[AArch64] Swap 'lsl(val1,small-shmt)' to right hand side for ADD(lsl(val1,small-shmt), lsl(val2,large-shmt))

[mingmingl]

Summary

On many aarch64 processors (Cortex A78, Neoverse N1/N2/V1, etc), ADD with LSL shift (shift-amount <= 4) has smaller latency and higher
throughput than ADD with larger shift (shift-amunt > 4). This is at least no-op for the rest of the processors.

[mingmingl]

Run 'clang-format' and fix typo (AND->ADD)

[dmgreen]

Sorry for the delay. This sounds OK to me. Like you say, it could be based on LSLFast, but it shouldn't be worse for other CPUs and would be worthwhile for generic at least.

Other than some nitpicks, LGTM

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

16618

AND -> ADD

16619

> 4

16620

correctness ;)

[mingmingl]

Address the comments.

[mingmingl]

Thanks for the reviews!

[mingmingl]

Fix the typo (ADDD -> ADD)

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

EVT AArch64TargetLowering::getSetCCResultType(const DataLayout &,
                                               LLVMContext &C, EVT VT) const {
   if (!VT.isVector())
     return MVT::i32;
   if (VT.isScalableVector())
     return EVT::getVectorVT(C, MVT::i1, VT.getVectorElementCount());
   return VT.changeVectorElementTypeToInteger();
 }
 
+// isIntImmediate - This method tests to see if the node is a constant
+// operand. If so Imm will receive the value.
+static bool isIntImmediate(const SDNode *N, uint64_t &Imm) {
+  if (const ConstantSDNode *C = dyn_cast<const ConstantSDNode>(N)) {
+    Imm = C->getZExtValue();
+    return true;
+  }
+  return false;
+}
+
+// isOpcWithIntImmediate - This method tests to see if the node is a specific
+// opcode and that it has a immediate integer right operand.
+// If so Imm will receive the value.
+static bool isOpcWithIntImmediate(const SDNode *N, unsigned Opc,
+                                  uint64_t &Imm) {
+  return N->getOpcode() == Opc &&
+         isIntImmediate(N->getOperand(1).getNode(), Imm);
+}
+
 static bool optimizeLogicalImm(SDValue Op, unsigned Size, uint64_t Imm,
                                const APInt &Demanded,
                                TargetLowering::TargetLoweringOpt &TLO,
                                unsigned NewOpc) {
   uint64_t OldImm = Imm, NewImm, Enc;
   uint64_t Mask = ((uint64_t)(-1LL) >> (64 - Size)), OrigMask = Mask;
 
   // Return if the immediate is already all zeros, all ones, a bimm32 or a
if (Elt0->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
     SDValue Ext = DAG.getNode(ISD::ANY_EXTEND, DL, ExtVT, VecToExtend);
     return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Ext,
                        SubvectorIdx);
   }
 
   return SDValue();
 }
 
+static SDValue performAddCombineForShiftedOperands(SDNode *N,
+                                                   SelectionDAG &DAG) {
+  // NOTE: Swapping LHS and RHS is not done for SUB, since SUB is not
+  // commutative.
+  if (N->getOpcode() != ISD::ADD)
+    return SDValue();
+
+  // Bail out when value type is not one of {i32, i64}, since AArch64 ADD with
+  // shifted register is only available for i32 and i64.
+  EVT VT = N->getValueType(0);
+  if (VT != MVT::i32 && VT != MVT::i64)
+    return SDValue();
+
+  SDLoc DL(N);
+  SDValue LHS = N->getOperand(0);
+  SDValue RHS = N->getOperand(1);
+
+  uint64_t LHSImm = 0, RHSImm = 0;
+  // If both operand are shifted by imm and shift amount is not greater than 4
+  // for one operand, swap LHS and RHS to put operand with smaller shift amount
+  // on RHS.
+  //
+  // On many AArch64 processors (Cortex A78, Neoverse N1/N2/V1, etc), ADD with
+  // LSL shift (shift <= 4) has smaller latency and larger throughput than ADD
+  // with LSL (shift > 4). For the rest of processors, this is no-op for
+  // performance or correctness.
+  if (isOpcWithIntImmediate(LHS.getNode(), ISD::SHL, LHSImm) &&
+      isOpcWithIntImmediate(RHS.getNode(), ISD::SHL, RHSImm) && LHSImm <= 4 &&
+      RHSImm > 4 && LHS.hasOneUse())
+    return DAG.getNode(ISD::ADD, DL, VT, RHS, LHS);
+
+  return SDValue();
+}
+
 static SDValue performAddSubCombine(SDNode *N,
                                     TargetLowering::DAGCombinerInfo &DCI,
                                     SelectionDAG &DAG) {
   // Try to change sum of two reductions.
   if (SDValue Val = performAddUADDVCombine(N, DAG))
     return Val;
   if (SDValue Val = performAddDotCombine(N, DAG))
     return Val;
   if (SDValue Val = performAddCSelIntoCSinc(N, DAG))
     return Val;
   if (SDValue Val = performNegCSelCombine(N, DAG))
     return Val;
   if (SDValue Val = performVectorAddSubExtCombine(N, DAG))
     return Val;
+  if (SDValue Val = performAddCombineForShiftedOperands(N, DAG))
+    return Val;
 
   return performAddSubLongCombine(N, DCI, DAG);
 }
 
 // Massage DAGs which we can use the high-half "long" operations on into
 // something isel will recognize better. E.g.
 //
 // (aarch64_neon_umull (extract_high vec) (dupv64 scalar)) -->

llvm/test/CodeGen/AArch64/logical_shifted_reg.ll

other_exit:
   ret void
 ret:
   ret void
 }
 
 define i64 @add_swap_rhs_lhs_i64(i64 %0, i64 %1) {
 ; CHECK-LABEL: add_swap_rhs_lhs_i64:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    lsl x8, x1, #3
-; CHECK-NEXT:    add x0, x8, x0, lsl #8
+; CHECK-NEXT:    lsl x8, x0, #8
+; CHECK-NEXT:    add x0, x8, x1, lsl #3
 ; CHECK-NEXT:    ret
   %3 = shl i64 %0, 8
   %4 = shl i64 %1, 3
   %5 = add i64 %4, %3
   ret i64 %5
 }
 
 define i64 @add_swap_no_op_i64(i64 %0, i64 %1, i64* %2) {
 ; CHECK-LABEL: add_swap_no_op_i64:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    lsl x8, x1, #3
 ; CHECK-NEXT:    add x0, x8, x0, lsl #8
 ; CHECK-NEXT:    str x8, [x2]
 ; CHECK-NEXT:    ret
   %4 = shl i64 %0, 8
   %5 = shl i64 %1, 3
   store i64 %5, i64* %2
   %6 = add i64 %5, %4
   ret i64 %6
 }
 
 define i32 @add_swap_rhs_lhs_i32(i32 %0, i32 %1) {
 ; CHECK-LABEL: add_swap_rhs_lhs_i32:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    lsl w8, w1, #3
-; CHECK-NEXT:    add w0, w8, w0, lsl #8
+; CHECK-NEXT:    lsl w8, w0, #8
+; CHECK-NEXT:    add w0, w8, w1, lsl #3
 ; CHECK-NEXT:    ret
   %3 = shl i32 %0, 8
   %4 = shl i32 %1, 3
   %5 = add i32 %4, %3
   ret i32 %5
 }
 
 define i32 @add_swap_no_op_i32(i32 %0, i32 %1, i32* %2) {