Some more optimizations, mostly in involuntary task switches. Doesn not really help...

diff --git a/components/freertos/portasm.S b/components/freertos/portasm.S
index 3a45b19a326339f18624deb612f5b15e799c3a84..ad65a103ed626e53440cf807f33fca4860746901 100644 (file)
--- a/components/freertos/portasm.S
+++ b/components/freertos/portasm.S
@@ -69,9 +69,8 @@ _frxt_setup_switch:
     ENTRY(16)
 
        getcoreid a3
-       slli    a3,  a3, 2
     movi    a2, port_switch_flag
-       add             a2,  a2, a3
+       addx4   a2,  a3, a2
 
     movi    a3, 1
     s32i    a3, a2, 0
@@ -116,12 +115,11 @@ _frxt_int_enter:
     Manage nesting directly rather than call the generic IntEnter()
     (in windowed ABI we can't call a C function here anyway because PS.EXCM is still set).
     */
-       getcoreid a3
-       slli    a4,  a3, 2                                      /* a4 = cpuid * 4 */
+       getcoreid a4
     movi    a2,  port_xSchedulerRunning
-       add             a2,  a2, a4
+       addx4   a2,  a4, a2
     movi    a3,  port_interruptNesting
-       add             a3,  a3, a4
+       addx4   a3,  a4, a3
     l32i    a2,  a2, 0                  /* a2 = port_xSchedulerRunning     */
     beqz    a2,  1f                     /* scheduler not running, no tasks */
     l32i    a2,  a3, 0                  /* a2 = port_interruptNesting      */
@@ -130,14 +128,13 @@ _frxt_int_enter:
     bnei    a2,  1, .Lnested            /* !=0 before incr, so nested      */
 
     movi    a2,  pxCurrentTCB
-       add             a2,  a2, a4
+       addx4   a2,  a4, a2
     l32i    a2,  a2, 0                  /* a2 = current TCB                */
     beqz    a2,  1f
     s32i    a1,  a2, TOPOFSTACK_OFFS    /* pxCurrentTCB->pxTopOfStack = SP */
     movi    a1,  port_IntStackTop       /* a1 = top of intr stack          */
        movi    a2,  configISR_STACK_SIZE
-       getcoreid a3
-       mull    a2,  a3, a2
+       mull    a2,  a4, a2
        add     a1,  a1, a2                                     /* for current proc */
 
 .Lnested:
@@ -165,12 +162,11 @@ _frxt_int_enter:
     .align  4
 _frxt_int_exit:
 
-       getcoreid a3
-       slli    a4,  a3, 2                                      /* a4 is core * 4 */
+       getcoreid a4
     movi    a2,  port_xSchedulerRunning
-       add             a2,  a2, a4
+       addx4   a2,  a4, a2
     movi    a3,  port_interruptNesting
-       add             a3,  a3, a4
+       addx4   a3,  a4, a3
     rsil    a0,  XCHAL_EXCM_LEVEL       /* lock out interrupts             */
     l32i    a2,  a2, 0                  /* a2 = port_xSchedulerRunning     */
     beqz    a2,  .Lnoswitch             /* scheduler not running, no tasks */
@@ -180,13 +176,13 @@ _frxt_int_exit:
     bnez    a2,  .Lnesting              /* !=0 after decr so still nested  */
 
     movi    a2,  pxCurrentTCB
-       add             a2,  a2, a4
+       addx4   a2,  a4, a2
     l32i    a2,  a2, 0                  /* a2 = current TCB                */
     beqz    a2,  1f                     /* no task ? go to dispatcher      */
     l32i    a1,  a2, TOPOFSTACK_OFFS    /* SP = pxCurrentTCB->pxTopOfStack */
 
     movi    a2,  port_switch_flag       /* address of switch flag          */
-       add             a2,  a2, a4                                     /* point to flag for this cpu      */
+       addx4   a2,  a4, a2                                     /* point to flag for this cpu      */
     l32i    a3,  a2, 0                  /* a3 = port_switch_flag           */
     beqz    a3,  .Lnoswitch             /* flag = 0 means no switch reqd   */
     movi    a3,  0
@@ -392,14 +388,12 @@ _frxt_dispatch:
     call0   vTaskSwitchContext  // Get next TCB to resume
     movi    a2, pxCurrentTCB
        getcoreid a3
-       slli    a3,  a3, 2
-       add             a2,  a2, a3
+       addx4   a2,  a3, a2
     #else
     call4   vTaskSwitchContext  // Get next TCB to resume
     movi    a2, pxCurrentTCB
        getcoreid a3
-       slli    a3,  a3, 2
-       add             a2,  a2, a3
+       addx4   a2,  a3, a2
     #endif
     l32i    a3,  a2, 0
     l32i    sp,  a3, TOPOFSTACK_OFFS     /* SP = next_TCB->pxTopOfStack;  */
@@ -439,8 +433,7 @@ _frxt_dispatch:
     /* Restore CPENABLE from task's co-processor save area. */
     movi    a3, pxCurrentTCB            /* cp_state =                       */
        getcoreid a2
-       slli    a2,  a2, 2
-       add             a3,  a2, a3
+       addx4   a3,  a2, a3
     l32i    a3, a3, 0
     l32i    a2, a3, CP_TOPOFSTACK_OFFS     /* StackType_t                       *pxStack; */
     l16ui   a3, a2, XT_CPENABLE         /* CPENABLE = cp_state->cpenable;   */
@@ -529,8 +522,7 @@ vPortYield:
 
     movi    a2,  pxCurrentTCB
        getcoreid a3
-       slli    a3,  a3, 2
-       add             a2,  a2, a3
+       addx4   a2,  a3, a2
     l32i    a2,  a2, 0                  /* a2 = pxCurrentTCB                */
     movi    a3,  0
     s32i    a3,  sp, XT_SOL_EXIT        /* 0 to flag as solicited frame     */
@@ -581,8 +573,7 @@ vPortYieldFromInt:
     /* Save CPENABLE in task's co-processor save area, and clear CPENABLE.  */
     movi    a3, pxCurrentTCB            /* cp_state =                       */
        getcoreid a2
-       slli    a2,  a2, 2
-       add             a3,  a2, a3
+       addx4   a3,  a2, a3
     l32i    a3, a3, 0
 
     l32i    a2, a3, CP_TOPOFSTACK_OFFS
@@ -625,18 +616,17 @@ _frxt_task_coproc_state:
 
        /* We can use a3 as a scratchpad, the instances of code calling XT_RTOS_CP_STATE don't seem to need it saved. */
        getcoreid a3
-       slli    a3,  a3, 2                      /* a3=coreid*4 */
     movi    a15, port_xSchedulerRunning /* if (port_xSchedulerRunning              */
-       add             a15, a15, a3
+       addx4   a15, a3,a15
     l32i    a15, a15, 0
     beqz    a15, 1f
     movi    a15, port_interruptNesting  /* && port_interruptNesting == 0           */
-       add             a15, a15, a3
+       addx4   a15, a3, a15
     l32i    a15, a15, 0
     bnez    a15, 1f
 
     movi    a15, pxCurrentTCB
-       add             a15,  a3, a15
+       addx4   a15, a3, a15
     l32i    a15, a15, 0                 /* && pxCurrentTCB != 0) {                 */
 
 

diff --git a/components/freertos/xtensa_vectors.S b/components/freertos/xtensa_vectors.S
index f0d874a59c2ba7e99c07025861c0ac0c9a024617..7c2fc29607e359db5b25bacb1b703827a3ea183b 100644 (file)
--- a/components/freertos/xtensa_vectors.S
+++ b/components/freertos/xtensa_vectors.S
@@ -904,16 +904,13 @@ _xt_coproc_exc:
        core we're running on now. */
     movi    a2, pxCurrentTCB
     getcoreid a3
-    slli    a3,  a3, 2
-    add     a2,  a2, a3
+    addx4     a2,  a3, a2
     l32i    a2, a2, 0                       /* a2 = start of pxCurrentTCB[cpuid] */
     addi    a2, a2, TASKTCB_XCOREID_OFFSET  /* offset to xCoreID in tcb struct */
-    getcoreid a3
     s32i    a3, a2, 0                       /* store current cpuid */
 
        /* Grab correct xt_coproc_owner_sa for this core */
-       getcoreid a2
-       movi    a3, XCHAL_CP_MAX << 2
+       movi    a2, XCHAL_CP_MAX << 2
        mull    a2, a2, a3
     movi    a3, _xt_coproc_owner_sa         /* a3 = base of owner array */
        add     a3, a3, a2