iovisor / ubpf / 21450218157

78%

Build Type

push

github

Committed by web-flow

Commit Message

Add ARM64 JIT atomic instruction support (#727)

* Add ARM64 JIT atomic instruction support with LDXR/STXR

Co-authored-by: Alan-Jowett <20480683+Alan-Jowett@users.noreply.github.com>

* Fix atomic operation logic to properly handle all ALU operations

Co-authored-by: Alan-Jowett <20480683+Alan-Jowett@users.noreply.github.com>

* Fix atomic operation bugs: register corruption, CMPXCHG logic, FETCH semantics

Co-authored-by: Alan-Jowett <20480683+Alan-Jowett@users.noreply.github.com>

* Fix register aliasing and negative offset handling in atomic operations

Co-authored-by: Alan-Jowett <20480683+Alan-Jowett@users.noreply.github.com>

* fix: ARM64 atomic operation register aliasing for FETCH operations

Two issues fixed:

1. Register aliasing between op_result_reg and status_reg for FETCH:
   - For FETCH ops, op_result_reg was R25 which conflicts with status_reg
   - STXR writes status to status_reg, clobbering the computed result
   - Fix: For FETCH ops only, use R8 (caller-saved, not used elsewhere)
   - Non-FETCH retains original behavior (R25) which works correctly
     since the store happens before the status clobber

2. Undefined behavior when negating INT16_MIN:
   - (int16_t)-offset is UB when offset == INT16_MIN
   - Fix: Use int32_t for abs_offset and cast before negation

* fix: ARM64 atomic op_result_reg aliasing with status_reg

The previous code used R25 for op_result_reg when temp_reg was R24,
but R25 is also used as status_reg for STXR. This caused the STXR
instruction to have the same register for both the value to store
and the status output, which is undefined behavior on ARM64.

Fixed by always using R8 (caller-saved, unused in this context)
for the operation result register.

* test: enable lock tests on arm64/macos

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: Alan-Jowett <20480683+Alan-Jowett@users.noreply.github.com>
Co-authored-by: Alan Jowett <alan.jowett@microsoft.com>

Run Details

128 of 140 new or added lines in 1 file covered. (91.43%)

6 existing lines in 2 files now uncovered.

6332 of 8113 relevant lines covered (78.05%)

551568.38 hits per line