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
	JNZ	ssesuccess

failure:
	MOVQ $-1, (R8)
	RET

	MOVL 	CX, DX
	MOVL	(BX), AX
	ANDL	AX, DX
	LOCK
	CMPXCHGL	DX, (BX)
	JNZ casloop
	MOVL 	AX, ret+8(FP)
	RET

// func Or32(addr *uint32, v uint32) old uint32
TEXT ·Or32(SB), NOSPLIT, $0-12
	MOVL	ptr+0(FP), BX
	MOVL	val+4(FP), CX
casloop:
	MOVL 	CX, DX
	MOVL	(BX), AX
	ORL	AX, DX
	LOCK
	CMPXCHGL	DX, (BX)
	JNZ casloop
	MOVL 	AX, ret+8(FP)
	RET

TEXT ·xorBytes(SB), NOSPLIT, $0
	MOVQ  dst+0(FP), BX
	MOVQ  a+8(FP), SI
	MOVQ  b+16(FP), CX
	MOVQ  n+24(FP), DX
	TESTQ $15, DX            // AND 15 & len, if not zero jump to not_aligned.
	JNZ   not_aligned

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

	MOVL 	CX, DX
	MOVL	(BX), AX
	ORL	AX, DX
	LOCK
	CMPXCHGL	DX, (BX)
	JNZ casloop
	MOVL 	AX, ret+16(FP)
	RET

// func And32(addr *uint32, v uint32) old uint32
TEXT ·And32(SB), NOSPLIT, $0-20
	MOVQ	ptr+0(FP), BX
	MOVL	val+8(FP), CX
casloop:
	MOVL 	CX, DX
	MOVL	(BX), AX
	ANDL	AX, DX
	LOCK
	CMPXCHGL	DX, (BX)
	JNZ casloop
	MOVL 	AX, ret+16(FP)
	RET

	TESTQ R15, R15
	JZ    done
	MOVQ  $1, BX
	XORQ  CX, CX
	XORQ  R13, R13
	ADDQ  R15, SI

flush_buffer:
	SHLQ $8, BX, CX
	SHLQ $8, BX
	MOVB -1(SI), R13
	XORQ R13, BX
	DECQ SI
	DECQ R15
	JNZ  flush_buffer

	ADDQ BX, R8
	ADCQ CX, R9
	ADCQ $0, R10
	MOVQ $16, R15
	JMP  multiply

done:
	MOVQ R8, 0(DI)
	MOVQ R9, 8(DI)
	MOVQ R10, 16(DI)

sigtrampnog:
	// Signal arrived on a non-Go thread. If this is SIGPROF, get a
	// stack trace.
	CMPL	DI, $27 // 27 == SIGPROF
	JNZ	sigtramp

	// Lock sigprofCallersUse.
	MOVL	$0, AX
	MOVL	$1, CX
	MOVQ	$runtime·sigprofCallersUse(SB), R11
	LOCK
	CMPXCHGL	CX, 0(R11)
	JNZ	sigtramp  // Skip stack trace if already locked.

	// Jump to the traceback function in runtime/cgo.

sigtrampnog:
	// Signal arrived on a non-Go thread. If this is SIGPROF, get a
	// stack trace.
	CMPL	DI, $27 // 27 == SIGPROF
	JNZ	sigtramp

	// Lock sigprofCallersUse.
	MOVL	$0, AX
	MOVL	$1, CX
	MOVQ	$runtime·sigprofCallersUse(SB), R11
	LOCK
	CMPXCHGL	CX, 0(R11)
	JNZ	sigtramp  // Skip stack trace if already locked.

	// Jump to the traceback function in runtime/cgo.

Lexp_dec:
	// dec
	SUBQ $16, BX
	MOVUPS (BX), X1
	MOVUPS X1, (DX)
	DECQ CX
Lexp_dec_loop:
	MOVUPS -16(BX), X1
	AESIMC X1, X0
	MOVUPS X0, 16(DX)
	SUBQ $16, BX
	ADDQ $16, DX
	DECQ CX
	JNZ Lexp_dec_loop
	MOVUPS -16(BX), X0
	MOVUPS X0, 16(DX)
	RET

TEXT _expand_key_128<>(SB),NOSPLIT,$0
	PSHUFD $0xff, X1, X1
	SHUFPS $0x10, X0, X4
	PXOR X4, X0
	SHUFPS $0x8c, X0, X4
	PXOR X4, X0

	QR(X0, X5, X10, X15, X4)
	MOVOU X15, (15*16)(BX) // save X15
	QR(X1, X6, X11, X12, X4)
	MOVOU (4*16)(BX), X4  // reload X4; temp now X15
	QR(X2, X7, X8, X13, X15)
	QR(X3, X4, X9, X14, X15)

	DECL DX
	JNZ loop

	// Store interlaced blocks back to output buffer,
	// adding original seed along the way.

	// First the top and bottom rows.
	MOVOU X0, (0*16)(BX)
	MOVOU X1, (1*16)(BX)
	MOVOU X2, (2*16)(BX)

sigtrampnog:
	// Signal arrived on a non-Go thread. If this is SIGPROF, get a
	// stack trace.
	CMPL	DI, $27 // 27 == SIGPROF
	JNZ	sigtramp

	// Lock sigprofCallersUse.
	MOVL	$0, AX
	MOVL	$1, CX
	MOVQ	$runtime·sigprofCallersUse(SB), R11
	LOCK
	CMPXCHGL	CX, 0(R11)
	JNZ	sigtramp  // Skip stack trace if already locked.

	// Jump to the traceback function in runtime/cgo.