[clang] [llvm] [PowerPC] Support local-dynamic TLS relocation on AIX (PR #66316)

[Amy Kwan via llvm-commits]

================
@@ -145,9 +164,94 @@ namespace {
                                                               .addImm(0);
 
         if (IsAIX) {
-          // The variable offset and region handle are copied in r4 and r3. The
-          // copies are followed by GETtlsADDR32AIX/GETtlsADDR64AIX.
-          if (!IsTLSTPRelMI) {
+          if (IsTLSLDAIXMI) {
+            // The relative order between the LoadOffset at toc node (for the
+            // variable offset), and the .__tls_get_mod node is being tuned
+            // here. It is better to put the LoadOffset at toc node after the call,
+            // since the LoadOffset at toc node can use clobbers r4/r5. Search for
+            // the pattern of two Load at toc nodes (either for the variable offset
+            // or for the module handle), and then move the LoadOffset at toc node
+            // right before the node that uses the OutReg of the .__tls_get_mod
+            // node.
+            unsigned LDTocOp =
+                Is64Bit ? (IsLargeModel ? PPC::LDtocL : PPC::LDtoc)
+                        : (IsLargeModel ? PPC::LWZtocL : PPC::LWZtoc);
+            if (!RegInfo.use_empty(OutReg)) {
+              std::set<MachineInstr *> Uses;
+              // Collect all instructions that use the OutReg.
+              for (MachineOperand &MO : RegInfo.use_operands(OutReg))
+                Uses.insert(MO.getParent());
+              // Find the first user (e.g.: lwax/stfdx) of the OutReg within the
+              // current BB.
+              MachineBasicBlock::iterator UseIter = MBB.begin();
+              for (MachineBasicBlock::iterator IE = MBB.end(); UseIter != IE;
+                   ++UseIter)
+                if (Uses.count(&*UseIter))
+                  break;
+
+              // Additional handling is required when UserIter (the first user
+              // of OutReg) is pointing to a valid node. Check the pattern and
+              // do the movement if the pattern matches.
+              if (UseIter != MBB.end()) {
+                // Collect associated Load at toc nodes. Use hasOneDef() to guard
+                // against unexpected scenarios.
+                std::set<MachineInstr *> LoadFromTocs;
+                for (MachineOperand &MO : UseIter->operands())
+                  if (MO.isReg() && MO.isUse()) {
+                    Register MOReg = MO.getReg();
+                    if (RegInfo.hasOneDef(MOReg)) {
+                      MachineInstr *Temp =
+                          RegInfo.getOneDef(MOReg)->getParent();
+                      // For the current TLSLDAIX node, get the Load at toc node
+                      // for the InReg. Otherwise, Temp probably pointed to the
+                      // LoadOffset at toc node that we would like to move.
+                      if (Temp == &MI && RegInfo.hasOneDef(InReg))
+                        Temp = RegInfo.getOneDef(InReg)->getParent();
+                      if (Temp->getOpcode() == LDTocOp)
+                        LoadFromTocs.insert(Temp);
+                    } else {
+                      // FIXME: analyze this scenario if there is one.
+                      LoadFromTocs.clear();
+                      break;
+                    }
+                  }
+
+                // Check the two Load at toc: one should be _$TLSML, and the other
+                // will be moved before the node that uses the OutReg of the
+                // .__tls_get_mod node.
+                if (LoadFromTocs.size() == 2) {
+                  MachineBasicBlock::iterator TLSMLIter = MBB.end();
+                  MachineBasicBlock::iterator OffsetIter = MBB.end();
+                  // Make sure the two LoadFromTocs are within current BB, and
+                  // one of them from the "_$TLSML" pseudo symbol, while the
+                  // other from the variable.
+                  for (MachineBasicBlock::iterator I = MBB.begin(),
+                                                   IE = MBB.end();
+                       I != IE; ++I)
+                    if (LoadFromTocs.count(&*I)) {
+                      MachineOperand MO = I->getOperand(1);
+                      if (MO.isGlobal() && MO.getGlobal()->hasName() &&
+                          MO.getGlobal()->getName() == "_$TLSML")
+                        TLSMLIter = I;
+                      else
+                        OffsetIter = I;
+                    }
+                  // If both two iterators are valid, we should have identified
+                  // the scenario, and do the movement.
+                  if (TLSMLIter != MBB.end() && OffsetIter != MBB.end())
+                    OffsetIter->moveBefore(&*UseIter);
+                }
+              }
+            }
+            // The module-handle is copied into r3. The copy is followed by
+            // GETtlsMOD32AIX/GETtlsMOD64AIX.
+            BuildMI(MBB, I, DL, TII->get(TargetOpcode::COPY), GPR3)
+                .addReg(InReg);
+            // The call to .__tls_get_mod.
+            BuildMI(MBB, I, DL, TII->get(Opc2), GPR3).addReg(GPR3);
+          } else if (!IsTLSTPRelMI) {
+            // The variable offset and region handle are copied in r4 and r3.
----------------
amy-kwan wrote:

nit: We use module handle in the other places but region handle here. I think we should use the same terminology.

https://github.com/llvm/llvm-project/pull/66316