[clang] [llvm] [PowerPC] Tune AIX shared library TLS model at function level by heuristic (PR #84132)

[Hubert Tong via llvm-commits]

================
@@ -3362,6 +3362,64 @@ SDValue PPCTargetLowering::LowerGlobalTLSAddress(SDValue Op,
   return LowerGlobalTLSAddressLinux(Op, DAG);
 }
 
+/// updateForAIXShLibTLSModelOpt - Helper to initialize TLS model opt settings,
+/// and then apply the update.
+static void updateForAIXShLibTLSModelOpt(TLSModel::Model &Model,
+                                         SelectionDAG &DAG,
+                                         const TargetMachine &TM) {
+  // Initialize TLS model opt setting lazily:
+  // (1) Use initial-exec for single TLS var references within current function.
+  // (2) Use local-dynamic for multiple TLS var references within current
+  // function.
+  PPCFunctionInfo *FuncInfo =
+      DAG.getMachineFunction().getInfo<PPCFunctionInfo>();
+  if (!FuncInfo->isAIXFuncUseInitDone()) {
+    SmallPtrSet<const GlobalValue *, 8> TLSGV;
+    // Iterate over all instructions within current function, collect all TLS
+    // global variables (global variables taken as the first parameter to
+    // Intrinsic::threadlocal_address).
+    const Function &Func = DAG.getMachineFunction().getFunction();
+    for (Function::const_iterator BI = Func.begin(), BE = Func.end(); BI != BE;
+         ++BI)
+      for (BasicBlock::const_iterator II = BI->begin(), IE = BI->end();
+           II != IE; ++II)
+        if (II->getOpcode() == Instruction::Call)
+          if (const CallInst *CI = dyn_cast<const CallInst>(&*II))
+            if (Function *CF = CI->getCalledFunction())
+              if (CF->isDeclaration() &&
+                  CF->getIntrinsicID() == Intrinsic::threadlocal_address)
+                if (const GlobalValue *GV =
+                        dyn_cast<GlobalValue>(II->getOperand(0))) {
+                  TLSModel::Model GVModel = TM.getTLSModel(GV);
+                  if (GVModel == TLSModel::InitialExec ||
----------------
hubert-reinterpretcast wrote:

We should only count the number of `local-dynamic` accesses.

Rationale:
The option can be taken as providing the information that the shared object is loaded with the executable (not `dlopen`ed later). In that case, all variables accessed using the `local-dynamic` model can be accessed using `initial-exec`; however, there would still be variables accessed using `initial-exec` that cannot be accessed using `local-dynamic`.

https://github.com/llvm/llvm-project/pull/84132