Bring up WAMR on esp32-s3 device (#2348)

esp32-s3's instruction memory and data memory can be accessed through mutual mirroring way,
so we define a new feature named as WASM_MEM_DUAL_BUS_MIRROR.

core/config.h

@@ -461,4 +461,15 @@
 #define WASM_CONFIGURABLE_BOUNDS_CHECKS 0
 #endif
 
+/* Some chip cannot support external ram with rwx attr at the same time,
+   it has to map it into 2 spaces of idbus and dbus, code in dbus can be
+   read/written and read/executed in ibus. so there are 2 steps to execute
+   the code, first, copy&do relocaiton in dbus space, and second execute
+   it in ibus space, since in the 2 spaces the contents are the same,
+   so we call it bus mirror.
+ */
+#ifndef WASM_MEM_DUAL_BUS_MIRROR
+#define WASM_MEM_DUAL_BUS_MIRROR 0
+#endif
+
 #endif /* end of _CONFIG_H_ */

core/iwasm/aot/aot_loader.c

@@ -3013,6 +3013,9 @@ create_sections(AOTModule *module, const uint8 *buf, uint32 size,
     uint32 section_size;
     uint64 total_size;
     uint8 *aot_text;
+#if (WASM_MEM_DUAL_BUS_MIRROR != 0)
+    uint8 *mirrored_text;
+#endif
 
     if (!resolve_execute_mode(buf, size, &is_indirect_mode, error_buf,
                               error_buf_size)) {
@@ -3071,8 +3074,17 @@ create_sections(AOTModule *module, const uint8 *buf, uint32 size,
                     bh_assert((uintptr_t)aot_text < INT32_MAX);
 #endif
 #endif
+
+#if (WASM_MEM_DUAL_BUS_MIRROR != 0)
+                    mirrored_text = os_get_dbus_mirror(aot_text);
+                    bh_assert(mirrored_text != NULL);
+                    bh_memcpy_s(mirrored_text, (uint32)total_size,
+                                section->section_body, (uint32)section_size);
+                    os_dcache_flush();
+#else
                     bh_memcpy_s(aot_text, (uint32)total_size,
                                 section->section_body, (uint32)section_size);
+#endif
                     section->section_body = aot_text;
                     destroy_aot_text = true;
 

core/iwasm/aot/arch/aot_reloc_xtensa.c

@@ -207,6 +207,10 @@ apply_relocation(AOTModule *module, uint8 *target_section_addr,
         case R_XTENSA_32:
         {
             uint8 *insn_addr = target_section_addr + reloc_offset;
+#if (WASM_MEM_DUAL_BUS_MIRROR != 0)
+            insn_addr = os_get_dbus_mirror((void *)insn_addr);
+            bh_assert(insn_addr != NULL);
+#endif
             int32 initial_addend;
             /* (S + A) */
             if ((intptr_t)insn_addr & 3) {
@@ -265,6 +269,11 @@ apply_relocation(AOTModule *module, uint8 *target_section_addr,
                 return false;
             }
 
+#if (WASM_MEM_DUAL_BUS_MIRROR != 0)
+            insn_addr = os_get_dbus_mirror((void *)insn_addr);
+            bh_assert(insn_addr != NULL);
+            l32r_insn = (l32r_insn_t *)insn_addr;
+#endif
             imm16 = (int16)(relative_offset >> 2);
 
             /* write back the imm16 to the l32r instruction */
@@ -285,7 +294,6 @@ apply_relocation(AOTModule *module, uint8 *target_section_addr,
 #if __GNUC__ >= 9
 #pragma GCC diagnostic pop
 #endif
-
             break;
         }
 

core/shared/platform/include/platform_api_vmcore.h

@@ -129,6 +129,11 @@ os_munmap(void *addr, size_t size);
 int
 os_mprotect(void *addr, size_t size, int prot);
 
+#if (WASM_MEM_DUAL_BUS_MIRROR != 0)
+void *
+os_get_dbus_mirror(void *ibus);
+#endif
+
 /**
  * Flush cpu data cache, in some CPUs, after applying relocation to the
  * AOT code, the code may haven't been written back to the cpu data cache,

core/shared/platform/nuttx/nuttx_platform.c

@@ -10,6 +10,46 @@
 #include <nuttx/arch.h>
 #endif
 
+#if defined(CONFIG_ARCH_CHIP_ESP32S3)
+/*
+ * TODO: Move these methods below the operating system level
+ */
+#define MEM_DUAL_BUS_OFFSET (0x42000000 - 0x3C000000)
+#define IRAM0_CACHE_ADDRESS_LOW 0x42000000
+#define IRAM0_CACHE_ADDRESS_HIGH 0x44000000
+#define IRAM_ATTR locate_data(".iram1")
+
+#define in_ibus_ext(addr)                      \
+    (((uint32)addr >= IRAM0_CACHE_ADDRESS_LOW) \
+     && ((uint32)addr < IRAM0_CACHE_ADDRESS_HIGH))
+void IRAM_ATTR
+bus_sync(void)
+{
+    extern void cache_writeback_all(void);
+    extern uint32_t Cache_Disable_ICache(void);
+    extern void Cache_Enable_ICache(uint32_t autoload);
+
+    irqstate_t flags;
+    uint32_t preload;
+
+    flags = enter_critical_section();
+
+    cache_writeback_all();
+    preload = Cache_Disable_ICache();
+    Cache_Enable_ICache(preload);
+
+    leave_critical_section(flags);
+}
+#else
+#define MEM_DUAL_BUS_OFFSET (0)
+#define IRAM0_CACHE_ADDRESS_LOW (0)
+#define IRAM0_CACHE_ADDRESS_HIGH (0)
+#define in_ibus_ext(addr) (0)
+static void
+bus_sync(void)
+{}
+#endif
+
 int
 bh_platform_init()
 {
@@ -47,6 +87,10 @@ os_dumps_proc_mem_info(char *out, unsigned int size)
 void *
 os_mmap(void *hint, size_t size, int prot, int flags)
 {
+#if (WASM_MEM_DUAL_BUS_MIRROR != 0)
+    void *i_addr, *d_addr;
+#endif
+
 #if defined(CONFIG_ARCH_USE_TEXT_HEAP)
     if ((prot & MMAP_PROT_EXEC) != 0) {
         return up_textheap_memalign(sizeof(void *), size);
@@ -55,6 +99,17 @@ os_mmap(void *hint, size_t size, int prot, int flags)
 
     if ((uint64)size >= UINT32_MAX)
         return NULL;
+
+#if (WASM_MEM_DUAL_BUS_MIRROR != 0)
+    if ((prot & MMAP_PROT_EXEC) != 0) {
+        d_addr = malloc((uint32)size);
+        if (d_addr == NULL) {
+            return NULL;
+        }
+        i_addr = (void *)((uint8 *)d_addr + MEM_DUAL_BUS_OFFSET);
+        return in_ibus_ext(i_addr) ? i_addr : d_addr;
+    }
+#endif
     return malloc((uint32)size);
 }
 
@@ -67,7 +122,14 @@ os_munmap(void *addr, size_t size)
         return;
     }
 #endif
-    return free(addr);
+
+#if (WASM_MEM_DUAL_BUS_MIRROR != 0)
+    if (in_ibus_ext(addr)) {
+        free((void *)((uint8 *)addr - MEM_DUAL_BUS_OFFSET));
+        return;
+    }
+#endif
+    free(addr);
 }
 
 int
@@ -78,7 +140,22 @@ os_mprotect(void *addr, size_t size, int prot)
 
 void
 os_dcache_flush()
-{}
+{
+    bus_sync();
+}
+
+#if (WASM_MEM_DUAL_BUS_MIRROR != 0)
+void *
+os_get_dbus_mirror(void *ibus)
+{
+    if (in_ibus_ext(ibus)) {
+        return (void *)((uint8 *)ibus - MEM_DUAL_BUS_OFFSET);
+    }
+    else {
+        return ibus;
+    }
+}
+#endif
 
 /* If AT_FDCWD is provided, maybe we have openat family */
 #if !defined(AT_FDCWD)

product-mini/platforms/nuttx/wamr.mk

@@ -141,6 +141,12 @@ else
 CFLAGS += -DWASM_ENABLE_WORD_ALIGN_READ=0
 endif
 
+ifeq ($(CONFIG_INTERPRETERS_WAMR_MEM_DUAL_BUS_MIRROR),y)
+CFLAGS += -DWASM_MEM_DUAL_BUS_MIRROR=1
+else
+CFLAGS += -DWASM_MEM_DUAL_BUS_MIRROR=0
+endif
+
 ifeq ($(CONFIG_INTERPRETERS_WAMR_FAST), y)
 CFLAGS += -DWASM_ENABLE_FAST_INTERP=1
 CFLAGS += -DWASM_ENABLE_INTERP=1