TableGen: Support target specialized pseudoinstructions

arsenm · arsenm · commit 8f62ad4e96a0 · 2025-09-20T08:50:37.000+09:00
Allow a target to steal the definition of a generic pseudoinstruction
and remap the operands. This works by defining a new instruction, which
will simply swap out the emitted entry in the InstrInfo table.

This is intended to eliminate the C++ half of the implementation
of PointerLikeRegClass. With RegClassByHwMode, the remaining usecase
for PointerLikeRegClass are the common codegen pseudoinstructions.
Every target maintains its own copy of the generic pseudo operand
definitions anyway, so we can stub out the register operands with
an appropriate class instead of waiting for runtime resolution.

In the future we could probably take this a bit further. For example,
there is a similar problem for ADJCALLSTACKUP/DOWN since they depend
on target register definitions for the stack pointer register.
diff --git a/llvm/include/llvm/Target/Target.td b/llvm/include/llvm/Target/Target.td
@@ -1574,6 +1574,99 @@ def CONVERGENCECTRL_GLUE : StandardPseudoInstruction {
 }
 }
 
+/// Allow a target to replace the instruction definition of a
+/// StandardPseudoInstruction. A target should only define one
+/// instance of this per instruction.
+///
+/// This is intended to allow targets to specify the register class
+/// used for pointers. It should not be used to change the fundamental
+/// operand structure (e.g., this should not add or remove operands,
+/// or change the operand types).
+class TargetSpecializedStandardPseudoInstruction<
+  StandardPseudoInstruction base_inst> : Instruction {
+
+  StandardPseudoInstruction Instruction = base_inst;
+  let OutOperandList = base_inst.OutOperandList;
+  let InOperandList = base_inst.InOperandList;
+
+  // TODO: Copy everything
+  let usesCustomInserter = base_inst.usesCustomInserter;
+  let hasSideEffects = base_inst.hasSideEffects;
+  let mayLoad = base_inst.mayLoad;
+  let mayStore = base_inst.mayStore;
+  let isTerminator = base_inst.isTerminator;
+  let isBranch = base_inst.isBranch;
+  let isIndirectBranch = base_inst.isIndirectBranch;
+  let isEHScopeReturn = base_inst.isEHScopeReturn;
+  let isReturn = base_inst.isReturn;
+  let isCall = base_inst.isCall;
+  let hasCtrlDep = base_inst.hasCtrlDep;
+  let isReMaterializable = base_inst.isReMaterializable;
+  let isMeta = base_inst.isMeta;
+  let Size = base_inst.Size;
+  let isAsCheapAsAMove = base_inst.isAsCheapAsAMove;
+  let isPseudo = true;
+  let hasNoSchedulingInfo = true;
+  let isNotDuplicable = base_inst.isNotDuplicable;
+  let isConvergent = base_inst.isConvergent;
+  let hasExtraSrcRegAllocReq = base_inst.hasExtraSrcRegAllocReq;
+  let hasExtraDefRegAllocReq = base_inst.hasExtraDefRegAllocReq;
+}
+
+// All pseudo instructions which need a pointer register class, which
+// should be specialized by a target.
+defvar PseudosWithPtrOps = [
+  LOAD_STACK_GUARD,
+  PREALLOCATED_ARG,
+  PATCHABLE_EVENT_CALL,
+  PATCHABLE_TYPED_EVENT_CALL
+];
+
+
+/// Replace PointerLikeRegClass operands in OperandList with new_rc.
+class RemapPointerOperandList<dag OperandList, RegisterClassLike new_rc> {
+  // Collect the set of names so we can query and rewrite them.
+  list<string> op_names = !foreach(i, !range(!size(OperandList)),
+                                      !getdagname(OperandList, i));
+
+  // Beautiful language. This would be a lot easier if !getdagarg
+  // didn't require a specific type. We can't just collect a list of
+  // the operand values and reconstruct the dag, since there isn't a
+  // common base class for all the field kinds used in
+  // pseudoinstruction definitions; therefore everything must be
+  // maintained as a dag, so use a foldl. Additionally, ? doesn't
+  // evaluate as false so we get even more noise.
+  dag ret =
+    !foldl(OperandList, op_names, acc, name,
+      !cond(
+        !initialized(!getdagarg<PointerLikeRegClass>(OperandList, name))
+          : !setdagarg(acc, name, new_rc),
+        !initialized(!getdagarg<unknown_class>(OperandList, name)) : acc,
+        !initialized(!getdagarg<DAGOperand>(OperandList, name)) : acc
+      )
+    );
+}
+
+/// Define an override for a pseudoinstruction which uses a pointer
+/// register class, specialized to the target's pointer type.
+class RemapPointerOperands<StandardPseudoInstruction inst,
+                           RegisterClassLike new_rc> :
+  TargetSpecializedStandardPseudoInstruction<inst> {
+  let OutOperandList =
+    RemapPointerOperandList<inst.OutOperandList, new_rc>.ret;
+  let InOperandList =
+    RemapPointerOperandList<inst.InOperandList, new_rc>.ret;
+}
+
+/// Helper to replace all pseudoinstructions using pointers to a
+/// target register class. Most targets should use this
+multiclass RemapAllTargetPseudoPointerOperands<
+  RegisterClassLike default_ptr_rc> {
+  foreach inst = PseudosWithPtrOps in {
+    def : RemapPointerOperands<inst, default_ptr_rc>;
+  }
+}
+
 // Generic opcodes used in GlobalISel.
 include "llvm/Target/GenericOpcodes.td"
 
diff --git a/llvm/test/TableGen/target-specialized-pseudos.td b/llvm/test/TableGen/target-specialized-pseudos.td
@@ -0,0 +1,101 @@
+// RUN: llvm-tblgen -gen-instr-info -I %p/../../include %s -DONECASE -o - | FileCheck -check-prefixes=CHECK,ONECASE %s
+// RUN: llvm-tblgen -gen-instr-info -I %p/../../include %s -DALLCASES -o - | FileCheck -check-prefixes=CHECK,ALLCASES %s
+// RUN: not llvm-tblgen -gen-instr-info -I %p/../../include %s -DERROR -o /dev/null 2>&1 | FileCheck -check-prefix=ERROR %s
+
+// CHECK: namespace llvm::MyTarget {
+// CHECK: enum {
+// CHECK: LOAD_STACK_GUARD = [[LOAD_STACK_GUARD_OPCODE:[0-9]+]],
+// CHECK: PREALLOCATED_ARG = [[PREALLOCATED_ARG_OPCODE:[0-9]+]],
+// CHECK: PATCHABLE_EVENT_CALL = [[PATCHABLE_EVENT_CALL_OPCODE:[0-9]+]],
+// CHECK: PATCHABLE_TYPED_EVENT_CALL = [[PATCHABLE_TYPED_EVENT_CALL_OPCODE:[0-9]+]],
+
+// Make sure no enum entry is emitted for MY_LOAD_STACK_GUARD
+// CHECK: G_UBFX = [[G_UBFX_OPCODE:[0-9]+]],
+// CHECK-NEXT: MY_MOV = [[MY_MOV_OPCODE:[0-9]+]],
+// CHECK-NEXT: INSTRUCTION_LIST_END = [[INSTR_LIST_END_OPCODE:[0-9]+]]
+
+
+// CHECK: extern const MyTargetInstrTable MyTargetDescs = {
+// CHECK-NEXT: {
+// CHECK-NEXT: { [[MY_MOV_OPCODE]],	2,	1,	2,	0,	0,	0,	{{[0-9]+}},	MyTargetImpOpBase + 0,	0|(1ULL<<MCID::ExtraSrcRegAllocReq)|(1ULL<<MCID::ExtraDefRegAllocReq), 0x0ULL },  // MY_MOV
+// CHECK-NEXT: { [[G_UBFX_OPCODE]],	4,	1,	0,	0,	0,	0,	{{[0-9]+}},	MyTargetImpOpBase + 0,	0|(1ULL<<MCID::PreISelOpcode)|(1ULL<<MCID::Pseudo)|(1ULL<<MCID::ExtraSrcRegAllocReq)|(1ULL<<MCID::ExtraDefRegAllocReq), 0x0ULL },  // G_UBFX
+
+// ONECASE: { [[LOAD_STACK_GUARD_OPCODE]],	1,	1,	0,	0,	0,	0,	[[LOAD_STACK_GUARD_OP_ENTRY:[0-9]+]],	MyTargetImpOpBase + 0,	0|(1ULL<<MCID::Pseudo)|(1ULL<<MCID::MayLoad)|(1ULL<<MCID::Rematerializable)|(1ULL<<MCID::ExtraSrcRegAllocReq)|(1ULL<<MCID::ExtraDefRegAllocReq), 0x0ULL },  // MY_LOAD_STACK_GUARD
+
+// ALLCASES: { [[PATCHABLE_TYPED_EVENT_CALL_OPCODE]],	3,	0,	0,	0,	0,	0,	[[PATCHABLE_TYPED_EVENT_CALL_OP_ENTRY:[0-9]+]],	MyTargetImpOpBase + 0,	0|(1ULL<<MCID::Pseudo)|(1ULL<<MCID::Call)|(1ULL<<MCID::MayLoad)|(1ULL<<MCID::MayStore)|(1ULL<<MCID::UsesCustomInserter)|(1ULL<<MCID::UnmodeledSideEffects)|(1ULL<<MCID::ExtraSrcRegAllocReq)|(1ULL<<MCID::ExtraDefRegAllocReq), 0x0ULL },  // anonymous_
+// ALLCASES: { [[PATCHABLE_EVENT_CALL_OPCODE]],	2,	0,	0,	0,	0,	0,	[[PATCHABLE_EVENT_CALL_OP_ENTRY:[0-9]+]],	MyTargetImpOpBase + 0, 0|(1ULL<<MCID::Pseudo)|(1ULL<<MCID::Call)|(1ULL<<MCID::MayLoad)|(1ULL<<MCID::MayStore)|(1ULL<<MCID::UsesCustomInserter)|(1ULL<<MCID::UnmodeledSideEffects)|(1ULL<<MCID::ExtraSrcRegAllocReq)|(1ULL<<MCID::ExtraDefRegAllocReq), 0x0ULL },  // anonymous_
+// ALLCASES: { [[PREALLOCATED_ARG_OPCODE]],	3,	1,	0,	0,	0,	0,	[[PREALLOCATED_ARG_OP_ENTRY:[0-9]+]],	MyTargetImpOpBase + 0,	0|(1ULL<<MCID::Pseudo)|(1ULL<<MCID::UsesCustomInserter)|(1ULL<<MCID::UnmodeledSideEffects)|(1ULL<<MCID::ExtraSrcRegAllocReq)|(1ULL<<MCID::ExtraDefRegAllocReq), 0x0ULL },  // anonymous_
+// ALLCASES: { [[LOAD_STACK_GUARD_OPCODE]],	1,	1,	0,	0,	0,	0,	[[LOAD_STACK_GUARD_OP_ENTRY:[0-9]+]],	MyTargetImpOpBase + 0,	0|(1ULL<<MCID::Pseudo)|(1ULL<<MCID::MayLoad)|(1ULL<<MCID::Rematerializable)|(1ULL<<MCID::ExtraSrcRegAllocReq)|(1ULL<<MCID::ExtraDefRegAllocReq), 0x0ULL },  // anonymous_
+
+// CHECK: /* 0 */ { -1, 0, MCOI::OPERAND_UNKNOWN, 0 },
+
+// ONECASE: /* [[LOAD_STACK_GUARD_OP_ENTRY]] */ { MyTarget::XRegsRegClassID, 0, MCOI::OPERAND_REGISTER, 0 },
+
+// ALLCASES: /* [[LOAD_STACK_GUARD_OP_ENTRY]] */ { MyTarget::XRegsRegClassID, 0, MCOI::OPERAND_REGISTER, 0 },
+// ALLCASES: /* [[PREALLOCATED_ARG_OP_ENTRY]] */ { MyTarget::XRegsRegClassID, 0, MCOI::OPERAND_REGISTER, 0 }, { -1, 0, MCOI::OPERAND_IMMEDIATE, 0 }, { -1, 0, MCOI::OPERAND_IMMEDIATE, 0 },
+// ALLCASES: /* [[PATCHABLE_EVENT_CALL_OP_ENTRY]] */ { MyTarget::XRegsRegClassID, 0, MCOI::OPERAND_REGISTER, 0 }, { -1, 0, MCOI::OPERAND_UNKNOWN, 0 },
+// ALLCASES: /* [[PATCHABLE_TYPED_EVENT_CALL_OP_ENTRY]] */ { -1, 0, MCOI::OPERAND_UNKNOWN, 0 }, { MyTarget::XRegsRegClassID, 0, MCOI::OPERAND_REGISTER, 0 }, { -1, 0, MCOI::OPERAND_UNKNOWN, 0 },
+
+
+// CHECK: const char MyTargetInstrNameData[] = {
+// CHECK: /* {{[0-9]+}} */ "LOAD_STACK_GUARD\000"
+
+include "llvm/Target/Target.td"
+
+class MyReg<string n>
+  : Register<n> {
+  let Namespace = "MyTarget";
+}
+
+class MyClass<int size, list<ValueType> types, dag registers>
+  : RegisterClass<"MyTarget", types, size, registers> {
+  let Size = size;
+}
+
+def X0 : MyReg<"x0">;
+def X1 : MyReg<"x1">;
+def XRegs : RegisterClass<"MyTarget", [i64], 64, (add X0, X1)>;
+
+
+class TestInstruction : Instruction {
+  let Size = 2;
+  let Namespace = "MyTarget";
+  let hasSideEffects = false;
+}
+
+#ifdef ONECASE
+
+// Example setting the pointer register class manually
+def MY_LOAD_STACK_GUARD :
+  TargetSpecializedStandardPseudoInstruction<LOAD_STACK_GUARD> {
+  let Namespace = "MyTarget";
+  let OutOperandList = (outs XRegs:$dst);
+}
+
+#endif
+
+#ifdef ALLCASES
+
+defm my_remaps : RemapAllTargetPseudoPointerOperands<XRegs>;
+
+#endif
+
+
+#ifdef ERROR
+
+def MY_LOAD_STACK_GUARD_0 : TargetSpecializedStandardPseudoInstruction<LOAD_STACK_GUARD>;
+
+// ERROR: :[[@LINE+1]]:5: error: multiple overrides of 'LOAD_STACK_GUARD' defined
+def MY_LOAD_STACK_GUARD_1 : TargetSpecializedStandardPseudoInstruction<LOAD_STACK_GUARD>;
+
+#endif
+
+def MY_MOV : TestInstruction {
+  let OutOperandList = (outs XRegs:$dst);
+  let InOperandList = (ins XRegs:$src);
+  let AsmString = "my_mov $dst, $src";
+}
+
+
+def MyTargetISA : InstrInfo;
+def MyTarget : Target { let InstructionSet = MyTargetISA; }
diff --git a/llvm/utils/TableGen/Common/CodeGenTarget.cpp b/llvm/utils/TableGen/Common/CodeGenTarget.cpp
@@ -278,21 +278,32 @@ void CodeGenTarget::ComputeInstrsByEnum() const {
     const CodeGenInstruction *Instr = GetInstByName(Name, InstMap, Records);
     assert(Instr && "Missing target independent instruction");
     assert(Instr->Namespace == "TargetOpcode" && "Bad namespace");
+
     InstrsByEnum.push_back(Instr);
   }
   unsigned EndOfPredefines = InstrsByEnum.size();
   assert(EndOfPredefines == getNumFixedInstructions() &&
          "Missing generic opcode");
 
+  unsigned SkippedInsts = 0;
+
   for (const auto &[_, CGIUp] : InstMap) {
     const CodeGenInstruction *CGI = CGIUp.get();
     if (CGI->Namespace != "TargetOpcode") {
+
+      if (CGI->TheDef->isSubClassOf(
+              "TargetSpecializedStandardPseudoInstruction")) {
+        ++SkippedInsts;
+        continue;
+      }
+
       InstrsByEnum.push_back(CGI);
       NumPseudoInstructions += CGI->TheDef->getValueAsBit("isPseudo");
     }
   }
 
-  assert(InstrsByEnum.size() == InstMap.size() && "Missing predefined instr");
+  assert(InstrsByEnum.size() + SkippedInsts == InstMap.size() &&
+         "Missing predefined instr");
 
   // All of the instructions are now in random order based on the map iteration.
   llvm::sort(
diff --git a/llvm/utils/TableGen/InstrInfoEmitter.cpp b/llvm/utils/TableGen/InstrInfoEmitter.cpp
@@ -71,6 +71,13 @@ class InstrInfoEmitter {
   typedef std::vector<OperandInfoTy> OperandInfoListTy;
   typedef std::map<OperandInfoTy, unsigned> OperandInfoMapTy;
 
+  DenseMap<const CodeGenInstruction *, const CodeGenInstruction *>
+      TargetSpecializedPseudoInsts;
+
+  /// Compute mapping of opcodes which should have their definitions overridden
+  /// by a target version.
+  void buildTargetSpecializedPseudoInstsMap();
+
   /// Generate member functions in the target-specific GenInstrInfo class.
   ///
   /// This method is used to custom expand TIIPredicate definitions.
@@ -215,6 +222,10 @@ InstrInfoEmitter::CollectOperandInfo(OperandInfoListTy &OperandInfoList,
   const CodeGenTarget &Target = CDP.getTargetInfo();
   unsigned Offset = 0;
   for (const CodeGenInstruction *Inst : Target.getInstructions()) {
+    auto OverrideEntry = TargetSpecializedPseudoInsts.find(Inst);
+    if (OverrideEntry != TargetSpecializedPseudoInsts.end())
+      Inst = OverrideEntry->second;
+
     OperandInfoTy OperandInfo = GetOperandInfo(*Inst);
     if (OperandInfoMap.try_emplace(OperandInfo, Offset).second) {
       OperandInfoList.push_back(OperandInfo);
@@ -882,6 +893,25 @@ void InstrInfoEmitter::emitTIIHelperMethods(raw_ostream &OS,
   }
 }
 
+void InstrInfoEmitter::buildTargetSpecializedPseudoInstsMap() {
+  ArrayRef<const Record *> SpecializedInsts = Records.getAllDerivedDefinitions(
+      "TargetSpecializedStandardPseudoInstruction");
+  const CodeGenTarget &Target = CDP.getTargetInfo();
+
+  for (const Record *SpecializedRec : SpecializedInsts) {
+    const CodeGenInstruction &SpecializedInst =
+        Target.getInstruction(SpecializedRec);
+    const Record *BaseInstRec = SpecializedRec->getValueAsDef("Instruction");
+
+    const CodeGenInstruction &BaseInst = Target.getInstruction(BaseInstRec);
+
+    if (!TargetSpecializedPseudoInsts.insert({&BaseInst, &SpecializedInst})
+             .second)
+      PrintFatalError(SpecializedRec, "multiple overrides of '" +
+                                          BaseInst.getName() + "' defined");
+  }
+}
+
 //===----------------------------------------------------------------------===//
 // Main Output.
 //===----------------------------------------------------------------------===//
@@ -904,6 +934,8 @@ void InstrInfoEmitter::run(raw_ostream &OS) {
 
   // Collect all of the operand info records.
   Timer.startTimer("Collect operand info");
+  buildTargetSpecializedPseudoInstsMap();
+
   OperandInfoListTy OperandInfoList;
   OperandInfoMapTy OperandInfoMap;
   unsigned OperandInfoSize =
@@ -970,6 +1002,11 @@ void InstrInfoEmitter::run(raw_ostream &OS) {
   for (const CodeGenInstruction *Inst : reverse(NumberedInstructions)) {
     // Keep a list of the instruction names.
     InstrNames.add(Inst->getName());
+
+    auto OverrideEntry = TargetSpecializedPseudoInsts.find(Inst);
+    if (OverrideEntry != TargetSpecializedPseudoInsts.end())
+      Inst = OverrideEntry->second;
+
     // Emit the record into the table.
     emitRecord(*Inst, --Num, InstrInfo, EmittedLists, OperandInfoMap, OS);
   }
