[llvm-exegesis] Refactor how forbidden registers are computed.

[courbet]

Summary

Right now latency generation can incorrectly select the scratch register
as a dependency-carrying register.

• Move the logic for preventing register selection from Uops implementation to common SnippetGenerator class.
• Aliasing detection now takes a set of forbidden registers just like random register assignment does.

[gchatelet]

llvm/tools/llvm-exegesis/lib/MCInstrDescView.cpp

275 ↗

(On Diff #221932)

State.getRATC().emptyRegisters()

[courbet]

Thanks.

[courbet]

Address comments

llvm/trunk/tools/llvm-exegesis/lib/Latency.h

 namespace exegesis {
 
 class LatencySnippetGenerator : public SnippetGenerator {
 public:
   LatencySnippetGenerator(const LLVMState &State) : SnippetGenerator(State) {}
   ~LatencySnippetGenerator() override;
 
   llvm::Expected<std::vector<CodeTemplate>>
-  generateCodeTemplates(const Instruction &Instr) const override;
+  generateCodeTemplates(const Instruction &Instr,
+                        const BitVector &ForbiddenRegisters) const override;
 };
 
 class LatencyBenchmarkRunner : public BenchmarkRunner {
 public:
   LatencyBenchmarkRunner(const LLVMState &State,
                          InstructionBenchmark::ModeE Mode);
   ~LatencyBenchmarkRunner() override;
 
 private:
   llvm::Expected<std::vector<BenchmarkMeasure>>
   runMeasurements(const FunctionExecutor &Executor) const override;
 };
 } // namespace exegesis
 } // namespace llvm
 
 #endif // LLVM_TOOLS_LLVM_EXEGESIS_LATENCY_H

llvm/trunk/tools/llvm-exegesis/lib/Latency.cpp

"Repeating a single explicitly serial instruction"},
          ExecutionMode::SERIAL_VIA_NON_MEMORY_INSTR,
      "Repeating two instructions"},
 };
 
 static constexpr size_t kMaxAliasingInstructions = 10;
 
 static std::vector<Instruction>
 computeAliasingInstructions(const LLVMState &State, const Instruction &Instr,
-                            size_t MaxAliasingInstructions) {
+                            size_t MaxAliasingInstructions,
+                            const BitVector &ForbiddenRegisters) {
   // Randomly iterate the set of instructions.
   std::vector<unsigned> Opcodes;
   Opcodes.resize(State.getInstrInfo().getNumOpcodes());
   std::iota(Opcodes.begin(), Opcodes.end(), 0U);
   std::shuffle(Opcodes.begin(), Opcodes.end(), randomGenerator());
 
   std::vector<Instruction> AliasingInstructions;
   for (const unsigned OtherOpcode : Opcodes) {
     if (OtherOpcode == Instr.Description->getOpcode())
       continue;
     const Instruction &OtherInstr = State.getIC().getInstr(OtherOpcode);
     if (OtherInstr.hasMemoryOperands())
       continue;
-    if (Instr.hasAliasingRegistersThrough(OtherInstr))
-      AliasingInstructions.push_back(std::move(OtherInstr));
+    if (Instr.hasAliasingRegistersThrough(OtherInstr, ForbiddenRegisters))
+      AliasingInstructions.push_back(OtherInstr);
     if (AliasingInstructions.size() >= MaxAliasingInstructions)
       break;
   }
   return AliasingInstructions;
 }
 
-static ExecutionMode getExecutionModes(const Instruction &Instr) {
+static ExecutionMode getExecutionModes(const Instruction &Instr,
+                                       const BitVector &ForbiddenRegisters) {
   ExecutionMode EM = ExecutionMode::UNKNOWN;
   if (Instr.hasAliasingImplicitRegisters())
     EM |= ExecutionMode::ALWAYS_SERIAL_IMPLICIT_REGS_ALIAS;
   if (Instr.hasTiedRegisters())
     EM |= ExecutionMode::ALWAYS_SERIAL_TIED_REGS_ALIAS;
   if (Instr.hasMemoryOperands())
     EM |= ExecutionMode::SERIAL_VIA_MEMORY_INSTR;
   else {
-    if (Instr.hasAliasingRegisters())
+    if (Instr.hasAliasingRegisters(ForbiddenRegisters))
       EM |= ExecutionMode::SERIAL_VIA_EXPLICIT_REGS;
     if (Instr.hasOneUseOrOneDef())
       EM |= ExecutionMode::SERIAL_VIA_NON_MEMORY_INSTR;
   }
   return EM;
 }
 
 static void appendCodeTemplates(const LLVMState &State,
                                 const Instruction &Instr,
+                                const BitVector &ForbiddenRegisters,
                                 ExecutionMode ExecutionModeBit,
                                 llvm::StringRef ExecutionClassDescription,
                                 std::vector<CodeTemplate> &CodeTemplates) {
   assert(isEnumValue(ExecutionModeBit) && "Bit must be a power of two");
   switch (ExecutionModeBit) {
   case ExecutionMode::ALWAYS_SERIAL_IMPLICIT_REGS_ALIAS:
     // Nothing to do, the instruction is always serial.
     LLVM_FALLTHROUGH;
case ExecutionMode::SERIAL_VIA_EXPLICIT_REGS: {
     CT.Execution = ExecutionModeBit;
     CT.Info = ExecutionClassDescription;
     CT.Instructions.push_back(std::move(IT));
     CodeTemplates.push_back(std::move(CT));
     return;
   }
   case ExecutionMode::SERIAL_VIA_NON_MEMORY_INSTR: {
     // Select back-to-back non-memory instruction.
-    for (const auto OtherInstr :
-         computeAliasingInstructions(State, Instr, kMaxAliasingInstructions)) {
+    for (const auto OtherInstr : computeAliasingInstructions(
+             State, Instr, kMaxAliasingInstructions, ForbiddenRegisters)) {
       const AliasingConfigurations Forward(Instr, OtherInstr);
       const AliasingConfigurations Back(OtherInstr, Instr);
       InstructionTemplate ThisIT(Instr);
       InstructionTemplate OtherIT(OtherInstr);
       if (!Forward.hasImplicitAliasing())
         setRandomAliasing(Forward, ThisIT, OtherIT);
       if (!Back.hasImplicitAliasing())
         setRandomAliasing(Back, OtherIT, ThisIT);
static void appendCodeTemplates(const LLVMState &State,
   default:
     llvm_unreachable("Unhandled enum value");
   }
 }
 
 LatencySnippetGenerator::~LatencySnippetGenerator() = default;
 
 llvm::Expected<std::vector<CodeTemplate>>
-LatencySnippetGenerator::generateCodeTemplates(const Instruction &Instr) const {
+LatencySnippetGenerator::generateCodeTemplates(
+    const Instruction &Instr, const BitVector &ForbiddenRegisters) const {
   std::vector<CodeTemplate> Results;
-  const ExecutionMode EM = getExecutionModes(Instr);
+  const ExecutionMode EM = getExecutionModes(Instr, ForbiddenRegisters);
   for (const auto EC : kExecutionClasses) {
     for (const auto ExecutionModeBit : getExecutionModeBits(EM & EC.Mask))
-      appendCodeTemplates(State, Instr, ExecutionModeBit, EC.Description,
-                          Results);
+      appendCodeTemplates(State, Instr, ForbiddenRegisters, ExecutionModeBit,
+                          EC.Description, Results);
     if (!Results.empty())
       break;
   }
   if (Results.empty())
     return llvm::make_error<BenchmarkFailure>(
         "No strategy found to make the execution serial");
   return std::move(Results);
 }

llvm/trunk/tools/llvm-exegesis/lib/MCInstrDescView.h

struct Instruction {
   // Whether this instruction is self aliasing through its implicit registers.
   // Repeating this instruction is guaranteed to executes sequentially.
   bool hasAliasingImplicitRegisters() const;
 
   // Whether this instruction is self aliasing through some registers.
   // Repeating this instruction may execute sequentially by picking aliasing
   // Use and Def registers. It may also execute in parallel by picking non
   // aliasing Use and Def registers.
-  bool hasAliasingRegisters() const;
+  bool hasAliasingRegisters(const BitVector &ForbiddenRegisters) const;
 
   // Whether this instruction's registers alias with OtherInstr's registers.
-  bool hasAliasingRegistersThrough(const Instruction &OtherInstr) const;
+  bool hasAliasingRegistersThrough(const Instruction &OtherInstr,
+                                   const BitVector &ForbiddenRegisters) const;
 
   // Returns whether this instruction has Memory Operands.
   // Repeating this instruction executes sequentially with an instruction that
   // reads or write the same memory region.
   bool hasMemoryOperands() const;
 
   // Returns whether this instruction as at least one use or one def.
   // Repeating this instruction may execute sequentially by adding an
   // instruction that aliases one of these.
   bool hasOneUseOrOneDef() const;
 
   // Convenient function to help with debugging.
   void dump(const llvm::MCRegisterInfo &RegInfo,
+            const RegisterAliasingTrackerCache &RATC,
             llvm::raw_ostream &Stream) const;
 
   const llvm::MCInstrDesc *Description; // Never nullptr.
   llvm::StringRef Name;                 // The name of this instruction.
   llvm::SmallVector<Operand, 8> Operands;
   llvm::SmallVector<Variable, 4> Variables;
   llvm::BitVector ImplDefRegs; // The set of aliased implicit def registers.
   llvm::BitVector ImplUseRegs; // The set of aliased implicit use registers.

llvm/trunk/tools/llvm-exegesis/lib/MCInstrDescView.cpp

return any_of(Operands, [](const Operand &Op) {
     return Op.isReg() && Op.isExplicit() && Op.isMemory();
   });
 }
 
 bool Instruction::hasAliasingImplicitRegisters() const {
   return ImplDefRegs.anyCommon(ImplUseRegs);
 }
 
+// Returns true if there are registers that are both in `A` and `B` but not in
+// `Forbidden`.
+static bool anyCommonExcludingForbidden(const BitVector &A, const BitVector &B,
+                                        const BitVector &Forbidden) {
+  assert(A.size() == B.size() && B.size() == Forbidden.size());
+  const auto Size = A.size();
+  for (int AIndex = A.find_first(); AIndex != -1;) {
+    const int BIndex = B.find_first_in(AIndex, Size);
+    if (BIndex == -1)
+      return false;
+    if (AIndex == BIndex && !Forbidden.test(AIndex))
+      return true;
+    AIndex = A.find_first_in(BIndex + 1, Size);
+  }
+  return false;
+}
+
 bool Instruction::hasAliasingRegistersThrough(
-    const Instruction &OtherInstr) const {
-  return AllDefRegs.anyCommon(OtherInstr.AllUseRegs) &&
-         OtherInstr.AllDefRegs.anyCommon(AllUseRegs);
+    const Instruction &OtherInstr, const BitVector &ForbiddenRegisters) const {
+  return anyCommonExcludingForbidden(AllDefRegs, OtherInstr.AllUseRegs,
+                                     ForbiddenRegisters) &&
+         anyCommonExcludingForbidden(OtherInstr.AllDefRegs, AllUseRegs,
+                                     ForbiddenRegisters);
 }
 
 bool Instruction::hasTiedRegisters() const {
   return llvm::any_of(
       Variables, [](const Variable &Var) { return Var.hasTiedOperands(); });
 }
 
-bool Instruction::hasAliasingRegisters() const {
-  return AllDefRegs.anyCommon(AllUseRegs);
+bool Instruction::hasAliasingRegisters(
+    const BitVector &ForbiddenRegisters) const {
+  return anyCommonExcludingForbidden(AllDefRegs, AllUseRegs,
+                                     ForbiddenRegisters);
 }
 
 bool Instruction::hasOneUseOrOneDef() const {
   return AllDefRegs.count() || AllUseRegs.count();
 }
 
 void Instruction::dump(const llvm::MCRegisterInfo &RegInfo,
+                       const RegisterAliasingTrackerCache &RATC,
                        llvm::raw_ostream &Stream) const {
   Stream << "- " << Name << "\n";
   for (const auto &Op : Operands) {
     Stream << "- Op" << Op.getIndex();
     if (Op.isExplicit())
       Stream << " Explicit";
     if (Op.isImplicit())
       Stream << " Implicit";
for (const auto &Var : Variables) {
     Stream << "\n";
   }
   if (hasMemoryOperands())
     Stream << "- hasMemoryOperands\n";
   if (hasAliasingImplicitRegisters())
     Stream << "- hasAliasingImplicitRegisters (execution is always serial)\n";
   if (hasTiedRegisters())
     Stream << "- hasTiedRegisters (execution is always serial)\n";
-  if (hasAliasingRegisters())
+  if (hasAliasingRegisters(RATC.emptyRegisters()))
     Stream << "- hasAliasingRegisters\n";
 }
 
 InstructionsCache::InstructionsCache(const llvm::MCInstrInfo &InstrInfo,
                                      const RegisterAliasingTrackerCache &RATC)
     : InstrInfo(InstrInfo), RATC(RATC) {}
 
 const Instruction &InstructionsCache::getInstr(unsigned Opcode) const {

llvm/trunk/tools/llvm-exegesis/lib/SnippetGenerator.h

std::vector<RegisterValue> computeRegisterInitialValues(
       const std::vector<InstructionTemplate> &Snippet) const;
 
 protected:
   const LLVMState &State;
 
 private:
   // API to be implemented by subclasses.
   virtual llvm::Expected<std::vector<CodeTemplate>>
-  generateCodeTemplates(const Instruction &Instr) const = 0;
+  generateCodeTemplates(const Instruction &Instr,
+                        const BitVector &ForbiddenRegisters) const = 0;
 };
 
 // A global Random Number Generator to randomize configurations.
 // FIXME: Move random number generation into an object and make it seedable for
 // unit tests.
 std::mt19937 &randomGenerator();
 
 // Picks a random unsigned integer from 0 to Max (inclusive).

llvm/trunk/tools/llvm-exegesis/lib/SnippetGenerator.cpp

SnippetGeneratorFailure::SnippetGeneratorFailure(const llvm::Twine &S)
     : llvm::StringError(S, llvm::inconvertibleErrorCode()) {}
 
 SnippetGenerator::SnippetGenerator(const LLVMState &State) : State(State) {}
 
 SnippetGenerator::~SnippetGenerator() = default;
 
 llvm::Expected<std::vector<BenchmarkCode>>
 SnippetGenerator::generateConfigurations(const Instruction &Instr) const {
-  if (auto E = generateCodeTemplates(Instr)) {
-    const auto &RATC = State.getRATC();
+  llvm::BitVector ForbiddenRegs = State.getRATC().reservedRegisters();
+
+  // If the instruction has memory registers, prevent the generator from
+  // using the scratch register and its aliasing registers.
+  if (Instr.hasMemoryOperands()) {
+    const auto &ET = State.getExegesisTarget();
+    unsigned ScratchSpacePointerInReg =
+        ET.getScratchMemoryRegister(State.getTargetMachine().getTargetTriple());
+    if (ScratchSpacePointerInReg == 0)
+      return llvm::make_error<BenchmarkFailure>(
+          "Infeasible : target does not support memory instructions");
+    const auto &ScratchRegAliases =
+        State.getRATC().getRegister(ScratchSpacePointerInReg).aliasedBits();
+    // If the instruction implicitly writes to ScratchSpacePointerInReg , abort.
+    // FIXME: We could make a copy of the scratch register.
+    for (const auto &Op : Instr.Operands) {
+      if (Op.isDef() && Op.isImplicitReg() &&
+          ScratchRegAliases.test(Op.getImplicitReg()))
+        return llvm::make_error<BenchmarkFailure>(
+            "Infeasible : memory instruction uses scratch memory register");
+    }
+    ForbiddenRegs |= ScratchRegAliases;
+  }
+
+  if (auto E = generateCodeTemplates(Instr, ForbiddenRegs)) {
     std::vector<BenchmarkCode> Output;
     for (CodeTemplate &CT : E.get()) {
-      const llvm::BitVector &ForbiddenRegs =
-          CT.ScratchSpacePointerInReg
-              ? RATC.getRegister(CT.ScratchSpacePointerInReg).aliasedBits()
-              : RATC.emptyRegisters();
       // TODO: Generate as many BenchmarkCode as needed.
       {
         BenchmarkCode BC;
         BC.Info = CT.Info;
         for (InstructionTemplate &IT : CT.Instructions) {
           randomizeUnsetVariables(State.getExegesisTarget(), ForbiddenRegs, IT);
           BC.Instructions.push_back(IT.build());
         }

llvm/trunk/tools/llvm-exegesis/lib/Uops.h

 namespace exegesis {
 
 class UopsSnippetGenerator : public SnippetGenerator {
 public:
   UopsSnippetGenerator(const LLVMState &State) : SnippetGenerator(State) {}
   ~UopsSnippetGenerator() override;
 
   llvm::Expected<std::vector<CodeTemplate>>
-  generateCodeTemplates(const Instruction &Instr) const override;
+  generateCodeTemplates(const Instruction &Instr,
+                        const BitVector &ForbiddenRegisters) const override;
 
   static constexpr const size_t kMinNumDifferentAddresses = 6;
 
 private:
   // Instantiates memory operands within a snippet.
   // To make computations as parallel as possible, we generate independant
   // memory locations for instructions that load and store. If there are less
   // than kMinNumDifferentAddresses in the original snippet, we duplicate

llvm/trunk/tools/llvm-exegesis/lib/Uops.cpp

void UopsSnippetGenerator::instantiateMemoryOperands(
   }
   assert(I * MemStep < BenchmarkRunner::ScratchSpace::kSize &&
          "not enough scratch space");
 }
 
 static std::vector<InstructionTemplate> generateSnippetUsingStaticRenaming(
     const LLVMState &State, const InstructionTemplate &IT,
     const ArrayRef<const Variable *> TiedVariables,
-    const BitVector *ScratchSpaceAliasedRegs) {
+    const BitVector &ForbiddenRegisters) {
   std::vector<InstructionTemplate> Instructions;
   // Assign registers to variables in a round-robin manner. This is simple but
   // ensures that the most register-constrained variable does not get starved.
   std::vector<BitVector> PossibleRegsForVar;
   for (const Variable *Var : TiedVariables) {
     assert(Var);
     const Operand &Op = IT.Instr.getPrimaryOperand(*Var);
     assert(Op.isReg());
-    BitVector PossibleRegs = State.getRATC().emptyRegisters();
-    if (ScratchSpaceAliasedRegs) {
-      PossibleRegs |= *ScratchSpaceAliasedRegs;
-    }
-    PossibleRegs.flip();
-    PossibleRegs &= Op.getRegisterAliasing().sourceBits();
+    BitVector PossibleRegs = Op.getRegisterAliasing().sourceBits();
+    remove(PossibleRegs, ForbiddenRegisters);
     PossibleRegsForVar.push_back(std::move(PossibleRegs));
   }
   SmallVector<int, 2> Iterators(TiedVariables.size(), 0);
   while (true) {
     InstructionTemplate TmpIT = IT;
     // Find a possible register for each variable in turn, marking the
     // register as taken.
     for (size_t VarId = 0; VarId < TiedVariables.size(); ++VarId) {
for (size_t VarId = 0; VarId < TiedVariables.size(); ++VarId) {
         OtherPossibleRegs.reset(NextPossibleReg);
       }
     }
     Instructions.push_back(std::move(TmpIT));
   }
 }
 
 llvm::Expected<std::vector<CodeTemplate>>
-UopsSnippetGenerator::generateCodeTemplates(const Instruction &Instr) const {
+UopsSnippetGenerator::generateCodeTemplates(
+    const Instruction &Instr, const BitVector &ForbiddenRegisters) const {
   CodeTemplate CT;
-  const llvm::BitVector *ScratchSpaceAliasedRegs = nullptr;
-  if (Instr.hasMemoryOperands()) {
-    const auto &ET = State.getExegesisTarget();
-    CT.ScratchSpacePointerInReg =
-        ET.getScratchMemoryRegister(State.getTargetMachine().getTargetTriple());
-    if (CT.ScratchSpacePointerInReg == 0)
-      return llvm::make_error<BenchmarkFailure>(
-          "Infeasible : target does not support memory instructions");
-    ScratchSpaceAliasedRegs =
-        &State.getRATC().getRegister(CT.ScratchSpacePointerInReg).aliasedBits();
-    // If the instruction implicitly writes to ScratchSpacePointerInReg , abort.
-    // FIXME: We could make a copy of the scratch register.
-    for (const auto &Op : Instr.Operands) {
-      if (Op.isDef() && Op.isImplicitReg() &&
-          ScratchSpaceAliasedRegs->test(Op.getImplicitReg()))
-        return llvm::make_error<BenchmarkFailure>(
-            "Infeasible : memory instruction uses scratch memory register");
-    }
-  }
-
+      Instr.hasMemoryOperands()
+          ? State.getExegesisTarget().getScratchMemoryRegister(
+                State.getTargetMachine().getTargetTriple())
+          : 0;
   const AliasingConfigurations SelfAliasing(Instr, Instr);
   InstructionTemplate IT(Instr);
   if (SelfAliasing.empty()) {
     CT.Info = "instruction is parallel, repeating a random one.";
     CT.Instructions.push_back(std::move(IT));
     instantiateMemoryOperands(CT.ScratchSpacePointerInReg, CT.Instructions);
     return getSingleton(std::move(CT));
   }
   if (SelfAliasing.hasImplicitAliasing()) {
     CT.Info = "instruction is serial, repeating a random one.";
     CT.Instructions.push_back(std::move(IT));
     instantiateMemoryOperands(CT.ScratchSpacePointerInReg, CT.Instructions);
     return getSingleton(std::move(CT));
   }
   const auto TiedVariables = getVariablesWithTiedOperands(Instr);
   if (!TiedVariables.empty()) {
     CT.Info = "instruction has tied variables, using static renaming.";
     CT.Instructions = generateSnippetUsingStaticRenaming(
-        State, IT, TiedVariables, ScratchSpaceAliasedRegs);
+        State, IT, TiedVariables, ForbiddenRegisters);
     instantiateMemoryOperands(CT.ScratchSpacePointerInReg, CT.Instructions);
     return getSingleton(std::move(CT));
   }
-  const auto &ReservedRegisters = State.getRATC().reservedRegisters();
   // No tied variables, we pick random values for defs.
   llvm::BitVector Defs(State.getRegInfo().getNumRegs());
   for (const auto &Op : Instr.Operands) {
     if (Op.isReg() && Op.isExplicit() && Op.isDef() && !Op.isMemory()) {
       auto PossibleRegisters = Op.getRegisterAliasing().sourceBits();
-      remove(PossibleRegisters, ReservedRegisters);
-      // Do not use the scratch memory address register.
-      if (ScratchSpaceAliasedRegs)
-        remove(PossibleRegisters, *ScratchSpaceAliasedRegs);
+      // Do not use forbidden registers.
+      remove(PossibleRegisters, ForbiddenRegisters);
       assert(PossibleRegisters.any() && "No register left to choose from");
       const auto RandomReg = randomBit(PossibleRegisters);
       Defs.set(RandomReg);
       IT.getValueFor(Op) = llvm::MCOperand::createReg(RandomReg);
     }
   }
   // And pick random use values that are not reserved and don't alias with defs.
   const auto DefAliases = getAliasedBits(State.getRegInfo(), Defs);
   for (const auto &Op : Instr.Operands) {
     if (Op.isReg() && Op.isExplicit() && Op.isUse() && !Op.isMemory()) {
       auto PossibleRegisters = Op.getRegisterAliasing().sourceBits();
-      remove(PossibleRegisters, ReservedRegisters);
-      // Do not use the scratch memory address register.
-      if (ScratchSpaceAliasedRegs)
-        remove(PossibleRegisters, *ScratchSpaceAliasedRegs);
+      remove(PossibleRegisters, ForbiddenRegisters);
       remove(PossibleRegisters, DefAliases);
       assert(PossibleRegisters.any() && "No register left to choose from");
       const auto RandomReg = randomBit(PossibleRegisters);
       IT.getValueFor(Op) = llvm::MCOperand::createReg(RandomReg);
     }
   }
   CT.Info =
       "instruction has no tied variables picking Uses different from defs";

llvm/trunk/tools/llvm-exegesis/lib/X86/Target.cpp

 }
 
 namespace {
 class X86LatencySnippetGenerator : public LatencySnippetGenerator {
 public:
   using LatencySnippetGenerator::LatencySnippetGenerator;
 
   llvm::Expected<std::vector<CodeTemplate>>
-  generateCodeTemplates(const Instruction &Instr) const override;
+  generateCodeTemplates(const Instruction &Instr,
+                        const BitVector &ForbiddenRegisters) const override;
 };
 } // namespace
 
 llvm::Expected<std::vector<CodeTemplate>>
 X86LatencySnippetGenerator::generateCodeTemplates(
-    const Instruction &Instr) const {
+    const Instruction &Instr, const BitVector &ForbiddenRegisters) const {
   if (auto E = IsInvalidOpcode(Instr))
     return std::move(E);
 
   switch (getX86FPFlags(Instr)) {
   case llvm::X86II::NotFP:
-    return LatencySnippetGenerator::generateCodeTemplates(Instr);
+    return LatencySnippetGenerator::generateCodeTemplates(Instr,
+                                                          ForbiddenRegisters);
   case llvm::X86II::ZeroArgFP:
   case llvm::X86II::OneArgFP:
   case llvm::X86II::SpecialFP:
   case llvm::X86II::CompareFP:
   case llvm::X86II::CondMovFP:
     return llvm::make_error<BenchmarkFailure>("Unsupported x87 Instruction");
   case llvm::X86II::OneArgFPRW:
   case llvm::X86II::TwoArgFP:
     // These are instructions like
     //   - `ST(0) = fsqrt(ST(0))` (OneArgFPRW)
     //   - `ST(0) = ST(0) + ST(i)` (TwoArgFP)
     // They are intrinsically serial and do not modify the state of the stack.
     return generateSelfAliasingCodeTemplates(Instr);
   default:
     llvm_unreachable("Unknown FP Type!");
   }
 }
 
 namespace {
 class X86UopsSnippetGenerator : public UopsSnippetGenerator {
 public:
   using UopsSnippetGenerator::UopsSnippetGenerator;
 
   llvm::Expected<std::vector<CodeTemplate>>
-  generateCodeTemplates(const Instruction &Instr) const override;
+  generateCodeTemplates(const Instruction &Instr,
+                        const BitVector &ForbiddenRegisters) const override;
 };
 } // namespace
 
 llvm::Expected<std::vector<CodeTemplate>>
 X86UopsSnippetGenerator::generateCodeTemplates(
-    const Instruction &Instr) const {
+    const Instruction &Instr, const BitVector &ForbiddenRegisters) const {
   if (auto E = IsInvalidOpcode(Instr))
     return std::move(E);
 
   switch (getX86FPFlags(Instr)) {
   case llvm::X86II::NotFP:
-    return UopsSnippetGenerator::generateCodeTemplates(Instr);
+    return UopsSnippetGenerator::generateCodeTemplates(Instr,
+                                                       ForbiddenRegisters);
   case llvm::X86II::ZeroArgFP:
   case llvm::X86II::OneArgFP:
   case llvm::X86II::SpecialFP:
     return llvm::make_error<BenchmarkFailure>("Unsupported x87 Instruction");
   case llvm::X86II::OneArgFPRW:
   case llvm::X86II::TwoArgFP:
     // These are instructions like
     //   - `ST(0) = fsqrt(ST(0))` (OneArgFPRW)

llvm/trunk/unittests/tools/llvm-exegesis/X86/SnippetGeneratorTest.cpp

 template <typename SnippetGeneratorT>
 class SnippetGeneratorTest : public X86SnippetGeneratorTest {
 protected:
   SnippetGeneratorTest() : Generator(State) {}
 
   std::vector<CodeTemplate> checkAndGetCodeTemplates(unsigned Opcode) {
     randomGenerator().seed(0); // Initialize seed.
     const Instruction &Instr = State.getIC().getInstr(Opcode);
-    auto CodeTemplateOrError = Generator.generateCodeTemplates(Instr);
+    auto CodeTemplateOrError = Generator.generateCodeTemplates(
+        Instr, State.getRATC().emptyRegisters());
     EXPECT_FALSE(CodeTemplateOrError.takeError()); // Valid configuration.
     return std::move(CodeTemplateOrError.get());
   }
 
   SnippetGeneratorT Generator;
 };
 
 using LatencySnippetGeneratorTest =
TEST_F(LatencySnippetGeneratorTest, ImplicitSelfDependencyThroughExplicitRegs) {
   EXPECT_THAT(IT.getOpcode(), Opcode);
   ASSERT_THAT(IT.VariableValues, SizeIs(3));
   EXPECT_THAT(IT.VariableValues,
               AnyOf(ElementsAre(IsReg(), IsInvalid(), IsReg()),
                     ElementsAre(IsReg(), IsReg(), IsInvalid())))
       << "Op0 is either set to Op1 or to Op2";
 }
 
+TEST_F(LatencySnippetGeneratorTest,
+       ImplicitSelfDependencyThroughExplicitRegsForbidAll) {
+  // - VXORPSrr
+  // - Op0 Explicit Def RegClass(VR128)
+  // - Op1 Explicit Use RegClass(VR128)
+  // - Op2 Explicit Use RegClass(VR128)
+  // - Var0 [Op0]
+  // - Var1 [Op1]
+  // - Var2 [Op2]
+  // - hasAliasingRegisters
+  const unsigned Opcode = llvm::X86::VXORPSrr;
+  randomGenerator().seed(0); // Initialize seed.
+  const Instruction &Instr = State.getIC().getInstr(Opcode);
+  auto AllRegisters = State.getRATC().emptyRegisters();
+  AllRegisters.flip();
+  auto Error = Generator.generateCodeTemplates(Instr, AllRegisters).takeError();
+  EXPECT_TRUE((bool)Error);
+  llvm::consumeError(std::move(Error));
+}
+
 TEST_F(LatencySnippetGeneratorTest, DependencyThroughOtherOpcode) {
   // - CMP64rr
   // - Op0 Explicit Use RegClass(GR64)
   // - Op1 Explicit Use RegClass(GR64)
   // - Op2 Implicit Def Reg(EFLAGS)
   // - Var0 [Op0]
   // - Var1 [Op1]
   const unsigned Opcode = llvm::X86::CMP64rr;
TEST_F(UopsSnippetGeneratorTest, MemoryUse) {
   EXPECT_THAT(IT.getOpcode(), Opcode);
   ASSERT_THAT(IT.VariableValues, SizeIs(6));
   EXPECT_EQ(IT.VariableValues[2].getImm(), 1);
   EXPECT_EQ(IT.VariableValues[3].getReg(), 0u);
   EXPECT_EQ(IT.VariableValues[4].getImm(), 0);
   EXPECT_EQ(IT.VariableValues[5].getReg(), 0u);
 }
 
-TEST_F(UopsSnippetGeneratorTest, MemoryUse_Movsb) {
-  // MOVSB writes to scratch memory register.
-  // - MOVSB
-  // - Op0 Explicit Use Memory RegClass(GR8)
-  // - Op1 Explicit Use Memory RegClass(GR8)
-  // - Op2 Explicit Use Memory RegClass(SEGMENT_REG)
-  // - Op3 Implicit Def Reg(EDI)
-  // - Op4 Implicit Def Reg(ESI)
-  // - Op5 Implicit Use Reg(EDI)
-  // - Op6 Implicit Use Reg(ESI)
-  // - Op7 Implicit Use Reg(DF)
-  // - Var0 [Op0]
-  // - Var1 [Op1]
-  // - Var2 [Op2]
-  // - hasMemoryOperands
-  // - hasAliasingImplicitRegisters (execution is always serial)
-  // - hasAliasingRegisters
-  const unsigned Opcode = llvm::X86::MOVSB;
-  const Instruction &Instr = State.getIC().getInstr(Opcode);
-  auto Error = Generator.generateCodeTemplates(Instr).takeError();
-  EXPECT_TRUE((bool)Error);
-  llvm::consumeError(std::move(Error));
-}
-
 class FakeSnippetGenerator : public SnippetGenerator {
 public:
   FakeSnippetGenerator(const LLVMState &State) : SnippetGenerator(State) {}
 
   Instruction createInstruction(unsigned Opcode) {
     return State.getIC().getInstr(Opcode);
   }
 
 private:
   llvm::Expected<std::vector<CodeTemplate>>
-  generateCodeTemplates(const Instruction &Instr) const override {
+  generateCodeTemplates(const Instruction &, const BitVector &) const override {
     return llvm::make_error<llvm::StringError>("not implemented",
                                                llvm::inconvertibleErrorCode());
   }
 };
 
 using FakeSnippetGeneratorTest = SnippetGeneratorTest<FakeSnippetGenerator>;
 
 testing::Matcher<const RegisterValue &> IsRegisterValue(unsigned Reg,
                                                         llvm::APInt Value) {
   return testing::AllOf(testing::Field(&RegisterValue::Register, Reg),
                         testing::Field(&RegisterValue::Value, Value));
 }
 
+TEST_F(FakeSnippetGeneratorTest, MemoryUse_Movsb) {
+  // MOVSB writes to scratch memory register.
+  // - MOVSB
+  // - Op0 Explicit Use Memory RegClass(GR8)
+  // - Op1 Explicit Use Memory RegClass(GR8)
+  // - Op2 Explicit Use Memory RegClass(SEGMENT_REG)
+  // - Op3 Implicit Def Reg(EDI)
+  // - Op4 Implicit Def Reg(ESI)
+  // - Op5 Implicit Use Reg(EDI)
+  // - Op6 Implicit Use Reg(ESI)
+  // - Op7 Implicit Use Reg(DF)
+  // - Var0 [Op0]
+  // - Var1 [Op1]
+  // - Var2 [Op2]
+  // - hasMemoryOperands
+  // - hasAliasingImplicitRegisters (execution is always serial)
+  // - hasAliasingRegisters
+  const unsigned Opcode = llvm::X86::MOVSB;
+  const Instruction &Instr = State.getIC().getInstr(Opcode);
+  auto Error = Generator.generateConfigurations(Instr).takeError();
+  EXPECT_TRUE((bool)Error);
+  llvm::consumeError(std::move(Error));
+}
+
 TEST_F(FakeSnippetGeneratorTest, ComputeRegisterInitialValuesAdd16ri) {
   // ADD16ri:
   // explicit def 0       : reg RegClass=GR16
   // explicit use 1       : reg RegClass=GR16 | TIED_TO:0
   // explicit use 2       : imm
   // implicit def         : EFLAGS
   InstructionTemplate IT(Generator.createInstruction(llvm::X86::ADD16ri));
   IT.getValueFor(IT.Instr.Variables[0]) =