FS#10048 : enable MMU and data cache on Sansa AMS to give a major speed up

- cache IRAM and DRAM
- map IRAM just next to DRAM to remove the need for -mlong-calls and reduce binsize
- tweak delays in Fuze button code
- tweak delays in Clip button code (down button sometimes doesn't respond anyway : an alternate driver is being worked on)

Before reporting any problem, please check your filesystem or format your player from the OF

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@21228 a1c6a512-1295-4272-9138-f99709370657

9 files changed, 67 insertions, 44 deletions

diff --git a/firmware/target/arm/as3525/app.lds b/firmware/target/arm/as3525/app.lds
index 45ae30f9f3..eb5b28f852 100644
--- a/firmware/target/arm/as3525/app.lds
+++ b/firmware/target/arm/as3525/app.lds
@@ -55,6 +55,12 @@ SECTIONS
 {
     loadaddress = DRAM_ORIG;
 
+    .vectors :
+    {
+        _vectors_start = .;
+        *(.init.text)
+    } > DRAM
+
     .text :
     {
         _loadaddress = .;
@@ -83,14 +89,6 @@ SECTIONS
         *(.eh_frame)
     }
 
-    .vectors IRAMORIG:
-    {
-        _vectors_start = .;
-        *(.init.text)
-    } > IRAM AT > DRAM
-
-    _vectorscopy = LOADADDR(.vectors);
-
     .iram :
     {
         _iramstart = .;
diff --git a/firmware/target/arm/as3525/ata_sd_as3525.c b/firmware/target/arm/as3525/ata_sd_as3525.c
index b3ab31b293..91d65235ef 100644
--- a/firmware/target/arm/as3525/ata_sd_as3525.c
+++ b/firmware/target/arm/as3525/ata_sd_as3525.c
@@ -595,7 +595,7 @@ static int sd_select_bank(signed char bank)
     dma_enable_channel(0, card_data, MCI_FIFO(INTERNAL_AS3525), DMA_PERI_SD,
         DMAC_FLOWCTRL_PERI_MEM_TO_PERI, true, false, 0, DMA_S8, NULL);
 
-    MCI_DATA_TIMER(INTERNAL_AS3525) = 0x1000000; /* FIXME: arbitrary */
+    MCI_DATA_TIMER(INTERNAL_AS3525) = 0xffffffff;/* FIXME: arbitrary */
     MCI_DATA_LENGTH(INTERNAL_AS3525) = 512;
     MCI_DATA_CTRL(INTERNAL_AS3525) =  (1<<0) /* enable */   |
                             (0<<1) /* transfer direction */ |
@@ -618,7 +618,8 @@ static int sd_select_bank(signed char bank)
 }
 
 #define UNALIGNED_NUM_SECTORS 10
-static int32_t aligned_buffer[UNALIGNED_NUM_SECTORS* (SECTOR_SIZE / 4)];
+static unsigned char aligned_buffer[UNALIGNED_NUM_SECTORS* SECTOR_SIZE] __attribute__((aligned(32)));   /* align on cache line size */
+static unsigned char *uncached_buffer = UNCACHED_ADDR(aligned_buffer);
 
 static int sd_transfer_sectors(IF_MV2(int drive,) unsigned long start,
                                int count, void* buf, const bool write)
@@ -627,7 +628,6 @@ static int sd_transfer_sectors(IF_MV2(int drive,) unsigned long start,
     const int drive = 0;
 #endif
     int ret = 0;
-    bool unaligned_transfer = (int)buf & 3;
 
     /* skip SanDisk OF */
     if (drive == INTERNAL_AS3525)
@@ -693,16 +693,11 @@ static int sd_transfer_sectors(IF_MV2(int drive,) unsigned long start,
                 transfer = BLOCKS_PER_BANK - bank_start;
         }
 
-        if(unaligned_transfer)
-        {
-            dma_buf = aligned_buffer;
-            if(transfer > UNALIGNED_NUM_SECTORS)
-                transfer = UNALIGNED_NUM_SECTORS;
-            if(write)
-                memcpy(aligned_buffer, buf, transfer * SECTOR_SIZE);
-        }
-        else    /* Aligned transfers are faster : no memcpy */
-            dma_buf = buf;
+        dma_buf = aligned_buffer;
+        if(transfer > UNALIGNED_NUM_SECTORS)
+            transfer = UNALIGNED_NUM_SECTORS;
+        if(write)
+            memcpy(uncached_buffer, buf, transfer * SECTOR_SIZE);
 
         /* Set bank_start to the correct unit (blocks or bytes) */
         if(!(card_info[drive].ocr & (1<<30)))   /* not SDHC */
@@ -734,8 +729,8 @@ static int sd_transfer_sectors(IF_MV2(int drive,) unsigned long start,
         wakeup_wait(&transfer_completion_signal, TIMEOUT_BLOCK);
         if(!retry)
         {
-            if(unaligned_transfer && !write)
-                memcpy(buf, aligned_buffer, transfer * SECTOR_SIZE);
+            if(!write)
+                memcpy(buf, uncached_buffer, transfer * SECTOR_SIZE);
             buf += transfer * SECTOR_SIZE;
             start += transfer;
             count -= transfer;
diff --git a/firmware/target/arm/as3525/boot.lds b/firmware/target/arm/as3525/boot.lds
index a98763e8f6..9c6d6faf11 100644
--- a/firmware/target/arm/as3525/boot.lds
+++ b/firmware/target/arm/as3525/boot.lds
@@ -6,17 +6,11 @@ OUTPUT_FORMAT(elf32-littlearm)
 OUTPUT_ARCH(arm)
 STARTUP(target/arm/crt0.o)
 
-/*
-No need for DRAM in our bootloader
-#define DRAMSIZE (MEMORYSIZE * 0x100000) - TTB_SIZE
-#define DRAMORIG 0x30000000
-*/
 #define IRAMORIG 0x81000000
 #define IRAMSIZE 0x50000
 
 MEMORY
 {
-    /*DRAM : ORIGIN = DRAMORIG, LENGTH = DRAMSIZE*/
     IRAM : ORIGIN = IRAMORIG, LENGTH = IRAMSIZE
 }
 
diff --git a/firmware/target/arm/as3525/pcm-as3525.c b/firmware/target/arm/as3525/pcm-as3525.c
index 5074941f4a..4f9e18523d 100644
--- a/firmware/target/arm/as3525/pcm-as3525.c
+++ b/firmware/target/arm/as3525/pcm-as3525.c
@@ -28,6 +28,7 @@
 #include "panic.h"
 #include "as3514.h"
 #include "audiohw.h"
+#include "mmu-arm.h"
 
 #define MAX_TRANSFER (4*((1<<11)-1)) /* maximum data we can transfer via DMA
                                       * i.e. 32 bits at once (size of I2SO_DATA)
@@ -69,6 +70,7 @@ static void play_start_pcm(void)
     CGU_PERI |= CGU_I2SOUT_APB_CLOCK_ENABLE;
     CGU_AUDIO |= (1<<11);
 
+    clean_dcache_range((void*)addr, size);  /* force write back */
     dma_enable_channel(1, (void*)addr, (void*)I2SOUT_DATA, DMA_PERI_I2SOUT,
                 DMAC_FLOWCTRL_DMAC_MEM_TO_PERI, true, false, size >> 2, DMA_S1,
                 dma_callback);
@@ -164,6 +166,15 @@ const void * pcm_play_dma_get_peak_buffer(int *count)
     return (const void*)dma_start_addr;
 }
 
+#ifdef HAVE_PCM_DMA_ADDRESS
+void * pcm_dma_addr(void *addr)
+{
+    if (addr != NULL)
+        addr = UNCACHED_ADDR(addr);
+    return addr;
+}
+#endif
+
 
 /****************************************************************************
  ** Recording DMA transfer
diff --git a/firmware/target/arm/as3525/sansa-clip/button-clip.c b/firmware/target/arm/as3525/sansa-clip/button-clip.c
index b6da130d2f..2d2a3a661e 100644
--- a/firmware/target/arm/as3525/sansa-clip/button-clip.c
+++ b/firmware/target/arm/as3525/sansa-clip/button-clip.c
@@ -46,8 +46,9 @@ int button_read_device(void)
 
     /* This is a keypad using C4-C6 as columns and B0-B2 as rows */
     GPIOC_PIN(4) = (1<<4);
+    asm volatile("nop\nnop\nnop\nnop\nnop\n"); /* small delay */
 
-    /* C4B0 is unused */
+    (void)GPIOB_PIN(0); /* C4B0 is unused */
 
     if (GPIOB_PIN(1))
         result |= BUTTON_VOL_UP;
@@ -58,6 +59,7 @@ int button_read_device(void)
     GPIOC_PIN(4) = 0x00;
 
     GPIOC_PIN(5) = (1<<5);
+    asm volatile("nop\nnop\nnop\nnop\nnop\n"); /* small delay */
 
     if (GPIOB_PIN(0))
         result |= BUTTON_LEFT;
@@ -71,6 +73,7 @@ int button_read_device(void)
     GPIOC_PIN(5) = 0x00;
 
     GPIOC_PIN(6) = (1<<6);
+    asm volatile("nop\nnop\nnop\nnop\nnop\n"); /* small delay */
 
     if (GPIOB_PIN(0))
         result |= BUTTON_DOWN;
diff --git a/firmware/target/arm/as3525/sansa-fuze/button-fuze.c b/firmware/target/arm/as3525/sansa-fuze/button-fuze.c
index f8c489c4dd..266ca79deb 100644
--- a/firmware/target/arm/as3525/sansa-fuze/button-fuze.c
+++ b/firmware/target/arm/as3525/sansa-fuze/button-fuze.c
@@ -150,7 +150,7 @@ bool button_hold(void)
 
 static void button_delay(void)
 {
-    int i = 24;
+    int i = 50;
     while(i--) asm volatile ("nop\n");
 }
 
diff --git a/firmware/target/arm/as3525/system-as3525.c b/firmware/target/arm/as3525/system-as3525.c
index 2f320e91ab..60fe7032e6 100644
--- a/firmware/target/arm/as3525/system-as3525.c
+++ b/firmware/target/arm/as3525/system-as3525.c
@@ -28,6 +28,9 @@
 #include "clock-target.h"
 #include "fmradio_i2c.h"
 #include "button-target.h"
+#ifndef BOOTLOADER
+#include "mmu-arm.h"
+#endif
 
 #define default_interrupt(name) \
   extern __attribute__((weak,alias("UIRQ"))) void name (void)
@@ -214,6 +217,26 @@ static void sdram_init(void)
 
     MPMC_DYNAMIC_CONFIG_0 |= (1<<19); /* buffer enable */
 }
+#else
+void memory_init(void)
+{
+    ttb_init();
+    /* map every region to itself, uncached */
+    map_section(0, 0, 4096, CACHE_NONE);
+
+    /* IRAM */
+    map_section(0, IRAM_ORIG, 1, CACHE_ALL);
+    map_section(0, UNCACHED_ADDR(IRAM_ORIG), 1, CACHE_NONE);
+
+    /* DRAM */
+    map_section(0x30000000, DRAM_ORIG, MEMORYSIZE, CACHE_ALL);
+    map_section(0x30000000, UNCACHED_ADDR(DRAM_ORIG), MEMORYSIZE, CACHE_NONE);
+
+    /* map 1st mbyte of DRAM at 0x0 to have exception vectors available */
+    map_section(0x30000000, 0, 1, CACHE_ALL);
+
+    enable_mmu();
+}
 #endif
 
 void system_init(void)
@@ -226,6 +249,9 @@ void system_init(void)
         CCU_SRL = CCU_SRL_MAGIC_NUMBER;
     CCU_SRC = CCU_SRL = 0;
 
+    CCU_SCON = 1; /* AHB master's priority configuration :
+                     TIC (Test Interface Controller) > DMA > USB > IDE > ARM */
+
     CGU_PROC = 0;           /* fclk 24 MHz */
     CGU_PERI &= ~0x7f;      /* pclk 24 MHz */
 
@@ -244,10 +270,8 @@ void system_init(void)
 
     asm volatile(
         "mov r0, #0               \n"
-        "mcr p15, 0, r0, c7, c7   \n"      /* invalidate icache & dcache */
         "mrc p15, 0, r0, c1, c0   \n"      /* control register */
-        "bic r0, r0, #3<<30       \n"      /* clears bus bits & sets fastbus */
-        "orr r0, r0, #1<<12       \n"      /* enable icache */
+        "bic r0, r0, #3<<30       \n"      /* clears bus bits : sets fastbus */
         "mcr p15, 0, r0, c1, c0   \n"
         : : : "r0" );
 
diff --git a/firmware/target/arm/as3525/system-target.h b/firmware/target/arm/as3525/system-target.h
index 713e96c4b8..daea180a64 100644
--- a/firmware/target/arm/as3525/system-target.h
+++ b/firmware/target/arm/as3525/system-target.h
@@ -25,4 +25,10 @@
 
 #include "clock-target.h" /* CPUFREQ_* are defined here */
 
+#ifdef BOOTLOADER
+#define UNCACHED_ADDR(a) (a)
+#else
+#define UNCACHED_ADDR(a) (a + 0x10000000)
+#endif
+
 #endif /* SYSTEM_TARGET_H */
diff --git a/firmware/target/arm/crt0.S b/firmware/target/arm/crt0.S
index 35d0aecccb..4bd01e4952 100644
--- a/firmware/target/arm/crt0.S
+++ b/firmware/target/arm/crt0.S
@@ -58,16 +58,8 @@ newstart:
     msr     cpsr_c, #0xd3 /* enter supervisor mode, disable IRQ/FIQ */
 
 #if CONFIG_CPU==AS3525 && !defined(BOOTLOADER)
-
-    /* relocate vectors */
-    mov     r1, #0              @ destination
-    ldr     r2, =_vectorscopy   @ source
-    ldr     r3, =_vectorsend    @ end
-
-1:  ldr     r0, [r2], #4
-    str     r0, [r1], #4
-    cmp     r1, r3
-    bne     1b
+    /* Setup MMU : has to be done before accessing IRAM ! */
+    bl      memory_init
 
     /* Zero out IBSS */
     ldr     r2, =_iedata