Fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n)), to simplify some code

that SCEVExpander can produce when running on behalf of LSR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@93949 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 7fe908c730248d907fda34c94ece0e972369987e..8883064df999c71012128a455f714aa8f64d73a1 100644 (file)
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -1088,6 +1088,26 @@ SDValue DAGCombiner::visitADD(SDNode *N) {
     if (Result.getNode()) return Result;
   }
 
+  // fold (add x, shl(0 - y, n)) -> sub(x, shl(y, n))
+  if (N1.getOpcode() == ISD::SHL &&
+      N1.getOperand(0).getOpcode() == ISD::SUB)
+    if (ConstantSDNode *C =
+          dyn_cast<ConstantSDNode>(N1.getOperand(0).getOperand(0)))
+      if (C->getAPIntValue() == 0)
+        return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N0,
+                           DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
+                                       N1.getOperand(0).getOperand(1),
+                                       N1.getOperand(1)));
+  if (N0.getOpcode() == ISD::SHL &&
+      N0.getOperand(0).getOpcode() == ISD::SUB)
+    if (ConstantSDNode *C =
+          dyn_cast<ConstantSDNode>(N0.getOperand(0).getOperand(0)))
+      if (C->getAPIntValue() == 0)
+        return DAG.getNode(ISD::SUB, N->getDebugLoc(), VT, N1,
+                           DAG.getNode(ISD::SHL, N->getDebugLoc(), VT,
+                                       N0.getOperand(0).getOperand(1),
+                                       N0.getOperand(1)));
+
   return SDValue();
 }
 

diff --git a/test/CodeGen/X86/neg-shl-add.ll b/test/CodeGen/X86/neg-shl-add.ll
new file mode 100644 (file)
index 0000000..7aebc38
--- /dev/null
+++ b/test/CodeGen/X86/neg-shl-add.ll
@@ -0,0 +1,17 @@
+; RUN: llc -march=x86-64 < %s | not grep negq
+
+; These sequences don't need neg instructions; they can be done with
+; a single shift and sub each.
+
+define i64 @foo(i64 %x, i64 %y, i64 %n) nounwind {
+  %a = sub i64 0, %y
+  %b = shl i64 %a, %n
+  %c = add i64 %b, %x
+  ret i64 %c
+}
+define i64 @boo(i64 %x, i64 %y, i64 %n) nounwind {
+  %a = sub i64 0, %y
+  %b = shl i64 %a, %n
+  %c = add i64 %x, %b
+  ret i64 %c
+}