[slowriot]

Motivation

Emscripten currently compiles all its allocators as C. This makes no difference for dlmalloc and emmalloc, as they have no C++ feature detection and don't supply their own new/delete, but mimalloc does. As a result, building it as C causes abort on bad allocation instead of throwing std::bad_alloc, even when exceptions are enabled and -sABORTING_MALLOC=0.

This change makes sure that mimalloc is built as C++, only for C++ programs with exceptions enabled.

Changes

• tools/system_libs.py: Make libmimalloc inherit from ExceptionLibrary, giving it separately cached variants for disabled exceptions, Emscripten/JavaScript exceptions, legacy Wasm exceptions and standardized Wasm exceptions. Compile mimalloc's sources with -x c++ -std=c++17 for exception-enabled C++ links. All mimalloc translation units use the same language to keep internal linkage consistent; the underlying emmalloc.c and sbrk.c remain C. C-only programs explicitly select the C-built, no-exception variant, even when Wasm exception support is enabled. C++ programs without exceptions also retain the C build.
• embuilder.py: Add the single-threaded and multithreaded no-exception variants, plus the legacy and standardized Wasm exception variants, to the minimal prebuild task list.
• alloc-override.c: Allow Clang through the existing C++ aligned-allocation guard. Clang defines __GNUC__ as 4 for compatibility, so the previous GCC-version check excluded it despite support for aligned allocation. The guard still requires C++17 and __cpp_aligned_new. This preserves mimalloc's aligned new/delete implementations when switching to C++ compilation.
• prim/emscripten/prim.c: Use mi_decl_externc for the declarations of emmalloc_free and emmalloc_memalign. This gives them extern "C" linkage in the C++ build, matching the underlying C implementation. Their behaviour is unchanged.
• test/core/test_new_oom.cpp and .out: Add regression coverage for scalar, array and aligned allocation: zero-sized allocation, std::bad_alloc on failure, null returns from nothrow allocation, retrying a returning new_handler, propagating or catching handler exceptions as appropriate, and successful allocation after handling a failure. The test deliberately exceeds MAXIMUM_MEMORY without exhausting the memory needed to construct the exception.
• test/test_core.py: Run the new test with dlmalloc, emmalloc and mimalloc across all three exception-handling modes. Add mimalloc cases to the existing aborting-new and nothrow-new tests, with and without memory growth.
• test/test_other.py: Add single-threaded and pthread C allocation tests with Wasm exception support enabled, checking that C-only links do not acquire a C++ runtime dependency.

[sbc100]

I wonder if we should make this the default in ExceptionLibrary?

[slowriot]

That sounds sensible. Would you like me to include that into this change or raise a separate PR?

[sbc100]

Lets try to land that change first / separately.

[kripken]

This looks like a change to upstream mimalloc (not the emscripten-specific parts). Has it been discussed upstream?

[kleisauke]

The C path also uses || defined(__clang__), so this looks like a good candidate for upstreaming. /cc @daanx

emscripten/system/lib/mimalloc/src/alloc-override.c

Line 275 in 4539e75

#elif (defined(__GNUC__) || defined(__clang__))

Note that you could also fix this by compiling with -fgnuc-version=5.0.0, though this is probably safer.

$ emcc -dM -E - < /dev/null | grep -e GNUC
#define __GNUC_MINOR__ 2
#define __GNUC_PATCHLEVEL__ 1
#define __GNUC_STDC_INLINE__ 1
#define __GNUC__ 4
$ emcc -fgnuc-version=5.0.0 -dM -E - < /dev/null | grep -e GNUC
#define __GNUC_MINOR__ 0
#define __GNUC_PATCHLEVEL__ 0
#define __GNUC_STDC_INLINE__ 1
#define __GNUC__ 5

(Clang claims compatibility with GCC 4.2.1 by default)