diff options
author | Luke Nelson <lukenels@cs.washington.edu> | 2020-05-06 03:03:20 +0300 |
---|---|---|
committer | Daniel Borkmann <daniel@iogearbox.net> | 2020-05-06 10:48:15 +0300 |
commit | 073ca6a0369e09c586a103e665f2dd67f1c71444 (patch) | |
tree | f1ca6237c59417e3626b2e1c583fb525683dd52e /arch/riscv | |
parent | ca349a6a104e58479defdc08ce56472a48f7cb81 (diff) | |
download | linux-073ca6a0369e09c586a103e665f2dd67f1c71444.tar.xz |
bpf, riscv: Optimize BPF_JSET BPF_K using andi on RV64
This patch optimizes BPF_JSET BPF_K by using a RISC-V andi instruction
when the BPF immediate fits in 12 bits, instead of first loading the
immediate to a temporary register.
Examples of generated code with and without this optimization:
BPF_JMP_IMM(BPF_JSET, R1, 2, 1) without optimization:
20: li t1,2
24: and t1,a0,t1
28: bnez t1,0x30
BPF_JMP_IMM(BPF_JSET, R1, 2, 1) with optimization:
20: andi t1,a0,2
24: bnez t1,0x2c
BPF_JMP32_IMM(BPF_JSET, R1, 2, 1) without optimization:
20: li t1,2
24: mv t2,a0
28: slli t2,t2,0x20
2c: srli t2,t2,0x20
30: slli t1,t1,0x20
34: srli t1,t1,0x20
38: and t1,t2,t1
3c: bnez t1,0x44
BPF_JMP32_IMM(BPF_JSET, R1, 2, 1) with optimization:
20: andi t1,a0,2
24: bnez t1,0x2c
In these examples, because the upper 32 bits of the sign-extended
immediate are 0, BPF_JMP BPF_JSET and BPF_JMP32 BPF_JSET are equivalent
and therefore the JIT produces identical code for them.
Co-developed-by: Xi Wang <xi.wang@gmail.com>
Signed-off-by: Xi Wang <xi.wang@gmail.com>
Signed-off-by: Luke Nelson <luke.r.nels@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Björn Töpel <bjorn.topel@gmail.com>
Acked-by: Björn Töpel <bjorn.topel@gmail.com>
Link: https://lore.kernel.org/bpf/20200506000320.28965-5-luke.r.nels@gmail.com
Diffstat (limited to 'arch/riscv')
-rw-r--r-- | arch/riscv/net/bpf_jit_comp64.c | 27 |
1 files changed, 19 insertions, 8 deletions
diff --git a/arch/riscv/net/bpf_jit_comp64.c b/arch/riscv/net/bpf_jit_comp64.c index b07cef952019..6cfd164cbe88 100644 --- a/arch/riscv/net/bpf_jit_comp64.c +++ b/arch/riscv/net/bpf_jit_comp64.c @@ -792,8 +792,6 @@ out_be: case BPF_JMP32 | BPF_JSGE | BPF_K: case BPF_JMP | BPF_JSLE | BPF_K: case BPF_JMP32 | BPF_JSLE | BPF_K: - case BPF_JMP | BPF_JSET | BPF_K: - case BPF_JMP32 | BPF_JSET | BPF_K: rvoff = rv_offset(i, off, ctx); s = ctx->ninsns; if (imm) { @@ -813,15 +811,28 @@ out_be: /* Adjust for extra insns */ rvoff -= (e - s) << 2; + emit_branch(BPF_OP(code), rd, rs, rvoff, ctx); + break; - if (BPF_OP(code) == BPF_JSET) { - /* Adjust for and */ - rvoff -= 4; - emit(rv_and(rs, rd, rs), ctx); - emit_branch(BPF_JNE, rs, RV_REG_ZERO, rvoff, ctx); + case BPF_JMP | BPF_JSET | BPF_K: + case BPF_JMP32 | BPF_JSET | BPF_K: + rvoff = rv_offset(i, off, ctx); + s = ctx->ninsns; + if (is_12b_int(imm)) { + emit(rv_andi(RV_REG_T1, rd, imm), ctx); } else { - emit_branch(BPF_OP(code), rd, rs, rvoff, ctx); + emit_imm(RV_REG_T1, imm, ctx); + emit(rv_and(RV_REG_T1, rd, RV_REG_T1), ctx); } + /* For jset32, we should clear the upper 32 bits of t1, but + * sign-extension is sufficient here and saves one instruction, + * as t1 is used only in comparison against zero. + */ + if (!is64 && imm < 0) + emit(rv_addiw(RV_REG_T1, RV_REG_T1, 0), ctx); + e = ctx->ninsns; + rvoff -= (e - s) << 2; + emit_branch(BPF_JNE, RV_REG_T1, RV_REG_ZERO, rvoff, ctx); break; /* function call */ |