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 +1,209 @@
+// Differential kernel: the arith_add vector slice against the Go
+// toolchain's loong64enc1.s rows (gasm v. go tool asm, byte for byte).
+
+#include "textflag.h"
+
+TEXT ·VADDB(SB), NOSPLIT, $0
+	VADDB V1, V2, V3
+	RET
+
+TEXT ·VADDH(SB), NOSPLIT, $0
+	VADDH V1, V2, V3
+	RET
+
+TEXT ·VADDQ(SB), NOSPLIT, $0
+	VADDQ V1, V2, V3
+	RET
+
+TEXT ·XVADDB(SB), NOSPLIT, $0
+	XVADDB X3, X2, X1
+	RET
+
+TEXT ·XVADDH(SB), NOSPLIT, $0
+	XVADDH X3, X2, X1
+	RET
+
+TEXT ·XVADDW(SB), NOSPLIT, $0
+	XVADDW X3, X2, X1
+	RET
+
+TEXT ·XVADDQ(SB), NOSPLIT, $0
+	XVADDQ X3, X2, X1
+	RET
+
+TEXT ·VADDBU(SB), NOSPLIT, $0
+	VADDBU $1, V2
+	VADDBU $1, V2, V1
+	RET
+
+TEXT ·VADDHU(SB), NOSPLIT, $0
+	VADDHU $2, V2, V1
+	RET
+
+TEXT ·VADDWU(SB), NOSPLIT, $0
+	VADDWU $3, V2, V1
+	RET
+
+TEXT ·VADDVU(SB), NOSPLIT, $0
+	VADDVU $4, V2, V1
+	RET
+
+TEXT ·XVADDBU(SB), NOSPLIT, $0
+	XVADDBU $9, X1, X2
+	RET
+
+TEXT ·XVADDHU(SB), NOSPLIT, $0
+	XVADDHU $10, X1, X2
+	RET
+
+TEXT ·XVADDWU(SB), NOSPLIT, $0
+	XVADDWU $11, X1, X2
+	RET
+
+TEXT ·XVADDVU(SB), NOSPLIT, $0
+	XVADDVU $12, X1, X2
+	RET
+
+TEXT ·VADDWEVHB(SB), NOSPLIT, $0
+	VADDWEVHB V1, V2, V3
+	RET
+
+TEXT ·VADDWEVWH(SB), NOSPLIT, $0
+	VADDWEVWH V1, V2, V3
+	RET
+
+TEXT ·VADDWEVVW(SB), NOSPLIT, $0
+	VADDWEVVW V1, V2, V3
+	RET
+
+TEXT ·VADDWEVQV(SB), NOSPLIT, $0
+	VADDWEVQV V1, V2, V3
+	RET
+
+TEXT ·VADDWODHB(SB), NOSPLIT, $0
+	VADDWODHB V1, V2, V3
+	RET
+
+TEXT ·VADDWODWH(SB), NOSPLIT, $0
+	VADDWODWH V1, V2, V3
+	RET
+
+TEXT ·VADDWODVW(SB), NOSPLIT, $0
+	VADDWODVW V1, V2, V3
+	RET
+
+TEXT ·VADDWODQV(SB), NOSPLIT, $0
+	VADDWODQV V1, V2, V3
+	RET
+
+TEXT ·XVADDWEVHB(SB), NOSPLIT, $0
+	XVADDWEVHB X1, X2, X3
+	RET
+
+TEXT ·XVADDWEVWH(SB), NOSPLIT, $0
+	XVADDWEVWH X1, X2, X3
+	RET
+
+TEXT ·XVADDWEVVW(SB), NOSPLIT, $0
+	XVADDWEVVW X1, X2, X3
+	RET
+
+TEXT ·XVADDWEVQV(SB), NOSPLIT, $0
+	XVADDWEVQV X1, X2, X3
+	RET
+
+TEXT ·XVADDWODHB(SB), NOSPLIT, $0
+	XVADDWODHB X1, X2, X3
+	RET
+
+TEXT ·XVADDWODWH(SB), NOSPLIT, $0
+	XVADDWODWH X1, X2, X3
+	RET
+
+TEXT ·XVADDWODVW(SB), NOSPLIT, $0
+	XVADDWODVW X1, X2, X3
+	RET
+
+TEXT ·XVADDWODQV(SB), NOSPLIT, $0
+	XVADDWODQV X1, X2, X3
+	RET
+
+TEXT ·VADDWEVHBU(SB), NOSPLIT, $0
+	VADDWEVHBU V1, V2, V3
+	RET
+
+TEXT ·VADDWEVWHU(SB), NOSPLIT, $0
+	VADDWEVWHU V1, V2, V3
+	RET
+
+TEXT ·VADDWEVVWU(SB), NOSPLIT, $0
+	VADDWEVVWU V1, V2, V3
+	RET
+
+TEXT ·VADDWEVQVU(SB), NOSPLIT, $0
+	VADDWEVQVU V1, V2, V3
+	RET
+
+TEXT ·VADDWODHBU(SB), NOSPLIT, $0
+	VADDWODHBU V1, V2, V3
+	RET
+
+TEXT ·VADDWODWHU(SB), NOSPLIT, $0
+	VADDWODWHU V1, V2, V3
+	RET
+
+TEXT ·VADDWODVWU(SB), NOSPLIT, $0
+	VADDWODVWU V1, V2, V3
+	RET
+
+TEXT ·VADDWODQVU(SB), NOSPLIT, $0
+	VADDWODQVU V1, V2, V3
+	RET
+
+TEXT ·XVADDWEVHBU(SB), NOSPLIT, $0
+	XVADDWEVHBU X1, X2, X3
+	RET
+
+TEXT ·XVADDWEVWHU(SB), NOSPLIT, $0
+	XVADDWEVWHU X1, X2, X3
+	RET
+
+TEXT ·XVADDWEVVWU(SB), NOSPLIT, $0
+	XVADDWEVVWU X1, X2, X3
+	RET
+
+TEXT ·XVADDWEVQVU(SB), NOSPLIT, $0
+	XVADDWEVQVU X1, X2, X3
+	RET
+
+TEXT ·XVADDWODHBU(SB), NOSPLIT, $0
+	XVADDWODHBU X1, X2, X3
+	RET
+
+TEXT ·XVADDWODWHU(SB), NOSPLIT, $0
+	XVADDWODWHU X1, X2, X3
+	RET
+
+TEXT ·XVADDWODVWU(SB), NOSPLIT, $0
+	XVADDWODVWU X1, X2, X3
+	RET
+
+TEXT ·XVADDWODQVU(SB), NOSPLIT, $0
+	XVADDWODQVU X1, X2, X3
+	RET
+
+TEXT ·VADDF(SB), NOSPLIT, $0
+	VADDF V1, V2, V3
+	RET
+
+TEXT ·VADDD(SB), NOSPLIT, $0
+	VADDD V1, V2, V3
+	RET
+
+TEXT ·XVADDF(SB), NOSPLIT, $0
+	XVADDF X1, X2, X3
+	RET
+
+TEXT ·XVADDD(SB), NOSPLIT, $0
+	XVADDD X1, X2, X3
+	RET