[llvm] [AMDGPU] Fix VOP3P NEG fold to check the use opcode, not the def (PR #219728)

[Arseniy Obolenskiy via llvm-commits]

================
@@ -5283,16 +5300,19 @@ static bool isValidToPack(SrcStatus HiStat, SrcStatus LoStat, Register NewReg,
          IsHalfState(HiStat);
 }
 
-std::pair<Register, unsigned> AMDGPUInstructionSelector::selectVOP3PModsImpl(
-    Register RootReg, const MachineRegisterInfo &MRI, bool IsDOT) const {
+std::pair<Register, unsigned>
+AMDGPUInstructionSelector::selectVOP3PModsImpl(const MachineOperand &Root,
+                                               const MachineRegisterInfo &MRI,
+                                               bool IsDOT) const {
+  Register RootReg = Root.getReg();
   unsigned Mods = 0;
   // No modification if Root type is not form of <2 x Type>.
   if (isVectorOfTwoOrScalar(RootReg, MRI) != TypeClass::VECTOR_OF_TWO) {
     Mods |= SISrcMods::OP_SEL_1;
     return {RootReg, Mods};
   }
 
-  SearchOptions SO(RootReg, MRI);
+  SearchOptions SO(*Root.getParent());
----------------
aobolensk wrote:

reverted

https://github.com/llvm/llvm-project/pull/219728