[PATCH] LoongArch: BPF: Optimize doubleword loads with ld.d and ldptr.d
From: Tiezhu Yang
Date: Wed Nov 18 2026 - 03:12:16 EST
Currently, 64-bit doubleword loads (BPF_DW) unconditionally load the
offset into a temporary register with move_imm() and emit ldx.d.
Optimize these loads by emitting ld.d for signed 12-bit offsets and
ldptr.d for 4-byte aligned signed 16-bit offsets, or falling back to
move_imm() + ldx.d when necessary.
With this patch, only a single load instruction is emitted in most
cases.
Signed-off-by: Tiezhu Yang <yangtiezhu@xxxxxxxxxxx>
---
This is based on the patch:
LoongArch: BPF: Fix the offset of stptr.w and stptr.d stores
https://lore.kernel.org/loongarch/20261118071444.10258-1-dongtai.guo@xxxxxxxxx/T/#u
Note that BPF_W is left unchanged because LoongArch provides ldptr.w
(which sign-extends to 64 bits) but lacks an unsigned zero-extending
ldptr.wu instruction required by standard 32-bit BPF loads.
arch/loongarch/net/bpf_jit.c | 10 ++++++++--
1 file changed, 8 insertions(+), 2 deletions(-)
diff --git a/arch/loongarch/net/bpf_jit.c b/arch/loongarch/net/bpf_jit.c
index 906f34f420f1..d36891db1179 100644
--- a/arch/loongarch/net/bpf_jit.c
+++ b/arch/loongarch/net/bpf_jit.c
@@ -1352,8 +1352,14 @@ static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx, bool ext
}
break;
case BPF_DW:
- move_imm(ctx, t1, off);
- emit_insn(ctx, ldxd, dst, src, t1);
+ if (is_signed_imm12(off)) {
+ emit_insn(ctx, ldd, dst, src, off);
+ } else if (is_aligned_signed_imm16(off)) {
+ emit_insn(ctx, ldptrd, dst, src, off >> 2);
+ } else {
+ move_imm(ctx, t1, off);
+ emit_insn(ctx, ldxd, dst, src, t1);
+ }
break;
}
--
2.42.0