optimize(t544): drop unsafe (#2076)

Updates #2075 #2072 #2051

internal/encryption/t544/encryption.go

@@ -50,4 +50,4 @@ func (c *state) init(key []byte, data []byte, counter uint64, nr uint8) {
 	initState(c, key, data, counter)
 }
 
-func sub_ad([]uint32)
+func refreshState(c *state)

internal/encryption/t544/encryption_amd64.s

@@ -10,8 +10,8 @@ DATA  LC0<>+12(SB)/4, $1797285236
 GLOBL LC0<>(SB), NOPTR, $16
 
 TEXT ·sub_a(SB), NOSPLIT, $0-48
-    MOVQ ·a+0(FP), DI
-    MOVQ ·b+24(FP), CX
+    MOVQ a+0(FP), DI
+    MOVQ b+24(FP), CX
     MOVQ    CX, DX
     MOVBLZX 3(CX), CX
     XORB    CX,  (DI)
@@ -48,8 +48,8 @@ TEXT ·sub_a(SB), NOSPLIT, $0-48
     RET
 
 TEXT ·sub_b(SB), NOSPLIT, $0-48
-    MOVQ ·a+0(FP), DI
-    MOVQ ·b+24(FP), CX
+    MOVQ a+0(FP), DI
+    MOVQ b+24(FP), CX
     MOVQ    CX, DX
     MOVBLZX 3(CX), CX
     XORB    CX, (DI)
@@ -87,8 +87,8 @@ TEXT ·sub_b(SB), NOSPLIT, $0-48
 
 
 TEXT ·sub_c(SB), NOSPLIT, $0-32
-    MOVQ ·a+0(FP), DI
-    MOVQ ·b+8(FP), SI
+    MOVQ a+0(FP), DI
+    MOVQ b+8(FP), SI
     MOVQ SI, AX
     MOVBLZX (SI), SI
     MOVL SI, CX
@@ -236,26 +236,27 @@ TEXT ·sub_c(SB), NOSPLIT, $0-32
     MOVB CX, 15(AX)
     RET 
 
-TEXT ·sub_d(SB), NOSPLIT, $16-32
-    MOVQ ·t+0(FP), BX
-    MOVQ ·s+8(FP), SI
-    MOVOU   (SI), X0
+TEXT ·sub_d(SB), NOSPLIT, $24-32
+    MOVQ t+0(FP), BX
+    MOVQ s+8(FP), DI
+    MOVOU   (DI), X0
     MOVOU   X0, in-16(SP)
-    MOVQ    SI, DI
+    MOVQ    $16, CX
     ADDQ    $15, DI
-    MOVB    $16, CX
+    PUSHFQ
+    STD
 lop:
     LEAQ    -1(CX), AX
     XLAT
-    MOVBLZX in-16(SP)(AX*1), AX
-    STD
-    STOSB
+    LEAQ    in-16(SP)(AX*1), SI
+    MOVSB
     LOOP    lop
+    POPFQ
     RET
 
 TEXT ·sub_e(SB), NOSPLIT, $0-32
-    MOVQ ·a+0(FP), DI
-    MOVQ ·n+8(FP), SI
+    MOVQ a+0(FP), DI
+    MOVQ n+8(FP), SI
     MOVQ $4, AX
 lop:
     MOVBQZX -4(SI)(AX*4), DX
@@ -295,9 +296,9 @@ lop:
     JNZ lop
     RET
 
-TEXT sub_ab(SB), NOSPLIT, $0-24
-    MOVQ ·s+0(FP), DI
-    MOVQ ·w+8(FP), SI
+TEXT sub_ab<>(SB), NOSPLIT, $0-24
+    MOVQ s+0(FP), DI
+    MOVQ w+8(FP), SI
     MOVL SI, AX
     MOVL SI, CX
     MOVL SI, DX
@@ -329,14 +330,14 @@ TEXT sub_ab(SB), NOSPLIT, $0-24
     MOVBLZX (DI)(DX*1), DX
     SALL $16, DX
     ORL DX, AX
-    MOVQ AX, ·retval+16(FP)
+    MOVQ AX, retval+16(FP)
     RET
 
 TEXT ·sub_f(SB), NOSPLIT, $24-68
-    MOVQ ·k+0(FP), DI
-    MOVQ ·r+8(FP), SI
-    MOVQ ·s+16(FP), DX
-    MOVQ $·w+24(FP), CX
+    MOVQ k+0(FP), DI
+    MOVQ r+8(FP), SI
+    MOVQ s+16(FP), DX
+    MOVQ $w+24(FP), CX
     MOVQ CX, R10
     MOVQ SI, R9
     MOVQ DX, R8
@@ -367,7 +368,7 @@ inner:
     ROLL $8, AX
     MOVQ R8, 0(SP)
     MOVL AX, 8(SP)
-    CALL sub_ab(SB)
+    CALL sub_ab<>(SB)
     MOVQ 16(SP), AX
     LEAL -1(BX), DX
     SARL $2, DX
@@ -378,10 +379,10 @@ end:
     RET
 
 TEXT ·sub_aa(SB), NOSPLIT, $0-56
-    MOVQ ·i+0(FP), DI
-    MOVQ ·t+8(FP), SI
-    MOVQ ·b+16(FP), DX
-    MOVQ ·m+24(FP), CX
+    MOVQ i+0(FP), DI
+    MOVQ t+8(FP), SI
+    MOVQ b+16(FP), DX
+    MOVQ m+24(FP), CX
     MOVL DI, AX
     MOVLQSX DI, DI
     MOVQ SI, R8
@@ -407,13 +408,13 @@ TEXT ·sub_aa(SB), NOSPLIT, $0-56
     MOVBLZX (AX)(DI*1), AX
     SALL $4, AX
     ORB 256(SI)(DX*1), AX
-    MOVQ AX, ·retval+48(FP)
+    MOVQ AX, retval+48(FP)
     RET
 
 // func transformInner(x *[0x15]byte, tab *[32][16]byte)
 TEXT ·transformInner(SB), NOSPLIT, $0-16
-    MOVQ ·x+0(FP), DI
-    MOVQ ·tab+8(FP), SI
+    MOVQ x+0(FP), DI
+    MOVQ tab+8(FP), SI
     MOVQ    DI, AX
     MOVL    $1, CX
     MOVQ    SI, DI
@@ -446,10 +447,10 @@ lop:
     RET
 
 TEXT ·initState(SB), NOSPLIT, $0-64
-    MOVQ ·c+0(FP), DI
-    MOVQ ·key+8(FP), SI
-    MOVQ ·data+32(FP), R8
-    MOVQ ·counter+56(FP), AX
+    MOVQ c+0(FP), DI
+    MOVQ key+8(FP), SI
+    MOVQ data+32(FP), R8
+    MOVQ counter+56(FP), AX
     MOVOA   LC0<>(SB), X0
     MOVUPS  X0, (DI)
     MOVOU   (SI), X1
@@ -467,8 +468,8 @@ TEXT ·initState(SB), NOSPLIT, $0-64
     MOVUPS  X6,112(DI)
     RET
 
-TEXT ·sub_ad(SB), NOSPLIT, $8-24
-    MOVQ ·a+0(FP), DI
+TEXT sub_ad<>(SB), NOSPLIT, $8-8
+    MOVQ a+0(FP), DI
     MOVQ DI, AX
     MOVL 40(DI), R10
     MOVL 12(DI), R12
@@ -610,11 +611,42 @@ TEXT ·sub_ad(SB), NOSPLIT, $8-24
     MOVUPS X0, 32(AX)
     RET
 
+TEXT ·refreshState(SB), NOSPLIT, $16-8
+    MOVQ    i+0(FP), BX
+    MOVB    128(BX), CX
+    JE      ad
+    SHRQ    $1, CX
+fr:
+    MOVQ    BX, 0(SP)
+    MOVQ    CX, c-8(SP)
+    CALL    sub_ad<>(SB)
+    MOVQ    c-8(SP), CX
+    MOVQ    i+0(FP), BX
+    LOOP    fr
+ad:
+    MOVOU   (BX), X0
+    MOVOU   64(BX), X1
+    MOVOU   80(BX), X2
+    MOVOU   96(BX), X3
+    PADDD   X1, X0
+    MOVOU   48(BX), X4
+    MOVUPS  X0, (BX)
+    MOVOU   16(BX), X0
+    PADDD   X2, X0
+    MOVUPS  X0, 16(BX)
+    MOVOU   32(BX), X0
+    PADDD   X3, X0
+    MOVUPS  X0, 32(BX)
+    MOVOU   112(BX), X0
+    PADDD   X4, X0
+    MOVUPS  X0, 48(BX)
+    RET
+
 // func tencentCrc32(tab *crc32.Table, b []byte) uint32
 TEXT ·tencentCrc32(SB), NOSPLIT, $0-40
-    MOVQ ·tab+0(FP), DI
-    MOVQ ·bptr+8(FP), SI
-    MOVQ ·bngas+16(FP), DX
+    MOVQ tab+0(FP), DI
+    MOVQ bptr+8(FP), SI
+    MOVQ bngas+16(FP), DX
     TESTQ   DX, DX
     JE      quickend
     ADDQ    SI, DX
@@ -629,7 +661,7 @@ lop:
     CMPQ    SI, DX
     JNE     lop
     NOTL    AX
-    MOVQ    AX, ·bngas+32(FP)
+    MOVQ    AX, bngas+32(FP)
     RET
 quickend:
     XORL    AX, AX

internal/encryption/t544/encryption_generic.go

@@ -91,12 +91,7 @@ func (c *state) encrypt(data []byte) {
 	dataLen := uint32(len(data))
 	for dataLen > 0 {
 		if c.p == 0 {
-			for i := uint8(0); i < c.nr; i += 2 {
-				sub_ad(c.state[:])
-			}
-			for i := 0; i < 16; i++ {
-				c.state[i] += c.orgstate[i]
-			}
+			refreshState(c)
 		}
 		var sb [16 * 4]byte
 		for i, v := range c.state {

internal/encryption/t544/t544.go

@@ -7,7 +7,6 @@ import (
 	"crypto/rc4"
 	"encoding/binary"
 	"math/rand"
-	"unsafe"
 
 	"github.com/Mrs4s/go-cqhttp/internal/encryption"
 )
@@ -31,8 +30,10 @@ func init() {
 // sign t544 algorithm
 // special thanks to the anonymous contributor who provided the algorithm
 func sign(curr int64, input []byte) []byte {
+	var crcData [0x15]byte
 	curr %= 1000000
-	input = append(input, []byte{byte(curr >> 24), byte(curr >> 16), byte(curr >> 8), byte(curr)}...)
+	binary.BigEndian.PutUint32(crcData[:4], uint32(curr))
+	input = append(input, crcData[:4]...)
 	var kt [4 + 32 + 4]byte
 	r := rand.New(rand.NewSource(curr))
 	for i := 0; i < 2; i++ {
@@ -51,8 +52,7 @@ func sign(curr int64, input []byte) []byte {
 	k3calc[6], k3calc[7] = 0, 0
 	rc4Cipher, _ := rc4.NewCipher(key3)
 	rc4Cipher.XORKeyStream(key3, key3)
-	var crcData [0x15]byte
-	copy(crcData[4:4+8], (*[8]byte)(unsafe.Pointer(&magic))[:])
+	binary.LittleEndian.PutUint64(crcData[4:4+8], magic)
 	tencentEncryptionA(input, kt[4:4+32], crcData[4:4+8])
 	result := md5.Sum(input)
 	crcData[2] = 1
@@ -61,7 +61,7 @@ func sign(curr int64, input []byte) []byte {
 	binary.BigEndian.PutUint32(crcData[9:13], uint32(curr))
 	copy(crcData[13:], result[:8])
 	calcCrc := tencentCrc32(&crc32Table, crcData[2:])
-	copy(kt[4+32:4+32+4], (*[4]byte)(unsafe.Pointer(&calcCrc))[:])
+	binary.LittleEndian.PutUint32(kt[4+32:4+32+4], calcCrc)
 	crcData[0] = kt[4+32]
 	crcData[1] = kt[4+32+3]
 	nonce := uint32(r.Int() ^ r.Int() ^ r.Int())

internal/encryption/t544/t544_test.go

@@ -0,0 +1,22 @@
+package t544
+
+import (
+	"crypto/rand"
+	"encoding/hex"
+	"testing"
+)
+
+func TestT544(t *testing.T) {
+	r := hex.EncodeToString(sign(0, []byte{}))
+	if r != "0c05d28b405bce1595c70ffa694ff163d4b600f229482e07de32c8000000003525382c00000000" {
+		t.Fatal(r)
+	}
+}
+
+func TestCrash(t *testing.T) {
+	brand := make([]byte, 4096)
+	for i := 1; i <= 1024; i++ {
+		rand.Reader.Read(brand)
+		sign(123, brand)
+	}
+}