[llvm-commits] [PATCH] Add DAGCombine for A-(-B) -> A+B

[Eli Friedman]

Attached.  I wanted to ask before committing because there might be
some non-obvious reason we don't already do this transformation...

This affects IR like the following:
define i32 @a(i32 %x, i32 %y) nounwind readnone {
entry:
  %mul = mul i32 %y, -8
  %sub = sub i32 %x, %mul
  ret i32 %sub
}

-Eli
-------------- next part --------------
Index: DAGCombiner.cpp
===================================================================
--- DAGCombiner.cpp	(revision 106979)
+++ DAGCombiner.cpp	(working copy)
@@ -1540,6 +1540,13 @@
       N0.getOperand(1).getOperand(1) == N1)
     return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT,
                        N0.getOperand(0), N0.getOperand(1).getOperand(0));
+  // fold A-(0-B) -> A+B
+  if (N1.getOpcode() == ISD::SUB)
+    if (ConstantSDNode *N01C =
+        dyn_cast<ConstantSDNode>(N1.getOperand(0).getNode()))
+      if (N01C->isNullValue())
+        return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, N0,
+                           N1.getOperand(1));
 
   // If either operand of a sub is undef, the result is undef
   if (N0.getOpcode() == ISD::UNDEF)