Le 24/11/2022 à 11:13, Naveen N. Rao a écrit : > Christophe Leroy wrote: >> ldimm64 is not only used for loading function addresses, and > > That's probably true today, but I worry that that can change upstream > and we may not notice at all. Not sure what you mean. Today POWERPC considers that ldimm64 is _always_ loading a function address whereas upstream BPF considers that ldimm64 is a function only when it is flagged BPF_PSEUDO_FUNC. In what direction could that change in the future ? For me if they change that it becomes an API change. Christophe > >> the NOPs added for padding are impacting performance, so avoid >> them when not necessary. >> >> On QEMU mac99, with the patch: >> >> test_bpf: #829 ALU64_MOV_K: all immediate value magnitudes jited:1 >> 167436810 PASS >> test_bpf: #831 ALU64_OR_K: all immediate value magnitudes jited:1 >> 170702940 PASS >> >> Without the patch: >> >> test_bpf: #829 ALU64_MOV_K: all immediate value magnitudes jited:1 >> 173012360 PASS >> test_bpf: #831 ALU64_OR_K: all immediate value magnitudes jited:1 >> 176424090 PASS >> >> That's a 3.5% performance improvement. > > A better approach would be to do a full JIT during the extra pass. > That's what most other architectures do today. And, as long as we can > ensure that the JIT'ed program size can never increase during the extra > pass, we should be ok to do a single extra pass. > > > - Naveen
