1 | //===- HexagonConstPropagation.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 | #include "HexagonInstrInfo.h" |
10 | #include "HexagonRegisterInfo.h" |
11 | #include "HexagonSubtarget.h" |
12 | #include "llvm/ADT/APFloat.h" |
13 | #include "llvm/ADT/APInt.h" |
14 | #include "llvm/ADT/PostOrderIterator.h" |
15 | #include "llvm/ADT/SetVector.h" |
16 | #include "llvm/ADT/SmallVector.h" |
17 | #include "llvm/ADT/StringRef.h" |
18 | #include "llvm/CodeGen/MachineBasicBlock.h" |
19 | #include "llvm/CodeGen/MachineFunction.h" |
20 | #include "llvm/CodeGen/MachineFunctionPass.h" |
21 | #include "llvm/CodeGen/MachineInstr.h" |
22 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
23 | #include "llvm/CodeGen/MachineOperand.h" |
24 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
25 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
26 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
27 | #include "llvm/IR/Constants.h" |
28 | #include "llvm/IR/Type.h" |
29 | #include "llvm/Pass.h" |
30 | #include "llvm/Support/Casting.h" |
31 | #include "llvm/Support/Compiler.h" |
32 | #include "llvm/Support/Debug.h" |
33 | #include "llvm/Support/ErrorHandling.h" |
34 | #include "llvm/Support/MathExtras.h" |
35 | #include "llvm/Support/raw_ostream.h" |
36 | #include <cassert> |
37 | #include <cstdint> |
38 | #include <cstring> |
39 | #include <iterator> |
40 | #include <map> |
41 | #include <queue> |
42 | #include <set> |
43 | #include <utility> |
44 | #include <vector> |
45 | |
46 | #define DEBUG_TYPE "hcp" |
47 | |
48 | using namespace llvm; |
49 | |
50 | namespace { |
51 | |
52 | // Properties of a value that are tracked by the propagation. |
53 | // A property that is marked as present (i.e. bit is set) dentes that the |
54 | // value is known (proven) to have this property. Not all combinations |
55 | // of bits make sense, for example Zero and NonZero are mutually exclusive, |
56 | // but on the other hand, Zero implies Finite. In this case, whenever |
57 | // the Zero property is present, Finite should also be present. |
58 | class ConstantProperties { |
59 | public: |
60 | enum { |
61 | Unknown = 0x0000, |
62 | Zero = 0x0001, |
63 | NonZero = 0x0002, |
64 | Finite = 0x0004, |
65 | Infinity = 0x0008, |
66 | NaN = 0x0010, |
67 | SignedZero = 0x0020, |
68 | NumericProperties = (Zero|NonZero|Finite|Infinity|NaN|SignedZero), |
69 | PosOrZero = 0x0100, |
70 | NegOrZero = 0x0200, |
71 | SignProperties = (PosOrZero|NegOrZero), |
72 | Everything = (NumericProperties|SignProperties) |
73 | }; |
74 | |
75 | // For a given constant, deduce the set of trackable properties that this |
76 | // constant has. |
77 | static uint32_t deduce(const Constant *C); |
78 | }; |
79 | |
80 | // A representation of a register as it can appear in a MachineOperand, |
81 | // i.e. a pair register:subregister. |
82 | |
83 | // FIXME: Use TargetInstrInfo::RegSubRegPair. Also duplicated in |
84 | // HexagonGenPredicate |
85 | struct RegisterSubReg { |
86 | Register Reg; |
87 | unsigned SubReg; |
88 | |
89 | explicit RegisterSubReg(unsigned R, unsigned SR = 0) : Reg(R), SubReg(SR) {} |
90 | explicit RegisterSubReg(const MachineOperand &MO) |
91 | : Reg(MO.getReg()), SubReg(MO.getSubReg()) {} |
92 | |
93 | void print(const TargetRegisterInfo *TRI = nullptr) const { |
94 | dbgs() << printReg(Reg, TRI, SubIdx: SubReg); |
95 | } |
96 | |
97 | bool operator== (const RegisterSubReg &R) const { |
98 | return (Reg == R.Reg) && (SubReg == R.SubReg); |
99 | } |
100 | }; |
101 | |
102 | // Lattice cell, based on that was described in the W-Z paper on constant |
103 | // propagation. |
104 | // Latice cell will be allowed to hold multiple constant values. While |
105 | // multiple values would normally indicate "bottom", we can still derive |
106 | // some useful information from them. For example, comparison X > 0 |
107 | // could be folded if all the values in the cell associated with X are |
108 | // positive. |
109 | class LatticeCell { |
110 | private: |
111 | enum { Normal, Top, Bottom }; |
112 | |
113 | static const unsigned MaxCellSize = 4; |
114 | |
115 | unsigned Kind:2; |
116 | unsigned Size:3; |
117 | unsigned IsSpecial:1; |
118 | unsigned :0; |
119 | |
120 | public: |
121 | union { |
122 | uint32_t Properties; |
123 | const Constant *Value; |
124 | const Constant *Values[MaxCellSize]; |
125 | }; |
126 | |
127 | LatticeCell() : Kind(Top), Size(0), IsSpecial(false) { |
128 | for (const Constant *&Value : Values) |
129 | Value = nullptr; |
130 | } |
131 | |
132 | bool meet(const LatticeCell &L); |
133 | bool add(const Constant *C); |
134 | bool add(uint32_t Property); |
135 | uint32_t properties() const; |
136 | unsigned size() const { return Size; } |
137 | |
138 | LatticeCell(const LatticeCell &L) { |
139 | // This memcpy also copies Properties (when L.Size == 0). |
140 | uint32_t N = |
141 | L.IsSpecial ? sizeof L.Properties : L.Size * sizeof(const Constant *); |
142 | memcpy(dest: Values, src: L.Values, n: N); |
143 | Kind = L.Kind; |
144 | Size = L.Size; |
145 | IsSpecial = L.IsSpecial; |
146 | } |
147 | |
148 | LatticeCell &operator=(const LatticeCell &L) { |
149 | if (this != &L) { |
150 | // This memcpy also copies Properties (when L.Size == 0). |
151 | uint32_t N = L.IsSpecial ? sizeof L.Properties |
152 | : L.Size * sizeof(const Constant *); |
153 | memcpy(dest: Values, src: L.Values, n: N); |
154 | Kind = L.Kind; |
155 | Size = L.Size; |
156 | IsSpecial = L.IsSpecial; |
157 | } |
158 | return *this; |
159 | } |
160 | |
161 | bool isSingle() const { return size() == 1; } |
162 | bool isProperty() const { return IsSpecial; } |
163 | bool isTop() const { return Kind == Top; } |
164 | bool isBottom() const { return Kind == Bottom; } |
165 | |
166 | bool setBottom() { |
167 | bool Changed = (Kind != Bottom); |
168 | Kind = Bottom; |
169 | Size = 0; |
170 | IsSpecial = false; |
171 | return Changed; |
172 | } |
173 | |
174 | void print(raw_ostream &os) const; |
175 | |
176 | private: |
177 | void setProperty() { |
178 | IsSpecial = true; |
179 | Size = 0; |
180 | Kind = Normal; |
181 | } |
182 | |
183 | bool convertToProperty(); |
184 | }; |
185 | |
186 | #ifndef NDEBUG |
187 | raw_ostream &operator<< (raw_ostream &os, const LatticeCell &L) { |
188 | L.print(os); |
189 | return os; |
190 | } |
191 | #endif |
192 | |
193 | class MachineConstEvaluator; |
194 | |
195 | class MachineConstPropagator { |
196 | public: |
197 | MachineConstPropagator(MachineConstEvaluator &E) : MCE(E) { |
198 | Bottom.setBottom(); |
199 | } |
200 | |
201 | // Mapping: vreg -> cell |
202 | // The keys are registers _without_ subregisters. This won't allow |
203 | // definitions in the form of "vreg:subreg = ...". Such definitions |
204 | // would be questionable from the point of view of SSA, since the "vreg" |
205 | // could not be initialized in its entirety (specifically, an instruction |
206 | // defining the "other part" of "vreg" would also count as a definition |
207 | // of "vreg", which would violate the SSA). |
208 | // If a value of a pair vreg:subreg needs to be obtained, the cell for |
209 | // "vreg" needs to be looked up, and then the value of subregister "subreg" |
210 | // needs to be evaluated. |
211 | class CellMap { |
212 | public: |
213 | CellMap() { |
214 | assert(Top.isTop()); |
215 | Bottom.setBottom(); |
216 | } |
217 | |
218 | void clear() { Map.clear(); } |
219 | |
220 | bool has(Register R) const { |
221 | // All non-virtual registers are considered "bottom". |
222 | if (!R.isVirtual()) |
223 | return true; |
224 | MapType::const_iterator F = Map.find(x: R); |
225 | return F != Map.end(); |
226 | } |
227 | |
228 | const LatticeCell &get(Register R) const { |
229 | if (!R.isVirtual()) |
230 | return Bottom; |
231 | MapType::const_iterator F = Map.find(x: R); |
232 | if (F != Map.end()) |
233 | return F->second; |
234 | return Top; |
235 | } |
236 | |
237 | // Invalidates any const references. |
238 | void update(Register R, const LatticeCell &L) { Map[R] = L; } |
239 | |
240 | void print(raw_ostream &os, const TargetRegisterInfo &TRI) const; |
241 | |
242 | private: |
243 | using MapType = std::map<Register, LatticeCell>; |
244 | |
245 | MapType Map; |
246 | // To avoid creating "top" entries, return a const reference to |
247 | // this cell in "get". Also, have a "Bottom" cell to return from |
248 | // get when a value of a physical register is requested. |
249 | LatticeCell Top, Bottom; |
250 | |
251 | public: |
252 | using const_iterator = MapType::const_iterator; |
253 | |
254 | const_iterator begin() const { return Map.begin(); } |
255 | const_iterator end() const { return Map.end(); } |
256 | }; |
257 | |
258 | bool run(MachineFunction &MF); |
259 | |
260 | private: |
261 | void visitPHI(const MachineInstr &PN); |
262 | void visitNonBranch(const MachineInstr &MI); |
263 | void visitBranchesFrom(const MachineInstr &BrI); |
264 | void visitUsesOf(unsigned R); |
265 | bool computeBlockSuccessors(const MachineBasicBlock *MB, |
266 | SetVector<const MachineBasicBlock*> &Targets); |
267 | void removeCFGEdge(MachineBasicBlock *From, MachineBasicBlock *To); |
268 | |
269 | void propagate(MachineFunction &MF); |
270 | bool rewrite(MachineFunction &MF); |
271 | |
272 | MachineRegisterInfo *MRI = nullptr; |
273 | MachineConstEvaluator &MCE; |
274 | |
275 | using CFGEdge = std::pair<unsigned, unsigned>; |
276 | using SetOfCFGEdge = std::set<CFGEdge>; |
277 | using SetOfInstr = std::set<const MachineInstr *>; |
278 | using QueueOfCFGEdge = std::queue<CFGEdge>; |
279 | |
280 | LatticeCell Bottom; |
281 | CellMap Cells; |
282 | SetOfCFGEdge EdgeExec; |
283 | SetOfInstr InstrExec; |
284 | QueueOfCFGEdge FlowQ; |
285 | }; |
286 | |
287 | // The "evaluator/rewriter" of machine instructions. This is an abstract |
288 | // base class that provides the interface that the propagator will use, |
289 | // as well as some helper functions that are target-independent. |
290 | class MachineConstEvaluator { |
291 | public: |
292 | MachineConstEvaluator(MachineFunction &Fn) |
293 | : TRI(*Fn.getSubtarget().getRegisterInfo()), |
294 | MF(Fn), CX(Fn.getFunction().getContext()) {} |
295 | virtual ~MachineConstEvaluator() = default; |
296 | |
297 | // The required interface: |
298 | // - A set of three "evaluate" functions. Each returns "true" if the |
299 | // computation succeeded, "false" otherwise. |
300 | // (1) Given an instruction MI, and the map with input values "Inputs", |
301 | // compute the set of output values "Outputs". An example of when |
302 | // the computation can "fail" is if MI is not an instruction that |
303 | // is recognized by the evaluator. |
304 | // (2) Given a register R (as reg:subreg), compute the cell that |
305 | // corresponds to the "subreg" part of the given register. |
306 | // (3) Given a branch instruction BrI, compute the set of target blocks. |
307 | // If the branch can fall-through, add null (0) to the list of |
308 | // possible targets. |
309 | // - A function "rewrite", that given the cell map after propagation, |
310 | // could rewrite instruction MI in a more beneficial form. Return |
311 | // "true" if a change has been made, "false" otherwise. |
312 | using CellMap = MachineConstPropagator::CellMap; |
313 | virtual bool evaluate(const MachineInstr &MI, const CellMap &Inputs, |
314 | CellMap &Outputs) = 0; |
315 | virtual bool evaluate(const RegisterSubReg &R, const LatticeCell &SrcC, |
316 | LatticeCell &Result) = 0; |
317 | virtual bool evaluate(const MachineInstr &BrI, const CellMap &Inputs, |
318 | SetVector<const MachineBasicBlock*> &Targets, |
319 | bool &CanFallThru) = 0; |
320 | virtual bool rewrite(MachineInstr &MI, const CellMap &Inputs) = 0; |
321 | |
322 | const TargetRegisterInfo &TRI; |
323 | |
324 | protected: |
325 | MachineFunction &MF; |
326 | LLVMContext &CX; |
327 | |
328 | struct Comparison { |
329 | enum { |
330 | Unk = 0x00, |
331 | EQ = 0x01, |
332 | NE = 0x02, |
333 | L = 0x04, // Less-than property. |
334 | G = 0x08, // Greater-than property. |
335 | U = 0x40, // Unsigned property. |
336 | LTs = L, |
337 | LEs = L | EQ, |
338 | GTs = G, |
339 | GEs = G | EQ, |
340 | LTu = L | U, |
341 | LEu = L | EQ | U, |
342 | GTu = G | U, |
343 | GEu = G | EQ | U |
344 | }; |
345 | |
346 | static uint32_t negate(uint32_t Cmp) { |
347 | if (Cmp == EQ) |
348 | return NE; |
349 | if (Cmp == NE) |
350 | return EQ; |
351 | assert((Cmp & (L|G)) != (L|G)); |
352 | return Cmp ^ (L|G); |
353 | } |
354 | }; |
355 | |
356 | // Helper functions. |
357 | |
358 | bool getCell(const RegisterSubReg &R, const CellMap &Inputs, LatticeCell &RC); |
359 | bool constToInt(const Constant *C, APInt &Val) const; |
360 | const ConstantInt *intToConst(const APInt &Val) const; |
361 | |
362 | // Compares. |
363 | bool evaluateCMPrr(uint32_t Cmp, const RegisterSubReg &R1, const RegisterSubReg &R2, |
364 | const CellMap &Inputs, bool &Result); |
365 | bool evaluateCMPri(uint32_t Cmp, const RegisterSubReg &R1, const APInt &A2, |
366 | const CellMap &Inputs, bool &Result); |
367 | bool evaluateCMPrp(uint32_t Cmp, const RegisterSubReg &R1, uint64_t Props2, |
368 | const CellMap &Inputs, bool &Result); |
369 | bool evaluateCMPii(uint32_t Cmp, const APInt &A1, const APInt &A2, |
370 | bool &Result); |
371 | bool evaluateCMPpi(uint32_t Cmp, uint32_t Props, const APInt &A2, |
372 | bool &Result); |
373 | bool evaluateCMPpp(uint32_t Cmp, uint32_t Props1, uint32_t Props2, |
374 | bool &Result); |
375 | |
376 | bool evaluateCOPY(const RegisterSubReg &R1, const CellMap &Inputs, |
377 | LatticeCell &Result); |
378 | |
379 | // Logical operations. |
380 | bool evaluateANDrr(const RegisterSubReg &R1, const RegisterSubReg &R2, |
381 | const CellMap &Inputs, LatticeCell &Result); |
382 | bool evaluateANDri(const RegisterSubReg &R1, const APInt &A2, |
383 | const CellMap &Inputs, LatticeCell &Result); |
384 | bool evaluateANDii(const APInt &A1, const APInt &A2, APInt &Result); |
385 | bool evaluateORrr(const RegisterSubReg &R1, const RegisterSubReg &R2, |
386 | const CellMap &Inputs, LatticeCell &Result); |
387 | bool evaluateORri(const RegisterSubReg &R1, const APInt &A2, |
388 | const CellMap &Inputs, LatticeCell &Result); |
389 | bool evaluateORii(const APInt &A1, const APInt &A2, APInt &Result); |
390 | bool evaluateXORrr(const RegisterSubReg &R1, const RegisterSubReg &R2, |
391 | const CellMap &Inputs, LatticeCell &Result); |
392 | bool evaluateXORri(const RegisterSubReg &R1, const APInt &A2, |
393 | const CellMap &Inputs, LatticeCell &Result); |
394 | bool evaluateXORii(const APInt &A1, const APInt &A2, APInt &Result); |
395 | |
396 | // Extensions. |
397 | bool evaluateZEXTr(const RegisterSubReg &R1, unsigned Width, unsigned Bits, |
398 | const CellMap &Inputs, LatticeCell &Result); |
399 | bool evaluateZEXTi(const APInt &A1, unsigned Width, unsigned Bits, |
400 | APInt &Result); |
401 | bool evaluateSEXTr(const RegisterSubReg &R1, unsigned Width, unsigned Bits, |
402 | const CellMap &Inputs, LatticeCell &Result); |
403 | bool evaluateSEXTi(const APInt &A1, unsigned Width, unsigned Bits, |
404 | APInt &Result); |
405 | |
406 | // Leading/trailing bits. |
407 | bool evaluateCLBr(const RegisterSubReg &R1, bool Zeros, bool Ones, |
408 | const CellMap &Inputs, LatticeCell &Result); |
409 | bool evaluateCLBi(const APInt &A1, bool Zeros, bool Ones, APInt &Result); |
410 | bool evaluateCTBr(const RegisterSubReg &R1, bool Zeros, bool Ones, |
411 | const CellMap &Inputs, LatticeCell &Result); |
412 | bool evaluateCTBi(const APInt &A1, bool Zeros, bool Ones, APInt &Result); |
413 | |
414 | // Bitfield extract. |
415 | bool evaluateEXTRACTr(const RegisterSubReg &R1, unsigned Width, unsigned Bits, |
416 | unsigned Offset, bool Signed, const CellMap &Inputs, |
417 | LatticeCell &Result); |
418 | bool evaluateEXTRACTi(const APInt &A1, unsigned Bits, unsigned Offset, |
419 | bool Signed, APInt &Result); |
420 | // Vector operations. |
421 | bool evaluateSplatr(const RegisterSubReg &R1, unsigned Bits, unsigned Count, |
422 | const CellMap &Inputs, LatticeCell &Result); |
423 | bool evaluateSplati(const APInt &A1, unsigned Bits, unsigned Count, |
424 | APInt &Result); |
425 | }; |
426 | |
427 | } // end anonymous namespace |
428 | |
429 | uint32_t ConstantProperties::deduce(const Constant *C) { |
430 | if (isa<ConstantInt>(Val: C)) { |
431 | const ConstantInt *CI = cast<ConstantInt>(Val: C); |
432 | if (CI->isZero()) |
433 | return Zero | PosOrZero | NegOrZero | Finite; |
434 | uint32_t Props = (NonZero | Finite); |
435 | if (CI->isNegative()) |
436 | return Props | NegOrZero; |
437 | return Props | PosOrZero; |
438 | } |
439 | |
440 | if (isa<ConstantFP>(Val: C)) { |
441 | const ConstantFP *CF = cast<ConstantFP>(Val: C); |
442 | uint32_t Props = CF->isNegative() ? (NegOrZero|NonZero) |
443 | : PosOrZero; |
444 | if (CF->isZero()) |
445 | return (Props & ~NumericProperties) | (Zero|Finite); |
446 | Props = (Props & ~NumericProperties) | NonZero; |
447 | if (CF->isNaN()) |
448 | return (Props & ~NumericProperties) | NaN; |
449 | const APFloat &Val = CF->getValueAPF(); |
450 | if (Val.isInfinity()) |
451 | return (Props & ~NumericProperties) | Infinity; |
452 | Props |= Finite; |
453 | return Props; |
454 | } |
455 | |
456 | return Unknown; |
457 | } |
458 | |
459 | // Convert a cell from a set of specific values to a cell that tracks |
460 | // properties. |
461 | bool LatticeCell::convertToProperty() { |
462 | if (isProperty()) |
463 | return false; |
464 | // Corner case: converting a fresh (top) cell to "special". |
465 | // This can happen, when adding a property to a top cell. |
466 | uint32_t Everything = ConstantProperties::Everything; |
467 | uint32_t Ps = !isTop() ? properties() |
468 | : Everything; |
469 | if (Ps != ConstantProperties::Unknown) { |
470 | Properties = Ps; |
471 | setProperty(); |
472 | } else { |
473 | setBottom(); |
474 | } |
475 | return true; |
476 | } |
477 | |
478 | #ifndef NDEBUG |
479 | void LatticeCell::print(raw_ostream &os) const { |
480 | if (isProperty()) { |
481 | os << "{ " ; |
482 | uint32_t Ps = properties(); |
483 | if (Ps & ConstantProperties::Zero) |
484 | os << "zero " ; |
485 | if (Ps & ConstantProperties::NonZero) |
486 | os << "nonzero " ; |
487 | if (Ps & ConstantProperties::Finite) |
488 | os << "finite " ; |
489 | if (Ps & ConstantProperties::Infinity) |
490 | os << "infinity " ; |
491 | if (Ps & ConstantProperties::NaN) |
492 | os << "nan " ; |
493 | if (Ps & ConstantProperties::PosOrZero) |
494 | os << "poz " ; |
495 | if (Ps & ConstantProperties::NegOrZero) |
496 | os << "nez " ; |
497 | os << '}'; |
498 | return; |
499 | } |
500 | |
501 | os << "{ " ; |
502 | if (isBottom()) { |
503 | os << "bottom" ; |
504 | } else if (isTop()) { |
505 | os << "top" ; |
506 | } else { |
507 | for (unsigned i = 0; i < size(); ++i) { |
508 | const Constant *C = Values[i]; |
509 | if (i != 0) |
510 | os << ", " ; |
511 | C->print(os); |
512 | } |
513 | } |
514 | os << " }" ; |
515 | } |
516 | #endif |
517 | |
518 | // "Meet" operation on two cells. This is the key of the propagation |
519 | // algorithm. |
520 | bool LatticeCell::meet(const LatticeCell &L) { |
521 | bool Changed = false; |
522 | if (L.isBottom()) |
523 | Changed = setBottom(); |
524 | if (isBottom() || L.isTop()) |
525 | return Changed; |
526 | if (isTop()) { |
527 | *this = L; |
528 | // L can be neither Top nor Bottom, so *this must have changed. |
529 | return true; |
530 | } |
531 | |
532 | // Top/bottom cases covered. Need to integrate L's set into ours. |
533 | if (L.isProperty()) |
534 | return add(Property: L.properties()); |
535 | for (unsigned i = 0; i < L.size(); ++i) { |
536 | const Constant *LC = L.Values[i]; |
537 | Changed |= add(C: LC); |
538 | } |
539 | return Changed; |
540 | } |
541 | |
542 | // Add a new constant to the cell. This is actually where the cell update |
543 | // happens. If a cell has room for more constants, the new constant is added. |
544 | // Otherwise, the cell is converted to a "property" cell (i.e. a cell that |
545 | // will track properties of the associated values, and not the values |
546 | // themselves. Care is taken to handle special cases, like "bottom", etc. |
547 | bool LatticeCell::add(const Constant *LC) { |
548 | assert(LC); |
549 | if (isBottom()) |
550 | return false; |
551 | |
552 | if (!isProperty()) { |
553 | // Cell is not special. Try to add the constant here first, |
554 | // if there is room. |
555 | unsigned Index = 0; |
556 | while (Index < Size) { |
557 | const Constant *C = Values[Index]; |
558 | // If the constant is already here, no change is needed. |
559 | if (C == LC) |
560 | return false; |
561 | Index++; |
562 | } |
563 | if (Index < MaxCellSize) { |
564 | Values[Index] = LC; |
565 | Kind = Normal; |
566 | Size++; |
567 | return true; |
568 | } |
569 | } |
570 | |
571 | bool Changed = false; |
572 | |
573 | // This cell is special, or is not special, but is full. After this |
574 | // it will be special. |
575 | Changed = convertToProperty(); |
576 | uint32_t Ps = properties(); |
577 | uint32_t NewPs = Ps & ConstantProperties::deduce(C: LC); |
578 | if (NewPs == ConstantProperties::Unknown) { |
579 | setBottom(); |
580 | return true; |
581 | } |
582 | if (Ps != NewPs) { |
583 | Properties = NewPs; |
584 | Changed = true; |
585 | } |
586 | return Changed; |
587 | } |
588 | |
589 | // Add a property to the cell. This will force the cell to become a property- |
590 | // tracking cell. |
591 | bool LatticeCell::add(uint32_t Property) { |
592 | bool Changed = convertToProperty(); |
593 | uint32_t Ps = properties(); |
594 | if (Ps == (Ps & Property)) |
595 | return Changed; |
596 | Properties = Property & Ps; |
597 | return true; |
598 | } |
599 | |
600 | // Return the properties of the values in the cell. This is valid for any |
601 | // cell, and does not alter the cell itself. |
602 | uint32_t LatticeCell::properties() const { |
603 | if (isProperty()) |
604 | return Properties; |
605 | assert(!isTop() && "Should not call this for a top cell" ); |
606 | if (isBottom()) |
607 | return ConstantProperties::Unknown; |
608 | |
609 | assert(size() > 0 && "Empty cell" ); |
610 | uint32_t Ps = ConstantProperties::deduce(C: Values[0]); |
611 | for (unsigned i = 1; i < size(); ++i) { |
612 | if (Ps == ConstantProperties::Unknown) |
613 | break; |
614 | Ps &= ConstantProperties::deduce(C: Values[i]); |
615 | } |
616 | return Ps; |
617 | } |
618 | |
619 | #ifndef NDEBUG |
620 | void MachineConstPropagator::CellMap::print(raw_ostream &os, |
621 | const TargetRegisterInfo &TRI) const { |
622 | for (auto &I : Map) |
623 | dbgs() << " " << printReg(I.first, &TRI) << " -> " << I.second << '\n'; |
624 | } |
625 | #endif |
626 | |
627 | void MachineConstPropagator::visitPHI(const MachineInstr &PN) { |
628 | const MachineBasicBlock *MB = PN.getParent(); |
629 | unsigned MBN = MB->getNumber(); |
630 | LLVM_DEBUG(dbgs() << "Visiting FI(" << printMBBReference(*MB) << "): " << PN); |
631 | |
632 | const MachineOperand &MD = PN.getOperand(i: 0); |
633 | RegisterSubReg DefR(MD); |
634 | assert(DefR.Reg.isVirtual()); |
635 | |
636 | bool Changed = false; |
637 | |
638 | // If the def has a sub-register, set the corresponding cell to "bottom". |
639 | if (DefR.SubReg) { |
640 | Bottomize: |
641 | const LatticeCell &T = Cells.get(R: DefR.Reg); |
642 | Changed = !T.isBottom(); |
643 | Cells.update(R: DefR.Reg, L: Bottom); |
644 | if (Changed) |
645 | visitUsesOf(R: DefR.Reg); |
646 | return; |
647 | } |
648 | |
649 | LatticeCell DefC = Cells.get(R: DefR.Reg); |
650 | |
651 | for (unsigned i = 1, n = PN.getNumOperands(); i < n; i += 2) { |
652 | const MachineBasicBlock *PB = PN.getOperand(i: i+1).getMBB(); |
653 | unsigned PBN = PB->getNumber(); |
654 | if (!EdgeExec.count(x: CFGEdge(PBN, MBN))) { |
655 | LLVM_DEBUG(dbgs() << " edge " << printMBBReference(*PB) << "->" |
656 | << printMBBReference(*MB) << " not executable\n" ); |
657 | continue; |
658 | } |
659 | const MachineOperand &SO = PN.getOperand(i); |
660 | RegisterSubReg UseR(SO); |
661 | // If the input is not a virtual register, we don't really know what |
662 | // value it holds. |
663 | if (!UseR.Reg.isVirtual()) |
664 | goto Bottomize; |
665 | // If there is no cell for an input register, it means top. |
666 | if (!Cells.has(R: UseR.Reg)) |
667 | continue; |
668 | |
669 | LatticeCell SrcC; |
670 | bool Eval = MCE.evaluate(R: UseR, SrcC: Cells.get(R: UseR.Reg), Result&: SrcC); |
671 | LLVM_DEBUG(dbgs() << " edge from " << printMBBReference(*PB) << ": " |
672 | << printReg(UseR.Reg, &MCE.TRI, UseR.SubReg) << SrcC |
673 | << '\n'); |
674 | Changed |= Eval ? DefC.meet(L: SrcC) |
675 | : DefC.setBottom(); |
676 | Cells.update(R: DefR.Reg, L: DefC); |
677 | if (DefC.isBottom()) |
678 | break; |
679 | } |
680 | if (Changed) |
681 | visitUsesOf(R: DefR.Reg); |
682 | } |
683 | |
684 | void MachineConstPropagator::visitNonBranch(const MachineInstr &MI) { |
685 | LLVM_DEBUG(dbgs() << "Visiting MI(" << printMBBReference(*MI.getParent()) |
686 | << "): " << MI); |
687 | CellMap Outputs; |
688 | bool Eval = MCE.evaluate(MI, Inputs: Cells, Outputs); |
689 | LLVM_DEBUG({ |
690 | if (Eval) { |
691 | dbgs() << " outputs:" ; |
692 | for (auto &I : Outputs) |
693 | dbgs() << ' ' << I.second; |
694 | dbgs() << '\n'; |
695 | } |
696 | }); |
697 | |
698 | // Update outputs. If the value was not computed, set all the |
699 | // def cells to bottom. |
700 | for (const MachineOperand &MO : MI.operands()) { |
701 | if (!MO.isReg() || !MO.isDef()) |
702 | continue; |
703 | RegisterSubReg DefR(MO); |
704 | // Only track virtual registers. |
705 | if (!DefR.Reg.isVirtual()) |
706 | continue; |
707 | bool Changed = false; |
708 | // If the evaluation failed, set cells for all output registers to bottom. |
709 | if (!Eval) { |
710 | const LatticeCell &T = Cells.get(R: DefR.Reg); |
711 | Changed = !T.isBottom(); |
712 | Cells.update(R: DefR.Reg, L: Bottom); |
713 | } else { |
714 | // Find the corresponding cell in the computed outputs. |
715 | // If it's not there, go on to the next def. |
716 | if (!Outputs.has(R: DefR.Reg)) |
717 | continue; |
718 | LatticeCell RC = Cells.get(R: DefR.Reg); |
719 | Changed = RC.meet(L: Outputs.get(R: DefR.Reg)); |
720 | Cells.update(R: DefR.Reg, L: RC); |
721 | } |
722 | if (Changed) |
723 | visitUsesOf(R: DefR.Reg); |
724 | } |
725 | } |
726 | |
727 | // Starting at a given branch, visit remaining branches in the block. |
728 | // Traverse over the subsequent branches for as long as the preceding one |
729 | // can fall through. Add all the possible targets to the flow work queue, |
730 | // including the potential fall-through to the layout-successor block. |
731 | void MachineConstPropagator::visitBranchesFrom(const MachineInstr &BrI) { |
732 | const MachineBasicBlock &B = *BrI.getParent(); |
733 | unsigned MBN = B.getNumber(); |
734 | MachineBasicBlock::const_iterator It = BrI.getIterator(); |
735 | MachineBasicBlock::const_iterator End = B.end(); |
736 | |
737 | SetVector<const MachineBasicBlock*> Targets; |
738 | bool EvalOk = true, FallsThru = true; |
739 | while (It != End) { |
740 | const MachineInstr &MI = *It; |
741 | InstrExec.insert(x: &MI); |
742 | LLVM_DEBUG(dbgs() << "Visiting " << (EvalOk ? "BR" : "br" ) << "(" |
743 | << printMBBReference(B) << "): " << MI); |
744 | // Do not evaluate subsequent branches if the evaluation of any of the |
745 | // previous branches failed. Keep iterating over the branches only |
746 | // to mark them as executable. |
747 | EvalOk = EvalOk && MCE.evaluate(BrI: MI, Inputs: Cells, Targets, CanFallThru&: FallsThru); |
748 | if (!EvalOk) |
749 | FallsThru = true; |
750 | if (!FallsThru) |
751 | break; |
752 | ++It; |
753 | } |
754 | |
755 | if (B.mayHaveInlineAsmBr()) |
756 | EvalOk = false; |
757 | |
758 | if (EvalOk) { |
759 | // Need to add all CFG successors that lead to EH landing pads. |
760 | // There won't be explicit branches to these blocks, but they must |
761 | // be processed. |
762 | for (const MachineBasicBlock *SB : B.successors()) { |
763 | if (SB->isEHPad()) |
764 | Targets.insert(X: SB); |
765 | } |
766 | if (FallsThru) { |
767 | const MachineFunction &MF = *B.getParent(); |
768 | MachineFunction::const_iterator BI = B.getIterator(); |
769 | MachineFunction::const_iterator Next = std::next(x: BI); |
770 | if (Next != MF.end()) |
771 | Targets.insert(X: &*Next); |
772 | } |
773 | } else { |
774 | // If the evaluation of the branches failed, make "Targets" to be the |
775 | // set of all successors of the block from the CFG. |
776 | // If the evaluation succeeded for all visited branches, then if the |
777 | // last one set "FallsThru", then add an edge to the layout successor |
778 | // to the targets. |
779 | Targets.clear(); |
780 | LLVM_DEBUG(dbgs() << " failed to evaluate a branch...adding all CFG " |
781 | "successors\n" ); |
782 | for (const MachineBasicBlock *SB : B.successors()) |
783 | Targets.insert(X: SB); |
784 | } |
785 | |
786 | for (const MachineBasicBlock *TB : Targets) { |
787 | unsigned TBN = TB->getNumber(); |
788 | LLVM_DEBUG(dbgs() << " pushing edge " << printMBBReference(B) << " -> " |
789 | << printMBBReference(*TB) << "\n" ); |
790 | FlowQ.push(x: CFGEdge(MBN, TBN)); |
791 | } |
792 | } |
793 | |
794 | void MachineConstPropagator::visitUsesOf(unsigned Reg) { |
795 | LLVM_DEBUG(dbgs() << "Visiting uses of " << printReg(Reg, &MCE.TRI) |
796 | << Cells.get(Reg) << '\n'); |
797 | for (MachineInstr &MI : MRI->use_nodbg_instructions(Reg)) { |
798 | // Do not process non-executable instructions. They can become exceutable |
799 | // later (via a flow-edge in the work queue). In such case, the instruc- |
800 | // tion will be visited at that time. |
801 | if (!InstrExec.count(x: &MI)) |
802 | continue; |
803 | if (MI.isPHI()) |
804 | visitPHI(PN: MI); |
805 | else if (!MI.isBranch()) |
806 | visitNonBranch(MI); |
807 | else |
808 | visitBranchesFrom(BrI: MI); |
809 | } |
810 | } |
811 | |
812 | bool MachineConstPropagator::computeBlockSuccessors(const MachineBasicBlock *MB, |
813 | SetVector<const MachineBasicBlock*> &Targets) { |
814 | Targets.clear(); |
815 | |
816 | MachineBasicBlock::const_iterator FirstBr = MB->end(); |
817 | for (const MachineInstr &MI : *MB) { |
818 | if (MI.getOpcode() == TargetOpcode::INLINEASM_BR) |
819 | return false; |
820 | if (MI.isDebugInstr()) |
821 | continue; |
822 | if (MI.isBranch()) { |
823 | FirstBr = MI.getIterator(); |
824 | break; |
825 | } |
826 | } |
827 | |
828 | MachineBasicBlock::const_iterator End = MB->end(); |
829 | |
830 | bool DoNext = true; |
831 | for (MachineBasicBlock::const_iterator I = FirstBr; I != End; ++I) { |
832 | const MachineInstr &MI = *I; |
833 | // Can there be debug instructions between branches? |
834 | if (MI.isDebugInstr()) |
835 | continue; |
836 | if (!InstrExec.count(x: &MI)) |
837 | continue; |
838 | bool Eval = MCE.evaluate(BrI: MI, Inputs: Cells, Targets, CanFallThru&: DoNext); |
839 | if (!Eval) |
840 | return false; |
841 | if (!DoNext) |
842 | break; |
843 | } |
844 | // If the last branch could fall-through, add block's layout successor. |
845 | if (DoNext) { |
846 | MachineFunction::const_iterator BI = MB->getIterator(); |
847 | MachineFunction::const_iterator NextI = std::next(x: BI); |
848 | if (NextI != MB->getParent()->end()) |
849 | Targets.insert(X: &*NextI); |
850 | } |
851 | |
852 | // Add all the EH landing pads. |
853 | for (const MachineBasicBlock *SB : MB->successors()) |
854 | if (SB->isEHPad()) |
855 | Targets.insert(X: SB); |
856 | |
857 | return true; |
858 | } |
859 | |
860 | void MachineConstPropagator::removeCFGEdge(MachineBasicBlock *From, |
861 | MachineBasicBlock *To) { |
862 | // First, remove the CFG successor/predecessor information. |
863 | From->removeSuccessor(Succ: To); |
864 | // Remove all corresponding PHI operands in the To block. |
865 | for (MachineInstr &PN : To->phis()) { |
866 | // reg0 = PHI reg1, bb2, reg3, bb4, ... |
867 | int N = PN.getNumOperands() - 2; |
868 | while (N > 0) { |
869 | if (PN.getOperand(i: N + 1).getMBB() == From) { |
870 | PN.removeOperand(OpNo: N + 1); |
871 | PN.removeOperand(OpNo: N); |
872 | } |
873 | N -= 2; |
874 | } |
875 | } |
876 | } |
877 | |
878 | void MachineConstPropagator::propagate(MachineFunction &MF) { |
879 | MachineBasicBlock *Entry = GraphTraits<MachineFunction*>::getEntryNode(F: &MF); |
880 | unsigned EntryNum = Entry->getNumber(); |
881 | |
882 | // Start with a fake edge, just to process the entry node. |
883 | FlowQ.push(x: CFGEdge(EntryNum, EntryNum)); |
884 | |
885 | while (!FlowQ.empty()) { |
886 | CFGEdge Edge = FlowQ.front(); |
887 | FlowQ.pop(); |
888 | |
889 | LLVM_DEBUG( |
890 | dbgs() << "Picked edge " |
891 | << printMBBReference(*MF.getBlockNumbered(Edge.first)) << "->" |
892 | << printMBBReference(*MF.getBlockNumbered(Edge.second)) << '\n'); |
893 | if (Edge.first != EntryNum) |
894 | if (EdgeExec.count(x: Edge)) |
895 | continue; |
896 | EdgeExec.insert(x: Edge); |
897 | MachineBasicBlock *SB = MF.getBlockNumbered(N: Edge.second); |
898 | |
899 | // Process the block in three stages: |
900 | // - visit all PHI nodes, |
901 | // - visit all non-branch instructions, |
902 | // - visit block branches. |
903 | MachineBasicBlock::const_iterator It = SB->begin(), End = SB->end(); |
904 | |
905 | // Visit PHI nodes in the successor block. |
906 | while (It != End && It->isPHI()) { |
907 | InstrExec.insert(x: &*It); |
908 | visitPHI(PN: *It); |
909 | ++It; |
910 | } |
911 | |
912 | // If the successor block just became executable, visit all instructions. |
913 | // To see if this is the first time we're visiting it, check the first |
914 | // non-debug instruction to see if it is executable. |
915 | while (It != End && It->isDebugInstr()) |
916 | ++It; |
917 | assert(It == End || !It->isPHI()); |
918 | // If this block has been visited, go on to the next one. |
919 | if (It != End && InstrExec.count(x: &*It)) |
920 | continue; |
921 | // For now, scan all non-branch instructions. Branches require different |
922 | // processing. |
923 | while (It != End && !It->isBranch()) { |
924 | if (!It->isDebugInstr()) { |
925 | InstrExec.insert(x: &*It); |
926 | visitNonBranch(MI: *It); |
927 | } |
928 | ++It; |
929 | } |
930 | |
931 | // Time to process the end of the block. This is different from |
932 | // processing regular (non-branch) instructions, because there can |
933 | // be multiple branches in a block, and they can cause the block to |
934 | // terminate early. |
935 | if (It != End) { |
936 | visitBranchesFrom(BrI: *It); |
937 | } else { |
938 | // If the block didn't have a branch, add all successor edges to the |
939 | // work queue. (There should really be only one successor in such case.) |
940 | unsigned SBN = SB->getNumber(); |
941 | for (const MachineBasicBlock *SSB : SB->successors()) |
942 | FlowQ.push(x: CFGEdge(SBN, SSB->getNumber())); |
943 | } |
944 | } // while (FlowQ) |
945 | |
946 | LLVM_DEBUG({ |
947 | dbgs() << "Cells after propagation:\n" ; |
948 | Cells.print(dbgs(), MCE.TRI); |
949 | dbgs() << "Dead CFG edges:\n" ; |
950 | for (const MachineBasicBlock &B : MF) { |
951 | unsigned BN = B.getNumber(); |
952 | for (const MachineBasicBlock *SB : B.successors()) { |
953 | unsigned SN = SB->getNumber(); |
954 | if (!EdgeExec.count(CFGEdge(BN, SN))) |
955 | dbgs() << " " << printMBBReference(B) << " -> " |
956 | << printMBBReference(*SB) << '\n'; |
957 | } |
958 | } |
959 | }); |
960 | } |
961 | |
962 | bool MachineConstPropagator::rewrite(MachineFunction &MF) { |
963 | bool Changed = false; |
964 | // Rewrite all instructions based on the collected cell information. |
965 | // |
966 | // Traverse the instructions in a post-order, so that rewriting an |
967 | // instruction can make changes "downstream" in terms of control-flow |
968 | // without affecting the rewriting process. (We should not change |
969 | // instructions that have not yet been visited by the rewriter.) |
970 | // The reason for this is that the rewriter can introduce new vregs, |
971 | // and replace uses of old vregs (which had corresponding cells |
972 | // computed during propagation) with these new vregs (which at this |
973 | // point would not have any cells, and would appear to be "top"). |
974 | // If an attempt was made to evaluate an instruction with a fresh |
975 | // "top" vreg, it would cause an error (abend) in the evaluator. |
976 | |
977 | // Collect the post-order-traversal block ordering. The subsequent |
978 | // traversal/rewrite will update block successors, so it's safer |
979 | // if the visiting order it computed ahead of time. |
980 | std::vector<MachineBasicBlock*> POT; |
981 | for (MachineBasicBlock *B : post_order(G: &MF)) |
982 | if (!B->empty()) |
983 | POT.push_back(x: B); |
984 | |
985 | for (MachineBasicBlock *B : POT) { |
986 | // Walk the block backwards (which usually begin with the branches). |
987 | // If any branch is rewritten, we may need to update the successor |
988 | // information for this block. Unless the block's successors can be |
989 | // precisely determined (which may not be the case for indirect |
990 | // branches), we cannot modify any branch. |
991 | |
992 | // Compute the successor information. |
993 | SetVector<const MachineBasicBlock*> Targets; |
994 | bool HaveTargets = computeBlockSuccessors(MB: B, Targets); |
995 | // Rewrite the executable instructions. Skip branches if we don't |
996 | // have block successor information. |
997 | for (MachineInstr &MI : llvm::reverse(C&: *B)) { |
998 | if (InstrExec.count(x: &MI)) { |
999 | if (MI.isBranch() && !HaveTargets) |
1000 | continue; |
1001 | Changed |= MCE.rewrite(MI, Inputs: Cells); |
1002 | } |
1003 | } |
1004 | // The rewriting could rewrite PHI nodes to non-PHI nodes, causing |
1005 | // regular instructions to appear in between PHI nodes. Bring all |
1006 | // the PHI nodes to the beginning of the block. |
1007 | for (auto I = B->begin(), E = B->end(); I != E; ++I) { |
1008 | if (I->isPHI()) |
1009 | continue; |
1010 | // I is not PHI. Find the next PHI node P. |
1011 | auto P = I; |
1012 | while (++P != E) |
1013 | if (P->isPHI()) |
1014 | break; |
1015 | // Not found. |
1016 | if (P == E) |
1017 | break; |
1018 | // Splice P right before I. |
1019 | B->splice(Where: I, Other: B, From: P); |
1020 | // Reset I to point at the just spliced PHI node. |
1021 | --I; |
1022 | } |
1023 | // Update the block successor information: remove unnecessary successors. |
1024 | if (HaveTargets) { |
1025 | SmallVector<MachineBasicBlock*,2> ToRemove; |
1026 | for (MachineBasicBlock *SB : B->successors()) { |
1027 | if (!Targets.count(key: SB)) |
1028 | ToRemove.push_back(Elt: const_cast<MachineBasicBlock*>(SB)); |
1029 | Targets.remove(X: SB); |
1030 | } |
1031 | for (MachineBasicBlock *MBB : ToRemove) |
1032 | removeCFGEdge(From: B, To: MBB); |
1033 | // If there are any blocks left in the computed targets, it means that |
1034 | // we think that the block could go somewhere, but the CFG does not. |
1035 | // This could legitimately happen in blocks that have non-returning |
1036 | // calls---we would think that the execution can continue, but the |
1037 | // CFG will not have a successor edge. |
1038 | } |
1039 | } |
1040 | // Need to do some final post-processing. |
1041 | // If a branch was not executable, it will not get rewritten, but should |
1042 | // be removed (or replaced with something equivalent to a A2_nop). We can't |
1043 | // erase instructions during rewriting, so this needs to be delayed until |
1044 | // now. |
1045 | for (MachineBasicBlock &B : MF) { |
1046 | for (MachineInstr &MI : llvm::make_early_inc_range(Range&: B)) |
1047 | if (MI.isBranch() && !InstrExec.count(x: &MI)) |
1048 | B.erase(I: &MI); |
1049 | } |
1050 | return Changed; |
1051 | } |
1052 | |
1053 | // This is the constant propagation algorithm as described by Wegman-Zadeck. |
1054 | // Most of the terminology comes from there. |
1055 | bool MachineConstPropagator::run(MachineFunction &MF) { |
1056 | LLVM_DEBUG(MF.print(dbgs() << "Starting MachineConstPropagator\n" , nullptr)); |
1057 | |
1058 | MRI = &MF.getRegInfo(); |
1059 | |
1060 | Cells.clear(); |
1061 | EdgeExec.clear(); |
1062 | InstrExec.clear(); |
1063 | assert(FlowQ.empty()); |
1064 | |
1065 | propagate(MF); |
1066 | bool Changed = rewrite(MF); |
1067 | |
1068 | LLVM_DEBUG({ |
1069 | dbgs() << "End of MachineConstPropagator (Changed=" << Changed << ")\n" ; |
1070 | if (Changed) |
1071 | MF.print(dbgs(), nullptr); |
1072 | }); |
1073 | return Changed; |
1074 | } |
1075 | |
1076 | // -------------------------------------------------------------------- |
1077 | // Machine const evaluator. |
1078 | |
1079 | bool MachineConstEvaluator::getCell(const RegisterSubReg &R, const CellMap &Inputs, |
1080 | LatticeCell &RC) { |
1081 | if (!R.Reg.isVirtual()) |
1082 | return false; |
1083 | const LatticeCell &L = Inputs.get(R: R.Reg); |
1084 | if (!R.SubReg) { |
1085 | RC = L; |
1086 | return !RC.isBottom(); |
1087 | } |
1088 | bool Eval = evaluate(R, SrcC: L, Result&: RC); |
1089 | return Eval && !RC.isBottom(); |
1090 | } |
1091 | |
1092 | bool MachineConstEvaluator::constToInt(const Constant *C, |
1093 | APInt &Val) const { |
1094 | const ConstantInt *CI = dyn_cast<ConstantInt>(Val: C); |
1095 | if (!CI) |
1096 | return false; |
1097 | Val = CI->getValue(); |
1098 | return true; |
1099 | } |
1100 | |
1101 | const ConstantInt *MachineConstEvaluator::intToConst(const APInt &Val) const { |
1102 | return ConstantInt::get(Context&: CX, V: Val); |
1103 | } |
1104 | |
1105 | bool MachineConstEvaluator::evaluateCMPrr(uint32_t Cmp, const RegisterSubReg &R1, |
1106 | const RegisterSubReg &R2, const CellMap &Inputs, bool &Result) { |
1107 | assert(Inputs.has(R1.Reg) && Inputs.has(R2.Reg)); |
1108 | LatticeCell LS1, LS2; |
1109 | if (!getCell(R: R1, Inputs, RC&: LS1) || !getCell(R: R2, Inputs, RC&: LS2)) |
1110 | return false; |
1111 | |
1112 | bool IsProp1 = LS1.isProperty(); |
1113 | bool IsProp2 = LS2.isProperty(); |
1114 | if (IsProp1) { |
1115 | uint32_t Prop1 = LS1.properties(); |
1116 | if (IsProp2) |
1117 | return evaluateCMPpp(Cmp, Props1: Prop1, Props2: LS2.properties(), Result); |
1118 | uint32_t NegCmp = Comparison::negate(Cmp); |
1119 | return evaluateCMPrp(Cmp: NegCmp, R1: R2, Props2: Prop1, Inputs, Result); |
1120 | } |
1121 | if (IsProp2) { |
1122 | uint32_t Prop2 = LS2.properties(); |
1123 | return evaluateCMPrp(Cmp, R1, Props2: Prop2, Inputs, Result); |
1124 | } |
1125 | |
1126 | APInt A; |
1127 | bool IsTrue = true, IsFalse = true; |
1128 | for (unsigned i = 0; i < LS2.size(); ++i) { |
1129 | bool Res; |
1130 | bool Computed = constToInt(C: LS2.Values[i], Val&: A) && |
1131 | evaluateCMPri(Cmp, R1, A2: A, Inputs, Result&: Res); |
1132 | if (!Computed) |
1133 | return false; |
1134 | IsTrue &= Res; |
1135 | IsFalse &= !Res; |
1136 | } |
1137 | assert(!IsTrue || !IsFalse); |
1138 | // The actual logical value of the comparison is same as IsTrue. |
1139 | Result = IsTrue; |
1140 | // Return true if the result was proven to be true or proven to be false. |
1141 | return IsTrue || IsFalse; |
1142 | } |
1143 | |
1144 | bool MachineConstEvaluator::evaluateCMPri(uint32_t Cmp, const RegisterSubReg &R1, |
1145 | const APInt &A2, const CellMap &Inputs, bool &Result) { |
1146 | assert(Inputs.has(R1.Reg)); |
1147 | LatticeCell LS; |
1148 | if (!getCell(R: R1, Inputs, RC&: LS)) |
1149 | return false; |
1150 | if (LS.isProperty()) |
1151 | return evaluateCMPpi(Cmp, Props: LS.properties(), A2, Result); |
1152 | |
1153 | APInt A; |
1154 | bool IsTrue = true, IsFalse = true; |
1155 | for (unsigned i = 0; i < LS.size(); ++i) { |
1156 | bool Res; |
1157 | bool Computed = constToInt(C: LS.Values[i], Val&: A) && |
1158 | evaluateCMPii(Cmp, A1: A, A2, Result&: Res); |
1159 | if (!Computed) |
1160 | return false; |
1161 | IsTrue &= Res; |
1162 | IsFalse &= !Res; |
1163 | } |
1164 | assert(!IsTrue || !IsFalse); |
1165 | // The actual logical value of the comparison is same as IsTrue. |
1166 | Result = IsTrue; |
1167 | // Return true if the result was proven to be true or proven to be false. |
1168 | return IsTrue || IsFalse; |
1169 | } |
1170 | |
1171 | bool MachineConstEvaluator::evaluateCMPrp(uint32_t Cmp, const RegisterSubReg &R1, |
1172 | uint64_t Props2, const CellMap &Inputs, bool &Result) { |
1173 | assert(Inputs.has(R1.Reg)); |
1174 | LatticeCell LS; |
1175 | if (!getCell(R: R1, Inputs, RC&: LS)) |
1176 | return false; |
1177 | if (LS.isProperty()) |
1178 | return evaluateCMPpp(Cmp, Props1: LS.properties(), Props2, Result); |
1179 | |
1180 | APInt A; |
1181 | uint32_t NegCmp = Comparison::negate(Cmp); |
1182 | bool IsTrue = true, IsFalse = true; |
1183 | for (unsigned i = 0; i < LS.size(); ++i) { |
1184 | bool Res; |
1185 | bool Computed = constToInt(C: LS.Values[i], Val&: A) && |
1186 | evaluateCMPpi(Cmp: NegCmp, Props: Props2, A2: A, Result&: Res); |
1187 | if (!Computed) |
1188 | return false; |
1189 | IsTrue &= Res; |
1190 | IsFalse &= !Res; |
1191 | } |
1192 | assert(!IsTrue || !IsFalse); |
1193 | Result = IsTrue; |
1194 | return IsTrue || IsFalse; |
1195 | } |
1196 | |
1197 | bool MachineConstEvaluator::evaluateCMPii(uint32_t Cmp, const APInt &A1, |
1198 | const APInt &A2, bool &Result) { |
1199 | // NE is a special kind of comparison (not composed of smaller properties). |
1200 | if (Cmp == Comparison::NE) { |
1201 | Result = !APInt::isSameValue(I1: A1, I2: A2); |
1202 | return true; |
1203 | } |
1204 | if (Cmp == Comparison::EQ) { |
1205 | Result = APInt::isSameValue(I1: A1, I2: A2); |
1206 | return true; |
1207 | } |
1208 | if (Cmp & Comparison::EQ) { |
1209 | if (APInt::isSameValue(I1: A1, I2: A2)) |
1210 | return (Result = true); |
1211 | } |
1212 | assert((Cmp & (Comparison::L | Comparison::G)) && "Malformed comparison" ); |
1213 | Result = false; |
1214 | |
1215 | unsigned W1 = A1.getBitWidth(); |
1216 | unsigned W2 = A2.getBitWidth(); |
1217 | unsigned MaxW = (W1 >= W2) ? W1 : W2; |
1218 | if (Cmp & Comparison::U) { |
1219 | APInt Zx1 = A1.zext(width: MaxW); |
1220 | APInt Zx2 = A2.zext(width: MaxW); |
1221 | if (Cmp & Comparison::L) |
1222 | Result = Zx1.ult(RHS: Zx2); |
1223 | else if (Cmp & Comparison::G) |
1224 | Result = Zx2.ult(RHS: Zx1); |
1225 | return true; |
1226 | } |
1227 | |
1228 | // Signed comparison. |
1229 | APInt Sx1 = A1.sext(width: MaxW); |
1230 | APInt Sx2 = A2.sext(width: MaxW); |
1231 | if (Cmp & Comparison::L) |
1232 | Result = Sx1.slt(RHS: Sx2); |
1233 | else if (Cmp & Comparison::G) |
1234 | Result = Sx2.slt(RHS: Sx1); |
1235 | return true; |
1236 | } |
1237 | |
1238 | bool MachineConstEvaluator::evaluateCMPpi(uint32_t Cmp, uint32_t Props, |
1239 | const APInt &A2, bool &Result) { |
1240 | if (Props == ConstantProperties::Unknown) |
1241 | return false; |
1242 | |
1243 | // Should never see NaN here, but check for it for completeness. |
1244 | if (Props & ConstantProperties::NaN) |
1245 | return false; |
1246 | // Infinity could theoretically be compared to a number, but the |
1247 | // presence of infinity here would be very suspicious. If we don't |
1248 | // know for sure that the number is finite, bail out. |
1249 | if (!(Props & ConstantProperties::Finite)) |
1250 | return false; |
1251 | |
1252 | // Let X be a number that has properties Props. |
1253 | |
1254 | if (Cmp & Comparison::U) { |
1255 | // In case of unsigned comparisons, we can only compare against 0. |
1256 | if (A2 == 0) { |
1257 | // Any x!=0 will be considered >0 in an unsigned comparison. |
1258 | if (Props & ConstantProperties::Zero) |
1259 | Result = (Cmp & Comparison::EQ); |
1260 | else if (Props & ConstantProperties::NonZero) |
1261 | Result = (Cmp & Comparison::G) || (Cmp == Comparison::NE); |
1262 | else |
1263 | return false; |
1264 | return true; |
1265 | } |
1266 | // A2 is not zero. The only handled case is if X = 0. |
1267 | if (Props & ConstantProperties::Zero) { |
1268 | Result = (Cmp & Comparison::L) || (Cmp == Comparison::NE); |
1269 | return true; |
1270 | } |
1271 | return false; |
1272 | } |
1273 | |
1274 | // Signed comparisons are different. |
1275 | if (Props & ConstantProperties::Zero) { |
1276 | if (A2 == 0) |
1277 | Result = (Cmp & Comparison::EQ); |
1278 | else |
1279 | Result = (Cmp == Comparison::NE) || |
1280 | ((Cmp & Comparison::L) && !A2.isNegative()) || |
1281 | ((Cmp & Comparison::G) && A2.isNegative()); |
1282 | return true; |
1283 | } |
1284 | if (Props & ConstantProperties::PosOrZero) { |
1285 | // X >= 0 and !(A2 < 0) => cannot compare |
1286 | if (!A2.isNegative()) |
1287 | return false; |
1288 | // X >= 0 and A2 < 0 |
1289 | Result = (Cmp & Comparison::G) || (Cmp == Comparison::NE); |
1290 | return true; |
1291 | } |
1292 | if (Props & ConstantProperties::NegOrZero) { |
1293 | // X <= 0 and Src1 < 0 => cannot compare |
1294 | if (A2 == 0 || A2.isNegative()) |
1295 | return false; |
1296 | // X <= 0 and A2 > 0 |
1297 | Result = (Cmp & Comparison::L) || (Cmp == Comparison::NE); |
1298 | return true; |
1299 | } |
1300 | |
1301 | return false; |
1302 | } |
1303 | |
1304 | bool MachineConstEvaluator::evaluateCMPpp(uint32_t Cmp, uint32_t Props1, |
1305 | uint32_t Props2, bool &Result) { |
1306 | using P = ConstantProperties; |
1307 | |
1308 | if ((Props1 & P::NaN) && (Props2 & P::NaN)) |
1309 | return false; |
1310 | if (!(Props1 & P::Finite) || !(Props2 & P::Finite)) |
1311 | return false; |
1312 | |
1313 | bool Zero1 = (Props1 & P::Zero), Zero2 = (Props2 & P::Zero); |
1314 | bool NonZero1 = (Props1 & P::NonZero), NonZero2 = (Props2 & P::NonZero); |
1315 | if (Zero1 && Zero2) { |
1316 | Result = (Cmp & Comparison::EQ); |
1317 | return true; |
1318 | } |
1319 | if (Cmp == Comparison::NE) { |
1320 | if ((Zero1 && NonZero2) || (NonZero1 && Zero2)) |
1321 | return (Result = true); |
1322 | return false; |
1323 | } |
1324 | |
1325 | if (Cmp & Comparison::U) { |
1326 | // In unsigned comparisons, we can only compare against a known zero, |
1327 | // or a known non-zero. |
1328 | if (Zero1 && NonZero2) { |
1329 | Result = (Cmp & Comparison::L); |
1330 | return true; |
1331 | } |
1332 | if (NonZero1 && Zero2) { |
1333 | Result = (Cmp & Comparison::G); |
1334 | return true; |
1335 | } |
1336 | return false; |
1337 | } |
1338 | |
1339 | // Signed comparison. The comparison is not NE. |
1340 | bool Poz1 = (Props1 & P::PosOrZero), Poz2 = (Props2 & P::PosOrZero); |
1341 | bool Nez1 = (Props1 & P::NegOrZero), Nez2 = (Props2 & P::NegOrZero); |
1342 | if (Nez1 && Poz2) { |
1343 | if (NonZero1 || NonZero2) { |
1344 | Result = (Cmp & Comparison::L); |
1345 | return true; |
1346 | } |
1347 | // Either (or both) could be zero. Can only say that X <= Y. |
1348 | if ((Cmp & Comparison::EQ) && (Cmp & Comparison::L)) |
1349 | return (Result = true); |
1350 | } |
1351 | if (Poz1 && Nez2) { |
1352 | if (NonZero1 || NonZero2) { |
1353 | Result = (Cmp & Comparison::G); |
1354 | return true; |
1355 | } |
1356 | // Either (or both) could be zero. Can only say that X >= Y. |
1357 | if ((Cmp & Comparison::EQ) && (Cmp & Comparison::G)) |
1358 | return (Result = true); |
1359 | } |
1360 | |
1361 | return false; |
1362 | } |
1363 | |
1364 | bool MachineConstEvaluator::evaluateCOPY(const RegisterSubReg &R1, |
1365 | const CellMap &Inputs, LatticeCell &Result) { |
1366 | return getCell(R: R1, Inputs, RC&: Result); |
1367 | } |
1368 | |
1369 | bool MachineConstEvaluator::evaluateANDrr(const RegisterSubReg &R1, |
1370 | const RegisterSubReg &R2, const CellMap &Inputs, LatticeCell &Result) { |
1371 | assert(Inputs.has(R1.Reg) && Inputs.has(R2.Reg)); |
1372 | const LatticeCell &L1 = Inputs.get(R: R2.Reg); |
1373 | const LatticeCell &L2 = Inputs.get(R: R2.Reg); |
1374 | // If both sources are bottom, exit. Otherwise try to evaluate ANDri |
1375 | // with the non-bottom argument passed as the immediate. This is to |
1376 | // catch cases of ANDing with 0. |
1377 | if (L2.isBottom()) { |
1378 | if (L1.isBottom()) |
1379 | return false; |
1380 | return evaluateANDrr(R1: R2, R2: R1, Inputs, Result); |
1381 | } |
1382 | LatticeCell LS2; |
1383 | if (!evaluate(R: R2, SrcC: L2, Result&: LS2)) |
1384 | return false; |
1385 | if (LS2.isBottom() || LS2.isProperty()) |
1386 | return false; |
1387 | |
1388 | APInt A; |
1389 | for (unsigned i = 0; i < LS2.size(); ++i) { |
1390 | LatticeCell RC; |
1391 | bool Eval = constToInt(C: LS2.Values[i], Val&: A) && |
1392 | evaluateANDri(R1, A2: A, Inputs, Result&: RC); |
1393 | if (!Eval) |
1394 | return false; |
1395 | Result.meet(L: RC); |
1396 | } |
1397 | return !Result.isBottom(); |
1398 | } |
1399 | |
1400 | bool MachineConstEvaluator::evaluateANDri(const RegisterSubReg &R1, |
1401 | const APInt &A2, const CellMap &Inputs, LatticeCell &Result) { |
1402 | assert(Inputs.has(R1.Reg)); |
1403 | if (A2 == -1) |
1404 | return getCell(R: R1, Inputs, RC&: Result); |
1405 | if (A2 == 0) { |
1406 | LatticeCell RC; |
1407 | RC.add(LC: intToConst(Val: A2)); |
1408 | // Overwrite Result. |
1409 | Result = RC; |
1410 | return true; |
1411 | } |
1412 | LatticeCell LS1; |
1413 | if (!getCell(R: R1, Inputs, RC&: LS1)) |
1414 | return false; |
1415 | if (LS1.isBottom() || LS1.isProperty()) |
1416 | return false; |
1417 | |
1418 | APInt A, ResA; |
1419 | for (unsigned i = 0; i < LS1.size(); ++i) { |
1420 | bool Eval = constToInt(C: LS1.Values[i], Val&: A) && |
1421 | evaluateANDii(A1: A, A2, Result&: ResA); |
1422 | if (!Eval) |
1423 | return false; |
1424 | const Constant *C = intToConst(Val: ResA); |
1425 | Result.add(LC: C); |
1426 | } |
1427 | return !Result.isBottom(); |
1428 | } |
1429 | |
1430 | bool MachineConstEvaluator::evaluateANDii(const APInt &A1, |
1431 | const APInt &A2, APInt &Result) { |
1432 | Result = A1 & A2; |
1433 | return true; |
1434 | } |
1435 | |
1436 | bool MachineConstEvaluator::evaluateORrr(const RegisterSubReg &R1, |
1437 | const RegisterSubReg &R2, const CellMap &Inputs, LatticeCell &Result) { |
1438 | assert(Inputs.has(R1.Reg) && Inputs.has(R2.Reg)); |
1439 | const LatticeCell &L1 = Inputs.get(R: R2.Reg); |
1440 | const LatticeCell &L2 = Inputs.get(R: R2.Reg); |
1441 | // If both sources are bottom, exit. Otherwise try to evaluate ORri |
1442 | // with the non-bottom argument passed as the immediate. This is to |
1443 | // catch cases of ORing with -1. |
1444 | if (L2.isBottom()) { |
1445 | if (L1.isBottom()) |
1446 | return false; |
1447 | return evaluateORrr(R1: R2, R2: R1, Inputs, Result); |
1448 | } |
1449 | LatticeCell LS2; |
1450 | if (!evaluate(R: R2, SrcC: L2, Result&: LS2)) |
1451 | return false; |
1452 | if (LS2.isBottom() || LS2.isProperty()) |
1453 | return false; |
1454 | |
1455 | APInt A; |
1456 | for (unsigned i = 0; i < LS2.size(); ++i) { |
1457 | LatticeCell RC; |
1458 | bool Eval = constToInt(C: LS2.Values[i], Val&: A) && |
1459 | evaluateORri(R1, A2: A, Inputs, Result&: RC); |
1460 | if (!Eval) |
1461 | return false; |
1462 | Result.meet(L: RC); |
1463 | } |
1464 | return !Result.isBottom(); |
1465 | } |
1466 | |
1467 | bool MachineConstEvaluator::evaluateORri(const RegisterSubReg &R1, |
1468 | const APInt &A2, const CellMap &Inputs, LatticeCell &Result) { |
1469 | assert(Inputs.has(R1.Reg)); |
1470 | if (A2 == 0) |
1471 | return getCell(R: R1, Inputs, RC&: Result); |
1472 | if (A2 == -1) { |
1473 | LatticeCell RC; |
1474 | RC.add(LC: intToConst(Val: A2)); |
1475 | // Overwrite Result. |
1476 | Result = RC; |
1477 | return true; |
1478 | } |
1479 | LatticeCell LS1; |
1480 | if (!getCell(R: R1, Inputs, RC&: LS1)) |
1481 | return false; |
1482 | if (LS1.isBottom() || LS1.isProperty()) |
1483 | return false; |
1484 | |
1485 | APInt A, ResA; |
1486 | for (unsigned i = 0; i < LS1.size(); ++i) { |
1487 | bool Eval = constToInt(C: LS1.Values[i], Val&: A) && |
1488 | evaluateORii(A1: A, A2, Result&: ResA); |
1489 | if (!Eval) |
1490 | return false; |
1491 | const Constant *C = intToConst(Val: ResA); |
1492 | Result.add(LC: C); |
1493 | } |
1494 | return !Result.isBottom(); |
1495 | } |
1496 | |
1497 | bool MachineConstEvaluator::evaluateORii(const APInt &A1, |
1498 | const APInt &A2, APInt &Result) { |
1499 | Result = A1 | A2; |
1500 | return true; |
1501 | } |
1502 | |
1503 | bool MachineConstEvaluator::evaluateXORrr(const RegisterSubReg &R1, |
1504 | const RegisterSubReg &R2, const CellMap &Inputs, LatticeCell &Result) { |
1505 | assert(Inputs.has(R1.Reg) && Inputs.has(R2.Reg)); |
1506 | LatticeCell LS1, LS2; |
1507 | if (!getCell(R: R1, Inputs, RC&: LS1) || !getCell(R: R2, Inputs, RC&: LS2)) |
1508 | return false; |
1509 | if (LS1.isProperty()) { |
1510 | if (LS1.properties() & ConstantProperties::Zero) |
1511 | return !(Result = LS2).isBottom(); |
1512 | return false; |
1513 | } |
1514 | if (LS2.isProperty()) { |
1515 | if (LS2.properties() & ConstantProperties::Zero) |
1516 | return !(Result = LS1).isBottom(); |
1517 | return false; |
1518 | } |
1519 | |
1520 | APInt A; |
1521 | for (unsigned i = 0; i < LS2.size(); ++i) { |
1522 | LatticeCell RC; |
1523 | bool Eval = constToInt(C: LS2.Values[i], Val&: A) && |
1524 | evaluateXORri(R1, A2: A, Inputs, Result&: RC); |
1525 | if (!Eval) |
1526 | return false; |
1527 | Result.meet(L: RC); |
1528 | } |
1529 | return !Result.isBottom(); |
1530 | } |
1531 | |
1532 | bool MachineConstEvaluator::evaluateXORri(const RegisterSubReg &R1, |
1533 | const APInt &A2, const CellMap &Inputs, LatticeCell &Result) { |
1534 | assert(Inputs.has(R1.Reg)); |
1535 | LatticeCell LS1; |
1536 | if (!getCell(R: R1, Inputs, RC&: LS1)) |
1537 | return false; |
1538 | if (LS1.isProperty()) { |
1539 | if (LS1.properties() & ConstantProperties::Zero) { |
1540 | const Constant *C = intToConst(Val: A2); |
1541 | Result.add(LC: C); |
1542 | return !Result.isBottom(); |
1543 | } |
1544 | return false; |
1545 | } |
1546 | |
1547 | APInt A, XA; |
1548 | for (unsigned i = 0; i < LS1.size(); ++i) { |
1549 | bool Eval = constToInt(C: LS1.Values[i], Val&: A) && |
1550 | evaluateXORii(A1: A, A2, Result&: XA); |
1551 | if (!Eval) |
1552 | return false; |
1553 | const Constant *C = intToConst(Val: XA); |
1554 | Result.add(LC: C); |
1555 | } |
1556 | return !Result.isBottom(); |
1557 | } |
1558 | |
1559 | bool MachineConstEvaluator::evaluateXORii(const APInt &A1, |
1560 | const APInt &A2, APInt &Result) { |
1561 | Result = A1 ^ A2; |
1562 | return true; |
1563 | } |
1564 | |
1565 | bool MachineConstEvaluator::evaluateZEXTr(const RegisterSubReg &R1, unsigned Width, |
1566 | unsigned Bits, const CellMap &Inputs, LatticeCell &Result) { |
1567 | assert(Inputs.has(R1.Reg)); |
1568 | LatticeCell LS1; |
1569 | if (!getCell(R: R1, Inputs, RC&: LS1)) |
1570 | return false; |
1571 | if (LS1.isProperty()) |
1572 | return false; |
1573 | |
1574 | APInt A, XA; |
1575 | for (unsigned i = 0; i < LS1.size(); ++i) { |
1576 | bool Eval = constToInt(C: LS1.Values[i], Val&: A) && |
1577 | evaluateZEXTi(A1: A, Width, Bits, Result&: XA); |
1578 | if (!Eval) |
1579 | return false; |
1580 | const Constant *C = intToConst(Val: XA); |
1581 | Result.add(LC: C); |
1582 | } |
1583 | return true; |
1584 | } |
1585 | |
1586 | bool MachineConstEvaluator::evaluateZEXTi(const APInt &A1, unsigned Width, |
1587 | unsigned Bits, APInt &Result) { |
1588 | unsigned BW = A1.getBitWidth(); |
1589 | (void)BW; |
1590 | assert(Width >= Bits && BW >= Bits); |
1591 | APInt Mask = APInt::getLowBitsSet(numBits: Width, loBitsSet: Bits); |
1592 | Result = A1.zextOrTrunc(width: Width) & Mask; |
1593 | return true; |
1594 | } |
1595 | |
1596 | bool MachineConstEvaluator::evaluateSEXTr(const RegisterSubReg &R1, unsigned Width, |
1597 | unsigned Bits, const CellMap &Inputs, LatticeCell &Result) { |
1598 | assert(Inputs.has(R1.Reg)); |
1599 | LatticeCell LS1; |
1600 | if (!getCell(R: R1, Inputs, RC&: LS1)) |
1601 | return false; |
1602 | if (LS1.isBottom() || LS1.isProperty()) |
1603 | return false; |
1604 | |
1605 | APInt A, XA; |
1606 | for (unsigned i = 0; i < LS1.size(); ++i) { |
1607 | bool Eval = constToInt(C: LS1.Values[i], Val&: A) && |
1608 | evaluateSEXTi(A1: A, Width, Bits, Result&: XA); |
1609 | if (!Eval) |
1610 | return false; |
1611 | const Constant *C = intToConst(Val: XA); |
1612 | Result.add(LC: C); |
1613 | } |
1614 | return true; |
1615 | } |
1616 | |
1617 | bool MachineConstEvaluator::evaluateSEXTi(const APInt &A1, unsigned Width, |
1618 | unsigned Bits, APInt &Result) { |
1619 | unsigned BW = A1.getBitWidth(); |
1620 | assert(Width >= Bits && BW >= Bits); |
1621 | // Special case to make things faster for smaller source widths. |
1622 | // Sign extension of 0 bits generates 0 as a result. This is consistent |
1623 | // with what the HW does. |
1624 | if (Bits == 0) { |
1625 | Result = APInt(Width, 0); |
1626 | return true; |
1627 | } |
1628 | // In C, shifts by 64 invoke undefined behavior: handle that case in APInt. |
1629 | if (BW <= 64 && Bits != 0) { |
1630 | int64_t V = A1.getSExtValue(); |
1631 | switch (Bits) { |
1632 | case 8: |
1633 | V = static_cast<int8_t>(V); |
1634 | break; |
1635 | case 16: |
1636 | V = static_cast<int16_t>(V); |
1637 | break; |
1638 | case 32: |
1639 | V = static_cast<int32_t>(V); |
1640 | break; |
1641 | default: |
1642 | // Shift left to lose all bits except lower "Bits" bits, then shift |
1643 | // the value back, replicating what was a sign bit after the first |
1644 | // shift. |
1645 | V = (V << (64-Bits)) >> (64-Bits); |
1646 | break; |
1647 | } |
1648 | // V is a 64-bit sign-extended value. Convert it to APInt of desired |
1649 | // width. |
1650 | Result = APInt(Width, V, true); |
1651 | return true; |
1652 | } |
1653 | // Slow case: the value doesn't fit in int64_t. |
1654 | if (Bits < BW) |
1655 | Result = A1.trunc(width: Bits).sext(width: Width); |
1656 | else // Bits == BW |
1657 | Result = A1.sext(width: Width); |
1658 | return true; |
1659 | } |
1660 | |
1661 | bool MachineConstEvaluator::evaluateCLBr(const RegisterSubReg &R1, bool Zeros, |
1662 | bool Ones, const CellMap &Inputs, LatticeCell &Result) { |
1663 | assert(Inputs.has(R1.Reg)); |
1664 | LatticeCell LS1; |
1665 | if (!getCell(R: R1, Inputs, RC&: LS1)) |
1666 | return false; |
1667 | if (LS1.isBottom() || LS1.isProperty()) |
1668 | return false; |
1669 | |
1670 | APInt A, CA; |
1671 | for (unsigned i = 0; i < LS1.size(); ++i) { |
1672 | bool Eval = constToInt(C: LS1.Values[i], Val&: A) && |
1673 | evaluateCLBi(A1: A, Zeros, Ones, Result&: CA); |
1674 | if (!Eval) |
1675 | return false; |
1676 | const Constant *C = intToConst(Val: CA); |
1677 | Result.add(LC: C); |
1678 | } |
1679 | return true; |
1680 | } |
1681 | |
1682 | bool MachineConstEvaluator::evaluateCLBi(const APInt &A1, bool Zeros, |
1683 | bool Ones, APInt &Result) { |
1684 | unsigned BW = A1.getBitWidth(); |
1685 | if (!Zeros && !Ones) |
1686 | return false; |
1687 | unsigned Count = 0; |
1688 | if (Zeros && (Count == 0)) |
1689 | Count = A1.countl_zero(); |
1690 | if (Ones && (Count == 0)) |
1691 | Count = A1.countl_one(); |
1692 | Result = APInt(BW, static_cast<uint64_t>(Count), false); |
1693 | return true; |
1694 | } |
1695 | |
1696 | bool MachineConstEvaluator::evaluateCTBr(const RegisterSubReg &R1, bool Zeros, |
1697 | bool Ones, const CellMap &Inputs, LatticeCell &Result) { |
1698 | assert(Inputs.has(R1.Reg)); |
1699 | LatticeCell LS1; |
1700 | if (!getCell(R: R1, Inputs, RC&: LS1)) |
1701 | return false; |
1702 | if (LS1.isBottom() || LS1.isProperty()) |
1703 | return false; |
1704 | |
1705 | APInt A, CA; |
1706 | for (unsigned i = 0; i < LS1.size(); ++i) { |
1707 | bool Eval = constToInt(C: LS1.Values[i], Val&: A) && |
1708 | evaluateCTBi(A1: A, Zeros, Ones, Result&: CA); |
1709 | if (!Eval) |
1710 | return false; |
1711 | const Constant *C = intToConst(Val: CA); |
1712 | Result.add(LC: C); |
1713 | } |
1714 | return true; |
1715 | } |
1716 | |
1717 | bool MachineConstEvaluator::evaluateCTBi(const APInt &A1, bool Zeros, |
1718 | bool Ones, APInt &Result) { |
1719 | unsigned BW = A1.getBitWidth(); |
1720 | if (!Zeros && !Ones) |
1721 | return false; |
1722 | unsigned Count = 0; |
1723 | if (Zeros && (Count == 0)) |
1724 | Count = A1.countr_zero(); |
1725 | if (Ones && (Count == 0)) |
1726 | Count = A1.countr_one(); |
1727 | Result = APInt(BW, static_cast<uint64_t>(Count), false); |
1728 | return true; |
1729 | } |
1730 | |
1731 | bool MachineConstEvaluator::(const RegisterSubReg &R1, |
1732 | unsigned Width, unsigned Bits, unsigned Offset, bool Signed, |
1733 | const CellMap &Inputs, LatticeCell &Result) { |
1734 | assert(Inputs.has(R1.Reg)); |
1735 | assert(Bits+Offset <= Width); |
1736 | LatticeCell LS1; |
1737 | if (!getCell(R: R1, Inputs, RC&: LS1)) |
1738 | return false; |
1739 | if (LS1.isBottom()) |
1740 | return false; |
1741 | if (LS1.isProperty()) { |
1742 | uint32_t Ps = LS1.properties(); |
1743 | if (Ps & ConstantProperties::Zero) { |
1744 | const Constant *C = intToConst(Val: APInt(Width, 0, false)); |
1745 | Result.add(LC: C); |
1746 | return true; |
1747 | } |
1748 | return false; |
1749 | } |
1750 | |
1751 | APInt A, CA; |
1752 | for (unsigned i = 0; i < LS1.size(); ++i) { |
1753 | bool Eval = constToInt(C: LS1.Values[i], Val&: A) && |
1754 | evaluateEXTRACTi(A1: A, Bits, Offset, Signed, Result&: CA); |
1755 | if (!Eval) |
1756 | return false; |
1757 | const Constant *C = intToConst(Val: CA); |
1758 | Result.add(LC: C); |
1759 | } |
1760 | return true; |
1761 | } |
1762 | |
1763 | bool MachineConstEvaluator::(const APInt &A1, unsigned Bits, |
1764 | unsigned Offset, bool Signed, APInt &Result) { |
1765 | unsigned BW = A1.getBitWidth(); |
1766 | assert(Bits+Offset <= BW); |
1767 | // Extracting 0 bits generates 0 as a result (as indicated by the HW people). |
1768 | if (Bits == 0) { |
1769 | Result = APInt(BW, 0); |
1770 | return true; |
1771 | } |
1772 | if (BW <= 64) { |
1773 | int64_t V = A1.getZExtValue(); |
1774 | V <<= (64-Bits-Offset); |
1775 | if (Signed) |
1776 | V >>= (64-Bits); |
1777 | else |
1778 | V = static_cast<uint64_t>(V) >> (64-Bits); |
1779 | Result = APInt(BW, V, Signed); |
1780 | return true; |
1781 | } |
1782 | if (Signed) |
1783 | Result = A1.shl(shiftAmt: BW-Bits-Offset).ashr(ShiftAmt: BW-Bits); |
1784 | else |
1785 | Result = A1.shl(shiftAmt: BW-Bits-Offset).lshr(shiftAmt: BW-Bits); |
1786 | return true; |
1787 | } |
1788 | |
1789 | bool MachineConstEvaluator::evaluateSplatr(const RegisterSubReg &R1, |
1790 | unsigned Bits, unsigned Count, const CellMap &Inputs, |
1791 | LatticeCell &Result) { |
1792 | assert(Inputs.has(R1.Reg)); |
1793 | LatticeCell LS1; |
1794 | if (!getCell(R: R1, Inputs, RC&: LS1)) |
1795 | return false; |
1796 | if (LS1.isBottom() || LS1.isProperty()) |
1797 | return false; |
1798 | |
1799 | APInt A, SA; |
1800 | for (unsigned i = 0; i < LS1.size(); ++i) { |
1801 | bool Eval = constToInt(C: LS1.Values[i], Val&: A) && |
1802 | evaluateSplati(A1: A, Bits, Count, Result&: SA); |
1803 | if (!Eval) |
1804 | return false; |
1805 | const Constant *C = intToConst(Val: SA); |
1806 | Result.add(LC: C); |
1807 | } |
1808 | return true; |
1809 | } |
1810 | |
1811 | bool MachineConstEvaluator::evaluateSplati(const APInt &A1, unsigned Bits, |
1812 | unsigned Count, APInt &Result) { |
1813 | assert(Count > 0); |
1814 | unsigned BW = A1.getBitWidth(), SW = Count*Bits; |
1815 | APInt LoBits = (Bits < BW) ? A1.trunc(width: Bits) : A1.zext(width: Bits); |
1816 | if (Count > 1) |
1817 | LoBits = LoBits.zext(width: SW); |
1818 | |
1819 | APInt Res(SW, 0, false); |
1820 | for (unsigned i = 0; i < Count; ++i) { |
1821 | Res <<= Bits; |
1822 | Res |= LoBits; |
1823 | } |
1824 | Result = Res; |
1825 | return true; |
1826 | } |
1827 | |
1828 | // ---------------------------------------------------------------------- |
1829 | // Hexagon-specific code. |
1830 | |
1831 | namespace llvm { |
1832 | |
1833 | FunctionPass *createHexagonConstPropagationPass(); |
1834 | void initializeHexagonConstPropagationPass(PassRegistry &Registry); |
1835 | |
1836 | } // end namespace llvm |
1837 | |
1838 | namespace { |
1839 | |
1840 | class HexagonConstEvaluator : public MachineConstEvaluator { |
1841 | public: |
1842 | HexagonConstEvaluator(MachineFunction &Fn); |
1843 | |
1844 | bool evaluate(const MachineInstr &MI, const CellMap &Inputs, |
1845 | CellMap &Outputs) override; |
1846 | bool evaluate(const RegisterSubReg &R, const LatticeCell &SrcC, |
1847 | LatticeCell &Result) override; |
1848 | bool evaluate(const MachineInstr &BrI, const CellMap &Inputs, |
1849 | SetVector<const MachineBasicBlock*> &Targets, bool &FallsThru) |
1850 | override; |
1851 | bool rewrite(MachineInstr &MI, const CellMap &Inputs) override; |
1852 | |
1853 | private: |
1854 | unsigned getRegBitWidth(unsigned Reg) const; |
1855 | |
1856 | static uint32_t getCmp(unsigned Opc); |
1857 | static APInt getCmpImm(unsigned Opc, unsigned OpX, |
1858 | const MachineOperand &MO); |
1859 | void replaceWithNop(MachineInstr &MI); |
1860 | |
1861 | bool evaluateHexRSEQ32(RegisterSubReg RL, RegisterSubReg RH, const CellMap &Inputs, |
1862 | LatticeCell &Result); |
1863 | bool evaluateHexCompare(const MachineInstr &MI, const CellMap &Inputs, |
1864 | CellMap &Outputs); |
1865 | // This is suitable to be called for compare-and-jump instructions. |
1866 | bool evaluateHexCompare2(uint32_t Cmp, const MachineOperand &Src1, |
1867 | const MachineOperand &Src2, const CellMap &Inputs, bool &Result); |
1868 | bool evaluateHexLogical(const MachineInstr &MI, const CellMap &Inputs, |
1869 | CellMap &Outputs); |
1870 | bool evaluateHexCondMove(const MachineInstr &MI, const CellMap &Inputs, |
1871 | CellMap &Outputs); |
1872 | bool evaluateHexExt(const MachineInstr &MI, const CellMap &Inputs, |
1873 | CellMap &Outputs); |
1874 | bool evaluateHexVector1(const MachineInstr &MI, const CellMap &Inputs, |
1875 | CellMap &Outputs); |
1876 | bool evaluateHexVector2(const MachineInstr &MI, const CellMap &Inputs, |
1877 | CellMap &Outputs); |
1878 | |
1879 | void replaceAllRegUsesWith(Register FromReg, Register ToReg); |
1880 | bool rewriteHexBranch(MachineInstr &BrI, const CellMap &Inputs); |
1881 | bool rewriteHexConstDefs(MachineInstr &MI, const CellMap &Inputs, |
1882 | bool &AllDefs); |
1883 | bool rewriteHexConstUses(MachineInstr &MI, const CellMap &Inputs); |
1884 | |
1885 | MachineRegisterInfo *MRI; |
1886 | const HexagonInstrInfo &HII; |
1887 | const HexagonRegisterInfo &HRI; |
1888 | }; |
1889 | |
1890 | class HexagonConstPropagation : public MachineFunctionPass { |
1891 | public: |
1892 | static char ID; |
1893 | |
1894 | HexagonConstPropagation() : MachineFunctionPass(ID) {} |
1895 | |
1896 | StringRef getPassName() const override { |
1897 | return "Hexagon Constant Propagation" ; |
1898 | } |
1899 | |
1900 | bool runOnMachineFunction(MachineFunction &MF) override { |
1901 | const Function &F = MF.getFunction(); |
1902 | if (skipFunction(F)) |
1903 | return false; |
1904 | |
1905 | HexagonConstEvaluator HCE(MF); |
1906 | return MachineConstPropagator(HCE).run(MF); |
1907 | } |
1908 | }; |
1909 | |
1910 | } // end anonymous namespace |
1911 | |
1912 | char HexagonConstPropagation::ID = 0; |
1913 | |
1914 | INITIALIZE_PASS(HexagonConstPropagation, "hexagon-constp" , |
1915 | "Hexagon Constant Propagation" , false, false) |
1916 | |
1917 | HexagonConstEvaluator::HexagonConstEvaluator(MachineFunction &Fn) |
1918 | : MachineConstEvaluator(Fn), |
1919 | HII(*Fn.getSubtarget<HexagonSubtarget>().getInstrInfo()), |
1920 | HRI(*Fn.getSubtarget<HexagonSubtarget>().getRegisterInfo()) { |
1921 | MRI = &Fn.getRegInfo(); |
1922 | } |
1923 | |
1924 | bool HexagonConstEvaluator::evaluate(const MachineInstr &MI, |
1925 | const CellMap &Inputs, CellMap &Outputs) { |
1926 | if (MI.isCall()) |
1927 | return false; |
1928 | if (MI.getNumOperands() == 0 || !MI.getOperand(i: 0).isReg()) |
1929 | return false; |
1930 | const MachineOperand &MD = MI.getOperand(i: 0); |
1931 | if (!MD.isDef()) |
1932 | return false; |
1933 | |
1934 | unsigned Opc = MI.getOpcode(); |
1935 | RegisterSubReg DefR(MD); |
1936 | assert(!DefR.SubReg); |
1937 | if (!DefR.Reg.isVirtual()) |
1938 | return false; |
1939 | |
1940 | if (MI.isCopy()) { |
1941 | LatticeCell RC; |
1942 | RegisterSubReg SrcR(MI.getOperand(i: 1)); |
1943 | bool Eval = evaluateCOPY(R1: SrcR, Inputs, Result&: RC); |
1944 | if (!Eval) |
1945 | return false; |
1946 | Outputs.update(R: DefR.Reg, L: RC); |
1947 | return true; |
1948 | } |
1949 | if (MI.isRegSequence()) { |
1950 | unsigned Sub1 = MI.getOperand(i: 2).getImm(); |
1951 | unsigned Sub2 = MI.getOperand(i: 4).getImm(); |
1952 | const TargetRegisterClass &DefRC = *MRI->getRegClass(Reg: DefR.Reg); |
1953 | unsigned SubLo = HRI.getHexagonSubRegIndex(RC: DefRC, GenIdx: Hexagon::ps_sub_lo); |
1954 | unsigned SubHi = HRI.getHexagonSubRegIndex(RC: DefRC, GenIdx: Hexagon::ps_sub_hi); |
1955 | if (Sub1 != SubLo && Sub1 != SubHi) |
1956 | return false; |
1957 | if (Sub2 != SubLo && Sub2 != SubHi) |
1958 | return false; |
1959 | assert(Sub1 != Sub2); |
1960 | bool LoIs1 = (Sub1 == SubLo); |
1961 | const MachineOperand &OpLo = LoIs1 ? MI.getOperand(i: 1) : MI.getOperand(i: 3); |
1962 | const MachineOperand &OpHi = LoIs1 ? MI.getOperand(i: 3) : MI.getOperand(i: 1); |
1963 | LatticeCell RC; |
1964 | RegisterSubReg SrcRL(OpLo), SrcRH(OpHi); |
1965 | bool Eval = evaluateHexRSEQ32(RL: SrcRL, RH: SrcRH, Inputs, Result&: RC); |
1966 | if (!Eval) |
1967 | return false; |
1968 | Outputs.update(R: DefR.Reg, L: RC); |
1969 | return true; |
1970 | } |
1971 | if (MI.isCompare()) { |
1972 | bool Eval = evaluateHexCompare(MI, Inputs, Outputs); |
1973 | return Eval; |
1974 | } |
1975 | |
1976 | switch (Opc) { |
1977 | default: |
1978 | return false; |
1979 | case Hexagon::A2_tfrsi: |
1980 | case Hexagon::A2_tfrpi: |
1981 | case Hexagon::CONST32: |
1982 | case Hexagon::CONST64: |
1983 | { |
1984 | const MachineOperand &VO = MI.getOperand(i: 1); |
1985 | // The operand of CONST32 can be a blockaddress, e.g. |
1986 | // %0 = CONST32 <blockaddress(@eat, %l)> |
1987 | // Do this check for all instructions for safety. |
1988 | if (!VO.isImm()) |
1989 | return false; |
1990 | int64_t V = MI.getOperand(i: 1).getImm(); |
1991 | unsigned W = getRegBitWidth(Reg: DefR.Reg); |
1992 | if (W != 32 && W != 64) |
1993 | return false; |
1994 | IntegerType *Ty = (W == 32) ? Type::getInt32Ty(C&: CX) |
1995 | : Type::getInt64Ty(C&: CX); |
1996 | const ConstantInt *CI = ConstantInt::get(Ty, V, IsSigned: true); |
1997 | LatticeCell RC = Outputs.get(R: DefR.Reg); |
1998 | RC.add(LC: CI); |
1999 | Outputs.update(R: DefR.Reg, L: RC); |
2000 | break; |
2001 | } |
2002 | |
2003 | case Hexagon::PS_true: |
2004 | case Hexagon::PS_false: |
2005 | { |
2006 | LatticeCell RC = Outputs.get(R: DefR.Reg); |
2007 | bool NonZero = (Opc == Hexagon::PS_true); |
2008 | uint32_t P = NonZero ? ConstantProperties::NonZero |
2009 | : ConstantProperties::Zero; |
2010 | RC.add(Property: P); |
2011 | Outputs.update(R: DefR.Reg, L: RC); |
2012 | break; |
2013 | } |
2014 | |
2015 | case Hexagon::A2_and: |
2016 | case Hexagon::A2_andir: |
2017 | case Hexagon::A2_andp: |
2018 | case Hexagon::A2_or: |
2019 | case Hexagon::A2_orir: |
2020 | case Hexagon::A2_orp: |
2021 | case Hexagon::A2_xor: |
2022 | case Hexagon::A2_xorp: |
2023 | { |
2024 | bool Eval = evaluateHexLogical(MI, Inputs, Outputs); |
2025 | if (!Eval) |
2026 | return false; |
2027 | break; |
2028 | } |
2029 | |
2030 | case Hexagon::A2_combineii: // combine(#s8Ext, #s8) |
2031 | case Hexagon::A4_combineii: // combine(#s8, #u6Ext) |
2032 | { |
2033 | if (!MI.getOperand(i: 1).isImm() || !MI.getOperand(i: 2).isImm()) |
2034 | return false; |
2035 | uint64_t Hi = MI.getOperand(i: 1).getImm(); |
2036 | uint64_t Lo = MI.getOperand(i: 2).getImm(); |
2037 | uint64_t Res = (Hi << 32) | (Lo & 0xFFFFFFFF); |
2038 | IntegerType *Ty = Type::getInt64Ty(C&: CX); |
2039 | const ConstantInt *CI = ConstantInt::get(Ty, V: Res, IsSigned: false); |
2040 | LatticeCell RC = Outputs.get(R: DefR.Reg); |
2041 | RC.add(LC: CI); |
2042 | Outputs.update(R: DefR.Reg, L: RC); |
2043 | break; |
2044 | } |
2045 | |
2046 | case Hexagon::S2_setbit_i: |
2047 | { |
2048 | int64_t B = MI.getOperand(i: 2).getImm(); |
2049 | assert(B >=0 && B < 32); |
2050 | APInt A(32, (1ull << B), false); |
2051 | RegisterSubReg R(MI.getOperand(i: 1)); |
2052 | LatticeCell RC = Outputs.get(R: DefR.Reg); |
2053 | bool Eval = evaluateORri(R1: R, A2: A, Inputs, Result&: RC); |
2054 | if (!Eval) |
2055 | return false; |
2056 | Outputs.update(R: DefR.Reg, L: RC); |
2057 | break; |
2058 | } |
2059 | |
2060 | case Hexagon::C2_mux: |
2061 | case Hexagon::C2_muxir: |
2062 | case Hexagon::C2_muxri: |
2063 | case Hexagon::C2_muxii: |
2064 | { |
2065 | bool Eval = evaluateHexCondMove(MI, Inputs, Outputs); |
2066 | if (!Eval) |
2067 | return false; |
2068 | break; |
2069 | } |
2070 | |
2071 | case Hexagon::A2_sxtb: |
2072 | case Hexagon::A2_sxth: |
2073 | case Hexagon::A2_sxtw: |
2074 | case Hexagon::A2_zxtb: |
2075 | case Hexagon::A2_zxth: |
2076 | { |
2077 | bool Eval = evaluateHexExt(MI, Inputs, Outputs); |
2078 | if (!Eval) |
2079 | return false; |
2080 | break; |
2081 | } |
2082 | |
2083 | case Hexagon::S2_ct0: |
2084 | case Hexagon::S2_ct0p: |
2085 | case Hexagon::S2_ct1: |
2086 | case Hexagon::S2_ct1p: |
2087 | { |
2088 | using namespace Hexagon; |
2089 | |
2090 | bool Ones = (Opc == S2_ct1) || (Opc == S2_ct1p); |
2091 | RegisterSubReg R1(MI.getOperand(i: 1)); |
2092 | assert(Inputs.has(R1.Reg)); |
2093 | LatticeCell T; |
2094 | bool Eval = evaluateCTBr(R1, Zeros: !Ones, Ones, Inputs, Result&: T); |
2095 | if (!Eval) |
2096 | return false; |
2097 | // All of these instructions return a 32-bit value. The evaluate |
2098 | // will generate the same type as the operand, so truncate the |
2099 | // result if necessary. |
2100 | APInt C; |
2101 | LatticeCell RC = Outputs.get(R: DefR.Reg); |
2102 | for (unsigned i = 0; i < T.size(); ++i) { |
2103 | const Constant *CI = T.Values[i]; |
2104 | if (constToInt(C: CI, Val&: C) && C.getBitWidth() > 32) |
2105 | CI = intToConst(Val: C.trunc(width: 32)); |
2106 | RC.add(LC: CI); |
2107 | } |
2108 | Outputs.update(R: DefR.Reg, L: RC); |
2109 | break; |
2110 | } |
2111 | |
2112 | case Hexagon::S2_cl0: |
2113 | case Hexagon::S2_cl0p: |
2114 | case Hexagon::S2_cl1: |
2115 | case Hexagon::S2_cl1p: |
2116 | case Hexagon::S2_clb: |
2117 | case Hexagon::S2_clbp: |
2118 | { |
2119 | using namespace Hexagon; |
2120 | |
2121 | bool OnlyZeros = (Opc == S2_cl0) || (Opc == S2_cl0p); |
2122 | bool OnlyOnes = (Opc == S2_cl1) || (Opc == S2_cl1p); |
2123 | RegisterSubReg R1(MI.getOperand(i: 1)); |
2124 | assert(Inputs.has(R1.Reg)); |
2125 | LatticeCell T; |
2126 | bool Eval = evaluateCLBr(R1, Zeros: !OnlyOnes, Ones: !OnlyZeros, Inputs, Result&: T); |
2127 | if (!Eval) |
2128 | return false; |
2129 | // All of these instructions return a 32-bit value. The evaluate |
2130 | // will generate the same type as the operand, so truncate the |
2131 | // result if necessary. |
2132 | APInt C; |
2133 | LatticeCell RC = Outputs.get(R: DefR.Reg); |
2134 | for (unsigned i = 0; i < T.size(); ++i) { |
2135 | const Constant *CI = T.Values[i]; |
2136 | if (constToInt(C: CI, Val&: C) && C.getBitWidth() > 32) |
2137 | CI = intToConst(Val: C.trunc(width: 32)); |
2138 | RC.add(LC: CI); |
2139 | } |
2140 | Outputs.update(R: DefR.Reg, L: RC); |
2141 | break; |
2142 | } |
2143 | |
2144 | case Hexagon::S4_extract: |
2145 | case Hexagon::S4_extractp: |
2146 | case Hexagon::S2_extractu: |
2147 | case Hexagon::S2_extractup: |
2148 | { |
2149 | bool Signed = (Opc == Hexagon::S4_extract) || |
2150 | (Opc == Hexagon::S4_extractp); |
2151 | RegisterSubReg R1(MI.getOperand(i: 1)); |
2152 | unsigned BW = getRegBitWidth(Reg: R1.Reg); |
2153 | unsigned Bits = MI.getOperand(i: 2).getImm(); |
2154 | unsigned Offset = MI.getOperand(i: 3).getImm(); |
2155 | LatticeCell RC = Outputs.get(R: DefR.Reg); |
2156 | if (Offset >= BW) { |
2157 | APInt Zero(BW, 0, false); |
2158 | RC.add(LC: intToConst(Val: Zero)); |
2159 | break; |
2160 | } |
2161 | if (Offset+Bits > BW) { |
2162 | // If the requested bitfield extends beyond the most significant bit, |
2163 | // the extra bits are treated as 0s. To emulate this behavior, reduce |
2164 | // the number of requested bits, and make the extract unsigned. |
2165 | Bits = BW-Offset; |
2166 | Signed = false; |
2167 | } |
2168 | bool Eval = evaluateEXTRACTr(R1, Width: BW, Bits, Offset, Signed, Inputs, Result&: RC); |
2169 | if (!Eval) |
2170 | return false; |
2171 | Outputs.update(R: DefR.Reg, L: RC); |
2172 | break; |
2173 | } |
2174 | |
2175 | case Hexagon::S2_vsplatrb: |
2176 | case Hexagon::S2_vsplatrh: |
2177 | // vabsh, vabsh:sat |
2178 | // vabsw, vabsw:sat |
2179 | // vconj:sat |
2180 | // vrndwh, vrndwh:sat |
2181 | // vsathb, vsathub, vsatwuh |
2182 | // vsxtbh, vsxthw |
2183 | // vtrunehb, vtrunohb |
2184 | // vzxtbh, vzxthw |
2185 | { |
2186 | bool Eval = evaluateHexVector1(MI, Inputs, Outputs); |
2187 | if (!Eval) |
2188 | return false; |
2189 | break; |
2190 | } |
2191 | |
2192 | // TODO: |
2193 | // A2_vaddh |
2194 | // A2_vaddhs |
2195 | // A2_vaddw |
2196 | // A2_vaddws |
2197 | } |
2198 | |
2199 | return true; |
2200 | } |
2201 | |
2202 | bool HexagonConstEvaluator::evaluate(const RegisterSubReg &R, |
2203 | const LatticeCell &Input, LatticeCell &Result) { |
2204 | if (!R.SubReg) { |
2205 | Result = Input; |
2206 | return true; |
2207 | } |
2208 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: R.Reg); |
2209 | if (RC != &Hexagon::DoubleRegsRegClass) |
2210 | return false; |
2211 | if (R.SubReg != Hexagon::isub_lo && R.SubReg != Hexagon::isub_hi) |
2212 | return false; |
2213 | |
2214 | assert(!Input.isTop()); |
2215 | if (Input.isBottom()) |
2216 | return false; |
2217 | |
2218 | using P = ConstantProperties; |
2219 | |
2220 | if (Input.isProperty()) { |
2221 | uint32_t Ps = Input.properties(); |
2222 | if (Ps & (P::Zero|P::NaN)) { |
2223 | uint32_t Ns = (Ps & (P::Zero|P::NaN|P::SignProperties)); |
2224 | Result.add(Property: Ns); |
2225 | return true; |
2226 | } |
2227 | if (R.SubReg == Hexagon::isub_hi) { |
2228 | uint32_t Ns = (Ps & P::SignProperties); |
2229 | Result.add(Property: Ns); |
2230 | return true; |
2231 | } |
2232 | return false; |
2233 | } |
2234 | |
2235 | // The Input cell contains some known values. Pick the word corresponding |
2236 | // to the subregister. |
2237 | APInt A; |
2238 | for (unsigned i = 0; i < Input.size(); ++i) { |
2239 | const Constant *C = Input.Values[i]; |
2240 | if (!constToInt(C, Val&: A)) |
2241 | return false; |
2242 | if (!A.isIntN(N: 64)) |
2243 | return false; |
2244 | uint64_t U = A.getZExtValue(); |
2245 | if (R.SubReg == Hexagon::isub_hi) |
2246 | U >>= 32; |
2247 | U &= 0xFFFFFFFFULL; |
2248 | uint32_t U32 = Lo_32(Value: U); |
2249 | int32_t V32; |
2250 | memcpy(dest: &V32, src: &U32, n: sizeof V32); |
2251 | IntegerType *Ty = Type::getInt32Ty(C&: CX); |
2252 | const ConstantInt *C32 = ConstantInt::get(Ty, V: static_cast<int64_t>(V32)); |
2253 | Result.add(LC: C32); |
2254 | } |
2255 | return true; |
2256 | } |
2257 | |
2258 | bool HexagonConstEvaluator::evaluate(const MachineInstr &BrI, |
2259 | const CellMap &Inputs, SetVector<const MachineBasicBlock*> &Targets, |
2260 | bool &FallsThru) { |
2261 | // We need to evaluate one branch at a time. TII::analyzeBranch checks |
2262 | // all the branches in a basic block at once, so we cannot use it. |
2263 | unsigned Opc = BrI.getOpcode(); |
2264 | bool SimpleBranch = false; |
2265 | bool Negated = false; |
2266 | switch (Opc) { |
2267 | case Hexagon::J2_jumpf: |
2268 | case Hexagon::J2_jumpfnew: |
2269 | case Hexagon::J2_jumpfnewpt: |
2270 | Negated = true; |
2271 | [[fallthrough]]; |
2272 | case Hexagon::J2_jumpt: |
2273 | case Hexagon::J2_jumptnew: |
2274 | case Hexagon::J2_jumptnewpt: |
2275 | // Simple branch: if([!]Pn) jump ... |
2276 | // i.e. Op0 = predicate, Op1 = branch target. |
2277 | SimpleBranch = true; |
2278 | break; |
2279 | case Hexagon::J2_jump: |
2280 | Targets.insert(X: BrI.getOperand(i: 0).getMBB()); |
2281 | FallsThru = false; |
2282 | return true; |
2283 | default: |
2284 | Undetermined: |
2285 | // If the branch is of unknown type, assume that all successors are |
2286 | // executable. |
2287 | FallsThru = !BrI.isUnconditionalBranch(); |
2288 | return false; |
2289 | } |
2290 | |
2291 | if (SimpleBranch) { |
2292 | const MachineOperand &MD = BrI.getOperand(i: 0); |
2293 | RegisterSubReg PR(MD); |
2294 | // If the condition operand has a subregister, this is not something |
2295 | // we currently recognize. |
2296 | if (PR.SubReg) |
2297 | goto Undetermined; |
2298 | assert(Inputs.has(PR.Reg)); |
2299 | const LatticeCell &PredC = Inputs.get(R: PR.Reg); |
2300 | if (PredC.isBottom()) |
2301 | goto Undetermined; |
2302 | |
2303 | uint32_t Props = PredC.properties(); |
2304 | bool CTrue = false, CFalse = false; |
2305 | if (Props & ConstantProperties::Zero) |
2306 | CFalse = true; |
2307 | else if (Props & ConstantProperties::NonZero) |
2308 | CTrue = true; |
2309 | // If the condition is not known to be either, bail out. |
2310 | if (!CTrue && !CFalse) |
2311 | goto Undetermined; |
2312 | |
2313 | const MachineBasicBlock *BranchTarget = BrI.getOperand(i: 1).getMBB(); |
2314 | |
2315 | FallsThru = false; |
2316 | if ((!Negated && CTrue) || (Negated && CFalse)) |
2317 | Targets.insert(X: BranchTarget); |
2318 | else if ((!Negated && CFalse) || (Negated && CTrue)) |
2319 | FallsThru = true; |
2320 | else |
2321 | goto Undetermined; |
2322 | } |
2323 | |
2324 | return true; |
2325 | } |
2326 | |
2327 | bool HexagonConstEvaluator::rewrite(MachineInstr &MI, const CellMap &Inputs) { |
2328 | if (MI.isBranch()) |
2329 | return rewriteHexBranch(BrI&: MI, Inputs); |
2330 | |
2331 | unsigned Opc = MI.getOpcode(); |
2332 | switch (Opc) { |
2333 | default: |
2334 | break; |
2335 | case Hexagon::A2_tfrsi: |
2336 | case Hexagon::A2_tfrpi: |
2337 | case Hexagon::CONST32: |
2338 | case Hexagon::CONST64: |
2339 | case Hexagon::PS_true: |
2340 | case Hexagon::PS_false: |
2341 | return false; |
2342 | } |
2343 | |
2344 | unsigned NumOp = MI.getNumOperands(); |
2345 | if (NumOp == 0) |
2346 | return false; |
2347 | |
2348 | bool AllDefs, Changed; |
2349 | Changed = rewriteHexConstDefs(MI, Inputs, AllDefs); |
2350 | // If not all defs have been rewritten (i.e. the instruction defines |
2351 | // a register that is not compile-time constant), then try to rewrite |
2352 | // register operands that are known to be constant with immediates. |
2353 | if (!AllDefs) |
2354 | Changed |= rewriteHexConstUses(MI, Inputs); |
2355 | |
2356 | return Changed; |
2357 | } |
2358 | |
2359 | unsigned HexagonConstEvaluator::getRegBitWidth(unsigned Reg) const { |
2360 | const TargetRegisterClass *RC = MRI->getRegClass(Reg); |
2361 | if (Hexagon::IntRegsRegClass.hasSubClassEq(RC)) |
2362 | return 32; |
2363 | if (Hexagon::DoubleRegsRegClass.hasSubClassEq(RC)) |
2364 | return 64; |
2365 | if (Hexagon::PredRegsRegClass.hasSubClassEq(RC)) |
2366 | return 8; |
2367 | llvm_unreachable("Invalid register" ); |
2368 | return 0; |
2369 | } |
2370 | |
2371 | uint32_t HexagonConstEvaluator::getCmp(unsigned Opc) { |
2372 | switch (Opc) { |
2373 | case Hexagon::C2_cmpeq: |
2374 | case Hexagon::C2_cmpeqp: |
2375 | case Hexagon::A4_cmpbeq: |
2376 | case Hexagon::A4_cmpheq: |
2377 | case Hexagon::A4_cmpbeqi: |
2378 | case Hexagon::A4_cmpheqi: |
2379 | case Hexagon::C2_cmpeqi: |
2380 | case Hexagon::J4_cmpeqn1_t_jumpnv_nt: |
2381 | case Hexagon::J4_cmpeqn1_t_jumpnv_t: |
2382 | case Hexagon::J4_cmpeqi_t_jumpnv_nt: |
2383 | case Hexagon::J4_cmpeqi_t_jumpnv_t: |
2384 | case Hexagon::J4_cmpeq_t_jumpnv_nt: |
2385 | case Hexagon::J4_cmpeq_t_jumpnv_t: |
2386 | return Comparison::EQ; |
2387 | |
2388 | case Hexagon::C4_cmpneq: |
2389 | case Hexagon::C4_cmpneqi: |
2390 | case Hexagon::J4_cmpeqn1_f_jumpnv_nt: |
2391 | case Hexagon::J4_cmpeqn1_f_jumpnv_t: |
2392 | case Hexagon::J4_cmpeqi_f_jumpnv_nt: |
2393 | case Hexagon::J4_cmpeqi_f_jumpnv_t: |
2394 | case Hexagon::J4_cmpeq_f_jumpnv_nt: |
2395 | case Hexagon::J4_cmpeq_f_jumpnv_t: |
2396 | return Comparison::NE; |
2397 | |
2398 | case Hexagon::C2_cmpgt: |
2399 | case Hexagon::C2_cmpgtp: |
2400 | case Hexagon::A4_cmpbgt: |
2401 | case Hexagon::A4_cmphgt: |
2402 | case Hexagon::A4_cmpbgti: |
2403 | case Hexagon::A4_cmphgti: |
2404 | case Hexagon::C2_cmpgti: |
2405 | case Hexagon::J4_cmpgtn1_t_jumpnv_nt: |
2406 | case Hexagon::J4_cmpgtn1_t_jumpnv_t: |
2407 | case Hexagon::J4_cmpgti_t_jumpnv_nt: |
2408 | case Hexagon::J4_cmpgti_t_jumpnv_t: |
2409 | case Hexagon::J4_cmpgt_t_jumpnv_nt: |
2410 | case Hexagon::J4_cmpgt_t_jumpnv_t: |
2411 | return Comparison::GTs; |
2412 | |
2413 | case Hexagon::C4_cmplte: |
2414 | case Hexagon::C4_cmpltei: |
2415 | case Hexagon::J4_cmpgtn1_f_jumpnv_nt: |
2416 | case Hexagon::J4_cmpgtn1_f_jumpnv_t: |
2417 | case Hexagon::J4_cmpgti_f_jumpnv_nt: |
2418 | case Hexagon::J4_cmpgti_f_jumpnv_t: |
2419 | case Hexagon::J4_cmpgt_f_jumpnv_nt: |
2420 | case Hexagon::J4_cmpgt_f_jumpnv_t: |
2421 | return Comparison::LEs; |
2422 | |
2423 | case Hexagon::C2_cmpgtu: |
2424 | case Hexagon::C2_cmpgtup: |
2425 | case Hexagon::A4_cmpbgtu: |
2426 | case Hexagon::A4_cmpbgtui: |
2427 | case Hexagon::A4_cmphgtu: |
2428 | case Hexagon::A4_cmphgtui: |
2429 | case Hexagon::C2_cmpgtui: |
2430 | case Hexagon::J4_cmpgtui_t_jumpnv_nt: |
2431 | case Hexagon::J4_cmpgtui_t_jumpnv_t: |
2432 | case Hexagon::J4_cmpgtu_t_jumpnv_nt: |
2433 | case Hexagon::J4_cmpgtu_t_jumpnv_t: |
2434 | return Comparison::GTu; |
2435 | |
2436 | case Hexagon::J4_cmpltu_f_jumpnv_nt: |
2437 | case Hexagon::J4_cmpltu_f_jumpnv_t: |
2438 | return Comparison::GEu; |
2439 | |
2440 | case Hexagon::J4_cmpltu_t_jumpnv_nt: |
2441 | case Hexagon::J4_cmpltu_t_jumpnv_t: |
2442 | return Comparison::LTu; |
2443 | |
2444 | case Hexagon::J4_cmplt_f_jumpnv_nt: |
2445 | case Hexagon::J4_cmplt_f_jumpnv_t: |
2446 | return Comparison::GEs; |
2447 | |
2448 | case Hexagon::C4_cmplteu: |
2449 | case Hexagon::C4_cmplteui: |
2450 | case Hexagon::J4_cmpgtui_f_jumpnv_nt: |
2451 | case Hexagon::J4_cmpgtui_f_jumpnv_t: |
2452 | case Hexagon::J4_cmpgtu_f_jumpnv_nt: |
2453 | case Hexagon::J4_cmpgtu_f_jumpnv_t: |
2454 | return Comparison::LEu; |
2455 | |
2456 | case Hexagon::J4_cmplt_t_jumpnv_nt: |
2457 | case Hexagon::J4_cmplt_t_jumpnv_t: |
2458 | return Comparison::LTs; |
2459 | |
2460 | default: |
2461 | break; |
2462 | } |
2463 | return Comparison::Unk; |
2464 | } |
2465 | |
2466 | APInt HexagonConstEvaluator::getCmpImm(unsigned Opc, unsigned OpX, |
2467 | const MachineOperand &MO) { |
2468 | bool Signed = false; |
2469 | switch (Opc) { |
2470 | case Hexagon::A4_cmpbgtui: // u7 |
2471 | case Hexagon::A4_cmphgtui: // u7 |
2472 | break; |
2473 | case Hexagon::A4_cmpheqi: // s8 |
2474 | case Hexagon::C4_cmpneqi: // s8 |
2475 | Signed = true; |
2476 | break; |
2477 | case Hexagon::A4_cmpbeqi: // u8 |
2478 | break; |
2479 | case Hexagon::C2_cmpgtui: // u9 |
2480 | case Hexagon::C4_cmplteui: // u9 |
2481 | break; |
2482 | case Hexagon::C2_cmpeqi: // s10 |
2483 | case Hexagon::C2_cmpgti: // s10 |
2484 | case Hexagon::C4_cmpltei: // s10 |
2485 | Signed = true; |
2486 | break; |
2487 | case Hexagon::J4_cmpeqi_f_jumpnv_nt: // u5 |
2488 | case Hexagon::J4_cmpeqi_f_jumpnv_t: // u5 |
2489 | case Hexagon::J4_cmpeqi_t_jumpnv_nt: // u5 |
2490 | case Hexagon::J4_cmpeqi_t_jumpnv_t: // u5 |
2491 | case Hexagon::J4_cmpgti_f_jumpnv_nt: // u5 |
2492 | case Hexagon::J4_cmpgti_f_jumpnv_t: // u5 |
2493 | case Hexagon::J4_cmpgti_t_jumpnv_nt: // u5 |
2494 | case Hexagon::J4_cmpgti_t_jumpnv_t: // u5 |
2495 | case Hexagon::J4_cmpgtui_f_jumpnv_nt: // u5 |
2496 | case Hexagon::J4_cmpgtui_f_jumpnv_t: // u5 |
2497 | case Hexagon::J4_cmpgtui_t_jumpnv_nt: // u5 |
2498 | case Hexagon::J4_cmpgtui_t_jumpnv_t: // u5 |
2499 | break; |
2500 | default: |
2501 | llvm_unreachable("Unhandled instruction" ); |
2502 | break; |
2503 | } |
2504 | |
2505 | uint64_t Val = MO.getImm(); |
2506 | return APInt(32, Val, Signed); |
2507 | } |
2508 | |
2509 | void HexagonConstEvaluator::replaceWithNop(MachineInstr &MI) { |
2510 | MI.setDesc(HII.get(Opcode: Hexagon::A2_nop)); |
2511 | while (MI.getNumOperands() > 0) |
2512 | MI.removeOperand(OpNo: 0); |
2513 | } |
2514 | |
2515 | bool HexagonConstEvaluator::evaluateHexRSEQ32(RegisterSubReg RL, RegisterSubReg RH, |
2516 | const CellMap &Inputs, LatticeCell &Result) { |
2517 | assert(Inputs.has(RL.Reg) && Inputs.has(RH.Reg)); |
2518 | LatticeCell LSL, LSH; |
2519 | if (!getCell(R: RL, Inputs, RC&: LSL) || !getCell(R: RH, Inputs, RC&: LSH)) |
2520 | return false; |
2521 | if (LSL.isProperty() || LSH.isProperty()) |
2522 | return false; |
2523 | |
2524 | unsigned LN = LSL.size(), HN = LSH.size(); |
2525 | SmallVector<APInt,4> LoVs(LN), HiVs(HN); |
2526 | for (unsigned i = 0; i < LN; ++i) { |
2527 | bool Eval = constToInt(C: LSL.Values[i], Val&: LoVs[i]); |
2528 | if (!Eval) |
2529 | return false; |
2530 | assert(LoVs[i].getBitWidth() == 32); |
2531 | } |
2532 | for (unsigned i = 0; i < HN; ++i) { |
2533 | bool Eval = constToInt(C: LSH.Values[i], Val&: HiVs[i]); |
2534 | if (!Eval) |
2535 | return false; |
2536 | assert(HiVs[i].getBitWidth() == 32); |
2537 | } |
2538 | |
2539 | for (unsigned i = 0; i < HiVs.size(); ++i) { |
2540 | APInt HV = HiVs[i].zext(width: 64) << 32; |
2541 | for (unsigned j = 0; j < LoVs.size(); ++j) { |
2542 | APInt LV = LoVs[j].zext(width: 64); |
2543 | const Constant *C = intToConst(Val: HV | LV); |
2544 | Result.add(LC: C); |
2545 | if (Result.isBottom()) |
2546 | return false; |
2547 | } |
2548 | } |
2549 | return !Result.isBottom(); |
2550 | } |
2551 | |
2552 | bool HexagonConstEvaluator::evaluateHexCompare(const MachineInstr &MI, |
2553 | const CellMap &Inputs, CellMap &Outputs) { |
2554 | unsigned Opc = MI.getOpcode(); |
2555 | bool Classic = false; |
2556 | switch (Opc) { |
2557 | case Hexagon::C2_cmpeq: |
2558 | case Hexagon::C2_cmpeqp: |
2559 | case Hexagon::C2_cmpgt: |
2560 | case Hexagon::C2_cmpgtp: |
2561 | case Hexagon::C2_cmpgtu: |
2562 | case Hexagon::C2_cmpgtup: |
2563 | case Hexagon::C2_cmpeqi: |
2564 | case Hexagon::C2_cmpgti: |
2565 | case Hexagon::C2_cmpgtui: |
2566 | // Classic compare: Dst0 = CMP Src1, Src2 |
2567 | Classic = true; |
2568 | break; |
2569 | default: |
2570 | // Not handling other compare instructions now. |
2571 | return false; |
2572 | } |
2573 | |
2574 | if (Classic) { |
2575 | const MachineOperand &Src1 = MI.getOperand(i: 1); |
2576 | const MachineOperand &Src2 = MI.getOperand(i: 2); |
2577 | |
2578 | bool Result; |
2579 | unsigned Opc = MI.getOpcode(); |
2580 | bool Computed = evaluateHexCompare2(Cmp: Opc, Src1, Src2, Inputs, Result); |
2581 | if (Computed) { |
2582 | // Only create a zero/non-zero cell. At this time there isn't really |
2583 | // much need for specific values. |
2584 | RegisterSubReg DefR(MI.getOperand(i: 0)); |
2585 | LatticeCell L = Outputs.get(R: DefR.Reg); |
2586 | uint32_t P = Result ? ConstantProperties::NonZero |
2587 | : ConstantProperties::Zero; |
2588 | L.add(Property: P); |
2589 | Outputs.update(R: DefR.Reg, L); |
2590 | return true; |
2591 | } |
2592 | } |
2593 | |
2594 | return false; |
2595 | } |
2596 | |
2597 | bool HexagonConstEvaluator::evaluateHexCompare2(unsigned Opc, |
2598 | const MachineOperand &Src1, const MachineOperand &Src2, |
2599 | const CellMap &Inputs, bool &Result) { |
2600 | uint32_t Cmp = getCmp(Opc); |
2601 | bool Reg1 = Src1.isReg(), Reg2 = Src2.isReg(); |
2602 | bool Imm1 = Src1.isImm(), Imm2 = Src2.isImm(); |
2603 | if (Reg1) { |
2604 | RegisterSubReg R1(Src1); |
2605 | if (Reg2) { |
2606 | RegisterSubReg R2(Src2); |
2607 | return evaluateCMPrr(Cmp, R1, R2, Inputs, Result); |
2608 | } else if (Imm2) { |
2609 | APInt A2 = getCmpImm(Opc, OpX: 2, MO: Src2); |
2610 | return evaluateCMPri(Cmp, R1, A2, Inputs, Result); |
2611 | } |
2612 | } else if (Imm1) { |
2613 | APInt A1 = getCmpImm(Opc, OpX: 1, MO: Src1); |
2614 | if (Reg2) { |
2615 | RegisterSubReg R2(Src2); |
2616 | uint32_t NegCmp = Comparison::negate(Cmp); |
2617 | return evaluateCMPri(Cmp: NegCmp, R1: R2, A2: A1, Inputs, Result); |
2618 | } else if (Imm2) { |
2619 | APInt A2 = getCmpImm(Opc, OpX: 2, MO: Src2); |
2620 | return evaluateCMPii(Cmp, A1, A2, Result); |
2621 | } |
2622 | } |
2623 | // Unknown kind of comparison. |
2624 | return false; |
2625 | } |
2626 | |
2627 | bool HexagonConstEvaluator::evaluateHexLogical(const MachineInstr &MI, |
2628 | const CellMap &Inputs, CellMap &Outputs) { |
2629 | unsigned Opc = MI.getOpcode(); |
2630 | if (MI.getNumOperands() != 3) |
2631 | return false; |
2632 | const MachineOperand &Src1 = MI.getOperand(i: 1); |
2633 | const MachineOperand &Src2 = MI.getOperand(i: 2); |
2634 | RegisterSubReg R1(Src1); |
2635 | bool Eval = false; |
2636 | LatticeCell RC; |
2637 | switch (Opc) { |
2638 | default: |
2639 | return false; |
2640 | case Hexagon::A2_and: |
2641 | case Hexagon::A2_andp: |
2642 | Eval = evaluateANDrr(R1, R2: RegisterSubReg(Src2), Inputs, Result&: RC); |
2643 | break; |
2644 | case Hexagon::A2_andir: { |
2645 | if (!Src2.isImm()) |
2646 | return false; |
2647 | APInt A(32, Src2.getImm(), true); |
2648 | Eval = evaluateANDri(R1, A2: A, Inputs, Result&: RC); |
2649 | break; |
2650 | } |
2651 | case Hexagon::A2_or: |
2652 | case Hexagon::A2_orp: |
2653 | Eval = evaluateORrr(R1, R2: RegisterSubReg(Src2), Inputs, Result&: RC); |
2654 | break; |
2655 | case Hexagon::A2_orir: { |
2656 | if (!Src2.isImm()) |
2657 | return false; |
2658 | APInt A(32, Src2.getImm(), true); |
2659 | Eval = evaluateORri(R1, A2: A, Inputs, Result&: RC); |
2660 | break; |
2661 | } |
2662 | case Hexagon::A2_xor: |
2663 | case Hexagon::A2_xorp: |
2664 | Eval = evaluateXORrr(R1, R2: RegisterSubReg(Src2), Inputs, Result&: RC); |
2665 | break; |
2666 | } |
2667 | if (Eval) { |
2668 | RegisterSubReg DefR(MI.getOperand(i: 0)); |
2669 | Outputs.update(R: DefR.Reg, L: RC); |
2670 | } |
2671 | return Eval; |
2672 | } |
2673 | |
2674 | bool HexagonConstEvaluator::evaluateHexCondMove(const MachineInstr &MI, |
2675 | const CellMap &Inputs, CellMap &Outputs) { |
2676 | // Dst0 = Cond1 ? Src2 : Src3 |
2677 | RegisterSubReg CR(MI.getOperand(i: 1)); |
2678 | assert(Inputs.has(CR.Reg)); |
2679 | LatticeCell LS; |
2680 | if (!getCell(R: CR, Inputs, RC&: LS)) |
2681 | return false; |
2682 | uint32_t Ps = LS.properties(); |
2683 | unsigned TakeOp; |
2684 | if (Ps & ConstantProperties::Zero) |
2685 | TakeOp = 3; |
2686 | else if (Ps & ConstantProperties::NonZero) |
2687 | TakeOp = 2; |
2688 | else |
2689 | return false; |
2690 | |
2691 | const MachineOperand &ValOp = MI.getOperand(i: TakeOp); |
2692 | RegisterSubReg DefR(MI.getOperand(i: 0)); |
2693 | LatticeCell RC = Outputs.get(R: DefR.Reg); |
2694 | |
2695 | if (ValOp.isImm()) { |
2696 | int64_t V = ValOp.getImm(); |
2697 | unsigned W = getRegBitWidth(Reg: DefR.Reg); |
2698 | APInt A(W, V, true); |
2699 | const Constant *C = intToConst(Val: A); |
2700 | RC.add(LC: C); |
2701 | Outputs.update(R: DefR.Reg, L: RC); |
2702 | return true; |
2703 | } |
2704 | if (ValOp.isReg()) { |
2705 | RegisterSubReg R(ValOp); |
2706 | const LatticeCell &LR = Inputs.get(R: R.Reg); |
2707 | LatticeCell LSR; |
2708 | if (!evaluate(R, Input: LR, Result&: LSR)) |
2709 | return false; |
2710 | RC.meet(L: LSR); |
2711 | Outputs.update(R: DefR.Reg, L: RC); |
2712 | return true; |
2713 | } |
2714 | return false; |
2715 | } |
2716 | |
2717 | bool HexagonConstEvaluator::evaluateHexExt(const MachineInstr &MI, |
2718 | const CellMap &Inputs, CellMap &Outputs) { |
2719 | // Dst0 = ext R1 |
2720 | RegisterSubReg R1(MI.getOperand(i: 1)); |
2721 | assert(Inputs.has(R1.Reg)); |
2722 | |
2723 | unsigned Opc = MI.getOpcode(); |
2724 | unsigned Bits; |
2725 | switch (Opc) { |
2726 | case Hexagon::A2_sxtb: |
2727 | case Hexagon::A2_zxtb: |
2728 | Bits = 8; |
2729 | break; |
2730 | case Hexagon::A2_sxth: |
2731 | case Hexagon::A2_zxth: |
2732 | Bits = 16; |
2733 | break; |
2734 | case Hexagon::A2_sxtw: |
2735 | Bits = 32; |
2736 | break; |
2737 | default: |
2738 | llvm_unreachable("Unhandled extension opcode" ); |
2739 | } |
2740 | |
2741 | bool Signed = false; |
2742 | switch (Opc) { |
2743 | case Hexagon::A2_sxtb: |
2744 | case Hexagon::A2_sxth: |
2745 | case Hexagon::A2_sxtw: |
2746 | Signed = true; |
2747 | break; |
2748 | } |
2749 | |
2750 | RegisterSubReg DefR(MI.getOperand(i: 0)); |
2751 | unsigned BW = getRegBitWidth(Reg: DefR.Reg); |
2752 | LatticeCell RC = Outputs.get(R: DefR.Reg); |
2753 | bool Eval = Signed ? evaluateSEXTr(R1, Width: BW, Bits, Inputs, Result&: RC) |
2754 | : evaluateZEXTr(R1, Width: BW, Bits, Inputs, Result&: RC); |
2755 | if (!Eval) |
2756 | return false; |
2757 | Outputs.update(R: DefR.Reg, L: RC); |
2758 | return true; |
2759 | } |
2760 | |
2761 | bool HexagonConstEvaluator::evaluateHexVector1(const MachineInstr &MI, |
2762 | const CellMap &Inputs, CellMap &Outputs) { |
2763 | // DefR = op R1 |
2764 | RegisterSubReg DefR(MI.getOperand(i: 0)); |
2765 | RegisterSubReg R1(MI.getOperand(i: 1)); |
2766 | assert(Inputs.has(R1.Reg)); |
2767 | LatticeCell RC = Outputs.get(R: DefR.Reg); |
2768 | bool Eval; |
2769 | |
2770 | unsigned Opc = MI.getOpcode(); |
2771 | switch (Opc) { |
2772 | case Hexagon::S2_vsplatrb: |
2773 | // Rd = 4 times Rs:0..7 |
2774 | Eval = evaluateSplatr(R1, Bits: 8, Count: 4, Inputs, Result&: RC); |
2775 | break; |
2776 | case Hexagon::S2_vsplatrh: |
2777 | // Rdd = 4 times Rs:0..15 |
2778 | Eval = evaluateSplatr(R1, Bits: 16, Count: 4, Inputs, Result&: RC); |
2779 | break; |
2780 | default: |
2781 | return false; |
2782 | } |
2783 | |
2784 | if (!Eval) |
2785 | return false; |
2786 | Outputs.update(R: DefR.Reg, L: RC); |
2787 | return true; |
2788 | } |
2789 | |
2790 | bool HexagonConstEvaluator::rewriteHexConstDefs(MachineInstr &MI, |
2791 | const CellMap &Inputs, bool &AllDefs) { |
2792 | AllDefs = false; |
2793 | |
2794 | // Some diagnostics. |
2795 | // LLVM_DEBUG({...}) gets confused with all this code as an argument. |
2796 | #ifndef NDEBUG |
2797 | bool Debugging = DebugFlag && isCurrentDebugType(DEBUG_TYPE); |
2798 | if (Debugging) { |
2799 | bool Const = true, HasUse = false; |
2800 | for (const MachineOperand &MO : MI.operands()) { |
2801 | if (!MO.isReg() || !MO.isUse() || MO.isImplicit()) |
2802 | continue; |
2803 | RegisterSubReg R(MO); |
2804 | if (!R.Reg.isVirtual()) |
2805 | continue; |
2806 | HasUse = true; |
2807 | // PHIs can legitimately have "top" cells after propagation. |
2808 | if (!MI.isPHI() && !Inputs.has(R.Reg)) { |
2809 | dbgs() << "Top " << printReg(R.Reg, &HRI, R.SubReg) |
2810 | << " in MI: " << MI; |
2811 | continue; |
2812 | } |
2813 | const LatticeCell &L = Inputs.get(R.Reg); |
2814 | Const &= L.isSingle(); |
2815 | if (!Const) |
2816 | break; |
2817 | } |
2818 | if (HasUse && Const) { |
2819 | if (!MI.isCopy()) { |
2820 | dbgs() << "CONST: " << MI; |
2821 | for (const MachineOperand &MO : MI.operands()) { |
2822 | if (!MO.isReg() || !MO.isUse() || MO.isImplicit()) |
2823 | continue; |
2824 | Register R = MO.getReg(); |
2825 | dbgs() << printReg(R, &TRI) << ": " << Inputs.get(R) << "\n" ; |
2826 | } |
2827 | } |
2828 | } |
2829 | } |
2830 | #endif |
2831 | |
2832 | // Avoid generating TFRIs for register transfers---this will keep the |
2833 | // coalescing opportunities. |
2834 | if (MI.isCopy()) |
2835 | return false; |
2836 | |
2837 | MachineFunction *MF = MI.getParent()->getParent(); |
2838 | auto &HST = MF->getSubtarget<HexagonSubtarget>(); |
2839 | |
2840 | // Collect all virtual register-def operands. |
2841 | SmallVector<unsigned,2> DefRegs; |
2842 | for (const MachineOperand &MO : MI.operands()) { |
2843 | if (!MO.isReg() || !MO.isDef()) |
2844 | continue; |
2845 | Register R = MO.getReg(); |
2846 | if (!R.isVirtual()) |
2847 | continue; |
2848 | assert(!MO.getSubReg()); |
2849 | assert(Inputs.has(R)); |
2850 | DefRegs.push_back(Elt: R); |
2851 | } |
2852 | |
2853 | MachineBasicBlock &B = *MI.getParent(); |
2854 | const DebugLoc &DL = MI.getDebugLoc(); |
2855 | unsigned ChangedNum = 0; |
2856 | #ifndef NDEBUG |
2857 | SmallVector<const MachineInstr*,4> NewInstrs; |
2858 | #endif |
2859 | |
2860 | // For each defined register, if it is a constant, create an instruction |
2861 | // NewR = const |
2862 | // and replace all uses of the defined register with NewR. |
2863 | for (unsigned R : DefRegs) { |
2864 | const LatticeCell &L = Inputs.get(R); |
2865 | if (L.isBottom()) |
2866 | continue; |
2867 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: R); |
2868 | MachineBasicBlock::iterator At = MI.getIterator(); |
2869 | |
2870 | if (!L.isSingle()) { |
2871 | // If this a zero/non-zero cell, we can fold a definition |
2872 | // of a predicate register. |
2873 | using P = ConstantProperties; |
2874 | |
2875 | uint64_t Ps = L.properties(); |
2876 | if (!(Ps & (P::Zero|P::NonZero))) |
2877 | continue; |
2878 | const TargetRegisterClass *PredRC = &Hexagon::PredRegsRegClass; |
2879 | if (RC != PredRC) |
2880 | continue; |
2881 | const MCInstrDesc *NewD = (Ps & P::Zero) ? |
2882 | &HII.get(Opcode: Hexagon::PS_false) : |
2883 | &HII.get(Opcode: Hexagon::PS_true); |
2884 | Register NewR = MRI->createVirtualRegister(RegClass: PredRC); |
2885 | const MachineInstrBuilder &MIB = BuildMI(BB&: B, I: At, MIMD: DL, MCID: *NewD, DestReg: NewR); |
2886 | (void)MIB; |
2887 | #ifndef NDEBUG |
2888 | NewInstrs.push_back(&*MIB); |
2889 | #endif |
2890 | replaceAllRegUsesWith(FromReg: R, ToReg: NewR); |
2891 | } else { |
2892 | // This cell has a single value. |
2893 | APInt A; |
2894 | if (!constToInt(C: L.Value, Val&: A) || !A.isSignedIntN(N: 64)) |
2895 | continue; |
2896 | const TargetRegisterClass *NewRC; |
2897 | const MCInstrDesc *NewD; |
2898 | |
2899 | unsigned W = getRegBitWidth(Reg: R); |
2900 | int64_t V = A.getSExtValue(); |
2901 | assert(W == 32 || W == 64); |
2902 | if (W == 32) |
2903 | NewRC = &Hexagon::IntRegsRegClass; |
2904 | else |
2905 | NewRC = &Hexagon::DoubleRegsRegClass; |
2906 | Register NewR = MRI->createVirtualRegister(RegClass: NewRC); |
2907 | const MachineInstr *NewMI; |
2908 | |
2909 | if (W == 32) { |
2910 | NewD = &HII.get(Opcode: Hexagon::A2_tfrsi); |
2911 | NewMI = BuildMI(BB&: B, I: At, MIMD: DL, MCID: *NewD, DestReg: NewR) |
2912 | .addImm(Val: V); |
2913 | } else { |
2914 | if (A.isSignedIntN(N: 8)) { |
2915 | NewD = &HII.get(Opcode: Hexagon::A2_tfrpi); |
2916 | NewMI = BuildMI(BB&: B, I: At, MIMD: DL, MCID: *NewD, DestReg: NewR) |
2917 | .addImm(Val: V); |
2918 | } else { |
2919 | int32_t Hi = V >> 32; |
2920 | int32_t Lo = V & 0xFFFFFFFFLL; |
2921 | if (isInt<8>(x: Hi) && isInt<8>(x: Lo)) { |
2922 | NewD = &HII.get(Opcode: Hexagon::A2_combineii); |
2923 | NewMI = BuildMI(BB&: B, I: At, MIMD: DL, MCID: *NewD, DestReg: NewR) |
2924 | .addImm(Val: Hi) |
2925 | .addImm(Val: Lo); |
2926 | } else if (MF->getFunction().hasOptSize() || !HST.isTinyCore()) { |
2927 | // Disable CONST64 for tiny core since it takes a LD resource. |
2928 | NewD = &HII.get(Opcode: Hexagon::CONST64); |
2929 | NewMI = BuildMI(BB&: B, I: At, MIMD: DL, MCID: *NewD, DestReg: NewR) |
2930 | .addImm(Val: V); |
2931 | } else |
2932 | return false; |
2933 | } |
2934 | } |
2935 | (void)NewMI; |
2936 | #ifndef NDEBUG |
2937 | NewInstrs.push_back(NewMI); |
2938 | #endif |
2939 | replaceAllRegUsesWith(FromReg: R, ToReg: NewR); |
2940 | } |
2941 | ChangedNum++; |
2942 | } |
2943 | |
2944 | LLVM_DEBUG({ |
2945 | if (!NewInstrs.empty()) { |
2946 | MachineFunction &MF = *MI.getParent()->getParent(); |
2947 | dbgs() << "In function: " << MF.getName() << "\n" ; |
2948 | dbgs() << "Rewrite: for " << MI << " created " << *NewInstrs[0]; |
2949 | for (unsigned i = 1; i < NewInstrs.size(); ++i) |
2950 | dbgs() << " " << *NewInstrs[i]; |
2951 | } |
2952 | }); |
2953 | |
2954 | AllDefs = (ChangedNum == DefRegs.size()); |
2955 | return ChangedNum > 0; |
2956 | } |
2957 | |
2958 | bool HexagonConstEvaluator::rewriteHexConstUses(MachineInstr &MI, |
2959 | const CellMap &Inputs) { |
2960 | bool Changed = false; |
2961 | unsigned Opc = MI.getOpcode(); |
2962 | MachineBasicBlock &B = *MI.getParent(); |
2963 | const DebugLoc &DL = MI.getDebugLoc(); |
2964 | MachineBasicBlock::iterator At = MI.getIterator(); |
2965 | MachineInstr *NewMI = nullptr; |
2966 | |
2967 | switch (Opc) { |
2968 | case Hexagon::M2_maci: |
2969 | // Convert DefR += mpyi(R2, R3) |
2970 | // to DefR += mpyi(R, #imm), |
2971 | // or DefR -= mpyi(R, #imm). |
2972 | { |
2973 | RegisterSubReg DefR(MI.getOperand(i: 0)); |
2974 | assert(!DefR.SubReg); |
2975 | RegisterSubReg R2(MI.getOperand(i: 2)); |
2976 | RegisterSubReg R3(MI.getOperand(i: 3)); |
2977 | assert(Inputs.has(R2.Reg) && Inputs.has(R3.Reg)); |
2978 | LatticeCell LS2, LS3; |
2979 | // It is enough to get one of the input cells, since we will only try |
2980 | // to replace one argument---whichever happens to be a single constant. |
2981 | bool HasC2 = getCell(R: R2, Inputs, RC&: LS2), HasC3 = getCell(R: R3, Inputs, RC&: LS3); |
2982 | if (!HasC2 && !HasC3) |
2983 | return false; |
2984 | bool Zero = ((HasC2 && (LS2.properties() & ConstantProperties::Zero)) || |
2985 | (HasC3 && (LS3.properties() & ConstantProperties::Zero))); |
2986 | // If one of the operands is zero, eliminate the multiplication. |
2987 | if (Zero) { |
2988 | // DefR == R1 (tied operands). |
2989 | MachineOperand &Acc = MI.getOperand(i: 1); |
2990 | RegisterSubReg R1(Acc); |
2991 | unsigned NewR = R1.Reg; |
2992 | if (R1.SubReg) { |
2993 | // Generate COPY. FIXME: Replace with the register:subregister. |
2994 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: DefR.Reg); |
2995 | NewR = MRI->createVirtualRegister(RegClass: RC); |
2996 | NewMI = BuildMI(BB&: B, I: At, MIMD: DL, MCID: HII.get(Opcode: TargetOpcode::COPY), DestReg: NewR) |
2997 | .addReg(RegNo: R1.Reg, flags: getRegState(RegOp: Acc), SubReg: R1.SubReg); |
2998 | } |
2999 | replaceAllRegUsesWith(FromReg: DefR.Reg, ToReg: NewR); |
3000 | MRI->clearKillFlags(Reg: NewR); |
3001 | Changed = true; |
3002 | break; |
3003 | } |
3004 | |
3005 | bool Swap = false; |
3006 | if (!LS3.isSingle()) { |
3007 | if (!LS2.isSingle()) |
3008 | return false; |
3009 | Swap = true; |
3010 | } |
3011 | const LatticeCell &LI = Swap ? LS2 : LS3; |
3012 | const MachineOperand &OpR2 = Swap ? MI.getOperand(i: 3) |
3013 | : MI.getOperand(i: 2); |
3014 | // LI is single here. |
3015 | APInt A; |
3016 | if (!constToInt(C: LI.Value, Val&: A) || !A.isSignedIntN(N: 8)) |
3017 | return false; |
3018 | int64_t V = A.getSExtValue(); |
3019 | const MCInstrDesc &D = (V >= 0) ? HII.get(Opcode: Hexagon::M2_macsip) |
3020 | : HII.get(Opcode: Hexagon::M2_macsin); |
3021 | if (V < 0) |
3022 | V = -V; |
3023 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: DefR.Reg); |
3024 | Register NewR = MRI->createVirtualRegister(RegClass: RC); |
3025 | const MachineOperand &Src1 = MI.getOperand(i: 1); |
3026 | NewMI = BuildMI(BB&: B, I: At, MIMD: DL, MCID: D, DestReg: NewR) |
3027 | .addReg(RegNo: Src1.getReg(), flags: getRegState(RegOp: Src1), SubReg: Src1.getSubReg()) |
3028 | .addReg(RegNo: OpR2.getReg(), flags: getRegState(RegOp: OpR2), SubReg: OpR2.getSubReg()) |
3029 | .addImm(Val: V); |
3030 | replaceAllRegUsesWith(FromReg: DefR.Reg, ToReg: NewR); |
3031 | Changed = true; |
3032 | break; |
3033 | } |
3034 | |
3035 | case Hexagon::A2_and: |
3036 | { |
3037 | RegisterSubReg R1(MI.getOperand(i: 1)); |
3038 | RegisterSubReg R2(MI.getOperand(i: 2)); |
3039 | assert(Inputs.has(R1.Reg) && Inputs.has(R2.Reg)); |
3040 | LatticeCell LS1, LS2; |
3041 | unsigned CopyOf = 0; |
3042 | // Check if any of the operands is -1 (i.e. all bits set). |
3043 | if (getCell(R: R1, Inputs, RC&: LS1) && LS1.isSingle()) { |
3044 | APInt M1; |
3045 | if (constToInt(C: LS1.Value, Val&: M1) && !~M1) |
3046 | CopyOf = 2; |
3047 | } |
3048 | else if (getCell(R: R2, Inputs, RC&: LS2) && LS2.isSingle()) { |
3049 | APInt M1; |
3050 | if (constToInt(C: LS2.Value, Val&: M1) && !~M1) |
3051 | CopyOf = 1; |
3052 | } |
3053 | if (!CopyOf) |
3054 | return false; |
3055 | MachineOperand &SO = MI.getOperand(i: CopyOf); |
3056 | RegisterSubReg SR(SO); |
3057 | RegisterSubReg DefR(MI.getOperand(i: 0)); |
3058 | unsigned NewR = SR.Reg; |
3059 | if (SR.SubReg) { |
3060 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: DefR.Reg); |
3061 | NewR = MRI->createVirtualRegister(RegClass: RC); |
3062 | NewMI = BuildMI(BB&: B, I: At, MIMD: DL, MCID: HII.get(Opcode: TargetOpcode::COPY), DestReg: NewR) |
3063 | .addReg(RegNo: SR.Reg, flags: getRegState(RegOp: SO), SubReg: SR.SubReg); |
3064 | } |
3065 | replaceAllRegUsesWith(FromReg: DefR.Reg, ToReg: NewR); |
3066 | MRI->clearKillFlags(Reg: NewR); |
3067 | Changed = true; |
3068 | } |
3069 | break; |
3070 | |
3071 | case Hexagon::A2_or: |
3072 | { |
3073 | RegisterSubReg R1(MI.getOperand(i: 1)); |
3074 | RegisterSubReg R2(MI.getOperand(i: 2)); |
3075 | assert(Inputs.has(R1.Reg) && Inputs.has(R2.Reg)); |
3076 | LatticeCell LS1, LS2; |
3077 | unsigned CopyOf = 0; |
3078 | |
3079 | using P = ConstantProperties; |
3080 | |
3081 | if (getCell(R: R1, Inputs, RC&: LS1) && (LS1.properties() & P::Zero)) |
3082 | CopyOf = 2; |
3083 | else if (getCell(R: R2, Inputs, RC&: LS2) && (LS2.properties() & P::Zero)) |
3084 | CopyOf = 1; |
3085 | if (!CopyOf) |
3086 | return false; |
3087 | MachineOperand &SO = MI.getOperand(i: CopyOf); |
3088 | RegisterSubReg SR(SO); |
3089 | RegisterSubReg DefR(MI.getOperand(i: 0)); |
3090 | unsigned NewR = SR.Reg; |
3091 | if (SR.SubReg) { |
3092 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: DefR.Reg); |
3093 | NewR = MRI->createVirtualRegister(RegClass: RC); |
3094 | NewMI = BuildMI(BB&: B, I: At, MIMD: DL, MCID: HII.get(Opcode: TargetOpcode::COPY), DestReg: NewR) |
3095 | .addReg(RegNo: SR.Reg, flags: getRegState(RegOp: SO), SubReg: SR.SubReg); |
3096 | } |
3097 | replaceAllRegUsesWith(FromReg: DefR.Reg, ToReg: NewR); |
3098 | MRI->clearKillFlags(Reg: NewR); |
3099 | Changed = true; |
3100 | } |
3101 | break; |
3102 | } |
3103 | |
3104 | if (NewMI) { |
3105 | // clear all the kill flags of this new instruction. |
3106 | for (MachineOperand &MO : NewMI->operands()) |
3107 | if (MO.isReg() && MO.isUse()) |
3108 | MO.setIsKill(false); |
3109 | } |
3110 | |
3111 | LLVM_DEBUG({ |
3112 | if (NewMI) { |
3113 | dbgs() << "Rewrite: for " << MI; |
3114 | if (NewMI != &MI) |
3115 | dbgs() << " created " << *NewMI; |
3116 | else |
3117 | dbgs() << " modified the instruction itself and created:" << *NewMI; |
3118 | } |
3119 | }); |
3120 | |
3121 | return Changed; |
3122 | } |
3123 | |
3124 | void HexagonConstEvaluator::replaceAllRegUsesWith(Register FromReg, |
3125 | Register ToReg) { |
3126 | assert(FromReg.isVirtual()); |
3127 | assert(ToReg.isVirtual()); |
3128 | for (MachineOperand &O : |
3129 | llvm::make_early_inc_range(Range: MRI->use_operands(Reg: FromReg))) |
3130 | O.setReg(ToReg); |
3131 | } |
3132 | |
3133 | bool HexagonConstEvaluator::rewriteHexBranch(MachineInstr &BrI, |
3134 | const CellMap &Inputs) { |
3135 | MachineBasicBlock &B = *BrI.getParent(); |
3136 | unsigned NumOp = BrI.getNumOperands(); |
3137 | if (!NumOp) |
3138 | return false; |
3139 | |
3140 | bool FallsThru; |
3141 | SetVector<const MachineBasicBlock*> Targets; |
3142 | bool Eval = evaluate(BrI, Inputs, Targets, FallsThru); |
3143 | unsigned NumTargets = Targets.size(); |
3144 | if (!Eval || NumTargets > 1 || (NumTargets == 1 && FallsThru)) |
3145 | return false; |
3146 | if (BrI.getOpcode() == Hexagon::J2_jump) |
3147 | return false; |
3148 | |
3149 | LLVM_DEBUG(dbgs() << "Rewrite(" << printMBBReference(B) << "):" << BrI); |
3150 | bool Rewritten = false; |
3151 | if (NumTargets > 0) { |
3152 | assert(!FallsThru && "This should have been checked before" ); |
3153 | // MIB.addMBB needs non-const pointer. |
3154 | MachineBasicBlock *TargetB = const_cast<MachineBasicBlock*>(Targets[0]); |
3155 | bool Moot = B.isLayoutSuccessor(MBB: TargetB); |
3156 | if (!Moot) { |
3157 | // If we build a branch here, we must make sure that it won't be |
3158 | // erased as "non-executable". We can't mark any new instructions |
3159 | // as executable here, so we need to overwrite the BrI, which we |
3160 | // know is executable. |
3161 | const MCInstrDesc &JD = HII.get(Opcode: Hexagon::J2_jump); |
3162 | auto NI = BuildMI(BB&: B, I: BrI.getIterator(), MIMD: BrI.getDebugLoc(), MCID: JD) |
3163 | .addMBB(MBB: TargetB); |
3164 | BrI.setDesc(JD); |
3165 | while (BrI.getNumOperands() > 0) |
3166 | BrI.removeOperand(OpNo: 0); |
3167 | // This ensures that all implicit operands (e.g. implicit-def %r31, etc) |
3168 | // are present in the rewritten branch. |
3169 | for (auto &Op : NI->operands()) |
3170 | BrI.addOperand(Op); |
3171 | NI->eraseFromParent(); |
3172 | Rewritten = true; |
3173 | } |
3174 | } |
3175 | |
3176 | // Do not erase instructions. A newly created instruction could get |
3177 | // the same address as an instruction marked as executable during the |
3178 | // propagation. |
3179 | if (!Rewritten) |
3180 | replaceWithNop(MI&: BrI); |
3181 | return true; |
3182 | } |
3183 | |
3184 | FunctionPass *llvm::createHexagonConstPropagationPass() { |
3185 | return new HexagonConstPropagation(); |
3186 | } |
3187 | |