not use VREPIF #19

emmansun · Oct 17, 2024 · ff84ed0 · ff84ed0
1 parent f66a91e
commit ff84ed0
Showing 1 changed file with 14 additions and 24 deletions.
diff --git a/base64_s390x.s b/base64_s390x.s
@@ -10,28 +10,22 @@ DATA base64_const<>+0x00(SB)/8, $0x0f0e0d0c0b0a0908 // enccoding, byte order
 DATA base64_const<>+0x08(SB)/8, $0x0706050403020100
 DATA base64_const<>+0x10(SB)/8, $0x0a0b090a07080607 // enccoding, reshufling
 DATA base64_const<>+0x18(SB)/8, $0x0405030401020001
-DATA base64_const<>+0x20(SB)/8, $0x1010010204080408 // STD LUT HI
-DATA base64_const<>+0x28(SB)/8, $0x1010101010101010
-DATA base64_const<>+0x30(SB)/8, $0x1511111111111111 // STD LUT LO
-DATA base64_const<>+0x38(SB)/8, $0x1111131A1B1B1B1A
-DATA base64_const<>+0x40(SB)/8, $0x00101304BFBFB9B9 // STD LUT ROLL
-DATA base64_const<>+0x48(SB)/8, $0x0000000000000000
-DATA base64_const<>+0x50(SB)/8, $0x1010010204080428 // URL LUT HI
-DATA base64_const<>+0x58(SB)/8, $0x1010101010101010
-DATA base64_const<>+0x60(SB)/8, $0x1511111111111111 // URL LUT LO
-DATA base64_const<>+0x68(SB)/8, $0x1111131B1B1A1B33
-DATA base64_const<>+0x70(SB)/8, $0x00001104BFBFE0B9 // URL LUT ROLL
-DATA base64_const<>+0x78(SB)/8, $0xB900000000000000
-DATA base64_const<>+0x80(SB)/8, $0x010203050607090a // RESHUFFLE MASK
-DATA base64_const<>+0x88(SB)/8, $0x0b0d0e0f00000000
-GLOBL base64_const<>(SB), (NOPTR+RODATA), $144
+DATA base64_const<>+0x20(SB)/8, $0x0fc0fc000fc0fc00 // mulhi mask
+DATA base64_const<>+0x28(SB)/8, $0x0fc0fc000fc0fc00
+DATA base64_const<>+0x30(SB)/8, $0x0400004004000040 // mulhi const
+DATA base64_const<>+0x38(SB)/8, $0x0400004004000040
+DATA base64_const<>+0x40(SB)/8, $0x003F03F0003F03F0 // mullo mask
+DATA base64_const<>+0x48(SB)/8, $0x003F03F0003F03F0
+DATA base64_const<>+0x50(SB)/8, $0x0100001001000010 // mullo const
+DATA base64_const<>+0x58(SB)/8, $0x0100001001000010
+GLOBL base64_const<>(SB), (NOPTR+RODATA), $96
 
 #define REV_BYTES V0
 #define RESHUFFLE_MASK V1
-#define MULHI_CONST V2
-#define MULHI_MASK V3
-#define MULLO_CONST V4
-#define MULLO_MASK V5
+#define MULHI_MASK V2
+#define MULHI_CONST V3
+#define MULLO_MASK V4
+#define MULLO_CONST V5
 #define RANGE1_END V6
 #define RANGE0_END V7
 #define LUT V8
@@ -50,11 +44,7 @@ TEXT ·encodeAsm(SB),NOSPLIT,$0
 	VL (R4), LUT
 
 	MOVD $base64_const<>(SB), R5
-	VLM (R5), REV_BYTES, RESHUFFLE_MASK
-	VREPIF $0x0fc0fc00, MULHI_MASK
-	VREPIF $0x04000040, MULHI_CONST
-	VREPIF $0x003F03F0, MULLO_MASK
-	VREPIF $0x01000010, MULLO_CONST
+	VLM (R5), REV_BYTES, MULLO_CONST
 	VREPIB $0x33, RANGE1_END
 	VREPIB $0x19, RANGE0_END
 	VZERO ZERO