aboutsummaryrefslogtreecommitdiff
path: root/lib/CodeGen/MachineLICM.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'lib/CodeGen/MachineLICM.cpp')
-rw-r--r--lib/CodeGen/MachineLICM.cpp156
1 files changed, 128 insertions, 28 deletions
diff --git a/lib/CodeGen/MachineLICM.cpp b/lib/CodeGen/MachineLICM.cpp
index 722ceb202439..a1f80d5282e0 100644
--- a/lib/CodeGen/MachineLICM.cpp
+++ b/lib/CodeGen/MachineLICM.cpp
@@ -37,10 +37,16 @@
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/Statistic.h"
+#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;
+static cl::opt<bool>
+AvoidSpeculation("avoid-speculation",
+ cl::desc("MachineLICM should avoid speculation"),
+ cl::init(false), cl::Hidden);
+
STATISTIC(NumHoisted,
"Number of machine instructions hoisted out of loops");
STATISTIC(NumLowRP,
@@ -91,6 +97,17 @@ namespace {
// For each opcode, keep a list of potential CSE instructions.
DenseMap<unsigned, std::vector<const MachineInstr*> > CSEMap;
+ enum {
+ SpeculateFalse = 0,
+ SpeculateTrue = 1,
+ SpeculateUnknown = 2
+ };
+
+ // If a MBB does not dominate loop exiting blocks then it may not safe
+ // to hoist loads from this block.
+ // Tri-state: 0 - false, 1 - true, 2 - unknown
+ unsigned SpeculationState;
+
public:
static char ID; // Pass identification, replacement for typeid
MachineLICM() :
@@ -194,6 +211,10 @@ namespace {
/// hoist the given loop invariant.
bool IsProfitableToHoist(MachineInstr &MI);
+ /// IsGuaranteedToExecute - Check if this mbb is guaranteed to execute.
+ /// If not then a load from this mbb may not be safe to hoist.
+ bool IsGuaranteedToExecute(MachineBasicBlock *BB);
+
/// HoistRegion - Walk the specified region of the CFG (defined by all
/// blocks dominated by the specified block, and that are in the current
/// loop) in depth first order w.r.t the DominatorTree. This allows us to
@@ -202,6 +223,13 @@ namespace {
///
void HoistRegion(MachineDomTreeNode *N, bool IsHeader = false);
+ /// getRegisterClassIDAndCost - For a given MI, register, and the operand
+ /// index, return the ID and cost of its representative register class by
+ /// reference.
+ void getRegisterClassIDAndCost(const MachineInstr *MI,
+ unsigned Reg, unsigned OpIdx,
+ unsigned &RCId, unsigned &RCCost) const;
+
/// InitRegPressure - Find all virtual register references that are liveout
/// of the preheader to initialize the starting "register pressure". Note
/// this does not count live through (livein but not used) registers.
@@ -229,6 +257,10 @@ namespace {
bool EliminateCSE(MachineInstr *MI,
DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator &CI);
+ /// MayCSE - Return true if the given instruction will be CSE'd if it's
+ /// hoisted out of the loop.
+ bool MayCSE(MachineInstr *MI);
+
/// Hoist - When an instruction is found to only use loop invariant operands
/// that is safe to hoist, this instruction is called to do the dirty work.
/// It returns true if the instruction is hoisted.
@@ -441,6 +473,12 @@ void MachineLICM::HoistRegionPostRA() {
const std::vector<MachineBasicBlock*> Blocks = CurLoop->getBlocks();
for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
MachineBasicBlock *BB = Blocks[i];
+
+ // If the header of the loop containing this basic block is a landing pad,
+ // then don't try to hoist instructions out of this loop.
+ const MachineLoop *ML = MLI->getLoopFor(BB);
+ if (ML && ML->getHeader()->isLandingPad()) continue;
+
// Conservatively treat live-in's as an external def.
// FIXME: That means a reload that're reused in successor block(s) will not
// be LICM'ed.
@@ -452,6 +490,7 @@ void MachineLICM::HoistRegionPostRA() {
++PhysRegDefs[*AS];
}
+ SpeculationState = SpeculateUnknown;
for (MachineBasicBlock::iterator
MII = BB->begin(), E = BB->end(); MII != E; ++MII) {
MachineInstr *MI = &*MII;
@@ -545,6 +584,27 @@ void MachineLICM::HoistPostRA(MachineInstr *MI, unsigned Def) {
Changed = true;
}
+// IsGuaranteedToExecute - Check if this mbb is guaranteed to execute.
+// If not then a load from this mbb may not be safe to hoist.
+bool MachineLICM::IsGuaranteedToExecute(MachineBasicBlock *BB) {
+ if (SpeculationState != SpeculateUnknown)
+ return SpeculationState == SpeculateFalse;
+
+ if (BB != CurLoop->getHeader()) {
+ // Check loop exiting blocks.
+ SmallVector<MachineBasicBlock*, 8> CurrentLoopExitingBlocks;
+ CurLoop->getExitingBlocks(CurrentLoopExitingBlocks);
+ for (unsigned i = 0, e = CurrentLoopExitingBlocks.size(); i != e; ++i)
+ if (!DT->dominates(BB, CurrentLoopExitingBlocks[i])) {
+ SpeculationState = SpeculateTrue;
+ return false;
+ }
+ }
+
+ SpeculationState = SpeculateFalse;
+ return true;
+}
+
/// HoistRegion - Walk the specified region of the CFG (defined by all blocks
/// dominated by the specified block, and that are in the current loop) in depth
/// first order w.r.t the DominatorTree. This allows us to visit definitions
@@ -554,6 +614,11 @@ void MachineLICM::HoistRegion(MachineDomTreeNode *N, bool IsHeader) {
assert(N != 0 && "Null dominator tree node?");
MachineBasicBlock *BB = N->getBlock();
+ // If the header of the loop containing this basic block is a landing pad,
+ // then don't try to hoist instructions out of this loop.
+ const MachineLoop *ML = MLI->getLoopFor(BB);
+ if (ML && ML->getHeader()->isLandingPad()) return;
+
// If this subregion is not in the top level loop at all, exit.
if (!CurLoop->contains(BB)) return;
@@ -571,6 +636,7 @@ void MachineLICM::HoistRegion(MachineDomTreeNode *N, bool IsHeader) {
// Remember livein register pressure.
BackTrace.push_back(RegPressure);
+ SpeculationState = SpeculateUnknown;
for (MachineBasicBlock::iterator
MII = BB->begin(), E = BB->end(); MII != E; ) {
MachineBasicBlock::iterator NextMII = MII; ++NextMII;
@@ -596,6 +662,23 @@ static bool isOperandKill(const MachineOperand &MO, MachineRegisterInfo *MRI) {
return MO.isKill() || MRI->hasOneNonDBGUse(MO.getReg());
}
+/// getRegisterClassIDAndCost - For a given MI, register, and the operand
+/// index, return the ID and cost of its representative register class.
+void
+MachineLICM::getRegisterClassIDAndCost(const MachineInstr *MI,
+ unsigned Reg, unsigned OpIdx,
+ unsigned &RCId, unsigned &RCCost) const {
+ const TargetRegisterClass *RC = MRI->getRegClass(Reg);
+ EVT VT = *RC->vt_begin();
+ if (VT == MVT::untyped) {
+ RCId = RC->getID();
+ RCCost = 1;
+ } else {
+ RCId = TLI->getRepRegClassFor(VT)->getID();
+ RCCost = TLI->getRepRegClassCostFor(VT);
+ }
+}
+
/// InitRegPressure - Find all virtual register references that are liveout of
/// the preheader to initialize the starting "register pressure". Note this
/// does not count live through (livein but not used) registers.
@@ -625,18 +708,17 @@ void MachineLICM::InitRegPressure(MachineBasicBlock *BB) {
continue;
bool isNew = RegSeen.insert(Reg);
- const TargetRegisterClass *RC = MRI->getRegClass(Reg);
- EVT VT = *RC->vt_begin();
- unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
+ unsigned RCId, RCCost;
+ getRegisterClassIDAndCost(MI, Reg, i, RCId, RCCost);
if (MO.isDef())
- RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
+ RegPressure[RCId] += RCCost;
else {
bool isKill = isOperandKill(MO, MRI);
if (isNew && !isKill)
// Haven't seen this, it must be a livein.
- RegPressure[RCId] += TLI->getRepRegClassCostFor(VT);
+ RegPressure[RCId] += RCCost;
else if (!isNew && isKill)
- RegPressure[RCId] -= TLI->getRepRegClassCostFor(VT);
+ RegPressure[RCId] -= RCCost;
}
}
}
@@ -661,11 +743,8 @@ void MachineLICM::UpdateRegPressure(const MachineInstr *MI) {
if (MO.isDef())
Defs.push_back(Reg);
else if (!isNew && isOperandKill(MO, MRI)) {
- const TargetRegisterClass *RC = MRI->getRegClass(Reg);
- EVT VT = *RC->vt_begin();
- unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
- unsigned RCCost = TLI->getRepRegClassCostFor(VT);
-
+ unsigned RCId, RCCost;
+ getRegisterClassIDAndCost(MI, Reg, i, RCId, RCCost);
if (RCCost > RegPressure[RCId])
RegPressure[RCId] = 0;
else
@@ -673,13 +752,13 @@ void MachineLICM::UpdateRegPressure(const MachineInstr *MI) {
}
}
+ unsigned Idx = 0;
while (!Defs.empty()) {
unsigned Reg = Defs.pop_back_val();
- const TargetRegisterClass *RC = MRI->getRegClass(Reg);
- EVT VT = *RC->vt_begin();
- unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
- unsigned RCCost = TLI->getRepRegClassCostFor(VT);
+ unsigned RCId, RCCost;
+ getRegisterClassIDAndCost(MI, Reg, Idx, RCId, RCCost);
RegPressure[RCId] += RCCost;
+ ++Idx;
}
}
@@ -691,7 +770,14 @@ bool MachineLICM::IsLICMCandidate(MachineInstr &I) {
bool DontMoveAcrossStore = true;
if (!I.isSafeToMove(TII, AA, DontMoveAcrossStore))
return false;
-
+
+ // If it is load then check if it is guaranteed to execute by making sure that
+ // it dominates all exiting blocks. If it doesn't, then there is a path out of
+ // the loop which does not execute this load, so we can't hoist it.
+ // Stores and side effects are already checked by isSafeToMove.
+ if (I.getDesc().mayLoad() && !IsGuaranteedToExecute(I.getParent()))
+ return false;
+
return true;
}
@@ -879,10 +965,8 @@ void MachineLICM::UpdateBackTraceRegPressure(const MachineInstr *MI) {
if (!TargetRegisterInfo::isVirtualRegister(Reg))
continue;
- const TargetRegisterClass *RC = MRI->getRegClass(Reg);
- EVT VT = *RC->vt_begin();
- unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
- unsigned RCCost = TLI->getRepRegClassCostFor(VT);
+ unsigned RCId, RCCost;
+ getRegisterClassIDAndCost(MI, Reg, i, RCId, RCCost);
if (MO.isDef()) {
DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
if (CI != Cost.end())
@@ -941,16 +1025,15 @@ bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) {
unsigned Reg = MO.getReg();
if (!TargetRegisterInfo::isVirtualRegister(Reg))
continue;
+
+ unsigned RCId, RCCost;
+ getRegisterClassIDAndCost(&MI, Reg, i, RCId, RCCost);
if (MO.isDef()) {
if (HasHighOperandLatency(MI, i, Reg)) {
++NumHighLatency;
return true;
}
- const TargetRegisterClass *RC = MRI->getRegClass(Reg);
- EVT VT = *RC->vt_begin();
- unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
- unsigned RCCost = TLI->getRepRegClassCostFor(VT);
DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
if (CI != Cost.end())
CI->second += RCCost;
@@ -960,10 +1043,6 @@ bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) {
// Is a virtual register use is a kill, hoisting it out of the loop
// may actually reduce register pressure or be register pressure
// neutral.
- const TargetRegisterClass *RC = MRI->getRegClass(Reg);
- EVT VT = *RC->vt_begin();
- unsigned RCId = TLI->getRepRegClassFor(VT)->getID();
- unsigned RCCost = TLI->getRepRegClassCostFor(VT);
DenseMap<unsigned, int>::iterator CI = Cost.find(RCId);
if (CI != Cost.end())
CI->second -= RCCost;
@@ -979,6 +1058,13 @@ bool MachineLICM::IsProfitableToHoist(MachineInstr &MI) {
return true;
}
+ // Do not "speculate" in high register pressure situation. If an
+ // instruction is not guaranteed to be executed in the loop, it's best to be
+ // conservative.
+ if (AvoidSpeculation &&
+ (!IsGuaranteedToExecute(MI.getParent()) && !MayCSE(&MI)))
+ return false;
+
// High register pressure situation, only hoist if the instruction is going to
// be remat'ed.
if (!TII->isTriviallyReMaterializable(&MI, AA) &&
@@ -1116,6 +1202,20 @@ bool MachineLICM::EliminateCSE(MachineInstr *MI,
return false;
}
+/// MayCSE - Return true if the given instruction will be CSE'd if it's
+/// hoisted out of the loop.
+bool MachineLICM::MayCSE(MachineInstr *MI) {
+ unsigned Opcode = MI->getOpcode();
+ DenseMap<unsigned, std::vector<const MachineInstr*> >::iterator
+ CI = CSEMap.find(Opcode);
+ // Do not CSE implicit_def so ProcessImplicitDefs can properly propagate
+ // the undef property onto uses.
+ if (CI == CSEMap.end() || MI->isImplicitDef())
+ return false;
+
+ return LookForDuplicate(MI, CI->second) != 0;
+}
+
/// Hoist - When an instruction is found to use only loop invariant operands
/// that are safe to hoist, this instruction is called to do the dirty work.
///
generated by cgit v1.2.3 (git 2.25.1) at 2024-09-19 15:44:24 +0000