[GlobalISel] Add a localizer pass for target to use

This reverts commit r299287 plus clean-ups.

The localizer pass is a helper pass that could be run at O0 in the GISel
pipeline to work around the deficiency of the fast register allocator.
It basically shortens the live-ranges of the constants so that the
allocator does not spill all over the place.

Long term fix would be to make the greedy allocator fast.

llvm-svn: 304051

llvm/include/llvm/CodeGen/GlobalISel/Localizer.h

@@ -0,0 +1,78 @@
+//== llvm/CodeGen/GlobalISel/Localizer.h - Localizer -------------*- C++ -*-==//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file This file describes the interface of the Localizer pass.
+/// This pass moves/duplicates constant-like instructions close to their uses.
+/// Its primarily goal is to workaround the deficiencies of the fast register
+/// allocator.
+/// With GlobalISel constants are all materialized in the entry block of
+/// a function. However, the fast allocator cannot rematerialize constants and
+/// has a lot more live-ranges to deal with and will most likely end up
+/// spilling a lot.
+/// By pushing the constants close to their use, we only create small
+/// live-ranges.
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_CODEGEN_GLOBALISEL_LOCALIZER_H
+#define LLVM_CODEGEN_GLOBALISEL_LOCALIZER_H
+
+#include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+
+namespace llvm {
+// Forward declarations.
+class MachineRegisterInfo;
+
+/// This pass implements the localization mechanism described at the
+/// top of this file. One specificity of the implementation is that
+/// it will materialize one and only one instance of a constant per
+/// basic block, thus enabling reuse of that constant within that block.
+/// Moreover, it only materializes constants in blocks where they
+/// are used. PHI uses are considered happening at the end of the
+/// related predecessor.
+class Localizer : public MachineFunctionPass {
+public:
+  static char ID;
+
+private:
+  /// MRI contains all the register class/bank information that this
+  /// pass uses and updates.
+  MachineRegisterInfo *MRI;
+
+  /// Check whether or not \p MI needs to be moved close to its uses.
+  static bool shouldLocalize(const MachineInstr &MI);
+
+  /// Check if \p MOUse is used in the same basic block as \p Def.
+  /// If the use is in the same block, we say it is local.
+  /// When the use is not local, \p InsertMBB will contain the basic
+  /// block when to insert \p Def to have a local use.
+  static bool isLocalUse(MachineOperand &MOUse, const MachineInstr &Def,
+                         MachineBasicBlock *&InsertMBB);
+
+  /// Initialize the field members using \p MF.
+  void init(MachineFunction &MF);
+
+public:
+  Localizer();
+
+  StringRef getPassName() const override { return "Localizer"; }
+
+  MachineFunctionProperties getRequiredProperties() const override {
+    return MachineFunctionProperties()
+        .set(MachineFunctionProperties::Property::IsSSA)
+        .set(MachineFunctionProperties::Property::Legalized)
+        .set(MachineFunctionProperties::Property::RegBankSelected);
+  }
+
+  bool runOnMachineFunction(MachineFunction &MF) override;
+};
+
+} // End namespace llvm.
+
+#endif

llvm/include/llvm/CodeGen/MachineRegisterInfo.h

@@ -642,6 +642,11 @@ public:
   ///
   void setRegBank(unsigned Reg, const RegisterBank &RegBank);
 
+  void setRegClassOrRegBank(unsigned Reg,
+                            const RegClassOrRegBank &RCOrRB){
+    VRegInfo[Reg].first = RCOrRB;
+  }
+
   /// constrainRegClass - Constrain the register class of the specified virtual
   /// register to be a common subclass of RC and the current register class,
   /// but only if the new class has at least MinNumRegs registers.  Return the

llvm/include/llvm/InitializePasses.h

@@ -194,6 +194,7 @@ void initializeLiveVariablesPass(PassRegistry&);
 void initializeLoadCombinePass(PassRegistry&);
 void initializeLoadStoreVectorizerPass(PassRegistry&);
 void initializeLoaderPassPass(PassRegistry&);
+void initializeLocalizerPass(PassRegistry&);
 void initializeLocalStackSlotPassPass(PassRegistry&);
 void initializeLoopAccessLegacyAnalysisPass(PassRegistry&);
 void initializeLoopDataPrefetchLegacyPassPass(PassRegistry&);

llvm/lib/CodeGen/GlobalISel/CMakeLists.txt

@@ -8,6 +8,7 @@ set(GLOBAL_ISEL_FILES
       LegalizerHelper.cpp
       Legalizer.cpp
       LegalizerInfo.cpp
+      Localizer.cpp
       RegBankSelect.cpp
       RegisterBank.cpp
       RegisterBankInfo.cpp

llvm/lib/CodeGen/GlobalISel/GlobalISel.cpp

@@ -26,6 +26,7 @@ void llvm::initializeGlobalISel(PassRegistry &Registry) {
 void llvm::initializeGlobalISel(PassRegistry &Registry) {
   initializeIRTranslatorPass(Registry);
   initializeLegalizerPass(Registry);
+  initializeLocalizerPass(Registry);
   initializeRegBankSelectPass(Registry);
   initializeInstructionSelectPass(Registry);
 }

llvm/lib/CodeGen/GlobalISel/Localizer.cpp

@@ -0,0 +1,125 @@
+//===- Localizer.cpp ---------------------- Localize some instrs -*- C++ -*-==//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+/// \file
+/// This file implements the Localizer class.
+//===----------------------------------------------------------------------===//
+
+#include "llvm/CodeGen/GlobalISel/Localizer.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/Support/Debug.h"
+
+#define DEBUG_TYPE "localizer"
+
+using namespace llvm;
+
+char Localizer::ID = 0;
+INITIALIZE_PASS(Localizer, DEBUG_TYPE,
+                "Move/duplicate certain instructions close to their use", false,
+                false);
+
+Localizer::Localizer() : MachineFunctionPass(ID) {
+  initializeLocalizerPass(*PassRegistry::getPassRegistry());
+}
+
+void Localizer::init(MachineFunction &MF) { MRI = &MF.getRegInfo(); }
+
+bool Localizer::shouldLocalize(const MachineInstr &MI) {
+  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:
+    return true;
+  }
+}
+
+bool Localizer::isLocalUse(MachineOperand &MOUse, const MachineInstr &Def,
+                           MachineBasicBlock *&InsertMBB) {
+  MachineInstr &MIUse = *MOUse.getParent();
+  InsertMBB = MIUse.getParent();
+  if (MIUse.isPHI())
+    InsertMBB = MIUse.getOperand(MIUse.getOperandNo(&MOUse) + 1).getMBB();
+  return InsertMBB == Def.getParent();
+}
+
+bool Localizer::runOnMachineFunction(MachineFunction &MF) {
+  // If the ISel pipeline failed, do not bother running that pass.
+  if (MF.getProperties().hasProperty(
+          MachineFunctionProperties::Property::FailedISel))
+    return false;
+
+  DEBUG(dbgs() << "Localize instructions for: " << MF.getName() << '\n');
+
+  init(MF);
+
+  bool Changed = false;
+  // Keep track of the instructions we localized.
+  // We won't need to process them if we see them later in the CFG.
+  SmallPtrSet<MachineInstr *, 16> LocalizedInstrs;
+  DenseMap<std::pair<MachineBasicBlock *, unsigned>, unsigned> MBBWithLocalDef;
+  // TODO: Do bottom up traversal.
+  for (MachineBasicBlock &MBB : MF) {
+    for (MachineInstr &MI : MBB) {
+      if (LocalizedInstrs.count(&MI) || !shouldLocalize(MI))
+        continue;
+      DEBUG(dbgs() << "Should localize: " << MI);
+      assert(MI.getDesc().getNumDefs() == 1 &&
+             "More than one definition not supported yet");
+      unsigned 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.
+      for (auto MOIt = MRI->use_begin(Reg), MOItEnd = MRI->use_end();
+           MOIt != MOItEnd;) {
+        MachineOperand &MOUse = *MOIt++;
+        // Check if the use is already local.
+        MachineBasicBlock *InsertMBB;
+        DEBUG(MachineInstr &MIUse = *MOUse.getParent();
+              dbgs() << "Checking use: " << MIUse
+                     << " #Opd: " << MIUse.getOperandNo(&MOUse) << '\n');
+        if (isLocalUse(MOUse, MI, InsertMBB))
+          continue;
+        DEBUG(dbgs() << "Fixing non-local use\n");
+        Changed = true;
+        auto MBBAndReg = std::make_pair(InsertMBB, Reg);
+        auto NewVRegIt = MBBWithLocalDef.find(MBBAndReg);
+        if (NewVRegIt == MBBWithLocalDef.end()) {
+          // Create the localized instruction.
+          MachineInstr *LocalizedMI = MF.CloneMachineInstr(&MI);
+          LocalizedInstrs.insert(LocalizedMI);
+          // Move it at the right place.
+          MachineInstr &MIUse = *MOUse.getParent();
+          if (MIUse.getParent() == InsertMBB)
+            InsertMBB->insert(MIUse, LocalizedMI);
+          else
+            InsertMBB->insert(InsertMBB->getFirstNonPHI(), LocalizedMI);
+
+          // Set a new register for the definition.
+          unsigned NewReg =
+              MRI->createGenericVirtualRegister(MRI->getType(Reg));
+          MRI->setRegClassOrRegBank(NewReg, MRI->getRegClassOrRegBank(Reg));
+          LocalizedMI->getOperand(0).setReg(NewReg);
+          NewVRegIt =
+              MBBWithLocalDef.insert(std::make_pair(MBBAndReg, NewReg)).first;
+          DEBUG(dbgs() << "Inserted: " << *LocalizedMI);
+        }
+        DEBUG(dbgs() << "Update use with: " << PrintReg(NewVRegIt->second)
+                     << '\n');
+        // Update the user reg.
+        MOUse.setReg(NewVRegIt->second);
+      }
+    }
+  }
+  return Changed;
+}

llvm/test/CodeGen/AArch64/GlobalISel/localizer.mir

@@ -0,0 +1,312 @@
+# RUN: llc -O0 -mtriple=aarch64-apple-ios -run-pass=localizer -verify-machineinstrs -global-isel %s -o - | FileCheck %s -check-prefix=CHECK
+
+# Test the localizer.
+
+--- |
+  target datalayout = "e-m:o-i64:64-i128:128-n32:64-S128"
+
+  define void @local_use() { ret void }
+  define void @non_local_1use() { ret void }
+  define void @non_local_2uses() { ret void }
+  define void @non_local_phi_use() { ret void }
+  define void @non_local_phi_use_followed_by_use() { ret void }
+  define void @non_local_phi_use_followed_by_use_fi() { ret void }
+  define void @float_non_local_phi_use_followed_by_use_fi() { ret void }
+...
+
+---
+# CHECK-LABEL: name: local_use
+name:            local_use
+legalized:       true
+regBankSelected: true
+
+# CHECK:      registers:
+registers:
+  - { id: 0, class: gpr }
+  - { id: 1, class: gpr }
+  - { id: 2, class: gpr }
+
+# CHECK:  body:
+# CHECK:    %0(s32) = G_CONSTANT 1
+# CHECK-NEXT: %1(s32) = G_ADD %0, %0
+body:             |
+  bb.0:
+    %0(s32) = G_CONSTANT 1
+    %1(s32) = G_ADD %0, %0
+...
+
+---
+# CHECK-LABEL: name: non_local_1use
+name:            non_local_1use
+legalized:       true
+regBankSelected: true
+
+# CHECK:      registers:
+# Existing registers should be left untouched
+# CHECK:  - { id: 0, class: gpr }
+#CHECK-NEXT:  - { id: 1, class: gpr }
+#CHECK-NEXT:  - { id: 2, class: gpr }
+# The newly created reg should be on the same regbank/regclass as its origin.
+#CHECK-NEXT:  - { id: 3, class: gpr }
+
+registers:
+  - { id: 0, class: gpr }
+  - { id: 1, class: gpr }
+  - { id: 2, class: gpr }
+
+# CHECK:  body:
+# CHECK:    %0(s32) = G_CONSTANT 1
+# CHECK-NEXT: %1(s32) = G_ADD %0, %0
+
+# CHECK: bb.1:
+# CHECK: %3(s32) = G_CONSTANT 1
+# CHECK-NEXT: %2(s32) = G_ADD %3, %1
+body:             |
+  bb.0:
+    successors: %bb.1
+
+    %0(s32) = G_CONSTANT 1
+    %1(s32) = G_ADD %0, %0
+
+  bb.1:
+    %2(s32) = G_ADD %0, %1
+...
+
+
+---
+# CHECK-LABEL: name: non_local_2uses
+name:            non_local_2uses
+legalized:       true
+regBankSelected: true
+
+# CHECK:      registers:
+# Existing registers should be left untouched
+# CHECK:  - { id: 0, class: gpr }
+#CHECK-NEXT:  - { id: 1, class: gpr }
+#CHECK-NEXT:  - { id: 2, class: gpr }
+# The newly created reg should be on the same regbank/regclass as its origin.
+#CHECK-NEXT:  - { id: 3, class: gpr }
+
+registers:
+  - { id: 0, class: gpr }
+  - { id: 1, class: gpr }
+  - { id: 2, class: gpr }
+
+# CHECK:  body:
+# CHECK:    %0(s32) = G_CONSTANT 1
+# CHECK-NEXT: %1(s32) = G_ADD %0, %0
+
+# CHECK: bb.1:
+# CHECK: %3(s32) = G_CONSTANT 1
+# CHECK-NEXT: %2(s32) = G_ADD %3, %3
+body:             |
+  bb.0:
+    successors: %bb.1
+
+    %0(s32) = G_CONSTANT 1
+    %1(s32) = G_ADD %0, %0
+
+  bb.1:
+    %2(s32) = G_ADD %0, %0
+...
+
+---
+# CHECK-LABEL: name: non_local_phi_use
+name:            non_local_phi_use
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+
+# CHECK:      registers:
+# Existing registers should be left untouched
+# CHECK:  - { id: 0, class: gpr }
+#CHECK-NEXT:  - { id: 1, class: gpr }
+#CHECK-NEXT:  - { id: 2, class: gpr }
+#CHECK-NEXT:  - { id: 3, class: gpr }
+#CHECK-NEXT:  - { id: 4, class: gpr }
+# The newly created reg should be on the same regbank/regclass as its origin.
+#CHECK-NEXT:  - { id: 5, class: gpr }
+
+registers:
+  - { id: 0, class: gpr }
+  - { id: 1, class: gpr }
+  - { id: 2, class: gpr }
+  - { id: 3, class: gpr }
+  - { id: 4, class: gpr }
+
+# CHECK:  body:
+# CHECK:    %0(s32) = G_CONSTANT 1
+# CHECK-NEXT: %1(s32) = G_ADD %0, %0
+
+# CHECK: bb.1:
+# CHECK: %5(s32) = G_CONSTANT 1
+
+# CHECK: bb.2:
+# CHECK: %3(s32) = PHI %5(s32), %bb.1
+body:             |
+  bb.0:
+    successors: %bb.1
+
+    %0(s32) = G_CONSTANT 1
+    %1(s32) = G_ADD %0, %0
+
+  bb.1:
+    successors: %bb.2
+
+  bb.2:
+    %3(s32) = PHI %0(s32), %bb.1
+    %2(s32) = G_ADD %3, %3
+...
+
+---
+# CHECK-LABEL: name: non_local_phi_use_followed_by_use
+name:            non_local_phi_use_followed_by_use
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+
+# CHECK:      registers:
+# Existing registers should be left untouched
+# CHECK:  - { id: 0, class: gpr }
+#CHECK-NEXT:  - { id: 1, class: gpr }
+#CHECK-NEXT:  - { id: 2, class: gpr }
+#CHECK-NEXT:  - { id: 3, class: gpr }
+#CHECK-NEXT:  - { id: 4, class: gpr }
+# The newly created regs should be on the same regbank/regclass as its origin.
+#CHECK-NEXT:  - { id: 5, class: gpr }
+#CHECK-NEXT:  - { id: 6, class: gpr }
+
+registers:
+  - { id: 0, class: gpr }
+  - { id: 1, class: gpr }
+  - { id: 2, class: gpr }
+  - { id: 3, class: gpr }
+  - { id: 4, class: gpr }
+
+# CHECK:  body:
+# CHECK:    %0(s32) = G_CONSTANT 1
+# CHECK-NEXT: %1(s32) = G_ADD %0, %0
+
+# CHECK: bb.1:
+# CHECK: %5(s32) = G_CONSTANT 1
+
+# CHECK: bb.2:
+# CHECK: %3(s32) = PHI %5(s32), %bb.1
+# CHECK-NEXT: %6(s32) = G_CONSTANT 1
+# CHECK-NEXT: %2(s32) = G_ADD %3, %6
+body:             |
+  bb.0:
+    successors: %bb.1
+
+    %0(s32) = G_CONSTANT 1
+    %1(s32) = G_ADD %0, %0
+
+  bb.1:
+    successors: %bb.2
+
+  bb.2:
+    %3(s32) = PHI %0(s32), %bb.1
+    %2(s32) = G_ADD %3, %0
+...
+
+---
+# CHECK-LABEL: name: non_local_phi_use_followed_by_use_fi
+name:            non_local_phi_use_followed_by_use_fi
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+
+# CHECK:      registers:
+# Existing registers should be left untouched
+# CHECK:  - { id: 0, class: gpr }
+#CHECK-NEXT:  - { id: 1, class: gpr }
+#CHECK-NEXT:  - { id: 2, class: gpr }
+#CHECK-NEXT:  - { id: 3, class: gpr }
+#CHECK-NEXT:  - { id: 4, class: gpr }
+# The newly created reg should be on the same regbank/regclass as its origin.
+#CHECK-NEXT:  - { id: 5, class: gpr }
+#CHECK-NEXT:  - { id: 6, class: gpr }
+
+registers:
+  - { id: 0, class: gpr }
+  - { id: 1, class: gpr }
+  - { id: 2, class: gpr }
+  - { id: 3, class: gpr }
+  - { id: 4, class: gpr }
+
+# CHECK:  body:
+# CHECK:    %0(s32) = G_FRAME_INDEX 1
+# CHECK-NEXT: %1(s32) = G_ADD %0, %0
+
+# CHECK: bb.1:
+# CHECK: %5(s32) = G_FRAME_INDEX 1
+
+# CHECK: bb.2:
+# CHECK: %3(s32) = PHI %5(s32), %bb.1
+# CHECK-NEXT: %6(s32) = G_FRAME_INDEX 1
+# CHECK-NEXT: %2(s32) = G_ADD %3, %6
+body:             |
+  bb.0:
+    successors: %bb.1
+
+    %0(s32) = G_FRAME_INDEX 1
+    %1(s32) = G_ADD %0, %0
+
+  bb.1:
+    successors: %bb.2
+
+  bb.2:
+    %3(s32) = PHI %0(s32), %bb.1
+    %2(s32) = G_ADD %3, %0
+...
+
+---
+# CHECK-LABEL: name: float_non_local_phi_use_followed_by_use_fi
+name:            float_non_local_phi_use_followed_by_use_fi
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+
+# CHECK:      registers:
+# Existing registers should be left untouched
+# CHECK:  - { id: 0, class: fpr }
+#CHECK-NEXT:  - { id: 1, class: fpr }
+#CHECK-NEXT:  - { id: 2, class: fpr }
+#CHECK-NEXT:  - { id: 3, class: fpr }
+#CHECK-NEXT:  - { id: 4, class: fpr }
+# The newly created reg should be on the same regbank/regclass as its origin.
+#CHECK-NEXT:  - { id: 5, class: fpr }
+#CHECK-NEXT:  - { id: 6, class: fpr }
+
+registers:
+  - { id: 0, class: fpr }
+  - { id: 1, class: fpr }
+  - { id: 2, class: fpr }
+  - { id: 3, class: fpr }
+  - { id: 4, class: fpr }
+
+# CHECK:  body:
+# CHECK:    %0(s32) = G_FCONSTANT float 1.0
+# CHECK-NEXT: %1(s32) = G_FADD %0, %0
+
+# CHECK: bb.1:
+# CHECK: %5(s32) = G_FCONSTANT float 1.0
+
+# CHECK: bb.2:
+# CHECK: %3(s32) = PHI %5(s32), %bb.1
+# CHECK-NEXT: %6(s32) = G_FCONSTANT float 1.0
+# CHECK-NEXT: %2(s32) = G_FADD %3, %6
+body:             |
+  bb.0:
+    successors: %bb.1
+
+    %0(s32) = G_FCONSTANT float 1.0
+    %1(s32) = G_FADD %0, %0
+
+  bb.1:
+    successors: %bb.2
+
+  bb.2:
+    %3(s32) = PHI %0(s32), %bb.1
+    %2(s32) = G_FADD %3, %0
+...