Malloc support for allocating to DRAM and/or IRAM.

core/app_main.c

@@ -137,6 +137,9 @@ static void IRAM default_putc(char c) {
 void init_newlib_locks(void);
 extern uint8_t sdk_wDevCtrl[];
 void nano_malloc_insert_chunk(void *start, size_t size);
+extern uint8_t _heap_start[];
+extern uint8_t _text_end[];
+extern uint8_t enable_low_icache;
 
 // .text+0x258
 void IRAM sdk_user_start(void) {
@@ -209,8 +212,13 @@ void IRAM sdk_user_start(void) {
     cksum_value = buf32[5 + boot_slot];
     ic_flash_addr = (flash_sectors - 3 + boot_slot) * sdk_flashchip.sector_size;
     sdk_SPIRead(ic_flash_addr, buf32, sizeof(struct sdk_g_ic_saved_st));
+
+#ifdef ESP8266_ENABLE_LOW_ICACHE
+    enable_low_icache = ESP8266_ENABLE_LOW_ICACHE;
+#endif
     Cache_Read_Enable(0, 0, 1);
     zero_bss();
+
     sdk_os_install_putc1(default_putc);
 
     /* HACK Reclaim a region of unused bss from wdev.o. This would not be
@@ -219,6 +227,26 @@ void IRAM sdk_user_start(void) {
      * it is in very useful dram. */
     nano_malloc_insert_chunk((void *)(sdk_wDevCtrl + 0x2190), 8000);
 
+    /* Use all the used DRAM is for the dynamic heap. */
+    nano_malloc_insert_chunk(_heap_start, 0x3FFFC000 - (uintptr_t)_heap_start);
+
+    /* Add unused IRAM to the malloc free list. */
+    if (enable_low_icache) {
+        /* The memory region 0x40108000 to 0x4010C000 is used for icache so can
+         * not be used, but there might still be some unused IRAM */
+        nano_malloc_insert_chunk(_text_end, 0x40108000 - (uintptr_t)_text_end);
+    } else {
+        /* The memory region 0x40108000 to 0x4010C000 is not used as part of the
+         * instruction cache and is usable as extra IRAM. */
+        nano_malloc_insert_chunk(_text_end, 0x4010C000 - (uintptr_t)_text_end);
+    }
+
+    /* The preferred memory region to start allocate the early data. If the app
+     * has ample memory the use the DRAM, other if the app is running low on
+     * DRAM then it might help the allocated to the IRAM when possible. */
+    set_malloc_regions(MALLOC_MASK_PREFER_DRAM);
+    //set_malloc_regions(MALLOC_MASK_PREFER_IRAM);
+
     init_newlib_locks();
 
     if (cksum_magic == 0xffffffff) {
@@ -368,6 +396,7 @@ void sdk_user_init_task(void *params) {
     /* The start up stack is not used after scheduling has started, so all of
      * the top area of RAM which was stack can be used for the dynamic heap. */
     xPortSupervisorStackPointer = (void *)0x40000000;
+    nano_malloc_insert_chunk((void *)0x3FFFC000, 0x4000);
 
     sdk_ets_timer_init();
     printf("\nESP-Open-SDK ver: %s compiled @ %s %s\n", OS_VERSION_STR, __DATE__, __TIME__);

core/include/common_macros.h

@@ -117,5 +117,10 @@
     #define IROM __attribute__((section(".irom0.literal"))) const
 #endif
 
+uint32_t set_malloc_regions(uint32_t mask);
+#define MALLOC_MASK_PREFER_IRAM 0xfffdfffc
+#define MALLOC_MASK_PREFER_DRAM 0
+#define MALLOC_MASK_DRAM 0xfffffffe
+#define MALLOC_MASK_IRAM 0xfffffffd
 
 #endif

core/newlib_syscalls.c

@@ -34,39 +34,30 @@
 #error Too many lwip sockets for the FD_SETSIZE.
 #endif
 
-extern void *xPortSupervisorStackPointer;
-
-IRAM void *_sbrk_r (struct _reent *r, ptrdiff_t incr)
+void *_sbrk_r (struct _reent *r, ptrdiff_t incr)
 {
-    extern char   _heap_start; /* linker script defined */
-    static char * heap_end;
-    char *        prev_heap_end;
-
-    if (heap_end == NULL)
-	heap_end = &_heap_start;
-    prev_heap_end = heap_end;
-
-    intptr_t sp = (intptr_t)xPortSupervisorStackPointer;
-    if(sp == 0) /* scheduler not started */
-        SP(sp);
-
-    if ((intptr_t)heap_end + incr >= sp)
-    {
-        r->_errno = ENOMEM;
-        return (caddr_t)-1;
-    }
-
-    heap_end += incr;
-
-    return (caddr_t) prev_heap_end;
+    r->_errno = ENOMEM;
+    return (caddr_t)-1;
 }
 
+/* If there is a restriction on the dram usage then skip this chunk if in dram,
+ * and if there is a restriction on the iram usage then skip this chunk if in
+ * iram */
+IRAM int _malloc_region_masked(void *r, unsigned int mask)
+{
+    if ( ((mask & 1) && (uint32_t)r < 0x40000000) ||
+         ((mask & 2) && (uint32_t)r >= 0x40100000) ) {
+        return 1;
+    }
 
-/* Insert a disjoint region into the nano malloc pool. Create a malloc chunk,
- * filling the size as newlib nano malloc expects, and then free it. */
-void nano_malloc_insert_chunk(void *start, size_t size) {
-    *(uint32_t *)start = size;
-    free(start + sizeof(size_t));
+    return 0;
+}
+
+uint32_t set_malloc_regions(uint32_t mask)
+{
+    uint32_t malloc_mask = _REENT->malloc_region_mask;
+    _REENT->malloc_region_mask = mask;
+    return malloc_mask;
 }
 
 /* syscall implementation for stdio write to UART */