1 //===- InstrRefBasedImpl.cpp - Tracking Debug Value MIs -------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 /// \file InstrRefBasedImpl.cpp
9 ///
10 /// This is a separate implementation of LiveDebugValues, see
11 /// LiveDebugValues.cpp and VarLocBasedImpl.cpp for more information.
12 ///
13 /// This pass propagates variable locations between basic blocks, resolving
14 /// control flow conflicts between them. The problem is much like SSA
15 /// construction, where each DBG_VALUE instruction assigns the *value* that
16 /// a variable has, and every instruction where the variable is in scope uses
17 /// that variable. The resulting map of instruction-to-value is then translated
18 /// into a register (or spill) location for each variable over each instruction.
19 ///
20 /// This pass determines which DBG_VALUE dominates which instructions, or if
21 /// none do, where values must be merged (like PHI nodes). The added
22 /// complication is that because codegen has already finished, a PHI node may
23 /// be needed for a variable location to be correct, but no register or spill
24 /// slot merges the necessary values. In these circumstances, the variable
25 /// location is dropped.
26 ///
27 /// What makes this analysis non-trivial is loops: we cannot tell in advance
28 /// whether a variable location is live throughout a loop, or whether its
29 /// location is clobbered (or redefined by another DBG_VALUE), without
30 /// exploring all the way through.
31 ///
32 /// To make this simpler we perform two kinds of analysis. First, we identify
33 /// every value defined by every instruction (ignoring those that only move
34 /// another value), then compute a map of which values are available for each
35 /// instruction. This is stronger than a reaching-def analysis, as we create
36 /// PHI values where other values merge.
37 ///
38 /// Secondly, for each variable, we effectively re-construct SSA using each
39 /// DBG_VALUE as a def. The DBG_VALUEs read a value-number computed by the
40 /// first analysis from the location they refer to. We can then compute the
41 /// dominance frontiers of where a variable has a value, and create PHI nodes
42 /// where they merge.
43 /// This isn't precisely SSA-construction though, because the function shape
44 /// is pre-defined. If a variable location requires a PHI node, but no
45 /// PHI for the relevant values is present in the function (as computed by the
46 /// first analysis), the location must be dropped.
47 ///
48 /// Once both are complete, we can pass back over all instructions knowing:
49 ///  * What _value_ each variable should contain, either defined by an
50 ///    instruction or where control flow merges
51 ///  * What the location of that value is (if any).
52 /// Allowing us to create appropriate live-in DBG_VALUEs, and DBG_VALUEs when
53 /// a value moves location. After this pass runs, all variable locations within
54 /// a block should be specified by DBG_VALUEs within that block, allowing
55 /// DbgEntityHistoryCalculator to focus on individual blocks.
56 ///
57 /// This pass is able to go fast because the size of the first
58 /// reaching-definition analysis is proportional to the working-set size of
59 /// the function, which the compiler tries to keep small. (It's also
60 /// proportional to the number of blocks). Additionally, we repeatedly perform
61 /// the second reaching-definition analysis with only the variables and blocks
62 /// in a single lexical scope, exploiting their locality.
63 ///
64 /// Determining where PHIs happen is trickier with this approach, and it comes
65 /// to a head in the major problem for LiveDebugValues: is a value live-through
66 /// a loop, or not? Your garden-variety dataflow analysis aims to build a set of
67 /// facts about a function, however this analysis needs to generate new value
68 /// numbers at joins.
69 ///
70 /// To do this, consider a lattice of all definition values, from instructions
71 /// and from PHIs. Each PHI is characterised by the RPO number of the block it
72 /// occurs in. Each value pair A, B can be ordered by RPO(A) < RPO(B):
73 /// with non-PHI values at the top, and any PHI value in the last block (by RPO
74 /// order) at the bottom.
75 ///
76 /// (Awkwardly: lower-down-the _lattice_ means a greater RPO _number_. Below,
77 /// "rank" always refers to the former).
78 ///
79 /// At any join, for each register, we consider:
80 ///  * All incoming values, and
81 ///  * The PREVIOUS live-in value at this join.
82 /// If all incoming values agree: that's the live-in value. If they do not, the
83 /// incoming values are ranked according to the partial order, and the NEXT
84 /// LOWEST rank after the PREVIOUS live-in value is picked (multiple values of
85 /// the same rank are ignored as conflicting). If there are no candidate values,
86 /// or if the rank of the live-in would be lower than the rank of the current
87 /// blocks PHIs, create a new PHI value.
88 ///
89 /// Intuitively: if it's not immediately obvious what value a join should result
90 /// in, we iteratively descend from instruction-definitions down through PHI
91 /// values, getting closer to the current block each time. If the current block
92 /// is a loop head, this ordering is effectively searching outer levels of
93 /// loops, to find a value that's live-through the current loop.
94 ///
95 /// If there is no value that's live-through this loop, a PHI is created for
96 /// this location instead. We can't use a lower-ranked PHI because by definition
97 /// it doesn't dominate the current block. We can't create a PHI value any
98 /// earlier, because we risk creating a PHI value at a location where values do
99 /// not in fact merge, thus misrepresenting the truth, and not making the true
100 /// live-through value for variable locations.
101 ///
102 /// This algorithm applies to both calculating the availability of values in
103 /// the first analysis, and the location of variables in the second. However
104 /// for the second we add an extra dimension of pain: creating a variable
105 /// location PHI is only valid if, for each incoming edge,
106 ///  * There is a value for the variable on the incoming edge, and
107 ///  * All the edges have that value in the same register.
108 /// Or put another way: we can only create a variable-location PHI if there is
109 /// a matching machine-location PHI, each input to which is the variables value
110 /// in the predecessor block.
111 ///
112 /// To accomodate this difference, each point on the lattice is split in
113 /// two: a "proposed" PHI and "definite" PHI. Any PHI that can immediately
114 /// have a location determined are "definite" PHIs, and no further work is
115 /// needed. Otherwise, a location that all non-backedge predecessors agree
116 /// on is picked and propagated as a "proposed" PHI value. If that PHI value
117 /// is truly live-through, it'll appear on the loop backedges on the next
118 /// dataflow iteration, after which the block live-in moves to be a "definite"
119 /// PHI. If it's not truly live-through, the variable value will be downgraded
120 /// further as we explore the lattice, or remains "proposed" and is considered
121 /// invalid once dataflow completes.
122 ///
123 /// ### Terminology
124 ///
125 /// A machine location is a register or spill slot, a value is something that's
126 /// defined by an instruction or PHI node, while a variable value is the value
127 /// assigned to a variable. A variable location is a machine location, that must
128 /// contain the appropriate variable value. A value that is a PHI node is
129 /// occasionally called an mphi.
130 ///
131 /// The first dataflow problem is the "machine value location" problem,
132 /// because we're determining which machine locations contain which values.
133 /// The "locations" are constant: what's unknown is what value they contain.
134 ///
135 /// The second dataflow problem (the one for variables) is the "variable value
136 /// problem", because it's determining what values a variable has, rather than
137 /// what location those values are placed in. Unfortunately, it's not that
138 /// simple, because producing a PHI value always involves picking a location.
139 /// This is an imperfection that we just have to accept, at least for now.
140 ///
141 /// TODO:
142 ///   Overlapping fragments
143 ///   Entry values
144 ///   Add back DEBUG statements for debugging this
145 ///   Collect statistics
146 ///
147 //===----------------------------------------------------------------------===//
148 
149 #include "llvm/ADT/DenseMap.h"
150 #include "llvm/ADT/PostOrderIterator.h"
151 #include "llvm/ADT/SmallPtrSet.h"
152 #include "llvm/ADT/SmallSet.h"
153 #include "llvm/ADT/SmallVector.h"
154 #include "llvm/ADT/Statistic.h"
155 #include "llvm/ADT/UniqueVector.h"
156 #include "llvm/CodeGen/LexicalScopes.h"
157 #include "llvm/CodeGen/MachineBasicBlock.h"
158 #include "llvm/CodeGen/MachineFrameInfo.h"
159 #include "llvm/CodeGen/MachineFunction.h"
160 #include "llvm/CodeGen/MachineFunctionPass.h"
161 #include "llvm/CodeGen/MachineInstr.h"
162 #include "llvm/CodeGen/MachineInstrBuilder.h"
163 #include "llvm/CodeGen/MachineMemOperand.h"
164 #include "llvm/CodeGen/MachineOperand.h"
165 #include "llvm/CodeGen/PseudoSourceValue.h"
166 #include "llvm/CodeGen/RegisterScavenging.h"
167 #include "llvm/CodeGen/TargetFrameLowering.h"
168 #include "llvm/CodeGen/TargetInstrInfo.h"
169 #include "llvm/CodeGen/TargetLowering.h"
170 #include "llvm/CodeGen/TargetPassConfig.h"
171 #include "llvm/CodeGen/TargetRegisterInfo.h"
172 #include "llvm/CodeGen/TargetSubtargetInfo.h"
173 #include "llvm/Config/llvm-config.h"
174 #include "llvm/IR/DIBuilder.h"
175 #include "llvm/IR/DebugInfoMetadata.h"
176 #include "llvm/IR/DebugLoc.h"
177 #include "llvm/IR/Function.h"
178 #include "llvm/IR/Module.h"
179 #include "llvm/InitializePasses.h"
180 #include "llvm/MC/MCRegisterInfo.h"
181 #include "llvm/Pass.h"
182 #include "llvm/Support/Casting.h"
183 #include "llvm/Support/Compiler.h"
184 #include "llvm/Support/Debug.h"
185 #include "llvm/Support/raw_ostream.h"
186 #include <algorithm>
187 #include <cassert>
188 #include <cstdint>
189 #include <functional>
190 #include <queue>
191 #include <tuple>
192 #include <utility>
193 #include <vector>
194 #include <limits.h>
195 #include <limits>
196 
197 #include "LiveDebugValues.h"
198 
199 using namespace llvm;
200 
201 #define DEBUG_TYPE "livedebugvalues"
202 
203 STATISTIC(NumInserted, "Number of DBG_VALUE instructions inserted");
204 STATISTIC(NumRemoved, "Number of DBG_VALUE instructions removed");
205 
206 // Act more like the VarLoc implementation, by propagating some locations too
207 // far and ignoring some transfers.
208 static cl::opt<bool> EmulateOldLDV("emulate-old-livedebugvalues", cl::Hidden,
209                                    cl::desc("Act like old LiveDebugValues did"),
210                                    cl::init(false));
211 
212 // Rely on isStoreToStackSlotPostFE and similar to observe all stack spills.
213 static cl::opt<bool>
214     ObserveAllStackops("observe-all-stack-ops", cl::Hidden,
215                        cl::desc("Allow non-kill spill and restores"),
216                        cl::init(false));
217 
218 namespace {
219 
220 // The location at which a spilled value resides. It consists of a register and
221 // an offset.
222 struct SpillLoc {
223   unsigned SpillBase;
224   int SpillOffset;
225   bool operator==(const SpillLoc &Other) const {
226     return std::tie(SpillBase, SpillOffset) ==
227            std::tie(Other.SpillBase, Other.SpillOffset);
228   }
229   bool operator<(const SpillLoc &Other) const {
230     return std::tie(SpillBase, SpillOffset) <
231            std::tie(Other.SpillBase, Other.SpillOffset);
232   }
233 };
234 
235 class LocIdx {
236   unsigned Location;
237 
238   // Default constructor is private, initializing to an illegal location number.
239   // Use only for "not an entry" elements in IndexedMaps.
240   LocIdx() : Location(UINT_MAX) { }
241 
242 public:
243   #define NUM_LOC_BITS 24
244   LocIdx(unsigned L) : Location(L) {
245     assert(L < (1 << NUM_LOC_BITS) && "Machine locations must fit in 24 bits");
246   }
247 
248   static LocIdx MakeIllegalLoc() {
249     return LocIdx();
250   }
251 
252   bool isIllegal() const {
253     return Location == UINT_MAX;
254   }
255 
256   uint64_t asU64() const {
257     return Location;
258   }
259 
260   bool operator==(unsigned L) const {
261     return Location == L;
262   }
263 
264   bool operator==(const LocIdx &L) const {
265     return Location == L.Location;
266   }
267 
268   bool operator!=(unsigned L) const {
269     return !(*this == L);
270   }
271 
272   bool operator!=(const LocIdx &L) const {
273     return !(*this == L);
274   }
275 
276   bool operator<(const LocIdx &Other) const {
277     return Location < Other.Location;
278   }
279 };
280 
281 class LocIdxToIndexFunctor {
282 public:
283   using argument_type = LocIdx;
284   unsigned operator()(const LocIdx &L) const {
285     return L.asU64();
286   }
287 };
288 
289 /// Unique identifier for a value defined by an instruction, as a value type.
290 /// Casts back and forth to a uint64_t. Probably replacable with something less
291 /// bit-constrained. Each value identifies the instruction and machine location
292 /// where the value is defined, although there may be no corresponding machine
293 /// operand for it (ex: regmasks clobbering values). The instructions are
294 /// one-based, and definitions that are PHIs have instruction number zero.
295 ///
296 /// The obvious limits of a 1M block function or 1M instruction blocks are
297 /// problematic; but by that point we should probably have bailed out of
298 /// trying to analyse the function.
299 class ValueIDNum {
300   uint64_t BlockNo : 20;         /// The block where the def happens.
301   uint64_t InstNo : 20;          /// The Instruction where the def happens.
302                                  /// One based, is distance from start of block.
303   uint64_t LocNo : NUM_LOC_BITS; /// The machine location where the def happens.
304 
305 public:
306   // XXX -- temporarily enabled while the live-in / live-out tables are moved
307   // to something more type-y
308   ValueIDNum() : BlockNo(0xFFFFF),
309                  InstNo(0xFFFFF),
310                  LocNo(0xFFFFFF) { }
311 
312   ValueIDNum(uint64_t Block, uint64_t Inst, uint64_t Loc)
313     : BlockNo(Block), InstNo(Inst), LocNo(Loc) { }
314 
315   ValueIDNum(uint64_t Block, uint64_t Inst, LocIdx Loc)
316     : BlockNo(Block), InstNo(Inst), LocNo(Loc.asU64()) { }
317 
318   uint64_t getBlock() const { return BlockNo; }
319   uint64_t getInst() const { return InstNo; }
320   uint64_t getLoc() const { return LocNo; }
321   bool isPHI() const { return InstNo == 0; }
322 
323   uint64_t asU64() const {
324     uint64_t TmpBlock = BlockNo;
325     uint64_t TmpInst = InstNo;
326     return TmpBlock << 44ull | TmpInst << NUM_LOC_BITS | LocNo;
327   }
328 
329   static ValueIDNum fromU64(uint64_t v) {
330     uint64_t L = (v & 0x3FFF);
331     return {v >> 44ull, ((v >> NUM_LOC_BITS) & 0xFFFFF), L};
332   }
333 
334   bool operator<(const ValueIDNum &Other) const {
335     return asU64() < Other.asU64();
336   }
337 
338   bool operator==(const ValueIDNum &Other) const {
339     return std::tie(BlockNo, InstNo, LocNo) ==
340            std::tie(Other.BlockNo, Other.InstNo, Other.LocNo);
341   }
342 
343   bool operator!=(const ValueIDNum &Other) const { return !(*this == Other); }
344 
345   std::string asString(const std::string &mlocname) const {
346     return Twine("bb ")
347         .concat(Twine(BlockNo).concat(Twine(" inst ").concat(
348             Twine(InstNo).concat(Twine(" loc ").concat(Twine(mlocname))))))
349         .str();
350   }
351 
352   static ValueIDNum EmptyValue;
353 };
354 
355 } // end anonymous namespace
356 
357 namespace {
358 
359 /// Meta qualifiers for a value. Pair of whatever expression is used to qualify
360 /// the the value, and Boolean of whether or not it's indirect.
361 class DbgValueProperties {
362 public:
363   DbgValueProperties(const DIExpression *DIExpr, bool Indirect)
364       : DIExpr(DIExpr), Indirect(Indirect) {}
365 
366   /// Extract properties from an existing DBG_VALUE instruction.
367   DbgValueProperties(const MachineInstr &MI) {
368     assert(MI.isDebugValue());
369     DIExpr = MI.getDebugExpression();
370     Indirect = MI.getOperand(1).isImm();
371   }
372 
373   bool operator==(const DbgValueProperties &Other) const {
374     return std::tie(DIExpr, Indirect) == std::tie(Other.DIExpr, Other.Indirect);
375   }
376 
377   bool operator!=(const DbgValueProperties &Other) const {
378     return !(*this == Other);
379   }
380 
381   const DIExpression *DIExpr;
382   bool Indirect;
383 };
384 
385 /// Tracker for what values are in machine locations. Listens to the Things
386 /// being Done by various instructions, and maintains a table of what machine
387 /// locations have what values (as defined by a ValueIDNum).
388 ///
389 /// There are potentially a much larger number of machine locations on the
390 /// target machine than the actual working-set size of the function. On x86 for
391 /// example, we're extremely unlikely to want to track values through control
392 /// or debug registers. To avoid doing so, MLocTracker has several layers of
393 /// indirection going on, with two kinds of ``location'':
394 ///  * A LocID uniquely identifies a register or spill location, with a
395 ///    predictable value.
396 ///  * A LocIdx is a key (in the database sense) for a LocID and a ValueIDNum.
397 /// Whenever a location is def'd or used by a MachineInstr, we automagically
398 /// create a new LocIdx for a location, but not otherwise. This ensures we only
399 /// account for locations that are actually used or defined. The cost is another
400 /// vector lookup (of LocID -> LocIdx) over any other implementation. This is
401 /// fairly cheap, and the compiler tries to reduce the working-set at any one
402 /// time in the function anyway.
403 ///
404 /// Register mask operands completely blow this out of the water; I've just
405 /// piled hacks on top of hacks to get around that.
406 class MLocTracker {
407 public:
408   MachineFunction &MF;
409   const TargetInstrInfo &TII;
410   const TargetRegisterInfo &TRI;
411   const TargetLowering &TLI;
412 
413   /// IndexedMap type, mapping from LocIdx to ValueIDNum.
414   typedef IndexedMap<ValueIDNum, LocIdxToIndexFunctor> LocToValueType;
415 
416   /// Map of LocIdxes to the ValueIDNums that they store. This is tightly
417   /// packed, entries only exist for locations that are being tracked.
418   LocToValueType LocIdxToIDNum;
419 
420   /// "Map" of machine location IDs (i.e., raw register or spill number) to the
421   /// LocIdx key / number for that location. There are always at least as many
422   /// as the number of registers on the target -- if the value in the register
423   /// is not being tracked, then the LocIdx value will be zero. New entries are
424   /// appended if a new spill slot begins being tracked.
425   /// This, and the corresponding reverse map persist for the analysis of the
426   /// whole function, and is necessarying for decoding various vectors of
427   /// values.
428   std::vector<LocIdx> LocIDToLocIdx;
429 
430   /// Inverse map of LocIDToLocIdx.
431   IndexedMap<unsigned, LocIdxToIndexFunctor> LocIdxToLocID;
432 
433   /// Unique-ification of spill slots. Used to number them -- their LocID
434   /// number is the index in SpillLocs minus one plus NumRegs.
435   UniqueVector<SpillLoc> SpillLocs;
436 
437   // If we discover a new machine location, assign it an mphi with this
438   // block number.
439   unsigned CurBB;
440 
441   /// Cached local copy of the number of registers the target has.
442   unsigned NumRegs;
443 
444   /// Collection of register mask operands that have been observed. Second part
445   /// of pair indicates the instruction that they happened in. Used to
446   /// reconstruct where defs happened if we start tracking a location later
447   /// on.
448   SmallVector<std::pair<const MachineOperand *, unsigned>, 32> Masks;
449 
450   /// Iterator for locations and the values they contain. Dereferencing
451   /// produces a struct/pair containing the LocIdx key for this location,
452   /// and a reference to the value currently stored. Simplifies the process
453   /// of seeking a particular location.
454   class MLocIterator {
455     LocToValueType &ValueMap;
456     LocIdx Idx;
457 
458   public:
459     class value_type {
460       public:
461       value_type(LocIdx Idx, ValueIDNum &Value) : Idx(Idx), Value(Value) { }
462       const LocIdx Idx;  /// Read-only index of this location.
463       ValueIDNum &Value; /// Reference to the stored value at this location.
464     };
465 
466     MLocIterator(LocToValueType &ValueMap, LocIdx Idx)
467       : ValueMap(ValueMap), Idx(Idx) { }
468 
469     bool operator==(const MLocIterator &Other) const {
470       assert(&ValueMap == &Other.ValueMap);
471       return Idx == Other.Idx;
472     }
473 
474     bool operator!=(const MLocIterator &Other) const {
475       return !(*this == Other);
476     }
477 
478     void operator++() {
479       Idx = LocIdx(Idx.asU64() + 1);
480     }
481 
482     value_type operator*() {
483       return value_type(Idx, ValueMap[LocIdx(Idx)]);
484     }
485   };
486 
487   MLocTracker(MachineFunction &MF, const TargetInstrInfo &TII,
488               const TargetRegisterInfo &TRI, const TargetLowering &TLI)
489       : MF(MF), TII(TII), TRI(TRI), TLI(TLI),
490         LocIdxToIDNum(ValueIDNum::EmptyValue),
491         LocIdxToLocID(0) {
492     NumRegs = TRI.getNumRegs();
493     reset();
494     LocIDToLocIdx.resize(NumRegs, LocIdx::MakeIllegalLoc());
495     assert(NumRegs < (1u << NUM_LOC_BITS)); // Detect bit packing failure
496 
497     // Always track SP. This avoids the implicit clobbering caused by regmasks
498     // from affectings its values. (LiveDebugValues disbelieves calls and
499     // regmasks that claim to clobber SP).
500     unsigned SP = TLI.getStackPointerRegisterToSaveRestore();
501     if (SP) {
502       unsigned ID = getLocID(SP, false);
503       (void)lookupOrTrackRegister(ID);
504     }
505   }
506 
507   /// Produce location ID number for indexing LocIDToLocIdx. Takes the register
508   /// or spill number, and flag for whether it's a spill or not.
509   unsigned getLocID(unsigned RegOrSpill, bool isSpill) {
510     return (isSpill) ? RegOrSpill + NumRegs - 1 : RegOrSpill;
511   }
512 
513   /// Accessor for reading the value at Idx.
514   ValueIDNum getNumAtPos(LocIdx Idx) const {
515     assert(Idx.asU64() < LocIdxToIDNum.size());
516     return LocIdxToIDNum[Idx];
517   }
518 
519   unsigned getNumLocs(void) const { return LocIdxToIDNum.size(); }
520 
521   /// Reset all locations to contain a PHI value at the designated block. Used
522   /// sometimes for actual PHI values, othertimes to indicate the block entry
523   /// value (before any more information is known).
524   void setMPhis(unsigned NewCurBB) {
525     CurBB = NewCurBB;
526     for (auto Location : locations())
527       Location.Value = {CurBB, 0, Location.Idx};
528   }
529 
530   /// Load values for each location from array of ValueIDNums. Take current
531   /// bbnum just in case we read a value from a hitherto untouched register.
532   void loadFromArray(ValueIDNum *Locs, unsigned NewCurBB) {
533     CurBB = NewCurBB;
534     // Iterate over all tracked locations, and load each locations live-in
535     // value into our local index.
536     for (auto Location : locations())
537       Location.Value = Locs[Location.Idx.asU64()];
538   }
539 
540   /// Wipe any un-necessary location records after traversing a block.
541   void reset(void) {
542     // We could reset all the location values too; however either loadFromArray
543     // or setMPhis should be called before this object is re-used. Just
544     // clear Masks, they're definitely not needed.
545     Masks.clear();
546   }
547 
548   /// Clear all data. Destroys the LocID <=> LocIdx map, which makes most of
549   /// the information in this pass uninterpretable.
550   void clear(void) {
551     reset();
552     LocIDToLocIdx.clear();
553     LocIdxToLocID.clear();
554     LocIdxToIDNum.clear();
555     //SpillLocs.reset(); XXX UniqueVector::reset assumes a SpillLoc casts from 0
556     SpillLocs = decltype(SpillLocs)();
557 
558     LocIDToLocIdx.resize(NumRegs, LocIdx::MakeIllegalLoc());
559   }
560 
561   /// Set a locaiton to a certain value.
562   void setMLoc(LocIdx L, ValueIDNum Num) {
563     assert(L.asU64() < LocIdxToIDNum.size());
564     LocIdxToIDNum[L] = Num;
565   }
566 
567   /// Create a LocIdx for an untracked register ID. Initialize it to either an
568   /// mphi value representing a live-in, or a recent register mask clobber.
569   LocIdx trackRegister(unsigned ID) {
570     assert(ID != 0);
571     LocIdx NewIdx = LocIdx(LocIdxToIDNum.size());
572     LocIdxToIDNum.grow(NewIdx);
573     LocIdxToLocID.grow(NewIdx);
574 
575     // Default: it's an mphi.
576     ValueIDNum ValNum = {CurBB, 0, NewIdx};
577     // Was this reg ever touched by a regmask?
578     for (const auto &MaskPair : reverse(Masks)) {
579       if (MaskPair.first->clobbersPhysReg(ID)) {
580         // There was an earlier def we skipped.
581         ValNum = {CurBB, MaskPair.second, NewIdx};
582         break;
583       }
584     }
585 
586     LocIdxToIDNum[NewIdx] = ValNum;
587     LocIdxToLocID[NewIdx] = ID;
588     return NewIdx;
589   }
590 
591   LocIdx lookupOrTrackRegister(unsigned ID) {
592     LocIdx &Index = LocIDToLocIdx[ID];
593     if (Index.isIllegal())
594       Index = trackRegister(ID);
595     return Index;
596   }
597 
598   /// Record a definition of the specified register at the given block / inst.
599   /// This doesn't take a ValueIDNum, because the definition and its location
600   /// are synonymous.
601   void defReg(Register R, unsigned BB, unsigned Inst) {
602     unsigned ID = getLocID(R, false);
603     LocIdx Idx = lookupOrTrackRegister(ID);
604     ValueIDNum ValueID = {BB, Inst, Idx};
605     LocIdxToIDNum[Idx] = ValueID;
606   }
607 
608   /// Set a register to a value number. To be used if the value number is
609   /// known in advance.
610   void setReg(Register R, ValueIDNum ValueID) {
611     unsigned ID = getLocID(R, false);
612     LocIdx Idx = lookupOrTrackRegister(ID);
613     LocIdxToIDNum[Idx] = ValueID;
614   }
615 
616   ValueIDNum readReg(Register R) {
617     unsigned ID = getLocID(R, false);
618     LocIdx Idx = lookupOrTrackRegister(ID);
619     return LocIdxToIDNum[Idx];
620   }
621 
622   /// Reset a register value to zero / empty. Needed to replicate the
623   /// VarLoc implementation where a copy to/from a register effectively
624   /// clears the contents of the source register. (Values can only have one
625   ///  machine location in VarLocBasedImpl).
626   void wipeRegister(Register R) {
627     unsigned ID = getLocID(R, false);
628     LocIdx Idx = LocIDToLocIdx[ID];
629     LocIdxToIDNum[Idx] = ValueIDNum::EmptyValue;
630   }
631 
632   /// Determine the LocIdx of an existing register.
633   LocIdx getRegMLoc(Register R) {
634     unsigned ID = getLocID(R, false);
635     return LocIDToLocIdx[ID];
636   }
637 
638   /// Record a RegMask operand being executed. Defs any register we currently
639   /// track, stores a pointer to the mask in case we have to account for it
640   /// later.
641   void writeRegMask(const MachineOperand *MO, unsigned CurBB, unsigned InstID) {
642     // Ensure SP exists, so that we don't override it later.
643     unsigned SP = TLI.getStackPointerRegisterToSaveRestore();
644 
645     // Def any register we track have that isn't preserved. The regmask
646     // terminates the liveness of a register, meaning its value can't be
647     // relied upon -- we represent this by giving it a new value.
648     for (auto Location : locations()) {
649       unsigned ID = LocIdxToLocID[Location.Idx];
650       // Don't clobber SP, even if the mask says it's clobbered.
651       if (ID < NumRegs && ID != SP && MO->clobbersPhysReg(ID))
652         defReg(ID, CurBB, InstID);
653     }
654     Masks.push_back(std::make_pair(MO, InstID));
655   }
656 
657   /// Find LocIdx for SpillLoc \p L, creating a new one if it's not tracked.
658   LocIdx getOrTrackSpillLoc(SpillLoc L) {
659     unsigned SpillID = SpillLocs.idFor(L);
660     if (SpillID == 0) {
661       SpillID = SpillLocs.insert(L);
662       unsigned L = getLocID(SpillID, true);
663       LocIdx Idx = LocIdx(LocIdxToIDNum.size()); // New idx
664       LocIdxToIDNum.grow(Idx);
665       LocIdxToLocID.grow(Idx);
666       LocIDToLocIdx.push_back(Idx);
667       LocIdxToLocID[Idx] = L;
668       return Idx;
669     } else {
670       unsigned L = getLocID(SpillID, true);
671       LocIdx Idx = LocIDToLocIdx[L];
672       return Idx;
673     }
674   }
675 
676   /// Set the value stored in a spill slot.
677   void setSpill(SpillLoc L, ValueIDNum ValueID) {
678     LocIdx Idx = getOrTrackSpillLoc(L);
679     LocIdxToIDNum[Idx] = ValueID;
680   }
681 
682   /// Read whatever value is in a spill slot, or None if it isn't tracked.
683   Optional<ValueIDNum> readSpill(SpillLoc L) {
684     unsigned SpillID = SpillLocs.idFor(L);
685     if (SpillID == 0)
686       return None;
687 
688     unsigned LocID = getLocID(SpillID, true);
689     LocIdx Idx = LocIDToLocIdx[LocID];
690     return LocIdxToIDNum[Idx];
691   }
692 
693   /// Determine the LocIdx of a spill slot. Return None if it previously
694   /// hasn't had a value assigned.
695   Optional<LocIdx> getSpillMLoc(SpillLoc L) {
696     unsigned SpillID = SpillLocs.idFor(L);
697     if (SpillID == 0)
698       return None;
699     unsigned LocNo = getLocID(SpillID, true);
700     return LocIDToLocIdx[LocNo];
701   }
702 
703   /// Return true if Idx is a spill machine location.
704   bool isSpill(LocIdx Idx) const {
705     return LocIdxToLocID[Idx] >= NumRegs;
706   }
707 
708   MLocIterator begin() {
709     return MLocIterator(LocIdxToIDNum, 0);
710   }
711 
712   MLocIterator end() {
713     return MLocIterator(LocIdxToIDNum, LocIdxToIDNum.size());
714   }
715 
716   /// Return a range over all locations currently tracked.
717   iterator_range<MLocIterator> locations() {
718     return llvm::make_range(begin(), end());
719   }
720 
721   std::string LocIdxToName(LocIdx Idx) const {
722     unsigned ID = LocIdxToLocID[Idx];
723     if (ID >= NumRegs)
724       return Twine("slot ").concat(Twine(ID - NumRegs)).str();
725     else
726       return TRI.getRegAsmName(ID).str();
727   }
728 
729   std::string IDAsString(const ValueIDNum &Num) const {
730     std::string DefName = LocIdxToName(Num.getLoc());
731     return Num.asString(DefName);
732   }
733 
734   LLVM_DUMP_METHOD
735   void dump() {
736     for (auto Location : locations()) {
737       std::string MLocName = LocIdxToName(Location.Value.getLoc());
738       std::string DefName = Location.Value.asString(MLocName);
739       dbgs() << LocIdxToName(Location.Idx) << " --> " << DefName << "\n";
740     }
741   }
742 
743   LLVM_DUMP_METHOD
744   void dump_mloc_map() {
745     for (auto Location : locations()) {
746       std::string foo = LocIdxToName(Location.Idx);
747       dbgs() << "Idx " << Location.Idx.asU64() << " " << foo << "\n";
748     }
749   }
750 
751   /// Create a DBG_VALUE based on  machine location \p MLoc. Qualify it with the
752   /// information in \pProperties, for variable Var. Don't insert it anywhere,
753   /// just return the builder for it.
754   MachineInstrBuilder emitLoc(Optional<LocIdx> MLoc, const DebugVariable &Var,
755                               const DbgValueProperties &Properties) {
756     DebugLoc DL =
757         DebugLoc::get(0, 0, Var.getVariable()->getScope(), Var.getInlinedAt());
758     auto MIB = BuildMI(MF, DL, TII.get(TargetOpcode::DBG_VALUE));
759 
760     const DIExpression *Expr = Properties.DIExpr;
761     if (!MLoc) {
762       // No location -> DBG_VALUE $noreg
763       MIB.addReg(0, RegState::Debug);
764       MIB.addReg(0, RegState::Debug);
765     } else if (LocIdxToLocID[*MLoc] >= NumRegs) {
766       unsigned LocID = LocIdxToLocID[*MLoc];
767       const SpillLoc &Spill = SpillLocs[LocID - NumRegs + 1];
768       Expr = DIExpression::prepend(Expr, DIExpression::ApplyOffset,
769                                    Spill.SpillOffset);
770       unsigned Base = Spill.SpillBase;
771       MIB.addReg(Base, RegState::Debug);
772       MIB.addImm(0);
773     } else {
774       unsigned LocID = LocIdxToLocID[*MLoc];
775       MIB.addReg(LocID, RegState::Debug);
776       if (Properties.Indirect)
777         MIB.addImm(0);
778       else
779         MIB.addReg(0, RegState::Debug);
780     }
781 
782     MIB.addMetadata(Var.getVariable());
783     MIB.addMetadata(Expr);
784     return MIB;
785   }
786 };
787 
788 /// Class recording the (high level) _value_ of a variable. Identifies either
789 /// the value of the variable as a ValueIDNum, or a constant MachineOperand.
790 /// This class also stores meta-information about how the value is qualified.
791 /// Used to reason about variable values when performing the second
792 /// (DebugVariable specific) dataflow analysis.
793 class DbgValue {
794 public:
795   union {
796     /// If Kind is Def, the value number that this value is based on.
797     ValueIDNum ID;
798     /// If Kind is Const, the MachineOperand defining this value.
799     MachineOperand MO;
800     /// For a NoVal DbgValue, which block it was generated in.
801     unsigned BlockNo;
802   };
803   /// Qualifiers for the ValueIDNum above.
804   DbgValueProperties Properties;
805 
806   typedef enum {
807     Undef,     // Represents a DBG_VALUE $noreg in the transfer function only.
808     Def,       // This value is defined by an inst, or is a PHI value.
809     Const,     // A constant value contained in the MachineOperand field.
810     Proposed,  // This is a tentative PHI value, which may be confirmed or
811                // invalidated later.
812     NoVal      // Empty DbgValue, generated during dataflow. BlockNo stores
813                // which block this was generated in.
814    } KindT;
815   /// Discriminator for whether this is a constant or an in-program value.
816   KindT Kind;
817 
818   DbgValue(const ValueIDNum &Val, const DbgValueProperties &Prop, KindT Kind)
819     : ID(Val), Properties(Prop), Kind(Kind) {
820     assert(Kind == Def || Kind == Proposed);
821   }
822 
823   DbgValue(unsigned BlockNo, const DbgValueProperties &Prop, KindT Kind)
824     : BlockNo(BlockNo), Properties(Prop), Kind(Kind) {
825     assert(Kind == NoVal);
826   }
827 
828   DbgValue(const MachineOperand &MO, const DbgValueProperties &Prop, KindT Kind)
829     : MO(MO), Properties(Prop), Kind(Kind) {
830     assert(Kind == Const);
831   }
832 
833   DbgValue(const DbgValueProperties &Prop, KindT Kind)
834     : Properties(Prop), Kind(Kind) {
835     assert(Kind == Undef &&
836            "Empty DbgValue constructor must pass in Undef kind");
837   }
838 
839   void dump(const MLocTracker *MTrack) const {
840     if (Kind == Const) {
841       MO.dump();
842     } else if (Kind == NoVal) {
843       dbgs() << "NoVal(" << BlockNo << ")";
844     } else if (Kind == Proposed) {
845       dbgs() << "VPHI(" << MTrack->IDAsString(ID) << ")";
846     } else {
847       assert(Kind == Def);
848       dbgs() << MTrack->IDAsString(ID);
849     }
850     if (Properties.Indirect)
851       dbgs() << " indir";
852     if (Properties.DIExpr)
853       dbgs() << " " << *Properties.DIExpr;
854   }
855 
856   bool operator==(const DbgValue &Other) const {
857     if (std::tie(Kind, Properties) != std::tie(Other.Kind, Other.Properties))
858       return false;
859     else if (Kind == Proposed && ID != Other.ID)
860       return false;
861     else if (Kind == Def && ID != Other.ID)
862       return false;
863     else if (Kind == NoVal && BlockNo != Other.BlockNo)
864       return false;
865     else if (Kind == Const)
866       return MO.isIdenticalTo(Other.MO);
867 
868     return true;
869   }
870 
871   bool operator!=(const DbgValue &Other) const { return !(*this == Other); }
872 };
873 
874 /// Types for recording sets of variable fragments that overlap. For a given
875 /// local variable, we record all other fragments of that variable that could
876 /// overlap it, to reduce search time.
877 using FragmentOfVar =
878     std::pair<const DILocalVariable *, DIExpression::FragmentInfo>;
879 using OverlapMap =
880     DenseMap<FragmentOfVar, SmallVector<DIExpression::FragmentInfo, 1>>;
881 
882 /// Collection of DBG_VALUEs observed when traversing a block. Records each
883 /// variable and the value the DBG_VALUE refers to. Requires the machine value
884 /// location dataflow algorithm to have run already, so that values can be
885 /// identified.
886 class VLocTracker {
887 public:
888   /// Map DebugVariable to the latest Value it's defined to have.
889   /// Needs to be a MapVector because we determine order-in-the-input-MIR from
890   /// the order in this container.
891   /// We only retain the last DbgValue in each block for each variable, to
892   /// determine the blocks live-out variable value. The Vars container forms the
893   /// transfer function for this block, as part of the dataflow analysis. The
894   /// movement of values between locations inside of a block is handled at a
895   /// much later stage, in the TransferTracker class.
896   MapVector<DebugVariable, DbgValue> Vars;
897   DenseMap<DebugVariable, const DILocation *> Scopes;
898   MachineBasicBlock *MBB;
899 
900 public:
901   VLocTracker() {}
902 
903   void defVar(const MachineInstr &MI, Optional<ValueIDNum> ID) {
904     // XXX skipping overlapping fragments for now.
905     assert(MI.isDebugValue());
906     DebugVariable Var(MI.getDebugVariable(), MI.getDebugExpression(),
907                       MI.getDebugLoc()->getInlinedAt());
908     DbgValueProperties Properties(MI);
909     DbgValue Rec = (ID) ? DbgValue(*ID, Properties, DbgValue::Def)
910                         : DbgValue(Properties, DbgValue::Undef);
911 
912     // Attempt insertion; overwrite if it's already mapped.
913     auto Result = Vars.insert(std::make_pair(Var, Rec));
914     if (!Result.second)
915       Result.first->second = Rec;
916     Scopes[Var] = MI.getDebugLoc().get();
917   }
918 
919   void defVar(const MachineInstr &MI, const MachineOperand &MO) {
920     // XXX skipping overlapping fragments for now.
921     assert(MI.isDebugValue());
922     DebugVariable Var(MI.getDebugVariable(), MI.getDebugExpression(),
923                       MI.getDebugLoc()->getInlinedAt());
924     DbgValueProperties Properties(MI);
925     DbgValue Rec = DbgValue(MO, Properties, DbgValue::Const);
926 
927     // Attempt insertion; overwrite if it's already mapped.
928     auto Result = Vars.insert(std::make_pair(Var, Rec));
929     if (!Result.second)
930       Result.first->second = Rec;
931     Scopes[Var] = MI.getDebugLoc().get();
932   }
933 };
934 
935 /// Tracker for converting machine value locations and variable values into
936 /// variable locations (the output of LiveDebugValues), recorded as DBG_VALUEs
937 /// specifying block live-in locations and transfers within blocks.
938 ///
939 /// Operating on a per-block basis, this class takes a (pre-loaded) MLocTracker
940 /// and must be initialized with the set of variable values that are live-in to
941 /// the block. The caller then repeatedly calls process(). TransferTracker picks
942 /// out variable locations for the live-in variable values (if there _is_ a
943 /// location) and creates the corresponding DBG_VALUEs. Then, as the block is
944 /// stepped through, transfers of values between machine locations are
945 /// identified and if profitable, a DBG_VALUE created.
946 ///
947 /// This is where debug use-before-defs would be resolved: a variable with an
948 /// unavailable value could materialize in the middle of a block, when the
949 /// value becomes available. Or, we could detect clobbers and re-specify the
950 /// variable in a backup location. (XXX these are unimplemented).
951 class TransferTracker {
952 public:
953   const TargetInstrInfo *TII;
954   /// This machine location tracker is assumed to always contain the up-to-date
955   /// value mapping for all machine locations. TransferTracker only reads
956   /// information from it. (XXX make it const?)
957   MLocTracker *MTracker;
958   MachineFunction &MF;
959 
960   /// Record of all changes in variable locations at a block position. Awkwardly
961   /// we allow inserting either before or after the point: MBB != nullptr
962   /// indicates it's before, otherwise after.
963   struct Transfer {
964     MachineBasicBlock::iterator Pos; /// Position to insert DBG_VALUes
965     MachineBasicBlock *MBB;          /// non-null if we should insert after.
966     SmallVector<MachineInstr *, 4> Insts; /// Vector of DBG_VALUEs to insert.
967   };
968 
969   typedef struct {
970     LocIdx Loc;
971     DbgValueProperties Properties;
972   } LocAndProperties;
973 
974   /// Collection of transfers (DBG_VALUEs) to be inserted.
975   SmallVector<Transfer, 32> Transfers;
976 
977   /// Local cache of what-value-is-in-what-LocIdx. Used to identify differences
978   /// between TransferTrackers view of variable locations and MLocTrackers. For
979   /// example, MLocTracker observes all clobbers, but TransferTracker lazily
980   /// does not.
981   std::vector<ValueIDNum> VarLocs;
982 
983   /// Map from LocIdxes to which DebugVariables are based that location.
984   /// Mantained while stepping through the block. Not accurate if
985   /// VarLocs[Idx] != MTracker->LocIdxToIDNum[Idx].
986   std::map<LocIdx, SmallSet<DebugVariable, 4>> ActiveMLocs;
987 
988   /// Map from DebugVariable to it's current location and qualifying meta
989   /// information. To be used in conjunction with ActiveMLocs to construct
990   /// enough information for the DBG_VALUEs for a particular LocIdx.
991   DenseMap<DebugVariable, LocAndProperties> ActiveVLocs;
992 
993   /// Temporary cache of DBG_VALUEs to be entered into the Transfers collection.
994   SmallVector<MachineInstr *, 4> PendingDbgValues;
995 
996   const TargetRegisterInfo &TRI;
997   const BitVector &CalleeSavedRegs;
998 
999   TransferTracker(const TargetInstrInfo *TII, MLocTracker *MTracker,
1000                   MachineFunction &MF, const TargetRegisterInfo &TRI,
1001                   const BitVector &CalleeSavedRegs)
1002       : TII(TII), MTracker(MTracker), MF(MF), TRI(TRI),
1003         CalleeSavedRegs(CalleeSavedRegs) {}
1004 
1005   /// Load object with live-in variable values. \p mlocs contains the live-in
1006   /// values in each machine location, while \p vlocs the live-in variable
1007   /// values. This method picks variable locations for the live-in variables,
1008   /// creates DBG_VALUEs and puts them in #Transfers, then prepares the other
1009   /// object fields to track variable locations as we step through the block.
1010   /// FIXME: could just examine mloctracker instead of passing in \p mlocs?
1011   void loadInlocs(MachineBasicBlock &MBB, ValueIDNum *MLocs,
1012                   SmallVectorImpl<std::pair<DebugVariable, DbgValue>> &VLocs,
1013                   unsigned NumLocs) {
1014     ActiveMLocs.clear();
1015     ActiveVLocs.clear();
1016     VarLocs.clear();
1017     VarLocs.reserve(NumLocs);
1018 
1019     auto isCalleeSaved = [&](LocIdx L) {
1020       unsigned Reg = MTracker->LocIdxToLocID[L];
1021       if (Reg >= MTracker->NumRegs)
1022         return false;
1023       for (MCRegAliasIterator RAI(Reg, &TRI, true); RAI.isValid(); ++RAI)
1024         if (CalleeSavedRegs.test(*RAI))
1025           return true;
1026       return false;
1027     };
1028 
1029     // Map of the preferred location for each value.
1030     std::map<ValueIDNum, LocIdx> ValueToLoc;
1031 
1032     // Produce a map of value numbers to the current machine locs they live
1033     // in. When emulating VarLocBasedImpl, there should only be one
1034     // location; when not, we get to pick.
1035     for (auto Location : MTracker->locations()) {
1036       LocIdx Idx = Location.Idx;
1037       ValueIDNum &VNum = MLocs[Idx.asU64()];
1038       VarLocs.push_back(VNum);
1039       auto it = ValueToLoc.find(VNum);
1040       // In order of preference, pick:
1041       //  * Callee saved registers,
1042       //  * Other registers,
1043       //  * Spill slots.
1044       if (it == ValueToLoc.end() || MTracker->isSpill(it->second) ||
1045           (!isCalleeSaved(it->second) && isCalleeSaved(Idx.asU64()))) {
1046         // Insert, or overwrite if insertion failed.
1047         auto PrefLocRes = ValueToLoc.insert(std::make_pair(VNum, Idx));
1048         if (!PrefLocRes.second)
1049           PrefLocRes.first->second = Idx;
1050       }
1051     }
1052 
1053     // Now map variables to their picked LocIdxes.
1054     for (auto Var : VLocs) {
1055       if (Var.second.Kind == DbgValue::Const) {
1056         PendingDbgValues.push_back(
1057             emitMOLoc(Var.second.MO, Var.first, Var.second.Properties));
1058         continue;
1059       }
1060 
1061       // If the value has no location, we can't make a variable location.
1062       auto ValuesPreferredLoc = ValueToLoc.find(Var.second.ID);
1063       if (ValuesPreferredLoc == ValueToLoc.end())
1064         continue;
1065 
1066       LocIdx M = ValuesPreferredLoc->second;
1067       auto NewValue = LocAndProperties{M, Var.second.Properties};
1068       auto Result = ActiveVLocs.insert(std::make_pair(Var.first, NewValue));
1069       if (!Result.second)
1070         Result.first->second = NewValue;
1071       ActiveMLocs[M].insert(Var.first);
1072       PendingDbgValues.push_back(
1073           MTracker->emitLoc(M, Var.first, Var.second.Properties));
1074     }
1075     flushDbgValues(MBB.begin(), &MBB);
1076   }
1077 
1078   /// Helper to move created DBG_VALUEs into Transfers collection.
1079   void flushDbgValues(MachineBasicBlock::iterator Pos, MachineBasicBlock *MBB) {
1080     if (PendingDbgValues.size() > 0) {
1081       Transfers.push_back({Pos, MBB, PendingDbgValues});
1082       PendingDbgValues.clear();
1083     }
1084   }
1085 
1086   /// Handle a DBG_VALUE within a block. Terminate the variables current
1087   /// location, and record the value its DBG_VALUE refers to, so that we can
1088   /// detect location transfers later on.
1089   void redefVar(const MachineInstr &MI) {
1090     DebugVariable Var(MI.getDebugVariable(), MI.getDebugExpression(),
1091                       MI.getDebugLoc()->getInlinedAt());
1092     const MachineOperand &MO = MI.getOperand(0);
1093 
1094     // Erase any previous location,
1095     auto It = ActiveVLocs.find(Var);
1096     if (It != ActiveVLocs.end()) {
1097       ActiveMLocs[It->second.Loc].erase(Var);
1098     }
1099 
1100     // Insert a new variable location. Ignore non-register locations, we don't
1101     // transfer those, and can't currently describe spill locs independently of
1102     // regs.
1103     // (This is because a spill location is a DBG_VALUE of the stack pointer).
1104     if (!MO.isReg() || MO.getReg() == 0) {
1105       if (It != ActiveVLocs.end())
1106         ActiveVLocs.erase(It);
1107       return;
1108     }
1109 
1110     Register Reg = MO.getReg();
1111     LocIdx MLoc = MTracker->getRegMLoc(Reg);
1112     DbgValueProperties Properties(MI);
1113 
1114     // Check whether our local copy of values-by-location in #VarLocs is out of
1115     // date. Wipe old tracking data for the location if it's been clobbered in
1116     // the meantime.
1117     if (MTracker->getNumAtPos(MLoc) != VarLocs[MLoc.asU64()]) {
1118       for (auto &P : ActiveMLocs[MLoc.asU64()]) {
1119         ActiveVLocs.erase(P);
1120       }
1121       ActiveMLocs[MLoc].clear();
1122       VarLocs[MLoc.asU64()] = MTracker->getNumAtPos(MLoc);
1123     }
1124 
1125     ActiveMLocs[MLoc].insert(Var);
1126     if (It == ActiveVLocs.end()) {
1127       ActiveVLocs.insert(std::make_pair(Var, LocAndProperties{MLoc, Properties}));
1128     } else {
1129       It->second.Loc = MLoc;
1130       It->second.Properties = Properties;
1131     }
1132   }
1133 
1134   /// Explicitly terminate variable locations based on \p mloc. Creates undef
1135   /// DBG_VALUEs for any variables that were located there, and clears
1136   /// #ActiveMLoc / #ActiveVLoc tracking information for that location.
1137   void clobberMloc(LocIdx MLoc, MachineBasicBlock::iterator Pos) {
1138     assert(MTracker->isSpill(MLoc));
1139     auto ActiveMLocIt = ActiveMLocs.find(MLoc);
1140     if (ActiveMLocIt == ActiveMLocs.end())
1141       return;
1142 
1143     VarLocs[MLoc.asU64()] = ValueIDNum::EmptyValue;
1144 
1145     for (auto &Var : ActiveMLocIt->second) {
1146       auto ActiveVLocIt = ActiveVLocs.find(Var);
1147       // Create an undef. We can't feed in a nullptr DIExpression alas,
1148       // so use the variables last expression. Pass None as the location.
1149       const DIExpression *Expr = ActiveVLocIt->second.Properties.DIExpr;
1150       DbgValueProperties Properties(Expr, false);
1151       PendingDbgValues.push_back(MTracker->emitLoc(None, Var, Properties));
1152       ActiveVLocs.erase(ActiveVLocIt);
1153     }
1154     flushDbgValues(Pos, nullptr);
1155 
1156     ActiveMLocIt->second.clear();
1157   }
1158 
1159   /// Transfer variables based on \p Src to be based on \p Dst. This handles
1160   /// both register copies as well as spills and restores. Creates DBG_VALUEs
1161   /// describing the movement.
1162   void transferMlocs(LocIdx Src, LocIdx Dst, MachineBasicBlock::iterator Pos) {
1163     // Does Src still contain the value num we expect? If not, it's been
1164     // clobbered in the meantime, and our variable locations are stale.
1165     if (VarLocs[Src.asU64()] != MTracker->getNumAtPos(Src))
1166       return;
1167 
1168     // assert(ActiveMLocs[Dst].size() == 0);
1169     //^^^ Legitimate scenario on account of un-clobbered slot being assigned to?
1170     ActiveMLocs[Dst] = ActiveMLocs[Src];
1171     VarLocs[Dst.asU64()] = VarLocs[Src.asU64()];
1172 
1173     // For each variable based on Src; create a location at Dst.
1174     for (auto &Var : ActiveMLocs[Src]) {
1175       auto ActiveVLocIt = ActiveVLocs.find(Var);
1176       assert(ActiveVLocIt != ActiveVLocs.end());
1177       ActiveVLocIt->second.Loc = Dst;
1178 
1179       assert(Dst != 0);
1180       MachineInstr *MI =
1181           MTracker->emitLoc(Dst, Var, ActiveVLocIt->second.Properties);
1182       PendingDbgValues.push_back(MI);
1183     }
1184     ActiveMLocs[Src].clear();
1185     flushDbgValues(Pos, nullptr);
1186 
1187     // XXX XXX XXX "pretend to be old LDV" means dropping all tracking data
1188     // about the old location.
1189     if (EmulateOldLDV)
1190       VarLocs[Src.asU64()] = ValueIDNum::EmptyValue;
1191   }
1192 
1193   MachineInstrBuilder emitMOLoc(const MachineOperand &MO,
1194                                 const DebugVariable &Var,
1195                                 const DbgValueProperties &Properties) {
1196     DebugLoc DL =
1197         DebugLoc::get(0, 0, Var.getVariable()->getScope(), Var.getInlinedAt());
1198     auto MIB = BuildMI(MF, DL, TII->get(TargetOpcode::DBG_VALUE));
1199     MIB.add(MO);
1200     if (Properties.Indirect)
1201       MIB.addImm(0);
1202     else
1203       MIB.addReg(0);
1204     MIB.addMetadata(Var.getVariable());
1205     MIB.addMetadata(Properties.DIExpr);
1206     return MIB;
1207   }
1208 };
1209 
1210 class InstrRefBasedLDV : public LDVImpl {
1211 private:
1212   using FragmentInfo = DIExpression::FragmentInfo;
1213   using OptFragmentInfo = Optional<DIExpression::FragmentInfo>;
1214 
1215   // Helper while building OverlapMap, a map of all fragments seen for a given
1216   // DILocalVariable.
1217   using VarToFragments =
1218       DenseMap<const DILocalVariable *, SmallSet<FragmentInfo, 4>>;
1219 
1220   /// Machine location/value transfer function, a mapping of which locations
1221   // are assigned which new values.
1222   typedef std::map<LocIdx, ValueIDNum> MLocTransferMap;
1223 
1224   /// Live in/out structure for the variable values: a per-block map of
1225   /// variables to their values. XXX, better name?
1226   typedef DenseMap<const MachineBasicBlock *,
1227                    DenseMap<DebugVariable, DbgValue> *>
1228       LiveIdxT;
1229 
1230   typedef std::pair<DebugVariable, DbgValue> VarAndLoc;
1231 
1232   /// Type for a live-in value: the predecessor block, and its value.
1233   typedef std::pair<MachineBasicBlock *, DbgValue *> InValueT;
1234 
1235   /// Vector (per block) of a collection (inner smallvector) of live-ins.
1236   /// Used as the result type for the variable value dataflow problem.
1237   typedef SmallVector<SmallVector<VarAndLoc, 8>, 8> LiveInsT;
1238 
1239   const TargetRegisterInfo *TRI;
1240   const TargetInstrInfo *TII;
1241   const TargetFrameLowering *TFI;
1242   BitVector CalleeSavedRegs;
1243   LexicalScopes LS;
1244   TargetPassConfig *TPC;
1245 
1246   /// Object to track machine locations as we step through a block. Could
1247   /// probably be a field rather than a pointer, as it's always used.
1248   MLocTracker *MTracker;
1249 
1250   /// Number of the current block LiveDebugValues is stepping through.
1251   unsigned CurBB;
1252 
1253   /// Number of the current instruction LiveDebugValues is evaluating.
1254   unsigned CurInst;
1255 
1256   /// Variable tracker -- listens to DBG_VALUEs occurring as InstrRefBasedImpl
1257   /// steps through a block. Reads the values at each location from the
1258   /// MLocTracker object.
1259   VLocTracker *VTracker;
1260 
1261   /// Tracker for transfers, listens to DBG_VALUEs and transfers of values
1262   /// between locations during stepping, creates new DBG_VALUEs when values move
1263   /// location.
1264   TransferTracker *TTracker;
1265 
1266   /// Blocks which are artificial, i.e. blocks which exclusively contain
1267   /// instructions without DebugLocs, or with line 0 locations.
1268   SmallPtrSet<const MachineBasicBlock *, 16> ArtificialBlocks;
1269 
1270   // Mapping of blocks to and from their RPOT order.
1271   DenseMap<unsigned int, MachineBasicBlock *> OrderToBB;
1272   DenseMap<MachineBasicBlock *, unsigned int> BBToOrder;
1273   DenseMap<unsigned, unsigned> BBNumToRPO;
1274 
1275   // Map of overlapping variable fragments.
1276   OverlapMap OverlapFragments;
1277   VarToFragments SeenFragments;
1278 
1279   /// Tests whether this instruction is a spill to a stack slot.
1280   bool isSpillInstruction(const MachineInstr &MI, MachineFunction *MF);
1281 
1282   /// Decide if @MI is a spill instruction and return true if it is. We use 2
1283   /// criteria to make this decision:
1284   /// - Is this instruction a store to a spill slot?
1285   /// - Is there a register operand that is both used and killed?
1286   /// TODO: Store optimization can fold spills into other stores (including
1287   /// other spills). We do not handle this yet (more than one memory operand).
1288   bool isLocationSpill(const MachineInstr &MI, MachineFunction *MF,
1289                        unsigned &Reg);
1290 
1291   /// If a given instruction is identified as a spill, return the spill slot
1292   /// and set \p Reg to the spilled register.
1293   Optional<SpillLoc> isRestoreInstruction(const MachineInstr &MI,
1294                                           MachineFunction *MF, unsigned &Reg);
1295 
1296   /// Given a spill instruction, extract the register and offset used to
1297   /// address the spill slot in a target independent way.
1298   SpillLoc extractSpillBaseRegAndOffset(const MachineInstr &MI);
1299 
1300   /// Observe a single instruction while stepping through a block.
1301   void process(MachineInstr &MI);
1302 
1303   /// Examines whether \p MI is a DBG_VALUE and notifies trackers.
1304   /// \returns true if MI was recognized and processed.
1305   bool transferDebugValue(const MachineInstr &MI);
1306 
1307   /// Examines whether \p MI is copy instruction, and notifies trackers.
1308   /// \returns true if MI was recognized and processed.
1309   bool transferRegisterCopy(MachineInstr &MI);
1310 
1311   /// Examines whether \p MI is stack spill or restore  instruction, and
1312   /// notifies trackers. \returns true if MI was recognized and processed.
1313   bool transferSpillOrRestoreInst(MachineInstr &MI);
1314 
1315   /// Examines \p MI for any registers that it defines, and notifies trackers.
1316   /// \returns true if MI was recognized and processed.
1317   void transferRegisterDef(MachineInstr &MI);
1318 
1319   /// Copy one location to the other, accounting for movement of subregisters
1320   /// too.
1321   void performCopy(Register Src, Register Dst);
1322 
1323   void accumulateFragmentMap(MachineInstr &MI);
1324 
1325   /// Step through the function, recording register definitions and movements
1326   /// in an MLocTracker. Convert the observations into a per-block transfer
1327   /// function in \p MLocTransfer, suitable for using with the machine value
1328   /// location dataflow problem. Do the same with VLoc trackers in \p VLocs,
1329   /// although the precise machine value numbers can't be known until after
1330   /// the machine value number problem is solved.
1331   void produceTransferFunctions(MachineFunction &MF,
1332                                 SmallVectorImpl<MLocTransferMap> &MLocTransfer,
1333                                 unsigned MaxNumBlocks,
1334                                 SmallVectorImpl<VLocTracker> &VLocs);
1335 
1336   /// Solve the machine value location dataflow problem. Takes as input the
1337   /// transfer functions in \p MLocTransfer. Writes the output live-in and
1338   /// live-out arrays to the (initialized to zero) multidimensional arrays in
1339   /// \p MInLocs and \p MOutLocs. The outer dimension is indexed by block
1340   /// number, the inner by LocIdx.
1341   void mlocDataflow(ValueIDNum **MInLocs, ValueIDNum **MOutLocs,
1342                     SmallVectorImpl<MLocTransferMap> &MLocTransfer);
1343 
1344   /// Perform a control flow join (lattice value meet) of the values in machine
1345   /// locations at \p MBB. Follows the algorithm described in the file-comment,
1346   /// reading live-outs of predecessors from \p OutLocs, the current live ins
1347   /// from \p InLocs, and assigning the newly computed live ins back into
1348   /// \p InLocs. \returns two bools -- the first indicates whether a change
1349   /// was made, the second whether a lattice downgrade occurred. If the latter
1350   /// is true, revisiting this block is necessary.
1351   std::tuple<bool, bool>
1352   mlocJoin(MachineBasicBlock &MBB,
1353            SmallPtrSet<const MachineBasicBlock *, 16> &Visited,
1354            ValueIDNum **OutLocs, ValueIDNum *InLocs);
1355 
1356   /// Solve the variable value dataflow problem, for a single lexical scope.
1357   /// Uses the algorithm from the file comment to resolve control flow joins,
1358   /// although there are extra hacks, see vlocJoin. Reads the
1359   /// locations of values from the \p MInLocs and \p MOutLocs arrays (see
1360   /// mlocDataflow) and reads the variable values transfer function from
1361   /// \p AllTheVlocs. Live-in and Live-out variable values are stored locally,
1362   /// with the live-ins permanently stored to \p Output once the fixedpoint is
1363   /// reached.
1364   /// \p VarsWeCareAbout contains a collection of the variables in \p Scope
1365   /// that we should be tracking.
1366   /// \p AssignBlocks contains the set of blocks that aren't in \p Scope, but
1367   /// which do contain DBG_VALUEs, which VarLocBasedImpl tracks locations
1368   /// through.
1369   void vlocDataflow(const LexicalScope *Scope, const DILocation *DILoc,
1370                     const SmallSet<DebugVariable, 4> &VarsWeCareAbout,
1371                     SmallPtrSetImpl<MachineBasicBlock *> &AssignBlocks,
1372                     LiveInsT &Output, ValueIDNum **MOutLocs,
1373                     ValueIDNum **MInLocs,
1374                     SmallVectorImpl<VLocTracker> &AllTheVLocs);
1375 
1376   /// Compute the live-ins to a block, considering control flow merges according
1377   /// to the method in the file comment. Live out and live in variable values
1378   /// are stored in \p VLOCOutLocs and \p VLOCInLocs. The live-ins for \p MBB
1379   /// are computed and stored into \p VLOCInLocs. \returns true if the live-ins
1380   /// are modified.
1381   /// \p InLocsT Output argument, storage for calculated live-ins.
1382   /// \returns two bools -- the first indicates whether a change
1383   /// was made, the second whether a lattice downgrade occurred. If the latter
1384   /// is true, revisiting this block is necessary.
1385   std::tuple<bool, bool>
1386   vlocJoin(MachineBasicBlock &MBB, LiveIdxT &VLOCOutLocs, LiveIdxT &VLOCInLocs,
1387            SmallPtrSet<const MachineBasicBlock *, 16> *VLOCVisited,
1388            unsigned BBNum, const SmallSet<DebugVariable, 4> &AllVars,
1389            ValueIDNum **MOutLocs, ValueIDNum **MInLocs,
1390            SmallPtrSet<const MachineBasicBlock *, 8> &InScopeBlocks,
1391            SmallPtrSet<const MachineBasicBlock *, 8> &BlocksToExplore,
1392            DenseMap<DebugVariable, DbgValue> &InLocsT);
1393 
1394   /// Continue exploration of the variable-value lattice, as explained in the
1395   /// file-level comment. \p OldLiveInLocation contains the current
1396   /// exploration position, from which we need to descend further. \p Values
1397   /// contains the set of live-in values, \p CurBlockRPONum the RPO number of
1398   /// the current block, and \p CandidateLocations a set of locations that
1399   /// should be considered as PHI locations, if we reach the bottom of the
1400   /// lattice. \returns true if we should downgrade; the value is the agreeing
1401   /// value number in a non-backedge predecessor.
1402   bool vlocDowngradeLattice(const MachineBasicBlock &MBB,
1403                             const DbgValue &OldLiveInLocation,
1404                             const SmallVectorImpl<InValueT> &Values,
1405                             unsigned CurBlockRPONum);
1406 
1407   /// For the given block and live-outs feeding into it, try to find a
1408   /// machine location where they all join. If a solution for all predecessors
1409   /// can't be found, a location where all non-backedge-predecessors join
1410   /// will be returned instead. While this method finds a join location, this
1411   /// says nothing as to whether it should be used.
1412   /// \returns Pair of value ID if found, and true when the correct value
1413   /// is available on all predecessor edges, or false if it's only available
1414   /// for non-backedge predecessors.
1415   std::tuple<Optional<ValueIDNum>, bool>
1416   pickVPHILoc(MachineBasicBlock &MBB, const DebugVariable &Var,
1417               const LiveIdxT &LiveOuts, ValueIDNum **MOutLocs,
1418               ValueIDNum **MInLocs,
1419               const SmallVectorImpl<MachineBasicBlock *> &BlockOrders);
1420 
1421   /// Given the solutions to the two dataflow problems, machine value locations
1422   /// in \p MInLocs and live-in variable values in \p SavedLiveIns, runs the
1423   /// TransferTracker class over the function to produce live-in and transfer
1424   /// DBG_VALUEs, then inserts them. Groups of DBG_VALUEs are inserted in the
1425   /// order given by AllVarsNumbering -- this could be any stable order, but
1426   /// right now "order of appearence in function, when explored in RPO", so
1427   /// that we can compare explictly against VarLocBasedImpl.
1428   void emitLocations(MachineFunction &MF, LiveInsT SavedLiveIns,
1429                      ValueIDNum **MInLocs,
1430                      DenseMap<DebugVariable, unsigned> &AllVarsNumbering);
1431 
1432   /// Boilerplate computation of some initial sets, artifical blocks and
1433   /// RPOT block ordering.
1434   void initialSetup(MachineFunction &MF);
1435 
1436   bool ExtendRanges(MachineFunction &MF, TargetPassConfig *TPC) override;
1437 
1438 public:
1439   /// Default construct and initialize the pass.
1440   InstrRefBasedLDV();
1441 
1442   LLVM_DUMP_METHOD
1443   void dump_mloc_transfer(const MLocTransferMap &mloc_transfer) const;
1444 
1445   bool isCalleeSaved(LocIdx L) {
1446     unsigned Reg = MTracker->LocIdxToLocID[L];
1447     for (MCRegAliasIterator RAI(Reg, TRI, true); RAI.isValid(); ++RAI)
1448       if (CalleeSavedRegs.test(*RAI))
1449         return true;
1450     return false;
1451   }
1452 };
1453 
1454 } // end anonymous namespace
1455 
1456 //===----------------------------------------------------------------------===//
1457 //            Implementation
1458 //===----------------------------------------------------------------------===//
1459 
1460 ValueIDNum ValueIDNum::EmptyValue = {UINT_MAX, UINT_MAX, UINT_MAX};
1461 
1462 /// Default construct and initialize the pass.
1463 InstrRefBasedLDV::InstrRefBasedLDV() {}
1464 
1465 //===----------------------------------------------------------------------===//
1466 //            Debug Range Extension Implementation
1467 //===----------------------------------------------------------------------===//
1468 
1469 #ifndef NDEBUG
1470 // Something to restore in the future.
1471 // void InstrRefBasedLDV::printVarLocInMBB(..)
1472 #endif
1473 
1474 SpillLoc
1475 InstrRefBasedLDV::extractSpillBaseRegAndOffset(const MachineInstr &MI) {
1476   assert(MI.hasOneMemOperand() &&
1477          "Spill instruction does not have exactly one memory operand?");
1478   auto MMOI = MI.memoperands_begin();
1479   const PseudoSourceValue *PVal = (*MMOI)->getPseudoValue();
1480   assert(PVal->kind() == PseudoSourceValue::FixedStack &&
1481          "Inconsistent memory operand in spill instruction");
1482   int FI = cast<FixedStackPseudoSourceValue>(PVal)->getFrameIndex();
1483   const MachineBasicBlock *MBB = MI.getParent();
1484   Register Reg;
1485   int Offset = TFI->getFrameIndexReference(*MBB->getParent(), FI, Reg);
1486   return {Reg, Offset};
1487 }
1488 
1489 /// End all previous ranges related to @MI and start a new range from @MI
1490 /// if it is a DBG_VALUE instr.
1491 bool InstrRefBasedLDV::transferDebugValue(const MachineInstr &MI) {
1492   if (!MI.isDebugValue())
1493     return false;
1494 
1495   const DILocalVariable *Var = MI.getDebugVariable();
1496   const DIExpression *Expr = MI.getDebugExpression();
1497   const DILocation *DebugLoc = MI.getDebugLoc();
1498   const DILocation *InlinedAt = DebugLoc->getInlinedAt();
1499   assert(Var->isValidLocationForIntrinsic(DebugLoc) &&
1500          "Expected inlined-at fields to agree");
1501 
1502   DebugVariable V(Var, Expr, InlinedAt);
1503 
1504   // If there are no instructions in this lexical scope, do no location tracking
1505   // at all, this variable shouldn't get a legitimate location range.
1506   auto *Scope = LS.findLexicalScope(MI.getDebugLoc().get());
1507   if (Scope == nullptr)
1508     return true; // handled it; by doing nothing
1509 
1510   const MachineOperand &MO = MI.getOperand(0);
1511 
1512   // MLocTracker needs to know that this register is read, even if it's only
1513   // read by a debug inst.
1514   if (MO.isReg() && MO.getReg() != 0)
1515     (void)MTracker->readReg(MO.getReg());
1516 
1517   // If we're preparing for the second analysis (variables), the machine value
1518   // locations are already solved, and we report this DBG_VALUE and the value
1519   // it refers to to VLocTracker.
1520   if (VTracker) {
1521     if (MO.isReg()) {
1522       // Feed defVar the new variable location, or if this is a
1523       // DBG_VALUE $noreg, feed defVar None.
1524       if (MO.getReg())
1525         VTracker->defVar(MI, MTracker->readReg(MO.getReg()));
1526       else
1527         VTracker->defVar(MI, None);
1528     } else if (MI.getOperand(0).isImm() || MI.getOperand(0).isFPImm() ||
1529                MI.getOperand(0).isCImm()) {
1530       VTracker->defVar(MI, MI.getOperand(0));
1531     }
1532   }
1533 
1534   // If performing final tracking of transfers, report this variable definition
1535   // to the TransferTracker too.
1536   if (TTracker)
1537     TTracker->redefVar(MI);
1538   return true;
1539 }
1540 
1541 void InstrRefBasedLDV::transferRegisterDef(MachineInstr &MI) {
1542   // Meta Instructions do not affect the debug liveness of any register they
1543   // define.
1544   if (MI.isImplicitDef()) {
1545     // Except when there's an implicit def, and the location it's defining has
1546     // no value number. The whole point of an implicit def is to announce that
1547     // the register is live, without be specific about it's value. So define
1548     // a value if there isn't one already.
1549     ValueIDNum Num = MTracker->readReg(MI.getOperand(0).getReg());
1550     // Has a legitimate value -> ignore the implicit def.
1551     if (Num.getLoc() != 0)
1552       return;
1553     // Otherwise, def it here.
1554   } else if (MI.isMetaInstruction())
1555     return;
1556 
1557   MachineFunction *MF = MI.getMF();
1558   const TargetLowering *TLI = MF->getSubtarget().getTargetLowering();
1559   unsigned SP = TLI->getStackPointerRegisterToSaveRestore();
1560 
1561   // Find the regs killed by MI, and find regmasks of preserved regs.
1562   // Max out the number of statically allocated elements in `DeadRegs`, as this
1563   // prevents fallback to std::set::count() operations.
1564   SmallSet<uint32_t, 32> DeadRegs;
1565   SmallVector<const uint32_t *, 4> RegMasks;
1566   SmallVector<const MachineOperand *, 4> RegMaskPtrs;
1567   for (const MachineOperand &MO : MI.operands()) {
1568     // Determine whether the operand is a register def.
1569     if (MO.isReg() && MO.isDef() && MO.getReg() &&
1570         Register::isPhysicalRegister(MO.getReg()) &&
1571         !(MI.isCall() && MO.getReg() == SP)) {
1572       // Remove ranges of all aliased registers.
1573       for (MCRegAliasIterator RAI(MO.getReg(), TRI, true); RAI.isValid(); ++RAI)
1574         // FIXME: Can we break out of this loop early if no insertion occurs?
1575         DeadRegs.insert(*RAI);
1576     } else if (MO.isRegMask()) {
1577       RegMasks.push_back(MO.getRegMask());
1578       RegMaskPtrs.push_back(&MO);
1579     }
1580   }
1581 
1582   // Tell MLocTracker about all definitions, of regmasks and otherwise.
1583   for (uint32_t DeadReg : DeadRegs)
1584     MTracker->defReg(DeadReg, CurBB, CurInst);
1585 
1586   for (auto *MO : RegMaskPtrs)
1587     MTracker->writeRegMask(MO, CurBB, CurInst);
1588 }
1589 
1590 void InstrRefBasedLDV::performCopy(Register SrcRegNum, Register DstRegNum) {
1591   ValueIDNum SrcValue = MTracker->readReg(SrcRegNum);
1592 
1593   MTracker->setReg(DstRegNum, SrcValue);
1594 
1595   // In all circumstances, re-def the super registers. It's definitely a new
1596   // value now. This doesn't uniquely identify the composition of subregs, for
1597   // example, two identical values in subregisters composed in different
1598   // places would not get equal value numbers.
1599   for (MCSuperRegIterator SRI(DstRegNum, TRI); SRI.isValid(); ++SRI)
1600     MTracker->defReg(*SRI, CurBB, CurInst);
1601 
1602   // If we're emulating VarLocBasedImpl, just define all the subregisters.
1603   // DBG_VALUEs of them will expect to be tracked from the DBG_VALUE, not
1604   // through prior copies.
1605   if (EmulateOldLDV) {
1606     for (MCSubRegIndexIterator DRI(DstRegNum, TRI); DRI.isValid(); ++DRI)
1607       MTracker->defReg(DRI.getSubReg(), CurBB, CurInst);
1608     return;
1609   }
1610 
1611   // Otherwise, actually copy subregisters from one location to another.
1612   // XXX: in addition, any subregisters of DstRegNum that don't line up with
1613   // the source register should be def'd.
1614   for (MCSubRegIndexIterator SRI(SrcRegNum, TRI); SRI.isValid(); ++SRI) {
1615     unsigned SrcSubReg = SRI.getSubReg();
1616     unsigned SubRegIdx = SRI.getSubRegIndex();
1617     unsigned DstSubReg = TRI->getSubReg(DstRegNum, SubRegIdx);
1618     if (!DstSubReg)
1619       continue;
1620 
1621     // Do copy. There are two matching subregisters, the source value should
1622     // have been def'd when the super-reg was, the latter might not be tracked
1623     // yet.
1624     // This will force SrcSubReg to be tracked, if it isn't yet.
1625     (void)MTracker->readReg(SrcSubReg);
1626     LocIdx SrcL = MTracker->getRegMLoc(SrcSubReg);
1627     assert(SrcL.asU64());
1628     (void)MTracker->readReg(DstSubReg);
1629     LocIdx DstL = MTracker->getRegMLoc(DstSubReg);
1630     assert(DstL.asU64());
1631     (void)DstL;
1632     ValueIDNum CpyValue = {SrcValue.getBlock(), SrcValue.getInst(), SrcL};
1633 
1634     MTracker->setReg(DstSubReg, CpyValue);
1635   }
1636 }
1637 
1638 bool InstrRefBasedLDV::isSpillInstruction(const MachineInstr &MI,
1639                                           MachineFunction *MF) {
1640   // TODO: Handle multiple stores folded into one.
1641   if (!MI.hasOneMemOperand())
1642     return false;
1643 
1644   if (!MI.getSpillSize(TII) && !MI.getFoldedSpillSize(TII))
1645     return false; // This is not a spill instruction, since no valid size was
1646                   // returned from either function.
1647 
1648   return true;
1649 }
1650 
1651 bool InstrRefBasedLDV::isLocationSpill(const MachineInstr &MI,
1652                                        MachineFunction *MF, unsigned &Reg) {
1653   if (!isSpillInstruction(MI, MF))
1654     return false;
1655 
1656   // XXX FIXME: On x86, isStoreToStackSlotPostFE returns '1' instead of an
1657   // actual register number.
1658   if (ObserveAllStackops) {
1659     int FI;
1660     Reg = TII->isStoreToStackSlotPostFE(MI, FI);
1661     return Reg != 0;
1662   }
1663 
1664   auto isKilledReg = [&](const MachineOperand MO, unsigned &Reg) {
1665     if (!MO.isReg() || !MO.isUse()) {
1666       Reg = 0;
1667       return false;
1668     }
1669     Reg = MO.getReg();
1670     return MO.isKill();
1671   };
1672 
1673   for (const MachineOperand &MO : MI.operands()) {
1674     // In a spill instruction generated by the InlineSpiller the spilled
1675     // register has its kill flag set.
1676     if (isKilledReg(MO, Reg))
1677       return true;
1678     if (Reg != 0) {
1679       // Check whether next instruction kills the spilled register.
1680       // FIXME: Current solution does not cover search for killed register in
1681       // bundles and instructions further down the chain.
1682       auto NextI = std::next(MI.getIterator());
1683       // Skip next instruction that points to basic block end iterator.
1684       if (MI.getParent()->end() == NextI)
1685         continue;
1686       unsigned RegNext;
1687       for (const MachineOperand &MONext : NextI->operands()) {
1688         // Return true if we came across the register from the
1689         // previous spill instruction that is killed in NextI.
1690         if (isKilledReg(MONext, RegNext) && RegNext == Reg)
1691           return true;
1692       }
1693     }
1694   }
1695   // Return false if we didn't find spilled register.
1696   return false;
1697 }
1698 
1699 Optional<SpillLoc>
1700 InstrRefBasedLDV::isRestoreInstruction(const MachineInstr &MI,
1701                                        MachineFunction *MF, unsigned &Reg) {
1702   if (!MI.hasOneMemOperand())
1703     return None;
1704 
1705   // FIXME: Handle folded restore instructions with more than one memory
1706   // operand.
1707   if (MI.getRestoreSize(TII)) {
1708     Reg = MI.getOperand(0).getReg();
1709     return extractSpillBaseRegAndOffset(MI);
1710   }
1711   return None;
1712 }
1713 
1714 bool InstrRefBasedLDV::transferSpillOrRestoreInst(MachineInstr &MI) {
1715   // XXX -- it's too difficult to implement VarLocBasedImpl's  stack location
1716   // limitations under the new model. Therefore, when comparing them, compare
1717   // versions that don't attempt spills or restores at all.
1718   if (EmulateOldLDV)
1719     return false;
1720 
1721   MachineFunction *MF = MI.getMF();
1722   unsigned Reg;
1723   Optional<SpillLoc> Loc;
1724 
1725   LLVM_DEBUG(dbgs() << "Examining instruction: "; MI.dump(););
1726 
1727   // First, if there are any DBG_VALUEs pointing at a spill slot that is
1728   // written to, terminate that variable location. The value in memory
1729   // will have changed. DbgEntityHistoryCalculator doesn't try to detect this.
1730   if (isSpillInstruction(MI, MF)) {
1731     Loc = extractSpillBaseRegAndOffset(MI);
1732 
1733     if (TTracker) {
1734       Optional<LocIdx> MLoc = MTracker->getSpillMLoc(*Loc);
1735       if (MLoc)
1736         TTracker->clobberMloc(*MLoc, MI.getIterator());
1737     }
1738   }
1739 
1740   // Try to recognise spill and restore instructions that may transfer a value.
1741   if (isLocationSpill(MI, MF, Reg)) {
1742     Loc = extractSpillBaseRegAndOffset(MI);
1743     auto ValueID = MTracker->readReg(Reg);
1744 
1745     // If the location is empty, produce a phi, signify it's the live-in value.
1746     if (ValueID.getLoc() == 0)
1747       ValueID = {CurBB, 0, MTracker->getRegMLoc(Reg)};
1748 
1749     MTracker->setSpill(*Loc, ValueID);
1750     auto OptSpillLocIdx = MTracker->getSpillMLoc(*Loc);
1751     assert(OptSpillLocIdx && "Spill slot set but has no LocIdx?");
1752     LocIdx SpillLocIdx = *OptSpillLocIdx;
1753 
1754     // Tell TransferTracker about this spill, produce DBG_VALUEs for it.
1755     if (TTracker)
1756       TTracker->transferMlocs(MTracker->getRegMLoc(Reg), SpillLocIdx,
1757                               MI.getIterator());
1758 
1759     // VarLocBasedImpl would, at this point, stop tracking the source
1760     // register of the store.
1761     if (EmulateOldLDV) {
1762       for (MCRegAliasIterator RAI(Reg, TRI, true); RAI.isValid(); ++RAI)
1763         MTracker->defReg(*RAI, CurBB, CurInst);
1764     }
1765   } else {
1766     if (!(Loc = isRestoreInstruction(MI, MF, Reg)))
1767       return false;
1768 
1769     // Is there a value to be restored?
1770     auto OptValueID = MTracker->readSpill(*Loc);
1771     if (OptValueID) {
1772       ValueIDNum ValueID = *OptValueID;
1773       LocIdx SpillLocIdx = *MTracker->getSpillMLoc(*Loc);
1774       // XXX -- can we recover sub-registers of this value? Until we can, first
1775       // overwrite all defs of the register being restored to.
1776       for (MCRegAliasIterator RAI(Reg, TRI, true); RAI.isValid(); ++RAI)
1777         MTracker->defReg(*RAI, CurBB, CurInst);
1778 
1779       // Now override the reg we're restoring to.
1780       MTracker->setReg(Reg, ValueID);
1781 
1782       // Report this restore to the transfer tracker too.
1783       if (TTracker)
1784         TTracker->transferMlocs(SpillLocIdx, MTracker->getRegMLoc(Reg),
1785                                 MI.getIterator());
1786     } else {
1787       // There isn't anything in the location; not clear if this is a code path
1788       // that still runs. Def this register anyway just in case.
1789       for (MCRegAliasIterator RAI(Reg, TRI, true); RAI.isValid(); ++RAI)
1790         MTracker->defReg(*RAI, CurBB, CurInst);
1791 
1792       // Force the spill slot to be tracked.
1793       LocIdx L = MTracker->getOrTrackSpillLoc(*Loc);
1794 
1795       // Set the restored value to be a machine phi number, signifying that it's
1796       // whatever the spills live-in value is in this block. Definitely has
1797       // a LocIdx due to the setSpill above.
1798       ValueIDNum ValueID = {CurBB, 0, L};
1799       MTracker->setReg(Reg, ValueID);
1800       MTracker->setSpill(*Loc, ValueID);
1801     }
1802   }
1803   return true;
1804 }
1805 
1806 bool InstrRefBasedLDV::transferRegisterCopy(MachineInstr &MI) {
1807   auto DestSrc = TII->isCopyInstr(MI);
1808   if (!DestSrc)
1809     return false;
1810 
1811   const MachineOperand *DestRegOp = DestSrc->Destination;
1812   const MachineOperand *SrcRegOp = DestSrc->Source;
1813 
1814   auto isCalleeSavedReg = [&](unsigned Reg) {
1815     for (MCRegAliasIterator RAI(Reg, TRI, true); RAI.isValid(); ++RAI)
1816       if (CalleeSavedRegs.test(*RAI))
1817         return true;
1818     return false;
1819   };
1820 
1821   Register SrcReg = SrcRegOp->getReg();
1822   Register DestReg = DestRegOp->getReg();
1823 
1824   // Ignore identity copies. Yep, these make it as far as LiveDebugValues.
1825   if (SrcReg == DestReg)
1826     return true;
1827 
1828   // For emulating VarLocBasedImpl:
1829   // We want to recognize instructions where destination register is callee
1830   // saved register. If register that could be clobbered by the call is
1831   // included, there would be a great chance that it is going to be clobbered
1832   // soon. It is more likely that previous register, which is callee saved, is
1833   // going to stay unclobbered longer, even if it is killed.
1834   //
1835   // For InstrRefBasedImpl, we can track multiple locations per value, so
1836   // ignore this condition.
1837   if (EmulateOldLDV && !isCalleeSavedReg(DestReg))
1838     return false;
1839 
1840   // InstrRefBasedImpl only followed killing copies.
1841   if (EmulateOldLDV && !SrcRegOp->isKill())
1842     return false;
1843 
1844   // Copy MTracker info, including subregs if available.
1845   InstrRefBasedLDV::performCopy(SrcReg, DestReg);
1846 
1847   // Only produce a transfer of DBG_VALUE within a block where old LDV
1848   // would have. We might make use of the additional value tracking in some
1849   // other way, later.
1850   if (TTracker && isCalleeSavedReg(DestReg) && SrcRegOp->isKill())
1851     TTracker->transferMlocs(MTracker->getRegMLoc(SrcReg),
1852                             MTracker->getRegMLoc(DestReg), MI.getIterator());
1853 
1854   // VarLocBasedImpl would quit tracking the old location after copying.
1855   if (EmulateOldLDV && SrcReg != DestReg)
1856     MTracker->defReg(SrcReg, CurBB, CurInst);
1857 
1858   return true;
1859 }
1860 
1861 /// Accumulate a mapping between each DILocalVariable fragment and other
1862 /// fragments of that DILocalVariable which overlap. This reduces work during
1863 /// the data-flow stage from "Find any overlapping fragments" to "Check if the
1864 /// known-to-overlap fragments are present".
1865 /// \param MI A previously unprocessed DEBUG_VALUE instruction to analyze for
1866 ///           fragment usage.
1867 void InstrRefBasedLDV::accumulateFragmentMap(MachineInstr &MI) {
1868   DebugVariable MIVar(MI.getDebugVariable(), MI.getDebugExpression(),
1869                       MI.getDebugLoc()->getInlinedAt());
1870   FragmentInfo ThisFragment = MIVar.getFragmentOrDefault();
1871 
1872   // If this is the first sighting of this variable, then we are guaranteed
1873   // there are currently no overlapping fragments either. Initialize the set
1874   // of seen fragments, record no overlaps for the current one, and return.
1875   auto SeenIt = SeenFragments.find(MIVar.getVariable());
1876   if (SeenIt == SeenFragments.end()) {
1877     SmallSet<FragmentInfo, 4> OneFragment;
1878     OneFragment.insert(ThisFragment);
1879     SeenFragments.insert({MIVar.getVariable(), OneFragment});
1880 
1881     OverlapFragments.insert({{MIVar.getVariable(), ThisFragment}, {}});
1882     return;
1883   }
1884 
1885   // If this particular Variable/Fragment pair already exists in the overlap
1886   // map, it has already been accounted for.
1887   auto IsInOLapMap =
1888       OverlapFragments.insert({{MIVar.getVariable(), ThisFragment}, {}});
1889   if (!IsInOLapMap.second)
1890     return;
1891 
1892   auto &ThisFragmentsOverlaps = IsInOLapMap.first->second;
1893   auto &AllSeenFragments = SeenIt->second;
1894 
1895   // Otherwise, examine all other seen fragments for this variable, with "this"
1896   // fragment being a previously unseen fragment. Record any pair of
1897   // overlapping fragments.
1898   for (auto &ASeenFragment : AllSeenFragments) {
1899     // Does this previously seen fragment overlap?
1900     if (DIExpression::fragmentsOverlap(ThisFragment, ASeenFragment)) {
1901       // Yes: Mark the current fragment as being overlapped.
1902       ThisFragmentsOverlaps.push_back(ASeenFragment);
1903       // Mark the previously seen fragment as being overlapped by the current
1904       // one.
1905       auto ASeenFragmentsOverlaps =
1906           OverlapFragments.find({MIVar.getVariable(), ASeenFragment});
1907       assert(ASeenFragmentsOverlaps != OverlapFragments.end() &&
1908              "Previously seen var fragment has no vector of overlaps");
1909       ASeenFragmentsOverlaps->second.push_back(ThisFragment);
1910     }
1911   }
1912 
1913   AllSeenFragments.insert(ThisFragment);
1914 }
1915 
1916 void InstrRefBasedLDV::process(MachineInstr &MI) {
1917   // Try to interpret an MI as a debug or transfer instruction. Only if it's
1918   // none of these should we interpret it's register defs as new value
1919   // definitions.
1920   if (transferDebugValue(MI))
1921     return;
1922   if (transferRegisterCopy(MI))
1923     return;
1924   if (transferSpillOrRestoreInst(MI))
1925     return;
1926   transferRegisterDef(MI);
1927 }
1928 
1929 void InstrRefBasedLDV::produceTransferFunctions(
1930     MachineFunction &MF, SmallVectorImpl<MLocTransferMap> &MLocTransfer,
1931     unsigned MaxNumBlocks, SmallVectorImpl<VLocTracker> &VLocs) {
1932   // Because we try to optimize around register mask operands by ignoring regs
1933   // that aren't currently tracked, we set up something ugly for later: RegMask
1934   // operands that are seen earlier than the first use of a register, still need
1935   // to clobber that register in the transfer function. But this information
1936   // isn't actively recorded. Instead, we track each RegMask used in each block,
1937   // and accumulated the clobbered but untracked registers in each block into
1938   // the following bitvector. Later, if new values are tracked, we can add
1939   // appropriate clobbers.
1940   SmallVector<BitVector, 32> BlockMasks;
1941   BlockMasks.resize(MaxNumBlocks);
1942 
1943   // Reserve one bit per register for the masks described above.
1944   unsigned BVWords = MachineOperand::getRegMaskSize(TRI->getNumRegs());
1945   for (auto &BV : BlockMasks)
1946     BV.resize(TRI->getNumRegs(), true);
1947 
1948   // Step through all instructions and inhale the transfer function.
1949   for (auto &MBB : MF) {
1950     // Object fields that are read by trackers to know where we are in the
1951     // function.
1952     CurBB = MBB.getNumber();
1953     CurInst = 1;
1954 
1955     // Set all machine locations to a PHI value. For transfer function
1956     // production only, this signifies the live-in value to the block.
1957     MTracker->reset();
1958     MTracker->setMPhis(CurBB);
1959 
1960     VTracker = &VLocs[CurBB];
1961     VTracker->MBB = &MBB;
1962 
1963     // Step through each instruction in this block.
1964     for (auto &MI : MBB) {
1965       process(MI);
1966       // Also accumulate fragment map.
1967       if (MI.isDebugValue())
1968         accumulateFragmentMap(MI);
1969       ++CurInst;
1970     }
1971 
1972     // Produce the transfer function, a map of machine location to new value. If
1973     // any machine location has the live-in phi value from the start of the
1974     // block, it's live-through and doesn't need recording in the transfer
1975     // function.
1976     for (auto Location : MTracker->locations()) {
1977       LocIdx Idx = Location.Idx;
1978       ValueIDNum &P = Location.Value;
1979       if (P.isPHI() && P.getLoc() == Idx.asU64())
1980         continue;
1981 
1982       // Insert-or-update.
1983       auto &TransferMap = MLocTransfer[CurBB];
1984       auto Result = TransferMap.insert(std::make_pair(Idx.asU64(), P));
1985       if (!Result.second)
1986         Result.first->second = P;
1987     }
1988 
1989     // Accumulate any bitmask operands into the clobberred reg mask for this
1990     // block.
1991     for (auto &P : MTracker->Masks) {
1992       BlockMasks[CurBB].clearBitsNotInMask(P.first->getRegMask(), BVWords);
1993     }
1994   }
1995 
1996   // Compute a bitvector of all the registers that are tracked in this block.
1997   const TargetLowering *TLI = MF.getSubtarget().getTargetLowering();
1998   unsigned SP = TLI->getStackPointerRegisterToSaveRestore();
1999   BitVector UsedRegs(TRI->getNumRegs());
2000   for (auto Location : MTracker->locations()) {
2001     unsigned ID = MTracker->LocIdxToLocID[Location.Idx];
2002     if (ID >= TRI->getNumRegs() || ID == SP)
2003       continue;
2004     UsedRegs.set(ID);
2005   }
2006 
2007   // Check that any regmask-clobber of a register that gets tracked, is not
2008   // live-through in the transfer function. It needs to be clobbered at the
2009   // very least.
2010   for (unsigned int I = 0; I < MaxNumBlocks; ++I) {
2011     BitVector &BV = BlockMasks[I];
2012     BV.flip();
2013     BV &= UsedRegs;
2014     // This produces all the bits that we clobber, but also use. Check that
2015     // they're all clobbered or at least set in the designated transfer
2016     // elem.
2017     for (unsigned Bit : BV.set_bits()) {
2018       unsigned ID = MTracker->getLocID(Bit, false);
2019       LocIdx Idx = MTracker->LocIDToLocIdx[ID];
2020       auto &TransferMap = MLocTransfer[I];
2021 
2022       // Install a value representing the fact that this location is effectively
2023       // written to in this block. As there's no reserved value, instead use
2024       // a value number that is never generated. Pick the value number for the
2025       // first instruction in the block, def'ing this location, which we know
2026       // this block never used anyway.
2027       ValueIDNum NotGeneratedNum = ValueIDNum(I, 1, Idx);
2028       auto Result =
2029         TransferMap.insert(std::make_pair(Idx.asU64(), NotGeneratedNum));
2030       if (!Result.second) {
2031         ValueIDNum &ValueID = Result.first->second;
2032         if (ValueID.getBlock() == I && ValueID.isPHI())
2033           // It was left as live-through. Set it to clobbered.
2034           ValueID = NotGeneratedNum;
2035       }
2036     }
2037   }
2038 }
2039 
2040 std::tuple<bool, bool>
2041 InstrRefBasedLDV::mlocJoin(MachineBasicBlock &MBB,
2042                            SmallPtrSet<const MachineBasicBlock *, 16> &Visited,
2043                            ValueIDNum **OutLocs, ValueIDNum *InLocs) {
2044   LLVM_DEBUG(dbgs() << "join MBB: " << MBB.getNumber() << "\n");
2045   bool Changed = false;
2046   bool DowngradeOccurred = false;
2047 
2048   // Collect predecessors that have been visited. Anything that hasn't been
2049   // visited yet is a backedge on the first iteration, and the meet of it's
2050   // lattice value for all locations will be unaffected.
2051   SmallVector<const MachineBasicBlock *, 8> BlockOrders;
2052   for (auto Pred : MBB.predecessors()) {
2053     if (Visited.count(Pred)) {
2054       BlockOrders.push_back(Pred);
2055     }
2056   }
2057 
2058   // Visit predecessors in RPOT order.
2059   auto Cmp = [&](const MachineBasicBlock *A, const MachineBasicBlock *B) {
2060     return BBToOrder.find(A)->second < BBToOrder.find(B)->second;
2061   };
2062   llvm::sort(BlockOrders.begin(), BlockOrders.end(), Cmp);
2063 
2064   // Skip entry block.
2065   if (BlockOrders.size() == 0)
2066     return std::tuple<bool, bool>(false, false);
2067 
2068   // Step through all machine locations, then look at each predecessor and
2069   // detect disagreements.
2070   unsigned ThisBlockRPO = BBToOrder.find(&MBB)->second;
2071   for (auto Location : MTracker->locations()) {
2072     LocIdx Idx = Location.Idx;
2073     // Pick out the first predecessors live-out value for this location. It's
2074     // guaranteed to be not a backedge, as we order by RPO.
2075     ValueIDNum BaseVal = OutLocs[BlockOrders[0]->getNumber()][Idx.asU64()];
2076 
2077     // Some flags for whether there's a disagreement, and whether it's a
2078     // disagreement with a backedge or not.
2079     bool Disagree = false;
2080     bool NonBackEdgeDisagree = false;
2081 
2082     // Loop around everything that wasn't 'base'.
2083     for (unsigned int I = 1; I < BlockOrders.size(); ++I) {
2084       auto *MBB = BlockOrders[I];
2085       if (BaseVal != OutLocs[MBB->getNumber()][Idx.asU64()]) {
2086         // Live-out of a predecessor disagrees with the first predecessor.
2087         Disagree = true;
2088 
2089         // Test whether it's a disagreemnt in the backedges or not.
2090         if (BBToOrder.find(MBB)->second < ThisBlockRPO) // might be self b/e
2091           NonBackEdgeDisagree = true;
2092       }
2093     }
2094 
2095     bool OverRide = false;
2096     if (Disagree && !NonBackEdgeDisagree) {
2097       // Only the backedges disagree. Consider demoting the livein
2098       // lattice value, as per the file level comment. The value we consider
2099       // demoting to is the value that the non-backedge predecessors agree on.
2100       // The order of values is that non-PHIs are \top, a PHI at this block
2101       // \bot, and phis between the two are ordered by their RPO number.
2102       // If there's no agreement, or we've already demoted to this PHI value
2103       // before, replace with a PHI value at this block.
2104 
2105       // Calculate order numbers: zero means normal def, nonzero means RPO
2106       // number.
2107       unsigned BaseBlockRPONum = BBNumToRPO[BaseVal.getBlock()] + 1;
2108       if (!BaseVal.isPHI())
2109         BaseBlockRPONum = 0;
2110 
2111       ValueIDNum &InLocID = InLocs[Idx.asU64()];
2112       unsigned InLocRPONum = BBNumToRPO[InLocID.getBlock()] + 1;
2113       if (!InLocID.isPHI())
2114         InLocRPONum = 0;
2115 
2116       // Should we ignore the disagreeing backedges, and override with the
2117       // value the other predecessors agree on (in "base")?
2118       unsigned ThisBlockRPONum = BBNumToRPO[MBB.getNumber()] + 1;
2119       if (BaseBlockRPONum > InLocRPONum && BaseBlockRPONum < ThisBlockRPONum) {
2120         // Override.
2121         OverRide = true;
2122         DowngradeOccurred = true;
2123       }
2124     }
2125     // else: if we disagree in the non-backedges, then this is definitely
2126     // a control flow merge where different values merge. Make it a PHI.
2127 
2128     // Generate a phi...
2129     ValueIDNum PHI = {(uint64_t)MBB.getNumber(), 0, Idx};
2130     ValueIDNum NewVal = (Disagree && !OverRide) ? PHI : BaseVal;
2131     if (InLocs[Idx.asU64()] != NewVal) {
2132       Changed |= true;
2133       InLocs[Idx.asU64()] = NewVal;
2134     }
2135   }
2136 
2137   // Uhhhhhh, reimplement NumInserted and NumRemoved pls.
2138   return std::tuple<bool, bool>(Changed, DowngradeOccurred);
2139 }
2140 
2141 void InstrRefBasedLDV::mlocDataflow(
2142     ValueIDNum **MInLocs, ValueIDNum **MOutLocs,
2143     SmallVectorImpl<MLocTransferMap> &MLocTransfer) {
2144   std::priority_queue<unsigned int, std::vector<unsigned int>,
2145                       std::greater<unsigned int>>
2146       Worklist, Pending;
2147 
2148   // We track what is on the current and pending worklist to avoid inserting
2149   // the same thing twice. We could avoid this with a custom priority queue,
2150   // but this is probably not worth it.
2151   SmallPtrSet<MachineBasicBlock *, 16> OnPending, OnWorklist;
2152 
2153   // Initialize worklist with every block to be visited.
2154   for (unsigned int I = 0; I < BBToOrder.size(); ++I) {
2155     Worklist.push(I);
2156     OnWorklist.insert(OrderToBB[I]);
2157   }
2158 
2159   MTracker->reset();
2160 
2161   // Set inlocs for entry block -- each as a PHI at the entry block. Represents
2162   // the incoming value to the function.
2163   MTracker->setMPhis(0);
2164   for (auto Location : MTracker->locations())
2165     MInLocs[0][Location.Idx.asU64()] = Location.Value;
2166 
2167   SmallPtrSet<const MachineBasicBlock *, 16> Visited;
2168   while (!Worklist.empty() || !Pending.empty()) {
2169     // Vector for storing the evaluated block transfer function.
2170     SmallVector<std::pair<LocIdx, ValueIDNum>, 32> ToRemap;
2171 
2172     while (!Worklist.empty()) {
2173       MachineBasicBlock *MBB = OrderToBB[Worklist.top()];
2174       CurBB = MBB->getNumber();
2175       Worklist.pop();
2176 
2177       // Join the values in all predecessor blocks.
2178       bool InLocsChanged, DowngradeOccurred;
2179       std::tie(InLocsChanged, DowngradeOccurred) =
2180           mlocJoin(*MBB, Visited, MOutLocs, MInLocs[CurBB]);
2181       InLocsChanged |= Visited.insert(MBB).second;
2182 
2183       // If a downgrade occurred, book us in for re-examination on the next
2184       // iteration.
2185       if (DowngradeOccurred && OnPending.insert(MBB).second)
2186         Pending.push(BBToOrder[MBB]);
2187 
2188       // Don't examine transfer function if we've visited this loc at least
2189       // once, and inlocs haven't changed.
2190       if (!InLocsChanged)
2191         continue;
2192 
2193       // Load the current set of live-ins into MLocTracker.
2194       MTracker->loadFromArray(MInLocs[CurBB], CurBB);
2195 
2196       // Each element of the transfer function can be a new def, or a read of
2197       // a live-in value. Evaluate each element, and store to "ToRemap".
2198       ToRemap.clear();
2199       for (auto &P : MLocTransfer[CurBB]) {
2200         if (P.second.getBlock() == CurBB && P.second.isPHI()) {
2201           // This is a movement of whatever was live in. Read it.
2202           ValueIDNum NewID = MTracker->getNumAtPos(P.second.getLoc());
2203           ToRemap.push_back(std::make_pair(P.first, NewID));
2204         } else {
2205           // It's a def. Just set it.
2206           assert(P.second.getBlock() == CurBB);
2207           ToRemap.push_back(std::make_pair(P.first, P.second));
2208         }
2209       }
2210 
2211       // Commit the transfer function changes into mloc tracker, which
2212       // transforms the contents of the MLocTracker into the live-outs.
2213       for (auto &P : ToRemap)
2214         MTracker->setMLoc(P.first, P.second);
2215 
2216       // Now copy out-locs from mloc tracker into out-loc vector, checking
2217       // whether changes have occurred. These changes can have come from both
2218       // the transfer function, and mlocJoin.
2219       bool OLChanged = false;
2220       for (auto Location : MTracker->locations()) {
2221         OLChanged |= MOutLocs[CurBB][Location.Idx.asU64()] != Location.Value;
2222         MOutLocs[CurBB][Location.Idx.asU64()] = Location.Value;
2223       }
2224 
2225       MTracker->reset();
2226 
2227       // No need to examine successors again if out-locs didn't change.
2228       if (!OLChanged)
2229         continue;
2230 
2231       // All successors should be visited: put any back-edges on the pending
2232       // list for the next dataflow iteration, and any other successors to be
2233       // visited this iteration, if they're not going to be already.
2234       for (auto s : MBB->successors()) {
2235         // Does branching to this successor represent a back-edge?
2236         if (BBToOrder[s] > BBToOrder[MBB]) {
2237           // No: visit it during this dataflow iteration.
2238           if (OnWorklist.insert(s).second)
2239             Worklist.push(BBToOrder[s]);
2240         } else {
2241           // Yes: visit it on the next iteration.
2242           if (OnPending.insert(s).second)
2243             Pending.push(BBToOrder[s]);
2244         }
2245       }
2246     }
2247 
2248     Worklist.swap(Pending);
2249     std::swap(OnPending, OnWorklist);
2250     OnPending.clear();
2251     // At this point, pending must be empty, since it was just the empty
2252     // worklist
2253     assert(Pending.empty() && "Pending should be empty");
2254   }
2255 
2256   // Once all the live-ins don't change on mlocJoin(), we've reached a
2257   // fixedpoint.
2258 }
2259 
2260 bool InstrRefBasedLDV::vlocDowngradeLattice(
2261     const MachineBasicBlock &MBB, const DbgValue &OldLiveInLocation,
2262     const SmallVectorImpl<InValueT> &Values, unsigned CurBlockRPONum) {
2263   // Ranking value preference: see file level comment, the highest rank is
2264   // a plain def, followed by PHI values in reverse post-order. Numerically,
2265   // we assign all defs the rank '0', all PHIs their blocks RPO number plus
2266   // one, and consider the lowest value the highest ranked.
2267   int OldLiveInRank = BBNumToRPO[OldLiveInLocation.ID.getBlock()] + 1;
2268   if (!OldLiveInLocation.ID.isPHI())
2269     OldLiveInRank = 0;
2270 
2271   // Allow any unresolvable conflict to be over-ridden.
2272   if (OldLiveInLocation.Kind == DbgValue::NoVal) {
2273     // Although if it was an unresolvable conflict from _this_ block, then
2274     // all other seeking of downgrades and PHIs must have failed before hand.
2275     if (OldLiveInLocation.BlockNo == (unsigned)MBB.getNumber())
2276       return false;
2277     OldLiveInRank = INT_MIN;
2278   }
2279 
2280   auto &InValue = *Values[0].second;
2281 
2282   if (InValue.Kind == DbgValue::Const || InValue.Kind == DbgValue::NoVal)
2283     return false;
2284 
2285   unsigned ThisRPO = BBNumToRPO[InValue.ID.getBlock()];
2286   int ThisRank = ThisRPO + 1;
2287   if (!InValue.ID.isPHI())
2288     ThisRank = 0;
2289 
2290   // Too far down the lattice?
2291   if (ThisRPO >= CurBlockRPONum)
2292     return false;
2293 
2294   // Higher in the lattice than what we've already explored?
2295   if (ThisRank <= OldLiveInRank)
2296     return false;
2297 
2298   return true;
2299 }
2300 
2301 std::tuple<Optional<ValueIDNum>, bool> InstrRefBasedLDV::pickVPHILoc(
2302     MachineBasicBlock &MBB, const DebugVariable &Var, const LiveIdxT &LiveOuts,
2303     ValueIDNum **MOutLocs, ValueIDNum **MInLocs,
2304     const SmallVectorImpl<MachineBasicBlock *> &BlockOrders) {
2305   // Collect a set of locations from predecessor where its live-out value can
2306   // be found.
2307   SmallVector<SmallVector<LocIdx, 4>, 8> Locs;
2308   unsigned NumLocs = MTracker->getNumLocs();
2309   unsigned BackEdgesStart = 0;
2310 
2311   for (auto p : BlockOrders) {
2312     // Pick out where backedges start in the list of predecessors. Relies on
2313     // BlockOrders being sorted by RPO.
2314     if (BBToOrder[p] < BBToOrder[&MBB])
2315       ++BackEdgesStart;
2316 
2317     // For each predecessor, create a new set of locations.
2318     Locs.resize(Locs.size() + 1);
2319     unsigned ThisBBNum = p->getNumber();
2320     auto LiveOutMap = LiveOuts.find(p);
2321     if (LiveOutMap == LiveOuts.end())
2322       // This predecessor isn't in scope, it must have no live-in/live-out
2323       // locations.
2324       continue;
2325 
2326     auto It = LiveOutMap->second->find(Var);
2327     if (It == LiveOutMap->second->end())
2328       // There's no value recorded for this variable in this predecessor,
2329       // leave an empty set of locations.
2330       continue;
2331 
2332     const DbgValue &OutVal = It->second;
2333 
2334     if (OutVal.Kind == DbgValue::Const || OutVal.Kind == DbgValue::NoVal)
2335       // Consts and no-values cannot have locations we can join on.
2336       continue;
2337 
2338     assert(OutVal.Kind == DbgValue::Proposed || OutVal.Kind == DbgValue::Def);
2339     ValueIDNum ValToLookFor = OutVal.ID;
2340 
2341     // Search the live-outs of the predecessor for the specified value.
2342     for (unsigned int I = 0; I < NumLocs; ++I) {
2343       if (MOutLocs[ThisBBNum][I] == ValToLookFor)
2344         Locs.back().push_back(LocIdx(I));
2345     }
2346   }
2347 
2348   // If there were no locations at all, return an empty result.
2349   if (Locs.empty())
2350     return std::tuple<Optional<ValueIDNum>, bool>(None, false);
2351 
2352   // Lambda for seeking a common location within a range of location-sets.
2353   typedef SmallVector<SmallVector<LocIdx, 4>, 8>::iterator LocsIt;
2354   auto SeekLocation =
2355       [&Locs](llvm::iterator_range<LocsIt> SearchRange) -> Optional<LocIdx> {
2356     // Starting with the first set of locations, take the intersection with
2357     // subsequent sets.
2358     SmallVector<LocIdx, 4> base = Locs[0];
2359     for (auto &S : SearchRange) {
2360       SmallVector<LocIdx, 4> new_base;
2361       std::set_intersection(base.begin(), base.end(), S.begin(), S.end(),
2362                             std::inserter(new_base, new_base.begin()));
2363       base = new_base;
2364     }
2365     if (base.empty())
2366       return None;
2367 
2368     // We now have a set of LocIdxes that contain the right output value in
2369     // each of the predecessors. Pick the lowest; if there's a register loc,
2370     // that'll be it.
2371     return *base.begin();
2372   };
2373 
2374   // Search for a common location for all predecessors. If we can't, then fall
2375   // back to only finding a common location between non-backedge predecessors.
2376   bool ValidForAllLocs = true;
2377   auto TheLoc = SeekLocation(Locs);
2378   if (!TheLoc) {
2379     ValidForAllLocs = false;
2380     TheLoc =
2381         SeekLocation(make_range(Locs.begin(), Locs.begin() + BackEdgesStart));
2382   }
2383 
2384   if (!TheLoc)
2385     return std::tuple<Optional<ValueIDNum>, bool>(None, false);
2386 
2387   // Return a PHI-value-number for the found location.
2388   LocIdx L = *TheLoc;
2389   ValueIDNum PHIVal = {(unsigned)MBB.getNumber(), 0, L};
2390   return std::tuple<Optional<ValueIDNum>, bool>(PHIVal, ValidForAllLocs);
2391 }
2392 
2393 std::tuple<bool, bool> InstrRefBasedLDV::vlocJoin(
2394     MachineBasicBlock &MBB, LiveIdxT &VLOCOutLocs, LiveIdxT &VLOCInLocs,
2395     SmallPtrSet<const MachineBasicBlock *, 16> *VLOCVisited, unsigned BBNum,
2396     const SmallSet<DebugVariable, 4> &AllVars, ValueIDNum **MOutLocs,
2397     ValueIDNum **MInLocs,
2398     SmallPtrSet<const MachineBasicBlock *, 8> &InScopeBlocks,
2399     SmallPtrSet<const MachineBasicBlock *, 8> &BlocksToExplore,
2400     DenseMap<DebugVariable, DbgValue> &InLocsT) {
2401   bool DowngradeOccurred = false;
2402 
2403   // To emulate VarLocBasedImpl, process this block if it's not in scope but
2404   // _does_ assign a variable value. No live-ins for this scope are transferred
2405   // in though, so we can return immediately.
2406   if (InScopeBlocks.count(&MBB) == 0 && !ArtificialBlocks.count(&MBB)) {
2407     if (VLOCVisited)
2408       return std::tuple<bool, bool>(true, false);
2409     return std::tuple<bool, bool>(false, false);
2410   }
2411 
2412   LLVM_DEBUG(dbgs() << "join MBB: " << MBB.getNumber() << "\n");
2413   bool Changed = false;
2414 
2415   // Find any live-ins computed in a prior iteration.
2416   auto ILSIt = VLOCInLocs.find(&MBB);
2417   assert(ILSIt != VLOCInLocs.end());
2418   auto &ILS = *ILSIt->second;
2419 
2420   // Order predecessors by RPOT order, for exploring them in that order.
2421   SmallVector<MachineBasicBlock *, 8> BlockOrders;
2422   for (auto p : MBB.predecessors())
2423     BlockOrders.push_back(p);
2424 
2425   auto Cmp = [&](MachineBasicBlock *A, MachineBasicBlock *B) {
2426     return BBToOrder[A] < BBToOrder[B];
2427   };
2428 
2429   llvm::sort(BlockOrders.begin(), BlockOrders.end(), Cmp);
2430 
2431   unsigned CurBlockRPONum = BBToOrder[&MBB];
2432 
2433   // Force a re-visit to loop heads in the first dataflow iteration.
2434   // FIXME: if we could "propose" Const values this wouldn't be needed,
2435   // because they'd need to be confirmed before being emitted.
2436   if (!BlockOrders.empty() &&
2437       BBToOrder[BlockOrders[BlockOrders.size() - 1]] >= CurBlockRPONum &&
2438       VLOCVisited)
2439     DowngradeOccurred = true;
2440 
2441   auto ConfirmValue = [&InLocsT](const DebugVariable &DV, DbgValue VR) {
2442     auto Result = InLocsT.insert(std::make_pair(DV, VR));
2443     (void)Result;
2444     assert(Result.second);
2445   };
2446 
2447   auto ConfirmNoVal = [&ConfirmValue, &MBB](const DebugVariable &Var, const DbgValueProperties &Properties) {
2448     DbgValue NoLocPHIVal(MBB.getNumber(), Properties, DbgValue::NoVal);
2449 
2450     ConfirmValue(Var, NoLocPHIVal);
2451   };
2452 
2453   // Attempt to join the values for each variable.
2454   for (auto &Var : AllVars) {
2455     // Collect all the DbgValues for this variable.
2456     SmallVector<InValueT, 8> Values;
2457     bool Bail = false;
2458     unsigned BackEdgesStart = 0;
2459     for (auto p : BlockOrders) {
2460       // If the predecessor isn't in scope / to be explored, we'll never be
2461       // able to join any locations.
2462       if (BlocksToExplore.find(p) == BlocksToExplore.end()) {
2463         Bail = true;
2464         break;
2465       }
2466 
2467       // Don't attempt to handle unvisited predecessors: they're implicitly
2468       // "unknown"s in the lattice.
2469       if (VLOCVisited && !VLOCVisited->count(p))
2470         continue;
2471 
2472       // If the predecessors OutLocs is absent, there's not much we can do.
2473       auto OL = VLOCOutLocs.find(p);
2474       if (OL == VLOCOutLocs.end()) {
2475         Bail = true;
2476         break;
2477       }
2478 
2479       // No live-out value for this predecessor also means we can't produce
2480       // a joined value.
2481       auto VIt = OL->second->find(Var);
2482       if (VIt == OL->second->end()) {
2483         Bail = true;
2484         break;
2485       }
2486 
2487       // Keep track of where back-edges begin in the Values vector. Relies on
2488       // BlockOrders being sorted by RPO.
2489       unsigned ThisBBRPONum = BBToOrder[p];
2490       if (ThisBBRPONum < CurBlockRPONum)
2491         ++BackEdgesStart;
2492 
2493       Values.push_back(std::make_pair(p, &VIt->second));
2494     }
2495 
2496     // If there were no values, or one of the predecessors couldn't have a
2497     // value, then give up immediately. It's not safe to produce a live-in
2498     // value.
2499     if (Bail || Values.size() == 0)
2500       continue;
2501 
2502     // Enumeration identifying the current state of the predecessors values.
2503     enum {
2504       Unset = 0,
2505       Agreed,       // All preds agree on the variable value.
2506       PropDisagree, // All preds agree, but the value kind is Proposed in some.
2507       BEDisagree,   // Only back-edges disagree on variable value.
2508       PHINeeded,    // Non-back-edge predecessors have conflicing values.
2509       NoSolution    // Conflicting Value metadata makes solution impossible.
2510     } OurState = Unset;
2511 
2512     // All (non-entry) blocks have at least one non-backedge predecessor.
2513     // Pick the variable value from the first of these, to compare against
2514     // all others.
2515     const DbgValue &FirstVal = *Values[0].second;
2516     const ValueIDNum &FirstID = FirstVal.ID;
2517 
2518     // Scan for variable values that can't be resolved: if they have different
2519     // DIExpressions, different indirectness, or are mixed constants /
2520     // non-constants.
2521     for (auto &V : Values) {
2522       if (V.second->Properties != FirstVal.Properties)
2523         OurState = NoSolution;
2524       if (V.second->Kind == DbgValue::Const && FirstVal.Kind != DbgValue::Const)
2525         OurState = NoSolution;
2526     }
2527 
2528     // Flags diagnosing _how_ the values disagree.
2529     bool NonBackEdgeDisagree = false;
2530     bool DisagreeOnPHINess = false;
2531     bool IDDisagree = false;
2532     bool Disagree = false;
2533     if (OurState == Unset) {
2534       for (auto &V : Values) {
2535         if (*V.second == FirstVal)
2536           continue; // No disagreement.
2537 
2538         Disagree = true;
2539 
2540         // Flag whether the value number actually diagrees.
2541         if (V.second->ID != FirstID)
2542           IDDisagree = true;
2543 
2544         // Distinguish whether disagreement happens in backedges or not.
2545         // Relies on Values (and BlockOrders) being sorted by RPO.
2546         unsigned ThisBBRPONum = BBToOrder[V.first];
2547         if (ThisBBRPONum < CurBlockRPONum)
2548           NonBackEdgeDisagree = true;
2549 
2550         // Is there a difference in whether the value is definite or only
2551         // proposed?
2552         if (V.second->Kind != FirstVal.Kind &&
2553             (V.second->Kind == DbgValue::Proposed ||
2554              V.second->Kind == DbgValue::Def) &&
2555             (FirstVal.Kind == DbgValue::Proposed ||
2556              FirstVal.Kind == DbgValue::Def))
2557           DisagreeOnPHINess = true;
2558       }
2559 
2560       // Collect those flags together and determine an overall state for
2561       // what extend the predecessors agree on a live-in value.
2562       if (!Disagree)
2563         OurState = Agreed;
2564       else if (!IDDisagree && DisagreeOnPHINess)
2565         OurState = PropDisagree;
2566       else if (!NonBackEdgeDisagree)
2567         OurState = BEDisagree;
2568       else
2569         OurState = PHINeeded;
2570     }
2571 
2572     // An extra indicator: if we only disagree on whether the value is a
2573     // Def, or proposed, then also flag whether that disagreement happens
2574     // in backedges only.
2575     bool PropOnlyInBEs = Disagree && !IDDisagree && DisagreeOnPHINess &&
2576                          !NonBackEdgeDisagree && FirstVal.Kind == DbgValue::Def;
2577 
2578     const auto &Properties = FirstVal.Properties;
2579 
2580     auto OldLiveInIt = ILS.find(Var);
2581     const DbgValue *OldLiveInLocation =
2582         (OldLiveInIt != ILS.end()) ? &OldLiveInIt->second : nullptr;
2583 
2584     bool OverRide = false;
2585     if (OurState == BEDisagree && OldLiveInLocation) {
2586       // Only backedges disagree: we can consider downgrading. If there was a
2587       // previous live-in value, use it to work out whether the current
2588       // incoming value represents a lattice downgrade or not.
2589       OverRide =
2590           vlocDowngradeLattice(MBB, *OldLiveInLocation, Values, CurBlockRPONum);
2591     }
2592 
2593     // Use the current state of predecessor agreement and other flags to work
2594     // out what to do next. Possibilities include:
2595     //  * Accept a value all predecessors agree on, or accept one that
2596     //    represents a step down the exploration lattice,
2597     //  * Use a PHI value number, if one can be found,
2598     //  * Propose a PHI value number, and see if it gets confirmed later,
2599     //  * Emit a 'NoVal' value, indicating we couldn't resolve anything.
2600     if (OurState == Agreed) {
2601       // Easiest solution: all predecessors agree on the variable value.
2602       ConfirmValue(Var, FirstVal);
2603     } else if (OurState == BEDisagree && OverRide) {
2604       // Only backedges disagree, and the other predecessors have produced
2605       // a new live-in value further down the exploration lattice.
2606       DowngradeOccurred = true;
2607       ConfirmValue(Var, FirstVal);
2608     } else if (OurState == PropDisagree) {
2609       // Predecessors agree on value, but some say it's only a proposed value.
2610       // Propagate it as proposed: unless it was proposed in this block, in
2611       // which case we're able to confirm the value.
2612       if (FirstID.getBlock() == (uint64_t)MBB.getNumber() && FirstID.isPHI()) {
2613         ConfirmValue(Var, DbgValue(FirstID, Properties, DbgValue::Def));
2614       } else if (PropOnlyInBEs) {
2615         // If only backedges disagree, a higher (in RPO) block confirmed this
2616         // location, and we need to propagate it into this loop.
2617         ConfirmValue(Var, DbgValue(FirstID, Properties, DbgValue::Def));
2618       } else {
2619         // Otherwise; a Def meeting a Proposed is still a Proposed.
2620         ConfirmValue(Var, DbgValue(FirstID, Properties, DbgValue::Proposed));
2621       }
2622     } else if ((OurState == PHINeeded || OurState == BEDisagree)) {
2623       // Predecessors disagree and can't be downgraded: this can only be
2624       // solved with a PHI. Use pickVPHILoc to go look for one.
2625       Optional<ValueIDNum> VPHI;
2626       bool AllEdgesVPHI = false;
2627       std::tie(VPHI, AllEdgesVPHI) =
2628           pickVPHILoc(MBB, Var, VLOCOutLocs, MOutLocs, MInLocs, BlockOrders);
2629 
2630       if (VPHI && AllEdgesVPHI) {
2631         // There's a PHI value that's valid for all predecessors -- we can use
2632         // it. If any of the non-backedge predecessors have proposed values
2633         // though, this PHI is also only proposed, until the predecessors are
2634         // confirmed.
2635         DbgValue::KindT K = DbgValue::Def;
2636         for (unsigned int I = 0; I < BackEdgesStart; ++I)
2637           if (Values[I].second->Kind == DbgValue::Proposed)
2638             K = DbgValue::Proposed;
2639 
2640         ConfirmValue(Var, DbgValue(*VPHI, Properties, K));
2641       } else if (VPHI) {
2642         // There's a PHI value, but it's only legal for backedges. Leave this
2643         // as a proposed PHI value: it might come back on the backedges,
2644         // and allow us to confirm it in the future.
2645         DbgValue NoBEValue = DbgValue(*VPHI, Properties, DbgValue::Proposed);
2646         ConfirmValue(Var, NoBEValue);
2647       } else {
2648         ConfirmNoVal(Var, Properties);
2649       }
2650     } else {
2651       // Otherwise: we don't know. Emit a "phi but no real loc" phi.
2652       ConfirmNoVal(Var, Properties);
2653     }
2654   }
2655 
2656   // Store newly calculated in-locs into VLOCInLocs, if they've changed.
2657   Changed = ILS != InLocsT;
2658   if (Changed)
2659     ILS = InLocsT;
2660 
2661   return std::tuple<bool, bool>(Changed, DowngradeOccurred);
2662 }
2663 
2664 void InstrRefBasedLDV::vlocDataflow(
2665     const LexicalScope *Scope, const DILocation *DILoc,
2666     const SmallSet<DebugVariable, 4> &VarsWeCareAbout,
2667     SmallPtrSetImpl<MachineBasicBlock *> &AssignBlocks, LiveInsT &Output,
2668     ValueIDNum **MOutLocs, ValueIDNum **MInLocs,
2669     SmallVectorImpl<VLocTracker> &AllTheVLocs) {
2670   // This method is much like mlocDataflow: but focuses on a single
2671   // LexicalScope at a time. Pick out a set of blocks and variables that are
2672   // to have their value assignments solved, then run our dataflow algorithm
2673   // until a fixedpoint is reached.
2674   std::priority_queue<unsigned int, std::vector<unsigned int>,
2675                       std::greater<unsigned int>>
2676       Worklist, Pending;
2677   SmallPtrSet<MachineBasicBlock *, 16> OnWorklist, OnPending;
2678 
2679   // The set of blocks we'll be examining.
2680   SmallPtrSet<const MachineBasicBlock *, 8> BlocksToExplore;
2681 
2682   // The order in which to examine them (RPO).
2683   SmallVector<MachineBasicBlock *, 8> BlockOrders;
2684 
2685   // RPO ordering function.
2686   auto Cmp = [&](MachineBasicBlock *A, MachineBasicBlock *B) {
2687     return BBToOrder[A] < BBToOrder[B];
2688   };
2689 
2690   LS.getMachineBasicBlocks(DILoc, BlocksToExplore);
2691 
2692   // A separate container to distinguish "blocks we're exploring" versus
2693   // "blocks that are potentially in scope. See comment at start of vlocJoin.
2694   SmallPtrSet<const MachineBasicBlock *, 8> InScopeBlocks = BlocksToExplore;
2695 
2696   // Old LiveDebugValues tracks variable locations that come out of blocks
2697   // not in scope, where DBG_VALUEs occur. This is something we could
2698   // legitimately ignore, but lets allow it for now.
2699   if (EmulateOldLDV)
2700     BlocksToExplore.insert(AssignBlocks.begin(), AssignBlocks.end());
2701 
2702   // We also need to propagate variable values through any artificial blocks
2703   // that immediately follow blocks in scope.
2704   DenseSet<const MachineBasicBlock *> ToAdd;
2705 
2706   // Helper lambda: For a given block in scope, perform a depth first search
2707   // of all the artificial successors, adding them to the ToAdd collection.
2708   auto AccumulateArtificialBlocks =
2709       [this, &ToAdd, &BlocksToExplore,
2710        &InScopeBlocks](const MachineBasicBlock *MBB) {
2711         // Depth-first-search state: each node is a block and which successor
2712         // we're currently exploring.
2713         SmallVector<std::pair<const MachineBasicBlock *,
2714                               MachineBasicBlock::const_succ_iterator>,
2715                     8>
2716             DFS;
2717 
2718         // Find any artificial successors not already tracked.
2719         for (auto *succ : MBB->successors()) {
2720           if (BlocksToExplore.count(succ) || InScopeBlocks.count(succ))
2721             continue;
2722           if (!ArtificialBlocks.count(succ))
2723             continue;
2724           DFS.push_back(std::make_pair(succ, succ->succ_begin()));
2725           ToAdd.insert(succ);
2726         }
2727 
2728         // Search all those blocks, depth first.
2729         while (!DFS.empty()) {
2730           const MachineBasicBlock *CurBB = DFS.back().first;
2731           MachineBasicBlock::const_succ_iterator &CurSucc = DFS.back().second;
2732           // Walk back if we've explored this blocks successors to the end.
2733           if (CurSucc == CurBB->succ_end()) {
2734             DFS.pop_back();
2735             continue;
2736           }
2737 
2738           // If the current successor is artificial and unexplored, descend into
2739           // it.
2740           if (!ToAdd.count(*CurSucc) && ArtificialBlocks.count(*CurSucc)) {
2741             DFS.push_back(std::make_pair(*CurSucc, (*CurSucc)->succ_begin()));
2742             ToAdd.insert(*CurSucc);
2743             continue;
2744           }
2745 
2746           ++CurSucc;
2747         }
2748       };
2749 
2750   // Search in-scope blocks and those containing a DBG_VALUE from this scope
2751   // for artificial successors.
2752   for (auto *MBB : BlocksToExplore)
2753     AccumulateArtificialBlocks(MBB);
2754   for (auto *MBB : InScopeBlocks)
2755     AccumulateArtificialBlocks(MBB);
2756 
2757   BlocksToExplore.insert(ToAdd.begin(), ToAdd.end());
2758   InScopeBlocks.insert(ToAdd.begin(), ToAdd.end());
2759 
2760   // Single block scope: not interesting! No propagation at all. Note that
2761   // this could probably go above ArtificialBlocks without damage, but
2762   // that then produces output differences from original-live-debug-values,
2763   // which propagates from a single block into many artificial ones.
2764   if (BlocksToExplore.size() == 1)
2765     return;
2766 
2767   // Picks out relevants blocks RPO order and sort them.
2768   for (auto *MBB : BlocksToExplore)
2769     BlockOrders.push_back(const_cast<MachineBasicBlock *>(MBB));
2770 
2771   llvm::sort(BlockOrders.begin(), BlockOrders.end(), Cmp);
2772   unsigned NumBlocks = BlockOrders.size();
2773 
2774   // Allocate some vectors for storing the live ins and live outs. Large.
2775   SmallVector<DenseMap<DebugVariable, DbgValue>, 32> LiveIns, LiveOuts;
2776   LiveIns.resize(NumBlocks);
2777   LiveOuts.resize(NumBlocks);
2778 
2779   // Produce by-MBB indexes of live-in/live-outs, to ease lookup within
2780   // vlocJoin.
2781   LiveIdxT LiveOutIdx, LiveInIdx;
2782   LiveOutIdx.reserve(NumBlocks);
2783   LiveInIdx.reserve(NumBlocks);
2784   for (unsigned I = 0; I < NumBlocks; ++I) {
2785     LiveOutIdx[BlockOrders[I]] = &LiveOuts[I];
2786     LiveInIdx[BlockOrders[I]] = &LiveIns[I];
2787   }
2788 
2789   for (auto *MBB : BlockOrders) {
2790     Worklist.push(BBToOrder[MBB]);
2791     OnWorklist.insert(MBB);
2792   }
2793 
2794   // Iterate over all the blocks we selected, propagating variable values.
2795   bool FirstTrip = true;
2796   SmallPtrSet<const MachineBasicBlock *, 16> VLOCVisited;
2797   while (!Worklist.empty() || !Pending.empty()) {
2798     while (!Worklist.empty()) {
2799       auto *MBB = OrderToBB[Worklist.top()];
2800       CurBB = MBB->getNumber();
2801       Worklist.pop();
2802 
2803       DenseMap<DebugVariable, DbgValue> JoinedInLocs;
2804 
2805       // Join values from predecessors. Updates LiveInIdx, and writes output
2806       // into JoinedInLocs.
2807       bool InLocsChanged, DowngradeOccurred;
2808       std::tie(InLocsChanged, DowngradeOccurred) = vlocJoin(
2809           *MBB, LiveOutIdx, LiveInIdx, (FirstTrip) ? &VLOCVisited : nullptr,
2810           CurBB, VarsWeCareAbout, MOutLocs, MInLocs, InScopeBlocks,
2811           BlocksToExplore, JoinedInLocs);
2812 
2813       auto &VTracker = AllTheVLocs[MBB->getNumber()];
2814       bool FirstVisit = VLOCVisited.insert(MBB).second;
2815 
2816       // Always explore transfer function if inlocs changed, or if we've not
2817       // visited this block before.
2818       InLocsChanged |= FirstVisit;
2819 
2820       // If a downgrade occurred, book us in for re-examination on the next
2821       // iteration.
2822       if (DowngradeOccurred && OnPending.insert(MBB).second)
2823         Pending.push(BBToOrder[MBB]);
2824 
2825       // Patch up the variable value transfer function to use the live-in
2826       // machine values, now that that problem is solved.
2827       if (FirstVisit) {
2828         for (auto &Transfer : VTracker.Vars) {
2829           if (Transfer.second.Kind == DbgValue::Def &&
2830               Transfer.second.ID.getBlock() == CurBB &&
2831               Transfer.second.ID.isPHI()) {
2832             LocIdx Loc = Transfer.second.ID.getLoc();
2833             Transfer.second.ID = MInLocs[CurBB][Loc.asU64()];
2834           }
2835         }
2836       }
2837 
2838       if (!InLocsChanged)
2839         continue;
2840 
2841       // Do transfer function.
2842       for (auto &Transfer : VTracker.Vars) {
2843         // Is this var we're mangling in this scope?
2844         if (VarsWeCareAbout.count(Transfer.first)) {
2845           // Erase on empty transfer (DBG_VALUE $noreg).
2846           if (Transfer.second.Kind == DbgValue::Undef) {
2847             JoinedInLocs.erase(Transfer.first);
2848           } else {
2849             // Insert new variable value; or overwrite.
2850             auto NewValuePair = std::make_pair(Transfer.first, Transfer.second);
2851             auto Result = JoinedInLocs.insert(NewValuePair);
2852             if (!Result.second)
2853               Result.first->second = Transfer.second;
2854           }
2855         }
2856       }
2857 
2858       // Did the live-out locations change?
2859       bool OLChanged = JoinedInLocs != *LiveOutIdx[MBB];
2860 
2861       // If they haven't changed, there's no need to explore further.
2862       if (!OLChanged)
2863         continue;
2864 
2865       // Commit to the live-out record.
2866       *LiveOutIdx[MBB] = JoinedInLocs;
2867 
2868       // We should visit all successors. Ensure we'll visit any non-backedge
2869       // successors during this dataflow iteration; book backedge successors
2870       // to be visited next time around.
2871       for (auto s : MBB->successors()) {
2872         // Ignore out of scope / not-to-be-explored successors.
2873         if (LiveInIdx.find(s) == LiveInIdx.end())
2874           continue;
2875 
2876         if (BBToOrder[s] > BBToOrder[MBB]) {
2877           if (OnWorklist.insert(s).second)
2878             Worklist.push(BBToOrder[s]);
2879         } else if (OnPending.insert(s).second && (FirstTrip || OLChanged)) {
2880           Pending.push(BBToOrder[s]);
2881         }
2882       }
2883     }
2884     Worklist.swap(Pending);
2885     std::swap(OnWorklist, OnPending);
2886     OnPending.clear();
2887     assert(Pending.empty());
2888     FirstTrip = false;
2889   }
2890 
2891   // Dataflow done. Now what? Save live-ins. Ignore any that are still marked
2892   // as being variable-PHIs, because those did not have their machine-PHI
2893   // value confirmed. Such variable values are places that could have been
2894   // PHIs, but are not.
2895   for (auto *MBB : BlockOrders) {
2896     auto &VarMap = *LiveInIdx[MBB];
2897     for (auto &P : VarMap) {
2898       if (P.second.Kind == DbgValue::Proposed ||
2899           P.second.Kind == DbgValue::NoVal)
2900         continue;
2901       Output[MBB->getNumber()].push_back(P);
2902     }
2903   }
2904 
2905   BlockOrders.clear();
2906   BlocksToExplore.clear();
2907 }
2908 
2909 void InstrRefBasedLDV::dump_mloc_transfer(
2910     const MLocTransferMap &mloc_transfer) const {
2911   for (auto &P : mloc_transfer) {
2912     std::string foo = MTracker->LocIdxToName(P.first);
2913     std::string bar = MTracker->IDAsString(P.second);
2914     dbgs() << "Loc " << foo << " --> " << bar << "\n";
2915   }
2916 }
2917 
2918 void InstrRefBasedLDV::emitLocations(
2919     MachineFunction &MF, LiveInsT SavedLiveIns, ValueIDNum **MInLocs,
2920     DenseMap<DebugVariable, unsigned> &AllVarsNumbering) {
2921   TTracker = new TransferTracker(TII, MTracker, MF, *TRI, CalleeSavedRegs);
2922   unsigned NumLocs = MTracker->getNumLocs();
2923 
2924   // For each block, load in the machine value locations and variable value
2925   // live-ins, then step through each instruction in the block. New DBG_VALUEs
2926   // to be inserted will be created along the way.
2927   for (MachineBasicBlock &MBB : MF) {
2928     unsigned bbnum = MBB.getNumber();
2929     MTracker->reset();
2930     MTracker->loadFromArray(MInLocs[bbnum], bbnum);
2931     TTracker->loadInlocs(MBB, MInLocs[bbnum], SavedLiveIns[MBB.getNumber()],
2932                          NumLocs);
2933 
2934     CurBB = bbnum;
2935     CurInst = 1;
2936     for (auto &MI : MBB) {
2937       process(MI);
2938       ++CurInst;
2939     }
2940   }
2941 
2942   // We have to insert DBG_VALUEs in a consistent order, otherwise they appeaer
2943   // in DWARF in different orders. Use the order that they appear when walking
2944   // through each block / each instruction, stored in AllVarsNumbering.
2945   auto OrderDbgValues = [&](const MachineInstr *A,
2946                             const MachineInstr *B) -> bool {
2947     DebugVariable VarA(A->getDebugVariable(), A->getDebugExpression(),
2948                        A->getDebugLoc()->getInlinedAt());
2949     DebugVariable VarB(B->getDebugVariable(), B->getDebugExpression(),
2950                        B->getDebugLoc()->getInlinedAt());
2951     return AllVarsNumbering.find(VarA)->second <
2952            AllVarsNumbering.find(VarB)->second;
2953   };
2954 
2955   // Go through all the transfers recorded in the TransferTracker -- this is
2956   // both the live-ins to a block, and any movements of values that happen
2957   // in the middle.
2958   for (auto &P : TTracker->Transfers) {
2959     // Sort them according to appearance order.
2960     llvm::sort(P.Insts.begin(), P.Insts.end(), OrderDbgValues);
2961     // Insert either before or after the designated point...
2962     if (P.MBB) {
2963       MachineBasicBlock &MBB = *P.MBB;
2964       for (auto *MI : P.Insts) {
2965         MBB.insert(P.Pos, MI);
2966       }
2967     } else {
2968       MachineBasicBlock &MBB = *P.Pos->getParent();
2969       for (auto *MI : P.Insts) {
2970         MBB.insertAfter(P.Pos, MI);
2971       }
2972     }
2973   }
2974 }
2975 
2976 void InstrRefBasedLDV::initialSetup(MachineFunction &MF) {
2977   // Build some useful data structures.
2978   auto hasNonArtificialLocation = [](const MachineInstr &MI) -> bool {
2979     if (const DebugLoc &DL = MI.getDebugLoc())
2980       return DL.getLine() != 0;
2981     return false;
2982   };
2983   // Collect a set of all the artificial blocks.
2984   for (auto &MBB : MF)
2985     if (none_of(MBB.instrs(), hasNonArtificialLocation))
2986       ArtificialBlocks.insert(&MBB);
2987 
2988   // Compute mappings of block <=> RPO order.
2989   ReversePostOrderTraversal<MachineFunction *> RPOT(&MF);
2990   unsigned int RPONumber = 0;
2991   for (auto RI = RPOT.begin(), RE = RPOT.end(); RI != RE; ++RI) {
2992     OrderToBB[RPONumber] = *RI;
2993     BBToOrder[*RI] = RPONumber;
2994     BBNumToRPO[(*RI)->getNumber()] = RPONumber;
2995     ++RPONumber;
2996   }
2997 }
2998 
2999 /// Calculate the liveness information for the given machine function and
3000 /// extend ranges across basic blocks.
3001 bool InstrRefBasedLDV::ExtendRanges(MachineFunction &MF,
3002                                     TargetPassConfig *TPC) {
3003   // No subprogram means this function contains no debuginfo.
3004   if (!MF.getFunction().getSubprogram())
3005     return false;
3006 
3007   LLVM_DEBUG(dbgs() << "\nDebug Range Extension\n");
3008   this->TPC = TPC;
3009 
3010   TRI = MF.getSubtarget().getRegisterInfo();
3011   TII = MF.getSubtarget().getInstrInfo();
3012   TFI = MF.getSubtarget().getFrameLowering();
3013   TFI->getCalleeSaves(MF, CalleeSavedRegs);
3014   LS.initialize(MF);
3015 
3016   MTracker =
3017       new MLocTracker(MF, *TII, *TRI, *MF.getSubtarget().getTargetLowering());
3018   VTracker = nullptr;
3019   TTracker = nullptr;
3020 
3021   SmallVector<MLocTransferMap, 32> MLocTransfer;
3022   SmallVector<VLocTracker, 8> vlocs;
3023   LiveInsT SavedLiveIns;
3024 
3025   int MaxNumBlocks = -1;
3026   for (auto &MBB : MF)
3027     MaxNumBlocks = std::max(MBB.getNumber(), MaxNumBlocks);
3028   assert(MaxNumBlocks >= 0);
3029   ++MaxNumBlocks;
3030 
3031   MLocTransfer.resize(MaxNumBlocks);
3032   vlocs.resize(MaxNumBlocks);
3033   SavedLiveIns.resize(MaxNumBlocks);
3034 
3035   initialSetup(MF);
3036 
3037   produceTransferFunctions(MF, MLocTransfer, MaxNumBlocks, vlocs);
3038 
3039   // Allocate and initialize two array-of-arrays for the live-in and live-out
3040   // machine values. The outer dimension is the block number; while the inner
3041   // dimension is a LocIdx from MLocTracker.
3042   ValueIDNum **MOutLocs = new ValueIDNum *[MaxNumBlocks];
3043   ValueIDNum **MInLocs = new ValueIDNum *[MaxNumBlocks];
3044   unsigned NumLocs = MTracker->getNumLocs();
3045   for (int i = 0; i < MaxNumBlocks; ++i) {
3046     MOutLocs[i] = new ValueIDNum[NumLocs];
3047     MInLocs[i] = new ValueIDNum[NumLocs];
3048   }
3049 
3050   // Solve the machine value dataflow problem using the MLocTransfer function,
3051   // storing the computed live-ins / live-outs into the array-of-arrays. We use
3052   // both live-ins and live-outs for decision making in the variable value
3053   // dataflow problem.
3054   mlocDataflow(MInLocs, MOutLocs, MLocTransfer);
3055 
3056   // Number all variables in the order that they appear, to be used as a stable
3057   // insertion order later.
3058   DenseMap<DebugVariable, unsigned> AllVarsNumbering;
3059 
3060   // Map from one LexicalScope to all the variables in that scope.
3061   DenseMap<const LexicalScope *, SmallSet<DebugVariable, 4>> ScopeToVars;
3062 
3063   // Map from One lexical scope to all blocks in that scope.
3064   DenseMap<const LexicalScope *, SmallPtrSet<MachineBasicBlock *, 4>>
3065       ScopeToBlocks;
3066 
3067   // Store a DILocation that describes a scope.
3068   DenseMap<const LexicalScope *, const DILocation *> ScopeToDILocation;
3069 
3070   // To mirror old LiveDebugValues, enumerate variables in RPOT order. Otherwise
3071   // the order is unimportant, it just has to be stable.
3072   for (unsigned int I = 0; I < OrderToBB.size(); ++I) {
3073     auto *MBB = OrderToBB[I];
3074     auto *VTracker = &vlocs[MBB->getNumber()];
3075     // Collect each variable with a DBG_VALUE in this block.
3076     for (auto &idx : VTracker->Vars) {
3077       const auto &Var = idx.first;
3078       const DILocation *ScopeLoc = VTracker->Scopes[Var];
3079       assert(ScopeLoc != nullptr);
3080       auto *Scope = LS.findLexicalScope(ScopeLoc);
3081 
3082       // No insts in scope -> shouldn't have been recorded.
3083       assert(Scope != nullptr);
3084 
3085       AllVarsNumbering.insert(std::make_pair(Var, AllVarsNumbering.size()));
3086       ScopeToVars[Scope].insert(Var);
3087       ScopeToBlocks[Scope].insert(VTracker->MBB);
3088       ScopeToDILocation[Scope] = ScopeLoc;
3089     }
3090   }
3091 
3092   // OK. Iterate over scopes: there might be something to be said for
3093   // ordering them by size/locality, but that's for the future. For each scope,
3094   // solve the variable value problem, producing a map of variables to values
3095   // in SavedLiveIns.
3096   for (auto &P : ScopeToVars) {
3097     vlocDataflow(P.first, ScopeToDILocation[P.first], P.second,
3098                  ScopeToBlocks[P.first], SavedLiveIns, MOutLocs, MInLocs,
3099                  vlocs);
3100   }
3101 
3102   // Using the computed value locations and variable values for each block,
3103   // create the DBG_VALUE instructions representing the extended variable
3104   // locations.
3105   emitLocations(MF, SavedLiveIns, MInLocs, AllVarsNumbering);
3106 
3107   for (int Idx = 0; Idx < MaxNumBlocks; ++Idx) {
3108     delete[] MOutLocs[Idx];
3109     delete[] MInLocs[Idx];
3110   }
3111   delete[] MOutLocs;
3112   delete[] MInLocs;
3113 
3114   // Did we actually make any changes? If we created any DBG_VALUEs, then yes.
3115   bool Changed = TTracker->Transfers.size() != 0;
3116 
3117   delete MTracker;
3118   VTracker = nullptr;
3119   TTracker = nullptr;
3120 
3121   ArtificialBlocks.clear();
3122   OrderToBB.clear();
3123   BBToOrder.clear();
3124   BBNumToRPO.clear();
3125 
3126   return Changed;
3127 }
3128 
3129 LDVImpl *llvm::makeInstrRefBasedLiveDebugValues() {
3130   return new InstrRefBasedLDV();
3131 }
3132