diff --git a/arch/blackfin/mach-bf561/secondary.S b/arch/blackfin/mach-bf561/secondary.S
index 4c462838f4e1..348c9191d19f 100644
--- a/arch/blackfin/mach-bf561/secondary.S
+++ b/arch/blackfin/mach-bf561/secondary.S
@@ -23,74 +23,56 @@
 #define INITIAL_STACK	(COREB_L1_SCRATCH_START + L1_SCRATCH_LENGTH - 12)
 
 ENTRY(_coreb_trampoline_start)
-	/* Set the SYSCFG register */
-	R0 = 0x36;
-	SYSCFG = R0; /*Enable Cycle Counter and Nesting Of Interrupts(3rd Bit)*/
-	R0 = 0;
+	/* Enable Cycle Counter and Nesting Of Interrupts */
+#ifdef CONFIG_BFIN_SCRATCH_REG_CYCLES
+	R0 = SYSCFG_SNEN;
+#else
+	R0 = SYSCFG_SNEN | SYSCFG_CCEN;
+#endif
+	SYSCFG = R0;
 
-	/*Clear Out All the data and pointer  Registers*/
-	R1 = R0;
-	R2 = R0;
-	R3 = R0;
-	R4 = R0;
-	R5 = R0;
-	R6 = R0;
-	R7 = R0;
+	/* Optimization register tricks: keep a base value in the
+	 * reserved P registers so we use the load/store with an
+	 * offset syntax.  R0 = [P5 + <constant>];
+	 *   P5 - core MMR base
+	 *   R6 - 0
+	 */
+	r6 = 0;
+	p5.l = 0;
+	p5.h = hi(COREMMR_BASE);
 
-	P0 = R0;
-	P1 = R0;
-	P2 = R0;
-	P3 = R0;
-	P4 = R0;
-	P5 = R0;
+	/* Zero out registers required by Blackfin ABI */
 
-	LC0 = r0;
-	LC1 = r0;
-	L0 = r0;
-	L1 = r0;
-	L2 = r0;
-	L3 = r0;
+	/* Disable circular buffers */
+	L0 = r6;
+	L1 = r6;
+	L2 = r6;
+	L3 = r6;
 
-	/* Clear Out All the DAG Registers*/
-	B0 = r0;
-	B1 = r0;
-	B2 = r0;
-	B3 = r0;
+	/* Disable hardware loops in case we were started by 'go' */
+	LC0 = r6;
+	LC1 = r6;
 
-	I0 = r0;
-	I1 = r0;
-	I2 = r0;
-	I3 = r0;
-
-	M0 = r0;
-	M1 = r0;
-	M2 = r0;
-	M3 = r0;
+	/*
+	 * Clear ITEST_COMMAND and DTEST_COMMAND registers,
+	 * Leaving these as non-zero can confuse the emulator
+	 */
+	[p5 + (DTEST_COMMAND - COREMMR_BASE)] = r6;
+	[p5 + (ITEST_COMMAND - COREMMR_BASE)] = r6;
+	CSYNC;
 
 	trace_buffer_init(p0,r0);
 
 	/* Turn off the icache */
-	p0.l = LO(IMEM_CONTROL);
-	p0.h = HI(IMEM_CONTROL);
-	R1 = [p0];
-	R0 = ~ENICPLB;
-	R0 = R0 & R1;
-
-	/* Disabling of CPLBs should be proceeded by a CSYNC */
-	CSYNC;
-	[p0] = R0;
+	r1 = [p5 + (IMEM_CONTROL - COREMMR_BASE)];
+	BITCLR (r1, ENICPLB_P);
+	[p5 + (IMEM_CONTROL - COREMMR_BASE)] = r1;
 	SSYNC;
 
 	/* Turn off the dcache */
-	p0.l = LO(DMEM_CONTROL);
-	p0.h = HI(DMEM_CONTROL);
-	R1 = [p0];
-	R0 = ~ENDCPLB;
-	R0 = R0 & R1;
-
-	/* Disabling of CPLBs should be proceeded by a CSYNC */
-	CSYNC;
-	[p0] = R0;
+	r1 = [p5 + (DMEM_CONTROL - COREMMR_BASE)];
+	BITCLR (r1, ENDCPLB_P);
+	[p5 + (DMEM_CONTROL - COREMMR_BASE)] = r1;
 	SSYNC;
 
 	/* in case of double faults, save a few things */
@@ -105,25 +87,25 @@ ENTRY(_coreb_trampoline_start)
 	 * below
 	 */
 	GET_PDA(p0, r0);
-	r7 = [p0 + PDA_DF_RETX];
+	r5 = [p0 + PDA_DF_RETX];
 	p1.l = _init_saved_retx_coreb;
 	p1.h = _init_saved_retx_coreb;
-	[p1] = r7;
+	[p1] = r5;
 
-	r7 = [p0 + PDA_DF_DCPLB];
+	r5 = [p0 + PDA_DF_DCPLB];
 	p1.l = _init_saved_dcplb_fault_addr_coreb;
 	p1.h = _init_saved_dcplb_fault_addr_coreb;
-	[p1] = r7;
+	[p1] = r5;
 
-	r7 = [p0 + PDA_DF_ICPLB];
+	r5 = [p0 + PDA_DF_ICPLB];
 	p1.l = _init_saved_icplb_fault_addr_coreb;
 	p1.h = _init_saved_icplb_fault_addr_coreb;
-	[p1] = r7;
+	[p1] = r5;
 
-	r7 = [p0 + PDA_DF_SEQSTAT];
+	r5 = [p0 + PDA_DF_SEQSTAT];
 	p1.l = _init_saved_seqstat_coreb;
 	p1.h = _init_saved_seqstat_coreb;
-	[p1] = r7;
+	[p1] = r5;
 #endif
 
 	/* Initialize stack pointer */
@@ -138,19 +120,13 @@ ENTRY(_coreb_trampoline_start)
 
 	/* EVT15 = _real_start */
 
-	p0.l = lo(EVT15);
-	p0.h = hi(EVT15);
 	p1.l = _coreb_start;
 	p1.h = _coreb_start;
-	[p0] = p1;
+	[p5 + (EVT15 - COREMMR_BASE)] = p1;
 	csync;
 
-	p0.l = lo(IMASK);
-	p0.h = hi(IMASK);
-	p1.l = IMASK_IVG15;
-	p1.h = 0x0;
-	[p0] = p1;
-	csync;
+	r0 = EVT_IVG15 (z);
+	sti r0;
 
 	raise 15;
 	p0.l = .LWAIT_HERE;