Blackfin arch: Zero out bss region in L1/L2 memory.

	w[p0] = r0;

	ssync;

#if L1_DATA_A_LENGTH > 0

	r1.l = __sbss_l1;

	r1.h = __sbss_l1;

	r2.l = __ebss_l1;

	r2.h = __ebss_l1;

	r0 = 0 (z);

	r2 = r2 - r1;

	cc = r2 == 0;

	if cc jump .L_a_l1_done;

	r2 >>= 2;

	p1 = r1;

	p2 = r2;

	lsetup (.L_clear_a_l1, .L_clear_a_l1 ) lc0 = p2;

.L_clear_a_l1:

	[p1++] = r0;

.L_a_l1_done:

#endif

#if L1_DATA_B_LENGTH > 0

	r1.l = __sbss_b_l1;

	r1.h = __sbss_b_l1;

	r2.l = __ebss_b_l1;

	r2.h = __ebss_b_l1;

	r0 = 0 (z);

	r2 = r2 - r1;

	cc = r2 == 0;

	if cc jump .L_b_l1_done;

	r2 >>= 2;

	p1 = r1;

	p2 = r2;

	lsetup (.L_clear_b_l1, .L_clear_b_l1 ) lc0 = p2;

.L_clear_b_l1:

	[p1++] = r0;

.L_b_l1_done:

#endif

#if L2_LENGTH > 0

	r1.l = __sbss_l2;

	r1.h = __sbss_l2;

	r2.l = __ebss_l2;

	r2.h = __ebss_l2;

	r0 = 0 (z);

	r2 = r2 - r1;

	cc = r2 == 0;

	if cc jump .L_l2_done;

	r2 >>= 2;

	p1 = r1;

	p2 = r2;

	lsetup (.L_clear_l2, .L_clear_l2 ) lc0 = p2;

.L_clear_l2:

	[p1++] = r0;

.L_l2_done:

#endif

	/* Zero out the bss region

	 * Note: this will fail if bss is 0 bytes ...

	 */