add big-endian support

decoder.go

@@ -273,39 +273,33 @@ func (e *Encoding) decode(dst []byte, src []byte) int {
 			}
 		}
 	}
-	up := (*[3]byte)(unsafe.Pointer(&u))
+	up := (*[4]byte)(unsafe.Pointer(&u))
 	switch l {
 	case 4:
 		if !e.pad && op-opstart+3 > dstlen {
 			return 0
 		}
 		u = ctou32(ip)
 		u = (e.lutXd0[byte(u)] | e.lutXd1[byte(u>>8)] | e.lutXd2[byte(u>>16)] | e.lutXd3[u>>24])
-		*(*byte)(unsafe.Pointer(op)) = up[0]
-		op++
-		*(*byte)(unsafe.Pointer(op)) = up[1]
-		op++
-		*(*byte)(unsafe.Pointer(op)) = up[2]
-		op++
+		putTail(op, up, 3)
+		op += 3
 		cu |= u
 		break
 	case 3:
 		if !e.pad && op-opstart+2 > dstlen {
 			return 0
 		}
 		u = e.lutXd0[*(*byte)(unsafe.Pointer(ip + 0))] | e.lutXd1[*(*byte)(unsafe.Pointer(ip + 1))] | e.lutXd2[*(*byte)(unsafe.Pointer(ip + 2))]
-		*(*byte)(unsafe.Pointer(op)) = up[0]
-		op++
-		*(*byte)(unsafe.Pointer(op)) = up[1]
-		op++
+		putTail(op, up, 2)
+		op += 2
 		cu |= u
 		break
 	case 2:
 		if !e.pad && op-opstart >= dstlen {
 			return 0
 		}
 		u = e.lutXd0[*(*byte)(unsafe.Pointer(ip + 0))] | e.lutXd1[*(*byte)(unsafe.Pointer(ip + 1))]
-		*(*byte)(unsafe.Pointer(op)) = up[0]
+		putTail(op, up, 1)
 		op++
 		cu |= u
 		break

decoder_be.go

@@ -0,0 +1,20 @@
+//go:build armbe || arm64be || mips || mips64 || mips64p32 || ppc || ppc64 || sparc || sparc64 || s390 || s390x
+// +build armbe arm64be mips mips64 mips64p32 ppc ppc64 sparc sparc64 s390 s390x
+
+package base64
+
+import "unsafe"
+
+func putTail(ptr uintptr, tail *[4]byte, n int) {
+	switch n {
+	case 3:
+		*(*byte)(unsafe.Pointer(ptr)) = tail[3]
+		*(*byte)(unsafe.Pointer(ptr + 1)) = tail[2]
+		*(*byte)(unsafe.Pointer(ptr + 2)) = tail[1]
+	case 2:
+		*(*byte)(unsafe.Pointer(ptr)) = tail[3]
+		*(*byte)(unsafe.Pointer(ptr + 1)) = tail[2]
+	case 1:
+		*(*byte)(unsafe.Pointer(ptr)) = tail[3]
+	}
+}

decoder_le.go

@@ -0,0 +1,20 @@
+//go:build 386 || amd64 || amd64p32 || arm || arm64 || mipsle || mis64le || mips64p32le || ppc64le || riscv || riscv64 || wasm
+// +build 386 amd64 amd64p32 arm arm64 mipsle mis64le mips64p32le ppc64le riscv riscv64 wasm
+
+package base64
+
+import "unsafe"
+
+func putTail(ptr uintptr, tail *[4]byte, n int) {
+	switch n {
+	case 3:
+		*(*byte)(unsafe.Pointer(ptr)) = tail[0]
+		*(*byte)(unsafe.Pointer(ptr + 1)) = tail[1]
+		*(*byte)(unsafe.Pointer(ptr + 2)) = tail[2]
+	case 2:
+		*(*byte)(unsafe.Pointer(ptr)) = tail[0]
+		*(*byte)(unsafe.Pointer(ptr + 1)) = tail[1]
+	case 1:
+		*(*byte)(unsafe.Pointer(ptr)) = tail[0]
+	}
+}

utils.go

@@ -2,7 +2,6 @@ package base64
 
 import (
 	"math"
-	"math/bits"
 	"unsafe"
 )
 
@@ -30,21 +29,6 @@ func s2b(value string) (b []byte) {
 	return b
 }
 
-//go:nosplit
-func bswap32(ptr uintptr) uint32 {
-	return bits.ReverseBytes32(*(*uint32)(unsafe.Pointer(ptr)))
-}
-
-//go:nosplit
-func stou32(cp uintptr, x uint32) {
-	*(*uint32)(unsafe.Pointer(cp)) = x
-}
-
-//go:nosplit
-func ctou32(cp uintptr) uint32 {
-	return *(*uint32)(unsafe.Pointer(cp))
-}
-
 func makeLuts(lutSe [64]byte) ([4096]uint32, [256]uint32, [256]uint32, [256]uint32, [256]uint32) {
 	lutXe := [4096]uint32{}
 	lutXd0 := [256]uint32{}

utils_be.go

@@ -0,0 +1,24 @@
+//go:build armbe || arm64be || mips || mips64 || mips64p32 || ppc || ppc64 || sparc || sparc64 || s390 || s390x
+// +build armbe arm64be mips mips64 mips64p32 ppc ppc64 sparc sparc64 s390 s390x
+
+package base64
+
+import (
+	"math/bits"
+	"unsafe"
+)
+
+//go:nosplit
+func bswap32(ptr uintptr) uint32 {
+	return *(*uint32)(unsafe.Pointer(ptr))
+}
+
+//go:nosplit
+func stou32(cp uintptr, x uint32) {
+	*(*uint32)(unsafe.Pointer(cp)) = bits.ReverseBytes32(x)
+}
+
+//go:nosplit
+func ctou32(cp uintptr) uint32 {
+	return bits.ReverseBytes32(*(*uint32)(unsafe.Pointer(cp)))
+}

utils_le.go

@@ -0,0 +1,24 @@
+//go:build 386 || amd64 || amd64p32 || arm || arm64 || mipsle || mis64le || mips64p32le || ppc64le || riscv || riscv64 || wasm
+// +build 386 amd64 amd64p32 arm arm64 mipsle mis64le mips64p32le ppc64le riscv riscv64 wasm
+
+package base64
+
+import (
+	"math/bits"
+	"unsafe"
+)
+
+//go:nosplit
+func bswap32(ptr uintptr) uint32 {
+	return bits.ReverseBytes32(*(*uint32)(unsafe.Pointer(ptr)))
+}
+
+//go:nosplit
+func stou32(cp uintptr, x uint32) {
+	*(*uint32)(unsafe.Pointer(cp)) = x
+}
+
+//go:nosplit
+func ctou32(cp uintptr) uint32 {
+	return *(*uint32)(unsafe.Pointer(cp))
+}