[BOLT] Add support for Linux kernel static keys jump table (#86090)

Runtime code modification used by static keys is the most ubiquitous
self-modifying feature of the Linux kernel. The idea is to to eliminate
the condition check and associated conditional jump on a hot path if
that condition (based on a boolean value of a static key) does not
change often. Whenever they condition changes, the kernel runtime
modifies all code paths associated with that key flipping the code
between nop and (unconditional) jump.

bolt/include/bolt/Core/MCPlus.h

@@ -73,6 +73,7 @@ class MCAnnotation {
     kOffset,              /// Offset in the function.
     kLabel,               /// MCSymbol pointing to this instruction.
     kSize,                /// Size of the instruction.
+    kDynamicBranch,       /// Jit instruction patched at runtime.
     kGeneric              /// First generic annotation.
   };
 

bolt/include/bolt/Core/MCPlusBuilder.h

@@ -1199,6 +1199,16 @@ class MCPlusBuilder {
   /// Set instruction size.
   void setSize(MCInst &Inst, uint32_t Size) const;
 
+  /// Check if the branch instruction could be modified at runtime.
+  bool isDynamicBranch(const MCInst &Inst) const;
+
+  /// Return ID for runtime-modifiable instruction.
+  std::optional<uint32_t> getDynamicBranchID(const MCInst &Inst) const;
+
+  /// Mark instruction as a dynamic branch, i.e. a branch that can be
+  /// overwritten at runtime.
+  void setDynamicBranch(MCInst &Inst, uint32_t ID) const;
+
   /// Return MCSymbol that represents a target of this instruction at a given
   /// operand number \p OpNum. If there's no symbol associated with
   /// the operand - return nullptr.
@@ -1688,6 +1698,13 @@ class MCPlusBuilder {
     llvm_unreachable("not implemented");
   }
 
+  /// Create long conditional branch with a target-specific conditional code
+  /// \p CC.
+  virtual void createLongCondBranch(MCInst &Inst, const MCSymbol *Target,
+                                    unsigned CC, MCContext *Ctx) const {
+    llvm_unreachable("not implemented");
+  }
+
   /// Reverses the branch condition in Inst and update its taken target to TBB.
   ///
   /// Returns true on success.

bolt/lib/Core/BinaryContext.cpp

@@ -1939,7 +1939,13 @@ void BinaryContext::printInstruction(raw_ostream &OS, const MCInst &Instruction,
     OS << Endl;
     return;
   }
-  InstPrinter->printInst(&Instruction, 0, "", *STI, OS);
+  if (std::optional<uint32_t> DynamicID =
+          MIB->getDynamicBranchID(Instruction)) {
+    OS << "\tjit\t" << MIB->getTargetSymbol(Instruction)->getName()
+       << " # ID: " << DynamicID;
+  } else {
+    InstPrinter->printInst(&Instruction, 0, "", *STI, OS);
+  }
   if (MIB->isCall(Instruction)) {
     if (MIB->isTailCall(Instruction))
       OS << " # TAILCALL ";

bolt/lib/Core/BinaryFunction.cpp

@@ -3350,6 +3350,16 @@ void BinaryFunction::fixBranches() {
 
       // Eliminate unnecessary conditional branch.
       if (TSuccessor == FSuccessor) {
+        // FIXME: at the moment, we cannot safely remove static key branches.
+        if (MIB->isDynamicBranch(*CondBranch)) {
+          if (opts::Verbosity) {
+            BC.outs()
+                << "BOLT-INFO: unable to remove redundant dynamic branch in "
+                << *this << '\n';
+          }
+          continue;
+        }
+
         BB->removeDuplicateConditionalSuccessor(CondBranch);
         if (TSuccessor != NextBB)
           BB->addBranchInstruction(TSuccessor);
@@ -3358,8 +3368,13 @@ void BinaryFunction::fixBranches() {
 
       // Reverse branch condition and swap successors.
       auto swapSuccessors = [&]() {
-        if (MIB->isUnsupportedBranch(*CondBranch))
+        if (MIB->isUnsupportedBranch(*CondBranch)) {
+          if (opts::Verbosity) {
+            BC.outs() << "BOLT-INFO: unable to swap successors in " << *this
+                      << '\n';
+          }
           return false;
+        }
         std::swap(TSuccessor, FSuccessor);
         BB->swapConditionalSuccessors();
         auto L = BC.scopeLock();

bolt/lib/Core/MCPlusBuilder.cpp

@@ -303,6 +303,28 @@ void MCPlusBuilder::setSize(MCInst &Inst, uint32_t Size) const {
   setAnnotationOpValue(Inst, MCAnnotation::kSize, Size);
 }
 
+bool MCPlusBuilder::isDynamicBranch(const MCInst &Inst) const {
+  if (!hasAnnotation(Inst, MCAnnotation::kDynamicBranch))
+    return false;
+  assert(isBranch(Inst) && "Branch expected.");
+  return true;
+}
+
+std::optional<uint32_t>
+MCPlusBuilder::getDynamicBranchID(const MCInst &Inst) const {
+  if (std::optional<int64_t> Value =
+          getAnnotationOpValue(Inst, MCAnnotation::kDynamicBranch)) {
+    assert(isBranch(Inst) && "Branch expected.");
+    return static_cast<uint32_t>(*Value);
+  }
+  return std::nullopt;
+}
+
+void MCPlusBuilder::setDynamicBranch(MCInst &Inst, uint32_t ID) const {
+  assert(isBranch(Inst) && "Branch expected.");
+  setAnnotationOpValue(Inst, MCAnnotation::kDynamicBranch, ID);
+}
+
 bool MCPlusBuilder::hasAnnotation(const MCInst &Inst, unsigned Index) const {
   return (bool)getAnnotationOpValue(Inst, Index);
 }

bolt/lib/Passes/BinaryPasses.cpp

@@ -107,6 +107,12 @@ static cl::opt<unsigned>
                   cl::desc("print statistics about basic block ordering"),
                   cl::init(0), cl::cat(BoltOptCategory));
 
+static cl::opt<bool> PrintLargeFunctions(
+    "print-large-functions",
+    cl::desc("print functions that could not be overwritten due to excessive "
+             "size"),
+    cl::init(false), cl::cat(BoltOptCategory));
+
 static cl::list<bolt::DynoStats::Category>
     PrintSortedBy("print-sorted-by", cl::CommaSeparated,
                   cl::desc("print functions sorted by order of dyno stats"),
@@ -570,8 +576,12 @@ Error CheckLargeFunctions::runOnFunctions(BinaryContext &BC) {
     uint64_t HotSize, ColdSize;
     std::tie(HotSize, ColdSize) =
         BC.calculateEmittedSize(BF, /*FixBranches=*/false);
-    if (HotSize > BF.getMaxSize())
+    if (HotSize > BF.getMaxSize()) {
+      if (opts::PrintLargeFunctions)
+        BC.outs() << "BOLT-INFO: " << BF << " size exceeds allocated space by "
+                  << (HotSize - BF.getMaxSize()) << " bytes\n";
       BF.setSimple(false);
+    }
   };
 
   ParallelUtilities::PredicateTy SkipFunc = [&](const BinaryFunction &BF) {
@@ -852,6 +862,10 @@ uint64_t SimplifyConditionalTailCalls::fixTailCalls(BinaryFunction &BF) {
       assert(Result && "internal error analyzing conditional branch");
       assert(CondBranch && "conditional branch expected");
 
+      // Skip dynamic branches for now.
+      if (BF.getBinaryContext().MIB->isDynamicBranch(*CondBranch))
+        continue;
+
       // It's possible that PredBB is also a successor to BB that may have
       // been processed by a previous iteration of the SCTC loop, in which
       // case it may have been marked invalid.  We should skip rewriting in
@@ -1012,6 +1026,10 @@ uint64_t ShortenInstructions::shortenInstructions(BinaryFunction &Function) {
   const BinaryContext &BC = Function.getBinaryContext();
   for (BinaryBasicBlock &BB : Function) {
     for (MCInst &Inst : BB) {
+      // Skip shortening instructions with Size annotation.
+      if (BC.MIB->getSize(Inst))
+        continue;
+
       MCInst OriginalInst;
       if (opts::Verbosity > 2)
         OriginalInst = Inst;