From: Mikko Rasa <tdb@tdb.fi>
Date: Fri, 29 Dec 2023 20:36:05 +0000 (+0200)
Subject: Optimize memory accesses in Regex::run
X-Git-Url: http://git.tdb.fi/?a=commitdiff_plain;h=5b5776a76b24e26a46d3bacefffa57485cb94abc;p=libs%2Fcore.git

Optimize memory accesses in Regex::run

Use a vector instead of a linked list to store execution contexts, and
store matched groups in contiguous memory as well.
---

diff --git a/source/strings/regex.cpp b/source/strings/regex.cpp
index 5307c69..c61fdfe 100644
--- a/source/strings/regex.cpp
+++ b/source/strings/regex.cpp
@@ -375,15 +375,16 @@ RegMatch Regex::match(const string &str) const
 	return RegMatch();
 }
 
-bool Regex::run(const string &str, const string::const_iterator &begin, vector<RegMatch::Group> &groups) const
+bool Regex::run(const string &str, const string::const_iterator &begin, vector<RegMatch::Group> &out_groups) const
 {
 	bool result = false;
-	list<RunContext> ctx;
-	ctx.push_back(RunContext());
-	ctx.front().citer = code.begin();
-	ctx.front().groups.resize(groups.size());
+	vector<RunContext> ctx(1);
+	ctx.front().code_iter = code.begin();
+	vector<RegMatch::Group> groups(out_groups.size());
+	size_t ctx_count = 1;
+	size_t best_groups = 0;
 
-	for(auto i=begin;;)
+	for(auto i=begin;; ++i)
 	{
 		int c;
 		if(i!=str.end())
@@ -391,55 +392,65 @@ bool Regex::run(const string &str, const string::const_iterator &begin, vector<R
 		else
 			c = -1;
 
-		for(auto j=ctx.begin(); j!=ctx.end();)
+		for(size_t j=0; j<ctx_count; ++j)
 		{
-			bool terminate = false;
 			bool negate_match = false;
-			for(; j->citer!=code.end();)
+			while(ctx[j].code_iter!=code.end())
 			{
-				Instruction instr = static_cast<Instruction>(*j->citer++);
+				Instruction instr = static_cast<Instruction>(*ctx[j].code_iter++);
 
 				if(instr==NEGATE)
 					negate_match = true;
 				else if(instr==JUMP)
 				{
-					Offset offset = read_int<Offset>(j->citer);
-					j->citer += offset;
+					Offset offset = read_int<Offset>(ctx[j].code_iter);
+					ctx[j].code_iter += offset;
 				}
 				else if(instr==ND_JUMP)
 				{
-					Offset offset = read_int<Offset>(j->citer);
-					ctx.push_back(*j);
-					ctx.back().citer += offset;
+					Offset offset = read_int<Offset>(ctx[j].code_iter);
+					if(ctx_count>=ctx.size())
+					{
+						ctx.emplace_back();
+						ctx[ctx_count].groups_index = groups.size();
+						groups.resize(groups.size()+out_groups.size());
+					}
+					ctx[ctx_count].code_iter = ctx[j].code_iter+offset;
+					RegMatch::Group *groups_ptr = groups.data()+ctx[j].groups_index;
+					copy(groups_ptr, groups_ptr+out_groups.size(), groups.data()+ctx[ctx_count].groups_index);
+					++ctx_count;
 				}
 				else if(instr==GROUP_BEGIN)
 				{
-					Index n = read_int<Index>(j->citer);
-					if(!j->groups[n].match)
-						j->groups[n].begin = i-str.begin();
+					RegMatch::Group *groups_ptr = groups.data()+ctx[j].groups_index;
+					Index n = read_int<Index>(ctx[j].code_iter);
+					if(!groups_ptr[n].match)
+						groups_ptr[n].begin = i-str.begin();
 				}
 				else if(instr==GROUP_END)
 				{
-					Index n = read_int<Index>(j->citer);
-					if(!j->groups[n].match)
+					RegMatch::Group *groups_ptr = groups.data()+ctx[j].groups_index;
+					Index n = read_int<Index>(ctx[j].code_iter);
+					if(!groups_ptr[n].match)
 					{
-						j->groups[n].match = true;
-						j->groups[n].end = i-str.begin();
-						j->groups[n].length = j->groups[n].end-j->groups[n].begin;
+						groups_ptr[n].match = true;
+						groups_ptr[n].end = i-str.begin();
+						groups_ptr[n].length = groups_ptr[n].end-groups_ptr[n].begin;
 					}
 
 					if(n==0)
 					{
 						result = true;
 						bool better = false;
-						for(unsigned k=0; (k<groups.size() && !better); ++k)
+						const RegMatch::Group *best_ptr = groups.data()+best_groups;
+						for(unsigned k=0; (k<out_groups.size() && !better); ++k)
 						{
-							better = group_compare(j->groups[k], groups[k]);
-							if(group_compare(groups[k], j->groups[k]))
+							better = group_compare(groups_ptr[k], best_ptr[k]);
+							if(group_compare(best_ptr[k], groups_ptr[k]))
 								break;
 						}
 						if(better)
-							groups = j->groups;
+							best_groups = ctx[j].groups_index;
 					}
 				}
 				else
@@ -452,13 +463,13 @@ bool Regex::run(const string &str, const string::const_iterator &begin, vector<R
 						match_result = (i==str.end());
 					else if(instr==MATCH_CHAR)
 					{
-						match_result = (c==*j->citer++);
+						match_result = (c==*ctx[j].code_iter++);
 						input_consumed = true;
 					}
 					else if(instr==MATCH_RANGE)
 					{
-						unsigned char first = *j->citer++;
-						unsigned char last = *j->citer++;
+						unsigned char first = *ctx[j].code_iter++;
+						unsigned char last = *ctx[j].code_iter++;
 						match_result = (c>=first && c<=last);
 						input_consumed = true;
 					}
@@ -466,11 +477,11 @@ bool Regex::run(const string &str, const string::const_iterator &begin, vector<R
 					{
 						if(c>=0 && c<=0xFF)
 						{
-							unsigned char m = *(j->citer+(c>>3));
+							unsigned char m = *(ctx[j].code_iter+(c>>3));
 							match_result = m&(1<<(c&7));
 						}
 						input_consumed = true;
-						j->citer += 32;
+						ctx[j].code_iter += 32;
 					}
 					else if(instr==MATCH_ANY)
 					{
@@ -481,25 +492,34 @@ bool Regex::run(const string &str, const string::const_iterator &begin, vector<R
 						throw internal_error("invalid instruction in regex bytecode");
 
 					if(match_result==negate_match)
-						terminate = true;
+						ctx[j].code_iter = code.end();
 					negate_match = false;
 
-					if(input_consumed || terminate)
+					if(input_consumed)
 						break;
 				}
 			}
+		}
 
-			if(terminate || j->citer==code.end())
-				j = ctx.erase(j);
+		for(size_t j=0; j<ctx_count; )
+		{
+			if(ctx[j].code_iter==code.end())
+			{
+				if(j!=ctx_count-1)
+					swap(ctx[j], ctx[ctx_count-1]);
+				--ctx_count;
+			}
 			else
 				++j;
 		}
 
-		if(i==str.end() || ctx.empty())
+		if(i==str.end() || !ctx_count)
 			break;
-		++i;
 	}
 
+	const RegMatch::Group *best_ptr = groups.data()+best_groups;
+	copy(best_ptr, best_ptr+out_groups.size(), out_groups.begin());
+
 	return result;
 }
 
diff --git a/source/strings/regex.h b/source/strings/regex.h
index 42c1e01..e0ef08d 100644
--- a/source/strings/regex.h
+++ b/source/strings/regex.h
@@ -102,8 +102,8 @@ private:
 
 	struct RunContext
 	{
-		Code::const_iterator citer;
-		std::vector<RegMatch::Group> groups;
+		Code::const_iterator code_iter;
+		std::size_t groups_index = 0;
 	};
 
 	Code code;