diff --git a/demo/stest.c b/demo/stest.c
index c241e51..747ed89 100644
--- a/demo/stest.c
+++ b/demo/stest.c
@@ -11,8 +11,29 @@
    #define DISPLAY_P(...) printf(__VA_ARGS__)
 #else
    #define DISPLAY_P(...) (void)0
+   #define fp_dump(n,p) do{}while(0)
 #endif
 
+#ifndef fp_dump
+void fp_dump(const char* n, fp_int* p)
+{
+  int sz;
+  if (fp_radix_size(p, 2, &sz) != FP_OKAY)
+    return;
+  char* str = malloc(sz);
+  if (!str)
+    return;
+#ifdef STEST_VERBOSE
+  fp_toradix(p, str, 2);
+  DISPLAY_P("%s = 0b%s\n", n, str);
+  fp_toradix(p, str, 16);
+  DISPLAY_P("%s = 0x%s\n", n, str);
+#endif
+  fp_toradix(p, str, 10);
+  DISPLAY_P("%s = %s\n", n, str);
+  free(str);
+}
+#endif
 
 #ifdef GBA_MODE
 int c_main(void)
@@ -45,7 +66,7 @@ int main(void)
    fp_mul(&a, &b, &d);
    if (fp_cmp(&c, &d)) {
       DISPLAY("mul failed\n");
-      return 0;
+      return -1;
    } else {
       DISPLAY("mul passed\n");
    }
@@ -57,7 +78,7 @@ int main(void)
    fp_mul(&a, &b, &d);
    if (fp_cmp(&c, &d)) {
       DISPLAY("mul failed\n");
-      return 0;
+      return -1;
    } else {
       DISPLAY("mul passed\n");
    }
@@ -69,7 +90,7 @@ int main(void)
    fp_mul(&a, &b, &d);
    if (fp_cmp(&c, &d)) {
       DISPLAY("mul failed\n");
-      return 0;
+      return -1;
    } else {
       DISPLAY("mul passed\n");
    }
@@ -80,7 +101,7 @@ int main(void)
    fp_sqr(&a, &c);
    if (fp_cmp(&c, &b)) {
       DISPLAY("sqr failed\n");
-      return 0;
+      return -1;
    } else {
       DISPLAY("sqr passed\n");
    }
@@ -90,7 +111,7 @@ int main(void)
    fp_sqr(&a, &c);
    if (fp_cmp(&c, &b)) {
       DISPLAY("sqr failed\n");
-      return 0;
+      return -1;
    } else {
       DISPLAY("sqr passed\n");
    }
@@ -100,7 +121,7 @@ int main(void)
    fp_sqr(&a, &c);
    if (fp_cmp(&c, &b)) {
       DISPLAY("sqr failed\n");
-      return 0;
+      return -1;
    } else {
       DISPLAY("sqr passed\n");
    }
@@ -109,12 +130,19 @@ int main(void)
    /* montgomery reductions */
    fp_read_radix(&a, "234892374892374893489123428937892781237863278637826327367637836278362783627836783678363", 10);
    fp_read_radix(&b, "4447823492749823749234123489273987393983289319382762756425425425642727352327452374521", 10);
+#ifdef FP_64BIT
+   fp_read_radix(&c, "942974496560863503657226741422301598807235487941674147660989764036913926327577165648", 10);
+#else
    fp_read_radix(&c, "2396271882990732698083317035605836523697277786556053771759862552557086442129695099100", 10);
-   fp_montgomery_setup(&b, &dp);
+#endif
+   if (fp_montgomery_setup(&b, &dp) != FP_OKAY)
+      DISPLAY("mont setup failed\n");
    fp_montgomery_reduce(&a, &b, dp);
    if (fp_cmp(&a, &c)) {
       DISPLAY("mont failed\n");
-      return 0;
+      fp_dump("a (is    )", &a);
+      fp_dump("c (should)", &c);
+      return -1;
    } else {
       DISPLAY("mont passed\n");
    }
@@ -122,11 +150,14 @@ int main(void)
    fp_read_radix(&a, "2348923748923748934891234456645654645645684576353428937892781237863278637826327367637836278362783627836783678363", 10);
    fp_read_radix(&b, "444782349274982374923412348927398739398328931938276275642542542564272735232745237452123424324324444121111119", 10);
    fp_read_radix(&c, "45642613844554582908652603086180267403823312390990082328515008314514368668691233331246183943400359349283420", 10);
-   fp_montgomery_setup(&b, &dp);
+   if (fp_montgomery_setup(&b, &dp) != FP_OKAY)
+      DISPLAY("mont setup failed\n");
    fp_montgomery_reduce(&a, &b, dp);
    if (fp_cmp(&a, &c)) {
       DISPLAY("mont failed\n");
-      return 0;
+      fp_dump("a (is    )", &a);
+      fp_dump("c (should)", &c);
+      return -1;
    } else {
       DISPLAY("mont passed\n");
    }
@@ -134,11 +165,14 @@ int main(void)
    fp_read_radix(&a, "234823424242342923748923748934891234456645654645645684576353424972378234762378623891236834132352375235378462378489378927812378632786378263273676378362783627555555555539568389052478124618461834763837685723645827529034853490580134568947341278498542893481762349723907847892983627836783678363", 10);
    fp_read_radix(&b, "44478234927456563455982374923412348927398739398328931938276275642485623481638279025465891276312903262837562349056234783648712314678120389173890128905425242424239784256427", 10);
    fp_read_radix(&c, "33160865265453361650564031464519042126185632333462754084489985719613480783282357410514898819797738034600484519472656152351777186694609218202276509271061460265488348645081", 10);
-   fp_montgomery_setup(&b, &dp);
+   if (fp_montgomery_setup(&b, &dp) != FP_OKAY)
+      DISPLAY("mont setup failed\n");
    fp_montgomery_reduce(&a, &b, dp);
    if (fp_cmp(&a, &c)) {
       DISPLAY("mont failed\n");
-      return 0;
+      fp_dump("a (is    )", &a);
+      fp_dump("c (should)", &c);
+      return -1;
    } else {
       DISPLAY("mont passed\n");
    }
diff --git a/demo/test.c b/demo/test.c
index 596f7bd..43981d5 100644
--- a/demo/test.c
+++ b/demo/test.c
@@ -8,7 +8,11 @@ void draw(fp_int *a)
   int x;
   printf("%d, %d, ", a->used, a->sign);
   for (x = a->used - 1; x >= 0; x--) {
+#if SIZEOF_FP_DIGIT == 4
       printf("%08lx ", a->dp[x]);
+#else
+      printf("%016llx ", a->dp[x]);
+#endif
   }
   printf("\n");
 }
@@ -16,10 +20,12 @@ void draw(fp_int *a)
 int myrng(unsigned char *dst, int len, void *dat)
 {
    int x;
+   (void)dat;
    for (x = 0; x < len; x++) dst[x] = rand() & 0xFF;
    return len;
 }
 
+#ifndef TESTING
 /* RDTSC from Scott Duplichan */
 static ulong64 TIMFUNC (void)
    {
@@ -68,22 +74,30 @@ static ulong64 TIMFUNC (void)
      #error need rdtsc function for this build
    #endif
    }
+#endif
 
    char cmd[4096], buf[4096];
 
 int main(void)
 {
   fp_int a,b,c,d,e,f;
+  unsigned long expt_n, add_n, sub_n, mul_n, div_n, sqr_n, mul2d_n, div2d_n, gcd_n, lcm_n, inv_n,
+                 div2_n, mul2_n, add_d_n, sub_d_n, mul_d_n, cnt, rr, ix;
+#ifndef TESTING
+  unsigned long t;
   fp_digit fp;
   int n, err;
-   unsigned long expt_n, add_n, sub_n, mul_n, div_n, sqr_n, mul2d_n, div2d_n, gcd_n, lcm_n, inv_n,
-                 div2_n, mul2_n, add_d_n, sub_d_n, mul_d_n, t, cnt, rr, ix;
-   ulong64 t1, t2;
+  ulong64 t1, t2;
+#endif
 
   srand(time(NULL));
   printf("TFM Ident string:\n%s\n\n", fp_ident());
   fp_zero(&b); fp_zero(&c); fp_zero(&d); fp_zero(&e); fp_zero(&f);
-  fp_zero(&a); draw(&a);
+  fp_zero(&a);
+
+#ifndef TESTING
+
+  draw(&a);
 
   /* test set and simple shifts */
   printf("Testing mul/div 2\n");
@@ -217,17 +231,12 @@ int main(void)
    }
    printf("\n\n");
 
-#ifdef TESTING
-goto testing;
-#endif
-
 #if 1
 
 t1 = TIMFUNC();
 sleep(1);
 printf("Ticks per second: %llu\n", TIMFUNC() - t1);
 
-goto multtime;
  /* do some timings... */
   printf("Addition:\n");
   for (t = 2; t <= FP_SIZE/2; t += 2) {
@@ -252,7 +261,6 @@ goto multtime;
       }
       printf("%5lu-bit: %9llu\n", t * DIGIT_BIT, t2);
   }
-multtime:
   printf("Multiplication:\n");
   for (t = 2; t < FP_SIZE/2; t += 2) {
       fp_zero(&a);
@@ -336,8 +344,7 @@ multtime:
       }
       printf("%5lu-bit: %9llu\n", t * DIGIT_BIT, t2);
   }
-//#else
-sqrtime:
+
   printf("Squaring:\n");
   for (t = 2; t < FP_SIZE/2; t += 2) {
       fp_zero(&a);
@@ -418,7 +425,7 @@ sqrtime:
       }
       printf("%5lu-bit: %9llu\n", t * DIGIT_BIT, t2);
   }
-invmodtime:
+
   printf("Invmod:\n");
   for (t = 2; t < FP_SIZE/2; t += 2) {
      fp_zero(&a);
@@ -505,8 +512,7 @@ invmodtime:
       }
       printf("%5lu-bit: %9llu\n", t * DIGIT_BIT, t2);
   }
-//#else
-monttime:
+
   printf("Montgomery:\n");
   for (t = 2; t <= (FP_SIZE/2)-4; t += 2) {
 //      printf("%5lu-bit: %9llu\n", t * DIGIT_BIT, t2);
@@ -596,8 +602,7 @@ monttime:
       }
       printf("%5lu-bit: %9llu\n", t * DIGIT_BIT, t2);
   }
-//#else
-expttime:
+
   printf("Exptmod:\n");
 
   for (t = 512/DIGIT_BIT; t <= (FP_SIZE/2)-2; t += 256/DIGIT_BIT) {
@@ -629,7 +634,7 @@ expttime:
 #endif
 
 return 0;
-testing:
+#endif
 
   fp_zero(&b); fp_zero(&c); fp_zero(&d); fp_zero(&e); fp_zero(&f); fp_zero(&a);
 
@@ -650,7 +655,7 @@ testing:
           fp_mul_2d(&a, rr, &a);
           a.sign = b.sign;
           if (fp_cmp(&a, &b) != FP_EQ) {
-             printf("mul2d failed, rr == %lu\n",rr);
+             printf("\nmul2d failed, rr == %lu\n",rr);
              draw(&a);
              draw(&b);
              return 0;
@@ -664,7 +669,7 @@ testing:
           a.sign = b.sign;
           if (a.used == b.used && a.used == 0) { a.sign = b.sign = FP_ZPOS; }
           if (fp_cmp(&a, &b) != FP_EQ) {
-             printf("div2d failed, rr == %lu\n",rr);
+             printf("\ndiv2d failed, rr == %lu\n",rr);
              draw(&a);
              draw(&b);
              return 0;
@@ -676,7 +681,7 @@ testing:
           fp_copy(&a, &d);
           fp_add(&d, &b, &d);
           if (fp_cmp(&c, &d) != FP_EQ) {
-             printf("add %lu failure!\n", add_n);
+             printf("\nadd %lu failure!\n", add_n);
 draw(&a);draw(&b);draw(&c);draw(&d);
              return 0;
           }
@@ -688,7 +693,7 @@ draw(&a);draw(&b);draw(&c);draw(&d);
           memset(cmd+rr, rand()&255, sizeof(cmd)-rr);
           fp_read_signed_bin(&d, (unsigned char *)cmd, rr);
           if (fp_cmp(&c, &d) != FP_EQ) {
-             printf("fp_signed_bin failure!\n");
+             printf("f\np_signed_bin failure!\n");
              draw(&c);
              draw(&d);
              return 0;
@@ -699,7 +704,7 @@ draw(&a);draw(&b);draw(&c);draw(&d);
           memset(cmd+rr, rand()&255, sizeof(cmd)-rr);
           fp_read_unsigned_bin(&d, (unsigned char *)cmd, rr);
           if (fp_cmp_mag(&c, &d) != FP_EQ) {
-             printf("fp_unsigned_bin failure!\n");
+             printf("\nfp_unsigned_bin failure!\n");
              draw(&c);
              draw(&d);
              return 0;
@@ -712,98 +717,98 @@ draw(&a);draw(&b);draw(&c);draw(&d);
           fp_copy(&a, &d);
           fp_sub(&d, &b, &d);
           if (fp_cmp(&c, &d) != FP_EQ) {
-             printf("sub %lu failure!\n", sub_n);
+             printf("\nsub %lu failure!\n", sub_n);
 draw(&a);draw(&b);draw(&c);draw(&d);
              return 0;
           }
-       } else if (!strcmp(cmd, "mul")) {
+       } else if (!strcmp(cmd, "mul")) { ++mul_n;
           fgets(buf, 4095, stdin);  fp_read_radix(&a, buf, 64);
           fgets(buf, 4095, stdin);  fp_read_radix(&b, buf, 64);
           fgets(buf, 4095, stdin);  fp_read_radix(&c, buf, 64);
 //continue;
           fp_copy(&a, &d);
-          fp_mul(&d, &b, &d); ++mul_n;
+          fp_mul(&d, &b, &d);
           if (fp_cmp(&c, &d) != FP_EQ) {
-             printf("mul %lu failure!\n", mul_n);
+             printf("\nmul %lu failure!\n", mul_n);
 draw(&a);draw(&b);draw(&c);draw(&d);
              return 0;
           }
-       } else if (!strcmp(cmd, "div")) {
+       } else if (!strcmp(cmd, "div")) { ++div_n;
           fgets(buf, 4095, stdin); fp_read_radix(&a, buf, 64);
           fgets(buf, 4095, stdin); fp_read_radix(&b, buf, 64);
           fgets(buf, 4095, stdin); fp_read_radix(&c, buf, 64);
           fgets(buf, 4095, stdin); fp_read_radix(&d, buf, 64);
 // continue;
-          fp_div(&a, &b, &e, &f); ++div_n;
+          fp_div(&a, &b, &e, &f);
           if (fp_cmp(&c, &e) != FP_EQ || fp_cmp(&d, &f) != FP_EQ) {
-             printf("div %lu failure!\n", div_n);
+             printf("\ndiv %lu failure!\n", div_n);
 draw(&a);draw(&b);draw(&c);draw(&d); draw(&e); draw(&f);
              return 0;
           }
 
-       } else if (!strcmp(cmd, "sqr")) {
+       } else if (!strcmp(cmd, "sqr")) { ++sqr_n;
           fgets(buf, 4095, stdin);  fp_read_radix(&a, buf, 64);
           fgets(buf, 4095, stdin);  fp_read_radix(&b, buf, 64);
 // continue;
           fp_copy(&a, &c);
-          fp_sqr(&c, &c); ++sqr_n;
+          fp_sqr(&c, &c);
           if (fp_cmp(&b, &c) != FP_EQ) {
-             printf("sqr %lu failure!\n", sqr_n);
+             printf("\nsqr %lu failure!\n", sqr_n);
 draw(&a);draw(&b);draw(&c);
              return 0;
           }
-       } else if (!strcmp(cmd, "gcd")) {
+       } else if (!strcmp(cmd, "gcd")) { ++gcd_n;
           fgets(buf, 4095, stdin);  fp_read_radix(&a, buf, 64);
           fgets(buf, 4095, stdin);  fp_read_radix(&b, buf, 64);
           fgets(buf, 4095, stdin);  fp_read_radix(&c, buf, 64);
 // continue;
           fp_copy(&a, &d);
-          fp_gcd(&d, &b, &d); ++gcd_n;
+          fp_gcd(&d, &b, &d);
           d.sign = c.sign;
           if (fp_cmp(&c, &d) != FP_EQ) {
-             printf("gcd %lu failure!\n", gcd_n);
+             printf("\ngcd %lu failure!\n", gcd_n);
 draw(&a);draw(&b);draw(&c);draw(&d);
              return 0;
           }
-       } else if (!strcmp(cmd, "lcm")) {
+       } else if (!strcmp(cmd, "lcm")) { ++lcm_n;
              fgets(buf, 4095, stdin);  fp_read_radix(&a, buf, 64);
              fgets(buf, 4095, stdin);  fp_read_radix(&b, buf, 64);
              fgets(buf, 4095, stdin);  fp_read_radix(&c, buf, 64);
 //continue;
              fp_copy(&a, &d);
-             fp_lcm(&d, &b, &d); ++lcm_n;
+             fp_lcm(&d, &b, &d);
              d.sign = c.sign;
              if (fp_cmp(&c, &d) != FP_EQ) {
-                printf("lcm %lu failure!\n", lcm_n);
+                printf("\nlcm %lu failure!\n", lcm_n);
    draw(&a);draw(&b);draw(&c);draw(&d);
                 return 0;
              }
-       } else if (!strcmp(cmd, "expt")) {
+       } else if (!strcmp(cmd, "expt")) { ++expt_n;
              fgets(buf, 4095, stdin);  fp_read_radix(&a, buf, 64);
              fgets(buf, 4095, stdin);  fp_read_radix(&b, buf, 64);
              fgets(buf, 4095, stdin);  fp_read_radix(&c, buf, 64);
              fgets(buf, 4095, stdin);  fp_read_radix(&d, buf, 64);
 // continue;
              fp_copy(&a, &e);
-             fp_exptmod(&e, &b, &c, &e); ++expt_n;
+             fp_exptmod(&e, &b, &c, &e);
              if (fp_cmp(&d, &e) != FP_EQ) {
-                printf("expt %lu failure!\n", expt_n);
+                printf("\nexpt %lu failure!\n", expt_n);
    draw(&a);draw(&b);draw(&c);draw(&d); draw(&e);
                 return 0;
              }
-       } else if (!strcmp(cmd, "invmod")) {
+       } else if (!strcmp(cmd, "invmod")) { ++inv_n;
              fgets(buf, 4095, stdin);  fp_read_radix(&a, buf, 64);
              fgets(buf, 4095, stdin);  fp_read_radix(&b, buf, 64);
              fgets(buf, 4095, stdin);  fp_read_radix(&c, buf, 64);
 //continue;
              fp_invmod(&a, &b, &d);
 #if 1
-             fp_mulmod(&d,&a,&b,&e); ++inv_n;
+             fp_mulmod(&d,&a,&b,&e);
              if (fp_cmp_d(&e, 1) != FP_EQ) {
 #else
              if (fp_cmp(&d, &c) != FP_EQ) {
 #endif
-                printf("inv [wrong value from MPI?!] failure\n");
+                printf("\ninv [wrong value from MPI?!] failure\n");
                 draw(&a);draw(&b);draw(&c);draw(&d);
                 return 0;
              }
@@ -813,7 +818,7 @@ draw(&a);draw(&b);draw(&c);draw(&d);
              fgets(buf, 4095, stdin);  fp_read_radix(&b, buf, 64);
              fp_div_2(&a, &c);
              if (fp_cmp(&c, &b) != FP_EQ) {
-                 printf("div_2 %lu failure\n", div2_n);
+                 printf("\ndiv_2 %lu failure\n", div2_n);
                  draw(&a);
                  draw(&b);
                  draw(&c);
@@ -824,7 +829,7 @@ draw(&a);draw(&b);draw(&c);draw(&d);
              fgets(buf, 4095, stdin);  fp_read_radix(&b, buf, 64);
              fp_mul_2(&a, &c);
              if (fp_cmp(&c, &b) != FP_EQ) {
-                 printf("mul_2 %lu failure\n", mul2_n);
+                 printf("\nmul_2 %lu failure\n", mul2_n);
                  draw(&a);
                  draw(&b);
                  draw(&c);
@@ -836,7 +841,7 @@ draw(&a);draw(&b);draw(&c);draw(&d);
               fgets(buf, 4095, stdin); fp_read_radix(&b, buf, 64);
               fp_add_d(&a, ix, &c);
               if (fp_cmp(&b, &c) != FP_EQ) {
-                 printf("add_d %lu failure\n", add_d_n);
+                 printf("\nadd_d %lu failure\n", add_d_n);
                  draw(&a);
                  draw(&b);
                  draw(&c);
@@ -849,7 +854,7 @@ draw(&a);draw(&b);draw(&c);draw(&d);
               fgets(buf, 4095, stdin); fp_read_radix(&b, buf, 64);
               fp_sub_d(&a, ix, &c);
               if (fp_cmp(&b, &c) != FP_EQ) {
-                 printf("sub_d %lu failure\n", sub_d_n);
+                 printf("\nsub_d %lu failure\n", sub_d_n);
                  draw(&a);
                  draw(&b);
                  draw(&c);
@@ -862,7 +867,7 @@ draw(&a);draw(&b);draw(&c);draw(&d);
               fgets(buf, 4095, stdin); fp_read_radix(&b, buf, 64);
               fp_mul_d(&a, ix, &c);
               if (fp_cmp(&b, &c) != FP_EQ) {
-                 printf("mul_d %lu failure\n", sub_d_n);
+                 printf("\nmul_d %lu failure\n", mul_d_n);
                  draw(&a);
                  draw(&b);
                  draw(&c);
diff --git a/makefile b/makefile
index b513d0f..9b6de65 100644
--- a/makefile
+++ b/makefile
@@ -101,6 +101,8 @@ install: $(LIBNAME)
 mtest: $(LIBNAME)
 	cd mtest; CC="$(CC)" CFLAGS="$(CFLAGS) -I../" MAKE=${MAKE} ${MAKE} mtest
 
+demo/test.o: CFLAGS+=-Wno-unused-result
+
 .PHONY: test
 test: $(LIBNAME) demo/test.o
 	$(CC) $(CFLAGS) demo/test.o $(LIBNAME) $(PROF) -o test
diff --git a/mtest/makefile b/mtest/makefile
index de6bf58..be512d0 100644
--- a/mtest/makefile
+++ b/mtest/makefile
@@ -1,9 +1,10 @@
-CFLAGS += -Wall -W -O3 
+CFLAGS += -Wall -W -O3 -Wno-unused-result
 
 default: mtest
 
+.PHONY: mtest
 mtest: mtest.o
 	$(CC) $(CFLAGS) mtest.o -ltommath -o mtest
 
 clean:
-	rm -f *.o mtest *~
+	rm -f *.o mtest *~ mtest.exe
diff --git a/mtest/mtest.c b/mtest/mtest.c
index b19a0fc..f8a62a8 100644
--- a/mtest/mtest.c
+++ b/mtest/mtest.c
@@ -39,6 +39,7 @@ mulmod
 #include <time.h>
 #include <tommath.h>
 #define CRYPT
+#undef DIGIT_BIT
 #include "../src/headers/tfm.h"
 
 FILE *rng;
@@ -46,8 +47,8 @@ FILE *rng;
 /* 1-2048 bit numbers */
 void rand_num(mp_int *a)
 {
-   int n, size;
-   unsigned char buf[2048];
+   int size;
+   unsigned char buf[(FP_MAX_SIZE/16 - DIGIT_BIT/2) + 1];
 
    size = 1 + ((fgetc(rng)<<8) + fgetc(rng)) % (FP_MAX_SIZE/16 - DIGIT_BIT/2);
    buf[0] = (fgetc(rng)&1)?1:0;
@@ -59,8 +60,8 @@ void rand_num(mp_int *a)
 /* 1-256 bit numbers (to test things like exptmod) */
 void rand_num2(mp_int *a)
 {
-   int n, size;
-   unsigned char buf[2048];
+   int size;
+   unsigned char buf[(FP_MAX_SIZE/16 - DIGIT_BIT/2) + 1];
 
    size = 1 + ((fgetc(rng)<<8) + fgetc(rng)) % (FP_MAX_SIZE/16 - DIGIT_BIT/2);
    buf[0] = (fgetc(rng)&1)?1:0;
@@ -69,13 +70,15 @@ void rand_num2(mp_int *a)
    mp_read_raw(a, buf, 1+size);
 }
 
-#define mp_to64(a, b) mp_toradix(a, b, 64)
+#define mp_to64(a, b) mp_toradix_n(a, b, 64, sizeof(b))
 
 int main(void)
 {
    int n, tmp;
    mp_int a, b, c, d, e;
+#ifdef MTEST_NO_FULLSPEED
    clock_t t1;
+#endif
    char buf[4096];
 
    mp_init(&a);
@@ -88,7 +91,7 @@ int main(void)
    /* initial (2^n - 1)^2 testing, makes sure the comba multiplier works [it has the new carry code] */
 /*
    mp_set(&a, 1);
-   for (n = 1; n < 8192; n++) {
+   for (n = 1; n < ((FP_MAX_SIZE-(8*DIGIT_BIT))/2); n++) {
        mp_mul(&a, &a, &c);
        printf("mul\n");
        mp_to64(&a, buf);
@@ -111,9 +114,11 @@ int main(void)
       }
    }
 
+#ifdef MTEST_NO_FULLSPEED
    t1 = clock();
+#endif
    for (;;) {
-#if 0
+#ifdef MTEST_NO_FULLSPEED
       if (clock() - t1 > CLOCKS_PER_SEC) {
          sleep(2);
          t1 = clock();