diff options
author | Jakub Kicinski <jakub.kicinski@netronome.com> | 2018-04-24 21:22:37 -0700 |
---|---|---|
committer | Daniel Borkmann <daniel@iogearbox.net> | 2018-04-25 09:56:10 +0200 |
commit | 6c59500c2dbfae0e3c90854ef443948d2889495d (patch) | |
tree | 0ad0b3cf7ace13bc02338e6408b4bd45013f5b86 /drivers/net/ethernet/netronome | |
parent | 9c9e53233c281ed6c00425b165c4dd7fd0881cd5 (diff) | |
download | blackbird-obmc-linux-6c59500c2dbfae0e3c90854ef443948d2889495d.tar.gz blackbird-obmc-linux-6c59500c2dbfae0e3c90854ef443948d2889495d.zip |
nfp: bpf: optimize add/sub of a negative constant
NFP instruction set can fit small immediates into the instruction.
Negative integers, however, will never fit because they will have
highest bit set. If we swap the ALU op between ADD and SUB and
negate the constant we have a better chance of fitting small negative
integers into the instruction itself and saving one or two cycles.
immed[gprB_21, 0xfffffffc]
alu[gprA_4, gprA_4, +, gprB_21], gpr_wrboth
immed[gprB_21, 0xffffffff]
alu[gprA_5, gprA_5, +carry, gprB_21], gpr_wrboth
now becomes:
alu[gprA_4, gprA_4, -, 4], gpr_wrboth
alu[gprA_5, gprA_5, -carry, 0], gpr_wrboth
Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Diffstat (limited to 'drivers/net/ethernet/netronome')
-rw-r--r-- | drivers/net/ethernet/netronome/nfp/bpf/jit.c | 35 |
1 files changed, 35 insertions, 0 deletions
diff --git a/drivers/net/ethernet/netronome/nfp/bpf/jit.c b/drivers/net/ethernet/netronome/nfp/bpf/jit.c index 9cc638718272..a5590988fc69 100644 --- a/drivers/net/ethernet/netronome/nfp/bpf/jit.c +++ b/drivers/net/ethernet/netronome/nfp/bpf/jit.c @@ -2777,6 +2777,40 @@ static void nfp_bpf_opt_reg_init(struct nfp_prog *nfp_prog) } } +/* abs(insn.imm) will fit better into unrestricted reg immediate - + * convert add/sub of a negative number into a sub/add of a positive one. + */ +static void nfp_bpf_opt_neg_add_sub(struct nfp_prog *nfp_prog) +{ + struct nfp_insn_meta *meta; + + list_for_each_entry(meta, &nfp_prog->insns, l) { + struct bpf_insn insn = meta->insn; + + if (meta->skip) + continue; + + if (BPF_CLASS(insn.code) != BPF_ALU && + BPF_CLASS(insn.code) != BPF_ALU64) + continue; + if (BPF_SRC(insn.code) != BPF_K) + continue; + if (insn.imm >= 0) + continue; + + if (BPF_OP(insn.code) == BPF_ADD) + insn.code = BPF_CLASS(insn.code) | BPF_SUB; + else if (BPF_OP(insn.code) == BPF_SUB) + insn.code = BPF_CLASS(insn.code) | BPF_ADD; + else + continue; + + meta->insn.code = insn.code | BPF_K; + + meta->insn.imm = -insn.imm; + } +} + /* Remove masking after load since our load guarantees this is not needed */ static void nfp_bpf_opt_ld_mask(struct nfp_prog *nfp_prog) { @@ -3212,6 +3246,7 @@ static int nfp_bpf_optimize(struct nfp_prog *nfp_prog) { nfp_bpf_opt_reg_init(nfp_prog); + nfp_bpf_opt_neg_add_sub(nfp_prog); nfp_bpf_opt_ld_mask(nfp_prog); nfp_bpf_opt_ld_shift(nfp_prog); nfp_bpf_opt_ldst_gather(nfp_prog); |