From: Matt Corallo Date: Thu, 9 Dec 2021 21:14:52 +0000 (+0000) Subject: Combine redundant rule conditions to work around LLVM bug 52455 fully X-Git-Url: http://git.bitcoin.ninja/index.cgi?p=flowspec-xdp;a=commitdiff_plain;h=822037d58f379cca575bbeb74647cc60ba8113c5 Combine redundant rule conditions to work around LLVM bug 52455 fully This adds a trivial optimization pass to combine redundant rule conditions in back-to-back rules before printing them. --- diff --git a/genrules.py b/genrules.py index 10f138f..3865b6b 100755 --- a/genrules.py +++ b/genrules.py @@ -256,6 +256,69 @@ def flow_label_to_rule(rules): return f"""if (ip6 == NULL) break; if (!( {ast.write("((((uint32_t)(ip6->flow_lbl[0] & 0xf)) << 2*8) | (((uint32_t)ip6->flow_lbl[1]) << 1*8) | (uint32_t)ip6->flow_lbl[0])")} )) break;""" +class RuleAction(Enum): + CONDITIONS = 1 + ACTION = 2 + LIST = 3 +class RuleNode: + def __init__(self, ty, action, inner): + self.ty = ty + self.action = action + self.inner = inner + if ty == RuleAction.ACTION: + assert inner is None + assert type(action) == str + elif ty == RuleAction.LIST: + assert type(inner) == list + assert action is None + for item in inner: + assert type(item) == RuleNode + else: + assert ty == RuleAction.CONDITIONS + assert type(action) == list + assert type(inner) == RuleNode + + def maybe_join(self, neighbor): + if self.ty == RuleAction.CONDITIONS and neighbor.ty == RuleAction.CONDITIONS: + overlapping_conditions = [x for x in self.action if x in neighbor.action] + if len(overlapping_conditions) != 0: + us = RuleNode(RuleAction.CONDITIONS, [x for x in self.action if x not in overlapping_conditions], self.inner) + them = RuleNode(RuleAction.CONDITIONS, [x for x in neighbor.action if x not in overlapping_conditions], neighbor.inner) + self.action = overlapping_conditions + if self.inner.ty == RuleAction.LIST and us.action == []: + self.inner.inner.append(them) + else: + self.inner = RuleNode(RuleAction.LIST, None, [us, them]) + self.inner.flatten() + return True + return False + + def flatten(self): + # LLVM can be pretty bad at optimizing out common subexpressions. Thus, we have to do a + # pass here to toptimize out common subexpressions in back-to-back rules. + # See https://bugs.llvm.org/show_bug.cgi?id=52455 + assert self.ty == RuleAction.LIST + did_update = True + while did_update: + did_update = False + for i in range(0, len(self.inner) - 1): + if self.inner[i].maybe_join(self.inner[i + 1]): + del self.inner[i + 1] + did_update = True + break + + def write(self, out, pfx="\t"): + if self.ty == RuleAction.CONDITIONS: + out.write(pfx + "do {\\\n") + for cond in self.action: + out.write("\t" + pfx + cond.strip().replace("\n", " \\\n\t" + pfx) + " \\\n") + self.inner.write(out, pfx) + out.write(pfx + "} while(0);\\\n") + elif self.ty == RuleAction.LIST: + for item in self.inner: + item.write(out, pfx + "\t") + else: + out.write("\t" + pfx + self.action.strip().replace("\n", " \\\n\t" + pfx) + " \\\n") with open("rules.h", "w") as out: parse = argparse.ArgumentParser() @@ -291,8 +354,8 @@ with open("rules.h", "w") as out: assert False out.write("#define REQ_8021Q " + args.vlan_tag + "\n") - rules6 = "" - rules4 = "" + rules6 = [] + rules4 = [] use_v6_frags = False stats_rulecnt = 0 ratelimitcnt = 0 @@ -313,29 +376,15 @@ with open("rules.h", "w") as out: t = lastrule.split("{") if t[0].strip() == "flow4": proto = 4 - rules4 += "\tdo {\\\n" elif t[0].strip() == "flow6": proto = 6 - rules6 += "\tdo {\\\n" else: continue - # LLVM can be pretty bad at optimizing out common subexpressions. Ideally we'd optimize - # by pulling common subexpressions in back-to-back rules out into a single check, but - # that's a bunch of work that LLVM really should do for us. Instead, we blindly guess - # that source-address is the least likely to be a common subexpression and rely on LLVM - # managing to pull out common subexpressions as long as they're the first check(s). By - # placing source-address checks last, LLVM should do at least some work for us. - # See https://bugs.llvm.org/show_bug.cgi?id=52455 - last_checks = "" - def write_rule(r, place_at_end=False): - global rules4, rules6, last_checks - if place_at_end: - last_checks += "\t\t" + r.replace("\n", " \\\n\t\t") + " \\\n" - elif proto == 6: - rules6 += "\t\t" + r.replace("\n", " \\\n\t\t") + " \\\n" - else: - rules4 += "\t\t" + r.replace("\n", " \\\n\t\t") + " \\\n" + conditions = [] + def write_rule(r): + global conditions + conditions.append(r + "\n") rule = t[1].split("}")[0].strip() for step in rule.split(";"): @@ -347,7 +396,7 @@ with open("rules.h", "w") as out: else: offset = None if step.strip().startswith("src"): - write_rule(ip_to_rule(proto, nets[0], "saddr", offset), True) + write_rule(ip_to_rule(proto, nets[0], "saddr", offset)) else: write_rule(ip_to_rule(proto, nets[0], "daddr", offset)) elif step.strip().startswith("proto") and proto == 4: @@ -377,10 +426,10 @@ with open("rules.h", "w") as out: else: assert False - if proto == 6: - rules6 += last_checks - else: - rules4 += last_checks + actions = "" + def write_rule(r): + global actions + actions += r + "\n" # Now write the match handling! first_action = None @@ -485,9 +534,9 @@ with open("rules.h", "w") as out: if last_action is not None: write_rule(last_action) if proto == 6: - rules6 += "\t} while(0);\\\n" + rules6.append(RuleNode(RuleAction.CONDITIONS, conditions, RuleNode(RuleAction.ACTION, actions, None))) else: - rules4 += "\t} while(0);\\\n" + rules4.append(RuleNode(RuleAction.CONDITIONS, conditions, RuleNode(RuleAction.ACTION, actions, None))) if stats_action != "": print(rule) stats_rulecnt += 1 @@ -497,12 +546,26 @@ with open("rules.h", "w") as out: out.write(f"#define STATS_RULECNT {stats_rulecnt}\n") if ratelimitcnt != 0: out.write(f"#define RATE_CNT {ratelimitcnt}\n") - if rules4 != "": + + # Here we should probably sort the rules according to flowspec's sorting rules. We don't bother + # however, because its annoying. + + if len(rules4) != 0: out.write("#define NEED_V4_PARSE\n") - out.write("#define RULES4 {\\\n" + rules4 + "}\n") - if rules6: + out.write("#define RULES4 {\\\n") + rules4 = RuleNode(RuleAction.LIST, None, rules4) + rules4.flatten() + rules4.write(out) + out.write("}\n") + + if len(rules6) != 0: out.write("#define NEED_V6_PARSE\n") - out.write("#define RULES6 {\\\n" + rules6 + "}\n") + out.write("#define RULES6 {\\\n") + rules6 = RuleNode(RuleAction.LIST, None, rules6) + rules6.flatten() + rules6.write(out) + out.write("}\n") + if args.v6frag == "ignore-parse-if-rule": if use_v6_frags: out.write("#define PARSE_V6_FRAG PARSE\n")