FEX-2107

Compatibility & Bug Fixes

• Fixes bugs in unaligned atomic signal handlers
• CPUID cleanups
• Fixed a memory leak in Register Allocation
• Several syscall fixes (pidfd_send_signal, arch_prctl, fcntl, send(m)msg, shmdt, recvmsg, *chown32, edge cases around time syscalls, 32/tmpfile)
• Implement more syscalls (setfsuid32, setfsgid32, getgroups32, setgroups32, settimeofday, futimesat, utimes, 32/sigpending, 32/truncate64)
• Deferred signal handler registration, fixes bash (pts handover)
• Fixed handling of some rare elf files
• Implemented 32 bit iret

Usability

• Adds support for squashfs based rootfses
• cpack support for debian packaging

Performance

• Huge page support in our VA allocator

Misc

• Reduce warnings
• Several assorted Cleanups
• Relocate ELF handling logic to the os frontend
• Improve logging SNR

FEX-2106

Compatibility & Bug Fixes

• Several syscall bug fixes (pselect, epool, semctl, msgctl, pidfd_getfd, poll, fadvise64, sigaction, mmap, epoll, uname, openat, ...)
• Implemented SSE4.1

Performance

• Multi threaded AOTGen
• Optimized ioctl/drm marshaling

Misc

• Several minor cleanups & refactors

FEX-2105

Compatibility & Bug Fixes

• New, simpler ELF Loader
• ioctl32 marshaling for several devices
• Handle more cases of an application pinging self
• binfmt_misc fixes to support AppImage
• Several system call fixes (eventfd, eventfd2, openat, creat, trucnate, clock_nanosleep, pselect6, sendmsg, recvmsg)
• Signal fixes
• Better handling of /proc/self/, /proc/pid-self/
• Implements support for CLFLUSH

Performance

• Faster AOTIR file loading
• AOTIR offline generation
• Switch to xxhash from fasthash
• IR structure optimizations
• Adds Long Divide removal pass
• Zero-cost asserts

Misc

• Fixes host thread stacks ending up in lower 32-bit VA
• Remove reliance on librt and libnuma
• Default hidden visibility and strip symbols
• FEX now uses jemalloc

FEX-2104

Compatibility & Bug Fixes

• Disables RCPC on ARM64 JIT
• Implements CPUID 0x8000'0005 for L1 cacheline information
• Implements CPUID 0x8000'0006 for cacheline information
• Implements MOVNTDQA
• Restrict imm code motion around selects to matching sizes, fixes dav1d
• Validate LOCK handling, add missing segment offsets
• Add atomic logic for SecondaryALUOp
• Adds support for locked NOT
• Adds support for locked ADC and SBB
• Adds a couple new 32bit syscalls
• Fixes an edge case of 32bit cmpxchg <reg>, <reg>
• Lock around FDToNameMap accesses (Fixes Geekbench 4 stability issues)
• Flush context around OP_SYSCALLs, Syscalls might read it
• SA_NOCLDSTOP only blocks CLD_CONTINUED/STOPPED/TRAPPED
• Add missing break for UD2 in INTOp
• Init on X87FNSAVE, fix FNINIT
• Switches FEXCore over to pthreads implementation

Usability

• Support for global application profiles
• Thunks can be configured with json as an overlay
• FEXConfig improvements
• Adds support for Named RootFS folders in FEXConfig
• Cleanup threads when they exit
• Default to no logging
• Allows installing of FEXThunks in our data directory

Documentation

• New Readme.md & auto generated SourceOutline.md to help newcomers to the codebase
• Man Pages

Internal restructuring

• Unify all four dispatchers
• Separate thread and state
• Allow both ARM64 and X86_64 jits to be compiled at the same time

FEX-2103

Compatibility

• Support for unaligned atomic memory ops
• Thread local mman cache invalidation
• glib 2.32 support (cpuid fixes, faccessat2)
• cmpxchg flag fixes
• fixed cvtt* ops to actually truncate, both for x87 and sse
• Implemented BTC, BTR and BTS atomic variants
• Workaround crashes around exit_group for graceful exit
• Fixed SAR8 & SAR16 sign extension
• Added support for x87 rounding modes
• Added support for x87 precision modes
• Temporary fix for x87's frem
• Fixed select system calls to update the descriptors
• Fixed handling of invalid ops when multiblock is enabled
• Fixed an edge case bug in CreateElementPair
• Fixed BRK handling
• Fixed /proc/cpuinfo topology
• Fixed compilation of superblocks with more than 256 spills
• Fixed a bug in in the MUL -> SHL optimization

Performance

• Added ir-cache that halves loading times when enabled (--aotir-capture && --aotir-load)
• x87 ops no longer force interpreter for the entire superblock

Usability

• uname now returns the host name
• Added checks for python in the CMake files

Misc

• ThunkLibs now generate asm wrappers, so they can be compiled with older GCCs
• Asserts on unsupported atomic ops
• Removed erratic asserts in RA

FEX-2102

Compatibility

• Unaligned cmpxchg & cmpxchg8b on ARMv8.1+
• Mutliblock now gracefully handles unsupported / invalid instructions. This makes it safe to always enable multi block
• Several fixes for 32-bit binaries: branch handling, signal return, memory allocation, ioctl32 for x86-64 host
• Improved BRK handling

Perfomance

• Much reduced stuttering during JIT compilation. The JIT optimizer is now over 3x as fast.
• Per-thread IR Caching with cached RA, for faster recovery from code cache resets

Usability

• Defaults now to jit, multiblock, 5000 instructions.
• CPUID returns fex version

Misc

• Cleaned up IR printing, extended asm tests to test IR dumping & IR printing
• Integrated gcc target tests to our CI
• Removed many warnings

The detailed change log is available here