1 //===- StackColoring.cpp --------------------------------------------------===//
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 //
9 // This pass implements the stack-coloring optimization that looks for
10 // lifetime markers machine instructions (LIFESTART_BEGIN and LIFESTART_END),
11 // which represent the possible lifetime of stack slots. It attempts to
12 // merge disjoint stack slots and reduce the used stack space.
13 // NOTE: This pass is not StackSlotColoring, which optimizes spill slots.
14 //
15 // TODO: In the future we plan to improve stack coloring in the following ways:
16 // 1. Allow merging multiple small slots into a single larger slot at different
17 //    offsets.
18 // 2. Merge this pass with StackSlotColoring and allow merging of allocas with
19 //    spill slots.
20 //
21 //===----------------------------------------------------------------------===//
22 
23 #include "llvm/ADT/BitVector.h"
24 #include "llvm/ADT/DenseMap.h"
25 #include "llvm/ADT/DepthFirstIterator.h"
26 #include "llvm/ADT/SmallPtrSet.h"
27 #include "llvm/ADT/SmallVector.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/Analysis/ValueTracking.h"
30 #include "llvm/CodeGen/LiveInterval.h"
31 #include "llvm/CodeGen/MachineBasicBlock.h"
32 #include "llvm/CodeGen/MachineFrameInfo.h"
33 #include "llvm/CodeGen/MachineFunction.h"
34 #include "llvm/CodeGen/MachineFunctionPass.h"
35 #include "llvm/CodeGen/MachineInstr.h"
36 #include "llvm/CodeGen/MachineMemOperand.h"
37 #include "llvm/CodeGen/MachineOperand.h"
38 #include "llvm/CodeGen/Passes.h"
39 #include "llvm/CodeGen/SlotIndexes.h"
40 #include "llvm/CodeGen/TargetOpcodes.h"
41 #include "llvm/CodeGen/WinEHFuncInfo.h"
42 #include "llvm/Config/llvm-config.h"
43 #include "llvm/IR/Constants.h"
44 #include "llvm/IR/DebugInfoMetadata.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/Metadata.h"
47 #include "llvm/IR/Use.h"
48 #include "llvm/IR/Value.h"
49 #include "llvm/InitializePasses.h"
50 #include "llvm/Pass.h"
51 #include "llvm/Support/Casting.h"
52 #include "llvm/Support/CommandLine.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/Support/Debug.h"
55 #include "llvm/Support/raw_ostream.h"
56 #include <algorithm>
57 #include <cassert>
58 #include <limits>
59 #include <memory>
60 #include <utility>
61 
62 using namespace llvm;
63 
64 #define DEBUG_TYPE "stack-coloring"
65 
66 static cl::opt<bool>
67 DisableColoring("no-stack-coloring",
68         cl::init(false), cl::Hidden,
69         cl::desc("Disable stack coloring"));
70 
71 /// The user may write code that uses allocas outside of the declared lifetime
72 /// zone. This can happen when the user returns a reference to a local
73 /// data-structure. We can detect these cases and decide not to optimize the
74 /// code. If this flag is enabled, we try to save the user. This option
75 /// is treated as overriding LifetimeStartOnFirstUse below.
76 static cl::opt<bool>
77 ProtectFromEscapedAllocas("protect-from-escaped-allocas",
78                           cl::init(false), cl::Hidden,
79                           cl::desc("Do not optimize lifetime zones that "
80                                    "are broken"));
81 
82 /// Enable enhanced dataflow scheme for lifetime analysis (treat first
83 /// use of stack slot as start of slot lifetime, as opposed to looking
84 /// for LIFETIME_START marker). See "Implementation notes" below for
85 /// more info.
86 static cl::opt<bool>
87 LifetimeStartOnFirstUse("stackcoloring-lifetime-start-on-first-use",
88         cl::init(true), cl::Hidden,
89         cl::desc("Treat stack lifetimes as starting on first use, not on START marker."));
90 
91 
92 STATISTIC(NumMarkerSeen,  "Number of lifetime markers found.");
93 STATISTIC(StackSpaceSaved, "Number of bytes saved due to merging slots.");
94 STATISTIC(StackSlotMerged, "Number of stack slot merged.");
95 STATISTIC(EscapedAllocas, "Number of allocas that escaped the lifetime region");
96 
97 //===----------------------------------------------------------------------===//
98 //                           StackColoring Pass
99 //===----------------------------------------------------------------------===//
100 //
101 // Stack Coloring reduces stack usage by merging stack slots when they
102 // can't be used together. For example, consider the following C program:
103 //
104 //     void bar(char *, int);
105 //     void foo(bool var) {
106 //         A: {
107 //             char z[4096];
108 //             bar(z, 0);
109 //         }
110 //
111 //         char *p;
112 //         char x[4096];
113 //         char y[4096];
114 //         if (var) {
115 //             p = x;
116 //         } else {
117 //             bar(y, 1);
118 //             p = y + 1024;
119 //         }
120 //     B:
121 //         bar(p, 2);
122 //     }
123 //
124 // Naively-compiled, this program would use 12k of stack space. However, the
125 // stack slot corresponding to `z` is always destroyed before either of the
126 // stack slots for `x` or `y` are used, and then `x` is only used if `var`
127 // is true, while `y` is only used if `var` is false. So in no time are 2
128 // of the stack slots used together, and therefore we can merge them,
129 // compiling the function using only a single 4k alloca:
130 //
131 //     void foo(bool var) { // equivalent
132 //         char x[4096];
133 //         char *p;
134 //         bar(x, 0);
135 //         if (var) {
136 //             p = x;
137 //         } else {
138 //             bar(x, 1);
139 //             p = x + 1024;
140 //         }
141 //         bar(p, 2);
142 //     }
143 //
144 // This is an important optimization if we want stack space to be under
145 // control in large functions, both open-coded ones and ones created by
146 // inlining.
147 //
148 // Implementation Notes:
149 // ---------------------
150 //
151 // An important part of the above reasoning is that `z` can't be accessed
152 // while the latter 2 calls to `bar` are running. This is justified because
153 // `z`'s lifetime is over after we exit from block `A:`, so any further
154 // accesses to it would be UB. The way we represent this information
155 // in LLVM is by having frontends delimit blocks with `lifetime.start`
156 // and `lifetime.end` intrinsics.
157 //
158 // The effect of these intrinsics seems to be as follows (maybe I should
159 // specify this in the reference?):
160 //
161 //   L1) at start, each stack-slot is marked as *out-of-scope*, unless no
162 //   lifetime intrinsic refers to that stack slot, in which case
163 //   it is marked as *in-scope*.
164 //   L2) on a `lifetime.start`, a stack slot is marked as *in-scope* and
165 //   the stack slot is overwritten with `undef`.
166 //   L3) on a `lifetime.end`, a stack slot is marked as *out-of-scope*.
167 //   L4) on function exit, all stack slots are marked as *out-of-scope*.
168 //   L5) `lifetime.end` is a no-op when called on a slot that is already
169 //   *out-of-scope*.
170 //   L6) memory accesses to *out-of-scope* stack slots are UB.
171 //   L7) when a stack-slot is marked as *out-of-scope*, all pointers to it
172 //   are invalidated, unless the slot is "degenerate". This is used to
173 //   justify not marking slots as in-use until the pointer to them is
174 //   used, but feels a bit hacky in the presence of things like LICM. See
175 //   the "Degenerate Slots" section for more details.
176 //
177 // Now, let's ground stack coloring on these rules. We'll define a slot
178 // as *in-use* at a (dynamic) point in execution if it either can be
179 // written to at that point, or if it has a live and non-undef content
180 // at that point.
181 //
182 // Obviously, slots that are never *in-use* together can be merged, and
183 // in our example `foo`, the slots for `x`, `y` and `z` are never
184 // in-use together (of course, sometimes slots that *are* in-use together
185 // might still be mergable, but we don't care about that here).
186 //
187 // In this implementation, we successively merge pairs of slots that are
188 // not *in-use* together. We could be smarter - for example, we could merge
189 // a single large slot with 2 small slots, or we could construct the
190 // interference graph and run a "smart" graph coloring algorithm, but with
191 // that aside, how do we find out whether a pair of slots might be *in-use*
192 // together?
193 //
194 // From our rules, we see that *out-of-scope* slots are never *in-use*,
195 // and from (L7) we see that "non-degenerate" slots remain non-*in-use*
196 // until their address is taken. Therefore, we can approximate slot activity
197 // using dataflow.
198 //
199 // A subtle point: naively, we might try to figure out which pairs of
200 // stack-slots interfere by propagating `S in-use` through the CFG for every
201 // stack-slot `S`, and having `S` and `T` interfere if there is a CFG point in
202 // which they are both *in-use*.
203 //
204 // That is sound, but overly conservative in some cases: in our (artificial)
205 // example `foo`, either `x` or `y` might be in use at the label `B:`, but
206 // as `x` is only in use if we came in from the `var` edge and `y` only
207 // if we came from the `!var` edge, they still can't be in use together.
208 // See PR32488 for an important real-life case.
209 //
210 // If we wanted to find all points of interference precisely, we could
211 // propagate `S in-use` and `S&T in-use` predicates through the CFG. That
212 // would be precise, but requires propagating `O(n^2)` dataflow facts.
213 //
214 // However, we aren't interested in the *set* of points of interference
215 // between 2 stack slots, only *whether* there *is* such a point. So we
216 // can rely on a little trick: for `S` and `T` to be in-use together,
217 // one of them needs to become in-use while the other is in-use (or
218 // they might both become in use simultaneously). We can check this
219 // by also keeping track of the points at which a stack slot might *start*
220 // being in-use.
221 //
222 // Exact first use:
223 // ----------------
224 //
225 // Consider the following motivating example:
226 //
227 //     int foo() {
228 //       char b1[1024], b2[1024];
229 //       if (...) {
230 //         char b3[1024];
231 //         <uses of b1, b3>;
232 //         return x;
233 //       } else {
234 //         char b4[1024], b5[1024];
235 //         <uses of b2, b4, b5>;
236 //         return y;
237 //       }
238 //     }
239 //
240 // In the code above, "b3" and "b4" are declared in distinct lexical
241 // scopes, meaning that it is easy to prove that they can share the
242 // same stack slot. Variables "b1" and "b2" are declared in the same
243 // scope, meaning that from a lexical point of view, their lifetimes
244 // overlap. From a control flow pointer of view, however, the two
245 // variables are accessed in disjoint regions of the CFG, thus it
246 // should be possible for them to share the same stack slot. An ideal
247 // stack allocation for the function above would look like:
248 //
249 //     slot 0: b1, b2
250 //     slot 1: b3, b4
251 //     slot 2: b5
252 //
253 // Achieving this allocation is tricky, however, due to the way
254 // lifetime markers are inserted. Here is a simplified view of the
255 // control flow graph for the code above:
256 //
257 //                +------  block 0 -------+
258 //               0| LIFETIME_START b1, b2 |
259 //               1| <test 'if' condition> |
260 //                +-----------------------+
261 //                   ./              \.
262 //   +------  block 1 -------+   +------  block 2 -------+
263 //  2| LIFETIME_START b3     |  5| LIFETIME_START b4, b5 |
264 //  3| <uses of b1, b3>      |  6| <uses of b2, b4, b5>  |
265 //  4| LIFETIME_END b3       |  7| LIFETIME_END b4, b5   |
266 //   +-----------------------+   +-----------------------+
267 //                   \.              /.
268 //                +------  block 3 -------+
269 //               8| <cleanupcode>         |
270 //               9| LIFETIME_END b1, b2   |
271 //              10| return                |
272 //                +-----------------------+
273 //
274 // If we create live intervals for the variables above strictly based
275 // on the lifetime markers, we'll get the set of intervals on the
276 // left. If we ignore the lifetime start markers and instead treat a
277 // variable's lifetime as beginning with the first reference to the
278 // var, then we get the intervals on the right.
279 //
280 //            LIFETIME_START      First Use
281 //     b1:    [0,9]               [3,4] [8,9]
282 //     b2:    [0,9]               [6,9]
283 //     b3:    [2,4]               [3,4]
284 //     b4:    [5,7]               [6,7]
285 //     b5:    [5,7]               [6,7]
286 //
287 // For the intervals on the left, the best we can do is overlap two
288 // variables (b3 and b4, for example); this gives us a stack size of
289 // 4*1024 bytes, not ideal. When treating first-use as the start of a
290 // lifetime, we can additionally overlap b1 and b5, giving us a 3*1024
291 // byte stack (better).
292 //
293 // Degenerate Slots:
294 // -----------------
295 //
296 // Relying entirely on first-use of stack slots is problematic,
297 // however, due to the fact that optimizations can sometimes migrate
298 // uses of a variable outside of its lifetime start/end region. Here
299 // is an example:
300 //
301 //     int bar() {
302 //       char b1[1024], b2[1024];
303 //       if (...) {
304 //         <uses of b2>
305 //         return y;
306 //       } else {
307 //         <uses of b1>
308 //         while (...) {
309 //           char b3[1024];
310 //           <uses of b3>
311 //         }
312 //       }
313 //     }
314 //
315 // Before optimization, the control flow graph for the code above
316 // might look like the following:
317 //
318 //                +------  block 0 -------+
319 //               0| LIFETIME_START b1, b2 |
320 //               1| <test 'if' condition> |
321 //                +-----------------------+
322 //                   ./              \.
323 //   +------  block 1 -------+    +------- block 2 -------+
324 //  2| <uses of b2>          |   3| <uses of b1>          |
325 //   +-----------------------+    +-----------------------+
326 //              |                            |
327 //              |                 +------- block 3 -------+ <-\.
328 //              |                4| <while condition>     |    |
329 //              |                 +-----------------------+    |
330 //              |               /          |                   |
331 //              |              /  +------- block 4 -------+
332 //              \             /  5| LIFETIME_START b3     |    |
333 //               \           /   6| <uses of b3>          |    |
334 //                \         /    7| LIFETIME_END b3       |    |
335 //                 \        |    +------------------------+    |
336 //                  \       |                 \                /
337 //                +------  block 5 -----+      \---------------
338 //               8| <cleanupcode>       |
339 //               9| LIFETIME_END b1, b2 |
340 //              10| return              |
341 //                +---------------------+
342 //
343 // During optimization, however, it can happen that an instruction
344 // computing an address in "b3" (for example, a loop-invariant GEP) is
345 // hoisted up out of the loop from block 4 to block 2.  [Note that
346 // this is not an actual load from the stack, only an instruction that
347 // computes the address to be loaded]. If this happens, there is now a
348 // path leading from the first use of b3 to the return instruction
349 // that does not encounter the b3 LIFETIME_END, hence b3's lifetime is
350 // now larger than if we were computing live intervals strictly based
351 // on lifetime markers. In the example above, this lengthened lifetime
352 // would mean that it would appear illegal to overlap b3 with b2.
353 //
354 // To deal with this such cases, the code in ::collectMarkers() below
355 // tries to identify "degenerate" slots -- those slots where on a single
356 // forward pass through the CFG we encounter a first reference to slot
357 // K before we hit the slot K lifetime start marker. For such slots,
358 // we fall back on using the lifetime start marker as the beginning of
359 // the variable's lifetime.  NB: with this implementation, slots can
360 // appear degenerate in cases where there is unstructured control flow:
361 //
362 //    if (q) goto mid;
363 //    if (x > 9) {
364 //         int b[100];
365 //         memcpy(&b[0], ...);
366 //    mid: b[k] = ...;
367 //         abc(&b);
368 //    }
369 //
370 // If in RPO ordering chosen to walk the CFG  we happen to visit the b[k]
371 // before visiting the memcpy block (which will contain the lifetime start
372 // for "b" then it will appear that 'b' has a degenerate lifetime.
373 
374 namespace {
375 
376 /// StackColoring - A machine pass for merging disjoint stack allocations,
377 /// marked by the LIFETIME_START and LIFETIME_END pseudo instructions.
378 class StackColoring : public MachineFunctionPass {
379   MachineFrameInfo *MFI = nullptr;
380   MachineFunction *MF = nullptr;
381 
382   /// A class representing liveness information for a single basic block.
383   /// Each bit in the BitVector represents the liveness property
384   /// for a different stack slot.
385   struct BlockLifetimeInfo {
386     /// Which slots BEGINs in each basic block.
387     BitVector Begin;
388 
389     /// Which slots ENDs in each basic block.
390     BitVector End;
391 
392     /// Which slots are marked as LIVE_IN, coming into each basic block.
393     BitVector LiveIn;
394 
395     /// Which slots are marked as LIVE_OUT, coming out of each basic block.
396     BitVector LiveOut;
397   };
398 
399   /// Maps active slots (per bit) for each basic block.
400   using LivenessMap = DenseMap<const MachineBasicBlock *, BlockLifetimeInfo>;
401   LivenessMap BlockLiveness;
402 
403   /// Maps serial numbers to basic blocks.
404   DenseMap<const MachineBasicBlock *, int> BasicBlocks;
405 
406   /// Maps basic blocks to a serial number.
407   SmallVector<const MachineBasicBlock *, 8> BasicBlockNumbering;
408 
409   /// Maps slots to their use interval. Outside of this interval, slots
410   /// values are either dead or `undef` and they will not be written to.
411   SmallVector<std::unique_ptr<LiveInterval>, 16> Intervals;
412 
413   /// Maps slots to the points where they can become in-use.
414   SmallVector<SmallVector<SlotIndex, 4>, 16> LiveStarts;
415 
416   /// VNInfo is used for the construction of LiveIntervals.
417   VNInfo::Allocator VNInfoAllocator;
418 
419   /// SlotIndex analysis object.
420   SlotIndexes *Indexes = nullptr;
421 
422   /// The list of lifetime markers found. These markers are to be removed
423   /// once the coloring is done.
424   SmallVector<MachineInstr*, 8> Markers;
425 
426   /// Record the FI slots for which we have seen some sort of
427   /// lifetime marker (either start or end).
428   BitVector InterestingSlots;
429 
430   /// FI slots that need to be handled conservatively (for these
431   /// slots lifetime-start-on-first-use is disabled).
432   BitVector ConservativeSlots;
433 
434   /// Number of iterations taken during data flow analysis.
435   unsigned NumIterations;
436 
437 public:
438   static char ID;
439 
440   StackColoring() : MachineFunctionPass(ID) {
441     initializeStackColoringPass(*PassRegistry::getPassRegistry());
442   }
443 
444   void getAnalysisUsage(AnalysisUsage &AU) const override;
445   bool runOnMachineFunction(MachineFunction &Func) override;
446 
447 private:
448   /// Used in collectMarkers
449   using BlockBitVecMap = DenseMap<const MachineBasicBlock *, BitVector>;
450 
451   /// Debug.
452   void dump() const;
453   void dumpIntervals() const;
454   void dumpBB(MachineBasicBlock *MBB) const;
455   void dumpBV(const char *tag, const BitVector &BV) const;
456 
457   /// Removes all of the lifetime marker instructions from the function.
458   /// \returns true if any markers were removed.
459   bool removeAllMarkers();
460 
461   /// Scan the machine function and find all of the lifetime markers.
462   /// Record the findings in the BEGIN and END vectors.
463   /// \returns the number of markers found.
464   unsigned collectMarkers(unsigned NumSlot);
465 
466   /// Perform the dataflow calculation and calculate the lifetime for each of
467   /// the slots, based on the BEGIN/END vectors. Set the LifetimeLIVE_IN and
468   /// LifetimeLIVE_OUT maps that represent which stack slots are live coming
469   /// in and out blocks.
470   void calculateLocalLiveness();
471 
472   /// Returns TRUE if we're using the first-use-begins-lifetime method for
473   /// this slot (if FALSE, then the start marker is treated as start of lifetime).
474   bool applyFirstUse(int Slot) {
475     if (!LifetimeStartOnFirstUse || ProtectFromEscapedAllocas)
476       return false;
477     if (ConservativeSlots.test(Slot))
478       return false;
479     return true;
480   }
481 
482   /// Examines the specified instruction and returns TRUE if the instruction
483   /// represents the start or end of an interesting lifetime. The slot or slots
484   /// starting or ending are added to the vector "slots" and "isStart" is set
485   /// accordingly.
486   /// \returns True if inst contains a lifetime start or end
487   bool isLifetimeStartOrEnd(const MachineInstr &MI,
488                             SmallVector<int, 4> &slots,
489                             bool &isStart);
490 
491   /// Construct the LiveIntervals for the slots.
492   void calculateLiveIntervals(unsigned NumSlots);
493 
494   /// Go over the machine function and change instructions which use stack
495   /// slots to use the joint slots.
496   void remapInstructions(DenseMap<int, int> &SlotRemap);
497 
498   /// The input program may contain instructions which are not inside lifetime
499   /// markers. This can happen due to a bug in the compiler or due to a bug in
500   /// user code (for example, returning a reference to a local variable).
501   /// This procedure checks all of the instructions in the function and
502   /// invalidates lifetime ranges which do not contain all of the instructions
503   /// which access that frame slot.
504   void removeInvalidSlotRanges();
505 
506   /// Map entries which point to other entries to their destination.
507   ///   A->B->C becomes A->C.
508   void expungeSlotMap(DenseMap<int, int> &SlotRemap, unsigned NumSlots);
509 };
510 
511 } // end anonymous namespace
512 
513 char StackColoring::ID = 0;
514 
515 char &llvm::StackColoringID = StackColoring::ID;
516 
517 INITIALIZE_PASS_BEGIN(StackColoring, DEBUG_TYPE,
518                       "Merge disjoint stack slots", false, false)
519 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
520 INITIALIZE_PASS_END(StackColoring, DEBUG_TYPE,
521                     "Merge disjoint stack slots", false, false)
522 
523 void StackColoring::getAnalysisUsage(AnalysisUsage &AU) const {
524   AU.addRequired<SlotIndexes>();
525   MachineFunctionPass::getAnalysisUsage(AU);
526 }
527 
528 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
529 LLVM_DUMP_METHOD void StackColoring::dumpBV(const char *tag,
530                                             const BitVector &BV) const {
531   dbgs() << tag << " : { ";
532   for (unsigned I = 0, E = BV.size(); I != E; ++I)
533     dbgs() << BV.test(I) << " ";
534   dbgs() << "}\n";
535 }
536 
537 LLVM_DUMP_METHOD void StackColoring::dumpBB(MachineBasicBlock *MBB) const {
538   LivenessMap::const_iterator BI = BlockLiveness.find(MBB);
539   assert(BI != BlockLiveness.end() && "Block not found");
540   const BlockLifetimeInfo &BlockInfo = BI->second;
541 
542   dumpBV("BEGIN", BlockInfo.Begin);
543   dumpBV("END", BlockInfo.End);
544   dumpBV("LIVE_IN", BlockInfo.LiveIn);
545   dumpBV("LIVE_OUT", BlockInfo.LiveOut);
546 }
547 
548 LLVM_DUMP_METHOD void StackColoring::dump() const {
549   for (MachineBasicBlock *MBB : depth_first(MF)) {
550     dbgs() << "Inspecting block #" << MBB->getNumber() << " ["
551            << MBB->getName() << "]\n";
552     dumpBB(MBB);
553   }
554 }
555 
556 LLVM_DUMP_METHOD void StackColoring::dumpIntervals() const {
557   for (unsigned I = 0, E = Intervals.size(); I != E; ++I) {
558     dbgs() << "Interval[" << I << "]:\n";
559     Intervals[I]->dump();
560   }
561 }
562 #endif
563 
564 static inline int getStartOrEndSlot(const MachineInstr &MI)
565 {
566   assert((MI.getOpcode() == TargetOpcode::LIFETIME_START ||
567           MI.getOpcode() == TargetOpcode::LIFETIME_END) &&
568          "Expected LIFETIME_START or LIFETIME_END op");
569   const MachineOperand &MO = MI.getOperand(0);
570   int Slot = MO.getIndex();
571   if (Slot >= 0)
572     return Slot;
573   return -1;
574 }
575 
576 // At the moment the only way to end a variable lifetime is with
577 // a VARIABLE_LIFETIME op (which can't contain a start). If things
578 // change and the IR allows for a single inst that both begins
579 // and ends lifetime(s), this interface will need to be reworked.
580 bool StackColoring::isLifetimeStartOrEnd(const MachineInstr &MI,
581                                          SmallVector<int, 4> &slots,
582                                          bool &isStart) {
583   if (MI.getOpcode() == TargetOpcode::LIFETIME_START ||
584       MI.getOpcode() == TargetOpcode::LIFETIME_END) {
585     int Slot = getStartOrEndSlot(MI);
586     if (Slot < 0)
587       return false;
588     if (!InterestingSlots.test(Slot))
589       return false;
590     slots.push_back(Slot);
591     if (MI.getOpcode() == TargetOpcode::LIFETIME_END) {
592       isStart = false;
593       return true;
594     }
595     if (!applyFirstUse(Slot)) {
596       isStart = true;
597       return true;
598     }
599   } else if (LifetimeStartOnFirstUse && !ProtectFromEscapedAllocas) {
600     if (!MI.isDebugInstr()) {
601       bool found = false;
602       for (const MachineOperand &MO : MI.operands()) {
603         if (!MO.isFI())
604           continue;
605         int Slot = MO.getIndex();
606         if (Slot<0)
607           continue;
608         if (InterestingSlots.test(Slot) && applyFirstUse(Slot)) {
609           slots.push_back(Slot);
610           found = true;
611         }
612       }
613       if (found) {
614         isStart = true;
615         return true;
616       }
617     }
618   }
619   return false;
620 }
621 
622 unsigned StackColoring::collectMarkers(unsigned NumSlot) {
623   unsigned MarkersFound = 0;
624   BlockBitVecMap SeenStartMap;
625   InterestingSlots.clear();
626   InterestingSlots.resize(NumSlot);
627   ConservativeSlots.clear();
628   ConservativeSlots.resize(NumSlot);
629 
630   // number of start and end lifetime ops for each slot
631   SmallVector<int, 8> NumStartLifetimes(NumSlot, 0);
632   SmallVector<int, 8> NumEndLifetimes(NumSlot, 0);
633 
634   // Step 1: collect markers and populate the "InterestingSlots"
635   // and "ConservativeSlots" sets.
636   for (MachineBasicBlock *MBB : depth_first(MF)) {
637     // Compute the set of slots for which we've seen a START marker but have
638     // not yet seen an END marker at this point in the walk (e.g. on entry
639     // to this bb).
640     BitVector BetweenStartEnd;
641     BetweenStartEnd.resize(NumSlot);
642     for (const MachineBasicBlock *Pred : MBB->predecessors()) {
643       BlockBitVecMap::const_iterator I = SeenStartMap.find(Pred);
644       if (I != SeenStartMap.end()) {
645         BetweenStartEnd |= I->second;
646       }
647     }
648 
649     // Walk the instructions in the block to look for start/end ops.
650     for (MachineInstr &MI : *MBB) {
651       if (MI.isDebugInstr())
652         continue;
653       if (MI.getOpcode() == TargetOpcode::LIFETIME_START ||
654           MI.getOpcode() == TargetOpcode::LIFETIME_END) {
655         int Slot = getStartOrEndSlot(MI);
656         if (Slot < 0)
657           continue;
658         InterestingSlots.set(Slot);
659         if (MI.getOpcode() == TargetOpcode::LIFETIME_START) {
660           BetweenStartEnd.set(Slot);
661           NumStartLifetimes[Slot] += 1;
662         } else {
663           BetweenStartEnd.reset(Slot);
664           NumEndLifetimes[Slot] += 1;
665         }
666         const AllocaInst *Allocation = MFI->getObjectAllocation(Slot);
667         if (Allocation) {
668           LLVM_DEBUG(dbgs() << "Found a lifetime ");
669           LLVM_DEBUG(dbgs() << (MI.getOpcode() == TargetOpcode::LIFETIME_START
670                                     ? "start"
671                                     : "end"));
672           LLVM_DEBUG(dbgs() << " marker for slot #" << Slot);
673           LLVM_DEBUG(dbgs()
674                      << " with allocation: " << Allocation->getName() << "\n");
675         }
676         Markers.push_back(&MI);
677         MarkersFound += 1;
678       } else {
679         for (const MachineOperand &MO : MI.operands()) {
680           if (!MO.isFI())
681             continue;
682           int Slot = MO.getIndex();
683           if (Slot < 0)
684             continue;
685           if (! BetweenStartEnd.test(Slot)) {
686             ConservativeSlots.set(Slot);
687           }
688         }
689       }
690     }
691     BitVector &SeenStart = SeenStartMap[MBB];
692     SeenStart |= BetweenStartEnd;
693   }
694   if (!MarkersFound) {
695     return 0;
696   }
697 
698   // PR27903: slots with multiple start or end lifetime ops are not
699   // safe to enable for "lifetime-start-on-first-use".
700   for (unsigned slot = 0; slot < NumSlot; ++slot) {
701     if (NumStartLifetimes[slot] > 1 || NumEndLifetimes[slot] > 1)
702       ConservativeSlots.set(slot);
703   }
704 
705   // The write to the catch object by the personality function is not propely
706   // modeled in IR: It happens before any cleanuppads are executed, even if the
707   // first mention of the catch object is in a catchpad. As such, mark catch
708   // object slots as conservative, so they are excluded from first-use analysis.
709   if (WinEHFuncInfo *EHInfo = MF->getWinEHFuncInfo())
710     for (WinEHTryBlockMapEntry &TBME : EHInfo->TryBlockMap)
711       for (WinEHHandlerType &H : TBME.HandlerArray)
712         if (H.CatchObj.FrameIndex != std::numeric_limits<int>::max() &&
713             H.CatchObj.FrameIndex >= 0)
714           ConservativeSlots.set(H.CatchObj.FrameIndex);
715 
716   LLVM_DEBUG(dumpBV("Conservative slots", ConservativeSlots));
717 
718   // Step 2: compute begin/end sets for each block
719 
720   // NOTE: We use a depth-first iteration to ensure that we obtain a
721   // deterministic numbering.
722   for (MachineBasicBlock *MBB : depth_first(MF)) {
723     // Assign a serial number to this basic block.
724     BasicBlocks[MBB] = BasicBlockNumbering.size();
725     BasicBlockNumbering.push_back(MBB);
726 
727     // Keep a reference to avoid repeated lookups.
728     BlockLifetimeInfo &BlockInfo = BlockLiveness[MBB];
729 
730     BlockInfo.Begin.resize(NumSlot);
731     BlockInfo.End.resize(NumSlot);
732 
733     SmallVector<int, 4> slots;
734     for (MachineInstr &MI : *MBB) {
735       bool isStart = false;
736       slots.clear();
737       if (isLifetimeStartOrEnd(MI, slots, isStart)) {
738         if (!isStart) {
739           assert(slots.size() == 1 && "unexpected: MI ends multiple slots");
740           int Slot = slots[0];
741           if (BlockInfo.Begin.test(Slot)) {
742             BlockInfo.Begin.reset(Slot);
743           }
744           BlockInfo.End.set(Slot);
745         } else {
746           for (auto Slot : slots) {
747             LLVM_DEBUG(dbgs() << "Found a use of slot #" << Slot);
748             LLVM_DEBUG(dbgs()
749                        << " at " << printMBBReference(*MBB) << " index ");
750             LLVM_DEBUG(Indexes->getInstructionIndex(MI).print(dbgs()));
751             const AllocaInst *Allocation = MFI->getObjectAllocation(Slot);
752             if (Allocation) {
753               LLVM_DEBUG(dbgs()
754                          << " with allocation: " << Allocation->getName());
755             }
756             LLVM_DEBUG(dbgs() << "\n");
757             if (BlockInfo.End.test(Slot)) {
758               BlockInfo.End.reset(Slot);
759             }
760             BlockInfo.Begin.set(Slot);
761           }
762         }
763       }
764     }
765   }
766 
767   // Update statistics.
768   NumMarkerSeen += MarkersFound;
769   return MarkersFound;
770 }
771 
772 void StackColoring::calculateLocalLiveness() {
773   unsigned NumIters = 0;
774   bool changed = true;
775   while (changed) {
776     changed = false;
777     ++NumIters;
778 
779     for (const MachineBasicBlock *BB : BasicBlockNumbering) {
780       // Use an iterator to avoid repeated lookups.
781       LivenessMap::iterator BI = BlockLiveness.find(BB);
782       assert(BI != BlockLiveness.end() && "Block not found");
783       BlockLifetimeInfo &BlockInfo = BI->second;
784 
785       // Compute LiveIn by unioning together the LiveOut sets of all preds.
786       BitVector LocalLiveIn;
787       for (MachineBasicBlock *Pred : BB->predecessors()) {
788         LivenessMap::const_iterator I = BlockLiveness.find(Pred);
789         // PR37130: transformations prior to stack coloring can
790         // sometimes leave behind statically unreachable blocks; these
791         // can be safely skipped here.
792         if (I != BlockLiveness.end())
793           LocalLiveIn |= I->second.LiveOut;
794       }
795 
796       // Compute LiveOut by subtracting out lifetimes that end in this
797       // block, then adding in lifetimes that begin in this block.  If
798       // we have both BEGIN and END markers in the same basic block
799       // then we know that the BEGIN marker comes after the END,
800       // because we already handle the case where the BEGIN comes
801       // before the END when collecting the markers (and building the
802       // BEGIN/END vectors).
803       BitVector LocalLiveOut = LocalLiveIn;
804       LocalLiveOut.reset(BlockInfo.End);
805       LocalLiveOut |= BlockInfo.Begin;
806 
807       // Update block LiveIn set, noting whether it has changed.
808       if (LocalLiveIn.test(BlockInfo.LiveIn)) {
809         changed = true;
810         BlockInfo.LiveIn |= LocalLiveIn;
811       }
812 
813       // Update block LiveOut set, noting whether it has changed.
814       if (LocalLiveOut.test(BlockInfo.LiveOut)) {
815         changed = true;
816         BlockInfo.LiveOut |= LocalLiveOut;
817       }
818     }
819   } // while changed.
820 
821   NumIterations = NumIters;
822 }
823 
824 void StackColoring::calculateLiveIntervals(unsigned NumSlots) {
825   SmallVector<SlotIndex, 16> Starts;
826   SmallVector<bool, 16> DefinitelyInUse;
827 
828   // For each block, find which slots are active within this block
829   // and update the live intervals.
830   for (const MachineBasicBlock &MBB : *MF) {
831     Starts.clear();
832     Starts.resize(NumSlots);
833     DefinitelyInUse.clear();
834     DefinitelyInUse.resize(NumSlots);
835 
836     // Start the interval of the slots that we previously found to be 'in-use'.
837     BlockLifetimeInfo &MBBLiveness = BlockLiveness[&MBB];
838     for (int pos = MBBLiveness.LiveIn.find_first(); pos != -1;
839          pos = MBBLiveness.LiveIn.find_next(pos)) {
840       Starts[pos] = Indexes->getMBBStartIdx(&MBB);
841     }
842 
843     // Create the interval for the basic blocks containing lifetime begin/end.
844     for (const MachineInstr &MI : MBB) {
845       SmallVector<int, 4> slots;
846       bool IsStart = false;
847       if (!isLifetimeStartOrEnd(MI, slots, IsStart))
848         continue;
849       SlotIndex ThisIndex = Indexes->getInstructionIndex(MI);
850       for (auto Slot : slots) {
851         if (IsStart) {
852           // If a slot is already definitely in use, we don't have to emit
853           // a new start marker because there is already a pre-existing
854           // one.
855           if (!DefinitelyInUse[Slot]) {
856             LiveStarts[Slot].push_back(ThisIndex);
857             DefinitelyInUse[Slot] = true;
858           }
859           if (!Starts[Slot].isValid())
860             Starts[Slot] = ThisIndex;
861         } else {
862           if (Starts[Slot].isValid()) {
863             VNInfo *VNI = Intervals[Slot]->getValNumInfo(0);
864             Intervals[Slot]->addSegment(
865                 LiveInterval::Segment(Starts[Slot], ThisIndex, VNI));
866             Starts[Slot] = SlotIndex(); // Invalidate the start index
867             DefinitelyInUse[Slot] = false;
868           }
869         }
870       }
871     }
872 
873     // Finish up started segments
874     for (unsigned i = 0; i < NumSlots; ++i) {
875       if (!Starts[i].isValid())
876         continue;
877 
878       SlotIndex EndIdx = Indexes->getMBBEndIdx(&MBB);
879       VNInfo *VNI = Intervals[i]->getValNumInfo(0);
880       Intervals[i]->addSegment(LiveInterval::Segment(Starts[i], EndIdx, VNI));
881     }
882   }
883 }
884 
885 bool StackColoring::removeAllMarkers() {
886   unsigned Count = 0;
887   for (MachineInstr *MI : Markers) {
888     MI->eraseFromParent();
889     Count++;
890   }
891   Markers.clear();
892 
893   LLVM_DEBUG(dbgs() << "Removed " << Count << " markers.\n");
894   return Count;
895 }
896 
897 void StackColoring::remapInstructions(DenseMap<int, int> &SlotRemap) {
898   unsigned FixedInstr = 0;
899   unsigned FixedMemOp = 0;
900   unsigned FixedDbg = 0;
901 
902   // Remap debug information that refers to stack slots.
903   for (auto &VI : MF->getVariableDbgInfo()) {
904     if (!VI.Var || !VI.inStackSlot())
905       continue;
906     int Slot = VI.getStackSlot();
907     if (SlotRemap.count(Slot)) {
908       LLVM_DEBUG(dbgs() << "Remapping debug info for ["
909                         << cast<DILocalVariable>(VI.Var)->getName() << "].\n");
910       VI.updateStackSlot(SlotRemap[Slot]);
911       FixedDbg++;
912     }
913   }
914 
915   // Keep a list of *allocas* which need to be remapped.
916   DenseMap<const AllocaInst*, const AllocaInst*> Allocas;
917 
918   // Keep a list of allocas which has been affected by the remap.
919   SmallPtrSet<const AllocaInst*, 32> MergedAllocas;
920 
921   for (const std::pair<int, int> &SI : SlotRemap) {
922     const AllocaInst *From = MFI->getObjectAllocation(SI.first);
923     const AllocaInst *To = MFI->getObjectAllocation(SI.second);
924     assert(To && From && "Invalid allocation object");
925     Allocas[From] = To;
926 
927     // If From is before wo, its possible that there is a use of From between
928     // them.
929     if (From->comesBefore(To))
930       const_cast<AllocaInst*>(To)->moveBefore(const_cast<AllocaInst*>(From));
931 
932     // AA might be used later for instruction scheduling, and we need it to be
933     // able to deduce the correct aliasing releationships between pointers
934     // derived from the alloca being remapped and the target of that remapping.
935     // The only safe way, without directly informing AA about the remapping
936     // somehow, is to directly update the IR to reflect the change being made
937     // here.
938     Instruction *Inst = const_cast<AllocaInst *>(To);
939     if (From->getType() != To->getType()) {
940       BitCastInst *Cast = new BitCastInst(Inst, From->getType());
941       Cast->insertAfter(Inst);
942       Inst = Cast;
943     }
944 
945     // We keep both slots to maintain AliasAnalysis metadata later.
946     MergedAllocas.insert(From);
947     MergedAllocas.insert(To);
948 
949     // Transfer the stack protector layout tag, but make sure that SSPLK_AddrOf
950     // does not overwrite SSPLK_SmallArray or SSPLK_LargeArray, and make sure
951     // that SSPLK_SmallArray does not overwrite SSPLK_LargeArray.
952     MachineFrameInfo::SSPLayoutKind FromKind
953         = MFI->getObjectSSPLayout(SI.first);
954     MachineFrameInfo::SSPLayoutKind ToKind = MFI->getObjectSSPLayout(SI.second);
955     if (FromKind != MachineFrameInfo::SSPLK_None &&
956         (ToKind == MachineFrameInfo::SSPLK_None ||
957          (ToKind != MachineFrameInfo::SSPLK_LargeArray &&
958           FromKind != MachineFrameInfo::SSPLK_AddrOf)))
959       MFI->setObjectSSPLayout(SI.second, FromKind);
960 
961     // The new alloca might not be valid in a llvm.dbg.declare for this
962     // variable, so undef out the use to make the verifier happy.
963     AllocaInst *FromAI = const_cast<AllocaInst *>(From);
964     if (FromAI->isUsedByMetadata())
965       ValueAsMetadata::handleRAUW(FromAI, UndefValue::get(FromAI->getType()));
966     for (auto &Use : FromAI->uses()) {
967       if (BitCastInst *BCI = dyn_cast<BitCastInst>(Use.get()))
968         if (BCI->isUsedByMetadata())
969           ValueAsMetadata::handleRAUW(BCI, UndefValue::get(BCI->getType()));
970     }
971 
972     // Note that this will not replace uses in MMOs (which we'll update below),
973     // or anywhere else (which is why we won't delete the original
974     // instruction).
975     FromAI->replaceAllUsesWith(Inst);
976   }
977 
978   // Remap all instructions to the new stack slots.
979   std::vector<std::vector<MachineMemOperand *>> SSRefs(
980       MFI->getObjectIndexEnd());
981   for (MachineBasicBlock &BB : *MF)
982     for (MachineInstr &I : BB) {
983       // Skip lifetime markers. We'll remove them soon.
984       if (I.getOpcode() == TargetOpcode::LIFETIME_START ||
985           I.getOpcode() == TargetOpcode::LIFETIME_END)
986         continue;
987 
988       // Update the MachineMemOperand to use the new alloca.
989       for (MachineMemOperand *MMO : I.memoperands()) {
990         // We've replaced IR-level uses of the remapped allocas, so we only
991         // need to replace direct uses here.
992         const AllocaInst *AI = dyn_cast_or_null<AllocaInst>(MMO->getValue());
993         if (!AI)
994           continue;
995 
996         if (!Allocas.count(AI))
997           continue;
998 
999         MMO->setValue(Allocas[AI]);
1000         FixedMemOp++;
1001       }
1002 
1003       // Update all of the machine instruction operands.
1004       for (MachineOperand &MO : I.operands()) {
1005         if (!MO.isFI())
1006           continue;
1007         int FromSlot = MO.getIndex();
1008 
1009         // Don't touch arguments.
1010         if (FromSlot<0)
1011           continue;
1012 
1013         // Only look at mapped slots.
1014         if (!SlotRemap.count(FromSlot))
1015           continue;
1016 
1017         // In a debug build, check that the instruction that we are modifying is
1018         // inside the expected live range. If the instruction is not inside
1019         // the calculated range then it means that the alloca usage moved
1020         // outside of the lifetime markers, or that the user has a bug.
1021         // NOTE: Alloca address calculations which happen outside the lifetime
1022         // zone are okay, despite the fact that we don't have a good way
1023         // for validating all of the usages of the calculation.
1024 #ifndef NDEBUG
1025         bool TouchesMemory = I.mayLoadOrStore();
1026         // If we *don't* protect the user from escaped allocas, don't bother
1027         // validating the instructions.
1028         if (!I.isDebugInstr() && TouchesMemory && ProtectFromEscapedAllocas) {
1029           SlotIndex Index = Indexes->getInstructionIndex(I);
1030           const LiveInterval *Interval = &*Intervals[FromSlot];
1031           assert(Interval->find(Index) != Interval->end() &&
1032                  "Found instruction usage outside of live range.");
1033         }
1034 #endif
1035 
1036         // Fix the machine instructions.
1037         int ToSlot = SlotRemap[FromSlot];
1038         MO.setIndex(ToSlot);
1039         FixedInstr++;
1040       }
1041 
1042       // We adjust AliasAnalysis information for merged stack slots.
1043       SmallVector<MachineMemOperand *, 2> NewMMOs;
1044       bool ReplaceMemOps = false;
1045       for (MachineMemOperand *MMO : I.memoperands()) {
1046         // Collect MachineMemOperands which reference
1047         // FixedStackPseudoSourceValues with old frame indices.
1048         if (const auto *FSV = dyn_cast_or_null<FixedStackPseudoSourceValue>(
1049                 MMO->getPseudoValue())) {
1050           int FI = FSV->getFrameIndex();
1051           auto To = SlotRemap.find(FI);
1052           if (To != SlotRemap.end())
1053             SSRefs[FI].push_back(MMO);
1054         }
1055 
1056         // If this memory location can be a slot remapped here,
1057         // we remove AA information.
1058         bool MayHaveConflictingAAMD = false;
1059         if (MMO->getAAInfo()) {
1060           if (const Value *MMOV = MMO->getValue()) {
1061             SmallVector<Value *, 4> Objs;
1062             getUnderlyingObjectsForCodeGen(MMOV, Objs);
1063 
1064             if (Objs.empty())
1065               MayHaveConflictingAAMD = true;
1066             else
1067               for (Value *V : Objs) {
1068                 // If this memory location comes from a known stack slot
1069                 // that is not remapped, we continue checking.
1070                 // Otherwise, we need to invalidate AA infomation.
1071                 const AllocaInst *AI = dyn_cast_or_null<AllocaInst>(V);
1072                 if (AI && MergedAllocas.count(AI)) {
1073                   MayHaveConflictingAAMD = true;
1074                   break;
1075                 }
1076               }
1077           }
1078         }
1079         if (MayHaveConflictingAAMD) {
1080           NewMMOs.push_back(MF->getMachineMemOperand(MMO, AAMDNodes()));
1081           ReplaceMemOps = true;
1082         } else {
1083           NewMMOs.push_back(MMO);
1084         }
1085       }
1086 
1087       // If any memory operand is updated, set memory references of
1088       // this instruction.
1089       if (ReplaceMemOps)
1090         I.setMemRefs(*MF, NewMMOs);
1091     }
1092 
1093   // Rewrite MachineMemOperands that reference old frame indices.
1094   for (auto E : enumerate(SSRefs))
1095     if (!E.value().empty()) {
1096       const PseudoSourceValue *NewSV =
1097           MF->getPSVManager().getFixedStack(SlotRemap.find(E.index())->second);
1098       for (MachineMemOperand *Ref : E.value())
1099         Ref->setValue(NewSV);
1100     }
1101 
1102   // Update the location of C++ catch objects for the MSVC personality routine.
1103   if (WinEHFuncInfo *EHInfo = MF->getWinEHFuncInfo())
1104     for (WinEHTryBlockMapEntry &TBME : EHInfo->TryBlockMap)
1105       for (WinEHHandlerType &H : TBME.HandlerArray)
1106         if (H.CatchObj.FrameIndex != std::numeric_limits<int>::max() &&
1107             SlotRemap.count(H.CatchObj.FrameIndex))
1108           H.CatchObj.FrameIndex = SlotRemap[H.CatchObj.FrameIndex];
1109 
1110   LLVM_DEBUG(dbgs() << "Fixed " << FixedMemOp << " machine memory operands.\n");
1111   LLVM_DEBUG(dbgs() << "Fixed " << FixedDbg << " debug locations.\n");
1112   LLVM_DEBUG(dbgs() << "Fixed " << FixedInstr << " machine instructions.\n");
1113   (void) FixedMemOp;
1114   (void) FixedDbg;
1115   (void) FixedInstr;
1116 }
1117 
1118 void StackColoring::removeInvalidSlotRanges() {
1119   for (MachineBasicBlock &BB : *MF)
1120     for (MachineInstr &I : BB) {
1121       if (I.getOpcode() == TargetOpcode::LIFETIME_START ||
1122           I.getOpcode() == TargetOpcode::LIFETIME_END || I.isDebugInstr())
1123         continue;
1124 
1125       // Some intervals are suspicious! In some cases we find address
1126       // calculations outside of the lifetime zone, but not actual memory
1127       // read or write. Memory accesses outside of the lifetime zone are a clear
1128       // violation, but address calculations are okay. This can happen when
1129       // GEPs are hoisted outside of the lifetime zone.
1130       // So, in here we only check instructions which can read or write memory.
1131       if (!I.mayLoad() && !I.mayStore())
1132         continue;
1133 
1134       // Check all of the machine operands.
1135       for (const MachineOperand &MO : I.operands()) {
1136         if (!MO.isFI())
1137           continue;
1138 
1139         int Slot = MO.getIndex();
1140 
1141         if (Slot<0)
1142           continue;
1143 
1144         if (Intervals[Slot]->empty())
1145           continue;
1146 
1147         // Check that the used slot is inside the calculated lifetime range.
1148         // If it is not, warn about it and invalidate the range.
1149         LiveInterval *Interval = &*Intervals[Slot];
1150         SlotIndex Index = Indexes->getInstructionIndex(I);
1151         if (Interval->find(Index) == Interval->end()) {
1152           Interval->clear();
1153           LLVM_DEBUG(dbgs() << "Invalidating range #" << Slot << "\n");
1154           EscapedAllocas++;
1155         }
1156       }
1157     }
1158 }
1159 
1160 void StackColoring::expungeSlotMap(DenseMap<int, int> &SlotRemap,
1161                                    unsigned NumSlots) {
1162   // Expunge slot remap map.
1163   for (unsigned i=0; i < NumSlots; ++i) {
1164     // If we are remapping i
1165     if (SlotRemap.count(i)) {
1166       int Target = SlotRemap[i];
1167       // As long as our target is mapped to something else, follow it.
1168       while (SlotRemap.count(Target)) {
1169         Target = SlotRemap[Target];
1170         SlotRemap[i] = Target;
1171       }
1172     }
1173   }
1174 }
1175 
1176 bool StackColoring::runOnMachineFunction(MachineFunction &Func) {
1177   LLVM_DEBUG(dbgs() << "********** Stack Coloring **********\n"
1178                     << "********** Function: " << Func.getName() << '\n');
1179   MF = &Func;
1180   MFI = &MF->getFrameInfo();
1181   Indexes = &getAnalysis<SlotIndexes>();
1182   BlockLiveness.clear();
1183   BasicBlocks.clear();
1184   BasicBlockNumbering.clear();
1185   Markers.clear();
1186   Intervals.clear();
1187   LiveStarts.clear();
1188   VNInfoAllocator.Reset();
1189 
1190   unsigned NumSlots = MFI->getObjectIndexEnd();
1191 
1192   // If there are no stack slots then there are no markers to remove.
1193   if (!NumSlots)
1194     return false;
1195 
1196   SmallVector<int, 8> SortedSlots;
1197   SortedSlots.reserve(NumSlots);
1198   Intervals.reserve(NumSlots);
1199   LiveStarts.resize(NumSlots);
1200 
1201   unsigned NumMarkers = collectMarkers(NumSlots);
1202 
1203   unsigned TotalSize = 0;
1204   LLVM_DEBUG(dbgs() << "Found " << NumMarkers << " markers and " << NumSlots
1205                     << " slots\n");
1206   LLVM_DEBUG(dbgs() << "Slot structure:\n");
1207 
1208   for (int i=0; i < MFI->getObjectIndexEnd(); ++i) {
1209     LLVM_DEBUG(dbgs() << "Slot #" << i << " - " << MFI->getObjectSize(i)
1210                       << " bytes.\n");
1211     TotalSize += MFI->getObjectSize(i);
1212   }
1213 
1214   LLVM_DEBUG(dbgs() << "Total Stack size: " << TotalSize << " bytes\n\n");
1215 
1216   // Don't continue because there are not enough lifetime markers, or the
1217   // stack is too small, or we are told not to optimize the slots.
1218   if (NumMarkers < 2 || TotalSize < 16 || DisableColoring ||
1219       skipFunction(Func.getFunction())) {
1220     LLVM_DEBUG(dbgs() << "Will not try to merge slots.\n");
1221     return removeAllMarkers();
1222   }
1223 
1224   for (unsigned i=0; i < NumSlots; ++i) {
1225     std::unique_ptr<LiveInterval> LI(new LiveInterval(i, 0));
1226     LI->getNextValue(Indexes->getZeroIndex(), VNInfoAllocator);
1227     Intervals.push_back(std::move(LI));
1228     SortedSlots.push_back(i);
1229   }
1230 
1231   // Calculate the liveness of each block.
1232   calculateLocalLiveness();
1233   LLVM_DEBUG(dbgs() << "Dataflow iterations: " << NumIterations << "\n");
1234   LLVM_DEBUG(dump());
1235 
1236   // Propagate the liveness information.
1237   calculateLiveIntervals(NumSlots);
1238   LLVM_DEBUG(dumpIntervals());
1239 
1240   // Search for allocas which are used outside of the declared lifetime
1241   // markers.
1242   if (ProtectFromEscapedAllocas)
1243     removeInvalidSlotRanges();
1244 
1245   // Maps old slots to new slots.
1246   DenseMap<int, int> SlotRemap;
1247   unsigned RemovedSlots = 0;
1248   unsigned ReducedSize = 0;
1249 
1250   // Do not bother looking at empty intervals.
1251   for (unsigned I = 0; I < NumSlots; ++I) {
1252     if (Intervals[SortedSlots[I]]->empty())
1253       SortedSlots[I] = -1;
1254   }
1255 
1256   // This is a simple greedy algorithm for merging allocas. First, sort the
1257   // slots, placing the largest slots first. Next, perform an n^2 scan and look
1258   // for disjoint slots. When you find disjoint slots, merge the smaller one
1259   // into the bigger one and update the live interval. Remove the small alloca
1260   // and continue.
1261 
1262   // Sort the slots according to their size. Place unused slots at the end.
1263   // Use stable sort to guarantee deterministic code generation.
1264   llvm::stable_sort(SortedSlots, [this](int LHS, int RHS) {
1265     // We use -1 to denote a uninteresting slot. Place these slots at the end.
1266     if (LHS == -1)
1267       return false;
1268     if (RHS == -1)
1269       return true;
1270     // Sort according to size.
1271     return MFI->getObjectSize(LHS) > MFI->getObjectSize(RHS);
1272   });
1273 
1274   for (auto &s : LiveStarts)
1275     llvm::sort(s);
1276 
1277   bool Changed = true;
1278   while (Changed) {
1279     Changed = false;
1280     for (unsigned I = 0; I < NumSlots; ++I) {
1281       if (SortedSlots[I] == -1)
1282         continue;
1283 
1284       for (unsigned J=I+1; J < NumSlots; ++J) {
1285         if (SortedSlots[J] == -1)
1286           continue;
1287 
1288         int FirstSlot = SortedSlots[I];
1289         int SecondSlot = SortedSlots[J];
1290 
1291         // Objects with different stack IDs cannot be merged.
1292         if (MFI->getStackID(FirstSlot) != MFI->getStackID(SecondSlot))
1293           continue;
1294 
1295         LiveInterval *First = &*Intervals[FirstSlot];
1296         LiveInterval *Second = &*Intervals[SecondSlot];
1297         auto &FirstS = LiveStarts[FirstSlot];
1298         auto &SecondS = LiveStarts[SecondSlot];
1299         assert(!First->empty() && !Second->empty() && "Found an empty range");
1300 
1301         // Merge disjoint slots. This is a little bit tricky - see the
1302         // Implementation Notes section for an explanation.
1303         if (!First->isLiveAtIndexes(SecondS) &&
1304             !Second->isLiveAtIndexes(FirstS)) {
1305           Changed = true;
1306           First->MergeSegmentsInAsValue(*Second, First->getValNumInfo(0));
1307 
1308           int OldSize = FirstS.size();
1309           FirstS.append(SecondS.begin(), SecondS.end());
1310           auto Mid = FirstS.begin() + OldSize;
1311           std::inplace_merge(FirstS.begin(), Mid, FirstS.end());
1312 
1313           SlotRemap[SecondSlot] = FirstSlot;
1314           SortedSlots[J] = -1;
1315           LLVM_DEBUG(dbgs() << "Merging #" << FirstSlot << " and slots #"
1316                             << SecondSlot << " together.\n");
1317           Align MaxAlignment = std::max(MFI->getObjectAlign(FirstSlot),
1318                                         MFI->getObjectAlign(SecondSlot));
1319 
1320           assert(MFI->getObjectSize(FirstSlot) >=
1321                  MFI->getObjectSize(SecondSlot) &&
1322                  "Merging a small object into a larger one");
1323 
1324           RemovedSlots+=1;
1325           ReducedSize += MFI->getObjectSize(SecondSlot);
1326           MFI->setObjectAlignment(FirstSlot, MaxAlignment);
1327           MFI->RemoveStackObject(SecondSlot);
1328         }
1329       }
1330     }
1331   }// While changed.
1332 
1333   // Record statistics.
1334   StackSpaceSaved += ReducedSize;
1335   StackSlotMerged += RemovedSlots;
1336   LLVM_DEBUG(dbgs() << "Merge " << RemovedSlots << " slots. Saved "
1337                     << ReducedSize << " bytes\n");
1338 
1339   // Scan the entire function and update all machine operands that use frame
1340   // indices to use the remapped frame index.
1341   expungeSlotMap(SlotRemap, NumSlots);
1342   remapInstructions(SlotRemap);
1343 
1344   return removeAllMarkers();
1345 }
1346