curve25519: avoid loss of R15 in -dynlink mode

Fixes golang/go#18820.

Change-Id: I4b3a49b3bbbecc4e1008989fefd39da9725a28ea
Reviewed-on: https://go-review.googlesource.com/36359
Run-TryBot: Russ Cox <rsc@golang.org>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: Austin Clements <austin@google.com>
Reviewed-by: Adam Langley <agl@golang.org>

curve25519/const_amd64.h

@@ -0,0 +1,8 @@
+// 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
+
+#define REDMASK51     0x0007FFFFFFFFFFFF

curve25519/const_amd64.s

@@ -7,8 +7,8 @@
 
 // +build amd64,!gccgo,!appengine
 
-DATA ·REDMASK51(SB)/8, $0x0007FFFFFFFFFFFF
-GLOBL ·REDMASK51(SB), 8, $8
+// These constants cannot be encoded in non-MOVQ immediates.
+// We access them directly from memory instead.
 
 DATA ·_121666_213(SB)/8, $996687872
 GLOBL ·_121666_213(SB), 8, $8

curve25519/freeze_amd64.s

@@ -7,6 +7,8 @@
 
 // +build amd64,!gccgo,!appengine
 
+#include "const_amd64.h"
+
 // func freeze(inout *[5]uint64)
 TEXT ·freeze(SB),7,$0-8
 	MOVQ inout+0(FP), DI
@@ -16,7 +18,7 @@ TEXT ·freeze(SB),7,$0-8
 	MOVQ 16(DI),CX
 	MOVQ 24(DI),R8
 	MOVQ 32(DI),R9
-	MOVQ ·REDMASK51(SB),AX
+	MOVQ $REDMASK51,AX
 	MOVQ AX,R10
 	SUBQ $18,R10
 	MOVQ $3,R11

curve25519/ladderstep_amd64.s

@@ -7,6 +7,8 @@
 
 // +build amd64,!gccgo,!appengine
 
+#include "const_amd64.h"
+
 // func ladderstep(inout *[5][5]uint64)
 TEXT ·ladderstep(SB),0,$296-8
 	MOVQ inout+0(FP),DI
@@ -118,7 +120,7 @@ TEXT ·ladderstep(SB),0,$296-8
 	MULQ 72(SP)
 	ADDQ AX,R12
 	ADCQ DX,R13
-	MOVQ ·REDMASK51(SB),DX
+	MOVQ $REDMASK51,DX
 	SHLQ $13,CX:SI
 	ANDQ DX,SI
 	SHLQ $13,R9:R8
@@ -233,7 +235,7 @@ TEXT ·ladderstep(SB),0,$296-8
 	MULQ 32(SP)
 	ADDQ AX,R12
 	ADCQ DX,R13
-	MOVQ ·REDMASK51(SB),DX
+	MOVQ $REDMASK51,DX
 	SHLQ $13,CX:SI
 	ANDQ DX,SI
 	SHLQ $13,R9:R8
@@ -438,7 +440,7 @@ TEXT ·ladderstep(SB),0,$296-8
 	MULQ 72(SP)
 	ADDQ AX,R12
 	ADCQ DX,R13
-	MOVQ ·REDMASK51(SB),DX
+	MOVQ $REDMASK51,DX
 	SHLQ $13,CX:SI
 	ANDQ DX,SI
 	SHLQ $13,R9:R8
@@ -588,7 +590,7 @@ TEXT ·ladderstep(SB),0,$296-8
 	MULQ 32(SP)
 	ADDQ AX,R12
 	ADCQ DX,R13
-	MOVQ ·REDMASK51(SB),DX
+	MOVQ $REDMASK51,DX
 	SHLQ $13,CX:SI
 	ANDQ DX,SI
 	SHLQ $13,R9:R8
@@ -728,7 +730,7 @@ TEXT ·ladderstep(SB),0,$296-8
 	MULQ 152(DI)
 	ADDQ AX,R12
 	ADCQ DX,R13
-	MOVQ ·REDMASK51(SB),DX
+	MOVQ $REDMASK51,DX
 	SHLQ $13,CX:SI
 	ANDQ DX,SI
 	SHLQ $13,R9:R8
@@ -843,7 +845,7 @@ TEXT ·ladderstep(SB),0,$296-8
 	MULQ 192(DI)
 	ADDQ AX,R12
 	ADCQ DX,R13
-	MOVQ ·REDMASK51(SB),DX
+	MOVQ $REDMASK51,DX
 	SHLQ $13,CX:SI
 	ANDQ DX,SI
 	SHLQ $13,R9:R8
@@ -993,7 +995,7 @@ TEXT ·ladderstep(SB),0,$296-8
 	MULQ 32(DI)
 	ADDQ AX,R12
 	ADCQ DX,R13
-	MOVQ ·REDMASK51(SB),DX
+	MOVQ $REDMASK51,DX
 	SHLQ $13,CX:SI
 	ANDQ DX,SI
 	SHLQ $13,R9:R8
@@ -1143,7 +1145,7 @@ TEXT ·ladderstep(SB),0,$296-8
 	MULQ 112(SP)
 	ADDQ AX,R12
 	ADCQ DX,R13
-	MOVQ ·REDMASK51(SB),DX
+	MOVQ $REDMASK51,DX
 	SHLQ $13,CX:SI
 	ANDQ DX,SI
 	SHLQ $13,R9:R8
@@ -1329,7 +1331,7 @@ TEXT ·ladderstep(SB),0,$296-8
 	MULQ 192(SP)
 	ADDQ AX,R12
 	ADCQ DX,R13
-	MOVQ ·REDMASK51(SB),DX
+	MOVQ $REDMASK51,DX
 	SHLQ $13,CX:SI
 	ANDQ DX,SI
 	SHLQ $13,R9:R8

curve25519/mul_amd64.s

@@ -7,6 +7,8 @@
 
 // +build amd64,!gccgo,!appengine
 
+#include "const_amd64.h"
+
 // func mul(dest, a, b *[5]uint64)
 TEXT ·mul(SB),0,$16-24
 	MOVQ dest+0(FP), DI
@@ -121,7 +123,7 @@ TEXT ·mul(SB),0,$16-24
 	MULQ 32(CX)
 	ADDQ AX,R14
 	ADCQ DX,R15
-	MOVQ ·REDMASK51(SB),SI
+	MOVQ $REDMASK51,SI
 	SHLQ $13,R9:R8
 	ANDQ SI,R8
 	SHLQ $13,R11:R10

curve25519/square_amd64.s

@@ -7,6 +7,8 @@
 
 // +build amd64,!gccgo,!appengine
 
+#include "const_amd64.h"
+
 // func square(out, in *[5]uint64)
 TEXT ·square(SB),7,$0-16
 	MOVQ out+0(FP), DI
@@ -84,7 +86,7 @@ TEXT ·square(SB),7,$0-16
 	MULQ 32(SI)
 	ADDQ AX,R13
 	ADCQ DX,R14
-	MOVQ ·REDMASK51(SB),SI
+	MOVQ $REDMASK51,SI
 	SHLQ $13,R8:CX
 	ANDQ SI,CX
 	SHLQ $13,R10:R9