Make performance counters always 64 bits wide (#178)

This prevents the counters from easily overflowing for a 32-bit pk.

5 files changed, 93 insertions, 21 deletions

diff --git a/pk/boot.h b/pk/boot.h
index f16ab15..f8bc51c 100644
--- a/pk/boot.h
+++ b/pk/boot.h
@@ -20,9 +20,9 @@ typedef struct {
   size_t brk_max;
   size_t mmap_max;
   size_t stack_top;
-  size_t time0;
-  size_t cycle0;
-  size_t instret0;
+  uint64_t time0;
+  uint64_t cycle0;
+  uint64_t instret0;
 } elf_info;
 
 extern elf_info current;
diff --git a/pk/pk.c b/pk/pk.c
index b8b0b1c..8ba9dc3 100644
--- a/pk/pk.c
+++ b/pk/pk.c
@@ -132,9 +132,9 @@ static void run_loaded_program(size_t argc, char** argv, uintptr_t kstack_top)
   STACK_INIT(uintptr_t);
 
   if (current.cycle0) { // start timer if so requested
-    current.time0 = rdtime();
-    current.cycle0 = rdcycle();
-    current.instret0 = rdinstret();
+    current.time0 = rdtime64();
+    current.cycle0 = rdcycle64();
+    current.instret0 = rdinstret64();
   }
 
   trapframe_t tf;
diff --git a/pk/pk.h b/pk/pk.h
index 1167f6e..a55ab6c 100644
--- a/pk/pk.h
+++ b/pk/pk.h
@@ -41,6 +41,64 @@ static inline int insn_len(long insn)
   return (insn & 0x3) < 0x3 ? 2 : 4;
 }
 
+#if __riscv_xlen == 32
+
+static inline uint64_t rdtime64()
+{
+  uint32_t time;
+  uint32_t timeh1;
+  uint32_t timeh2;
+
+  do
+  {
+    timeh1 = read_csr(timeh);
+    time = read_csr(time);
+    timeh2 = read_csr(timeh);
+  } while(timeh1 != timeh2);
+
+  return (((uint64_t) timeh1) << 32) | time;
+}
+
+static inline uint64_t rdcycle64()
+{
+  uint32_t cycle;
+  uint32_t cycleh1;
+  uint32_t cycleh2;
+
+  do
+  {
+    cycleh1 = read_csr(cycleh);
+    cycle = read_csr(cycle);
+    cycleh2 = read_csr(cycleh);
+  } while(cycleh1 != cycleh2);
+
+  return (((uint64_t) cycleh1) << 32) | cycle;
+}
+
+static inline uint64_t rdinstret64()
+{
+  uint32_t instret;
+  uint32_t instreth1;
+  uint32_t instreth2;
+
+  do
+  {
+    instreth1 = read_csr(instreth);
+    instret = read_csr(instret);
+    instreth2 = read_csr(instreth);
+  } while(instreth1 != instreth2);
+
+  return (((uint64_t) instreth1) << 32) | instret;
+}
+
+#else
+
+#define rdtime64 rdtime
+#define rdcycle64 rdcycle
+#define rdinstret64 rdinstret
+
+#endif
+
 #define ARRAY_SIZE(x) (sizeof(x)/sizeof((x)[0]))
 
 #ifdef __cplusplus
diff --git a/pk/syscall.c b/pk/syscall.c
index 8903ecc..07fa41d 100644
--- a/pk/syscall.c
+++ b/pk/syscall.c
@@ -16,14 +16,15 @@ typedef long (*syscall_t)(long, long, long, long, long, long, long);
 void sys_exit(int code)
 {
   if (current.cycle0) {
-    size_t dt = rdtime() - current.time0;
-    size_t dc = rdcycle() - current.cycle0;
-    size_t di = rdinstret() - current.instret0;
-
-    printk("%ld ticks\n", dt);
-    printk("%ld cycles\n", dc);
-    printk("%ld instructions\n", di);
-    printk("%d.%d%d CPI\n", dc/di, 10ULL*dc/di % 10, (100ULL*dc + di/2)/di % 10);
+    uint64_t dt = rdtime64() - current.time0;
+    uint64_t dc = rdcycle64() - current.cycle0;
+    uint64_t di = rdinstret64() - current.instret0;
+
+    printk("%lld ticks\n", dt);
+    printk("%lld cycles\n", dc);
+    printk("%lld instructions\n", di);
+    printk("%d.%d%d CPI\n", (int)(dc/di), (int)(10ULL*dc/di % 10),
+        (int)((100ULL*dc + di/2)/di % 10));
   }
   shutdown(code);
 }
@@ -353,7 +354,7 @@ int sys_rt_sigaction(int sig, const void* act, void* oact, size_t sssz)
 
 long sys_time(long* loc)
 {
-  uintptr_t t = rdcycle() / CLOCK_FREQ;
+  uint64_t t = rdcycle64() / CLOCK_FREQ;
   if (loc)
     *loc = t;
   return t;
@@ -361,7 +362,7 @@ long sys_time(long* loc)
 
 int sys_times(long* loc)
 {
-  uintptr_t t = rdcycle();
+  uint64_t t = rdcycle64();
   kassert(CLOCK_FREQ % 1000000 == 0);
   loc[0] = t / (CLOCK_FREQ / 1000000);
   loc[1] = 0;
@@ -373,7 +374,7 @@ int sys_times(long* loc)
 
 int sys_gettimeofday(long* loc)
 {
-  uintptr_t t = rdcycle();
+  uint64_t t = rdcycle64();
   loc[0] = t / CLOCK_FREQ;
   loc[1] = (t % CLOCK_FREQ) / (CLOCK_FREQ / 1000000);
   
@@ -382,7 +383,7 @@ int sys_gettimeofday(long* loc)
 
 long sys_clock_gettime(int clk_id, long *loc)
 {
-  uintptr_t t = rdcycle();
+  uint64_t t = rdcycle64();
   loc[0] = t / CLOCK_FREQ;
   loc[1] = (t % CLOCK_FREQ) / (CLOCK_FREQ / 1000000000);
 
diff --git a/util/snprintf.c b/util/snprintf.c
index 1544a6c..a11a4f7 100644
--- a/util/snprintf.c
+++ b/util/snprintf.c
@@ -9,6 +9,7 @@ int vsnprintf(char* out, size_t n, const char* s, va_list vl)
 {
   bool format = false;
   bool longarg = false;
+  bool longlongarg = false;
   size_t pos = 0;
   for( ; *s; s++)
   {
@@ -17,7 +18,12 @@ int vsnprintf(char* out, size_t n, const char* s, va_list vl)
       switch(*s)
       {
         case 'l':
-          longarg = true;
+          if (s[1] == 'l') {
+              longlongarg = true;
+              s++;
+          }
+          else
+              longarg = true;
           break;
         case 'p':
           longarg = true;
@@ -36,13 +42,19 @@ int vsnprintf(char* out, size_t n, const char* s, va_list vl)
         }
         case 'd':
         {
-          long num = longarg ? va_arg(vl, long) : va_arg(vl, int);
+          long long num;
+          if (longarg)
+              num = va_arg(vl, long);
+          else if (longlongarg)
+              num = va_arg(vl, long long);
+          else
+              num = va_arg(vl, int);
           if (num < 0) {
             num = -num;
             if (++pos < n) out[pos-1] = '-';
           }
           long digits = 1;
-          for (long nn = num; nn /= 10; digits++)
+          for (long long nn = num; nn /= 10; digits++)
             ;
           for (int i = digits-1; i >= 0; i--) {
             if (pos + i + 1 < n) out[pos + i] = '0' + (num % 10);
@@ -50,6 +62,7 @@ int vsnprintf(char* out, size_t n, const char* s, va_list vl)
           }
           pos += digits;
           longarg = false;
+          longlongarg = false;
           format = false;
           break;
         }