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 +1,214 @@
+// Differential kernel: the shift vector slice against the Go
+// toolchain's loong64enc1.s rows (gasm v. go tool asm, byte for byte).
+
+#include "textflag.h"
+
+TEXT ·VSLLB(SB), NOSPLIT, $0
+	VSLLB V1, V2, V3
+	VSLLB $0, V1, V2
+	VSLLB $7, V1, V2
+	VSLLB $5, V1
+	RET
+
+TEXT ·VSLLH(SB), NOSPLIT, $0
+	VSLLH V1, V2, V3
+	VSLLH $0, V1, V2
+	VSLLH $15, V1, V2
+	VSLLH $10, V1
+	RET
+
+TEXT ·VSLLW(SB), NOSPLIT, $0
+	VSLLW V1, V2, V3
+	VSLLW $0, V1, V2
+	VSLLW $31, V1, V2
+	VSLLW $11, V1
+	RET
+
+TEXT ·VSLLV(SB), NOSPLIT, $0
+	VSLLV V1, V2, V3
+	VSLLV $0, V1, V2
+	VSLLV $63, V1, V2
+	VSLLV $30, V1
+	RET
+
+TEXT ·VSRLB(SB), NOSPLIT, $0
+	VSRLB V1, V2, V3
+	VSRLB $0, V1, V2
+	VSRLB $7, V1, V2
+	VSRLB $4, V1
+	RET
+
+TEXT ·VSRLH(SB), NOSPLIT, $0
+	VSRLH V1, V2, V3
+	VSRLH $0, V1, V2
+	VSRLH $15, V1, V2
+	VSRLH $9, V1
+	RET
+
+TEXT ·VSRLW(SB), NOSPLIT, $0
+	VSRLW V1, V2, V3
+	VSRLW $0, V1, V2
+	VSRLW $31, V1, V2
+	VSRLW $16, V1
+	RET
+
+TEXT ·VSRLV(SB), NOSPLIT, $0
+	VSRLV V1, V2, V3
+	VSRLV $0, V1, V2
+	VSRLV $63, V1, V2
+	VSRLV $40, V1
+	RET
+
+TEXT ·VSRAH(SB), NOSPLIT, $0
+	VSRAH V1, V2, V3
+	VSRAH $0, V1, V2
+	VSRAH $15, V1, V2
+	VSRAH $8, V1
+	RET
+
+TEXT ·VSRAW(SB), NOSPLIT, $0
+	VSRAW V1, V2, V3
+	VSRAW $0, V1, V2
+	VSRAW $31, V1, V2
+	VSRAW $12, V1
+	RET
+
+TEXT ·VSRAV(SB), NOSPLIT, $0
+	VSRAV V1, V2, V3
+	VSRAV $0, V1, V2
+	VSRAV $63, V1, V2
+	VSRAV $50, V1
+	RET
+
+TEXT ·VROTRB(SB), NOSPLIT, $0
+	VROTRB V1, V2, V3
+	VROTRB $0, V1, V2
+	VROTRB $7, V1, V2
+	VROTRB $3, V1
+	RET
+
+TEXT ·VROTRH(SB), NOSPLIT, $0
+	VROTRH V1, V2, V3
+	VROTRH $0, V1, V2
+	VROTRH $15, V1, V2
+	VROTRH $5, V1
+	RET
+
+TEXT ·VROTRV(SB), NOSPLIT, $0
+	VROTRV V1, V2, V3
+	VROTRV $0, V1, V2
+	VROTRV $63, V1, V2
+	VROTRV $52, V1
+	RET
+
+TEXT ·XVSLLB(SB), NOSPLIT, $0
+	XVSLLB X3, X2, X1
+	XVSLLB $0, X2, X1
+	XVSLLB $7, X2, X1
+	XVSLLB $4, X2
+	RET
+
+TEXT ·XVSLLH(SB), NOSPLIT, $0
+	XVSLLH X3, X2, X1
+	XVSLLH $0, X2, X1
+	XVSLLH $15, X2, X1
+	XVSLLH $8, X2
+	RET
+
+TEXT ·XVSLLW(SB), NOSPLIT, $0
+	XVSLLW X3, X2, X1
+	XVSLLW $0, X2, X1
+	XVSLLW $31, X2, X1
+	XVSLLW $13, X2
+	RET
+
+TEXT ·XVSLLV(SB), NOSPLIT, $0
+	XVSLLV X3, X2, X1
+	XVSLLV $0, X2, X1
+	XVSLLV $63, X2, X1
+	XVSLLV $36, X2
+	RET
+
+TEXT ·XVSRLB(SB), NOSPLIT, $0
+	XVSRLB X3, X2, X1
+	XVSRLB $0, X2, X1
+	XVSRLB $7, X2, X1
+	XVSRLB $5, X2
+	RET
+
+TEXT ·XVSRLH(SB), NOSPLIT, $0
+	XVSRLH X3, X2, X1
+	XVSRLH $0, X2, X1
+	XVSRLH $15, X2, X1
+	XVSRLH $9, X2
+	RET
+
+TEXT ·XVSRLW(SB), NOSPLIT, $0
+	XVSRLW X3, X2, X1
+	XVSRLW $0, X2, X1
+	XVSRLW $31, X2, X1
+	XVSRLW $14, X2
+	RET
+
+TEXT ·XVSRLV(SB), NOSPLIT, $0
+	XVSRLV X3, X2, X1
+	XVSRLV $0, X2, X1
+	XVSRLV $63, X2, X1
+	XVSRLV $45, X2
+	RET
+
+TEXT ·XVSRAB(SB), NOSPLIT, $0
+	XVSRAB X3, X2, X1
+	XVSRAB $0, X2, X1
+	XVSRAB $7, X2, X1
+	XVSRAB $6, X2
+	RET
+
+TEXT ·XVSRAH(SB), NOSPLIT, $0
+	XVSRAH X3, X2, X1
+	XVSRAH $0, X2, X1
+	XVSRAH $15, X2, X1
+	XVSRAH $10, X2
+	RET
+
+TEXT ·XVSRAW(SB), NOSPLIT, $0
+	XVSRAW X3, X2, X1
+	XVSRAW $0, X2, X1
+	XVSRAW $31, X2, X1
+	XVSRAW $16, X2
+	RET
+
+TEXT ·XVSRAV(SB), NOSPLIT, $0
+	XVSRAV X3, X2, X1
+	XVSRAV $0, X2, X1
+	XVSRAV $63, X2, X1
+	XVSRAV $48, X2
+	RET
+
+TEXT ·XVROTRB(SB), NOSPLIT, $0
+	XVROTRB X3, X2, X1
+	XVROTRB $0, X2, X1
+	XVROTRB $7, X2, X1
+	XVROTRB $3, X2
+	RET
+
+TEXT ·XVROTRH(SB), NOSPLIT, $0
+	XVROTRH X3, X2, X1
+	XVROTRH $0, X2, X1
+	XVROTRH $15, X2, X1
+	XVROTRH $13, X2
+	RET
+
+TEXT ·XVROTRW(SB), NOSPLIT, $0
+	XVROTRW X3, X2, X1
+	XVROTRW $0, X2, X1
+	XVROTRW $31, X2, X1
+	XVROTRW $24, X2
+	RET
+
+TEXT ·XVROTRV(SB), NOSPLIT, $0
+	XVROTRV X3, X2, X1
+	XVROTRV $0, X2, X1
+	XVROTRV $63, X2, X1
+	XVROTRV $52, X2
+	RET