[llvm] [BOLT] Add support for safe-icf (PR #116275)

[Maksim Panchenko via llvm-commits]

================
@@ -341,6 +341,61 @@ typedef std::unordered_map<BinaryFunction *, std::vector<BinaryFunction *>,
 namespace llvm {
 namespace bolt {
 
+Error IdenticalCodeFolding::processDataRelocations(
+    BinaryContext &BC, const SectionRef &SecRefRelData) {
+  for (const RelocationRef &Rel : SecRefRelData.relocations()) {
+    symbol_iterator SymbolIter = Rel.getSymbol();
+    const ObjectFile *OwningObj = Rel.getObject();
+    assert(SymbolIter != OwningObj->symbol_end() &&
+           "relocation Symbol expected");
+    const SymbolRef &Symbol = *SymbolIter;
+    const uint64_t SymbolAddress = cantFail(Symbol.getAddress());
+    const ELFObjectFileBase *ELFObj = dyn_cast<ELFObjectFileBase>(OwningObj);
+    if (!ELFObj)
+      return createFatalBOLTError(
+          Twine("BOLT-ERROR: Only ELFObjectFileBase is supported"));
+    const int64_t Addend = getRelocationAddend(ELFObj, Rel);
+    BinaryFunction *BF = BC.getBinaryFunctionAtAddress(SymbolAddress + Addend);
+    if (!BF)
+      continue;
+    BF->setUnsafeICF();
+  }
+  return Error::success();
+}
+
+Error IdenticalCodeFolding::markFunctionsUnsafeToFold(BinaryContext &BC) {
+  ErrorOr<BinarySection &> SecRelData = BC.getUniqueSectionByName(".rela.data");
----------------
maksfb wrote:

Are you saying it's impossible to have and application that loads a function address from read-only section without having the same function address being referenced directly from code?

Orthogonally, with pragma directives and linker scripts it's possible to assign arbitrary names to sections. Relying on section name matching in this case is wrong. If you want to check for read-only vs writable section, you should check section attributes.

https://github.com/llvm/llvm-project/pull/116275