Generate instructions to perform the endian conversion using registers,
rather than generating two memory accesses.
The "way easier and faster" comment was obviously for the author, not
the processor.
Signed-off-by: Naveen N. Rao <naveen.n....@linux.vnet.ibm.com>
---
arch/powerpc/net/bpf_jit_comp64.c | 22 ++++++++++++++--------
1 file changed, 14 insertions(+), 8 deletions(-)
diff --git a/arch/powerpc/net/bpf_jit_comp64.c
b/arch/powerpc/net/bpf_jit_comp64.c
index 1e313db..0413a89 100644
--- a/arch/powerpc/net/bpf_jit_comp64.c
+++ b/arch/powerpc/net/bpf_jit_comp64.c
@@ -599,16 +599,22 @@ static int bpf_jit_build_body(struct bpf_prog *fp, u32
*image,
break;
case 64:
/*
- * Way easier and faster(?) to store the value
- * into stack and then use ldbrx
+ * We'll split it up into two words, swap those
+ * independently and then merge them back.
*
- * ctx->seen will be reliable in pass2, but
- * the instructions generated will remain the
- * same across all passes
+ * First up, let's swap the most-significant
word.
*/
- PPC_STD(dst_reg, 1, bpf_jit_stack_local(ctx));
- PPC_ADDI(b2p[TMP_REG_1], 1,
bpf_jit_stack_local(ctx));
- PPC_LDBRX(dst_reg, 0, b2p[TMP_REG_1]);
+ PPC_RLDICL(b2p[TMP_REG_1], dst_reg, 32, 32);
+ PPC_RLWINM(b2p[TMP_REG_2], b2p[TMP_REG_1], 8,
0, 31);
+ PPC_RLWIMI(b2p[TMP_REG_2], b2p[TMP_REG_1], 24,
0, 7);
+ PPC_RLWIMI(b2p[TMP_REG_2], b2p[TMP_REG_1], 24,
16, 23);
+ /* Then, the second half */
+ PPC_RLWINM(b2p[TMP_REG_1], dst_reg, 8, 0, 31);
+ PPC_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 0, 7);
+ PPC_RLWIMI(b2p[TMP_REG_1], dst_reg, 24, 16, 23);
+ /* Merge back */
+ PPC_RLDICR(dst_reg, b2p[TMP_REG_1], 32, 31);
+ PPC_OR(dst_reg, dst_reg, b2p[TMP_REG_2]);
break;
}
break;
--
2.10.2
