Changed, again, RDTSC and Hardware counter, introducing auto calibration when hardware counter is too slow for modern standard (and removed BOX64_RDTSC env. var.)

6 files changed, 58 insertions, 42 deletions

diff --git a/src/dynarec/arm64/dynarec_arm64_0f.c b/src/dynarec/arm64/dynarec_arm64_0f.c
index 0bb9dffc..8e8ec00e 100644
--- a/src/dynarec/arm64/dynarec_arm64_0f.c
+++ b/src/dynarec/arm64/dynarec_arm64_0f.c
@@ -98,6 +98,9 @@ uintptr_t dynarec64_0F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
                     } else {

                         MRS_cntvct_el0(x1);

                     }

+                    if(box64_rdtsc_shift) {

+                        LSLx(x1, x1, box64_rdtsc_shift);

+                    }

                     LSRx(xRDX, x1, 32);

                     MOVw_REG(xRAX, x1);   // wipe upper part

                     MOVw_REG(xRCX, xZR);    // IA32_TSC, 0 for now

@@ -499,6 +502,9 @@ uintptr_t dynarec64_0F(dynarec_arm_t* dyn, uintptr_t addr, uintptr_t ip, int nin
             } else {

                 MRS_cntvct_el0(x1);

             }

+            if(box64_rdtsc_shift) {

+                LSLx(x1, x1, box64_rdtsc_shift);

+            }

             LSRx(xRDX, x1, 32);

             MOVw_REG(xRAX, x1);   // wipe upper part

             break;

diff --git a/src/dynarec/rv64/dynarec_rv64_0f.c b/src/dynarec/rv64/dynarec_rv64_0f.c
index bfe29a2c..d19ecd4e 100644
--- a/src/dynarec/rv64/dynarec_rv64_0f.c
+++ b/src/dynarec/rv64/dynarec_rv64_0f.c
@@ -87,6 +87,9 @@ uintptr_t dynarec64_0F(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
                         } else {
                             CSRRS(x3, xZR, 0xC01); // RDTIME
                         }
+                        if(box64_rdtsc_shift) {
+                            SRLI(x3, x3, box64_rdtsc_shift);
+                        }
                         SRLI(xRDX, x3, 32);
                         AND(xRAX, x3, xMASK); // wipe upper part
                         MV(xRCX, xZR);    // IA32_TSC, 0 for now
@@ -412,6 +415,9 @@ uintptr_t dynarec64_0F(dynarec_rv64_t* dyn, uintptr_t addr, uintptr_t ip, int ni
             } else {
                 CSRRS(x3, xZR, 0xC01); // RDTIME
             }
+            if(box64_rdtsc_shift) {
+                SRLI(x3, x3, box64_rdtsc_shift);
+            }
             SRLI(xRDX, x3, 32);
             AND(xRAX, x3, xMASK); // wipe upper part
             break;
diff --git a/src/emu/x64run0f.c b/src/emu/x64run0f.c
index f14b9dd5..e9c2c761 100644
--- a/src/emu/x64run0f.c
+++ b/src/emu/x64run0f.c
@@ -107,6 +107,8 @@ uintptr_t Run0F(x64emu_t *emu, rex_t rex, uintptr_t addr, int *step)
                     break;

                 case 0xF9:  /* RDTSCP */

                     tmp64u = ReadTSC(emu);

+                    if(box64_rdtsc_shift)

+                        tmp64u<<=box64_rdtsc_shift;

                     R_RAX = tmp64u & 0xffffffff;

                     R_RDX = tmp64u >> 32;

                     R_RCX = 0;  // should be low of IA32_TSC

@@ -368,6 +370,8 @@ uintptr_t Run0F(x64emu_t *emu, rex_t rex, uintptr_t addr, int *step)
 

         case 0x31:                   /* RDTSC */

             tmp64u = ReadTSC(emu);

+            if(box64_rdtsc_shift)

+                tmp64u<<=box64_rdtsc_shift;

             R_RDX = tmp64u>>32;

             R_RAX = tmp64u&0xFFFFFFFF;

             break;

diff --git a/src/include/debug.h b/src/include/debug.h
index a6660d7d..422a8f6e 100644
--- a/src/include/debug.h
+++ b/src/include/debug.h
@@ -14,6 +14,7 @@ extern int box64_maxcpu;
 extern int box64_mmap32;
 extern int box64_ignoreint3;
 extern int box64_rdtsc;
+extern uint8_t box64_rdtsc_shift;
 #ifdef DYNAREC
 extern int box64_dynarec_dump;
 extern int box64_dynarec_trace;
diff --git a/src/main.c b/src/main.c
index 4d9eb6ee..38c44df1 100644
--- a/src/main.c
+++ b/src/main.c
@@ -63,6 +63,7 @@ int box64_mmap32 = 0;
 #endif
 int box64_ignoreint3 = 0;
 int box64_rdtsc = 0;
+uint8_t box64_rdtsc_shift = 0;
 #ifdef DYNAREC
 int box64_dynarec = 1;
 int box64_dynarec_dump = 0;
@@ -1021,39 +1022,54 @@ void LoadLogEnv()
         if(box64_ignoreint3)
             printf_log(LOG_INFO, "Will silently ignore INT3 in the code\n");
     }
-    p = getenv("BOX64_RDTSC");
-        if(p) {
-        if(strlen(p)==1) {
-            if(p[0]>='0' && p[0]<='0'+2)
-                box64_rdtsc = p[0]-'0';
-        }
-        if(box64_rdtsc==2) {
-            #if defined(ARM64) || defined(RV64)
-            box64_rdtsc = 0;    // allow hardxare counter
-            uint64_t freq = ReadTSCFrequency(NULL);
-            printf_log(LOG_INFO, "Hardware counter measured at %d Mhz, ", freq/1000);
-            if(freq>1000000000) {
-                printf_log(LOG_INFO, "keeping it\n");
-            } else {
-                box64_rdtsc = 1;
-                printf_log(LOG_INFO, "not using it\n");
-            }
-            #else
-            box64_rdtsc = 1;
-            printf_log(LOG_INFO, "Will use time-based emulation for rdtsc, even if hardware counter are available\n");
-            #endif
-        } else if(box64_rdtsc)
-            printf_log(LOG_INFO, "Will use time-based emulation for rdtsc, even if hardware counter are available\n");
-    }
+    // grab pagesize
     box64_pagesize = sysconf(_SC_PAGESIZE);
     if(!box64_pagesize)
         box64_pagesize = 4096;
 #ifdef DYNAREC
+    // grab cpu extensions for dynarec usage
     GatherDynarecExtensions();
 #endif
+    // grab cpu name
     int ncpu = getNCpu();
     const char* cpuname = getCpuName();
     printf_log(LOG_INFO, " PageSize:%zd Running on %s with %d Cores\n", box64_pagesize, cpuname, ncpu);
+    // grab and calibrate hardware counter
+    int hardware  = 0;
+    #if defined(ARM64) || defined(RV64)
+    hardware = 1;
+    box64_rdtsc = 0;    // allow hardxare counter
+    #else
+    box64_rdtsc = 1;
+    printf_log(LOG_INFO, "Will use time-based emulation for rdtsc, even if hardware counter are available\n");
+    #endif
+    uint64_t freq = ReadTSCFrequency(NULL);
+    if(freq<1000000) {
+        box64_rdtsc = 1;
+        if(hardware) printf_log(LOG_INFO, "Hardware counter to slow (%d kHz), not using it\n", freq/1000);
+        hardware = 0;
+        freq = ReadTSCFrequency(NULL);
+    }
+    uint64_t efreq = freq;
+    while(efreq<500000000) {    // minium 500MHz
+        ++box64_rdtsc_shift;
+        efreq = freq<<box64_rdtsc_shift;
+    }
+    printf_log(LOG_INFO, "Will use %s counter measured at ", box64_rdtsc?"Software":"Hardware");
+    int ghz = freq>=1000000000LL;
+    if(ghz) freq/=100000000LL; else freq/=100000;
+    if(ghz) printf_log(LOG_INFO, "%d.%d GHz", freq/10, freq%10);
+    if(!ghz & freq>=1000) printf_log(LOG_INFO, "%d MHz", freq/10);
+    if(!ghz & freq<1000) printf_log(LOG_INFO, "%d.%d MHz", freq/10, freq%10);
+    if(box64_rdtsc_shift) {
+        printf_log(LOG_INFO, " emulating ");
+        ghz = efreq>=1000000000LL;
+        if(ghz) efreq/=100000000LL; else efreq/=100000;
+        if(ghz) printf_log(LOG_INFO, "%d.%d GHz", efreq/10, efreq%10);
+        if(!ghz & efreq>=1000) printf_log(LOG_INFO, "%d MHz", efreq/10);
+        if(!ghz & efreq<1000) printf_log(LOG_INFO, "%d.%d MHz", efreq/10, efreq%10);
+    }
+    printf_log(LOG_INFO, "\n");
 }
 
 EXPORTDYN
@@ -1186,7 +1202,6 @@ void PrintFlags() {
     printf(" BOX64_ENV1='XXX=yyyy' will add XXX=yyyy env. var. and continue with BOX86_ENV2 ... until var doesn't exist\n");
     printf(" BOX64_JITGDB with 1 to launch \"gdb\" when a segfault is trapped, attached to the offending process\n");
     printf(" BOX64_MMAP32=1 to use 32bits address space mmap in priority for external mmap as soon a 32bits process are detected (default for Snapdragon build)\n");
-    printf(" BOX64_RDTSC to use a monotonic timer for rdtsc even if hardware counter are available (or check if precision is >=1Ghz for 2)\n");
 }
 
 void PrintHelp() {
diff --git a/src/tools/rcfile.c b/src/tools/rcfile.c
index 87cc3dd9..e6c40f2e 100644
--- a/src/tools/rcfile.c
+++ b/src/tools/rcfile.c
@@ -81,7 +81,7 @@ ENTRYBOOL(BOX64_SHOWSEGV, box64_showsegv)               \
 ENTRYBOOL(BOX64_SHOWBT, box64_showbt)                   \
 ENTRYBOOL(BOX64_MMAP32, box64_mmap32)                   \
 ENTRYBOOL(BOX64_IGNOREINT3, box64_ignoreint3)           \
-ENTRYINT(BOX64_RDTSC, box64_rdtsc, 0, 2, 2)             \
+IGNORE(BOX64_RDTSC)                                     \
 ENTRYBOOL(BOX64_X11THREADS, box64_x11threads)           \
 ENTRYBOOL(BOX64_X11GLX, box64_x11glx)                   \
 ENTRYDSTRING(BOX64_LIBGL, box64_libGL)                  \
@@ -593,22 +593,6 @@ void ApplyParams(const char* name)
         my_context->bashpath = strdup(param->bash);
         printf_log(LOG_INFO, "Applying %s=%s\n", "BOX64_BASH", param->bash);
     }
-    if(param->is_box64_rdtsc_present && (box64_rdtsc==2)) {
-        #if defined(ARM64) || defined(RV64)
-        box64_rdtsc = 0;    // allow hardxware counter
-        uint64_t freq = ReadTSCFrequency(NULL);
-        printf_log(LOG_INFO, "Applying RDTSC: Hardware counter measured at %d Mhz, ", freq/1000);
-        if(freq>1000000000) {
-            printf_log(LOG_INFO, "keeping it\n");
-        } else {
-            box64_rdtsc = 1;
-            printf_log(LOG_INFO, "not using it\n");
-        }
-        #else
-        box64_rdtsc = 1;
-        printf_log(LOG_INFO, "Applying RDTSC: Will use time-based emulation for rdtsc, even if hardware counter are available\n");
-        #endif
-    }
     #ifdef HAVE_TRACE
     int old_x64trace = my_context->x64trace;
     if(param->is_trace_present) {