From f6d737d0a618014f79bc05209a246e918a68752e Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E6=BA=90=E6=96=87=E9=9B=A8?=
 <41315874+fumiama@users.noreply.github.com>
Date: Sun, 2 Oct 2022 22:49:06 +0800
Subject: [PATCH] add amd64 asm ver enc16blk1

---
 encode.go       | 17 -------------
 encode_amd64.go |  8 ++++++
 encode_amd64.s  | 66 +++++++++++++++++++++++++++++++++++++++++++++++++
 encode_noasm.go | 23 +++++++++++++++++
 4 files changed, 97 insertions(+), 17 deletions(-)
 create mode 100644 encode_amd64.go
 create mode 100644 encode_amd64.s
 create mode 100644 encode_noasm.go

diff --git a/encode.go b/encode.go
index 7042007..fcaee96 100644
--- a/encode.go
+++ b/encode.go
@@ -36,23 +36,6 @@ func (bs Base) Encode(data []byte) (out []byte) {
 	return
 }
 
-// enc16blk1 for bit 1 (actual enc128blk1)
-//    len(in)!=0, len(out)==len(in)*16
-//go:nosplit
-func enc16blk1(mask uint16, in, out []byte) {
-	for i, n := range in {
-		c := i * 16
-		binary.BigEndian.PutUint16(out[c:c+2], uint16(n>>7)+mask)
-		binary.BigEndian.PutUint16(out[c+2:c+4], uint16(n>>6&1)+mask)
-		binary.BigEndian.PutUint16(out[c+4:c+6], uint16(n>>5&1)+mask)
-		binary.BigEndian.PutUint16(out[c+6:c+8], uint16(n>>4&1)+mask)
-		binary.BigEndian.PutUint16(out[c+8:c+10], uint16(n>>3&1)+mask)
-		binary.BigEndian.PutUint16(out[c+10:c+12], uint16(n>>2&1)+mask)
-		binary.BigEndian.PutUint16(out[c+12:c+14], uint16(n>>1&1)+mask)
-		binary.BigEndian.PutUint16(out[c+14:c+16], uint16(n&1)+mask)
-	}
-}
-
 // enc64blk2 for bit 2
 //    len(in)!=0, len(out)==len(in)*8
 //go:nosplit
diff --git a/encode_amd64.go b/encode_amd64.go
new file mode 100644
index 0000000..03b7b5f
--- /dev/null
+++ b/encode_amd64.go
@@ -0,0 +1,8 @@
+//go:build amd64
+// +build amd64
+
+package unibase2n
+
+// enc16blk1 for bit 1 (actual enc128blk1)
+//    len(in)!=0, len(out)==len(in)*16
+func enc16blk1(mask uint16, in, out []byte)
diff --git a/encode_amd64.s b/encode_amd64.s
new file mode 100644
index 0000000..4cb0cf9
--- /dev/null
+++ b/encode_amd64.s
@@ -0,0 +1,66 @@
+//go:build amd64
+// +build amd64
+
+#include "textflag.h"
+
+// enc16blk1(mask uint16, in, out []byte)
+//    for bit 1 (actual enc128blk1)
+TEXT ·enc16blk1(SB), NOSPLIT, $0-56
+    // create mask
+    MOVQ ·mask+0(FP), BX
+    ANDQ $0xffff, BX
+    BSWAPQ BX
+    MOVQ BX, AX
+    SHRQ $16, AX
+    ORQ  AX, BX
+    SHRQ $16, AX
+    ORQ  AX, BX
+    SHRQ $16, AX
+    ORQ  AX, BX
+    // load source addr
+    MOVQ ·in+8(FP), SI
+    // load source len
+    MOVQ ·inlen+16(FP), CX
+    // load dest addr
+    MOVQ ·out+32(FP), DI
+    // go forward
+    CLD
+lop:
+    LODSB
+    // 8 -> 64
+    XORQ DX, DX
+    SHLB $1, AX
+    SETCS DX
+    SHLQ $16, DX
+    SHLB $1, AX
+    SETCS DX
+    SHLQ $16, DX
+    SHLB $1, AX
+    SETCS DX
+    SHLQ $16, DX
+    SHLB $1, AX
+    SETCS DX
+    // add mask
+    MOVQ AX, R8
+    LEAQ 0(DX)(BX*1), AX
+    BSWAPQ AX
+    STOSQ
+    // 8 -> 64
+    XORQ AX, AX
+    SHLB $1, R8
+    SETCS AX
+    SHLQ $16, AX
+    SHLB $1, R8
+    SETCS AX
+    SHLQ $16, AX
+    SHLB $1, R8
+    SETCS AX
+    SHLQ $16, AX
+    SHLB $1, R8
+    SETCS AX
+    // add mask
+    ADDQ BX, AX
+    BSWAPQ AX
+    STOSQ
+    LOOP lop
+    RET
diff --git a/encode_noasm.go b/encode_noasm.go
new file mode 100644
index 0000000..919d351
--- /dev/null
+++ b/encode_noasm.go
@@ -0,0 +1,23 @@
+//go:build !amd64
+// +build !amd64
+
+package unibase2n
+
+import "encoding/binary"
+
+// enc16blk1 for bit 1 (actual enc128blk1)
+//    len(in)!=0, len(out)==len(in)*16
+//go:nosplit
+func enc16blk1(mask uint16, in, out []byte) {
+	for i, n := range in {
+		c := i * 16
+		binary.BigEndian.PutUint16(out[c:c+2], uint16(n>>7)+mask)
+		binary.BigEndian.PutUint16(out[c+2:c+4], uint16(n>>6&1)+mask)
+		binary.BigEndian.PutUint16(out[c+4:c+6], uint16(n>>5&1)+mask)
+		binary.BigEndian.PutUint16(out[c+6:c+8], uint16(n>>4&1)+mask)
+		binary.BigEndian.PutUint16(out[c+8:c+10], uint16(n>>3&1)+mask)
+		binary.BigEndian.PutUint16(out[c+10:c+12], uint16(n>>2&1)+mask)
+		binary.BigEndian.PutUint16(out[c+12:c+14], uint16(n>>1&1)+mask)
+		binary.BigEndian.PutUint16(out[c+14:c+16], uint16(n&1)+mask)
+	}
+}