Merging r330037:

------------------------------------------------------------------------
r330037 | sdardis | 2018-04-13 09:09:07 -0700 (Fri, 13 Apr 2018) | 21 lines

[mips] Materialize constants for multiplication

Previously, the MIPS backend would alwyas break down constant multiplications
into a series of shifts, adds, and subs. This patch changes that so the cost of
doing so is estimated.

The cost is estimated against worst case constant materialization and retrieving
the results from the HI/LO registers.

For cases where the value type of the multiplication is not legal, the cost of
legalization is estimated and is accounted for before performing the
optimization of breaking down the constant

This resolves PR36884.

Thanks to npl for reporting the issue!

Reviewers: abeserminji, smaksimovic

Differential Revision: https://reviews.llvm.org/D45316

------------------------------------------------------------------------

llvm-svn: 332782

llvm/lib/Target/Mips/MipsSEISelLowering.cpp

@@ -701,6 +701,77 @@ static SDValue performORCombine(SDNode *N, SelectionDAG &DAG,
   return SDValue();
 }
 
+static bool shouldTransformMulToShiftsAddsSubs(APInt C, EVT VT,
+                                               SelectionDAG &DAG,
+                                               const MipsSubtarget &Subtarget) {
+  // Estimate the number of operations the below transform will turn a
+  // constant multiply into. The number is approximately how many powers
+  // of two summed together that the constant can be broken down into.
+
+  SmallVector<APInt, 16> WorkStack(1, C);
+  unsigned Steps = 0;
+  unsigned BitWidth = C.getBitWidth();
+
+  while (!WorkStack.empty()) {
+    APInt Val = WorkStack.pop_back_val();
+
+    if (Val == 0 || Val == 1)
+      continue;
+
+    if (Val.isPowerOf2()) {
+      ++Steps;
+      continue;
+    }
+
+    APInt Floor = APInt(BitWidth, 1) << Val.logBase2();
+    APInt Ceil = Val.isNegative() ? APInt(BitWidth, 0)
+                                  : APInt(BitWidth, 1) << C.ceilLogBase2();
+
+    if ((Val - Floor).ule(Ceil - Val)) {
+      WorkStack.push_back(Floor);
+      WorkStack.push_back(Val - Floor);
+      ++Steps;
+      continue;
+    }
+
+    WorkStack.push_back(Ceil);
+    WorkStack.push_back(Ceil - Val);
+    ++Steps;
+
+    // If we have taken more than 12[1] / 8[2] steps to attempt the
+    // optimization for a native sized value, it is more than likely that this
+    // optimization will make things worse.
+    //
+    // [1] MIPS64 requires 6 instructions at most to materialize any constant,
+    //     multiplication requires at least 4 cycles, but another cycle (or two)
+    //     to retrieve the result from the HI/LO registers.
+    //
+    // [2] For MIPS32, more than 8 steps is expensive as the constant could be
+    //     materialized in 2 instructions, multiplication requires at least 4
+    //     cycles, but another cycle (or two) to retrieve the result from the
+    //     HI/LO registers.
+
+    if (Steps > 12 && (Subtarget.isABI_N32() || Subtarget.isABI_N64()))
+      return false;
+
+    if (Steps > 8 && Subtarget.isABI_O32())
+      return false;
+  }
+
+  // If the value being multiplied is not supported natively, we have to pay
+  // an additional legalization cost, conservatively assume an increase in the
+  // cost of 3 instructions per step. This values for this heuristic were
+  // determined experimentally.
+  unsigned RegisterSize = DAG.getTargetLoweringInfo()
+                              .getRegisterType(*DAG.getContext(), VT)
+                              .getSizeInBits();
+  Steps *= (VT.getSizeInBits() != RegisterSize) * 3;
+  if (Steps > 27)
+    return false;
+
+  return true;
+}
+
 static SDValue genConstMult(SDValue X, APInt C, const SDLoc &DL, EVT VT,
                             EVT ShiftTy, SelectionDAG &DAG) {
   // Return 0.
@@ -739,11 +810,13 @@ static SDValue genConstMult(SDValue X, APInt C, const SDLoc &DL, EVT VT,
 
 static SDValue performMULCombine(SDNode *N, SelectionDAG &DAG,
                                  const TargetLowering::DAGCombinerInfo &DCI,
-                                 const MipsSETargetLowering *TL) {
+                                 const MipsSETargetLowering *TL,
+                                 const MipsSubtarget &Subtarget) {
   EVT VT = N->getValueType(0);
 
   if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)))
-    if (!VT.isVector())
+    if (!VT.isVector() && shouldTransformMulToShiftsAddsSubs(
+                              C->getAPIntValue(), VT, DAG, Subtarget))
       return genConstMult(N->getOperand(0), C->getAPIntValue(), SDLoc(N), VT,
                           TL->getScalarShiftAmountTy(DAG.getDataLayout(), VT),
                           DAG);
@@ -983,7 +1056,7 @@ MipsSETargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const {
     Val = performORCombine(N, DAG, DCI, Subtarget);
     break;
   case ISD::MUL:
-    return performMULCombine(N, DAG, DCI, this);
+    return performMULCombine(N, DAG, DCI, this, Subtarget);
   case ISD::SHL:
     Val = performSHLCombine(N, DAG, DCI, Subtarget);
     break;