diff --git a/include/common/aarch32/asm_macros.S b/include/common/aarch32/asm_macros.S
index 3adcbf5..f573744 100644
--- a/include/common/aarch32/asm_macros.S
+++ b/include/common/aarch32/asm_macros.S
@@ -80,6 +80,19 @@
 	.endm
 
 	/*
+	 * Helper macro to generate the best mov/movw/movt combinations
+	 * according to the value to be moved.
+	 */
+	.macro mov_imm _reg, _val
+		.if ((\_val) & 0xffff0000) == 0
+			mov	\_reg, #(\_val)
+		.else
+			movw	\_reg, #((\_val) & 0xffff)
+			movt	\_reg, #((\_val) >> 16)
+		.endif
+	.endm
+
+	/*
 	 * Macro to mark instances where we're jumping to a function and don't
 	 * expect a return. To provide the function being jumped to with
 	 * additional information, we use 'bl' instruction to jump rather than
diff --git a/include/lib/el3_runtime/cpu_data.h b/include/lib/el3_runtime/cpu_data.h
index 1e8bfa7..c0c3a19 100644
--- a/include/lib/el3_runtime/cpu_data.h
+++ b/include/lib/el3_runtime/cpu_data.h
@@ -7,12 +7,15 @@
 #ifndef __CPU_DATA_H__
 #define __CPU_DATA_H__
 
+#include <platform_def.h>	/* CACHE_WRITEBACK_GRANULE required */
+
 #ifdef AARCH32
 
 #if CRASH_REPORTING
 #error "Crash reporting is not supported in AArch32"
 #endif
 #define CPU_DATA_CPU_OPS_PTR		0x0
+#define CPU_DATA_CRASH_BUF_OFFSET	0x4
 
 #else /* AARCH32 */
 
@@ -25,14 +28,18 @@
 #endif /* AARCH32 */
 
 #if CRASH_REPORTING
-#define CPU_DATA_LOG2SIZE		7
 #define CPU_DATA_CRASH_BUF_END		(CPU_DATA_CRASH_BUF_OFFSET + \
 						CPU_DATA_CRASH_BUF_SIZE)
 #else
-#define CPU_DATA_LOG2SIZE		6
 #define CPU_DATA_CRASH_BUF_END		CPU_DATA_CRASH_BUF_OFFSET
 #endif
 
+/* cpu_data size is the data size rounded up to the platform cache line size */
+#define CPU_DATA_SIZE			(((CPU_DATA_CRASH_BUF_END + \
+					CACHE_WRITEBACK_GRANULE - 1) / \
+						CACHE_WRITEBACK_GRANULE) * \
+							CACHE_WRITEBACK_GRANULE)
+
 #if ENABLE_RUNTIME_INSTRUMENTATION
 /* Temporary space to store PMF timestamps from assembly code */
 #define CPU_DATA_PMF_TS_COUNT		1
@@ -98,8 +105,8 @@
 	assert_cpu_data_crash_stack_offset_mismatch);
 #endif
 
-CASSERT((1 << CPU_DATA_LOG2SIZE) == sizeof(cpu_data_t),
-	assert_cpu_data_log2size_mismatch);
+CASSERT(CPU_DATA_SIZE == sizeof(cpu_data_t),
+		assert_cpu_data_size_mismatch);
 
 CASSERT(CPU_DATA_CPU_OPS_PTR == __builtin_offsetof
 		(cpu_data_t, cpu_ops_ptr),
diff --git a/lib/el3_runtime/aarch32/cpu_data.S b/lib/el3_runtime/aarch32/cpu_data.S
index 3d6b806..68d6415 100644
--- a/lib/el3_runtime/aarch32/cpu_data.S
+++ b/lib/el3_runtime/aarch32/cpu_data.S
@@ -34,7 +34,9 @@
  * -----------------------------------------------------------------
  */
 func _cpu_data_by_index
+	mov_imm	r1, CPU_DATA_SIZE
+	mul	r0, r0, r1
 	ldr	r1, =percpu_data
-	add	r0, r1, r0, LSL #CPU_DATA_LOG2SIZE
+	add	r0, r0, r1
 	bx	lr
 endfunc _cpu_data_by_index
diff --git a/lib/el3_runtime/aarch64/cpu_data.S b/lib/el3_runtime/aarch64/cpu_data.S
index de48816..96be081 100644
--- a/lib/el3_runtime/aarch64/cpu_data.S
+++ b/lib/el3_runtime/aarch64/cpu_data.S
@@ -39,7 +39,9 @@
  * -----------------------------------------------------------------
  */
 func _cpu_data_by_index
+	mov_imm	x1, CPU_DATA_SIZE
+	mul	x0, x0, x1
 	adr	x1, percpu_data
-	add	x0, x1, x0, LSL #CPU_DATA_LOG2SIZE
+	add	x0, x0, x1
 	ret
 endfunc _cpu_data_by_index