[DAG] ExpandIntRes_ADDSUB - create UADDO/USUBO instead of ADDCARRY/SUBCARRY if overflow is known to be zero

As noticed on D127115, when splitting ADD/SUB nodes we often end up with cases where overflow from the lower bits is impossible - in such cases we're better off breaking the carry chain dependency as soon as possible.

This path is being exercised by llvm/test/CodeGen/ARM/dsp-mlal.ll, although I haven't been able to get any codegen diff without a topological worklist.

1 files changed, 6 insertions, 2 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
index 282cc5f9ae38..228d4a43ccde 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
@@ -2912,11 +2912,15 @@ void DAGTypeLegalizer::ExpandIntRes_ADDSUB(SDNode *N,
     if (N->getOpcode() == ISD::ADD) {
       Lo = DAG.getNode(ISD::UADDO, dl, VTList, LoOps);
       HiOps[2] = Lo.getValue(1);
-      Hi = DAG.getNode(ISD::ADDCARRY, dl, VTList, HiOps);
+      Hi = DAG.computeKnownBits(HiOps[2]).isZero()
+               ? DAG.getNode(ISD::UADDO, dl, VTList, makeArrayRef(HiOps, 2))
+               : DAG.getNode(ISD::ADDCARRY, dl, VTList, HiOps);
     } else {
       Lo = DAG.getNode(ISD::USUBO, dl, VTList, LoOps);
       HiOps[2] = Lo.getValue(1);
-      Hi = DAG.getNode(ISD::SUBCARRY, dl, VTList, HiOps);
+      Hi = DAG.computeKnownBits(HiOps[2]).isZero()
+               ? DAG.getNode(ISD::USUBO, dl, VTList, makeArrayRef(HiOps, 2))
+               : DAG.getNode(ISD::SUBCARRY, dl, VTList, HiOps);
     }
     return;
   }