diff --git a/Include/Hallocy/Core/Memory.h b/Include/Hallocy/Core/Memory.h index b001d22..29e78d3 100644 --- a/Include/Hallocy/Core/Memory.h +++ b/Include/Hallocy/Core/Memory.h @@ -31,5 +31,6 @@ HallocyError hallocy_set_memory(void *destination, int value, const size_t size); HallocyError hallocy_copy_memory(void *destination, void *source, const size_t size); HallocyError hallocy_move_memory(void *destination, void *source, const size_t size); +bool hallocy_compare_memory(void *left_side, void *right_side, const size_t size); #endif \ No newline at end of file diff --git a/Src/Core/Memory.c b/Src/Core/Memory.c index fd2e7e6..3fc4417 100644 --- a/Src/Core/Memory.c +++ b/Src/Core/Memory.c @@ -172,7 +172,7 @@ HallocyError hallocy_copy_memory(void *destination, void *source, const size_t s __m512i simd_value; while (destination_bytes - end_address >= 64) { simd_value = _mm512_load_si512((__m512i*)source_bytes); - _mm512_store_si512((__m512i*)destination, simd_value); + _mm512_store_si512((__m512i*)destination_bytes, simd_value); destination_bytes += 64; source_bytes += 64; @@ -181,7 +181,7 @@ HallocyError hallocy_copy_memory(void *destination, void *source, const size_t s __m512i simd_value; while (destination_bytes - end_address >= 64) { simd_value = _mm512_loadu_si512((__m512i*)source_bytes); - _mm512_storeu_si512((__m512i*)destination, simd_value); + _mm512_storeu_si512((__m512i*)destination_bytes, simd_value); destination_bytes += 64; source_bytes += 64; @@ -202,7 +202,7 @@ HallocyError hallocy_copy_memory(void *destination, void *source, const size_t s __m256i simd_value; while (destination_bytes - end_address >= 32) { simd_value = _mm256_load_si256((__m256i*)source_bytes); - _mm256_store_si256((__m256i*)destination, simd_value); + _mm256_store_si256((__m256i*)destination_bytes, simd_value); destination_bytes += 32; source_bytes += 32; @@ -211,7 +211,7 @@ HallocyError hallocy_copy_memory(void *destination, void *source, const size_t s __m256i simd_value; while (destination_bytes - end_address >= 32) { simd_value = _mm256_loadu_si256((__m256i*)source_bytes); - _mm256_storeu_si256((__m256i*)destination, simd_value); + _mm256_storeu_si256((__m256i*)destination_bytes, simd_value); destination_bytes += 32; source_bytes += 32; @@ -232,7 +232,7 @@ HallocyError hallocy_copy_memory(void *destination, void *source, const size_t s __m128i simd_value; while (destination_bytes - end_address >= 16) { simd_value = _mm_load_si128((__m128i*)source_bytes); - _mm_store_si128((__m128i*)destination, simd_value); + _mm_store_si128((__m128i*)destination_bytes, simd_value); destination_bytes += 16; source_bytes += 16; @@ -241,7 +241,7 @@ HallocyError hallocy_copy_memory(void *destination, void *source, const size_t s __m128i simd_value; while (destination_bytes - end_address >= 16) { simd_value = _mm_loadu_si128((__m128i*)source_bytes); - _mm_storeu_si128((__m128i*)destination, simd_value); + _mm_storeu_si128((__m128i*)destination_bytes, simd_value); destination_bytes += 16; source_bytes += 16; @@ -334,7 +334,7 @@ HallocyError hallocy_move_memory(void *destination, void *source, const size_t s source_bytes -= 64; simd_value = _mm512_load_si512((__m512i*)source_bytes); - _mm512_store_si512((__m512i*)destination, simd_value); + _mm512_store_si512((__m512i*)destination_bytes, simd_value); } } else { __m512i simd_value; @@ -343,7 +343,7 @@ HallocyError hallocy_move_memory(void *destination, void *source, const size_t s source_bytes -= 64; simd_value = _mm512_loadu_si512((__m512i*)source_bytes); - _mm512_storeu_si512((__m512i*)destination, simd_value); + _mm512_storeu_si512((__m512i*)destination_bytes, simd_value); } } break; @@ -365,7 +365,7 @@ HallocyError hallocy_move_memory(void *destination, void *source, const size_t s source_bytes -= 32; simd_value = _mm256_load_si256((__m256i*)source_bytes); - _mm256_store_si256((__m256i*)destination, simd_value); + _mm256_store_si256((__m256i*)destination_bytes, simd_value); } } else { __m256i simd_value; @@ -374,7 +374,7 @@ HallocyError hallocy_move_memory(void *destination, void *source, const size_t s source_bytes -= 32; simd_value = _mm256_loadu_si256((__m256i*)source_bytes); - _mm256_storeu_si256((__m256i*)destination, simd_value); + _mm256_storeu_si256((__m256i*)destination_bytes, simd_value); } } break; @@ -395,7 +395,7 @@ HallocyError hallocy_move_memory(void *destination, void *source, const size_t s source_bytes -= 16; simd_value = _mm_load_si128((__m128i*)source_bytes); - _mm_store_si128((__m128i*)destination, simd_value); + _mm_store_si128((__m128i*)destination_bytes, simd_value); } } else { __m128i simd_value; @@ -404,7 +404,7 @@ HallocyError hallocy_move_memory(void *destination, void *source, const size_t s source_bytes -= 16; simd_value = _mm_loadu_si128((__m128i*)source_bytes); - _mm_storeu_si128((__m128i*)destination, simd_value); + _mm_storeu_si128((__m128i*)destination_bytes, simd_value); } } break; @@ -442,4 +442,205 @@ HallocyError hallocy_move_memory(void *destination, void *source, const size_t s } return HALLOCY_ERROR_NONE; +} + +bool hallocy_compare_memory(void *left_side, void *right_side, const size_t size) { + if (left_side == NULL || right_side == NULL) { + return false; + } + + unsigned char *left_side_bytes = (unsigned char*)left_side; + unsigned char *right_side_bytes = (unsigned char*)right_side; + + unsigned char *end_address = left_side_bytes + size; + + switch (hallocy_is_simd_supported()) { + #if defined(_M_ARM64) || defined(__aarch64__) || defined(__arm__) + case HALLOCY_SIMD_NEON: { + if ((size_t)left_side_bytes % 16 == (size_t)right_side_bytes % 16) { + while (((size_t)left_side_bytes % 16) != 0 && left_side_bytes != end_address) { + if (*left_side_bytes != *right_side_bytes) { + return false; + } + + left_side_bytes += 1; + right_side_bytes += 1; + } + } + + while (end_address - destination_bytes >= 16) { + uint8x16_t simd_left_side_value = vdupq_n_u8(left_side_bytes); + uint8x16_t simd_right_side_value = vdupq_n_u8(right_side_bytes); + + uint8x16_t result = vceqq_u8(simd_left_side_value, simd_right_side_value); + if (vmaxvq_u8(result) != 0xFF) { + return false; + } + + left_side_bytes += 16; + right_side_bytes += 16 + } + break; + } + #else + case HALLOCY_SIMD_AVX512: { + if ((size_t)left_side_bytes % 64 == (size_t)right_side_bytes % 64) { + while (((size_t)left_side_bytes % 64) != 0 && left_side_bytes != end_address) { + if (*left_side_bytes != *right_side_bytes) { + return false; + } + + left_side_bytes += 1; + right_side_bytes += 1; + } + + while (left_side_bytes - end_address >= 64) { + __m512i simd_left_side_value = _mm512_load_si512((__m512i*)left_side_bytes); + __m512i simd_right_side_value = _mm512_load_si512((__m512i*)right_side_bytes); + + __m512i result = _mm512_xor_si512(simd_left_side_value, simd_right_side_value); + if (_mm512_test_epi64_mask(result, result) != 0) { + return false; + } + + left_side_bytes += 64; + right_side_bytes += 64; + } + } else { + while (left_side_bytes - end_address >= 64) { + __m512i simd_left_side_value = _mm512_loadu_si512((__m512i*)left_side_bytes); + __m512i simd_right_side_value = _mm512_loadu_si512((__m512i*)right_side_bytes); + + __m512i result = _mm512_xor_si512(simd_left_side_value, simd_right_side_value); + if (_mm512_test_epi64_mask(result, result) != 0) { + return false; + } + + left_side_bytes += 64; + right_side_bytes += 64; + } + } + break; + } + + case HALLOCY_SIMD_AVX2: + case HALLOCY_SIMD_AVX: { + if ((size_t)left_side_bytes % 32 == (size_t)right_side_bytes % 32) { + while (((size_t)left_side_bytes % 32) != 0 && left_side_bytes != end_address) { + if (*left_side_bytes != *right_side_bytes) { + return false; + } + + left_side_bytes += 1; + right_side_bytes += 1; + } + + while (left_side_bytes - end_address >= 32) { + __m256i simd_left_side_value = _mm256_load_si256((__m256i*)left_side_bytes); + __m256i simd_right_side_value = _mm256_load_si256((__m256i*)right_side_bytes); + + __m256i result = _mm256_xor_si256(simd_left_side_value, simd_right_side_value); + if (_mm256_testz_si256(result, result) == 0) { + return false; + } + + left_side_bytes += 32; + right_side_bytes += 32; + } + } else { + while (left_side_bytes - end_address >= 32) { + __m256i simd_left_side_value = _mm256_loadu_si256((__m256i*)left_side_bytes); + __m256i simd_right_side_value = _mm256_loadu_si256((__m256i*)right_side_bytes); + + __m256i result = _mm256_xor_si256(simd_left_side_value, simd_right_side_value); + if (_mm256_testz_si256(result, result) == 0) { + return false; + } + + left_side_bytes += 32; + right_side_bytes += 32; + } + } + break; + } + + case HALLOCY_SIMD_SSE2: + case HALLOCY_SIMD_SSE: { + if ((size_t)left_side_bytes % 16 == (size_t)right_side_bytes % 16) { + while (((size_t)left_side_bytes % 16) != 0 && left_side_bytes != end_address) { + if (*left_side_bytes != *right_side_bytes) { + return false; + } + + left_side_bytes += 1; + right_side_bytes += 1; + } + + while (left_side_bytes - end_address >= 16) { + __m128i simd_left_side_value = _mm_load_si128((__m128i*)left_side_bytes); + __m128i simd_right_side_value = _mm_load_si128((__m128i*)right_side_bytes); + + __m128i result = _mm_xor_si128(simd_left_side_value, simd_right_side_value); + if (_mm_testz_si128(result, result) == 0) { + return false; + } + + left_side_bytes += 16; + right_side_bytes += 16; + } + } else { + while (left_side_bytes - end_address >= 16) { + __m128i simd_left_side_value = _mm_loadu_si128((__m128i*)left_side_bytes); + __m128i simd_right_side_value = _mm_loadu_si128((__m128i*)right_side_bytes); + + __m128i result = _mm_xor_si128(simd_left_side_value, simd_right_side_value); + if (_mm_testz_si128(result, result) == 0) { + return false; + } + + left_side_bytes += 16; + right_side_bytes += 16; + } + } + break; + } + #endif + default: { + size_t word_size = sizeof(size_t); + if ((size_t)left_side_bytes % word_size == (size_t)right_side_bytes % word_size) { + while (((size_t)left_side_bytes % word_size) != 0 && left_side_bytes != end_address) { + *left_side_bytes = *right_side_bytes; + left_side_bytes += 1; + right_side_bytes += 1; + } + + size_t *left_side_word = (size_t*)left_side_bytes; + size_t *right_side_word = (size_t*)right_side_bytes; + while ((unsigned char*)(left_side_word + 1) < end_address) { + if (*left_side_word != *right_side_word) { + return false; + } + + left_side_word += 1; + right_side_word += 1; + } + + right_side_bytes = (unsigned char*)right_side_word; + left_side_bytes = (unsigned char*)left_side_word; + } + + break; + } + } + + while (left_side_bytes != end_address) { + if (*left_side_bytes != *right_side_bytes) { + return false; + } + + left_side_bytes += 1; + right_side_bytes += 1; + } + + return true; } \ No newline at end of file diff --git a/Tests/Main.c b/Tests/Main.c index 09645e3..934c082 100644 --- a/Tests/Main.c +++ b/Tests/Main.c @@ -59,6 +59,19 @@ hallocy_move_memory(text_copy + 5, text_copy + 10, 15); printf("%s\n", text_copy); + if (!hallocy_compare_memory(text, text_copy, 36)) { + printf("text and text_copy are not equal\n"); + } + + char *text_copy_copy = (char*)hallocy_malloc(36); + hallocy_copy_memory(text_copy_copy, text_copy, 36); + + if (hallocy_compare_memory(text_copy_copy, text_copy, 36)) { + printf("text_copy_copy and text_copy are equal\n"); + } + + hallocy_free(text_copy_copy); + hallocy_free(text_copy); hallocy_free(text); return 0;