Two new sections for IRAM usage: .irodata (selectable with the ICONST_ATTR attribute macro), allowing to put 'const' data into IRAM without causing a section type conflict, and .ibss (selectable with the IBSS_ATTR attribute macro) for uninitialised data. * Rockbox core: Adjusted the linker scripts and init code to not include the .ibss section in the binary, it is cleared instead. Saves ~500 bytes on archos and ~30KB on iriver. Codecs and plugins don't handle .ibss in a special way yet. * The init code properly handles empty sections now (except .stack, which should never be empty). * Unified the init code for SH1 and coldfire a bit.

git-svn-id: svn://svn.rockbox.org/rockbox/trunk@7644 a1c6a512-1295-4272-9138-f99709370657

apps/dsp.c

@@ -130,9 +130,9 @@ struct dither_data
     long random;
 };
 
-static struct dsp_config dsp_conf[2] IDATA_ATTR;
+static struct dsp_config dsp_conf[2] IBSS_ATTR;
-static struct dither_data dither_data[2] IDATA_ATTR;
+static struct dither_data dither_data[2] IBSS_ATTR;
-static struct resample_data resample_data[2][2] IDATA_ATTR;
+static struct resample_data resample_data[2][2] IBSS_ATTR;
 
 extern int current_codec;
 struct dsp_config *dsp;
@@ -142,8 +142,8 @@ struct dsp_config *dsp;
  * of copying needed is minimized for that case.
  */
 
-static long sample_buf[SAMPLE_BUF_SIZE] IDATA_ATTR;
+static long sample_buf[SAMPLE_BUF_SIZE] IBSS_ATTR;
-static long resample_buf[RESAMPLE_BUF_SIZE] IDATA_ATTR;
+static long resample_buf[RESAMPLE_BUF_SIZE] IBSS_ATTR;
 
 
 /* Convert at most count samples to the internal format, if needed. Returns

apps/playback.c

@@ -112,13 +112,13 @@ static const char audio_thread_name[] = "audio";
 
 /* Codec thread. */
 static struct event_queue codec_queue;
-static long codec_stack[(DEFAULT_STACK_SIZE + 0x2000)/sizeof(long)] IDATA_ATTR;
+static long codec_stack[(DEFAULT_STACK_SIZE + 0x2000)/sizeof(long)] IBSS_ATTR;
 static const char codec_thread_name[] = "codec";
 
 /* Voice codec thread. */
 static struct event_queue voice_codec_queue;
 /* Not enough IRAM for this. */
-static long voice_codec_stack[(DEFAULT_STACK_SIZE + 0x2000)/sizeof(long)] IDATA_ATTR;
+static long voice_codec_stack[(DEFAULT_STACK_SIZE + 0x2000)/sizeof(long)] IBSS_ATTR;
 static const char voice_codec_thread_name[] = "voice codec";
 
 static struct mutex mutex_bufferfill;

apps/plugins/plugin.lds

@@ -80,7 +80,9 @@ SECTIONS
     {
         iramstart = .;
         *(.icode)
+        *(.irodata)
         *(.idata)
+        *(.ibss)
         iramend = .;
     } > PLUGIN_IRAM
 #endif

firmware/app.lds

@@ -188,9 +188,18 @@ SECTIONS
     {
         _iramstart = .;
         *(.icode)
+        *(.irodata)
         *(.idata)
         _iramend = .;
     } > IRAM
+    
+    .ibss (NOLOAD) :
+    {
+        _iedata = .;
+        *(.ibss)
+        . = ALIGN(0x4);
+        _iend = .;
+    } > IRAM
 
 #ifdef CPU_COLDFIRE
     .stack :

firmware/boot.lds

@@ -51,9 +51,10 @@ SECTIONS
         *(.resetvectors);
         *(.vectors);
         . = ALIGN(0x200);
-        *(.data)
         *(.icode)
+        *(.irodata)
         *(.idata)
+        *(.data)
         . = ALIGN(0x4);
         _dataend = .;         
         . = ALIGN(0x10);  /* Maintain proper alignment for .text section */
@@ -91,6 +92,7 @@ SECTIONS
     .bss :
     {
        _edata = .;
+        *(.ibss)
         *(.bss)
         *(COMMON)
        _end = .;

firmware/crt0.S

@@ -246,7 +246,7 @@ irq_handler:
 
     move.l  #0xffffffbf,%d0
     and.l  %d0,(0x108,%a0)  /* Back to normal, the DRAM is now ready */
-#endif
+#endif /* BOOTLOADER */
 
     /* Invalicate cache */
     move.l  #0x01000000,%d0
@@ -263,38 +263,57 @@ irq_handler:
     movec.l %d0,%acr1
 
 #ifndef BOOTLOADER
+    /* zero out .ibss */
+    lea     _iedata,%a2
+    lea     _iend,%a4
+    bra.b   .iedatastart
+.iedataloop:
+    clr.l   (%a2)+
+.iedatastart:
+    cmp.l   %a2,%a4
+    bhi.b   .iedataloop
+
+    /* copy the .iram section */
     lea     _iramcopy,%a2
     lea     _iramstart,%a3
     lea     _iramend,%a4
+    bra.b   .iramstart
 .iramloop:
     move.l  (%a2)+,(%a3)+
+.iramstart:
     cmp.l   %a3,%a4
     bhi.b   .iramloop
-#endif
+#endif /* !BOOTLOADER */
 
+    /* zero out bss */
     lea     _edata,%a2
     lea     _end,%a4
+    bra.b   .edatastart
 .edataloop:
     clr.l   (%a2)+
+.edatastart:
     cmp.l   %a2,%a4
     bhi.b   .edataloop
 
+    /* copy the .data section */
     lea     _datacopy,%a2
     lea     _datastart,%a3
-    lea     _dataend,%a4
     cmp.l   %a2,%a3
-    beq.b   .nodatacopy
+    beq.b   .nodatacopy   /* Don't copy if src and dest are equal */
+    lea     _dataend,%a4
+    bra.b   .datastart
 .dataloop:
     move.l  (%a2)+,(%a3)+
+.datastart:
     cmp.l   %a3,%a4
     bhi.b   .dataloop
 .nodatacopy:
 
     /* Munge the main stack */
-    move.l  #0xdeadbeef,%d0
     lea     stackbegin,%a2
     lea     stackend,%a4
     move.l  %a4,%sp
+    move.l  #0xdeadbeef,%d0
 .mungeloop:
     move.l  %d0,(%a2)+
     cmp.l   %a2,%a4
@@ -315,12 +334,12 @@ vectors:
 #else
     /* Platform:     Archos Jukebox */
 
-    mov.l   vbr_k,r1
+    mov.l   .vbr_k,r1
 #ifdef DEBUG
     /* If we have built our code to be loaded via the standalone GDB
      * stub, we will have out VBR at some other location than 0x9000000.
      * We must copy the trap vectors for the GDB stub to our vector table. */
-    mov.l   orig_vbr_k,r2
+    mov.l   .orig_vbr_k,r2
 
     /* Move the invalid instruction vector (4) */
     mov     #4,r0
@@ -384,58 +403,77 @@ vectors:
 #endif /* DEBUG */
     ldc     r1,vbr
 
-    /* Now let's get on with the normal business */
-    mov.l   stackend_k,r15
-
-    /* zero out bss */
-    mov.l   edata_k,r0
-    mov.l   end_k,r1
-    mov     #0,r2
-edata_l:    /* backwards is faster and shorter */
-    mov.l   r2,@-r1
-    cmp/hi  r0,r1
-    bt      edata_l
-
-    /* copy the .iram section */
-    mov.l   iramcopy_k,r0
-    mov.l   iram_k,r1
-    mov.l   iramend_k,r2
-iramcopy_l:
-    mov.l   @r0+,r3
-    mov.l   r3,@r1
-    add     #4,r1
-    cmp/hi  r1,r2
-    bt      iramcopy_l
-
-    /* copy the .data section, for rombased execution */
-    mov.l   datacopy_k,r0
-    mov.l   data_k,r1
-    mov.l   dataend_k,r2
-    /* Don't copy if src and dest are equal */
-    cmp/eq  r0,r1
-    bt      nodatacopy
-data_l:
-    mov.l   @r0+,r3
-    mov.l   r3,@r1
-    add     #4,r1
-    cmp/hi  r1,r2
-    bt      data_l
-nodatacopy:
-
-    /* Munge the main thread stack */
-    mov.l   stackbegin_k,r0
-    mov.l   stackend_k,r1
-    mov.l   deadbeef_k,r2
-munge_l:    /* backwards is faster and shorter */
-    mov.l   r2,@-r1
-    cmp/hi  r0,r1
-    bt      munge_l
-
     mov     #0,r0
     ldc     r0,gbr
 
-    ! call the mainline
+    /* zero out .ibss */
-    mov.l   main_k,r0
+    mov.l   .iedata_k,r0
+    mov.l   .iend_k,r1
+    bra     .iedatastart
+    mov     #0,r2
+.iedataloop: /* backwards is faster and shorter */
+    mov.l   r2,@-r1
+.iedatastart:
+    cmp/hi  r0,r1
+    bt      .iedataloop
+
+    /* copy the .iram section */
+    mov.l   .iramcopy_k,r0
+    mov.l   .iram_k,r1
+    mov.l   .iramend_k,r2
+    /* Note: We cannot put a PC relative load into the delay slot of a 'bra'
+       instruction (the offset would be wrong), but there is nothing else to
+       do before the loop, so the delay slot would be 'nop'. The cmp / bf
+       sequence is the same length, but more efficient. */
+    cmp/hi  r1,r2
+    bf      .noiramcopy
+.iramloop:
+    mov.l   @r0+,r3
+    mov.l   r3,@r1
+    add     #4,r1
+    cmp/hi  r1,r2
+    bt      .iramloop
+.noiramcopy:
+
+    /* zero out bss */
+    mov.l   .edata_k,r0
+    mov.l   .end_k,r1
+    bra     .edatastart
+    mov     #0,r2
+.edataloop: /* backwards is faster and shorter */
+    mov.l   r2,@-r1
+.edatastart:
+    cmp/hi  r0,r1
+    bt      .edataloop
+
+    /* copy the .data section, for rombased execution */
+    mov.l   .datacopy_k,r0
+    mov.l   .data_k,r1
+    cmp/eq  r0,r1
+    bt      .nodatacopy   /* Don't copy if src and dest are equal */
+    mov.l   .dataend_k,r2
+    cmp/hi  r1,r2
+    bf      .nodatacopy
+.dataloop:
+    mov.l   @r0+,r3
+    mov.l   r3,@r1
+    add     #4,r1
+    cmp/hi  r1,r2
+    bt      .dataloop
+.nodatacopy:
+
+    /* Munge the main thread stack */
+    mov.l   .stackbegin_k,r0
+    mov.l   .stackend_k,r1
+    mov     r1,r15
+    mov.l   .deadbeef_k,r2
+.mungeloop: /* backwards is faster and shorter */
+    mov.l   r2,@-r1
+    cmp/hi  r0,r1
+    bt      .mungeloop
+
+    /* call the mainline */
+    mov.l   .main_k,r0
     jsr     @r0
     nop
 .hoo:
@@ -443,36 +481,40 @@ munge_l:    /* backwards is faster and shorter */
     nop
 
     .align  2
-stackend_k:
+.vbr_k:
-    .long   _stackend
-stackbegin_k:
-    .long   _stackbegin
-deadbeef_k:
-    .long   0xdeadbeef
-edata_k:
-    .long   _edata
-end_k:
-    .long   _end
-iramcopy_k:
-    .long   _iramcopy
-iram_k:
-    .long   _iramstart
-iramend_k:
-    .long   _iramend
-datacopy_k:
-    .long   _datacopy
-data_k:
-    .long   _datastart
-dataend_k:
-    .long   _dataend
-main_k:
-    .long   _main
-vbr_k:
     .long   vectors
 #ifdef DEBUG
-orig_vbr_k:
+.orig_vbr_k:
-    .long   0x9000000
+    .long   0x09000000
 #endif
+.iedata_k:
+    .long   _iedata
+.iend_k:
+    .long   _iend
+.iramcopy_k:
+    .long   _iramcopy
+.iram_k:
+    .long   _iramstart
+.iramend_k:
+    .long   _iramend
+.edata_k:
+    .long   _edata
+.end_k:
+    .long   _end
+.datacopy_k:
+    .long   _datacopy
+.data_k:
+    .long   _datastart
+.dataend_k:
+    .long   _dataend
+.stackbegin_k:
+    .long   _stackbegin
+.stackend_k:
+    .long   _stackend
+.deadbeef_k:
+    .long   0xdeadbeef
+.main_k:
+    .long   _main
 
     .section .resetvectors
 vectors:

firmware/drivers/lcd-h100-remote.c

@@ -67,7 +67,7 @@
 /*** globals ***/
 
 unsigned char lcd_remote_framebuffer[LCD_REMOTE_HEIGHT/8][LCD_REMOTE_WIDTH]
-                                     IDATA_ATTR;
+                                     IBSS_ATTR;
 
 static int drawmode = DRMODE_SOLID;
 static int xmargin = 0;

firmware/drivers/lcd-h100.c

@@ -61,10 +61,9 @@
 
 /*** globals ***/
 
-unsigned char lcd_framebuffer[LCD_HEIGHT/4][LCD_WIDTH] IDATA_ATTR;
+unsigned char lcd_framebuffer[LCD_HEIGHT/4][LCD_WIDTH] IBSS_ATTR;
 
-/* should be 'const', but this causes a section type conflict */
+static const unsigned char dibits[16] ICONST_ATTR = {
-static unsigned char dibits[16] IDATA_ATTR = {
     0x00, 0x03, 0x0C, 0x0F, 0x30, 0x33, 0x3C, 0x3F,
     0xC0, 0xC3, 0xCC, 0xCF, 0xF0, 0xF3, 0xFC, 0xFF
 };

firmware/export/config.h

@@ -139,11 +139,15 @@
     (CONFIG_CPU == MCF5249) || /* Coldfire: core, plugins, codecs */ \
     (CONFIG_CPU == TCC730))  /* CalmRISC16: core, (plugins, codecs) */
 #define ICODE_ATTR	__attribute__ ((section(".icode")))
+#define ICONST_ATTR __attribute__ ((section(".irodata")))
 #define IDATA_ATTR	__attribute__ ((section(".idata")))
+#define IBSS_ATTR	__attribute__ ((section(".ibss")))
 #define USE_IRAM
 #else
 #define ICODE_ATTR
+#define ICONST_ATTR
 #define IDATA_ATTR
+#define IBSS_ATTR
 #endif
 
 #endif

firmware/rom.lds

@@ -90,10 +90,19 @@ SECTIONS
     {
         _iramstart = .;
         *(.icode)
+        *(.irodata)
         *(.idata)
         _iramend = .;
     } > IRAM
   
+    .ibss (NOLOAD) :
+    {
+        _iedata = .;
+        *(.ibss)
+        . = ALIGN(0x4);
+        _iend = .;
+    } > IRAM
+
     .stack :
     {
        *(.stack)

firmware/thread.c

@@ -57,7 +57,7 @@ struct regs
 int num_threads;
 static volatile int num_sleepers;
 static int current_thread;
-static struct regs thread_contexts[MAXTHREADS] __attribute__ ((section(".idata")));
+static struct regs thread_contexts[MAXTHREADS] IBSS_ATTR;
 const char *thread_name[MAXTHREADS];
 void *thread_stack[MAXTHREADS];
 int thread_stack_size[MAXTHREADS];
@@ -66,7 +66,7 @@ static const char main_thread_name[] = "main";
 extern int stackbegin[];
 extern int stackend[];
 
-void switch_thread(void) __attribute__ ((section(".icode")));
+void switch_thread(void) ICODE_ATTR;
 static inline void store_context(void* addr) __attribute__ ((always_inline));
 static inline void load_context(const void* addr) __attribute__ ((always_inline));