Revert "[AArch64][SME] Port all SME routines to RuntimeLibcalls" (#153392)

This introduced a 5% compile-time regression on AArch64, see
https://llvm-compile-time-tracker.com/compare.php?from=b9138bde3562de5c28a239dbd303caf2406678c6&to=271688b87abe7cf45aceaff8266270a25eb7b436&stat=instructions:u.

Reverts #152505.

Author: nikic

llvm/include/llvm/CodeGen/TargetLowering.h

@@ -3560,12 +3560,6 @@ class LLVM_ABI TargetLoweringBase {
     return Libcalls.getLibcallImplName(Call);
   }
 
-  /// Check if this is valid libcall for the current module, otherwise
-  /// RTLIB::Unsupported.
-  RTLIB::LibcallImpl getSupportedLibcallImpl(StringRef FuncName) const {
-    return Libcalls.getSupportedLibcallImpl(FuncName);
-  }
-
   const char *getMemcpyName() const { return Libcalls.getMemcpyName(); }
 
   /// Get the comparison predicate that's to be used to test the result of the

llvm/include/llvm/IR/RuntimeLibcalls.td

@@ -406,17 +406,6 @@ multiclass LibmLongDoubleLibCall<string libcall_basename = !toupper(NAME),
 def SC_MEMCPY : RuntimeLibcall;
 def SC_MEMMOVE : RuntimeLibcall;
 def SC_MEMSET : RuntimeLibcall;
-def SC_MEMCHR: RuntimeLibcall;
-
-// AArch64 SME ABI calls
-def SMEABI_SME_STATE : RuntimeLibcall;
-def SMEABI_TPIDR2_SAVE : RuntimeLibcall;
-def SMEABI_ZA_DISABLE : RuntimeLibcall;
-def SMEABI_TPIDR2_RESTORE : RuntimeLibcall;
-def SMEABI_GET_CURRENT_VG : RuntimeLibcall;
-def SMEABI_SME_STATE_SIZE : RuntimeLibcall;
-def SMEABI_SME_SAVE : RuntimeLibcall;
-def SMEABI_SME_RESTORE : RuntimeLibcall;
 
 // ARM EABI calls
 def AEABI_MEMCPY4 : RuntimeLibcall; // Align 4
@@ -1234,35 +1223,8 @@ defset list<RuntimeLibcallImpl> AArch64LibcallImpls = {
   def __arm_sc_memcpy : RuntimeLibcallImpl<SC_MEMCPY>;
   def __arm_sc_memmove : RuntimeLibcallImpl<SC_MEMMOVE>;
   def __arm_sc_memset : RuntimeLibcallImpl<SC_MEMSET>;
-  def __arm_sc_memchr : RuntimeLibcallImpl<SC_MEMCHR>;
 } // End AArch64LibcallImpls
 
-def __arm_sme_state : RuntimeLibcallImpl<SMEABI_SME_STATE>;
-def __arm_tpidr2_save : RuntimeLibcallImpl<SMEABI_TPIDR2_SAVE>;
-def __arm_za_disable : RuntimeLibcallImpl<SMEABI_ZA_DISABLE>;
-def __arm_tpidr2_restore : RuntimeLibcallImpl<SMEABI_TPIDR2_RESTORE>;
-def __arm_get_current_vg : RuntimeLibcallImpl<SMEABI_GET_CURRENT_VG>;
-def __arm_sme_state_size : RuntimeLibcallImpl<SMEABI_SME_STATE_SIZE>;
-def __arm_sme_save : RuntimeLibcallImpl<SMEABI_SME_SAVE>;
-def __arm_sme_restore : RuntimeLibcallImpl<SMEABI_SME_RESTORE>;
-
-def SMEABI_LibCalls_PreserveMost_From_X0 : LibcallsWithCC<(add
-  __arm_tpidr2_save,
-  __arm_za_disable,
-  __arm_tpidr2_restore),
-  SMEABI_PreserveMost_From_X0>;
-
-def SMEABI_LibCalls_PreserveMost_From_X1 : LibcallsWithCC<(add
-  __arm_get_current_vg,
-  __arm_sme_state_size,
-  __arm_sme_save,
-  __arm_sme_restore),
-  SMEABI_PreserveMost_From_X1>;
-
-def SMEABI_LibCalls_PreserveMost_From_X2 : LibcallsWithCC<(add
-  __arm_sme_state),
-  SMEABI_PreserveMost_From_X2>;
-
 def isAArch64_ExceptArm64EC
     : RuntimeLibcallPredicate<"(TT.isAArch64() && !TT.isWindowsArm64EC())">;
 def isWindowsArm64EC : RuntimeLibcallPredicate<"TT.isWindowsArm64EC()">;
@@ -1282,10 +1244,7 @@ def AArch64SystemLibrary : SystemRuntimeLibrary<
        LibmHasSinCosF32, LibmHasSinCosF64, LibmHasSinCosF128,
        DefaultLibmExp10,
        DefaultStackProtector,
-       SecurityCheckCookieIfWinMSVC,
-       SMEABI_LibCalls_PreserveMost_From_X0,
-       SMEABI_LibCalls_PreserveMost_From_X1,
-       SMEABI_LibCalls_PreserveMost_From_X2)
+       SecurityCheckCookieIfWinMSVC)
 >;
 
 // Prepend a # to every name

llvm/include/llvm/IR/RuntimeLibcallsImpl.td

@@ -36,9 +36,6 @@ def ARM_AAPCS : LibcallCallingConv<[{CallingConv::ARM_AAPCS}]>;
 def ARM_AAPCS_VFP : LibcallCallingConv<[{CallingConv::ARM_AAPCS_VFP}]>;
 def X86_STDCALL : LibcallCallingConv<[{CallingConv::X86_StdCall}]>;
 def AVR_BUILTIN : LibcallCallingConv<[{CallingConv::AVR_BUILTIN}]>;
-def SMEABI_PreserveMost_From_X0 : LibcallCallingConv<[{CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X0}]>;
-def SMEABI_PreserveMost_From_X1 : LibcallCallingConv<[{CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1}]>;
-def SMEABI_PreserveMost_From_X2 : LibcallCallingConv<[{CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X2}]>;
 
 /// Abstract definition for functionality the compiler may need to
 /// emit a call to. Emits the RTLIB::Libcall enum - This enum defines

llvm/lib/Target/AArch64/AArch64FrameLowering.cpp

@@ -1487,11 +1487,8 @@ bool isVGInstruction(MachineBasicBlock::iterator MBBI) {
 
     if (Opc == AArch64::BL) {
       auto Op1 = MBBI->getOperand(0);
-      auto &TLI =
-          *MBBI->getMF()->getSubtarget<AArch64Subtarget>().getTargetLowering();
-      char const *GetCurrentVG =
-          TLI.getLibcallName(RTLIB::SMEABI_GET_CURRENT_VG);
-      return Op1.isSymbol() && StringRef(Op1.getSymbolName()) == GetCurrentVG;
+      return Op1.isSymbol() &&
+             (StringRef(Op1.getSymbolName()) == "__arm_get_current_vg");
     }
   }
 
@@ -3471,7 +3468,6 @@ bool AArch64FrameLowering::spillCalleeSavedRegisters(
     MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
     ArrayRef<CalleeSavedInfo> CSI, const TargetRegisterInfo *TRI) const {
   MachineFunction &MF = *MBB.getParent();
-  auto &TLI = *MF.getSubtarget<AArch64Subtarget>().getTargetLowering();
   const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
   AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
   bool NeedsWinCFI = needsWinCFI(MF);
@@ -3585,11 +3581,11 @@ bool AArch64FrameLowering::spillCalleeSavedRegisters(
               .addReg(AArch64::X0, RegState::Implicit)
               .setMIFlag(MachineInstr::FrameSetup);
 
-        RTLIB::Libcall LC = RTLIB::SMEABI_GET_CURRENT_VG;
-        const uint32_t *RegMask =
-            TRI->getCallPreservedMask(MF, TLI.getLibcallCallingConv(LC));
+        const uint32_t *RegMask = TRI->getCallPreservedMask(
+            MF,
+            CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1);
         BuildMI(MBB, MI, DL, TII.get(AArch64::BL))
-            .addExternalSymbol(TLI.getLibcallName(LC))
+            .addExternalSymbol("__arm_get_current_vg")
             .addRegMask(RegMask)
             .addReg(AArch64::X0, RegState::ImplicitDefine)
             .setMIFlag(MachineInstr::FrameSetup);

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -3083,12 +3083,13 @@ AArch64TargetLowering::EmitGetSMESaveSize(MachineInstr &MI,
   AArch64FunctionInfo *FuncInfo = MF->getInfo<AArch64FunctionInfo>();
   const TargetInstrInfo *TII = Subtarget->getInstrInfo();
   if (FuncInfo->isSMESaveBufferUsed()) {
-    RTLIB::Libcall LC = RTLIB::SMEABI_SME_STATE_SIZE;
     const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
     BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(AArch64::BL))
-        .addExternalSymbol(getLibcallName(LC))
+        .addExternalSymbol("__arm_sme_state_size")
         .addReg(AArch64::X0, RegState::ImplicitDefine)
-        .addRegMask(TRI->getCallPreservedMask(*MF, getLibcallCallingConv(LC)));
+        .addRegMask(TRI->getCallPreservedMask(
+            *MF, CallingConv::
+                     AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1));
     BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(TargetOpcode::COPY),
             MI.getOperand(0).getReg())
         .addReg(AArch64::X0);
@@ -3108,12 +3109,13 @@ AArch64TargetLowering::EmitEntryPStateSM(MachineInstr &MI,
   const TargetInstrInfo *TII = Subtarget->getInstrInfo();
   Register ResultReg = MI.getOperand(0).getReg();
   if (FuncInfo->isPStateSMRegUsed()) {
-    RTLIB::Libcall LC = RTLIB::SMEABI_SME_STATE;
     const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
     BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(AArch64::BL))
-        .addExternalSymbol(getLibcallName(LC))
+        .addExternalSymbol("__arm_sme_state")
         .addReg(AArch64::X0, RegState::ImplicitDefine)
-        .addRegMask(TRI->getCallPreservedMask(*MF, getLibcallCallingConv(LC)));
+        .addRegMask(TRI->getCallPreservedMask(
+            *MF, CallingConv::
+                     AArch64_SME_ABI_Support_Routines_PreserveMost_From_X2));
     BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(TargetOpcode::COPY), ResultReg)
         .addReg(AArch64::X0);
   } else {
@@ -5737,15 +5739,15 @@ static SDValue getSVEPredicateBitCast(EVT VT, SDValue Op, SelectionDAG &DAG) {
 SDValue AArch64TargetLowering::getRuntimePStateSM(SelectionDAG &DAG,
                                                   SDValue Chain, SDLoc DL,
                                                   EVT VT) const {
-  RTLIB::Libcall LC = RTLIB::SMEABI_SME_STATE;
-  SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
+  SDValue Callee = DAG.getExternalSymbol("__arm_sme_state",
                                          getPointerTy(DAG.getDataLayout()));
   Type *Int64Ty = Type::getInt64Ty(*DAG.getContext());
   Type *RetTy = StructType::get(Int64Ty, Int64Ty);
   TargetLowering::CallLoweringInfo CLI(DAG);
   ArgListTy Args;
   CLI.setDebugLoc(DL).setChain(Chain).setLibCallee(
-      getLibcallCallingConv(LC), RetTy, Callee, std::move(Args));
+      CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X2,
+      RetTy, Callee, std::move(Args));
   std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
   SDValue Mask = DAG.getConstant(/*PSTATE.SM*/ 1, DL, MVT::i64);
   return DAG.getNode(ISD::AND, DL, MVT::i64, CallResult.first.getOperand(0),
@@ -8598,12 +8600,12 @@ static void analyzeCallOperands(const AArch64TargetLowering &TLI,
 }
 
 static SMECallAttrs
-getSMECallAttrs(const Function &Caller, const TargetLowering &TLI,
+getSMECallAttrs(const Function &Caller,
                 const TargetLowering::CallLoweringInfo &CLI) {
   if (CLI.CB)
-    return SMECallAttrs(*CLI.CB, &TLI);
+    return SMECallAttrs(*CLI.CB);
   if (auto *ES = dyn_cast<ExternalSymbolSDNode>(CLI.Callee))
-    return SMECallAttrs(SMEAttrs(Caller), SMEAttrs(ES->getSymbol(), TLI));
+    return SMECallAttrs(SMEAttrs(Caller), SMEAttrs(ES->getSymbol()));
   return SMECallAttrs(SMEAttrs(Caller), SMEAttrs(SMEAttrs::Normal));
 }
 
@@ -8625,7 +8627,7 @@ bool AArch64TargetLowering::isEligibleForTailCallOptimization(
 
   // SME Streaming functions are not eligible for TCO as they may require
   // the streaming mode or ZA to be restored after returning from the call.
-  SMECallAttrs CallAttrs = getSMECallAttrs(CallerF, *this, CLI);
+  SMECallAttrs CallAttrs = getSMECallAttrs(CallerF, CLI);
   if (CallAttrs.requiresSMChange() || CallAttrs.requiresLazySave() ||
       CallAttrs.requiresPreservingAllZAState() ||
       CallAttrs.caller().hasStreamingBody())
@@ -8919,14 +8921,14 @@ static SDValue emitSMEStateSaveRestore(const AArch64TargetLowering &TLI,
       DAG.getCopyFromReg(Chain, DL, Info->getSMESaveBufferAddr(), MVT::i64);
   Args.push_back(Entry);
 
-  RTLIB::Libcall LC =
-      IsSave ? RTLIB::SMEABI_SME_SAVE : RTLIB::SMEABI_SME_RESTORE;
-  SDValue Callee = DAG.getExternalSymbol(TLI.getLibcallName(LC),
-                                         TLI.getPointerTy(DAG.getDataLayout()));
+  SDValue Callee =
+      DAG.getExternalSymbol(IsSave ? "__arm_sme_save" : "__arm_sme_restore",
+                            TLI.getPointerTy(DAG.getDataLayout()));
   auto *RetTy = Type::getVoidTy(*DAG.getContext());
   TargetLowering::CallLoweringInfo CLI(DAG);
   CLI.setDebugLoc(DL).setChain(Chain).setLibCallee(
-      TLI.getLibcallCallingConv(LC), RetTy, Callee, std::move(Args));
+      CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1, RetTy,
+      Callee, std::move(Args));
   return TLI.LowerCallTo(CLI).second;
 }
 
@@ -9114,7 +9116,7 @@ AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
   }
 
   // Determine whether we need any streaming mode changes.
-  SMECallAttrs CallAttrs = getSMECallAttrs(MF.getFunction(), *this, CLI);
+  SMECallAttrs CallAttrs = getSMECallAttrs(MF.getFunction(), CLI);
 
   auto DescribeCallsite =
       [&](OptimizationRemarkAnalysis &R) -> OptimizationRemarkAnalysis & {
@@ -9691,12 +9693,11 @@ AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
 
   if (RequiresLazySave) {
     // Conditionally restore the lazy save using a pseudo node.
-    RTLIB::Libcall LC = RTLIB::SMEABI_TPIDR2_RESTORE;
     TPIDR2Object &TPIDR2 = FuncInfo->getTPIDR2Obj();
     SDValue RegMask = DAG.getRegisterMask(
-        TRI->getCallPreservedMask(MF, getLibcallCallingConv(LC)));
+        TRI->SMEABISupportRoutinesCallPreservedMaskFromX0());
     SDValue RestoreRoutine = DAG.getTargetExternalSymbol(
-        getLibcallName(LC), getPointerTy(DAG.getDataLayout()));
+        "__arm_tpidr2_restore", getPointerTy(DAG.getDataLayout()));
     SDValue TPIDR2_EL0 = DAG.getNode(
         ISD::INTRINSIC_W_CHAIN, DL, MVT::i64, Result,
         DAG.getConstant(Intrinsic::aarch64_sme_get_tpidr2, DL, MVT::i32));
@@ -29035,7 +29036,7 @@ bool AArch64TargetLowering::fallBackToDAGISel(const Instruction &Inst) const {
 
   // Checks to allow the use of SME instructions
   if (auto *Base = dyn_cast<CallBase>(&Inst)) {
-    auto CallAttrs = SMECallAttrs(*Base, this);
+    auto CallAttrs = SMECallAttrs(*Base);
     if (CallAttrs.requiresSMChange() || CallAttrs.requiresLazySave() ||
         CallAttrs.requiresPreservingZT0() ||
         CallAttrs.requiresPreservingAllZAState())

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

@@ -220,16 +220,20 @@ static cl::opt<bool> EnableFixedwidthAutovecInStreamingMode(
 static cl::opt<bool> EnableScalableAutovecInStreamingMode(
     "enable-scalable-autovec-in-streaming-mode", cl::init(false), cl::Hidden);
 
-static bool isSMEABIRoutineCall(const CallInst &CI, const TargetLowering &TLI) {
+static bool isSMEABIRoutineCall(const CallInst &CI) {
   const auto *F = CI.getCalledFunction();
-  return F && SMEAttrs(F->getName(), TLI).isSMEABIRoutine();
+  return F && StringSwitch<bool>(F->getName())
+                  .Case("__arm_sme_state", true)
+                  .Case("__arm_tpidr2_save", true)
+                  .Case("__arm_tpidr2_restore", true)
+                  .Case("__arm_za_disable", true)
+                  .Default(false);
 }
 
 /// Returns true if the function has explicit operations that can only be
 /// lowered using incompatible instructions for the selected mode. This also
 /// returns true if the function F may use or modify ZA state.
-static bool hasPossibleIncompatibleOps(const Function *F,
-                                       const TargetLowering &TLI) {
+static bool hasPossibleIncompatibleOps(const Function *F) {
   for (const BasicBlock &BB : *F) {
     for (const Instruction &I : BB) {
       // Be conservative for now and assume that any call to inline asm or to
@@ -238,7 +242,7 @@ static bool hasPossibleIncompatibleOps(const Function *F,
       // all native LLVM instructions can be lowered to compatible instructions.
       if (isa<CallInst>(I) && !I.isDebugOrPseudoInst() &&
           (cast<CallInst>(I).isInlineAsm() || isa<IntrinsicInst>(I) ||
-           isSMEABIRoutineCall(cast<CallInst>(I), TLI)))
+           isSMEABIRoutineCall(cast<CallInst>(I))))
         return true;
     }
   }
@@ -286,7 +290,7 @@ bool AArch64TTIImpl::areInlineCompatible(const Function *Caller,
   if (CallAttrs.requiresLazySave() || CallAttrs.requiresSMChange() ||
       CallAttrs.requiresPreservingZT0() ||
       CallAttrs.requiresPreservingAllZAState()) {
-    if (hasPossibleIncompatibleOps(Callee, *getTLI()))
+    if (hasPossibleIncompatibleOps(Callee))
       return false;
   }
 
@@ -353,7 +357,7 @@ AArch64TTIImpl::getInlineCallPenalty(const Function *F, const CallBase &Call,
   // change only once and avoid inlining of G into F.
 
   SMEAttrs FAttrs(*F);
-  SMECallAttrs CallAttrs(Call, getTLI());
+  SMECallAttrs CallAttrs(Call);
 
   if (SMECallAttrs(FAttrs, CallAttrs.callee()).requiresSMChange()) {
     if (F == Call.getCaller()) // (1)