LLVM  10.0.0svn
FunctionComparator.h
Go to the documentation of this file.
1 //===- FunctionComparator.h - Function Comparator ---------------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file defines the FunctionComparator and GlobalNumberState classes which
10 // are used by the MergeFunctions pass for comparing functions.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef LLVM_TRANSFORMS_UTILS_FUNCTIONCOMPARATOR_H
15 #define LLVM_TRANSFORMS_UTILS_FUNCTIONCOMPARATOR_H
16 
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/StringRef.h"
19 #include "llvm/IR/Attributes.h"
20 #include "llvm/IR/Instructions.h"
21 #include "llvm/IR/Operator.h"
22 #include "llvm/IR/ValueMap.h"
24 #include "llvm/Support/Casting.h"
25 #include <cstdint>
26 #include <tuple>
27 
28 namespace llvm {
29 
30 class APFloat;
31 class APInt;
32 class BasicBlock;
33 class Constant;
34 class Function;
35 class GlobalValue;
36 class InlineAsm;
37 class Instruction;
38 class MDNode;
39 class Type;
40 class Value;
41 
42 /// GlobalNumberState assigns an integer to each global value in the program,
43 /// which is used by the comparison routine to order references to globals. This
44 /// state must be preserved throughout the pass, because Functions and other
45 /// globals need to maintain their relative order. Globals are assigned a number
46 /// when they are first visited. This order is deterministic, and so the
47 /// assigned numbers are as well. When two functions are merged, neither number
48 /// is updated. If the symbols are weak, this would be incorrect. If they are
49 /// strong, then one will be replaced at all references to the other, and so
50 /// direct callsites will now see one or the other symbol, and no update is
51 /// necessary. Note that if we were guaranteed unique names, we could just
52 /// compare those, but this would not work for stripped bitcodes or for those
53 /// few symbols without a name.
55  struct Config : ValueMapConfig<GlobalValue *> {
56  enum { FollowRAUW = false };
57  };
58 
59  // Each GlobalValue is mapped to an identifier. The Config ensures when RAUW
60  // occurs, the mapping does not change. Tracking changes is unnecessary, and
61  // also problematic for weak symbols (which may be overwritten).
63  ValueNumberMap GlobalNumbers;
64 
65  // The next unused serial number to assign to a global.
66  uint64_t NextNumber = 0;
67 
68 public:
69  GlobalNumberState() = default;
70 
71  uint64_t getNumber(GlobalValue* Global) {
73  bool Inserted;
74  std::tie(MapIter, Inserted) = GlobalNumbers.insert({Global, NextNumber});
75  if (Inserted)
76  NextNumber++;
77  return MapIter->second;
78  }
79 
80  void erase(GlobalValue *Global) {
81  GlobalNumbers.erase(Global);
82  }
83 
84  void clear() {
85  GlobalNumbers.clear();
86  }
87 };
88 
89 /// FunctionComparator - Compares two functions to determine whether or not
90 /// they will generate machine code with the same behaviour. DataLayout is
91 /// used if available. The comparator always fails conservatively (erring on the
92 /// side of claiming that two functions are different).
94 public:
95  FunctionComparator(const Function *F1, const Function *F2,
97  : FnL(F1), FnR(F2), GlobalNumbers(GN) {}
98 
99  /// Test whether the two functions have equivalent behaviour.
100  int compare();
101 
102  /// Hash a function. Equivalent functions will have the same hash, and unequal
103  /// functions will have different hashes with high probability.
104  using FunctionHash = uint64_t;
105  static FunctionHash functionHash(Function &);
106 
107 protected:
108  /// Start the comparison.
109  void beginCompare() {
110  sn_mapL.clear();
111  sn_mapR.clear();
112  }
113 
114  /// Compares the signature and other general attributes of the two functions.
115  int compareSignature() const;
116 
117  /// Test whether two basic blocks have equivalent behaviour.
118  int cmpBasicBlocks(const BasicBlock *BBL, const BasicBlock *BBR) const;
119 
120  /// Constants comparison.
121  /// Its analog to lexicographical comparison between hypothetical numbers
122  /// of next format:
123  /// <bitcastability-trait><raw-bit-contents>
124  ///
125  /// 1. Bitcastability.
126  /// Check whether L's type could be losslessly bitcasted to R's type.
127  /// On this stage method, in case when lossless bitcast is not possible
128  /// method returns -1 or 1, thus also defining which type is greater in
129  /// context of bitcastability.
130  /// Stage 0: If types are equal in terms of cmpTypes, then we can go straight
131  /// to the contents comparison.
132  /// If types differ, remember types comparison result and check
133  /// whether we still can bitcast types.
134  /// Stage 1: Types that satisfies isFirstClassType conditions are always
135  /// greater then others.
136  /// Stage 2: Vector is greater then non-vector.
137  /// If both types are vectors, then vector with greater bitwidth is
138  /// greater.
139  /// If both types are vectors with the same bitwidth, then types
140  /// are bitcastable, and we can skip other stages, and go to contents
141  /// comparison.
142  /// Stage 3: Pointer types are greater than non-pointers. If both types are
143  /// pointers of the same address space - go to contents comparison.
144  /// Different address spaces: pointer with greater address space is
145  /// greater.
146  /// Stage 4: Types are neither vectors, nor pointers. And they differ.
147  /// We don't know how to bitcast them. So, we better don't do it,
148  /// and return types comparison result (so it determines the
149  /// relationship among constants we don't know how to bitcast).
150  ///
151  /// Just for clearance, let's see how the set of constants could look
152  /// on single dimension axis:
153  ///
154  /// [NFCT], [FCT, "others"], [FCT, pointers], [FCT, vectors]
155  /// Where: NFCT - Not a FirstClassType
156  /// FCT - FirstClassTyp:
157  ///
158  /// 2. Compare raw contents.
159  /// It ignores types on this stage and only compares bits from L and R.
160  /// Returns 0, if L and R has equivalent contents.
161  /// -1 or 1 if values are different.
162  /// Pretty trivial:
163  /// 2.1. If contents are numbers, compare numbers.
164  /// Ints with greater bitwidth are greater. Ints with same bitwidths
165  /// compared by their contents.
166  /// 2.2. "And so on". Just to avoid discrepancies with comments
167  /// perhaps it would be better to read the implementation itself.
168  /// 3. And again about overall picture. Let's look back at how the ordered set
169  /// of constants will look like:
170  /// [NFCT], [FCT, "others"], [FCT, pointers], [FCT, vectors]
171  ///
172  /// Now look, what could be inside [FCT, "others"], for example:
173  /// [FCT, "others"] =
174  /// [
175  /// [double 0.1], [double 1.23],
176  /// [i32 1], [i32 2],
177  /// { double 1.0 }, ; StructTyID, NumElements = 1
178  /// { i32 1 }, ; StructTyID, NumElements = 1
179  /// { double 1, i32 1 }, ; StructTyID, NumElements = 2
180  /// { i32 1, double 1 } ; StructTyID, NumElements = 2
181  /// ]
182  ///
183  /// Let's explain the order. Float numbers will be less than integers, just
184  /// because of cmpType terms: FloatTyID < IntegerTyID.
185  /// Floats (with same fltSemantics) are sorted according to their value.
186  /// Then you can see integers, and they are, like a floats,
187  /// could be easy sorted among each others.
188  /// The structures. Structures are grouped at the tail, again because of their
189  /// TypeID: StructTyID > IntegerTyID > FloatTyID.
190  /// Structures with greater number of elements are greater. Structures with
191  /// greater elements going first are greater.
192  /// The same logic with vectors, arrays and other possible complex types.
193  ///
194  /// Bitcastable constants.
195  /// Let's assume, that some constant, belongs to some group of
196  /// "so-called-equal" values with different types, and at the same time
197  /// belongs to another group of constants with equal types
198  /// and "really" equal values.
199  ///
200  /// Now, prove that this is impossible:
201  ///
202  /// If constant A with type TyA is bitcastable to B with type TyB, then:
203  /// 1. All constants with equal types to TyA, are bitcastable to B. Since
204  /// those should be vectors (if TyA is vector), pointers
205  /// (if TyA is pointer), or else (if TyA equal to TyB), those types should
206  /// be equal to TyB.
207  /// 2. All constants with non-equal, but bitcastable types to TyA, are
208  /// bitcastable to B.
209  /// Once again, just because we allow it to vectors and pointers only.
210  /// This statement could be expanded as below:
211  /// 2.1. All vectors with equal bitwidth to vector A, has equal bitwidth to
212  /// vector B, and thus bitcastable to B as well.
213  /// 2.2. All pointers of the same address space, no matter what they point to,
214  /// bitcastable. So if C is pointer, it could be bitcasted to A and to B.
215  /// So any constant equal or bitcastable to A is equal or bitcastable to B.
216  /// QED.
217  ///
218  /// In another words, for pointers and vectors, we ignore top-level type and
219  /// look at their particular properties (bit-width for vectors, and
220  /// address space for pointers).
221  /// If these properties are equal - compare their contents.
222  int cmpConstants(const Constant *L, const Constant *R) const;
223 
224  /// Compares two global values by number. Uses the GlobalNumbersState to
225  /// identify the same gobals across function calls.
226  int cmpGlobalValues(GlobalValue *L, GlobalValue *R) const;
227 
228  /// Assign or look up previously assigned numbers for the two values, and
229  /// return whether the numbers are equal. Numbers are assigned in the order
230  /// visited.
231  /// Comparison order:
232  /// Stage 0: Value that is function itself is always greater then others.
233  /// If left and right values are references to their functions, then
234  /// they are equal.
235  /// Stage 1: Constants are greater than non-constants.
236  /// If both left and right are constants, then the result of
237  /// cmpConstants is used as cmpValues result.
238  /// Stage 2: InlineAsm instances are greater than others. If both left and
239  /// right are InlineAsm instances, InlineAsm* pointers casted to
240  /// integers and compared as numbers.
241  /// Stage 3: For all other cases we compare order we meet these values in
242  /// their functions. If right value was met first during scanning,
243  /// then left value is greater.
244  /// In another words, we compare serial numbers, for more details
245  /// see comments for sn_mapL and sn_mapR.
246  int cmpValues(const Value *L, const Value *R) const;
247 
248  /// Compare two Instructions for equivalence, similar to
249  /// Instruction::isSameOperationAs.
250  ///
251  /// Stages are listed in "most significant stage first" order:
252  /// On each stage below, we do comparison between some left and right
253  /// operation parts. If parts are non-equal, we assign parts comparison
254  /// result to the operation comparison result and exit from method.
255  /// Otherwise we proceed to the next stage.
256  /// Stages:
257  /// 1. Operations opcodes. Compared as numbers.
258  /// 2. Number of operands.
259  /// 3. Operation types. Compared with cmpType method.
260  /// 4. Compare operation subclass optional data as stream of bytes:
261  /// just convert it to integers and call cmpNumbers.
262  /// 5. Compare in operation operand types with cmpType in
263  /// most significant operand first order.
264  /// 6. Last stage. Check operations for some specific attributes.
265  /// For example, for Load it would be:
266  /// 6.1.Load: volatile (as boolean flag)
267  /// 6.2.Load: alignment (as integer numbers)
268  /// 6.3.Load: ordering (as underlying enum class value)
269  /// 6.4.Load: synch-scope (as integer numbers)
270  /// 6.5.Load: range metadata (as integer ranges)
271  /// On this stage its better to see the code, since its not more than 10-15
272  /// strings for particular instruction, and could change sometimes.
273  ///
274  /// Sets \p needToCmpOperands to true if the operands of the instructions
275  /// still must be compared afterwards. In this case it's already guaranteed
276  /// that both instructions have the same number of operands.
277  int cmpOperations(const Instruction *L, const Instruction *R,
278  bool &needToCmpOperands) const;
279 
280  /// cmpType - compares two types,
281  /// defines total ordering among the types set.
282  ///
283  /// Return values:
284  /// 0 if types are equal,
285  /// -1 if Left is less than Right,
286  /// +1 if Left is greater than Right.
287  ///
288  /// Description:
289  /// Comparison is broken onto stages. Like in lexicographical comparison
290  /// stage coming first has higher priority.
291  /// On each explanation stage keep in mind total ordering properties.
292  ///
293  /// 0. Before comparison we coerce pointer types of 0 address space to
294  /// integer.
295  /// We also don't bother with same type at left and right, so
296  /// just return 0 in this case.
297  ///
298  /// 1. If types are of different kind (different type IDs).
299  /// Return result of type IDs comparison, treating them as numbers.
300  /// 2. If types are integers, check that they have the same width. If they
301  /// are vectors, check that they have the same count and subtype.
302  /// 3. Types have the same ID, so check whether they are one of:
303  /// * Void
304  /// * Float
305  /// * Double
306  /// * X86_FP80
307  /// * FP128
308  /// * PPC_FP128
309  /// * Label
310  /// * Metadata
311  /// We can treat these types as equal whenever their IDs are same.
312  /// 4. If Left and Right are pointers, return result of address space
313  /// comparison (numbers comparison). We can treat pointer types of same
314  /// address space as equal.
315  /// 5. If types are complex.
316  /// Then both Left and Right are to be expanded and their element types will
317  /// be checked with the same way. If we get Res != 0 on some stage, return it.
318  /// Otherwise return 0.
319  /// 6. For all other cases put llvm_unreachable.
320  int cmpTypes(Type *TyL, Type *TyR) const;
321 
322  int cmpNumbers(uint64_t L, uint64_t R) const;
323  int cmpAPInts(const APInt &L, const APInt &R) const;
324  int cmpAPFloats(const APFloat &L, const APFloat &R) const;
325  int cmpMem(StringRef L, StringRef R) const;
326 
327  // The two functions undergoing comparison.
328  const Function *FnL, *FnR;
329 
330 private:
331  int cmpOrderings(AtomicOrdering L, AtomicOrdering R) const;
332  int cmpInlineAsm(const InlineAsm *L, const InlineAsm *R) const;
333  int cmpAttrs(const AttributeList L, const AttributeList R) const;
334  int cmpRangeMetadata(const MDNode *L, const MDNode *R) const;
335  int cmpOperandBundlesSchema(const Instruction *L, const Instruction *R) const;
336 
337  /// Compare two GEPs for equivalent pointer arithmetic.
338  /// Parts to be compared for each comparison stage,
339  /// most significant stage first:
340  /// 1. Address space. As numbers.
341  /// 2. Constant offset, (using GEPOperator::accumulateConstantOffset method).
342  /// 3. Pointer operand type (using cmpType method).
343  /// 4. Number of operands.
344  /// 5. Compare operands, using cmpValues method.
345  int cmpGEPs(const GEPOperator *GEPL, const GEPOperator *GEPR) const;
346  int cmpGEPs(const GetElementPtrInst *GEPL,
347  const GetElementPtrInst *GEPR) const {
348  return cmpGEPs(cast<GEPOperator>(GEPL), cast<GEPOperator>(GEPR));
349  }
350 
351  /// Assign serial numbers to values from left function, and values from
352  /// right function.
353  /// Explanation:
354  /// Being comparing functions we need to compare values we meet at left and
355  /// right sides.
356  /// Its easy to sort things out for external values. It just should be
357  /// the same value at left and right.
358  /// But for local values (those were introduced inside function body)
359  /// we have to ensure they were introduced at exactly the same place,
360  /// and plays the same role.
361  /// Let's assign serial number to each value when we meet it first time.
362  /// Values that were met at same place will be with same serial numbers.
363  /// In this case it would be good to explain few points about values assigned
364  /// to BBs and other ways of implementation (see below).
365  ///
366  /// 1. Safety of BB reordering.
367  /// It's safe to change the order of BasicBlocks in function.
368  /// Relationship with other functions and serial numbering will not be
369  /// changed in this case.
370  /// As follows from FunctionComparator::compare(), we do CFG walk: we start
371  /// from the entry, and then take each terminator. So it doesn't matter how in
372  /// fact BBs are ordered in function. And since cmpValues are called during
373  /// this walk, the numbering depends only on how BBs located inside the CFG.
374  /// So the answer is - yes. We will get the same numbering.
375  ///
376  /// 2. Impossibility to use dominance properties of values.
377  /// If we compare two instruction operands: first is usage of local
378  /// variable AL from function FL, and second is usage of local variable AR
379  /// from FR, we could compare their origins and check whether they are
380  /// defined at the same place.
381  /// But, we are still not able to compare operands of PHI nodes, since those
382  /// could be operands from further BBs we didn't scan yet.
383  /// So it's impossible to use dominance properties in general.
384  mutable DenseMap<const Value*, int> sn_mapL, sn_mapR;
385 
386  // The global state we will use
387  GlobalNumberState* GlobalNumbers;
388 };
389 
390 } // end namespace llvm
391 
392 #endif // LLVM_TRANSFORMS_UTILS_FUNCTIONCOMPARATOR_H
void clear()
Definition: ValueMap.h:151
Type
MessagePack types as defined in the standard, with the exception of Integer being divided into a sign...
Definition: MsgPackReader.h:48
Atomic ordering constants.
This class represents lattice values for constants.
Definition: AllocatorList.h:23
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: ValueMap.h:178
Various leaf nodes.
Definition: ISDOpcodes.h:59
This class defines the default behavior for configurable aspects of ValueMap<>.
Definition: ValueMap.h:57
Metadata node.
Definition: Metadata.h:863
uint64_t FunctionHash
Hash a function.
FunctionComparator(const Function *F1, const Function *F2, GlobalNumberState *GN)
This file contains the simple types necessary to represent the attributes associated with functions a...
AtomicOrdering
Atomic ordering for LLVM&#39;s memory model.
void beginCompare()
Start the comparison.
uint64_t getNumber(GlobalValue *Global)
void erase(GlobalValue *Global)
an instruction for type-safe pointer arithmetic to access elements of arrays and structs ...
Definition: Instructions.h:875
FunctionComparator - Compares two functions to determine whether or not they will generate machine co...
LLVM Basic Block Representation.
Definition: BasicBlock.h:57
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:45
This is an important base class in LLVM.
Definition: Constant.h:41
GlobalNumberState assigns an integer to each global value in the program, which is used by the compar...
Class for arbitrary precision integers.
Definition: APInt.h:69
int compare(DigitsT LDigits, int16_t LScale, DigitsT RDigits, int16_t RScale)
Compare two scaled numbers.
Definition: ScaledNumber.h:251
LLVM Value Representation.
Definition: Value.h:73
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:48
bool erase(const KeyT &Val)
Definition: ValueMap.h:196