1 //===- subzero/src/IceTargetLowering.cpp - Basic lowering implementation --===//
3 // The Subzero Code Generator
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the skeleton of the TargetLowering class,
11 // specifically invoking the appropriate lowering method for a given
12 // instruction kind and driving global register allocation. It also
13 // implements the non-deleted instruction iteration in
16 //===----------------------------------------------------------------------===//
18 #include "assembler_arm32.h"
19 #include "assembler_ia32.h"
20 #include "IceCfg.h" // setError()
21 #include "IceCfgNode.h"
22 #include "IceOperand.h"
23 #include "IceRegAlloc.h"
24 #include "IceTargetLowering.h"
25 #include "IceTargetLoweringARM32.h"
26 #include "IceTargetLoweringX8632.h"
30 void LoweringContext::init(CfgNode *N) {
32 End = getNode()->getInsts().end();
37 void LoweringContext::rewind() {
38 Begin = getNode()->getInsts().begin();
44 void LoweringContext::insert(Inst *Inst) {
45 getNode()->getInsts().insert(Next, Inst);
49 void LoweringContext::skipDeleted(InstList::iterator &I) const {
50 while (I != End && I->isDeleted())
54 void LoweringContext::advanceForward(InstList::iterator &I) const {
61 Inst *LoweringContext::getLastInserted() const {
66 TargetLowering *TargetLowering::createLowering(TargetArch Target, Cfg *Func) {
67 #define SUBZERO_TARGET(X) \
68 if (Target == Target_##X) \
69 return Target##X::create(Func);
70 #include "llvm/Config/SZTargets.def"
72 Func->setError("Unsupported target");
76 TargetLowering::TargetLowering(Cfg *Func)
77 : Func(Func), Ctx(Func->getContext()), HasComputedFrame(false),
78 CallsReturnsTwice(false), StackAdjustment(0), NextLabelNumber(0),
79 Context(), SnapshotStackAdjustment(0) {}
81 std::unique_ptr<Assembler> TargetLowering::createAssembler(TargetArch Target,
83 #define SUBZERO_TARGET(X) \
84 if (Target == Target_##X) \
85 return std::unique_ptr<Assembler>(new X::Assembler##X());
86 #include "llvm/Config/SZTargets.def"
88 Func->setError("Unsupported target assembler");
92 void TargetLowering::doAddressOpt() {
93 if (llvm::isa<InstLoad>(*Context.getCur()))
95 else if (llvm::isa<InstStore>(*Context.getCur()))
98 Context.advanceNext();
101 void TargetLowering::doNopInsertion() {
102 Inst *I = Context.getCur();
103 bool ShouldSkip = llvm::isa<InstFakeUse>(I) || llvm::isa<InstFakeDef>(I) ||
104 llvm::isa<InstFakeKill>(I) || I->isRedundantAssign() ||
107 int Probability = Ctx->getFlags().getNopProbabilityAsPercentage();
108 for (int I = 0; I < Ctx->getFlags().getMaxNopsPerInstruction(); ++I) {
109 randomlyInsertNop(Probability / 100.0);
114 // Lowers a single instruction according to the information in
115 // Context, by checking the Context.Cur instruction kind and calling
116 // the appropriate lowering method. The lowering method should insert
117 // target instructions at the Cur.Next insertion point, and should not
118 // delete the Context.Cur instruction or advance Context.Cur.
120 // The lowering method may look ahead in the instruction stream as
121 // desired, and lower additional instructions in conjunction with the
122 // current one, for example fusing a compare and branch. If it does,
123 // it should advance Context.Cur to point to the next non-deleted
124 // instruction to process, and it should delete any additional
125 // instructions it consumes.
126 void TargetLowering::lower() {
127 assert(!Context.atEnd());
128 Inst *Inst = Context.getCur();
129 Inst->deleteIfDead();
130 if (!Inst->isDeleted()) {
131 // Mark the current instruction as deleted before lowering,
132 // otherwise the Dest variable will likely get marked as non-SSA.
133 // See Variable::setDefinition().
135 switch (Inst->getKind()) {
137 lowerAlloca(llvm::cast<InstAlloca>(Inst));
139 case Inst::Arithmetic:
140 lowerArithmetic(llvm::cast<InstArithmetic>(Inst));
143 lowerAssign(llvm::cast<InstAssign>(Inst));
146 lowerBr(llvm::cast<InstBr>(Inst));
149 lowerCall(llvm::cast<InstCall>(Inst));
152 lowerCast(llvm::cast<InstCast>(Inst));
154 case Inst::ExtractElement:
155 lowerExtractElement(llvm::cast<InstExtractElement>(Inst));
158 lowerFcmp(llvm::cast<InstFcmp>(Inst));
161 lowerIcmp(llvm::cast<InstIcmp>(Inst));
163 case Inst::InsertElement:
164 lowerInsertElement(llvm::cast<InstInsertElement>(Inst));
166 case Inst::IntrinsicCall: {
167 InstIntrinsicCall *Call = llvm::cast<InstIntrinsicCall>(Inst);
168 if (Call->getIntrinsicInfo().ReturnsTwice)
169 setCallsReturnsTwice(true);
170 lowerIntrinsicCall(Call);
174 lowerLoad(llvm::cast<InstLoad>(Inst));
177 lowerPhi(llvm::cast<InstPhi>(Inst));
180 lowerRet(llvm::cast<InstRet>(Inst));
183 lowerSelect(llvm::cast<InstSelect>(Inst));
186 lowerStore(llvm::cast<InstStore>(Inst));
189 lowerSwitch(llvm::cast<InstSwitch>(Inst));
191 case Inst::Unreachable:
192 lowerUnreachable(llvm::cast<InstUnreachable>(Inst));
194 case Inst::BundleLock:
195 case Inst::BundleUnlock:
200 // These are all Target instruction types and shouldn't be
201 // encountered at this stage.
202 Func->setError("Can't lower unsupported instruction type");
209 Context.advanceCur();
210 Context.advanceNext();
213 // Drives register allocation, allowing all physical registers (except
214 // perhaps for the frame pointer) to be allocated. This set of
215 // registers could potentially be parameterized if we want to restrict
216 // registers e.g. for performance testing.
217 void TargetLowering::regAlloc(RegAllocKind Kind) {
218 TimerMarker T(TimerStack::TT_regAlloc, Func);
219 LinearScan LinearScan(Func);
220 RegSetMask RegInclude = RegSet_None;
221 RegSetMask RegExclude = RegSet_None;
222 RegInclude |= RegSet_CallerSave;
223 RegInclude |= RegSet_CalleeSave;
224 if (hasFramePointer())
225 RegExclude |= RegSet_FramePointer;
226 LinearScan.init(Kind);
227 llvm::SmallBitVector RegMask = getRegisterSet(RegInclude, RegExclude);
228 LinearScan.scan(RegMask, Ctx->getFlags().shouldRandomizeRegAlloc());
231 void TargetLowering::inferTwoAddress() {
232 // Find two-address non-SSA instructions where Dest==Src0, and set
233 // the DestNonKillable flag to keep liveness analysis consistent.
234 for (auto Inst = Context.getCur(), E = Context.getNext(); Inst != E; ++Inst) {
235 if (Inst->isDeleted())
237 if (Variable *Dest = Inst->getDest()) {
238 // TODO(stichnot): We may need to consider all source
239 // operands, not just the first one, if using 3-address
241 if (Inst->getSrcSize() > 0 && Inst->getSrc(0) == Dest)
242 Inst->setDestNonKillable();
247 void TargetLowering::sortVarsByAlignment(VarList &Dest,
248 const VarList &Source) const {
250 // Instead of std::sort, we could do a bucket sort with log2(alignment)
251 // as the buckets, if performance is an issue.
252 std::sort(Dest.begin(), Dest.end(),
253 [this](const Variable *V1, const Variable *V2) {
254 return typeWidthInBytesOnStack(V1->getType()) >
255 typeWidthInBytesOnStack(V2->getType());
259 void TargetLowering::getVarStackSlotParams(
260 VarList &SortedSpilledVariables, llvm::SmallBitVector &RegsUsed,
261 size_t *GlobalsSize, size_t *SpillAreaSizeBytes,
262 uint32_t *SpillAreaAlignmentBytes, uint32_t *LocalsSlotsAlignmentBytes,
263 std::function<bool(Variable *)> TargetVarHook) {
264 const VariablesMetadata *VMetadata = Func->getVMetadata();
265 llvm::BitVector IsVarReferenced(Func->getNumVariables());
266 for (CfgNode *Node : Func->getNodes()) {
267 for (Inst &Inst : Node->getInsts()) {
268 if (Inst.isDeleted())
270 if (const Variable *Var = Inst.getDest())
271 IsVarReferenced[Var->getIndex()] = true;
272 for (SizeT I = 0; I < Inst.getSrcSize(); ++I) {
273 Operand *Src = Inst.getSrc(I);
274 SizeT NumVars = Src->getNumVars();
275 for (SizeT J = 0; J < NumVars; ++J) {
276 const Variable *Var = Src->getVar(J);
277 IsVarReferenced[Var->getIndex()] = true;
283 // If SimpleCoalescing is false, each variable without a register
284 // gets its own unique stack slot, which leads to large stack
285 // frames. If SimpleCoalescing is true, then each "global" variable
286 // without a register gets its own slot, but "local" variable slots
287 // are reused across basic blocks. E.g., if A and B are local to
288 // block 1 and C is local to block 2, then C may share a slot with A or B.
290 // We cannot coalesce stack slots if this function calls a "returns twice"
291 // function. In that case, basic blocks may be revisited, and variables
292 // local to those basic blocks are actually live until after the
293 // called function returns a second time.
294 const bool SimpleCoalescing = !callsReturnsTwice();
296 std::vector<size_t> LocalsSize(Func->getNumNodes());
297 const VarList &Variables = Func->getVariables();
298 VarList SpilledVariables;
299 for (Variable *Var : Variables) {
301 RegsUsed[Var->getRegNum()] = true;
304 // An argument either does not need a stack slot (if passed in a
305 // register) or already has one (if passed on the stack).
308 // An unreferenced variable doesn't need a stack slot.
309 if (!IsVarReferenced[Var->getIndex()])
311 // Check a target-specific variable (it may end up sharing stack slots)
312 // and not need accounting here.
313 if (TargetVarHook(Var))
315 SpilledVariables.push_back(Var);
318 SortedSpilledVariables.reserve(SpilledVariables.size());
319 sortVarsByAlignment(SortedSpilledVariables, SpilledVariables);
321 for (Variable *Var : SortedSpilledVariables) {
322 size_t Increment = typeWidthInBytesOnStack(Var->getType());
323 // We have sorted by alignment, so the first variable we encounter that
324 // is located in each area determines the max alignment for the area.
325 if (!*SpillAreaAlignmentBytes)
326 *SpillAreaAlignmentBytes = Increment;
327 if (SimpleCoalescing && VMetadata->isTracked(Var)) {
328 if (VMetadata->isMultiBlock(Var)) {
329 *GlobalsSize += Increment;
331 SizeT NodeIndex = VMetadata->getLocalUseNode(Var)->getIndex();
332 LocalsSize[NodeIndex] += Increment;
333 if (LocalsSize[NodeIndex] > *SpillAreaSizeBytes)
334 *SpillAreaSizeBytes = LocalsSize[NodeIndex];
335 if (!*LocalsSlotsAlignmentBytes)
336 *LocalsSlotsAlignmentBytes = Increment;
339 *SpillAreaSizeBytes += Increment;
344 void TargetLowering::alignStackSpillAreas(uint32_t SpillAreaStartOffset,
345 uint32_t SpillAreaAlignmentBytes,
347 uint32_t LocalsSlotsAlignmentBytes,
348 uint32_t *SpillAreaPaddingBytes,
349 uint32_t *LocalsSlotsPaddingBytes) {
350 if (SpillAreaAlignmentBytes) {
351 uint32_t PaddingStart = SpillAreaStartOffset;
352 uint32_t SpillAreaStart =
353 Utils::applyAlignment(PaddingStart, SpillAreaAlignmentBytes);
354 *SpillAreaPaddingBytes = SpillAreaStart - PaddingStart;
357 // If there are separate globals and locals areas, make sure the
358 // locals area is aligned by padding the end of the globals area.
359 if (LocalsSlotsAlignmentBytes) {
360 uint32_t GlobalsAndSubsequentPaddingSize = GlobalsSize;
361 GlobalsAndSubsequentPaddingSize =
362 Utils::applyAlignment(GlobalsSize, LocalsSlotsAlignmentBytes);
363 *LocalsSlotsPaddingBytes = GlobalsAndSubsequentPaddingSize - GlobalsSize;
367 void TargetLowering::assignVarStackSlots(VarList &SortedSpilledVariables,
368 size_t SpillAreaPaddingBytes,
369 size_t SpillAreaSizeBytes,
370 size_t GlobalsAndSubsequentPaddingSize,
371 bool UsesFramePointer) {
372 const VariablesMetadata *VMetadata = Func->getVMetadata();
373 size_t GlobalsSpaceUsed = SpillAreaPaddingBytes;
374 size_t NextStackOffset = SpillAreaPaddingBytes;
375 std::vector<size_t> LocalsSize(Func->getNumNodes());
376 const bool SimpleCoalescing = !callsReturnsTwice();
377 for (Variable *Var : SortedSpilledVariables) {
378 size_t Increment = typeWidthInBytesOnStack(Var->getType());
379 if (SimpleCoalescing && VMetadata->isTracked(Var)) {
380 if (VMetadata->isMultiBlock(Var)) {
381 GlobalsSpaceUsed += Increment;
382 NextStackOffset = GlobalsSpaceUsed;
384 SizeT NodeIndex = VMetadata->getLocalUseNode(Var)->getIndex();
385 LocalsSize[NodeIndex] += Increment;
386 NextStackOffset = SpillAreaPaddingBytes +
387 GlobalsAndSubsequentPaddingSize +
388 LocalsSize[NodeIndex];
391 NextStackOffset += Increment;
393 if (UsesFramePointer)
394 Var->setStackOffset(-NextStackOffset);
396 Var->setStackOffset(SpillAreaSizeBytes - NextStackOffset);
400 InstCall *TargetLowering::makeHelperCall(const IceString &Name, Variable *Dest,
402 const bool HasTailCall = false;
403 Constant *CallTarget = Ctx->getConstantExternSym(Name);
405 InstCall::create(Func, MaxSrcs, Dest, CallTarget, HasTailCall);
409 void TargetLowering::emitWithoutPrefix(const ConstantRelocatable *C) const {
412 Ostream &Str = Ctx->getStrEmit();
413 if (C->getSuppressMangling())
416 Str << Ctx->mangleName(C->getName());
417 RelocOffsetT Offset = C->getOffset();
425 void TargetLowering::emit(const ConstantRelocatable *C) const {
428 Ostream &Str = Ctx->getStrEmit();
429 Str << getConstantPrefix();
430 emitWithoutPrefix(C);
433 std::unique_ptr<TargetDataLowering>
434 TargetDataLowering::createLowering(GlobalContext *Ctx) {
435 TargetArch Target = Ctx->getFlags().getTargetArch();
436 #define SUBZERO_TARGET(X) \
437 if (Target == Target_##X) \
438 return std::unique_ptr<TargetDataLowering>(TargetData##X::create(Ctx));
439 #include "llvm/Config/SZTargets.def"
441 llvm_unreachable("Unsupported target data lowering");
445 TargetDataLowering::~TargetDataLowering() {}
447 } // end of namespace Ice