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 +1,220 @@
+// Differential kernel: the arith_sub vector slice against the Go
+// toolchain's loong64enc1.s rows (gasm v. go tool asm, byte for byte).
+
+#include "textflag.h"
+
+TEXT ·VSUBB(SB), NOSPLIT, $0
+	VSUBB V1, V2, V3
+	RET
+
+TEXT ·VSUBH(SB), NOSPLIT, $0
+	VSUBH V1, V2, V3
+	RET
+
+TEXT ·VSUBW(SB), NOSPLIT, $0
+	VSUBW V1, V2, V3
+	RET
+
+TEXT ·VSUBV(SB), NOSPLIT, $0
+	VSUBV V1, V2, V3
+	RET
+
+TEXT ·VSUBQ(SB), NOSPLIT, $0
+	VSUBQ V1, V2, V3
+	RET
+
+TEXT ·XVSUBB(SB), NOSPLIT, $0
+	XVSUBB X3, X2, X1
+	RET
+
+TEXT ·XVSUBH(SB), NOSPLIT, $0
+	XVSUBH X3, X2, X1
+	RET
+
+TEXT ·XVSUBW(SB), NOSPLIT, $0
+	XVSUBW X3, X2, X1
+	RET
+
+TEXT ·XVSUBV(SB), NOSPLIT, $0
+	XVSUBV X3, X2, X1
+	RET
+
+TEXT ·XVSUBQ(SB), NOSPLIT, $0
+	XVSUBQ X3, X2, X1
+	RET
+
+TEXT ·VSUBBU(SB), NOSPLIT, $0
+	VSUBBU $5, V2, V1
+	RET
+
+TEXT ·VSUBHU(SB), NOSPLIT, $0
+	VSUBHU $6, V2, V1
+	RET
+
+TEXT ·VSUBWU(SB), NOSPLIT, $0
+	VSUBWU $7, V2, V1
+	RET
+
+TEXT ·VSUBVU(SB), NOSPLIT, $0
+	VSUBVU $8, V2, V1
+	RET
+
+TEXT ·XVSUBBU(SB), NOSPLIT, $0
+	XVSUBBU $13, X1, X2
+	RET
+
+TEXT ·XVSUBHU(SB), NOSPLIT, $0
+	XVSUBHU $14, X1, X2
+	RET
+
+TEXT ·XVSUBWU(SB), NOSPLIT, $0
+	XVSUBWU $15, X1, X2
+	RET
+
+TEXT ·XVSUBVU(SB), NOSPLIT, $0
+	XVSUBVU $16, X1, X2
+	RET
+
+TEXT ·VSUBWEVHB(SB), NOSPLIT, $0
+	VSUBWEVHB V1, V2, V3
+	RET
+
+TEXT ·VSUBWEVWH(SB), NOSPLIT, $0
+	VSUBWEVWH V1, V2, V3
+	RET
+
+TEXT ·VSUBWEVVW(SB), NOSPLIT, $0
+	VSUBWEVVW V1, V2, V3
+	RET
+
+TEXT ·VSUBWEVQV(SB), NOSPLIT, $0
+	VSUBWEVQV V1, V2, V3
+	RET
+
+TEXT ·VSUBWODHB(SB), NOSPLIT, $0
+	VSUBWODHB V1, V2, V3
+	RET
+
+TEXT ·VSUBWODWH(SB), NOSPLIT, $0
+	VSUBWODWH V1, V2, V3
+	RET
+
+TEXT ·VSUBWODVW(SB), NOSPLIT, $0
+	VSUBWODVW V1, V2, V3
+	RET
+
+TEXT ·VSUBWODQV(SB), NOSPLIT, $0
+	VSUBWODQV V1, V2, V3
+	RET
+
+TEXT ·XVSUBWEVHB(SB), NOSPLIT, $0
+	XVSUBWEVHB X1, X2, X3
+	RET
+
+TEXT ·XVSUBWEVWH(SB), NOSPLIT, $0
+	XVSUBWEVWH X1, X2, X3
+	RET
+
+TEXT ·XVSUBWEVVW(SB), NOSPLIT, $0
+	XVSUBWEVVW X1, X2, X3
+	RET
+
+TEXT ·XVSUBWEVQV(SB), NOSPLIT, $0
+	XVSUBWEVQV X1, X2, X3
+	RET
+
+TEXT ·XVSUBWODHB(SB), NOSPLIT, $0
+	XVSUBWODHB X1, X2, X3
+	RET
+
+TEXT ·XVSUBWODWH(SB), NOSPLIT, $0
+	XVSUBWODWH X1, X2, X3
+	RET
+
+TEXT ·XVSUBWODVW(SB), NOSPLIT, $0
+	XVSUBWODVW X1, X2, X3
+	RET
+
+TEXT ·XVSUBWODQV(SB), NOSPLIT, $0
+	XVSUBWODQV X1, X2, X3
+	RET
+
+TEXT ·VSUBWEVHBU(SB), NOSPLIT, $0
+	VSUBWEVHBU V1, V2, V3
+	RET
+
+TEXT ·VSUBWEVWHU(SB), NOSPLIT, $0
+	VSUBWEVWHU V1, V2, V3
+	RET
+
+TEXT ·VSUBWEVVWU(SB), NOSPLIT, $0
+	VSUBWEVVWU V1, V2, V3
+	RET
+
+TEXT ·VSUBWEVQVU(SB), NOSPLIT, $0
+	VSUBWEVQVU V1, V2, V3
+	RET
+
+TEXT ·VSUBWODHBU(SB), NOSPLIT, $0
+	VSUBWODHBU V1, V2, V3
+	RET
+
+TEXT ·VSUBWODWHU(SB), NOSPLIT, $0
+	VSUBWODWHU V1, V2, V3
+	RET
+
+TEXT ·VSUBWODVWU(SB), NOSPLIT, $0
+	VSUBWODVWU V1, V2, V3
+	RET
+
+TEXT ·VSUBWODQVU(SB), NOSPLIT, $0
+	VSUBWODQVU V1, V2, V3
+	RET
+
+TEXT ·XVSUBWEVHBU(SB), NOSPLIT, $0
+	XVSUBWEVHBU X1, X2, X3
+	RET
+
+TEXT ·XVSUBWEVWHU(SB), NOSPLIT, $0
+	XVSUBWEVWHU X1, X2, X3
+	RET
+
+TEXT ·XVSUBWEVVWU(SB), NOSPLIT, $0
+	XVSUBWEVVWU X1, X2, X3
+	RET
+
+TEXT ·XVSUBWEVQVU(SB), NOSPLIT, $0
+	XVSUBWEVQVU X1, X2, X3
+	RET
+
+TEXT ·XVSUBWODHBU(SB), NOSPLIT, $0
+	XVSUBWODHBU X1, X2, X3
+	RET
+
+TEXT ·XVSUBWODWHU(SB), NOSPLIT, $0
+	XVSUBWODWHU X1, X2, X3
+	RET
+
+TEXT ·XVSUBWODVWU(SB), NOSPLIT, $0
+	XVSUBWODVWU X1, X2, X3
+	RET
+
+TEXT ·XVSUBWODQVU(SB), NOSPLIT, $0
+	XVSUBWODQVU X1, X2, X3
+	RET
+
+TEXT ·VSUBF(SB), NOSPLIT, $0
+	VSUBF V1, V2, V3
+	RET
+
+TEXT ·VSUBD(SB), NOSPLIT, $0
+	VSUBD V1, V2, V3
+	RET
+
+TEXT ·XVSUBF(SB), NOSPLIT, $0
+	XVSUBF X1, X2, X3
+	RET
+
+TEXT ·XVSUBD(SB), NOSPLIT, $0
+	XVSUBD X1, X2, X3
+	RET