GlobalISel: Help reduce operation width for instruction with two results.

[cdevadas]

Summary

The function reduceOperationWidth helps to legalize a vector
operation either by narrowing its type or by scalarizing the
operation itself. It currently supports instructions with one result.
This patch, in addition, allows the same for instructions with two
results (for instance, G_SDIVREM).

[cdevadas]

Ping

[cdevadas]

Ping

[cdevadas]

Ping

[arsenm]

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

4050

This is assuming the two results have the same type, which I guess is fine for now

4063

You don't need to call build undef for each time. I think just append(Num, Reg) works?

[cdevadas]

Comment addressed.

[cdevadas]

Ping

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

 }
 
 LegalizerHelper::LegalizeResult
 LegalizerHelper::reduceOperationWidth(MachineInstr &MI, unsigned int TypeIdx,
                                       LLT NarrowTy) {
   assert(TypeIdx == 0 && "only one type index expected");
 
   const unsigned Opc = MI.getOpcode();
-  const int NumOps = MI.getNumOperands() - 1;
-  const Register DstReg = MI.getOperand(0).getReg();
+  const int NumDefOps = MI.getNumExplicitDefs();
+  const int NumSrcOps = MI.getNumOperands() - NumDefOps;
   const unsigned Flags = MI.getFlags();
   const unsigned NarrowSize = NarrowTy.getSizeInBits();
   const LLT NarrowScalarTy = LLT::scalar(NarrowSize);
 
-  assert(NumOps <= 3 && "expected instruction with 1 result and 1-3 sources");
+  assert(MI.getNumOperands() <= 4 && "expected instruction with either 1 "
+                                     "result and 1-3 sources or 2 results and "
+                                     "1-2 sources");
+
+  SmallVector<Register, 2> DstRegs;
+  for (int I = 0; I < NumDefOps; ++I)
+    DstRegs.push_back(MI.getOperand(I).getReg());
 
   // First of all check whether we are narrowing (changing the element type)
   // or reducing the vector elements
-  const LLT DstTy = MRI.getType(DstReg);
+  const LLT DstTy = MRI.getType(DstRegs[0]);
   const bool IsNarrow = NarrowTy.getScalarType() != DstTy.getScalarType();
 
   SmallVector<Register, 8> ExtractedRegs[3];
   SmallVector<Register, 8> Parts;
 
   unsigned NarrowElts = NarrowTy.isVector() ? NarrowTy.getNumElements() : 1;
 
   // Break down all the sources into NarrowTy pieces we can operate on. This may
   // involve creating merges to a wider type, padded with undef.
-  for (int I = 0; I != NumOps; ++I) {
-    Register SrcReg = MI.getOperand(I + 1).getReg();
+  for (int I = 0; I != NumSrcOps; ++I) {
+    Register SrcReg = MI.getOperand(I + NumDefOps).getReg();
     LLT SrcTy = MRI.getType(SrcReg);
 
     // The type to narrow SrcReg to. For narrowing, this is a smaller scalar.
     // For fewerElements, this is a smaller vector with the same element type.
     LLT OpNarrowTy;
     if (IsNarrow) {
       OpNarrowTy = NarrowScalarTy;
 
for (int I = 0; I != NumSrcOps; ++I) {
 
     LLT GCDTy = extractGCDType(ExtractedRegs[I], SrcTy, OpNarrowTy, SrcReg);
 
     // Build a sequence of NarrowTy pieces in ExtractedRegs for this operand.
     buildLCMMergePieces(SrcTy, OpNarrowTy, GCDTy, ExtractedRegs[I],
                         TargetOpcode::G_ANYEXT);
   }
 
-  SmallVector<Register, 8> ResultRegs;
+  SmallVector<Register, 8> ResultRegs[2];
 
   // Input operands for each sub-instruction.
-  SmallVector<SrcOp, 4> InputRegs(NumOps, Register());
+  SmallVector<SrcOp, 4> InputRegs(NumSrcOps, Register());
 
   int NumParts = ExtractedRegs[0].size();
   const unsigned DstSize = DstTy.getSizeInBits();
   const LLT DstScalarTy = LLT::scalar(DstSize);
 
   // Narrowing needs to use scalar types
   LLT DstLCMTy, NarrowDstTy;
   if (IsNarrow) {
     DstLCMTy = getLCMType(DstScalarTy, NarrowScalarTy);
     NarrowDstTy = NarrowScalarTy;
   } else {
     DstLCMTy = getLCMType(DstTy, NarrowTy);
     NarrowDstTy = NarrowTy;
   }
 
   // We widened the source registers to satisfy merge/unmerge size
   // constraints. We'll have some extra fully undef parts.
   const int NumRealParts = (DstSize + NarrowSize - 1) / NarrowSize;
 
   for (int I = 0; I != NumRealParts; ++I) {
     // Emit this instruction on each of the split pieces.
-    for (int J = 0; J != NumOps; ++J)
+    for (int J = 0; J != NumSrcOps; ++J)
       InputRegs[J] = ExtractedRegs[J][I];
 
-    auto Inst = MIRBuilder.buildInstr(Opc, {NarrowDstTy}, InputRegs, Flags);
-    ResultRegs.push_back(Inst.getReg(0));
+    MachineInstrBuilder Inst;
+    if (NumDefOps == 1)
+      Inst = MIRBuilder.buildInstr(Opc, {NarrowDstTy}, InputRegs, Flags);
+    else
+      Inst = MIRBuilder.buildInstr(Opc, {NarrowDstTy, NarrowDstTy}, InputRegs,
+                                   Flags);
+
+    for (int J = 0; J != NumDefOps; ++J)
+      ResultRegs[J].push_back(Inst.getReg(J));
   }
 
   // Fill out the widened result with undef instead of creating instructions
   // with undef inputs.
   int NumUndefParts = NumParts - NumRealParts;
-  if (NumUndefParts != 0)
-    ResultRegs.append(NumUndefParts,
+  if (NumUndefParts != 0) {
+    for (int I = 0; I != NumDefOps; ++I)
+      ResultRegs[I].append(NumUndefParts,
                            MIRBuilder.buildUndef(NarrowDstTy).getReg(0));
+  }
 
   // Extract the possibly padded result. Use a scratch register if we need to do
   // a final bitcast, otherwise use the original result register.
   Register MergeDstReg;
+  for (int I = 0; I != NumDefOps; ++I) {
     if (IsNarrow && DstTy.isVector())
       MergeDstReg = MRI.createGenericVirtualRegister(DstScalarTy);
     else
-    MergeDstReg = DstReg;
+      MergeDstReg = DstRegs[I];
 
-  buildWidenedRemergeToDst(MergeDstReg, DstLCMTy, ResultRegs);
+    buildWidenedRemergeToDst(MergeDstReg, DstLCMTy, ResultRegs[I]);
 
     // Recast to vector if we narrowed a vector
     if (IsNarrow && DstTy.isVector())
-    MIRBuilder.buildBitcast(DstReg, MergeDstReg);
+      MIRBuilder.buildBitcast(DstRegs[I], MergeDstReg);
+  }
 
   MI.eraseFromParent();
   return Legalized;
 }
 
 LegalizerHelper::LegalizeResult
 LegalizerHelper::fewerElementsVectorSextInReg(MachineInstr &MI, unsigned TypeIdx,
                                               LLT NarrowTy) {