perf: 6 performance optimizations for STM32WB55 (Cortex-M4 @ 64MHz)

applications/main/subghz/scenes/subghz_scene_save_name.c

@@ -70,7 +70,7 @@ void subghz_scene_save_name_on_enter(void* context) {
         furi_string_set(subghz->file_path, dir_name);
     }
 
-    strncpy(subghz->file_name_tmp, furi_string_get_cstr(file_name), SUBGHZ_MAX_LEN_NAME);
+    strlcpy(subghz->file_name_tmp, furi_string_get_cstr(file_name), SUBGHZ_MAX_LEN_NAME);
     text_input_set_header_text(text_input, "Name signal");
     text_input_set_result_callback(
         text_input,

furi/core/check.h

@@ -59,11 +59,11 @@ FURI_NORETURN void __furi_halt_implementation(void);
 #define furi_halt(...) M_APPLY(__furi_halt, M_IF_EMPTY(__VA_ARGS__)((NULL), (__VA_ARGS__)))
 
 /** Check condition and crash if check failed */
-#define __furi_check(__e, __m) \
-    do {                       \
-        if(!(__e)) {           \
-            __furi_crash(__m); \
-        }                      \
+#define __furi_check(__e, __m)               \
+    do {                                      \
+        if(__builtin_expect(!(__e), 0)) {     \
+            __furi_crash(__m);                \
+        }                                     \
     } while(0)
 
 /** Check condition and crash if failed
@@ -75,11 +75,11 @@ FURI_NORETURN void __furi_halt_implementation(void);
 
 /** Only in debug build: Assert condition and crash if assert failed  */
 #ifdef FURI_DEBUG
-#define __furi_assert(__e, __m) \
-    do {                        \
-        if(!(__e)) {            \
-            __furi_crash(__m);  \
-        }                       \
+#define __furi_assert(__e, __m)               \
+    do {                                      \
+        if(__builtin_expect(!(__e), 0)) {     \
+            __furi_crash(__m);                \
+        }                                     \
     } while(0)
 #else
 #define __furi_assert(__e, __m) \
@@ -98,11 +98,11 @@ FURI_NORETURN void __furi_halt_implementation(void);
 #define furi_assert(...) \
     M_APPLY(__furi_assert, M_DEFAULT_ARGS(2, (__FURI_ASSERT_MESSAGE_FLAG), __VA_ARGS__))
 
-#define furi_break(__e)             \
-    do {                            \
-        if(!(__e)) {                \
-            asm volatile("bkpt 0"); \
-        }                           \
+#define furi_break(__e)                          \
+    do {                                          \
+        if(__builtin_expect(!(__e), 0)) {         \
+            asm volatile("bkpt 0");               \
+        }                                         \
     } while(0)
 
 #ifdef __cplusplus

furi/core/kernel.c

@@ -166,7 +166,7 @@ FuriStatus furi_delay_until_tick(uint32_t tick) {
     return stat;
 }
 
-uint32_t furi_get_tick(void) {
+__attribute__((flatten)) uint32_t furi_get_tick(void) {
     TickType_t ticks;
 
     if(furi_kernel_is_irq_or_masked() != 0U) {

furi/core/memmgr.c

@@ -1,4 +1,5 @@
 #include "memmgr.h"
+#include "memmgr_heap.h"
 #include <string.h>
 #include <furi_hal_memory.h>
 #include <FreeRTOS.h>
@@ -25,7 +26,9 @@ void* realloc(void* ptr, size_t size) {
 
     void* p = pvPortMalloc(size);
     if(ptr != NULL) {
-        memcpy(p, ptr, size);
+        size_t old_size = memmgr_heap_get_block_size(ptr);
+        size_t copy_size = old_size < size ? old_size : size;
+        memcpy(p, ptr, copy_size);
         vPortFree(ptr);
     }
 

furi/core/memmgr_heap.c

@@ -539,6 +539,24 @@ size_t xPortGetMinimumEverFreeHeapSize(void) {
 }
 /*-----------------------------------------------------------*/
 
+size_t memmgr_heap_get_block_size(const void* pv) {
+    const uint8_t* puc = (const uint8_t*)pv;
+    BlockLink_t* pxLink;
+
+    if(pv == NULL) {
+        return 0;
+    }
+
+    puc -= xHeapStructSize;
+    pxLink = (BlockLink_t*)(uintptr_t)puc;
+
+    heapVALIDATE_BLOCK_POINTER(pxLink);
+    configASSERT(heapBLOCK_IS_ALLOCATED(pxLink) != 0);
+
+    return (pxLink->xBlockSize & ~heapBLOCK_ALLOCATED_BITMASK) - xHeapStructSize;
+}
+/*-----------------------------------------------------------*/
+
 void xPortResetHeapMinimumEverFreeHeapSize(void) {
     xMinimumEverFreeBytesRemaining = xFreeBytesRemaining;
 }
@@ -734,3 +752,5 @@ void vPortHeapResetState(void) {
     xNumberOfSuccessfulFrees = (size_t)0U;
 }
 /*-----------------------------------------------------------*/
+
+/*-----------------------------------------------------------*/

furi/core/memmgr_heap.h

@@ -44,6 +44,13 @@ size_t memmgr_heap_get_max_free_block(void);
  */
 void memmgr_heap_printf_free_blocks(void);
 
+/** Get usable size of an allocated heap block
+ *
+ * @param      ptr   pointer to allocated memory
+ * @return     usable size in bytes
+ */
+size_t memmgr_heap_get_block_size(const void* ptr);
+
 #ifdef __cplusplus
 }
 #endif

furi/core/string.c

@@ -1,5 +1,6 @@
 #include "string.h"
 #include <m-string.h>
+#include <stdio.h>
 
 struct FuriString {
     string_t string;
@@ -175,11 +176,33 @@ int furi_string_cat_printf(FuriString* v, const char format[], ...) {
 }
 
 int furi_string_cat_vprintf(FuriString* v, const char format[], va_list args) {
-    FuriString* string = furi_string_alloc();
-    int ret = furi_string_vprintf(string, format, args);
-    furi_string_cat(v, string);
-    furi_string_free(string);
-    return ret;
+    // In-place append: format directly into destination buffer at current offset
+    // Eliminates temporary string allocation (malloc + free) per call
+    va_list args_copy;
+    va_copy(args_copy, args);
+
+    size_t old_size = string_size(v->string);
+    char* ptr = m_str1ng_get_cstr(v->string);
+    size_t alloc = string_capacity(v->string);
+
+    int size = vsnprintf(&ptr[old_size], alloc - old_size, format, args);
+
+    if(size > 0 && (old_size + (size_t)size + 1 > alloc)) {
+        // Buffer too small — grow and retry
+        ptr = m_str1ng_fit2size(v->string, old_size + (size_t)size + 1);
+        size = vsnprintf(
+            &ptr[old_size], string_capacity(v->string) - old_size, format, args_copy);
+    }
+    va_end(args_copy);
+
+    if(size >= 0) {
+        m_str1ng_set_size(v->string, old_size + (size_t)size);
+    } else {
+        // vsnprintf error — restore original string
+        ptr[old_size] = 0;
+    }
+
+    return size;
 }
 
 bool furi_string_empty(const FuriString* v) {

furi/core/thread.c

@@ -458,11 +458,11 @@ int32_t furi_thread_get_return_code(FuriThread* thread) {
     return thread->ret;
 }
 
-FuriThreadId furi_thread_get_current_id(void) {
+__attribute__((flatten)) FuriThreadId furi_thread_get_current_id(void) {
     return (FuriThreadId)xTaskGetCurrentTaskHandle();
 }
 
-FuriThread* furi_thread_get_current(void) {
+__attribute__((flatten)) FuriThread* furi_thread_get_current(void) {
     FuriThread* thread = pvTaskGetThreadLocalStoragePointer(NULL, 0);
     return thread;
 }
@@ -538,7 +538,7 @@ uint32_t furi_thread_flags_clear(uint32_t flags) {
     return rflags;
 }
 
-uint32_t furi_thread_flags_get(void) {
+__attribute__((flatten)) uint32_t furi_thread_flags_get(void) {
     TaskHandle_t hTask;
     uint32_t rflags;
 

site_scons/firmwareopts.scons

@@ -28,7 +28,7 @@ else:
             "NDEBUG",
         ],
         CCFLAGS=[
-            "-Og",
+            "-Os",
         ],
     )
 

targets/f7/api_symbols.csv

@@ -1785,7 +1785,7 @@ Function,+,furi_hal_spi_bus_handle_init,void,const FuriHalSpiBusHandle*
 Function,+,furi_hal_spi_bus_init,void,FuriHalSpiBus*
 Function,+,furi_hal_spi_bus_rx,_Bool,"const FuriHalSpiBusHandle*, uint8_t*, size_t, uint32_t"
 Function,+,furi_hal_spi_bus_trx,_Bool,"const FuriHalSpiBusHandle*, const uint8_t*, uint8_t*, size_t, uint32_t"
-Function,+,furi_hal_spi_bus_trx_dma,_Bool,"const FuriHalSpiBusHandle*, uint8_t*, uint8_t*, size_t, uint32_t"
+Function,+,furi_hal_spi_bus_trx_dma,_Bool,"const FuriHalSpiBusHandle*, const uint8_t*, uint8_t*, size_t, uint32_t"
 Function,+,furi_hal_spi_bus_tx,_Bool,"const FuriHalSpiBusHandle*, const uint8_t*, size_t, uint32_t"
 Function,-,furi_hal_spi_config_deinit_early,void,
 Function,-,furi_hal_spi_config_init,void,
@@ -2711,6 +2711,7 @@ Function,+,memmgr_get_minimum_free_heap,size_t,
 Function,+,memmgr_get_total_heap,size_t,
 Function,+,memmgr_heap_disable_thread_trace,void,FuriThreadId
 Function,+,memmgr_heap_enable_thread_trace,void,FuriThreadId
+Function,+,memmgr_heap_get_block_size,size_t,const void*
 Function,+,memmgr_heap_get_max_free_block,size_t,
 Function,+,memmgr_heap_get_thread_memory,size_t,FuriThreadId
 Function,+,memmgr_heap_printf_free_blocks,void,

targets/f7/furi_hal/furi_hal_spi.c

@@ -111,6 +111,14 @@ bool furi_hal_spi_bus_tx(
     furi_check(buffer);
     furi_check(size > 0);
 
+    if(furi_kernel_is_running()) {
+        // Use DMA for TX when scheduler is running, freeing the CPU during transfer
+        bool ret = furi_hal_spi_bus_trx_dma(handle, buffer, NULL, size, timeout);
+        LL_SPI_ClearFlag_OVR(handle->bus->spi);
+        return ret;
+    }
+
+    // Polling fallback for pre-scheduler context
     bool ret = true;
 
     while(size > 0) {
@@ -193,7 +201,7 @@ static void spi_dma_isr(void* context) {
 
 bool furi_hal_spi_bus_trx_dma(
     const FuriHalSpiBusHandle* handle,
-    uint8_t* tx_buffer,
+    const uint8_t* tx_buffer,
     uint8_t* rx_buffer,
     size_t size,
     uint32_t timeout_ms) {
@@ -203,6 +211,9 @@ bool furi_hal_spi_bus_trx_dma(
 
     // If scheduler is not running, use blocking mode
     if(!furi_kernel_is_running()) {
+        if(rx_buffer == NULL) {
+            return furi_hal_spi_bus_tx(handle, tx_buffer, size, timeout_ms);
+        }
         return furi_hal_spi_bus_trx(handle, tx_buffer, rx_buffer, size, timeout_ms);
     }
 
@@ -227,9 +238,26 @@ bool furi_hal_spi_bus_trx_dma(
     }
 
     if(rx_buffer == NULL) {
-        // Only TX mode, do not use RX channel
+        // TX-only mode: set up RX DMA to drain incoming bytes (prevents OVR)
+        uint8_t dma_rx_dummy;
 
         LL_DMA_InitTypeDef dma_config = {0};
+
+        // RX DMA channel: drain SPI RX FIFO into dummy byte (no increment)
+        dma_config.PeriphOrM2MSrcAddress = (uint32_t) & (spi->DR);
+        dma_config.MemoryOrM2MDstAddress = (uint32_t)&dma_rx_dummy;
+        dma_config.Direction = LL_DMA_DIRECTION_PERIPH_TO_MEMORY;
+        dma_config.Mode = LL_DMA_MODE_NORMAL;
+        dma_config.PeriphOrM2MSrcIncMode = LL_DMA_PERIPH_NOINCREMENT;
+        dma_config.MemoryOrM2MDstIncMode = LL_DMA_MEMORY_NOINCREMENT;
+        dma_config.PeriphOrM2MSrcDataSize = LL_DMA_PDATAALIGN_BYTE;
+        dma_config.MemoryOrM2MDstDataSize = LL_DMA_MDATAALIGN_BYTE;
+        dma_config.NbData = size;
+        dma_config.PeriphRequest = dma_rx_req;
+        dma_config.Priority = LL_DMA_PRIORITY_MEDIUM;
+        LL_DMA_Init(SPI_DMA_RX_DEF, &dma_config);
+
+        // TX DMA channel
         dma_config.PeriphOrM2MSrcAddress = (uint32_t) & (spi->DR);
         dma_config.MemoryOrM2MDstAddress = (uint32_t)tx_buffer;
         dma_config.Direction = LL_DMA_DIRECTION_MEMORY_TO_PERIPH;
@@ -243,48 +271,70 @@ bool furi_hal_spi_bus_trx_dma(
         dma_config.Priority = LL_DMA_PRIORITY_MEDIUM;
         LL_DMA_Init(SPI_DMA_TX_DEF, &dma_config);
 
-#if SPI_DMA_TX_CHANNEL == LL_DMA_CHANNEL_7
+#if SPI_DMA_RX_CHANNEL == LL_DMA_CHANNEL_6 && SPI_DMA_TX_CHANNEL == LL_DMA_CHANNEL_7
+        LL_DMA_ClearFlag_TC6(SPI_DMA);
         LL_DMA_ClearFlag_TC7(SPI_DMA);
 #else
 #error Update this code. Would you kindly?
 #endif
 
-        furi_hal_interrupt_set_isr(SPI_DMA_TX_IRQ, spi_dma_isr, NULL);
+        furi_hal_interrupt_set_isr(SPI_DMA_RX_IRQ, spi_dma_isr, NULL);
 
         bool dma_tx_was_enabled = LL_SPI_IsEnabledDMAReq_TX(spi);
+        bool dma_rx_was_enabled = LL_SPI_IsEnabledDMAReq_RX(spi);
         if(!dma_tx_was_enabled) {
             LL_SPI_EnableDMAReq_TX(spi);
         }
+        if(!dma_rx_was_enabled) {
+            LL_SPI_EnableDMAReq_RX(spi);
+        }
 
         // acquire semaphore before enabling DMA
         furi_check(furi_semaphore_acquire(spi_dma_completed, timeout_ms) == FuriStatusOk);
 
-        LL_DMA_EnableIT_TC(SPI_DMA_TX_DEF);
+        LL_DMA_EnableIT_TC(SPI_DMA_RX_DEF);
+        LL_DMA_EnableChannel(SPI_DMA_RX_DEF);
         LL_DMA_EnableChannel(SPI_DMA_TX_DEF);
 
-        // and wait for it to be released (DMA transfer complete)
+        // wait for RX DMA complete (all bytes transmitted and drained)
         if(furi_semaphore_acquire(spi_dma_completed, timeout_ms) != FuriStatusOk) {
             ret = false;
             FURI_LOG_E(TAG, "DMA timeout\r\n");
         }
+
+        // Disable TC IRQ and clear pending TC flag BEFORE releasing the
+        // semaphore. On timeout the ISR may still fire and would try to
+        // re-release the binary semaphore, crashing furi_check.
+        LL_DMA_DisableIT_TC(SPI_DMA_RX_DEF);
+#if SPI_DMA_RX_CHANNEL == LL_DMA_CHANNEL_6 && SPI_DMA_TX_CHANNEL == LL_DMA_CHANNEL_7
+        LL_DMA_ClearFlag_TC6(SPI_DMA);
+        LL_DMA_ClearFlag_TC7(SPI_DMA);
+#else
+#error Update this code. Would you kindly?
+#endif
+
         // release semaphore, because we are using it as a flag
         furi_semaphore_release(spi_dma_completed);
 
-        LL_DMA_DisableIT_TC(SPI_DMA_TX_DEF);
         LL_DMA_DisableChannel(SPI_DMA_TX_DEF);
+        LL_DMA_DisableChannel(SPI_DMA_RX_DEF);
         if(!dma_tx_was_enabled) {
             LL_SPI_DisableDMAReq_TX(spi);
         }
-        furi_hal_interrupt_set_isr(SPI_DMA_TX_IRQ, NULL, NULL);
+        if(!dma_rx_was_enabled) {
+            LL_SPI_DisableDMAReq_RX(spi);
+        }
+        furi_hal_interrupt_set_isr(SPI_DMA_RX_IRQ, NULL, NULL);
 
         LL_DMA_DeInit(SPI_DMA_TX_DEF);
+        LL_DMA_DeInit(SPI_DMA_RX_DEF);
     } else {
         // TRX or RX mode, use both channels
         uint32_t tx_mem_increase_mode;
 
         if(tx_buffer == NULL) {
             // RX mode, use dummy data instead of TX buffer
-            tx_buffer = (uint8_t*)&dma_dummy_u32;
+            tx_buffer = (const uint8_t*)&dma_dummy_u32;
             tx_mem_increase_mode = LL_DMA_MEMORY_NOINCREMENT;
         } else {
             tx_mem_increase_mode = LL_DMA_MEMORY_INCREMENT;
@@ -317,8 +367,9 @@ bool furi_hal_spi_bus_trx_dma(
         dma_config.Priority = LL_DMA_PRIORITY_MEDIUM;
         LL_DMA_Init(SPI_DMA_RX_DEF, &dma_config);
 
-#if SPI_DMA_RX_CHANNEL == LL_DMA_CHANNEL_6
+#if SPI_DMA_RX_CHANNEL == LL_DMA_CHANNEL_6 && SPI_DMA_TX_CHANNEL == LL_DMA_CHANNEL_7
         LL_DMA_ClearFlag_TC6(SPI_DMA);
+        LL_DMA_ClearFlag_TC7(SPI_DMA);
 #else
 #error Update this code. Would you kindly?
 #endif
@@ -348,10 +399,20 @@ bool furi_hal_spi_bus_trx_dma(
             ret = false;
             FURI_LOG_E(TAG, "DMA timeout\r\n");
         }
-        // release semaphore, because we are using it as a flag
-        furi_semaphore_release(spi_dma_completed);
 
+        // Disable TC IRQ and clear pending TC flag BEFORE releasing the
+        // semaphore. On timeout the ISR may still fire and would try to
+        // re-release the binary semaphore, crashing furi_check.
         LL_DMA_DisableIT_TC(SPI_DMA_RX_DEF);
+#if SPI_DMA_RX_CHANNEL == LL_DMA_CHANNEL_6 && SPI_DMA_TX_CHANNEL == LL_DMA_CHANNEL_7
+        LL_DMA_ClearFlag_TC6(SPI_DMA);
+        LL_DMA_ClearFlag_TC7(SPI_DMA);
+#else
+#error Update this code. Would you kindly?
+#endif
+
+        // release semaphore, because we are using it as a flag
+        furi_semaphore_release(spi_dma_completed);
 
         LL_DMA_DisableChannel(SPI_DMA_TX_DEF);
         LL_DMA_DisableChannel(SPI_DMA_RX_DEF);