winapi: Less assembly, good enough(tm)

XboxDev · Oct 24, 2023 · 2f6d75b · 2f6d75b
1 parent eb4385b
commit 2f6d75b
Showing 1 changed file with 19 additions and 38 deletions.
diff --git a/lib/winapi/profiling.c b/lib/winapi/profiling.c
@@ -2,6 +2,7 @@
 
 // SPDX-FileCopyrightText: 2019 Stefan Schmidt
 
+#define USE_RDTSC_FOR_FREQ
 #include <profileapi.h>
 #ifdef USE_RDTSC_FOR_FREQ
 #include <synchapi.h>
@@ -19,6 +20,7 @@ static void __attribute__((constructor)) PrimeQueryPerformanceFrequency ()
     #define NV_PTIMER_NUM *(volatile ULONG *)0xFD009200
     #define NV_PTIMER_DEN *(volatile ULONG *)0xFD009210
     #define NV_PTIMER_COUNT 0xFD009400
+    #define ptr_PTIMER_COUNT *(volatile ULONG *)0xFD009400
     #define ASM_LOOPS 1024 * 4
 
     ULARGE_INTEGER rdtsc_count_1 = {{0, 0}}, rdtsc_count_2 = {{0, 0}};
@@ -33,60 +35,38 @@ static void __attribute__((constructor)) PrimeQueryPerformanceFrequency ()
 
     KeEnterCriticalRegion();
 
+    // Turn off caches
     __asm
     {
-        push eax
-        push edx
-        push ecx
-
         cli
         sfence
-
-        // Turn off caches
         mov eax, cr0
         or eax, 1 << 30 // Set CD bit
         mov cr0, eax
         wbinvd
 
-        // Reset PTIMER
-        mov eax, [NV_PTIMER_COUNT]
-        and eax, ~(0xFFFFFFE0) // First 5 bits are not used
-        mov [NV_PTIMER_COUNT], eax
-
-        rdtsc
-        mov rdtsc_count_1.LowPart, eax
-        mov rdtsc_count_1.HighPart, edx
-
-        mov eax, [NV_PTIMER_COUNT]
-        mov ptimer_count_1, eax
-
-        // Spin for a bit
-        mov eax, ASM_LOOPS
-        loop_1:
-        dec eax
-        jnz loop_1
-
-        rdtsc
-        mov rdtsc_count_2.LowPart, eax
-        mov rdtsc_count_2.HighPart, edx
-
-        mov eax, [NV_PTIMER_COUNT]
-        mov ptimer_count_2, eax
-
-        // Without this, invaldidating the cache below will crash the system
-        sfence
+    }
+
+    ptr_PTIMER_COUNT &= ~(0xFFFFFFE0); // First 5 bits are not used
+
+    rdtsc_count_1.QuadPart = __rdtsc();
+    ptimer_count_1 = ptr_PTIMER_COUNT;
+
+    KeStallExecutionProcessor(10);
 
+    rdtsc_count_2.QuadPart = __rdtsc();
+    ptimer_count_2 = ptr_PTIMER_COUNT;
+
+    __asm
+    {
+        sfence
         mov eax, cr0
         and eax, ~(1 << 30) // Clear CD bit
         mov cr0, eax
         wbinvd
-
         sti
-
-        pop ecx
-        pop edx
-        pop eax
     }
+
     KeLeaveCriticalRegion();
 
     double ptimer_diff = (ptimer_count_2 >> 5) - (ptimer_count_1 >> 5);
@@ -116,6 +96,7 @@ BOOL QueryPerformanceFrequency (LARGE_INTEGER *lpFrequency)
     assert(lpFrequency != NULL);
 
 #ifdef USE_RDTSC_FOR_FREQ
+    PrimeQueryPerformanceFrequency();
     lpFrequency->QuadPart = frequency.QuadPart;
 #else
     lpFrequency->QuadPart = 733333333;