From fe68ff121910bf15097abae2360a9d6baffabc3e Mon Sep 17 00:00:00 2001
From: aineoae86-sys <ai.neo.ae86@gmail.com>
Date: Sat, 4 Jul 2026 15:45:37 +0800
Subject: [PATCH] Fix SH2A_FPU FPU context cleanup

---
 portable/Renesas/SH2A_FPU/port.c      | 46 ++++++++++++++-------------
 portable/Renesas/SH2A_FPU/portmacro.h | 22 +++++++++++++
 2 files changed, 46 insertions(+), 22 deletions(-)

diff --git a/portable/Renesas/SH2A_FPU/port.c b/portable/Renesas/SH2A_FPU/port.c
index c14a0a851c3..1c66691ecee 100644
--- a/portable/Renesas/SH2A_FPU/port.c
+++ b/portable/Renesas/SH2A_FPU/port.c
@@ -43,13 +43,6 @@
  * value is for all interrupts to be enabled. */
 #define portINITIAL_SR                 ( 0UL )
 
-/* Dimensions the array into which the floating point context is saved.
- * Allocate enough space for FPR0 to FPR15, FPUL and FPSCR, each of which is 4
- * bytes big.  If this number is changed then the 72 in portasm.src also needs
- * changing. */
-#define portFLOP_REGISTERS_TO_STORE    ( 18 )
-#define portFLOP_STORAGE_SIZE          ( portFLOP_REGISTERS_TO_STORE * 4 )
-
 #if ( configSUPPORT_DYNAMIC_ALLOCATION == 0 )
     #error configSUPPORT_DYNAMIC_ALLOCATION must be 1 to use this port.
 #endif
@@ -245,26 +238,35 @@ BaseType_t xPortUsesFloatingPoint( TaskHandle_t xTask )
         xTask = ( TaskHandle_t ) pxCurrentTCB;
     }
 
-    /* Allocate a buffer large enough to hold all the flop registers. */
-    pulFlopBuffer = ( uint32_t * ) pvPortMalloc( portFLOP_STORAGE_SIZE );
-
-    if( pulFlopBuffer != NULL )
+    /* The task tag already owns the FPU buffer for this port.  Do not replace
+     * it, or the original allocation would become unreachable. */
+    if( xTaskGetApplicationTaskTag( xTask ) != NULL )
     {
-        /* Start with the registers in a benign state. */
-        memset( ( void * ) pulFlopBuffer, 0x00, portFLOP_STORAGE_SIZE );
-
-        /* The first thing to get saved in the buffer is the FPSCR value -
-         * initialise this to the current FPSCR value. */
-        *pulFlopBuffer = get_fpscr();
-
-        /* Use the task tag to point to the flop buffer.  Pass pointer to just
-         * above the buffer because the flop save routine uses a pre-decrement. */
-        vTaskSetApplicationTaskTag( xTask, ( void * ) ( pulFlopBuffer + portFLOP_REGISTERS_TO_STORE ) );
         xReturn = pdPASS;
     }
     else
     {
-        xReturn = pdFAIL;
+        /* Allocate a buffer large enough to hold all the flop registers. */
+        pulFlopBuffer = ( uint32_t * ) pvPortMalloc( portFLOP_STORAGE_SIZE );
+
+        if( pulFlopBuffer != NULL )
+        {
+            /* Start with the registers in a benign state. */
+            memset( ( void * ) pulFlopBuffer, 0x00, portFLOP_STORAGE_SIZE );
+
+            /* The first thing to get saved in the buffer is the FPSCR value -
+             * initialise this to the current FPSCR value. */
+            *pulFlopBuffer = get_fpscr();
+
+            /* Use the task tag to point to the flop buffer.  Pass pointer to just
+             * above the buffer because the flop save routine uses a pre-decrement. */
+            vTaskSetApplicationTaskTag( xTask, ( void * ) ( pulFlopBuffer + portFLOP_REGISTERS_TO_STORE ) );
+            xReturn = pdPASS;
+        }
+        else
+        {
+            xReturn = pdFAIL;
+        }
     }
 
     return xReturn;
diff --git a/portable/Renesas/SH2A_FPU/portmacro.h b/portable/Renesas/SH2A_FPU/portmacro.h
index 2229fd3f66b..b9d751b4b70 100644
--- a/portable/Renesas/SH2A_FPU/portmacro.h
+++ b/portable/Renesas/SH2A_FPU/portmacro.h
@@ -84,6 +84,13 @@ typedef unsigned long    UBaseType_t;
 #define portYIELD_TRAP_NO                ( 33 )
 #define portKERNEL_INTERRUPT_PRIORITY    ( 1 )
 
+/* Dimensions the array into which the floating point context is saved.
+ * Allocate enough space for FPR0 to FPR15, FPUL and FPSCR, each of which is 4
+ * bytes big.  If this number is changed then the 72 in portasm.src also needs
+ * changing. */
+#define portFLOP_REGISTERS_TO_STORE      ( 18 )
+#define portFLOP_STORAGE_SIZE            ( portFLOP_REGISTERS_TO_STORE * 4 )
+
 void vPortYield( void );
 #define portYIELD()                vPortYield()
 
@@ -112,6 +119,21 @@ void vPortRestoreFlopRegisters( void * pulBuffer );
 #define traceTASK_SWITCHED_OUT()    do { if( pxCurrentTCB->pxTaskTag != NULL ) vPortSaveFlopRegisters( pxCurrentTCB->pxTaskTag ); } while( 0 )
 #define traceTASK_SWITCHED_IN()     do { if( pxCurrentTCB->pxTaskTag != NULL ) vPortRestoreFlopRegisters( pxCurrentTCB->pxTaskTag ); } while( 0 )
 
+/* pxTaskTag points just above the FPU context buffer because the save routine
+ * uses a pre-decrement.  Recover the allocation base before freeing it. */
+#define portCLEAN_UP_TCB( pxTCB )                                         \
+    do                                                                    \
+    {                                                                     \
+        if( ( pxTCB )->pxTaskTag != NULL )                                \
+        {                                                                 \
+            uint32_t * pulFlopBufferEnd =                                 \
+                ( uint32_t * ) ( pxTCB )->pxTaskTag;                      \
+            vPortFree( ( void * ) ( pulFlopBufferEnd -                    \
+                                    portFLOP_REGISTERS_TO_STORE ) );      \
+            ( pxTCB )->pxTaskTag = NULL;                                  \
+        }                                                                 \
+    } while( 0 )
+
 /*
  * These macros should be called directly, but through the taskENTER_CRITICAL()
  * and taskEXIT_CRITICAL() macros.