pseh2_64 relied on inline-asm clobbers across a C/asm funclet jump, which is not a reliable ABI boundary.
On AMD64 gcc path, this can corrupt non-volatile state during exception filtering/unwind.
This change makes the trampoline ABI-safe by explicitly preserving and restoring non-volatile registers and keeping unwind metadata/prologue consistent.
Changes:
- Save/restore rbx, rdi, rsi, r12-r15 in `__seh2_global_filter_func`.
- Adjust stack allocation/unwind annotation to match the new prologue/epilogue.
- Keep filter return in eax and jump back through `__seh2_global_filter_func_exit`.
Note: This is exception-only path: extra stack usage is acceptable for correctness.
How to test: Build and run ReactOS with gcc 13/15; you should no longer see stack exhaustion.
PSEH has 2 scopes: an outer scope, that starts with _SEH3_TRY and ends with _SEH3_END, and an inner scope that goes around the try block only. PSEH tracks try-levels by using an enum, which increments at every new TRY scrope, and which needs to be accessible in the outer scope for all registration and cleanup to work. But it should only increment when nested within the inner try scope, not from the except scope. This is fixed by adding an additional _SEH3$_InnerTryLevel, which increments only within the try block, but resets to the parent-value outside of it.
Should fix CORE-20316.
On Windows the function that calls the language specific exception handler function registers it's own exception handler, so the top of the exception chain does not point to our own handler. We need to take that into account when unwinding the stack and removing each handler as we go.
Previously, when leave was used in the except or finally block of a nested try block, it would jump back to the start of the except/finally block, resulting in an endless loop. This is fixed by jumping back to a label at the beginning of the try block, which is only visible from within the try block itself and from there to the end of the SEH block.
Fixes seh0055 testcase of MS SEH tests.
This is used with Clang, which does have native SEH, but it's broken and can cause the compiler to crash. Previously some code was not handling this and instead used native SEH for clang. Fix this by always using _SEH2_* macros instead of relying on __USE_PSEH2__ to be defined.
Try to improve things
The change is needed, because the previous version emitted multiple ".seh_handlerdata" blocks and GAS isn't able to merge them into one, but will instead only include the first one. This is fixed by generating "asm defines" in the code and pass the line number to the "REACTOS seh" pragma, which is then used to create all handler data by referring to the predefined symbols, that include the line number.
Also the EstablisherFrame parameter passed to the filter/finally function is the original stack pointer, not the frame pointer! Take this into account by passing __builtin_frame_address(0) from the filter context to the global wrapper function, which includes the frame-offset, and use that to recalculate the frame-pointer from the passed in stack pointer.
Also, put include directory next to the library and use
target_include_directories(.. INTERFACE ..) to get this right.
This is because :
- Having includes & implementation in two different places buggers me
- This makes sure that there is no "if it compiles everything is fine" behaviour from anyone
because now even static libraries need it for GCC amd64 build
Also add __USE_PSEH2__ define for the non SEH-aware compilers out there and use it in a few headers
where we define macros involving __try
