diff --git a/arch/arm/core/offsets/offsets.c b/arch/arm/core/offsets/offsets.c
index 69b7471f468..b91a5c95cfc 100644
--- a/arch/arm/core/offsets/offsets.c
+++ b/arch/arm/core/offsets/offsets.c
@@ -38,6 +38,7 @@
 #include <kernel_offsets.h>
 
 GEN_OFFSET_SYM(_thread_arch_t, basepri);
+GEN_OFFSET_SYM(_thread_arch_t, swap_return_value);
 
 #ifdef CONFIG_FLOAT
 GEN_OFFSET_SYM(_thread_arch_t, preempt_float);
diff --git a/arch/arm/core/swap.S b/arch/arm/core/swap.S
index b1f0aecfe0b..d7648742f67 100644
--- a/arch/arm/core/swap.S
+++ b/arch/arm/core/swap.S
@@ -234,16 +234,6 @@ SECTION_FUNC(TEXT, __svc)
 _context_switch:
 #endif
 
-    /*
-     * Set _Swap()'s default return code to -EAGAIN. This eliminates the
-     * need for the timeout code to invoke fiberRtnValueSet().
-     */
-
-    mrs  r2, PSP   /* thread mode, stack frame is on PSP */
-    ldr  r3, =_k_neg_eagain
-    ldr  r3, [r3, #0]
-    str  r3, [r2, #___esf_t_a1_OFFSET]
-
     /*
      * Unlock interrupts:
      * - in a SVC call, so protected against context switches
@@ -305,17 +295,21 @@ SECTION_FUNC(TEXT, _Swap)
     ldr r2, [r1, #_kernel_offset_to_current]
     str r0, [r2, #_thread_offset_to_basepri]
 
+    /*
+     * Set _Swap()'s default return code to -EAGAIN. This eliminates the need
+     * for the timeout code to set it itself.
+     */
+    ldr r1, =_k_neg_eagain
+    ldr r1, [r1]
+    str r1, [r2, #_thread_offset_to_swap_return_value]
+
 #if defined(CONFIG_CPU_CORTEX_M0_M0PLUS)
     /* No priority-based interrupt masking on M0/M0+,
      * pending PendSV is used instead of svc
      */
     ldr r1, =_SCS_ICSR
-    ldr r2, =_SCS_ICSR_PENDSV
-    str r2, [r1, #0]
-
-    /* load -EAGAIN as the default return value */
-    ldr r0, =_k_neg_eagain
-    ldr r0, [r0]
+    ldr r3, =_SCS_ICSR_PENDSV
+    str r3, [r1, #0]
 
     /* Unlock interrupts to allow PendSV, since it's running at prio 0xff
      *
@@ -323,12 +317,10 @@ SECTION_FUNC(TEXT, _Swap)
      * of a higher priority pending.
      */
     cpsie i
-
-    /* PC stored in stack frame by the hw */
-    bx lr
 #else /* CONFIG_CPU_CORTEX_M3_M4 */
     svc #0
-
-    /* r0 contains the return value if needed */
-    bx lr
 #endif
+
+    /* coming back from exception, r2 still holds the pointer to _current */
+    ldr r0, [r2, #_thread_offset_to_swap_return_value]
+    bx lr
diff --git a/arch/arm/core/thread.c b/arch/arm/core/thread.c
index 485897bdb6f..6d56edfcd63 100644
--- a/arch/arm/core/thread.c
+++ b/arch/arm/core/thread.c
@@ -140,6 +140,8 @@ void _new_thread(char *pStackMem, unsigned stackSize,
 	tcs->callee_saved.psp = (uint32_t)pInitCtx;
 	tcs->arch.basepri = 0;
 
+	/* swap_return_value can contain garbage */
+
 	_nano_timeout_thread_init(tcs);
 
 	/* initial values in all other registers/TCS entries are irrelevant */
diff --git a/arch/arm/include/kernel_arch_data.h b/arch/arm/include/kernel_arch_data.h
index 3fe0b868383..967092ff1a4 100644
--- a/arch/arm/include/kernel_arch_data.h
+++ b/arch/arm/include/kernel_arch_data.h
@@ -142,6 +142,9 @@ struct _thread_arch {
 	/* interrupt locking key */
 	uint32_t basepri;
 
+	/* r0 in stack frame cannot be written to reliably */
+	uint32_t swap_return_value;
+
 #ifdef CONFIG_FLOAT
 	/*
 	 * No cooperative floating point register set structure exists for
diff --git a/arch/arm/include/kernel_arch_func.h b/arch/arm/include/kernel_arch_func.h
index f298a0cdacd..4a254174b6c 100644
--- a/arch/arm/include/kernel_arch_func.h
+++ b/arch/arm/include/kernel_arch_func.h
@@ -47,25 +47,10 @@ static ALWAYS_INLINE void nanoArchInit(void)
 	_CpuIdleInit();
 }
 
-/**
- *
- * @brief Set the return value for the specified fiber (inline)
- *
- * The register used to store the return value from a function call invocation
- * to <value>.  It is assumed that the specified <fiber> is pending, and thus
- * the fiber's thread is stored in its struct tcs structure.
- *
- * @param fiber pointer to the fiber
- * @param value is the value to set as a return value
- *
- * @return N/A
- */
 static ALWAYS_INLINE void
 _set_thread_return_value(struct k_thread *thread, unsigned int value)
 {
-	struct __esf *esf = (struct __esf *)thread->callee_saved.psp;
-
-	esf->a1 = value;
+	thread->arch.swap_return_value = value;
 }
 
 extern void nano_cpu_atomic_idle(unsigned int);
diff --git a/arch/arm/include/offsets_short_arch.h b/arch/arm/include/offsets_short_arch.h
index a568737247a..983d0693805 100644
--- a/arch/arm/include/offsets_short_arch.h
+++ b/arch/arm/include/offsets_short_arch.h
@@ -30,6 +30,9 @@
 #define _thread_offset_to_basepri \
 	(___thread_t_arch_OFFSET + ___thread_arch_t_basepri_OFFSET)
 
+#define _thread_offset_to_swap_return_value \
+	(___thread_t_arch_OFFSET + ___thread_arch_t_swap_return_value_OFFSET)
+
 #define _thread_offset_to_preempt_float \
 	(___thread_t_arch_OFFSET + ___thread_arch_t_preempt_float_OFFSET)