feat(loong64): encode the full LSX and LASX table

Assisted-by: GLM 5.3 Flash
2026-09-20 19:14:43 +02:00
parent d6cf7cfa44
commit 289cabe993
22 changed files with 3528 additions and 12 deletions
@@ -0,0 +1,132 @@
+// Differential kernel: the shuffle vector slice against the Go
+// toolchain's loong64enc1.s rows (gasm v. go tool asm, byte for byte).
+
+#include "textflag.h"
+
+TEXT ·VSHUF4IB(SB), NOSPLIT, $0
+	VSHUF4IB $0, V2, V1
+	VSHUF4IB $16, V2, V1
+	VSHUF4IB $255, V2, V1
+	RET
+
+TEXT ·VSHUF4IH(SB), NOSPLIT, $0
+	VSHUF4IH $0, V2, V1
+	VSHUF4IH $128, V2, V1
+	VSHUF4IH $255, V2, V1
+	RET
+
+TEXT ·VSHUF4IW(SB), NOSPLIT, $0
+	VSHUF4IW $0, V2, V1
+	VSHUF4IW $96, V2, V1
+	VSHUF4IW $255, V2, V1
+	RET
+
+TEXT ·VSHUF4IV(SB), NOSPLIT, $0
+	VSHUF4IV $0, V2, V1
+	VSHUF4IV $8, V2, V1
+	VSHUF4IV $15, V2, V1
+	RET
+
+TEXT ·XVSHUF4IB(SB), NOSPLIT, $0
+	XVSHUF4IB $0, X1, X2
+	XVSHUF4IB $16, X1, X2
+	XVSHUF4IB $255, X1, X2
+	RET
+
+TEXT ·XVSHUF4IH(SB), NOSPLIT, $0
+	XVSHUF4IH $0, X1, X2
+	XVSHUF4IH $128, X1, X2
+	XVSHUF4IH $255, X1, X2
+	RET
+
+TEXT ·XVSHUF4IW(SB), NOSPLIT, $0
+	XVSHUF4IW $0, X1, X2
+	XVSHUF4IW $96, X1, X2
+	XVSHUF4IW $255, X1, X2
+	RET
+
+TEXT ·XVSHUF4IV(SB), NOSPLIT, $0
+	XVSHUF4IV $0, X1, X2
+	XVSHUF4IV $8, X1, X2
+	XVSHUF4IV $15, X1, X2
+	RET
+
+TEXT ·VSHUFH(SB), NOSPLIT, $0
+	VSHUFH V1, V2, V3
+	RET
+
+TEXT ·VSHUFW(SB), NOSPLIT, $0
+	VSHUFW V1, V2, V3
+	RET
+
+TEXT ·VSHUFV(SB), NOSPLIT, $0
+	VSHUFV V1, V2, V3
+	RET
+
+TEXT ·XVSHUFH(SB), NOSPLIT, $0
+	XVSHUFH X1, X2, X3
+	RET
+
+TEXT ·XVSHUFW(SB), NOSPLIT, $0
+	XVSHUFW X1, X2, X3
+	RET
+
+TEXT ·XVSHUFV(SB), NOSPLIT, $0
+	XVSHUFV X1, X2, X3
+	RET
+
+TEXT ·VSHUFB(SB), NOSPLIT, $0
+	VSHUFB V1, V2, V3, V4
+	RET
+
+TEXT ·XVSHUFB(SB), NOSPLIT, $0
+	XVSHUFB X1, X2, X3, X4
+	RET
+
+TEXT ·VPERMIW(SB), NOSPLIT, $0
+	VPERMIW $0x1B, V1, V2
+	RET
+
+TEXT ·XVPERMIW(SB), NOSPLIT, $0
+	XVPERMIW $0x2B, X1, X2
+	RET
+
+TEXT ·XVPERMIV(SB), NOSPLIT, $0
+	XVPERMIV $0x3B, X1, X2
+	RET
+
+TEXT ·XVPERMIQ(SB), NOSPLIT, $0
+	XVPERMIQ $0x4B, X1, X2
+	RET
+
+TEXT ·VEXTRINSB(SB), NOSPLIT, $0
+	VEXTRINSB $0x18, V1, V2
+	RET
+
+TEXT ·VEXTRINSH(SB), NOSPLIT, $0
+	VEXTRINSH $0x27, V1, V2
+	RET
+
+TEXT ·VEXTRINSW(SB), NOSPLIT, $0
+	VEXTRINSW $0x36, V1, V2
+	RET
+
+TEXT ·VEXTRINSV(SB), NOSPLIT, $0
+	VEXTRINSV $0x45, V1, V2
+	RET
+
+TEXT ·XVEXTRINSB(SB), NOSPLIT, $0
+	XVEXTRINSB $0x54, X1, X2
+	RET
+
+TEXT ·XVEXTRINSH(SB), NOSPLIT, $0
+	XVEXTRINSH $0x63, X1, X2
+	RET
+
+TEXT ·XVEXTRINSW(SB), NOSPLIT, $0
+	XVEXTRINSW $0x72, X1, X2
+	RET
+
+TEXT ·XVEXTRINSV(SB), NOSPLIT, $0
+	XVEXTRINSV $0x81, X1, X2
+	RET