some peepholes that should match horizontal add/sub operations.

llvm-svn: 163103

llvm/lib/Target/X86/README-SSE.txt

@@ -941,3 +941,15 @@ and inversion with an rsqrtss instruction, which computes 1/sqrt faster at the
 cost of reduced accuracy.
 
 //===---------------------------------------------------------------------===//
+
+This function should be matched to haddpd when the appropriate CPU is enabled:
+
+#include <x86intrin.h>
+double f (__m128d p) {
+  return p[0] + p[1];
+}
+
+similarly, v[0]-v[1] should match to hsubpd, and {v[0]-v[1], w[0]-w[1]} should
+turn into hsubpd also.
+
+//===---------------------------------------------------------------------===//