llvm.org GIT mirror llvm / 191bd64
Revert 112442 and 112440 until the compile time problems introduced by 112440 are resolved. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@112692 91177308-0d34-0410-b5e6-96231b3b80d8 Dan Gohman 10 years ago
4 changed file(s) with 138 addition(s) and 208 deletion(s). Raw diff Collapse all Expand all
2626 class IVUsers;
2727 class ScalarEvolution;
2828 class SCEV;
29 class SCEVAddRecExpr;
3029 class IVUsers;
3130
3231 /// IVStrideUse - Keep track of one use of a strided induction variable.
122121 LoopInfo *LI;
123122 DominatorTree *DT;
124123 ScalarEvolution *SE;
125 SmallPtrSet *, 16> Processed;
124 SmallPtrSet*,16> Processed;
126125
127126 /// IVUses - A list of all tracked IV uses of induction variable expressions
128127 /// we are interested in.
134133
135134 virtual void releaseMemory();
136135
137 const SCEVAddRecExpr *findInterestingAddRec(const SCEV *S) const;
138 bool isInterestingUser(const Instruction *User) const;
139
140136 public:
141137 static char ID; // Pass ID, replacement for typeid
142138 IVUsers();
143139
144 /// AddUsersIfInteresting - Inspect the def-use graph starting at the
145 /// specified Instruction and add IVUsers.
146 void AddUsersIfInteresting(Instruction *I);
140 /// AddUsersIfInteresting - Inspect the specified Instruction. If it is a
141 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
142 /// return true. Otherwise, return false.
143 bool AddUsersIfInteresting(Instruction *I);
147144
148145 IVStrideUse &AddUser(Instruction *User, Value *Operand);
149146
3434 return new IVUsers();
3535 }
3636
37 /// findInterestingAddRec - Test whether the given expression is interesting.
38 /// Return the addrec with the current loop which makes it interesting, or
39 /// null if it is not interesting.
40 const SCEVAddRecExpr *IVUsers::findInterestingAddRec(const SCEV *S) const {
37 /// isInteresting - Test whether the given expression is "interesting" when
38 /// used by the given expression, within the context of analyzing the
39 /// given loop.
40 static bool isInteresting(const SCEV *S, const Instruction *I, const Loop *L,
41 ScalarEvolution *SE) {
4142 // An addrec is interesting if it's affine or if it has an interesting start.
4243 if (const SCEVAddRecExpr *AR = dyn_cast(S)) {
4344 // Keep things simple. Don't touch loop-variant strides.
4445 if (AR->getLoop() == L)
45 return AR;
46 // We don't yet know how to do effective SCEV expansions for addrecs
47 // with interesting steps.
48 if (findInterestingAddRec(AR->getStepRecurrence(*SE)))
49 return 0;
50 // Otherwise recurse to see if the start value is interesting.
51 return findInterestingAddRec(AR->getStart());
46 return AR->isAffine() || !L->contains(I);
47 // Otherwise recurse to see if the start value is interesting, and that
48 // the step value is not interesting, since we don't yet know how to
49 // do effective SCEV expansions for addrecs with interesting steps.
50 return isInteresting(AR->getStart(), I, L, SE) &&
51 !isInteresting(AR->getStepRecurrence(*SE), I, L, SE);
5252 }
5353
5454 // An add is interesting if exactly one of its operands is interesting.
5555 if (const SCEVAddExpr *Add = dyn_cast(S)) {
56 bool AnyInterestingYet = false;
5657 for (SCEVAddExpr::op_iterator OI = Add->op_begin(), OE = Add->op_end();
5758 OI != OE; ++OI)
58 if (const SCEVAddRecExpr *AR = findInterestingAddRec(*OI))
59 return AR;
60 return 0;
59 if (isInteresting(*OI, I, L, SE)) {
60 if (AnyInterestingYet)
61 return false;
62 AnyInterestingYet = true;
63 }
64 return AnyInterestingYet;
6165 }
6266
6367 // Nothing else is interesting here.
64 return 0;
65 }
66
67 bool IVUsers::isInterestingUser(const Instruction *User) const {
68 // Void and FP expressions cannot be reduced.
69 if (!SE->isSCEVable(User->getType()))
70 return false;
71
72 // LSR is not APInt clean, do not touch integers bigger than 64-bits.
73 if (SE->getTypeSizeInBits(User->getType()) > 64)
74 return false;
75
76 // Don't descend into PHI nodes outside the current loop.
77 if (LI->getLoopFor(User->getParent()) != L &&
78 isa(User))
79 return false;
80
81 // Otherwise, it may be interesting.
82 return true;
68 return false;
8369 }
8470
8571 /// AddUsersIfInteresting - Inspect the specified instruction. If it is a
8672 /// reducible SCEV, recursively add its users to the IVUsesByStride set and
8773 /// return true. Otherwise, return false.
88 void IVUsers::AddUsersIfInteresting(Instruction *I) {
89 // Stop if we've seen this before.
74 bool IVUsers::AddUsersIfInteresting(Instruction *I) {
75 if (!SE->isSCEVable(I->getType()))
76 return false; // Void and FP expressions cannot be reduced.
77
78 // LSR is not APInt clean, do not touch integers bigger than 64-bits.
79 if (SE->getTypeSizeInBits(I->getType()) > 64)
80 return false;
81
9082 if (!Processed.insert(I))
91 return;
92
93 // If this PHI node is not SCEVable, ignore it.
94 if (!SE->isSCEVable(I->getType()))
95 return;
96
97 // If this PHI node is not an addrec for this loop, ignore it.
98 const SCEVAddRecExpr *Expr = findInterestingAddRec(SE->getSCEV(I));
99 if (!Expr)
100 return;
101
102 // Walk the def-use graph.
103 SmallVector, 16> Worklist;
104 Worklist.push_back(std::make_pair(I, Expr));
105 do {
106 std::pair P =
107 Worklist.pop_back_val();
108 Instruction *Op = P.first;
109 const SCEVAddRecExpr *OpAR = P.second;
110
111 // Visit Op's users.
112 SmallPtrSet VisitedUsers;
113 for (Value::use_iterator UI = Op->use_begin(), E = Op->use_end();
114 UI != E; ++UI) {
115 // Don't visit any individual user more than once.
116 Instruction *User = cast(*UI);
117 if (!VisitedUsers.insert(User))
118 continue;
119
120 // If it's an affine addrec (which we can pretty safely re-expand) inside
121 // the loop, or a potentially non-affine addrec outside the loop (which
122 // we can evaluate outside of the loop), follow it.
123 if (OpAR->isAffine() || !L->contains(User)) {
124 if (isInterestingUser(User)) {
125 const SCEV *UserExpr = SE->getSCEV(User);
126
127 if (const SCEVAddRecExpr *AR = findInterestingAddRec(UserExpr)) {
128 // Interesting. Keep searching.
129 if (Processed.insert(User))
130 Worklist.push_back(std::make_pair(User, AR));
131 continue;
132 }
133 }
83 return true; // Instruction already handled.
84
85 // Get the symbolic expression for this instruction.
86 const SCEV *ISE = SE->getSCEV(I);
87
88 // If we've come to an uninteresting expression, stop the traversal and
89 // call this a user.
90 if (!isInteresting(ISE, I, L, SE))
91 return false;
92
93 SmallPtrSet UniqueUsers;
94 for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
95 UI != E; ++UI) {
96 Instruction *User = cast(*UI);
97 if (!UniqueUsers.insert(User))
98 continue;
99
100 // Do not infinitely recurse on PHI nodes.
101 if (isa(User) && Processed.count(User))
102 continue;
103
104 // Descend recursively, but not into PHI nodes outside the current loop.
105 // It's important to see the entire expression outside the loop to get
106 // choices that depend on addressing mode use right, although we won't
107 // consider references outside the loop in all cases.
108 // If User is already in Processed, we don't want to recurse into it again,
109 // but do want to record a second reference in the same instruction.
110 bool AddUserToIVUsers = false;
111 if (LI->getLoopFor(User->getParent()) != L) {
112 if (isa(User) || Processed.count(User) ||
113 !AddUsersIfInteresting(User)) {
114 DEBUG(dbgs() << "FOUND USER in other loop: " << *User << '\n'
115 << " OF SCEV: " << *ISE << '\n');
116 AddUserToIVUsers = true;
134117 }
135
136 // Otherwise, this is the point where the def-use chain
137 // becomes uninteresting. Call it an IV User.
138 AddUser(User, Op);
118 } else if (Processed.count(User) ||
119 !AddUsersIfInteresting(User)) {
120 DEBUG(dbgs() << "FOUND USER: " << *User << '\n'
121 << " OF SCEV: " << *ISE << '\n');
122 AddUserToIVUsers = true;
139123 }
140 } while (!Worklist.empty());
124
125 if (AddUserToIVUsers) {
126 // Okay, we found a user that we cannot reduce.
127 IVUses.push_back(new IVStrideUse(this, User, I));
128 IVStrideUse &NewUse = IVUses.back();
129 // Transform the expression into a normalized form.
130 ISE = TransformForPostIncUse(NormalizeAutodetect,
131 ISE, User, I,
132 NewUse.PostIncLoops,
133 *SE, *DT);
134 DEBUG(dbgs() << " NORMALIZED TO: " << *ISE << '\n');
135 }
136 }
137 return true;
141138 }
142139
143140 IVStrideUse &IVUsers::AddUser(Instruction *User, Value *Operand) {
144141 IVUses.push_back(new IVStrideUse(this, User, Operand));
145 IVStrideUse &NewUse = IVUses.back();
146
147 // Auto-detect and remember post-inc loops for this expression.
148 const SCEV *S = SE->getSCEV(Operand);
149 (void)TransformForPostIncUse(NormalizeAutodetect,
150 S, User, Operand,
151 NewUse.PostIncLoops,
152 *SE, *DT);
153 return NewUse;
142 return IVUses.back();
154143 }
155144
156145 IVUsers::IVUsers()
175164 // them by stride. Start by finding all of the PHI nodes in the header for
176165 // this loop. If they are induction variables, inspect their uses.
177166 for (BasicBlock::iterator I = L->getHeader()->begin(); isa(I); ++I)
178 AddUsersIfInteresting(I);
167 (void)AddUsersIfInteresting(I);
179168
180169 return false;
181170 }
112112 public:
113113 void CountRegister(const SCEV *Reg, size_t LUIdx);
114114 void DropRegister(const SCEV *Reg, size_t LUIdx);
115 void DropUse(size_t LUIdx, size_t NewLUIdx);
116115 void DropUse(size_t LUIdx);
117116
118117 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
151150 RSD.UsedByIndices.reset(LUIdx);
152151 }
153152
154 /// DropUse - Clear out reference by use LUIdx, and prepare for use NewLUIdx
155 /// to be swapped into LUIdx's position.
156 void
157 RegUseTracker::DropUse(size_t LUIdx, size_t NewLUIdx) {
158 // Remove the use index from every register's use list.
159 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
160 I != E; ++I) {
161 SmallBitVector &UsedByIndices = I->second.UsedByIndices;
162 UsedByIndices.resize(std::max(UsedByIndices.size(), NewLUIdx + 1));
163 if (LUIdx < UsedByIndices.size()) {
164 UsedByIndices[LUIdx] = UsedByIndices[NewLUIdx];
165 UsedByIndices.reset(NewLUIdx);
166 } else
167 UsedByIndices.reset(LUIdx);
168 }
169 }
170
171 /// DropUse - Clear out reference by use LUIdx.
172153 void
173154 RegUseTracker::DropUse(size_t LUIdx) {
174155 // Remove the use index from every register's use list.
13521333 UseMapDenseMapInfo> UseMapTy;
13531334 UseMapTy UseMap;
13541335
1355 bool reconcileNewOffset(LSRUse &LU,
1356 int64_t NewMinOffset, int64_t NewMaxOffset,
1357 bool HasBaseReg,
1336 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
13581337 LSRUse::KindType Kind, const Type *AccessTy);
13591338
13601339 std::pair getUse(const SCEV *&Expr,
13631342
13641343 void DeleteUse(LSRUse &LU);
13651344
1366 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU,
1367 int64_t &NewBaseOffs);
1345 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
13681346
13691347 public:
13701348 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
18651843 /// at the given offset and other details. If so, update the use and
18661844 /// return true.
18671845 bool
1868 LSRInstance::reconcileNewOffset(LSRUse &LU,
1869 int64_t NewMinOffset, int64_t NewMaxOffset,
1870 bool HasBaseReg,
1846 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
18711847 LSRUse::KindType Kind, const Type *AccessTy) {
1872 int64_t ResultMinOffset = LU.MinOffset;
1873 int64_t ResultMaxOffset = LU.MaxOffset;
1874 const Type *ResultAccessTy = AccessTy;
1848 int64_t NewMinOffset = LU.MinOffset;
1849 int64_t NewMaxOffset = LU.MaxOffset;
1850 const Type *NewAccessTy = AccessTy;
18751851
18761852 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to
18771853 // something conservative, however this can pessimize in the case that one of
18791855 if (LU.Kind != Kind)
18801856 return false;
18811857 // Conservatively assume HasBaseReg is true for now.
1882 if (NewMinOffset < LU.MinOffset) {
1883 if (!isAlwaysFoldable(LU.MaxOffset - NewMinOffset, 0, HasBaseReg,
1858 if (NewOffset < LU.MinOffset) {
1859 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
18841860 Kind, AccessTy, TLI))
18851861 return false;
1886 ResultMinOffset = NewMinOffset;
1887 } else if (NewMaxOffset > LU.MaxOffset) {
1888 if (!isAlwaysFoldable(NewMaxOffset - LU.MinOffset, 0, HasBaseReg,
1862 NewMinOffset = NewOffset;
1863 } else if (NewOffset > LU.MaxOffset) {
1864 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
18891865 Kind, AccessTy, TLI))
18901866 return false;
1891 ResultMaxOffset = NewMaxOffset;
1867 NewMaxOffset = NewOffset;
18921868 }
18931869 // Check for a mismatched access type, and fall back conservatively as needed.
18941870 // TODO: Be less conservative when the type is similar and can use the same
18951871 // addressing modes.
18961872 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
1897 ResultAccessTy = Type::getVoidTy(AccessTy->getContext());
1873 NewAccessTy = Type::getVoidTy(AccessTy->getContext());
18981874
18991875 // Update the use.
1900 LU.MinOffset = ResultMinOffset;
1901 LU.MaxOffset = ResultMaxOffset;
1902 LU.AccessTy = ResultAccessTy;
1876 LU.MinOffset = NewMinOffset;
1877 LU.MaxOffset = NewMaxOffset;
1878 LU.AccessTy = NewAccessTy;
1879 if (NewOffset != LU.Offsets.back())
1880 LU.Offsets.push_back(NewOffset);
19031881 return true;
19041882 }
19051883
19241902 // A use already existed with this base.
19251903 size_t LUIdx = P.first->second;
19261904 LSRUse &LU = Uses[LUIdx];
1927 if (reconcileNewOffset(LU, Offset, Offset,
1928 /*HasBaseReg=*/true, Kind, AccessTy)) {
1929 LU.Offsets.push_back(Offset);
1905 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
19301906 // Reuse this use.
19311907 return std::make_pair(LUIdx, Offset);
1932 }
19331908 }
19341909
19351910 // Create a new use.
19381913 Uses.push_back(LSRUse(Kind, AccessTy));
19391914 LSRUse &LU = Uses[LUIdx];
19401915
1941 LU.Offsets.push_back(Offset);
1916 // We don't need to track redundant offsets, but we don't need to go out
1917 // of our way here to avoid them.
1918 if (LU.Offsets.empty() || Offset != LU.Offsets.back())
1919 LU.Offsets.push_back(Offset);
1920
19421921 LU.MinOffset = Offset;
19431922 LU.MaxOffset = Offset;
19441923 return std::make_pair(LUIdx, Offset);
19461925
19471926 /// DeleteUse - Delete the given use from the Uses list.
19481927 void LSRInstance::DeleteUse(LSRUse &LU) {
1949 if (&LU != &Uses.back()) {
1928 if (&LU != &Uses.back())
19501929 std::swap(LU, Uses.back());
1951 RegUses.DropUse(&LU - Uses.begin(), Uses.size() - 1);
1952 } else {
1953 RegUses.DropUse(&LU - Uses.begin());
1954 }
19551930 Uses.pop_back();
19561931 }
19571932
19591934 /// a formula that has the same registers as the given formula.
19601935 LSRUse *
19611936 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
1962 const LSRUse &OrigLU,
1963 int64_t &NewBaseOffs) {
1964 // Search all uses for a formula similar to OrigF. This could be more clever.
1937 const LSRUse &OrigLU) {
1938 // Search all uses for the formula. This could be more clever.
19651939 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
19661940 LSRUse &LU = Uses[LUIdx];
19671941 // Check whether this use is close enough to OrigLU, to see whether it's
19841958 F.ScaledReg == OrigF.ScaledReg &&
19851959 F.AM.BaseGV == OrigF.AM.BaseGV &&
19861960 F.AM.Scale == OrigF.AM.Scale) {
1987 // Ok, all the registers and symbols matched. Check to see if the
1988 // immediate looks nicer than our old one.
1989 if (OrigF.AM.BaseOffs == INT64_MIN ||
1990 (F.AM.BaseOffs != INT64_MIN &&
1991 abs64(F.AM.BaseOffs) < abs64(OrigF.AM.BaseOffs))) {
1992 // Looks good. Take it.
1993 NewBaseOffs = F.AM.BaseOffs;
1961 if (F.AM.BaseOffs == 0)
19941962 return &LU;
1995 }
19961963 // This is the formula where all the registers and symbols matched;
19971964 // there aren't going to be any others. Since we declined it, we
19981965 // can skip the rest of the formulae and procede to the next LSRUse.
26332600 WorkItem(size_t LI, int64_t I, const SCEV *R)
26342601 : LUIdx(LI), Imm(I), OrigReg(R) {}
26352602
2636 bool operator==(const WorkItem &that) const {
2637 return LUIdx == that.LUIdx && Imm == that.Imm && OrigReg == that.OrigReg;
2638 }
2639 bool operator<(const WorkItem &that) const {
2640 if (LUIdx != that.LUIdx)
2641 return LUIdx < that.LUIdx;
2642 if (Imm != that.Imm)
2643 return Imm < that.Imm;
2644 return OrigReg < that.OrigReg;
2645 }
2646
26472603 void print(raw_ostream &OS) const;
26482604 void dump() const;
26492605 };
26832639 // Now examine each set of registers with the same base value. Build up
26842640 // a list of work to do and do the work in a separate step so that we're
26852641 // not adding formulae and register counts while we're searching.
2686 SmallSetVector WorkItems;
2642 SmallVector WorkItems;
2643 SmallSet, 32> UniqueItems;
26872644 for (SmallVectorImpl::const_iterator I = Sequence.begin(),
26882645 E = Sequence.end(); I != E; ++I) {
26892646 const SCEV *Reg = *I;
27262683 // Compute the difference between the two.
27272684 int64_t Imm = (uint64_t)JImm - M->first;
27282685 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
2729 LUIdx = UsedByIndices.find_next(LUIdx)) {
2686 LUIdx = UsedByIndices.find_next(LUIdx))
27302687 // Make a memo of this use, offset, and register tuple.
2731 WorkItems.insert(WorkItem(LUIdx, Imm, OrigReg));
2732 }
2688 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
2689 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
27332690 }
27342691 }
27352692 }
27372694 Map.clear();
27382695 Sequence.clear();
27392696 UsedByIndicesMap.clear();
2697 UniqueItems.clear();
27402698
27412699 // Now iterate through the worklist and add new formulae.
27422700 for (SmallVectorImpl::const_iterator I = WorkItems.begin(),
30332991 E = LU.Formulae.end(); I != E; ++I) {
30342992 const Formula &F = *I;
30352993 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
3036 int64_t NewBaseOffs;
3037 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU,
3038 NewBaseOffs)) {
3039 if (reconcileNewOffset(*LUThatHas,
3040 F.AM.BaseOffs + LU.MinOffset - NewBaseOffs,
3041 F.AM.BaseOffs + LU.MaxOffset - NewBaseOffs,
2994 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
2995 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
30422996 /*HasBaseReg=*/false,
30432997 LU.Kind, LU.AccessTy)) {
30442998 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
30452999 dbgs() << '\n');
30463000
30473001 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
3048
3049 // Update the relocs to reference the new use.
3050 // Do this first so that MinOffset and MaxOffset are updated
3051 // before we begin to determine which formulae to delete.
3052 for (SmallVectorImpl::iterator I = Fixups.begin(),
3053 E = Fixups.end(); I != E; ++I) {
3054 LSRFixup &Fixup = *I;
3055 if (Fixup.LUIdx == LUIdx) {
3056 Fixup.LUIdx = LUThatHas - &Uses.front();
3057 Fixup.Offset += F.AM.BaseOffs - NewBaseOffs;
3058 DEBUG(dbgs() << "New fixup has offset "
3059 << Fixup.Offset << '\n');
3060 LUThatHas->Offsets.push_back(Fixup.Offset);
3061 if (Fixup.Offset > LUThatHas->MaxOffset)
3062 LUThatHas->MaxOffset = Fixup.Offset;
3063 if (Fixup.Offset < LUThatHas->MinOffset)
3064 LUThatHas->MinOffset = Fixup.Offset;
3065 }
3066 // DeleteUse will do a swap+pop_back, so if this fixup is
3067 // now pointing to the last LSRUse, update it to point to the
3068 // position it'll be swapped to.
3069 if (Fixup.LUIdx == NumUses-1)
3070 Fixup.LUIdx = LUIdx;
3071 }
30723002
30733003 // Delete formulae from the new use which are no longer legal.
30743004 bool Any = false;
30873017 }
30883018 if (Any)
30893019 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
3020
3021 // Update the relocs to reference the new use.
3022 for (SmallVectorImpl::iterator I = Fixups.begin(),
3023 E = Fixups.end(); I != E; ++I) {
3024 LSRFixup &Fixup = *I;
3025 if (Fixup.LUIdx == LUIdx) {
3026 Fixup.LUIdx = LUThatHas - &Uses.front();
3027 Fixup.Offset += F.AM.BaseOffs;
3028 DEBUG(dbgs() << "New fixup has offset "
3029 << Fixup.Offset << '\n');
3030 }
3031 if (Fixup.LUIdx == NumUses-1)
3032 Fixup.LUIdx = LUIdx;
3033 }
30903034
30913035 // Delete the old use.
30923036 DeleteUse(LU);
451451 ; CHECK-NEXT: addss %xmm{{.*}}, %xmm{{.*}}
452452 ; CHECK-NEXT: mulss (%r{{[^,]*}}), %xmm{{.*}}
453453 ; CHECK-NEXT: movss %xmm{{.*}}, (%r{{[^,]*}})
454 ; CHECK-NEXT: addq $4, %r{{.*}}
454455 ; CHECK-NEXT: decq %r{{.*}}
455 ; CHECK-NEXT: addq $4, %r{{.*}}
456456 ; CHECK-NEXT: addq $4, %r{{.*}}
457457 ; CHECK-NEXT: movaps %xmm{{.*}}, %xmm{{.*}}
458458 ; CHECK-NEXT: BB10_2: