Change alignment macros to allow further performance optimization. Define the CACHEALIGN macros for all ARM CPUs, the used alignment size is derived from CACHEALIGN_BITS which has been defined for each supported ARM CPU with r28619. The default alignment size for ARM is set to 32 bytes as new -- not yet supported -- ARM CPUs will most probably need this alignment. To be able to differ between ARM and other CPUs a new macro called MEM_ALIGN_ATTR is introduced. This equals CACHEALIGN_ATTR for ARM, 16 byte alignment for Coldfire and is kept empty for other CPUs. MEM_ALIGN_ATTR is available system wide. From measurements it is expected that the usage of MEM_ALIGN_ATTR can give significant performance gain on ARM11 CPUs.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@28625 a1c6a512-1295-4272-9138-f99709370657

2 files changed, 65 insertions, 50 deletions

diff --git a/firmware/export/system.h b/firmware/export/system.h
index ed10c84a58..af70152299 100644
--- a/firmware/export/system.h
+++ b/firmware/export/system.h
@@ -309,58 +309,77 @@ static inline void cpucache_flush(void)
 }
 #endif
 
-#ifndef CACHEALIGN_SIZE /* could be elsewhere for a particular reason */
+/* Define this, if the CPU may take advantage of cache aligment. Is enabled
+ * for all ARM CPUs. */
+#ifdef CPU_ARM
+    #define HAVE_CPU_CACHE_ALIGN
+#endif
+
+/* Calculate CACHEALIGN_SIZE from CACHEALIGN_BITS */
+#ifdef CACHEALIGN_SIZE
+    /* undefine, if defined. always calculate from CACHEALIGN_BITS */
+    #undef CACHEALIGN_SIZE
+#endif
 #ifdef CACHEALIGN_BITS
-/* 2^CACHEALIGN_BITS = the byte size */
-#define CACHEALIGN_SIZE (1u << CACHEALIGN_BITS)
+    /* CACHEALIGN_BITS = 2 ^ CACHEALIGN_BITS */
+    #define CACHEALIGN_SIZE (1u << CACHEALIGN_BITS)
+#else
+    /* FIXME: set to maximum known cache alignment of supported CPUs */
+    #define CACHEALIGN_BITS  5
+    #define CACHEALIGN_SIZE 32
+#endif
+
+#ifdef HAVE_CPU_CACHE_ALIGN
+    /* Cache alignment attributes and sizes are enabled */
+    #define CACHEALIGN_ATTR __attribute__((aligned(CACHEALIGN_SIZE)))
+    /* Aligns x up to a CACHEALIGN_SIZE boundary */
+    #define CACHEALIGN_UP(x) \
+        ((typeof (x))ALIGN_UP_P2((uintptr_t)(x), CACHEALIGN_BITS))
+    /* Aligns x down to a CACHEALIGN_SIZE boundary */
+    #define CACHEALIGN_DOWN(x) \
+        ((typeof (x))ALIGN_DOWN_P2((uintptr_t)(x), CACHEALIGN_BITS))
+    /* Aligns at least to the greater of size x or CACHEALIGN_SIZE */
+    #define CACHEALIGN_AT_LEAST_ATTR(x) \
+        __attribute__((aligned(CACHEALIGN_UP(x))))
+    /* Aligns a buffer pointer and size to proper boundaries */
+    #define CACHEALIGN_BUFFER(start, size) \
+        ALIGN_BUFFER((start), (size), CACHEALIGN_SIZE)
+#else
+    /* Cache alignment attributes and sizes are not enabled */
+    #define CACHEALIGN_ATTR
+    #define CACHEALIGN_AT_LEAST_ATTR(x) __attribute__((aligned(x)))
+    #define CACHEALIGN_UP(x) (x)
+    #define CACHEALIGN_DOWN(x) (x)
+    /* Make no adjustments */
+    #define CACHEALIGN_BUFFER(start, size)
+#endif
+
+/* Define MEM_ALIGN_ATTR which may be used to align e.g. buffers for faster
+ * access. */
+#if   defined(CPU_ARM)
+    /* Use ARMs cache alignment. */
+    #define MEM_ALIGN_ATTR CACHEALIGN_ATTR
+#elif defined(CPU_COLDFIRE)
+    /* Use fixed alignment of 16 bytes. Speed up only for 'movem' in DRAM. */
+    #define MEM_ALIGN_ATTR __attribute__((aligned(16)))
 #else
-#define CACHEALIGN_SIZE 16  /* FIXME */
+    /* Do nothing. */
+    #define MEM_ALIGN_ATTR
 #endif
-#endif /* CACHEALIGN_SIZE */
-
-#ifdef PROC_NEEDS_CACHEALIGN
-/* Cache alignment attributes and sizes are enabled */
-
-#define CACHEALIGN_ATTR __attribute__((aligned(CACHEALIGN_SIZE)))
-/* Aligns x up to a CACHEALIGN_SIZE boundary */
-#define CACHEALIGN_UP(x) \
-    ((typeof (x))ALIGN_UP_P2((uintptr_t)(x), CACHEALIGN_BITS))
-/* Aligns x down to a CACHEALIGN_SIZE boundary */
-#define CACHEALIGN_DOWN(x) \
-    ((typeof (x))ALIGN_DOWN_P2((uintptr_t)(x), CACHEALIGN_BITS))
-/* Aligns at least to the greater of size x or CACHEALIGN_SIZE */
-#define CACHEALIGN_AT_LEAST_ATTR(x) \
-    __attribute__((aligned(CACHEALIGN_UP(x))))
-/* Aligns a buffer pointer and size to proper boundaries */
-#define CACHEALIGN_BUFFER(start, size) \
-    ALIGN_BUFFER((start), (size), CACHEALIGN_SIZE)
-
-#else /* ndef PROC_NEEDS_CACHEALIGN */
-
-/* Cache alignment attributes and sizes are not enabled */
-#define CACHEALIGN_ATTR
-#define CACHEALIGN_AT_LEAST_ATTR(x) \
-    __attribute__((aligned(x)))
-#define CACHEALIGN_UP(x) (x)
-#define CACHEALIGN_DOWN(x) (x)
-/* Make no adjustments */
-#define CACHEALIGN_BUFFER(start, size)
-
-#endif /* PROC_NEEDS_CACHEALIGN */
 
 #ifdef STORAGE_WANTS_ALIGN
-#define STORAGE_ALIGN_ATTR __attribute__((aligned(CACHEALIGN_SIZE)))
-#define STORAGE_ALIGN_DOWN(x) \
-    ((typeof (x))ALIGN_DOWN_P2((uintptr_t)(x), CACHEALIGN_BITS))
-/* Pad a size so the buffer can be aligned later */
-#define STORAGE_PAD(x) ((x) + CACHEALIGN_SIZE - 1)
-/* Number of bytes in the last cacheline assuming buffer of size x is aligned */
-#define STORAGE_OVERLAP(x) ((x) & (CACHEALIGN_SIZE - 1))
+    #define STORAGE_ALIGN_ATTR __attribute__((aligned(CACHEALIGN_SIZE)))
+    #define STORAGE_ALIGN_DOWN(x) \
+        ((typeof (x))ALIGN_DOWN_P2((uintptr_t)(x), CACHEALIGN_BITS))
+    /* Pad a size so the buffer can be aligned later */
+    #define STORAGE_PAD(x) ((x) + CACHEALIGN_SIZE - 1)
+    /* Number of bytes in the last cacheline assuming buffer of size x is aligned */
+    #define STORAGE_OVERLAP(x) ((x) & (CACHEALIGN_SIZE - 1))
 #else
-#define STORAGE_ALIGN_ATTR
-#define STORAGE_ALIGN_DOWN(x) (x)
-#define STORAGE_PAD(x) (x)
-#define STORAGE_OVERLAP(x) 0
+    #define STORAGE_ALIGN_ATTR
+    #define STORAGE_ALIGN_DOWN(x) (x)
+    #define STORAGE_PAD(x) (x)
+    #define STORAGE_OVERLAP(x) 0
 #endif
 
 /* Double-cast to avoid 'dereferencing type-punned pointer will
diff --git a/firmware/target/arm/system-target.h b/firmware/target/arm/system-target.h
index 1e573be291..3e3eab8f8d 100644
--- a/firmware/target/arm/system-target.h
+++ b/firmware/target/arm/system-target.h
@@ -158,10 +158,6 @@ static inline void wake_core(int core)
     ((typeof (a))((uintptr_t)(a) | UNCACHED_BASE_ADDR))
 #endif /* BOOTLOADER */
 
-/* Certain data needs to be out of the way of cache line interference
- * such as data for COP use or for use with UNCACHED_ADDR */
-#define PROC_NEEDS_CACHEALIGN
-
 #if defined(CPU_PP502x) && defined(HAVE_ATA_DMA)
 #define STORAGE_WANTS_ALIGN
 #endif