[PATCH RFC 1/2] LoongArch: BPF: Eliminate zero-extension for memory load operations

[Tiezhu Yang]

The unsigned memory load instructions naturally clear the upper 32 bits
of the destination register during sub-word and word memory accesses.

Add a look-ahead optimization for the load instructions by checking the
next instruction. This is achieved by overwriting bpf_jit_needs_zext()
to return true.

If insn_is_zext() detects a redundant zero-extension, return 1 to inform
the outer build_body() to skip the instruction.

For normal loads, this look-ahead is safely restricted to unsigned modes
(!sign_extend) to preserve signed contexts.

With this patch, directly reduce the final JITted image size.

Signed-off-by: Tiezhu Yang <yangtiezhu@xxxxxxxxxxx>
---
 arch/loongarch/net/bpf_jit.c | 21 +++++++++++++++++++++
 1 file changed, 21 insertions(+)

diff --git a/arch/loongarch/net/bpf_jit.c b/arch/loongarch/net/bpf_jit.c
index 69bc30f88c44..b8cef3ea5c6d 100644
--- a/arch/loongarch/net/bpf_jit.c
+++ b/arch/loongarch/net/bpf_jit.c
@@ -579,6 +579,14 @@ static int emit_atomic_ld_st(const struct bpf_insn *insn, struct jit_ctx *ctx)
 			break;
 		}
 		emit_insn(ctx, dbar, 0b10100);
+
+		/*
+		 * If our next insn is a redundant zext, return 1 to tell
+		 * build_body() to skip it, as atomic load-acquires always
+		 * zero-extend sub-word results.
+		 */
+		if (BPF_SIZE(code) != BPF_DW && insn_is_zext(&insn[1]))
+			return 1;
 		break;
 	/* store_release(dst_reg + off16, src_reg) */
 	case BPF_STORE_REL:
@@ -1318,6 +1326,14 @@ static int build_insn(const struct bpf_insn *insn, struct jit_ctx *ctx, bool ext
 		ret = add_exception_handler(insn, ctx, dst);
 		if (ret)
 			return ret;
+
+		/*
+		 * If our next insn is a redundant zext, and this is a standard
+		 * unsigned sub-word load where hardware already zero-extends,
+		 * return 1 to tell build_body() to skip it.
+		 */
+		if (BPF_SIZE(code) != BPF_DW && !sign_extend && insn_is_zext(&insn[1]))
+			return 1;
 		break;
 
 	/* *(size *)(dst + off) = imm */
@@ -2397,6 +2413,11 @@ bool bpf_jit_supports_subprog_tailcalls(void)
 	return true;
 }
 
+bool bpf_jit_needs_zext(void)
+{
+	return true;
+}
+
 bool bpf_jit_inlines_helper_call(s32 imm)
 {
 	switch (imm) {
-- 
2.42.0