[PATCH bpf-next 0/4] nfp: bpf: optimize negative sums

[PATCH bpf-next 0/4] nfp: bpf: optimize negative sums

[Jakub Kicinski]

Hi!

This set adds an optimization run to the NFP jit to turn ADD and SUB
instructions with negative immediate into the opposite operation with
a positive immediate.  NFP can fit small immediates into the instructions
but it can't ever fit negative immediates.  Addition of small negative
immediates is quite common in BPF programs for stack address calculations,
therefore this optimization gives us non-negligible savings in instruction
count (up to 4%).


Jakub Kicinski (4):
  nfp: bpf: remove double space
  nfp: bpf: optimize add/sub of a negative constant
  nfp: bpf: tabularize generations of compare operations
  nfp: bpf: optimize comparisons to negative constants

 drivers/net/ethernet/netronome/nfp/bpf/jit.c  | 231 +++++++++++++-------------
 drivers/net/ethernet/netronome/nfp/bpf/main.h |   6 +-
 2 files changed, 124 insertions(+), 113 deletions(-)

-- 
2.16.2

[PATCH bpf-next 1/4] nfp: bpf: remove double space

[Jakub Kicinski]

Whitespace cleanup - remove double space.

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
---
 drivers/net/ethernet/netronome/nfp/bpf/jit.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/netronome/nfp/bpf/jit.c b/drivers/net/ethernet/netronome/nfp/bpf/jit.c
index 29b4e5f8c102..9cc638718272 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/jit.c
+++ b/drivers/net/ethernet/netronome/nfp/bpf/jit.c
@@ -1400,7 +1400,7 @@ map_call_stack_common(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 	if (!load_lm_ptr)
 		return 0;
 
-	emit_csr_wr(nfp_prog, stack_reg(nfp_prog),  NFP_CSR_ACT_LM_ADDR0);
+	emit_csr_wr(nfp_prog, stack_reg(nfp_prog), NFP_CSR_ACT_LM_ADDR0);
 	wrp_nops(nfp_prog, 3);
 
 	return 0;
-- 
2.16.2

[PATCH bpf-next 2/4] nfp: bpf: optimize add/sub of a negative constant

[Jakub Kicinski]

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>
---
 drivers/net/ethernet/netronome/nfp/bpf/jit.c | 35 ++++++++++++++++++++++++++++
 1 file changed, 35 insertions(+)

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);
-- 
2.16.2

[PATCH bpf-next 3/4] nfp: bpf: tabularize generations of compare operations

[Jakub Kicinski]

There are quite a few compare instructions now, use a table
to translate BPF instruction code to NFP instruction parameters
instead of parameterizing helpers.  This saves LOC and makes
future extensions easier.

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
---
 drivers/net/ethernet/netronome/nfp/bpf/jit.c | 168 ++++++++++-----------------
 1 file changed, 61 insertions(+), 107 deletions(-)

diff --git a/drivers/net/ethernet/netronome/nfp/bpf/jit.c b/drivers/net/ethernet/netronome/nfp/bpf/jit.c
index a5590988fc69..5b8da7a67df4 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/jit.c
+++ b/drivers/net/ethernet/netronome/nfp/bpf/jit.c
@@ -1214,45 +1214,79 @@ wrp_test_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
 	return 0;
 }
 
-static int
-wrp_cmp_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
-	    enum br_mask br_mask, bool swap)
+static const struct jmp_code_map {
+	enum br_mask br_mask;
+	bool swap;
+} jmp_code_map[] = {
+	[BPF_JGT >> 4]	= { BR_BLO, true },
+	[BPF_JGE >> 4]	= { BR_BHS, false },
+	[BPF_JLT >> 4]	= { BR_BLO, false },
+	[BPF_JLE >> 4]	= { BR_BHS, true },
+	[BPF_JSGT >> 4]	= { BR_BLT, true },
+	[BPF_JSGE >> 4]	= { BR_BGE, false },
+	[BPF_JSLT >> 4]	= { BR_BLT, false },
+	[BPF_JSLE >> 4]	= { BR_BGE, true },
+};
+
+static const struct jmp_code_map *nfp_jmp_code_get(struct nfp_insn_meta *meta)
+{
+	unsigned int op;
+
+	op = BPF_OP(meta->insn.code) >> 4;
+	/* br_mask of 0 is BR_BEQ which we don't use in jump code table */
+	if (WARN_ONCE(op >= ARRAY_SIZE(jmp_code_map) ||
+		      !jmp_code_map[op].br_mask,
+		      "no code found for jump instruction"))
+		return NULL;
+
+	return &jmp_code_map[op];
+}
+
+static int cmp_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 {
 	const struct bpf_insn *insn = &meta->insn;
 	u64 imm = insn->imm; /* sign extend */
+	const struct jmp_code_map *code;
 	u8 reg = insn->dst_reg * 2;
 	swreg tmp_reg;
 
+	code = nfp_jmp_code_get(meta);
+	if (!code)
+		return -EINVAL;
+
 	tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog));
-	if (!swap)
+	if (!code->swap)
 		emit_alu(nfp_prog, reg_none(), reg_a(reg), ALU_OP_SUB, tmp_reg);
 	else
 		emit_alu(nfp_prog, reg_none(), tmp_reg, ALU_OP_SUB, reg_a(reg));
 
 	tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog));
-	if (!swap)
+	if (!code->swap)
 		emit_alu(nfp_prog, reg_none(),
 			 reg_a(reg + 1), ALU_OP_SUB_C, tmp_reg);
 	else
 		emit_alu(nfp_prog, reg_none(),
 			 tmp_reg, ALU_OP_SUB_C, reg_a(reg + 1));
 
-	emit_br(nfp_prog, br_mask, insn->off, 0);
+	emit_br(nfp_prog, code->br_mask, insn->off, 0);
 
 	return 0;
 }
 
-static int
-wrp_cmp_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
-	    enum br_mask br_mask, bool swap)
+static int cmp_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 {
 	const struct bpf_insn *insn = &meta->insn;
+	const struct jmp_code_map *code;
 	u8 areg, breg;
 
+	code = nfp_jmp_code_get(meta);
+	if (!code)
+		return -EINVAL;
+
 	areg = insn->dst_reg * 2;
 	breg = insn->src_reg * 2;
 
-	if (swap) {
+	if (code->swap) {
 		areg ^= breg;
 		breg ^= areg;
 		areg ^= breg;
@@ -1261,7 +1295,7 @@ wrp_cmp_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta,
 	emit_alu(nfp_prog, reg_none(), reg_a(areg), ALU_OP_SUB, reg_b(breg));
 	emit_alu(nfp_prog, reg_none(),
 		 reg_a(areg + 1), ALU_OP_SUB_C, reg_b(breg + 1));
-	emit_br(nfp_prog, br_mask, insn->off, 0);
+	emit_br(nfp_prog, code->br_mask, insn->off, 0);
 
 	return 0;
 }
@@ -2283,46 +2317,6 @@ static int jeq_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 	return 0;
 }
 
-static int jgt_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_imm(nfp_prog, meta, BR_BLO, true);
-}
-
-static int jge_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_imm(nfp_prog, meta, BR_BHS, false);
-}
-
-static int jlt_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_imm(nfp_prog, meta, BR_BLO, false);
-}
-
-static int jle_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_imm(nfp_prog, meta, BR_BHS, true);
-}
-
-static int jsgt_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_imm(nfp_prog, meta, BR_BLT, true);
-}
-
-static int jsge_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_imm(nfp_prog, meta, BR_BGE, false);
-}
-
-static int jslt_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_imm(nfp_prog, meta, BR_BLT, false);
-}
-
-static int jsle_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_imm(nfp_prog, meta, BR_BGE, true);
-}
-
 static int jset_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 {
 	const struct bpf_insn *insn = &meta->insn;
@@ -2392,46 +2386,6 @@ static int jeq_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 	return 0;
 }
 
-static int jgt_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_reg(nfp_prog, meta, BR_BLO, true);
-}
-
-static int jge_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_reg(nfp_prog, meta, BR_BHS, false);
-}
-
-static int jlt_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_reg(nfp_prog, meta, BR_BLO, false);
-}
-
-static int jle_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_reg(nfp_prog, meta, BR_BHS, true);
-}
-
-static int jsgt_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_reg(nfp_prog, meta, BR_BLT, true);
-}
-
-static int jsge_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_reg(nfp_prog, meta, BR_BGE, false);
-}
-
-static int jslt_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_reg(nfp_prog, meta, BR_BLT, false);
-}
-
-static int jsle_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
-{
-	return wrp_cmp_reg(nfp_prog, meta, BR_BGE, true);
-}
-
 static int jset_reg(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 {
 	return wrp_test_reg(nfp_prog, meta, ALU_OP_AND, BR_BNE);
@@ -2520,25 +2474,25 @@ static const instr_cb_t instr_cb[256] = {
 	[BPF_ST | BPF_MEM | BPF_DW] =	mem_st8,
 	[BPF_JMP | BPF_JA | BPF_K] =	jump,
 	[BPF_JMP | BPF_JEQ | BPF_K] =	jeq_imm,
-	[BPF_JMP | BPF_JGT | BPF_K] =	jgt_imm,
-	[BPF_JMP | BPF_JGE | BPF_K] =	jge_imm,
-	[BPF_JMP | BPF_JLT | BPF_K] =	jlt_imm,
-	[BPF_JMP | BPF_JLE | BPF_K] =	jle_imm,
-	[BPF_JMP | BPF_JSGT | BPF_K] =  jsgt_imm,
-	[BPF_JMP | BPF_JSGE | BPF_K] =  jsge_imm,
-	[BPF_JMP | BPF_JSLT | BPF_K] =  jslt_imm,
-	[BPF_JMP | BPF_JSLE | BPF_K] =  jsle_imm,
+	[BPF_JMP | BPF_JGT | BPF_K] =	cmp_imm,
+	[BPF_JMP | BPF_JGE | BPF_K] =	cmp_imm,
+	[BPF_JMP | BPF_JLT | BPF_K] =	cmp_imm,
+	[BPF_JMP | BPF_JLE | BPF_K] =	cmp_imm,
+	[BPF_JMP | BPF_JSGT | BPF_K] =  cmp_imm,
+	[BPF_JMP | BPF_JSGE | BPF_K] =  cmp_imm,
+	[BPF_JMP | BPF_JSLT | BPF_K] =  cmp_imm,
+	[BPF_JMP | BPF_JSLE | BPF_K] =  cmp_imm,
 	[BPF_JMP | BPF_JSET | BPF_K] =	jset_imm,
 	[BPF_JMP | BPF_JNE | BPF_K] =	jne_imm,
 	[BPF_JMP | BPF_JEQ | BPF_X] =	jeq_reg,
-	[BPF_JMP | BPF_JGT | BPF_X] =	jgt_reg,
-	[BPF_JMP | BPF_JGE | BPF_X] =	jge_reg,
-	[BPF_JMP | BPF_JLT | BPF_X] =	jlt_reg,
-	[BPF_JMP | BPF_JLE | BPF_X] =	jle_reg,
-	[BPF_JMP | BPF_JSGT | BPF_X] =  jsgt_reg,
-	[BPF_JMP | BPF_JSGE | BPF_X] =  jsge_reg,
-	[BPF_JMP | BPF_JSLT | BPF_X] =  jslt_reg,
-	[BPF_JMP | BPF_JSLE | BPF_X] =  jsle_reg,
+	[BPF_JMP | BPF_JGT | BPF_X] =	cmp_reg,
+	[BPF_JMP | BPF_JGE | BPF_X] =	cmp_reg,
+	[BPF_JMP | BPF_JLT | BPF_X] =	cmp_reg,
+	[BPF_JMP | BPF_JLE | BPF_X] =	cmp_reg,
+	[BPF_JMP | BPF_JSGT | BPF_X] =  cmp_reg,
+	[BPF_JMP | BPF_JSGE | BPF_X] =  cmp_reg,
+	[BPF_JMP | BPF_JSLT | BPF_X] =  cmp_reg,
+	[BPF_JMP | BPF_JSLE | BPF_X] =  cmp_reg,
 	[BPF_JMP | BPF_JSET | BPF_X] =	jset_reg,
 	[BPF_JMP | BPF_JNE | BPF_X] =	jne_reg,
 	[BPF_JMP | BPF_CALL] =		call,
-- 
2.16.2

[PATCH bpf-next 4/4] nfp: bpf: optimize comparisons to negative constants

[Jakub Kicinski]

Comparison instruction requires a subtraction.  If the constant
is negative we are more likely to fit it into a NFP instruction
directly if we change the sign and use addition.

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
---
 drivers/net/ethernet/netronome/nfp/bpf/jit.c  | 42 +++++++++++++++++++--------
 drivers/net/ethernet/netronome/nfp/bpf/main.h |  6 +++-
 2 files changed, 35 insertions(+), 13 deletions(-)

diff --git a/drivers/net/ethernet/netronome/nfp/bpf/jit.c b/drivers/net/ethernet/netronome/nfp/bpf/jit.c
index 5b8da7a67df4..65f0791cae0c 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/jit.c
+++ b/drivers/net/ethernet/netronome/nfp/bpf/jit.c
@@ -1247,6 +1247,7 @@ static int cmp_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 	const struct bpf_insn *insn = &meta->insn;
 	u64 imm = insn->imm; /* sign extend */
 	const struct jmp_code_map *code;
+	enum alu_op alu_op, carry_op;
 	u8 reg = insn->dst_reg * 2;
 	swreg tmp_reg;
 
@@ -1254,19 +1255,22 @@ static int cmp_imm(struct nfp_prog *nfp_prog, struct nfp_insn_meta *meta)
 	if (!code)
 		return -EINVAL;
 
+	alu_op = meta->jump_neg_op ? ALU_OP_ADD : ALU_OP_SUB;
+	carry_op = meta->jump_neg_op ? ALU_OP_ADD_C : ALU_OP_SUB_C;
+
 	tmp_reg = ur_load_imm_any(nfp_prog, imm & ~0U, imm_b(nfp_prog));
 	if (!code->swap)
-		emit_alu(nfp_prog, reg_none(), reg_a(reg), ALU_OP_SUB, tmp_reg);
+		emit_alu(nfp_prog, reg_none(), reg_a(reg), alu_op, tmp_reg);
 	else
-		emit_alu(nfp_prog, reg_none(), tmp_reg, ALU_OP_SUB, reg_a(reg));
+		emit_alu(nfp_prog, reg_none(), tmp_reg, alu_op, reg_a(reg));
 
 	tmp_reg = ur_load_imm_any(nfp_prog, imm >> 32, imm_b(nfp_prog));
 	if (!code->swap)
 		emit_alu(nfp_prog, reg_none(),
-			 reg_a(reg + 1), ALU_OP_SUB_C, tmp_reg);
+			 reg_a(reg + 1), carry_op, tmp_reg);
 	else
 		emit_alu(nfp_prog, reg_none(),
-			 tmp_reg, ALU_OP_SUB_C, reg_a(reg + 1));
+			 tmp_reg, carry_op, reg_a(reg + 1));
 
 	emit_br(nfp_prog, code->br_mask, insn->off, 0);
 
@@ -2745,21 +2749,35 @@ static void nfp_bpf_opt_neg_add_sub(struct nfp_prog *nfp_prog)
 			continue;
 
 		if (BPF_CLASS(insn.code) != BPF_ALU &&
-		    BPF_CLASS(insn.code) != BPF_ALU64)
+		    BPF_CLASS(insn.code) != BPF_ALU64 &&
+		    BPF_CLASS(insn.code) != BPF_JMP)
 			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;
+		if (BPF_CLASS(insn.code) == BPF_JMP) {
+			switch (BPF_OP(insn.code)) {
+			case BPF_JGE:
+			case BPF_JSGE:
+			case BPF_JLT:
+			case BPF_JSLT:
+				meta->jump_neg_op = true;
+				break;
+			default:
+				continue;
+			}
+		} else {
+			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.code = insn.code | BPF_K;
+		}
 
 		meta->insn.imm = -insn.imm;
 	}
diff --git a/drivers/net/ethernet/netronome/nfp/bpf/main.h b/drivers/net/ethernet/netronome/nfp/bpf/main.h
index 4981c8944ca3..68b5d326483d 100644
--- a/drivers/net/ethernet/netronome/nfp/bpf/main.h
+++ b/drivers/net/ethernet/netronome/nfp/bpf/main.h
@@ -236,6 +236,7 @@ struct nfp_bpf_reg_state {
  * @xadd_over_16bit: 16bit immediate is not guaranteed
  * @xadd_maybe_16bit: 16bit immediate is possible
  * @jmp_dst: destination info for jump instructions
+ * @jump_neg_op: jump instruction has inverted immediate, use ADD instead of SUB
  * @func_id: function id for call instructions
  * @arg1: arg1 for call instructions
  * @arg2: arg2 for call instructions
@@ -264,7 +265,10 @@ struct nfp_insn_meta {
 			bool xadd_maybe_16bit;
 		};
 		/* jump */
-		struct nfp_insn_meta *jmp_dst;
+		struct {
+			struct nfp_insn_meta *jmp_dst;
+			bool jump_neg_op;
+		};
 		/* function calls */
 		struct {
 			u32 func_id;
-- 
2.16.2

Re: [PATCH bpf-next 0/4] nfp: bpf: optimize negative sums

[Daniel Borkmann]

On 04/25/2018 06:22 AM, Jakub Kicinski wrote:
> Hi!
> 
> This set adds an optimization run to the NFP jit to turn ADD and SUB
> instructions with negative immediate into the opposite operation with
> a positive immediate.  NFP can fit small immediates into the instructions
> but it can't ever fit negative immediates.  Addition of small negative
> immediates is quite common in BPF programs for stack address calculations,
> therefore this optimization gives us non-negligible savings in instruction
> count (up to 4%).

Applied to bpf-next, thanks Jakub!