Two new Poly1305 implementations: ctmul32 uses pure 32-bit multiplications (MUL15, constant-time on about everything); i15 uses the generic i15 big integers (MUL15 again), which is quite slow but also small.

Makefile

@@ -57,7 +57,7 @@ OBJMAC = $(BUILD)/hmac.o $(BUILD)/hmac_ct.o
 OBJRAND = $(BUILD)/hmac_drbg.o
 OBJRSA = $(BUILD)/rsa_i15_pkcs1_sign.o $(BUILD)/rsa_i15_pkcs1_vrfy.o $(BUILD)/rsa_i15_priv.o $(BUILD)/rsa_i15_pub.o $(BUILD)/rsa_i31_pkcs1_sign.o $(BUILD)/rsa_i31_pkcs1_vrfy.o $(BUILD)/rsa_i31_priv.o $(BUILD)/rsa_i31_pub.o $(BUILD)/rsa_i32_pkcs1_sign.o $(BUILD)/rsa_i32_pkcs1_vrfy.o $(BUILD)/rsa_i32_priv.o $(BUILD)/rsa_i32_pub.o $(BUILD)/rsa_pkcs1_sig_pad.o $(BUILD)/rsa_pkcs1_sig_unpad.o $(BUILD)/rsa_ssl_decrypt.o
 OBJSSL = $(BUILD)/prf.o $(BUILD)/prf_md5sha1.o $(BUILD)/prf_sha256.o $(BUILD)/prf_sha384.o $(BUILD)/ssl_ccert_single_ec.o $(BUILD)/ssl_ccert_single_rsa.o $(BUILD)/ssl_client.o $(BUILD)/ssl_client_full.o $(BUILD)/ssl_engine.o $(BUILD)/ssl_hashes.o $(BUILD)/ssl_hs_client.o $(BUILD)/ssl_hs_server.o $(BUILD)/ssl_io.o $(BUILD)/ssl_lru.o $(BUILD)/ssl_rec_cbc.o $(BUILD)/ssl_rec_chapol.o $(BUILD)/ssl_rec_gcm.o $(BUILD)/ssl_server.o $(BUILD)/ssl_server_mine2c.o $(BUILD)/ssl_server_mine2g.o $(BUILD)/ssl_server_minf2c.o $(BUILD)/ssl_server_minf2g.o $(BUILD)/ssl_server_minr2g.o $(BUILD)/ssl_server_minu2g.o $(BUILD)/ssl_server_minv2g.o $(BUILD)/ssl_server_full_ec.o $(BUILD)/ssl_server_full_rsa.o $(BUILD)/ssl_scert_single_ec.o $(BUILD)/ssl_scert_single_rsa.o
-OBJSYMCIPHER = $(BUILD)/aes_big_cbcdec.o $(BUILD)/aes_big_cbcenc.o $(BUILD)/aes_big_ctr.o $(BUILD)/aes_big_dec.o $(BUILD)/aes_big_enc.o $(BUILD)/aes_common.o $(BUILD)/aes_ct.o $(BUILD)/aes_ct64.o $(BUILD)/aes_ct64_cbcdec.o $(BUILD)/aes_ct64_cbcenc.o $(BUILD)/aes_ct64_ctr.o $(BUILD)/aes_ct64_dec.o $(BUILD)/aes_ct64_enc.o $(BUILD)/aes_ct_cbcdec.o $(BUILD)/aes_ct_cbcenc.o $(BUILD)/aes_ct_ctr.o $(BUILD)/aes_ct_dec.o $(BUILD)/aes_ct_enc.o $(BUILD)/aes_small_cbcdec.o $(BUILD)/aes_small_cbcenc.o $(BUILD)/aes_small_ctr.o $(BUILD)/aes_small_dec.o $(BUILD)/aes_small_enc.o $(BUILD)/chacha20_ct.o $(BUILD)/des_ct.o $(BUILD)/des_ct_cbcdec.o $(BUILD)/des_ct_cbcenc.o $(BUILD)/des_support.o $(BUILD)/des_tab.o $(BUILD)/des_tab_cbcdec.o $(BUILD)/des_tab_cbcenc.o $(BUILD)/poly1305_ctmul.o
+OBJSYMCIPHER = $(BUILD)/aes_big_cbcdec.o $(BUILD)/aes_big_cbcenc.o $(BUILD)/aes_big_ctr.o $(BUILD)/aes_big_dec.o $(BUILD)/aes_big_enc.o $(BUILD)/aes_common.o $(BUILD)/aes_ct.o $(BUILD)/aes_ct64.o $(BUILD)/aes_ct64_cbcdec.o $(BUILD)/aes_ct64_cbcenc.o $(BUILD)/aes_ct64_ctr.o $(BUILD)/aes_ct64_dec.o $(BUILD)/aes_ct64_enc.o $(BUILD)/aes_ct_cbcdec.o $(BUILD)/aes_ct_cbcenc.o $(BUILD)/aes_ct_ctr.o $(BUILD)/aes_ct_dec.o $(BUILD)/aes_ct_enc.o $(BUILD)/aes_small_cbcdec.o $(BUILD)/aes_small_cbcenc.o $(BUILD)/aes_small_ctr.o $(BUILD)/aes_small_dec.o $(BUILD)/aes_small_enc.o $(BUILD)/chacha20_ct.o $(BUILD)/des_ct.o $(BUILD)/des_ct_cbcdec.o $(BUILD)/des_ct_cbcenc.o $(BUILD)/des_support.o $(BUILD)/des_tab.o $(BUILD)/des_tab_cbcdec.o $(BUILD)/des_tab_cbcenc.o $(BUILD)/poly1305_ctmul.o $(BUILD)/poly1305_ctmul32.o $(BUILD)/poly1305_i15.o
 OBJX509 = $(BUILD)/skey_decoder.o $(BUILD)/x509_decoder.o $(BUILD)/x509_knownkey.o $(BUILD)/x509_minimal.o $(BUILD)/x509_minimal_full.o
 OBJ = $(OBJCODEC) $(OBJEC) $(OBJHASH) $(OBJINT15) $(OBJINT31) $(OBJINT32) $(OBJMAC) $(OBJRAND) $(OBJRSA) $(OBJSSL) $(OBJSYMCIPHER) $(OBJX509)
 OBJBRSSL = $(BUILD)/brssl.o $(BUILD)/certs.o $(BUILD)/chain.o $(BUILD)/client.o $(BUILD)/errors.o $(BUILD)/files.o $(BUILD)/keys.o $(BUILD)/names.o $(BUILD)/server.o $(BUILD)/skey.o $(BUILD)/sslio.o $(BUILD)/ta.o $(BUILD)/vector.o $(BUILD)/verify.o $(BUILD)/xmem.o
@@ -606,6 +606,12 @@ $(BUILD)/des_tab_cbcenc.o: src/symcipher/des_tab_cbcenc.c $(HEADERS)
 $(BUILD)/poly1305_ctmul.o: src/symcipher/poly1305_ctmul.c $(HEADERS)
 	$(CC) $(CFLAGS) -c -o $(BUILD)/poly1305_ctmul.o src/symcipher/poly1305_ctmul.c
 
+$(BUILD)/poly1305_ctmul32.o: src/symcipher/poly1305_ctmul32.c $(HEADERS)
+	$(CC) $(CFLAGS) -c -o $(BUILD)/poly1305_ctmul32.o src/symcipher/poly1305_ctmul32.c
+
+$(BUILD)/poly1305_i15.o: src/symcipher/poly1305_i15.c $(HEADERS)
+	$(CC) $(CFLAGS) -c -o $(BUILD)/poly1305_i15.o src/symcipher/poly1305_i15.c
+
 $(BUILD)/skey_decoder.o: src/x509/skey_decoder.c $(HEADERS)
 	$(CC) $(CFLAGS) -c -o $(BUILD)/skey_decoder.o src/x509/skey_decoder.c
 

inc/bearssl_block.h

@@ -1323,4 +1323,48 @@ void br_poly1305_ctmul_run(const void *key, const void *iv,
 	void *data, size_t len, const void *aad, size_t aad_len,
 	void *tag, br_chacha20_run ichacha, int encrypt);
 
+/**
+ * \brief ChaCha20+Poly1305 AEAD implementation (pure 32-bit multiplications).
+ *
+ * \see br_poly1305_run
+ *
+ * \param key       secret key (32 bytes).
+ * \param iv        nonce (12 bytes).
+ * \param data      data to encrypt or decrypt.
+ * \param len       data length (in bytes).
+ * \param aad       additional authenticated data.
+ * \param aad_len   length of additional authenticated data (in bytes).
+ * \param tag       output buffer for the authentication tag.
+ * \param ichacha   implementation of ChaCha20.
+ * \param encrypt   non-zero for encryption, zero for decryption.
+ */
+void br_poly1305_ctmul32_run(const void *key, const void *iv,
+	void *data, size_t len, const void *aad, size_t aad_len,
+	void *tag, br_chacha20_run ichacha, int encrypt);
+
+/**
+ * \brief ChaCha20+Poly1305 AEAD implementation (i15).
+ *
+ * This implementation relies on the generic big integer code "i15"
+ * (which uses pure 32-bit multiplications). As such, it may save a
+ * little code footprint in a context where "i15" is already included
+ * (e.g. for elliptic curves or for RSA); however, it is also
+ * substantially slower than the ctmul and ctmul32 implementations.
+ *
+ * \see br_poly1305_run
+ *
+ * \param key       secret key (32 bytes).
+ * \param iv        nonce (12 bytes).
+ * \param data      data to encrypt or decrypt.
+ * \param len       data length (in bytes).
+ * \param aad       additional authenticated data.
+ * \param aad_len   length of additional authenticated data (in bytes).
+ * \param tag       output buffer for the authentication tag.
+ * \param ichacha   implementation of ChaCha20.
+ * \param encrypt   non-zero for encryption, zero for decryption.
+ */
+void br_poly1305_i15_run(const void *key, const void *iv,
+	void *data, size_t len, const void *aad, size_t aad_len,
+	void *tag, br_chacha20_run ichacha, int encrypt);
+
 #endif

src/inner.h

@@ -571,16 +571,17 @@ MUL31_lo(uint32_t x, uint32_t y)
 #endif
 
 /*
- * Multiply two words together; each word may contain up to 15 bits of
+ * Multiply two words together; the sum of the lengths of the two
- * data. If BR_CT_MUL15 is non-zero, then the macro will contain some
+ * operands must not exceed 31 (for instance, one operand may use 16
- * extra operations that help in making the operation constant-time on
+ * bits if the other fits on 15). If BR_CT_MUL15 is non-zero, then the
- * some platforms, where the basic 32-bit multiplication is not
+ * macro will contain some extra operations that help in making the
- * constant-time.
+ * operation constant-time on some platforms, where the basic 32-bit
+ * multiplication is not constant-time.
  */
 #if BR_CT_MUL15
 #define MUL15(x, y)   (((uint32_t)(x) | (uint32_t)0x80000000) \
                        * ((uint32_t)(y) | (uint32_t)0x80000000) \
-		       & (uint32_t)0x3FFFFFFF)
+		       & (uint32_t)0x7FFFFFFF)
 #else
 #define MUL15(x, y)   ((uint32_t)(x) * (uint32_t)(y))
 #endif

src/symcipher/poly1305_ctmul32.c

@@ -0,0 +1,297 @@
+/*
+ * Copyright (c) 2017 Thomas Pornin <pornin@bolet.org>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining 
+ * a copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sublicense, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be 
+ * included in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "inner.h"
+
+/*
+ * Perform the inner processing of blocks for Poly1305.
+ */
+static void
+poly1305_inner(uint32_t *a, const uint32_t *r, const void *data, size_t len)
+{
+	/*
+	 * Implementation notes: we split the 130-bit values into ten
+	 * 13-bit words. This gives us some space for carries and allows
+	 * using only 32x32->32 multiplications, which are way faster than
+	 * 32x32->64 multiplications on the ARM Cortex-M0/M0+, and also
+	 * help in making constant-time code on the Cortex-M3.
+	 *
+	 * Since we compute modulo 2^130-5, the "upper words" become
+	 * low words with a factor of 5; that is, x*2^130 = x*5 mod p.
+	 * This has already been integrated in the r[] array, which
+	 * is extended to the 0..18 range.
+	 *
+	 * In each loop iteration, a[] and r[] words are 13-bit each,
+	 * except a[1] which may use 14 bits.
+	 */
+	const unsigned char *buf;
+
+	buf = data;
+	while (len > 0) {
+		unsigned char tmp[16];
+		uint32_t b[10];
+		unsigned u, v;
+		uint32_t z, cc1, cc2;
+
+		/*
+		 * If there is a partial block, right-pad it with zeros.
+		 */
+		if (len < 16) {
+			memset(tmp, 0, sizeof tmp);
+			memcpy(tmp, buf, len);
+			buf = tmp;
+			len = 16;
+		}
+
+		/*
+		 * Decode next block and apply the "high bit"; that value
+		 * is added to the accumulator.
+		 */
+		v = br_dec16le(buf);
+		a[0] += v & 0x01FFF;
+		v >>= 13;
+		v |= buf[2] << 3;
+		v |= buf[3] << 11;
+		a[1] += v & 0x01FFF;
+		v >>= 13;
+		v |= buf[4] << 6;
+		a[2] += v & 0x01FFF;
+		v >>= 13;
+		v |= buf[5] << 1;
+		v |= buf[6] << 9;
+		a[3] += v & 0x01FFF;
+		v >>= 13;
+		v |= buf[7] << 4;
+		v |= buf[8] << 12;
+		a[4] += v & 0x01FFF;
+		v >>= 13;
+		v |= buf[9] << 7;
+		a[5] += v & 0x01FFF;
+		v >>= 13;
+		v |= buf[10] << 2;
+		v |= buf[11] << 10;
+		a[6] += v & 0x01FFF;
+		v >>= 13;
+		v |= buf[12] << 5;
+		a[7] += v & 0x01FFF;
+		v = br_dec16le(buf + 13);
+		a[8] += v & 0x01FFF;
+		v >>= 13;
+		v |= buf[15] << 3;
+		a[9] += v | 0x00800;
+
+		/*
+		 * At that point, all a[] values fit on 14 bits, while
+		 * all r[] values fit on 13 bits. Thus products fit on
+		 * 27 bits, and we can accumulate up to 31 of them in
+		 * a 32-bit word and still have some room for carries.
+		 */
+
+		/*
+		 * Now a[] contains words with values up to 14 bits each.
+		 * We perform the multiplication with r[].
+		 *
+		 * The extended words of r[] may be larger than 13 bits
+		 * (they are 5 times a 13-bit word) so the full summation
+		 * may yield values up to 46 times a 27-bit word, which
+		 * does not fit on a 32-bit word. To avoid that issue, we
+		 * must split the loop below in two, with a carry
+		 * propagation operation in the middle.
+		 */
+		cc1 = 0;
+		for (u = 0; u < 10; u ++) {
+			uint32_t s;
+
+			s = cc1
+				+ MUL15(a[0], r[u + 9 - 0])
+				+ MUL15(a[1], r[u + 9 - 1])
+				+ MUL15(a[2], r[u + 9 - 2])
+				+ MUL15(a[3], r[u + 9 - 3])
+				+ MUL15(a[4], r[u + 9 - 4]);
+			b[u] = s & 0x1FFF;
+			cc1 = s >> 13;
+		}
+		cc2 = 0;
+		for (u = 0; u < 10; u ++) {
+			uint32_t s;
+
+			s = b[u] + cc2
+				+ MUL15(a[5], r[u + 9 - 5])
+				+ MUL15(a[6], r[u + 9 - 6])
+				+ MUL15(a[7], r[u + 9 - 7])
+				+ MUL15(a[8], r[u + 9 - 8])
+				+ MUL15(a[9], r[u + 9 - 9]);
+			b[u] = s & 0x1FFF;
+			cc2 = s >> 13;
+		}
+		memcpy(a, b, sizeof b);
+
+		/*
+		 * The two carries "loop back" with a factor of 5. We
+		 * propagate them into a[0] and a[1].
+		 */
+		z = cc1 + cc2;
+		z += (z << 2) + a[0];
+		a[0] = z & 0x1FFF;
+		a[1] += z >> 13;
+
+		buf += 16;
+		len -= 16;
+	}
+}
+
+/* see bearssl_block.h */
+void
+br_poly1305_ctmul32_run(const void *key, const void *iv,
+	void *data, size_t len, const void *aad, size_t aad_len,
+	void *tag, br_chacha20_run ichacha, int encrypt)
+{
+	unsigned char pkey[32], foot[16];
+	uint32_t z, r[19], acc[10], cc, ctl;
+	int i;
+
+	/*
+	 * Compute the MAC key. The 'r' value is the first 16 bytes of
+	 * pkey[].
+	 */
+	memset(pkey, 0, sizeof pkey);
+	ichacha(key, iv, 0, pkey, sizeof pkey);
+
+	/*
+	 * If encrypting, ChaCha20 must run first, followed by Poly1305.
+	 * When decrypting, the operations are reversed.
+	 */
+	if (encrypt) {
+		ichacha(key, iv, 1, data, len);
+	}
+
+	/*
+	 * Run Poly1305. We must process the AAD, then ciphertext, then
+	 * the footer (with the lengths). Note that the AAD and ciphertext
+	 * are meant to be padded with zeros up to the next multiple of 16,
+	 * and the length of the footer is 16 bytes as well.
+	 */
+
+	/*
+	 * Decode the 'r' value into 13-bit words, with the "clamping"
+	 * operation applied.
+	 */
+	z = br_dec32le(pkey) & 0x03FFFFFF;
+	r[9] = z & 0x1FFF;
+	r[10] = z >> 13;
+	z = (br_dec32le(pkey +  3) >> 2) & 0x03FFFF03;
+	r[11] = z & 0x1FFF;
+	r[12] = z >> 13;
+	z = (br_dec32le(pkey +  6) >> 4) & 0x03FFC0FF;
+	r[13] = z & 0x1FFF;
+	r[14] = z >> 13;
+	z = (br_dec32le(pkey +  9) >> 6) & 0x03F03FFF;
+	r[15] = z & 0x1FFF;
+	r[16] = z >> 13;
+	z = (br_dec32le(pkey + 12) >> 8) & 0x000FFFFF;
+	r[17] = z & 0x1FFF;
+	r[18] = z >> 13;
+
+	/*
+	 * Extend r[] with the 5x factor pre-applied.
+	 */
+	for (i = 0; i < 9; i ++) {
+		r[i] = MUL15(5, r[i + 10]);
+	}
+
+	/*
+	 * Accumulator is 0.
+	 */
+	memset(acc, 0, sizeof acc);
+
+	/*
+	 * Process the additional authenticated data, ciphertext, and
+	 * footer in due order.
+	 */
+	br_enc64le(foot, (uint64_t)aad_len);
+	br_enc64le(foot + 8, (uint64_t)len);
+	poly1305_inner(acc, r, aad, aad_len);
+	poly1305_inner(acc, r, data, len);
+	poly1305_inner(acc, r, foot, sizeof foot);
+
+	/*
+	 * Finalise modular reduction. This is done with carry propagation
+	 * and applying the '2^130 = -5 mod p' rule. Note that the output
+	 * of poly1035_inner() is already mostly reduced, since only
+	 * acc[1] may be (very slightly) above 2^13. A single loop back
+	 * to acc[1] will be enough to make the value fit in 130 bits.
+	 */
+	cc = 0;
+	for (i = 1; i < 10; i ++) {
+		z = acc[i] + cc;
+		acc[i] = z & 0x1FFF;
+		cc = z >> 13;
+	}
+	z = acc[0] + cc + (cc << 2);
+	acc[0] = z & 0x1FFF;
+	acc[1] += z >> 13;
+
+	/*
+	 * We may still have a value in the 2^130-5..2^130-1 range, in
+	 * which case we must reduce it again. The code below selects,
+	 * in constant-time, between 'acc' and 'acc-p',
+	 */
+	ctl = GT(acc[0], 0x1FFA);
+	for (i = 1; i < 10; i ++) {
+		ctl &= EQ(acc[i], 0x1FFF);
+	}
+	acc[0] = MUX(ctl, acc[0] - 0x1FFB, acc[0]);
+	for (i = 1; i < 10; i ++) {
+		acc[i] &= ~(-ctl);
+	}
+
+	/*
+	 * Convert back the accumulator to 32-bit words, and add the
+	 * 's' value (second half of pkey[]). That addition is done
+	 * modulo 2^128.
+	 */
+	z = acc[0] + (acc[1] << 13) + br_dec16le(pkey + 16);
+	br_enc16le((unsigned char *)tag, z & 0xFFFF);
+	z = (z >> 16) + (acc[2] << 10) + br_dec16le(pkey + 18);
+	br_enc16le((unsigned char *)tag + 2, z & 0xFFFF);
+	z = (z >> 16) + (acc[3] << 7) + br_dec16le(pkey + 20);
+	br_enc16le((unsigned char *)tag + 4, z & 0xFFFF);
+	z = (z >> 16) + (acc[4] << 4) + br_dec16le(pkey + 22);
+	br_enc16le((unsigned char *)tag + 6, z & 0xFFFF);
+	z = (z >> 16) + (acc[5] << 1) + (acc[6] << 14) + br_dec16le(pkey + 24);
+	br_enc16le((unsigned char *)tag + 8, z & 0xFFFF);
+	z = (z >> 16) + (acc[7] << 11) + br_dec16le(pkey + 26);
+	br_enc16le((unsigned char *)tag + 10, z & 0xFFFF);
+	z = (z >> 16) + (acc[8] << 8) + br_dec16le(pkey + 28);
+	br_enc16le((unsigned char *)tag + 12, z & 0xFFFF);
+	z = (z >> 16) + (acc[9] << 5) + br_dec16le(pkey + 30);
+	br_enc16le((unsigned char *)tag + 14, z & 0xFFFF);
+
+	/*
+	 * If decrypting, then ChaCha20 runs _after_ Poly1305.
+	 */
+	if (!encrypt) {
+		ichacha(key, iv, 1, data, len);
+	}
+}

src/symcipher/poly1305_i15.c

@@ -0,0 +1,221 @@
+/*
+ * Copyright (c) 2017 Thomas Pornin <pornin@bolet.org>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining 
+ * a copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sublicense, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be 
+ * included in all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "inner.h"
+
+/*
+ * This is a "reference" implementation of Poly1305 that uses the
+ * generic "i15" code for big integers. It is slow, but it handles all
+ * big-integer operations with generic code, thereby avoiding most
+ * tricky situations with carry propagation and modular reduction.
+ */
+
+/*
+ * Modulus: 2^130-5.
+ */
+static const uint16_t P1305[] = {
+	0x008A,
+	0x7FFB, 0x7FFF, 0x7FFF, 0x7FFF, 0x7FFF, 0x7FFF, 0x7FFF, 0x7FFF, 0x03FF
+};
+
+/*
+ * -p mod 2^15.
+ */
+#define P0I   0x4CCD
+
+/*
+ * R^2 mod p, for conversion to Montgomery representation (R = 2^135,
+ * since we use 9 words of 15 bits each, and 15*9 = 135).
+ */
+static const uint16_t R2[] = {
+	0x008A,
+	0x6400, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000, 0x0000
+};
+
+/*
+ * Perform the inner processing of blocks for Poly1305. The "r" array
+ * is in Montgomery representation, while the "a" array is not.
+ */
+static void
+poly1305_inner(uint16_t *a, const uint16_t *r, const void *data, size_t len)
+{
+	const unsigned char *buf;
+
+	buf = data;
+	while (len > 0) {
+		unsigned char tmp[16], rev[16];
+		uint16_t b[10];
+		uint32_t ctl;
+		int i;
+
+		/*
+		 * If there is a partial block, right-pad it with zeros.
+		 */
+		if (len < 16) {
+			memset(tmp, 0, sizeof tmp);
+			memcpy(tmp, buf, len);
+			buf = tmp;
+			len = 16;
+		}
+
+		/*
+		 * Decode next block and apply the "high bit". Since
+		 * decoding is little-endian, we must byte-swap the buffer.
+		 */
+		for (i = 0; i < 16; i ++) {
+			rev[i] = buf[15 - i];
+		}
+		br_i15_decode_mod(b, rev, sizeof rev, P1305);
+		b[9] |= 0x0100;
+
+		/*
+		 * Add the accumulator to the decoded block (modular
+		 * addition).
+		 */
+		ctl = br_i15_add(b, a, 1);
+		ctl |= NOT(br_i15_sub(b, P1305, 0));
+		br_i15_sub(b, P1305, ctl);
+
+		/*
+		 * Multiply by r, result is the new accumulator value.
+		 */
+		br_i15_montymul(a, b, r, P1305, P0I);
+
+		buf += 16;
+		len -= 16;
+	}
+}
+
+/*
+ * Byteswap a 16-byte value.
+ */
+static void
+byteswap16(unsigned char *buf)
+{
+	int i;
+
+	for (i = 0; i < 8; i ++) {
+		unsigned x;
+
+		x = buf[i];
+		buf[i] = buf[15 - i];
+		buf[15 - i] = x;
+	}
+}
+
+/* see bearssl_block.h */
+void
+br_poly1305_i15_run(const void *key, const void *iv,
+	void *data, size_t len, const void *aad, size_t aad_len,
+	void *tag, br_chacha20_run ichacha, int encrypt)
+{
+	unsigned char pkey[32], foot[16];
+	uint16_t t[10], r[10], acc[10];
+
+	/*
+	 * Compute the MAC key. The 'r' value is the first 16 bytes of
+	 * pkey[].
+	 */
+	memset(pkey, 0, sizeof pkey);
+	ichacha(key, iv, 0, pkey, sizeof pkey);
+
+	/*
+	 * If encrypting, ChaCha20 must run first, followed by Poly1305.
+	 * When decrypting, the operations are reversed.
+	 */
+	if (encrypt) {
+		ichacha(key, iv, 1, data, len);
+	}
+
+	/*
+	 * Run Poly1305. We must process the AAD, then ciphertext, then
+	 * the footer (with the lengths). Note that the AAD and ciphertext
+	 * are meant to be padded with zeros up to the next multiple of 16,
+	 * and the length of the footer is 16 bytes as well.
+	 */
+
+	/*
+	 * Apply the "clamping" operation on the encoded 'r' value.
+	 */
+	pkey[ 3] &= 0x0F;
+	pkey[ 7] &= 0x0F;
+	pkey[11] &= 0x0F;
+	pkey[15] &= 0x0F;
+	pkey[ 4] &= 0xFC;
+	pkey[ 8] &= 0xFC;
+	pkey[12] &= 0xFC;
+
+	/*
+	 * Decode the clamped 'r' value. Decoding should use little-endian
+	 * so we must byteswap the value first.
+	 */
+	byteswap16(pkey);
+	br_i15_decode_mod(t, pkey, 16, P1305);
+
+	/*
+	 * Convert 'r' to Montgomery representation.
+	 */
+	br_i15_montymul(r, t, R2, P1305, P0I);
+
+	/*
+	 * Accumulator is 0.
+	 */
+	br_i15_zero(acc, 0x8A);
+
+	/*
+	 * Process the additional authenticated data, ciphertext, and
+	 * footer in due order.
+	 */
+	br_enc64le(foot, (uint64_t)aad_len);
+	br_enc64le(foot + 8, (uint64_t)len);
+	poly1305_inner(acc, r, aad, aad_len);
+	poly1305_inner(acc, r, data, len);
+	poly1305_inner(acc, r, foot, sizeof foot);
+
+	/*
+	 * Decode the value 's'. Again, a byteswap is needed.
+	 */
+	byteswap16(pkey + 16);
+	br_i15_decode_mod(t, pkey + 16, 16, P1305);
+
+	/*
+	 * Add the value 's' to the accumulator. That addition is done
+	 * modulo 2^128, so we just ignore the carry.
+	 */
+	br_i15_add(acc, t, 1);
+
+	/*
+	 * Encode the result (128 low bits) to the tag. Encoding should
+	 * be little-endian.
+	 */
+	br_i15_encode(tag, 16, acc);
+	byteswap16(tag);
+
+	/*
+	 * If decrypting, then ChaCha20 runs _after_ Poly1305.
+	 */
+	if (!encrypt) {
+		ichacha(key, iv, 1, data, len);
+	}
+}

test/test_crypto.c

@@ -4113,11 +4113,13 @@ static const struct {
 };
 
 static void
-test_Poly1305_ctmul(void)
+test_Poly1305_inner(const char *name, br_poly1305_run ipoly,
+	br_poly1305_run iref)
 {
 	size_t u;
+	br_hmac_drbg_context rng;
 
-	printf("Test Poly1305_ctmul: ");
+	printf("Test %s: ", name);
 	fflush(stdout);
 
 	for (u = 0; KAT_POLY1305[u].skey; u ++) {
@@ -4133,7 +4135,7 @@ test_Poly1305_ctmul(void)
 		hextobin(tag, KAT_POLY1305[u].stag);
 
 		memcpy(data, plain, len);
-		br_poly1305_ctmul_run(key, nonce, data, len,
+		ipoly(key, nonce, data, len,
 			aad, aad_len, tmp, br_chacha20_ct_run, 1);
 		if (memcmp(data, cipher, len) != 0) {
 			fprintf(stderr, "ChaCha20+Poly1305 KAT failed (1)\n");
@@ -4143,7 +4145,7 @@ test_Poly1305_ctmul(void)
 			fprintf(stderr, "ChaCha20+Poly1305 KAT failed (2)\n");
 			exit(EXIT_FAILURE);
 		}
-		br_poly1305_ctmul_run(key, nonce, data, len,
+		ipoly(key, nonce, data, len,
 			aad, aad_len, tmp, br_chacha20_ct_run, 0);
 		if (memcmp(data, plain, len) != 0) {
 			fprintf(stderr, "ChaCha20+Poly1305 KAT failed (3)\n");
@@ -4158,10 +4160,66 @@ test_Poly1305_ctmul(void)
 		fflush(stdout);
 	}
 
+	printf(" ");
+	fflush(stdout);
+
+	/*
+	 * We compare the "ipoly" and "iref" implementations together on
+	 * a bunch of pseudo-random messages.
+	 */
+	br_hmac_drbg_init(&rng, &br_sha256_vtable, "seed for Poly1305", 17);
+	for (u = 0; u < 100; u ++) {
+		unsigned char plain[100], aad[100], tmp[100];
+		unsigned char key[32], iv[12], tag1[16], tag2[16];
+
+		br_hmac_drbg_generate(&rng, key, sizeof key);
+		br_hmac_drbg_generate(&rng, iv, sizeof iv);
+		br_hmac_drbg_generate(&rng, plain, u);
+		br_hmac_drbg_generate(&rng, aad, u);
+		memcpy(tmp, plain, u);
+		memset(tmp + u, 0xFF, (sizeof tmp) - u);
+		ipoly(key, iv, tmp, u, aad, u, tag1,
+			&br_chacha20_ct_run, 1);
+		memset(tmp + u, 0x00, (sizeof tmp) - u);
+		iref(key, iv, tmp, u, aad, u, tag2,
+			&br_chacha20_ct_run, 0);
+		if (memcmp(tmp, plain, u) != 0) {
+			fprintf(stderr, "cross enc/dec failed\n");
+			exit(EXIT_FAILURE);
+		}
+		if (memcmp(tag1, tag2, sizeof tag1) != 0) {
+			fprintf(stderr, "cross MAC failed\n");
+			exit(EXIT_FAILURE);
+		}
+		printf(".");
+		fflush(stdout);
+	}
+
 	printf(" done.\n");
 	fflush(stdout);
 }
 
+static void
+test_Poly1305_ctmul(void)
+{
+	test_Poly1305_inner("Poly1305_ctmul", &br_poly1305_ctmul_run,
+		&br_poly1305_i15_run);
+}
+
+static void
+test_Poly1305_ctmul32(void)
+{
+	test_Poly1305_inner("Poly1305_ctmul32", &br_poly1305_ctmul32_run,
+		&br_poly1305_i15_run);
+}
+
+static void
+test_Poly1305_i15(void)
+{
+	test_Poly1305_inner("Poly1305_i15", &br_poly1305_i15_run,
+		&br_poly1305_ctmul_run);
+}
+
 /*
  * A 1024-bit RSA key, generated with OpenSSL.
  */
@@ -5380,6 +5438,8 @@ static const struct {
 	STU(DES_ct),
 	STU(ChaCha20_ct),
 	STU(Poly1305_ctmul),
+	STU(Poly1305_ctmul32),
+	STU(Poly1305_i15),
 	STU(RSA_i15),
 	STU(RSA_i31),
 	STU(RSA_i32),

test/test_speed.c

@@ -347,6 +347,19 @@ test_speed_poly1305_ctmul(void)
 	test_speed_poly1305_inner("Poly1305 (ctmul)", &br_poly1305_ctmul_run);
 }
 
+static void
+test_speed_poly1305_ctmul32(void)
+{
+	test_speed_poly1305_inner("Poly1305 (ctmul32)",
+		&br_poly1305_ctmul32_run);
+}
+
+static void
+test_speed_poly1305_i15(void)
+{
+	test_speed_poly1305_inner("Poly1305 (i15)", &br_poly1305_i15_run);
+}
+
 static const unsigned char RSA_N[] = {
 	0xE9, 0xF2, 0x4A, 0x2F, 0x96, 0xDF, 0x0A, 0x23,
 	0x01, 0x85, 0xF1, 0x2C, 0xB2, 0xA8, 0xEF, 0x23,
@@ -1166,6 +1179,8 @@ static const struct {
 	STU(ghash_ctmul64),
 
 	STU(poly1305_ctmul),
+	STU(poly1305_ctmul32),
+	STU(poly1305_i15),
 
 	STU(rsa_i15),
 	STU(rsa_i31),