ADD	$(1024 * 8), RSP

	// Ask signaller to setgid
	MOVD	$·Baton(SB), R0
	MOVD	$1, R1
storeloop:
	LDAXRW	(R0), R2
	STLXRW	R1, (R0), R3
	CBNZ	R3, storeloop

	// Wait for setgid completion
	MOVW	$0, R1
	MOVW	$0, R2
loop:
	LDAXRW	(R0), R3
	CMPW	R1, R3
	BNE	loop
	STLXRW	R2, (R0), R3
	CBNZ	R3, loop

	// Restore stack
	SUB	$(1024 * 8), RSP

	MOVD	R9, R30

#ifndef GOARM64_LSE
	MOVBU	internal∕cpu·ARM64+const_offsetARM64HasATOMICS(SB), R4
	CBZ 	R4, load_store_loop
#endif
	SWPALW	R1, (R0), R2
	MOVW	R2, ret+16(FP)
	RET
#ifndef GOARM64_LSE
load_store_loop:
	LDAXRW	(R0), R2
	STLXRW	R1, (R0), R3
	CBNZ	R3, load_store_loop
	MOVW	R2, ret+16(FP)
	RET
#endif

// uint64 Xchg64(ptr *uint64, new uint64)
// Atomically:
//	old := *ptr;
//	*ptr = new;

	"LDADDH",
	"LDADDLB",
	"LDADDLD",
	"LDADDLH",
	"LDADDLW",
	"LDADDW",
	"LDAR",
	"LDARB",
	"LDARH",
	"LDARW",
	"LDAXP",
	"LDAXPW",
	"LDAXR",
	"LDAXRB",
	"LDAXRH",
	"LDAXRW",
	"LDCLRAB",
	"LDCLRAD",
	"LDCLRAH",
	"LDCLRALB",
	"LDCLRALD",
	"LDCLRALH",
	"LDCLRALW",
	"LDCLRAW",
	"LDCLRB",
	"LDCLRD",
	"LDCLRH",
	"LDCLRLB",
	"LDCLRLD",
	"LDCLRLH",