Berry stores compiled bytecode into IRAM, freeing space in heap

CHANGELOG.md

@@ -11,6 +11,7 @@ All notable changes to this project will be documented in this file.
 - PubSubClient library from v2.8.12 to v2.8.13
 - From Semantic Versioning (SemVer) to Calendar Versioning (CalVer)
 - Set ESP32 stack size with ``#define SET_ESP32_STACK_SIZE``, added ``StackLowMark`` metrics
+- Berry stores compiled bytecode into IRAM, freeing space in heap
 
 ### Fixed
 - Intermittent exceptions and heap corruption due to PubSubClient library buffer overflow (#13700)

lib/libesp32/berry/default/berry_conf.h

@@ -214,6 +214,9 @@ extern "C" {
   extern void *berry_malloc(size_t size);
   extern void  berry_free(void *ptr);
   extern void *berry_realloc(void *ptr, size_t size);
+  extern void *berry_malloc32(size_t size);
+  extern void  berry_free32(void *ptr);
+  extern void *berry_realloc32(void *ptr, size_t size);
 #ifdef __cplusplus
 }
 #endif
@@ -241,7 +244,7 @@ extern "C" {
 /* Tasmota debug specific */
 #ifdef USE_BERRY_DEBUG
   #undef BE_DEBUG_RUNTIME_INFO
-  #define BE_DEBUG_RUNTIME_INFO 2 /* record line information in 16 bits */
+  #define BE_DEBUG_RUNTIME_INFO 1 /* record line information in 32 bits to be places in IRAM */
 #endif // USE_BERRY_DEBUG
 
 #endif

lib/libesp32/berry/src/be_parser.c

@@ -313,14 +313,14 @@ static void end_func(bparser *parser)
     be_code_ret(finfo, NULL); /* append a return to last code */
     end_block(parser); /* close block */
     setupvals(finfo); /* close upvals */
-    proto->code = be_vector_release(vm, &finfo->code); /* compact all vectors and return NULL if empty */
+    proto->code = be_vector_release_32(vm, &finfo->code); /* compact all vectors and return NULL if empty */
     proto->codesize = finfo->pc;
-    proto->ktab = be_vector_release(vm, &finfo->kvec);
+    proto->ktab = be_vector_release_32(vm, &finfo->kvec);
     proto->nconst = be_vector_count(&finfo->kvec);
     proto->ptab = be_vector_release(vm, &finfo->pvec);
     proto->nproto = be_vector_count(&finfo->pvec);
 #if BE_DEBUG_RUNTIME_INFO
-    proto->lineinfo = be_vector_release(vm, &finfo->linevec);
+    proto->lineinfo = be_vector_release_32(vm, &finfo->linevec);
     proto->nlineinfo = be_vector_count(&finfo->linevec);
 #endif
 #if BE_DEBUG_VAR_INFO

lib/libesp32/berry/src/be_vector.c

@@ -110,6 +110,32 @@ void* be_vector_release(bvm *vm, bvector *vector)
     return vector->data;
 }
 
+/* free not used */
+void* be_vector_release_32(bvm *vm, bvector *vector)
+{
+    size_t size = vector->size;
+    int count = be_vector_count(vector);
+    if (count == 0) {
+        be_free(vm, vector->data, vector->capacity * size);
+        vector->capacity = 0;
+        vector->data = NULL;
+        vector->end = NULL;
+    } else if (count < vector->capacity) {
+        vector->data = be_realloc(vm,
+            vector->data, vector->capacity * size, count * size);       // TODO - can we skip that step?
+        void* iram = berry_malloc32(count * size);
+        if (iram) {
+            memcpy(iram, vector->data, count * size);
+            free(vector->data);
+            vector->data = iram;
+        }
+        // vm->gc.usage = vm->gc.usage + count * size - vector->capacity * size; /* update allocated count */
+        vector->end = (char*)vector->data + ((size_t)count - 1) * size;
+        vector->capacity = count;
+    }
+    return vector->data;
+}
+
 /* use binary search to find the vector capacity between 0-1024 */
 static int binary_search(int value)
 {

lib/libesp32/berry/src/be_vector.h

@@ -38,6 +38,7 @@ void be_vector_remove_end(bvector *vector);
 void be_vector_resize(bvm *vm, bvector *vector, int count);
 void be_vector_clear(bvector *vector);
 void* be_vector_release(bvm *vm, bvector *vector);
+void* be_vector_release_32(bvm *vm, bvector *vector);   /* specialized call for 32 bits aligned accesses */
 int be_nextsize(int value);
 
 #endif

tasmota/my_user_config.h

@@ -1009,6 +1009,7 @@
   #define USE_BERRY_PYTHON_COMPAT                // Enable by default `import python_compat`
   #define USE_BERRY_TIMEOUT             4000     // Timeout in ms, will raise an exception if running time exceeds this timeout
   #define USE_BERRY_PSRAM                        // Allocate Berry memory in PSRAM if PSRAM is connected - this might be slightly slower but leaves main memory intact
+  #define USE_BERRY_IRAM                         // Allocate some data structures in IRAM (which is ususally unused) when possible and if no PSRAM is available
   // #define USE_BERRY_DEBUG                        // Compile Berry bytecode with line number information, makes exceptions easier to debug. Adds +8% of memory consumption for compiled code
   #define USE_WEBCLIENT                          // Enable `webclient` to make HTTP/HTTPS requests. Can be disabled for security reasons.
     // #define USE_WEBCLIENT_HTTPS                  // Enable HTTPS outgoing requests based on BearSSL (much ligher then mbedTLS, 42KB vs 150KB) in insecure mode (no verification of server's certificate)

tasmota/support_esp.ino

@@ -503,6 +503,17 @@ void *special_calloc(size_t num, size_t size) {
   }
 }
 
+// Variants for IRAM heap, which need all accesses to be 32 bits aligned
+void *special_malloc32(uint32_t size) {
+  return heap_caps_malloc(size, UsePSRAM() ? MALLOC_CAP_SPIRAM | MALLOC_CAP_8BIT : MALLOC_CAP_32BIT);
+}
+void *special_realloc32(void *ptr, size_t size) {
+  return heap_caps_realloc(ptr, size, UsePSRAM() ? MALLOC_CAP_SPIRAM | MALLOC_CAP_8BIT : MALLOC_CAP_32BIT);
+}
+void *special_calloc32(size_t num, size_t size) {
+  return heap_caps_calloc(num, size, UsePSRAM() ? MALLOC_CAP_SPIRAM | MALLOC_CAP_8BIT : MALLOC_CAP_32BIT);
+}
+
 float CpuTemperature(void) {
 #ifdef CONFIG_IDF_TARGET_ESP32
   return (float)temperatureRead();  // In Celsius

tasmota/xdrv_52_9_berry.ino

@@ -84,6 +84,29 @@ extern "C" {
   }
 #endif // USE_BERRY_PSRAM
 
+
+  void *berry_malloc32(uint32_t size) {
+  #ifdef USE_BERRY_IRAM
+    return special_malloc32(size);
+  #else
+    return special_malloc(size);
+  #endif
+  }
+  void *berry_realloc32(void *ptr, size_t size) {
+  #ifdef USE_BERRY_IRAM
+    return special_realloc32(ptr, size);
+  #else
+    return special_realloc(ptr, size);
+  #endif
+  }
+  void *berry_calloc32(size_t num, size_t size) {
+  #ifdef USE_BERRY_IRAM
+    return special_calloc32(num, size);
+  #else
+    return special_calloc(num, size);
+  #endif
+  }
+
   void berry_free(void *ptr) {
     free(ptr);
   }