1 //===- FunctionComparator.h - Function Comparator ---------------*- C++ -*-===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines the FunctionComparator and GlobalNumberState classes which
11 // are used by the MergeFunctions pass for comparing functions.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #ifndef LLVM_TRANSFORMS_UTILS_FUNCTIONCOMPARATOR_H
16 #define LLVM_TRANSFORMS_UTILS_FUNCTIONCOMPARATOR_H
17 
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/StringRef.h"
20 #include "llvm/IR/Attributes.h"
21 #include "llvm/IR/Instructions.h"
22 #include "llvm/IR/Operator.h"
23 #include "llvm/IR/ValueMap.h"
24 #include "llvm/Support/AtomicOrdering.h"
25 #include "llvm/Support/Casting.h"
26 #include <cstdint>
27 #include <tuple>
28 
29 namespace llvm {
30 
31 class APFloat;
32 class APInt;
33 class BasicBlock;
34 class Constant;
35 class Function;
36 class GlobalValue;
37 class InlineAsm;
38 class Instruction;
39 class MDNode;
40 class Type;
41 class Value;
42 
43 /// GlobalNumberState assigns an integer to each global value in the program,
44 /// which is used by the comparison routine to order references to globals. This
45 /// state must be preserved throughout the pass, because Functions and other
46 /// globals need to maintain their relative order. Globals are assigned a number
47 /// when they are first visited. This order is deterministic, and so the
48 /// assigned numbers are as well. When two functions are merged, neither number
49 /// is updated. If the symbols are weak, this would be incorrect. If they are
50 /// strong, then one will be replaced at all references to the other, and so
51 /// direct callsites will now see one or the other symbol, and no update is
52 /// necessary. Note that if we were guaranteed unique names, we could just
53 /// compare those, but this would not work for stripped bitcodes or for those
54 /// few symbols without a name.
55 class GlobalNumberState {
56   struct Config : ValueMapConfig<GlobalValue *> {
57     enum { FollowRAUW = false };
58   };
59 
60   // Each GlobalValue is mapped to an identifier. The Config ensures when RAUW
61   // occurs, the mapping does not change. Tracking changes is unnecessary, and
62   // also problematic for weak symbols (which may be overwritten).
63   using ValueNumberMap = ValueMap<GlobalValue *, uint64_t, Config>;
64   ValueNumberMap GlobalNumbers;
65 
66   // The next unused serial number to assign to a global.
67   uint64_t NextNumber = 0;
68 
69 public:
70   GlobalNumberState() = default;
71 
getNumber(GlobalValue * Global)72   uint64_t getNumber(GlobalValue* Global) {
73     ValueNumberMap::iterator MapIter;
74     bool Inserted;
75     std::tie(MapIter, Inserted) = GlobalNumbers.insert({Global, NextNumber});
76     if (Inserted)
77       NextNumber++;
78     return MapIter->second;
79   }
80 
erase(GlobalValue * Global)81   void erase(GlobalValue *Global) {
82     GlobalNumbers.erase(Global);
83   }
84 
clear()85   void clear() {
86     GlobalNumbers.clear();
87   }
88 };
89 
90 /// FunctionComparator - Compares two functions to determine whether or not
91 /// they will generate machine code with the same behaviour. DataLayout is
92 /// used if available. The comparator always fails conservatively (erring on the
93 /// side of claiming that two functions are different).
94 class FunctionComparator {
95 public:
FunctionComparator(const Function * F1,const Function * F2,GlobalNumberState * GN)96   FunctionComparator(const Function *F1, const Function *F2,
97                      GlobalNumberState* GN)
98       : FnL(F1), FnR(F2), GlobalNumbers(GN) {}
99 
100   /// Test whether the two functions have equivalent behaviour.
101   int compare();
102 
103   /// Hash a function. Equivalent functions will have the same hash, and unequal
104   /// functions will have different hashes with high probability.
105   using FunctionHash = uint64_t;
106   static FunctionHash functionHash(Function &);
107 
108 protected:
109   /// Start the comparison.
beginCompare()110   void beginCompare() {
111     sn_mapL.clear();
112     sn_mapR.clear();
113   }
114 
115   /// Compares the signature and other general attributes of the two functions.
116   int compareSignature() const;
117 
118   /// Test whether two basic blocks have equivalent behaviour.
119   int cmpBasicBlocks(const BasicBlock *BBL, const BasicBlock *BBR) const;
120 
121   /// Constants comparison.
122   /// Its analog to lexicographical comparison between hypothetical numbers
123   /// of next format:
124   /// <bitcastability-trait><raw-bit-contents>
125   ///
126   /// 1. Bitcastability.
127   /// Check whether L's type could be losslessly bitcasted to R's type.
128   /// On this stage method, in case when lossless bitcast is not possible
129   /// method returns -1 or 1, thus also defining which type is greater in
130   /// context of bitcastability.
131   /// Stage 0: If types are equal in terms of cmpTypes, then we can go straight
132   ///          to the contents comparison.
133   ///          If types differ, remember types comparison result and check
134   ///          whether we still can bitcast types.
135   /// Stage 1: Types that satisfies isFirstClassType conditions are always
136   ///          greater then others.
137   /// Stage 2: Vector is greater then non-vector.
138   ///          If both types are vectors, then vector with greater bitwidth is
139   ///          greater.
140   ///          If both types are vectors with the same bitwidth, then types
141   ///          are bitcastable, and we can skip other stages, and go to contents
142   ///          comparison.
143   /// Stage 3: Pointer types are greater than non-pointers. If both types are
144   ///          pointers of the same address space - go to contents comparison.
145   ///          Different address spaces: pointer with greater address space is
146   ///          greater.
147   /// Stage 4: Types are neither vectors, nor pointers. And they differ.
148   ///          We don't know how to bitcast them. So, we better don't do it,
149   ///          and return types comparison result (so it determines the
150   ///          relationship among constants we don't know how to bitcast).
151   ///
152   /// Just for clearance, let's see how the set of constants could look
153   /// on single dimension axis:
154   ///
155   /// [NFCT], [FCT, "others"], [FCT, pointers], [FCT, vectors]
156   /// Where: NFCT - Not a FirstClassType
157   ///        FCT - FirstClassTyp:
158   ///
159   /// 2. Compare raw contents.
160   /// It ignores types on this stage and only compares bits from L and R.
161   /// Returns 0, if L and R has equivalent contents.
162   /// -1 or 1 if values are different.
163   /// Pretty trivial:
164   /// 2.1. If contents are numbers, compare numbers.
165   ///    Ints with greater bitwidth are greater. Ints with same bitwidths
166   ///    compared by their contents.
167   /// 2.2. "And so on". Just to avoid discrepancies with comments
168   /// perhaps it would be better to read the implementation itself.
169   /// 3. And again about overall picture. Let's look back at how the ordered set
170   /// of constants will look like:
171   /// [NFCT], [FCT, "others"], [FCT, pointers], [FCT, vectors]
172   ///
173   /// Now look, what could be inside [FCT, "others"], for example:
174   /// [FCT, "others"] =
175   /// [
176   ///   [double 0.1], [double 1.23],
177   ///   [i32 1], [i32 2],
178   ///   { double 1.0 },       ; StructTyID, NumElements = 1
179   ///   { i32 1 },            ; StructTyID, NumElements = 1
180   ///   { double 1, i32 1 },  ; StructTyID, NumElements = 2
181   ///   { i32 1, double 1 }   ; StructTyID, NumElements = 2
182   /// ]
183   ///
184   /// Let's explain the order. Float numbers will be less than integers, just
185   /// because of cmpType terms: FloatTyID < IntegerTyID.
186   /// Floats (with same fltSemantics) are sorted according to their value.
187   /// Then you can see integers, and they are, like a floats,
188   /// could be easy sorted among each others.
189   /// The structures. Structures are grouped at the tail, again because of their
190   /// TypeID: StructTyID > IntegerTyID > FloatTyID.
191   /// Structures with greater number of elements are greater. Structures with
192   /// greater elements going first are greater.
193   /// The same logic with vectors, arrays and other possible complex types.
194   ///
195   /// Bitcastable constants.
196   /// Let's assume, that some constant, belongs to some group of
197   /// "so-called-equal" values with different types, and at the same time
198   /// belongs to another group of constants with equal types
199   /// and "really" equal values.
200   ///
201   /// Now, prove that this is impossible:
202   ///
203   /// If constant A with type TyA is bitcastable to B with type TyB, then:
204   /// 1. All constants with equal types to TyA, are bitcastable to B. Since
205   ///    those should be vectors (if TyA is vector), pointers
206   ///    (if TyA is pointer), or else (if TyA equal to TyB), those types should
207   ///    be equal to TyB.
208   /// 2. All constants with non-equal, but bitcastable types to TyA, are
209   ///    bitcastable to B.
210   ///    Once again, just because we allow it to vectors and pointers only.
211   ///    This statement could be expanded as below:
212   /// 2.1. All vectors with equal bitwidth to vector A, has equal bitwidth to
213   ///      vector B, and thus bitcastable to B as well.
214   /// 2.2. All pointers of the same address space, no matter what they point to,
215   ///      bitcastable. So if C is pointer, it could be bitcasted to A and to B.
216   /// So any constant equal or bitcastable to A is equal or bitcastable to B.
217   /// QED.
218   ///
219   /// In another words, for pointers and vectors, we ignore top-level type and
220   /// look at their particular properties (bit-width for vectors, and
221   /// address space for pointers).
222   /// If these properties are equal - compare their contents.
223   int cmpConstants(const Constant *L, const Constant *R) const;
224 
225   /// Compares two global values by number. Uses the GlobalNumbersState to
226   /// identify the same gobals across function calls.
227   int cmpGlobalValues(GlobalValue *L, GlobalValue *R) const;
228 
229   /// Assign or look up previously assigned numbers for the two values, and
230   /// return whether the numbers are equal. Numbers are assigned in the order
231   /// visited.
232   /// Comparison order:
233   /// Stage 0: Value that is function itself is always greater then others.
234   ///          If left and right values are references to their functions, then
235   ///          they are equal.
236   /// Stage 1: Constants are greater than non-constants.
237   ///          If both left and right are constants, then the result of
238   ///          cmpConstants is used as cmpValues result.
239   /// Stage 2: InlineAsm instances are greater than others. If both left and
240   ///          right are InlineAsm instances, InlineAsm* pointers casted to
241   ///          integers and compared as numbers.
242   /// Stage 3: For all other cases we compare order we meet these values in
243   ///          their functions. If right value was met first during scanning,
244   ///          then left value is greater.
245   ///          In another words, we compare serial numbers, for more details
246   ///          see comments for sn_mapL and sn_mapR.
247   int cmpValues(const Value *L, const Value *R) const;
248 
249   /// Compare two Instructions for equivalence, similar to
250   /// Instruction::isSameOperationAs.
251   ///
252   /// Stages are listed in "most significant stage first" order:
253   /// On each stage below, we do comparison between some left and right
254   /// operation parts. If parts are non-equal, we assign parts comparison
255   /// result to the operation comparison result and exit from method.
256   /// Otherwise we proceed to the next stage.
257   /// Stages:
258   /// 1. Operations opcodes. Compared as numbers.
259   /// 2. Number of operands.
260   /// 3. Operation types. Compared with cmpType method.
261   /// 4. Compare operation subclass optional data as stream of bytes:
262   /// just convert it to integers and call cmpNumbers.
263   /// 5. Compare in operation operand types with cmpType in
264   /// most significant operand first order.
265   /// 6. Last stage. Check operations for some specific attributes.
266   /// For example, for Load it would be:
267   /// 6.1.Load: volatile (as boolean flag)
268   /// 6.2.Load: alignment (as integer numbers)
269   /// 6.3.Load: ordering (as underlying enum class value)
270   /// 6.4.Load: synch-scope (as integer numbers)
271   /// 6.5.Load: range metadata (as integer ranges)
272   /// On this stage its better to see the code, since its not more than 10-15
273   /// strings for particular instruction, and could change sometimes.
274   ///
275   /// Sets \p needToCmpOperands to true if the operands of the instructions
276   /// still must be compared afterwards. In this case it's already guaranteed
277   /// that both instructions have the same number of operands.
278   int cmpOperations(const Instruction *L, const Instruction *R,
279                     bool &needToCmpOperands) const;
280 
281   /// cmpType - compares two types,
282   /// defines total ordering among the types set.
283   ///
284   /// Return values:
285   /// 0 if types are equal,
286   /// -1 if Left is less than Right,
287   /// +1 if Left is greater than Right.
288   ///
289   /// Description:
290   /// Comparison is broken onto stages. Like in lexicographical comparison
291   /// stage coming first has higher priority.
292   /// On each explanation stage keep in mind total ordering properties.
293   ///
294   /// 0. Before comparison we coerce pointer types of 0 address space to
295   /// integer.
296   /// We also don't bother with same type at left and right, so
297   /// just return 0 in this case.
298   ///
299   /// 1. If types are of different kind (different type IDs).
300   ///    Return result of type IDs comparison, treating them as numbers.
301   /// 2. If types are integers, check that they have the same width. If they
302   /// are vectors, check that they have the same count and subtype.
303   /// 3. Types have the same ID, so check whether they are one of:
304   /// * Void
305   /// * Float
306   /// * Double
307   /// * X86_FP80
308   /// * FP128
309   /// * PPC_FP128
310   /// * Label
311   /// * Metadata
312   /// We can treat these types as equal whenever their IDs are same.
313   /// 4. If Left and Right are pointers, return result of address space
314   /// comparison (numbers comparison). We can treat pointer types of same
315   /// address space as equal.
316   /// 5. If types are complex.
317   /// Then both Left and Right are to be expanded and their element types will
318   /// be checked with the same way. If we get Res != 0 on some stage, return it.
319   /// Otherwise return 0.
320   /// 6. For all other cases put llvm_unreachable.
321   int cmpTypes(Type *TyL, Type *TyR) const;
322 
323   int cmpNumbers(uint64_t L, uint64_t R) const;
324   int cmpAPInts(const APInt &L, const APInt &R) const;
325   int cmpAPFloats(const APFloat &L, const APFloat &R) const;
326   int cmpMem(StringRef L, StringRef R) const;
327 
328   // The two functions undergoing comparison.
329   const Function *FnL, *FnR;
330 
331 private:
332   int cmpOrderings(AtomicOrdering L, AtomicOrdering R) const;
333   int cmpInlineAsm(const InlineAsm *L, const InlineAsm *R) const;
334   int cmpAttrs(const AttributeList L, const AttributeList R) const;
335   int cmpRangeMetadata(const MDNode *L, const MDNode *R) const;
336   int cmpOperandBundlesSchema(const Instruction *L, const Instruction *R) const;
337 
338   /// Compare two GEPs for equivalent pointer arithmetic.
339   /// Parts to be compared for each comparison stage,
340   /// most significant stage first:
341   /// 1. Address space. As numbers.
342   /// 2. Constant offset, (using GEPOperator::accumulateConstantOffset method).
343   /// 3. Pointer operand type (using cmpType method).
344   /// 4. Number of operands.
345   /// 5. Compare operands, using cmpValues method.
346   int cmpGEPs(const GEPOperator *GEPL, const GEPOperator *GEPR) const;
cmpGEPs(const GetElementPtrInst * GEPL,const GetElementPtrInst * GEPR)347   int cmpGEPs(const GetElementPtrInst *GEPL,
348               const GetElementPtrInst *GEPR) const {
349     return cmpGEPs(cast<GEPOperator>(GEPL), cast<GEPOperator>(GEPR));
350   }
351 
352   /// Assign serial numbers to values from left function, and values from
353   /// right function.
354   /// Explanation:
355   /// Being comparing functions we need to compare values we meet at left and
356   /// right sides.
357   /// Its easy to sort things out for external values. It just should be
358   /// the same value at left and right.
359   /// But for local values (those were introduced inside function body)
360   /// we have to ensure they were introduced at exactly the same place,
361   /// and plays the same role.
362   /// Let's assign serial number to each value when we meet it first time.
363   /// Values that were met at same place will be with same serial numbers.
364   /// In this case it would be good to explain few points about values assigned
365   /// to BBs and other ways of implementation (see below).
366   ///
367   /// 1. Safety of BB reordering.
368   /// It's safe to change the order of BasicBlocks in function.
369   /// Relationship with other functions and serial numbering will not be
370   /// changed in this case.
371   /// As follows from FunctionComparator::compare(), we do CFG walk: we start
372   /// from the entry, and then take each terminator. So it doesn't matter how in
373   /// fact BBs are ordered in function. And since cmpValues are called during
374   /// this walk, the numbering depends only on how BBs located inside the CFG.
375   /// So the answer is - yes. We will get the same numbering.
376   ///
377   /// 2. Impossibility to use dominance properties of values.
378   /// If we compare two instruction operands: first is usage of local
379   /// variable AL from function FL, and second is usage of local variable AR
380   /// from FR, we could compare their origins and check whether they are
381   /// defined at the same place.
382   /// But, we are still not able to compare operands of PHI nodes, since those
383   /// could be operands from further BBs we didn't scan yet.
384   /// So it's impossible to use dominance properties in general.
385   mutable DenseMap<const Value*, int> sn_mapL, sn_mapR;
386 
387   // The global state we will use
388   GlobalNumberState* GlobalNumbers;
389 };
390 
391 } // end namespace llvm
392 
393 #endif // LLVM_TRANSFORMS_UTILS_FUNCTIONCOMPARATOR_H
394