[AArch64] Only create called thunks when hardening against SLS

In preparation for implementing hardening of BLRA* instructions,
restrict thunk function generation to only the thunks being actually
called from any function. As described in the existing comments,
emitting all possible thunks for BLRAA and BLRAB instructions would
mean adding about 1800 functions in total, most of which are likely
not to be called.

This commit merges AArch64SLSHardening class into SLSBLRThunkInserter,
so thunks can be created as needed while rewriting a machine function.
The usages of TII, TRI and ST fields of AArch64SLSHardening class are
replaced with requesting them in-place, as ThunkInserter assumes
multiple "entry points" in contrast to the only runOnMachineFunction
method of AArch64SLSHardening.

The runOnMachineFunction method essentially replaces pre-existing
insertThunks implementation as there is no more need to insert all
possible thunks unconditionally. Instead, thunks are created on first
use from inside of insertThunks method.

Author: atrosinenko

llvm/lib/Target/AArch64/AArch64.h

@@ -40,7 +40,6 @@ FunctionPass *createAArch64ISelDag(AArch64TargetMachine &TM,
 FunctionPass *createAArch64StorePairSuppressPass();
 FunctionPass *createAArch64ExpandPseudoPass();
 FunctionPass *createAArch64SLSHardeningPass();
-FunctionPass *createAArch64IndirectThunks();
 FunctionPass *createAArch64SpeculationHardeningPass();
 FunctionPass *createAArch64LoadStoreOptimizationPass();
 ModulePass *createAArch64LowerHomogeneousPrologEpilogPass();

llvm/lib/Target/AArch64/AArch64SLSHardening.cpp

@@ -13,20 +13,16 @@
 
 #include "AArch64InstrInfo.h"
 #include "AArch64Subtarget.h"
-#include "Utils/AArch64BaseInfo.h"
 #include "llvm/CodeGen/IndirectThunks.h"
 #include "llvm/CodeGen/MachineBasicBlock.h"
 #include "llvm/CodeGen/MachineFunction.h"
-#include "llvm/CodeGen/MachineFunctionPass.h"
 #include "llvm/CodeGen/MachineInstr.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineOperand.h"
-#include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/RegisterScavenging.h"
 #include "llvm/IR/DebugLoc.h"
 #include "llvm/Pass.h"
-#include "llvm/Support/CodeGen.h"
-#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
 #include "llvm/Target/TargetMachine.h"
 #include <cassert>
 
@@ -36,38 +32,42 @@ using namespace llvm;
 
 #define AARCH64_SLS_HARDENING_NAME "AArch64 sls hardening pass"
 
+static const char SLSBLRNamePrefix[] = "__llvm_slsblr_thunk_";
+
 namespace {
 
-class AArch64SLSHardening : public MachineFunctionPass {
-public:
-  const TargetInstrInfo *TII;
-  const TargetRegisterInfo *TRI;
-  const AArch64Subtarget *ST;
+// Set of inserted thunks: bitmask with bits corresponding to
+// indexes in SLSBLRThunks array.
+typedef uint32_t ThunksSet;
 
-  static char ID;
-
-  AArch64SLSHardening() : MachineFunctionPass(ID) {
-    initializeAArch64SLSHardeningPass(*PassRegistry::getPassRegistry());
+struct SLSHardeningInserter : ThunkInserter<SLSHardeningInserter, ThunksSet> {
+public:
+  const char *getThunkPrefix() { return SLSBLRNamePrefix; }
+  bool mayUseThunk(const MachineFunction &MF) {
+    ComdatThunks &= !MF.getSubtarget<AArch64Subtarget>().hardenSlsNoComdat();
+    // We are inserting barriers aside from thunk calls, so
+    // check hardenSlsRetBr() as well.
+    return MF.getSubtarget<AArch64Subtarget>().hardenSlsBlr() ||
+           MF.getSubtarget<AArch64Subtarget>().hardenSlsRetBr();
   }
+  ThunksSet insertThunks(MachineModuleInfo &MMI, MachineFunction &MF,
+                         ThunksSet ExistingThunks);
+  void populateThunk(MachineFunction &MF);
 
-  bool runOnMachineFunction(MachineFunction &Fn) override;
+private:
+  bool ComdatThunks = true;
 
-  StringRef getPassName() const override { return AARCH64_SLS_HARDENING_NAME; }
+  bool hardenReturnsAndBRs(MachineModuleInfo &MMI, MachineBasicBlock &MBB);
+  bool hardenBLRs(MachineModuleInfo &MMI, MachineBasicBlock &MBB,
+                  ThunksSet &Thunks);
 
-private:
-  bool hardenReturnsAndBRs(MachineBasicBlock &MBB) const;
-  bool hardenBLRs(MachineBasicBlock &MBB) const;
-  MachineBasicBlock &ConvertBLRToBL(MachineBasicBlock &MBB,
-                                    MachineBasicBlock::instr_iterator) const;
+  void convertBLRToBL(MachineModuleInfo &MMI, MachineBasicBlock &MBB,
+                      MachineBasicBlock::instr_iterator MBBI,
+                      ThunksSet &Thunks);
 };
 
 } // end anonymous namespace
 
-char AArch64SLSHardening::ID = 0;
-
-INITIALIZE_PASS(AArch64SLSHardening, "aarch64-sls-hardening",
-                AARCH64_SLS_HARDENING_NAME, false, false)
-
 static void insertSpeculationBarrier(const AArch64Subtarget *ST,
                                      MachineBasicBlock &MBB,
                                      MachineBasicBlock::iterator MBBI,
@@ -90,18 +90,18 @@ static void insertSpeculationBarrier(const AArch64Subtarget *ST,
     BuildMI(MBB, MBBI, DL, TII->get(BarrierOpc));
 }
 
-bool AArch64SLSHardening::runOnMachineFunction(MachineFunction &MF) {
-  ST = &MF.getSubtarget<AArch64Subtarget>();
-  TII = MF.getSubtarget().getInstrInfo();
-  TRI = MF.getSubtarget().getRegisterInfo();
+ThunksSet SLSHardeningInserter::insertThunks(MachineModuleInfo &MMI,
+                                             MachineFunction &MF,
+                                             ThunksSet ExistingThunks) {
+  const AArch64Subtarget *ST = &MF.getSubtarget<AArch64Subtarget>();
 
-  bool Modified = false;
   for (auto &MBB : MF) {
-    Modified |= hardenReturnsAndBRs(MBB);
-    Modified |= hardenBLRs(MBB);
+    if (ST->hardenSlsRetBr())
+      hardenReturnsAndBRs(MMI, MBB);
+    if (ST->hardenSlsBlr())
+      hardenBLRs(MMI, MBB, ExistingThunks);
   }
-
-  return Modified;
+  return ExistingThunks;
 }
 
 static bool isBLR(const MachineInstr &MI) {
@@ -120,9 +120,10 @@ static bool isBLR(const MachineInstr &MI) {
   return false;
 }
 
-bool AArch64SLSHardening::hardenReturnsAndBRs(MachineBasicBlock &MBB) const {
-  if (!ST->hardenSlsRetBr())
-    return false;
+bool SLSHardeningInserter::hardenReturnsAndBRs(MachineModuleInfo &MMI,
+                                               MachineBasicBlock &MBB) {
+  const AArch64Subtarget *ST =
+      &MBB.getParent()->getSubtarget<AArch64Subtarget>();
   bool Modified = false;
   MachineBasicBlock::iterator MBBI = MBB.getFirstTerminator(), E = MBB.end();
   MachineBasicBlock::iterator NextMBBI;
@@ -138,78 +139,55 @@ bool AArch64SLSHardening::hardenReturnsAndBRs(MachineBasicBlock &MBB) const {
   return Modified;
 }
 
-static const char SLSBLRNamePrefix[] = "__llvm_slsblr_thunk_";
-
+static const unsigned NumPermittedRegs = 29;
 static const struct ThunkNameAndReg {
   const char* Name;
   Register Reg;
-} SLSBLRThunks[] = {
-  { "__llvm_slsblr_thunk_x0",  AArch64::X0},
-  { "__llvm_slsblr_thunk_x1",  AArch64::X1},
-  { "__llvm_slsblr_thunk_x2",  AArch64::X2},
-  { "__llvm_slsblr_thunk_x3",  AArch64::X3},
-  { "__llvm_slsblr_thunk_x4",  AArch64::X4},
-  { "__llvm_slsblr_thunk_x5",  AArch64::X5},
-  { "__llvm_slsblr_thunk_x6",  AArch64::X6},
-  { "__llvm_slsblr_thunk_x7",  AArch64::X7},
-  { "__llvm_slsblr_thunk_x8",  AArch64::X8},
-  { "__llvm_slsblr_thunk_x9",  AArch64::X9},
-  { "__llvm_slsblr_thunk_x10",  AArch64::X10},
-  { "__llvm_slsblr_thunk_x11",  AArch64::X11},
-  { "__llvm_slsblr_thunk_x12",  AArch64::X12},
-  { "__llvm_slsblr_thunk_x13",  AArch64::X13},
-  { "__llvm_slsblr_thunk_x14",  AArch64::X14},
-  { "__llvm_slsblr_thunk_x15",  AArch64::X15},
-  // X16 and X17 are deliberately missing, as the mitigation requires those
-  // register to not be used in BLR. See comment in ConvertBLRToBL for more
-  // details.
-  { "__llvm_slsblr_thunk_x18",  AArch64::X18},
-  { "__llvm_slsblr_thunk_x19",  AArch64::X19},
-  { "__llvm_slsblr_thunk_x20",  AArch64::X20},
-  { "__llvm_slsblr_thunk_x21",  AArch64::X21},
-  { "__llvm_slsblr_thunk_x22",  AArch64::X22},
-  { "__llvm_slsblr_thunk_x23",  AArch64::X23},
-  { "__llvm_slsblr_thunk_x24",  AArch64::X24},
-  { "__llvm_slsblr_thunk_x25",  AArch64::X25},
-  { "__llvm_slsblr_thunk_x26",  AArch64::X26},
-  { "__llvm_slsblr_thunk_x27",  AArch64::X27},
-  { "__llvm_slsblr_thunk_x28",  AArch64::X28},
-  { "__llvm_slsblr_thunk_x29",  AArch64::FP},
-  // X30 is deliberately missing, for similar reasons as X16 and X17 are
-  // missing.
-  { "__llvm_slsblr_thunk_x31",  AArch64::XZR},
+} SLSBLRThunks[NumPermittedRegs] = {
+    {"__llvm_slsblr_thunk_x0", AArch64::X0},
+    {"__llvm_slsblr_thunk_x1", AArch64::X1},
+    {"__llvm_slsblr_thunk_x2", AArch64::X2},
+    {"__llvm_slsblr_thunk_x3", AArch64::X3},
+    {"__llvm_slsblr_thunk_x4", AArch64::X4},
+    {"__llvm_slsblr_thunk_x5", AArch64::X5},
+    {"__llvm_slsblr_thunk_x6", AArch64::X6},
+    {"__llvm_slsblr_thunk_x7", AArch64::X7},
+    {"__llvm_slsblr_thunk_x8", AArch64::X8},
+    {"__llvm_slsblr_thunk_x9", AArch64::X9},
+    {"__llvm_slsblr_thunk_x10", AArch64::X10},
+    {"__llvm_slsblr_thunk_x11", AArch64::X11},
+    {"__llvm_slsblr_thunk_x12", AArch64::X12},
+    {"__llvm_slsblr_thunk_x13", AArch64::X13},
+    {"__llvm_slsblr_thunk_x14", AArch64::X14},
+    {"__llvm_slsblr_thunk_x15", AArch64::X15},
+    // X16 and X17 are deliberately missing, as the mitigation requires those
+    // register to not be used in BLR. See comment in ConvertBLRToBL for more
+    // details.
+    {"__llvm_slsblr_thunk_x18", AArch64::X18},
+    {"__llvm_slsblr_thunk_x19", AArch64::X19},
+    {"__llvm_slsblr_thunk_x20", AArch64::X20},
+    {"__llvm_slsblr_thunk_x21", AArch64::X21},
+    {"__llvm_slsblr_thunk_x22", AArch64::X22},
+    {"__llvm_slsblr_thunk_x23", AArch64::X23},
+    {"__llvm_slsblr_thunk_x24", AArch64::X24},
+    {"__llvm_slsblr_thunk_x25", AArch64::X25},
+    {"__llvm_slsblr_thunk_x26", AArch64::X26},
+    {"__llvm_slsblr_thunk_x27", AArch64::X27},
+    {"__llvm_slsblr_thunk_x28", AArch64::X28},
+    {"__llvm_slsblr_thunk_x29", AArch64::FP},
+    // X30 is deliberately missing, for similar reasons as X16 and X17 are
+    // missing.
+    {"__llvm_slsblr_thunk_x31", AArch64::XZR},
 };
 
-namespace {
-struct SLSBLRThunkInserter : ThunkInserter<SLSBLRThunkInserter> {
-  const char *getThunkPrefix() { return SLSBLRNamePrefix; }
-  bool mayUseThunk(const MachineFunction &MF) {
-    ComdatThunks &= !MF.getSubtarget<AArch64Subtarget>().hardenSlsNoComdat();
-    return MF.getSubtarget<AArch64Subtarget>().hardenSlsBlr();
-  }
-  bool insertThunks(MachineModuleInfo &MMI, MachineFunction &MF,
-                    bool ExistingThunks);
-  void populateThunk(MachineFunction &MF);
-
-private:
-  bool ComdatThunks = true;
-};
-} // namespace
-
-bool SLSBLRThunkInserter::insertThunks(MachineModuleInfo &MMI,
-                                       MachineFunction &MF,
-                                       bool ExistingThunks) {
-  if (ExistingThunks)
-    return false;
-  // FIXME: It probably would be possible to filter which thunks to produce
-  // based on which registers are actually used in BLR instructions in this
-  // function. But would that be a worthwhile optimization?
-  for (auto T : SLSBLRThunks)
-    createThunkFunction(MMI, T.Name, ComdatThunks);
-  return true;
+unsigned getThunkIndex(Register Reg) {
+  for (unsigned I = 0; I < NumPermittedRegs; ++I)
+    if (SLSBLRThunks[I].Reg == Reg)
+      return I;
+  llvm_unreachable("Unexpected register");
 }
 
-void SLSBLRThunkInserter::populateThunk(MachineFunction &MF) {
+void SLSHardeningInserter::populateThunk(MachineFunction &MF) {
   assert(MF.getFunction().hasComdat() == ComdatThunks &&
          "ComdatThunks value changed since MF creation");
   // FIXME: How to better communicate Register number, rather than through
@@ -258,8 +236,9 @@ void SLSBLRThunkInserter::populateThunk(MachineFunction &MF) {
                            Entry->end(), DebugLoc(), true /*AlwaysUseISBDSB*/);
 }
 
-MachineBasicBlock &AArch64SLSHardening::ConvertBLRToBL(
-    MachineBasicBlock &MBB, MachineBasicBlock::instr_iterator MBBI) const {
+void SLSHardeningInserter::convertBLRToBL(
+    MachineModuleInfo &MMI, MachineBasicBlock &MBB,
+    MachineBasicBlock::instr_iterator MBBI, ThunksSet &Thunks) {
   // Transform a BLR to a BL as follows:
   // Before:
   //   |-----------------------------|
@@ -285,7 +264,6 @@ MachineBasicBlock &AArch64SLSHardening::ConvertBLRToBL(
   //   |  barrierInsts               |
   //   |-----------------------------|
   //
-  // The __llvm_slsblr_thunk_xN thunks are created by the SLSBLRThunkInserter.
   // This function merely needs to transform BLR xN into BL
   // __llvm_slsblr_thunk_xN.
   //
@@ -318,37 +296,16 @@ MachineBasicBlock &AArch64SLSHardening::ConvertBLRToBL(
   }
   DebugLoc DL = BLR.getDebugLoc();
 
-  // If we'd like to support also BLRAA and BLRAB instructions, we'd need
-  // a lot more different kind of thunks.
-  // For example, a
-  //
-  // BLRAA xN, xM
-  //
-  // instruction probably would need to be transformed to something like:
-  //
-  // BL __llvm_slsblraa_thunk_x<N>_x<M>
-  //
-  // __llvm_slsblraa_thunk_x<N>_x<M>:
-  //   BRAA x<N>, x<M>
-  //   barrierInsts
-  //
-  // Given that about 30 different values of N are possible and about 30
-  // different values of M are possible in the above, with the current way
-  // of producing indirect thunks, we'd be producing about 30 times 30, i.e.
-  // about 900 thunks (where most might not be actually called). This would
-  // multiply further by two to support both BLRAA and BLRAB variants of those
-  // instructions.
-  // If we'd want to support this, we'd probably need to look into a different
-  // way to produce thunk functions, based on which variants are actually
-  // needed, rather than producing all possible variants.
-  // So far, LLVM does never produce BLRA* instructions, so let's leave this
-  // for the future when LLVM can start producing BLRA* instructions.
   MachineFunction &MF = *MBBI->getMF();
   MCContext &Context = MBB.getParent()->getContext();
-  auto ThunkIt =
-      llvm::find_if(SLSBLRThunks, [Reg](auto T) { return T.Reg == Reg; });
-  assert (ThunkIt != std::end(SLSBLRThunks));
-  MCSymbol *Sym = Context.getOrCreateSymbol(ThunkIt->Name);
+  const TargetInstrInfo *TII = MF.getSubtarget().getInstrInfo();
+  unsigned ThunkIndex = getThunkIndex(Reg);
+  StringRef ThunkName = SLSBLRThunks[ThunkIndex].Name;
+  MCSymbol *Sym = Context.getOrCreateSymbol(ThunkName);
+  if (!(Thunks & (1u << ThunkIndex))) {
+    Thunks |= 1u << ThunkIndex;
+    createThunkFunction(MMI, ThunkName, ComdatThunks);
+  }
 
   MachineInstr *BL = BuildMI(MBB, MBBI, DL, TII->get(BLOpcode)).addSym(Sym);
 
@@ -386,13 +343,11 @@ MachineBasicBlock &AArch64SLSHardening::ConvertBLRToBL(
                                            RegIsKilled /*isKill*/));
   // Remove BLR instruction
   MBB.erase(MBBI);
-
-  return MBB;
 }
 
-bool AArch64SLSHardening::hardenBLRs(MachineBasicBlock &MBB) const {
-  if (!ST->hardenSlsBlr())
-    return false;
+bool SLSHardeningInserter::hardenBLRs(MachineModuleInfo &MMI,
+                                      MachineBasicBlock &MBB,
+                                      ThunksSet &Thunks) {
   bool Modified = false;
   MachineBasicBlock::instr_iterator MBBI = MBB.instr_begin(),
                                     E = MBB.instr_end();
@@ -401,31 +356,30 @@ bool AArch64SLSHardening::hardenBLRs(MachineBasicBlock &MBB) const {
     MachineInstr &MI = *MBBI;
     NextMBBI = std::next(MBBI);
     if (isBLR(MI)) {
-      ConvertBLRToBL(MBB, MBBI);
+      convertBLRToBL(MMI, MBB, MBBI, Thunks);
       Modified = true;
     }
   }
   return Modified;
 }
 
-FunctionPass *llvm::createAArch64SLSHardeningPass() {
-  return new AArch64SLSHardening();
-}
-
 namespace {
-class AArch64IndirectThunks : public ThunkInserterPass<SLSBLRThunkInserter> {
+class AArch64SLSHardening : public ThunkInserterPass<SLSHardeningInserter> {
 public:
   static char ID;
 
-  AArch64IndirectThunks() : ThunkInserterPass(ID) {}
+  AArch64SLSHardening() : ThunkInserterPass(ID) {}
 
-  StringRef getPassName() const override { return "AArch64 Indirect Thunks"; }
+  StringRef getPassName() const override { return AARCH64_SLS_HARDENING_NAME; }
 };
 
 } // end anonymous namespace
 
-char AArch64IndirectThunks::ID = 0;
+char AArch64SLSHardening::ID = 0;
+
+INITIALIZE_PASS(AArch64SLSHardening, "aarch64-sls-hardening",
+                AARCH64_SLS_HARDENING_NAME, false, false)
 
-FunctionPass *llvm::createAArch64IndirectThunks() {
-  return new AArch64IndirectThunks();
+FunctionPass *llvm::createAArch64SLSHardeningPass() {
+  return new AArch64SLSHardening();
 }

llvm/lib/Target/AArch64/AArch64TargetMachine.cpp

@@ -861,7 +861,6 @@ void AArch64PassConfig::addPreEmitPass() {
 }
 
 void AArch64PassConfig::addPostBBSections() {
-  addPass(createAArch64IndirectThunks());
   addPass(createAArch64SLSHardeningPass());
   addPass(createAArch64PointerAuthPass());
   if (EnableBranchTargets)

llvm/test/CodeGen/AArch64/O0-pipeline.ll

@@ -74,7 +74,6 @@
 ; CHECK-NEXT:       StackMap Liveness Analysis
 ; CHECK-NEXT:       Live DEBUG_VALUE analysis
 ; CHECK-NEXT:       Machine Sanitizer Binary Metadata
-; CHECK-NEXT:       AArch64 Indirect Thunks
 ; CHECK-NEXT:       AArch64 sls hardening pass
 ; CHECK-NEXT:       AArch64 Pointer Authentication
 ; CHECK-NEXT:       AArch64 Branch Targets