bsps: Rework cache manager implementation

The previous cache manager support used a single souce file
(cache_manager.c) which included an implementation header (cache_.h).
This required the use of specialized include paths to find the right
header file.  Change this to include a generic implementation header
(cacheimpl.h) in specialized source files.

Use the following directories and files:

* bsps/shared/cache

* bsps/@RTEMS_CPU@/shared/cache

* bsps/@RTEMS_CPU@/@RTEMS_BSP_FAMILY/start/cache.c

Update #3285.

bsps/bfin/shared/cache/cache.c

@@ -0,0 +1,134 @@
+/*  Blackfin Cache Support
+ *
+ *  Copyright (c) 2008 Kallisti Labs, Los Gatos, CA, USA
+ *             written by Allan Hessenflow <allanh@kallisti.com>
+ *
+ *  The license and distribution terms for this file may be
+ *  found in the file LICENSE in this distribution or at
+ *  http://www.rtems.org/license/LICENSE.
+ */
+
+
+#include <rtems.h>
+#include <bsp.h>
+#include <libcpu/memoryRegs.h>
+
+#define CPU_DATA_CACHE_ALIGNMENT          32
+#define CPU_INSTRUCTION_CACHE_ALIGNMENT   32
+
+#ifdef BSP_DATA_CACHE_CONFIG
+#define LIBCPU_DATA_CACHE_CONFIG BSP_DATA_CACHE_CONFIG
+#else
+/* use 16K of each SRAM bank */
+#define LIBCPU_DATA_CACHE_CONFIG (3 << DMEM_CONTROL_DMC_SHIFT)
+#endif
+
+/* There are many syncs in the following code because they should be
+   harmless except for wasting time, and this is easier than figuring out
+   exactly where they're needed to protect from the effects of write
+   buffers and queued reads.  Many of them are likely unnecessary. */
+
+
+static void _CPU_cache_flush_1_data_line(const void *d_addr) {
+
+  __asm__ __volatile__ ("ssync; flush [%0]; ssync" :: "a" (d_addr));
+}
+
+/* Blackfins can't just invalidate cache; they can only do flush +
+   invalidate.  If the line isn't dirty then this is equivalent to
+   just an invalidate.  Even if it is dirty, this should still be
+   okay since with a pure invalidate method the caller would have no
+   way to insure the dirty line hadn't been written out anyway prior
+   to the invalidate. */
+static void _CPU_cache_invalidate_1_data_line(const void *d_addr) {
+
+  __asm__ __volatile__ ("ssync; flushinv [%0]; ssync" :: "a" (d_addr));
+}
+
+static void _CPU_cache_freeze_data(void) {
+}
+
+static void _CPU_cache_unfreeze_data(void) {
+}
+
+static void _CPU_cache_invalidate_1_instruction_line(const void *d_addr) {
+
+  __asm__ __volatile__ ("ssync; iflush [%0]; ssync" :: "a" (d_addr));
+}
+
+static void _CPU_cache_freeze_instruction(void) {
+}
+
+static void _CPU_cache_unfreeze_instruction(void) {
+}
+
+/* incredibly inefficient...  It would be better to make use of the
+   DTEST_COMMAND/DTEST_DATAx registers to find the addresses in each
+   cache line and flush just those.  However the documentation I've
+   seen on those is a bit sketchy, and I sure wouldn't want to get it
+   wrong. */
+static void _CPU_cache_flush_entire_data(void) {
+  uint32_t i;
+
+  i = 0;
+  __asm__ __volatile__ ("ssync");
+  do {
+      __asm__ __volatile__ ("flush [%0]" :: "a" (i));
+      i += CPU_DATA_CACHE_ALIGNMENT;
+  } while (i);
+  __asm__ __volatile__ ("ssync");
+}
+
+static void _CPU_cache_invalidate_entire_data(void) {
+  uint32_t dmemControl;
+
+  __asm__ __volatile__ ("ssync");
+  dmemControl = *(uint32_t volatile *) DMEM_CONTROL;
+  *(uint32_t volatile *) DMEM_CONTROL = dmemControl & ~DMEM_CONTROL_DMC_MASK;
+  *(uint32_t volatile *) DMEM_CONTROL = dmemControl;
+  __asm__ __volatile__ ("ssync");
+}
+
+/* this does not actually enable data cache unless CPLBs are also enabled.
+   LIBCPU_DATA_CACHE_CONFIG contains the DMEM_CONTROL_DMC bits to set. */
+static void _CPU_cache_enable_data(void) {
+
+  __asm__ __volatile__ ("ssync");
+  *(uint32_t volatile *) DMEM_CONTROL |= LIBCPU_DATA_CACHE_CONFIG;
+  __asm__ __volatile__ ("ssync");
+}
+
+static void _CPU_cache_disable_data(void) {
+
+  __asm__ __volatile__ ("ssync");
+  *(uint32_t volatile *) DMEM_CONTROL &= ~DMEM_CONTROL_DMC_MASK;
+  __asm__ __volatile__ ("ssync");
+}
+
+static void _CPU_cache_invalidate_entire_instruction(void) {
+  uint32_t imemControl;
+
+  __asm__ __volatile__ ("ssync");
+  imemControl = *(uint32_t volatile *) IMEM_CONTROL;
+  *(uint32_t volatile *) IMEM_CONTROL = imemControl & ~IMEM_CONTROL_IMC;
+  *(uint32_t volatile *) IMEM_CONTROL = imemControl;
+  __asm__ __volatile__ ("ssync");
+}
+
+/* this only actually enables the instruction cache if the CPLBs are also
+   enabled. */
+static void _CPU_cache_enable_instruction(void) {
+
+  __asm__ __volatile__ ("ssync");
+  *(uint32_t volatile *) IMEM_CONTROL |= IMEM_CONTROL_IMC;
+  __asm__ __volatile__ ("ssync");
+}
+
+static void _CPU_cache_disable_instruction(void) {
+
+  __asm__ __volatile__ ("ssync");
+  *(uint32_t volatile *) IMEM_CONTROL &= ~IMEM_CONTROL_IMC;
+  __asm__ __volatile__ ("ssync");
+}
+
+#include "../../../shared/cache/cacheimpl.h"