1 //===- subzero/src/IceTargetLowering.h - Lowering interface -----*- C++ -*-===//
2 //
3 //                        The Subzero Code Generator
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 ///
10 /// \file
11 /// \brief Declares the TargetLowering, LoweringContext, and TargetDataLowering
12 /// classes.
13 ///
14 /// TargetLowering is an abstract class used to drive the translation/lowering
15 /// process. LoweringContext maintains a context for lowering each instruction,
16 /// offering conveniences such as iterating over non-deleted instructions.
17 /// TargetDataLowering is an abstract class used to drive the lowering/emission
18 /// of global initializers, external global declarations, and internal constant
19 /// pools.
20 ///
21 //===----------------------------------------------------------------------===//
22 
23 #ifndef SUBZERO_SRC_ICETARGETLOWERING_H
24 #define SUBZERO_SRC_ICETARGETLOWERING_H
25 
26 #include "IceBitVector.h"
27 #include "IceCfgNode.h"
28 #include "IceDefs.h"
29 #include "IceInst.h" // for the names of the Inst subtypes
30 #include "IceOperand.h"
31 #include "IceRegAlloc.h"
32 #include "IceTypes.h"
33 
34 #include <utility>
35 
36 namespace Ice {
37 
38 // UnimplementedError is defined as a macro so that we can get actual line
39 // numbers.
40 #define UnimplementedError(Flags)                                              \
41   do {                                                                         \
42     if (!static_cast<const ClFlags &>(Flags).getSkipUnimplemented()) {         \
43       /* Use llvm_unreachable instead of report_fatal_error, which gives       \
44          better stack traces. */                                               \
45       llvm_unreachable("Not yet implemented");                                 \
46       abort();                                                                 \
47     }                                                                          \
48   } while (0)
49 
50 // UnimplementedLoweringError is similar in style to UnimplementedError.  Given
51 // a TargetLowering object pointer and an Inst pointer, it adds appropriate
52 // FakeDef and FakeUse instructions to try maintain liveness consistency.
53 #define UnimplementedLoweringError(Target, Instr)                              \
54   do {                                                                         \
55     if (getFlags().getSkipUnimplemented()) {                                   \
56       (Target)->addFakeDefUses(Instr);                                         \
57     } else {                                                                   \
58       /* Use llvm_unreachable instead of report_fatal_error, which gives       \
59          better stack traces. */                                               \
60       llvm_unreachable(                                                        \
61           (std::string("Not yet implemented: ") + Instr->getInstName())        \
62               .c_str());                                                       \
63       abort();                                                                 \
64     }                                                                          \
65   } while (0)
66 
67 /// LoweringContext makes it easy to iterate through non-deleted instructions in
68 /// a node, and insert new (lowered) instructions at the current point. Along
69 /// with the instruction list container and associated iterators, it holds the
70 /// current node, which is needed when inserting new instructions in order to
71 /// track whether variables are used as single-block or multi-block.
72 class LoweringContext {
73   LoweringContext(const LoweringContext &) = delete;
74   LoweringContext &operator=(const LoweringContext &) = delete;
75 
76 public:
77   LoweringContext() = default;
78   ~LoweringContext() = default;
79   void init(CfgNode *Node);
getNextInst()80   Inst *getNextInst() const {
81     if (Next == End)
82       return nullptr;
83     return iteratorToInst(Next);
84   }
getNextInst(InstList::iterator & Iter)85   Inst *getNextInst(InstList::iterator &Iter) const {
86     advanceForward(Iter);
87     if (Iter == End)
88       return nullptr;
89     return iteratorToInst(Iter);
90   }
getNode()91   CfgNode *getNode() const { return Node; }
atEnd()92   bool atEnd() const { return Cur == End; }
getCur()93   InstList::iterator getCur() const { return Cur; }
getNext()94   InstList::iterator getNext() const { return Next; }
getEnd()95   InstList::iterator getEnd() const { return End; }
96   void insert(Inst *Instr);
insert(Args &&...A)97   template <typename Inst, typename... Args> Inst *insert(Args &&... A) {
98     auto *New = Inst::create(Node->getCfg(), std::forward<Args>(A)...);
99     insert(New);
100     return New;
101   }
102   Inst *getLastInserted() const;
advanceCur()103   void advanceCur() { Cur = Next; }
advanceNext()104   void advanceNext() { advanceForward(Next); }
setCur(InstList::iterator C)105   void setCur(InstList::iterator C) { Cur = C; }
setNext(InstList::iterator N)106   void setNext(InstList::iterator N) { Next = N; }
107   void rewind();
setInsertPoint(const InstList::iterator & Position)108   void setInsertPoint(const InstList::iterator &Position) { Next = Position; }
109   void availabilityReset();
110   void availabilityUpdate();
111   Variable *availabilityGet(Operand *Src) const;
112 
113 private:
114   /// Node is the argument to Inst::updateVars().
115   CfgNode *Node = nullptr;
116   Inst *LastInserted = nullptr;
117   /// Cur points to the current instruction being considered. It is guaranteed
118   /// to point to a non-deleted instruction, or to be End.
119   InstList::iterator Cur;
120   /// Next doubles as a pointer to the next valid instruction (if any), and the
121   /// new-instruction insertion point. It is also updated for the caller in case
122   /// the lowering consumes more than one high-level instruction. It is
123   /// guaranteed to point to a non-deleted instruction after Cur, or to be End.
124   // TODO: Consider separating the notion of "next valid instruction" and "new
125   // instruction insertion point", to avoid confusion when previously-deleted
126   // instructions come between the two points.
127   InstList::iterator Next;
128   /// Begin is a copy of Insts.begin(), used if iterators are moved backward.
129   InstList::iterator Begin;
130   /// End is a copy of Insts.end(), used if Next needs to be advanced.
131   InstList::iterator End;
132   /// LastDest and LastSrc capture the parameters of the last "Dest=Src" simple
133   /// assignment inserted (provided Src is a variable).  This is used for simple
134   /// availability analysis.
135   Variable *LastDest = nullptr;
136   Variable *LastSrc = nullptr;
137 
138   void skipDeleted(InstList::iterator &I) const;
139   void advanceForward(InstList::iterator &I) const;
140 };
141 
142 /// A helper class to advance the LoweringContext at each loop iteration.
143 class PostIncrLoweringContext {
144   PostIncrLoweringContext() = delete;
145   PostIncrLoweringContext(const PostIncrLoweringContext &) = delete;
146   PostIncrLoweringContext &operator=(const PostIncrLoweringContext &) = delete;
147 
148 public:
PostIncrLoweringContext(LoweringContext & Context)149   explicit PostIncrLoweringContext(LoweringContext &Context)
150       : Context(Context) {}
~PostIncrLoweringContext()151   ~PostIncrLoweringContext() {
152     Context.advanceCur();
153     Context.advanceNext();
154   }
155 
156 private:
157   LoweringContext &Context;
158 };
159 
160 /// TargetLowering is the base class for all backends in Subzero. In addition to
161 /// implementing the abstract methods in this class, each concrete target must
162 /// also implement a named constructor in its own namespace. For instance, for
163 /// X8632 we have:
164 ///
165 ///  namespace X8632 {
166 ///    void createTargetLowering(Cfg *Func);
167 ///  }
168 class TargetLowering {
169   TargetLowering() = delete;
170   TargetLowering(const TargetLowering &) = delete;
171   TargetLowering &operator=(const TargetLowering &) = delete;
172 
173 public:
174   static void staticInit(GlobalContext *Ctx);
175   // Each target must define a public static method:
176   //   static void staticInit(GlobalContext *Ctx);
177   static bool shouldBePooled(const class Constant *C);
178   static Type getPointerType();
179 
180   static std::unique_ptr<TargetLowering> createLowering(TargetArch Target,
181                                                         Cfg *Func);
182 
183   virtual std::unique_ptr<Assembler> createAssembler() const = 0;
184 
translate()185   void translate() {
186     switch (Func->getOptLevel()) {
187     case Opt_m1:
188       translateOm1();
189       break;
190     case Opt_0:
191       translateO0();
192       break;
193     case Opt_1:
194       translateO1();
195       break;
196     case Opt_2:
197       translateO2();
198       break;
199     }
200   }
translateOm1()201   virtual void translateOm1() {
202     Func->setError("Target doesn't specify Om1 lowering steps.");
203   }
translateO0()204   virtual void translateO0() {
205     Func->setError("Target doesn't specify O0 lowering steps.");
206   }
translateO1()207   virtual void translateO1() {
208     Func->setError("Target doesn't specify O1 lowering steps.");
209   }
translateO2()210   virtual void translateO2() {
211     Func->setError("Target doesn't specify O2 lowering steps.");
212   }
213 
214   /// Generates calls to intrinsics for operations the Target can't handle.
215   void genTargetHelperCalls();
216   /// Tries to do address mode optimization on a single instruction.
217   void doAddressOpt();
218   /// Lowers a single non-Phi instruction.
219   void lower();
220   /// Inserts and lowers a single high-level instruction at a specific insertion
221   /// point.
222   void lowerInst(CfgNode *Node, InstList::iterator Next, InstHighLevel *Instr);
223   /// Does preliminary lowering of the set of Phi instructions in the current
224   /// node. The main intention is to do what's needed to keep the unlowered Phi
225   /// instructions consistent with the lowered non-Phi instructions, e.g. to
226   /// lower 64-bit operands on a 32-bit target.
prelowerPhis()227   virtual void prelowerPhis() {}
228   /// Tries to do branch optimization on a single instruction. Returns true if
229   /// some optimization was done.
doBranchOpt(Inst *,const CfgNode *)230   virtual bool doBranchOpt(Inst * /*I*/, const CfgNode * /*NextNode*/) {
231     return false;
232   }
233 
234   virtual SizeT getNumRegisters() const = 0;
235   /// Returns a variable pre-colored to the specified physical register. This is
236   /// generally used to get very direct access to the register such as in the
237   /// prolog or epilog or for marking scratch registers as killed by a call. If
238   /// a Type is not provided, a target-specific default type is used.
239   virtual Variable *getPhysicalRegister(RegNumT RegNum,
240                                         Type Ty = IceType_void) = 0;
241   /// Returns a printable name for the register.
242   virtual const char *getRegName(RegNumT RegNum, Type Ty) const = 0;
243 
hasFramePointer()244   virtual bool hasFramePointer() const { return false; }
245   virtual void setHasFramePointer() = 0;
246   virtual RegNumT getStackReg() const = 0;
247   virtual RegNumT getFrameReg() const = 0;
248   virtual RegNumT getFrameOrStackReg() const = 0;
249   virtual size_t typeWidthInBytesOnStack(Type Ty) const = 0;
250   virtual uint32_t getStackAlignment() const = 0;
needsStackPointerAlignment()251   virtual bool needsStackPointerAlignment() const { return false; }
252   virtual void reserveFixedAllocaArea(size_t Size, size_t Align) = 0;
253   virtual int32_t getFrameFixedAllocaOffset() const = 0;
maxOutArgsSizeBytes()254   virtual uint32_t maxOutArgsSizeBytes() const { return 0; }
255   // Addressing relative to frame pointer differs in MIPS compared to X86/ARM
256   // since MIPS decrements its stack pointer prior to saving it in the frame
257   // pointer register.
getFramePointerOffset(uint32_t CurrentOffset,uint32_t Size)258   virtual uint32_t getFramePointerOffset(uint32_t CurrentOffset,
259                                          uint32_t Size) const {
260     return -(CurrentOffset + Size);
261   }
262   /// Return whether a 64-bit Variable should be split into a Variable64On32.
263   virtual bool shouldSplitToVariable64On32(Type Ty) const = 0;
264 
265   /// Return whether a Vector Variable should be split into a VariableVecOn32.
shouldSplitToVariableVecOn32(Type Ty)266   virtual bool shouldSplitToVariableVecOn32(Type Ty) const {
267     (void)Ty;
268     return false;
269   }
270 
hasComputedFrame()271   bool hasComputedFrame() const { return HasComputedFrame; }
272   /// Returns true if this function calls a function that has the "returns
273   /// twice" attribute.
callsReturnsTwice()274   bool callsReturnsTwice() const { return CallsReturnsTwice; }
setCallsReturnsTwice(bool RetTwice)275   void setCallsReturnsTwice(bool RetTwice) { CallsReturnsTwice = RetTwice; }
makeNextLabelNumber()276   SizeT makeNextLabelNumber() { return NextLabelNumber++; }
makeNextJumpTableNumber()277   SizeT makeNextJumpTableNumber() { return NextJumpTableNumber++; }
getContext()278   LoweringContext &getContext() { return Context; }
getFunc()279   Cfg *getFunc() const { return Func; }
getGlobalContext()280   GlobalContext *getGlobalContext() const { return Ctx; }
281 
282   enum RegSet {
283     RegSet_None = 0,
284     RegSet_CallerSave = 1 << 0,
285     RegSet_CalleeSave = 1 << 1,
286     RegSet_StackPointer = 1 << 2,
287     RegSet_FramePointer = 1 << 3,
288     RegSet_All = ~RegSet_None
289   };
290   using RegSetMask = uint32_t;
291 
292   virtual SmallBitVector getRegisterSet(RegSetMask Include,
293                                         RegSetMask Exclude) const = 0;
294   /// Get the set of physical registers available for the specified Variable's
295   /// register class, applying register restrictions from the command line.
296   virtual const SmallBitVector &
297   getRegistersForVariable(const Variable *Var) const = 0;
298   /// Get the set of *all* physical registers available for the specified
299   /// Variable's register class, *not* applying register restrictions from the
300   /// command line.
301   virtual const SmallBitVector &
302   getAllRegistersForVariable(const Variable *Var) const = 0;
303   virtual const SmallBitVector &getAliasesForRegister(RegNumT) const = 0;
304 
305   void regAlloc(RegAllocKind Kind);
306   void postRegallocSplitting(const SmallBitVector &RegMask);
307 
308   /// Get the minimum number of clusters required for a jump table to be
309   /// considered.
310   virtual SizeT getMinJumpTableSize() const = 0;
311   virtual void emitJumpTable(const Cfg *Func,
312                              const InstJumpTable *JumpTable) const = 0;
313 
314   virtual void emitVariable(const Variable *Var) const = 0;
315 
316   void emitWithoutPrefix(const ConstantRelocatable *CR,
317                          const char *Suffix = "") const;
318 
319   virtual void emit(const ConstantInteger32 *C) const = 0;
320   virtual void emit(const ConstantInteger64 *C) const = 0;
321   virtual void emit(const ConstantFloat *C) const = 0;
322   virtual void emit(const ConstantDouble *C) const = 0;
323   virtual void emit(const ConstantUndef *C) const = 0;
324   virtual void emit(const ConstantRelocatable *CR) const = 0;
325 
326   /// Performs target-specific argument lowering.
327   virtual void lowerArguments() = 0;
328 
initNodeForLowering(CfgNode *)329   virtual void initNodeForLowering(CfgNode *) {}
330   virtual void addProlog(CfgNode *Node) = 0;
331   virtual void addEpilog(CfgNode *Node) = 0;
332 
333   /// Create a properly-typed "mov" instruction.  This is primarily for local
334   /// variable splitting.
createLoweredMove(Variable * Dest,Variable * SrcVar)335   virtual Inst *createLoweredMove(Variable *Dest, Variable *SrcVar) {
336     // TODO(stichnot): make pure virtual by implementing for all targets
337     (void)Dest;
338     (void)SrcVar;
339     llvm::report_fatal_error("createLoweredMove() unimplemented");
340     return nullptr;
341   }
342 
343   virtual ~TargetLowering() = default;
344 
345 private:
346   // This control variable is used by AutoBundle (RAII-style bundle
347   // locking/unlocking) to prevent nested bundles.
348   bool AutoBundling = false;
349 
350   /// This indicates whether we are in the genTargetHelperCalls phase, and
351   /// therefore can do things like scalarization.
352   bool GeneratingTargetHelpers = false;
353 
354   // _bundle_lock(), and _bundle_unlock(), were made private to force subtargets
355   // to use the AutoBundle helper.
356   void
357   _bundle_lock(InstBundleLock::Option BundleOption = InstBundleLock::Opt_None) {
358     Context.insert<InstBundleLock>(BundleOption);
359   }
_bundle_unlock()360   void _bundle_unlock() { Context.insert<InstBundleUnlock>(); }
361 
362 protected:
363   /// AutoBundle provides RIAA-style bundling. Sub-targets are expected to use
364   /// it when emitting NaCl Bundles to ensure proper bundle_unlocking, and
365   /// prevent nested bundles.
366   ///
367   /// AutoBundle objects will emit a _bundle_lock during construction (but only
368   /// if sandboxed code generation was requested), and a bundle_unlock() during
369   /// destruction. By carefully scoping objects of this type, Subtargets can
370   /// ensure proper bundle emission.
371   class AutoBundle {
372     AutoBundle() = delete;
373     AutoBundle(const AutoBundle &) = delete;
374     AutoBundle &operator=(const AutoBundle &) = delete;
375 
376   public:
377     explicit AutoBundle(TargetLowering *Target, InstBundleLock::Option Option =
378                                                     InstBundleLock::Opt_None);
379     ~AutoBundle();
380 
381   private:
382     TargetLowering *const Target;
383     const bool NeedSandboxing;
384   };
385 
386   explicit TargetLowering(Cfg *Func);
387   // Applies command line filters to TypeToRegisterSet array.
388   static void filterTypeToRegisterSet(
389       GlobalContext *Ctx, int32_t NumRegs, SmallBitVector TypeToRegisterSet[],
390       size_t TypeToRegisterSetSize,
391       std::function<std::string(RegNumT)> getRegName,
392       std::function<const char *(RegClass)> getRegClassName);
393   virtual void lowerAlloca(const InstAlloca *Instr) = 0;
394   virtual void lowerArithmetic(const InstArithmetic *Instr) = 0;
395   virtual void lowerAssign(const InstAssign *Instr) = 0;
396   virtual void lowerBr(const InstBr *Instr) = 0;
397   virtual void lowerBreakpoint(const InstBreakpoint *Instr) = 0;
398   virtual void lowerCall(const InstCall *Instr) = 0;
399   virtual void lowerCast(const InstCast *Instr) = 0;
400   virtual void lowerFcmp(const InstFcmp *Instr) = 0;
401   virtual void lowerExtractElement(const InstExtractElement *Instr) = 0;
402   virtual void lowerIcmp(const InstIcmp *Instr) = 0;
403   virtual void lowerInsertElement(const InstInsertElement *Instr) = 0;
404   virtual void lowerIntrinsic(const InstIntrinsic *Instr) = 0;
405   virtual void lowerLoad(const InstLoad *Instr) = 0;
406   virtual void lowerPhi(const InstPhi *Instr) = 0;
407   virtual void lowerRet(const InstRet *Instr) = 0;
408   virtual void lowerSelect(const InstSelect *Instr) = 0;
409   virtual void lowerShuffleVector(const InstShuffleVector *Instr) = 0;
410   virtual void lowerStore(const InstStore *Instr) = 0;
411   virtual void lowerSwitch(const InstSwitch *Instr) = 0;
412   virtual void lowerUnreachable(const InstUnreachable *Instr) = 0;
413   virtual void lowerOther(const Inst *Instr);
414 
415   virtual void genTargetHelperCallFor(Inst *Instr) = 0;
416   virtual uint32_t getCallStackArgumentsSizeBytes(const InstCall *Instr) = 0;
417 
418   /// Opportunity to modify other instructions to help Address Optimization
doAddressOptOther()419   virtual void doAddressOptOther() {}
doAddressOptLoad()420   virtual void doAddressOptLoad() {}
doAddressOptStore()421   virtual void doAddressOptStore() {}
doAddressOptLoadSubVector()422   virtual void doAddressOptLoadSubVector() {}
doAddressOptStoreSubVector()423   virtual void doAddressOptStoreSubVector() {}
doMockBoundsCheck(Operand *)424   virtual void doMockBoundsCheck(Operand *) {}
425   /// This gives the target an opportunity to post-process the lowered expansion
426   /// before returning.
postLower()427   virtual void postLower() {}
428 
429   /// When the SkipUnimplemented flag is set, addFakeDefUses() gets invoked by
430   /// the UnimplementedLoweringError macro to insert fake uses of all the
431   /// instruction variables and a fake def of the instruction dest, in order to
432   /// preserve integrity of liveness analysis.
433   void addFakeDefUses(const Inst *Instr);
434 
435   /// Find (non-SSA) instructions where the Dest variable appears in some source
436   /// operand, and set the IsDestRedefined flag.  This keeps liveness analysis
437   /// consistent.
438   void markRedefinitions();
439 
440   /// Make a pass over the Cfg to determine which variables need stack slots and
441   /// place them in a sorted list (SortedSpilledVariables). Among those, vars,
442   /// classify the spill variables as local to the basic block vs global
443   /// (multi-block) in order to compute the parameters GlobalsSize and
444   /// SpillAreaSizeBytes (represents locals or general vars if the coalescing of
445   /// locals is disallowed) along with alignments required for variables in each
446   /// area. We rely on accurate VMetadata in order to classify a variable as
447   /// global vs local (otherwise the variable is conservatively global). The
448   /// in-args should be initialized to 0.
449   ///
450   /// This is only a pre-pass and the actual stack slot assignment is handled
451   /// separately.
452   ///
453   /// There may be target-specific Variable types, which will be handled by
454   /// TargetVarHook. If the TargetVarHook returns true, then the variable is
455   /// skipped and not considered with the rest of the spilled variables.
456   void getVarStackSlotParams(VarList &SortedSpilledVariables,
457                              SmallBitVector &RegsUsed, size_t *GlobalsSize,
458                              size_t *SpillAreaSizeBytes,
459                              uint32_t *SpillAreaAlignmentBytes,
460                              uint32_t *LocalsSlotsAlignmentBytes,
461                              std::function<bool(Variable *)> TargetVarHook);
462 
463   /// Calculate the amount of padding needed to align the local and global areas
464   /// to the required alignment. This assumes the globals/locals layout used by
465   /// getVarStackSlotParams and assignVarStackSlots.
466   void alignStackSpillAreas(uint32_t SpillAreaStartOffset,
467                             uint32_t SpillAreaAlignmentBytes,
468                             size_t GlobalsSize,
469                             uint32_t LocalsSlotsAlignmentBytes,
470                             uint32_t *SpillAreaPaddingBytes,
471                             uint32_t *LocalsSlotsPaddingBytes);
472 
473   /// Make a pass through the SortedSpilledVariables and actually assign stack
474   /// slots. SpillAreaPaddingBytes takes into account stack alignment padding.
475   /// The SpillArea starts after that amount of padding. This matches the scheme
476   /// in getVarStackSlotParams, where there may be a separate multi-block global
477   /// var spill area and a local var spill area.
478   void assignVarStackSlots(VarList &SortedSpilledVariables,
479                            size_t SpillAreaPaddingBytes,
480                            size_t SpillAreaSizeBytes,
481                            size_t GlobalsAndSubsequentPaddingSize,
482                            bool UsesFramePointer);
483 
484   /// Sort the variables in Source based on required alignment. The variables
485   /// with the largest alignment need are placed in the front of the Dest list.
486   void sortVarsByAlignment(VarList &Dest, const VarList &Source) const;
487 
488   InstCall *makeHelperCall(RuntimeHelper FuncID, Variable *Dest, SizeT MaxSrcs);
489 
_set_dest_redefined()490   void _set_dest_redefined() { Context.getLastInserted()->setDestRedefined(); }
491 
492   bool shouldOptimizeMemIntrins();
493 
494   void scalarizeArithmetic(InstArithmetic::OpKind K, Variable *Dest,
495                            Operand *Src0, Operand *Src1);
496 
497   /// Generalizes scalarizeArithmetic to support other instruction types.
498   ///
499   /// insertScalarInstruction is a function-like object with signature
500   /// (Variable *Dest, Variable *Src0, Variable *Src1) -> Instr *.
501   template <typename... Operands,
502             typename F = std::function<Inst *(Variable *, Operands *...)>>
scalarizeInstruction(Variable * Dest,F insertScalarInstruction,Operands * ...Srcs)503   void scalarizeInstruction(Variable *Dest, F insertScalarInstruction,
504                             Operands *... Srcs) {
505     assert(GeneratingTargetHelpers &&
506            "scalarizeInstruction called during incorrect phase");
507     const Type DestTy = Dest->getType();
508     assert(isVectorType(DestTy));
509     const Type DestElementTy = typeElementType(DestTy);
510     const SizeT NumElements = typeNumElements(DestTy);
511 
512     Variable *T = Func->makeVariable(DestTy);
513     if (auto *VarVecOn32 = llvm::dyn_cast<VariableVecOn32>(T)) {
514       VarVecOn32->initVecElement(Func);
515       auto *Undef = ConstantUndef::create(Ctx, DestTy);
516       Context.insert<InstAssign>(T, Undef);
517     } else {
518       Context.insert<InstFakeDef>(T);
519     }
520 
521     for (SizeT I = 0; I < NumElements; ++I) {
522       auto *Index = Ctx->getConstantInt32(I);
523 
524       auto makeExtractThunk = [this, Index, NumElements](Operand *Src) {
525         return [this, Index, NumElements, Src]() {
526           (void)NumElements;
527           assert(typeNumElements(Src->getType()) == NumElements);
528 
529           const auto ElementTy = typeElementType(Src->getType());
530           auto *Op = Func->makeVariable(ElementTy);
531           Context.insert<InstExtractElement>(Op, Src, Index);
532           return Op;
533         };
534       };
535 
536       // Perform the operation as a scalar operation.
537       auto *Res = Func->makeVariable(DestElementTy);
538       auto *Arith = applyToThunkedArgs(insertScalarInstruction, Res,
539                                        makeExtractThunk(Srcs)...);
540       genTargetHelperCallFor(Arith);
541 
542       Variable *DestT = Func->makeVariable(DestTy);
543       Context.insert<InstInsertElement>(DestT, T, Res, Index);
544       T = DestT;
545     }
546     Context.insert<InstAssign>(Dest, T);
547   }
548 
549   // applyToThunkedArgs is used by scalarizeInstruction. Ideally, we would just
550   // call insertScalarInstruction(Res, Srcs...), but C++ does not specify
551   // evaluation order which means this leads to an unpredictable final
552   // output. Instead, we wrap each of the Srcs in a thunk and these
553   // applyToThunkedArgs functions apply the thunks in a well defined order so we
554   // still get well-defined output.
applyToThunkedArgs(std::function<Inst * (Variable *,Variable *)> insertScalarInstruction,Variable * Res,std::function<Variable * ()> thunk0)555   Inst *applyToThunkedArgs(
556       std::function<Inst *(Variable *, Variable *)> insertScalarInstruction,
557       Variable *Res, std::function<Variable *()> thunk0) {
558     auto *Src0 = thunk0();
559     return insertScalarInstruction(Res, Src0);
560   }
561 
562   Inst *
applyToThunkedArgs(std::function<Inst * (Variable *,Variable *,Variable *)> insertScalarInstruction,Variable * Res,std::function<Variable * ()> thunk0,std::function<Variable * ()> thunk1)563   applyToThunkedArgs(std::function<Inst *(Variable *, Variable *, Variable *)>
564                          insertScalarInstruction,
565                      Variable *Res, std::function<Variable *()> thunk0,
566                      std::function<Variable *()> thunk1) {
567     auto *Src0 = thunk0();
568     auto *Src1 = thunk1();
569     return insertScalarInstruction(Res, Src0, Src1);
570   }
571 
applyToThunkedArgs(std::function<Inst * (Variable *,Variable *,Variable *,Variable *)> insertScalarInstruction,Variable * Res,std::function<Variable * ()> thunk0,std::function<Variable * ()> thunk1,std::function<Variable * ()> thunk2)572   Inst *applyToThunkedArgs(
573       std::function<Inst *(Variable *, Variable *, Variable *, Variable *)>
574           insertScalarInstruction,
575       Variable *Res, std::function<Variable *()> thunk0,
576       std::function<Variable *()> thunk1, std::function<Variable *()> thunk2) {
577     auto *Src0 = thunk0();
578     auto *Src1 = thunk1();
579     auto *Src2 = thunk2();
580     return insertScalarInstruction(Res, Src0, Src1, Src2);
581   }
582 
583   /// SandboxType enumerates all possible sandboxing strategies that
584   enum SandboxType {
585     ST_None,
586     ST_NaCl,
587     ST_Nonsfi,
588   };
589 
590   static SandboxType determineSandboxTypeFromFlags(const ClFlags &Flags);
591 
592   Cfg *Func;
593   GlobalContext *Ctx;
594   bool HasComputedFrame = false;
595   bool CallsReturnsTwice = false;
596   SizeT NextLabelNumber = 0;
597   SizeT NextJumpTableNumber = 0;
598   LoweringContext Context;
599   const SandboxType SandboxingType = ST_None;
600 
601   const static constexpr char *H_getIP_prefix = "__Sz_getIP_";
602 };
603 
604 /// TargetDataLowering is used for "lowering" data including initializers for
605 /// global variables, and the internal constant pools. It is separated out from
606 /// TargetLowering because it does not require a Cfg.
607 class TargetDataLowering {
608   TargetDataLowering() = delete;
609   TargetDataLowering(const TargetDataLowering &) = delete;
610   TargetDataLowering &operator=(const TargetDataLowering &) = delete;
611 
612 public:
613   static std::unique_ptr<TargetDataLowering> createLowering(GlobalContext *Ctx);
614   virtual ~TargetDataLowering();
615 
616   virtual void lowerGlobals(const VariableDeclarationList &Vars,
617                             const std::string &SectionSuffix) = 0;
618   virtual void lowerConstants() = 0;
619   virtual void lowerJumpTables() = 0;
emitTargetRODataSections()620   virtual void emitTargetRODataSections() {}
621 
622 protected:
623   void emitGlobal(const VariableDeclaration &Var,
624                   const std::string &SectionSuffix);
625 
626   /// For now, we assume .long is the right directive for emitting 4 byte emit
627   /// global relocations. However, LLVM MIPS usually uses .4byte instead.
628   /// Perhaps there is some difference when the location is unaligned.
getEmit32Directive()629   static const char *getEmit32Directive() { return ".long"; }
630 
TargetDataLowering(GlobalContext * Ctx)631   explicit TargetDataLowering(GlobalContext *Ctx) : Ctx(Ctx) {}
632   GlobalContext *Ctx;
633 };
634 
635 /// TargetHeaderLowering is used to "lower" the header of an output file. It
636 /// writes out the target-specific header attributes. E.g., for ARM this writes
637 /// out the build attributes (float ABI, etc.).
638 class TargetHeaderLowering {
639   TargetHeaderLowering() = delete;
640   TargetHeaderLowering(const TargetHeaderLowering &) = delete;
641   TargetHeaderLowering &operator=(const TargetHeaderLowering &) = delete;
642 
643 public:
644   static std::unique_ptr<TargetHeaderLowering>
645   createLowering(GlobalContext *Ctx);
646   virtual ~TargetHeaderLowering();
647 
lower()648   virtual void lower() {}
649 
650 protected:
TargetHeaderLowering(GlobalContext * Ctx)651   explicit TargetHeaderLowering(GlobalContext *Ctx) : Ctx(Ctx) {}
652   GlobalContext *Ctx;
653 };
654 
655 } // end of namespace Ice
656 
657 #endif // SUBZERO_SRC_ICETARGETLOWERING_H
658