hackjob_bench.patch

diff --git a/demo/main.c b/demo/main.c
index 4cebb12..70bbcaf 100644
--- a/demo/main.c
+++ b/demo/main.c
@@ -1,50 +1,145 @@
 #include "shared.h"
+#include <time.h>
+#include <unistd.h>
+#include <inttypes.h>
 
-int mtest_opponent(void);
-int unit_tests(int argc, char **argv);
-
-void ndraw(mp_int *a, const char *name)
+/* RDTSC from Scott Duplichan */
+static uint64_t TIMFUNC(void)
 {
-   char *buf;
-   int size;
-
-   mp_radix_size(a, 10, &size);
-   buf = (char *)malloc((size_t) size);
-   if (buf == NULL) {
-      fprintf(stderr, "\nndraw: malloc(%d) failed\n", size);
-      exit(EXIT_FAILURE);
-   }
-
-   printf("%s: ", name);
-   mp_toradix(a, buf, 10);
-   printf("%s\n", buf);
-   mp_toradix(a, buf, 16);
-   printf("0x%s\n", buf);
+#if defined __GNUC__
+#if defined(__i386__) || defined(__x86_64__)
+   /* version from http://www.mcs.anl.gov/~kazutomo/rdtsc.html
+    * the old code always got a warning issued by gcc, clang did not complain...
+    */
+   unsigned hi, lo;
+   __asm__ __volatile__("rdtsc" : "=a"(lo), "=d"(hi));
+   return ((uint64_t)lo)|(((uint64_t)hi)<<32);
+#else /* gcc-IA64 version */
+   unsigned long result;
+   __asm__ __volatile__("mov %0=ar.itc":"=r"(result)::"memory");
 
-   free(buf);
-}
+   while (__builtin_expect((int) result == -1, 0))
+      __asm__ __volatile__("mov %0=ar.itc":"=r"(result)::"memory");
 
-int main(int argc, char **argv)
-{
-#ifdef MP_8BIT
-   printf("Digit size 8 Bit \n");
-#endif
-#ifdef MP_16BIT
-   printf("Digit size 16 Bit \n");
+   return result;
 #endif
-#ifdef MP_32BIT
-   printf("Digit size 32 Bit \n");
+
+   /* Microsoft and Intel Windows compilers */
+#elif defined _M_IX86
+   __asm rdtsc
+#elif defined _M_AMD64
+   return __rdtsc();
+#elif defined _M_IA64
+#if defined __INTEL_COMPILER
+#include <ia64intrin.h>
 #endif
-#ifdef MP_64BIT
-   printf("Digit size 64 Bit \n");
+   return __getReg(3116);
+#else
+#error need rdtsc function for this build
 #endif
-   printf("Size of mp_digit: %u\n", (unsigned int)sizeof(mp_digit));
-   printf("Size of mp_word: %u\n", (unsigned int)sizeof(mp_word));
-   printf("MP_DIGIT_BIT: %d\n", MP_DIGIT_BIT);
-   printf("MP_PREC: %d\n", MP_PREC);
+}
+
+int main(void)
+{
+   uint64_t tt, gg, CLK_PER_SEC;
+   CLK_PER_SEC = TIMFUNC();
+   sleep(1);
+   CLK_PER_SEC = TIMFUNC() - CLK_PER_SEC;
+
+   printf("CLK_PER_SEC == %" PRIu64 "\n", CLK_PER_SEC);
+   for (int i = 1; i <= 100001; i *= 2) {
+      mp_int a, b;
+      char *buf_a, *buf_b, *buf_c, *buf_a2, *buf_b2, *buf_c2;
+      int size_a, size_b, size_c, size_c2;
+      unsigned rr;
+      mp_init(&a);
+      mp_2expt(&a, i);
+      mp_init_i32(&b, 13);
+      mp_expt_u32(&b, i, &b);
 
-   if (LTM_DEMO_TEST_VS_MTEST) {
-      return mtest_opponent();
+      mp_radix_size(&a, 10, &size_a);
+      mp_radix_size(&b, 10, &size_b);
+      mp_radix_size(&a, 16, &size_c);
+      mp_radix_size(&b, 16, &size_c2);
+      buf_a = (char *)malloc((size_t) size_a);
+      buf_b = (char *)malloc((size_t) size_b);
+      buf_c = (char *)malloc((size_t) size_c);
+      buf_a2 = (char *)malloc((size_t) size_a);
+      buf_b2 = (char *)malloc((size_t) size_b);
+      buf_c2 = (char *)malloc((size_t) size_c2);
+      rr = 0u;
+      tt = UINT64_MAX;
+      do {
+         gg = TIMFUNC();
+         mp_tohex(&a, buf_c);
+         gg = (TIMFUNC() - gg) >> 1;
+         if (tt > gg)
+            tt = gg;
+      } while (++rr < 100u);
+      printf("mp_tohex of 2**%d => %9" PRIu64 "/sec, %9" PRIu64 " cycles, %0.4f seconds\n",
+             i, CLK_PER_SEC / tt, tt, (float) tt / (float) CLK_PER_SEC);
+      rr = 0u;
+      tt = UINT64_MAX;
+      do {
+         gg = TIMFUNC();
+         mp_toradix(&a, buf_a, 10);
+         gg = (TIMFUNC() - gg) >> 1;
+         if (tt > gg)
+            tt = gg;
+      } while (++rr < 100u);
+      printf("mp_torad of 2**%d => %9" PRIu64 "/sec, %9" PRIu64 " cycles, %0.4f seconds\n",
+             i, CLK_PER_SEC / tt, tt, (float) tt / (float) CLK_PER_SEC);
+      rr = 0u;
+      tt = UINT64_MAX;
+      do {
+         gg = TIMFUNC();
+         mp_todecimal(&a, buf_a2);
+         gg = (TIMFUNC() - gg) >> 1;
+         if (tt > gg)
+            tt = gg;
+      } while (++rr < 100u);
+      printf("mp_todec of 2**%d => %9" PRIu64 "/sec, %9" PRIu64 " cycles, %0.4f seconds\n",
+             i, CLK_PER_SEC / tt, tt, (float) tt / (float) CLK_PER_SEC);
+      rr = 0u;
+      tt = UINT64_MAX;
+      do {
+         gg = TIMFUNC();
+         mp_tohex(&b, buf_c2);
+         gg = (TIMFUNC() - gg) >> 1;
+         if (tt > gg)
+            tt = gg;
+      } while (++rr < 100u);
+      printf("mp_tohex of 13**%d => %9" PRIu64 "/sec, %9" PRIu64 " cycles, %0.4f seconds\n",
+             i, CLK_PER_SEC / tt, tt, (float) tt / (float) CLK_PER_SEC);
+      rr = 0u;
+      tt = UINT64_MAX;
+      do {
+         gg = TIMFUNC();
+         mp_toradix(&b, buf_b, 10);
+         gg = (TIMFUNC() - gg) >> 1;
+         if (tt > gg)
+            tt = gg;
+      } while (++rr < 100u);
+      printf("mp_torad of 13**%d => %9" PRIu64 "/sec, %9" PRIu64 " cycles, %0.4f seconds\n",
+             i, CLK_PER_SEC / tt, tt, (float) tt / (float) CLK_PER_SEC);
+      rr = 0u;
+      tt = UINT64_MAX;
+      do {
+         gg = TIMFUNC();
+         mp_todecimal(&b, buf_b2);
+         gg = (TIMFUNC() - gg) >> 1;
+         if (tt > gg)
+            tt = gg;
+      } while (++rr < 100u);
+      printf("mp_todec of 13**%d => %9" PRIu64 "/sec, %9" PRIu64 " cycles, %0.4f seconds\n\n",
+             i, CLK_PER_SEC / tt, tt, (float) tt / (float) CLK_PER_SEC);
+      free(buf_a);
+      free(buf_b);
+      free(buf_c);
+      free(buf_a2);
+      free(buf_b2);
+      free(buf_c2);
    }
-   return unit_tests(argc, argv);
+
+   return 0;
 }
diff --git a/makefile b/makefile
index fabd2c4..f5273bd 100644
--- a/makefile
+++ b/makefile
@@ -21,7 +21,7 @@ include makefile_include.mk
 ifneq ($V,1)
        @echo "   * ${CC} $@"
 endif
-       ${silent} ${CC} -c ${CFLAGS} $< -o $@
+       ${silent} ${CC} -g -c ${CFLAGS} $< -o $@
 
 LCOV_ARGS=--directory .
 
@@ -102,6 +102,9 @@ uninstall:
 test: demo/main.o demo/opponent.o demo/test.o $(LIBNAME)
        $(CC) $(CFLAGS) $^ $(LFLAGS) -o test
 
+main: demo/main.o $(LIBNAME)
+       $(CC) -g $(CFLAGS) $^ $(LFLAGS) -o main
+
 test_standalone: demo/main.o demo/opponent.o demo/test.o $(LIBNAME)
        $(CC) $(CFLAGS) $^ $(LFLAGS) -o test