New release.

Rework the directory structure into something saner, with src/ test/ include/ and build/

Beginning some tests.  Also, now support scan-build.

Now support 32-bit including vectorless ARM.  NEON is not yet supported, because I don't
have a test machine.

Many internal changes, improvements, and bug fixes.

Doxyfile

@@ -508,7 +508,7 @@ HIDE_SCOPE_NAMES       = NO
 # the files that are included by a file in the documentation of that file.
 # The default value is: YES.
 
-SHOW_INCLUDE_FILES     = YES
+SHOW_INCLUDE_FILES     = NO
 
 # If the SHOW_GROUPED_MEMB_INC tag is set to YES then Doxygen will add for each
 # grouped member an include statement to the documentation, telling the reader
@@ -777,7 +777,7 @@ FILE_PATTERNS          =
 # be searched for input files as well.
 # The default value is: NO.
 
-RECURSIVE              = NO
+RECURSIVE              = YES
 
 # The EXCLUDE tag can be used to specify files and/or directories that should be
 # excluded from the INPUT source files. This way you can easily exclude a

HISTORY.txt

+March 29, 2014:
+    Added a test directory with various tests.  Currently testing SHA512 Monte
+    Carlo, compatibility of the different scalarmul functions, and some
+    identities on EC point ops.  Began moving these tests out of benchmarker.
+    
+    Added scan-build support.
+    
+    Improved some internal interfaces.  Made a structure for Barrett primes
+    instead of passing parameters individually.  Moved some field operations
+    to places that make more sense, eg Barrett serialize and deserialize.  The
+    deserialize operation now checks that its argument is in [0,q).
+    
+    Added more documentation.
+    
+    Changed the names of a bunch of functions.  Still not entirely consistent,
+    but getting more so.
+    
+    Some minor speed improvements.  For example, multiply is now a couple cycles
+    faster.
+    
+    Added a hackish attempt at thread-safety and initialization sanity checking
+    in the Goldilocks top-level routines.
+    
+    Fixed some vector alignment bugs.  Compiling with -O0 should now work.
+    
+    Slightly simplified recode_wnaf.
+
+    Add a config.h file for future configuration.  EXPERIMENT flags moved here.
+    
+    I've decided against major changes to SHA512 for the moment.  They add speed
+    but also significantly bloat the code, which is going to hurt L1 cache
+    performance.  Perhaps we should link to OpenSSL if a faster SHA512 is desired.
+    
+    Reorganize the source tree into src, test; factor arch stuff into src/arch_*.
+    
+    Make most of the code 32-bit clean.  There's now a 32-bit generic and 32-bit
+    vectorless ARM version.  No NEON version yet because I don't have a test
+    machine (could use my phone in a pinch I guess?).  The 32-bit version still
+    isn't heavily optimized, but on ARM it's using a nicely reworked signed/phi-adic
+    multiplier.  The squaring is also based on this, but could really stand some
+    improvement.
+    
+    When passed an even exponent (or extra doubles), the Montgomery ladder should
+    now be accept points if and only if they lie on the curve.  This needs
+    additional testing, but it passes the zero bit exponent test.
+    
+    On 32-bit, use 8x4x14 instead of 5x5x18 table organization.  Probably there's
+    a better heuristic.
 
 March 5, 2014:
     First revision.

Makefile

@@ -2,61 +2,101 @@
 # Released under the MIT License.  See LICENSE.txt for license information.
 
 CC = clang
-CFLAGS = -O3 -std=c99 -pedantic -Wall -Wextra -Werror  \
-  -mssse3 -maes -mavx2 -DMUST_HAVE_AVX2 -mbmi2 \
-  -ffunction-sections -fdata-sections -fomit-frame-pointer -fPIC \
-  -DEXPERIMENT_ECDH_OBLITERATE_CT=1 -DEXPERIMENT_ECDH_STIR_IN_PUBKEYS=1
+LD = clang
 
-.PHONY: clean all runbench todo doc
+ARCH = arch_x86_64
+
+WARNFLAGS = -pedantic -Wall -Wextra -Werror -Wunreachable-code \
+	-Wgcc-compat -Wmissing-declarations
+INCFLAGS = -Isrc/include -Iinclude -Isrc/$(ARCH)
+LANGFLAGS = -std=c99
+GENFLAGS = -ffunction-sections -fdata-sections -fomit-frame-pointer -fPIC
+OFLAGS = -O3
+#XFLAGS = -DN_TESTS_BASE=1000
+ARCHFLAGS = -mssse3 -maes -mavx2 -DMUST_HAVE_AVX2 -mbmi2
+#ARCHFLAGS = -m32 -mcpu=cortex-a9 -mfpu=vfpv3-d16
+
+CFLAGS = $(LANGFLAGS) $(WARNFLAGS) $(INCFLAGS) $(OFLAGS) $(ARCHFLAGS) $(GENFLAGS) $(XFLAGS)
+LDFLAGS = $(ARCHFLAGS)
+ASFLAGS = $(ARCHFLAGS)
+
+.PHONY: clean all test bench todo doc lib
 .PRECIOUS: build/%.s
-	
+
 HEADERS= Makefile $(shell find . -name "*.h") build/timestamp
 
 LIBCOMPONENTS= build/goldilocks.o build/barrett_field.o build/crandom.o \
   build/p448.o build/ec_point.o build/scalarmul.o build/sha512.o
 
-all: bench
+TESTCOMPONENTS=build/test.o build/test_scalarmul.o build/test_sha512.o \
+	build/test_pointops.o
+
+BENCHCOMPONENTS=build/bench.o
+
+all: lib build/test build/bench
+
+scan: clean
+	scan-build --use-analyzer=`which clang` \
+		 -enable-checker deadcode -enable-checker llvm \
+		 -enable-checker osx -enable-checker security -enable-checker unix \
+		make build/bench build/test build/goldilocks.so
+
+build/bench: $(LIBCOMPONENTS) $(BENCHCOMPONENTS)
+	$(LD) $(LDFLAGS) -o $@ $^
+
+build/test: $(LIBCOMPONENTS) $(TESTCOMPONENTS)
+	$(LD) $(LDFLAGS) -o $@ $^
+
+lib: build/goldilocks.so
+
+build/goldilocks.so: $(LIBCOMPONENTS)
+	rm -f $@
+	libtool -macosx_version_min 10.6 -dynamic -dead_strip -lc -x -o $@ \
+		  -exported_symbols_list src/exported.sym \
+		  $(LIBCOMPONENTS)
 
-bench: *.h *.c
-	$(CC) $(CFLAGS) -o $@ *.c
-	
 build/timestamp:
 	mkdir -p build
 	touch $@
 
 build/%.o: build/%.s
-	$(CC) -c -o $@ $<
+	$(CC) $(ASFLAGS) -c -o $@ $<
 
-build/%.s: %.c $(HEADERS)
+build/%.s: src/%.c $(HEADERS)
 	$(CC) $(CFLAGS) -S -c -o $@ $<
 
-build/goldilocks.so: $(LIBCOMPONENTS)
-	rm -f $@
-	libtool -macosx_version_min 10.6 -dynamic -dead_strip -lc -x -o $@ \
-		  -exported_symbols_list exported.sym \
-		  $(LIBCOMPONENTS)
+build/%.s: test/%.c $(HEADERS)
+	$(CC) $(CFLAGS) -S -c -o $@ $<
+
+build/%.s: src/$(ARCH)/%.c $(HEADERS)
+	$(CC) $(CFLAGS) -S -c -o $@ $<
 
 doc/timestamp:
 	mkdir -p doc
 	touch $@
 
-doc: Doxyfile doc/timestamp *.c *.h
+doc: Doxyfile doc/timestamp src/*.c src/include/*.h src/$(ARCH)/*.c src/$(ARCH)/*.h
 	doxygen
 
 todo::
-	@egrep --color=auto -w -i 'hack|todo|fixme|bug|xxx|perf|future|remove' *.h *.c
+	@(find * -name '*.h'; find * -name '*.c') | xargs egrep --color=auto -w \
+		'HACK|TODO|FIXME|BUG|XXX|PERF|FUTURE|REMOVE'
 	@echo '============================='
 	@(for i in FIXME BUG XXX TODO HACK PERF FUTURE REMOVE; do \
-	  egrep -w -i $$i *.h *.c > /dev/null || continue; \
+	  (find * -name '*.h'; find * -name '*.c') | xargs egrep -w $$i > /dev/null || continue; \
 	  /bin/echo -n $$i'       ' | head -c 10; \
-	  egrep -w -i $$i *.h *.c | wc -l; \
+	  (find * -name '*.h'; find * -name '*.c') | xargs egrep -w $$i| wc -l; \
 	done)
 	@echo '============================='
 	@echo -n 'Total     '
-	@egrep -w -i 'hack|todo|fixme|bug|xxx|perf|future|remove' *.h *.c | wc -l
+	@(find * -name '*.h'; find * -name '*.c') | xargs egrep -w \
+		'HACK|TODO|FIXME|BUG|XXX|PERF|FUTURE|REMOVE' | wc -l
+
+bench: build/bench
+	./$<
 
-runbench: bench
+test: build/test
 	./$<
 
 clean:
-	rm -fr build bench *.o *.s
+	rm -fr build doc

TODO.txt

@@ -23,7 +23,7 @@ Important work items for Ed448-Goldilocks:
     * Word_t, mask_t, bigregister_t, etc.
     * Generate asm intrinsics with a script?
 
-* Bugfix: make sure that init() and randomization are thread-safe.
+* [DONE] Bugfix: make sure that init() and randomization are thread-safe.
 
 * Security: check on deserialization that points are < p.
     * Check also that they're nonzero or otherwise non-pathological?
@@ -80,30 +80,29 @@ Important work items for Ed448-Goldilocks:
 * Portability: make the inner layers of the code 32-bit clean.
     * Write new versions of the field code.
         * 28-bit limbs give less headroom for carries.
-        * NEON and vectorless ARM.
+        * Now have a vectorless ARM version; need NEON.
+        * Improve speed of 32-bit field code.
     
     * Run through the SAGE tool to generate new bias & bound.
 
-* Portability: make the outer layers of the code 32-bit clean.
-    * There are endian bugs in the signing algorithm.
-    * NEON and vectorless constant-time comparison.
+* [DONE] Portability: make the outer layers of the code 32-bit clean.
 
-* Performance: write and incorporate some extra routines
-    * Deserialize_and_isogeny
-    * Unconditional negate (or just plain subtract)
-
-* Performance: fixed parameters?
+* Performance/flexibility: decide which parameters should be hard-coded.
     * Perhaps useful for comb precomputation.
 
 * Performance: Improve SHA512.
-    * Improve portability.
+    * [DONE?] Improve portability.
     * Improve speed.
+        * Except not, because this adds too much code size.
+        * Link OpenSSL if a fast SHA is desired.
+
+* Protocol:
     * Decide what things to stir into hashes for various functions.
     
 * Performance: improve the Barrett field code.
     * Support other primes?
     * Capture prime shape into a struct instead of passing 3 params.
-    * Make 32-bit clean.  (SAGE?)
+    * [DONE] Make 32-bit clean.
 
 * Automation:
     * Improve the SAGE tool to cover more cases
@@ -111,6 +110,10 @@ Important work items for Ed448-Goldilocks:
         * Constant-time selection
         * Intrinsics code
         * Field code?
+    
+    * SAGE tool is impossibly slow on 32-bit
+         * Currently stuck on Elligator after 19 hours.
+         * [FIXED] at least for now.
         
     * Vector-mul-chains
     * Negation "bubble pushing" optimization

include/goldilocks.h

+/* Copyright (c) 2014 Cryptography Research, Inc.
+ * Released under the MIT License.  See LICENSE.txt for license information.
+ */
+
+/**
+ * @file goldilocks.h
+ * @author Mike Hamburg
+ * @brief Goldilocks high-level functions.
+ */
+#ifndef __GOLDILOCKS_H__
+#define __GOLDILOCKS_H__ 1
+
+#include <stdint.h>
+
+/**
+ * @brief Serialized form of a Goldilocks public key.
+ *
+ * @warning This isn't even my final form!
+ */
+struct goldilocks_public_key_t {
+    uint8_t opaque[56]; /**< Serialized data. */
+};
+
+/**
+ * @brief Serialized form of a Goldilocks private key.
+ *
+ * Contains 56 bytes of actual private key, 56 bytes of
+ * public key, and 32 bytes of symmetric key for randomization.
+ *
+ * @warning This isn't even my final form!
+ */
+struct goldilocks_private_key_t {
+    uint8_t opaque[144]; /**< Serialized data. */
+};
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/** @brief No error. */
+static const int GOLDI_EOK      = 0;
+
+/** @brief Error: your key or other state is corrupt. */
+static const int GOLDI_ECORRUPT = 44801;
+
+/** @brief Error: other party's key is corrupt. */
+static const int GOLDI_EINVAL   = 44802;
+
+/** @brief Error: not enough entropy. */
+static const int GOLDI_ENODICE  = 44804;
+
+/** @brief Error: you need to initialize the library first. */
+static const int GOLDI_EUNINIT  = 44805;
+
+/** @brief Error: called init() but we are already initialized. */
+static const int GOLDI_EALREADYINIT  = 44805;
+
+/**
+ * @brief Initialize Goldilocks' precomputed tables and
+ * random number generator.  This function must be called before
+ * any of the other Goldilocks routines (except
+ * goldilocks_shared_secret in the current version) and should be
+ * called only once per process.
+ *
+ * There is currently no way to tear down this state.  It is possible
+ * that a future version of this library will not require this function.
+ *
+ * @retval GOLDI_EOK Success.
+ * @retval GOLDI_EALREADYINIT Already initialized.
+ * @retval GOLDI_ECORRUPT Memory is corrupted, or another thread is already init'ing.
+ * @retval Nonzero An error occurred.
+ */
+int
+goldilocks_init ()
+__attribute__((warn_unused_result));
+
+
+/**
+ * @brief Generate a new random keypair.
+ * @param [out] privkey The generated private key.
+ * @param [out] pubkey The generated public key.
+ *
+ * @warning This isn't even my final form!
+ *
+ * @retval GOLDI_EOK Success.
+ * @retval GOLDI_ENODICE Insufficient entropy.
+ * @retval GOLDI_EUNINIT You must call goldilocks_init() first.
+ */
+int
+goldilocks_keygen (
+    struct goldilocks_private_key_t *privkey,
+    struct goldilocks_public_key_t *pubkey
+) __attribute__((warn_unused_result,nonnull(1,2)));
+
+/**
+ * @brief Extract the public key from a private key.
+ *
+ * This is essentially a memcpy from the public part of the privkey.
+ *    
+ * @param [out] pubkey The extracted private key.
+ * @param [in] privkey The private key.
+ *
+ * @retval GOLDI_EOK Success.
+ * @retval GOLDI_ECORRUPT The private key is corrupt.
+ */
+int
+goldilocks_private_to_public (
+    struct goldilocks_public_key_t *pubkey,
+    const struct goldilocks_private_key_t *privkey
+) __attribute__((nonnull(1,2)));
+
+/**
+ * @brief Generate a Diffie-Hellman shared secret in constant time.
+ *
+ * This function uses some compile-time flags whose merit remains to
+ * be decided.
+ *
+ * If the flag EXPERIMENT_ECDH_OBLITERATE_CT is set, prepend 40 bytes
+ * of zeros to the secret before hashing.  In the case that the other
+ * party's key is detectably corrupt, instead the symmetric part
+ * of the secret key is used to produce a pseudorandom value.
+ *
+ * If EXPERIMENT_ECDH_STIR_IN_PUBKEYS is set, the sum and product of
+ * the two parties' public keys is prepended to the hash.
+ *
+ * In the current version, this function can safely be run even without
+ * goldilocks_init().  But this property is not guaranteed for future
+ * versions, so call it anyway.
+ *
+ * @warning This isn't even my final form!
+ *
+ * @param [out] shared The shared secret established with the other party.
+ * @param [in] my_privkey My private key.
+ * @param [in] your_pubkey The other party's public key.
+ *
+ * @retval GOLDI_EOK Success.
+ * @retval GOLDI_ECORRUPT My key is corrupt.
+ * @retval GOLDI_EINVAL   The other party's key is corrupt.
+ * @retval GOLDI_EUNINIT You must call goldilocks_init() first.
+ */
+int
+goldilocks_shared_secret (
+    uint8_t shared[64],
+    const struct goldilocks_private_key_t *my_privkey,
+    const struct goldilocks_public_key_t *your_pubkey
+) __attribute__((warn_unused_result,nonnull(1,2,3)));
+    
+/**
+ * @brief Sign a message.
+ *
+ * The signature is deterministic, using the symmetric secret found in the
+ * secret key to form a nonce.
+ *
+ * The technique used in signing is a modified Schnorr system, like EdDSA.
+ *
+ * @warning This isn't even my final form!
+ *
+ * @param [out] signature_out Space for the output signature.
+ * @param [in] message The message to be signed.
+ * @param [in] message_len The length of the message to be signed.
+ * @param [in] privkey My private key.
+ *
+ * @retval GOLDI_EOK Success.
+ * @retval GOLDI_ECORRUPT My key is corrupt.
+ * @retval GOLDI_EUNINIT You must call goldilocks_init() first.
+ */
+int
+goldilocks_sign (
+    uint8_t signature_out[56*2],
+    const uint8_t *message,
+    uint64_t message_len,
+    const struct goldilocks_private_key_t *privkey
+) __attribute__((nonnull(1,2,4)));
+
+/**
+ * @brief Verify a signature.
+ *
+ * This function is fairly strict.  It will correctly detect when
+ * the signature has the wrong cofactor component, or when the sig
+ * values aren't less than p or q.
+ * 
+ * Currently this function does not detect when the public key is weird,
+ * eg 0, has cofactor, etc.  As a result, a party with a bogus public
+ * key could create signatures that succeed on some systems and fail on
+ * others.
+ *
+ * @warning This isn't even my final form!
+ *
+ * @param [in] signature The signature.
+ * @param [in] message The message to be verified.
+ * @param [in] message_len The length of the message to be verified.
+ * @param [in] pubkey The signer's public key.
+ *
+ * @retval GOLDI_EOK Success.
+ * @retval GOLDI_EINVAL The public key or signature is corrupt.
+ * @retval GOLDI_EUNINIT You must call goldilocks_init() first.
+ */
+int
+goldilocks_verify (
+    const uint8_t signature[56*2],
+    const uint8_t *message,
+    uint64_t message_len,
+    const struct goldilocks_public_key_t *pubkey
+) __attribute__((warn_unused_result,nonnull(1,2,4)));
+
+#ifdef __cplusplus
+}; /* extern "C" */
+#endif
+
+#endif /* __GOLDILOCKS_H__ */

src/arch_32/p448.c

+/* Copyright (c) 2014 Cryptography Research, Inc.
+ * Released under the MIT License.  See LICENSE.txt for license information.
+ */
+
+#include "word.h"
+#include "p448.h"
+//#include "x86-64-arith.h"
+
+static inline mask_t __attribute__((always_inline))
+is_zero (
+    word_t x
+) {
+    dword_t xx = x;
+    xx--;
+    return xx >> WORD_BITS;
+}
+
+static uint64_t widemul_32 (
+    const uint32_t a,
+    const uint32_t b
+) {
+    return ((uint64_t)a)* b;
+}
+
+void
+p448_mul (
+    p448_t *__restrict__ cs,
+    const p448_t *as,
+    const p448_t *bs
+) {
+    // p448_t ar, br;
+//     p448_copy(&ar,as);
+//     p448_copy(&br,bs);
+//     p448_weak_reduce(&ar);
+//     p448_weak_reduce(&br);
+    
+    const uint32_t *a = as->limb, *b = bs->limb;
+    uint32_t *c = cs->limb;
+
+    uint64_t accum0 = 0, accum1 = 0, accum2 = 0;
+    uint32_t mask = (1<<28) - 1;  
+
+    uint32_t aa[8], bb[8];
+
+    /* For some reason clang doesn't vectorize this without prompting? */
+    // unsigned int i;
+    // for (i=0; i<sizeof(aa)/sizeof(uint64xn_t); i++) {
+    //     ((uint64xn_t*)aa)[i] = ((const uint64xn_t*)a)[i] + ((const uint64xn_t*)(&a[4]))[i];
+    //     ((uint64xn_t*)bb)[i] = ((const uint64xn_t*)b)[i] + ((const uint64xn_t*)(&b[4]))[i];     
+    // }
+    int i,j;
+    for (i=0; i<8; i++) {
+        aa[i] = a[i] + a[i+8];
+        bb[i] = b[i] + b[i+8];
+    }
+    
+    for (j=0; j<8; j++) {
+        accum2 = 0;
+    
+        for (i=0; i<=j; i++) {      
+            accum2 += widemul_32(a[j-i],b[i]);
+            accum1 += widemul_32(aa[j-i],bb[i]);
+            accum0 += widemul_32(a[8+j-i], b[8+i]);
+        }
+        
+        accum1 -= accum2;
+        accum0 += accum2;
+        accum2 = 0;
+        
+        for (; i<8; i++) {
+            accum0 -= widemul_32(a[8+j-i], b[i]);
+            accum2 += widemul_32(aa[8+j-i], bb[i]);
+            accum1 += widemul_32(a[16+j-i], b[8+i]);
+        }
+
+        accum1 += accum2;
+        accum0 += accum2;
+
+        c[j] = ((uint32_t)(accum0)) & mask;
+        c[j+8] = ((uint32_t)(accum1)) & mask;
+
+        accum0 >>= 28;
+        accum1 >>= 28;
+    }
+    
+    accum0 += accum1;
+    accum0 += c[8];
+    accum1 += c[0];
+    c[8] = ((uint32_t)(accum0)) & mask;
+    c[0] = ((uint32_t)(accum1)) & mask;
+    
+    accum0 >>= 28;
+    accum1 >>= 28;
+    c[9] += ((uint32_t)(accum0));
+    c[1] += ((uint32_t)(accum1));
+}
+
+void
+p448_mulw (
+    p448_t *__restrict__ cs,
+    const p448_t *as,
+    uint64_t b
+) {
+    const uint32_t bhi = b>>28, blo = b & (1<<28)-1;
+    
+    const uint32_t *a = as->limb;
+    uint32_t *c = cs->limb;
+
+    uint64_t accum0, accum8;
+    uint32_t mask = (1ull<<28)-1;  
+
+    int i;
+
+    accum0 = widemul_32(blo, a[0]);
+    accum8 = widemul_32(blo, a[8]);
+    accum0 += widemul_32(bhi, a[15]);
+    accum8 += widemul_32(bhi, a[15] + a[7]);
+
+    c[0] = accum0 & mask; accum0 >>= 28;
+    c[8] = accum8 & mask; accum8 >>= 28;
+    
+    for (i=1; i<8; i++) {
+        accum0 += widemul_32(blo, a[i]);
+        accum8 += widemul_32(blo, a[i+8]);
+        
+        accum0 += widemul_32(bhi, a[i-1]);
+        accum8 += widemul_32(bhi, a[i+7]);
+
+        c[i] = accum0 & mask; accum0 >>= 28;
+        c[i+8] = accum8 & mask; accum8 >>= 28;
+    }
+
+    accum0 += accum8 + c[8];
+    c[8] = accum0 & mask;
+    c[9] += accum0 >> 28;
+
+    accum8 += c[0];
+    c[0] = accum8 & mask;
+    c[1] += accum8 >> 28;
+}
+
+void
+p448_sqr (
+    p448_t *__restrict__ cs,
+    const p448_t *as
+) {
+    p448_mul(cs,as,as); // PERF
+}
+
+void
+p448_strong_reduce (
+    p448_t *a
+) {
+    word_t mask = (1ull<<28)-1;
+
+    /* first, clear high */
+    a->limb[8] += a->limb[15]>>28;
+    a->limb[0] += a->limb[15]>>28;
+    a->limb[15] &= mask;
+
+    /* now the total is less than 2^448 - 2^(448-56) + 2^(448-56+8) < 2p */
+
+    /* compute total_value - p.  No need to reduce mod p. */