Alias Analysis (aka Pointer Analysis) is a class of techniques which attempt to determine whether or not two pointers ever can point to the same object in memory. There are many different algorithms for alias analysis and many different ways of classifying them: flow-sensitive vs flow-insensitive, context-sensitive vs context-insensitive, field-sensitive vs field-insensitive, unification-based vs subset-based, etc. Traditionally, alias analyses respond to a query with a Must, May, or No alias response, indicating that two pointers always point to the same object, might point to the same object, or are known to never point to the same object.
The LLVM AliasAnalysis class is the primary interface used by clients and implementations of alias analyses in the LLVM system. This class is the common interface between clients of alias analysis information and the implementations providing it, and is designed to support a wide range of implementations and clients (but currently all clients are assumed to be flow-insensitive). In addition to simple alias analysis information, this class exposes Mod/Ref information from those implementations which can provide it, allowing for powerful analyses and transformations to work well together.
This document contains information necessary to successfully implement this interface, use it, and to test both sides. It also explains some of the finer points about what exactly results mean. If you feel that something is unclear or should be added, please let me know.
The AliasAnalysis class defines the interface that the various alias analysis implementations should support. This class exports two important enums: AliasResult and ModRefResult which represent the result of an alias query or a mod/ref query, respectively.
The AliasAnalysis interface exposes information about memory, represented in several different ways. In particular, memory objects are represented as a starting address and size, and function calls are represented as the actual call or invoke instructions that performs the call. The AliasAnalysis interface also exposes some helper methods which allow you to get mod/ref information for arbitrary instructions.
All AliasAnalysis interfaces require that in queries involving multiple values, values which are not constants are all defined within the same function.
Most importantly, the AliasAnalysis class provides several methods which are used to query whether or not two memory objects alias, whether function calls can modify or read a memory object, etc. For all of these queries, memory objects are represented as a pair of their starting address (a symbolic LLVM Value*) and a static size.
Representing memory objects as a starting address and a size is critically important for correct Alias Analyses. For example, consider this (silly, but possible) C code:
int i;
char C[2];
char A[10];
/* ... */
for (i = 0; i != 10; ++i) {
C[0] = A[i]; /* One byte store */
C[1] = A[9-i]; /* One byte store */
}
In this case, the basicaa pass will disambiguate the stores to C[0] and C[1] because they are accesses to two distinct locations one byte apart, and the accesses are each one byte. In this case, the LICM pass can use store motion to remove the stores from the loop. In constrast, the following code:
int i;
char C[2];
char A[10];
/* ... */
for (i = 0; i != 10; ++i) {
((short*)C)[0] = A[i]; /* Two byte store! */
C[1] = A[9-i]; /* One byte store */
}
In this case, the two stores to C do alias each other, because the access to the &C[0] element is a two byte access. If size information wasn't available in the query, even the first case would have to conservatively assume that the accesses alias.
The alias method is the primary interface used to determine whether or not two memory objects alias each other. It takes two memory objects as input and returns MustAlias, PartialAlias, MayAlias, or NoAlias as appropriate.
Like all AliasAnalysis interfaces, the alias method requires that either the two pointer values be defined within the same function, or at least one of the values is a constant.
The NoAlias response may be used when there is never an immediate dependence between any memory reference based on one pointer and any memory reference based the other. The most obvious example is when the two pointers point to non-overlapping memory ranges. Another is when the two pointers are only ever used for reading memory. Another is when the memory is freed and reallocated between accesses through one pointer and accesses through the other -- in this case, there is a dependence, but it's mediated by the free and reallocation.
As an exception to this is with the noalias keyword; the "irrelevant" dependencies are ignored.
The MayAlias response is used whenever the two pointers might refer to the same object.
The PartialAlias response is used when the two memory objects are known to be overlapping in some way, but do not start at the same address.
The MustAlias response may only be returned if the two memory objects are guaranteed to always start at exactly the same location. A MustAlias response implies that the pointers compare equal.
The getModRefInfo methods return information about whether the execution of an instruction can read or modify a memory location. Mod/Ref information is always conservative: if an instruction might read or write a location, ModRef is returned.
The AliasAnalysis class also provides a getModRefInfo method for testing dependencies between function calls. This method takes two call sites (CS1 & CS2), returns NoModRef if neither call writes to memory read or written by the other, Ref if CS1 reads memory written by CS2, Mod if CS1 writes to memory read or written by CS2, or ModRef if CS1 might read or write memory written to by CS2. Note that this relation is not commutative.
Several other tidbits of information are often collected by various alias analysis implementations and can be put to good use by various clients.
The pointsToConstantMemory method returns true if and only if the analysis can prove that the pointer only points to unchanging memory locations (functions, constant global variables, and the null pointer). This information can be used to refine mod/ref information: it is impossible for an unchanging memory location to be modified.
These methods are used to provide very simple mod/ref information for function calls. The doesNotAccessMemory method returns true for a function if the analysis can prove that the function never reads or writes to memory, or if the function only reads from constant memory. Functions with this property are side-effect free and only depend on their input arguments, allowing them to be eliminated if they form common subexpressions or be hoisted out of loops. Many common functions behave this way (e.g., sin and cos) but many others do not (e.g., acos, which modifies the errno variable).
The onlyReadsMemory method returns true for a function if analysis can prove that (at most) the function only reads from non-volatile memory. Functions with this property are side-effect free, only depending on their input arguments and the state of memory when they are called. This property allows calls to these functions to be eliminated and moved around, as long as there is no store instruction that changes the contents of memory. Note that all functions that satisfy the doesNotAccessMemory method also satisfies onlyReadsMemory.
Writing a new alias analysis implementation for LLVM is quite straight-forward. There are already several implementations that you can use for examples, and the following information should help fill in any details. For a examples, take a look at the various alias analysis implementations included with LLVM.
The first step to determining what type of LLVM pass you need to use for your Alias Analysis. As is the case with most other analyses and transformations, the answer should be fairly obvious from what type of problem you are trying to solve:
In addition to the pass that you subclass, you should also inherit from the AliasAnalysis interface, of course, and use the RegisterAnalysisGroup template to register as an implementation of AliasAnalysis.
Your subclass of AliasAnalysis is required to invoke two methods on the AliasAnalysis base class: getAnalysisUsage and InitializeAliasAnalysis. In particular, your implementation of getAnalysisUsage should explicitly call into the AliasAnalysis::getAnalysisUsage method in addition to doing any declaring any pass dependencies your pass has. Thus you should have something like this:
void getAnalysisUsage(AnalysisUsage &AU) const {
AliasAnalysis::getAnalysisUsage(AU);
// declare your dependencies here.
}
Additionally, your must invoke the InitializeAliasAnalysis method from your analysis run method (run for a Pass, runOnFunction for a FunctionPass, or InitializePass for an ImmutablePass). For example (as part of a Pass):
bool run(Module &M) {
InitializeAliasAnalysis(this);
// Perform analysis here...
return false;
}
All of the AliasAnalysis virtual methods default to providing chaining to another alias analysis implementation, which ends up returning conservatively correct information (returning "May" Alias and "Mod/Ref" for alias and mod/ref queries respectively). Depending on the capabilities of the analysis you are implementing, you just override the interfaces you can improve.
With only two special exceptions (the basicaa and no-aa passes) every alias analysis pass chains to another alias analysis implementation (for example, the user can specify "-basicaa -ds-aa -licm" to get the maximum benefit from both alias analyses). The alias analysis class automatically takes care of most of this for methods that you don't override. For methods that you do override, in code paths that return a conservative MayAlias or Mod/Ref result, simply return whatever the superclass computes. For example:
AliasAnalysis::AliasResult alias(const Value *V1, unsigned V1Size,
const Value *V2, unsigned V2Size) {
if (...)
return NoAlias;
...
// Couldn't determine a must or no-alias result.
return AliasAnalysis::alias(V1, V1Size, V2, V2Size);
}
In addition to analysis queries, you must make sure to unconditionally pass LLVM update notification methods to the superclass as well if you override them, which allows all alias analyses in a change to be updated.
Alias analysis information is initially computed for a static snapshot of the program, but clients will use this information to make transformations to the code. All but the most trivial forms of alias analysis will need to have their analysis results updated to reflect the changes made by these transformations.
The AliasAnalysis interface exposes four methods which are used to communicate program changes from the clients to the analysis implementations. Various alias analysis implementations should use these methods to ensure that their internal data structures are kept up-to-date as the program changes (for example, when an instruction is deleted), and clients of alias analysis must be sure to call these interfaces appropriately.
The addEscapingUse method is used when the uses of a pointer value have changed in ways that may invalidate precomputed analysis information. Implementations may either use this callback to provide conservative responses for points whose uses have change since analysis time, or may recompute some or all of their internal state to continue providing accurate responses.
In general, any new use of a pointer value is considered an escaping use, and must be reported through this callback, except for the uses below:
From the LLVM perspective, the only thing you need to do to provide an efficient alias analysis is to make sure that alias analysis queries are serviced quickly. The actual calculation of the alias analysis results (the "run" method) is only performed once, but many (perhaps duplicate) queries may be performed. Because of this, try to move as much computation to the run method as possible (within reason).
The AliasAnalysis infrastructure has several limitations which make writing a new AliasAnalysis implementation difficult.
There is no way to override the default alias analysis. It would be very useful to be able to do something like "opt -my-aa -O2" and have it use -my-aa for all passes which need AliasAnalysis, but there is currently no support for that, short of changing the source code and recompiling. Similarly, there is also no way of setting a chain of analyses as the default.
There is no way for transform passes to declare that they preserve AliasAnalysis implementations. The AliasAnalysis interface includes deleteValue and copyValue methods which are intended to allow a pass to keep an AliasAnalysis consistent, however there's no way for a pass to declare in its getAnalysisUsage that it does so. Some passes attempt to use AU.addPreserved<AliasAnalysis>, however this doesn't actually have any effect.
AliasAnalysisCounter (-count-aa) and AliasDebugger (-debug-aa) are implemented as ModulePass classes, so if your alias analysis uses FunctionPass, it won't be able to use these utilities. If you try to use them, the pass manager will silently route alias analysis queries directly to BasicAliasAnalysis instead.
Similarly, the opt -p option introduces ModulePass passes between each pass, which prevents the use of FunctionPass alias analysis passes.
The AliasAnalysis API does have functions for notifying implementations when values are deleted or copied, however these aren't sufficient. There are many other ways that LLVM IR can be modified which could be relevant to AliasAnalysis implementations which can not be expressed.
The AliasAnalysisDebugger utility seems to suggest that AliasAnalysis implementations can expect that they will be informed of any relevant Value before it appears in an alias query. However, popular clients such as GVN don't support this, and are known to trigger errors when run with the AliasAnalysisDebugger.
Due to several of the above limitations, the most obvious use for the AliasAnalysisCounter utility, collecting stats on all alias queries in a compilation, doesn't work, even if the AliasAnalysis implementations don't use FunctionPass. There's no way to set a default, much less a default sequence, and there's no way to preserve it.
The AliasSetTracker class (which is used by LICM makes a non-deterministic number of alias queries. This can cause stats collected by AliasAnalysisCounter to have fluctuations among identical runs, for example. Another consequence is that debugging techniques involving pausing execution after a predetermined number of queries can be unreliable.
Many alias queries can be reformulated in terms of other alias queries. When multiple AliasAnalysis queries are chained together, it would make sense to start those queries from the beginning of the chain, with care taken to avoid infinite looping, however currently an implementation which wants to do this can only start such queries from itself.
There are several different ways to use alias analysis results. In order of preference, these are...
The memdep pass uses alias analysis to provide high-level dependence information about memory-using instructions. This will tell you which store feeds into a load, for example. It uses caching and other techniques to be efficient, and is used by Dead Store Elimination, GVN, and memcpy optimizations.
Many transformations need information about alias sets that are active in some scope, rather than information about pairwise aliasing. The AliasSetTracker class is used to efficiently build these Alias Sets from the pairwise alias analysis information provided by the AliasAnalysis interface.
First you initialize the AliasSetTracker by using the "add" methods to add information about various potentially aliasing instructions in the scope you are interested in. Once all of the alias sets are completed, your pass should simply iterate through the constructed alias sets, using the AliasSetTracker begin()/end() methods.
The AliasSets formed by the AliasSetTracker are guaranteed to be disjoint, calculate mod/ref information and volatility for the set, and keep track of whether or not all of the pointers in the set are Must aliases. The AliasSetTracker also makes sure that sets are properly folded due to call instructions, and can provide a list of pointers in each set.
As an example user of this, the Loop Invariant Code Motion pass uses AliasSetTrackers to calculate alias sets for each loop nest. If an AliasSet in a loop is not modified, then all load instructions from that set may be hoisted out of the loop. If any alias sets are stored to and are must alias sets, then the stores may be sunk to outside of the loop, promoting the memory location to a register for the duration of the loop nest. Both of these transformations only apply if the pointer argument is loop-invariant.
The AliasSetTracker class is implemented to be as efficient as possible. It uses the union-find algorithm to efficiently merge AliasSets when a pointer is inserted into the AliasSetTracker that aliases multiple sets. The primary data structure is a hash table mapping pointers to the AliasSet they are in.
The AliasSetTracker class must maintain a list of all of the LLVM Value*'s that are in each AliasSet. Since the hash table already has entries for each LLVM Value* of interest, the AliasesSets thread the linked list through these hash-table nodes to avoid having to allocate memory unnecessarily, and to make merging alias sets extremely efficient (the linked list merge is constant time).
You shouldn't need to understand these details if you are just a client of the AliasSetTracker, but if you look at the code, hopefully this brief description will help make sense of why things are designed the way they are.
If neither of these utility class are what your pass needs, you should use the interfaces exposed by the AliasAnalysis class directly. Try to use the higher-level methods when possible (e.g., use mod/ref information instead of the alias method directly if possible) to get the best precision and efficiency.
If you're going to be working with the LLVM alias analysis infrastructure, you should know what clients and implementations of alias analysis are available. In particular, if you are implementing an alias analysis, you should be aware of the the clients that are useful for monitoring and evaluating different implementations.
This section lists the various implementations of the AliasAnalysis interface. With the exception of the -no-aa implementation, all of these chain to other alias analysis implementations.
The -no-aa pass is just like what it sounds: an alias analysis that never returns any useful information. This pass can be useful if you think that alias analysis is doing something wrong and are trying to narrow down a problem.
The -basicaa pass is an aggressive local analysis that "knows" many important facts:
This pass implements a simple context-sensitive mod/ref and alias analysis for internal global variables that don't "have their address taken". If a global does not have its address taken, the pass knows that no pointers alias the global. This pass also keeps track of functions that it knows never access memory or never read memory. This allows certain optimizations (e.g. GVN) to eliminate call instructions entirely.
The real power of this pass is that it provides context-sensitive mod/ref information for call instructions. This allows the optimizer to know that calls to a function do not clobber or read the value of the global, allowing loads and stores to be eliminated.
Note that this pass is somewhat limited in its scope (only support non-address taken globals), but is very quick analysis.
The -steens-aa pass implements a variation on the well-known "Steensgaard's algorithm" for interprocedural alias analysis. Steensgaard's algorithm is a unification-based, flow-insensitive, context-insensitive, and field-insensitive alias analysis that is also very scalable (effectively linear time).
The LLVM -steens-aa pass implements a "speculatively field-sensitive" version of Steensgaard's algorithm using the Data Structure Analysis framework. This gives it substantially more precision than the standard algorithm while maintaining excellent analysis scalability.
Note that -steens-aa is available in the optional "poolalloc" module, it is not part of the LLVM core.
The -ds-aa pass implements the full Data Structure Analysis algorithm. Data Structure Analysis is a modular unification-based, flow-insensitive, context-sensitive, and speculatively field-sensitive alias analysis that is also quite scalable, usually at O(n*log(n)).
This algorithm is capable of responding to a full variety of alias analysis queries, and can provide context-sensitive mod/ref information as well. The only major facility not implemented so far is support for must-alias information.
Note that -ds-aa is available in the optional "poolalloc" module, it is not part of the LLVM core.
The -scev-aa pass implements AliasAnalysis queries by translating them into ScalarEvolution queries. This gives it a more complete understanding of getelementptr instructions and loop induction variables than other alias analyses have.
The -adce pass, which implements Aggressive Dead Code Elimination uses the AliasAnalysis interface to delete calls to functions that do not have side-effects and are not used.
The -licm pass implements various Loop Invariant Code Motion related transformations. It uses the AliasAnalysis interface for several different transformations:
The -argpromotion pass promotes by-reference arguments to be passed in by-value instead. In particular, if pointer arguments are only loaded from it passes in the value loaded instead of the address to the function. This pass uses alias information to make sure that the value loaded from the argument pointer is not modified between the entry of the function and any load of the pointer.
These passes use AliasAnalysis information to reason about loads and stores.
These passes are useful for evaluating the various alias analysis implementations. You can use them with commands like 'opt -ds-aa -aa-eval foo.bc -disable-output -stats'.
The -print-alias-sets pass is exposed as part of the opt tool to print out the Alias Sets formed by the AliasSetTracker class. This is useful if you're using the AliasSetTracker class. To use it, use something like:
% opt -ds-aa -print-alias-sets -disable-output
The -count-aa pass is useful to see how many queries a particular pass is making and what responses are returned by the alias analysis. As an example,
% opt -basicaa -count-aa -ds-aa -count-aa -licm
will print out how many queries (and what responses are returned) by the -licm pass (of the -ds-aa pass) and how many queries are made of the -basicaa pass by the -ds-aa pass. This can be useful when debugging a transformation or an alias analysis implementation.
The -aa-eval pass simply iterates through all pairs of pointers in a function and asks an alias analysis whether or not the pointers alias. This gives an indication of the precision of the alias analysis. Statistics are printed indicating the percent of no/may/must aliases found (a more precise algorithm will have a lower number of may aliases).
If you're just looking to be a client of alias analysis information, consider using the Memory Dependence Analysis interface instead. MemDep is a lazy, caching layer on top of alias analysis that is able to answer the question of what preceding memory operations a given instruction depends on, either at an intra- or inter-block level. Because of its laziness and caching policy, using MemDep can be a significant performance win over accessing alias analysis directly.
Branch Weight Metadata represents branch weights as its likeliness to be taken. Metadata is assigned to the TerminatorInst as a MDNode of the MD_prof kind. The first operator is always a MDString node with the string "branch_weights". Number of operators depends on the terminator type.
Branch weights might be fetch from the profiling file, or generated based on __builtin_expect instruction.
All weights are represented as an unsigned 32-bit values, where higher value indicates greater chance to be taken.
Metadata is only assign to the conditional branches. There are two extra operarands, for the true and the false branch.
!0 = metadata !{
metadata !"branch_weights",
i32 <TRUE_BRANCH_WEIGHT>,
i32 <FALSE_BRANCH_WEIGHT>
}
Branch weights are assign to every case (including default case which is always case #0).
!0 = metadata !{
metadata !"branch_weights",
i32 <DEFAULT_BRANCH_WEIGHT>
[ , i32 <CASE_BRANCH_WEIGHT> ... ]
}
Branch weights are assign to every destination.
!0 = metadata !{
metadata !"branch_weights",
i32 <LABEL_BRANCH_WEIGHT>
[ , i32 <LABEL_BRANCH_WEIGHT> ... ]
}
Other terminator instructions are not allowed to contain Branch Weight Metadata.
__builtin_expect(long exp, long c) instruction provides branch prediction information. The return value is the value of exp.
It is especially useful in conditional statements. Currently Clang supports two conditional statements:
The exp parameter is the condition. The c parameter is the expected comparision value. If it is equal to 1 (true), the condition is likely to be true, in other case condition is likely to be false. For example:
if (__builtin_expect(x > 0, 1)) {
// This block is likely to be taken.
}
The exp parameter is the value. The c parameter is the expected value. If the expected value doesn't show on the cases list, the default case is assumed to be likely taken.
switch (__builtin_expect(x, 5)) {
default: break;
case 0: // ...
case 3: // ...
case 5: // This case is likely to be taken.
}
Branch Weight Metatada is not proof against CFG changes. If terminator operands' are changed some action should be taken. In other case some misoptimizations may occur due to incorrent branch prediction information.
bugpoint narrows down the source of problems in LLVM tools and passes. It can be used to debug three types of failures: optimizer crashes, miscompilations by optimizers, or bad native code generation (including problems in the static and JIT compilers). It aims to reduce large test cases to small, useful ones. For example, if opt crashes while optimizing a file, it will identify the optimization (or combination of optimizations) that causes the crash, and reduce the file down to a small example which triggers the crash.
For detailed case scenarios, such as debugging opt, llvm-ld, or one of the LLVM code generators, see How To Submit a Bug Report document.
bugpoint is designed to be a useful tool without requiring any hooks into the LLVM infrastructure at all. It works with any and all LLVM passes and code generators, and does not need to "know" how they work. Because of this, it may appear to do stupid things or miss obvious simplifications. bugpoint is also designed to trade off programmer time for computer time in the compiler-debugging process; consequently, it may take a long period of (unattended) time to reduce a test case, but we feel it is still worth it. Note that bugpoint is generally very quick unless debugging a miscompilation where each test of the program (which requires executing it) takes a long time.
bugpoint reads each .bc or .ll file specified on the command line and links them together into a single module, called the test program. If any LLVM passes are specified on the command line, it runs these passes on the test program. If any of the passes crash, or if they produce malformed output (which causes the verifier to abort), bugpoint starts the crash debugger.
Otherwise, if the -output option was not specified, bugpoint runs the test program with the C backend (which is assumed to generate good code) to generate a reference output. Once bugpoint has a reference output for the test program, it tries executing it with the selected code generator. If the selected code generator crashes, bugpoint starts the crash debugger on the code generator. Otherwise, if the resulting output differs from the reference output, it assumes the difference resulted from a code generator failure, and starts the code generator debugger.
Finally, if the output of the selected code generator matches the reference output, bugpoint runs the test program after all of the LLVM passes have been applied to it. If its output differs from the reference output, it assumes the difference resulted from a failure in one of the LLVM passes, and enters the miscompilation debugger. Otherwise, there is no problem bugpoint can debug.
If an optimizer or code generator crashes, bugpoint will try as hard as it can to reduce the list of passes (for optimizer crashes) and the size of the test program. First, bugpoint figures out which combination of optimizer passes triggers the bug. This is useful when debugging a problem exposed by opt, for example, because it runs over 38 passes.
Next, bugpoint tries removing functions from the test program, to reduce its size. Usually it is able to reduce a test program to a single function, when debugging intraprocedural optimizations. Once the number of functions has been reduced, it attempts to delete various edges in the control flow graph, to reduce the size of the function as much as possible. Finally, bugpoint deletes any individual LLVM instructions whose absence does not eliminate the failure. At the end, bugpoint should tell you what passes crash, give you a bitcode file, and give you instructions on how to reproduce the failure with opt or llc.
The code generator debugger attempts to narrow down the amount of code that is being miscompiled by the selected code generator. To do this, it takes the test program and partitions it into two pieces: one piece which it compiles with the C backend (into a shared object), and one piece which it runs with either the JIT or the static LLC compiler. It uses several techniques to reduce the amount of code pushed through the LLVM code generator, to reduce the potential scope of the problem. After it is finished, it emits two bitcode files (called "test" [to be compiled with the code generator] and "safe" [to be compiled with the C backend], respectively), and instructions for reproducing the problem. The code generator debugger assumes that the C backend produces good code.
The miscompilation debugger works similarly to the code generator debugger. It works by splitting the test program into two pieces, running the optimizations specified on one piece, linking the two pieces back together, and then executing the result. It attempts to narrow down the list of passes to the one (or few) which are causing the miscompilation, then reduce the portion of the test program which is being miscompiled. The miscompilation debugger assumes that the selected code generator is working properly.
bugpoint can generate a lot of output and run for a long period of time. It is often useful to capture the output of the program to file. For example, in the C shell, you can run:
bugpoint ... |& tee bugpoint.log
to get a copy of bugpoint's output in the file bugpoint.log, as well as on your terminal.