tiny fbstring push_back(Char) optimization

Summary:
I noticed that push_back(Char) was slower for fbstring than
for std::string for long strings and found that I could make
it faster by inlining the call to mutable_data() and
exploit that it's always non-small and non-shared.

Benchmarks before:

./folly/test/FBStringTestBenchmarks.cpp.h       relative  time/iter  iters/s
BM_push_back_string(1)                                      69.42ns   14.41M
BM_push_back_string(23)                                    582.31ns    1.72M
BM_push_back_string(127)                                     1.47us  682.12K
BM_push_back_string(1024)                                    5.52us  181.07K
BM_push_back_fbstring(1)                                     9.55ns  104.74M
BM_push_back_fbstring(23)                                  212.45ns    4.71M
BM_push_back_fbstring(127)                                 864.00ns    1.16M
BM_push_back_fbstring(1024)                                  6.73us  148.52K

and after:

BM_push_back_fbstring(1)                                     9.55ns  104.74M
BM_push_back_fbstring(23)                                  212.45ns    4.71M
BM_push_back_fbstring(127)                                 783.08ns    1.28M
BM_push_back_fbstring(1024)                                  4.03us  248.05K

Test Plan: fbconfig folly/test && fbmake runtests

Reviewed By: tudorb@fb.com

FB internal diff: D646081

diff --git a/folly/FBString.h b/folly/FBString.h
index ef58b16c2d2e61a1dde591d11b4b0a0c8b703a56..11ec56022146a790227a0766388bcc55d3f1ea61 100644 (file)
--- a/folly/FBString.h
+++ b/folly/FBString.h
@@ -641,7 +641,7 @@ public:
 
   void push_back(Char c) {
     assert(capacity() >= size());
-    size_t sz, cp;
+    size_t sz;
     if (category() == isSmall) {
       sz = smallSize();
       if (sz < maxSmallSize) {
@@ -653,14 +653,16 @@ public:
       reserve(maxSmallSize * 2);
     } else {
       sz = ml_.size_;
-      cp = capacity();  // != ml_.capacity() for isShared()
-      if (sz == cp) reserve(cp * 3 / 2);
+      if (sz == capacity()) {  // always true for isShared()
+        reserve(sz * 3 / 2);  // ensures not shared
+      }
     }
+    assert(!isShared());
     assert(capacity() >= sz + 1);
     // Category can't be small - we took care of that above
     assert(category() == isMedium || category() == isLarge);
     ml_.size_ = sz + 1;
-    mutable_data()[sz] = c;
+    ml_.data_[sz] = c;
     writeTerminator();
   }