOVQ R9, 40(SI) + ADDQ 48(SI), R10 + MOVQ R10, 48(SI) + ADDQ 56(SI), R11 + MOVQ R11, 56(SI) + MOVQ 40(SP), DI + ADDQ $0x80, DI + CMPQ DI, 48(SP) + JNE loop0 done_hash: VZEROUPPER RET + +DATA MASK_YMM_LO<>+0(SB)/8, $0x0000000000000000 +DATA MASK_YMM_LO<>+8(SB)/8, $0x0000000000000000 +DATA MASK_YMM_LO<>+16(SB)/8, $0xffffffffffffffff +DATA MASK_YMM_LO<>+24(SB)/8, $0xffffffffffffffff +GLOBL MASK_YMM_LO<>(SB), RODATA|NOPTR, $32 diff --git a/src/go/types/stdlib_test.go b/src/go/types/stdlib_test.go index ac38826ecb6..b82fd0ddc1a 100644 --- a/src/go/types/stdlib_test.go +++ b/src/go/types/stdlib_test.go @@ -361,6 +361,7 @@ var excluded = map[string]bool{ "crypto/internal/edwards25519/field/_asm": true, "crypto/sha1/_asm": true, "crypto/sha256/_asm": true, + "crypto/sha512/_asm": true, } // printPackageMu synchronizes the printing of type-checked package files in