Don't include extension when measuring cur_file_prefix_len

Stripping the extension allows e.g. "a/foo.c" and "a/foo.h" to rank
equally for a current file of "foo.c" (or, more importantly, "a/foo.h"
to rank above "a/b/foo.c"). Retaining the "." ensures that "foo.cc" will
still rank above e.g. "foox.h".

README.md

@@ -22,7 +22,7 @@ cpsm is to enable a particular one based on CtrlP:
 
 4. Hit Enter to open the file you wanted in the current window.
 
-To achieve this, cpsm needs to deliver
+To achieve this, cpsm needs to deliver:
 
 - high quality search results (at sufficiently high levels of quality, it's
   possible to enter a short query, hit Enter without needing to look at and
@@ -33,7 +33,7 @@ To achieve this, cpsm needs to deliver
   common switching between files is)
 
 - with as little latency as possible (to support scaling to very large, and
-  especially very deeply nested, code bases with very long pathnames).
+  especially very deeply nested, code bases with very long pathnames)
 
 See the "Performance" section below for both search quality and time
 comparisons to other matchers.
@@ -168,40 +168,15 @@ Performance
 
   - Query "", current file "kernel/signal.c":
 
-    - cpsm: "arch/alpha/kernel/signal.c"; 4.936ms (15.769ms with 1 thread)
-
-    - "signal" is significantly more common (e.g. just about every arch has its
-      own signal.c), so this result is sane, but not likely to be useful.
+    - cpsm: "include/asm-generic/signal.c"; 5.048ms (16.013ms with 1 thread)
 
     - All others: same as above
 
-  - The next two cases simulate a user trying to get to files that are closely
-    related to the currently open file.
-
-  - Query ".h", current file "kernel/signal.c":
-
-    - cpsm: "include/asm-generic/signal.h"; 6.218ms (22.107ms with 1 thread)
-
-    - ctrlp-cmatcher: "fs/ext2/ext2.h"; 22.850ms
-
-    - ctrlp-py-matcher: "mm/slab.h"; 33.533ms
-
-    - ctrlp: "security/selinux/ss/sidtab.h"
-
-    - fzf: "Documentation/scsi/LICENSE.FlashPoint"
-
-    - Without using the current filename, there is nothing the other matchers
-      can do to disambiguate the query. (cpsm doesn't get what I would consider
-      the best match either - "include/linux/signal.h" - but it's impossible to
-      choose between these two results without knowledge of the Linux kernel's
-      source layout. cpsm does pick that file as the second-best match.) fzf's
-      top result seems to be particularly out of whack, since at least the
-      other matchers return a .h file, but this is perhaps understandable given
-      that fzf is the most generic matcher out of the group.
-
-    - cpsm is much faster than either of the other two benchmarkable matchers
-      with multithreading enabled, and competitive with ctrlp-cmatcher when
-      locked to a single thread.
+    - "signal" is a significantly more common prefix; cpsm doesn't get what I
+      would consider the best match ("include/linux/signal.h") but it's
+      impossible to choose between these two results without knowledge of the
+      Linux kernel's source layout. (cpsm does pick that file as the
+      second-best match.)
 
   - Query "x86/", current file "kernel/signal.c":
 
@@ -215,7 +190,8 @@ Performance
 
     - fzf: "Documentation/x86/early-microcode.txt"
 
-    - Similar story to the previous case.
+    - Without using the current filename, there is nothing the other matchers
+      can do to disambiguate the query.
 
   - The next set of cases simulate a user typing progressively more letters in
     a desired file's name ("include/linux/rcupdate.h"), when they happen to be
@@ -233,6 +209,10 @@ Performance
 
     - fzf: "CREDITS"
 
+    - cpsm is much faster than either of the other two benchmarkable matchers
+      with multithreading enabled, and competitive with ctrlp-cmatcher when
+      locked to a single thread.
+
   - Query "rc", current file "kernel/signal.c":
 
     - cpsm: "kernel/rcu/rcu.h"; 7.617ms (26.343ms with 1 thread)

src/matcher.cc

@@ -62,6 +62,14 @@ Matcher::Matcher(boost::string_ref const query, MatcherOpts opts,
                   [&](char32_t const c) { return strings_.is_uppercase(c); });
 
   cur_file_parts_ = path_components_of(opts_.cur_file);
+  if (!cur_file_parts_.empty()) {
+    cur_file_key_ = cur_file_parts_.back();
+    // Strip the extension from cur_file_key_, if any (but not the trailing .)
+    auto const ext_sep_pos = cur_file_key_.find_last_of('.');
+    if (ext_sep_pos != boost::string_ref::npos) {
+      cur_file_key_ = cur_file_key_.substr(0, ext_sep_pos + 1);
+    }
+  }
 }
 
 bool Matcher::match_base(boost::string_ref const item, MatchBase& m,
@@ -157,9 +165,8 @@ void Matcher::match_path(std::vector<boost::string_ref> const& item_parts,
   // We don't want to exclude cur_file as a match, but we also don't want it
   // to be the top match, so force cur_file_prefix_len to 0 for cur_file (i.e.
   // if path_distance is 0).
-  if (m.path_distance != 0 && !cur_file_parts_.empty() && !item_parts.empty()) {
-    m.cur_file_prefix_len =
-        common_prefix(cur_file_parts_.back(), item_parts.back());
+  if (m.path_distance != 0 && !item_parts.empty()) {
+    m.cur_file_prefix_len = common_prefix(cur_file_key_, item_parts.back());
   }
 }
 

src/matcher.h

@@ -85,6 +85,7 @@ class Matcher final {
   bool is_case_sensitive_;
   bool require_full_part_;
   std::vector<boost::string_ref> cur_file_parts_;
+  boost::string_ref cur_file_key_;
 };
 
 }  // namespace cpsm