zuc: eia256 ppc64x use VPERXOR for bit reverse #262

Author: emmansun

zuc/eia256_asm_ppc64x.s

@@ -20,7 +20,6 @@
 #define KS_H V11
 #define BIT_REV_TAB_L V12
 #define BIT_REV_TAB_H V13
-#define BIT_REV_AND_TAB V14
 #define ZERO V15
 #define PTR R7
 
@@ -43,26 +42,18 @@ TEXT ·eia256RoundTag8(SB),NOSPLIT,$0
 	VPERM XDATA, XDATA, XTMP1, XDATA
 #endif
 
-	LXVD2X (PTR)(R0), BIT_REV_AND_TAB
-	VAND	BIT_REV_AND_TAB, XDATA, XTMP3
 	VSPLTISB $4, XTMP2;
-	VSRW	XDATA, XTMP2, XTMP1
-	VAND	BIT_REV_AND_TAB, XTMP1, XTMP1
-
-	MOVD	$0x10, R8
-	LXVD2X (PTR)(R8), BIT_REV_TAB_L
+	LXVD2X (PTR)(R0), BIT_REV_TAB_L
 	VSLB  BIT_REV_TAB_L, XTMP2, BIT_REV_TAB_H
-	VPERM BIT_REV_TAB_L, BIT_REV_TAB_L, XTMP1, XTMP1
-	VPERM BIT_REV_TAB_H, BIT_REV_TAB_H, XTMP3, XTMP3
-	VXOR XTMP1, XTMP3, XTMP3 // XTMP3 - bit reverse data bytes
-
+	VPERMXOR BIT_REV_TAB_L, BIT_REV_TAB_H, XDATA, XTMP3 // XTMP3 - bit reverse data bytes
+	
 	// ZUC authentication part, 4x32 data bits
 	// setup data
 	VSPLTISB $0, ZERO
-	MOVD $0x20, R8
+	MOVD $0x10, R8
 	LXVD2X (PTR)(R8), XTMP4
 	VPERM ZERO, XTMP3, XTMP4, XTMP1
-	MOVD $0x30, R8
+	MOVD $0x20, R8
 	LXVD2X (PTR)(R8), XTMP4
 	VPERM ZERO, XTMP3, XTMP4, XTMP2
 
@@ -72,7 +63,7 @@ TEXT ·eia256RoundTag8(SB),NOSPLIT,$0
 	LXVW4X (R8)(R4), KS_M1
 	MOVD $16, R8
 	LXVW4X (R8)(R4), KS_M2
-	MOVD $0x40, R8
+	MOVD $0x30, R8
 	LXVD2X (PTR)(R8), XTMP4
 	VPERM KS_L, KS_L, XTMP4, KS_L
 	VPERM KS_M1, KS_M1, XTMP4, KS_M1
@@ -131,26 +122,18 @@ TEXT ·eia256RoundTag16(SB),NOSPLIT,$0
 	VPERM XDATA, XDATA, XTMP1, XDATA
 #endif
 
-	LXVD2X (PTR)(R0), BIT_REV_AND_TAB
-	VAND	BIT_REV_AND_TAB, XDATA, XTMP3
 	VSPLTISB $4, XTMP2;
-	VSRW	XDATA, XTMP2, XTMP1
-	VAND	BIT_REV_AND_TAB, XTMP1, XTMP1
-
-	MOVD	$0x10, R8
-	LXVD2X (PTR)(R8), BIT_REV_TAB_L
+	LXVD2X (PTR)(R0), BIT_REV_TAB_L
 	VSLB  BIT_REV_TAB_L, XTMP2, BIT_REV_TAB_H
-	VPERM BIT_REV_TAB_L, BIT_REV_TAB_L, XTMP1, XTMP1
-	VPERM BIT_REV_TAB_H, BIT_REV_TAB_H, XTMP3, XTMP3
-	VXOR XTMP1, XTMP3, XTMP3 // XTMP3 - bit reverse data bytes
+	VPERMXOR BIT_REV_TAB_L, BIT_REV_TAB_H, XDATA, XTMP3 // XTMP3 - bit reverse data bytes
 
 	// ZUC authentication part, 4x32 data bits
 	// setup data
 	VSPLTISB $0, ZERO
-	MOVD $0x20, R8
+	MOVD $0x10, R8
 	LXVD2X (PTR)(R8), XTMP4
 	VPERM ZERO, XTMP3, XTMP4, XTMP1
-	MOVD $0x30, R8
+	MOVD $0x20, R8
 	LXVD2X (PTR)(R8), XTMP4
 	VPERM ZERO, XTMP3, XTMP4, XTMP2
 
@@ -161,7 +144,7 @@ TEXT ·eia256RoundTag16(SB),NOSPLIT,$0
 	MOVD $16, R8
 	LXVW4X (R8)(R4), KS_M2
 	VOR KS_M2, KS_M2, KS_H
-	MOVD $0x40, R8
+	MOVD $0x30, R8
 	LXVD2X (PTR)(R8), XTMP4
 	VPERM KS_L, KS_L, XTMP4, KS_L
 	VPERM KS_M1, KS_M1, XTMP4, KS_M1