fixed bug with aligned alloc for macos

include/simsycl/detail/allocation.hh

@@ -15,9 +15,16 @@ namespace simsycl::detail {
 // NOTE: returned pointers must be freed with aligned_free
 inline void *aligned_alloc(size_t alignment, size_t size) {
 #if defined(_MSC_VER)
+    // MSVC does not have std::aligned_alloc because the pointers it returns cannot be freed with std::free
     return _aligned_malloc(size, alignment);
 #else
-    return std::aligned_alloc(alignment, size);
+    // POSIX and notably macOS requires a minimum alignment of sizeof(void*) for aligned_alloc, so we only use that for
+    // over-aligned allocations
+    if(alignment <= alignof(std::max_align_t)) {
+        return std::malloc(size);
+    } else {
+        return std::aligned_alloc(alignment, size);
+    }
 #endif
 }
 

src/simsycl/system.cc

@@ -182,19 +182,7 @@ void *usm_alloc(const sycl::context &context, sycl::usm::alloc kind, std::option
         if(*bytes_free < size_bytes) return nullptr;
     }
 
-    void *ptr;
-#if defined(_MSC_VER)
-    // MSVC does not have std::aligned_alloc because the pointers it returns cannot be freed with std::free
-    ptr = _aligned_malloc(size_bytes, alignment_bytes);
-#else
-    // POSIX and notably macOS requires a minimum alignment of sizeof(void*) for aligned_alloc, so we only use that for
-    // over-aligned allocations
-    if(alignment_bytes <= alignof(std::max_align_t)) {
-        ptr = std::malloc(size_bytes);
-    } else {
-        ptr = std::aligned_alloc(alignment_bytes, size_bytes);
-    }
-#endif
+    void *ptr = aligned_alloc(alignment_bytes, size_bytes);
 
     if(ptr == nullptr) return nullptr;
     std::memset(ptr, static_cast<int>(uninitialized_memory_pattern), size_bytes);

test/CMakeLists.txt

@@ -20,7 +20,7 @@ add_executable(tests
   math_tests.cc
   reduction_tests.cc
   simulation_tests.cc
-  usm_tests.cc
+  alloc_tests.cc
   vec_tests.cc
 )
 

test/usm_tests.cc → test/alloc_tests.cc

@@ -3,6 +3,19 @@
 
 #include <catch2/catch_test_macros.hpp>
 
+TEST_CASE("allocates memory of any alignment", "[aligned_alloc]") {
+    const size_t largest_sycl_align_bytes = alignof(sycl::long16);
+    const size_t size_bytes = 4096;
+    CAPTURE(size_bytes);
+    for(size_t align_bytes = 1; align_bytes <= largest_sycl_align_bytes; align_bytes *= 2) {
+        CAPTURE(align_bytes);
+        errno = 0;
+        const auto p = simsycl::detail::aligned_alloc(align_bytes, size_bytes);
+        CHECK(p != nullptr);
+        CHECK(reinterpret_cast<std::uintptr_t>(p) % align_bytes == 0);
+        simsycl::detail::aligned_free(p);
+    }
+}
 
 TEST_CASE("usm_alloc allocates memory of any alignment", "[usm]") {
     const size_t largest_sycl_align_bytes = alignof(sycl::long16);