GlobalISel: Move Localizer::shouldLocalize(..) to TargetLowering

[volkan]

Summary

Add a new target hook for shouldLocalize so that
targets can customize the logic.

[aemerson]

The change itself is ok but is there any way to have more granular overriding of this behavior, instead of overriding the whole shouldLocalize()?

[volkan]

In D75207#1894581, @aemerson wrote:

The change itself is ok but is there any way to have more granular overriding of this behavior, instead of overriding the whole shouldLocalize()?

We can have two functions (one for checking the opcode and another one for the cost), but this doesn't allow targets to localize target specific instruction. That's why I decided move the whole thing to TargetLowering. Do you think having two targets hooks would be better?

[aemerson]

In D75207#1896372, @volkan wrote:

In D75207#1894581, @aemerson wrote:

The change itself is ok but is there any way to have more granular overriding of this behavior, instead of overriding the whole shouldLocalize()?

We can have two functions (one for checking the opcode and another one for the cost), but this doesn't allow targets to localize target specific instruction. That's why I decided move the whole thing to TargetLowering. Do you think having two targets hooks would be better?

I suppose this is fine, overriding implementations can just call this base class version if the opcode is something they don't want to override.

[volkan]

Committed in 4167645d1e6a5ecc8790f0aba450799c4441882d.

llvm/include/llvm/CodeGen/TargetLowering.h

public:
   }
 
   /// Execute target specific actions to finalize target lowering.
   /// This is used to set extra flags in MachineFrameInformation and freezing
   /// the set of reserved registers.
   /// The default implementation just freezes the set of reserved registers.
   virtual void finalizeLowering(MachineFunction &MF) const;
 
+  //===----------------------------------------------------------------------===//
+  //  GlobalISel Hooks
+  //===----------------------------------------------------------------------===//
+  /// Check whether or not \p MI needs to be moved close to its uses.
+  virtual bool shouldLocalize(const MachineInstr &MI, const TargetTransformInfo *TTI) const;
+
+
 private:
   const TargetMachine &TM;
 
   /// Tells the code generator that the target has multiple (allocatable)
   /// condition registers that can be used to store the results of comparisons
   /// for use by selects and conditional branches. With multiple condition
   /// registers, the code generator will not aggressively sink comparisons into
   /// the blocks of their users.

llvm/lib/CodeGen/GlobalISel/Localizer.cpp

 //===- Localizer.cpp ---------------------- Localize some instrs -*- C++ -*-==//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 /// \file
 /// This file implements the Localizer class.
 //===----------------------------------------------------------------------===//
 
 #include "llvm/CodeGen/GlobalISel/Localizer.h"
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/TargetLowering.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/Support/Debug.h"
 
 #define DEBUG_TYPE "localizer"
 
 using namespace llvm;
 
 char Localizer::ID = 0;

 Localizer::Localizer()
     : Localizer([](const MachineFunction &) { return false; }) {}
 
 void Localizer::init(MachineFunction &MF) {
   MRI = &MF.getRegInfo();
   TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(MF.getFunction());
 }
 
-bool Localizer::shouldLocalize(const MachineInstr &MI) {
-  // Assuming a spill and reload of a value has a cost of 1 instruction each,
-  // this helper function computes the maximum number of uses we should consider
-  // for remat. E.g. on arm64 global addresses take 2 insts to materialize. We
-  // break even in terms of code size when the original MI has 2 users vs
-  // choosing to potentially spill. Any more than 2 users we we have a net code
-  // size increase. This doesn't take into account register pressure though.
-  auto maxUses = [](unsigned RematCost) {
-    // A cost of 1 means remats are basically free.
-    if (RematCost == 1)
-      return UINT_MAX;
-    if (RematCost == 2)
-      return 2U;
-
-    // Remat is too expensive, only sink if there's one user.
-    if (RematCost > 2)
-      return 1U;
-    llvm_unreachable("Unexpected remat cost");
-  };
-
-  // Helper to walk through uses and terminate if we've reached a limit. Saves
-  // us spending time traversing uses if all we want to know is if it's >= min.
-  auto isUsesAtMost = [&](unsigned Reg, unsigned MaxUses) {
-    unsigned NumUses = 0;
-    auto UI = MRI->use_instr_nodbg_begin(Reg), UE = MRI->use_instr_nodbg_end();
-    for (; UI != UE && NumUses < MaxUses; ++UI) {
-      NumUses++;
-    }
-    // If we haven't reached the end yet then there are more than MaxUses users.
-    return UI == UE;
-  };
-
-  switch (MI.getOpcode()) {
-  default:
-    return false;
-  // Constants-like instructions should be close to their users.
-  // We don't want long live-ranges for them.
-  case TargetOpcode::G_CONSTANT:
-  case TargetOpcode::G_FCONSTANT:
-  case TargetOpcode::G_FRAME_INDEX:
-  case TargetOpcode::G_INTTOPTR:
-    return true;
-  case TargetOpcode::G_GLOBAL_VALUE: {
-    unsigned RematCost = TTI->getGISelRematGlobalCost();
-    Register Reg = MI.getOperand(0).getReg();
-    unsigned MaxUses = maxUses(RematCost);
-    if (MaxUses == UINT_MAX)
-      return true; // Remats are "free" so always localize.
-    bool B = isUsesAtMost(Reg, MaxUses);
-    return B;
-  }
-  }
-}
-
 void Localizer::getAnalysisUsage(AnalysisUsage &AU) const {
   AU.addRequired<TargetTransformInfoWrapperPass>();
   getSelectionDAGFallbackAnalysisUsage(AU);
   MachineFunctionPass::getAnalysisUsage(AU);
 }
 
 bool Localizer::isLocalUse(MachineOperand &MOUse, const MachineInstr &Def,
                            MachineBasicBlock *&InsertMBB) {
bool Localizer::localizeInterBlock(MachineFunction &MF,
   bool Changed = false;
   DenseMap<std::pair<MachineBasicBlock *, unsigned>, unsigned> MBBWithLocalDef;
 
   // Since the IRTranslator only emits constants into the entry block, and the
   // rest of the GISel pipeline generally emits constants close to their users,
   // we only localize instructions in the entry block here. This might change if
   // we start doing CSE across blocks.
   auto &MBB = MF.front();
+  auto &TL = *MF.getSubtarget().getTargetLowering();
   for (auto RI = MBB.rbegin(), RE = MBB.rend(); RI != RE; ++RI) {
     MachineInstr &MI = *RI;
-    if (!shouldLocalize(MI))
+    if (!TL.shouldLocalize(MI, TTI))
       continue;
     LLVM_DEBUG(dbgs() << "Should localize: " << MI);
     assert(MI.getDesc().getNumDefs() == 1 &&
            "More than one definition not supported yet");
     Register Reg = MI.getOperand(0).getReg();
     // Check if all the users of MI are local.
     // We are going to invalidation the list of use operands, so we
     // can't use range iterator.

llvm/lib/CodeGen/TargetLoweringBase.cpp

 #include "llvm/ADT/BitVector.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringExtras.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/ADT/Triple.h"
 #include "llvm/ADT/Twine.h"
 #include "llvm/Analysis/Loads.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/CodeGen/Analysis.h"
 #include "llvm/CodeGen/ISDOpcodes.h"
 #include "llvm/CodeGen/MachineBasicBlock.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/MachineInstr.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineMemOperand.h"
if (CmpX->isVolatile())
       Flags |= MachineMemOperand::MOVolatile;
   } else
     llvm_unreachable("not an atomic instruction");
 
   // FIXME: Not preserving dereferenceable
   Flags |= getTargetMMOFlags(AI);
   return Flags;
 }
+
+//===----------------------------------------------------------------------===//
+//  GlobalISel Hooks
+//===----------------------------------------------------------------------===//
+
+bool TargetLoweringBase::shouldLocalize(const MachineInstr &MI,
+                                        const TargetTransformInfo *TTI) const {
+  auto &MF = *MI.getMF();
+  auto &MRI = MF.getRegInfo();
+  // Assuming a spill and reload of a value has a cost of 1 instruction each,
+  // this helper function computes the maximum number of uses we should consider
+  // for remat. E.g. on arm64 global addresses take 2 insts to materialize. We
+  // break even in terms of code size when the original MI has 2 users vs
+  // choosing to potentially spill. Any more than 2 users we we have a net code
+  // size increase. This doesn't take into account register pressure though.
+  auto maxUses = [](unsigned RematCost) {
+    // A cost of 1 means remats are basically free.
+    if (RematCost == 1)
+      return UINT_MAX;
+    if (RematCost == 2)
+      return 2U;
+
+    // Remat is too expensive, only sink if there's one user.
+    if (RematCost > 2)
+      return 1U;
+    llvm_unreachable("Unexpected remat cost");
+  };
+
+  // Helper to walk through uses and terminate if we've reached a limit. Saves
+  // us spending time traversing uses if all we want to know is if it's >= min.
+  auto isUsesAtMost = [&](unsigned Reg, unsigned MaxUses) {
+    unsigned NumUses = 0;
+    auto UI = MRI.use_instr_nodbg_begin(Reg), UE = MRI.use_instr_nodbg_end();
+    for (; UI != UE && NumUses < MaxUses; ++UI) {
+      NumUses++;
+    }
+    // If we haven't reached the end yet then there are more than MaxUses users.
+    return UI == UE;
+  };
+
+  switch (MI.getOpcode()) {
+  default:
+    return false;
+  // Constants-like instructions should be close to their users.
+  // We don't want long live-ranges for them.
+  case TargetOpcode::G_CONSTANT:
+  case TargetOpcode::G_FCONSTANT:
+  case TargetOpcode::G_FRAME_INDEX:
+  case TargetOpcode::G_INTTOPTR:
+    return true;
+  case TargetOpcode::G_GLOBAL_VALUE: {
+    unsigned RematCost = TTI->getGISelRematGlobalCost();
+    Register Reg = MI.getOperand(0).getReg();
+    unsigned MaxUses = maxUses(RematCost);
+    if (MaxUses == UINT_MAX)
+      return true; // Remats are "free" so always localize.
+    bool B = isUsesAtMost(Reg, MaxUses);
+    return B;
+  }
+  }
+}