Windows: modules other than gc share the process heap; attributes named once; Debug DEBUG_TYPE fix

[ASDAlexander77]

Summary

Three fixes found by running the Debug build locally on the cross-module (-shared, async, @linkname) tests. Those tests had 19 failures in Debug.

1. Windows: every module but gc allocates from the process heap

A block made in one module is often freed in another. Two examples:

• under -mm=rc and -mm=own, an object of a library's class is made in the library and destroyed by the program;
• a coroutine frame comes from the runtime and goes back through the program.

On Windows the C runtime does not give these modules one allocator:

• each module links its own copy of the static CRT;
• a debug CRT keeps its block list per copy, so a cross-module free asserts or corrupts the heap;
• with the prebuilt LLVM, malloc in tslang.exe and TypeScriptRuntime.dll is rpmalloc.

The fix:

• A new ProcessHeapPass runs on Windows (not wasm) when the model is not gc; gc already has GCPass.
  • It renames malloc, calloc, realloc, free, aligned_alloc and aligned_free to __tslang_heap_*. It changes both calls and address-of uses, such as a destructor slot.
  • The helpers take the same arguments as the C functions, so this is a rename, not a rewrite.
• include/TypeScript/ProcessHeap.h is header-only and wraps HeapAlloc, HeapReAlloc and HeapFree:
  • realloc(p, 0) frees the block and returns null, as UCRT does;
  • aligned_alloc fails for an alignment above MEMORY_ALLOCATION_ALIGNMENT rather than return under-aligned memory.
• lib/ProcessHeapExports.inc exports the helpers from three places, all from the same code:
  • the async runtime library that programs and libraries link;
  • TypeScriptRuntime.dll (listed in the .def);
  • the JIT, where a symbol table replaces the earlier ucrtbase lookup.
• MemRuntime.cpp (Alloc, AlignedAlloc, Free, AlignedFree) uses the process heap on Windows in every build.

The pass must mark the helpers as allocators. The first version renamed the functions and nothing else. After that, DefaultLib under rc double-freed. The renamed functions had lost what LLVM knows about malloc by name, so LLVM stopped deleting allocations whose only use is being freed, and that exposed double frees the deletion had been hiding. Each helper now gets the same attributes the C functions get:

• allockind;
• "alloc-family"="malloc";
• allocptr on the argument of free and realloc;
• memory effects.

2. Debug assert: "DictionaryAttr element names must be unique"

processFunctionAttributes could add the same attribute twice, for example dllname on a @linkname declaration used from a -shared module. Each attribute name is now added once.

• New test: test-declare-linkname-attribute-once.

3. Debug build: MLIRGenCast.cpp defines DEBUG_TYPE again after the RTTI helper

The RTTI helper undefines DEBUG_TYPE on its way out, which broke the Debug build. This is the same commit as in #411; whichever PR merges second will have nothing to apply for it.

Test results

• Windows Release ctest: 3254/3254
• Windows Debug, cross-module tests (shared|async|linkname): 352/352 (19 failures before)
• Linux (WSL): 3240/3240
• DefaultLib gc: 157/157
• DefaultLib rc: 52 failures of 314 runs, the same with or without this change. They are flaky Set/Map/string/regex tests (heap corruption, 0xC0000409, busy loops). They are not caused by this PR and will be looked at separately.

🤖 Generated with Claude Code