stk_arch_risc-v.cpp

Go to the documentation of this file.

/*
 * SuperTinyKernel(TM) RTOS: Lightweight High-Performance Deterministic C++ RTOS for Embedded Systems.
 *
 * Source: https://github.com/SuperTinyKernel-RTOS
 *
 * Copyright (c) 2022-2026 Neutron Code Limited <stk@neutroncode.com>. All Rights Reserved.
 * License: MIT License, see LICENSE for a full text.
 */
 
// note: If missing, this header must be customized (get it in the root of the source folder) and
//       copied to the /include folder manually.
#include "stk_config.h"
 
#ifdef _STK_ARCH_RISC_V
 
#include "stk_arch.h"
#include "arch/stk_arch_common.h"
#include "stk_helper.h"
 
using namespace stk;

//#define _STK_RISCV_USE_PENDSV
 
// CLINT
// Details: https://github.com/riscv/riscv-aclint/blob/main/riscv-aclint.adoc
#ifndef STK_RISCV_CLINT_BASE_ADDR
    #define STK_RISCV_CLINT_BASE_ADDR (0x2000000U)
#endif
#ifndef STK_RISCV_CLINT_MTIMECMP_ADDR
    #define STK_RISCV_CLINT_MTIMECMP_ADDR (STK_RISCV_CLINT_BASE_ADDR + 0x4000U) // 8-byte value, 1 per hart
#endif
#ifndef STK_RISCV_CLINT_MTIME_ADDR
    #define STK_RISCV_CLINT_MTIME_ADDR (STK_RISCV_CLINT_BASE_ADDR + 0xBFF8U) // 8-byte value, global
#endif

#define STK_RISCV_ISR_STACK_SIZE 256U

#ifndef STK_TIMER_CLOCK_FREQUENCY
    #define STK_TIMER_CLOCK_FREQUENCY 1000000U
#endif

#ifndef STK_RISCV_ISR_SECTION
    #define STK_RISCV_ISR_SECTION
#endif

#define STK_RISCV_ISR extern "C" STK_RISCV_ISR_SECTION __attribute__ ((interrupt("machine")))

#define STK_ASM_EXIT_FROM_HANDLER "mret"

#ifndef STK_RISCV_CLINT_MTIMECMP_PER_HART
    #define STK_RISCV_CLINT_MTIMECMP_PER_HART (1)
#endif

#ifndef STK_ARCH_GET_CPU_ID
    #define STK_ARCH_GET_CPU_ID() read_csr(mhartid)
#endif

#if (__riscv_flen == 0)
    #define STK_RISCV_FPU 0
#else
    #define STK_RISCV_FPU __riscv_flen
#endif
 
#define STR(x) #x
#define XSTR(s) STR(s)

#if (__riscv_xlen == 32)
    #define REGBYTES XSTR(4)
    #define LREG     XSTR(lw)
    #define SREG     XSTR(sw)
#elif (__riscv_xlen == 64)
    #define REGBYTES XSTR(8)
    #define LREG     XSTR(ld)
    #define SREG     XSTR(sd)
#else
    #error Unsupported RISC-V platform!
#endif
 
#if (STK_RISCV_FPU == 32)
    #define FREGBYTES XSTR(4)
    #define FLREG     XSTR(flw)
    #define FSREG     XSTR(fsw)
#elif (STK_RISCV_FPU == 64)
    #define FREGBYTES XSTR(8)
    #define FLREG     XSTR(fld)
    #define FSREG     XSTR(fsd)
#elif (STK_RISCV_FPU != 0)
#error Unsupported FP register count!
#endif
 
 
#if (__riscv_32e == 1)
    #define STK_RISCV_REGISTER_COUNT (15 + (STK_RISCV_FPU != 0 ? 31 : 0))
#else
    #define STK_RISCV_REGISTER_COUNT (31 + (STK_RISCV_FPU != 0 ? 31 : 0))
#endif
 
#define STK_SERVICE_SLOTS 2 // (0) mepc, (1) mstatus
 
#if (__riscv_32e == 1)
    #define FOFFSET XSTR(68) // FP stack offset = (17 * 4)
    #if (STK_RISCV_FPU == 0)
        #define REGSIZE XSTR(((15 + STK_SERVICE_SLOTS) * 4)) // STK_RISCV_REGISTER_COUNT + 2 for mepc, mstatus
    #else
        #if (STK_RISCV_FPU == 32)
        #define REGSIZE XSTR((((15 + STK_SERVICE_SLOTS) * 4) + (31 * 4))) // STK_RISCV_REGISTER_COUNT + 2 for mepc, mstatus + 32 fp registers
        #elif (STK_RISCV_FPU == 64)
        #define REGSIZE XSTR((((15 + STK_SERVICE_SLOTS) * 4) + (31 * 8))) // STK_RISCV_REGISTER_COUNT + 2 for mepc, mstatus + 32 fp registers
        #endif
    #endif
#elif (__riscv_xlen == 32)
    #define FOFFSET XSTR(132) // FP stack offset = (33 * 4)
    #if (STK_RISCV_FPU == 0)
        #define REGSIZE XSTR(((31 + STK_SERVICE_SLOTS) * 4)) // STK_RISCV_REGISTER_COUNT + 2 for mepc, mstatus
    #else
        #if (STK_RISCV_FPU == 32)
        #define REGSIZE XSTR((((31 + STK_SERVICE_SLOTS) * 4) + (31 * 4))) // STK_RISCV_REGISTER_COUNT + 2 for mepc, mstatus + 32 fp registers
        #elif (STK_RISCV_FPU == 64)
        #define REGSIZE XSTR((((31 + STK_SERVICE_SLOTS) * 4) + (31 * 8))) // STK_RISCV_REGISTER_COUNT + 2 for mepc, mstatus + 32 fp registers
        #endif
    #endif
#elif (__riscv_xlen == 64)
    #define FOFFSET XSTR(264) // FP stack offset = (33 * 8)
    #if (STK_RISCV_FPU == 0)
        #define REGSIZE XSTR(((31 + STK_SERVICE_SLOTS) * 8)) // STK_RISCV_REGISTER_COUNT + 2 for mepc, mstatus
    #else
        #if (STK_RISCV_FPU == 32)
        #define REGSIZE XSTR((((31 + STK_SERVICE_SLOTS) * 8) + (31 * 4))) // STK_RISCV_REGISTER_COUNT + 2 for mepc, mstatus + 32 fp registers
        #elif (STK_RISCV_FPU == 64)
        #define REGSIZE XSTR((((31 + STK_SERVICE_SLOTS) * 8) + (31 * 8))) // STK_RISCV_REGISTER_COUNT + 2 for mepc, mstatus + 32 fp registers
        #endif
    #endif
#endif
 
#if (__riscv_xlen == 32)
    #define REGBYTES_LOG2 "2" // log2(4) - used for hart-index shift
#elif (__riscv_xlen == 64)
    #define REGBYTES_LOG2 "3" // log2(8)
#endif

#ifndef STK_SYSTICK_HANDLER
    #define STK_SYSTICK_HANDLER riscv_mtvec_mti
#endif

#ifndef STK_SVC_HANDLER
    #define STK_SVC_HANDLER riscv_mtvec_exception
#endif

#ifndef STK_MSI_HANDLER
    #define STK_MSI_HANDLER riscv_mtvec_msi
#endif

struct TaskFrame
{
    // Service slots (indices 0, 1) - sit at sp+0 and sp+REGBYTES
    Word MEPC;    
    Word MSTATUS; 
 
    // General-purpose register slots (indices 2..N), one per xN
    Word X1_RA;   
    Word X2_SP;   
#if (STK_RISCV_FPU != 0)
    Word X3_FSR;  
#else
    Word X3_GP;   
#endif
    Word X4;      
    Word X5;      
    Word X6;      
    Word X7;      
    Word X8;      
    Word X9;      
    Word X10_A0;  
    Word X11;     
    Word X12;     
    Word X13;     
    Word X14;     
    Word X15;     
#if (__riscv_32e != 1)
    Word X16;     
    Word X17;     
    Word X18;     
    Word X19;     
    Word X20;     
    Word X21;     
    Word X22;     
    Word X23;     
    Word X24;     
    Word X25;     
    Word X26;     
    Word X27;     
    Word X28;     
    Word X29;     
    Word X30;     
    Word X31;     
#endif
#if (STK_RISCV_FPU != 0)
    // FP register slots - at FOFFSET from frame base, immediately after integer slots.
    // FREGBYTES may differ from REGBYTES (32-bit FP on a 64-bit integer machine).
    // Declared as Word arrays for uniform struct sizing; the FP load/store
    // instructions address them by byte offset and tolerate the type mismatch.
    Word F[32];   
#endif
};

static __stk_forceinline void __DSB()
{
    __asm volatile("fence rw, rw" ::: "memory");
}

static __stk_forceinline void __ISB()
{
#ifdef __riscv_zifencei
    __asm volatile("fence.i" ::: "memory");
#else
    __sync_synchronize();
#endif
}

static __stk_forceinline void __WFI()
{
    __asm volatile("wfi");
}

static __stk_forceinline void HW_StartScheduler()
{
    __asm volatile("ecall");
}

static __stk_forceinline uint8_t HW_GetHartId()
{
    return STK_ARCH_GET_CPU_ID();
}

static __stk_forceinline void HW_DisableInterrupts()
{
    __asm volatile("csrrci zero, mstatus, %0"
    : /* output: none */
    : "i"(MSTATUS_MIE)
    : /* clobbers: none */);
}

static __stk_forceinline void HW_EnableInterrupts()
{
    __asm volatile("csrrsi zero, mstatus, %0"
    : /* output: none */
    : "i"(MSTATUS_MIE)
    : /* clobbers: none */);
}

static __stk_forceinline Word HW_EnterCriticalSection()
{
    Word ses;
    __asm volatile("csrrci %0, mstatus, %1"
    : "=r"(ses)
    : "i"(MSTATUS_MIE)
    :  /* clobbers: none */);
 
    return ses;
}

static __stk_forceinline void HW_ExitCriticalSection(Word ses)
{
    __asm volatile("csrrs zero, mstatus, %0"
    : /* output: none */
    : "r"(ses)
    : /* clobbers: none */);
}

static __stk_forceinline void HW_StopMTimer()
{
    clear_csr(mie, MIP_MTIP);
}

static __stk_forceinline uint32_t HW_CoreClockFrequency()
{
    return SystemCoreClock; // CPU speed, e.g. 125/150 MHz
}

static __stk_forceinline uint32_t HW_MtimeClockFrequency()
{
    return STK_TIMER_CLOCK_FREQUENCY; // Timer frequency, e.g. 1 MHz
}

static __stk_forceinline uint64_t HW_GetMtime()
{
#if ( __riscv_xlen > 32)
    return *((volatile uint64_t *)STK_RISCV_CLINT_MTIME_ADDR);
#else
    volatile uint32_t *mtime_hi = ((uint32_t *)STK_RISCV_CLINT_MTIME_ADDR) + 1;
    volatile uint32_t *mtime_lo = ((uint32_t *)STK_RISCV_CLINT_MTIME_ADDR);
 
    uint32_t hi, lo;
    do
    {
        hi = (*mtime_hi);
        lo = (*mtime_lo);
    }
    while (hi != (*mtime_hi)); // make sure mtime_hi did not tick when read mtime_lo
 
    return ((uint64_t)hi << 32) | lo;
#endif
}

static __stk_forceinline void HW_SetMtimecmp(uint64_t time_next)
{
#if STK_RISCV_CLINT_MTIMECMP_PER_HART
    const uint8_t hart = HW_GetHartId();
#else
    const uint8_t hart = 0;
#endif
 
#if (__riscv_xlen == 64)
    ((volatile uint64_t *)STK_RISCV_CLINT_MTIMECMP_ADDR)[hart] = next;
#else
    volatile uint32_t *mtime_lo = (uint32_t *)((uint64_t *)STK_RISCV_CLINT_MTIMECMP_ADDR + hart);
    volatile uint32_t *mtime_hi = mtime_lo + 1;
 
    // expecting 4-byte aligned memory
    STK_ASSERT(((uintptr_t)mtime_lo & (4 - 1)) == 0);
    STK_ASSERT(((uintptr_t)mtime_hi & (4 - 1)) == 0);
 
    // prevent unexpected interrupt by setting some very large value to the high part
    // details: https://riscv.org/wp-content/uploads/2017/05/riscv-privileged-v1.10.pdf, page 31
    (*mtime_hi) = ~0;
 
    (*mtime_lo) = (uint32_t)(time_next & 0xFFFFFFFF);
    (*mtime_hi) = (uint32_t)(time_next >> 32);
#endif
}

static __stk_forceinline uint64_t HW_GetMtimeElapsed(uint64_t since)
{
    return HW_GetMtime() - since;
}

static __stk_forceinline void HW_EnableCycleCounter()
{
    __asm volatile("csrci mcountinhibit, 0x1");
}

static __stk_forceinline Cycles HW_GetCycleCounter()
{
    uint32_t high, low, check;
    do
    {
        __asm volatile("csrr %0, mcycleh" : "=r"(high));
        __asm volatile("csrr %0, mcycle"  : "=r"(low));
        __asm volatile("csrr %0, mcycleh" : "=r"(check));
    }
    while (high != check);
 
    return (static_cast<Cycles>(high) << 32) | low;
}

static __stk_forceinline Word HW_GetCallerSP()
{
    Word sp;
    __asm volatile("mv %0, sp"
    : "=r"(sp)
    : /* input: none */
    : /* clobbers: none */);
 
    return sp;
}

static __stk_forceinline void HW_CriticalSectionStart(Word &ses)
{
    ses = HW_EnterCriticalSection();
 
    // ensure the disable is recognized before subsequent code
    __DSB();
    __ISB();
}

static __stk_forceinline void HW_CriticalSectionEnd(Word ses)
{
    // ensure all memory work is finished before re-enabling
    __DSB();
 
    HW_ExitCriticalSection(ses);
 
    // synchronization point: any pending interrupt can be serviced immediately at this boundary
    __ISB();
}

static __stk_forceinline bool HW_SpinLockTryLock(volatile bool &lock)
{
    return !__atomic_test_and_set(&lock, __ATOMIC_ACQUIRE);
}

static __stk_forceinline void HW_SpinLockLock(volatile bool &lock)
{
    uint32_t timeout = 0xFFFFFF;
    while (!HW_SpinLockTryLock(lock))
    {
        if (--timeout == 0)
        {
            // Invariant violated: the lock owner exited without releasing,
            // Kernel state is suspect, enter defined safe state.
            STK_KERNEL_PANIC(KERNEL_PANIC_SPINLOCK_DEADLOCK);
        }
        __stk_relax_cpu();
    }
}

static __stk_forceinline void HW_SpinLockUnlock(volatile bool &lock)
{
    if (!lock)
        STK_KERNEL_PANIC(KERNEL_PANIC_SPINLOCK_DEADLOCK); // release attempt of unowned lock
 
    // ensure all data writes (like scheduling metadata) are flushed before the lock is released:
    // __atomic_clear with __ATOMIC_RELEASE provides the required store-release barrier,
    // the explicit fence rw,w is retained for toolchains that do not lower __ATOMIC_RELEASE
    // to a full release fence on all RISC-V targets
    __asm volatile("fence rw, w" ::: "memory");
 
    __atomic_clear(&lock, __ATOMIC_RELEASE);
}

static __stk_forceinline void HW_ScheduleContextSwitch(uint8_t hart)
{
#ifdef _STK_RISCV_USE_PENDSV
    // Pend Machine Software Interrupt (MSI) - equivalent of ARM's PENDSVSET
    volatile uint32_t *msip = (volatile uint32_t *)(STK_RISCV_CLINT_BASE_ADDR);
    msip[hart] = 1; // set pending
    __DSB();
#else
    (void)hart;
#endif
}

#ifndef _STK_SYSTEM_CORE_CLOCK_EXTERNAL
volatile uint32_t STK_SYSTEM_CORE_CLOCK_VAR = STK_SYSTEM_CORE_CLOCK_FREQUENCY;
#endif

static volatile bool s_StkRiscvCsuLock = false;


#ifdef _STK_RISCV_USE_PENDSV
Stack * volatile s_StkRiscvStackIdle[STK_ARCH_CPU_COUNT] = {};

volatile Word s_StkRiscvSpIsrInt[STK_ARCH_CPU_COUNT] = {};
#endif

Stack * volatile s_StkRiscvStackActive[STK_ARCH_CPU_COUNT] = {};

Stack * volatile s_StkRiscvStackIsr[STK_ARCH_CPU_COUNT] = {};



struct JmpFrame
{
    Word RA;   
    Word SP;   
    Word S0;   
    Word S1;   
    Word S2;   
    Word S3;   
    Word S4;   
    Word S5;   
    Word S6;   
    Word S7;   
    Word S8;   
    Word S9;   
    Word S10;  
    Word S11;  
#if (STK_RISCV_FPU != 0)
    Word FCSR; 
#endif
};

__attribute__((naked))
int32_t SaveJmp(JmpFrame &/*f*/)
{
    __asm volatile(
    // a0 = &f - no prologue has touched sp or s0 yet
    SREG " ra, 0*" REGBYTES "(a0)  \n" // save return address
    SREG " sp, 1*" REGBYTES "(a0)  \n" // save caller's stack pointer
    SREG " s0, 2*" REGBYTES "(a0)  \n"
    SREG " s1, 3*" REGBYTES "(a0)  \n"
    SREG " s2, 4*" REGBYTES "(a0)  \n"
    SREG " s3, 5*" REGBYTES "(a0)  \n"
    SREG " s4, 6*" REGBYTES "(a0)  \n"
    SREG " s5, 7*" REGBYTES "(a0)  \n"
    SREG " s6, 8*" REGBYTES "(a0)  \n"
    SREG " s7, 9*" REGBYTES "(a0)  \n"
    SREG " s8, 10*" REGBYTES "(a0) \n"
    SREG " s9, 11*" REGBYTES "(a0) \n"
    SREG " s10, 12*" REGBYTES "(a0) \n"
    SREG " s11, 13*" REGBYTES "(a0) \n"
#if (STK_RISCV_FPU != 0)
    "frcsr t0                       \n" // read fcsr (rounding mode + flags)
    SREG " t0, 14*" REGBYTES "(a0)  \n" // save to JmpFrame::FCSR
#endif
    "li    a0, 0                    \n" // return 0
    "ret                            \n" // explicit return (naked)
    );
}

__attribute__((naked, noreturn))
void RestoreJmp(JmpFrame &/*f*/, int32_t /*val*/)
{
    __asm volatile(
    // a0 = &f, a1 = val
    LREG " ra, 0*" REGBYTES "(a0)  \n"
    LREG " sp, 1*" REGBYTES "(a0)  \n"
    LREG " s0, 2*" REGBYTES "(a0)  \n"
    LREG " s1, 3*" REGBYTES "(a0)  \n"
    LREG " s2, 4*" REGBYTES "(a0)  \n"
    LREG " s3, 5*" REGBYTES "(a0)  \n"
    LREG " s4, 6*" REGBYTES "(a0)  \n"
    LREG " s5, 7*" REGBYTES "(a0)  \n"
    LREG " s6, 8*" REGBYTES "(a0)  \n"
    LREG " s7, 9*" REGBYTES "(a0)  \n"
    LREG " s8, 10*" REGBYTES "(a0) \n"
    LREG " s9, 11*" REGBYTES "(a0) \n"
    LREG " s10, 12*" REGBYTES "(a0) \n"
    LREG " s11, 13*" REGBYTES "(a0) \n"
#if (STK_RISCV_FPU != 0)
    LREG " t0, 14*" REGBYTES "(a0)  \n" // load saved fcsr into t0
    "fscsr t0                       \n" // restore rounding mode + flags
#endif
    "mv    a0, a1                   \n" // return val to SaveJmp's caller
    "ret                            \n" // jump to saved RA
    );
}


#if STK_SUBMICORSECOND_PRECISION_TIMER
class HiResClockCYCLE
{
public:
    static HiResClockCYCLE *GetInstance()
    {
        // keep declaration function-local to allow compiler stripping it from the binary if
        // it is unused by the user code
        static HiResClockCYCLE clock;
        return &clock;
    }
 
    Cycles GetCycles()
    {
        return HW_GetCycleCounter();
    }
 
    uint32_t GetFrequency()
    {
        return HW_CoreClockFrequency();
    }
};
typedef HiResClockCYCLE HiResClockImpl;
#else
class HiResClockMTIME
{
public:
    static HiResClockMTIME *GetInstance()
    {
        // keep declaration function-local to allow compiler stripping it from the binary if
        // it is unused by the user code
        static HiResClockMTIME clock;
        return &clock;
    }
 
    Cycles GetCycles()
    {
        return HW_GetMtime();
    }
 
    uint32_t GetFrequency()
    {
        return HW_MtimeClockFrequency();
    }
};
typedef HiResClockMTIME HiResClockImpl;
#endif // !STK_SUBMICORSECOND_PRECISION_TIMER

static struct Context : public PlatformContext
{
    Context() : PlatformContext(), m_stack_main(), m_stack_isr(), m_stack_isr_mem(),
        m_exit_buf(), m_overrider(nullptr), m_specific(nullptr), m_tick_period(0), m_last_mtime(0ULL),
    #if STK_TICKLESS_IDLE
        m_sleep_ticks(0),
    #endif
        m_csu(0), m_csu_nesting(0),
        m_starting(false), m_started(false), m_exiting(false)
 
    {}

    ~Context()
    {}
 
    void Initialize(IPlatform::IEventHandler *handler, IKernelService *service, Stack *exit_trap, int32_t resolution_us)
    {
        PlatformContext::Initialize(handler, service, exit_trap, resolution_us);
 
        // init ISR's stack
        {
            StackMemoryWrapper<STK_RISCV_ISR_STACK_SIZE> stack_isr_mem(&m_stack_isr_mem);
            m_stack_isr.SP   = hw::PtrToWord(InitStackMemory(&stack_isr_mem));
            m_stack_isr.mode = ACCESS_PRIVILEGED;
        }
 
        // init Main stack
        {
            m_stack_main.SP   = STK_STACK_MEMORY_FILLER;
            m_stack_main.mode = ACCESS_PRIVILEGED;
        }
 
        m_csu         = 0;
        m_csu_nesting = 0;
        m_overrider   = NULL;
        m_specific    = NULL;
        m_tick_period = ConvertTimeUsToClockCycles(STK_TIMER_CLOCK_FREQUENCY, resolution_us);
        m_last_mtime  = 0ULL;
        m_starting    = false;
        m_started     = false;
        m_exiting     = false;
 
        // mcycle counter must be enabled per-core
    #if STK_SUBMICORSECOND_PRECISION_TIMER
        HW_EnableCycleCounter();
    #endif
    }
 
    __stk_forceinline void OnTick()
    {
        // process tick - scheduler may update m_stack_active to point at a new task
        Word cs;
        HW_CriticalSectionStart(cs);
 
    #if STK_TICKLESS_IDLE
        Timeout ticks = m_sleep_ticks;
    #endif
 
        if (m_handler->OnTick(m_stack_idle, m_stack_active
        #if STK_TICKLESS_IDLE
            , ticks
        #endif
        ))
        {
            // refresh ISR asm pointer cache so the naked ISR reads the correct
            // (possibly new) active stack SP immediately when jal returns
            // s_StkRiscvStackActive[hart] always points to Context::m_stack_active, the pointer
            // itself is stable, but we reassign here so multi-core hart-indexed builds
            // stay correct if the hart mapping ever changes in future,
            // for single-core builds this is a simple store to a known address at index 0
            const uint8_t hart = HW_GetHartId();
            s_StkRiscvStackActive[hart] = m_stack_active;
        #ifdef _STK_RISCV_USE_PENDSV
            s_StkRiscvStackIdle[hart] = m_stack_idle;
        #endif
 
            HW_ScheduleContextSwitch(hart);
        }
 
    #if STK_TICKLESS_IDLE
        m_sleep_ticks = ticks;
    #endif
 
        HW_CriticalSectionEnd(cs);
    }
 
    __stk_forceinline void EnterCriticalSection()
    {
        // disable local interrupts and save state
        Word current_ses;
        HW_CriticalSectionStart(current_ses);
 
        if (m_csu_nesting == 0)
        {
            // ONLY attempt the global spinlock if we aren't already nested
            HW_SpinLockLock(s_StkRiscvCsuLock);
 
            // store the hardware interrupt state to restore later
            m_csu = current_ses;
        }
 
        // increase nesting count within a limit
        if (++m_csu_nesting > STK_CRITICAL_SECTION_NESTINGS_MAX)
        {
            // invariant violated: exceeded max allowed number of recursions
            STK_KERNEL_PANIC(KERNEL_PANIC_CS_NESTING_OVERFLOW);
        }
    }
 
    __stk_forceinline void ExitCriticalSection()
    {
        STK_ASSERT(m_csu_nesting != 0);
        --m_csu_nesting;
 
        if (m_csu_nesting == 0)
        {
            // capture the state before releasing lock
            Word ses_to_restore = m_csu;
 
            // release global lock
            HW_SpinLockUnlock(s_StkRiscvCsuLock);
 
            // restore hardware interrupts
            HW_CriticalSectionEnd(ses_to_restore);
        }
    }
 
    uint64_t GetSleepTicksPrev()
    {
    #if STK_TICKLESS_IDLE
        uint64_t ticks = (static_cast<uint64_t>(m_sleep_ticks) * static_cast<uint64_t>(m_tick_period));
    #else
        uint64_t ticks = (1U * static_cast<uint64_t>(m_tick_period));
    #endif
        return ticks;
    }
 
    __stk_forceinline void OnSwitchContext()
    {
        // capture mtime at ISR entry as the absolute base for the next period;
        // this eliminates drift from time spent inside OnTick regardless of how
        // long the scheduler takes to run
        const uint64_t mtime_now = HW_GetMtime();
        const uint64_t error = (mtime_now - m_last_mtime) - GetSleepTicksPrev();
        __stk_compiler_barrier(); // avoid compiler reordering, we count ticks from this point
 
        // make sure timer is enabled by the Kernel::Start(), disable its start anywhere else
        STK_ASSERT(m_started);
        STK_ASSERT(m_handler != NULL);
 
        // process tick - scheduler may update m_stack_active and m_sleep_ticks
        OnTick();
 
        // rearm timer: use the ISR-entry mtime snapshot as the absolute base so
        // any CPU cycles consumed by OnTick do not accumulate as period drift
    #if STK_TICKLESS_IDLE
        // guard against overflow (theoretical at normal tick periods and CPU frequencies)
        STK_ASSERT((static_cast<uint64_t>(m_sleep_ticks) * static_cast<uint64_t>(m_tick_period)) <= (UINT64_MAX - mtime_now));
        const uint64_t next_time = (static_cast<uint64_t>(m_sleep_ticks) * static_cast<uint64_t>(m_tick_period));
    #else
        const uint64_t next_time = (1U * static_cast<uint64_t>(m_tick_period));
    #endif
        HW_SetMtimecmp(mtime_now + next_time - error);
        m_last_mtime = mtime_now;
    }
 
    void Start();
    void OnStart();
    void OnStop();
 
    typedef IPlatform::IEventOverrider                               eovrd_t;
    typedef PlatformRiscV::ISpecificEventHandler                     sehndl_t;
    typedef StackMemoryWrapper<STK_RISCV_ISR_STACK_SIZE>::MemoryType isrmem_t;
 
    Stack     m_stack_main;    
    Stack     m_stack_isr;     
    isrmem_t  m_stack_isr_mem; 
    JmpFrame  m_exit_buf;      
    eovrd_t  *m_overrider;     
    sehndl_t *m_specific;      
    uint32_t  m_tick_period;   
    uint64_t  m_last_mtime;    
#if STK_TICKLESS_IDLE
    Timeout   m_sleep_ticks;   
#endif
    Word      m_csu;           
    uint8_t   m_csu_nesting;   
    bool      m_starting;      
    bool      m_started;       
    volatile bool m_exiting;   
}
s_StkPlatformContext[STK_ARCH_CPU_COUNT];
 
void PlatformRiscV::ProcessTick()
{
#ifdef _STK_RISCV_USE_PENDSV
    Word cs;
    HW_CriticalSectionStart(cs);
 
    GetContext().OnTick();
 
    HW_CriticalSectionEnd(cs);
#else
    // unsupported scenario
    STK_ASSERT(false);
#endif
}

static volatile EKernelPanicId g_LastPanicId = KERNEL_PANIC_NONE;
 
__stk_attr_noinline // keep out of inlining to preserve stack frame
__stk_attr_noreturn // never returns - a trap
void STK_PANIC_HANDLER_DEFAULT(EKernelPanicId id)
{
    g_LastPanicId = id;
 
    // disable all maskable interrupts: this prevents scheduler from running again and corrupting state further
    HW_DisableInterrupts();
 
    // spin forever: with a watchdog active this produces a clean reset, without a watchdog,
    // a debugger can attach and inspect 'id'
    for (;;)
    {
        __stk_relax_cpu();
    }
}
 
#define STK_ASM_SAVE_CONTEXT_BASE\
    SREG " x1, 2*" REGBYTES "(sp)    \n"\
    /*SREG " x2, 3*" REGBYTES "(sp)  \n" // skip saving sp, Stack pointer */\
    /*SREG " x3, 4*" REGBYTES "(sp)  \n" // skip saving gp, Global pointer (note: slot is used by fscsr) */\
    SREG " x4, 5*" REGBYTES "(sp)    \n"\
    SREG " x5, 6*" REGBYTES "(sp)    \n"\
    SREG " x6, 7*" REGBYTES "(sp)    \n"\
    SREG " x7, 8*" REGBYTES "(sp)    \n"\
    SREG " x8, 9*" REGBYTES "(sp)    \n"\
    SREG " x9, 10*" REGBYTES "(sp)   \n"\
    SREG " x10, 11*" REGBYTES "(sp)  \n"\
    SREG " x11, 12*" REGBYTES "(sp)  \n"\
    SREG " x12, 13*" REGBYTES "(sp)  \n"\
    SREG " x13, 14*" REGBYTES "(sp)  \n"\
    SREG " x14, 15*" REGBYTES "(sp)  \n"\
    SREG " x15, 16*" REGBYTES "(sp)  \n"
 
#if (__riscv_32e != 1)
#define STK_ASM_SAVE_CONTEXT_RV32I_EXT\
    SREG " x16, 17*" REGBYTES "(sp)  \n"\
    SREG " x17, 18*" REGBYTES "(sp)  \n"\
    SREG " x18, 19*" REGBYTES "(sp)  \n"\
    SREG " x19, 20*" REGBYTES "(sp)  \n"\
    SREG " x20, 21*" REGBYTES "(sp)  \n"\
    SREG " x21, 22*" REGBYTES "(sp)  \n"\
    SREG " x22, 23*" REGBYTES "(sp)  \n"\
    SREG " x23, 24*" REGBYTES "(sp)  \n"\
    SREG " x24, 25*" REGBYTES "(sp)  \n"\
    SREG " x25, 26*" REGBYTES "(sp)  \n"\
    SREG " x26, 27*" REGBYTES "(sp)  \n"\
    SREG " x27, 28*" REGBYTES "(sp)  \n"\
    SREG " x28, 29*" REGBYTES "(sp)  \n"\
    SREG " x29, 30*" REGBYTES "(sp)  \n"\
    SREG " x30, 31*" REGBYTES "(sp)  \n"\
    SREG " x31, 32*" REGBYTES "(sp)  \n"
#else
#define STK_ASM_SAVE_CONTEXT_RV32I_EXT
#endif
 
#if (STK_RISCV_FPU != 0)
#define STK_ASM_SAVE_CONTEXT_FP\
    FSREG " f0, " FOFFSET "+0*" FREGBYTES "(sp)  \n"\
    FSREG " f1, " FOFFSET "+1*" FREGBYTES "(sp)  \n"\
    FSREG " f2, " FOFFSET "+2*" FREGBYTES "(sp)  \n"\
    FSREG " f3, " FOFFSET "+3*" FREGBYTES "(sp)  \n"\
    FSREG " f4, " FOFFSET "+4*" FREGBYTES "(sp)  \n"\
    FSREG " f5, " FOFFSET "+5*" FREGBYTES "(sp)  \n"\
    FSREG " f6, " FOFFSET "+6*" FREGBYTES "(sp)  \n"\
    FSREG " f7, " FOFFSET "+7*" FREGBYTES "(sp)  \n"\
    FSREG " f8, " FOFFSET "+8*" FREGBYTES "(sp)  \n"\
    FSREG " f9, " FOFFSET "+9*" FREGBYTES "(sp)  \n"\
    FSREG " f10, " FOFFSET "+10*" FREGBYTES "(sp)  \n"\
    FSREG " f11, " FOFFSET "+11*" FREGBYTES "(sp)  \n"\
    FSREG " f12, " FOFFSET "+12*" FREGBYTES "(sp)  \n"\
    FSREG " f13, " FOFFSET "+13*" FREGBYTES "(sp)  \n"\
    FSREG " f14, " FOFFSET "+14*" FREGBYTES "(sp)  \n"\
    FSREG " f15, " FOFFSET "+15*" FREGBYTES "(sp)  \n"\
    FSREG " f16, " FOFFSET "+16*" FREGBYTES "(sp)  \n"\
    FSREG " f17, " FOFFSET "+17*" FREGBYTES "(sp)  \n"\
    FSREG " f18, " FOFFSET "+18*" FREGBYTES "(sp)  \n"\
    FSREG " f19, " FOFFSET "+19*" FREGBYTES "(sp)  \n"\
    FSREG " f20, " FOFFSET "+20*" FREGBYTES "(sp)  \n"\
    FSREG " f21, " FOFFSET "+21*" FREGBYTES "(sp)  \n"\
    FSREG " f22, " FOFFSET "+22*" FREGBYTES "(sp)  \n"\
    FSREG " f23, " FOFFSET "+23*" FREGBYTES "(sp)  \n"\
    FSREG " f24, " FOFFSET "+24*" FREGBYTES "(sp)  \n"\
    FSREG " f25, " FOFFSET "+25*" FREGBYTES "(sp)  \n"\
    FSREG " f26, " FOFFSET "+26*" FREGBYTES "(sp)  \n"\
    FSREG " f27, " FOFFSET "+27*" FREGBYTES "(sp)  \n"\
    FSREG " f28, " FOFFSET "+28*" FREGBYTES "(sp)  \n"\
    FSREG " f29, " FOFFSET "+29*" FREGBYTES "(sp)  \n"\
    FSREG " f30, " FOFFSET "+30*" FREGBYTES "(sp)  \n"\
    FSREG " f31, " FOFFSET "+31*" FREGBYTES "(sp)  \n"
#else
#define STK_ASM_SAVE_CONTEXT_FP
#endif
 
#define STK_ASM_SAVE_CONTEXT_PC_STATUS\
    "csrr t0, mepc                   \n"\
    "csrr t1, mstatus                \n"\
    SREG " t0, 0*" REGBYTES "(sp)    \n"\
    SREG " t1, 1*" REGBYTES "(sp)    \n"
 
#if (STK_RISCV_FPU != 0)
#define STK_ASM_SAVE_CONTEXT_FRCSR\
    "frcsr t0                        \n"\
    SREG " t0, 4*" REGBYTES "(sp)    \n"  /* use stack memory slot of gp  (see comment for x3 above) */
#else
#define STK_ASM_SAVE_CONTEXT_FRCSR
#endif
 
#define STK_ASM_SAVE_CONTEXT\
    "addi sp, sp, -" REGSIZE       " \n" /* allocate stack memory for registers */\
    STK_ASM_SAVE_CONTEXT_BASE\
    STK_ASM_SAVE_CONTEXT_RV32I_EXT\
    STK_ASM_SAVE_CONTEXT_FP\
    STK_ASM_SAVE_CONTEXT_PC_STATUS\
    STK_ASM_SAVE_CONTEXT_FRCSR
 
#define STK_ASM_LOAD_CONTEXT_BASE\
    LREG " x1, 2*" REGBYTES "(sp)    \n"\
    /*LREG " x2, 3*" REGBYTES "(sp)  \n" skip loading sp, Stack pointer */\
    /*LREG " x3, 4*" REGBYTES "(sp)  \n" skip loading gp, Global pointer (note: slot is used by fscsr) */\
    LREG " x4, 5*" REGBYTES "(sp)    \n"\
    LREG " x5, 6*" REGBYTES "(sp)    \n"\
    LREG " x6, 7*" REGBYTES "(sp)    \n"\
    LREG " x7, 8*" REGBYTES "(sp)    \n"\
    LREG " x8, 9*" REGBYTES "(sp)    \n"\
    LREG " x9, 10*" REGBYTES "(sp)   \n"\
    LREG " x10, 11*" REGBYTES "(sp)  \n"\
    LREG " x11, 12*" REGBYTES "(sp)  \n"\
    LREG " x12, 13*" REGBYTES "(sp)  \n"\
    LREG " x13, 14*" REGBYTES "(sp)  \n"\
    LREG " x14, 15*" REGBYTES "(sp)  \n"\
    LREG " x15, 16*" REGBYTES "(sp)  \n"
 
#if (__riscv_32e != 1)
#define STK_ASM_LOAD_CONTEXT_RV32I_EXT\
    LREG " x16, 17*" REGBYTES "(sp)  \n"\
    LREG " x17, 18*" REGBYTES "(sp)  \n"\
    LREG " x18, 19*" REGBYTES "(sp)  \n"\
    LREG " x19, 20*" REGBYTES "(sp)  \n"\
    LREG " x20, 21*" REGBYTES "(sp)  \n"\
    LREG " x21, 22*" REGBYTES "(sp)  \n"\
    LREG " x22, 23*" REGBYTES "(sp)  \n"\
    LREG " x23, 24*" REGBYTES "(sp)  \n"\
    LREG " x24, 25*" REGBYTES "(sp)  \n"\
    LREG " x25, 26*" REGBYTES "(sp)  \n"\
    LREG " x26, 27*" REGBYTES "(sp)  \n"\
    LREG " x27, 28*" REGBYTES "(sp)  \n"\
    LREG " x28, 29*" REGBYTES "(sp)  \n"\
    LREG " x29, 30*" REGBYTES "(sp)  \n"\
    LREG " x30, 31*" REGBYTES "(sp)  \n"\
    LREG " x31, 32*" REGBYTES "(sp)  \n"
#else
#define STK_ASM_LOAD_CONTEXT_RV32I_EXT
#endif
 
#if (STK_RISCV_FPU != 0)
#define STK_ASM_LOAD_CONTEXT_FP\
    FLREG " f0, " FOFFSET "+0*" FREGBYTES "(sp)  \n"\
    FLREG " f1, " FOFFSET "+1*" FREGBYTES "(sp)  \n"\
    FLREG " f2, " FOFFSET "+2*" FREGBYTES "(sp)  \n"\
    FLREG " f3, " FOFFSET "+3*" FREGBYTES "(sp)  \n"\
    FLREG " f4, " FOFFSET "+4*" FREGBYTES "(sp)  \n"\
    FLREG " f5, " FOFFSET "+5*" FREGBYTES "(sp)  \n"\
    FLREG " f6, " FOFFSET "+6*" FREGBYTES "(sp)  \n"\
    FLREG " f7, " FOFFSET "+7*" FREGBYTES "(sp)  \n"\
    FLREG " f8, " FOFFSET "+8*" FREGBYTES "(sp)  \n"\
    FLREG " f9, " FOFFSET "+9*" FREGBYTES "(sp)  \n"\
    FLREG " f10, " FOFFSET "+10*" FREGBYTES "(sp)  \n"\
    FLREG " f11, " FOFFSET "+11*" FREGBYTES "(sp)  \n"\
    FLREG " f12, " FOFFSET "+12*" FREGBYTES "(sp)  \n"\
    FLREG " f13, " FOFFSET "+13*" FREGBYTES "(sp)  \n"\
    FLREG " f14, " FOFFSET "+14*" FREGBYTES "(sp)  \n"\
    FLREG " f15, " FOFFSET "+15*" FREGBYTES "(sp)  \n"\
    FLREG " f16, " FOFFSET "+16*" FREGBYTES "(sp)  \n"\
    FLREG " f17, " FOFFSET "+17*" FREGBYTES "(sp)  \n"\
    FLREG " f18, " FOFFSET "+18*" FREGBYTES "(sp)  \n"\
    FLREG " f19, " FOFFSET "+19*" FREGBYTES "(sp)  \n"\
    FLREG " f20, " FOFFSET "+20*" FREGBYTES "(sp)  \n"\
    FLREG " f21, " FOFFSET "+21*" FREGBYTES "(sp)  \n"\
    FLREG " f22, " FOFFSET "+22*" FREGBYTES "(sp)  \n"\
    FLREG " f23, " FOFFSET "+23*" FREGBYTES "(sp)  \n"\
    FLREG " f24, " FOFFSET "+24*" FREGBYTES "(sp)  \n"\
    FLREG " f25, " FOFFSET "+25*" FREGBYTES "(sp)  \n"\
    FLREG " f26, " FOFFSET "+26*" FREGBYTES "(sp)  \n"\
    FLREG " f27, " FOFFSET "+27*" FREGBYTES "(sp)  \n"\
    FLREG " f28, " FOFFSET "+28*" FREGBYTES "(sp)  \n"\
    FLREG " f29, " FOFFSET "+29*" FREGBYTES "(sp)  \n"\
    FLREG " f30, " FOFFSET "+30*" FREGBYTES "(sp)  \n"\
    FLREG " f31, " FOFFSET "+31*" FREGBYTES "(sp)  \n"
#else
#define STK_ASM_LOAD_CONTEXT_FP
#endif
 
#define STK_ASM_LOAD_CONTEXT_PC_STATUS\
    LREG " t0, 0*" REGBYTES "(sp)    \n"\
    LREG " t1, 1*" REGBYTES "(sp)    \n"\
    "csrw mepc, t0                   \n"\
    "csrw mstatus, t1                \n"
 
#if (STK_RISCV_FPU != 0)
#define STK_ASM_LOAD_CONTEXT_FRCSR\
    LREG " t0, 4*" REGBYTES "(sp)    \n" /* use stack memory slot of gp (see comment for x3 below) */\
    "fscsr t0                        \n"
#else
#define STK_ASM_LOAD_CONTEXT_FRCSR
#endif
 
#define STK_ASM_LOAD_CONTEXT\
    STK_ASM_LOAD_CONTEXT_PC_STATUS\
    STK_ASM_LOAD_CONTEXT_FRCSR\
    STK_ASM_LOAD_CONTEXT_BASE\
    STK_ASM_LOAD_CONTEXT_RV32I_EXT\
    STK_ASM_LOAD_CONTEXT_FP\
    "addi sp, sp, " REGSIZE   " \n" /* shrink stack memory of registers */
 
static __stk_forceinline void HW_LoadContextAndExit()
{
    __asm volatile(
    LREG " t0, %0               \n" // load the first member (SP) into t0
    LREG " sp, 0(t0)            \n" // sp = t0
 
    STK_ASM_LOAD_CONTEXT
    STK_ASM_EXIT_FROM_HANDLER " \n"
 
    : /* output: none */
    : "m"(GetContext().m_stack_active)
    : "t0", "t1", "a2", "a3", "a4", "a5", "gp", "memory");
}
 
static __stk_forceinline void HW_EnableFullFpuAccess()
{
#if (STK_RISCV_FPU != 0)
    __asm volatile(
    "csrs mstatus, %0"
    : /* output: none */
    : "r"(MSTATUS_FS | MSTATUS_XS)
    : "memory" /* ensure no FP instructions are moved before this call */);
#endif
}
 
static __stk_forceinline void HW_ClearFpuState()
{
#if (STK_RISCV_FPU != 0)
    __asm volatile(
    "fssr x0"
    : /* output: none */
    : /* input: none */
    : "memory" /* ensure flags are cleared before next FP op */);
#endif
}
 
static __stk_forceinline void HW_SaveMainSP()
{
    __asm volatile(
    SREG " sp, %0"
    : "=m"(GetContext().m_stack_main)
    : /* input: none */
    : "memory" /* protect against compiler reordering */ );
}
 
static __stk_forceinline void HW_LoadMainSP()
{
    __asm volatile(
    LREG " sp, %0"
    : /* output: none */
    : "m"(GetContext().m_stack_main)
    : "memory" /* protect against compiler reordering */ );
}
 
static __stk_forceinline bool HW_IsHandlerMode()
{
    Word current_sp = HW_GetCallerSP();
 
    // get the bounds of the ISR stack from our Context
    // note: STK uses StackMemoryWrapper, so we check against that memory block
    const Word isr_stack_base = (Word)&GetContext().m_stack_isr_mem;
    const Word isr_stack_top  = isr_stack_base + STK_RISCV_ISR_STACK_SIZE;
 
    return ((current_sp >= isr_stack_base) && (current_sp < isr_stack_top));
}
 
static __stk_forceinline void OnTaskStart()
{
    HW_LoadContextAndExit();
}
 
// __stk_attr_used for LTO
extern "C" STK_RISCV_ISR_SECTION __stk_attr_used void TrySwitchContext()
{
    GetContext().OnSwitchContext();
}
 
#ifdef _STK_RISCV_USE_PENDSV
extern "C" STK_RISCV_ISR_SECTION __stk_attr_naked void STK_SYSTICK_HANDLER()
{
    __asm volatile(
    // 1. save full interrupted context onto the task stack
    STK_ASM_SAVE_CONTEXT
 
    // 2. store task SP into s_StkRiscvSpIsrInt[hart] directly (plain Word, no struct indirection)
#if (STK_ARCH_CPU_COUNT > 1)
    "csrr t0, mhartid                    \n"
    "la   t1, s_StkRiscvSpIsrInt         \n"
    "slli t0, t0, " REGBYTES_LOG2 "      \n" // t0 = hart * sizeof(Word)
    "add  t1, t1, t0                     \n" // t1 = &s_StkRiscvSpIsrInt[hart]
    SREG " sp, 0(t1)                     \n" // store sp directly - no pointer dereference
#else
    "la   t1, s_StkRiscvSpIsrInt         \n"
    SREG " sp, 0(t1)                     \n" // store sp directly - no pointer dereference
#endif
 
    // 3. switch to private ISR stack
#if (STK_ARCH_CPU_COUNT > 1)
    "csrr t0, mhartid                    \n"
    "la   t1, s_StkRiscvStackIsr         \n"
    "slli t0, t0, " REGBYTES_LOG2 "      \n"
    "add  t1, t1, t0                     \n"
    LREG " t1, 0(t1)                     \n"
#else
    "la   t1, s_StkRiscvStackIsr         \n"
    LREG " t1, 0(t1)                     \n"
#endif
    LREG " sp, 0(t1)                     \n" // sp = Stack::SP of ISR stack
 
    // 4. run scheduler
    "jal  ra, TrySwitchContext           \n"
 
    // 5. restore the interrupted task's SP from s_StkRiscvSpIsrInt[hart]
#if (STK_ARCH_CPU_COUNT > 1)
    "csrr t0, mhartid                    \n"
    "la   t1, s_StkRiscvSpIsrInt         \n"
    "slli t0, t0, " REGBYTES_LOG2 "      \n"
    "add  t1, t1, t0                     \n"
    LREG " sp, 0(t1)                     \n" // sp = saved task SP - direct load, no struct
#else
    "la   t1, s_StkRiscvSpIsrInt         \n"
    LREG " sp, 0(t1)                     \n" // sp = saved task SP - direct load, no struct
#endif
 
    // 6. restore context
    STK_ASM_LOAD_CONTEXT
 
    // 7. exit ISR handler
    STK_ASM_EXIT_FROM_HANDLER "          \n"
 
    : /* outputs:  none - naked, compiler emits nothing outside this asm */
    : /* inputs: all addresses loaded as linker symbols via "la" */
    : /* clobbers: none - the asm string owns all registers */);
}
extern "C" STK_RISCV_ISR_SECTION __stk_attr_naked void STK_MSI_HANDLER()
{
    __asm volatile(
    // 1. save context
    STK_ASM_SAVE_CONTEXT
 
    // 2. store task SP into s_StkRiscvStackIdle[hart]->SP
    // all integer registers are now saved. t0/t1 are free to use as scratch.
    // "la" loads the address of the global array - a linker-time constant,
    // no compiler-generated runtime code, safe to use here
#if (STK_ARCH_CPU_COUNT > 1)
    "csrr t0, mhartid                    \n"
    "la   t1, s_StkRiscvStackIdle        \n"
    "slli t0, t0, " REGBYTES_LOG2 "      \n" // t0 = hart * sizeof(Stack*)
    "add  t1, t1, t0                     \n" // t1 = &s_StkRiscvStackIdle[hart]
    LREG " t1, 0(t1)                     \n" // t1 = s_StkRiscvStackIdle[hart] (Stack*)
#else
    "la   t1, s_StkRiscvStackIdle        \n"
    LREG " t1, 0(t1)                     \n" // t1 = s_StkRiscvStackIdle[0] (Stack*)
#endif
    SREG " sp, 0(t1)                     \n" // Stack::SP = task's sp (SP is first member)
 
    // 3. clear exception: MSIP[hart] = 0
#if (STK_ARCH_CPU_COUNT > 1)
    "csrr  t0, mhartid                   \n"
    "slli  t0, t0, 2                     \n" // t0 = hart * 4
    "li    t1, %[clint_msip_base]        \n"
    "add   t0, t0, t1                    \n" // t0 = &MSIP[hart]
#else
    "li    t0, %[clint_msip_base]        \n" // t0 = &MSIP[0]
#endif
    "sw    zero, 0(t0)                   \n" // MSIP[hart] = 0
    "fence rw, rw                        \n" // fence rw,rw  - ensure the write is visible before re-enable
 
    // 4. load SP from s_StkRiscvStackActive[hart]->SP
#if (STK_ARCH_CPU_COUNT > 1)
    "csrr t0, mhartid                    \n"
    "la   t1, s_StkRiscvStackActive      \n"
    "slli t0, t0, " REGBYTES_LOG2 "      \n"
    "add  t1, t1, t0                     \n"
    LREG " t1, 0(t1)                     \n"
#else
    "la   t1, s_StkRiscvStackActive      \n"
    LREG " t1, 0(t1)                     \n"
#endif
    LREG " sp, 0(t1)                     \n"  // sp = active task's saved SP
 
    // 5. load context of the active task
    STK_ASM_LOAD_CONTEXT
 
    // 6. exit ISR handler
    STK_ASM_EXIT_FROM_HANDLER "          \n"
 
    : /* outputs:  none - naked, compiler emits nothing outside this asm */
    : [clint_msip_base] "i" (STK_RISCV_CLINT_BASE_ADDR) /* other inputs: all addresses loaded as linker symbols via "la" */
    : /* clobbers: none - the asm string owns all registers */);
}
#else // !_STK_RISCV_USE_PENDSV
/* STK_SYSTICK_HANDLER
 
RISC-V machine-timer ISR: Saves the interrupted task's full context, switches
to the private ISR stack, calls TrySwitchContext (which reschedules the timer
and runs the scheduler), then restores the (possibly new) task's context.
 
DESIGN RULES - must be obeyed to work correctly at all optimisation levels:
 
 1. Single asm volatile, no compiler operands.
    The function body is ONE __asm volatile("..." : : : ) with empty
    input/output/clobber lists. No "m" or "r" constraints are used because
    the compiler evaluates those as C expressions BEFORE emitting any asm
    text, i.e. before the register save - trashing uninitialized registers.
 
 2. All addresses are linker symbols loaded via "la" inside the asm.
    s_StkRiscvStackActive and s_StkRiscvStackIsr are plain file-scope globals. "la reg, sym"
    emits a PC-relative load that is resolved at link time, it produces no
    compiler-generated code outside the asm string.
 
 3. Stack pointer indexing uses sizeof(Stack*) == REGBYTES.
    For multi-hart builds the array index is hart * REGBYTES, which is a
    single left-shift by log2(REGBYTES): 2 for RV32 (4 bytes), 3 for RV64
    (8 bytes). REGBYTES_LOG2 is defined below accordingly.
 
 4. s_StkRiscvStackActive[hart]->SP is updated by TrySwitchContext.
    The naked asm reads it fresh after the jal returns, so it always sees
    the task the scheduler has chosen - even if it changed.
 
 Stack frame layout (offsets from sp after "addi sp,-REGSIZE"):
   [0*REGBYTES] mepc     (service slot 0)
   [1*REGBYTES] mstatus  (service slot 1)
   [2*REGBYTES] x1  / ra
   [3*REGBYTES] x2  / sp  - SKIPPED, managed explicitly
   [4*REGBYTES] x3  / gp  - SKIPPED, fixed register; slot reused for FCSR
   [5*REGBYTES] x4  / tp
   [6*REGBYTES] x5  / t0
   ...
   [32*REGBYTES] x31 / t6  (RV32I; absent on RV32E)
   [FOFFSET + n*FREGBYTES] fn  (FP registers, if STK_RISCV_FPU != 0)
*/
extern "C" STK_RISCV_ISR_SECTION __stk_attr_naked void STK_SYSTICK_HANDLER()
{
    __asm volatile(
    // 1. save context
    STK_ASM_SAVE_CONTEXT
 
    // 2. store task SP into s_StkRiscvStackActive[hart]->SP
    // all integer registers are now saved. t0/t1 are free to use as scratch.
    // "la" loads the address of the global array - a linker-time constant,
    // no compiler-generated runtime code, safe to use here
#if (STK_ARCH_CPU_COUNT > 1)
    "csrr t0, mhartid                    \n"
    "la   t1, s_StkRiscvStackActive      \n"
    "slli t0, t0, " REGBYTES_LOG2 "      \n"  // t0 = hart * sizeof(Stack*)
    "add  t1, t1, t0                     \n"  // t1 = &s_StkRiscvStackActive[hart]
    LREG " t1, 0(t1)                     \n"  // t1 = s_StkRiscvStackActive[hart] (Stack*)
#else
    "la   t1, s_StkRiscvStackActive      \n"
    LREG " t1, 0(t1)                     \n"  // t1 = s_StkRiscvStackActive[0] (Stack*)
#endif
    SREG " sp, 0(t1)                     \n"  // Stack::SP = task's sp (SP is first member)
 
    // 3. switch to private ISR stack
#if (STK_ARCH_CPU_COUNT > 1)
    "csrr t0, mhartid                    \n"
    "la   t1, s_StkRiscvStackIsr         \n"
    "slli t0, t0, " REGBYTES_LOG2 "      \n"
    "add  t1, t1, t0                     \n"
    LREG " t1, 0(t1)                     \n"  // t1 = s_StkRiscvStackIsr[hart] (Stack*)
#else
    "la   t1, s_StkRiscvStackIsr         \n"
    LREG " t1, 0(t1)                     \n"  // t1 = s_StkRiscvStackIsr[0] (Stack*)
#endif
    LREG " sp, 0(t1)                     \n"  // sp = Stack::SP of ISR stack
 
    // 4. call TrySwitchContext
    // runs on the ISR stack: reschedules timer, runs scheduler
    // (which may update m_stack_active to a new task), then updates
    // s_StkRiscvStackActive[hart] so step 5 below reads the correct new SP,
    // all caller-saved registers (a0-a7, t0-t6, ra) are trashed - expected
    "jal  ra, TrySwitchContext           \n"
 
    // 5. reload SP from s_StkRiscvStackActive[hart]->SP
    // TrySwitchContext updated s_StkRiscvStackActive[hart] before returning,
    // we re-read it fresh to pick up any task switch the scheduler made
#if (STK_ARCH_CPU_COUNT > 1)
    "csrr t0, mhartid                    \n"
    "la   t1, s_StkRiscvStackActive      \n"
    "slli t0, t0, " REGBYTES_LOG2 "      \n"
    "add  t1, t1, t0                     \n"
    LREG " t1, 0(t1)                     \n"
#else
    "la   t1, s_StkRiscvStackActive      \n"
    LREG " t1, 0(t1)                     \n"
#endif
    LREG " sp, 0(t1)                     \n"  // sp = active task's saved SP
 
    // 6. load context of the active task
    STK_ASM_LOAD_CONTEXT
 
    // 7. exit ISR handler
    STK_ASM_EXIT_FROM_HANDLER "          \n"
 
    : /* outputs:  none - naked, compiler emits nothing outside this asm */
    : /* inputs:   none - all addresses loaded as linker symbols via "la" */
    : /* clobbers: none - the asm string owns all registers */
    );
}
#endif // !_STK_RISCV_USE_PENDSV
 
void Context::OnStart()
{
    const uint8_t hart = HW_GetHartId();
 
    // save SP of main stack to reuse it for scheduler exit
    HW_SaveMainSP();
 
    // enable FPU (if available)
    HW_EnableFullFpuAccess();
 
    // clear FPU usage status if FPU was used before kernel start
    HW_ClearFpuState();
 
    // notify kernel
    m_handler->OnStart(m_stack_active);
 
#if STK_TICKLESS_IDLE
    // reset sleep ticks if kernel was restarted
    m_sleep_ticks = 1;
#endif
 
    // start timer with default periodicity
    m_last_mtime = HW_GetMtime();
    HW_SetMtimecmp(m_last_mtime + m_tick_period);
 
    // change state before enabling interrupt
    m_started  = true;
    m_starting = false;
 
    // initialize ISR asm pointer cache
    s_StkRiscvStackIsr[hart]    = &m_stack_isr; // set once here, the ISR stack never moves
    s_StkRiscvStackActive[hart] = m_stack_active;
#ifdef _STK_RISCV_USE_PENDSV
    s_StkRiscvStackIdle[hart]   = m_stack_idle;
#endif
 
    // enable timer interrupt
    set_csr(mie, MIP_MTIP
    #ifdef _STK_RISCV_USE_PENDSV
        | MIP_MSIP
    #endif
    );
}
 
STK_RISCV_ISR void STK_SVC_HANDLER()
{
    Word cause;
    __asm volatile("csrr %0, mcause"
    : "=r"(cause)
    : /* input : none */
    : /* clobbers: none */);
 
    /*if (cause & (1UL << (__riscv_xlen - 1)))
    {
        cause &= ~(1UL << (__riscv_xlen - 1));
 
        if (cause == IRQ_M_TIMER)
        {
 
        }
    }*/
 
    if (cause == IRQ_M_EXT)
    {
        // not starting scheduler, then try to forward ecall to user
        if (!GetContext().m_starting)
        {
            // forward event to user
            if (GetContext().m_specific != NULL)
                GetContext().m_specific->OnException(cause);
 
            // switch to the next instruction of the caller space (PC) after the return
            write_csr(mepc, read_csr(mepc) + sizeof(Word));
        }
        else
        {
            // make sure interrupts do not interfere
            HW_DisableInterrupts();
 
            // configure scheduling
            GetContext().OnStart();
 
            // start first task
            OnTaskStart();
        }
    }
    else
    {
        if (GetContext().m_specific != NULL)
        {
            // forward event to user
            GetContext().m_specific->OnException(cause);
        }
        else
        {
            // trap further execution
            // note: normally, if trapped here with cause 2 or 4 then check stack memory size of the
            // tasks, scheduler and ISR, they were likely overwritten if your code is 100% correct
            STK_KERNEL_PANIC(KERNEL_PANIC_CPU_EXCEPTION);
        }
    }
}
 
static void OnTaskRun(ITask *task)
{
    task->Run();
}
 
static void OnTaskExit()
{
    Word cs;
    HW_CriticalSectionStart(cs);
 
    GetContext().m_handler->OnTaskExit(GetContext().m_stack_active);
 
    HW_CriticalSectionEnd(cs);
 
    for (;;)
    {
        __DSB(); // data barrier
        __WFI(); // enter standby mode until time slot expires
    }
}
 
static STK_RISCV_ISR_SECTION void OnSchedulerSleep()
{
#if STK_SEGGER_SYSVIEW
    SEGGER_SYSVIEW_OnIdle();
#endif
 
    for (;;)
    {
        __DSB(); // data barrier
        __WFI(); // enter sleep until interrupt
    }
}
 
static STK_RISCV_ISR_SECTION void OnSchedulerSleepOverride()
{
    if (!GetContext().m_overrider->OnSleep())
        OnSchedulerSleep();
}
 
static void OnSchedulerExit()
{
    // switch to main stack
    HW_LoadMainSP();
 
    // jump to the exit from the IKernel::Start()
    RestoreJmp(GetContext().m_exit_buf, 0);
}
 
void PlatformRiscV::Initialize(IEventHandler *event_handler, IKernelService *service, uint32_t resolution_us, Stack *exit_trap)
{
    GetContext().Initialize(event_handler, service, exit_trap, resolution_us);
}
 
void Context::Start()
{
    m_exiting = false;
 
    // save jump location of the Exit trap
    SaveJmp(m_exit_buf);
    if (m_exiting)
    {
        // notify kernel about a full stop
        m_handler->OnStop();
        return;
    }
 
    // enable FPU (if available)
    HW_EnableFullFpuAccess();
 
    // start
    m_starting = true;
    HW_StartScheduler();
}
 
void PlatformRiscV::Start()
{
    GetContext().Start();
}
 
bool PlatformRiscV::InitStack(EStackType stack_type, Stack *stack, IStackMemory *stack_memory, ITask *user_task)
{
    // TaskFrame must map exactly onto the slot layout consumed by STK_ASM_SAVE_CONTEXT / STK_ASM_LOAD_CONTEXT - no padding allowed
    STK_STATIC_ASSERT_DESC(sizeof(TaskFrame) == (STK_RISCV_REGISTER_COUNT + STK_SERVICE_SLOTS) * sizeof(Word),
        "TaskFrame size must match REGSIZE: (REGISTER_COUNT + SERVICE_SLOTS) * REGBYTES");
 
    STK_ASSERT(stack_memory->GetStackSize() > (STK_RISCV_REGISTER_COUNT + STK_SERVICE_SLOTS));
 
    // initialize stack memory (fills all slots with STK_STACK_MEMORY_FILLER)
    Word *stack_top = PlatformContext::InitStackMemory(stack_memory);
 
    // initialize Stack Pointer (SP): frame sits at the bottom of the register window
    stack->SP = hw::PtrToWord(stack_top - (STK_RISCV_REGISTER_COUNT + STK_SERVICE_SLOTS));
 
    // place the task frame at SP directly at the base of the register window
    TaskFrame * const task_frame = reinterpret_cast<TaskFrame *>(stack->SP);
 
    // initialize registers for the user task's first start
    switch (stack_type)
    {
    case STACK_USER_TASK: {
        task_frame->MEPC   = hw::PtrToWord(&OnTaskRun);
        task_frame->X1_RA  = hw::PtrToWord(&OnTaskExit);
        task_frame->X10_A0 = hw::PtrToWord(user_task);
        break; }
 
    case STACK_SLEEP_TRAP: {
        task_frame->MEPC  = hw::PtrToWord(GetContext().m_overrider != NULL ? &OnSchedulerSleepOverride : &OnSchedulerSleep);
        task_frame->X1_RA = STK_STACK_MEMORY_FILLER; // should not attempt to exit
        break; }
 
    case STACK_EXIT_TRAP: {
        task_frame->MEPC  = hw::PtrToWord(&OnSchedulerExit);
        task_frame->X1_RA = STK_STACK_MEMORY_FILLER; // should not attempt to exit
        break; }
 
    default:
        return false;
    }
 
    // mstatus: return to M-mode (MPP), interrupts enabled on mret (MPIE),
    // FPU/extension state initial (FS/XS) if FPU present
    task_frame->MSTATUS = MSTATUS_MPP | MSTATUS_MPIE | (STK_RISCV_FPU != 0 ? (MSTATUS_FS | MSTATUS_XS) : 0);
 
#if (STK_RISCV_FPU != 0)
    task_frame->X3_FSR = 0; // FCSR = 0: round-to-nearest, no accrued exception flags
#endif
 
    return true;
}
 
void Context::OnStop()
{
    // stop timer
    HW_StopMTimer();
 
    // clear pending SV exception
#ifdef _STK_RISCV_USE_PENDSV
    clear_csr(mie, MIP_MSIP);
#endif
 
    m_started = false;
    m_exiting = true;
 
    // make sure all assignments are set and executed
    __DSB();
    __ISB();
}
 
void PlatformRiscV::Stop()
{
    GetContext().OnStop();
 
    // load context of the Exit trap
    HW_DisableInterrupts();
    OnTaskStart();
}
 
uint32_t PlatformRiscV::GetTickResolution() const
{
    return GetContext().m_tick_resolution;
}
 
void PlatformRiscV::SwitchToNext()
{
    GetContext().m_handler->OnTaskSwitch(HW_GetCallerSP());
}
 
void PlatformRiscV::Sleep(Timeout ticks)
{
    GetContext().m_handler->OnTaskSleep(HW_GetCallerSP(), ticks);
}
 
void PlatformRiscV::SleepUntil(Ticks timestamp)
{
    GetContext().m_handler->OnTaskSleepUntil(HW_GetCallerSP(), timestamp);
}
 
IWaitObject *PlatformRiscV::Wait(ISyncObject *sync_obj, IMutex *mutex, Timeout timeout)
{
    return GetContext().m_handler->OnTaskWait(HW_GetCallerSP(), sync_obj, mutex, timeout);
}
 
TId PlatformRiscV::GetTid() const
{
    return GetContext().m_handler->OnGetTid(HW_GetCallerSP());
}
 
void PlatformRiscV::ProcessHardFault()
{
    if ((GetContext().m_overrider == NULL) || !GetContext().m_overrider->OnHardFault())
    {
        STK_KERNEL_PANIC(KERNEL_PANIC_HRT_HARD_FAULT);
    }
}
 
void PlatformRiscV::SetEventOverrider(IEventOverrider *overrider)
{
    STK_ASSERT(!GetContext().m_started);
    GetContext().m_overrider = overrider;
}
 
Word PlatformRiscV::GetCallerSP() const
{
    return HW_GetCallerSP();
}
 
void PlatformRiscV::SetSpecificEventHandler(ISpecificEventHandler *handler)
{
    STK_ASSERT(!GetContext().m_started);
    GetContext().m_specific = handler;
}
 
IKernelService *IKernelService::GetInstance()
{
    return GetContext().m_service;
}
 
void stk::hw::CriticalSection::Enter()
{
    GetContext().EnterCriticalSection();
}
 
void stk::hw::CriticalSection::Exit()
{
    GetContext().ExitCriticalSection();
}
 
void stk::hw::SpinLock::Lock()
{
    HW_SpinLockLock(m_lock);
}
 
void stk::hw::SpinLock::Unlock()
{
    HW_SpinLockUnlock(m_lock);
}
 
bool stk::hw::SpinLock::TryLock()
{
    return HW_SpinLockTryLock(m_lock);
}
 
bool stk::hw::IsInsideISR()
{
    return HW_IsHandlerMode();
}
 
Cycles stk::hw::HiResClock::GetCycles()
{
    return HiResClockImpl::GetInstance()->GetCycles();
}
 
uint32_t stk::hw::HiResClock::GetFrequency()
{
    uint32_t freq = HiResClockImpl::GetInstance()->GetFrequency();
    STK_ASSERT(freq != 0);
    return freq;
}
 
#endif // _STK_ARCH_RISC_V