[libcxx-commits] [PATCH] D69459: Optimize std::midpoint for integers

[Jorg Brown via Phabricator via libcxx-commits]

jorgbrown created this revision.
jorgbrown added a reviewer: mclow.lists.
Herald added subscribers: libcxx-commits, ldionne.

Same idea as the current algorithm, that is, add (half of the difference between a and b) to a.

But we use a different technique for computing the difference: we compute b - a into a pair of integers that are named "__sign_bit" and "__diff".  We have to use a pair because subtracting two 32-bit integers produces a 33-bit result.

Computing half of that is a simple matter of shifting __diff right by 1, and adding __sign_bit shifted left by 31.  llvm knows how to do that with one instruction: shld.

The only tricky part is that if the difference is odd and negative, then shifting it by one isn't the same as dividing it by two - shifting a negative one produces a negative one, for example.  So there's one more adjustment: if the sign bit and the low bit of diff are one, we add one.

For a demonstration of the codegen difference, see https://godbolt.org/z/7ar3K9 , which also has a built-in test.


Repository:
  rCXX libc++

https://reviews.llvm.org/D69459

Files:
  include/numeric


Index: include/numeric
===================================================================
--- include/numeric
+++ include/numeric
@@ -532,17 +532,14 @@
 _LIBCPP_DISABLE_UBSAN_UNSIGNED_INTEGER_CHECK
 {
     using _Up = std::make_unsigned_t<_Tp>;
+    constexpr _Up __bitshift = std::numeric_limits<_Up>::digits - 1;
 
-    int __sign = 1;
-    _Up __m = __a;
-    _Up __M = __b;
-    if (__a > __b)
-    {
-        __sign = -1;
-        __m = __b;
-        __M = __a;
-    }
-     return __a + __sign * _Tp(_Up(__M-__m) >> 1);
+    _Up __diff = _Up(__b) - _Up(__a);
+    _Up __sign_bit = __b < __a;
+
+    _Up __half_diff = (__diff / 2) + (__sign_bit << __bitshift) + (__sign_bit & __diff);
+
+    return __a + __half_diff;
 }
 
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D69459.226530.patch
Type: text/x-patch
Size: 731 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/libcxx-commits/attachments/20191026/792e6281/attachment.bin>