summaryrefslogtreecommitdiff
path: root/lib/CodeGen/Spiller.cpp
diff options
context:
space:
mode:
authorEvan Cheng <evan.cheng@apple.com>2009-04-21 22:46:52 +0000
committerEvan Cheng <evan.cheng@apple.com>2009-04-21 22:46:52 +0000
commit5b69ebac857104770b1a751bf7a463fda4330a62 (patch)
tree4f04dd784d95779528ce075c994f65335631067c /lib/CodeGen/Spiller.cpp
parent578ccf81e5b94965e3ce8b4eadb6834e39f5f41d (diff)
downloadllvm-5b69ebac857104770b1a751bf7a463fda4330a62.tar.gz
llvm-5b69ebac857104770b1a751bf7a463fda4330a62.tar.bz2
llvm-5b69ebac857104770b1a751bf7a463fda4330a62.tar.xz
It has finally happened. Spiller is now using live interval info.
This fixes a very subtle bug. vr defined by an implicit_def is allowed overlap with any register since it doesn't actually modify anything. However, if it's used as a two-address use, its live range can be extended and it can be spilled. The spiller must take care not to emit a reload for the vn number that's defined by the implicit_def. This is both a correctness and performance issue. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@69743 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib/CodeGen/Spiller.cpp')
-rw-r--r--lib/CodeGen/Spiller.cpp85
1 files changed, 55 insertions, 30 deletions
diff --git a/lib/CodeGen/Spiller.cpp b/lib/CodeGen/Spiller.cpp
index 92bb785de6..26366d6042 100644
--- a/lib/CodeGen/Spiller.cpp
+++ b/lib/CodeGen/Spiller.cpp
@@ -23,6 +23,7 @@ STATISTIC(NumDRM , "Number of re-materializable defs elided");
STATISTIC(NumStores , "Number of stores added");
STATISTIC(NumPSpills , "Number of physical register spills");
STATISTIC(NumOmitted , "Number of reloads omited");
+STATISTIC(NumAvoided , "Number of reloads deemed unnecessary");
STATISTIC(NumCopified, "Number of available reloads turned into copies");
STATISTIC(NumReMats , "Number of re-materialization");
STATISTIC(NumLoads , "Number of loads added");
@@ -50,7 +51,8 @@ SpillerOpt("spiller",
Spiller::~Spiller() {}
-bool SimpleSpiller::runOnMachineFunction(MachineFunction &MF, VirtRegMap &VRM) {
+bool SimpleSpiller::runOnMachineFunction(MachineFunction &MF, VirtRegMap &VRM,
+ LiveIntervals* LIs) {
DOUT << "********** REWRITE MACHINE CODE **********\n";
DOUT << "********** Function: " << MF.getFunction()->getName() << '\n';
const TargetMachine &TM = MF.getTarget();
@@ -521,7 +523,8 @@ unsigned ReuseInfo::GetRegForReload(unsigned PhysReg, MachineInstr *MI,
// Local Spiller Implementation //
// ***************************** //
-bool LocalSpiller::runOnMachineFunction(MachineFunction &MF, VirtRegMap &VRM) {
+bool LocalSpiller::runOnMachineFunction(MachineFunction &MF, VirtRegMap &VRM,
+ LiveIntervals* LIs) {
RegInfo = &MF.getRegInfo();
TRI = MF.getTarget().getRegisterInfo();
TII = MF.getTarget().getInstrInfo();
@@ -555,7 +558,7 @@ bool LocalSpiller::runOnMachineFunction(MachineFunction &MF, VirtRegMap &VRM) {
DFI != E; ++DFI) {
MachineBasicBlock *MBB = *DFI;
if (!EarlyVisited.count(MBB))
- RewriteMBB(*MBB, VRM, Spills, RegKills, KillOps);
+ RewriteMBB(*MBB, VRM, LIs, Spills, RegKills, KillOps);
// If this MBB is the only predecessor of a successor. Keep the
// availability information and visit it next.
@@ -571,7 +574,7 @@ bool LocalSpiller::runOnMachineFunction(MachineFunction &MF, VirtRegMap &VRM) {
MBB = SinglePredSuccs[0];
if (!Visited.count(MBB) && EarlyVisited.insert(MBB)) {
Spills.AddAvailableRegsToLiveIn(*MBB, RegKills, KillOps);
- RewriteMBB(*MBB, VRM, Spills, RegKills, KillOps);
+ RewriteMBB(*MBB, VRM, LIs, Spills, RegKills, KillOps);
}
}
} while (MBB);
@@ -1109,6 +1112,7 @@ void LocalSpiller::TransferDeadness(MachineBasicBlock *MBB, unsigned CurDist,
/// rewriteMBB - Keep track of which spills are available even after the
/// register allocator is done with them. If possible, avid reloading vregs.
void LocalSpiller::RewriteMBB(MachineBasicBlock &MBB, VirtRegMap &VRM,
+ LiveIntervals *LIs,
AvailableSpills &Spills, BitVector &RegKills,
std::vector<MachineOperand*> &KillOps) {
DOUT << "\n**** Local spiller rewriting MBB '"
@@ -1339,6 +1343,22 @@ void LocalSpiller::RewriteMBB(MachineBasicBlock &MBB, VirtRegMap &VRM,
if (!MO.isUse())
continue; // Handle defs in the loop below (handle use&def here though)
+ bool AvoidReload = false;
+ if (LIs->hasInterval(VirtReg)) {
+ LiveInterval &LI = LIs->getInterval(VirtReg);
+ if (!LI.liveAt(LIs->getUseIndex(LI.beginNumber())))
+ // Must be defined by an implicit def. It should not be spilled. Note,
+ // this is for correctness reason. e.g.
+ // 8 %reg1024<def> = IMPLICIT_DEF
+ // 12 %reg1024<def> = INSERT_SUBREG %reg1024<kill>, %reg1025, 2
+ // The live range [12, 14) are not part of the r1024 live interval since
+ // it's defined by an implicit def. It will not conflicts with live
+ // interval of r1025. Now suppose both registers are spilled, you can
+ // easily see a situation where both registers are reloaded before
+ // the INSERT_SUBREG and both target registers that would overlap.
+ AvoidReload = true;
+ }
+
bool DoReMat = VRM.isReMaterialized(VirtReg);
int SSorRMId = DoReMat
? VRM.getReMatId(VirtReg) : VRM.getStackSlot(VirtReg);
@@ -1357,14 +1377,14 @@ void LocalSpiller::RewriteMBB(MachineBasicBlock &MBB, VirtRegMap &VRM,
// = EXTRACT_SUBREG fi#1
// fi#1 is available in EDI, but it cannot be reused because it's not in
// the right register file.
- if (PhysReg &&
+ if (PhysReg && !AvoidReload &&
(SubIdx || MI.getOpcode() == TargetInstrInfo::EXTRACT_SUBREG)) {
const TargetRegisterClass* RC = RegInfo->getRegClass(VirtReg);
if (!RC->contains(PhysReg))
PhysReg = 0;
}
- if (PhysReg) {
+ if (PhysReg && !AvoidReload) {
// This spilled operand might be part of a two-address operand. If this
// is the case, then changing it will necessarily require changing the
// def part of the instruction as well. However, in some cases, we
@@ -1513,34 +1533,39 @@ void LocalSpiller::RewriteMBB(MachineBasicBlock &MBB, VirtRegMap &VRM,
RegInfo->setPhysRegUsed(PhysReg);
ReusedOperands.markClobbered(PhysReg);
- if (DoReMat) {
- ReMaterialize(MBB, MII, PhysReg, VirtReg, TII, TRI, VRM);
- } else {
- const TargetRegisterClass* RC = RegInfo->getRegClass(VirtReg);
- TII->loadRegFromStackSlot(MBB, &MI, PhysReg, SSorRMId, RC);
- MachineInstr *LoadMI = prior(MII);
- VRM.addSpillSlotUse(SSorRMId, LoadMI);
- ++NumLoads;
- }
- // This invalidates PhysReg.
- Spills.ClobberPhysReg(PhysReg);
-
- // Any stores to this stack slot are not dead anymore.
- if (!DoReMat)
- MaybeDeadStores[SSorRMId] = NULL;
- Spills.addAvailable(SSorRMId, PhysReg);
- // Assumes this is the last use. IsKill will be unset if reg is reused
- // unless it's a two-address operand.
- if (!MI.isRegTiedToDefOperand(i) &&
- KilledMIRegs.count(VirtReg) == 0) {
- MI.getOperand(i).setIsKill();
- KilledMIRegs.insert(VirtReg);
+ if (AvoidReload)
+ ++NumAvoided;
+ else {
+ if (DoReMat) {
+ ReMaterialize(MBB, MII, PhysReg, VirtReg, TII, TRI, VRM);
+ } else {
+ const TargetRegisterClass* RC = RegInfo->getRegClass(VirtReg);
+ TII->loadRegFromStackSlot(MBB, &MI, PhysReg, SSorRMId, RC);
+ MachineInstr *LoadMI = prior(MII);
+ VRM.addSpillSlotUse(SSorRMId, LoadMI);
+ ++NumLoads;
+ }
+ // This invalidates PhysReg.
+ Spills.ClobberPhysReg(PhysReg);
+
+ // Any stores to this stack slot are not dead anymore.
+ if (!DoReMat)
+ MaybeDeadStores[SSorRMId] = NULL;
+ Spills.addAvailable(SSorRMId, PhysReg);
+ // Assumes this is the last use. IsKill will be unset if reg is reused
+ // unless it's a two-address operand.
+ if (!MI.isRegTiedToDefOperand(i) &&
+ KilledMIRegs.count(VirtReg) == 0) {
+ MI.getOperand(i).setIsKill();
+ KilledMIRegs.insert(VirtReg);
+ }
+
+ UpdateKills(*prior(MII), RegKills, KillOps, TRI);
+ DOUT << '\t' << *prior(MII);
}
unsigned RReg = SubIdx ? TRI->getSubReg(PhysReg, SubIdx) : PhysReg;
MI.getOperand(i).setReg(RReg);
MI.getOperand(i).setSubReg(0);
- UpdateKills(*prior(MII), RegKills, KillOps, TRI);
- DOUT << '\t' << *prior(MII);
}
// Ok - now we can remove stores that have been confirmed dead.