Implement stabilize_sig_stack

[Blackhex]

Implements stabilize_sig_stack  routine for AArch64 that is used both in setjmp and longjmp routines implementation. Please, refer to x64 implementation for comparison.

Validated by https://github.com/Windows-on-ARM-Experiments/mingw-woarm64-build/actions/runs/15304776330

[Copilot]

Pull Request Overview

Implements the stabilize_sig_stack routine on AArch64 to mirror the x64 behavior used by setjmp/longjmp.

• Adds AArch64 assembly prologue/epilogue and register saves/restores
• Implements an atomic lock loop, incy g counter management, and conditional signal handling
• Calls into call_signal_handler and handles cleanup before returning the TLS pointer

Comments suppressed due to low confidence (3)

winsup/cygwin/scripts/gendef:450

• This add instruction overwrites the TLS pointer in x0 instead of preparing the argument. It should likely be add x1, x10, x1 (or use a fresh register) to compute the address without clobbering x0.

add x0, x0, x1

winsup/cygwin/scripts/gendef:455

• This uses x1 instead of the offset in x11, so the address of incy g will be calculated incorrectly. It should read add x11, x10, x11.

add x11, x10, x1

winsup/cygwin/scripts/gendef:408

• There’s no visible test coverage for the new stabilize_sig_stack routine; consider adding a targeted test or integration scenario to validate lock acquisition, signal handling, and stack restoration.

.seh_proc stabilize_sig_stack

[Copilot]

Pull Request Overview

Implements the AArch64 version of stabilize_sig_stack to mirror existing x64 logic for use in setjmp/longjmp.

• Adds function prologue/epilogue and per-thread lock around signal processing
• Manages the incyg counter and invokes the C++-style signal handler in a loop
• Returns the TLS base pointer in x0 and releases locks appropriately

Comments suppressed due to low confidence (4)

winsup/cygwin/scripts/gendef:422

• [nitpick] Swapping with the same register for source and destination works but is subtle. For clarity and to avoid surprises, use separate registers for the value to store and the old value (e.g. mov w8, #1; swp w9, w8, [x11]).

swp	w9, w9, [x11]		// try to acquire the lock using atomic swap

winsup/cygwin/scripts/gendef:410

• [nitpick] This new AArch64 routine appears untested. Adding unit or integration tests for stabilize_sig_stack would help catch regressions and verify signal-handling behavior.

// prologue

winsup/cygwin/scripts/gendef:455

• The comment says "store incremented value back," but this is after decrementing the counter. It should read "store decremented value back."

str	w9, [x11]		// store incremented value back

winsup/cygwin/scripts/gendef:435

• The addressing mode ldr w9, [x10, =_cygtls.current_sig] is not valid AArch64 syntax. Use a separate literal load and offset addition (e.g. ldr x11, =_cygtls.current_sig; add x11, x10, x11; ldr w9, [x11]) to compute the correct address.

ldr	w9, [x10, =_cygtls.current_sig]	// load current value of current_sig

[Copilot]

Pull Request Overview

This PR adds the AArch64 implementation of the stabilize_sig_stack routine—used by both setjmp and longjmp—to manage TLS-based signal locking and dispatch.

• Introduces a new stabilize_sig_stack function in assembly for ARM64
• Implements a spinlock on stacklock, increments/decrements the incyg counter, and invokes the signal handler if needed
• Pushes/restores callee-saved registers and returns the TLS base in x0

Comments suppressed due to low confidence (2)

winsup/cygwin/scripts/gendef:414

• [nitpick] The comment could be clearer; consider rephrasing to "Save x13 (scratch) and x14 to maintain stack alignment" to concisely explain why both registers are pushed.

stp\tx13, x14, [sp, #-0x10]!\t\t// save x13 register used next and x14 just to keep stack orientation and alignment

winsup/cygwin/scripts/gendef:471

• The comment says "store incremented value back" but this path actually decrements the counter. Update it to "store decremented value back" for accuracy.

str\tw9, [x11]\t\t// store incremented value back

[eukarpov]

.seh commands are not needed here

[Blackhex]

OK, thanks. I believe that you know far more about this topic. Could you possibly share link to some documentation where I can learn why?

[eukarpov]

The only available documentation is https://learn.microsoft.com/en-us/cpp/build/arm64-exception-handling?view=msvc-170#unwind-codes
and binutils/gcc source code.

in short, for a single instruction like this, not having seh_endepilogue or duplicating seh commands is not changing anything. Both options are acceptable.

[Blackhex]

From this answer I'd deduce that the fact that it has no effect is an implementation detail of GCC/binutils so it makes sense to keep it for consistency with the prologue and accross the compilers (Regardled that Cygwin is build only with GCC, one never know who will copy&paste this code elsewhere).

[eukarpov]

Coming back to this thread, it is worth mentioning again that .seh_startepilogue is not needed here and unwinding should be fully covered by prolog. Current implementation emits duplicated unwinding codes.

[Copilot]

Pull Request Overview

This PR implements the stabilize_sig_stack routine for AArch64, mirroring the x64 implementation for improved support in setjmp/longjmp functionality.

• Introduces an assembly implementation for signal stack stabilization.
• Adds atomic operations for lock handling and incyg counter management in the new routine.

[Copilot]

Pull Request Overview

This PR implements the AArch64 version of the stabilize_sig_stack routine used by setjmp/longjmp, following the existing x64 implementation.

• Implements a new signal stack stabilization routine with atomic lock acquisition for thread local storage.
• Includes loop-based lock acquisition with yield and signal handler invocation.