m + add tests for optimisations

Signed-off-by: Mattias Andrée <maandree@kth.se>

3 files changed, 72 insertions, 7 deletions

diff --git a/libar2_decode_params.c b/libar2_decode_params.c
index 86fae6c..273de55 100644
--- a/libar2_decode_params.c
+++ b/libar2_decode_params.c
@@ -8,7 +8,7 @@ decode_u32(const char *s, uint_least32_t *outp)
 	uint_least32_t digit;
 	size_t i;
 
-	if (s[0] == '0' || !isdigit(s[0])) {
+	if ((s[0] == '0' && isdigit(s[1])) || !isdigit(s[0])) {
 		errno = EINVAL;
 		return 0;
 	}
diff --git a/libar2_hash.c b/libar2_hash.c
index 80fb6d3..4dc2e82 100644
--- a/libar2_hash.c
+++ b/libar2_hash.c
@@ -64,6 +64,7 @@ blockxor_avx2(struct block *a_, const struct block *b_)
 
 }
 
+/* $covered{$ (removing from coverage test because my machine does not spport AVX512F) */
 LIBAR2_TARGET__("avx512f")
 static void
 blockxor_avx512f(struct block *a_, const struct block *b_)
@@ -74,6 +75,7 @@ blockxor_avx512f(struct block *a_, const struct block *b_)
 	for (i = 0; i < sizeof(*a_) / (512 / 8); i++)
 		a[i] = _mm512_xor_si512(a[i], b[i]);
 }
+/* $covered}$ */
 
 static void
 blockxor_vanilla(struct block *a, const struct block *b)
@@ -98,6 +100,7 @@ blockxor3_avx2(struct block *a_, const struct block *b_, const struct block *c_)
 
 }
 
+/* $covered{$ (removing from coverage test because my machine does not spport AVX512F) */
 LIBAR2_TARGET__("avx512f")
 static void
 blockxor3_avx512f(struct block *a_, const struct block *b_, const struct block *c_)
@@ -109,6 +112,7 @@ blockxor3_avx512f(struct block *a_, const struct block *b_, const struct block *
 	for (i = 0; i < sizeof(*a_) / (512 / 8); i++)
 		a[i] = _mm512_xor_si512(b[i], c[i]);
 }
+/* $covered}$ */
 
 static void
 blockxor3_vanilla(struct block *a, const struct block *b, const struct block *c)
@@ -132,6 +136,7 @@ blockcpy_avx2(struct block *a_, const struct block *b_)
 		a[i] = _mm256_load_si256(&b[i]);
 }
 
+/* $covered{$ (removing from coverage test because my machine does not spport AVX512F) */
 LIBAR2_TARGET__("avx512f")
 static void
 blockcpy_avx512f(struct block *a_, const struct block *b_)
@@ -142,6 +147,7 @@ blockcpy_avx512f(struct block *a_, const struct block *b_)
 	for (i = 0; i < sizeof(*a_) / (512 / 8); i++)
 		a[i] = _mm512_load_si512(&b[i]);
 }
+/* $covered}$ */
 
 static void
 blockcpy_vanilla(struct block *a, const struct block *b)
@@ -674,6 +680,7 @@ libar2_init(void)
 		if (!initialised) {
 #if 0
 			__builtin_cpu_init();
+			/* $covered{$ (we know that it works, but the test cannot enter every branch) */
 			if (__builtin_cpu_supports("avx512f"))
 				libar2_internal_use_avx512f__();
 			else if (__builtin_cpu_supports("avx2"))
@@ -681,10 +688,12 @@ libar2_init(void)
 			else if (__builtin_cpu_supports("sse2"))
 				libar2_internal_use_sse2__();
 			else
-			libar2_internal_use_generic__();
+				libar2_internal_use_generic__();
+			/* $covered}$ */
 #else
 			uint32_t x;
 			__asm__ volatile("cpuid" : "=b"(x) : "a"(7), "c"(0) : "edx");
+			/* $covered{$ (we know that it works, but the test cannot enter every branch) */
 			if (x & ((uint32_t)1 << 16)) {
 				libar2_internal_use_avx512f__();
 			} else if (x & ((uint32_t)1 << 5)) {
@@ -696,6 +705,7 @@ libar2_init(void)
 				else
 					libar2_internal_use_generic__();
 			}
+			/* $covered}$ */
 #endif
 			initialised = 1;
 		}
diff --git a/test.c b/test.c
index 3aec24c..1e33c05 100644
--- a/test.c
+++ b/test.c
@@ -4,6 +4,9 @@
 #ifndef MEASURE_TIME
 # define MEASURE_TIME 0
 #endif
+#ifndef MEASURE_TIME_ONLY
+# define MEASURE_TIME_ONLY MEASURE_TIME
+#endif
 
 #include <stdlib.h>
 #if MEASURE_TIME
@@ -70,11 +73,11 @@ allocate(size_t num, size_t size, size_t alignment, struct libar2_context *ctx)
 		alignment = sizeof(void *);
 	err = posix_memalign(&ptr, alignment, num * size);
 	if (err)
-		goto fail;
+		goto fail; /* $covered$ */
 #elif defined(_ISOC11_SOURCE)
 	ptr = aligned_alloc(alignment, num * size);
 	if (!ptr)
-		goto fail;
+		goto fail; /* $covered$ */
 #else
 # error No implementation for aligned memory allocation available
 #endif
@@ -994,6 +997,57 @@ check_libar2_hash(void)
 }
 
 
+#if defined(__x86_64__) && defined(LIBAR2_TARGET__) && defined(__GNUC__)
+static void
+run_check_libar2_hash_optimisations(void)
+{
+
+#define CHECK(PWD, HASH)\
+	check_hash(MEM(PWD), HASH, NULL, 0, NULL, 0, &ctx_st, __LINE__)
+
+	CHECK("password", "$argon2i$m=256,t=2,p=2$c29tZXNhbHQ$tsEVYKap1h6scGt5ovl9aLRGOqOth+AMB+KwHpDFZPs");
+	CHECK("", "$argon2ds$v=16$m=8,t=1,p=2$ICAgICAgICA$+6+yBnWbuV7mLs6rKMhvi+SLbkzb5CB6Jd2pSWuC/Kw");
+	CHECK("", "$argon2d$v=16$m=8,t=1,p=1$ICAgICAgICA$X54KZYxUSfMUihzebb70sKbheabHilo8gsUldrVU4IU");
+	CHECK("password", "$argon2i$v=19$m=256,t=2,p=2$c29tZXNhbHQ$T/XOJ2mh1/TIpJHfCdQan76Q5esCFVoT5MAeIM1Oq2E");
+
+#undef CHECK
+}
+#endif
+
+
+static void
+check_libar2_hash_optimisations(void)
+{
+#if defined(__x86_64__) && defined(LIBAR2_TARGET__) && defined(__GNUC__)
+
+	__builtin_cpu_init();
+
+	libar2_internal_use_generic__();
+	run_check_libar2_hash_optimisations();
+
+	libar2_internal_use_sse2__();
+	if (__builtin_cpu_supports("sse2"))
+		run_check_libar2_hash_optimisations(); /* $covered$ */
+
+	libar2_internal_use_avx2__();
+	if (__builtin_cpu_supports("avx2"))
+		run_check_libar2_hash_optimisations(); /* $covered$ */
+
+	libar2_internal_use_avx512f__();
+	if (__builtin_cpu_supports("avx512f"))
+		run_check_libar2_hash_optimisations(); /* $covered$ */
+	/* $covered{$ */
+	else if (__builtin_cpu_supports("avx2"))
+		libar2_internal_use_avx2__();
+	else if (__builtin_cpu_supports("sse2"))
+		libar2_internal_use_sse2__();
+	else
+		libar2_internal_use_generic__();
+	/* $covered}$ */
+#endif
+}
+
+
 #ifdef LIBAR2_WEAKLY_LINKED__
 
 void
@@ -1212,7 +1266,7 @@ check_failures(void)
 int
 main(void)
 {
-#if 1
+#if !MEASURE_TIME_ONLY
 	check_libar2_type_to_string();
 	check_libar2_string_to_type();
 	check_libar2_version_to_string();
@@ -1227,9 +1281,7 @@ main(void)
 # ifdef LIBAR2_WEAKLY_LINKED__
 	check_libar2_hash_buf_size();
 # endif
-#endif
 
-#if 1
 	check_failures();
 #endif
 
@@ -1259,5 +1311,8 @@ main(void)
 	}
 #endif
 
+#if !MEASURE_TIME_ONLY
+	check_libar2_hash_optimisations();
+#endif
 	return 0;
 }