[RISCV] Add machine function pass to merge base + offset

[sabuasal]

Summary

In r333455 we added a peephole to fix the corner cases that result
from separating base + offset lowering of global address.The
peephole didn't handle some of the cases because it only has a basic
block view instead of a function level view.

This patch replaces that logic with a machine function pass. In
addition to handling the original cases it handles uses of the global
address across blocks in function and folding an offset from LW\SW
instruction. This pass won't run for OptNone compilation, so there
will be a negative impact overall vs the old approach at O0.

[asb]

Thanks, I'll have to finish stepping through the code tomorrow but this seems to be working great. For the GCC torture suite at O1, ~40 of the test cases see improved codegen (fewer instructions). Two gain a single instruction, but that's just noise due to slight regalloc differences.

One change it would be worth noting in the commit message / patch description: this pass won't run for OptNone compilation, so there will be a negative impact overall vs the old approach at O0. Not a problem - having this skipped at O0 makes sense, but worth mentioning all the same. Arguably the RISCVISelDAGToDAG peepholes should be skipped for OptNone too.

lib/Target/RISCV/RISCVMergeBaseOffset.cpp

1

This line is only 79 chars long. Probably nicer to keep the usual format (three =) and have a shorter description.

e.g. //== RISCVMergeBaseOffset.cpp - Fold offset into address lowering sequence ==//

143

Given that we know the MI is an Add, is looping over the operands the clearest way of handling this?

[asb]

lib/Target/RISCV/RISCVMergeBaseOffset.cpp

1

Oops, wrong paste in that comment. Corrected:

//===-- RISCVMergeBaseOffset.cpp - Optimise address calculations ---------===//

[asb]

Thanks, this looks good to me assuming the minor nits are addressed. The only one requiring code changes is the for loop in matchLargeOffset.

lib/Target/RISCV/RISCVMergeBaseOffset.cpp

25

There are a few miscellaneous minor formatting issues in this file that will be fixed with clang-format

28–38

A number of these includes seem totally unused. e.g. STLExtras., RISCVTargetObjctFile.h, TargetLoweringObjectFileIMpl.h

[sabuasal]

lib/Target/RISCV/RISCVMergeBaseOffset.cpp

1

I actually spent a long time pondering the cleanest way to add this comment, yours is better so I'l use it :)

28–38

Thanks for the comment, yeah I believe I need a second look here

143

I need to find the input argument that is coming from the offset creation; which can be a LUI or ADDI followed by LUI as explained in the comment above the function.

Do you think there is a cleaner way to do it?

[asb]

lib/Target/RISCV/RISCVMergeBaseOffset.cpp

143

I was imagining removing the loop with something like the below:

assert((TailAdd.getOpcode() == RISCV::ADD) && "Expected ADD instruction!");
 unsigned Rs = TailAdd.getOperand(1).getReg();
 unsigned Rt = TailAdd.getOperand(2).getReg();
 unsigned Reg = Rs == GAReg ? Rt : Rs;

 // Can't fold if the register has more than one use.
 if (!MRI->hasOneUse(Reg))
   return false;
 // This can point to an ADDI or a LUI:
 MachineInstr &OffsetTail = *MRI->getVRegDef(Reg);
 if (OffsetTail.getOpcode() == RISCV::ADDI) {
   // The offset value has non zero bits in both %hi and %lo parts.
   // Detect an ADDI that feeds from a LUI instruction.
   MachineOperand &AddiImmOp = OffsetTail.getOperand(2);
   if (AddiImmOp.getTargetFlags() != RISCVII::MO_None)
     return false;
   int64_t OffLo = AddiImmOp.getImm();
   MachineInstr &OffsetLui =
       *MRI->getVRegDef(OffsetTail.getOperand(1).getReg());
   MachineOperand &LuiImmOp = OffsetLui.getOperand(1);
   if (OffsetLui.getOpcode() != RISCV::LUI ||
       LuiImmOp.getTargetFlags() != RISCVII::MO_None ||
       !MRI->hasOneUse(OffsetLui.getOperand(0).getReg()))
     return false;
   int64_t OffHi = OffsetLui.getOperand(1).getImm();
   Offset = (OffHi << 12) + OffLo;
   LLVM_DEBUG(dbgs() << "  Offset Instrs: " << OffsetTail
                     << "                 " << OffsetLui);
   DeadInstrs.insert(&OffsetTail);
   DeadInstrs.insert(&OffsetLui);
   return true;
 } else if (OffsetTail.getOpcode() == RISCV::LUI) {
   // The offset value has all zero bits in the lower 12 bits. Only LUI
   // exists.
   LLVM_DEBUG(dbgs() << "  Offset Instr: " << OffsetTail);
   Offset = OffsetTail.getOperand(1).getImm() << 12;
   DeadInstrs.insert(&OffsetTail);
   return true;
 }
 return false;

[sabuasal]

Updated comments by @asb .

lib/Target/RISCV/CMakeLists.txt

 
 add_llvm_target(RISCVCodeGen
   RISCVAsmPrinter.cpp
   RISCVFrameLowering.cpp
   RISCVInstrInfo.cpp
   RISCVISelDAGToDAG.cpp
   RISCVISelLowering.cpp
   RISCVMCInstLower.cpp
+  RISCVMergeBaseOffset.cpp
   RISCVRegisterInfo.cpp
   RISCVSubtarget.cpp
   RISCVTargetMachine.cpp
   RISCVTargetObjectFile.cpp
   )
 
 add_subdirectory(AsmParser)
 add_subdirectory(Disassembler)
 add_subdirectory(InstPrinter)
 add_subdirectory(MCTargetDesc)
 add_subdirectory(TargetInfo)

lib/Target/RISCV/RISCV.h

 namespace llvm {
 class RISCVTargetMachine;
 class AsmPrinter;
 class FunctionPass;
 class MCInst;
 class MCOperand;
 class MachineInstr;
 class MachineOperand;
+class PassRegistry;
 
 void LowerRISCVMachineInstrToMCInst(const MachineInstr *MI, MCInst &OutMI,
                                     const AsmPrinter &AP);
 bool LowerRISCVMachineOperandToMCOperand(const MachineOperand &MO,
                                          MCOperand &MCOp, const AsmPrinter &AP);
 
 FunctionPass *createRISCVISelDag(RISCVTargetMachine &TM);
+
+FunctionPass *createRISCVMergeBaseOffsetOptPass();
+void initializeRISCVMergeBaseOffsetOptPass(PassRegistry &);
 }
 
 #endif

lib/Target/RISCV/RISCVISelDAGToDAG.cpp

public:
 
   bool SelectAddrFI(SDValue Addr, SDValue &Base);
 
 // Include the pieces autogenerated from the target description.
 #include "RISCVGenDAGISel.inc"
 
 private:
   void doPeepholeLoadStoreADDI();
-  void doPeepholeGlobalAddiLuiOffset();
   void doPeepholeBuildPairF64SplitF64();
 };
 }
 
 void RISCVDAGToDAGISel::PostprocessISelDAG() {
   doPeepholeLoadStoreADDI();
-  doPeepholeGlobalAddiLuiOffset();
   doPeepholeBuildPairF64SplitF64();
 }
 
 void RISCVDAGToDAGISel::Select(SDNode *Node) {
   unsigned Opcode = Node->getOpcode();
   MVT XLenVT = Subtarget->getXLenVT();
 
   // If we have a custom node, we have already selected

 bool RISCVDAGToDAGISel::SelectAddrFI(SDValue Addr, SDValue &Base) {
   if (auto FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
     Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), Subtarget->getXLenVT());
     return true;
   }
   return false;
 }
 
-// Detect the pattern lui %hi(global) --> ADDI %lo(global)
-//                          HiLUI          LoADDI
-static bool detectLuiAddiGlobal(SDNode *Tail, unsigned &Idx, SDValue &LoADDI,
-                                SDValue &HiLUI, GlobalAddressSDNode *&GAlo,
-                                GlobalAddressSDNode *&GAhi) {
-  // Try to detect the pattern on every operand of the tail instruction.
-  for (Idx = 0; Idx < Tail->getNumOperands(); Idx++) {
-    LoADDI = Tail->getOperand(Idx);
-    // LoADDI should only be used by one instruction (Tail).
-    if (!LoADDI->isMachineOpcode() ||
-        !(LoADDI->getMachineOpcode() == RISCV::ADDI) ||
-        !isa<GlobalAddressSDNode>(LoADDI->getOperand(1)) ||
-        !LoADDI->hasOneUse())
-      continue;
-    // Check for existence of %lo target flag.
-    GAlo = cast<GlobalAddressSDNode>(LoADDI->getOperand(1));
-    if (!(GAlo->getTargetFlags() == RISCVII::MO_LO) ||
-        !(GAlo->getOffset() == 0))
-      return false;
-    // Check for existence of %hi target flag.
-    HiLUI = LoADDI->getOperand(0);
-    if (!HiLUI->isMachineOpcode() ||
-        !(HiLUI->getMachineOpcode() == RISCV::LUI) ||
-        !isa<GlobalAddressSDNode>(HiLUI->getOperand(0)) || !HiLUI->hasOneUse())
-      return false;
-    GAhi = cast<GlobalAddressSDNode>(HiLUI->getOperand(0));
-    if (!(GAhi->getTargetFlags() == RISCVII::MO_HI) ||
-        !(GAhi->getOffset() == 0))
-      return false;
-    return true;
-  }
-  return false;
-}
-
-static bool matchLuiOffset(SDValue &OffsetLUI, int64_t &Offset) {
-  if (!OffsetLUI->isMachineOpcode() ||
-      !(OffsetLUI->getMachineOpcode() == RISCV::LUI) ||
-      !isa<ConstantSDNode>(OffsetLUI->getOperand(0)))
-    return false;
-  Offset = cast<ConstantSDNode>(OffsetLUI->getOperand(0))->getSExtValue();
-  Offset = Offset << 12;
-  LLVM_DEBUG(dbgs() << " Detected \" LUI Offset_hi\"\n");
-  return true;
-}
-
-static bool matchAddiLuiOffset(SDValue &OffsetLoADDI, int64_t &Offset) {
-  // LoADDI should only be used by the tail instruction only.
-  if (!OffsetLoADDI->isMachineOpcode() ||
-      !(OffsetLoADDI->getMachineOpcode() == RISCV::ADDI) ||
-      !isa<ConstantSDNode>(OffsetLoADDI->getOperand(1)) ||
-      !OffsetLoADDI->hasOneUse())
-    return false;
-  int64_t OffLo =
-      cast<ConstantSDNode>(OffsetLoADDI->getOperand(1))->getZExtValue();
-  // HiLUI should only be used by the loADDI.
-  SDValue OffsetHiLUI = (OffsetLoADDI->getOperand(0));
-  if (!OffsetHiLUI->isMachineOpcode() ||
-      !(OffsetHiLUI->getMachineOpcode() == RISCV::LUI) ||
-      !isa<ConstantSDNode>(OffsetHiLUI->getOperand(0)) ||
-      !OffsetHiLUI->hasOneUse())
-    return false;
-  int64_t OffHi =
-      cast<ConstantSDNode>(OffsetHiLUI->getOperand(0))->getSExtValue();
-  Offset = (OffHi << 12) + OffLo;
-  LLVM_DEBUG(dbgs() << " Detected \" ADDI (LUI Offset_hi), Offset_lo\"\n");
-  return true;
-}
-
-static void updateTailInstrUsers(SDNode *Tail, SelectionDAG *CurDAG,
-                                 GlobalAddressSDNode *GAhi,
-                                 GlobalAddressSDNode *GAlo,
-                                 SDValue &GlobalHiLUI, SDValue &GlobalLoADDI,
-                                 int64_t Offset) {
-  // Update the offset in GAhi and GAlo.
-  SDLoc DL(Tail->getOperand(1));
-  SDValue GAHiNew = CurDAG->getTargetGlobalAddress(GAhi->getGlobal(), DL,
-                                                   GlobalHiLUI.getValueType(),
-                                                   Offset, RISCVII::MO_HI);
-  SDValue GALoNew = CurDAG->getTargetGlobalAddress(GAlo->getGlobal(), DL,
-                                                   GlobalLoADDI.getValueType(),
-                                                   Offset, RISCVII::MO_LO);
-  CurDAG->UpdateNodeOperands(GlobalHiLUI.getNode(), GAHiNew);
-  CurDAG->UpdateNodeOperands(GlobalLoADDI.getNode(), GlobalHiLUI, GALoNew);
-  // Update all uses of the Tail with the GlobalLoADDI. After
-  // this Tail will be a dead node.
-  SDValue From = SDValue(Tail, 0);
-  CurDAG->ReplaceAllUsesOfValuesWith(&From, &GlobalLoADDI, 1);
-}
-
-// TODO: This transformation might be better implemeted in a Machine Funtion
-// Pass as discussed here: https://reviews.llvm.org/D45748.
-//
-// Merge the offset of address calculation into the offset field
-// of a global address node in a global address lowering sequence ("LUI
-// %hi(global) --> add %lo(global)") under the following conditions: 1) The
-// offset field in the global address lowering sequence is zero. 2) The lowered
-// global address is only used in one node, referred to as "Tail".
-
-// This peephole does the following transformations to merge the offset:
-
-// 1) ADDI (ADDI (LUI %hi(global)) %lo(global)), offset
-//     --->
-//      ADDI (LUI %hi(global + offset)) %lo(global + offset).
-//
-//  This generates:
-//  lui a0, hi (global + offset)
-//  add a0, a0, lo (global + offset)
-//  Instead of
-//  lui a0, hi (global)
-//  addi a0, hi (global)
-//  addi a0, offset
-// This pattern is for cases when the offset is small enough to fit in the
-// immediate filed of ADDI (less than 12 bits).
-
-// 2) ADD ((ADDI (LUI %hi(global)) %lo(global)), (LUI hi_offset))
-//   --->
-//    offset = hi_offset << 12
-//    ADDI (LUI %hi(global + offset)) %lo(global + offset)
-
-//  Which generates the ASM:
-//  lui  a0, hi(global + offset)
-//  addi a0, lo(global + offset)
-// Instead of:
-//  lui  a0, hi(global)
-//  addi a0, lo(global)
-//  lui a1, (offset)
-//  add a0, a0, a1
-
-// This pattern is for cases when the offset doesn't fit in an immediate field
-// of ADDI but the lower 12 bits are all zeros.
-
-// 3) ADD ((ADDI (LUI %hi(global)) %lo(global)), (ADDI lo_offset, (LUI
-// hi_offset)))
-//     --->
-//        ADDI (LUI %hi(global + offset)) %lo(global + offset)
-// Which generates the ASM:
-// lui  a1, %hi(global + offhi20<<12 + offlo12)
-// addi a1, %lo(global + offhi20<<12 + offlo12)
-// Instead of:
-//  lui  a0, hi(global)
-//  addi a0, lo(global)
-//  lui a1, (offhi20)
-//  addi a1, (offlo12)
-//  add a0, a0, a1
-// This pattern is for cases when the offset doesn't fit in an immediate field
-// of ADDI and both the lower 1 bits and high 20 bits are non zero.
-void RISCVDAGToDAGISel::doPeepholeGlobalAddiLuiOffset() {
-  SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
-  ++Position;
-  SelectionDAG::allnodes_iterator Begin(CurDAG->allnodes_begin());
-  while (Position != Begin) {
-    SDNode *Tail = &*--Position;
-    // Skip dead nodes and any non-machine opcodes.
-    if (Tail->use_empty() || !Tail->isMachineOpcode())
-      continue;
-    // The tail instruction can be an ADD or an ADDI.
-    if (!Tail->isMachineOpcode() || !(Tail->getMachineOpcode() == RISCV::ADD ||
-                                      Tail->getMachineOpcode() == RISCV::ADDI))
-      continue;
-    // First detect the global address part of pattern:
-    // (lui %hi(global) --> Addi %lo(global))
-    unsigned GlobalLoADDiIdx;
-    SDValue GlobalLoADDI;
-    SDValue GlobalHiLUI;
-    GlobalAddressSDNode *GAhi;
-    GlobalAddressSDNode *GAlo;
-    if (!detectLuiAddiGlobal(Tail, GlobalLoADDiIdx, GlobalLoADDI, GlobalHiLUI,
-                             GAlo, GAhi))
-      continue;
-    LLVM_DEBUG(dbgs() << " Detected \"ADDI LUI %hi(global), %lo(global)\n");
-    // Detect the offset part for the address calculation by looking at the
-    // other operand of the tail instruction:
-    int64_t Offset;
-    if (Tail->getMachineOpcode() == RISCV::ADD) {
-      // If the Tail is an ADD instruction, the offset can be in two forms:
-      // 1) LUI hi_Offset followed by:
-      //    ADDI lo_offset
-      //    This happens in case the offset has non zero bits in
-      //    both hi 20 and lo 12 bits.
-      // 2) LUI (offset20)
-      //    This happens in case the lower 12 bits of the offset are zeros.
-      SDValue OffsetVal = Tail->getOperand(1 - GlobalLoADDiIdx);
-      if (!matchAddiLuiOffset(OffsetVal, Offset) &&
-          !matchLuiOffset(OffsetVal, Offset))
-        continue;
-    } else
-      // The Tail is an ADDI instruction:
-      Offset = cast<ConstantSDNode>(Tail->getOperand(1 - GlobalLoADDiIdx))
-                   ->getSExtValue();
-
-    LLVM_DEBUG(
-        dbgs()
-        << " Fold offset value into global offset of LUI %hi and ADDI %lo\n");
-    LLVM_DEBUG(dbgs() << "\tTail:");
-    LLVM_DEBUG(Tail->dump(CurDAG));
-    LLVM_DEBUG(dbgs() << "\tGlobalHiLUI:");
-    LLVM_DEBUG(GlobalHiLUI->dump(CurDAG));
-    LLVM_DEBUG(dbgs() << "\tGlobalLoADDI:");
-    LLVM_DEBUG(GlobalLoADDI->dump(CurDAG));
-    LLVM_DEBUG(dbgs() << "\n");
-    updateTailInstrUsers(Tail, CurDAG, GAhi, GAlo, GlobalHiLUI, GlobalLoADDI,
-                         Offset);
-  }
-  CurDAG->RemoveDeadNodes();
-}
-
 // Merge an ADDI into the offset of a load/store instruction where possible.
 // (load (add base, off), 0) -> (load base, off)
 // (store val, (add base, off)) -> (store val, base, off)
 void RISCVDAGToDAGISel::doPeepholeLoadStoreADDI() {
   SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
   ++Position;
 
   while (Position != CurDAG->allnodes_begin()) {

lib/Target/RISCV/RISCVMergeBaseOffset.cpp

+//== RISCVMergeBaseOffset.cpp - Fold offset into address lowering sequence ==//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// Merge the offset of address calculation into the offset field
+// of instructions in a global address lowering sequence. This pass transforms:
+//   lui  vreg1, %hi(s)
+//   addi vreg2, vreg1, %lo(s)
+//   addi vreg3, verg2, Offset
+//
+//   Into:
+//   lui  vreg1, %hi(s+Offset)
+//   addi vreg2, vreg1, %lo(s+Offset)
+//
+// The transformation is carried out under certain conditions:
+// 1) The offset field in the base of global address lowering sequence is zero.
+// 2) The lowered global address has only one use.
+//
+// The offset field can be in a different form. This pass handles all of them.
+//===----------------------------------------------------------------------===//
+
+#include "RISCV.h"
+#include "RISCVTargetMachine.h"
+#include "RISCVTargetObjectFile.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/CodeGen/Passes.h"
+#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
+#include "llvm/CodeGen/TargetPassConfig.h"
+#include "llvm/IR/LegacyPassManager.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/FormattedStream.h"
+#include "llvm/Support/TargetRegistry.h"
+#include "llvm/Target/TargetOptions.h"
+#include <set>
+using namespace llvm;
+
+#define DEBUG_TYPE "riscv-merge-base-offset"
+#define RISCV_MERGE_BASE_OFFSET_NAME "RISCV Merge Base Offset"
+namespace {
+
+struct RISCVMergeBaseOffsetOpt : public MachineFunctionPass {
+  static char ID;
+  const MachineFunction *MF;
+  bool runOnMachineFunction(MachineFunction &Fn) override;
+  bool detectLuiAddiGlobal(MachineInstr &LUI, MachineInstr *&ADDI);
+
+  bool detectAndFoldOffset(MachineInstr &HiLUI, MachineInstr &LoADDI);
+  void foldOffset(MachineInstr &HiLUI, MachineInstr &LoADDI, MachineInstr &Tail,
+                  int64_t Offset);
+  bool matchLargeOffset(MachineInstr &TailAdd, unsigned GSReg, int64_t &Offset);
+  RISCVMergeBaseOffsetOpt() : MachineFunctionPass(ID) {}
+
+  MachineFunctionProperties getRequiredProperties() const override {
+    return MachineFunctionProperties().set(
+        MachineFunctionProperties::Property::IsSSA);
+  }
+
+  StringRef getPassName() const override {
+    return RISCV_MERGE_BASE_OFFSET_NAME;
+  }
+
+private:
+  MachineRegisterInfo *MRI;
+  std::set<MachineInstr *> DeadInstrs;
+};
+}; // end anonymous namespace
+
+char RISCVMergeBaseOffsetOpt::ID = 0;
+INITIALIZE_PASS(RISCVMergeBaseOffsetOpt, "riscv-merge-base-offset",
+                RISCV_MERGE_BASE_OFFSET_NAME, false, false)
+
+// Detect the pattern:
+//   lui   vreg1, %hi(s)
+//   addi  vreg2, vreg1, %lo(s)
+//
+//   Pattern only accepted if:
+//     1) ADDI has only one use.
+//     2) LUI has only one use; which is the ADDI.
+//     3) Both ADDI and LUI have GlobalAddress type which indicates that these
+//        are generated from global address lowering.
+//     4) Offset value in the Global Address is 0.
+bool RISCVMergeBaseOffsetOpt::detectLuiAddiGlobal(MachineInstr &HiLUI,
+                                                  MachineInstr *&LoADDI) {
+  if (HiLUI.getOpcode() != RISCV::LUI ||
+      HiLUI.getOperand(1).getTargetFlags() != RISCVII::MO_HI ||
+      HiLUI.getOperand(1).getType() != MachineOperand::MO_GlobalAddress ||
+      HiLUI.getOperand(1).getOffset() != 0 ||
+      !MRI->hasOneUse(HiLUI.getOperand(0).getReg()))
+    return false;
+  unsigned HiLuiDestReg = HiLUI.getOperand(0).getReg();
+  LoADDI = MRI->use_begin(HiLuiDestReg)->getParent();
+  if (LoADDI->getOpcode() != RISCV::ADDI ||
+      LoADDI->getOperand(2).getTargetFlags() != RISCVII::MO_LO ||
+      LoADDI->getOperand(2).getType() != MachineOperand::MO_GlobalAddress ||
+      LoADDI->getOperand(2).getOffset() != 0 ||
+      !MRI->hasOneUse(LoADDI->getOperand(0).getReg()))
+    return false;
+  return true;
+}
+
+// Update the offset in HiLUI and LoADDI instructions.
+// Delete the tail instruction and update all the uses to use the
+// output from LoADDI.
+void RISCVMergeBaseOffsetOpt::foldOffset(MachineInstr &HiLUI,
+                                         MachineInstr &LoADDI,
+                                         MachineInstr &Tail, int64_t Offset) {
+  // Put the offset back in HiLUI and the LoADDI
+  HiLUI.getOperand(1).setOffset(Offset);
+  LoADDI.getOperand(2).setOffset(Offset);
+  // Delete the tail instruction.
+  DeadInstrs.insert(&Tail);
+  MRI->replaceRegWith(Tail.getOperand(0).getReg(),
+                           LoADDI.getOperand(0).getReg());
+  LLVM_DEBUG(dbgs() << "  Merged offset " << Offset << " into base.\n"
+               << "     " << HiLUI << "     " << LoADDI;);
+}
+
+// Detect patterns for large offsets that are passed into an ADD instruction.
+//
+//                     Base address lowering is of the form:
+//                        HiLUI:  lui   vreg1, %hi(s)
+//                       LoADDI:  addi  vreg2, vreg1, %lo(s)
+//                       /                                  \
+//                      /                                    \
+//                     /                                      \
+//                    /  The large offset can be of two forms: \
+//  1) Offset that has non zero bits in lower      2) Offset that has non zero
+//     12 bits and upper 20 bits                      bits in upper 20 bits only
+//   OffseLUI: lui   vreg3, 4
+// OffsetTail: addi  voff, vreg3, 188                OffsetTail: lui  voff, 128
+//                    \                                        /
+//                     \                                      /
+//                      \                                    /
+//                       \                                  /
+//                         TailAdd: add  vreg4, vreg2, voff
+bool RISCVMergeBaseOffsetOpt::matchLargeOffset(MachineInstr &TailAdd,
+                                             unsigned GAReg, int64_t &Offset) {
+  assert((TailAdd.getOpcode() == RISCV::ADD) && "Expected ADD instruction!");
+  for (MachineOperand &Op : TailAdd.operands()) {
+    if (Op.isDef())
+      continue;
+    if (!Op.isReg())
+      continue;
+    unsigned Reg = Op.getReg();
+    // Ignore the register from Global Address base.
+    if (Reg == GAReg)
+      continue;
+    // Can't fold if the register has more than one use.
+    if (!MRI->hasOneUse(Reg))
+      continue;
+    // This can point to an ADDI or a LUI:
+    MachineInstr &OffsetTail = *MRI->getVRegDef(Reg);
+    if (OffsetTail.getOpcode() == RISCV::ADDI) {
+      // The offset value has non zero bits in both %hi and %lo parts.
+      // Detect an ADDI that feeds from a LUI instruction.
+      MachineOperand &AddiImmOp = OffsetTail.getOperand(2);
+      if (AddiImmOp.getTargetFlags() != RISCVII::MO_None)
+        continue;
+      int64_t OffLo = AddiImmOp.getImm();
+      MachineInstr &OffsetLui =
+          *MRI->getVRegDef(OffsetTail.getOperand(1).getReg());
+      MachineOperand &LuiImmOp = OffsetLui.getOperand(1);
+      if (OffsetLui.getOpcode() != RISCV::LUI ||
+          LuiImmOp.getTargetFlags() != RISCVII::MO_None ||
+          !MRI->hasOneUse(OffsetLui.getOperand(0).getReg()))
+        continue;
+      int64_t OffHi = OffsetLui.getOperand(1).getImm();
+      Offset = (OffHi << 12) + OffLo;
+      LLVM_DEBUG(dbgs() << "  Offset Instrs: " << OffsetTail
+                        << "                 " << OffsetLui);
+      DeadInstrs.insert(&OffsetTail);
+      DeadInstrs.insert(&OffsetLui);
+      return true;
+    } else if (OffsetTail.getOpcode() == RISCV::LUI) {
+      // The offset value has all zero bits in the lower 12 bits. Only LUI
+      // exists.
+      LLVM_DEBUG(dbgs() << "  Offset Instr: " << OffsetTail);
+      Offset = OffsetTail.getOperand(1).getImm() << 12;
+      DeadInstrs.insert(&OffsetTail);
+      return true;
+    }
+  }
+  return false;
+}
+
+bool RISCVMergeBaseOffsetOpt::detectAndFoldOffset(MachineInstr &HiLUI,
+                                                   MachineInstr &LoADDI) {
+  unsigned DestReg = LoADDI.getOperand(0).getReg();
+  assert(MRI->hasOneUse(DestReg) && "expected one use for LoADDI");
+  // LoADDI has only one use.
+  MachineInstr &Tail = *MRI->use_begin(DestReg)->getParent();
+  switch (Tail.getOpcode()) {
+  default:
+    LLVM_DEBUG(dbgs() << "Don't know how to get offset from this instr:"
+                      << Tail);
+    return false;
+  case RISCV::ADDI: {
+    // Offset is simply an immediate operand.
+    int64_t Offset = Tail.getOperand(2).getImm();
+    LLVM_DEBUG(dbgs() << "  Offset Instr: " << Tail);
+    foldOffset(HiLUI, LoADDI, Tail, Offset);
+    return true;
+  } break;
+  case RISCV::ADD: {
+    // The offset is too large to fit in the immediate field of ADDI.
+    // This can be in two forms:
+    // 1) LUI hi_Offset followed by:
+    //    ADDI lo_offset
+    //    This happens in case the offset has non zero bits in
+    //    both hi 20 and lo 12 bits.
+    // 2) LUI (offset20)
+    //    This happens in case the lower 12 bits of the offset are zeros.
+    int64_t Offset;
+    if (!matchLargeOffset(Tail, DestReg, Offset))
+      return false;
+    foldOffset(HiLUI, LoADDI, Tail, Offset);
+    return true;
+  } break;
+  case RISCV::LB:
+  case RISCV::LH:
+  case RISCV::LW:
+  case RISCV::LBU:
+  case RISCV::LHU:
+  case RISCV::LWU:
+  case RISCV::LD:
+  case RISCV::FLW:
+  case RISCV::FLD:
+  case RISCV::SB:
+  case RISCV::SH:
+  case RISCV::SW:
+  case RISCV::SD:
+  case RISCV::FSW:
+  case RISCV::FSD: {
+    // Transforms the sequence:            Into:
+    // HiLUI:  lui vreg1, %hi(foo)          --->  lui vreg1, %hi(foo+8)
+    // LoADDI: addi vreg2, vreg1, %lo(foo)  --->  lw vreg3, lo(foo+8)(vreg1)
+    // Tail:   lw vreg3, 8(vreg2)
+    if (Tail.getOperand(1).isFI())
+      return false;
+    // Register defined by LoADDI should be used in the base part of the
+    // load\store instruction. Otherwise, no folding possible.
+    unsigned BaseAddrReg = Tail.getOperand(1).getReg();
+    if (DestReg != BaseAddrReg)
+      return false;
+    MachineOperand &TailImmOp = Tail.getOperand(2);
+    int64_t Offset = TailImmOp.getImm();
+    // Update the offsets in global address lowering.
+    HiLUI.getOperand(1).setOffset(Offset);
+    // Update the immediate in the Tail instruction to add the offset.
+    Tail.RemoveOperand(2);
+    MachineOperand &ImmOp = LoADDI.getOperand(2);
+    ImmOp.setOffset(Offset);
+    Tail.addOperand(ImmOp);
+    // Update the base reg in the Tail instruction to feed from LUI.
+    // Output of HiLUI is only used in LoADDI, no need to use
+    // MRI->replaceRegWith().
+    Tail.getOperand(1).setReg(HiLUI.getOperand(0).getReg());
+    DeadInstrs.insert(&LoADDI);
+    return true;
+  } break;
+  }
+  return false;
+}
+
+bool RISCVMergeBaseOffsetOpt::runOnMachineFunction(MachineFunction &Fn) {
+  if (skipFunction(Fn.getFunction()))
+    return false;
+
+  DeadInstrs.clear();
+  MRI = &Fn.getRegInfo();
+  for (MachineBasicBlock &MBB : Fn) {
+    LLVM_DEBUG(dbgs() << "MBB: " << MBB.getName() << "\n");
+    for (MachineInstr &HiLUI : MBB) {
+      MachineInstr *LoADDI = nullptr;
+      if (!detectLuiAddiGlobal(HiLUI, LoADDI))
+        continue;
+      LLVM_DEBUG(dbgs() << "  Found lowered global address with one use: "
+                        << *LoADDI->getOperand(2).getGlobal() << "\n");
+      // If the use count is only one, merge the offset
+      detectAndFoldOffset(HiLUI, *LoADDI);
+    }
+  }
+  // Delete dead instructions.
+  for (auto *MI : DeadInstrs)
+    MI->eraseFromParent();
+  return true;
+}
+
+/// Returns an instance of the Merge Base Offset Optimization pass.
+FunctionPass *llvm::createRISCVMergeBaseOffsetOptPass() {
+  return new RISCVMergeBaseOffsetOpt();
+}

lib/Target/RISCV/RISCVTargetMachine.cpp

RISCVPassConfig(RISCVTargetMachine &TM, PassManagerBase &PM)
       : TargetPassConfig(TM, PM) {}
 
   RISCVTargetMachine &getRISCVTargetMachine() const {
     return getTM<RISCVTargetMachine>();
   }
 
   bool addInstSelector() override;
   void addPreEmitPass() override;
+  void addPreRegAlloc() override;
 };
 }
 
 TargetPassConfig *RISCVTargetMachine::createPassConfig(PassManagerBase &PM) {
   return new RISCVPassConfig(*this, PM);
 }
 
 bool RISCVPassConfig::addInstSelector() {
   addPass(createRISCVISelDag(getRISCVTargetMachine()));
 
   return false;
 }
 
 void RISCVPassConfig::addPreEmitPass() { addPass(&BranchRelaxationPassID); }
+
+void RISCVPassConfig::addPreRegAlloc() {
+  addPass(createRISCVMergeBaseOffsetOptPass());
+}

test/CodeGen/RISCV/hoist-global-addr-base.ll

 if.then:                                          ; preds = %entry
   store i32 10, i32* getelementptr inbounds (%struct.S, %struct.S* @s, i32 0, i32 1), align 4
   br label %if.end
 
 if.end:                                           ; preds = %if.then, %entry
   ret void
 }
 
+; This test checks that the offset is reconstructed correctly when
+; "addi" of the big offset has a negative immediate.
+; without peephole this generates:
+; lui  a1, %hi(g)
+; addi a1, a0, %lo(g)
+; lui  a0, 18     ---> offset
+; addi a0, a0, -160
+; add  a0, a0, a1  ---> base + offset.
+define i8* @big_offset_neg_addi() {
+; CHECK-LABEL: big_offset_neg_addi:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    lui a0, %hi(g+73568)
+; CHECK-NEXT:    addi a0, a0, %lo(g+73568)
+; CHECK-NEXT:    ret
+  ret i8* getelementptr inbounds ([1048576 x i8], [1048576 x i8]* @g, i32 0, i32 73568)
+}
+
 ; This test checks for the case where the offset is only an LUI.
 ; without peephole this generates:
 ; lui  a0, %hi(g)
 ; addi a0, a0, %lo(g)
 ; lui  a1, 128     ---> offset
 ; add  a0, a0, a1  ---> base + offset.
 define i8* @big_offset_lui_tail() {
 ; CHECK-LABEL: big_offset_lui_tail:

 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    lui a0, %hi(s+160)
 ; CHECK-NEXT:    addi a0, a0, %lo(s+160)
 ; CHECK-NEXT:    ret
 entry:
   ret i32* getelementptr inbounds (%struct.S, %struct.S* @s, i32 0, i32 1)
 }
 
-; TODO: In this case we get a better sequence if the offset didn't get didn't
-; get merged back in %if.end and %if.then. The current peephole is not able to
-; detect the shared global address node across blocks.
-; Without the peephole we can generate:
-;# %bb.0:                                # %entry
-;  lui   a0, %hi(s)
-;  addi  a0, a0, %lo(s)
-;  lw    a1, 164(a0)
-;  beqz  a1, .LBB0_2
-;# %bb.1:                                # %if.end
-;  addi  a0, a0, 168
-;  ret
-;.LBB0_2:                                # %if.then
-;  addi  a0, a0, 160
-;  ret
 ; Function Attrs: norecurse nounwind optsize readonly
 define dso_local i32* @control_flow_no_mem(i32 %n) local_unnamed_addr #1 {
 ; CHECK-LABEL: control_flow_no_mem:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    lui a0, %hi(s)
 ; CHECK-NEXT:    addi a0, a0, %lo(s)
-; CHECK-NEXT:    lw a0, 164(a0)
-; CHECK-NEXT:    beqz a0, .LBB5_2
+; CHECK-NEXT:    lw a1, 164(a0)
+; CHECK-NEXT:    beqz a1, .LBB6_2
 ; CHECK-NEXT:  # %bb.1: # %if.end
-; CHECK-NEXT:    lui a0, %hi(s+168)
-; CHECK-NEXT:    addi a0, a0, %lo(s+168)
+; CHECK-NEXT:    addi a0, a0, 168
 ; CHECK-NEXT:    ret
-; CHECK-NEXT:  .LBB5_2: # %if.then
-; CHECK-NEXT:    lui a0, %hi(s+160)
-; CHECK-NEXT:    addi a0, a0, %lo(s+160)
+; CHECK-NEXT:  .LBB6_2: # %if.then
+; CHECK-NEXT:    addi a0, a0, 160
 ; CHECK-NEXT:    ret
 entry:
   %0 = load i32, i32* getelementptr inbounds (%struct.S, %struct.S* @s, i32 0, i32 2), align 4
   %cmp = icmp eq i32 %0, 0
   br i1 %cmp, label %if.then, label %if.end
 if.then:                                          ; preds = %entry
   ret i32* getelementptr inbounds (%struct.S, %struct.S* @s, i32 0, i32 1)
 if.end:                                           ; preds = %if.then, %entry
   ret i32* getelementptr inbounds (%struct.S, %struct.S* @s, i32 0, i32 3)
 }
 
-;TODO: Offset shouln't be separated in this case. We get shorter sequence if it
-; is merged in the LUI %hi and the ADDI %lo, the "ADDI" could be folded in the
-; immediate part of "lhu" genertating the sequence:
-;  lui     a0, %hi(foo +8)
-;  lhu     a0, %lo(foo+8)(a0)
-; instead of:
-;  lui     a0, %hi(foo)
-;  addi    a0, a0, %lo(foo)
-;  lhu     a0, 8(a0)
-
 define dso_local i32 @load_half() nounwind {
 ; CHECK-LABEL: load_half:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    addi sp, sp, -16
 ; CHECK-NEXT:    sw ra, 12(sp)
-; CHECK-NEXT:    lui a0, %hi(foo)
-; CHECK-NEXT:    addi a0, a0, %lo(foo)
-; CHECK-NEXT:    lhu a0, 8(a0)
+; CHECK-NEXT:    lui a0, %hi(foo+8)
+; CHECK-NEXT:    lhu a0, %lo(foo+8)(a0)
 ; CHECK-NEXT:    addi a1, zero, 140
-; CHECK-NEXT:    bne a0, a1, .LBB6_2
+; CHECK-NEXT:    bne a0, a1, .LBB7_2
 ; CHECK-NEXT:  # %bb.1: # %if.end
 ; CHECK-NEXT:    mv a0, zero
 ; CHECK-NEXT:    lw ra, 12(sp)
 ; CHECK-NEXT:    addi sp, sp, 16
 ; CHECK-NEXT:    ret
-; CHECK-NEXT:  .LBB6_2: # %if.then
+; CHECK-NEXT:  .LBB7_2: # %if.then
 ; CHECK-NEXT:    call abort
 entry:
   %0 = load i16, i16* getelementptr inbounds ([6 x i16], [6 x i16]* @foo, i32 0, i32 4), align 2
   %cmp = icmp eq i16 %0, 140
   br i1 %cmp, label %if.end, label %if.then
 
 if.then:
   tail call void @abort()
   unreachable
 
 if.end:
   ret i32 0
 }
 
 declare void @abort()
+
+define dso_local void @one_store() local_unnamed_addr {
+; CHECK-LABEL: one_store:
+; CHECK:       # %bb.0: # %entry
+; CHECK-NEXT:    lui a0, %hi(s+160)
+; CHECK-NEXT:    addi a1, zero, 10
+; CHECK-NEXT:    sw a1, %lo(s+160)(a0)
+; CHECK-NEXT:    ret
+entry:
+  store i32 10, i32* getelementptr inbounds (%struct.S, %struct.S* @s, i32 0, i32 1), align 4
+  ret void
+}