// Take the top bit of each byte in X1 and put the result in DX.
	PMOVMSKB X1, DX
	// Find first set bit, if any.
	BSFL	DX, DX
	JNZ	ssesuccess
	// Advance to next block.
	ADDQ	$16, DI
sseloopentry:
	CMPQ	DI, AX
	JB	sseloop

	// Search the last 16-byte chunk. This chunk may overlap with the
	// chunks we've already searched, but that's ok.
	MOVQ	AX, DI
	MOVOU	(AX), X1
	PCMPEQB	X0, X1
	PMOVMSKB X1, DX
	BSFL	DX, DX

	// with some straightline code. The REP MOVSL instruction is really fast
	// 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
 */

	VMOVDQU XDWORD3, XDWORD1

	CMPQ SRND, $4*4*32
	JB   avx2_loop2

	MOVQ dig+0(FP), CTX // d.h[8]
	MOVQ _INP(SP), INP

	addm(  0(CTX), a)
	addm(  4(CTX), b)
	addm(  8(CTX), c)
	addm( 12(CTX), d)
	addm( 16(CTX), e)
	addm( 20(CTX), f)
	addm( 24(CTX), g)
	addm( 28(CTX), h)

	CMPQ _INP_END(SP), INP
	JB   done_hash

	XORQ SRND, SRND

	// into registers lets us ignore copy direction).
tail:
	// BSR+branch table make almost all memmove/memclr benchmarks worse. Not worth doing.
	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
#ifdef GO386_softfloat
	JMP	nosse2
#endif
	CMPL	BX, $32
	JBE	move_17through32
	CMPL	BX, $64
	JBE	move_33through64

	ROUND4(CX,DX,AX,BX, 9,0x2ad7d2bb,15);
	ROUND4(BX,CX,DX,AX, 0,0xeb86d391,21);

	ADDL	0(SP),	AX
	ADDL	4(SP),	BX
	ADDL	8(SP),	CX
	ADDL	12(SP),	DX

	ADDL	$64,		SI
	CMPL	SI,		16(SP)
	JB	loop

end:
	MOVL	dig+0(FP),	BP
	MOVL	AX,		(0*4)(BP)
	MOVL	BX,		(1*4)(BP)
	MOVL	CX,		(2*4)(BP)
	MOVL	DX,		(3*4)(BP)

	ROUND4(CX,DX,AX,BX, 9,0x2ad7d2bb,15);
	ROUND4(BX,CX,DX,AX, 0,0xeb86d391,21);

	ADDL	R12,	AX
	ADDL	R13,	BX
	ADDL	R14,	CX
	ADDL	R15,	DX

	ADDQ	$64,		SI
	CMPQ	SI,		DI
	JB	loop

end:
	MOVL	AX,		(0*4)(BP)
	MOVL	BX,		(1*4)(BP)
	MOVL	CX,		(2*4)(BP)
	MOVL	DX,		(3*4)(BP)

tail:
	// BSR+branch table make almost all memmove/memclr benchmarks worse. Not worth doing.
	TESTQ	BX, BX
	JEQ	_0
	CMPQ	BX, $2
	JBE	_1or2
	CMPQ	BX, $4
	JBE	_3or4
	CMPQ	BX, $8
	JB	_5through7
	JE	_8
	CMPQ	BX, $16
	JBE	_9through16
	CMPQ	BX, $32
	JBE	_17through32
	CMPQ	BX, $64
	JBE	_33through64
	CMPQ	BX, $128
	JBE	_65through128
	CMPQ	BX, $256
	JBE	_129through256

	MOVL	SP, BP
	MOVL	runtime·_TlsAlloc(SB), AX
	CALL	AX
	MOVL	BP, SP

	MOVL	AX, CX	// TLS index

	// Assert that slot is less than 64 so we can use _TEB->TlsSlots
	CMPL	CX, $64
	JB	ok
	// Fallback to the TEB arbitrary pointer.
	// TODO: don't use the arbitrary pointer (see go.dev/issue/59824)
	MOVL	$TEB_ArbitraryPtr, CX
	JMP	settls
ok:
	// Convert the TLS index at CX into

	// direction is forward.
	//
	// BSR+branch table make almost all memmove/memclr benchmarks worse. Not worth doing.
	TESTQ	BX, BX
	JEQ	move_0
	CMPQ	BX, $2
	JBE	move_1or2
	CMPQ	BX, $4
	JB	move_3
	JBE	move_4
	CMPQ	BX, $8
	JB	move_5through7
	JE	move_8
	CMPQ	BX, $16
	JBE	move_9through16
	CMPQ	BX, $32
	JBE	move_17through32
	CMPQ	BX, $64
	JBE	move_33through64
	CMPQ	BX, $128
	JBE	move_65through128

    //  it is explicitly spelled, e.g.
    //  Try<String> t = Try.successful("some")
    //  t.get().length()  // IDEA produces false positive "potential NPE" warning when JB annotations are used.
    //  Kotlin understands it correctly, though. JSpecify annotations of version 0.3 seem to work well too.
    private Try() {
    }

    /**