Improve libdemac SATURATE slightly on ARMv4/5, move filter buffers and code out of IRAM for sizes that aren't near realtime and extend udiv32_arm reciprocal table.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@24376 a1c6a512-1295-4272-9138-f99709370657

6 files changed, 55 insertions, 11 deletions

diff --git a/apps/codecs/demac/libdemac/decoder.c b/apps/codecs/demac/libdemac/decoder.c
index 0763c11037..09563e0112 100644
--- a/apps/codecs/demac/libdemac/decoder.c
+++ b/apps/codecs/demac/libdemac/decoder.c
@@ -33,10 +33,23 @@ Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110, USA
 
 /* Statically allocate the filter buffers */
 
+#ifdef FILTER256_IRAM
 static filter_int filterbuf32[(32*3 + FILTER_HISTORY_SIZE) * 2]   
                   IBSS_ATTR __attribute__((aligned(16))); /* 2432/4864 bytes */
 static filter_int filterbuf256[(256*3 + FILTER_HISTORY_SIZE) * 2]
                   IBSS_ATTR __attribute__((aligned(16))); /* 5120/10240 bytes */
+#define FILTERBUF64 filterbuf256
+#define FILTERBUF32 filterbuf32
+#define FILTERBUF16 filterbuf32
+#else
+static filter_int filterbuf64[(64*3 + FILTER_HISTORY_SIZE) * 2]   
+                  IBSS_ATTR __attribute__((aligned(16))); /* 2432/4864 bytes */
+static filter_int filterbuf256[(256*3 + FILTER_HISTORY_SIZE) * 2]
+                  __attribute__((aligned(16))); /* 5120/10240 bytes */
+#define FILTERBUF64 filterbuf64
+#define FILTERBUF32 filterbuf64
+#define FILTERBUF16 filterbuf64
+#endif
 
 /* This is only needed for "insane" files, and no current Rockbox targets
    can hope to decode them in realtime, although the Gigabeat S comes close. */
@@ -57,22 +70,22 @@ void init_frame_decoder(struct ape_ctx_t* ape_ctx,
     switch (ape_ctx->compressiontype)
     {
         case 2000:
-            init_filter_16_11(filterbuf32);
+            init_filter_16_11(FILTERBUF16);
             break;
 
         case 3000:
-            init_filter_64_11(filterbuf256);
+            init_filter_64_11(FILTERBUF64);
             break;
 
         case 4000:
             init_filter_256_13(filterbuf256);
-            init_filter_32_10(filterbuf32);
+            init_filter_32_10(FILTERBUF32);
             break;
 
         case 5000:
             init_filter_1280_15(filterbuf1280);
             init_filter_256_13(filterbuf256);
-            init_filter_16_11(filterbuf32);
+            init_filter_16_11(FILTERBUF32);
     }
 }
 
diff --git a/apps/codecs/demac/libdemac/demac_config.h b/apps/codecs/demac/libdemac/demac_config.h
index 1beda2b9cd..13166f69ae 100644
--- a/apps/codecs/demac/libdemac/demac_config.h
+++ b/apps/codecs/demac/libdemac/demac_config.h
@@ -44,6 +44,10 @@ Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110, USA
 #define FILTER_BITS 32
 #endif
 
+#if !defined(CPU_PP) && !defined(CPU_S5L870X)
+#define FILTER256_IRAM
+#endif
+
 #if CONFIG_CPU == PP5002
 /* Code in IRAM for speed, not enough IRAM for the insane filter buffer. */
 #define ICODE_SECTION_DEMAC_ARM   .icode
diff --git a/apps/codecs/demac/libdemac/filter.c b/apps/codecs/demac/libdemac/filter.c
index bab830a8bd..93edf39cb2 100644
--- a/apps/codecs/demac/libdemac/filter.c
+++ b/apps/codecs/demac/libdemac/filter.c
@@ -97,9 +97,26 @@ struct filter_t {
 #define FP_HALF  (1 << (FRACBITS - 1))   /* 0.5 in fixed-point format. */
 #define FP_TO_INT(x) ((x + FP_HALF) >> FRACBITS)  /* round(x) */
 
-#if defined(CPU_ARM) && (ARM_ARCH >= 6)
+#ifdef CPU_ARM
+#if ARM_ARCH >= 6
 #define SATURATE(x) ({int __res; asm("ssat %0, #16, %1" : "=r"(__res) : "r"(x)); __res; })
-#else
+#else /* ARM_ARCH < 6 */
+/* Keeping the asr #31 outside of the asm allows loads to be scheduled between
+   it and the rest of the block on ARM9E, with the load's result latency filled
+   by the other calculations. */
+#define SATURATE(x) ({ \
+    int __res = (x) >> 31; \
+    asm volatile ( \
+        "teq %0, %1, asr #15\n\t" \
+        "moveq %0, %1\n\t" \
+        "eorne %0, %0, #0xff\n\t" \
+        "eorne %0, %0, #0x7f00" \
+        : "+r" (__res) : "r" (x) : "cc" \
+    ); \
+    __res; \
+})
+#endif /* ARM_ARCH */
+#else /* CPU_ARM */
 #define SATURATE(x) (LIKELY((x) == (int16_t)(x)) ? (x) : ((x) >> 31) ^ 0x7FFF)
 #endif
 
diff --git a/apps/codecs/demac/libdemac/filter_1280_15.c b/apps/codecs/demac/libdemac/filter_1280_15.c
index 7077e0ee8e..f2301fb02a 100644
--- a/apps/codecs/demac/libdemac/filter_1280_15.c
+++ b/apps/codecs/demac/libdemac/filter_1280_15.c
@@ -22,6 +22,11 @@ Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110, USA
 
 */
 
+#include "demac_config.h"
+#ifndef FILTER256_IRAM
+#undef ICODE_ATTR_DEMAC
+#define ICODE_ATTR_DEMAC
+#endif
 #define ORDER 1280
 #define FRACBITS 15
 #include "filter.c"
diff --git a/apps/codecs/demac/libdemac/filter_256_13.c b/apps/codecs/demac/libdemac/filter_256_13.c
index 69cf638903..9e4b9fcb13 100644
--- a/apps/codecs/demac/libdemac/filter_256_13.c
+++ b/apps/codecs/demac/libdemac/filter_256_13.c
@@ -22,6 +22,11 @@ Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110, USA
 
 */
 
+#include "demac_config.h"
+#ifndef FILTER256_IRAM
+#undef ICODE_ATTR_DEMAC
+#define ICODE_ATTR_DEMAC
+#endif
 #define ORDER 256
 #define FRACBITS 13
 #include "filter.c"
diff --git a/apps/codecs/lib/udiv32_arm.S b/apps/codecs/lib/udiv32_arm.S
index 33ab7a43a5..8efc92c2e6 100644
--- a/apps/codecs/lib/udiv32_arm.S
+++ b/apps/codecs/lib/udiv32_arm.S
@@ -90,18 +90,18 @@
 
 #ifdef CPU_PP
 #if CONFIG_CPU == PP5020
-.set recip_max, 5952
+.set recip_max, 8384
 #elif CONFIG_CPU == PP5002
-.set recip_max, 1472
+.set recip_max, 4992
 #else
-.set recip_max, 14208
+.set recip_max, 16384
 #endif
 #elif CONFIG_CPU == AS3525
 .set recip_max, 42752
 #elif CONFIG_CPU == S5L8701
-.set recip_max, 9600
+.set recip_max, 13184
 #elif CONFIG_CPU == S5L8700
-.set recip_max, 5504
+.set recip_max, 9088
 #endif
 
 udiv32_arm: