// for large sizes. The cutover is approximately 1K.
tail:
	TESTL	BX, BX
	JEQ	move_0
	CMPL	BX, $2
	JBE	move_1or2
	CMPL	BX, $4
	JB	move_3
	JE	move_4
	CMPL	BX, $8
	JBE	move_5through8
	CMPL	BX, $16
	JBE	move_9through16

/*
 * check and set for backwards
 */
	CMPL	SI, DI
	JLS	back

/*
 * forward copy loop
 */
forward:
	MOVL	BX, CX

	MOVL	trap+0(FP), AX	// syscall entry
	// copy args down
	LEAL	a1+4(FP), SI
	LEAL	sysargs-144(SP), DI
	CLD
	MOVSL
	MOVSL
	MOVSL
	INT	$64
	MOVL	AX, r1+16(FP)
	MOVL	$0, r2+20(FP)
	CMPL	AX, $-1
	JNE	ok3

	LEAL	errbuf-128(SP), AX
	MOVL	AX, sysargs-144(SP)
	MOVL	$128, sysargs1-140(SP)
	MOVL	$SYS_ERRSTR, AX
	INT	$64
	CALL	runtime·exitsyscall(SB)
	MOVL	sysargs-144(SP), AX
	MOVL	AX, errbuf-148(SP)

func ptrEqual() {
	// amd64:-"JEQ",-"JNE"
	// ppc64x:-"BEQ",-"BNE"
	// s390x:-"BEQ",-"BNE"
	copy(x[:], x[:])
}

func ptrOneOffset() {
	// amd64:-"JEQ",-"JNE"
	// ppc64x:-"BEQ",-"BNE"
	// s390x:-"BEQ",-"BNE"
	copy(x[1:], x[:])
}

func ptrBothOffset() {
	// amd64:-"JEQ",-"JNE"
	// ppc64x:-"BEQ",-"BNE"
	// s390x:-"BEQ",-"BNE"
	copy(x[1:], x[2:])
}

	// will clobber the test pattern created by the caller
	ADDL	$(1024 * 8), SP

	// Ask signaller to setgid
	MOVL	$1, (BX)

	// Wait for setgid completion
loop:
	PAUSE
	MOVL	(BX), AX
	CMPL	AX, $0
	JNE	loop

	// Restore stack
	SUBL	$(1024 * 8), SP

aligned:
	MOVQ $0, AX // position in slices

	PCALIGN $16
loop16b:
	MOVOU (SI)(AX*1), X0   // XOR 16byte forwards.
	MOVOU (CX)(AX*1), X1
	PXOR  X1, X0
	MOVOU X0, (BX)(AX*1)
	ADDQ  $16, AX
	CMPQ  DX, AX
	JNE   loop16b
	RET

	PCALIGN $16
loop_1b:
	SUBQ  $1, DX           // XOR 1byte backwards.
	MOVB  (SI)(DX*1), DI
	MOVB  (CX)(DX*1), AX
	XORB  AX, DI
	MOVB  DI, (BX)(DX*1)

	MOVQ	trap+0(FP), BP	// syscall entry
	// copy args down
	LEAQ	a1+8(FP), SI
	LEAQ	sysargs-160(SP), DI
	CLD
	MOVSQ
	MOVSQ
	MOVSQ
	SYSCALL
	MOVQ	AX, r1+32(FP)
	MOVQ	$0, r2+40(FP)
	CMPL	AX, $-1
	JNE	ok3

	LEAQ	errbuf-128(SP), AX
	MOVQ	AX, sysargs-160(SP)
	MOVQ	$128, sysargs1-152(SP)
	MOVQ	$SYS_ERRSTR, BP
	SYSCALL
	CALL	runtime·exitsyscall(SB) // call via ABI wrapper, ensuring ABIInternal fixed registers are set

	MOVL	(_INTERRUPT_TIME+time_hi2), DI
	CMPL	AX, DI
	JNE	loop

	// w = DI:CX
	// multiply by 100
	MOVL	$100, AX
	MULL	CX
	IMULL	$100, DI
	ADDL	DI, DX
	// w*100 = DX:AX
	MOVL	AX, mono+12(FP)
	MOVL	DX, mono+16(FP)

wall:
	MOVL	(_SYSTEM_TIME+time_hi1), CX
	MOVL	(_SYSTEM_TIME+time_lo), AX
	MOVL	(_SYSTEM_TIME+time_hi2), DX
	CMPL	CX, DX
	JNE	wall

	// w = DX:AX

	ADDQ	$(1024 * 8), SP

	// Ask signaller to setgid
	MOVL	$1, ·Baton(SB)

	// Wait for setgid completion
loop:
	PAUSE
	MOVL	·Baton(SB), AX
	CMPL	AX, $0
	JNE	loop

	// Restore stack
	SUBQ	$(1024 * 8), SP

	PAND	X4, X0
	PMOVMSKB X0, DX
	ADDL	$64, SI
	ADDL	$64, DI
	SUBL	$64, BX
	CMPL	DX, $0xffff
	JEQ	hugeloop
	MOVB	$0, (AX)
	RET

	// 4 bytes at a time using 32-bit register
bigloop:
	CMPL	BX, $4
	JBE	leftover
	MOVL	(SI), CX
	MOVL	(DI), DX
	ADDL	$4, SI
	ADDL	$4, DI
	SUBL	$4, BX
	CMPL	CX, DX
	JEQ	bigloop
	MOVB	$0, (AX)
	RET

	// remaining 0-4 bytes
leftover:
	MOVL	-4(SI)(BX*1), CX

	MOVQ    BX, 24(SP)
	MOVQ    BP, 32(SP)
	MOVQ	R12, 40(SP)
	MOVQ	R13, 48(SP)
	MOVQ	R14, 56(SP)
	MOVQ	R15, 64(SP)

	get_tls(BX)
	// check that g exists
	MOVQ	g(BX), R10
	CMPQ	R10, $0
	JNE	allgood
	MOVQ	SI, 72(SP)
	MOVQ	DX, 80(SP)
	LEAQ	72(SP), AX
	MOVQ	DI, 0(SP)
	MOVQ	AX, 8(SP)
	MOVQ	$runtime·badsignal(SB), AX
	CALL	AX
	JMP	exit

allgood: