// Copyright 2012 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.

// This code was translated into a form compatible with 6a from the public
// domain sources in SUPERCOP: http://bench.cr.yp.to/supercop.html

// +build amd64,!gccgo,!appengine

// func poly1305(out *[16]byte, m *byte, mlen uint64, key *[32]key)
TEXT ·poly1305(SB),0,$224-32
	MOVQ out+0(FP),DI
	MOVQ m+8(FP),SI
	MOVQ mlen+16(FP),DX
	MOVQ key+24(FP),CX

	MOVQ SP,R11
	MOVQ $31,R9
	NOTQ R9
	ANDQ R9,SP
	ADDQ $32,SP

	MOVQ R11,32(SP)
	MOVQ R12,40(SP)
	MOVQ R13,48(SP)
	MOVQ R14,56(SP)
	MOVQ R15,64(SP)
	MOVQ BX,72(SP)
	MOVQ BP,80(SP)
	FLDCW ·ROUNDING(SB)
	MOVL 0(CX),R8
	MOVL 4(CX),R9
	MOVL 8(CX),AX
	MOVL 12(CX),R10
	MOVQ DI,88(SP)
	MOVQ CX,96(SP)
	MOVL $0X43300000,108(SP)
	MOVL $0X45300000,116(SP)
	MOVL $0X47300000,124(SP)
	MOVL $0X49300000,132(SP)
	ANDL $0X0FFFFFFF,R8
	ANDL $0X0FFFFFFC,R9
	ANDL $0X0FFFFFFC,AX
	ANDL $0X0FFFFFFC,R10
	MOVL R8,104(SP)
	MOVL R9,112(SP)
	MOVL AX,120(SP)
	MOVL R10,128(SP)
	FMOVD 104(SP), F0
	FSUBD ·DOFFSET0(SB), F0
	FMOVD 112(SP), F0
	FSUBD ·DOFFSET1(SB), F0
	FMOVD 120(SP), F0
	FSUBD ·DOFFSET2(SB), F0
	FMOVD 128(SP), F0
	FSUBD ·DOFFSET3(SB), F0
	FXCHD F0, F3
	FMOVDP F0, 136(SP)
	FXCHD F0, F1
	FMOVD F0, 144(SP)
	FMULD ·SCALE(SB), F0
	FMOVDP F0, 152(SP)
	FMOVD F0, 160(SP)
	FMULD ·SCALE(SB), F0
	FMOVDP F0, 168(SP)
	FMOVD F0, 176(SP)
	FMULD ·SCALE(SB), F0
	FMOVDP F0, 184(SP)
	FLDZ
	FLDZ
	FLDZ
	FLDZ
	CMPQ DX,$16
	JB ADDATMOST15BYTES
	INITIALATLEAST16BYTES:
	MOVL 12(SI),DI
	MOVL 8(SI),CX
	MOVL 4(SI),R8
	MOVL 0(SI),R9
	MOVL DI,128(SP)
	MOVL CX,120(SP)
	MOVL R8,112(SP)
	MOVL R9,104(SP)
	ADDQ $16,SI
	SUBQ $16,DX
	FXCHD F0, F3
	FADDD 128(SP), F0
	FSUBD ·DOFFSET3MINUSTWO128(SB), F0
	FXCHD F0, F1
	FADDD 112(SP), F0
	FSUBD ·DOFFSET1(SB), F0
	FXCHD F0, F2
	FADDD 120(SP), F0
	FSUBD ·DOFFSET2(SB), F0
	FXCHD F0, F3
	FADDD 104(SP), F0
	FSUBD ·DOFFSET0(SB), F0
	CMPQ DX,$16
	JB MULTIPLYADDATMOST15BYTES
	MULTIPLYADDATLEAST16BYTES:
	MOVL 12(SI),DI
	MOVL 8(SI),CX
	MOVL 4(SI),R8
	MOVL 0(SI),R9
	MOVL DI,128(SP)
	MOVL CX,120(SP)
	MOVL R8,112(SP)
	MOVL R9,104(SP)
	ADDQ $16,SI
	SUBQ $16,DX
	FMOVD ·ALPHA130(SB), F0
	FADDD F2,F0
	FSUBD ·ALPHA130(SB), F0
	FSUBD F0,F2
	FMULD ·SCALE(SB), F0
	FMOVD ·ALPHA32(SB), F0
	FADDD F2,F0
	FSUBD ·ALPHA32(SB), F0
	FSUBD F0,F2
	FXCHD F0, F2
	FADDDP F0,F1
	FMOVD ·ALPHA64(SB), F0
	FADDD F4,F0
	FSUBD ·ALPHA64(SB), F0
	FSUBD F0,F4
	FMOVD ·ALPHA96(SB), F0
	FADDD F6,F0
	FSUBD ·ALPHA96(SB), F0
	FSUBD F0,F6
	FXCHD F0, F6
	FADDDP F0,F1
	FXCHD F0, F3
	FADDDP F0,F5
	FXCHD F0, F3
	FADDDP F0,F1
	FMOVD 176(SP), F0
	FMULD F3,F0
	FMOVD 160(SP), F0
	FMULD F4,F0
	FMOVD 144(SP), F0
	FMULD F5,F0
	FMOVD 136(SP), F0
	FMULDP F0,F6
	FMOVD 160(SP), F0
	FMULD F4,F0
	FADDDP F0,F3
	FMOVD 144(SP), F0
	FMULD F4,F0
	FADDDP F0,F2
	FMOVD 136(SP), F0
	FMULD F4,F0
	FADDDP F0,F1
	FMOVD 184(SP), F0
	FMULDP F0,F4
	FXCHD F0, F3
	FADDDP F0,F5
	FMOVD 144(SP), F0
	FMULD F4,F0
	FADDDP F0,F2
	FMOVD 136(SP), F0
	FMULD F4,F0
	FADDDP F0,F1
	FMOVD 184(SP), F0
	FMULD F4,F0
	FADDDP F0,F3
	FMOVD 168(SP), F0
	FMULDP F0,F4
	FXCHD F0, F3
	FADDDP F0,F4
	FMOVD 136(SP), F0
	FMULD F5,F0
	FADDDP F0,F1
	FXCHD F0, F3
	FMOVD 184(SP), F0
	FMULD F5,F0
	FADDDP F0,F3
	FXCHD F0, F1
	FMOVD 168(SP), F0
	FMULD F5,F0
	FADDDP F0,F1
	FMOVD 152(SP), F0
	FMULDP F0,F5
	FXCHD F0, F4
	FADDDP F0,F1
	CMPQ DX,$16
	FXCHD F0, F2
	FMOVD 128(SP), F0
	FSUBD ·DOFFSET3MINUSTWO128(SB), F0
	FADDDP F0,F1
	FXCHD F0, F1
	FMOVD 120(SP), F0
	FSUBD ·DOFFSET2(SB), F0
	FADDDP F0,F1
	FXCHD F0, F3
	FMOVD 112(SP), F0
	FSUBD ·DOFFSET1(SB), F0
	FADDDP F0,F1
	FXCHD F0, F2
	FMOVD 104(SP), F0
	FSUBD ·DOFFSET0(SB), F0
	FADDDP F0,F1
	JAE MULTIPLYADDATLEAST16BYTES
	MULTIPLYADDATMOST15BYTES:
	FMOVD ·ALPHA130(SB), F0
	FADDD F2,F0
	FSUBD ·ALPHA130(SB), F0
	FSUBD F0,F2
	FMULD ·SCALE(SB), F0
	FMOVD ·ALPHA32(SB), F0
	FADDD F2,F0
	FSUBD ·ALPHA32(SB), F0
	FSUBD F0,F2
	FMOVD ·ALPHA64(SB), F0
	FADDD F5,F0
	FSUBD ·ALPHA64(SB), F0
	FSUBD F0,F5
	FMOVD ·ALPHA96(SB), F0
	FADDD F7,F0
	FSUBD ·ALPHA96(SB), F0
	FSUBD F0,F7
	FXCHD F0, F7
	FADDDP F0,F1
	FXCHD F0, F5
	FADDDP F0,F1
	FXCHD F0, F3
	FADDDP F0,F5
	FADDDP F0,F1
	FMOVD 176(SP), F0
	FMULD F1,F0
	FMOVD 160(SP), F0
	FMULD F2,F0
	FMOVD 144(SP), F0
	FMULD F3,F0
	FMOVD 136(SP), F0
	FMULDP F0,F4
	FMOVD 160(SP), F0
	FMULD F5,F0
	FADDDP F0,F3
	FMOVD 144(SP), F0
	FMULD F5,F0
	FADDDP F0,F2
	FMOVD 136(SP), F0
	FMULD F5,F0
	FADDDP F0,F1
	FMOVD 184(SP), F0
	FMULDP F0,F5
	FXCHD F0, F4
	FADDDP F0,F3
	FMOVD 144(SP), F0
	FMULD F5,F0
	FADDDP F0,F2
	FMOVD 136(SP), F0
	FMULD F5,F0
	FADDDP F0,F1
	FMOVD 184(SP), F0
	FMULD F5,F0
	FADDDP F0,F4
	FMOVD 168(SP), F0
	FMULDP F0,F5
	FXCHD F0, F4
	FADDDP F0,F2
	FMOVD 136(SP), F0
	FMULD F5,F0
	FADDDP F0,F1
	FMOVD 184(SP), F0
	FMULD F5,F0
	FADDDP F0,F4
	FMOVD 168(SP), F0
	FMULD F5,F0
	FADDDP F0,F3
	FMOVD 152(SP), F0
	FMULDP F0,F5
	FXCHD F0, F4
	FADDDP F0,F1
	ADDATMOST15BYTES:
	CMPQ DX,$0
	JE NOMOREBYTES
	MOVL $0,0(SP)
	MOVL $0, 4 (SP)
	MOVL $0, 8 (SP)
	MOVL $0, 12 (SP)
	LEAQ 0(SP),DI
	MOVQ DX,CX
	REP; MOVSB
	MOVB $1,0(DI)
	MOVL  12 (SP),DI
	MOVL  8 (SP),SI
	MOVL  4 (SP),DX
	MOVL 0(SP),CX
	MOVL DI,128(SP)
	MOVL SI,120(SP)
	MOVL DX,112(SP)
	MOVL CX,104(SP)
	FXCHD F0, F3
	FADDD 128(SP), F0
	FSUBD ·DOFFSET3(SB), F0
	FXCHD F0, F2
	FADDD 120(SP), F0
	FSUBD ·DOFFSET2(SB), F0
	FXCHD F0, F1
	FADDD 112(SP), F0
	FSUBD ·DOFFSET1(SB), F0
	FXCHD F0, F3
	FADDD 104(SP), F0
	FSUBD ·DOFFSET0(SB), F0
	FMOVD ·ALPHA130(SB), F0
	FADDD F3,F0
	FSUBD ·ALPHA130(SB), F0
	FSUBD F0,F3
	FMULD ·SCALE(SB), F0
	FMOVD ·ALPHA32(SB), F0
	FADDD F2,F0
	FSUBD ·ALPHA32(SB), F0
	FSUBD F0,F2
	FMOVD ·ALPHA64(SB), F0
	FADDD F6,F0
	FSUBD ·ALPHA64(SB), F0
	FSUBD F0,F6
	FMOVD ·ALPHA96(SB), F0
	FADDD F5,F0
	FSUBD ·ALPHA96(SB), F0
	FSUBD F0,F5
	FXCHD F0, F4
	FADDDP F0,F3
	FXCHD F0, F6
	FADDDP F0,F1
	FXCHD F0, F3
	FADDDP F0,F5
	FXCHD F0, F3
	FADDDP F0,F1
	FMOVD 176(SP), F0
	FMULD F3,F0
	FMOVD 160(SP), F0
	FMULD F4,F0
	FMOVD 144(SP), F0
	FMULD F5,F0
	FMOVD 136(SP), F0
	FMULDP F0,F6
	FMOVD 160(SP), F0
	FMULD F5,F0
	FADDDP F0,F3
	FMOVD 144(SP), F0
	FMULD F5,F0
	FADDDP F0,F2
	FMOVD 136(SP), F0
	FMULD F5,F0
	FADDDP F0,F1
	FMOVD 184(SP), F0
	FMULDP F0,F5
	FXCHD F0, F4
	FADDDP F0,F5
	FMOVD 144(SP), F0
	FMULD F6,F0
	FADDDP F0,F2
	FMOVD 136(SP), F0
	FMULD F6,F0
	FADDDP F0,F1
	FMOVD 184(SP), F0
	FMULD F6,F0
	FADDDP F0,F4
	FMOVD 168(SP), F0
	FMULDP F0,F6
	FXCHD F0, F5
	FADDDP F0,F4
	FMOVD 136(SP), F0
	FMULD F2,F0
	FADDDP F0,F1
	FMOVD 184(SP), F0
	FMULD F2,F0
	FADDDP F0,F5
	FMOVD 168(SP), F0
	FMULD F2,F0
	FADDDP F0,F3
	FMOVD 152(SP), F0
	FMULDP F0,F2
	FXCHD F0, F1
	FADDDP F0,F3
	FXCHD F0, F3
	FXCHD F0, F2
	NOMOREBYTES:
	MOVL $0,R10
	FMOVD ·ALPHA130(SB), F0
	FADDD F4,F0
	FSUBD ·ALPHA130(SB), F0
	FSUBD F0,F4
	FMULD ·SCALE(SB), F0
	FMOVD ·ALPHA32(SB), F0
	FADDD F2,F0
	FSUBD ·ALPHA32(SB), F0
	FSUBD F0,F2
	FMOVD ·ALPHA64(SB), F0
	FADDD F4,F0
	FSUBD ·ALPHA64(SB), F0
	FSUBD F0,F4
	FMOVD ·ALPHA96(SB), F0
	FADDD F6,F0
	FSUBD ·ALPHA96(SB), F0
	FXCHD F0, F6
	FSUBD F6,F0
	FXCHD F0, F4
	FADDDP F0,F3
	FXCHD F0, F4
	FADDDP F0,F1
	FXCHD F0, F2
	FADDDP F0,F3
	FXCHD F0, F4
	FADDDP F0,F3
	FXCHD F0, F3
	FADDD ·HOFFSET0(SB), F0
	FXCHD F0, F3
	FADDD ·HOFFSET1(SB), F0
	FXCHD F0, F1
	FADDD ·HOFFSET2(SB), F0
	FXCHD F0, F2
	FADDD ·HOFFSET3(SB), F0
	FXCHD F0, F3
	FMOVDP F0, 104(SP)
	FMOVDP F0, 112(SP)
	FMOVDP F0, 120(SP)
	FMOVDP F0, 128(SP)
	MOVL 108(SP),DI
	ANDL $63,DI
	MOVL 116(SP),SI
	ANDL $63,SI
	MOVL 124(SP),DX
	ANDL $63,DX
	MOVL 132(SP),CX
	ANDL $63,CX
	MOVL 112(SP),R8
	ADDL DI,R8
	MOVQ R8,112(SP)
	MOVL 120(SP),DI
	ADCL SI,DI
	MOVQ DI,120(SP)
	MOVL 128(SP),DI
	ADCL DX,DI
	MOVQ DI,128(SP)
	MOVL R10,DI
	ADCL CX,DI
	MOVQ DI,136(SP)
	MOVQ $5,DI
	MOVL 104(SP),SI
	ADDL SI,DI
	MOVQ DI,104(SP)
	MOVL R10,DI
	MOVQ 112(SP),DX
	ADCL DX,DI
	MOVQ DI,112(SP)
	MOVL R10,DI
	MOVQ 120(SP),CX
	ADCL CX,DI
	MOVQ DI,120(SP)
	MOVL R10,DI
	MOVQ 128(SP),R8
	ADCL R8,DI
	MOVQ DI,128(SP)
	MOVQ $0XFFFFFFFC,DI
	MOVQ 136(SP),R9
	ADCL R9,DI
	SARL $16,DI
	MOVQ DI,R9
	XORL $0XFFFFFFFF,R9
	ANDQ DI,SI
	MOVQ 104(SP),AX
	ANDQ R9,AX
	ORQ AX,SI
	ANDQ DI,DX
	MOVQ 112(SP),AX
	ANDQ R9,AX
	ORQ AX,DX
	ANDQ DI,CX
	MOVQ 120(SP),AX
	ANDQ R9,AX
	ORQ AX,CX
	ANDQ DI,R8
	MOVQ 128(SP),DI
	ANDQ R9,DI
	ORQ DI,R8
	MOVQ 88(SP),DI
	MOVQ 96(SP),R9
	ADDL 16(R9),SI
	ADCL 20(R9),DX
	ADCL 24(R9),CX
	ADCL 28(R9),R8
	MOVL SI,0(DI)
	MOVL DX,4(DI)
	MOVL CX,8(DI)
	MOVL R8,12(DI)
	MOVQ 32(SP),R11
	MOVQ 40(SP),R12
	MOVQ 48(SP),R13
	MOVQ 56(SP),R14
	MOVQ 64(SP),R15
	MOVQ 72(SP),BX
	MOVQ 80(SP),BP
	MOVQ R11,SP
	RET