MOVOU	(SI), X1
	// Compare target byte with each byte in data.
	PCMPEQB	X0, X1
	// Move result bits to integer register.
	PMOVMSKB X1, DX
	// Apply mask
	ANDQ R10, DX
	POPCNTL DX, DX
	// Directly return DX, we don't need to accumulate
	// since we have <16 bytes.
	MOVQ	DX, (R8)
	RET
endzero:
	MOVQ $0, (R8)
	RET

endofpage:
	// We must ignore low bytes as they aren't part of our slice.

	ROUND2(AX, BX, CX, DX, BP, 30)
	ROUND2(BP, AX, BX, CX, DX, 31)
	ROUND2(DX, BP, AX, BX, CX, 32)
	ROUND2(CX, DX, BP, AX, BX, 33)
	ROUND2(BX, CX, DX, BP, AX, 34)
	ROUND2(AX, BX, CX, DX, BP, 35)
	ROUND2(BP, AX, BX, CX, DX, 36)
	ROUND2(DX, BP, AX, BX, CX, 37)
	ROUND2(CX, DX, BP, AX, BX, 38)
	ROUND2(BX, CX, DX, BP, AX, 39)

	ROUND3(AX, BX, CX, DX, BP, 40)
	ROUND3(BP, AX, BX, CX, DX, 41)

	MOVQ	m_g0(CX), DX		// g

	// Layout new m scheduler stack on os stack.
	MOVQ	SP, AX
	MOVQ	AX, (g_stack+stack_hi)(DX)
	SUBQ	$(64*1024), AX		// initial stack size (adjusted later)
	MOVQ	AX, (g_stack+stack_lo)(DX)
	ADDQ	$const_stackGuard, AX
	MOVQ	AX, g_stackguard0(DX)
	MOVQ	AX, g_stackguard1(DX)

	// Set up tls.
	LEAQ	m_tls(CX), DI
	MOVQ	CX, g_m(DX)
	MOVQ	DX, g(DI)

	MOVL	val+4(FP), CX
casloop:
	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

	MOVL	((index-2)*4)(BP), AX; \
	MOVL	AX, CX; \
	RORL	$17, AX; \
	MOVL	CX, DX; \
	RORL	$19, CX; \
	SHRL	$10, DX; \
	MOVL	((index-15)*4)(BP), BX; \
	XORL	CX, AX; \
	MOVL	BX, CX; \
	XORL	DX, AX; \
	RORL	$7, BX; \
	MOVL	CX, DX; \
	SHRL	$3, DX; \
	RORL	$18, CX; \
	ADDL	((index-7)*4)(BP), AX; \
	XORL	CX, BX; \
	XORL	DX, BX; \
	ADDL	((index-16)*4)(BP), BX; \
	ADDL	BX, AX; \
	MOVL	AX, ((index)*4)(BP)

	MOVL	g(CX), BX	// BX = g
	MOVL	g_m(BX), DX	// DX = curm

	// set g to gcrash
	LEAL	runtime·gcrash(SB), BX // g = &gcrash
	MOVL	DX, g_m(BX)            // g.m = curm
	MOVL	BX, m_g0(DX)           // curm.g0 = g
	get_tls(CX)
	MOVL	BX, g(CX)

	// switch to crashstack
	MOVL	(g_stack+stack_hi)(BX), DX
	SUBL	$(4*8), DX
	MOVL	DX, SP

	// call target function
	MOVL	AX, DX
	MOVL	0(AX), AX
	CALL	AX

	MOVL	((index-2)*4)(BP), AX; \
	MOVL	AX, CX; \
	RORL	$17, AX; \
	MOVL	CX, DX; \
	RORL	$19, CX; \
	SHRL	$10, DX; \
	MOVL	((index-15)*4)(BP), BX; \
	XORL	CX, AX; \
	MOVL	BX, CX; \
	XORL	DX, AX; \
	RORL	$7, BX; \
	MOVL	CX, DX; \
	SHRL	$3, DX; \
	RORL	$18, CX; \
	ADDL	((index-7)*4)(BP), AX; \
	XORL	CX, BX; \
	XORL	DX, BX; \
	ADDL	((index-16)*4)(BP), BX; \
	ADDL	BX, AX; \
	MOVL	AX, ((index)*4)(BP)

	//	func f(x uint64) (uint64, uint64) { return x/1000000000, x%1000000000 }
	// adapted to reduce duplication
	MOVQ	AX, CX
	MOVQ	$1360296554856532783, AX
	MULQ	CX
	ADDQ	CX, DX
	RCRQ	$1, DX
	SHRQ	$29, DX
	MOVQ	DX, sec+0(FP)
	IMULQ	$1000000000, DX
	SUBQ	DX, CX
	MOVL	CX, nsec+8(FP)

	SBBL $(delta >> 32), DX

	// nano/100 = DX:AX
	// split into two decimal halves by div 1e9.
	// (decimal point is two spots over from correct place,
	// but we avoid overflow in the high word.)
	MOVL	$1000000000, CX
	DIVL	CX
	MOVL	AX, DI
	MOVL	DX, SI

	// DI = nano/100/1e9 = nano/1e11 = sec/100, DX = SI = nano/100%1e9
	// split DX into seconds and nanoseconds by div 1e7 magic multiply.
	MOVL	DX, AX
	MOVL	$1801439851, CX

	MOVQ	((index-2)*8)(BP), AX; \
	MOVQ	AX, CX; \
	RORQ	$19, AX; \
	MOVQ	CX, DX; \
	RORQ	$61, CX; \
	SHRQ	$6, DX; \
	MOVQ	((index-15)*8)(BP), BX; \
	XORQ	CX, AX; \
	MOVQ	BX, CX; \
	XORQ	DX, AX; \
	RORQ	$1, BX; \
	MOVQ	CX, DX; \
	SHRQ	$7, DX; \
	RORQ	$8, CX; \
	ADDQ	((index-7)*8)(BP), AX; \
	XORQ	CX, BX; \
	XORQ	DX, BX; \
	ADDQ	((index-16)*8)(BP), BX; \
	ADDQ	BX, AX; \
	MOVQ	AX, ((index)*8)(BP)