Conditionally save FPU context in IRQ, and preserve vApplicationFPUSafeIRQHandler logic

portable/GCC/ARM_CA9/port.c

@@ -168,7 +168,8 @@ static void prvTaskExitError( void );
 
 /*
  * If the application provides an implementation of vApplicationIRQHandler(),
- * then it will get called directly without saving the FPU registers on
+ * then it will get called directly without saving the FPU registers
+ * when there is no FPU context to preserve on
  * interrupt entry, and this weak implementation of
  * vApplicationFPUSafeIRQHandler() is just provided to remove linkage errors -
  * it should never actually get called so its implementation contains a
@@ -179,11 +180,16 @@ static void prvTaskExitError( void );
  * vApplicationIRQHandler() provided in portASM.S will save the FPU registers
  * before calling it.
  *
- * Therefore, if the application writer wants FPU registers to be saved on
+ * Therefore, if the application writer wants FPU registers to always be saved on
  * interrupt entry their IRQ handler must be called
  * vApplicationFPUSafeIRQHandler(), and if the application writer does not want
  * FPU registers to be saved on interrupt entry their IRQ handler must be
  * called vApplicationIRQHandler().
+ * 
+ * Note that for ARM Cortex A9, the FPU might be used for memory operations,
+ * depending on the C library used for memcpy, memcmp or memset. For instance,
+ * glibc uses the FPU, hence always saving it might be preferable when using it.
+ * See https://freertos.org/Using-FreeRTOS-on-Cortex-A-Embedded-Processors#important-note-for-gcc-and-possibly-other-compiler-users
  */
 void vApplicationFPUSafeIRQHandler( uint32_t ulICCIAR ) __attribute__( ( weak ) );
 

portable/GCC/ARM_CA9/portASM.S

@@ -197,6 +197,18 @@ FreeRTOS_IRQ_Handler:
     LDR     r2, [r2]
     LDR     r0, [r2]
 
+    /* Does the task have a floating point context that needs saving?  If
+    ulPortTaskHasFPUContext is 0 then no. */
+    LDR     r12, ulPortTaskHasFPUContextConst
+    LDR     r4, [r12]
+    CMP     r4, #0 /* r4 contains ulPortTaskHasFPUContext */
+
+    /* Save the floating point context, if any. */
+    FMRXNE  r12, FPSCR
+    VPUSHNE {d0-d15}
+    VPUSHNE {d16-d31}
+    PUSHNE  {r12} /* r12 contains FPSCR */
+
     /* Ensure bit 2 of the stack pointer is clear.  r2 holds the bit 2 value for
     future use.  _RB_ Does this ever actually need to be done provided the start
     of the stack is 8-byte aligned? */
@@ -204,7 +216,7 @@ FreeRTOS_IRQ_Handler:
     AND     r2, r2, #4
     SUB     sp, sp, r2
 
-    /* Call the interrupt handler.  r4 pushed to maintain alignment. */
+    /* Call the interrupt handler. */
     PUSH    {r0-r4, lr}
     LDR     r1, vApplicationIRQHandlerConst
     BLX     r1
@@ -215,6 +227,13 @@ FreeRTOS_IRQ_Handler:
     DSB
     ISB
 
+    /* Restore the floating point context, if any. */
+    CMP     r4, #0 /* r4 contains ulPortTaskHasFPUContext */
+    POPNE   {r12}  /* r12 contains FPSCR */
+    VPOPNE  {d16-d31}
+    VPOPNE  {d0-d15}
+    VMSRNE  FPSCR, r12
+
     /* Write the value read from ICCIAR to ICCEOIR. */
     LDR     r4, ulICCEOIRConst
     LDR     r4, [r4]
@@ -281,13 +300,13 @@ switch_before_exit:
 /******************************************************************************
  * If the application provides an implementation of vApplicationIRQHandler(),
  * then it will get called directly without saving the FPU registers on
- * interrupt entry, and this weak implementation of
+ * interrupt entry when the FPU is not used, and this weak implementation of
  * vApplicationIRQHandler() will not get called.
  *
  * If the application provides its own implementation of
  * vApplicationFPUSafeIRQHandler() then this implementation of
- * vApplicationIRQHandler() will be called, save the FPU registers, and then
- * call vApplicationFPUSafeIRQHandler().
+ * vApplicationIRQHandler() will be called, save the FPU registers if not done
+ * already, and then call vApplicationFPUSafeIRQHandler().
  *
  * Therefore, if the application writer wants FPU registers to be saved on
  * interrupt entry their IRQ handler must be called
@@ -301,18 +320,36 @@ switch_before_exit:
 .type vApplicationIRQHandler, %function
 vApplicationIRQHandler:
     PUSH    {LR}
-    FMRX    R1,  FPSCR
-    VPUSH   {D0-D7}
-    VPUSH   {D16-D31}
-    PUSH    {R1}
+
+    /* Does the task have a floating point context that has been saved already?  If
+    ulPortTaskHasFPUContext is 1 then yes. */
+    LDR     r12, ulPortTaskHasFPUContextConst
+    LDR     r4, [r12]
+
+    /* If there is an FPU context, it was already saved. */
+    CMP     r4, #1 /* If equal, the FPU context was saved by the caller. */
+    FMRXNE  r1,  FPSCR
+    VPUSHNE {d0-d7}
+    VPUSHNE {d16-d31}
+    PUSHNE  {r1}
+    PUSHNE  {r1} /* Push FPSCR twice to keep memory alignment */
+
+    /* Store ulPortTaskHasFPUContext */
+    PUSH    {r4}
 
     LDR     r1, vApplicationFPUSafeIRQHandlerConst
     BLX     r1
 
-    POP     {R0}
-    VPOP    {D16-D31}
-    VPOP    {D0-D7}
-    VMSR    FPSCR, R0
+    /* Restore ulPortTaskHasFPUContext */
+    POP     {r4}
+
+    /* Skip the FPU register reset if there is an FPU context, the IRQ handler will do it. */
+    CMP     r4, #1 /* If true, the FPU registers will be restored by the caller. */
+    POPNE   {r0}
+    POPNE   {r0} /* FPSCR was pushed twice for alignment purposes */
+    VPOPNE  {d16-d31}
+    VPOPNE  {d0-d7}
+    VMSRNE  FPSCR, r0
 
     POP {PC}