Bundle a copy of golang.org/x/crypto/curve25519 for testing.

Hopefully this can be replaced with a standard library version later. BUG=571231 Change-Id: I61ae1d9d057c6d9e1b92128042109758beccc7ff Reviewed-on: https://boringssl-review.googlesource.com/6776 Reviewed-by: Adam Langley <agl@google.com>
2015-12-18 21:15:22 -05:00 · 2015-12-18 21:15:22 -05:00 · ab14563022
commit ab14563022
parent a029ebc4c6
10 changed files with 3077 additions and 0 deletions
--- a/ssl/test/runner/curve25519/const_amd64.s
+++ b/ssl/test/runner/curve25519/const_amd64.s
@ -0,0 +1,20 @@
+// 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
+
+DATA ·REDMASK51(SB)/8, $0x0007FFFFFFFFFFFF
+GLOBL ·REDMASK51(SB), 8, $8
+
+DATA ·_121666_213(SB)/8, $996687872
+GLOBL ·_121666_213(SB), 8, $8
+
+DATA ·_2P0(SB)/8, $0xFFFFFFFFFFFDA
+GLOBL ·_2P0(SB), 8, $8
+
+DATA ·_2P1234(SB)/8, $0xFFFFFFFFFFFFE
+GLOBL ·_2P1234(SB), 8, $8
--- a/ssl/test/runner/curve25519/cswap_amd64.s
+++ b/ssl/test/runner/curve25519/cswap_amd64.s
@ -0,0 +1,88 @@
+// 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 cswap(inout *[5]uint64, v uint64)
+TEXT ·cswap(SB),7,$0
+	MOVQ inout+0(FP),DI
+	MOVQ v+8(FP),SI
+
+	CMPQ SI,$1
+	MOVQ 0(DI),SI
+	MOVQ 80(DI),DX
+	MOVQ 8(DI),CX
+	MOVQ 88(DI),R8
+	MOVQ SI,R9
+	CMOVQEQ DX,SI
+	CMOVQEQ R9,DX
+	MOVQ CX,R9
+	CMOVQEQ R8,CX
+	CMOVQEQ R9,R8
+	MOVQ SI,0(DI)
+	MOVQ DX,80(DI)
+	MOVQ CX,8(DI)
+	MOVQ R8,88(DI)
+	MOVQ 16(DI),SI
+	MOVQ 96(DI),DX
+	MOVQ 24(DI),CX
+	MOVQ 104(DI),R8
+	MOVQ SI,R9
+	CMOVQEQ DX,SI
+	CMOVQEQ R9,DX
+	MOVQ CX,R9
+	CMOVQEQ R8,CX
+	CMOVQEQ R9,R8
+	MOVQ SI,16(DI)
+	MOVQ DX,96(DI)
+	MOVQ CX,24(DI)
+	MOVQ R8,104(DI)
+	MOVQ 32(DI),SI
+	MOVQ 112(DI),DX
+	MOVQ 40(DI),CX
+	MOVQ 120(DI),R8
+	MOVQ SI,R9
+	CMOVQEQ DX,SI
+	CMOVQEQ R9,DX
+	MOVQ CX,R9
+	CMOVQEQ R8,CX
+	CMOVQEQ R9,R8
+	MOVQ SI,32(DI)
+	MOVQ DX,112(DI)
+	MOVQ CX,40(DI)
+	MOVQ R8,120(DI)
+	MOVQ 48(DI),SI
+	MOVQ 128(DI),DX
+	MOVQ 56(DI),CX
+	MOVQ 136(DI),R8
+	MOVQ SI,R9
+	CMOVQEQ DX,SI
+	CMOVQEQ R9,DX
+	MOVQ CX,R9
+	CMOVQEQ R8,CX
+	CMOVQEQ R9,R8
+	MOVQ SI,48(DI)
+	MOVQ DX,128(DI)
+	MOVQ CX,56(DI)
+	MOVQ R8,136(DI)
+	MOVQ 64(DI),SI
+	MOVQ 144(DI),DX
+	MOVQ 72(DI),CX
+	MOVQ 152(DI),R8
+	MOVQ SI,R9
+	CMOVQEQ DX,SI
+	CMOVQEQ R9,DX
+	MOVQ CX,R9
+	CMOVQEQ R8,CX
+	CMOVQEQ R9,R8
+	MOVQ SI,64(DI)
+	MOVQ DX,144(DI)
+	MOVQ CX,72(DI)
+	MOVQ R8,152(DI)
+	MOVQ DI,AX
+	MOVQ SI,DX
+	RET
--- a/ssl/test/runner/curve25519/curve25519.go
+++ b/ssl/test/runner/curve25519/curve25519.go
@ -0,0 +1,841 @@
+// 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.
+
+// We have a implementation in amd64 assembly so this code is only run on
+// non-amd64 platforms. The amd64 assembly does not support gccgo.
+// +build !amd64 gccgo appengine
+
+package curve25519
+
+// This code is a port of the public domain, "ref10" implementation of
+// curve25519 from SUPERCOP 20130419 by D. J. Bernstein.
+
+// fieldElement represents an element of the field GF(2^255 - 19). An element
+// t, entries t[0]...t[9], represents the integer t[0]+2^26 t[1]+2^51 t[2]+2^77
+// t[3]+2^102 t[4]+...+2^230 t[9]. Bounds on each t[i] vary depending on
+// context.
+type fieldElement [10]int32
+
+func feZero(fe *fieldElement) {
+	for i := range fe {
+		fe[i] = 0
+	}
+}
+
+func feOne(fe *fieldElement) {
+	feZero(fe)
+	fe[0] = 1
+}
+
+func feAdd(dst, a, b *fieldElement) {
+	for i := range dst {
+		dst[i] = a[i] + b[i]
+	}
+}
+
+func feSub(dst, a, b *fieldElement) {
+	for i := range dst {
+		dst[i] = a[i] - b[i]
+	}
+}
+
+func feCopy(dst, src *fieldElement) {
+	for i := range dst {
+		dst[i] = src[i]
+	}
+}
+
+// feCSwap replaces (f,g) with (g,f) if b == 1; replaces (f,g) with (f,g) if b == 0.
+//
+// Preconditions: b in {0,1}.
+func feCSwap(f, g *fieldElement, b int32) {
+	var x fieldElement
+	b = -b
+	for i := range x {
+		x[i] = b & (f[i] ^ g[i])
+	}
+
+	for i := range f {
+		f[i] ^= x[i]
+	}
+	for i := range g {
+		g[i] ^= x[i]
+	}
+}
+
+// load3 reads a 24-bit, little-endian value from in.
+func load3(in []byte) int64 {
+	var r int64
+	r = int64(in[0])
+	r |= int64(in[1]) << 8
+	r |= int64(in[2]) << 16
+	return r
+}
+
+// load4 reads a 32-bit, little-endian value from in.
+func load4(in []byte) int64 {
+	var r int64
+	r = int64(in[0])
+	r |= int64(in[1]) << 8
+	r |= int64(in[2]) << 16
+	r |= int64(in[3]) << 24
+	return r
+}
+
+func feFromBytes(dst *fieldElement, src *[32]byte) {
+	h0 := load4(src[:])
+	h1 := load3(src[4:]) << 6
+	h2 := load3(src[7:]) << 5
+	h3 := load3(src[10:]) << 3
+	h4 := load3(src[13:]) << 2
+	h5 := load4(src[16:])
+	h6 := load3(src[20:]) << 7
+	h7 := load3(src[23:]) << 5
+	h8 := load3(src[26:]) << 4
+	h9 := load3(src[29:]) << 2
+
+	var carry [10]int64
+	carry[9] = (h9 + 1<<24) >> 25
+	h0 += carry[9] * 19
+	h9 -= carry[9] << 25
+	carry[1] = (h1 + 1<<24) >> 25
+	h2 += carry[1]
+	h1 -= carry[1] << 25
+	carry[3] = (h3 + 1<<24) >> 25
+	h4 += carry[3]
+	h3 -= carry[3] << 25
+	carry[5] = (h5 + 1<<24) >> 25
+	h6 += carry[5]
+	h5 -= carry[5] << 25
+	carry[7] = (h7 + 1<<24) >> 25
+	h8 += carry[7]
+	h7 -= carry[7] << 25
+
+	carry[0] = (h0 + 1<<25) >> 26
+	h1 += carry[0]
+	h0 -= carry[0] << 26
+	carry[2] = (h2 + 1<<25) >> 26
+	h3 += carry[2]
+	h2 -= carry[2] << 26
+	carry[4] = (h4 + 1<<25) >> 26
+	h5 += carry[4]
+	h4 -= carry[4] << 26
+	carry[6] = (h6 + 1<<25) >> 26
+	h7 += carry[6]
+	h6 -= carry[6] << 26
+	carry[8] = (h8 + 1<<25) >> 26
+	h9 += carry[8]
+	h8 -= carry[8] << 26
+
+	dst[0] = int32(h0)
+	dst[1] = int32(h1)
+	dst[2] = int32(h2)
+	dst[3] = int32(h3)
+	dst[4] = int32(h4)
+	dst[5] = int32(h5)
+	dst[6] = int32(h6)
+	dst[7] = int32(h7)
+	dst[8] = int32(h8)
+	dst[9] = int32(h9)
+}
+
+// feToBytes marshals h to s.
+// Preconditions:
+//   |h| bounded by 1.1*2^25,1.1*2^24,1.1*2^25,1.1*2^24,etc.
+//
+// Write p=2^255-19; q=floor(h/p).
+// Basic claim: q = floor(2^(-255)(h + 19 2^(-25)h9 + 2^(-1))).
+//
+// Proof:
+//   Have |h|<=p so |q|<=1 so |19^2 2^(-255) q|<1/4.
+//   Also have |h-2^230 h9|<2^230 so |19 2^(-255)(h-2^230 h9)|<1/4.
+//
+//   Write y=2^(-1)-19^2 2^(-255)q-19 2^(-255)(h-2^230 h9).
+//   Then 0<y<1.
+//
+//   Write r=h-pq.
+//   Have 0<=r<=p-1=2^255-20.
+//   Thus 0<=r+19(2^-255)r<r+19(2^-255)2^255<=2^255-1.
+//
+//   Write x=r+19(2^-255)r+y.
+//   Then 0<x<2^255 so floor(2^(-255)x) = 0 so floor(q+2^(-255)x) = q.
+//
+//   Have q+2^(-255)x = 2^(-255)(h + 19 2^(-25) h9 + 2^(-1))
+//   so floor(2^(-255)(h + 19 2^(-25) h9 + 2^(-1))) = q.
+func feToBytes(s *[32]byte, h *fieldElement) {
+	var carry [10]int32
+
+	q := (19*h[9] + (1 << 24)) >> 25
+	q = (h[0] + q) >> 26
+	q = (h[1] + q) >> 25
+	q = (h[2] + q) >> 26
+	q = (h[3] + q) >> 25
+	q = (h[4] + q) >> 26
+	q = (h[5] + q) >> 25
+	q = (h[6] + q) >> 26
+	q = (h[7] + q) >> 25
+	q = (h[8] + q) >> 26
+	q = (h[9] + q) >> 25
+
+	// Goal: Output h-(2^255-19)q, which is between 0 and 2^255-20.
+	h[0] += 19 * q
+	// Goal: Output h-2^255 q, which is between 0 and 2^255-20.
+
+	carry[0] = h[0] >> 26
+	h[1] += carry[0]
+	h[0] -= carry[0] << 26
+	carry[1] = h[1] >> 25
+	h[2] += carry[1]
+	h[1] -= carry[1] << 25
+	carry[2] = h[2] >> 26
+	h[3] += carry[2]
+	h[2] -= carry[2] << 26
+	carry[3] = h[3] >> 25
+	h[4] += carry[3]
+	h[3] -= carry[3] << 25
+	carry[4] = h[4] >> 26
+	h[5] += carry[4]
+	h[4] -= carry[4] << 26
+	carry[5] = h[5] >> 25
+	h[6] += carry[5]
+	h[5] -= carry[5] << 25
+	carry[6] = h[6] >> 26
+	h[7] += carry[6]
+	h[6] -= carry[6] << 26
+	carry[7] = h[7] >> 25
+	h[8] += carry[7]
+	h[7] -= carry[7] << 25
+	carry[8] = h[8] >> 26
+	h[9] += carry[8]
+	h[8] -= carry[8] << 26
+	carry[9] = h[9] >> 25
+	h[9] -= carry[9] << 25
+	// h10 = carry9
+
+	// Goal: Output h[0]+...+2^255 h10-2^255 q, which is between 0 and 2^255-20.
+	// Have h[0]+...+2^230 h[9] between 0 and 2^255-1;
+	// evidently 2^255 h10-2^255 q = 0.
+	// Goal: Output h[0]+...+2^230 h[9].
+
+	s[0] = byte(h[0] >> 0)
+	s[1] = byte(h[0] >> 8)
+	s[2] = byte(h[0] >> 16)
+	s[3] = byte((h[0] >> 24) | (h[1] << 2))
+	s[4] = byte(h[1] >> 6)
+	s[5] = byte(h[1] >> 14)
+	s[6] = byte((h[1] >> 22) | (h[2] << 3))
+	s[7] = byte(h[2] >> 5)
+	s[8] = byte(h[2] >> 13)
+	s[9] = byte((h[2] >> 21) | (h[3] << 5))
+	s[10] = byte(h[3] >> 3)
+	s[11] = byte(h[3] >> 11)
+	s[12] = byte((h[3] >> 19) | (h[4] << 6))
+	s[13] = byte(h[4] >> 2)
+	s[14] = byte(h[4] >> 10)
+	s[15] = byte(h[4] >> 18)
+	s[16] = byte(h[5] >> 0)
+	s[17] = byte(h[5] >> 8)
+	s[18] = byte(h[5] >> 16)
+	s[19] = byte((h[5] >> 24) | (h[6] << 1))
+	s[20] = byte(h[6] >> 7)
+	s[21] = byte(h[6] >> 15)
+	s[22] = byte((h[6] >> 23) | (h[7] << 3))
+	s[23] = byte(h[7] >> 5)
+	s[24] = byte(h[7] >> 13)
+	s[25] = byte((h[7] >> 21) | (h[8] << 4))
+	s[26] = byte(h[8] >> 4)
+	s[27] = byte(h[8] >> 12)
+	s[28] = byte((h[8] >> 20) | (h[9] << 6))
+	s[29] = byte(h[9] >> 2)
+	s[30] = byte(h[9] >> 10)
+	s[31] = byte(h[9] >> 18)
+}
+
+// feMul calculates h = f * g
+// Can overlap h with f or g.
+//
+// Preconditions:
+//    |f| bounded by 1.1*2^26,1.1*2^25,1.1*2^26,1.1*2^25,etc.
+//    |g| bounded by 1.1*2^26,1.1*2^25,1.1*2^26,1.1*2^25,etc.
+//
+// Postconditions:
+//    |h| bounded by 1.1*2^25,1.1*2^24,1.1*2^25,1.1*2^24,etc.
+//
+// Notes on implementation strategy:
+//
+// Using schoolbook multiplication.
+// Karatsuba would save a little in some cost models.
+//
+// Most multiplications by 2 and 19 are 32-bit precomputations;
+// cheaper than 64-bit postcomputations.
+//
+// There is one remaining multiplication by 19 in the carry chain;
+// one *19 precomputation can be merged into this,
+// but the resulting data flow is considerably less clean.
+//
+// There are 12 carries below.
+// 10 of them are 2-way parallelizable and vectorizable.
+// Can get away with 11 carries, but then data flow is much deeper.
+//
+// With tighter constraints on inputs can squeeze carries into int32.
+func feMul(h, f, g *fieldElement) {
+	f0 := f[0]
+	f1 := f[1]
+	f2 := f[2]
+	f3 := f[3]
+	f4 := f[4]
+	f5 := f[5]
+	f6 := f[6]
+	f7 := f[7]
+	f8 := f[8]
+	f9 := f[9]
+	g0 := g[0]
+	g1 := g[1]
+	g2 := g[2]
+	g3 := g[3]
+	g4 := g[4]
+	g5 := g[5]
+	g6 := g[6]
+	g7 := g[7]
+	g8 := g[8]
+	g9 := g[9]
+	g1_19 := 19 * g1 // 1.4*2^29
+	g2_19 := 19 * g2 // 1.4*2^30; still ok
+	g3_19 := 19 * g3
+	g4_19 := 19 * g4
+	g5_19 := 19 * g5
+	g6_19 := 19 * g6
+	g7_19 := 19 * g7
+	g8_19 := 19 * g8
+	g9_19 := 19 * g9
+	f1_2 := 2 * f1
+	f3_2 := 2 * f3
+	f5_2 := 2 * f5
+	f7_2 := 2 * f7
+	f9_2 := 2 * f9
+	f0g0 := int64(f0) * int64(g0)
+	f0g1 := int64(f0) * int64(g1)
+	f0g2 := int64(f0) * int64(g2)
+	f0g3 := int64(f0) * int64(g3)
+	f0g4 := int64(f0) * int64(g4)
+	f0g5 := int64(f0) * int64(g5)
+	f0g6 := int64(f0) * int64(g6)
+	f0g7 := int64(f0) * int64(g7)
+	f0g8 := int64(f0) * int64(g8)
+	f0g9 := int64(f0) * int64(g9)
+	f1g0 := int64(f1) * int64(g0)
+	f1g1_2 := int64(f1_2) * int64(g1)
+	f1g2 := int64(f1) * int64(g2)
+	f1g3_2 := int64(f1_2) * int64(g3)
+	f1g4 := int64(f1) * int64(g4)
+	f1g5_2 := int64(f1_2) * int64(g5)
+	f1g6 := int64(f1) * int64(g6)
+	f1g7_2 := int64(f1_2) * int64(g7)
+	f1g8 := int64(f1) * int64(g8)
+	f1g9_38 := int64(f1_2) * int64(g9_19)
+	f2g0 := int64(f2) * int64(g0)
+	f2g1 := int64(f2) * int64(g1)
+	f2g2 := int64(f2) * int64(g2)
+	f2g3 := int64(f2) * int64(g3)
+	f2g4 := int64(f2) * int64(g4)
+	f2g5 := int64(f2) * int64(g5)
+	f2g6 := int64(f2) * int64(g6)
+	f2g7 := int64(f2) * int64(g7)
+	f2g8_19 := int64(f2) * int64(g8_19)
+	f2g9_19 := int64(f2) * int64(g9_19)
+	f3g0 := int64(f3) * int64(g0)
+	f3g1_2 := int64(f3_2) * int64(g1)
+	f3g2 := int64(f3) * int64(g2)
+	f3g3_2 := int64(f3_2) * int64(g3)
+	f3g4 := int64(f3) * int64(g4)
+	f3g5_2 := int64(f3_2) * int64(g5)
+	f3g6 := int64(f3) * int64(g6)
+	f3g7_38 := int64(f3_2) * int64(g7_19)
+	f3g8_19 := int64(f3) * int64(g8_19)
+	f3g9_38 := int64(f3_2) * int64(g9_19)
+	f4g0 := int64(f4) * int64(g0)
+	f4g1 := int64(f4) * int64(g1)
+	f4g2 := int64(f4) * int64(g2)
+	f4g3 := int64(f4) * int64(g3)
+	f4g4 := int64(f4) * int64(g4)
+	f4g5 := int64(f4) * int64(g5)
+	f4g6_19 := int64(f4) * int64(g6_19)
+	f4g7_19 := int64(f4) * int64(g7_19)
+	f4g8_19 := int64(f4) * int64(g8_19)
+	f4g9_19 := int64(f4) * int64(g9_19)
+	f5g0 := int64(f5) * int64(g0)
+	f5g1_2 := int64(f5_2) * int64(g1)
+	f5g2 := int64(f5) * int64(g2)
+	f5g3_2 := int64(f5_2) * int64(g3)
+	f5g4 := int64(f5) * int64(g4)
+	f5g5_38 := int64(f5_2) * int64(g5_19)
+	f5g6_19 := int64(f5) * int64(g6_19)
+	f5g7_38 := int64(f5_2) * int64(g7_19)
+	f5g8_19 := int64(f5) * int64(g8_19)
+	f5g9_38 := int64(f5_2) * int64(g9_19)
+	f6g0 := int64(f6) * int64(g0)
+	f6g1 := int64(f6) * int64(g1)
+	f6g2 := int64(f6) * int64(g2)
+	f6g3 := int64(f6) * int64(g3)
+	f6g4_19 := int64(f6) * int64(g4_19)
+	f6g5_19 := int64(f6) * int64(g5_19)
+	f6g6_19 := int64(f6) * int64(g6_19)
+	f6g7_19 := int64(f6) * int64(g7_19)
+	f6g8_19 := int64(f6) * int64(g8_19)
+	f6g9_19 := int64(f6) * int64(g9_19)
+	f7g0 := int64(f7) * int64(g0)
+	f7g1_2 := int64(f7_2) * int64(g1)
+	f7g2 := int64(f7) * int64(g2)
+	f7g3_38 := int64(f7_2) * int64(g3_19)
+	f7g4_19 := int64(f7) * int64(g4_19)
+	f7g5_38 := int64(f7_2) * int64(g5_19)
+	f7g6_19 := int64(f7) * int64(g6_19)
+	f7g7_38 := int64(f7_2) * int64(g7_19)
+	f7g8_19 := int64(f7) * int64(g8_19)
+	f7g9_38 := int64(f7_2) * int64(g9_19)
+	f8g0 := int64(f8) * int64(g0)
+	f8g1 := int64(f8) * int64(g1)
+	f8g2_19 := int64(f8) * int64(g2_19)
+	f8g3_19 := int64(f8) * int64(g3_19)
+	f8g4_19 := int64(f8) * int64(g4_19)
+	f8g5_19 := int64(f8) * int64(g5_19)
+	f8g6_19 := int64(f8) * int64(g6_19)
+	f8g7_19 := int64(f8) * int64(g7_19)
+	f8g8_19 := int64(f8) * int64(g8_19)
+	f8g9_19 := int64(f8) * int64(g9_19)
+	f9g0 := int64(f9) * int64(g0)
+	f9g1_38 := int64(f9_2) * int64(g1_19)
+	f9g2_19 := int64(f9) * int64(g2_19)
+	f9g3_38 := int64(f9_2) * int64(g3_19)
+	f9g4_19 := int64(f9) * int64(g4_19)
+	f9g5_38 := int64(f9_2) * int64(g5_19)
+	f9g6_19 := int64(f9) * int64(g6_19)
+	f9g7_38 := int64(f9_2) * int64(g7_19)
+	f9g8_19 := int64(f9) * int64(g8_19)
+	f9g9_38 := int64(f9_2) * int64(g9_19)
+	h0 := f0g0 + f1g9_38 + f2g8_19 + f3g7_38 + f4g6_19 + f5g5_38 + f6g4_19 + f7g3_38 + f8g2_19 + f9g1_38
+	h1 := f0g1 + f1g0 + f2g9_19 + f3g8_19 + f4g7_19 + f5g6_19 + f6g5_19 + f7g4_19 + f8g3_19 + f9g2_19
+	h2 := f0g2 + f1g1_2 + f2g0 + f3g9_38 + f4g8_19 + f5g7_38 + f6g6_19 + f7g5_38 + f8g4_19 + f9g3_38
+	h3 := f0g3 + f1g2 + f2g1 + f3g0 + f4g9_19 + f5g8_19 + f6g7_19 + f7g6_19 + f8g5_19 + f9g4_19
+	h4 := f0g4 + f1g3_2 + f2g2 + f3g1_2 + f4g0 + f5g9_38 + f6g8_19 + f7g7_38 + f8g6_19 + f9g5_38
+	h5 := f0g5 + f1g4 + f2g3 + f3g2 + f4g1 + f5g0 + f6g9_19 + f7g8_19 + f8g7_19 + f9g6_19
+	h6 := f0g6 + f1g5_2 + f2g4 + f3g3_2 + f4g2 + f5g1_2 + f6g0 + f7g9_38 + f8g8_19 + f9g7_38
+	h7 := f0g7 + f1g6 + f2g5 + f3g4 + f4g3 + f5g2 + f6g1 + f7g0 + f8g9_19 + f9g8_19
+	h8 := f0g8 + f1g7_2 + f2g6 + f3g5_2 + f4g4 + f5g3_2 + f6g2 + f7g1_2 + f8g0 + f9g9_38
+	h9 := f0g9 + f1g8 + f2g7 + f3g6 + f4g5 + f5g4 + f6g3 + f7g2 + f8g1 + f9g0
+	var carry [10]int64
+
+	// |h0| <= (1.1*1.1*2^52*(1+19+19+19+19)+1.1*1.1*2^50*(38+38+38+38+38))
+	//   i.e. |h0| <= 1.2*2^59; narrower ranges for h2, h4, h6, h8
+	// |h1| <= (1.1*1.1*2^51*(1+1+19+19+19+19+19+19+19+19))
+	//   i.e. |h1| <= 1.5*2^58; narrower ranges for h3, h5, h7, h9
+
+	carry[0] = (h0 + (1 << 25)) >> 26
+	h1 += carry[0]
+	h0 -= carry[0] << 26
+	carry[4] = (h4 + (1 << 25)) >> 26
+	h5 += carry[4]
+	h4 -= carry[4] << 26
+	// |h0| <= 2^25
+	// |h4| <= 2^25
+	// |h1| <= 1.51*2^58
+	// |h5| <= 1.51*2^58
+
+	carry[1] = (h1 + (1 << 24)) >> 25
+	h2 += carry[1]
+	h1 -= carry[1] << 25
+	carry[5] = (h5 + (1 << 24)) >> 25
+	h6 += carry[5]
+	h5 -= carry[5] << 25
+	// |h1| <= 2^24; from now on fits into int32
+	// |h5| <= 2^24; from now on fits into int32
+	// |h2| <= 1.21*2^59
+	// |h6| <= 1.21*2^59
+
+	carry[2] = (h2 + (1 << 25)) >> 26
+	h3 += carry[2]
+	h2 -= carry[2] << 26
+	carry[6] = (h6 + (1 << 25)) >> 26
+	h7 += carry[6]
+	h6 -= carry[6] << 26
+	// |h2| <= 2^25; from now on fits into int32 unchanged
+	// |h6| <= 2^25; from now on fits into int32 unchanged
+	// |h3| <= 1.51*2^58
+	// |h7| <= 1.51*2^58
+
+	carry[3] = (h3 + (1 << 24)) >> 25
+	h4 += carry[3]
+	h3 -= carry[3] << 25
+	carry[7] = (h7 + (1 << 24)) >> 25
+	h8 += carry[7]
+	h7 -= carry[7] << 25
+	// |h3| <= 2^24; from now on fits into int32 unchanged
+	// |h7| <= 2^24; from now on fits into int32 unchanged
+	// |h4| <= 1.52*2^33
+	// |h8| <= 1.52*2^33
+
+	carry[4] = (h4 + (1 << 25)) >> 26
+	h5 += carry[4]
+	h4 -= carry[4] << 26
+	carry[8] = (h8 + (1 << 25)) >> 26
+	h9 += carry[8]
+	h8 -= carry[8] << 26
+	// |h4| <= 2^25; from now on fits into int32 unchanged
+	// |h8| <= 2^25; from now on fits into int32 unchanged
+	// |h5| <= 1.01*2^24
+	// |h9| <= 1.51*2^58
+
+	carry[9] = (h9 + (1 << 24)) >> 25
+	h0 += carry[9] * 19
+	h9 -= carry[9] << 25
+	// |h9| <= 2^24; from now on fits into int32 unchanged
+	// |h0| <= 1.8*2^37
+
+	carry[0] = (h0 + (1 << 25)) >> 26
+	h1 += carry[0]
+	h0 -= carry[0] << 26
+	// |h0| <= 2^25; from now on fits into int32 unchanged
+	// |h1| <= 1.01*2^24
+
+	h[0] = int32(h0)
+	h[1] = int32(h1)
+	h[2] = int32(h2)
+	h[3] = int32(h3)
+	h[4] = int32(h4)
+	h[5] = int32(h5)
+	h[6] = int32(h6)
+	h[7] = int32(h7)
+	h[8] = int32(h8)
+	h[9] = int32(h9)
+}
+
+// feSquare calculates h = f*f. Can overlap h with f.
+//
+// Preconditions:
+//    |f| bounded by 1.1*2^26,1.1*2^25,1.1*2^26,1.1*2^25,etc.
+//
+// Postconditions:
+//    |h| bounded by 1.1*2^25,1.1*2^24,1.1*2^25,1.1*2^24,etc.
+func feSquare(h, f *fieldElement) {
+	f0 := f[0]
+	f1 := f[1]
+	f2 := f[2]
+	f3 := f[3]
+	f4 := f[4]
+	f5 := f[5]
+	f6 := f[6]
+	f7 := f[7]
+	f8 := f[8]
+	f9 := f[9]
+	f0_2 := 2 * f0
+	f1_2 := 2 * f1
+	f2_2 := 2 * f2
+	f3_2 := 2 * f3
+	f4_2 := 2 * f4
+	f5_2 := 2 * f5
+	f6_2 := 2 * f6
+	f7_2 := 2 * f7
+	f5_38 := 38 * f5 // 1.31*2^30
+	f6_19 := 19 * f6 // 1.31*2^30
+	f7_38 := 38 * f7 // 1.31*2^30
+	f8_19 := 19 * f8 // 1.31*2^30
+	f9_38 := 38 * f9 // 1.31*2^30
+	f0f0 := int64(f0) * int64(f0)
+	f0f1_2 := int64(f0_2) * int64(f1)
+	f0f2_2 := int64(f0_2) * int64(f2)
+	f0f3_2 := int64(f0_2) * int64(f3)
+	f0f4_2 := int64(f0_2) * int64(f4)
+	f0f5_2 := int64(f0_2) * int64(f5)
+	f0f6_2 := int64(f0_2) * int64(f6)
+	f0f7_2 := int64(f0_2) * int64(f7)
+	f0f8_2 := int64(f0_2) * int64(f8)
+	f0f9_2 := int64(f0_2) * int64(f9)
+	f1f1_2 := int64(f1_2) * int64(f1)
+	f1f2_2 := int64(f1_2) * int64(f2)
+	f1f3_4 := int64(f1_2) * int64(f3_2)
+	f1f4_2 := int64(f1_2) * int64(f4)
+	f1f5_4 := int64(f1_2) * int64(f5_2)
+	f1f6_2 := int64(f1_2) * int64(f6)
+	f1f7_4 := int64(f1_2) * int64(f7_2)
+	f1f8_2 := int64(f1_2) * int64(f8)
+	f1f9_76 := int64(f1_2) * int64(f9_38)
+	f2f2 := int64(f2) * int64(f2)
+	f2f3_2 := int64(f2_2) * int64(f3)
+	f2f4_2 := int64(f2_2) * int64(f4)
+	f2f5_2 := int64(f2_2) * int64(f5)
+	f2f6_2 := int64(f2_2) * int64(f6)
+	f2f7_2 := int64(f2_2) * int64(f7)
+	f2f8_38 := int64(f2_2) * int64(f8_19)
+	f2f9_38 := int64(f2) * int64(f9_38)
+	f3f3_2 := int64(f3_2) * int64(f3)
+	f3f4_2 := int64(f3_2) * int64(f4)
+	f3f5_4 := int64(f3_2) * int64(f5_2)
+	f3f6_2 := int64(f3_2) * int64(f6)
+	f3f7_76 := int64(f3_2) * int64(f7_38)
+	f3f8_38 := int64(f3_2) * int64(f8_19)
+	f3f9_76 := int64(f3_2) * int64(f9_38)
+	f4f4 := int64(f4) * int64(f4)
+	f4f5_2 := int64(f4_2) * int64(f5)
+	f4f6_38 := int64(f4_2) * int64(f6_19)
+	f4f7_38 := int64(f4) * int64(f7_38)
+	f4f8_38 := int64(f4_2) * int64(f8_19)
+	f4f9_38 := int64(f4) * int64(f9_38)
+	f5f5_38 := int64(f5) * int64(f5_38)
+	f5f6_38 := int64(f5_2) * int64(f6_19)
+	f5f7_76 := int64(f5_2) * int64(f7_38)
+	f5f8_38 := int64(f5_2) * int64(f8_19)
+	f5f9_76 := int64(f5_2) * int64(f9_38)
+	f6f6_19 := int64(f6) * int64(f6_19)
+	f6f7_38 := int64(f6) * int64(f7_38)
+	f6f8_38 := int64(f6_2) * int64(f8_19)
+	f6f9_38 := int64(f6) * int64(f9_38)
+	f7f7_38 := int64(f7) * int64(f7_38)
+	f7f8_38 := int64(f7_2) * int64(f8_19)
+	f7f9_76 := int64(f7_2) * int64(f9_38)
+	f8f8_19 := int64(f8) * int64(f8_19)
+	f8f9_38 := int64(f8) * int64(f9_38)
+	f9f9_38 := int64(f9) * int64(f9_38)
+	h0 := f0f0 + f1f9_76 + f2f8_38 + f3f7_76 + f4f6_38 + f5f5_38
+	h1 := f0f1_2 + f2f9_38 + f3f8_38 + f4f7_38 + f5f6_38
+	h2 := f0f2_2 + f1f1_2 + f3f9_76 + f4f8_38 + f5f7_76 + f6f6_19
+	h3 := f0f3_2 + f1f2_2 + f4f9_38 + f5f8_38 + f6f7_38
+	h4 := f0f4_2 + f1f3_4 + f2f2 + f5f9_76 + f6f8_38 + f7f7_38
+	h5 := f0f5_2 + f1f4_2 + f2f3_2 + f6f9_38 + f7f8_38
+	h6 := f0f6_2 + f1f5_4 + f2f4_2 + f3f3_2 + f7f9_76 + f8f8_19
+	h7 := f0f7_2 + f1f6_2 + f2f5_2 + f3f4_2 + f8f9_38
+	h8 := f0f8_2 + f1f7_4 + f2f6_2 + f3f5_4 + f4f4 + f9f9_38
+	h9 := f0f9_2 + f1f8_2 + f2f7_2 + f3f6_2 + f4f5_2
+	var carry [10]int64
+
+	carry[0] = (h0 + (1 << 25)) >> 26
+	h1 += carry[0]
+	h0 -= carry[0] << 26
+	carry[4] = (h4 + (1 << 25)) >> 26
+	h5 += carry[4]
+	h4 -= carry[4] << 26
+
+	carry[1] = (h1 + (1 << 24)) >> 25
+	h2 += carry[1]
+	h1 -= carry[1] << 25
+	carry[5] = (h5 + (1 << 24)) >> 25
+	h6 += carry[5]
+	h5 -= carry[5] << 25
+
+	carry[2] = (h2 + (1 << 25)) >> 26
+	h3 += carry[2]
+	h2 -= carry[2] << 26
+	carry[6] = (h6 + (1 << 25)) >> 26
+	h7 += carry[6]
+	h6 -= carry[6] << 26
+
+	carry[3] = (h3 + (1 << 24)) >> 25
+	h4 += carry[3]
+	h3 -= carry[3] << 25
+	carry[7] = (h7 + (1 << 24)) >> 25
+	h8 += carry[7]
+	h7 -= carry[7] << 25
+
+	carry[4] = (h4 + (1 << 25)) >> 26
+	h5 += carry[4]
+	h4 -= carry[4] << 26
+	carry[8] = (h8 + (1 << 25)) >> 26
+	h9 += carry[8]
+	h8 -= carry[8] << 26
+
+	carry[9] = (h9 + (1 << 24)) >> 25
+	h0 += carry[9] * 19
+	h9 -= carry[9] << 25
+
+	carry[0] = (h0 + (1 << 25)) >> 26
+	h1 += carry[0]
+	h0 -= carry[0] << 26
+
+	h[0] = int32(h0)
+	h[1] = int32(h1)
+	h[2] = int32(h2)
+	h[3] = int32(h3)
+	h[4] = int32(h4)
+	h[5] = int32(h5)
+	h[6] = int32(h6)
+	h[7] = int32(h7)
+	h[8] = int32(h8)
+	h[9] = int32(h9)
+}
+
+// feMul121666 calculates h = f * 121666. Can overlap h with f.
+//
+// Preconditions:
+//    |f| bounded by 1.1*2^26,1.1*2^25,1.1*2^26,1.1*2^25,etc.
+//
+// Postconditions:
+//    |h| bounded by 1.1*2^25,1.1*2^24,1.1*2^25,1.1*2^24,etc.
+func feMul121666(h, f *fieldElement) {
+	h0 := int64(f[0]) * 121666
+	h1 := int64(f[1]) * 121666
+	h2 := int64(f[2]) * 121666
+	h3 := int64(f[3]) * 121666
+	h4 := int64(f[4]) * 121666
+	h5 := int64(f[5]) * 121666
+	h6 := int64(f[6]) * 121666
+	h7 := int64(f[7]) * 121666
+	h8 := int64(f[8]) * 121666
+	h9 := int64(f[9]) * 121666
+	var carry [10]int64
+
+	carry[9] = (h9 + (1 << 24)) >> 25
+	h0 += carry[9] * 19
+	h9 -= carry[9] << 25
+	carry[1] = (h1 + (1 << 24)) >> 25
+	h2 += carry[1]
+	h1 -= carry[1] << 25
+	carry[3] = (h3 + (1 << 24)) >> 25
+	h4 += carry[3]
+	h3 -= carry[3] << 25
+	carry[5] = (h5 + (1 << 24)) >> 25
+	h6 += carry[5]
+	h5 -= carry[5] << 25
+	carry[7] = (h7 + (1 << 24)) >> 25
+	h8 += carry[7]
+	h7 -= carry[7] << 25
+
+	carry[0] = (h0 + (1 << 25)) >> 26
+	h1 += carry[0]
+	h0 -= carry[0] << 26
+	carry[2] = (h2 + (1 << 25)) >> 26
+	h3 += carry[2]
+	h2 -= carry[2] << 26
+	carry[4] = (h4 + (1 << 25)) >> 26
+	h5 += carry[4]
+	h4 -= carry[4] << 26
+	carry[6] = (h6 + (1 << 25)) >> 26
+	h7 += carry[6]
+	h6 -= carry[6] << 26
+	carry[8] = (h8 + (1 << 25)) >> 26
+	h9 += carry[8]
+	h8 -= carry[8] << 26
+
+	h[0] = int32(h0)
+	h[1] = int32(h1)
+	h[2] = int32(h2)
+	h[3] = int32(h3)
+	h[4] = int32(h4)
+	h[5] = int32(h5)
+	h[6] = int32(h6)
+	h[7] = int32(h7)
+	h[8] = int32(h8)
+	h[9] = int32(h9)
+}
+
+// feInvert sets out = z^-1.
+func feInvert(out, z *fieldElement) {
+	var t0, t1, t2, t3 fieldElement
+	var i int
+
+	feSquare(&t0, z)
+	for i = 1; i < 1; i++ {
+		feSquare(&t0, &t0)
+	}
+	feSquare(&t1, &t0)
+	for i = 1; i < 2; i++ {
+		feSquare(&t1, &t1)
+	}
+	feMul(&t1, z, &t1)
+	feMul(&t0, &t0, &t1)
+	feSquare(&t2, &t0)
+	for i = 1; i < 1; i++ {
+		feSquare(&t2, &t2)
+	}
+	feMul(&t1, &t1, &t2)
+	feSquare(&t2, &t1)
+	for i = 1; i < 5; i++ {
+		feSquare(&t2, &t2)
+	}
+	feMul(&t1, &t2, &t1)
+	feSquare(&t2, &t1)
+	for i = 1; i < 10; i++ {
+		feSquare(&t2, &t2)
+	}
+	feMul(&t2, &t2, &t1)
+	feSquare(&t3, &t2)
+	for i = 1; i < 20; i++ {
+		feSquare(&t3, &t3)
+	}
+	feMul(&t2, &t3, &t2)
+	feSquare(&t2, &t2)
+	for i = 1; i < 10; i++ {
+		feSquare(&t2, &t2)
+	}
+	feMul(&t1, &t2, &t1)
+	feSquare(&t2, &t1)
+	for i = 1; i < 50; i++ {
+		feSquare(&t2, &t2)
+	}
+	feMul(&t2, &t2, &t1)
+	feSquare(&t3, &t2)
+	for i = 1; i < 100; i++ {
+		feSquare(&t3, &t3)
+	}
+	feMul(&t2, &t3, &t2)
+	feSquare(&t2, &t2)
+	for i = 1; i < 50; i++ {
+		feSquare(&t2, &t2)
+	}
+	feMul(&t1, &t2, &t1)
+	feSquare(&t1, &t1)
+	for i = 1; i < 5; i++ {
+		feSquare(&t1, &t1)
+	}
+	feMul(out, &t1, &t0)
+}
+
+func scalarMult(out, in, base *[32]byte) {
+	var e [32]byte
+
+	copy(e[:], in[:])
+	e[0] &= 248
+	e[31] &= 127
+	e[31] |= 64
+
+	var x1, x2, z2, x3, z3, tmp0, tmp1 fieldElement
+	feFromBytes(&x1, base)
+	feOne(&x2)
+	feCopy(&x3, &x1)
+	feOne(&z3)
+
+	swap := int32(0)
+	for pos := 254; pos >= 0; pos-- {
+		b := e[pos/8] >> uint(pos&7)
+		b &= 1
+		swap ^= int32(b)
+		feCSwap(&x2, &x3, swap)
+		feCSwap(&z2, &z3, swap)
+		swap = int32(b)
+
+		feSub(&tmp0, &x3, &z3)
+		feSub(&tmp1, &x2, &z2)
+		feAdd(&x2, &x2, &z2)
+		feAdd(&z2, &x3, &z3)
+		feMul(&z3, &tmp0, &x2)
+		feMul(&z2, &z2, &tmp1)
+		feSquare(&tmp0, &tmp1)
+		feSquare(&tmp1, &x2)
+		feAdd(&x3, &z3, &z2)
+		feSub(&z2, &z3, &z2)
+		feMul(&x2, &tmp1, &tmp0)
+		feSub(&tmp1, &tmp1, &tmp0)
+		feSquare(&z2, &z2)
+		feMul121666(&z3, &tmp1)
+		feSquare(&x3, &x3)
+		feAdd(&tmp0, &tmp0, &z3)
+		feMul(&z3, &x1, &z2)
+		feMul(&z2, &tmp1, &tmp0)
+	}
+
+	feCSwap(&x2, &x3, swap)
+	feCSwap(&z2, &z3, swap)
+
+	feInvert(&z2, &z2)
+	feMul(&x2, &x2, &z2)
+	feToBytes(out, &x2)
+}
--- a/ssl/test/runner/curve25519/curve25519_test.go
+++ b/ssl/test/runner/curve25519/curve25519_test.go
@ -0,0 +1,29 @@
+// 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.
+
+package curve25519
+
+import (
+	"fmt"
+	"testing"
+)
+
+const expectedHex = "89161fde887b2b53de549af483940106ecc114d6982daa98256de23bdf77661a"
+
+func TestBaseScalarMult(t *testing.T) {
+	var a, b [32]byte
+	in := &a
+	out := &b
+	a[0] = 1
+
+	for i := 0; i < 200; i++ {
+		ScalarBaseMult(out, in)
+		in, out = out, in
+	}
+
+	result := fmt.Sprintf("%x", in[:])
+	if result != expectedHex {
+		t.Errorf("incorrect result: got %s, want %s", result, expectedHex)
+	}
+}
--- a/ssl/test/runner/curve25519/doc.go
+++ b/ssl/test/runner/curve25519/doc.go
@ -0,0 +1,23 @@
+// 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.
+
+// Package curve25519 provides an implementation of scalar multiplication on
+// the elliptic curve known as curve25519. See http://cr.yp.to/ecdh.html
+package curve25519 // import "golang.org/x/crypto/curve25519"
+
+// basePoint is the x coordinate of the generator of the curve.
+var basePoint = [32]byte{9, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}
+
+// ScalarMult sets dst to the product in*base where dst and base are the x
+// coordinates of group points and all values are in little-endian form.
+func ScalarMult(dst, in, base *[32]byte) {
+	scalarMult(dst, in, base)
+}
+
+// ScalarBaseMult sets dst to the product in*base where dst and base are the x
+// coordinates of group points, base is the standard generator and all values
+// are in little-endian form.
+func ScalarBaseMult(dst, in *[32]byte) {
+	ScalarMult(dst, in, &basePoint)
+}
--- a/ssl/test/runner/curve25519/freeze_amd64.s
+++ b/ssl/test/runner/curve25519/freeze_amd64.s
@ -0,0 +1,94 @@
+// 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 freeze(inout *[5]uint64)
+TEXT ·freeze(SB),7,$96-8
+	MOVQ inout+0(FP), DI
+
+	MOVQ SP,R11
+	MOVQ $31,CX
+	NOTQ CX
+	ANDQ CX,SP
+	ADDQ $32,SP
+
+	MOVQ R11,0(SP)
+	MOVQ R12,8(SP)
+	MOVQ R13,16(SP)
+	MOVQ R14,24(SP)
+	MOVQ R15,32(SP)
+	MOVQ BX,40(SP)
+	MOVQ BP,48(SP)
+	MOVQ 0(DI),SI
+	MOVQ 8(DI),DX
+	MOVQ 16(DI),CX
+	MOVQ 24(DI),R8
+	MOVQ 32(DI),R9
+	MOVQ ·REDMASK51(SB),AX
+	MOVQ AX,R10
+	SUBQ $18,R10
+	MOVQ $3,R11
+REDUCELOOP:
+	MOVQ SI,R12
+	SHRQ $51,R12
+	ANDQ AX,SI
+	ADDQ R12,DX
+	MOVQ DX,R12
+	SHRQ $51,R12
+	ANDQ AX,DX
+	ADDQ R12,CX
+	MOVQ CX,R12
+	SHRQ $51,R12
+	ANDQ AX,CX
+	ADDQ R12,R8
+	MOVQ R8,R12
+	SHRQ $51,R12
+	ANDQ AX,R8
+	ADDQ R12,R9
+	MOVQ R9,R12
+	SHRQ $51,R12
+	ANDQ AX,R9
+	IMUL3Q $19,R12,R12
+	ADDQ R12,SI
+	SUBQ $1,R11
+	JA REDUCELOOP
+	MOVQ $1,R12
+	CMPQ R10,SI
+	CMOVQLT R11,R12
+	CMPQ AX,DX
+	CMOVQNE R11,R12
+	CMPQ AX,CX
+	CMOVQNE R11,R12
+	CMPQ AX,R8
+	CMOVQNE R11,R12
+	CMPQ AX,R9
+	CMOVQNE R11,R12
+	NEGQ R12
+	ANDQ R12,AX
+	ANDQ R12,R10
+	SUBQ R10,SI
+	SUBQ AX,DX
+	SUBQ AX,CX
+	SUBQ AX,R8
+	SUBQ AX,R9
+	MOVQ SI,0(DI)
+	MOVQ DX,8(DI)
+	MOVQ CX,16(DI)
+	MOVQ R8,24(DI)
+	MOVQ R9,32(DI)
+	MOVQ 0(SP),R11
+	MOVQ 8(SP),R12
+	MOVQ 16(SP),R13
+	MOVQ 24(SP),R14
+	MOVQ 32(SP),R15
+	MOVQ 40(SP),BX
+	MOVQ 48(SP),BP
+	MOVQ R11,SP
+	MOVQ DI,AX
+	MOVQ SI,DX
+	RET
--- a/ssl/test/runner/curve25519/ladderstep_amd64.s
+++ b/ssl/test/runner/curve25519/ladderstep_amd64.s
--- a/ssl/test/runner/curve25519/mont25519_amd64.go
+++ b/ssl/test/runner/curve25519/mont25519_amd64.go
@ -0,0 +1,240 @@
+// 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.
+
+// +build amd64,!gccgo,!appengine
+
+package curve25519
+
+// These functions are implemented in the .s files. The names of the functions
+// in the rest of the file are also taken from the SUPERCOP sources to help
+// people following along.
+
+//go:noescape
+
+func cswap(inout *[5]uint64, v uint64)
+
+//go:noescape
+
+func ladderstep(inout *[5][5]uint64)
+
+//go:noescape
+
+func freeze(inout *[5]uint64)
+
+//go:noescape
+
+func mul(dest, a, b *[5]uint64)
+
+//go:noescape
+
+func square(out, in *[5]uint64)
+
+// mladder uses a Montgomery ladder to calculate (xr/zr) *= s.
+func mladder(xr, zr *[5]uint64, s *[32]byte) {
+	var work [5][5]uint64
+
+	work[0] = *xr
+	setint(&work[1], 1)
+	setint(&work[2], 0)
+	work[3] = *xr
+	setint(&work[4], 1)
+
+	j := uint(6)
+	var prevbit byte
+
+	for i := 31; i >= 0; i-- {
+		for j < 8 {
+			bit := ((*s)[i] >> j) & 1
+			swap := bit ^ prevbit
+			prevbit = bit
+			cswap(&work[1], uint64(swap))
+			ladderstep(&work)
+			j--
+		}
+		j = 7
+	}
+
+	*xr = work[1]
+	*zr = work[2]
+}
+
+func scalarMult(out, in, base *[32]byte) {
+	var e [32]byte
+	copy(e[:], (*in)[:])
+	e[0] &= 248
+	e[31] &= 127
+	e[31] |= 64
+
+	var t, z [5]uint64
+	unpack(&t, base)
+	mladder(&t, &z, &e)
+	invert(&z, &z)
+	mul(&t, &t, &z)
+	pack(out, &t)
+}
+
+func setint(r *[5]uint64, v uint64) {
+	r[0] = v
+	r[1] = 0
+	r[2] = 0
+	r[3] = 0
+	r[4] = 0
+}
+
+// unpack sets r = x where r consists of 5, 51-bit limbs in little-endian
+// order.
+func unpack(r *[5]uint64, x *[32]byte) {
+	r[0] = uint64(x[0]) |
+		uint64(x[1])<<8 |
+		uint64(x[2])<<16 |
+		uint64(x[3])<<24 |
+		uint64(x[4])<<32 |
+		uint64(x[5])<<40 |
+		uint64(x[6]&7)<<48
+
+	r[1] = uint64(x[6])>>3 |
+		uint64(x[7])<<5 |
+		uint64(x[8])<<13 |
+		uint64(x[9])<<21 |
+		uint64(x[10])<<29 |
+		uint64(x[11])<<37 |
+		uint64(x[12]&63)<<45
+
+	r[2] = uint64(x[12])>>6 |
+		uint64(x[13])<<2 |
+		uint64(x[14])<<10 |
+		uint64(x[15])<<18 |
+		uint64(x[16])<<26 |
+		uint64(x[17])<<34 |
+		uint64(x[18])<<42 |
+		uint64(x[19]&1)<<50
+
+	r[3] = uint64(x[19])>>1 |
+		uint64(x[20])<<7 |
+		uint64(x[21])<<15 |
+		uint64(x[22])<<23 |
+		uint64(x[23])<<31 |
+		uint64(x[24])<<39 |
+		uint64(x[25]&15)<<47
+
+	r[4] = uint64(x[25])>>4 |
+		uint64(x[26])<<4 |
+		uint64(x[27])<<12 |
+		uint64(x[28])<<20 |
+		uint64(x[29])<<28 |
+		uint64(x[30])<<36 |
+		uint64(x[31]&127)<<44
+}
+
+// pack sets out = x where out is the usual, little-endian form of the 5,
+// 51-bit limbs in x.
+func pack(out *[32]byte, x *[5]uint64) {
+	t := *x
+	freeze(&t)
+
+	out[0] = byte(t[0])
+	out[1] = byte(t[0] >> 8)
+	out[2] = byte(t[0] >> 16)
+	out[3] = byte(t[0] >> 24)
+	out[4] = byte(t[0] >> 32)
+	out[5] = byte(t[0] >> 40)
+	out[6] = byte(t[0] >> 48)
+
+	out[6] ^= byte(t[1]<<3) & 0xf8
+	out[7] = byte(t[1] >> 5)
+	out[8] = byte(t[1] >> 13)
+	out[9] = byte(t[1] >> 21)
+	out[10] = byte(t[1] >> 29)
+	out[11] = byte(t[1] >> 37)
+	out[12] = byte(t[1] >> 45)
+
+	out[12] ^= byte(t[2]<<6) & 0xc0
+	out[13] = byte(t[2] >> 2)
+	out[14] = byte(t[2] >> 10)
+	out[15] = byte(t[2] >> 18)
+	out[16] = byte(t[2] >> 26)
+	out[17] = byte(t[2] >> 34)
+	out[18] = byte(t[2] >> 42)
+	out[19] = byte(t[2] >> 50)
+
+	out[19] ^= byte(t[3]<<1) & 0xfe
+	out[20] = byte(t[3] >> 7)
+	out[21] = byte(t[3] >> 15)
+	out[22] = byte(t[3] >> 23)
+	out[23] = byte(t[3] >> 31)
+	out[24] = byte(t[3] >> 39)
+	out[25] = byte(t[3] >> 47)
+
+	out[25] ^= byte(t[4]<<4) & 0xf0
+	out[26] = byte(t[4] >> 4)
+	out[27] = byte(t[4] >> 12)
+	out[28] = byte(t[4] >> 20)
+	out[29] = byte(t[4] >> 28)
+	out[30] = byte(t[4] >> 36)
+	out[31] = byte(t[4] >> 44)
+}
+
+// invert calculates r = x^-1 mod p using Fermat's little theorem.
+func invert(r *[5]uint64, x *[5]uint64) {
+	var z2, z9, z11, z2_5_0, z2_10_0, z2_20_0, z2_50_0, z2_100_0, t [5]uint64
+
+	square(&z2, x)        /* 2 */
+	square(&t, &z2)       /* 4 */
+	square(&t, &t)        /* 8 */
+	mul(&z9, &t, x)       /* 9 */
+	mul(&z11, &z9, &z2)   /* 11 */
+	square(&t, &z11)      /* 22 */
+	mul(&z2_5_0, &t, &z9) /* 2^5 - 2^0 = 31 */
+
+	square(&t, &z2_5_0)      /* 2^6 - 2^1 */
+	for i := 1; i < 5; i++ { /* 2^20 - 2^10 */
+		square(&t, &t)
+	}
+	mul(&z2_10_0, &t, &z2_5_0) /* 2^10 - 2^0 */
+
+	square(&t, &z2_10_0)      /* 2^11 - 2^1 */
+	for i := 1; i < 10; i++ { /* 2^20 - 2^10 */
+		square(&t, &t)
+	}
+	mul(&z2_20_0, &t, &z2_10_0) /* 2^20 - 2^0 */
+
+	square(&t, &z2_20_0)      /* 2^21 - 2^1 */
+	for i := 1; i < 20; i++ { /* 2^40 - 2^20 */
+		square(&t, &t)
+	}
+	mul(&t, &t, &z2_20_0) /* 2^40 - 2^0 */
+
+	square(&t, &t)            /* 2^41 - 2^1 */
+	for i := 1; i < 10; i++ { /* 2^50 - 2^10 */
+		square(&t, &t)
+	}
+	mul(&z2_50_0, &t, &z2_10_0) /* 2^50 - 2^0 */
+
+	square(&t, &z2_50_0)      /* 2^51 - 2^1 */
+	for i := 1; i < 50; i++ { /* 2^100 - 2^50 */
+		square(&t, &t)
+	}
+	mul(&z2_100_0, &t, &z2_50_0) /* 2^100 - 2^0 */
+
+	square(&t, &z2_100_0)      /* 2^101 - 2^1 */
+	for i := 1; i < 100; i++ { /* 2^200 - 2^100 */
+		square(&t, &t)
+	}
+	mul(&t, &t, &z2_100_0) /* 2^200 - 2^0 */
+
+	square(&t, &t)            /* 2^201 - 2^1 */
+	for i := 1; i < 50; i++ { /* 2^250 - 2^50 */
+		square(&t, &t)
+	}
+	mul(&t, &t, &z2_50_0) /* 2^250 - 2^0 */
+
+	square(&t, &t) /* 2^251 - 2^1 */
+	square(&t, &t) /* 2^252 - 2^2 */
+	square(&t, &t) /* 2^253 - 2^3 */
+
+	square(&t, &t) /* 2^254 - 2^4 */
+
+	square(&t, &t)   /* 2^255 - 2^5 */
+	mul(r, &t, &z11) /* 2^255 - 21 */
+}
--- a/ssl/test/runner/curve25519/mul_amd64.s
+++ b/ssl/test/runner/curve25519/mul_amd64.s
@ -0,0 +1,191 @@
+// 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 mul(dest, a, b *[5]uint64)
+TEXT ·mul(SB),0,$128-24
+	MOVQ dest+0(FP), DI
+	MOVQ a+8(FP), SI
+	MOVQ b+16(FP), DX
+
+	MOVQ SP,R11
+	MOVQ $31,CX
+	NOTQ CX
+	ANDQ CX,SP
+	ADDQ $32,SP
+
+	MOVQ R11,0(SP)
+	MOVQ R12,8(SP)
+	MOVQ R13,16(SP)
+	MOVQ R14,24(SP)
+	MOVQ R15,32(SP)
+	MOVQ BX,40(SP)
+	MOVQ BP,48(SP)
+	MOVQ DI,56(SP)
+	MOVQ DX,CX
+	MOVQ 24(SI),DX
+	IMUL3Q $19,DX,AX
+	MOVQ AX,64(SP)
+	MULQ 16(CX)
+	MOVQ AX,R8
+	MOVQ DX,R9
+	MOVQ 32(SI),DX
+	IMUL3Q $19,DX,AX
+	MOVQ AX,72(SP)
+	MULQ 8(CX)
+	ADDQ AX,R8
+	ADCQ DX,R9
+	MOVQ 0(SI),AX
+	MULQ 0(CX)
+	ADDQ AX,R8
+	ADCQ DX,R9
+	MOVQ 0(SI),AX
+	MULQ 8(CX)
+	MOVQ AX,R10
+	MOVQ DX,R11
+	MOVQ 0(SI),AX
+	MULQ 16(CX)
+	MOVQ AX,R12
+	MOVQ DX,R13
+	MOVQ 0(SI),AX
+	MULQ 24(CX)
+	MOVQ AX,R14
+	MOVQ DX,R15
+	MOVQ 0(SI),AX
+	MULQ 32(CX)
+	MOVQ AX,BX
+	MOVQ DX,BP
+	MOVQ 8(SI),AX
+	MULQ 0(CX)
+	ADDQ AX,R10
+	ADCQ DX,R11
+	MOVQ 8(SI),AX
+	MULQ 8(CX)
+	ADDQ AX,R12
+	ADCQ DX,R13
+	MOVQ 8(SI),AX
+	MULQ 16(CX)
+	ADDQ AX,R14
+	ADCQ DX,R15
+	MOVQ 8(SI),AX
+	MULQ 24(CX)
+	ADDQ AX,BX
+	ADCQ DX,BP
+	MOVQ 8(SI),DX
+	IMUL3Q $19,DX,AX
+	MULQ 32(CX)
+	ADDQ AX,R8
+	ADCQ DX,R9
+	MOVQ 16(SI),AX
+	MULQ 0(CX)
+	ADDQ AX,R12
+	ADCQ DX,R13
+	MOVQ 16(SI),AX
+	MULQ 8(CX)
+	ADDQ AX,R14
+	ADCQ DX,R15
+	MOVQ 16(SI),AX
+	MULQ 16(CX)
+	ADDQ AX,BX
+	ADCQ DX,BP
+	MOVQ 16(SI),DX
+	IMUL3Q $19,DX,AX
+	MULQ 24(CX)
+	ADDQ AX,R8
+	ADCQ DX,R9
+	MOVQ 16(SI),DX
+	IMUL3Q $19,DX,AX
+	MULQ 32(CX)
+	ADDQ AX,R10
+	ADCQ DX,R11
+	MOVQ 24(SI),AX
+	MULQ 0(CX)
+	ADDQ AX,R14
+	ADCQ DX,R15
+	MOVQ 24(SI),AX
+	MULQ 8(CX)
+	ADDQ AX,BX
+	ADCQ DX,BP
+	MOVQ 64(SP),AX
+	MULQ 24(CX)
+	ADDQ AX,R10
+	ADCQ DX,R11
+	MOVQ 64(SP),AX
+	MULQ 32(CX)
+	ADDQ AX,R12
+	ADCQ DX,R13
+	MOVQ 32(SI),AX
+	MULQ 0(CX)
+	ADDQ AX,BX
+	ADCQ DX,BP
+	MOVQ 72(SP),AX
+	MULQ 16(CX)
+	ADDQ AX,R10
+	ADCQ DX,R11
+	MOVQ 72(SP),AX
+	MULQ 24(CX)
+	ADDQ AX,R12
+	ADCQ DX,R13
+	MOVQ 72(SP),AX
+	MULQ 32(CX)
+	ADDQ AX,R14
+	ADCQ DX,R15
+	MOVQ ·REDMASK51(SB),SI
+	SHLQ $13,R9:R8
+	ANDQ SI,R8
+	SHLQ $13,R11:R10
+	ANDQ SI,R10
+	ADDQ R9,R10
+	SHLQ $13,R13:R12
+	ANDQ SI,R12
+	ADDQ R11,R12
+	SHLQ $13,R15:R14
+	ANDQ SI,R14
+	ADDQ R13,R14
+	SHLQ $13,BP:BX
+	ANDQ SI,BX
+	ADDQ R15,BX
+	IMUL3Q $19,BP,DX
+	ADDQ DX,R8
+	MOVQ R8,DX
+	SHRQ $51,DX
+	ADDQ R10,DX
+	MOVQ DX,CX
+	SHRQ $51,DX
+	ANDQ SI,R8
+	ADDQ R12,DX
+	MOVQ DX,R9
+	SHRQ $51,DX
+	ANDQ SI,CX
+	ADDQ R14,DX
+	MOVQ DX,AX
+	SHRQ $51,DX
+	ANDQ SI,R9
+	ADDQ BX,DX
+	MOVQ DX,R10
+	SHRQ $51,DX
+	ANDQ SI,AX
+	IMUL3Q $19,DX,DX
+	ADDQ DX,R8
+	ANDQ SI,R10
+	MOVQ R8,0(DI)
+	MOVQ CX,8(DI)
+	MOVQ R9,16(DI)
+	MOVQ AX,24(DI)
+	MOVQ R10,32(DI)
+	MOVQ 0(SP),R11
+	MOVQ 8(SP),R12
+	MOVQ 16(SP),R13
+	MOVQ 24(SP),R14
+	MOVQ 32(SP),R15
+	MOVQ 40(SP),BX
+	MOVQ 48(SP),BP
+	MOVQ R11,SP
+	MOVQ DI,AX
+	MOVQ SI,DX
+	RET
--- a/ssl/test/runner/curve25519/square_amd64.s
+++ b/ssl/test/runner/curve25519/square_amd64.s
@ -0,0 +1,153 @@
+// 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 square(out, in *[5]uint64)
+TEXT ·square(SB),7,$96-16
+	MOVQ out+0(FP), DI
+	MOVQ in+8(FP), SI
+
+	MOVQ SP,R11
+	MOVQ $31,CX
+	NOTQ CX
+	ANDQ CX,SP
+	ADDQ $32, SP
+
+	MOVQ R11,0(SP)
+	MOVQ R12,8(SP)
+	MOVQ R13,16(SP)
+	MOVQ R14,24(SP)
+	MOVQ R15,32(SP)
+	MOVQ BX,40(SP)
+	MOVQ BP,48(SP)
+	MOVQ 0(SI),AX
+	MULQ 0(SI)
+	MOVQ AX,CX
+	MOVQ DX,R8
+	MOVQ 0(SI),AX
+	SHLQ $1,AX
+	MULQ 8(SI)
+	MOVQ AX,R9
+	MOVQ DX,R10
+	MOVQ 0(SI),AX
+	SHLQ $1,AX
+	MULQ 16(SI)
+	MOVQ AX,R11
+	MOVQ DX,R12
+	MOVQ 0(SI),AX
+	SHLQ $1,AX
+	MULQ 24(SI)
+	MOVQ AX,R13
+	MOVQ DX,R14
+	MOVQ 0(SI),AX
+	SHLQ $1,AX
+	MULQ 32(SI)
+	MOVQ AX,R15
+	MOVQ DX,BX
+	MOVQ 8(SI),AX
+	MULQ 8(SI)
+	ADDQ AX,R11
+	ADCQ DX,R12
+	MOVQ 8(SI),AX
+	SHLQ $1,AX
+	MULQ 16(SI)
+	ADDQ AX,R13
+	ADCQ DX,R14
+	MOVQ 8(SI),AX
+	SHLQ $1,AX
+	MULQ 24(SI)
+	ADDQ AX,R15
+	ADCQ DX,BX
+	MOVQ 8(SI),DX
+	IMUL3Q $38,DX,AX
+	MULQ 32(SI)
+	ADDQ AX,CX
+	ADCQ DX,R8
+	MOVQ 16(SI),AX
+	MULQ 16(SI)
+	ADDQ AX,R15
+	ADCQ DX,BX
+	MOVQ 16(SI),DX
+	IMUL3Q $38,DX,AX
+	MULQ 24(SI)
+	ADDQ AX,CX
+	ADCQ DX,R8
+	MOVQ 16(SI),DX
+	IMUL3Q $38,DX,AX
+	MULQ 32(SI)
+	ADDQ AX,R9
+	ADCQ DX,R10
+	MOVQ 24(SI),DX
+	IMUL3Q $19,DX,AX
+	MULQ 24(SI)
+	ADDQ AX,R9
+	ADCQ DX,R10
+	MOVQ 24(SI),DX
+	IMUL3Q $38,DX,AX
+	MULQ 32(SI)
+	ADDQ AX,R11
+	ADCQ DX,R12
+	MOVQ 32(SI),DX
+	IMUL3Q $19,DX,AX
+	MULQ 32(SI)
+	ADDQ AX,R13
+	ADCQ DX,R14
+	MOVQ ·REDMASK51(SB),SI
+	SHLQ $13,R8:CX
+	ANDQ SI,CX
+	SHLQ $13,R10:R9
+	ANDQ SI,R9
+	ADDQ R8,R9
+	SHLQ $13,R12:R11
+	ANDQ SI,R11
+	ADDQ R10,R11
+	SHLQ $13,R14:R13
+	ANDQ SI,R13
+	ADDQ R12,R13
+	SHLQ $13,BX:R15
+	ANDQ SI,R15
+	ADDQ R14,R15
+	IMUL3Q $19,BX,DX
+	ADDQ DX,CX
+	MOVQ CX,DX
+	SHRQ $51,DX
+	ADDQ R9,DX
+	ANDQ SI,CX
+	MOVQ DX,R8
+	SHRQ $51,DX
+	ADDQ R11,DX
+	ANDQ SI,R8
+	MOVQ DX,R9
+	SHRQ $51,DX
+	ADDQ R13,DX
+	ANDQ SI,R9
+	MOVQ DX,AX
+	SHRQ $51,DX
+	ADDQ R15,DX
+	ANDQ SI,AX
+	MOVQ DX,R10
+	SHRQ $51,DX
+	IMUL3Q $19,DX,DX
+	ADDQ DX,CX
+	ANDQ SI,R10
+	MOVQ CX,0(DI)
+	MOVQ R8,8(DI)
+	MOVQ R9,16(DI)
+	MOVQ AX,24(DI)
+	MOVQ R10,32(DI)
+	MOVQ 0(SP),R11
+	MOVQ 8(SP),R12
+	MOVQ 16(SP),R13
+	MOVQ 24(SP),R14
+	MOVQ 32(SP),R15
+	MOVQ 40(SP),BX
+	MOVQ 48(SP),BP
+	MOVQ R11,SP
+	MOVQ DI,AX
+	MOVQ SI,DX
+	RET