The scourge of x86 emulation

A deep technical dive into why emulating x86's strict Total Store Ordering (TSO) memory model on ARM's weak memory model is so hard, covering how FEX handles the problem through ARMv8.0's acquire/release instructions, the LRCPC extension family (LRCPC/LRCPC2/LRCPC3), alignment faults and patchpoints, split-lock emulation costs, and uncached/write-combine memory performance cliffs affecting PCIe GPU setups. Benchmarks compare AmpereOne, Cortex-X4, Cortex-X925, Qualcomm Oryon-3, and Apple M1, showing Apple's hardware TSO toggle gives near-native performance while other platforms suffer major penalties, especially on split-locks and write-combine stores (up to 816x slower). The piece closes by proposing a CASP-based approach to solve split-lock emulation correctness and calls for future ARM extensions like a hypothetical FEAT_LRCPC4.

Source: https://fex-emu.com/Scourge-of-emulation. 8 Sync News only summarizes and links out; content copyright belongs to the authors and original sources.