Add memory fence to bench_internal

Author: elichai

src/bench.h

@@ -12,6 +12,10 @@
 #include <math.h>
 #include "sys/time.h"
 
+static void escape(void *p) {
+    __asm__ __volatile__("": : "g"(p) : "memory");
+}
+
 static double gettimedouble(void) {
     struct timeval tv;
     gettimeofday(&tv, NULL);

src/bench_internal.c

@@ -27,6 +27,7 @@ typedef struct {
     int wnaf[256];
 } bench_inv;
 
+
 void bench_setup(void* arg) {
     bench_inv *data = (bench_inv*)arg;
 
@@ -62,6 +63,7 @@ void bench_scalar_add(void* arg) {
 
     for (i = 0; i < 2000000; i++) {
         secp256k1_scalar_add(&data->scalar_x, &data->scalar_x, &data->scalar_y);
+        escape(data);
     }
 }
 
@@ -71,6 +73,7 @@ void bench_scalar_negate(void* arg) {
 
     for (i = 0; i < 2000000; i++) {
         secp256k1_scalar_negate(&data->scalar_x, &data->scalar_x);
+        escape(data);
     }
 }
 
@@ -80,6 +83,7 @@ void bench_scalar_sqr(void* arg) {
 
     for (i = 0; i < 200000; i++) {
         secp256k1_scalar_sqr(&data->scalar_x, &data->scalar_x);
+        escape(data);
     }
 }
 
@@ -89,6 +93,7 @@ void bench_scalar_mul(void* arg) {
 
     for (i = 0; i < 200000; i++) {
         secp256k1_scalar_mul(&data->scalar_x, &data->scalar_x, &data->scalar_y);
+        escape(data);
     }
 }
 
@@ -101,6 +106,7 @@ void bench_scalar_split(void* arg) {
         secp256k1_scalar l, r;
         secp256k1_scalar_split_lambda(&l, &r, &data->scalar_x);
         secp256k1_scalar_add(&data->scalar_x, &data->scalar_x, &data->scalar_y);
+        escape(data);
     }
 }
 #endif
@@ -112,6 +118,7 @@ void bench_scalar_inverse(void* arg) {
     for (i = 0; i < 2000; i++) {
         secp256k1_scalar_inverse(&data->scalar_x, &data->scalar_x);
         secp256k1_scalar_add(&data->scalar_x, &data->scalar_x, &data->scalar_y);
+        escape(data);
     }
 }
 
@@ -122,6 +129,7 @@ void bench_scalar_inverse_var(void* arg) {
     for (i = 0; i < 2000; i++) {
         secp256k1_scalar_inverse_var(&data->scalar_x, &data->scalar_x);
         secp256k1_scalar_add(&data->scalar_x, &data->scalar_x, &data->scalar_y);
+        escape(data);
     }
 }
 
@@ -131,6 +139,7 @@ void bench_field_normalize(void* arg) {
 
     for (i = 0; i < 2000000; i++) {
         secp256k1_fe_normalize(&data->fe_x);
+        escape(data);
     }
 }
 
@@ -140,6 +149,7 @@ void bench_field_normalize_weak(void* arg) {
 
     for (i = 0; i < 2000000; i++) {
         secp256k1_fe_normalize_weak(&data->fe_x);
+        escape(data);
     }
 }
 
@@ -149,6 +159,7 @@ void bench_field_mul(void* arg) {
 
     for (i = 0; i < 200000; i++) {
         secp256k1_fe_mul(&data->fe_x, &data->fe_x, &data->fe_y);
+        escape(data);
     }
 }
 
@@ -158,6 +169,7 @@ void bench_field_sqr(void* arg) {
 
     for (i = 0; i < 200000; i++) {
         secp256k1_fe_sqr(&data->fe_x, &data->fe_x);
+        escape(data);
     }
 }
 
@@ -168,6 +180,7 @@ void bench_field_inverse(void* arg) {
     for (i = 0; i < 20000; i++) {
         secp256k1_fe_inv(&data->fe_x, &data->fe_x);
         secp256k1_fe_add(&data->fe_x, &data->fe_y);
+        escape(data);
     }
 }
 
@@ -178,6 +191,7 @@ void bench_field_inverse_var(void* arg) {
     for (i = 0; i < 20000; i++) {
         secp256k1_fe_inv_var(&data->fe_x, &data->fe_x);
         secp256k1_fe_add(&data->fe_x, &data->fe_y);
+        escape(data);
     }
 }
 
@@ -190,6 +204,7 @@ void bench_field_sqrt(void* arg) {
         t = data->fe_x;
         secp256k1_fe_sqrt(&data->fe_x, &t);
         secp256k1_fe_add(&data->fe_x, &data->fe_y);
+        escape(data);
     }
 }
 
@@ -199,6 +214,7 @@ void bench_group_double_var(void* arg) {
 
     for (i = 0; i < 200000; i++) {
         secp256k1_gej_double_var(&data->gej_x, &data->gej_x, NULL);
+        escape(data);
     }
 }
 
@@ -208,6 +224,7 @@ void bench_group_add_var(void* arg) {
 
     for (i = 0; i < 200000; i++) {
         secp256k1_gej_add_var(&data->gej_x, &data->gej_x, &data->gej_y, NULL);
+        escape(data);
     }
 }
 
@@ -217,6 +234,7 @@ void bench_group_add_affine(void* arg) {
 
     for (i = 0; i < 200000; i++) {
         secp256k1_gej_add_ge(&data->gej_x, &data->gej_x, &data->ge_y);
+        escape(data);
     }
 }
 
@@ -226,6 +244,7 @@ void bench_group_add_affine_var(void* arg) {
 
     for (i = 0; i < 200000; i++) {
         secp256k1_gej_add_ge_var(&data->gej_x, &data->gej_x, &data->ge_y, NULL);
+        escape(data);
     }
 }
 
@@ -235,6 +254,7 @@ void bench_group_jacobi_var(void* arg) {
 
     for (i = 0; i < 20000; i++) {
         secp256k1_gej_has_quad_y_var(&data->gej_x);
+        escape(data);
     }
 }
 
@@ -245,6 +265,7 @@ void bench_ecmult_wnaf(void* arg) {
     for (i = 0; i < 20000; i++) {
         secp256k1_ecmult_wnaf(data->wnaf, 256, &data->scalar_x, WINDOW_A);
         secp256k1_scalar_add(&data->scalar_x, &data->scalar_x, &data->scalar_y);
+        escape(data);
     }
 }
 
@@ -255,6 +276,7 @@ void bench_wnaf_const(void* arg) {
     for (i = 0; i < 20000; i++) {
         secp256k1_wnaf_const(data->wnaf, &data->scalar_x, WINDOW_A, 256);
         secp256k1_scalar_add(&data->scalar_x, &data->scalar_x, &data->scalar_y);
+        escape(data);
     }
 }
 
@@ -268,6 +290,7 @@ void bench_sha256(void* arg) {
         secp256k1_sha256_initialize(&sha);
         secp256k1_sha256_write(&sha, data->data, 32);
         secp256k1_sha256_finalize(&sha, data->data);
+        escape(data);
     }
 }
 
@@ -280,6 +303,7 @@ void bench_hmac_sha256(void* arg) {
         secp256k1_hmac_sha256_initialize(&hmac, data->data, 32);
         secp256k1_hmac_sha256_write(&hmac, data->data, 32);
         secp256k1_hmac_sha256_finalize(&hmac, data->data);
+        escape(data);
     }
 }
 
@@ -291,6 +315,7 @@ void bench_rfc6979_hmac_sha256(void* arg) {
     for (i = 0; i < 20000; i++) {
         secp256k1_rfc6979_hmac_sha256_initialize(&rng, data->data, 64);
         secp256k1_rfc6979_hmac_sha256_generate(&rng, data->data, 32);
+        escape(data);
     }
 }
 
@@ -312,7 +337,7 @@ void bench_context_sign(void* arg) {
 
 #ifndef USE_NUM_NONE
 void bench_num_jacobi(void* arg) {
-    int i;
+    int i, j;
     bench_inv *data = (bench_inv*)arg;
     secp256k1_num nx, norder;
 
@@ -321,7 +346,8 @@ void bench_num_jacobi(void* arg) {
     secp256k1_scalar_get_num(&norder, &data->scalar_y);
 
     for (i = 0; i < 200000; i++) {
-        secp256k1_num_jacobi(&nx, &norder);
+        j = secp256k1_num_jacobi(&nx, &norder);
+        escape(&j);
     }
 }
 #endif