From b75d6ab6d82ef5f6e99500e69a04f15a32b1fa26 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: Sat, 17 Sep 2022 20:38:32 +0800
Subject: [PATCH] fix generic panic, add movbe check

---
 base14_amd64.go   | 46 ++++++++++++++++++++++++++++++++++++++++++++++
 base14_asm.go     | 38 --------------------------------------
 base14_generic.go | 10 +++++-----
 base14_noasm.go   | 12 ++++++++++++
 cpuid.go          | 12 ++++++++++++
 cpuid_amd64.s     | 15 +++++++++++++++
 6 files changed, 90 insertions(+), 43 deletions(-)
 create mode 100644 base14_amd64.go
 delete mode 100644 base14_asm.go
 create mode 100644 base14_noasm.go
 create mode 100644 cpuid.go
 create mode 100644 cpuid_amd64.s

diff --git a/base14_amd64.go b/base14_amd64.go
new file mode 100644
index 0000000..80675d6
--- /dev/null
+++ b/base14_amd64.go
@@ -0,0 +1,46 @@
+//go:build amd64
+// +build amd64
+
+package base14
+
+import (
+	"encoding/binary"
+)
+
+//go:noescape
+//go:nosplit
+func _encode(offset, outlen int, b, encd []byte) (sum uint64, n uint64)
+
+//go:noescape
+//go:nosplit
+func _decode(offset, outlen int, b, decd []byte)
+
+func encode(offset, outlen int, b, encd []byte) {
+	if movbe {
+		if len(b) == 7 {
+			b = append(b, 0)
+		}
+		sum, n := _encode(offset, outlen, b, encd)
+		if offset == 0 {
+			return
+		}
+		var tmp [8]byte
+		binary.LittleEndian.PutUint64(tmp[:], sum)
+		copy(encd[n:], tmp[:])
+		encd[outlen-2] = '='
+		encd[outlen-1] = byte(offset)
+	} else {
+		encodeGeneric(offset, outlen, b, encd)
+	}
+}
+
+func decode(offset, outlen int, b, decd []byte) {
+	if movbe {
+		if offset != 0 && cap(b) == len(b) {
+			b = append(b, make([]byte, 8)...)
+		}
+		_decode(offset, outlen, b, decd)
+	} else {
+		decodeGeneric(offset, outlen, b, decd)
+	}
+}
diff --git a/base14_asm.go b/base14_asm.go
deleted file mode 100644
index ae64d06..0000000
--- a/base14_asm.go
+++ /dev/null
@@ -1,38 +0,0 @@
-//go:build amd64
-// +build amd64
-
-package base14
-
-import (
-	"encoding/binary"
-)
-
-//go:noescape
-//go:nosplit
-func _encode(offset, outlen int, b, encd []byte) (sum uint64, n uint64)
-
-//go:noescape
-//go:nosplit
-func _decode(offset, outlen int, b, decd []byte)
-
-func encode(offset, outlen int, b, encd []byte) {
-	if len(b) == 7 {
-		b = append(b, 0)
-	}
-	sum, n := _encode(offset, outlen, b, encd)
-	if offset == 0 {
-		return
-	}
-	var tmp [8]byte
-	binary.LittleEndian.PutUint64(tmp[:], sum)
-	copy(encd[n:], tmp[:])
-	encd[outlen-2] = '='
-	encd[outlen-1] = byte(offset)
-}
-
-func decode(offset, outlen int, b, decd []byte) {
-	if offset != 0 && cap(b) == len(b) {
-		b = append(b, make([]byte, 8)...)
-	}
-	_decode(offset, outlen, b, decd)
-}
diff --git a/base14_generic.go b/base14_generic.go
index 001ec94..2b00c86 100644
--- a/base14_generic.go
+++ b/base14_generic.go
@@ -1,12 +1,9 @@
-//go:build !amd64
-// +build !amd64
-
 package base14
 
 import "encoding/binary"
 
 //go:nosplit
-func encode(offset, outlen int, b, encd []byte) {
+func encodeGeneric(offset, outlen int, b, encd []byte) {
 	var n int
 	i := 0
 	if len(b) == 7 {
@@ -23,6 +20,9 @@ func encode(offset, outlen int, b, encd []byte) {
 		sum += 0x4e004e004e004e00
 		binary.BigEndian.PutUint64(encd[n:], sum)
 	} else {
+		if len(b)%7 == 0 {
+			b = append(b, 0)
+		}
 		for ; i <= len(b)-7; i += 7 {
 			shift := binary.BigEndian.Uint64(b[i:]) >> 2
 			sum := shift
@@ -71,7 +71,7 @@ func encode(offset, outlen int, b, encd []byte) {
 }
 
 //go:nosplit
-func decode(offset, outlen int, b, decd []byte) {
+func decodeGeneric(offset, outlen int, b, decd []byte) {
 	var n uintptr
 	i := 0
 	for ; i <= outlen-7; n += 8 {
diff --git a/base14_noasm.go b/base14_noasm.go
new file mode 100644
index 0000000..4c8d769
--- /dev/null
+++ b/base14_noasm.go
@@ -0,0 +1,12 @@
+//go:build !amd64
+// +build !amd64
+
+package base14
+
+import _ "unsafe"
+
+//go:linkname encode github.com/fumiama/go-base16384.encodeGeneric
+func encode(offset, outlen int, b, encd []byte)
+
+//go:linkname decode github.com/fumiama/go-base16384.decodeGeneric
+func decode(offset, outlen int, b, decd []byte)
diff --git a/cpuid.go b/cpuid.go
new file mode 100644
index 0000000..acf568f
--- /dev/null
+++ b/cpuid.go
@@ -0,0 +1,12 @@
+//go:build amd64
+// +build amd64
+
+package base14
+
+func cpuid(op uint32) (eax, ebx, ecx, edx uint32)
+
+// True when MOVBEx instructions are available.
+var movbe = func() bool {
+	_, _, c, _ := cpuid(1)
+	return c&(1<<22) > 0
+}()
diff --git a/cpuid_amd64.s b/cpuid_amd64.s
new file mode 100644
index 0000000..c27e32a
--- /dev/null
+++ b/cpuid_amd64.s
@@ -0,0 +1,15 @@
+//go:build amd64
+// +build amd64
+
+// Copyright (c) 2015 Klaus Post, released under MIT License. See LICENSE file.
+
+// func cpuid(op uint32) (eax, ebx, ecx, edx uint32)
+TEXT ·cpuid(SB), 7, $0
+        XORQ CX, CX
+        MOVL op+0(FP), AX
+        CPUID
+        MOVL AX, eax+8(FP)
+        MOVL BX, ebx+12(FP)
+        MOVL CX, ecx+16(FP)
+        MOVL DX, edx+20(FP)
+        RET