//===----------------------------------------------------------------------===//
CodeGenSubRegIndex::CodeGenSubRegIndex(Record *R, unsigned Enum)
- : TheDef(R),
- EnumValue(Enum)
-{}
-
-std::string CodeGenSubRegIndex::getNamespace() const {
- if (TheDef->getValue("Namespace"))
- return TheDef->getValueAsString("Namespace");
- else
- return "";
+ : TheDef(R), EnumValue(Enum), LaneMask(0) {
+ Name = R->getName();
+ if (R->getValue("Namespace"))
+ Namespace = R->getValueAsString("Namespace");
}
-const std::string &CodeGenSubRegIndex::getName() const {
- return TheDef->getName();
+CodeGenSubRegIndex::CodeGenSubRegIndex(StringRef N, StringRef Nspace,
+ unsigned Enum)
+ : TheDef(0), Name(N), Namespace(Nspace), EnumValue(Enum), LaneMask(0) {
}
std::string CodeGenSubRegIndex::getQualifiedName() const {
}
void CodeGenSubRegIndex::updateComponents(CodeGenRegBank &RegBank) {
- std::vector<Record*> Comps = TheDef->getValueAsListOfDefs("ComposedOf");
- if (Comps.empty())
+ if (!TheDef)
return;
- if (Comps.size() != 2)
- throw TGError(TheDef->getLoc(), "ComposedOf must have exactly two entries");
- CodeGenSubRegIndex *A = RegBank.getSubRegIdx(Comps[0]);
- CodeGenSubRegIndex *B = RegBank.getSubRegIdx(Comps[1]);
- CodeGenSubRegIndex *X = A->addComposite(B, this);
- if (X)
- throw TGError(TheDef->getLoc(), "Ambiguous ComposedOf entries");
-}
-void CodeGenSubRegIndex::cleanComposites() {
- // Clean out redundant mappings of the form this+X -> X.
- for (CompMap::iterator i = Composed.begin(), e = Composed.end(); i != e;) {
- CompMap::iterator j = i;
- ++i;
- if (j->first == j->second)
- Composed.erase(j);
+ std::vector<Record*> Comps = TheDef->getValueAsListOfDefs("ComposedOf");
+ if (!Comps.empty()) {
+ if (Comps.size() != 2)
+ throw TGError(TheDef->getLoc(), "ComposedOf must have exactly two entries");
+ CodeGenSubRegIndex *A = RegBank.getSubRegIdx(Comps[0]);
+ CodeGenSubRegIndex *B = RegBank.getSubRegIdx(Comps[1]);
+ CodeGenSubRegIndex *X = A->addComposite(B, this);
+ if (X)
+ throw TGError(TheDef->getLoc(), "Ambiguous ComposedOf entries");
+ }
+
+ std::vector<Record*> Parts =
+ TheDef->getValueAsListOfDefs("CoveringSubRegIndices");
+ if (!Parts.empty()) {
+ if (Parts.size() < 2)
+ throw TGError(TheDef->getLoc(),
+ "CoveredBySubRegs must have two or more entries");
+ SmallVector<CodeGenSubRegIndex*, 8> IdxParts;
+ for (unsigned i = 0, e = Parts.size(); i != e; ++i)
+ IdxParts.push_back(RegBank.getSubRegIdx(Parts[i]));
+ RegBank.addConcatSubRegIndex(IdxParts, this);
}
}
+unsigned CodeGenSubRegIndex::computeLaneMask() {
+ // Already computed?
+ if (LaneMask)
+ return LaneMask;
+
+ // Recursion guard, shouldn't be required.
+ LaneMask = ~0u;
+
+ // The lane mask is simply the union of all sub-indices.
+ unsigned M = 0;
+ for (CompMap::iterator I = Composed.begin(), E = Composed.end(); I != E; ++I)
+ M |= I->second->computeLaneMask();
+ assert(M && "Missing lane mask, sub-register cycle?");
+ LaneMask = M;
+ return LaneMask;
+}
+
//===----------------------------------------------------------------------===//
// CodeGenRegister
//===----------------------------------------------------------------------===//
EnumValue(Enum),
CostPerUse(R->getValueAsInt("CostPerUse")),
CoveredBySubRegs(R->getValueAsBit("CoveredBySubRegs")),
- SubRegsComplete(false)
+ NumNativeRegUnits(0),
+ SubRegsComplete(false),
+ SuperRegsComplete(false),
+ TopoSig(~0u)
{}
+void CodeGenRegister::buildObjectGraph(CodeGenRegBank &RegBank) {
+ std::vector<Record*> SRIs = TheDef->getValueAsListOfDefs("SubRegIndices");
+ std::vector<Record*> SRs = TheDef->getValueAsListOfDefs("SubRegs");
+
+ if (SRIs.size() != SRs.size())
+ throw TGError(TheDef->getLoc(),
+ "SubRegs and SubRegIndices must have the same size");
+
+ for (unsigned i = 0, e = SRIs.size(); i != e; ++i) {
+ ExplicitSubRegIndices.push_back(RegBank.getSubRegIdx(SRIs[i]));
+ ExplicitSubRegs.push_back(RegBank.getReg(SRs[i]));
+ }
+
+ // Also compute leading super-registers. Each register has a list of
+ // covered-by-subregs super-registers where it appears as the first explicit
+ // sub-register.
+ //
+ // This is used by computeSecondarySubRegs() to find candidates.
+ if (CoveredBySubRegs && !ExplicitSubRegs.empty())
+ ExplicitSubRegs.front()->LeadingSuperRegs.push_back(this);
+
+ // Add ad hoc alias links. This is a symmetric relationship between two
+ // registers, so build a symmetric graph by adding links in both ends.
+ std::vector<Record*> Aliases = TheDef->getValueAsListOfDefs("Aliases");
+ for (unsigned i = 0, e = Aliases.size(); i != e; ++i) {
+ CodeGenRegister *Reg = RegBank.getReg(Aliases[i]);
+ ExplicitAliases.push_back(Reg);
+ Reg->ExplicitAliases.push_back(this);
+ }
+}
+
const std::string &CodeGenRegister::getName() const {
return TheDef->getName();
}
bool isValid() const { return UnitI != UnitE; }
- unsigned operator* () const { assert(isValid()); return *UnitI; };
+ unsigned operator* () const { assert(isValid()); return *UnitI; }
const CodeGenRegister *getReg() const { assert(isValid()); return *RegI; }
unsigned OldNumUnits = RegUnits.size();
for (SubRegMap::const_iterator I = SubRegs.begin(), E = SubRegs.end();
I != E; ++I) {
- // Strangely a register may have itself as a subreg (self-cycle) e.g. XMM.
- // Only create a unit if no other subregs have units.
CodeGenRegister *SR = I->second;
- if (SR == this) {
- // RegUnits are only empty during computeSubRegs, prior to computing
- // weight.
- if (RegUnits.empty())
- RegUnits.push_back(RegBank.newRegUnit(0));
- continue;
- }
// Merge the subregister's units into this register's RegUnits.
mergeRegUnits(RegUnits, SR->RegUnits);
}
return SubRegs;
SubRegsComplete = true;
- std::vector<Record*> SubList = TheDef->getValueAsListOfDefs("SubRegs");
- std::vector<Record*> IdxList = TheDef->getValueAsListOfDefs("SubRegIndices");
- if (SubList.size() != IdxList.size())
- throw TGError(TheDef->getLoc(), "Register " + getName() +
- " SubRegIndices doesn't match SubRegs");
-
- // First insert the direct subregs and make sure they are fully indexed.
- SmallVector<CodeGenSubRegIndex*, 8> Indices;
- for (unsigned i = 0, e = SubList.size(); i != e; ++i) {
- CodeGenRegister *SR = RegBank.getReg(SubList[i]);
- CodeGenSubRegIndex *Idx = RegBank.getSubRegIdx(IdxList[i]);
- Indices.push_back(Idx);
+ // First insert the explicit subregs and make sure they are fully indexed.
+ for (unsigned i = 0, e = ExplicitSubRegs.size(); i != e; ++i) {
+ CodeGenRegister *SR = ExplicitSubRegs[i];
+ CodeGenSubRegIndex *Idx = ExplicitSubRegIndices[i];
if (!SubRegs.insert(std::make_pair(Idx, SR)).second)
throw TGError(TheDef->getLoc(), "SubRegIndex " + Idx->getName() +
" appears twice in Register " + getName());
// Clone inherited subregs and place duplicate entries in Orphans.
// Here the order is important - earlier subregs take precedence.
- for (unsigned i = 0, e = SubList.size(); i != e; ++i) {
- CodeGenRegister *SR = RegBank.getReg(SubList[i]);
+ for (unsigned i = 0, e = ExplicitSubRegs.size(); i != e; ++i) {
+ CodeGenRegister *SR = ExplicitSubRegs[i];
const SubRegMap &Map = SR->computeSubRegs(RegBank);
- // Add this as a super-register of SR now all sub-registers are in the list.
- // This creates a topological ordering, the exact order depends on the
- // order computeSubRegs is called on all registers.
- SR->SuperRegs.push_back(this);
-
for (SubRegMap::const_iterator SI = Map.begin(), SE = Map.end(); SI != SE;
++SI) {
if (!SubRegs.insert(*SI).second)
Orphans.insert(SI->second);
-
- // Noop sub-register indexes are possible, so avoid duplicates.
- if (SI->second != SR)
- SI->second->SuperRegs.push_back(this);
}
}
// If dsub_2 has ComposedOf = [qsub_1, dsub_0], and this register has a
// qsub_1 subreg, add a dsub_2 subreg. Keep growing Indices and process
// expanded subreg indices recursively.
+ SmallVector<CodeGenSubRegIndex*, 8> Indices = ExplicitSubRegIndices;
for (unsigned i = 0; i != Indices.size(); ++i) {
CodeGenSubRegIndex *Idx = Indices[i];
const CodeGenSubRegIndex::CompMap &Comps = Idx->getComposites();
}
}
- // Process the composites.
- ListInit *Comps = TheDef->getValueAsListInit("CompositeIndices");
- for (unsigned i = 0, e = Comps->size(); i != e; ++i) {
- DagInit *Pat = dynamic_cast<DagInit*>(Comps->getElement(i));
- if (!Pat)
- throw TGError(TheDef->getLoc(), "Invalid dag '" +
- Comps->getElement(i)->getAsString() +
- "' in CompositeIndices");
- DefInit *BaseIdxInit = dynamic_cast<DefInit*>(Pat->getOperator());
- if (!BaseIdxInit || !BaseIdxInit->getDef()->isSubClassOf("SubRegIndex"))
- throw TGError(TheDef->getLoc(), "Invalid SubClassIndex in " +
- Pat->getAsString());
- CodeGenSubRegIndex *BaseIdx = RegBank.getSubRegIdx(BaseIdxInit->getDef());
-
- // Resolve list of subreg indices into R2.
- CodeGenRegister *R2 = this;
- for (DagInit::const_arg_iterator di = Pat->arg_begin(),
- de = Pat->arg_end(); di != de; ++di) {
- DefInit *IdxInit = dynamic_cast<DefInit*>(*di);
- if (!IdxInit || !IdxInit->getDef()->isSubClassOf("SubRegIndex"))
- throw TGError(TheDef->getLoc(), "Invalid SubClassIndex in " +
- Pat->getAsString());
- CodeGenSubRegIndex *Idx = RegBank.getSubRegIdx(IdxInit->getDef());
- const SubRegMap &R2Subs = R2->computeSubRegs(RegBank);
- SubRegMap::const_iterator ni = R2Subs.find(Idx);
- if (ni == R2Subs.end())
- throw TGError(TheDef->getLoc(), "Composite " + Pat->getAsString() +
- " refers to bad index in " + R2->getName());
- R2 = ni->second;
- }
-
- // Insert composite index. Allow overriding inherited indices etc.
- SubRegs[BaseIdx] = R2;
-
- // R2 is no longer an orphan.
- Orphans.erase(R2);
- }
-
// Now Orphans contains the inherited subregisters without a direct index.
// Create inferred indexes for all missing entries.
// Work backwards in the Indices vector in order to compose subregs bottom-up.
// dsub_2 -> ssub_0
//
// We pick the latter composition because another register may have [dsub_0,
- // dsub_1, dsub_2] subregs without neccessarily having a qsub_1 subreg. The
+ // dsub_1, dsub_2] subregs without necessarily having a qsub_1 subreg. The
// dsub_2 -> ssub_0 composition can be shared.
while (!Indices.empty() && !Orphans.empty()) {
CodeGenSubRegIndex *Idx = Indices.pop_back_val();
// Compute the inverse SubReg -> Idx map.
for (SubRegMap::const_iterator SI = SubRegs.begin(), SE = SubRegs.end();
SI != SE; ++SI) {
- // Ignore idempotent sub-register indices.
- if (SI->second == this)
+ if (SI->second == this) {
+ ArrayRef<SMLoc> Loc;
+ if (TheDef)
+ Loc = TheDef->getLoc();
+ throw TGError(Loc, "Register " + getName() +
+ " has itself as a sub-register");
+ }
+ // Ensure that every sub-register has a unique name.
+ DenseMap<const CodeGenRegister*, CodeGenSubRegIndex*>::iterator Ins =
+ SubReg2Idx.insert(std::make_pair(SI->second, SI->first)).first;
+ if (Ins->second == SI->first)
continue;
- // Is is possible to have multiple names for the same sub-register.
- // For example, XMM0 appears as sub_xmm, sub_sd, and sub_ss in YMM0.
- // Eventually, this degeneration should go away, but for now we simply give
- // precedence to the explicit sub-register index over the inherited ones.
- SubReg2Idx.insert(std::make_pair(SI->second, SI->first));
+ // Trouble: Two different names for SI->second.
+ ArrayRef<SMLoc> Loc;
+ if (TheDef)
+ Loc = TheDef->getLoc();
+ throw TGError(Loc, "Sub-register can't have two names: " +
+ SI->second->getName() + " available as " +
+ SI->first->getName() + " and " + Ins->second->getName());
}
- // Initialize RegUnitList. A register with no subregisters creates its own
- // unit. Otherwise, it inherits all its subregister's units. Because
- // getSubRegs is called recursively, this processes the register hierarchy in
- // postorder.
+ // Derive possible names for sub-register concatenations from any explicit
+ // sub-registers. By doing this before computeSecondarySubRegs(), we ensure
+ // that getConcatSubRegIndex() won't invent any concatenated indices that the
+ // user already specified.
+ for (unsigned i = 0, e = ExplicitSubRegs.size(); i != e; ++i) {
+ CodeGenRegister *SR = ExplicitSubRegs[i];
+ if (!SR->CoveredBySubRegs || SR->ExplicitSubRegs.size() <= 1)
+ continue;
+
+ // SR is composed of multiple sub-regs. Find their names in this register.
+ SmallVector<CodeGenSubRegIndex*, 8> Parts;
+ for (unsigned j = 0, e = SR->ExplicitSubRegs.size(); j != e; ++j)
+ Parts.push_back(getSubRegIndex(SR->ExplicitSubRegs[j]));
+
+ // Offer this as an existing spelling for the concatenation of Parts.
+ RegBank.addConcatSubRegIndex(Parts, ExplicitSubRegIndices[i]);
+ }
+
+ // Initialize RegUnitList. Because getSubRegs is called recursively, this
+ // processes the register hierarchy in postorder.
//
- // TODO: We currently assume all register units correspond to a named "leaf"
- // register. We should also unify register units for ad-hoc register
- // aliases. This can be done by iteratively merging units for aliasing
- // registers using a worklist.
- assert(RegUnits.empty() && "Should only initialize RegUnits once");
- if (SubRegs.empty())
- RegUnits.push_back(RegBank.newRegUnit(0));
- else
- inheritRegUnits(RegBank);
+ // Inherit all sub-register units. It is good enough to look at the explicit
+ // sub-registers, the other registers won't contribute any more units.
+ for (unsigned i = 0, e = ExplicitSubRegs.size(); i != e; ++i) {
+ CodeGenRegister *SR = ExplicitSubRegs[i];
+ // Explicit sub-registers are usually disjoint, so this is a good way of
+ // computing the union. We may pick up a few duplicates that will be
+ // eliminated below.
+ unsigned N = RegUnits.size();
+ RegUnits.append(SR->RegUnits.begin(), SR->RegUnits.end());
+ std::inplace_merge(RegUnits.begin(), RegUnits.begin() + N, RegUnits.end());
+ }
+ RegUnits.erase(std::unique(RegUnits.begin(), RegUnits.end()), RegUnits.end());
+
+ // Absent any ad hoc aliasing, we create one register unit per leaf register.
+ // These units correspond to the maximal cliques in the register overlap
+ // graph which is optimal.
+ //
+ // When there is ad hoc aliasing, we simply create one unit per edge in the
+ // undirected ad hoc aliasing graph. Technically, we could do better by
+ // identifying maximal cliques in the ad hoc graph, but cliques larger than 2
+ // are extremely rare anyway (I've never seen one), so we don't bother with
+ // the added complexity.
+ for (unsigned i = 0, e = ExplicitAliases.size(); i != e; ++i) {
+ CodeGenRegister *AR = ExplicitAliases[i];
+ // Only visit each edge once.
+ if (AR->SubRegsComplete)
+ continue;
+ // Create a RegUnit representing this alias edge, and add it to both
+ // registers.
+ unsigned Unit = RegBank.newRegUnit(this, AR);
+ RegUnits.push_back(Unit);
+ AR->RegUnits.push_back(Unit);
+ }
+
+ // Finally, create units for leaf registers without ad hoc aliases. Note that
+ // a leaf register with ad hoc aliases doesn't get its own unit - it isn't
+ // necessary. This means the aliasing leaf registers can share a single unit.
+ if (RegUnits.empty())
+ RegUnits.push_back(RegBank.newRegUnit(this));
+
+ // We have now computed the native register units. More may be adopted later
+ // for balancing purposes.
+ NumNativeRegUnits = RegUnits.size();
+
return SubRegs;
}
+// In a register that is covered by its sub-registers, try to find redundant
+// sub-registers. For example:
+//
+// QQ0 = {Q0, Q1}
+// Q0 = {D0, D1}
+// Q1 = {D2, D3}
+//
+// We can infer that D1_D2 is also a sub-register, even if it wasn't named in
+// the register definition.
+//
+// The explicitly specified registers form a tree. This function discovers
+// sub-register relationships that would force a DAG.
+//
+void CodeGenRegister::computeSecondarySubRegs(CodeGenRegBank &RegBank) {
+ // Collect new sub-registers first, add them later.
+ SmallVector<SubRegMap::value_type, 8> NewSubRegs;
+
+ // Look at the leading super-registers of each sub-register. Those are the
+ // candidates for new sub-registers, assuming they are fully contained in
+ // this register.
+ for (SubRegMap::iterator I = SubRegs.begin(), E = SubRegs.end(); I != E; ++I){
+ const CodeGenRegister *SubReg = I->second;
+ const CodeGenRegister::SuperRegList &Leads = SubReg->LeadingSuperRegs;
+ for (unsigned i = 0, e = Leads.size(); i != e; ++i) {
+ CodeGenRegister *Cand = const_cast<CodeGenRegister*>(Leads[i]);
+ // Already got this sub-register?
+ if (Cand == this || getSubRegIndex(Cand))
+ continue;
+ // Check if each component of Cand is already a sub-register.
+ // We know that the first component is I->second, and is present with the
+ // name I->first.
+ SmallVector<CodeGenSubRegIndex*, 8> Parts(1, I->first);
+ assert(!Cand->ExplicitSubRegs.empty() &&
+ "Super-register has no sub-registers");
+ for (unsigned j = 1, e = Cand->ExplicitSubRegs.size(); j != e; ++j) {
+ if (CodeGenSubRegIndex *Idx = getSubRegIndex(Cand->ExplicitSubRegs[j]))
+ Parts.push_back(Idx);
+ else {
+ // Sub-register doesn't exist.
+ Parts.clear();
+ break;
+ }
+ }
+ // If some Cand sub-register is not part of this register, or if Cand only
+ // has one sub-register, there is nothing to do.
+ if (Parts.size() <= 1)
+ continue;
+
+ // Each part of Cand is a sub-register of this. Make the full Cand also
+ // a sub-register with a concatenated sub-register index.
+ CodeGenSubRegIndex *Concat= RegBank.getConcatSubRegIndex(Parts);
+ NewSubRegs.push_back(std::make_pair(Concat, Cand));
+ }
+ }
+
+ // Now add all the new sub-registers.
+ for (unsigned i = 0, e = NewSubRegs.size(); i != e; ++i) {
+ // Don't add Cand if another sub-register is already using the index.
+ if (!SubRegs.insert(NewSubRegs[i]).second)
+ continue;
+
+ CodeGenSubRegIndex *NewIdx = NewSubRegs[i].first;
+ CodeGenRegister *NewSubReg = NewSubRegs[i].second;
+ SubReg2Idx.insert(std::make_pair(NewSubReg, NewIdx));
+ }
+
+ // Create sub-register index composition maps for the synthesized indices.
+ for (unsigned i = 0, e = NewSubRegs.size(); i != e; ++i) {
+ CodeGenSubRegIndex *NewIdx = NewSubRegs[i].first;
+ CodeGenRegister *NewSubReg = NewSubRegs[i].second;
+ for (SubRegMap::const_iterator SI = NewSubReg->SubRegs.begin(),
+ SE = NewSubReg->SubRegs.end(); SI != SE; ++SI) {
+ CodeGenSubRegIndex *SubIdx = getSubRegIndex(SI->second);
+ if (!SubIdx)
+ throw TGError(TheDef->getLoc(), "No SubRegIndex for " +
+ SI->second->getName() + " in " + getName());
+ NewIdx->addComposite(SI->first, SubIdx);
+ }
+ }
+}
+
+void CodeGenRegister::computeSuperRegs(CodeGenRegBank &RegBank) {
+ // Only visit each register once.
+ if (SuperRegsComplete)
+ return;
+ SuperRegsComplete = true;
+
+ // Make sure all sub-registers have been visited first, so the super-reg
+ // lists will be topologically ordered.
+ for (SubRegMap::const_iterator I = SubRegs.begin(), E = SubRegs.end();
+ I != E; ++I)
+ I->second->computeSuperRegs(RegBank);
+
+ // Now add this as a super-register on all sub-registers.
+ // Also compute the TopoSigId in post-order.
+ TopoSigId Id;
+ for (SubRegMap::const_iterator I = SubRegs.begin(), E = SubRegs.end();
+ I != E; ++I) {
+ // Topological signature computed from SubIdx, TopoId(SubReg).
+ // Loops and idempotent indices have TopoSig = ~0u.
+ Id.push_back(I->first->EnumValue);
+ Id.push_back(I->second->TopoSig);
+
+ // Don't add duplicate entries.
+ if (!I->second->SuperRegs.empty() && I->second->SuperRegs.back() == this)
+ continue;
+ I->second->SuperRegs.push_back(this);
+ }
+ TopoSig = RegBank.getTopoSig(Id);
+}
+
void
CodeGenRegister::addSubRegsPreOrder(SetVector<const CodeGenRegister*> &OSet,
CodeGenRegBank &RegBank) const {
assert(SubRegsComplete && "Must precompute sub-registers");
- std::vector<Record*> Indices = TheDef->getValueAsListOfDefs("SubRegIndices");
- for (unsigned i = 0, e = Indices.size(); i != e; ++i) {
- CodeGenSubRegIndex *Idx = RegBank.getSubRegIdx(Indices[i]);
- CodeGenRegister *SR = SubRegs.find(Idx)->second;
+ for (unsigned i = 0, e = ExplicitSubRegs.size(); i != e; ++i) {
+ CodeGenRegister *SR = ExplicitSubRegs[i];
if (OSet.insert(SR))
SR->addSubRegsPreOrder(OSet, RegBank);
}
+ // Add any secondary sub-registers that weren't part of the explicit tree.
+ for (SubRegMap::const_iterator I = SubRegs.begin(), E = SubRegs.end();
+ I != E; ++I)
+ OSet.insert(I->second);
+}
+
+// Compute overlapping registers.
+//
+// The standard set is all super-registers and all sub-registers, but the
+// target description can add arbitrary overlapping registers via the 'Aliases'
+// field. This complicates things, but we can compute overlapping sets using
+// the following rules:
+//
+// 1. The relation overlap(A, B) is reflexive and symmetric but not transitive.
+//
+// 2. overlap(A, B) implies overlap(A, S) for all S in supers(B).
+//
+// Alternatively:
+//
+// overlap(A, B) if there exists:
+// A' in { A, subregs(A) } and B' in { B, subregs(B) } such that:
+// A' = B' or A' in aliases(B') or B' in aliases(A').
+//
+// Here subregs(A) is the full flattened sub-register set returned by
+// A.getSubRegs() while aliases(A) is simply the special 'Aliases' field in the
+// description of register A.
+//
+// This also implies that registers with a common sub-register are considered
+// overlapping. This can happen when forming register pairs:
+//
+// P0 = (R0, R1)
+// P1 = (R1, R2)
+// P2 = (R2, R3)
+//
+// In this case, we will infer an overlap between P0 and P1 because of the
+// shared sub-register R1. There is no overlap between P0 and P2.
+//
+void CodeGenRegister::computeOverlaps(CodeGenRegister::Set &Overlaps,
+ const CodeGenRegBank &RegBank) const {
+ assert(!RegUnits.empty() && "Compute register units before overlaps.");
+
+ // Register units are assigned such that the overlapping registers are the
+ // super-registers of the root registers of the register units.
+ for (unsigned rui = 0, rue = RegUnits.size(); rui != rue; ++rui) {
+ const RegUnit &RU = RegBank.getRegUnit(RegUnits[rui]);
+ ArrayRef<const CodeGenRegister*> Roots = RU.getRoots();
+ for (unsigned ri = 0, re = Roots.size(); ri != re; ++ri) {
+ const CodeGenRegister *Root = Roots[ri];
+ Overlaps.insert(Root);
+ ArrayRef<const CodeGenRegister*> Supers = Root->getSuperRegs();
+ Overlaps.insert(Supers.begin(), Supers.end());
+ }
+ }
}
// Get the sum of this register's unit weights.
unsigned Weight = 0;
for (RegUnitList::const_iterator I = RegUnits.begin(), E = RegUnits.end();
I != E; ++I) {
- Weight += RegBank.getRegUnitWeight(*I);
+ Weight += RegBank.getRegUnit(*I).Weight;
}
return Weight;
}
//===----------------------------------------------------------------------===//
CodeGenRegisterClass::CodeGenRegisterClass(CodeGenRegBank &RegBank, Record *R)
- : TheDef(R), Name(R->getName()), EnumValue(-1) {
+ : TheDef(R),
+ Name(R->getName()),
+ TopoSigs(RegBank.getNumTopoSigs()),
+ EnumValue(-1) {
// Rename anonymous register classes.
if (R->getName().size() > 9 && R->getName()[9] == '.') {
static unsigned AnonCounter = 0;
// Default allocation order always contains all registers.
for (unsigned i = 0, e = Elements->size(); i != e; ++i) {
Orders[0].push_back((*Elements)[i]);
- Members.insert(RegBank.getReg((*Elements)[i]));
+ const CodeGenRegister *Reg = RegBank.getReg((*Elements)[i]);
+ Members.insert(Reg);
+ TopoSigs.set(Reg->getTopoSig());
}
// Alternative allocation orders may be subsets.
// Create an inferred register class that was missing from the .td files.
// Most properties will be inherited from the closest super-class after the
// class structure has been computed.
-CodeGenRegisterClass::CodeGenRegisterClass(StringRef Name, Key Props)
+CodeGenRegisterClass::CodeGenRegisterClass(CodeGenRegBank &RegBank,
+ StringRef Name, Key Props)
: Members(*Props.Members),
TheDef(0),
Name(Name),
+ TopoSigs(RegBank.getNumTopoSigs()),
EnumValue(-1),
SpillSize(Props.SpillSize),
SpillAlignment(Props.SpillAlignment),
CopyCost(0),
Allocatable(true) {
+ for (CodeGenRegister::Set::iterator I = Members.begin(), E = Members.end();
+ I != E; ++I)
+ TopoSigs.set((*I)->getTopoSig());
}
// Compute inherited propertied for a synthesized register class.
RC.SubClasses |= SubRC->SubClasses;
}
- // Sweep up missed clique members. They will be immediately preceeding RC.
+ // Sweep up missed clique members. They will be immediately preceding RC.
for (unsigned s = rci - 1; s && testSubClass(&RC, RegClasses[s - 1]); --s)
RC.SubClasses.set(s - 1);
}
// CodeGenRegBank
//===----------------------------------------------------------------------===//
-CodeGenRegBank::CodeGenRegBank(RecordKeeper &Records) : Records(Records) {
+CodeGenRegBank::CodeGenRegBank(RecordKeeper &Records) {
// Configure register Sets to understand register classes and tuples.
Sets.addFieldExpander("RegisterClass", "MemberList");
Sets.addFieldExpander("CalleeSavedRegs", "SaveList");
// More indices will be synthesized later.
std::vector<Record*> SRIs = Records.getAllDerivedDefinitions("SubRegIndex");
std::sort(SRIs.begin(), SRIs.end(), LessRecord());
- NumNamedIndices = SRIs.size();
for (unsigned i = 0, e = SRIs.size(); i != e; ++i)
getSubRegIdx(SRIs[i]);
// Build composite maps from ComposedOf fields.
getReg((*TupRegs)[j]);
}
- // Precompute all sub-register maps now all the registers are known.
+ // Now all the registers are known. Build the object graph of explicit
+ // register-register references.
+ for (unsigned i = 0, e = Registers.size(); i != e; ++i)
+ Registers[i]->buildObjectGraph(*this);
+
+ // Compute register name map.
+ for (unsigned i = 0, e = Registers.size(); i != e; ++i)
+ RegistersByName.GetOrCreateValue(
+ Registers[i]->TheDef->getValueAsString("AsmName"),
+ Registers[i]);
+
+ // Precompute all sub-register maps.
// This will create Composite entries for all inferred sub-register indices.
- NumRegUnits = 0;
for (unsigned i = 0, e = Registers.size(); i != e; ++i)
Registers[i]->computeSubRegs(*this);
+ // Infer even more sub-registers by combining leading super-registers.
+ for (unsigned i = 0, e = Registers.size(); i != e; ++i)
+ if (Registers[i]->CoveredBySubRegs)
+ Registers[i]->computeSecondarySubRegs(*this);
+
+ // After the sub-register graph is complete, compute the topologically
+ // ordered SuperRegs list.
+ for (unsigned i = 0, e = Registers.size(); i != e; ++i)
+ Registers[i]->computeSuperRegs(*this);
+
// Native register units are associated with a leaf register. They've all been
// discovered now.
- NumNativeRegUnits = NumRegUnits;
+ NumNativeRegUnits = RegUnits.size();
// Read in register class definitions.
std::vector<Record*> RCs = Records.getAllDerivedDefinitions("RegisterClass");
CodeGenRegisterClass::computeSubClasses(*this);
}
+// Create a synthetic CodeGenSubRegIndex without a corresponding Record.
+CodeGenSubRegIndex*
+CodeGenRegBank::createSubRegIndex(StringRef Name, StringRef Namespace) {
+ CodeGenSubRegIndex *Idx = new CodeGenSubRegIndex(Name, Namespace,
+ SubRegIndices.size() + 1);
+ SubRegIndices.push_back(Idx);
+ return Idx;
+}
+
CodeGenSubRegIndex *CodeGenRegBank::getSubRegIdx(Record *Def) {
CodeGenSubRegIndex *&Idx = Def2SubRegIdx[Def];
if (Idx)
return FoundI->second;
// Sub-class doesn't exist, create a new one.
- CodeGenRegisterClass *NewRC = new CodeGenRegisterClass(Name, K);
+ CodeGenRegisterClass *NewRC = new CodeGenRegisterClass(*this, Name, K);
addToMaps(NewRC);
return NewRC;
}
// None exists, synthesize one.
std::string Name = A->getName() + "_then_" + B->getName();
- Comp = getSubRegIdx(new Record(Name, SMLoc(), Records));
+ Comp = createSubRegIndex(Name, A->getNamespace());
A->addComposite(B, Comp);
return Comp;
}
+CodeGenSubRegIndex *CodeGenRegBank::
+getConcatSubRegIndex(const SmallVector<CodeGenSubRegIndex*, 8> &Parts) {
+ assert(Parts.size() > 1 && "Need two parts to concatenate");
+
+ // Look for an existing entry.
+ CodeGenSubRegIndex *&Idx = ConcatIdx[Parts];
+ if (Idx)
+ return Idx;
+
+ // None exists, synthesize one.
+ std::string Name = Parts.front()->getName();
+ for (unsigned i = 1, e = Parts.size(); i != e; ++i) {
+ Name += '_';
+ Name += Parts[i]->getName();
+ }
+ return Idx = createSubRegIndex(Name, Parts.front()->getNamespace());
+}
+
void CodeGenRegBank::computeComposites() {
+ // Keep track of TopoSigs visited. We only need to visit each TopoSig once,
+ // and many registers will share TopoSigs on regular architectures.
+ BitVector TopoSigs(getNumTopoSigs());
+
for (unsigned i = 0, e = Registers.size(); i != e; ++i) {
CodeGenRegister *Reg1 = Registers[i];
+
+ // Skip identical subreg structures already processed.
+ if (TopoSigs.test(Reg1->getTopoSig()))
+ continue;
+ TopoSigs.set(Reg1->getTopoSig());
+
const CodeGenRegister::SubRegMap &SRM1 = Reg1->getSubRegs();
for (CodeGenRegister::SubRegMap::const_iterator i1 = SRM1.begin(),
e1 = SRM1.end(); i1 != e1; ++i1) {
// Try composing Idx1 with another SubRegIndex.
for (CodeGenRegister::SubRegMap::const_iterator i2 = SRM2.begin(),
e2 = SRM2.end(); i2 != e2; ++i2) {
- CodeGenSubRegIndex *Idx2 = i2->first;
+ CodeGenSubRegIndex *Idx2 = i2->first;
CodeGenRegister *Reg3 = i2->second;
// Ignore identity compositions.
if (Reg2 == Reg3)
continue;
// OK Reg1:IdxPair == Reg3. Find the index with Reg:Idx == Reg3.
- for (CodeGenRegister::SubRegMap::const_iterator i1d = SRM1.begin(),
- e1d = SRM1.end(); i1d != e1d; ++i1d) {
- if (i1d->second == Reg3) {
- // Conflicting composition? Emit a warning but allow it.
- if (CodeGenSubRegIndex *Prev = Idx1->addComposite(Idx2, i1d->first))
- PrintWarning(Twine("SubRegIndex ") + Idx1->getQualifiedName() +
- " and " + Idx2->getQualifiedName() +
- " compose ambiguously as " + Prev->getQualifiedName() +
- " or " + i1d->first->getQualifiedName());
- }
- }
+ CodeGenSubRegIndex *Idx3 = Reg1->getSubRegIndex(Reg3);
+ assert(Idx3 && "Sub-register doesn't have an index");
+
+ // Conflicting composition? Emit a warning but allow it.
+ if (CodeGenSubRegIndex *Prev = Idx1->addComposite(Idx2, Idx3))
+ PrintWarning(Twine("SubRegIndex ") + Idx1->getQualifiedName() +
+ " and " + Idx2->getQualifiedName() +
+ " compose ambiguously as " + Prev->getQualifiedName() +
+ " or " + Idx3->getQualifiedName());
}
}
}
+}
- // We don't care about the difference between (Idx1, Idx2) -> Idx2 and invalid
- // compositions, so remove any mappings of that form.
+// Compute lane masks. This is similar to register units, but at the
+// sub-register index level. Each bit in the lane mask is like a register unit
+// class, and two lane masks will have a bit in common if two sub-register
+// indices overlap in some register.
+//
+// Conservatively share a lane mask bit if two sub-register indices overlap in
+// some registers, but not in others. That shouldn't happen a lot.
+void CodeGenRegBank::computeSubRegIndexLaneMasks() {
+ // First assign individual bits to all the leaf indices.
+ unsigned Bit = 0;
+ for (unsigned i = 0, e = SubRegIndices.size(); i != e; ++i) {
+ CodeGenSubRegIndex *Idx = SubRegIndices[i];
+ if (Idx->getComposites().empty()) {
+ Idx->LaneMask = 1u << Bit;
+ // Share bit 31 in the unlikely case there are more than 32 leafs.
+ if (Bit < 31) ++Bit;
+ } else {
+ Idx->LaneMask = 0;
+ }
+ }
+
+ // FIXME: What if ad-hoc aliasing introduces overlaps that aren't represented
+ // by the sub-register graph? This doesn't occur in any known targets.
+
+ // Inherit lanes from composites.
for (unsigned i = 0, e = SubRegIndices.size(); i != e; ++i)
- SubRegIndices[i]->cleanComposites();
+ SubRegIndices[i]->computeLaneMask();
}
namespace {
Reg = UnitI.getReg();
Weight = 0;
}
- unsigned UWeight = RegBank.getRegUnitWeight(*UnitI);
+ unsigned UWeight = RegBank.getRegUnit(*UnitI).Weight;
if (!UWeight) {
UWeight = 1;
RegBank.increaseRegUnitWeight(*UnitI, UWeight);
static bool normalizeWeight(CodeGenRegister *Reg,
std::vector<UberRegSet> &UberSets,
std::vector<UberRegSet*> &RegSets,
+ std::set<unsigned> &NormalRegs,
CodeGenRegister::RegUnitList &NormalUnits,
CodeGenRegBank &RegBank) {
bool Changed = false;
+ if (!NormalRegs.insert(Reg->EnumValue).second)
+ return Changed;
+
const CodeGenRegister::SubRegMap &SRM = Reg->getSubRegs();
for (CodeGenRegister::SubRegMap::const_iterator SRI = SRM.begin(),
SRE = SRM.end(); SRI != SRE; ++SRI) {
if (SRI->second == Reg)
continue; // self-cycles happen
- Changed |=
- normalizeWeight(SRI->second, UberSets, RegSets, NormalUnits, RegBank);
+ Changed |= normalizeWeight(SRI->second, UberSets, RegSets,
+ NormalRegs, NormalUnits, RegBank);
}
// Postorder register normalization.
// The goal is that two registers in the same class will have the same weight,
// where each register's weight is defined as sum of its units' weights.
void CodeGenRegBank::computeRegUnitWeights() {
- assert(RegUnitWeights.empty() && "Only initialize RegUnitWeights once");
-
- // Only allocatable units will be initialized to nonzero weight.
- RegUnitWeights.resize(NumRegUnits);
-
std::vector<UberRegSet> UberSets;
std::vector<UberRegSet*> RegSets(Registers.size());
computeUberSets(UberSets, RegSets, *this);
Changed = false;
for (unsigned i = 0, e = Registers.size(); i != e; ++i) {
CodeGenRegister::RegUnitList NormalUnits;
- Changed |=
- normalizeWeight(Registers[i], UberSets, RegSets, NormalUnits, *this);
+ std::set<unsigned> NormalRegs;
+ Changed |= normalizeWeight(Registers[i], UberSets, RegSets,
+ NormalRegs, NormalUnits, *this);
}
}
}
}
}
-// Compute sets of overlapping registers.
-//
-// The standard set is all super-registers and all sub-registers, but the
-// target description can add arbitrary overlapping registers via the 'Aliases'
-// field. This complicates things, but we can compute overlapping sets using
-// the following rules:
-//
-// 1. The relation overlap(A, B) is reflexive and symmetric but not transitive.
-//
-// 2. overlap(A, B) implies overlap(A, S) for all S in supers(B).
-//
-// Alternatively:
-//
-// overlap(A, B) iff there exists:
-// A' in { A, subregs(A) } and B' in { B, subregs(B) } such that:
-// A' = B' or A' in aliases(B') or B' in aliases(A').
-//
-// Here subregs(A) is the full flattened sub-register set returned by
-// A.getSubRegs() while aliases(A) is simply the special 'Aliases' field in the
-// description of register A.
-//
-// This also implies that registers with a common sub-register are considered
-// overlapping. This can happen when forming register pairs:
-//
-// P0 = (R0, R1)
-// P1 = (R1, R2)
-// P2 = (R2, R3)
-//
-// In this case, we will infer an overlap between P0 and P1 because of the
-// shared sub-register R1. There is no overlap between P0 and P2.
-//
-void CodeGenRegBank::
-computeOverlaps(std::map<const CodeGenRegister*, CodeGenRegister::Set> &Map) {
- assert(Map.empty());
-
- // Collect overlaps that don't follow from rule 2.
- for (unsigned i = 0, e = Registers.size(); i != e; ++i) {
- CodeGenRegister *Reg = Registers[i];
- CodeGenRegister::Set &Overlaps = Map[Reg];
-
- // Reg overlaps itself.
- Overlaps.insert(Reg);
-
- // All super-registers overlap.
- const CodeGenRegister::SuperRegList &Supers = Reg->getSuperRegs();
- Overlaps.insert(Supers.begin(), Supers.end());
-
- // Form symmetrical relations from the special Aliases[] lists.
- std::vector<Record*> RegList = Reg->TheDef->getValueAsListOfDefs("Aliases");
- for (unsigned i2 = 0, e2 = RegList.size(); i2 != e2; ++i2) {
- CodeGenRegister *Reg2 = getReg(RegList[i2]);
- CodeGenRegister::Set &Overlaps2 = Map[Reg2];
- const CodeGenRegister::SuperRegList &Supers2 = Reg2->getSuperRegs();
- // Reg overlaps Reg2 which implies it overlaps supers(Reg2).
- Overlaps.insert(Reg2);
- Overlaps.insert(Supers2.begin(), Supers2.end());
- Overlaps2.insert(Reg);
- Overlaps2.insert(Supers.begin(), Supers.end());
- }
- }
-
- // Apply rule 2. and inherit all sub-register overlaps.
- for (unsigned i = 0, e = Registers.size(); i != e; ++i) {
- CodeGenRegister *Reg = Registers[i];
- CodeGenRegister::Set &Overlaps = Map[Reg];
- const CodeGenRegister::SubRegMap &SRM = Reg->getSubRegs();
- for (CodeGenRegister::SubRegMap::const_iterator i2 = SRM.begin(),
- e2 = SRM.end(); i2 != e2; ++i2) {
- CodeGenRegister::Set &Overlaps2 = Map[i2->second];
- Overlaps.insert(Overlaps2.begin(), Overlaps2.end());
- }
- }
-}
-
void CodeGenRegBank::computeDerivedInfo() {
computeComposites();
+ computeSubRegIndexLaneMasks();
// Compute a weight for each register unit created during getSubRegs.
// This may create adopted register units (with unit # >= NumNativeRegUnits).
unsigned FirstSubRegRC) {
SmallVector<std::pair<const CodeGenRegister*,
const CodeGenRegister*>, 16> SSPairs;
+ BitVector TopoSigs(getNumTopoSigs());
// Iterate in SubRegIndex numerical order to visit synthetic indices last.
for (unsigned sri = 0, sre = SubRegIndices.size(); sri != sre; ++sri) {
// Build list of (Super, Sub) pairs for this SubIdx.
SSPairs.clear();
+ TopoSigs.reset();
for (CodeGenRegister::Set::const_iterator RI = RC->getMembers().begin(),
RE = RC->getMembers().end(); RI != RE; ++RI) {
const CodeGenRegister *Super = *RI;
const CodeGenRegister *Sub = Super->getSubRegs().find(SubIdx)->second;
assert(Sub && "Missing sub-register");
SSPairs.push_back(std::make_pair(Super, Sub));
+ TopoSigs.set(Sub->getTopoSig());
}
// Iterate over sub-register class candidates. Ignore classes created by
for (unsigned rci = FirstSubRegRC, rce = RegClasses.size(); rci != rce;
++rci) {
CodeGenRegisterClass *SubRC = RegClasses[rci];
+ // Topological shortcut: SubRC members have the wrong shape.
+ if (!TopoSigs.anyCommon(SubRC->getTopoSigs()))
+ continue;
// Compute the subset of RC that maps into SubRC.
CodeGenRegister::Set SubSet;
for (unsigned i = 0, e = SSPairs.size(); i != e; ++i)
projects / oota-llvm.git / blobdiff