786 lines · plain
1=================================2MergeFunctions pass, how it works3=================================4 5.. contents::6 :local:7 8Introduction9============10Sometimes code contains equal functions, or functions that do exactly the same11thing even though they are non-equal on the IR level (e.g.,: multiplication on 212and ``shl 1``). This can happen for several reasons: mainly, the usage of13templates and automatic code generators. However, sometimes the user itself could14write the same thing twice :-)15 16The main purpose of this pass is to recognize such functions and merge them.17 18This document is an extension to pass comments and describes the pass logic. It19describes the algorithm used to compare functions and20explains how we could combine equal functions correctly to keep the module21valid.22 23The material is presented in a top-down form, so the reader could start to learn pass24from high level ideas and end with low-level algorithm details, thus preparing25him or her for reading the sources.26 27The main goal is to describe the algorithm and logic here and the concept. If28you *don't want* to read the source code, but want to understand pass29algorithms, this document is good for you. The author tries not to repeat the30source code and covers only common cases to avoid the cases of needing to31update this document after any minor code changes.32 33 34What should I know to be able to follow along with this document?35-----------------------------------------------------------------36 37The reader should be familiar with common compiler-engineering principles and38LLVM code fundamentals. In this article, we assume the reader is familiar with39`Single Static Assignment40<http://en.wikipedia.org/wiki/Static_single_assignment_form>`_41concept and has an understanding of42`IR structure <https://llvm.org/docs/LangRef.html#high-level-structure>`_.43 44We will use terms such as45"`module <https://llvm.org/docs/LangRef.html#high-level-structure>`_",46"`function <https://llvm.org/docs/ProgrammersManual.html#the-function-class>`_",47"`basic block <http://en.wikipedia.org/wiki/Basic_block>`_",48"`user <https://llvm.org/docs/ProgrammersManual.html#the-user-class>`_",49"`value <https://llvm.org/docs/ProgrammersManual.html#the-value-class>`_",50"`instruction51<https://llvm.org/docs/ProgrammersManual.html#the-instruction-class>`_".52 53As a good starting point, the Kaleidoscope tutorial can be used:54 55:doc:`tutorial/index`56 57It's especially important to understand Chapter 3 of the tutorial:58 59:doc:`tutorial/LangImpl03`60 61The reader should also know how passes work in LLVM. They can use this62article as a reference and start point here:63 64:doc:`WritingAnLLVMPass`65 66What else? Well perhaps the reader should also have some experience in LLVM pass67debugging and bug-fixing.68 69Narrative structure70-------------------71This article consists of three parts. The first part explains pass functionality72on the top-level. The second part describes the comparison procedure itself.73The third part describes the merging process.74 75In every part, the author tries to put the contents in the top-down form.76The top-level methods will first be described followed by the terminal ones at77the end, in the tail of each part. If the reader sees the reference to the78method that wasn't described yet, they will find its description a bit below.79 80Basics81======82 83How to do it?84-------------85Do we need to merge functions? The obvious answer is: Yes, that is quite a86possible case. We usually *do* have duplicates and it would be good to get rid87of them. But how do we detect duplicates? This is the idea: we split functions88into smaller bricks or parts and compare the "bricks" amount. If equal,89we compare the "bricks" themselves, and then do our conclusions about functions90themselves.91 92What could the difference be? For example, on a machine with 64-bit pointers93(let's assume we have only one address space), one function stores a 64-bit94integer, while another one stores a pointer. If the target is the machine95mentioned above, and if functions are identical, except the parameter type (we96could consider it as a part of function type), then we can treat a ``uint64_t``97and a ``void*`` as equal.98 99This is just an example; more possible details are described a bit below.100 101As another example, the reader may imagine two more functions. The first102function performs a multiplication by 2, while the second one performs a103logical left shift by 1.104 105Possible solutions106^^^^^^^^^^^^^^^^^^107Let's briefly consider possible options about how and what we have to implement108in order to create full-featured functions merging, and also what it would109mean for us.110 111Equal function detection obviously supposes that a "detector" method to be112implemented and latter should answer the question "whether functions are equal".113This "detector" method consists of tiny "sub-detectors", which each answers114exactly the same question, but for function parts.115 116As the second step, we should merge equal functions. So it should be a "merger"117method. "Merger" accepts two functions *F1* and *F2*, and produces *F1F2*118function, the result of merging.119 120Having such routines in our hands, we can process a whole module, and merge all121equal functions.122 123In this case, we have to compare every function with every another function. As124the reader may notice, this way seems to be quite expensive. Of course we could125introduce hashing and other helpers, but it is still just an optimization, and126thus the level of O(N*N) complexity.127 128Can we reach another level? Could we introduce logarithmical search, or random129access lookup? The answer is: "yes".130 131Random-access132"""""""""""""133How can this be done? Just convert each function to a number, and gather134all of them in a special hash table. Functions with equal hashes are equal.135Good hashing means, that every function part must be taken into account. That136means we have to convert every function part into some number, and then add it137into the hash. The lookup-up time would be small, but such an approach adds some138delay due to the hashing routine.139 140Logarithmical search141""""""""""""""""""""142We could introduce total ordering among the functions set, once ordered we143could then implement a logarithmical search. Lookup time still depends on N,144but adds a little of delay (*log(N)*).145 146Present state147"""""""""""""148Both of the approaches (random-access and logarithmical) have been implemented149and tested and both give a very good improvement. What was most150surprising is that logarithmical search was faster; sometimes by up to 15%. The151hashing method needs some extra CPU time, which is the main reason why it works152slower; in most cases, total "hashing" time is greater than total153"logarithmical-search" time.154 155So, preference has been granted to the "logarithmical search".156 157Though in the case of need, *logarithmical-search* (read "total-ordering") could158be used as a milestone on our way to the *random-access* implementation.159 160Every comparison is based either on the numbers or on the flags comparison. In161the *random-access* approach, we could use the same comparison algorithm.162During comparison, we exit once we find the difference, but here we might have163to scan the whole function body every time (note, it could be slower). Like in164"total-ordering", we will track every number and flag, but instead of165comparison, we should get the numbers sequence and then create the hash number.166So, once again, *total-ordering* could be considered as a milestone for even167faster (in theory) random-access approach.168 169MergeFunctions, main fields and runOnModule170^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^171There are two main important fields in the class:172 173``FnTree`` – the set of all unique functions. It keeps items that couldn't be174merged with each other. It is defined as:175 176``std::set<FunctionNode> FnTree;``177 178Here, ``FunctionNode`` is a wrapper for ``llvm::Function`` class, with an179implemented “<” operator among the functions set (below we explain how it works180exactly; this is a key point in fast functions comparison).181 182``Deferred`` – merging process can affect bodies of functions that are in183``FnTree`` already. Obviously, such functions should be rechecked again. In this184case, we remove them from ``FnTree``, and mark them to be rescanned, namely185put them into ``Deferred`` list.186 187runOnModule188"""""""""""189The algorithm is pretty simple:190 1911. Put all module's functions into the *worklist*.192 1932. Scan *worklist*'s functions twice: first, enumerate only strong functions and194then only weak ones:195 196 2.1. Loop body: take a function from *worklist* (call it *FCur*) and try to197 insert it into *FnTree*: check whether *FCur* is equal to one of functions198 in *FnTree*. If there *is* an equal function in *FnTree*199 (call it *FExists*): merge function *FCur* with *FExists*. Otherwise, add200 the function from the *worklist* to *FnTree*.201 2023. Once the *worklist* scanning and merging operations are complete, check the203*Deferred* list. If it is not empty, refill the *worklist* contents with204*Deferred* list and redo step 2, if the *Deferred* list is empty, then exit205from method.206 207Comparison and logarithmical search208"""""""""""""""""""""""""""""""""""209Let's recall our task: for every function *F* from module *M*, we have to find210equal functions *F`* in the shortest time possible and merge them into a211single function.212 213Defining total ordering among the functions set allows us to organize214functions into a binary tree. The lookup procedure complexity would be215estimated as O(log(N)) in this case. But how do we define *total-ordering*?216 217We have to introduce a single rule applicable to every pair of functions, and218following this rule, then evaluate which of them is greater. What kind of rule219could it be? Let's declare it as the "compare" method that returns one of 3220possible values:221 222-1, left is *less* than right,223 2240, left and right are *equal*,225 2261, left is *greater* than right.227 228Of course, it means that we have to maintain229*strict and non-strict order relation properties*:230 231* reflexivity (``a <= a``, ``a == a``, ``a >= a``),232* antisymmetry (if ``a <= b`` and ``b <= a`` then ``a == b``),233* transitivity (``a <= b`` and ``b <= c``, then ``a <= c``)234* asymmetry (if ``a < b``, then ``a > b`` or ``a == b``).235 236As mentioned before, the comparison routine consists of237"sub-comparison-routines", with each of them also consisting of238"sub-comparison-routines", and so on. Finally, it ends up with a primitive239comparison.240 241Below, we will use the following operations:242 243#. ``cmpNumbers(number1, number2)`` is a method that returns -1 if left is less244 than right; 0, if left and right are equal; and 1 otherwise.245 246#. ``cmpFlags(flag1, flag2)`` is a hypothetical method that compares two flags.247 The logic is the same as in ``cmpNumbers``, where ``true`` is 1, and248 ``false`` is 0.249 250The rest of the article is based on *MergeFunctions.cpp* source code251(found in *<llvm_dir>/lib/Transforms/IPO/MergeFunctions.cpp*). We would like252to ask the reader to keep this file open, so we could use it as a reference253for further explanations.254 255Now, we're ready to proceed to the next chapter and see how it works.256 257Functions comparison258====================259First, let's define exactly how we compare complex objects.260 261Complex object comparison (function, basic-block, etc) is mostly based on its262sub-object comparison results. It is similar to the next "tree" objects263comparison:264 265#. For two trees *T1* and *T2* we perform *depth-first-traversal* and have266 two sequences as a product: "*T1Items*" and "*T2Items*".267 268#. We then compare chains "*T1Items*" and "*T2Items*" in269 the most-significant-item-first order. The result of items comparison270 would be the result of *T1* and *T2* comparison itself.271 272FunctionComparator::compare(void)273---------------------------------274A brief look at the source code tells us that the comparison starts in the275“``int FunctionComparator::compare(void)``” method.276 2771. The first parts to be compared are the function's attributes and some278properties that are outside the “attributes” term, but still could make the279function different without changing its body. This part of the comparison is280usually done within simple *cmpNumbers* or *cmpFlags* operations (e.g.281``cmpFlags(F1->hasGC(), F2->hasGC())``). Below is a full list of function's282properties to be compared on this stage:283 284 * *Attributes* (those are returned by ``Function::getAttributes()``285 method).286 287 * *GC*, for equivalence, *RHS* and *LHS* should be both either without288 *GC* or with the same one.289 290 * *Section*, just like a *GC*: *RHS* and *LHS* should be defined in the291 same section.292 293 * *Variable arguments*. *LHS* and *RHS* should be both either with or294 without *var-args*.295 296 * *Calling convention* should be the same.297 2982. Function type. Checked by ``FunctionComparator::cmpType(Type*, Type*)``299method. It checks return type and parameters type; the method itself will be300described later.301 3023. Associate function formal parameters with each other. Then comparing function303bodies, if we see the usage of *LHS*'s *i*-th argument in *LHS*'s body, then,304we want to see usage of *RHS*'s *i*-th argument at the same place in *RHS*'s305body, otherwise functions are different. On this stage we grant the preference306to those we met later in function body (value we met first would be *less*).307This is done by “``FunctionComparator::cmpValues(const Value*, const Value*)``”308method (will be described a bit later).309 3104. Function body comparison. As written in method comments:311 312“We do a CFG-ordered walk since the actual ordering of the blocks in the linked313list is immaterial. Our walk starts at the entry block for both functions, then314takes each block from each terminator in order. As an artifact, this also means315that unreachable blocks are ignored.”316 317So, using this walk, we get BBs from *left* and *right* in the same order, and318compare them by “``FunctionComparator::compare(const BasicBlock*, const319BasicBlock*)``” method.320 321We also associate BBs with each other, like we did it with function formal322arguments (see ``cmpValues`` method below).323 324FunctionComparator::cmpType325---------------------------326Consider how type comparison works.327 3281. Coerce pointer to integer. If the left type is a pointer, try to coerce it to the329integer type. It could be done if its address space is 0, or if address spaces330are ignored at all. Do the same thing for the right type.331 3322. If the left and right types are equal, return 0. Otherwise, we need to give333preference to one of them. So proceed to the next step.334 3353. If the types are of different kind (different type IDs). Return result of type336IDs comparison, treating them as numbers (use ``cmpNumbers`` operation).337 3384. If the types are vectors or integers, return result of their pointers comparison,339comparing them as numbers.340 3415. Check whether type ID belongs to the next group (call it equivalent-group):342 343 * Void344 345 * Float346 347 * Double348 349 * X86_FP80350 351 * FP128352 353 * PPC_FP128354 355 * Label356 357 * Metadata.358 359 If ID belongs to group above, return 0. Since it's enough to see that360 types has the same ``TypeID``. No additional information is required.361 3626. Left and right are pointers. Return result of address space comparison363(numbers comparison).364 3657. Complex types (structures, arrays, etc.). Follow complex objects comparison366technique (see the very first paragraph of this chapter). Both *left* and367*right* are to be expanded and their element types will be checked the same368way. If we get -1 or 1 on some stage, return it. Otherwise, return 0.369 3708. Steps 1-6 describe all the possible cases, if we passed steps 1-6 and didn't371get any conclusions, then invoke ``llvm_unreachable``, since it's quite an372unexpectable case.373 374cmpValues(const Value*, const Value*)375-------------------------------------376Method that compares local values.377 378This method gives us an answer to a very curious question: whether we could379treat local values as equal, and which value is greater otherwise. It's380better to start from example:381 382Consider the situation when we're looking at the same place in left383function "*FL*" and in right function "*FR*". Every part of *left* place is384equal to the corresponding part of *right* place, and (!) both parts use385*Value* instances, for example:386 387.. code-block:: text388 389 instr0 i32 %LV ; left side, function FL390 instr0 i32 %RV ; right side, function FR391 392So, now our conclusion depends on *Value* instances comparison.393 394The main purpose of this method is to determine the relation between such values.395 396What can we expect from equal functions? At the same place, in functions397"*FL*" and "*FR*" we expect to see *equal* values, or values *defined* at398the same place in "*FL*" and "*FR*".399 400Consider a small example here:401 402.. code-block:: text403 404 define void %f(i32 %pf0, i32 %pf1) {405 instr0 i32 %pf0 instr1 i32 %pf1 instr2 i32 123406 }407 408.. code-block:: text409 410 define void %g(i32 %pg0, i32 %pg1) {411 instr0 i32 %pg0 instr1 i32 %pg0 instr2 i32 123412 }413 414In this example, *pf0* is associated with *pg0*, *pf1* is associated with415*pg1*, and we also declare that *pf0* < *pf1*, and thus *pg0* < *pf1*.416 417Instructions with opcode "*instr0*" would be *equal*, since their types and418opcodes are equal, and values are *associated*.419 420Instructions with opcode "*instr1*" from *f* is *greater* than instructions421with opcode "*instr1*" from *g*; here we have equal types and opcodes, but422"*pf1* is greater than "*pg0*".423 424Instructions with opcode "*instr2*" are equal, because their opcodes and425types are equal, and the same constant is used as a value.426 427What we associate in cmpValues?428^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^429* Function arguments. *i*-th argument from left function associated with430 *i*-th argument from right function.431* BasicBlock instances. In basic-block enumeration loop we associate *i*-th432 BasicBlock from the left function with *i*-th BasicBlock from the right433 function.434* Instructions.435* Instruction operands. Note, we can meet *Value* here we have never seen436 before. In this case it is not a function argument, nor *BasicBlock*, nor437 *Instruction*. It is a global value. It is a constant, since it's the only438 supposed global here. The method also compares: Constants that are of the439 same type and if right constant can be losslessly bit-casted to the left440 one, then we also compare them.441 442How to implement cmpValues?443^^^^^^^^^^^^^^^^^^^^^^^^^^^444*Association* is a case of equality for us. We just treat such values as equal,445but, in general, we need to implement antisymmetric relation. As mentioned446above, to understand what is *less*, we can use order in which we447meet values. If both values have the same order in a function (met at the same448time), we then treat values as *associated*. Otherwise, it depends on who was449first.450 451Every time we run the top-level compare method, we initialize two identical452maps (one for the left side, another one for the right side):453 454``map<Value, int> sn_mapL, sn_mapR;``455 456The key of the map is the *Value* itself; the *value* – is its order (call it457*serial number*).458 459To add value *V* we need to perform the next procedure:460 461``sn_map.insert(std::make_pair(V, sn_map.size()));``462 463For the first *Value*, the map will return *0*, for the second *Value*, the map will464return *1*, and so on.465 466We can then check whether the left and right values met at the same time with467a simple comparison:468 469``cmpNumbers(sn_mapL[Left], sn_mapR[Right]);``470 471Of course, we can combine insertion and comparison:472 473.. code-block:: c++474 475 std::pair<iterator, bool>476 LeftRes = sn_mapL.insert(std::make_pair(Left, sn_mapL.size())), RightRes477 = sn_mapR.insert(std::make_pair(Right, sn_mapR.size()));478 return cmpNumbers(LeftRes.first->second, RightRes.first->second);479 480Let's look at how the whole method could be implemented.481 4821. We have to start with the bad news. Consider function self and483cross-referencing cases:484 485.. code-block:: c++486 487 // self-reference unsigned fact0(unsigned n) { return n > 1 ? n488 * fact0(n-1) : 1; } unsigned fact1(unsigned n) { return n > 1 ? n *489 fact1(n-1) : 1; }490 491 // cross-reference unsigned ping(unsigned n) { return n!= 0 ? pong(n-1) : 0;492 } unsigned pong(unsigned n) { return n!= 0 ? ping(n-1) : 0; }493 494..495 496 This comparison has been implemented in initial *MergeFunctions* pass497 version. But, unfortunately, it is not transitive. And this is the only case498 we can't convert to less-equal-greater comparison. It is a seldom case, 4-5499 functions of 10000 (checked in test-suite), and, we hope, the reader would500 forgive us for such a sacrifice in order to get the O(log(N)) pass time.501 5022. If left/right *Value* is a constant, we have to compare them. Return 0 if it503is the same constant, or use ``cmpConstants`` method otherwise.504 5053. If left/right is *InlineAsm* instance. Return result of *Value* pointers506comparison.507 5084. Explicit association of *L* (left value) and *R* (right value). We need to509find out whether values met at the same time, and thus are *associated*. Or we510need to put the rule: when we treat *L* < *R*. Now it is easy: we just return511the result of numbers comparison:512 513.. code-block:: c++514 515 std::pair<iterator, bool>516 LeftRes = sn_mapL.insert(std::make_pair(Left, sn_mapL.size())),517 RightRes = sn_mapR.insert(std::make_pair(Right, sn_mapR.size()));518 if (LeftRes.first->second == RightRes.first->second) return 0;519 if (LeftRes.first->second < RightRes.first->second) return -1;520 return 1;521 522Now, when *cmpValues* returns 0, we can proceed with the comparison procedure.523Otherwise, if we get (-1 or 1), we need to pass this result to the top level,524and finish comparison procedure.525 526cmpConstants527------------528Performs a constant comparison as follows:529 5301. Compare constant types using ``cmpType`` method. If the result is -1 or 1,531goto step 2, otherwise proceed to step 3.532 5332. If types are different, we still can check whether constants could be534losslessly bitcasted to each other. The further explanation is modification of535``canLosslesslyBitCastTo`` method.536 537 2.1 Check whether constants are of the first class types538 (``isFirstClassType`` check):539 540 2.1.1. If both constants are *not* of the first class type: return result541 of ``cmpType``.542 543 2.1.2. Otherwise, if left type is not of the first class, return -1. If544 right type is not of the first class, return 1.545 546 2.1.3. If both types are of the first class type, proceed to the next step547 (2.1.3.1).548 549 2.1.3.1. If types are vectors, compare their bitwidth using the550 *cmpNumbers*. If result is not 0, return it.551 552 2.1.3.2. Different types, but not vectors:553 554 * if both of them are pointers, good for us, we can proceed to step 3.555 * if one of types is pointer, return result of *isPointer* flags556 comparison (*cmpFlags* operation).557 * otherwise we have no methods to prove bitcastability, and thus return558 result of types comparison (-1 or 1).559 560Steps below are for the case when types are equal, or case when constants are561bitcastable:562 5633. One of constants is a "*null*" value. Return the result of564``cmpFlags(L->isNullValue, R->isNullValue)`` comparison.565 5664. Compare value IDs, and return result if it is not 0:567 568.. code-block:: c++569 570 if (int Res = cmpNumbers(L->getValueID(), R->getValueID()))571 return Res;572 5735. Compare the contents of constants. The comparison depends on the kind of574constants, but on this stage it is just a lexicographical comparison. Just see575how it was described in the beginning of "*Functions comparison*" paragraph.576Mathematically, it is equal to the next case: we encode left constant and right577constant (with similar way *bitcode-writer* does). Then compare left code578sequence and right code sequence.579 580compare(const BasicBlock*, const BasicBlock*)581---------------------------------------------582Compares two *BasicBlock* instances.583 584It enumerates instructions from left *BB* and right *BB*.585 5861. It assigns serial numbers to the left and right instructions, using587``cmpValues`` method.588 5892. If one of left or right is *GEP* (``GetElementPtr``), then treat *GEP* as590greater than other instructions. If both instructions are *GEPs* use ``cmpGEP``591method for comparison. If result is -1 or 1, pass it to the top-level592comparison (return it).593 594 3.1. Compare operations. Call ``cmpOperation`` method. If result is -1 or595 1, return it.596 597 3.2. Compare number of operands, if result is -1 or 1, return it.598 599 3.3. Compare operands themselves, use ``cmpValues`` method. Return result600 if it is -1 or 1.601 602 3.4. Compare type of operands, using ``cmpType`` method. Return result if603 it is -1 or 1.604 605 3.5. Proceed to the next instruction.606 6074. We can finish instruction enumeration in 3 cases:608 609 4.1. We reached the end of both left and right basic-blocks. We didn't610 exit on steps 1-3, so contents are equal, return 0.611 612 4.2. We have reached the end of the left basic-block. Return -1.613 614 4.3. Return 1 (we reached the end of the right basic block).615 616cmpGEP617------618Compares two GEPs (``getelementptr`` instructions).619 620It differs from regular operations comparison with the only thing: possibility621to use ``accumulateConstantOffset`` method.622 623So, if we get constant offset for both left and right *GEPs*, then compare it as624numbers, and return comparison result.625 626Otherwise, treat it like a regular operation (see previous paragraph).627 628cmpOperation629------------630Compares instruction opcodes and some important operation properties.631 6321. Compare opcodes, if it differs return the result.633 6342. Compare number of operands. If it differs – return the result.635 6363. Compare operation types, use *cmpType*. All the same – if types are637different, return result.638 6394. Compare *subclassOptionalData*, get it with ``getRawSubclassOptionalData``640method, and compare it like a numbers.641 6425. Compare operand types.643 6446. For some particular instructions, check equivalence (relation in our case) of645some significant attributes. For example, we have to compare alignment for646``load`` instructions.647 648O(log(N))649---------650Methods described above implement order relationship. And latter, could be used651for nodes comparison in a binary tree. So we can organize functions set into652the binary tree and reduce the cost of lookup procedure from653O(N*N) to O(log(N)).654 655Merging process, mergeTwoFunctions656==================================657Once *MergeFunctions* detects that current function (*G*) is equal to one that658was analyzed before (function *F*) it calls ``mergeTwoFunctions(Function*,659Function*)``.660 661Operation affects ``FnTree`` contents in the following way: *F* will stay in662``FnTree``. *G* being equal to *F* will not be added to ``FnTree``. Calls of663*G* would be replaced with something else. It changes bodies of callers. So,664functions that calls *G* would be put into ``Deferred`` set and removed from665``FnTree``, and analyzed again.666 667The approach is as follows:668 6691. Most wished case: when we can use alias and both of *F* and *G* are weak. We670make both of them with aliases to the third strong function *H*. Actually *H*671is *F*. See below how it's made (but it's better to look straight into the672source code). Well, this is a case when we can just replace *G* with *F*673everywhere, we use ``replaceAllUsesWith`` operation here (*RAUW*).674 6752. *F* could not be overridden, while *G* could. It would be good to do the676next: after merging the places where overridable function were used, still use677overridable stub. So try to make *G* alias to *F*, or create overridable tail678call wrapper around *F* and replace *G* with that call.679 6803. Neither *F* nor *G* could be overridden. We can't use *RAUW*. We can just681change the callers: call *F* instead of *G*. That's what682``replaceDirectCallers`` does.683 684Below is a detailed body description.685 686If “F” may be overridden687------------------------688As follows from ``mayBeOverridden`` comments: “whether the definition of this689global may be replaced by something non-equivalent at link time”. If so, that's690ok: we can use alias to *F* instead of *G* or change call instructions itself.691 692HasGlobalAliases, removeUsers693^^^^^^^^^^^^^^^^^^^^^^^^^^^^^694First, consider the case when we have global aliases of one function name to695another. Our purpose is to make both of them with aliases to the third strong696function. However, if we keep *F* alive and without major changes, we can leave it697in ``FnTree``. Try to combine these two goals.698 699Do a stub replacement of *F* itself with an alias to *F*.700 7011. Create stub function *H*, with the same name and attributes like function702*F*. It takes maximum alignment of *F* and *G*.703 7042. Replace all uses of function *F* with uses of function *H*. It is a705two-step procedure instead. First of all, we must take into account that all functions706that call *F* would be changed because we change the call argument707(from *F* to *H*). If so, we must review these caller functions again after708this procedure. We remove callers from ``FnTree``, method with name709``removeUsers(F)`` does that (don't confuse with ``replaceAllUsesWith``):710 711 2.1. ``Inside removeUsers(Value*712 V)`` we go through the all values that use value *V* (or *F* in our context).713 If value is instruction, we go to function that holds this instruction and714 mark it as to-be-analyzed-again (put to ``Deferred`` set), we also remove715 caller from ``FnTree``.716 717 2.2. Now we can do the replacement: call ``F->replaceAllUsesWith(H)``.718 7193. *H* (that now "officially" plays *F*'s role) is replaced with alias to *F*.720Do the same with *G*: replace it with alias to *F*. So finally everywhere *F*721was used, we use *H* and it is alias to *F*, and everywhere *G* was used we722also have alias to *F*.723 7244. Set *F* linkage to private. Make it strong :-)725 726No global aliases, replaceDirectCallers727^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^728If global aliases are not supported, we call ``replaceDirectCallers``. Just729go through all calls of *G* and replace it with calls of *F*. If you look into730the method you will see that it scans all uses of *G* too, and if use is callee731(if user is call instruction and *G* is used as what to be called), we replace732it with use of *F*.733 734If “F” could not be overridden, fix it!735"""""""""""""""""""""""""""""""""""""""736 737We call ``writeThunkOrAlias(Function *F, Function *G)``. Here we try to replace738*G* with an alias to *F* first. The next conditions are essential:739 740* target should support global aliases,741* the address itself of *G* should be not significant, not named and not742 referenced anywhere,743* function should come with external, local or weak linkage.744 745Otherwise, we write thunk: some wrapper that has *G's* interface and calls *F*,746so *G* could be replaced with this wrapper.747 748*writeAlias*749 750As follows from *llvm* reference:751 752“Aliases act as *second name* for the aliasee value”. So we just want to create753a second name for *F* and use it instead of *G*:754 7551. create global alias itself (*GA*),756 7572. adjust alignment of *F* so it must be maximum of current and *G's* alignment;758 7593. replace uses of *G*:760 761 3.1. first mark all callers of *G* as to-be-analyzed-again, using762 ``removeUsers`` method (see chapter above),763 764 3.2. call ``G->replaceAllUsesWith(GA)``.765 7664. Get rid of *G*.767 768*writeThunk*769 770As it written in method comments:771 772“Replace G with a simple tail call to bitcast(F). Also replace direct uses of G773with bitcast(F). Deletes G.”774 775In general, it does the same as usual when we want to replace callee, except the776first point:777 7781. We generate tail call wrapper around *F*, but with an interface that allows using779it instead of *G*.780 7812. “As-usual”: ``removeUsers`` and ``replaceAllUsesWith`` then.782 7833. Get rid of *G*.784 785 786