[GlobalISel] Add re-association combine for G_PTR_ADD to allow better addressing mode usage.

[aemerson]

Summary

We're trying to match a few pointer computation patterns here for re-association opportunities.

1. Isolating a constant operand to be on the RHS, e.g.:

G_PTR_ADD(BASE, G_ADD(X, C)) -> G_PTR_ADD(G_PTR_ADD(BASE, X), C)

2. Folding two constants in each sub-tree as long as such folding doesn't break a legal addressing mode.

G_PTR_ADD(G_PTR_ADD(BASE, C1), C2) -> G_PTR_ADD(BASE, C1+C2)

AArch64 code size improvements on CTMark with -Os:

Program              before  after   diff
pairlocalalign      251048  251044 -0.0%
consumer-typeset    421820  421812 -0.0%
kc                  431348  431320 -0.0%
SPASS               413404  413300 -0.0%
clamscan            384396  384220 -0.0%
tramp3d-v4          370640  370412 -0.1%
lencod              432096  431772 -0.1%
bullet              479400  478796 -0.1%
sqlite3             288504  288072 -0.1%
7zip-benchmark      573796  570768 -0.5%
Geomean difference                 -0.1%

[arsenm]

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

1677

Check isLegalAddressingMode, or some similar hook?

4040–4041

This should be impossible?

4069–4070

Don't we retry combines like the DAG? This seems like a bigger structural problem to solve rather than specially treating this

[paquette]

llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h

530

maybe a doxygen comment?

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

4044

getOpcodeDef?

4055

maybe minorly faster

auto C1 = getConstantVRegVal(Src1Def->getOperand(2).getReg(), MRI);
if (!C1)
  return false

auto C2 = getConstantVRegVal(Src2Reg, MRI);
if (!C2)
  return false;

You could probably pull in some of the C1/C2APIntVal stuff below into there too, although that probably matters less.

[aemerson]

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

1677

I don't think it's necessary for this. We should avoid doing this combine no matter what with multiple users since doing so doesn't give any benefit since the def can't be deleted.

4040–4041

I'll change this to an assert for now.

4044

+1

4055

Sure.

4069–4070

We do, but this is a heuristic and not a combine. If we can't prove that we break an addressing mode (by detecting a memory-operation user), then we return false, and so the reassociation is done. Once it's done, we don't revisit the decision and undo it later, so we need to have a best effort to catch these cases here.

[aemerson]

Address some comments.

[paquette]

Just a couple nits and then this looks fine to me.

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

1677

I think a bunch of the matching logic here could be done with mi_match?

something like this?

if (!mi_match(Add2, MRI, m_OneNonDBGUse(m_GPtrAdd(m_Reg(Base), m_ICst(Imm2)))))
   return false;

although if looking through the constants is important, then maybe this won't work.

4065

I don't think the testcases cover this and the

if (C1APIntVal.getBitWidth() > 64 || C2APIntVal.getBitWidth() > 64)

case?

4080

might as well make ConvUseMI->getOpcode() a variable

4082

maybe slightly easier to read if you reduce the indentation level here

[aemerson]

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

1677

Yeah I think that wouldn't be NFC.

4065

Yep, and it seems it's not actually possible to trigger that at all. I'll remove this code and also put up a patch to do the same for the DAGCombiner counterpart on which this code was based.

[foad]

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

1677

We should avoid doing this combine no matter what with multiple users since doing so doesn't give any benefit since the def can't be deleted.

I don't understand. If Add2 is used by two ptr_adds, and both of them can be folded to give legal addresses, then we should do that and delete Add2. We are seeing some cases downstream where that is no longer happening because of this check.

[aemerson]

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

1677

Ok, fair enough. I’ll look into that case today.

[craig.topper]

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

4063

sorry to dig up an old patch, but I think this should be checking the users of PtrAdd. See https://reviews.llvm.org/D124644

llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h

public:
   bool matchExtractAllEltsFromBuildVector(
       MachineInstr &MI,
       SmallVectorImpl<std::pair<Register, MachineInstr *>> &MatchInfo);
   void applyExtractAllEltsFromBuildVector(
       MachineInstr &MI,
       SmallVectorImpl<std::pair<Register, MachineInstr *>> &MatchInfo);
 
   /// Use a function which takes in a MachineIRBuilder to perform a combine.
+  /// By default, it erases the instruction \p MI from the function.
   bool applyBuildFn(MachineInstr &MI,
                     std::function<void(MachineIRBuilder &)> &MatchInfo);
+  /// Use a function which takes in a MachineIRBuilder to perform a combine.
+  /// This variant does not erase \p MI after calling the build function.
+  bool applyBuildFnNoErase(MachineInstr &MI,
+                           std::function<void(MachineIRBuilder &)> &MatchInfo);
+
   bool matchFunnelShiftToRotate(MachineInstr &MI);
   void applyFunnelShiftToRotate(MachineInstr &MI);
   bool matchRotateOutOfRange(MachineInstr &MI);
   void applyRotateOutOfRange(MachineInstr &MI);
 
   /// \returns true if a G_ICMP instruction \p MI can be replaced with a true
   /// or false constant based off of KnownBits information.
   bool matchICmpToTrueFalseKnownBits(MachineInstr &MI, int64_t &MatchInfo);
 
   /// Match: and (lshr x, cst), mask -> ubfx x, cst, width
   bool matchBitfieldExtractFromAnd(
       MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo);
 
+  /// Reassociate pointer calculations with G_ADD involved, to allow better
+  /// addressing mode usage.
+  bool matchReassocPtrAdd(MachineInstr &MI,
+                          std::function<void(MachineIRBuilder &)> &MatchInfo);
+
   /// Try to transform \p MI by using all of the above
   /// combine functions. Returns true if changed.
   bool tryCombine(MachineInstr &MI);
 
 private:
   // Memcpy family optimization helpers.
   bool optimizeMemcpy(MachineInstr &MI, Register Dst, Register Src,
                       unsigned KnownLen, Align DstAlign, Align SrcAlign,
private:
   /// at to the index of the load.
   /// \param [in] MemSizeInBits - The number of bits each load should produce.
   ///
   /// \returns The lowest-index load found and the lowest index on success.
   Optional<std::pair<MachineInstr *, int64_t>> findLoadOffsetsForLoadOrCombine(
       SmallDenseMap<int64_t, int64_t, 8> &MemOffset2Idx,
       const SmallVector<Register, 8> &RegsToVisit,
       const unsigned MemSizeInBits);
+
+  /// Examines the G_PTR_ADD instruction \p PtrAdd and determines if performing
+  /// a re-association of its operands would break an existing legal addressing
+  /// mode that the address computation currently represents.
+  bool reassociationCanBreakAddressingModePattern(MachineInstr &PtrAdd);
 };
 } // namespace llvm
 
 #endif

llvm/include/llvm/Target/GlobalISel/Combine.td

 def bitfield_extract_from_and : GICombineRule<
   (defs root:$root, build_fn_matchinfo:$info),
   (match (wip_match_opcode G_AND):$root,
     [{ return Helper.matchBitfieldExtractFromAnd(*${root}, ${info}); }]),
   (apply [{ return Helper.applyBuildFn(*${root}, ${info}); }])>;
 
 def funnel_shift_combines : GICombineGroup<[funnel_shift_to_rotate]>;
 
+def reassoc_ptradd : GICombineRule<
+  (defs root:$root, build_fn_matchinfo:$matchinfo),
+  (match (wip_match_opcode G_PTR_ADD):$root,
+    [{ return Helper.matchReassocPtrAdd(*${root}, ${matchinfo}); }]),
+  (apply [{ Helper.applyBuildFnNoErase(*${root}, ${matchinfo}); }])>;
+
+def reassocs : GICombineGroup<[reassoc_ptradd]>;
+
 // FIXME: These should use the custom predicate feature once it lands.
 def undef_combines : GICombineGroup<[undef_to_fp_zero, undef_to_int_zero,
                                      undef_to_negative_one,
                                      binop_left_undef_to_zero,
                                      propagate_undef_any_op,
                                      propagate_undef_all_ops,
                                      propagate_undef_shuffle_mask,
                                      erase_undef_store]>;

 def phi_combines : GICombineGroup<[extend_through_phis]>;
 
 def select_combines : GICombineGroup<[select_undef_cmp, select_constant_cmp]>;
 
 def trivial_combines : GICombineGroup<[copy_prop, mul_to_shl, add_p2i_to_ptradd,
                                        mul_by_neg_one]>;
 
 def all_combines : GICombineGroup<[trivial_combines, insert_vec_elt_combines,
-    extract_vec_elt_combines, ptr_add_immed_chain, combines_for_extload,
+    extract_vec_elt_combines, combines_for_extload,
     combine_indexed_load_store, undef_combines, identity_combines, phi_combines,
     simplify_add_to_sub, hoist_logic_op_with_same_opcode_hands,
+    reassocs, ptr_add_immed_chain,
     shl_ashr_to_sext_inreg, sext_inreg_of_load,
     width_reduction_combines, select_combines,
     known_bits_simplifications, ext_ext_fold,
     not_cmp_fold, opt_brcond_by_inverting_cond,
     unmerge_merge, fabs_fabs_fold, unmerge_cst, unmerge_dead_to_trunc,
     unmerge_zext_to_zext, trunc_ext_fold, trunc_shl,
     const_combines, xor_of_and_with_same_reg, ptr_add_with_zero,
     shift_immed_chain, shift_of_shifted_logic_chain, load_or_combine,
     div_rem_to_divrem, funnel_shift_combines, bitfield_extract_from_and]>;
 
 // A combine group used to for prelegalizer combiners at -O0. The combines in
 // this group have been selected based on experiments to balance code size and
 // compile time performance.
 def optnone_combines : GICombineGroup<[trivial_combines,
     ptr_add_immed_chain, combines_for_extload,
     not_cmp_fold, opt_brcond_by_inverting_cond]>;

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

if (MI.getOpcode() != TargetOpcode::G_PTR_ADD)
     return false;
 
   Register Add2 = MI.getOperand(1).getReg();
   Register Imm1 = MI.getOperand(2).getReg();
   auto MaybeImmVal = getConstantVRegValWithLookThrough(Imm1, MRI);
   if (!MaybeImmVal)
     return false;
 
+  // Don't do this combine if there multiple uses of the first PTR_ADD,
+  // since we may be able to compute the second PTR_ADD as an immediate
+  // offset anyway. Folding the first offset into the second may cause us
+  // to go beyond the bounds of our legal addressing modes.
+  if (!MRI.hasOneNonDBGUse(Add2))
+    return false;
+
   MachineInstr *Add2Def = MRI.getUniqueVRegDef(Add2);
   if (!Add2Def || Add2Def->getOpcode() != TargetOpcode::G_PTR_ADD)
     return false;
 
   Register Base = Add2Def->getOperand(1).getReg();
   Register Imm2 = Add2Def->getOperand(2).getReg();
   auto MaybeImm2Val = getConstantVRegValWithLookThrough(Imm2, MRI);
   if (!MaybeImm2Val)

 bool CombinerHelper::applyBuildFn(
     MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
   Builder.setInstrAndDebugLoc(MI);
   MatchInfo(Builder);
   MI.eraseFromParent();
   return true;
 }
 
+bool CombinerHelper::applyBuildFnNoErase(
+    MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
+  Builder.setInstrAndDebugLoc(MI);
+  MatchInfo(Builder);
+  return true;
+}
+
 /// Match an FSHL or FSHR that can be combined to a ROTR or ROTL rotate.
 bool CombinerHelper::matchFunnelShiftToRotate(MachineInstr &MI) {
   unsigned Opc = MI.getOpcode();
   assert(Opc == TargetOpcode::G_FSHL || Opc == TargetOpcode::G_FSHR);
   Register X = MI.getOperand(1).getReg();
   Register Y = MI.getOperand(2).getReg();
   if (X != Y)
     return false;
bool CombinerHelper::matchBitfieldExtractFromAnd(
   MatchInfo = [=](MachineIRBuilder &B) {
     auto WidthCst = B.buildConstant(Ty, Width);
     auto LSBCst = B.buildConstant(Ty, LSBImm);
     B.buildInstr(TargetOpcode::G_UBFX, {Dst}, {ShiftSrc, LSBCst, WidthCst});
   };
   return true;
 }
 
+bool CombinerHelper::reassociationCanBreakAddressingModePattern(
+    MachineInstr &PtrAdd) {
+  assert(PtrAdd.getOpcode() == TargetOpcode::G_PTR_ADD);
+
+  Register Src1Reg = PtrAdd.getOperand(1).getReg();
+  MachineInstr *Src1Def = getOpcodeDef(TargetOpcode::G_PTR_ADD, Src1Reg, MRI);
+  if (!Src1Def)
+    return false;
+
+  Register Src2Reg = PtrAdd.getOperand(2).getReg();
+
+  if (MRI.hasOneNonDBGUse(Src1Reg))
+    return false;
+
+  auto C1 = getConstantVRegVal(Src1Def->getOperand(2).getReg(), MRI);
+  if (!C1)
+    return false;
+  auto C2 = getConstantVRegVal(Src2Reg, MRI);
+  if (!C2)
+    return false;
+
+  const APInt &C1APIntVal = *C1;
+  const APInt &C2APIntVal = *C2;
+  if (C1APIntVal.getBitWidth() > 64 || C2APIntVal.getBitWidth() > 64)
+    return false;
+
+  const APInt CombinedValueIntVal = C1APIntVal + C2APIntVal;
+  if (CombinedValueIntVal.getBitWidth() > 64)
+    return false;
+  const int64_t CombinedValue = CombinedValueIntVal.getSExtValue();
+
+  for (auto &UseMI : MRI.use_nodbg_instructions(Src1Reg)) {
+    // This combine may end up running before ptrtoint/inttoptr combines
+    // manage to eliminate redundant conversions, so try to look through them.
+    MachineInstr *ConvUseMI = &UseMI;
+    while (ConvUseMI->getOpcode() == TargetOpcode::G_INTTOPTR ||
+           ConvUseMI->getOpcode() == TargetOpcode::G_PTRTOINT) {
+      Register DefReg = ConvUseMI->getOperand(0).getReg();
+      if (!MRI.hasOneNonDBGUse(DefReg))
+        break;
+      ConvUseMI = &*MRI.use_instr_nodbg_begin(DefReg);
+    }
+    auto LoadStore = ConvUseMI->getOpcode() == TargetOpcode::G_LOAD ||
+                     ConvUseMI->getOpcode() == TargetOpcode::G_STORE;
+    if (LoadStore) {
+      // Is x[offset2] already not a legal addressing mode? If so then
+      // reassociating the constants breaks nothing (we test offset2 because
+      // that's the one we hope to fold into the load or store).
+      TargetLoweringBase::AddrMode AM;
+      AM.HasBaseReg = true;
+      AM.BaseOffs = C2APIntVal.getSExtValue();
+      unsigned AS =
+          MRI.getType(ConvUseMI->getOperand(1).getReg()).getAddressSpace();
+      Type *AccessTy =
+          getTypeForLLT(MRI.getType(ConvUseMI->getOperand(0).getReg()),
+                        PtrAdd.getMF()->getFunction().getContext());
+      const auto &TLI = *PtrAdd.getMF()->getSubtarget().getTargetLowering();
+      if (!TLI.isLegalAddressingMode(PtrAdd.getMF()->getDataLayout(), AM,
+                                     AccessTy, AS))
+        continue;
+
+      // Would x[offset1+offset2] still be a legal addressing mode?
+      AM.BaseOffs = CombinedValue;
+      if (!TLI.isLegalAddressingMode(PtrAdd.getMF()->getDataLayout(), AM,
+                                     AccessTy, AS))
+        return true;
+    }
+  }
+
+  return false;
+}
+
+bool CombinerHelper::matchReassocPtrAdd(
+    MachineInstr &MI, std::function<void(MachineIRBuilder &)> &MatchInfo) {
+  assert(MI.getOpcode() == TargetOpcode::G_PTR_ADD);
+  // We're trying to match a few pointer computation patterns here for
+  // re-association opportunities.
+  // 1) Isolating a constant operand to be on the RHS, e.g.:
+  // G_PTR_ADD(BASE, G_ADD(X, C)) -> G_PTR_ADD(G_PTR_ADD(BASE, X), C)
+  //
+  // 2) Folding two constants in each sub-tree as long as such folding
+  // doesn't break a legal addressing mode.
+  // G_PTR_ADD(G_PTR_ADD(BASE, C1), C2) -> G_PTR_ADD(BASE, C1+C2)
+  Register Src1Reg = MI.getOperand(1).getReg();
+  Register Src2Reg = MI.getOperand(2).getReg();
+  MachineInstr *LHS = MRI.getVRegDef(Src1Reg);
+  MachineInstr *RHS = MRI.getVRegDef(Src2Reg);
+
+  if (LHS->getOpcode() != TargetOpcode::G_PTR_ADD) {
+    // Try to match example 1).
+    if (RHS->getOpcode() != TargetOpcode::G_ADD)
+      return false;
+    auto C2 = getConstantVRegVal(RHS->getOperand(2).getReg(), MRI);
+    if (!C2)
+      return false;
+
+    MatchInfo = [=,&MI](MachineIRBuilder &B) {
+      LLT PtrTy = MRI.getType(MI.getOperand(0).getReg());
+
+      auto NewBase =
+          Builder.buildPtrAdd(PtrTy, Src1Reg, RHS->getOperand(1).getReg());
+      Observer.changingInstr(MI);
+      MI.getOperand(1).setReg(NewBase.getReg(0));
+      MI.getOperand(2).setReg(RHS->getOperand(2).getReg());
+      Observer.changedInstr(MI);
+    };
+  } else {
+    // Try to match example 2.
+    Register LHSSrc1 = LHS->getOperand(1).getReg();
+    Register LHSSrc2 = LHS->getOperand(2).getReg();
+    auto C1 = getConstantVRegVal(LHSSrc2, MRI);
+    if (!C1)
+      return false;
+    auto C2 = getConstantVRegVal(Src2Reg, MRI);
+    if (!C2)
+      return false;
+
+    MatchInfo = [=, &MI](MachineIRBuilder &B) {
+      auto NewCst = B.buildConstant(MRI.getType(Src2Reg), *C1 + *C2);
+      Observer.changingInstr(MI);
+      MI.getOperand(1).setReg(LHSSrc1);
+      MI.getOperand(2).setReg(NewCst.getReg(0));
+      Observer.changedInstr(MI);
+    };
+  }
+  return !reassociationCanBreakAddressingModePattern(MI);
+}
+
 bool CombinerHelper::tryCombine(MachineInstr &MI) {
   if (tryCombineCopy(MI))
     return true;
   if (tryCombineExtendingLoads(MI))
     return true;
   if (tryCombineIndexedLoadStore(MI))
     return true;
   return false;
 }

llvm/test/CodeGen/AArch64/GlobalISel/combine-ptradd-reassociation.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -run-pass=aarch64-prelegalizer-combiner -verify-machineinstrs -mtriple aarch64-unknown-unknown %s -o - | FileCheck %s
+---
+name:            test1_noreassoc_legal_already_new_is_illegal
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$x0' }
+body:             |
+  bb.1:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: test1_noreassoc_legal_already_new_is_illegal
+    ; CHECK: liveins: $x0
+    ; CHECK: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4777
+    ; CHECK: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 6
+    ; CHECK: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; CHECK: [[PTR_ADD1:%[0-9]+]]:_(p0) = G_PTR_ADD [[PTR_ADD]], [[C1]](s64)
+    ; CHECK: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p0) :: (load 4)
+    ; CHECK: G_STORE [[C2]](s32), [[PTR_ADD]](p0) :: (store 4)
+    ; CHECK: $w0 = COPY [[LOAD]](s32)
+    ; CHECK: RET_ReallyLR implicit $w0
+    %0:_(p0) = COPY $x0
+    %2:_(s64) = G_CONSTANT i64 4777
+    %4:_(s64) = G_CONSTANT i64 6
+    %9:_(s32) = G_CONSTANT i32 0
+    %10:_(p0) = G_PTR_ADD %0, %2(s64)
+    %11:_(p0) = G_PTR_ADD %10, %4(s64)
+    %7:_(s32) = G_LOAD %11(p0) :: (load 4)
+    G_STORE %9(s32), %10(p0) :: (store 4) ; other use of %10
+    $w0 = COPY %7(s32)
+    RET_ReallyLR implicit $w0
+
+...
+---
+name:            test2_reassoc_already_legal_new_also_legal
+alignment:       4
+liveins:
+  - { reg: '$x0' }
+body:             |
+  bb.1:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: test2_reassoc_already_legal_new_also_legal
+    ; CHECK: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 10
+    ; CHECK: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; CHECK: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 16
+    ; CHECK: [[PTR_ADD1:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[C2]](s64)
+    ; CHECK: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p0) :: (load 4)
+    ; CHECK: G_STORE [[C1]](s32), [[PTR_ADD]](p0) :: (store 4)
+    ; CHECK: $w0 = COPY [[LOAD]](s32)
+    ; CHECK: RET_ReallyLR implicit $w0
+    %0:_(p0) = COPY $x0
+    %2:_(s64) = G_CONSTANT i64 10
+    %4:_(s64) = G_CONSTANT i64 6
+    %9:_(s32) = G_CONSTANT i32 0
+    %10:_(p0) = G_PTR_ADD %0, %2(s64)
+    %11:_(p0) = G_PTR_ADD %10, %4(s64)
+    %7:_(s32) = G_LOAD %11(p0) :: (load 4)
+    G_STORE %9(s32), %10(p0) :: (store 4) ; other use of %10
+    $w0 = COPY %7(s32)
+    RET_ReallyLR implicit $w0
+
+...
+---
+name:            test3_noreassoc_only_oneuse
+alignment:       4
+liveins:
+  - { reg: '$x0' }
+body:             |
+  bb.1:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: test3_noreassoc_only_oneuse
+    ; CHECK: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 4783
+    ; CHECK: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; CHECK: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p0) :: (load 4)
+    ; CHECK: $w0 = COPY [[LOAD]](s32)
+    ; CHECK: RET_ReallyLR implicit $w0
+    %0:_(p0) = COPY $x0
+    %10:_(s64) = G_CONSTANT i64 4783
+    %9:_(p0) = G_PTR_ADD %0, %10(s64)
+    %7:_(s32) = G_LOAD %9(p0) :: (load 4)
+    $w0 = COPY %7(s32)
+    RET_ReallyLR implicit $w0
+
+...
+---
+name:            test4_reassoc_existing_is_already_illegal
+alignment:       4
+liveins:
+  - { reg: '$x0' }
+body:             |
+  bb.1:
+    liveins: $x0
+
+    ; CHECK-LABEL: name: test4_reassoc_existing_is_already_illegal
+    ; CHECK: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 17
+    ; CHECK: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 0
+    ; CHECK: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; CHECK: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 4096
+    ; CHECK: [[PTR_ADD1:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[C2]](s64)
+    ; CHECK: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p0) :: (load 4)
+    ; CHECK: G_STORE [[C1]](s32), [[PTR_ADD]](p0) :: (store 4)
+    ; CHECK: $w0 = COPY [[LOAD]](s32)
+    ; CHECK: RET_ReallyLR implicit $w0
+    %0:_(p0) = COPY $x0
+    %2:_(s64) = G_CONSTANT i64 17
+    %4:_(s64) = G_CONSTANT i64 4079
+    %9:_(s32) = G_CONSTANT i32 0
+    %10:_(p0) = G_PTR_ADD %0, %2(s64)
+    %11:_(p0) = G_PTR_ADD %10, %4(s64)
+    %7:_(s32) = G_LOAD %11(p0) :: (load 4)
+    G_STORE %9(s32), %10(p0) :: (store 4) ; other use of %10
+    $w0 = COPY %7(s32)
+    RET_ReallyLR implicit $w0
+
+...
+---
+name:            test5_add_on_rhs
+alignment:       4
+liveins:
+  - { reg: '$x0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $x0, $x1
+
+    ; CHECK-LABEL: name: test5_add_on_rhs
+    ; CHECK: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(s64) = COPY $x1
+    ; CHECK: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CHECK: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY]], [[COPY1]](s64)
+    ; CHECK: [[PTR_ADD1:%[0-9]+]]:_(p0) = G_PTR_ADD [[PTR_ADD]], [[C]](s64)
+    ; CHECK: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p0) :: (load 1)
+    ; CHECK: $w0 = COPY [[LOAD]](s32)
+    ; CHECK: RET_ReallyLR implicit $w0
+    %0:_(p0) = COPY $x0
+    %1:_(s64) = COPY $x1
+    %2:_(s64) = G_CONSTANT i64 1
+    %3:_(s64) = G_ADD %1, %2
+    %4:_(p0) = G_PTR_ADD %0, %3(s64)
+    %7:_(s32) = G_LOAD %4(p0) :: (load 1)
+    $w0 = COPY %7(s32)
+    RET_ReallyLR implicit $w0
+
+...