[GlobalISel] combine not symmetric merge/unmerge nodes.

[igorb]

Summary

In some cases legalization ends up with not symmetric merge/unmerge nodes.
Transform it to merge/unmerge nodes.

[igorb]

ping

[igorb]

ping

[t.p.northover]

This looks reasonable to me.

[igorb]

Thanks for the review!

llvm/trunk/include/llvm/CodeGen/GlobalISel/Legalizer.h

MachineFunctionProperties getSetProperties() const override {
     return MachineFunctionProperties().set(
         MachineFunctionProperties::Property::Legalized);
   }
 
   bool combineExtracts(MachineInstr &MI, MachineRegisterInfo &MRI,
                        const TargetInstrInfo &TII);
 
   bool combineMerges(MachineInstr &MI, MachineRegisterInfo &MRI,
-                     const TargetInstrInfo &TII);
+                     const TargetInstrInfo &TII, MachineIRBuilder &MIRBuilder);
 
   bool runOnMachineFunction(MachineFunction &MF) override;
 };
 } // End namespace llvm.
 
 #endif

llvm/trunk/lib/CodeGen/GlobalISel/Legalizer.cpp

bool Legalizer::combineExtracts(MachineInstr &MI, MachineRegisterInfo &MRI,
   bool Changed = false;
   if (MI.getOpcode() != TargetOpcode::G_EXTRACT)
     return Changed;
 
   unsigned NumDefs = (MI.getNumOperands() - 1) / 2;
   unsigned SrcReg = MI.getOperand(NumDefs).getReg();
   MachineInstr &SeqI = *MRI.def_instr_begin(SrcReg);
   if (SeqI.getOpcode() != TargetOpcode::G_SEQUENCE)
-      return Changed;
 
   unsigned NumSeqSrcs = (SeqI.getNumOperands() - 1) / 2;
   bool AllDefsReplaced = true;
 
   // Try to match each register extracted with a corresponding insertion formed
   // by the G_SEQUENCE.
   for (unsigned Idx = 0, SeqIdx = 0; Idx < NumDefs; ++Idx) {
     MachineOperand &ExtractMO = MI.getOperand(Idx);
if (MRI.use_empty(SrcReg))
       SeqI.eraseFromParent();
     Changed = true;
   }
 
   return Changed;
 }
 
 bool Legalizer::combineMerges(MachineInstr &MI, MachineRegisterInfo &MRI,
-                              const TargetInstrInfo &TII) {
+                              const TargetInstrInfo &TII,
+                              MachineIRBuilder &MIRBuilder) {
   if (MI.getOpcode() != TargetOpcode::G_UNMERGE_VALUES)
     return false;
 
   unsigned NumDefs = MI.getNumOperands() - 1;
   unsigned SrcReg = MI.getOperand(NumDefs).getReg();
   MachineInstr &MergeI = *MRI.def_instr_begin(SrcReg);
   if (MergeI.getOpcode() != TargetOpcode::G_MERGE_VALUES)
     return false;
 
-  if (MergeI.getNumOperands() - 1 != NumDefs)
+  const unsigned NumMergeRegs = MergeI.getNumOperands() - 1;
+
+  if (NumMergeRegs < NumDefs) {
+    if (NumDefs % NumMergeRegs != 0)
+      return false;
+
+    MIRBuilder.setInstr(MI);
+    // Transform to UNMERGEs, for example
+    //   %1 = G_MERGE_VALUES %4, %5
+    //   %9, %10, %11, %12 = G_UNMERGE_VALUES %1
+    // to
+    //   %9, %10 = G_UNMERGE_VALUES %4
+    //   %11, %12 = G_UNMERGE_VALUES %5
+
+    const unsigned NewNumDefs = NumDefs / NumMergeRegs;
+    for (unsigned Idx = 0; Idx < NumMergeRegs; ++Idx) {
+      SmallVector<unsigned, 2> DstRegs;
+      for (unsigned j = 0, DefIdx = Idx * NewNumDefs; j < NewNumDefs;
+           ++j, ++DefIdx)
+        DstRegs.push_back(MI.getOperand(DefIdx).getReg());
+
+      MIRBuilder.buildUnmerge(DstRegs, MergeI.getOperand(Idx + 1).getReg());
+    }
+
+  } else if (NumMergeRegs > NumDefs) {
+    if (NumMergeRegs % NumDefs != 0)
       return false;
 
-  // FIXME: is a COPY appropriate if the types mismatch? We know both registers
-  // are allocatable by now.
+    MIRBuilder.setInstr(MI);
+    // Transform to MERGEs
+    //   %6 = G_MERGE_VALUES %17, %18, %19, %20
+    //   %7, %8 = G_UNMERGE_VALUES %6
+    // to
+    //   %7 = G_MERGE_VALUES %17, %18
+    //   %8 = G_MERGE_VALUES %19, %20
+
+    const unsigned NumRegs = NumMergeRegs / NumDefs;
+    for (unsigned DefIdx = 0; DefIdx < NumDefs; ++DefIdx) {
+      SmallVector<unsigned, 2> Regs;
+      for (unsigned j = 0, Idx = NumRegs * DefIdx + 1; j < NumRegs; ++j, ++Idx)
+        Regs.push_back(MergeI.getOperand(Idx).getReg());
+
+      MIRBuilder.buildMerge(MI.getOperand(DefIdx).getReg(), Regs);
+    }
+
+  } else {
+    // FIXME: is a COPY appropriate if the types mismatch? We know both
+    // registers are allocatable by now.
     if (MRI.getType(MI.getOperand(0).getReg()) !=
         MRI.getType(MergeI.getOperand(1).getReg()))
       return false;
 
     for (unsigned Idx = 0; Idx < NumDefs; ++Idx)
       MRI.replaceRegWith(MI.getOperand(Idx).getReg(),
                          MergeI.getOperand(Idx + 1).getReg());
+  }
 
   MI.eraseFromParent();
   if (MRI.use_empty(MergeI.getOperand(0).getReg()))
     MergeI.eraseFromParent();
   return true;
 }
 
 bool Legalizer::runOnMachineFunction(MachineFunction &MF) {
for (auto MI = MBB.begin(); MI != MBB.end(); MI = NextMI) {
       // good chance MI will be deleted.
       NextMI = std::next(MI);
 
       // combineExtracts erases MI.
       if (combineExtracts(*MI, MRI, TII)) {
         Changed = true;
         continue;
       }
-      Changed |= combineMerges(*MI, MRI, TII);
+      Changed |= combineMerges(*MI, MRI, TII, Helper.MIRBuilder);
     }
   }
 
   return Changed;
 }

llvm/trunk/test/CodeGen/X86/GlobalISel/legalize-add-v512.mir

registers:
   - { id: 5, class: _ }
   - { id: 6, class: _ }
   - { id: 7, class: _ }
   - { id: 8, class: _ }
 # AVX1:              %2(<32 x s8>) = COPY %ymm0
 # AVX1-NEXT:         %3(<32 x s8>) = COPY %ymm1
 # AVX1-NEXT:         %4(<32 x s8>) = COPY %ymm2
 # AVX1-NEXT:         %5(<32 x s8>) = COPY %ymm3
-# AVX1-NEXT:         %0(<64 x s8>) = G_MERGE_VALUES %2(<32 x s8>), %3(<32 x s8>)
-# AVX1-NEXT:         %1(<64 x s8>) = G_MERGE_VALUES %4(<32 x s8>), %5(<32 x s8>)
-# AVX1-NEXT:         %9(<16 x s8>), %10(<16 x s8>), %11(<16 x s8>), %12(<16 x s8>) = G_UNMERGE_VALUES %0(<64 x s8>)
-# AVX1-NEXT:         %13(<16 x s8>), %14(<16 x s8>), %15(<16 x s8>), %16(<16 x s8>) = G_UNMERGE_VALUES %1(<64 x s8>)
+# AVX1-NEXT:         %9(<16 x s8>), %10(<16 x s8>) = G_UNMERGE_VALUES %2(<32 x s8>)
+# AVX1-NEXT:         %11(<16 x s8>), %12(<16 x s8>) = G_UNMERGE_VALUES %3(<32 x s8>)
+# AVX1-NEXT:         %13(<16 x s8>), %14(<16 x s8>) = G_UNMERGE_VALUES %4(<32 x s8>)
+# AVX1-NEXT:         %15(<16 x s8>), %16(<16 x s8>) = G_UNMERGE_VALUES %5(<32 x s8>)
 # AVX1-NEXT:         %17(<16 x s8>) = G_ADD %9, %13
 # AVX1-NEXT:         %18(<16 x s8>) = G_ADD %10, %14
 # AVX1-NEXT:         %19(<16 x s8>) = G_ADD %11, %15
 # AVX1-NEXT:         %20(<16 x s8>) = G_ADD %12, %16
-# AVX1-NEXT:         %6(<64 x s8>) = G_MERGE_VALUES %17(<16 x s8>), %18(<16 x s8>), %19(<16 x s8>), %20(<16 x s8>)
-# AVX1-NEXT:         %7(<32 x s8>), %8(<32 x s8>) = G_UNMERGE_VALUES %6(<64 x s8>)
+# AVX1-NEXT:         %7(<32 x s8>) = G_MERGE_VALUES %17(<16 x s8>), %18(<16 x s8>)
+# AVX1-NEXT:         %8(<32 x s8>) = G_MERGE_VALUES %19(<16 x s8>), %20(<16 x s8>)
 # AVX1-NEXT:         %ymm0 = COPY %7(<32 x s8>)
 # AVX1-NEXT:         %ymm1 = COPY %8(<32 x s8>)
 # AVX1-NEXT:         RET 0, implicit %ymm0, implicit %ymm1
 #
 # AVX512F:           %2(<32 x s8>) = COPY %ymm0
 # AVX512F-NEXT:      %3(<32 x s8>) = COPY %ymm1
 # AVX512F-NEXT:      %4(<32 x s8>) = COPY %ymm2
 # AVX512F-NEXT:      %5(<32 x s8>) = COPY %ymm3