Observation of rematerialization.
authorEvan Cheng <evan.cheng@apple.com>
Mon, 10 Sep 2007 22:11:18 +0000 (22:11 +0000)
committerEvan Cheng <evan.cheng@apple.com>
Mon, 10 Sep 2007 22:11:18 +0000 (22:11 +0000)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@41809 91177308-0d34-0410-b5e6-96231b3b80d8

lib/CodeGen/README.txt

index 4fff15c1e3fce8c652b105c78b0198262baabce4..87671913209cfb82c6331f6ac4adf7147d4a59a2 100644 (file)
@@ -143,3 +143,8 @@ load [T + 7]
 ...
 load [T + 15]
 //===---------------------------------------------------------------------===//
+
+It's not always a good idea to choose rematerialization over spilling. If all
+the load / store instructions would be folded then spilling is cheaper because
+it won't require new live intervals / registers. See 2003-05-31-LongShifts for
+an example.