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 (LIFETIME_START and LIFETIME_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/PseudoSourceValueManager.h"
40 #include "llvm/CodeGen/SlotIndexes.h"
41 #include "llvm/CodeGen/TargetOpcodes.h"
42 #include "llvm/CodeGen/WinEHFuncInfo.h"
43 #include "llvm/Config/llvm-config.h"
44 #include "llvm/IR/Constants.h"
45 #include "llvm/IR/DebugInfoMetadata.h"
46 #include "llvm/IR/Instructions.h"
47 #include "llvm/IR/Metadata.h"
48 #include "llvm/IR/Use.h"
49 #include "llvm/IR/Value.h"
50 #include "llvm/InitializePasses.h"
51 #include "llvm/Pass.h"
52 #include "llvm/Support/Casting.h"
53 #include "llvm/Support/CommandLine.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/Support/Debug.h"
56 #include "llvm/Support/raw_ostream.h"
57 #include <algorithm>
58 #include <cassert>
59 #include <limits>
60 #include <memory>
61 #include <utility>
62 
63 using namespace llvm;
64 
65 #define DEBUG_TYPE "stack-coloring"
66 
67 static cl::opt<bool>
68 DisableColoring("no-stack-coloring",
69         cl::init(false), cl::Hidden,
70         cl::desc("Disable stack coloring"));
71 
72 /// The user may write code that uses allocas outside of the declared lifetime
73 /// zone. This can happen when the user returns a reference to a local
74 /// data-structure. We can detect these cases and decide not to optimize the
75 /// code. If this flag is enabled, we try to save the user. This option
76 /// is treated as overriding LifetimeStartOnFirstUse below.
77 static cl::opt<bool>
78 ProtectFromEscapedAllocas("protect-from-escaped-allocas",
79                           cl::init(false), cl::Hidden,
80                           cl::desc("Do not optimize lifetime zones that "
81                                    "are broken"));
82 
83 /// Enable enhanced dataflow scheme for lifetime analysis (treat first
84 /// use of stack slot as start of slot lifetime, as opposed to looking
85 /// for LIFETIME_START marker). See "Implementation notes" below for
86 /// more info.
87 static cl::opt<bool>
88 LifetimeStartOnFirstUse("stackcoloring-lifetime-start-on-first-use",
89         cl::init(true), cl::Hidden,
90         cl::desc("Treat stack lifetimes as starting on first use, not on START marker."));
91 
92 
93 STATISTIC(NumMarkerSeen,  "Number of lifetime markers found.");
94 STATISTIC(StackSpaceSaved, "Number of bytes saved due to merging slots.");
95 STATISTIC(StackSlotMerged, "Number of stack slot merged.");
96 STATISTIC(EscapedAllocas, "Number of allocas that escaped the lifetime region");
97 
98 //===----------------------------------------------------------------------===//
99 //                           StackColoring Pass
100 //===----------------------------------------------------------------------===//
101 //
102 // Stack Coloring reduces stack usage by merging stack slots when they
103 // can't be used together. For example, consider the following C program:
104 //
105 //     void bar(char *, int);
106 //     void foo(bool var) {
107 //         A: {
108 //             char z[4096];
109 //             bar(z, 0);
110 //         }
111 //
112 //         char *p;
113 //         char x[4096];
114 //         char y[4096];
115 //         if (var) {
116 //             p = x;
117 //         } else {
118 //             bar(y, 1);
119 //             p = y + 1024;
120 //         }
121 //     B:
122 //         bar(p, 2);
123 //     }
124 //
125 // Naively-compiled, this program would use 12k of stack space. However, the
126 // stack slot corresponding to `z` is always destroyed before either of the
127 // stack slots for `x` or `y` are used, and then `x` is only used if `var`
128 // is true, while `y` is only used if `var` is false. So in no time are 2
129 // of the stack slots used together, and therefore we can merge them,
130 // compiling the function using only a single 4k alloca:
131 //
132 //     void foo(bool var) { // equivalent
133 //         char x[4096];
134 //         char *p;
135 //         bar(x, 0);
136 //         if (var) {
137 //             p = x;
138 //         } else {
139 //             bar(x, 1);
140 //             p = x + 1024;
141 //         }
142 //         bar(p, 2);
143 //     }
144 //
145 // This is an important optimization if we want stack space to be under
146 // control in large functions, both open-coded ones and ones created by
147 // inlining.
148 //
149 // Implementation Notes:
150 // ---------------------
151 //
152 // An important part of the above reasoning is that `z` can't be accessed
153 // while the latter 2 calls to `bar` are running. This is justified because
154 // `z`'s lifetime is over after we exit from block `A:`, so any further
155 // accesses to it would be UB. The way we represent this information
156 // in LLVM is by having frontends delimit blocks with `lifetime.start`
157 // and `lifetime.end` intrinsics.
158 //
159 // The effect of these intrinsics seems to be as follows (maybe I should
160 // specify this in the reference?):
161 //
162 //   L1) at start, each stack-slot is marked as *out-of-scope*, unless no
163 //   lifetime intrinsic refers to that stack slot, in which case
164 //   it is marked as *in-scope*.
165 //   L2) on a `lifetime.start`, a stack slot is marked as *in-scope* and
166 //   the stack slot is overwritten with `undef`.
167 //   L3) on a `lifetime.end`, a stack slot is marked as *out-of-scope*.
168 //   L4) on function exit, all stack slots are marked as *out-of-scope*.
169 //   L5) `lifetime.end` is a no-op when called on a slot that is already
170 //   *out-of-scope*.
171 //   L6) memory accesses to *out-of-scope* stack slots are UB.
172 //   L7) when a stack-slot is marked as *out-of-scope*, all pointers to it
173 //   are invalidated, unless the slot is "degenerate". This is used to
174 //   justify not marking slots as in-use until the pointer to them is
175 //   used, but feels a bit hacky in the presence of things like LICM. See
176 //   the "Degenerate Slots" section for more details.
177 //
178 // Now, let's ground stack coloring on these rules. We'll define a slot
179 // as *in-use* at a (dynamic) point in execution if it either can be
180 // written to at that point, or if it has a live and non-undef content
181 // at that point.
182 //
183 // Obviously, slots that are never *in-use* together can be merged, and
184 // in our example `foo`, the slots for `x`, `y` and `z` are never
185 // in-use together (of course, sometimes slots that *are* in-use together
186 // might still be mergable, but we don't care about that here).
187 //
188 // In this implementation, we successively merge pairs of slots that are
189 // not *in-use* together. We could be smarter - for example, we could merge
190 // a single large slot with 2 small slots, or we could construct the
191 // interference graph and run a "smart" graph coloring algorithm, but with
192 // that aside, how do we find out whether a pair of slots might be *in-use*
193 // together?
194 //
195 // From our rules, we see that *out-of-scope* slots are never *in-use*,
196 // and from (L7) we see that "non-degenerate" slots remain non-*in-use*
197 // until their address is taken. Therefore, we can approximate slot activity
198 // using dataflow.
199 //
200 // A subtle point: naively, we might try to figure out which pairs of
201 // stack-slots interfere by propagating `S in-use` through the CFG for every
202 // stack-slot `S`, and having `S` and `T` interfere if there is a CFG point in
203 // which they are both *in-use*.
204 //
205 // That is sound, but overly conservative in some cases: in our (artificial)
206 // example `foo`, either `x` or `y` might be in use at the label `B:`, but
207 // as `x` is only in use if we came in from the `var` edge and `y` only
208 // if we came from the `!var` edge, they still can't be in use together.
209 // See PR32488 for an important real-life case.
210 //
211 // If we wanted to find all points of interference precisely, we could
212 // propagate `S in-use` and `S&T in-use` predicates through the CFG. That
213 // would be precise, but requires propagating `O(n^2)` dataflow facts.
214 //
215 // However, we aren't interested in the *set* of points of interference
216 // between 2 stack slots, only *whether* there *is* such a point. So we
217 // can rely on a little trick: for `S` and `T` to be in-use together,
218 // one of them needs to become in-use while the other is in-use (or
219 // they might both become in use simultaneously). We can check this
220 // by also keeping track of the points at which a stack slot might *start*
221 // being in-use.
222 //
223 // Exact first use:
224 // ----------------
225 //
226 // Consider the following motivating example:
227 //
228 //     int foo() {
229 //       char b1[1024], b2[1024];
230 //       if (...) {
231 //         char b3[1024];
232 //         <uses of b1, b3>;
233 //         return x;
234 //       } else {
235 //         char b4[1024], b5[1024];
236 //         <uses of b2, b4, b5>;
237 //         return y;
238 //       }
239 //     }
240 //
241 // In the code above, "b3" and "b4" are declared in distinct lexical
242 // scopes, meaning that it is easy to prove that they can share the
243 // same stack slot. Variables "b1" and "b2" are declared in the same
244 // scope, meaning that from a lexical point of view, their lifetimes
245 // overlap. From a control flow pointer of view, however, the two
246 // variables are accessed in disjoint regions of the CFG, thus it
247 // should be possible for them to share the same stack slot. An ideal
248 // stack allocation for the function above would look like:
249 //
250 //     slot 0: b1, b2
251 //     slot 1: b3, b4
252 //     slot 2: b5
253 //
254 // Achieving this allocation is tricky, however, due to the way
255 // lifetime markers are inserted. Here is a simplified view of the
256 // control flow graph for the code above:
257 //
258 //                +------  block 0 -------+
259 //               0| LIFETIME_START b1, b2 |
260 //               1| <test 'if' condition> |
261 //                +-----------------------+
262 //                   ./              \.
263 //   +------  block 1 -------+   +------  block 2 -------+
264 //  2| LIFETIME_START b3     |  5| LIFETIME_START b4, b5 |
265 //  3| <uses of b1, b3>      |  6| <uses of b2, b4, b5>  |
266 //  4| LIFETIME_END b3       |  7| LIFETIME_END b4, b5   |
267 //   +-----------------------+   +-----------------------+
268 //                   \.              /.
269 //                +------  block 3 -------+
270 //               8| <cleanupcode>         |
271 //               9| LIFETIME_END b1, b2   |
272 //              10| return                |
273 //                +-----------------------+
274 //
275 // If we create live intervals for the variables above strictly based
276 // on the lifetime markers, we'll get the set of intervals on the
277 // left. If we ignore the lifetime start markers and instead treat a
278 // variable's lifetime as beginning with the first reference to the
279 // var, then we get the intervals on the right.
280 //
281 //            LIFETIME_START      First Use
282 //     b1:    [0,9]               [3,4] [8,9]
283 //     b2:    [0,9]               [6,9]
284 //     b3:    [2,4]               [3,4]
285 //     b4:    [5,7]               [6,7]
286 //     b5:    [5,7]               [6,7]
287 //
288 // For the intervals on the left, the best we can do is overlap two
289 // variables (b3 and b4, for example); this gives us a stack size of
290 // 4*1024 bytes, not ideal. When treating first-use as the start of a
291 // lifetime, we can additionally overlap b1 and b5, giving us a 3*1024
292 // byte stack (better).
293 //
294 // Degenerate Slots:
295 // -----------------
296 //
297 // Relying entirely on first-use of stack slots is problematic,
298 // however, due to the fact that optimizations can sometimes migrate
299 // uses of a variable outside of its lifetime start/end region. Here
300 // is an example:
301 //
302 //     int bar() {
303 //       char b1[1024], b2[1024];
304 //       if (...) {
305 //         <uses of b2>
306 //         return y;
307 //       } else {
308 //         <uses of b1>
309 //         while (...) {
310 //           char b3[1024];
311 //           <uses of b3>
312 //         }
313 //       }
314 //     }
315 //
316 // Before optimization, the control flow graph for the code above
317 // might look like the following:
318 //
319 //                +------  block 0 -------+
320 //               0| LIFETIME_START b1, b2 |
321 //               1| <test 'if' condition> |
322 //                +-----------------------+
323 //                   ./              \.
324 //   +------  block 1 -------+    +------- block 2 -------+
325 //  2| <uses of b2>          |   3| <uses of b1>          |
326 //   +-----------------------+    +-----------------------+
327 //              |                            |
328 //              |                 +------- block 3 -------+ <-\.
329 //              |                4| <while condition>     |    |
330 //              |                 +-----------------------+    |
331 //              |               /          |                   |
332 //              |              /  +------- block 4 -------+
333 //              \             /  5| LIFETIME_START b3     |    |
334 //               \           /   6| <uses of b3>          |    |
335 //                \         /    7| LIFETIME_END b3       |    |
336 //                 \        |    +------------------------+    |
337 //                  \       |                 \                /
338 //                +------  block 5 -----+      \---------------
339 //               8| <cleanupcode>       |
340 //               9| LIFETIME_END b1, b2 |
341 //              10| return              |
342 //                +---------------------+
343 //
344 // During optimization, however, it can happen that an instruction
345 // computing an address in "b3" (for example, a loop-invariant GEP) is
346 // hoisted up out of the loop from block 4 to block 2.  [Note that
347 // this is not an actual load from the stack, only an instruction that
348 // computes the address to be loaded]. If this happens, there is now a
349 // path leading from the first use of b3 to the return instruction
350 // that does not encounter the b3 LIFETIME_END, hence b3's lifetime is
351 // now larger than if we were computing live intervals strictly based
352 // on lifetime markers. In the example above, this lengthened lifetime
353 // would mean that it would appear illegal to overlap b3 with b2.
354 //
355 // To deal with this such cases, the code in ::collectMarkers() below
356 // tries to identify "degenerate" slots -- those slots where on a single
357 // forward pass through the CFG we encounter a first reference to slot
358 // K before we hit the slot K lifetime start marker. For such slots,
359 // we fall back on using the lifetime start marker as the beginning of
360 // the variable's lifetime.  NB: with this implementation, slots can
361 // appear degenerate in cases where there is unstructured control flow:
362 //
363 //    if (q) goto mid;
364 //    if (x > 9) {
365 //         int b[100];
366 //         memcpy(&b[0], ...);
367 //    mid: b[k] = ...;
368 //         abc(&b);
369 //    }
370 //
371 // If in RPO ordering chosen to walk the CFG  we happen to visit the b[k]
372 // before visiting the memcpy block (which will contain the lifetime start
373 // for "b" then it will appear that 'b' has a degenerate lifetime.
374 
375 namespace {
376 
377 /// StackColoring - A machine pass for merging disjoint stack allocations,
378 /// marked by the LIFETIME_START and LIFETIME_END pseudo instructions.
379 class StackColoring : public MachineFunctionPass {
380   MachineFrameInfo *MFI = nullptr;
381   MachineFunction *MF = nullptr;
382 
383   /// A class representing liveness information for a single basic block.
384   /// Each bit in the BitVector represents the liveness property
385   /// for a different stack slot.
386   struct BlockLifetimeInfo {
387     /// Which slots BEGINs in each basic block.
388     BitVector Begin;
389 
390     /// Which slots ENDs in each basic block.
391     BitVector End;
392 
393     /// Which slots are marked as LIVE_IN, coming into each basic block.
394     BitVector LiveIn;
395 
396     /// Which slots are marked as LIVE_OUT, coming out of each basic block.
397     BitVector LiveOut;
398   };
399 
400   /// Maps active slots (per bit) for each basic block.
401   using LivenessMap = DenseMap<const MachineBasicBlock *, BlockLifetimeInfo>;
402   LivenessMap BlockLiveness;
403 
404   /// Maps serial numbers to basic blocks.
405   DenseMap<const MachineBasicBlock *, int> BasicBlocks;
406 
407   /// Maps basic blocks to a serial number.
408   SmallVector<const MachineBasicBlock *, 8> BasicBlockNumbering;
409 
410   /// Maps slots to their use interval. Outside of this interval, slots
411   /// values are either dead or `undef` and they will not be written to.
412   SmallVector<std::unique_ptr<LiveInterval>, 16> Intervals;
413 
414   /// Maps slots to the points where they can become in-use.
415   SmallVector<SmallVector<SlotIndex, 4>, 16> LiveStarts;
416 
417   /// VNInfo is used for the construction of LiveIntervals.
418   VNInfo::Allocator VNInfoAllocator;
419 
420   /// SlotIndex analysis object.
421   SlotIndexes *Indexes = nullptr;
422 
423   /// The list of lifetime markers found. These markers are to be removed
424   /// once the coloring is done.
425   SmallVector<MachineInstr*, 8> Markers;
426 
427   /// Record the FI slots for which we have seen some sort of
428   /// lifetime marker (either start or end).
429   BitVector InterestingSlots;
430 
431   /// FI slots that need to be handled conservatively (for these
432   /// slots lifetime-start-on-first-use is disabled).
433   BitVector ConservativeSlots;
434 
435   /// Number of iterations taken during data flow analysis.
436   unsigned NumIterations;
437 
438 public:
439   static char ID;
440 
StackColoring()441   StackColoring() : MachineFunctionPass(ID) {
442     initializeStackColoringPass(*PassRegistry::getPassRegistry());
443   }
444 
445   void getAnalysisUsage(AnalysisUsage &AU) const override;
446   bool runOnMachineFunction(MachineFunction &Func) override;
447 
448 private:
449   /// Used in collectMarkers
450   using BlockBitVecMap = DenseMap<const MachineBasicBlock *, BitVector>;
451 
452   /// Debug.
453   void dump() const;
454   void dumpIntervals() const;
455   void dumpBB(MachineBasicBlock *MBB) const;
456   void dumpBV(const char *tag, const BitVector &BV) const;
457 
458   /// Removes all of the lifetime marker instructions from the function.
459   /// \returns true if any markers were removed.
460   bool removeAllMarkers();
461 
462   /// Scan the machine function and find all of the lifetime markers.
463   /// Record the findings in the BEGIN and END vectors.
464   /// \returns the number of markers found.
465   unsigned collectMarkers(unsigned NumSlot);
466 
467   /// Perform the dataflow calculation and calculate the lifetime for each of
468   /// the slots, based on the BEGIN/END vectors. Set the LifetimeLIVE_IN and
469   /// LifetimeLIVE_OUT maps that represent which stack slots are live coming
470   /// in and out blocks.
471   void calculateLocalLiveness();
472 
473   /// Returns TRUE if we're using the first-use-begins-lifetime method for
474   /// this slot (if FALSE, then the start marker is treated as start of lifetime).
applyFirstUse(int Slot)475   bool applyFirstUse(int Slot) {
476     if (!LifetimeStartOnFirstUse || ProtectFromEscapedAllocas)
477       return false;
478     if (ConservativeSlots.test(Slot))
479       return false;
480     return true;
481   }
482 
483   /// Examines the specified instruction and returns TRUE if the instruction
484   /// represents the start or end of an interesting lifetime. The slot or slots
485   /// starting or ending are added to the vector "slots" and "isStart" is set
486   /// accordingly.
487   /// \returns True if inst contains a lifetime start or end
488   bool isLifetimeStartOrEnd(const MachineInstr &MI,
489                             SmallVector<int, 4> &slots,
490                             bool &isStart);
491 
492   /// Construct the LiveIntervals for the slots.
493   void calculateLiveIntervals(unsigned NumSlots);
494 
495   /// Go over the machine function and change instructions which use stack
496   /// slots to use the joint slots.
497   void remapInstructions(DenseMap<int, int> &SlotRemap);
498 
499   /// The input program may contain instructions which are not inside lifetime
500   /// markers. This can happen due to a bug in the compiler or due to a bug in
501   /// user code (for example, returning a reference to a local variable).
502   /// This procedure checks all of the instructions in the function and
503   /// invalidates lifetime ranges which do not contain all of the instructions
504   /// which access that frame slot.
505   void removeInvalidSlotRanges();
506 
507   /// Map entries which point to other entries to their destination.
508   ///   A->B->C becomes A->C.
509   void expungeSlotMap(DenseMap<int, int> &SlotRemap, unsigned NumSlots);
510 };
511 
512 } // end anonymous namespace
513 
514 char StackColoring::ID = 0;
515 
516 char &llvm::StackColoringID = StackColoring::ID;
517 
518 INITIALIZE_PASS_BEGIN(StackColoring, DEBUG_TYPE,
519                       "Merge disjoint stack slots", false, false)
INITIALIZE_PASS_DEPENDENCY(SlotIndexes)520 INITIALIZE_PASS_DEPENDENCY(SlotIndexes)
521 INITIALIZE_PASS_END(StackColoring, DEBUG_TYPE,
522                     "Merge disjoint stack slots", false, false)
523 
524 void StackColoring::getAnalysisUsage(AnalysisUsage &AU) const {
525   AU.addRequired<SlotIndexes>();
526   MachineFunctionPass::getAnalysisUsage(AU);
527 }
528 
529 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dumpBV(const char * tag,const BitVector & BV) const530 LLVM_DUMP_METHOD void StackColoring::dumpBV(const char *tag,
531                                             const BitVector &BV) const {
532   dbgs() << tag << " : { ";
533   for (unsigned I = 0, E = BV.size(); I != E; ++I)
534     dbgs() << BV.test(I) << " ";
535   dbgs() << "}\n";
536 }
537 
dumpBB(MachineBasicBlock * MBB) const538 LLVM_DUMP_METHOD void StackColoring::dumpBB(MachineBasicBlock *MBB) const {
539   LivenessMap::const_iterator BI = BlockLiveness.find(MBB);
540   assert(BI != BlockLiveness.end() && "Block not found");
541   const BlockLifetimeInfo &BlockInfo = BI->second;
542 
543   dumpBV("BEGIN", BlockInfo.Begin);
544   dumpBV("END", BlockInfo.End);
545   dumpBV("LIVE_IN", BlockInfo.LiveIn);
546   dumpBV("LIVE_OUT", BlockInfo.LiveOut);
547 }
548 
dump() const549 LLVM_DUMP_METHOD void StackColoring::dump() const {
550   for (MachineBasicBlock *MBB : depth_first(MF)) {
551     dbgs() << "Inspecting block #" << MBB->getNumber() << " ["
552            << MBB->getName() << "]\n";
553     dumpBB(MBB);
554   }
555 }
556 
dumpIntervals() const557 LLVM_DUMP_METHOD void StackColoring::dumpIntervals() const {
558   for (unsigned I = 0, E = Intervals.size(); I != E; ++I) {
559     dbgs() << "Interval[" << I << "]:\n";
560     Intervals[I]->dump();
561   }
562 }
563 #endif
564 
getStartOrEndSlot(const MachineInstr & MI)565 static inline int getStartOrEndSlot(const MachineInstr &MI)
566 {
567   assert((MI.getOpcode() == TargetOpcode::LIFETIME_START ||
568           MI.getOpcode() == TargetOpcode::LIFETIME_END) &&
569          "Expected LIFETIME_START or LIFETIME_END op");
570   const MachineOperand &MO = MI.getOperand(0);
571   int Slot = MO.getIndex();
572   if (Slot >= 0)
573     return Slot;
574   return -1;
575 }
576 
577 // At the moment the only way to end a variable lifetime is with
578 // a VARIABLE_LIFETIME op (which can't contain a start). If things
579 // change and the IR allows for a single inst that both begins
580 // and ends lifetime(s), this interface will need to be reworked.
isLifetimeStartOrEnd(const MachineInstr & MI,SmallVector<int,4> & slots,bool & isStart)581 bool StackColoring::isLifetimeStartOrEnd(const MachineInstr &MI,
582                                          SmallVector<int, 4> &slots,
583                                          bool &isStart) {
584   if (MI.getOpcode() == TargetOpcode::LIFETIME_START ||
585       MI.getOpcode() == TargetOpcode::LIFETIME_END) {
586     int Slot = getStartOrEndSlot(MI);
587     if (Slot < 0)
588       return false;
589     if (!InterestingSlots.test(Slot))
590       return false;
591     slots.push_back(Slot);
592     if (MI.getOpcode() == TargetOpcode::LIFETIME_END) {
593       isStart = false;
594       return true;
595     }
596     if (!applyFirstUse(Slot)) {
597       isStart = true;
598       return true;
599     }
600   } else if (LifetimeStartOnFirstUse && !ProtectFromEscapedAllocas) {
601     if (!MI.isDebugInstr()) {
602       bool found = false;
603       for (const MachineOperand &MO : MI.operands()) {
604         if (!MO.isFI())
605           continue;
606         int Slot = MO.getIndex();
607         if (Slot<0)
608           continue;
609         if (InterestingSlots.test(Slot) && applyFirstUse(Slot)) {
610           slots.push_back(Slot);
611           found = true;
612         }
613       }
614       if (found) {
615         isStart = true;
616         return true;
617       }
618     }
619   }
620   return false;
621 }
622 
collectMarkers(unsigned NumSlot)623 unsigned StackColoring::collectMarkers(unsigned NumSlot) {
624   unsigned MarkersFound = 0;
625   BlockBitVecMap SeenStartMap;
626   InterestingSlots.clear();
627   InterestingSlots.resize(NumSlot);
628   ConservativeSlots.clear();
629   ConservativeSlots.resize(NumSlot);
630 
631   // number of start and end lifetime ops for each slot
632   SmallVector<int, 8> NumStartLifetimes(NumSlot, 0);
633   SmallVector<int, 8> NumEndLifetimes(NumSlot, 0);
634 
635   // Step 1: collect markers and populate the "InterestingSlots"
636   // and "ConservativeSlots" sets.
637   for (MachineBasicBlock *MBB : depth_first(MF)) {
638     // Compute the set of slots for which we've seen a START marker but have
639     // not yet seen an END marker at this point in the walk (e.g. on entry
640     // to this bb).
641     BitVector BetweenStartEnd;
642     BetweenStartEnd.resize(NumSlot);
643     for (const MachineBasicBlock *Pred : MBB->predecessors()) {
644       BlockBitVecMap::const_iterator I = SeenStartMap.find(Pred);
645       if (I != SeenStartMap.end()) {
646         BetweenStartEnd |= I->second;
647       }
648     }
649 
650     // Walk the instructions in the block to look for start/end ops.
651     for (MachineInstr &MI : *MBB) {
652       if (MI.isDebugInstr())
653         continue;
654       if (MI.getOpcode() == TargetOpcode::LIFETIME_START ||
655           MI.getOpcode() == TargetOpcode::LIFETIME_END) {
656         int Slot = getStartOrEndSlot(MI);
657         if (Slot < 0)
658           continue;
659         InterestingSlots.set(Slot);
660         if (MI.getOpcode() == TargetOpcode::LIFETIME_START) {
661           BetweenStartEnd.set(Slot);
662           NumStartLifetimes[Slot] += 1;
663         } else {
664           BetweenStartEnd.reset(Slot);
665           NumEndLifetimes[Slot] += 1;
666         }
667         const AllocaInst *Allocation = MFI->getObjectAllocation(Slot);
668         if (Allocation) {
669           LLVM_DEBUG(dbgs() << "Found a lifetime ");
670           LLVM_DEBUG(dbgs() << (MI.getOpcode() == TargetOpcode::LIFETIME_START
671                                     ? "start"
672                                     : "end"));
673           LLVM_DEBUG(dbgs() << " marker for slot #" << Slot);
674           LLVM_DEBUG(dbgs()
675                      << " with allocation: " << Allocation->getName() << "\n");
676         }
677         Markers.push_back(&MI);
678         MarkersFound += 1;
679       } else {
680         for (const MachineOperand &MO : MI.operands()) {
681           if (!MO.isFI())
682             continue;
683           int Slot = MO.getIndex();
684           if (Slot < 0)
685             continue;
686           if (! BetweenStartEnd.test(Slot)) {
687             ConservativeSlots.set(Slot);
688           }
689         }
690       }
691     }
692     BitVector &SeenStart = SeenStartMap[MBB];
693     SeenStart |= BetweenStartEnd;
694   }
695   if (!MarkersFound) {
696     return 0;
697   }
698 
699   // PR27903: slots with multiple start or end lifetime ops are not
700   // safe to enable for "lifetime-start-on-first-use".
701   for (unsigned slot = 0; slot < NumSlot; ++slot) {
702     if (NumStartLifetimes[slot] > 1 || NumEndLifetimes[slot] > 1)
703       ConservativeSlots.set(slot);
704   }
705 
706   // The write to the catch object by the personality function is not propely
707   // modeled in IR: It happens before any cleanuppads are executed, even if the
708   // first mention of the catch object is in a catchpad. As such, mark catch
709   // object slots as conservative, so they are excluded from first-use analysis.
710   if (WinEHFuncInfo *EHInfo = MF->getWinEHFuncInfo())
711     for (WinEHTryBlockMapEntry &TBME : EHInfo->TryBlockMap)
712       for (WinEHHandlerType &H : TBME.HandlerArray)
713         if (H.CatchObj.FrameIndex != std::numeric_limits<int>::max() &&
714             H.CatchObj.FrameIndex >= 0)
715           ConservativeSlots.set(H.CatchObj.FrameIndex);
716 
717   LLVM_DEBUG(dumpBV("Conservative slots", ConservativeSlots));
718 
719   // Step 2: compute begin/end sets for each block
720 
721   // NOTE: We use a depth-first iteration to ensure that we obtain a
722   // deterministic numbering.
723   for (MachineBasicBlock *MBB : depth_first(MF)) {
724     // Assign a serial number to this basic block.
725     BasicBlocks[MBB] = BasicBlockNumbering.size();
726     BasicBlockNumbering.push_back(MBB);
727 
728     // Keep a reference to avoid repeated lookups.
729     BlockLifetimeInfo &BlockInfo = BlockLiveness[MBB];
730 
731     BlockInfo.Begin.resize(NumSlot);
732     BlockInfo.End.resize(NumSlot);
733 
734     SmallVector<int, 4> slots;
735     for (MachineInstr &MI : *MBB) {
736       bool isStart = false;
737       slots.clear();
738       if (isLifetimeStartOrEnd(MI, slots, isStart)) {
739         if (!isStart) {
740           assert(slots.size() == 1 && "unexpected: MI ends multiple slots");
741           int Slot = slots[0];
742           if (BlockInfo.Begin.test(Slot)) {
743             BlockInfo.Begin.reset(Slot);
744           }
745           BlockInfo.End.set(Slot);
746         } else {
747           for (auto Slot : slots) {
748             LLVM_DEBUG(dbgs() << "Found a use of slot #" << Slot);
749             LLVM_DEBUG(dbgs()
750                        << " at " << printMBBReference(*MBB) << " index ");
751             LLVM_DEBUG(Indexes->getInstructionIndex(MI).print(dbgs()));
752             const AllocaInst *Allocation = MFI->getObjectAllocation(Slot);
753             if (Allocation) {
754               LLVM_DEBUG(dbgs()
755                          << " with allocation: " << Allocation->getName());
756             }
757             LLVM_DEBUG(dbgs() << "\n");
758             if (BlockInfo.End.test(Slot)) {
759               BlockInfo.End.reset(Slot);
760             }
761             BlockInfo.Begin.set(Slot);
762           }
763         }
764       }
765     }
766   }
767 
768   // Update statistics.
769   NumMarkerSeen += MarkersFound;
770   return MarkersFound;
771 }
772 
calculateLocalLiveness()773 void StackColoring::calculateLocalLiveness() {
774   unsigned NumIters = 0;
775   bool changed = true;
776   while (changed) {
777     changed = false;
778     ++NumIters;
779 
780     for (const MachineBasicBlock *BB : BasicBlockNumbering) {
781       // Use an iterator to avoid repeated lookups.
782       LivenessMap::iterator BI = BlockLiveness.find(BB);
783       assert(BI != BlockLiveness.end() && "Block not found");
784       BlockLifetimeInfo &BlockInfo = BI->second;
785 
786       // Compute LiveIn by unioning together the LiveOut sets of all preds.
787       BitVector LocalLiveIn;
788       for (MachineBasicBlock *Pred : BB->predecessors()) {
789         LivenessMap::const_iterator I = BlockLiveness.find(Pred);
790         // PR37130: transformations prior to stack coloring can
791         // sometimes leave behind statically unreachable blocks; these
792         // can be safely skipped here.
793         if (I != BlockLiveness.end())
794           LocalLiveIn |= I->second.LiveOut;
795       }
796 
797       // Compute LiveOut by subtracting out lifetimes that end in this
798       // block, then adding in lifetimes that begin in this block.  If
799       // we have both BEGIN and END markers in the same basic block
800       // then we know that the BEGIN marker comes after the END,
801       // because we already handle the case where the BEGIN comes
802       // before the END when collecting the markers (and building the
803       // BEGIN/END vectors).
804       BitVector LocalLiveOut = LocalLiveIn;
805       LocalLiveOut.reset(BlockInfo.End);
806       LocalLiveOut |= BlockInfo.Begin;
807 
808       // Update block LiveIn set, noting whether it has changed.
809       if (LocalLiveIn.test(BlockInfo.LiveIn)) {
810         changed = true;
811         BlockInfo.LiveIn |= LocalLiveIn;
812       }
813 
814       // Update block LiveOut set, noting whether it has changed.
815       if (LocalLiveOut.test(BlockInfo.LiveOut)) {
816         changed = true;
817         BlockInfo.LiveOut |= LocalLiveOut;
818       }
819     }
820   } // while changed.
821 
822   NumIterations = NumIters;
823 }
824 
calculateLiveIntervals(unsigned NumSlots)825 void StackColoring::calculateLiveIntervals(unsigned NumSlots) {
826   SmallVector<SlotIndex, 16> Starts;
827   SmallVector<bool, 16> DefinitelyInUse;
828 
829   // For each block, find which slots are active within this block
830   // and update the live intervals.
831   for (const MachineBasicBlock &MBB : *MF) {
832     Starts.clear();
833     Starts.resize(NumSlots);
834     DefinitelyInUse.clear();
835     DefinitelyInUse.resize(NumSlots);
836 
837     // Start the interval of the slots that we previously found to be 'in-use'.
838     BlockLifetimeInfo &MBBLiveness = BlockLiveness[&MBB];
839     for (int pos = MBBLiveness.LiveIn.find_first(); pos != -1;
840          pos = MBBLiveness.LiveIn.find_next(pos)) {
841       Starts[pos] = Indexes->getMBBStartIdx(&MBB);
842     }
843 
844     // Create the interval for the basic blocks containing lifetime begin/end.
845     for (const MachineInstr &MI : MBB) {
846       SmallVector<int, 4> slots;
847       bool IsStart = false;
848       if (!isLifetimeStartOrEnd(MI, slots, IsStart))
849         continue;
850       SlotIndex ThisIndex = Indexes->getInstructionIndex(MI);
851       for (auto Slot : slots) {
852         if (IsStart) {
853           // If a slot is already definitely in use, we don't have to emit
854           // a new start marker because there is already a pre-existing
855           // one.
856           if (!DefinitelyInUse[Slot]) {
857             LiveStarts[Slot].push_back(ThisIndex);
858             DefinitelyInUse[Slot] = true;
859           }
860           if (!Starts[Slot].isValid())
861             Starts[Slot] = ThisIndex;
862         } else {
863           if (Starts[Slot].isValid()) {
864             VNInfo *VNI = Intervals[Slot]->getValNumInfo(0);
865             Intervals[Slot]->addSegment(
866                 LiveInterval::Segment(Starts[Slot], ThisIndex, VNI));
867             Starts[Slot] = SlotIndex(); // Invalidate the start index
868             DefinitelyInUse[Slot] = false;
869           }
870         }
871       }
872     }
873 
874     // Finish up started segments
875     for (unsigned i = 0; i < NumSlots; ++i) {
876       if (!Starts[i].isValid())
877         continue;
878 
879       SlotIndex EndIdx = Indexes->getMBBEndIdx(&MBB);
880       VNInfo *VNI = Intervals[i]->getValNumInfo(0);
881       Intervals[i]->addSegment(LiveInterval::Segment(Starts[i], EndIdx, VNI));
882     }
883   }
884 }
885 
removeAllMarkers()886 bool StackColoring::removeAllMarkers() {
887   unsigned Count = 0;
888   for (MachineInstr *MI : Markers) {
889     MI->eraseFromParent();
890     Count++;
891   }
892   Markers.clear();
893 
894   LLVM_DEBUG(dbgs() << "Removed " << Count << " markers.\n");
895   return Count;
896 }
897 
remapInstructions(DenseMap<int,int> & SlotRemap)898 void StackColoring::remapInstructions(DenseMap<int, int> &SlotRemap) {
899   unsigned FixedInstr = 0;
900   unsigned FixedMemOp = 0;
901   unsigned FixedDbg = 0;
902 
903   // Remap debug information that refers to stack slots.
904   for (auto &VI : MF->getVariableDbgInfo()) {
905     if (!VI.Var || !VI.inStackSlot())
906       continue;
907     int Slot = VI.getStackSlot();
908     if (SlotRemap.count(Slot)) {
909       LLVM_DEBUG(dbgs() << "Remapping debug info for ["
910                         << cast<DILocalVariable>(VI.Var)->getName() << "].\n");
911       VI.updateStackSlot(SlotRemap[Slot]);
912       FixedDbg++;
913     }
914   }
915 
916   // Keep a list of *allocas* which need to be remapped.
917   DenseMap<const AllocaInst*, const AllocaInst*> Allocas;
918 
919   // Keep a list of allocas which has been affected by the remap.
920   SmallPtrSet<const AllocaInst*, 32> MergedAllocas;
921 
922   for (const std::pair<int, int> &SI : SlotRemap) {
923     const AllocaInst *From = MFI->getObjectAllocation(SI.first);
924     const AllocaInst *To = MFI->getObjectAllocation(SI.second);
925     assert(To && From && "Invalid allocation object");
926     Allocas[From] = To;
927 
928     // If From is before wo, its possible that there is a use of From between
929     // them.
930     if (From->comesBefore(To))
931       const_cast<AllocaInst*>(To)->moveBefore(const_cast<AllocaInst*>(From));
932 
933     // AA might be used later for instruction scheduling, and we need it to be
934     // able to deduce the correct aliasing releationships between pointers
935     // derived from the alloca being remapped and the target of that remapping.
936     // The only safe way, without directly informing AA about the remapping
937     // somehow, is to directly update the IR to reflect the change being made
938     // here.
939     Instruction *Inst = const_cast<AllocaInst *>(To);
940     if (From->getType() != To->getType()) {
941       BitCastInst *Cast = new BitCastInst(Inst, From->getType());
942       Cast->insertAfter(Inst);
943       Inst = Cast;
944     }
945 
946     // We keep both slots to maintain AliasAnalysis metadata later.
947     MergedAllocas.insert(From);
948     MergedAllocas.insert(To);
949 
950     // Transfer the stack protector layout tag, but make sure that SSPLK_AddrOf
951     // does not overwrite SSPLK_SmallArray or SSPLK_LargeArray, and make sure
952     // that SSPLK_SmallArray does not overwrite SSPLK_LargeArray.
953     MachineFrameInfo::SSPLayoutKind FromKind
954         = MFI->getObjectSSPLayout(SI.first);
955     MachineFrameInfo::SSPLayoutKind ToKind = MFI->getObjectSSPLayout(SI.second);
956     if (FromKind != MachineFrameInfo::SSPLK_None &&
957         (ToKind == MachineFrameInfo::SSPLK_None ||
958          (ToKind != MachineFrameInfo::SSPLK_LargeArray &&
959           FromKind != MachineFrameInfo::SSPLK_AddrOf)))
960       MFI->setObjectSSPLayout(SI.second, FromKind);
961 
962     // The new alloca might not be valid in a llvm.dbg.declare for this
963     // variable, so undef out the use to make the verifier happy.
964     AllocaInst *FromAI = const_cast<AllocaInst *>(From);
965     if (FromAI->isUsedByMetadata())
966       ValueAsMetadata::handleRAUW(FromAI, UndefValue::get(FromAI->getType()));
967     for (auto &Use : FromAI->uses()) {
968       if (BitCastInst *BCI = dyn_cast<BitCastInst>(Use.get()))
969         if (BCI->isUsedByMetadata())
970           ValueAsMetadata::handleRAUW(BCI, UndefValue::get(BCI->getType()));
971     }
972 
973     // Note that this will not replace uses in MMOs (which we'll update below),
974     // or anywhere else (which is why we won't delete the original
975     // instruction).
976     FromAI->replaceAllUsesWith(Inst);
977   }
978 
979   // Remap all instructions to the new stack slots.
980   std::vector<std::vector<MachineMemOperand *>> SSRefs(
981       MFI->getObjectIndexEnd());
982   for (MachineBasicBlock &BB : *MF)
983     for (MachineInstr &I : BB) {
984       // Skip lifetime markers. We'll remove them soon.
985       if (I.getOpcode() == TargetOpcode::LIFETIME_START ||
986           I.getOpcode() == TargetOpcode::LIFETIME_END)
987         continue;
988 
989       // Update the MachineMemOperand to use the new alloca.
990       for (MachineMemOperand *MMO : I.memoperands()) {
991         // We've replaced IR-level uses of the remapped allocas, so we only
992         // need to replace direct uses here.
993         const AllocaInst *AI = dyn_cast_or_null<AllocaInst>(MMO->getValue());
994         if (!AI)
995           continue;
996 
997         if (!Allocas.count(AI))
998           continue;
999 
1000         MMO->setValue(Allocas[AI]);
1001         FixedMemOp++;
1002       }
1003 
1004       // Update all of the machine instruction operands.
1005       for (MachineOperand &MO : I.operands()) {
1006         if (!MO.isFI())
1007           continue;
1008         int FromSlot = MO.getIndex();
1009 
1010         // Don't touch arguments.
1011         if (FromSlot<0)
1012           continue;
1013 
1014         // Only look at mapped slots.
1015         if (!SlotRemap.count(FromSlot))
1016           continue;
1017 
1018         // In a debug build, check that the instruction that we are modifying is
1019         // inside the expected live range. If the instruction is not inside
1020         // the calculated range then it means that the alloca usage moved
1021         // outside of the lifetime markers, or that the user has a bug.
1022         // NOTE: Alloca address calculations which happen outside the lifetime
1023         // zone are okay, despite the fact that we don't have a good way
1024         // for validating all of the usages of the calculation.
1025 #ifndef NDEBUG
1026         bool TouchesMemory = I.mayLoadOrStore();
1027         // If we *don't* protect the user from escaped allocas, don't bother
1028         // validating the instructions.
1029         if (!I.isDebugInstr() && TouchesMemory && ProtectFromEscapedAllocas) {
1030           SlotIndex Index = Indexes->getInstructionIndex(I);
1031           const LiveInterval *Interval = &*Intervals[FromSlot];
1032           assert(Interval->find(Index) != Interval->end() &&
1033                  "Found instruction usage outside of live range.");
1034         }
1035 #endif
1036 
1037         // Fix the machine instructions.
1038         int ToSlot = SlotRemap[FromSlot];
1039         MO.setIndex(ToSlot);
1040         FixedInstr++;
1041       }
1042 
1043       // We adjust AliasAnalysis information for merged stack slots.
1044       SmallVector<MachineMemOperand *, 2> NewMMOs;
1045       bool ReplaceMemOps = false;
1046       for (MachineMemOperand *MMO : I.memoperands()) {
1047         // Collect MachineMemOperands which reference
1048         // FixedStackPseudoSourceValues with old frame indices.
1049         if (const auto *FSV = dyn_cast_or_null<FixedStackPseudoSourceValue>(
1050                 MMO->getPseudoValue())) {
1051           int FI = FSV->getFrameIndex();
1052           auto To = SlotRemap.find(FI);
1053           if (To != SlotRemap.end())
1054             SSRefs[FI].push_back(MMO);
1055         }
1056 
1057         // If this memory location can be a slot remapped here,
1058         // we remove AA information.
1059         bool MayHaveConflictingAAMD = false;
1060         if (MMO->getAAInfo()) {
1061           if (const Value *MMOV = MMO->getValue()) {
1062             SmallVector<Value *, 4> Objs;
1063             getUnderlyingObjectsForCodeGen(MMOV, Objs);
1064 
1065             if (Objs.empty())
1066               MayHaveConflictingAAMD = true;
1067             else
1068               for (Value *V : Objs) {
1069                 // If this memory location comes from a known stack slot
1070                 // that is not remapped, we continue checking.
1071                 // Otherwise, we need to invalidate AA infomation.
1072                 const AllocaInst *AI = dyn_cast_or_null<AllocaInst>(V);
1073                 if (AI && MergedAllocas.count(AI)) {
1074                   MayHaveConflictingAAMD = true;
1075                   break;
1076                 }
1077               }
1078           }
1079         }
1080         if (MayHaveConflictingAAMD) {
1081           NewMMOs.push_back(MF->getMachineMemOperand(MMO, AAMDNodes()));
1082           ReplaceMemOps = true;
1083         } else {
1084           NewMMOs.push_back(MMO);
1085         }
1086       }
1087 
1088       // If any memory operand is updated, set memory references of
1089       // this instruction.
1090       if (ReplaceMemOps)
1091         I.setMemRefs(*MF, NewMMOs);
1092     }
1093 
1094   // Rewrite MachineMemOperands that reference old frame indices.
1095   for (auto E : enumerate(SSRefs))
1096     if (!E.value().empty()) {
1097       const PseudoSourceValue *NewSV =
1098           MF->getPSVManager().getFixedStack(SlotRemap.find(E.index())->second);
1099       for (MachineMemOperand *Ref : E.value())
1100         Ref->setValue(NewSV);
1101     }
1102 
1103   // Update the location of C++ catch objects for the MSVC personality routine.
1104   if (WinEHFuncInfo *EHInfo = MF->getWinEHFuncInfo())
1105     for (WinEHTryBlockMapEntry &TBME : EHInfo->TryBlockMap)
1106       for (WinEHHandlerType &H : TBME.HandlerArray)
1107         if (H.CatchObj.FrameIndex != std::numeric_limits<int>::max() &&
1108             SlotRemap.count(H.CatchObj.FrameIndex))
1109           H.CatchObj.FrameIndex = SlotRemap[H.CatchObj.FrameIndex];
1110 
1111   LLVM_DEBUG(dbgs() << "Fixed " << FixedMemOp << " machine memory operands.\n");
1112   LLVM_DEBUG(dbgs() << "Fixed " << FixedDbg << " debug locations.\n");
1113   LLVM_DEBUG(dbgs() << "Fixed " << FixedInstr << " machine instructions.\n");
1114   (void) FixedMemOp;
1115   (void) FixedDbg;
1116   (void) FixedInstr;
1117 }
1118 
removeInvalidSlotRanges()1119 void StackColoring::removeInvalidSlotRanges() {
1120   for (MachineBasicBlock &BB : *MF)
1121     for (MachineInstr &I : BB) {
1122       if (I.getOpcode() == TargetOpcode::LIFETIME_START ||
1123           I.getOpcode() == TargetOpcode::LIFETIME_END || I.isDebugInstr())
1124         continue;
1125 
1126       // Some intervals are suspicious! In some cases we find address
1127       // calculations outside of the lifetime zone, but not actual memory
1128       // read or write. Memory accesses outside of the lifetime zone are a clear
1129       // violation, but address calculations are okay. This can happen when
1130       // GEPs are hoisted outside of the lifetime zone.
1131       // So, in here we only check instructions which can read or write memory.
1132       if (!I.mayLoad() && !I.mayStore())
1133         continue;
1134 
1135       // Check all of the machine operands.
1136       for (const MachineOperand &MO : I.operands()) {
1137         if (!MO.isFI())
1138           continue;
1139 
1140         int Slot = MO.getIndex();
1141 
1142         if (Slot<0)
1143           continue;
1144 
1145         if (Intervals[Slot]->empty())
1146           continue;
1147 
1148         // Check that the used slot is inside the calculated lifetime range.
1149         // If it is not, warn about it and invalidate the range.
1150         LiveInterval *Interval = &*Intervals[Slot];
1151         SlotIndex Index = Indexes->getInstructionIndex(I);
1152         if (Interval->find(Index) == Interval->end()) {
1153           Interval->clear();
1154           LLVM_DEBUG(dbgs() << "Invalidating range #" << Slot << "\n");
1155           EscapedAllocas++;
1156         }
1157       }
1158     }
1159 }
1160 
expungeSlotMap(DenseMap<int,int> & SlotRemap,unsigned NumSlots)1161 void StackColoring::expungeSlotMap(DenseMap<int, int> &SlotRemap,
1162                                    unsigned NumSlots) {
1163   // Expunge slot remap map.
1164   for (unsigned i=0; i < NumSlots; ++i) {
1165     // If we are remapping i
1166     if (SlotRemap.count(i)) {
1167       int Target = SlotRemap[i];
1168       // As long as our target is mapped to something else, follow it.
1169       while (SlotRemap.count(Target)) {
1170         Target = SlotRemap[Target];
1171         SlotRemap[i] = Target;
1172       }
1173     }
1174   }
1175 }
1176 
runOnMachineFunction(MachineFunction & Func)1177 bool StackColoring::runOnMachineFunction(MachineFunction &Func) {
1178   LLVM_DEBUG(dbgs() << "********** Stack Coloring **********\n"
1179                     << "********** Function: " << Func.getName() << '\n');
1180   MF = &Func;
1181   MFI = &MF->getFrameInfo();
1182   Indexes = &getAnalysis<SlotIndexes>();
1183   BlockLiveness.clear();
1184   BasicBlocks.clear();
1185   BasicBlockNumbering.clear();
1186   Markers.clear();
1187   Intervals.clear();
1188   LiveStarts.clear();
1189   VNInfoAllocator.Reset();
1190 
1191   unsigned NumSlots = MFI->getObjectIndexEnd();
1192 
1193   // If there are no stack slots then there are no markers to remove.
1194   if (!NumSlots)
1195     return false;
1196 
1197   SmallVector<int, 8> SortedSlots;
1198   SortedSlots.reserve(NumSlots);
1199   Intervals.reserve(NumSlots);
1200   LiveStarts.resize(NumSlots);
1201 
1202   unsigned NumMarkers = collectMarkers(NumSlots);
1203 
1204   unsigned TotalSize = 0;
1205   LLVM_DEBUG(dbgs() << "Found " << NumMarkers << " markers and " << NumSlots
1206                     << " slots\n");
1207   LLVM_DEBUG(dbgs() << "Slot structure:\n");
1208 
1209   for (int i=0; i < MFI->getObjectIndexEnd(); ++i) {
1210     LLVM_DEBUG(dbgs() << "Slot #" << i << " - " << MFI->getObjectSize(i)
1211                       << " bytes.\n");
1212     TotalSize += MFI->getObjectSize(i);
1213   }
1214 
1215   LLVM_DEBUG(dbgs() << "Total Stack size: " << TotalSize << " bytes\n\n");
1216 
1217   // Don't continue because there are not enough lifetime markers, or the
1218   // stack is too small, or we are told not to optimize the slots.
1219   if (NumMarkers < 2 || TotalSize < 16 || DisableColoring ||
1220       skipFunction(Func.getFunction())) {
1221     LLVM_DEBUG(dbgs() << "Will not try to merge slots.\n");
1222     return removeAllMarkers();
1223   }
1224 
1225   for (unsigned i=0; i < NumSlots; ++i) {
1226     std::unique_ptr<LiveInterval> LI(new LiveInterval(i, 0));
1227     LI->getNextValue(Indexes->getZeroIndex(), VNInfoAllocator);
1228     Intervals.push_back(std::move(LI));
1229     SortedSlots.push_back(i);
1230   }
1231 
1232   // Calculate the liveness of each block.
1233   calculateLocalLiveness();
1234   LLVM_DEBUG(dbgs() << "Dataflow iterations: " << NumIterations << "\n");
1235   LLVM_DEBUG(dump());
1236 
1237   // Propagate the liveness information.
1238   calculateLiveIntervals(NumSlots);
1239   LLVM_DEBUG(dumpIntervals());
1240 
1241   // Search for allocas which are used outside of the declared lifetime
1242   // markers.
1243   if (ProtectFromEscapedAllocas)
1244     removeInvalidSlotRanges();
1245 
1246   // Maps old slots to new slots.
1247   DenseMap<int, int> SlotRemap;
1248   unsigned RemovedSlots = 0;
1249   unsigned ReducedSize = 0;
1250 
1251   // Do not bother looking at empty intervals.
1252   for (unsigned I = 0; I < NumSlots; ++I) {
1253     if (Intervals[SortedSlots[I]]->empty())
1254       SortedSlots[I] = -1;
1255   }
1256 
1257   // This is a simple greedy algorithm for merging allocas. First, sort the
1258   // slots, placing the largest slots first. Next, perform an n^2 scan and look
1259   // for disjoint slots. When you find disjoint slots, merge the smaller one
1260   // into the bigger one and update the live interval. Remove the small alloca
1261   // and continue.
1262 
1263   // Sort the slots according to their size. Place unused slots at the end.
1264   // Use stable sort to guarantee deterministic code generation.
1265   llvm::stable_sort(SortedSlots, [this](int LHS, int RHS) {
1266     // We use -1 to denote a uninteresting slot. Place these slots at the end.
1267     if (LHS == -1)
1268       return false;
1269     if (RHS == -1)
1270       return true;
1271     // Sort according to size.
1272     return MFI->getObjectSize(LHS) > MFI->getObjectSize(RHS);
1273   });
1274 
1275   for (auto &s : LiveStarts)
1276     llvm::sort(s);
1277 
1278   bool Changed = true;
1279   while (Changed) {
1280     Changed = false;
1281     for (unsigned I = 0; I < NumSlots; ++I) {
1282       if (SortedSlots[I] == -1)
1283         continue;
1284 
1285       for (unsigned J=I+1; J < NumSlots; ++J) {
1286         if (SortedSlots[J] == -1)
1287           continue;
1288 
1289         int FirstSlot = SortedSlots[I];
1290         int SecondSlot = SortedSlots[J];
1291 
1292         // Objects with different stack IDs cannot be merged.
1293         if (MFI->getStackID(FirstSlot) != MFI->getStackID(SecondSlot))
1294           continue;
1295 
1296         LiveInterval *First = &*Intervals[FirstSlot];
1297         LiveInterval *Second = &*Intervals[SecondSlot];
1298         auto &FirstS = LiveStarts[FirstSlot];
1299         auto &SecondS = LiveStarts[SecondSlot];
1300         assert(!First->empty() && !Second->empty() && "Found an empty range");
1301 
1302         // Merge disjoint slots. This is a little bit tricky - see the
1303         // Implementation Notes section for an explanation.
1304         if (!First->isLiveAtIndexes(SecondS) &&
1305             !Second->isLiveAtIndexes(FirstS)) {
1306           Changed = true;
1307           First->MergeSegmentsInAsValue(*Second, First->getValNumInfo(0));
1308 
1309           int OldSize = FirstS.size();
1310           FirstS.append(SecondS.begin(), SecondS.end());
1311           auto Mid = FirstS.begin() + OldSize;
1312           std::inplace_merge(FirstS.begin(), Mid, FirstS.end());
1313 
1314           SlotRemap[SecondSlot] = FirstSlot;
1315           SortedSlots[J] = -1;
1316           LLVM_DEBUG(dbgs() << "Merging #" << FirstSlot << " and slots #"
1317                             << SecondSlot << " together.\n");
1318           Align MaxAlignment = std::max(MFI->getObjectAlign(FirstSlot),
1319                                         MFI->getObjectAlign(SecondSlot));
1320 
1321           assert(MFI->getObjectSize(FirstSlot) >=
1322                  MFI->getObjectSize(SecondSlot) &&
1323                  "Merging a small object into a larger one");
1324 
1325           RemovedSlots+=1;
1326           ReducedSize += MFI->getObjectSize(SecondSlot);
1327           MFI->setObjectAlignment(FirstSlot, MaxAlignment);
1328           MFI->RemoveStackObject(SecondSlot);
1329         }
1330       }
1331     }
1332   }// While changed.
1333 
1334   // Record statistics.
1335   StackSpaceSaved += ReducedSize;
1336   StackSlotMerged += RemovedSlots;
1337   LLVM_DEBUG(dbgs() << "Merge " << RemovedSlots << " slots. Saved "
1338                     << ReducedSize << " bytes\n");
1339 
1340   // Scan the entire function and update all machine operands that use frame
1341   // indices to use the remapped frame index.
1342   if (!SlotRemap.empty()) {
1343     expungeSlotMap(SlotRemap, NumSlots);
1344     remapInstructions(SlotRemap);
1345   }
1346 
1347   return removeAllMarkers();
1348 }
1349