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 +1,204 @@
+// Differential kernel: the maddwiden vector slice against the Go
+// toolchain's loong64enc1.s rows (gasm v. go tool asm, byte for byte).
+
+#include "textflag.h"
+
+TEXT ·VMADDW(SB), NOSPLIT, $0
+	VMADDW V1, V2, V3
+	RET
+
+TEXT ·XVMADDW(SB), NOSPLIT, $0
+	XVMADDW X1, X2, X3
+	RET
+
+TEXT ·VMADDWEVHB(SB), NOSPLIT, $0
+	VMADDWEVHB V1, V2, V3
+	RET
+
+TEXT ·VMADDWEVWH(SB), NOSPLIT, $0
+	VMADDWEVWH V1, V2, V3
+	RET
+
+TEXT ·VMADDWEVVW(SB), NOSPLIT, $0
+	VMADDWEVVW V1, V2, V3
+	RET
+
+TEXT ·VMADDWEVQV(SB), NOSPLIT, $0
+	VMADDWEVQV V1, V2, V3
+	RET
+
+TEXT ·VMADDWODHB(SB), NOSPLIT, $0
+	VMADDWODHB V1, V2, V3
+	RET
+
+TEXT ·VMADDWODWH(SB), NOSPLIT, $0
+	VMADDWODWH V1, V2, V3
+	RET
+
+TEXT ·VMADDWODVW(SB), NOSPLIT, $0
+	VMADDWODVW V1, V2, V3
+	RET
+
+TEXT ·VMADDWODQV(SB), NOSPLIT, $0
+	VMADDWODQV V1, V2, V3
+	RET
+
+TEXT ·XVMADDWEVHB(SB), NOSPLIT, $0
+	XVMADDWEVHB X1, X2, X3
+	RET
+
+TEXT ·XVMADDWEVWH(SB), NOSPLIT, $0
+	XVMADDWEVWH X1, X2, X3
+	RET
+
+TEXT ·XVMADDWEVVW(SB), NOSPLIT, $0
+	XVMADDWEVVW X1, X2, X3
+	RET
+
+TEXT ·XVMADDWEVQV(SB), NOSPLIT, $0
+	XVMADDWEVQV X1, X2, X3
+	RET
+
+TEXT ·XVMADDWODHB(SB), NOSPLIT, $0
+	XVMADDWODHB X1, X2, X3
+	RET
+
+TEXT ·XVMADDWODWH(SB), NOSPLIT, $0
+	XVMADDWODWH X1, X2, X3
+	RET
+
+TEXT ·XVMADDWODVW(SB), NOSPLIT, $0
+	XVMADDWODVW X1, X2, X3
+	RET
+
+TEXT ·XVMADDWODQV(SB), NOSPLIT, $0
+	XVMADDWODQV X1, X2, X3
+	RET
+
+TEXT ·VMADDWEVHBU(SB), NOSPLIT, $0
+	VMADDWEVHBU V1, V2, V3
+	RET
+
+TEXT ·VMADDWEVWHU(SB), NOSPLIT, $0
+	VMADDWEVWHU V1, V2, V3
+	RET
+
+TEXT ·VMADDWEVVWU(SB), NOSPLIT, $0
+	VMADDWEVVWU V1, V2, V3
+	RET
+
+TEXT ·VMADDWEVQVU(SB), NOSPLIT, $0
+	VMADDWEVQVU V1, V2, V3
+	RET
+
+TEXT ·VMADDWODHBU(SB), NOSPLIT, $0
+	VMADDWODHBU V1, V2, V3
+	RET
+
+TEXT ·VMADDWODWHU(SB), NOSPLIT, $0
+	VMADDWODWHU V1, V2, V3
+	RET
+
+TEXT ·VMADDWODVWU(SB), NOSPLIT, $0
+	VMADDWODVWU V1, V2, V3
+	RET
+
+TEXT ·VMADDWODQVU(SB), NOSPLIT, $0
+	VMADDWODQVU V1, V2, V3
+	RET
+
+TEXT ·XVMADDWEVHBU(SB), NOSPLIT, $0
+	XVMADDWEVHBU X1, X2, X3
+	RET
+
+TEXT ·XVMADDWEVWHU(SB), NOSPLIT, $0
+	XVMADDWEVWHU X1, X2, X3
+	RET
+
+TEXT ·XVMADDWEVVWU(SB), NOSPLIT, $0
+	XVMADDWEVVWU X1, X2, X3
+	RET
+
+TEXT ·XVMADDWEVQVU(SB), NOSPLIT, $0
+	XVMADDWEVQVU X1, X2, X3
+	RET
+
+TEXT ·XVMADDWODHBU(SB), NOSPLIT, $0
+	XVMADDWODHBU X1, X2, X3
+	RET
+
+TEXT ·XVMADDWODWHU(SB), NOSPLIT, $0
+	XVMADDWODWHU X1, X2, X3
+	RET
+
+TEXT ·XVMADDWODVWU(SB), NOSPLIT, $0
+	XVMADDWODVWU X1, X2, X3
+	RET
+
+TEXT ·XVMADDWODQVU(SB), NOSPLIT, $0
+	XVMADDWODQVU X1, X2, X3
+	RET
+
+TEXT ·VMADDWEVHBUB(SB), NOSPLIT, $0
+	VMADDWEVHBUB V1, V2, V3
+	RET
+
+TEXT ·VMADDWEVWHUH(SB), NOSPLIT, $0
+	VMADDWEVWHUH V1, V2, V3
+	RET
+
+TEXT ·VMADDWEVVWUW(SB), NOSPLIT, $0
+	VMADDWEVVWUW V1, V2, V3
+	RET
+
+TEXT ·VMADDWEVQVUV(SB), NOSPLIT, $0
+	VMADDWEVQVUV V1, V2, V3
+	RET
+
+TEXT ·VMADDWODHBUB(SB), NOSPLIT, $0
+	VMADDWODHBUB V1, V2, V3
+	RET
+
+TEXT ·VMADDWODWHUH(SB), NOSPLIT, $0
+	VMADDWODWHUH V1, V2, V3
+	RET
+
+TEXT ·VMADDWODVWUW(SB), NOSPLIT, $0
+	VMADDWODVWUW V1, V2, V3
+	RET
+
+TEXT ·VMADDWODQVUV(SB), NOSPLIT, $0
+	VMADDWODQVUV V1, V2, V3
+	RET
+
+TEXT ·XVMADDWEVHBUB(SB), NOSPLIT, $0
+	XVMADDWEVHBUB X1, X2, X3
+	RET
+
+TEXT ·XVMADDWEVWHUH(SB), NOSPLIT, $0
+	XVMADDWEVWHUH X1, X2, X3
+	RET
+
+TEXT ·XVMADDWEVVWUW(SB), NOSPLIT, $0
+	XVMADDWEVVWUW X1, X2, X3
+	RET
+
+TEXT ·XVMADDWEVQVUV(SB), NOSPLIT, $0
+	XVMADDWEVQVUV X1, X2, X3
+	RET
+
+TEXT ·XVMADDWODHBUB(SB), NOSPLIT, $0
+	XVMADDWODHBUB X1, X2, X3
+	RET
+
+TEXT ·XVMADDWODWHUH(SB), NOSPLIT, $0
+	XVMADDWODWHUH X1, X2, X3
+	RET
+
+TEXT ·XVMADDWODVWUW(SB), NOSPLIT, $0
+	XVMADDWODVWUW X1, X2, X3
+	RET
+
+TEXT ·XVMADDWODQVUV(SB), NOSPLIT, $0
+	XVMADDWODQVUV X1, X2, X3
+	RET