umash.c: add support for aarch64 + crypto extensions

The only vector operations we use (all on 128 bits) are:

1. creating a vector from 64-bit values
2. shifting 64-bit lanes by one bit to the left
3. carry-less multiplication of two 64-bit scalars into 128-bit vector
4. carry-less multiplication of the low and high halves of a vector
5. xoring vectors together

That's easy to convert to neon with the crypto extension for
`vmull_p64` (and aarch64 for 64x64 -> 128 multiplications).

Author: pkhuong

.travis.yml

@@ -4,13 +4,20 @@ compiler:
   - gcc
 arch:
   - amd64
+  - arm64
 os: linux
 dist: bionic
 
 install:
   - pyenv shell 3.8 || echo "using default python"
   - sudo apt-get -y install python3-venv
   - pip3 install --upgrade pip
+
 script:
-  - export CFLAGS='-g -O2 -std=c99 -W -Wall -Werror -mpclmul';
+  - if [ "x$TRAVIS_CPU_ARCH" == "xarm64" ];
+    then
+      export CFLAGS='-g -O2 -std=c99 -W -Wall -Werror -march=armv8-a+crypto';
+    else
+      export CFLAGS='-g -O2 -std=c99 -W -Wall -Werror -mpclmul';
+    fi
   - t/run-tests-public.sh && t/run-tests.sh

umash.c

@@ -42,8 +42,43 @@ v128_clmul_cross(v128 x)
 {
 	return _mm_clmulepi64_si128(x, x, 1);
 }
+
+#elif defined(__ARM_FEATURE_CRYPTO)
+
+#include <arm_neon.h>
+
+typedef uint64x2_t v128;
+
+#define V128_ZERO { 0 };
+
+static inline v128
+v128_create(uint64_t lo, uint64_t hi)
+{
+	return vcombine_u64(vcreate_u64(lo), vcreate_u64(hi));
+}
+
+static inline v128
+v128_shift(v128 x)
+{
+	return vshlq_n_u64(x, 1);
+}
+
+static inline v128
+v128_clmul(uint64_t x, uint64_t y)
+{
+	return vreinterpretq_u64_p128(vmull_p64(x, y));
+}
+
+static inline v128
+v128_clmul_cross(v128 x)
+{
+	return v128_clmul(vgetq_lane_u64(x, 0), vgetq_lane_u64(x, 1));
+}
+
 #else
-#error "Unsupported platform: umash requires x86's SSE2 and CLMUL (-mpclmul)"
+
+#error                                                                  \
+    "Unsupported platform: umash requires CLMUL (-mpclmul) on x86-64, or crypto (-march=...+crypto) extensions on aarch64."
 #endif
 
 /*