save fpu state properly-ish, add a guard to doom_malloc_wrapper, fix QueryPerformance* and GetSystemTimeAsFileTime

code/doom/doom.c

@@ -27,11 +27,12 @@ static void doom_print_wrapper(const char* msg)
 
 // ---------------------------------------------------------------------------
 // doom_malloc_fn / doom_free_fn
-// Heap handle 0, same convention as video.c.
 // ---------------------------------------------------------------------------
 static void* doom_malloc_wrapper(int size)
 {
-    return HeapAlloc(0, HEAP_ZERO_MEMORY, size);
+    void* p = HeapAlloc(0, HEAP_ZERO_MEMORY, size);
+    if (!p) ExitProcess(GetLastError());
+    return p;
 }
 
 static void doom_free_wrapper(void* ptr)
@@ -126,7 +127,7 @@ static void doom_exit_wrapper(int code)
 static char* doom_getenv_wrapper(const char* name)
 {
     if (strcmp(name, "HOME") == 0) {
-        static char home[] = "fs0:\\";
+        static char home[] = "c:";
         doom_print_wrapper("getenv: Returning HOME="); doom_print_wrapper(home); doom_print_wrapper("\n");
         return home;
     }

code/kernel/ke_sched.c

@@ -53,6 +53,7 @@ void KeInitializeScheduler(void) {
     __stosq((unsigned __int64 *)ctx, 0, sizeof(KCONTEXT) / 8);
     ctx->Rip = (uint64_t)IdleLoop;
     idle->SavedContext = ctx;
+    KeInitThreadFpuContext(idle);
 
     gIdleThread    = idle;
     gCurrentThread = idle;

code/kernel/ke_switch.asm

@@ -9,6 +9,7 @@
 THREAD_SavedContext EQU 88
 THREAD_TssRsp0      EQU 104
 THREAD_Teb          EQU 112
+THREAD_FpuContext   EQU 160
 TSS_RSP0_OFFSET     EQU 4
 
 EXTERN gTss:QWORD          ; address of the single TSS, exported from ke_idt.c
@@ -20,6 +21,9 @@ EXTERN gTss:QWORD          ; address of the single TSS, exported from ke_idt.c
 ;   RCX = from,  RDX = to   (Microsoft x64)
 ; -------------------------------------------------------------------------
 KeSwitchContext PROC
+    mov     rax, [rcx + THREAD_FpuContext]
+    fxsave64 [rax]
+
     ; Save non-volatile registers onto the current (outgoing) kernel stack.
     push    rbx
     push    rbp
@@ -50,6 +54,9 @@ KeSwitchContext PROC
     mov     r8,  gTss
     mov     qword ptr [r8 + TSS_RSP0_OFFSET], rax
 
+    mov     rax, [rdx + THREAD_FpuContext]
+    fxrstor64 [rax]
+
     ; Restore non-volatile registers of the incoming thread.
     pop     r15
     pop     r14

code/kernel/ke_thread.c

@@ -7,6 +7,9 @@
 #include <mm_virt.h>
 #include <ob_object.h>
 
+#define FPU_FCW_DEFAULT   0x037Fu
+#define FPU_MXCSR_DEFAULT 0x1F80u
+
 static uint32_t gNextThreadId = 1;
 extern void KeThreadEntry(void);
 
@@ -80,8 +83,8 @@ __declspec(dllexport) KSTATUS KeCreateThread(PROCESS* owner, uintptr_t entry, ui
     // offset 0x10 is StackLimit, which __chkstk checks.
     uint64_t* teb = (uint64_t*)ustack;
     teb[0] = 0; // ExceptionList (GS:0x00)
-    teb[1] = 0; // StackBase     (GS:0x08)
-    teb[2] = 0; // StackLimit    (GS:0x10) - __chkstk will see 0 and safely exit
+    teb[1] = userStackTop;        // StackBase
+    teb[2] = (uintptr_t)ustack;   // StackLimit
     
     thread->Teb = teb;
 
@@ -124,6 +127,7 @@ __declspec(dllexport) KSTATUS KeCreateThread(PROCESS* owner, uintptr_t entry, ui
     owner->ThreadCount++;
     KeReleaseSpinlock(&owner->HandleLock, flags);
 
+    KeInitThreadFpuContext(thread);
     KeReadyThread(thread);
     *out = thread;
     return STATUS_SUCCESS;
@@ -211,6 +215,7 @@ __declspec(dllexport) KSTATUS KeCreateKernelThread(uintptr_t fn, uintptr_t arg,
 
     if (out) *out = thread;
 
+    KeInitThreadFpuContext(thread);
     KeReadyThread(thread);
     return STATUS_SUCCESS;
 }
@@ -233,3 +238,19 @@ void KeUnblockThread(THREAD* thread) {
     
     KeReadyThread(thread);
 }
+
+void KeInitThreadFpuContext(THREAD* thread)
+{
+    uintptr_t raw =
+        (uintptr_t)thread->FpuStorage;
+
+    uintptr_t aligned =
+        (raw + 15) & ~(uintptr_t)15;
+
+    thread->FpuContext = (void*)aligned;
+
+    memset(thread->FpuContext, 0, 512);
+
+    *(uint16_t*)((uint8_t*)thread->FpuContext + 0x00) = 0x037F;
+    *(uint32_t*)((uint8_t*)thread->FpuContext + 0x18) = 0x1F80;
+}

code/kernel/ke_thread.h

@@ -73,6 +73,9 @@ typedef struct THREAD {
     // Allocated in KeCreateThread, signaled in KeExitThread before teardown.
     // Waiters block on this via SYS_WAIT_THREAD -> KeWaitForEvent.
     EVENT*          ExitEvent;
+
+    void*   FpuContext;
+    uint8_t FpuStorage[528];
 } THREAD;
 
 __declspec(dllexport) KSTATUS KeCreateThread(PROCESS* owner, uintptr_t entry, uintptr_t arg,
@@ -94,3 +97,5 @@ void    KeUnblockThread(THREAD* thread);
 
 // Assembly stub - implemented in ke_switch.asm.
 void    KeSwitchContext(THREAD* from, THREAD* to);
+
+void    KeInitThreadFpuContext(THREAD* thread);

code/shell_win32/shell_win32.c

@@ -131,9 +131,11 @@ static int runCommand(const char* cmd) {
     if (strncmp(cmd, "type ", 5) == 0) { cmd_type(cmd + 5); return 0; }
     if (strncmp(cmd, "write ", 6) == 0) { cmd_write(cmd + 6); return 0; }
 
-    STARTUPINFOA si = { sizeof(si) }; PROCESS_INFORMATION pi;
-    char cmdBuf[MAX_CMD]; strncpy(cmdBuf, cmd, MAX_CMD-1);
-    if (CreateProcessA(NULL, cmdBuf, NULL, NULL, FALSE, 0, NULL, NULL, &si, &pi)) {
+    PROCESS_INFORMATION pi;
+    char cmdBuf[MAX_CMD];
+    strncpy(cmdBuf, cmd, MAX_CMD - 1);
+    cmdBuf[MAX_CMD - 1] = '\0';
+    if (CreateProcessA(NULL, cmdBuf, NULL, NULL, FALSE, 0, NULL, NULL, NULL, &pi)) {
         WaitForSingleObject(pi.hProcess, INFINITE);
         CloseHandle(pi.hProcess); CloseHandle(pi.hThread);
     } else {

code/win32/kernel32.c

@@ -1139,69 +1139,51 @@ __declspec(dllexport) void WakeAllConditionVariable(CONDITION_VARIABLE* cv) { (v
 
 // ---------------------------------------------------------------------------
 // Timing & Performance Counters
-//
-// We use the CPU's Time Stamp Counter (TSC) for QueryPerformanceCounter so
-// that callers like Bink get a truly high-resolution, continuously-advancing
-// clock.  The TSC frequency is measured once on first use.
-//
-// Calibration strategy:
-//   1. Try CPUID leaf 0x15 (TSC / Crystal-clock ratio).  If the crystal
-//      frequency (ECX) is non-zero, compute TSC Hz = crystal_hz * EBX / EAX
-//      with no blocking at all.
-//   2. Otherwise fall back to measuring TSC ticks over a known wall-clock
-//      interval supplied by SysGetTick().  SysGetTick() returns raw LAPIC
-//      interrupt counts; each unit represents QUANTUM_MS = 10 ms.  We wait
-//      for two consecutive tick-boundary crossings and divide.
 // ---------------------------------------------------------------------------
 
-#define QUANTUM_MS 10   // must match hal_timer.c QUANTUM_MS
-
 typedef union {
     struct { DWORD LowPart; long HighPart; } u;
     long long QuadPart;
 } LARGE_INTEGER;
 
-static volatile long long gTscFreq = 0;   // TSC ticks per second; 0 = uninitialised
+static volatile long long gTscFreq = 0;
 
-// Attempt to read TSC frequency from CPUID leaf 0x15.
-// Returns 0 if the leaf is absent or the crystal frequency is not enumerated.
+// Attempt to read TSC frequency from CPUID leaf 0x15 (Modern Intel/AMD).
 static long long TscFreqFromCpuid(void) {
     int info[4];
     __cpuid(info, 0);
     if ((unsigned)info[0] < 0x15) return 0;
 
     __cpuid(info, 0x15);
-    unsigned denom      = (unsigned)info[0];   // EAX – TSC/crystal denominator
-    unsigned numer      = (unsigned)info[1];   // EBX – TSC/crystal numerator
-    unsigned crystal_hz = (unsigned)info[2];   // ECX – crystal clock in Hz
+    unsigned denom      = (unsigned)info[0]; 
+    unsigned numer      = (unsigned)info[1]; 
+    unsigned crystal_hz = (unsigned)info[2]; 
 
     if (!denom || !numer || !crystal_hz) return 0;
     return (long long)crystal_hz * numer / denom;
 }
 
-// Calibrate TSC frequency by timing over two LAPIC-tick boundaries.
-// Each SysGetTick() unit = QUANTUM_MS ms; we measure across N_TICKS units
-// for a total gate time of N_TICKS * QUANTUM_MS ms.
+// Calibrate TSC against SysGetTick (1 unit = 1 microsecond).
 static long long TscFreqCalibrate(void) {
-    const uint64_t N_TICKS = 10;   // 10 * 10 ms = 100 ms gate time
+    // We'll wait for 100ms (100,000 microseconds) to get a stable sample.
+    const uint64_t CALIBRATION_US = 100000; 
 
-    // Synchronise to the leading edge of a tick.
-    uint64_t t = SysGetTick();
-    while (SysGetTick() == t) { _mm_pause(); }
-
-    uint64_t tsc0 = __rdtsc();
     uint64_t t0   = SysGetTick();
+    uint64_t tsc0 = __rdtsc();
 
-    // Wait for N_TICKS more ticks.
-    while ((SysGetTick() - t0) < N_TICKS) { _mm_pause(); }
+    // Simple busy wait for the duration
+    while ((SysGetTick() - t0) < CALIBRATION_US) { 
+        _mm_pause(); 
+    }
 
-    uint64_t tsc1      = __rdtsc();
-    uint64_t ticks_got = SysGetTick() - t0;   // actual ticks elapsed (>= N_TICKS)
+    uint64_t t1   = SysGetTick();
+    uint64_t tsc1 = __rdtsc();
 
-    // ticks_got LAPIC ticks = ticks_got * QUANTUM_MS ms  =  ticks_got * QUANTUM_MS / 1000 s
-    // TSC Hz = (tsc1-tsc0) / (ticks_got * QUANTUM_MS / 1000)
-    //        = (tsc1-tsc0) * 1000 / (ticks_got * QUANTUM_MS)
-    return (long long)((tsc1 - tsc0) * 1000ULL / (ticks_got * (uint64_t)QUANTUM_MS));
+    uint64_t delta_tsc = tsc1 - tsc0;
+    uint64_t delta_us  = t1 - t0;
+
+    // Hz = (Delta TSC / Delta Microseconds) * 1,000,000
+    return (long long)(delta_tsc * 1000000ULL / delta_us);
 }
 
 static long long TscFreqGet(void) {
@@ -1220,8 +1202,7 @@ __declspec(dllexport) BOOL QueryPerformanceFrequency(LARGE_INTEGER* lpFrequency)
 }
 
 __declspec(dllexport) BOOL QueryPerformanceCounter(LARGE_INTEGER* lpPerformanceCount) {
-    // Ensure frequency is measured before the first counter read so that the
-    // baseline TSC and the frequency come from the same calibration window.
+    // Ensure calibration happens before the first read.
     TscFreqGet();
     if (lpPerformanceCount)
         lpPerformanceCount->QuadPart = (long long)__rdtsc();
@@ -1229,15 +1210,19 @@ __declspec(dllexport) BOOL QueryPerformanceCounter(LARGE_INTEGER* lpPerformanceC
 }
 
 __declspec(dllexport) void GetSystemTimeAsFileTime(void* lpSystemTimeAsFileTime) {
-    // Return a fixed FILETIME representing 2024-01-01 00:00:00 UTC.
-    // FILETIME = 100-nanosecond intervals since 1601-01-01.
-    // 2024-01-01 offset: 133186080000000000 * 100 ns
     if (!lpSystemTimeAsFileTime) return;
-    // SysGetTick() units = QUANTUM_MS ms = QUANTUM_MS * 10000 * 100-ns intervals.
-    unsigned long long ft = 133186080000000000ULL
-                          + SysGetTick() * ((unsigned long long)QUANTUM_MS * 10000ULL);
-    unsigned char* p = (unsigned char*)lpSystemTimeAsFileTime;
-    for (int i = 0; i < 8; i++) p[i] = (unsigned char)(ft >> (i * 8));
+
+    // Baseline: 2024-01-01 00:00:00 UTC in 100-ns intervals
+    const unsigned long long BASELINE_FT = 133485408000000000ULL;
+
+    // SysGetTick() is in microseconds.
+    // 1 microsecond = 10 units of 100-nanoseconds.
+    unsigned long long now_us = SysGetTick();
+    unsigned long long ft = BASELINE_FT + (now_us * 10ULL);
+
+    // Copy the 64-bit value into the output buffer.
+    // We use memcpy to safely handle potential alignment issues with the void pointer.
+    memcpy(lpSystemTimeAsFileTime, &ft, sizeof(ft));
 }
 
 // ---------------------------------------------------------------------------