[ARM] Prefer lsls+lsrs over lsls+ands or lsrs+ands in Thumb1.

Saves materializing the immediate for the "ands".

Corresponding patterns exist for lsrs+lsls, but that seems less common
in practice.

Now implemented as a DAGCombine.

Differential Revision: https://reviews.llvm.org/D49585

llvm-svn: 337945

1 files changed, 81 insertions, 0 deletions

diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index 2e0f6096e43..47222a66f79 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -10676,6 +10676,83 @@ static SDValue PerformMULCombine(SDNode *N,
   return SDValue();
 }
 
+static SDValue CombineANDShift(SDNode *N,
+                               TargetLowering::DAGCombinerInfo &DCI,
+                               const ARMSubtarget *Subtarget) {
+  // Allow DAGCombine to pattern-match before we touch the canonical form.
+  if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
+    return SDValue();
+
+  if (N->getValueType(0) != MVT::i32)
+    return SDValue();
+
+  ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N->getOperand(1));
+  if (!N1C)
+    return SDValue();
+
+  uint32_t C1 = (uint32_t)N1C->getZExtValue();
+  // Don't transform uxtb/uxth.
+  if (C1 == 255 || C1 == 65535)
+    return SDValue();
+
+  SDNode *N0 = N->getOperand(0).getNode();
+  if (!N0->hasOneUse())
+    return SDValue();
+
+  if (N0->getOpcode() != ISD::SHL && N0->getOpcode() != ISD::SRL)
+    return SDValue();
+
+  bool LeftShift = N0->getOpcode() == ISD::SHL;
+
+  ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0->getOperand(1));
+  if (!N01C)
+    return SDValue();
+
+  uint32_t C2 = (uint32_t)N01C->getZExtValue();
+  if (!C2 || C2 >= 32)
+    return SDValue();
+
+  SelectionDAG &DAG = DCI.DAG;
+  SDLoc DL(N);
+
+  // We have a pattern of the form "(and (shl x, c2) c1)" or
+  // "(and (srl x, c2) c1)", where c1 is a shifted mask. Try to
+  // transform to a pair of shifts, to save materializing c1.
+
+  // First pattern: right shift, and c1+1 is a power of two.
+  // FIXME: Also check reversed pattern (left shift, and ~c1+1 is a power
+  // of two).
+  // FIXME: Use demanded bits?
+  if (!LeftShift && isMask_32(C1)) {
+    uint32_t C3 = countLeadingZeros(C1);
+    if (C2 < C3) {
+      SDValue SHL = DAG.getNode(ISD::SHL, DL, MVT::i32, N0->getOperand(0),
+                                DAG.getConstant(C3 - C2, DL, MVT::i32));
+      return DAG.getNode(ISD::SRL, DL, MVT::i32, SHL,
+                         DAG.getConstant(C3, DL, MVT::i32));
+    }
+  }
+
+  // Second pattern: left shift, and (c1>>c2)+1 is a power of two.
+  // FIXME: Also check reversed pattern (right shift, and ~(c1<<c2)+1
+  // is a power of two).
+  // FIXME: Use demanded bits?
+  if (LeftShift && isShiftedMask_32(C1)) {
+    uint32_t C3 = countLeadingZeros(C1);
+    if (C2 + C3 < 32 && C1 == ((-1U << (C2 + C3)) >> C3)) {
+      SDValue SHL = DAG.getNode(ISD::SHL, DL, MVT::i32, N0->getOperand(0),
+                                DAG.getConstant(C2 + C3, DL, MVT::i32));
+      return DAG.getNode(ISD::SRL, DL, MVT::i32, SHL,
+                        DAG.getConstant(C3, DL, MVT::i32));
+    }
+  }
+
+  // FIXME: Transform "(and (shl x, c2) c1)" ->
+  // "(shl (and x, c1>>c2), c2)" if "c1 >> c2" is a cheaper immediate than
+  // c1.
+  return SDValue();
+}
+
 static SDValue PerformANDCombine(SDNode *N,
                                  TargetLowering::DAGCombinerInfo &DCI,
                                  const ARMSubtarget *Subtarget) {
@@ -10717,6 +10794,10 @@ static SDValue PerformANDCombine(SDNode *N,
       return Result;
   }
 
+  if (Subtarget->isThumb1Only())
+    if (SDValue Result = CombineANDShift(N, DCI, Subtarget))
+      return Result;
+
   return SDValue();
 }