From: Matthew Wilcox > Sent: 21 February 2023 13:51 ... > > For this particular case, documenting such an access. Though I don't > > think there are any architectures which may tear a one byte read/write > > and merging/refetching is not an issue for this. > > Wouldn't a compiler be within its rights to implement a one byte store as: > > load-word > modify-byte-in-word > store-word > > and if this is a lockless store to a word which has an adjacent byte also > being modified by another CPU, one of those CPUs can lose its store? > And WRITE_ONCE would prevent the compiler from implementing the store > in that way. Some alpha cpu couldn't do byte memory accesses - so always did 32bit read-modify-write. But Linux doesn't support those ones any more. On arm 16bit structure members can be accessed with 32bit instructions because the 16bit ones have a smaller offset. On x86 the bit operations might access the (possibly misaligned) 32bit word containing the required bit - but they are locked. ISTR a problem where gcc was using wider instructions and doing a RMW on an adjacent volatile field. I really can't remember the justification for not marking fields that have unlocked accesses 'volatile' instead of requiring all the accesses be done as explicit volatile ones. David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
