set thread size

ed25519/ed25519.h

@@ -3,7 +3,22 @@
 #define ED25519_PUBLICKEYBYTES 32
 
 
+#ifndef ED25519_donna
+#  if defined(_MSC_VER)
+#    define ALIGN(x) __declspec(align(x))
+#  elif defined(__GNUC__)
+#    undef ALIGN
+#    define ALIGN(x) __attribute__((aligned(x)))
+#  else
+#    ifndef ALIGN
+#      define ALIGN(x)
+#    endif
+#  endif
+#endif
+
+
 #ifdef ED25519_ref10
+
 #include "ref10/ed25519.h"
 #define ed25519_seckey        ed25519_ref10_seckey
 #define ed25519_seckey_expand ed25519_ref10_seckey_expand
@@ -35,49 +50,47 @@ static const ge_cached ge_eightpoint = {
   }
 };
 inline static void ge_initeightpoint(void) {}
+
 #endif
 
 
 #ifdef ED25519_amd64_51_30k
-#define ED25519_amd64_common
-#else
-#ifdef ED25519_amd64_64_24k
-#define ED25519_amd64_common
-#endif
-#endif
 
-
-#ifdef ED25519_amd64_51_30k
 #include "amd64-51-30k/ed25519.h"
 #include "amd64-51-30k/ge25519.h"
 #define ed25519_seckey        ed25519_amd64_51_30k_seckey
 #define ed25519_seckey_expand ed25519_amd64_51_30k_seckey_expand
 #define ed25519_pubkey        ed25519_amd64_51_30k_pubkey
 #define ed25519_keygen        ed25519_amd64_51_30k_keygen
+
 #endif
 
+
 #ifdef ED25519_amd64_64_24k
+
 #include "amd64-64-24k/ed25519.h"
 #include "amd64-64-24k/ge25519.h"
 #define ed25519_seckey        ed25519_amd64_64_seckey
 #define ed25519_seckey_expand ed25519_amd64_64_seckey_expand
 #define ed25519_pubkey        ed25519_amd64_64_pubkey
 #define ed25519_keygen        ed25519_amd64_64_keygen
+
 #endif
 
 
 // common
-#ifdef ED25519_amd64_common
+#if defined(ED25519_amd64_51_30k) || defined(ED25519_amd64_64_24k)
+
 #define fe            fe25519
 #define ge_p1p1       ge25519_p1p1
 #define ge_p3         ge25519_p3
-#define ge_cached     ge25519_pniels
 #define ge_p1p1_to_p3 ge25519_p1p1_to_p3
 #define ge_p3_tobytes ge25519_pack
 #define ge_add        ge25519_pnielsadd_p1p1
 
 #define ge_p3_batchtobytes_destructive_1      ge25519_batchpack_destructive_1
 #define ge_p3_batchtobytes_destructive_finish ge25519_batchpack_destructive_finish
+
 #endif
 
 
@@ -130,6 +143,7 @@ inline static void ge_initeightpoint(void) {}
 
 
 #ifdef ED25519_donna
+
 #define ED25519_CUSTOMRANDOM
 #define ED25519_CUSTOMHASH
 #include <sodium/crypto_hash_sha512.h>
@@ -174,11 +188,10 @@ static int ed25519_keypair(unsigned char *pk,unsigned char *sk)
 
 	return 0;
 }
-// hacky, but works for current stuff in main.c
+
-#define fe            bignum25519 ALIGN(16)
+#define fe      bignum25519
-#define ge_p1p1       ge25519_p1p1 ALIGN(16)
+#define ge_p1p1 ge25519_p1p1
-#define ge_p3         ge25519 ALIGN(16)
+#define ge_p3   ge25519
-#define ge_cached     ge25519_pniels ALIGN(16)
 
 #define ge_p1p1_to_p3 ge25519_p1p1_to_full
 #define ge_p3_tobytes ge25519_pack

main.c

@@ -186,8 +186,6 @@ enum worker_type {
 	WT_BATCH,
 };
 
-#define TATTR 0
-
 int main(int argc,char **argv)
 {
 	const char *outfile = 0;
@@ -520,7 +518,6 @@ int main(int argc,char **argv)
 	VEC_ZERO(tstats);
 #endif
 
-#if TATTR
 	pthread_attr_t tattr,*tattrp = &tattr;
 	tret = pthread_attr_init(tattrp);
 	if (tret) {
@@ -528,11 +525,17 @@ int main(int argc,char **argv)
 		tattrp = 0;
 	}
 	else {
-		tret = pthread_attr_setstacksize(tattrp,2<<20);
+		// 256KiB plus whatever batch stuff uses if in batch mode
+		size_t ss = 256 << 10;
+		if (wt == WT_BATCH)
+			ss += worker_batch_memuse();
+		// align to 64KiB
+		ss = (ss + (64 << 10) - 1) & ~((64 << 10) - 1);
+		//printf("stack size: " FSZ "\n",ss);
+		tret = pthread_attr_setstacksize(tattrp,ss);
 		if (tret)
 			perror("pthread_attr_setstacksize");
 	}
-#endif
 
 	for (size_t i = 0;i < VEC_LENGTH(threads);++i) {
 		void *tp = 0;
@@ -562,13 +565,11 @@ int main(int argc,char **argv)
 		}
 	}
 
-#if TATTR
 	if (tattrp) {
 		tret = pthread_attr_destroy(tattrp);
 		if (tret)
 			perror("pthread_attr_destroy");
 	}
-#endif
 
 #ifdef STATISTICS
 	struct timespec nowtime;

worker.c

@@ -94,9 +94,9 @@ static void onionready(char *sname,const u8 *secret,const u8 *pubonion)
 	}
 
 	// disabled as this was never ever triggered as far as I'm aware
-#if 0
+#if 1
 	// Sanity check that the public key matches the private one.
-	ge_p3 point;
+	ge_p3 ALIGN(16) point;
 	u8 testpk[PUBLIC_LEN];
 	ge_scalarmult_base(&point, secret);
 	ge_p3_tobytes(testpk, &point);
@@ -220,6 +220,11 @@ static void reseedright(u8 sk[SECRET_LEN])
 	#define BATCHNUM 2048
 #endif
 
+size_t worker_batch_memuse(void)
+{
+	return (sizeof(ge_p3) + sizeof(fe) + sizeof(bytes32)) * BATCHNUM;
+}
+
 #include "worker_batch.inc.h"
 
 #include "worker_batch_pass.inc.h"

worker.h

@@ -38,6 +38,7 @@ extern u8 determseed[SEED_LEN];
 extern void worker_init(void);
 
 extern char *makesname(void);
+extern size_t worker_batch_memuse(void);
 
 extern void *worker_slow(void *task);
 extern void *worker_fast(void *task);

worker_batch.inc.h

@@ -8,13 +8,13 @@ void *worker_batch(void *task)
 	u8 seed[SEED_LEN];
 	u8 hashsrc[checksumstrlen + PUBLIC_LEN + 1];
 	u8 wpk[PUBLIC_LEN + 1];
-	ge_p3 ge_public;
+	ge_p3 ALIGN(16) ge_public;
 	char *sname;
 
 	// state to keep batch data
-	ge_p3 ge_batch[BATCHNUM];
+	ge_p3   ALIGN(16) ge_batch [BATCHNUM];
-	fe tmp_batch[BATCHNUM];
+	fe      ALIGN(16) tmp_batch[BATCHNUM];
-	bytes32 pk_batch[BATCHNUM];
+	bytes32 ALIGN(16) pk_batch [BATCHNUM];
 
 	size_t counter;
 	size_t i;
@@ -47,7 +47,7 @@ initseed:
 	ge_scalarmult_base(&ge_public,sk);
 
 	for (counter = 0;counter < SIZE_MAX-(8*BATCHNUM);counter += 8*BATCHNUM) {
-		ge_p1p1 sum;
+		ge_p1p1 ALIGN(16) sum;
 
 		if (unlikely(endwork))
 			goto end;

worker_batch_pass.inc.h

@@ -9,13 +9,13 @@ void *worker_batch_pass(void *task)
 	u8 seed[SEED_LEN];
 	u8 hashsrc[checksumstrlen + PUBLIC_LEN + 1];
 	u8 wpk[PUBLIC_LEN + 1];
-	ge_p3 ge_public;
+	ge_p3 ALIGN(16) ge_public;
 	char *sname;
 
 	// state to keep batch data
-	ge_p3 ge_batch[BATCHNUM];
+	ge_p3   ALIGN(16) ge_batch [BATCHNUM];
-	fe tmp_batch[BATCHNUM];
+	fe      ALIGN(16) tmp_batch[BATCHNUM];
-	bytes32 pk_batch[BATCHNUM];
+	bytes32 ALIGN(16) pk_batch [BATCHNUM];
 
 	size_t counter,oldcounter;
 	size_t i;
@@ -53,7 +53,7 @@ initseed:
 	ge_scalarmult_base(&ge_public,sk);
 
 	for (counter = oldcounter = 0;counter < DETERMINISTIC_LOOP_COUNT - (BATCHNUM - 1) * 8;counter += BATCHNUM * 8) {
-		ge_p1p1 sum;
+		ge_p1p1 ALIGN(16) sum;
 
 		if (unlikely(endwork))
 			goto end;
@@ -120,7 +120,7 @@ initseed:
 	// can't have leftovers in theory if BATCHNUM was power of 2 and smaller than DETERMINISTIC_LOOP_COUNT bound
 #if (BATCHNUM & (BATCHNUM - 1)) || (BATCHNUM * 8) > DETERMINISTIC_LOOP_COUNT
 	if (counter < DETERMINISTIC_LOOP_COUNT) {
-		ge_p1p1 sum;
+		ge_p1p1 ALIGN(16) sum;
 
 		if (unlikely(endwork))
 			goto end;

worker_fast.inc.h

@@ -8,7 +8,7 @@ void *worker_fast(void *task)
 	u8 seed[SEED_LEN];
 	u8 hashsrc[checksumstrlen + PUBLIC_LEN + 1];
 	u8 wpk[PUBLIC_LEN + 1];
-	ge_p3 ge_public;
+	ge_p3 ALIGN(16) ge_public;
 	char *sname;
 
 	size_t counter;
@@ -45,7 +45,7 @@ initseed:
 	ge_p3_tobytes(pk,&ge_public);
 
 	for (counter = 0;counter < SIZE_MAX-8;counter += 8) {
-		ge_p1p1 sum;
+		ge_p1p1 ALIGN(16) sum;
 
 		if (unlikely(endwork))
 			goto end;

worker_fast_pass.inc.h

@@ -9,7 +9,7 @@ void *worker_fast_pass(void *task)
 	u8 seed[SEED_LEN];
 	u8 hashsrc[checksumstrlen + PUBLIC_LEN + 1];
 	u8 wpk[PUBLIC_LEN + 1];
-	ge_p3 ge_public;
+	ge_p3 ALIGN(16) ge_public;
 	char *sname;
 
 	size_t counter,oldcounter;
@@ -51,7 +51,7 @@ initseed:
 	ge_p3_tobytes(pk,&ge_public);
 
 	for (counter = oldcounter = 0;counter < DETERMINISTIC_LOOP_COUNT;counter += 8) {
-		ge_p1p1 sum;
+		ge_p1p1 ALIGN(16) sum;
 
 		if (unlikely(endwork))
 			goto end;