Hi! This series brings stack support for offload.
We use the LMEM (Local memory) register file as memory to store the stack. Since this is a register file we need to do appropriate shifts on unaligned accesses. Verifier's state tracking helps us with that. LMEM can't be accessed directly, so we add support for setting pointer registers through which one can read/write LMEM. This set does not support accessing the stack when the alignment is not known. This can be added later (most likely using the byte_align instructions). There is also a number of optimizations which have been left out: - in more complex non aligned accesses, double shift and rotation can save us a cycle. This, however, leads to code explosion since all access sizes have to be coded separately; - since setting LM pointers costs around 5 cycles, we should be tracking their values to make sure we don't move them when they're already set correctly for earlier access; - in case of 8 byte access aligned to 4 bytes and crossing 32 byte boundary but not crossing a 64 byte boundary we don't have to increment the pointer, but this seems like a pretty rare case to justify the added complexity. Jakub Kicinski (9): nfp: bpf: add helper for emitting nops nfp: bpf: refactor nfp_bpf_check_ptr() nfp: bpf: add stack write support nfp: bpf: add stack read support nfp: bpf: optimize the RMW for stack accesses nfp: bpf: allow stack accesses via modified stack registers nfp: bpf: support accessing the stack beyond 64 bytes nfp: bpf: support stack accesses via non-constant pointers nfp: bpf: optimize mov64 a little drivers/net/ethernet/netronome/nfp/bpf/jit.c | 357 +++++++++++++++++++++- drivers/net/ethernet/netronome/nfp/bpf/main.h | 8 + drivers/net/ethernet/netronome/nfp/bpf/offload.c | 8 + drivers/net/ethernet/netronome/nfp/bpf/verifier.c | 64 +++- drivers/net/ethernet/netronome/nfp/nfp_asm.h | 5 + 5 files changed, 420 insertions(+), 22 deletions(-) -- 2.14.1
