RPCS3/llvm
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm.git synced 2025-01-10 22:46:25 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Revert 112442 and 112440 until the compile time problems introduced

Browse Source 
by 112440 are resolved.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@112692 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Dan Gohman 2010-09-01 01:45:53 +00:00
parent 9cfad89a68
commit 191bd64a39
4 changed files with 133 additions and 203 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
include/llvm/Analysis/IVUsers.h
View File

@ -27,7 +27,6 @@ class Value;
class IVUsers;
class ScalarEvolution;
class SCEV;
class SCEVAddRecExpr;
class IVUsers;
/// IVStrideUse - Keep track of one use of a strided induction variable.
@ -123,7 +122,7 @@ class IVUsers : public LoopPass {
LoopInfo *LI;
DominatorTree *DT;
ScalarEvolution *SE;
SmallPtrSet<Instruction *, 16> Processed;
SmallPtrSet<Instruction*,16> Processed;
/// IVUses - A list of all tracked IV uses of induction variable expressions
/// we are interested in.
@ -135,16 +134,14 @@ class IVUsers : public LoopPass {
virtual void releaseMemory();
const SCEVAddRecExpr *findInterestingAddRec(const SCEV *S) const;
bool isInterestingUser(const Instruction *User) const;
public:
static char ID; // Pass ID, replacement for typeid
IVUsers();
/// AddUsersIfInteresting - Inspect the def-use graph starting at the
/// specified Instruction and add IVUsers.
void AddUsersIfInteresting(Instruction *I);
/// AddUsersIfInteresting - Inspect the specified Instruction. If it is a
/// reducible SCEV, recursively add its users to the IVUsesByStride set and
/// return true. Otherwise, return false.
bool AddUsersIfInteresting(Instruction *I);
IVStrideUse &AddUser(Instruction *User, Value *Operand);

165
lib/Analysis/IVUsers.cpp
View File

@ -35,123 +35,112 @@ Pass *llvm::createIVUsersPass() {
return new IVUsers();
}
/// findInterestingAddRec - Test whether the given expression is interesting.
/// Return the addrec with the current loop which makes it interesting, or
/// null if it is not interesting.
const SCEVAddRecExpr *IVUsers::findInterestingAddRec(const SCEV *S) const {
/// isInteresting - Test whether the given expression is "interesting" when
/// used by the given expression, within the context of analyzing the
/// given loop.
static bool isInteresting(const SCEV *S, const Instruction *I, const Loop *L,
ScalarEvolution *SE) {
// An addrec is interesting if it's affine or if it has an interesting start.
if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
// Keep things simple. Don't touch loop-variant strides.
if (AR->getLoop() == L)
return AR;
// We don't yet know how to do effective SCEV expansions for addrecs
// with interesting steps.
if (findInterestingAddRec(AR->getStepRecurrence(*SE)))
return 0;
// Otherwise recurse to see if the start value is interesting.
return findInterestingAddRec(AR->getStart());
return AR->isAffine() || !L->contains(I);
// Otherwise recurse to see if the start value is interesting, and that
// the step value is not interesting, since we don't yet know how to
// do effective SCEV expansions for addrecs with interesting steps.
return isInteresting(AR->getStart(), I, L, SE) &&
!isInteresting(AR->getStepRecurrence(*SE), I, L, SE);
}
// An add is interesting if exactly one of its operands is interesting.
if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
bool AnyInterestingYet = false;
for (SCEVAddExpr::op_iterator OI = Add->op_begin(), OE = Add->op_end();
OI != OE; ++OI)
if (const SCEVAddRecExpr *AR = findInterestingAddRec(*OI))
return AR;
return 0;
if (isInteresting(*OI, I, L, SE)) {
if (AnyInterestingYet)
return false;
AnyInterestingYet = true;
}
return AnyInterestingYet;
}
// Nothing else is interesting here.
return 0;
}
bool IVUsers::isInterestingUser(const Instruction *User) const {
// Void and FP expressions cannot be reduced.
if (!SE->isSCEVable(User->getType()))
return false;
// LSR is not APInt clean, do not touch integers bigger than 64-bits.
if (SE->getTypeSizeInBits(User->getType()) > 64)
return false;
// Don't descend into PHI nodes outside the current loop.
if (LI->getLoopFor(User->getParent()) != L &&
isa<PHINode>(User))
return false;
// Otherwise, it may be interesting.
return true;
return false;
}
/// AddUsersIfInteresting - Inspect the specified instruction. If it is a
/// reducible SCEV, recursively add its users to the IVUsesByStride set and
/// return true. Otherwise, return false.
void IVUsers::AddUsersIfInteresting(Instruction *I) {
// Stop if we've seen this before.
if (!Processed.insert(I))
return;
// If this PHI node is not SCEVable, ignore it.
bool IVUsers::AddUsersIfInteresting(Instruction *I) {
if (!SE->isSCEVable(I->getType()))
return;
return false; // Void and FP expressions cannot be reduced.
// If this PHI node is not an addrec for this loop, ignore it.
const SCEVAddRecExpr *Expr = findInterestingAddRec(SE->getSCEV(I));
if (!Expr)
return;
// LSR is not APInt clean, do not touch integers bigger than 64-bits.
if (SE->getTypeSizeInBits(I->getType()) > 64)
return false;
// Walk the def-use graph.
SmallVector<std::pair<Instruction *, const SCEVAddRecExpr *>, 16> Worklist;
Worklist.push_back(std::make_pair(I, Expr));
do {
std::pair<Instruction *, const SCEVAddRecExpr *> P =
Worklist.pop_back_val();
Instruction *Op = P.first;
const SCEVAddRecExpr *OpAR = P.second;
if (!Processed.insert(I))
return true; // Instruction already handled.
// Visit Op's users.
SmallPtrSet<Instruction *, 8> VisitedUsers;
for (Value::use_iterator UI = Op->use_begin(), E = Op->use_end();
UI != E; ++UI) {
// Don't visit any individual user more than once.
Instruction *User = cast<Instruction>(*UI);
if (!VisitedUsers.insert(User))
continue;
// Get the symbolic expression for this instruction.
const SCEV *ISE = SE->getSCEV(I);
// If it's an affine addrec (which we can pretty safely re-expand) inside
// the loop, or a potentially non-affine addrec outside the loop (which
// we can evaluate outside of the loop), follow it.
if (OpAR->isAffine() || !L->contains(User)) {
if (isInterestingUser(User)) {
const SCEV *UserExpr = SE->getSCEV(User);
// If we've come to an uninteresting expression, stop the traversal and
// call this a user.
if (!isInteresting(ISE, I, L, SE))
return false;
if (const SCEVAddRecExpr *AR = findInterestingAddRec(UserExpr)) {
// Interesting. Keep searching.
if (Processed.insert(User))
Worklist.push_back(std::make_pair(User, AR));
continue;
}
}
SmallPtrSet<Instruction *, 4> UniqueUsers;
for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
UI != E; ++UI) {
Instruction *User = cast<Instruction>(*UI);
if (!UniqueUsers.insert(User))
continue;
// Do not infinitely recurse on PHI nodes.
if (isa<PHINode>(User) && Processed.count(User))
continue;
// Descend recursively, but not into PHI nodes outside the current loop.
// It's important to see the entire expression outside the loop to get
// choices that depend on addressing mode use right, although we won't
// consider references outside the loop in all cases.
// If User is already in Processed, we don't want to recurse into it again,
// but do want to record a second reference in the same instruction.
bool AddUserToIVUsers = false;
if (LI->getLoopFor(User->getParent()) != L) {
if (isa<PHINode>(User) || Processed.count(User) ||
!AddUsersIfInteresting(User)) {
DEBUG(dbgs() << "FOUND USER in other loop: " << *User << '\n'
<< " OF SCEV: " << *ISE << '\n');
AddUserToIVUsers = true;
}
// Otherwise, this is the point where the def-use chain
// becomes uninteresting. Call it an IV User.
AddUser(User, Op);
} else if (Processed.count(User) ||
!AddUsersIfInteresting(User)) {
DEBUG(dbgs() << "FOUND USER: " << *User << '\n'
<< " OF SCEV: " << *ISE << '\n');
AddUserToIVUsers = true;
}
} while (!Worklist.empty());
if (AddUserToIVUsers) {
// Okay, we found a user that we cannot reduce.
IVUses.push_back(new IVStrideUse(this, User, I));
IVStrideUse &NewUse = IVUses.back();
// Transform the expression into a normalized form.
ISE = TransformForPostIncUse(NormalizeAutodetect,
ISE, User, I,
NewUse.PostIncLoops,
*SE, *DT);
DEBUG(dbgs() << " NORMALIZED TO: " << *ISE << '\n');
}
}
return true;
}
IVStrideUse &IVUsers::AddUser(Instruction *User, Value *Operand) {
IVUses.push_back(new IVStrideUse(this, User, Operand));
IVStrideUse &NewUse = IVUses.back();
// Auto-detect and remember post-inc loops for this expression.
const SCEV *S = SE->getSCEV(Operand);
(void)TransformForPostIncUse(NormalizeAutodetect,
S, User, Operand,
NewUse.PostIncLoops,
*SE, *DT);
return NewUse;
return IVUses.back();
}
IVUsers::IVUsers()
@ -176,7 +165,7 @@ bool IVUsers::runOnLoop(Loop *l, LPPassManager &LPM) {
// them by stride. Start by finding all of the PHI nodes in the header for
// this loop. If they are induction variables, inspect their uses.
for (BasicBlock::iterator I = L->getHeader()->begin(); isa<PHINode>(I); ++I)
AddUsersIfInteresting(I);
(void)AddUsersIfInteresting(I);
return false;
}

156
lib/Transforms/Scalar/LoopStrengthReduce.cpp
View File

@ -113,7 +113,6 @@ class RegUseTracker {
public:
void CountRegister(const SCEV *Reg, size_t LUIdx);
void DropRegister(const SCEV *Reg, size_t LUIdx);
void DropUse(size_t LUIdx, size_t NewLUIdx);
void DropUse(size_t LUIdx);
bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const;
@ -152,24 +151,6 @@ RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) {
RSD.UsedByIndices.reset(LUIdx);
}
/// DropUse - Clear out reference by use LUIdx, and prepare for use NewLUIdx
/// to be swapped into LUIdx's position.
void
RegUseTracker::DropUse(size_t LUIdx, size_t NewLUIdx) {
// Remove the use index from every register's use list.
for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end();
I != E; ++I) {
SmallBitVector &UsedByIndices = I->second.UsedByIndices;
UsedByIndices.resize(std::max(UsedByIndices.size(), NewLUIdx + 1));
if (LUIdx < UsedByIndices.size()) {
UsedByIndices[LUIdx] = UsedByIndices[NewLUIdx];
UsedByIndices.reset(NewLUIdx);
} else
UsedByIndices.reset(LUIdx);
}
}
/// DropUse - Clear out reference by use LUIdx.
void
RegUseTracker::DropUse(size_t LUIdx) {
// Remove the use index from every register's use list.
@ -1353,9 +1334,7 @@ class LSRInstance {
UseMapDenseMapInfo> UseMapTy;
UseMapTy UseMap;
bool reconcileNewOffset(LSRUse &LU,
int64_t NewMinOffset, int64_t NewMaxOffset,
bool HasBaseReg,
bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
LSRUse::KindType Kind, const Type *AccessTy);
std::pair<size_t, int64_t> getUse(const SCEV *&Expr,
@ -1364,8 +1343,7 @@ class LSRInstance {
void DeleteUse(LSRUse &LU);
LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU,
int64_t &NewBaseOffs);
LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU);
public:
void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx);
@ -1866,13 +1844,11 @@ LSRInstance::OptimizeLoopTermCond() {
/// at the given offset and other details. If so, update the use and
/// return true.
bool
LSRInstance::reconcileNewOffset(LSRUse &LU,
int64_t NewMinOffset, int64_t NewMaxOffset,
bool HasBaseReg,
LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg,
LSRUse::KindType Kind, const Type *AccessTy) {
int64_t ResultMinOffset = LU.MinOffset;
int64_t ResultMaxOffset = LU.MaxOffset;
const Type *ResultAccessTy = AccessTy;
int64_t NewMinOffset = LU.MinOffset;
int64_t NewMaxOffset = LU.MaxOffset;
const Type *NewAccessTy = AccessTy;
// Check for a mismatched kind. It's tempting to collapse mismatched kinds to
// something conservative, however this can pessimize in the case that one of
@ -1880,27 +1856,29 @@ LSRInstance::reconcileNewOffset(LSRUse &LU,
if (LU.Kind != Kind)
return false;
// Conservatively assume HasBaseReg is true for now.
if (NewMinOffset < LU.MinOffset) {
if (!isAlwaysFoldable(LU.MaxOffset - NewMinOffset, 0, HasBaseReg,
if (NewOffset < LU.MinOffset) {
if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
Kind, AccessTy, TLI))
return false;
ResultMinOffset = NewMinOffset;
} else if (NewMaxOffset > LU.MaxOffset) {
if (!isAlwaysFoldable(NewMaxOffset - LU.MinOffset, 0, HasBaseReg,
NewMinOffset = NewOffset;
} else if (NewOffset > LU.MaxOffset) {
if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
Kind, AccessTy, TLI))
return false;
ResultMaxOffset = NewMaxOffset;
NewMaxOffset = NewOffset;
}
// Check for a mismatched access type, and fall back conservatively as needed.
// TODO: Be less conservative when the type is similar and can use the same
// addressing modes.
if (Kind == LSRUse::Address && AccessTy != LU.AccessTy)
ResultAccessTy = Type::getVoidTy(AccessTy->getContext());
NewAccessTy = Type::getVoidTy(AccessTy->getContext());
// Update the use.
LU.MinOffset = ResultMinOffset;
LU.MaxOffset = ResultMaxOffset;
LU.AccessTy = ResultAccessTy;
LU.MinOffset = NewMinOffset;
LU.MaxOffset = NewMaxOffset;
LU.AccessTy = NewAccessTy;
if (NewOffset != LU.Offsets.back())
LU.Offsets.push_back(NewOffset);
return true;
}
@ -1925,12 +1903,9 @@ LSRInstance::getUse(const SCEV *&Expr,
// A use already existed with this base.
size_t LUIdx = P.first->second;
LSRUse &LU = Uses[LUIdx];
if (reconcileNewOffset(LU, Offset, Offset,
/*HasBaseReg=*/true, Kind, AccessTy)) {
LU.Offsets.push_back(Offset);
if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy))
// Reuse this use.
return std::make_pair(LUIdx, Offset);
}
}
// Create a new use.
@ -1939,7 +1914,11 @@ LSRInstance::getUse(const SCEV *&Expr,
Uses.push_back(LSRUse(Kind, AccessTy));
LSRUse &LU = Uses[LUIdx];
LU.Offsets.push_back(Offset);
// We don't need to track redundant offsets, but we don't need to go out
// of our way here to avoid them.
if (LU.Offsets.empty() || Offset != LU.Offsets.back())
LU.Offsets.push_back(Offset);
LU.MinOffset = Offset;
LU.MaxOffset = Offset;
return std::make_pair(LUIdx, Offset);
@ -1947,12 +1926,8 @@ LSRInstance::getUse(const SCEV *&Expr,
/// DeleteUse - Delete the given use from the Uses list.
void LSRInstance::DeleteUse(LSRUse &LU) {
if (&LU != &Uses.back()) {
if (&LU != &Uses.back())
std::swap(LU, Uses.back());
RegUses.DropUse(&LU - Uses.begin(), Uses.size() - 1);
} else {
RegUses.DropUse(&LU - Uses.begin());
}
Uses.pop_back();
}
@ -1960,9 +1935,8 @@ void LSRInstance::DeleteUse(LSRUse &LU) {
/// a formula that has the same registers as the given formula.
LSRUse *
LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
const LSRUse &OrigLU,
int64_t &NewBaseOffs) {
// Search all uses for a formula similar to OrigF. This could be more clever.
const LSRUse &OrigLU) {
// Search all uses for the formula. This could be more clever.
for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) {
LSRUse &LU = Uses[LUIdx];
// Check whether this use is close enough to OrigLU, to see whether it's
@ -1985,15 +1959,8 @@ LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF,
F.ScaledReg == OrigF.ScaledReg &&
F.AM.BaseGV == OrigF.AM.BaseGV &&
F.AM.Scale == OrigF.AM.Scale) {
// Ok, all the registers and symbols matched. Check to see if the
// immediate looks nicer than our old one.
if (OrigF.AM.BaseOffs == INT64_MIN ||
(F.AM.BaseOffs != INT64_MIN &&
abs64(F.AM.BaseOffs) < abs64(OrigF.AM.BaseOffs))) {
// Looks good. Take it.
NewBaseOffs = F.AM.BaseOffs;
if (F.AM.BaseOffs == 0)
return &LU;
}
// This is the formula where all the registers and symbols matched;
// there aren't going to be any others. Since we declined it, we
// can skip the rest of the formulae and procede to the next LSRUse.
@ -2634,17 +2601,6 @@ struct WorkItem {
WorkItem(size_t LI, int64_t I, const SCEV *R)
: LUIdx(LI), Imm(I), OrigReg(R) {}
bool operator==(const WorkItem &that) const {
return LUIdx == that.LUIdx && Imm == that.Imm && OrigReg == that.OrigReg;
}
bool operator<(const WorkItem &that) const {
if (LUIdx != that.LUIdx)
return LUIdx < that.LUIdx;
if (Imm != that.Imm)
return Imm < that.Imm;
return OrigReg < that.OrigReg;
}
void print(raw_ostream &OS) const;
void dump() const;
};
@ -2684,7 +2640,8 @@ void LSRInstance::GenerateCrossUseConstantOffsets() {
// Now examine each set of registers with the same base value. Build up
// a list of work to do and do the work in a separate step so that we're
// not adding formulae and register counts while we're searching.
SmallSetVector<WorkItem, 32> WorkItems;
SmallVector<WorkItem, 32> WorkItems;
SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems;
for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(),
E = Sequence.end(); I != E; ++I) {
const SCEV *Reg = *I;
@ -2727,10 +2684,10 @@ void LSRInstance::GenerateCrossUseConstantOffsets() {
// Compute the difference between the two.
int64_t Imm = (uint64_t)JImm - M->first;
for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1;
LUIdx = UsedByIndices.find_next(LUIdx)) {
LUIdx = UsedByIndices.find_next(LUIdx))
// Make a memo of this use, offset, and register tuple.
WorkItems.insert(WorkItem(LUIdx, Imm, OrigReg));
}
if (UniqueItems.insert(std::make_pair(LUIdx, Imm)))
WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg));
}
}
}
@ -2738,6 +2695,7 @@ void LSRInstance::GenerateCrossUseConstantOffsets() {
Map.clear();
Sequence.clear();
UsedByIndicesMap.clear();
UniqueItems.clear();
// Now iterate through the worklist and add new formulae.
for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(),
@ -3034,12 +2992,8 @@ void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
E = LU.Formulae.end(); I != E; ++I) {
const Formula &F = *I;
if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) {
int64_t NewBaseOffs;
if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU,
NewBaseOffs)) {
if (reconcileNewOffset(*LUThatHas,
F.AM.BaseOffs + LU.MinOffset - NewBaseOffs,
F.AM.BaseOffs + LU.MaxOffset - NewBaseOffs,
if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) {
if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs,
/*HasBaseReg=*/false,
LU.Kind, LU.AccessTy)) {
DEBUG(dbgs() << " Deleting use "; LU.print(dbgs());
@ -3047,30 +3001,6 @@ void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop;
// Update the relocs to reference the new use.
// Do this first so that MinOffset and MaxOffset are updated
// before we begin to determine which formulae to delete.
for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
E = Fixups.end(); I != E; ++I) {
LSRFixup &Fixup = *I;
if (Fixup.LUIdx == LUIdx) {
Fixup.LUIdx = LUThatHas - &Uses.front();
Fixup.Offset += F.AM.BaseOffs - NewBaseOffs;
DEBUG(dbgs() << "New fixup has offset "
<< Fixup.Offset << '\n');
LUThatHas->Offsets.push_back(Fixup.Offset);
if (Fixup.Offset > LUThatHas->MaxOffset)
LUThatHas->MaxOffset = Fixup.Offset;
if (Fixup.Offset < LUThatHas->MinOffset)
LUThatHas->MinOffset = Fixup.Offset;
}
// DeleteUse will do a swap+pop_back, so if this fixup is
// now pointing to the last LSRUse, update it to point to the
// position it'll be swapped to.
if (Fixup.LUIdx == NumUses-1)
Fixup.LUIdx = LUIdx;
}
// Delete formulae from the new use which are no longer legal.
bool Any = false;
for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) {
@ -3089,6 +3019,20 @@ void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() {
if (Any)
LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses);
// Update the relocs to reference the new use.
for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(),
E = Fixups.end(); I != E; ++I) {
LSRFixup &Fixup = *I;
if (Fixup.LUIdx == LUIdx) {
Fixup.LUIdx = LUThatHas - &Uses.front();
Fixup.Offset += F.AM.BaseOffs;
DEBUG(dbgs() << "New fixup has offset "
<< Fixup.Offset << '\n');
}
if (Fixup.LUIdx == NumUses-1)
Fixup.LUIdx = LUIdx;
}
// Delete the old use.
DeleteUse(LU);
--LUIdx;

2
test/CodeGen/X86/lsr-reuse.ll
View File

@ -452,8 +452,8 @@ bb5: ; preds = %bb3, %entry
; CHECK-NEXT: addss %xmm{{.*}}, %xmm{{.*}}
; CHECK-NEXT: mulss (%r{{[^,]*}}), %xmm{{.*}}
; CHECK-NEXT: movss %xmm{{.*}}, (%r{{[^,]*}})
; CHECK-NEXT: decq %r{{.*}}
; CHECK-NEXT: addq $4, %r{{.*}}
; CHECK-NEXT: decq %r{{.*}}
; CHECK-NEXT: addq $4, %r{{.*}}
; CHECK-NEXT: movaps %xmm{{.*}}, %xmm{{.*}}
; CHECK-NEXT: BB10_2: