From 5838354d342f1cdd09da7099e86123b36ecec409 Mon Sep 17 00:00:00 2001
From: pennae <github@quasiparticle.net>
Date: Mon, 27 Dec 2021 02:04:49 +0100
Subject: [PATCH] optimize ExprConcatStrings::eval
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

constructing an ostringstream for non-string concats (like integer addition) is
a small constant cost that we can avoid. for string concats we can keep all the
string temporaries we get from coerceToString and concatenate them in one go,
which saves a lot of intermediate temporaries and copies in ostringstream. we
can also avoid copying the concatenated string again by directly allocating it
in GC memory and moving ownership of the concatenated string into the target
value.

saves about 2% on system eval.

before:

  Benchmark 1: nix eval --raw --impure --expr 'with import <nixpkgs/nixos> {}; system'
    Time (mean ± σ):      2.837 s ±  0.031 s    [User: 2.562 s, System: 0.191 s]
    Range (min … max):    2.796 s …  2.892 s    20 runs

after:

  Benchmark 1: nix eval --raw --impure --expr 'with import <nixpkgs/nixos> {}; system'
    Time (mean ± σ):      2.790 s ±  0.035 s    [User: 2.532 s, System: 0.187 s]
    Range (min … max):    2.722 s …  2.836 s    20 runs
---
 src/libexpr/eval.cc  | 75 ++++++++++++++++++++++++++++++++++++--------
 src/libexpr/value.hh |  2 ++
 2 files changed, 64 insertions(+), 13 deletions(-)

diff --git a/src/libexpr/eval.cc b/src/libexpr/eval.cc
index 81aa86641..61bccd6e2 100644
--- a/src/libexpr/eval.cc
+++ b/src/libexpr/eval.cc
@@ -36,6 +36,19 @@
 namespace nix {
 
 
+static char * allocString(size_t size)
+{
+    char * t;
+#if HAVE_BOEHMGC
+    t = (char *) GC_MALLOC_ATOMIC(size);
+#else
+    t = malloc(size);
+#endif
+    if (!t) throw std::bad_alloc();
+    return t;
+}
+
+
 static char * dupString(const char * s)
 {
     char * t;
@@ -771,17 +784,28 @@ void Value::mkString(std::string_view s)
 }
 
 
+static void copyContextToValue(Value & v, const PathSet & context)
+{
+    if (!context.empty()) {
+        size_t n = 0;
+        v.string.context = (const char * *)
+            allocBytes((context.size() + 1) * sizeof(char *));
+        for (auto & i : context)
+            v.string.context[n++] = dupString(i.c_str());
+        v.string.context[n] = 0;
+    }
+}
+
 void Value::mkString(std::string_view s, const PathSet & context)
 {
     mkString(s);
-    if (!context.empty()) {
-        size_t n = 0;
-        string.context = (const char * *)
-            allocBytes((context.size() + 1) * sizeof(char *));
-        for (auto & i : context)
-            string.context[n++] = dupString(i.c_str());
-        string.context[n] = 0;
-    }
+    copyContextToValue(*this, context);
+}
+
+void Value::mkStringMove(const char * s, const PathSet & context)
+{
+    mkString(s);
+    copyContextToValue(*this, context);
 }
 
 
@@ -1660,13 +1684,34 @@ void EvalState::concatLists(Value & v, size_t nrLists, Value * * lists, const Po
 void ExprConcatStrings::eval(EvalState & state, Env & env, Value & v)
 {
     PathSet context;
-    std::ostringstream s;
+    std::vector<std::string> s;
+    size_t sSize = 0;
     NixInt n = 0;
     NixFloat nf = 0;
 
     bool first = !forceString;
     ValueType firstType = nString;
 
+    const auto str = [&] {
+        std::string result;
+        result.reserve(sSize);
+        for (const auto & part : s) result += part;
+        return result;
+    };
+    /* c_str() is not str().c_str() because we want to create a string
+       Value. allocating a GC'd string directly and moving it into a
+       Value lets us avoid an allocation and copy. */
+    const auto c_str = [&] {
+        char * result = allocString(sSize + 1);
+        char * tmp = result;
+        for (const auto & part : s) {
+            memcpy(tmp, part.c_str(), part.size());
+            tmp += part.size();
+        }
+        *tmp = 0;
+        return result;
+    };
+
     for (auto & [i_pos, i] : *es) {
         Value vTmp;
         i->eval(state, env, vTmp);
@@ -1696,11 +1741,15 @@ void ExprConcatStrings::eval(EvalState & state, Env & env, Value & v)
                 nf += vTmp.fpoint;
             } else
                 throwEvalError(i_pos, "cannot add %1% to a float", showType(vTmp));
-        } else
+        } else {
+            if (s.empty()) s.reserve(es->size());
             /* skip canonization of first path, which would only be not
             canonized in the first place if it's coming from a ./${foo} type
             path */
-            s << state.coerceToString(i_pos, vTmp, context, false, firstType == nString, !first);
+            s.emplace_back(
+                state.coerceToString(i_pos, vTmp, context, false, firstType == nString, !first));
+            sSize += s.back().size();
+        }
 
         first = false;
     }
@@ -1712,9 +1761,9 @@ void ExprConcatStrings::eval(EvalState & state, Env & env, Value & v)
     else if (firstType == nPath) {
         if (!context.empty())
             throwEvalError(pos, "a string that refers to a store path cannot be appended to a path");
-        v.mkPath(canonPath(s.str()));
+        v.mkPath(canonPath(str()));
     } else
-        v.mkString(s.str(), context);
+        v.mkStringMove(c_str(), context);
 }
 
 
diff --git a/src/libexpr/value.hh b/src/libexpr/value.hh
index 7d007ebdc..1896c7563 100644
--- a/src/libexpr/value.hh
+++ b/src/libexpr/value.hh
@@ -241,6 +241,8 @@ public:
 
     void mkString(std::string_view s, const PathSet & context);
 
+    void mkStringMove(const char * s, const PathSet & context);
+
     inline void mkString(const Symbol & s)
     {
         mkString(((const std::string &) s).c_str());