2 title: "Pointers Are Complicated II, or: Language first, Optimizations second"
6 Some time ago, I wrote a blog post about how [there's more to a pointer than meets the eye]({% post_url 2018-07-24-pointers-and-bytes %}).
7 One key point I was trying to make is that
9 > *just because two pointers point to the same address, does not mean they are equal and can be used interchangeably.*
11 This "extra information" that distinguishes different pointers to the same address is typically called [*provenance*](https://rust-lang.github.io/unsafe-code-guidelines/glossary.html#pointer-provenance).
12 This post is a cautionary tale of what can happen when provenance is not considered sufficiently carefully in an optimizing compiler.
13 There is also a larger message here about how we should consider compiler IRs primarily from the perspective of a programming language, and only secondarily from the perspective of "which set of optimizations does this allow".
17 I will show a series of three compiler transformation that each seem "intuitively justified" based on wording in the C standard and some common understanding of how pointers work.
18 We will use LLVM for these examples, but the goal is not to pick on LLVM---other compilers suffer from similar issues.
19 The goal is to convince you that to build a correct compiler for languages permitting unsafe pointer manipulation such as C, C++, or Rust,
20 we need to take provenance specifically and IR semantics more generally seriously.
23 ## Warm-up: Why IRs need a precise semantics
25 As a warm-up, I will give a simple example showing that compiler IRs such as LLVM IR need a precise (and precisely documented) semantics.
26 If you are already familiar with the idea of treating compiler IRs as proper programming languages in their own right, or if you are just here for the pointers, you can skip to the next section.
28 Consider the following simple (and contrived, for the sake of this example) piece of C code:
30 int sum_up(int i, int j, int n) {
38 One transformation the compiler might want to do is to move the addition `i+j` out of the loop, to avoid computing the sum each time around the loop:
40 int sum_up(int i, int j, int n) { // optimized version
49 However, that transformation is actually incorrect.
50 If we imagine a caller using this function as `sum_up(INT_MAX, 1, 0)`, then this is a perfectly correct way to call `sum_up`: the loop is never entered, so the overflowing addition `INT_MAX+1` is never performed.
51 However, after the desired optimization, the program now causes a signed integer overflow, which is UB (Undefined Behavior) and thus May Never Happen!
53 One might be tempted to ignore this problem because the UB on integer overflow is a compiler-only concept; every target supported by the compiler will do the obvious thing and just produce an overflowing result.
54 However, there might be other compiler passes running after the optimization we are considering.
55 One such pass might inline `sum_up`, and another pass might notice the `INT_MAX+1` and replace it by `unreachable`, and another pass might then just remove all our code since it is unreachable.
56 Each of these passes has a good reason to exist (it can help real code become a lot faster or help prune dead code), but if we combine them all with our loop hoisting optimization, the result is a disaster.
58 I am convinced that the only way to avoid such problems is to find a way to justify the correctness of each optimization *in isolation*.
59 Each optimization must be *correct* for any possible program, where *correct* means that the optimized program must only "do things" that the original program could have done as well.
60 (This is basically the "as-if" rule in the C standard, and is typically called "refinement" in the academic literature.)
61 In particular, no optimization must ever introduce UB into a UB-free program.
63 It may seem now that under this premise, it is impossible to perform the loop hoisting optimization we are considering.
64 But that is not the case!
65 We can perform the optimization, *we just cannot perform it in C*.
66 Instead, we have to perform it in LLVM IR (or any other IR with a suitable semantics).
67 Specifically, signed integer overflow in LLVM yields a `poison` result.
68 It is not UB to produce `poison`, it is just UB to use `poison` in certain ways (the details do not matter here).
69 In a call to `sum_up(INT_MAX, 1, 0)`, the `s` variable introduced by loop hoisting is unused, so the fact that its value is `poison` does not matter!
71 Due to this behavior of signed integer overflow, the loop hoisting optimization is *correct* if we consider it as an optimization on programs that are written in LLVM IR.[^cheat]
72 In particular, this means we can freely combine this optimization with any other optimization that is *correct* in LLVM IR (such as inlining, replacing definite UB by `unreachable`, and removing unreachable code), and we can be sure that the result obtained after all these optimizations is a correct compilation of our original program.
74 However, to make the argument that an optimization is *correct*, the exact semantics of LLVM IR (what the behavior of all possible programs is and when they have UB) needs to be precisely and unambiguously documented.
75 All involved optimizations need to exactly agree on what is and is not UB, to ensure that whatever code they produce will not be considered UB by a later optimization.[^ub-difference]
76 This is exactly what we also expect from the specification of a programming language such as C, which is why I think we should consider compiler IRs as proper programming languages in their own right, and specify them with the same diligence as we would specify "normal" languages.
77 Sure, no human is going to write many programs in LLVM IR, but clang and rustc produce LLVM IR programs all the time, and as we have seen understanding the exact rules governing these programs is crucial to ensuring that the optimizations LLVM performs do not change program behavior.
79 [^cheat]: If now you feel like we somehow cheated, since we can always translate the program from C to LLVM IR, optimize there, and translate back, consider this: translating from LLVM IR to C is really hard! In particular, singed integer addition in LLVM IR can *not* be translated into signed integer addition in C, since the former is well-defined with `poison` result in case of overflow, but the latter says overflow is UB. C has, in a sense, strictly more UB than LLVM IR, which makes translation in one direction easy, while the other direction is hard.
81 [^ub-difference]: In fact, I would say that two different variants of the IR with different rules for UB are really *two different programming languages*. A program that is well-defined in one language may have UB in another, so great care needs to be taken when the program is moved from being governed by one set of rules to another.
83 *Take-away:* If we want to be able to justify the correctness of a compiler in a modular way, considering only one optimization at a time, we need to perform these optimizations in an IR that has a precise specification of all aspects of program behavior, including UB.
84 Then we can, for each optimization separately, consider the question: does the optimization ever change program behavior, and does it ever introduce UB into UB-free programs?
86 ## How 3 (seemingly) correct optimizations can be incorrect when used together
88 With the warm-up done, we are now ready to consider some more tricky optimizations.
89 We will look at three different optimizations LLVM can perform, and I will show that they *cannot all be correct* since the first and last program we are considering actually have *different behavior*.
90 (More precisely: the last program has a possible behavior that was not possible for the first program.)
91 This is only possible if at least one optimization changed program behavior in an incorrect way, but it is actually not entirely clear which optimization is the culprit.
92 The reasons this happens, I claim, is that LLVM IR and its optimizations are *not* treated the way our warm-up says they should be treated:
93 by precisely defining the behavior and UB of the IR, and arguing why each optimization is *correct* under this definition.
95 The sequence of examples is taken from [this talk](https://sf.snu.ac.kr/llvmtwin/files/presentation.pdf#page=32) by Chung-Kil Hur; it was discovered while working on a rigorous specification of LLVM.
97 Here is the source program:
99 char p[1], q[1] = {0};
107 We are using C syntax here, but remember that we just use C syntax as a convenient way to write programs in LLVM IR.
108 For simplicity, we assume that `int` has the right size to hold a pointer value; just imagine we used `uintptr_t` if you want to be more general.
110 This program has two possible behaviors: either `ip` (the address one-past-the-end of `p`) and `iq` (the address of `q`) are different, and nothing is printed.
111 Or the two are equal, in which case the program will print "10" (`iq` is the result of casting `q` to an integer, so casting it back will yield the original pointer, or at least a pointer pointing to the same object / location in memory).
113 The first "optimization" we will perform is to exploit that `iq == ip`, so we can replace all `iq` by `ip`, and subsequently inline the definition of `ip`:
115 char p[1], q[1] = {0};
119 *(char*)(int)(p+1) = 10; // <- This line changed
124 The second optimization notices that we are taking a pointer `p+1`, casting it to an integer, and casting it back, so we can remove the cast roundtrip:
126 char p[1], q[1] = {0};
130 *(p+1) = 10; // <- This line changed
135 The final optimization notices that `q` is never written to, so we can replace `q[0]` by its initial value `0`:
137 char p[1], q[1] = {0};
142 print(0); // <- This line changed
146 However, this final program is different from the first one!
147 Specifically, the final program will either print nothing or print "0", while the original program *could never print "0"*.
148 This shows that the sequence of three optimizations we performed, as a whole, is *not correct*.
150 #### What went wrong?
152 Clearly, one of the three optimizations is incorrect in the sense that it introduced a change in program behavior.
155 In an ideal world, we would have a sufficiently precise semantics for LLVM IR that we would just have to read the docs (or, even better, run some Miri-like tool) to figure out the answer.
156 However, describing language semantics at this level of precision is *hard*, and full of trade-offs.
157 That's why the LLVM LangRef will not give us a clear answer here, and indeed obtaining a clear answer requires some decisions that have not been explicitly made yet.
158 As a researcher, all I can do is to structure the design space and uncover the trade-offs; it will be up to the LLVM community to decide which option to pick.
159 But the key point is that *they will have to make a choice*, because the status quo of doing all three of these optimizations leads to incorrect compilation results.
161 To proceed, we will use the three optimizations that we considered above as cues: assuming that the optimization is correct for LLVM IR, what does that tell us about the semantics?
163 We start with the last optimization, where the `print` argument is changed from `q[0]` to `0`.
164 This optimization is based on alias analysis:
165 `q[0]` gets initialized to `0` at the beginning of the program, and the only write between that initialization and the `print` is to the pointer `p+1`.
166 Since `q` and `p` point to different local variables, a pointer derived from `p` cannot alias `q[0]`, and hence we know that this write cannot affect the value stored at `q[0]`.
168 Looking more closely, however, reveals that things are not quite so simple!
169 `p+1` is a one-past-the-end pointer, so it actually *can* have the same address as `q[0]`
170 (and, in fact, inside the conditional we know this to be the case).
171 However, LLVM IR (just like C) does not permit memory accesses through one-past-the-end pointers.
172 It makes a difference whether we use `p+1` or `q` inside the `if`, even though we know (in that branch) that both pointers point to the same memory location.
173 This demonstrates that in LLVM IR, there is more to a pointer than just the address it points to---it also matters how this address was computed.
174 This something extra is what we typically call *provenance*.
175 It is impossible to argue for the *correct*ness of the third optimization without acknowledging that provenance is a real part of the semantics of an LLVM IR program.
176 In a flat memory model where pointers are just integers (such as most assembly languages), this optimization is simply wrong.
178 Now that we know that provenance exists in pointers, we have to also consider what happens to provenance when a pointer gets cast to an integer and back.
179 The second optimization gives us a clue into this aspect of LLVM IR semantics: casting a pointer to an integer and back is optimized away, which means that *integers have provenance*.
180 To see why, consider the two expressions `(char*)(int)(p+1)` and `(char*)(int)q`:
181 if the optimization of removing pointer-integer-pointer roundtrips is correct, the first operation will output `p+1` and the second will output `q`, which we just established are two different pointers (they differ in their provenance).
182 The only way to explain this is to say that the input to the `(char*)` cast is different, since the Abstract Machine state is identical in both cases.
183 But we know that the integer value (i.e., the bit pattern of length 32) that serves as input to the `(char*)` cast is the same, and hence a difference can only arise if integers consist of more than just this bit pattern---just like pointers, integers have provenance.
185 Finally, let us consider the first optimization.
186 Here, a successful equality test `iq == ip` prompts the optimizer to replace one value by the other.
187 This optimization demonstrates that *integers do not have provenance*:
188 the optimization is only correct if a successful run-time equality test implies that the two values are equivalent in the Abstract Machine.
189 But this means that the Abstract Machine version of this value cannot have any "funny" extra parts that are not represented at run-time.
190 Of course, provenance is exactly such a "funny" extra part.
191 A different way to phrase the same argument is to say that this optimization is correct only if `iq == ip` implies that both values have the same provenance.
192 This would be a possible definition of `==` in LLVM IR, but only in principle---in practice this means the LLVM backends have to compile `==` in a way that pointer provenance is taken into account, which of course is impossible.
195 By considering each of these three optimizations in terms of what they tell us about the semantics of LLVM IR, we learned that pointers have provenance, that integers remember the provenance of the pointer they come from in case of a pointer-to-integer cast, and that integers do not have provenance.
196 This is a contradiction, and this contradiction explains why we saw incorrect compilation results when applying all three optimizations to the same program.
198 #### How can we fix this?
200 To fix the problem, we will have to declare one of the three optimizations incorrect and stop performing it.
201 Speaking in terms of the LLVM IR semantics, this corresponds to deciding whether pointers and/or integers have provenance:
202 * We could say both pointers and integers have provenance, which invalidates the first optimization.
203 * We could say pointers have provenance but integers do not, which invalidates the second optimization.
204 * We could say nothing has provenance, which invalidates the third optimization.
206 In my opinion, the first and last options are not tenable.
207 Removing provenance altogether kills all but the most simple alias analyses.[^alias]
208 On the other hand, declaring that integers have provenance does not just disable the first optimization in the chain shown above, it also disables common arithmetic optimizations such as `x - x` being equivalent to `0`.
209 Even achieving commutativity and associativity of `+` becomes non-trivial once integers have provenance.
211 [^alias]: Sadly, I do not know of a systematic study of the performance impact of this decision. It is my understanding that many compiler developers "obviously" consider this way too costly of a fix to the problem, but it would still be great to underpin this with some proper data.
213 So, I think that the issue should be resolved by saying that pointers have provenance but integers do not, which means that it is the second optimization that is wrong.
214 This also corresponds to [what has been recently proposed to the C standard committee](http://www.open-std.org/jtc1/sc22/wg14/www/docs/n2577.pdf).
215 That's why [LLVM bug #34548](https://bugs.llvm.org/show_bug.cgi?id=34548) says that optimizing away pointer-integer-pointer roundtrips is incorrect, and LLVM should stop doing this in the general case.
216 There might still be special cases where this can be done, but figuring out the limits of this really needs a more precise description of LLVM IR semantics such as what we proposed [in this paper](https://people.mpi-sws.org/~jung/twinsem/twinsem.pdf).
221 First of all, pointers are complicated.
222 Precisely describing their semantics in a way that is consistent with common alias analyses requires adding a notion of "provenance".
223 In a language such as Java or ML where pointers are opaque types whose representation cannot be observed, this is actually fairly easy to do.
224 But in a language such as Rust, C or C++ that supports pointer-integer casts, the introduction of provenance poses some really tricky questions, and at least one of the commonly performed optimizations in this space has to give.
226 We also learned that LLVM has a bug, but that was *not* the point of this blog post.
227 The GCC developers [made exactly the same mistake](https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82282), and I got word that MSVC and ICC have the same issue (though I do not know how to verify this).
228 And I cannot blame them; the way compiler development typically works nowadays, I think bugs like this are inevitable.
229 Pointer provenance is just a particularly good example of where the current approach failed, but it is not the only case.
230 For example, [ยง2.3 of this paper](https://plv.mpi-sws.org/validc/paper.pdf) (see Figure 3 for the code) shows how a sequence of two optimizations can lead to a miscompilation, where the first optimization is correct under the LLVM concurrency model, and the second optimization is correct under the C++11 concurrency model---but there is no concurrency model under which *both* optimizations are correct, so each compiler (or rather, each compiler IR) needs to pick one or the other.
232 Which brings me to my main conclusion for this post: I think the way optimizing compilers for these low-level languages are built is fundamentally flawed.
233 The current approach, which one might call "optimizations-first", is to largely think of the semantics of a compiler IR in terms of the set of optimizations that it enables.[^weak-mem]
234 However, each additional optimization puts another constraint on the IR semantics---how can we be sure that there even is a way to satisfy all constraints in a single language?
235 The only way I know to ensure that the constraint set remains satisfiable, i.e., to ensure that performing all these optimizations in any order does not introduce bugs, is to pick a consistent semantics for the IR and then show each optimization to be *correct* under those semantics, in the sense defined above.
237 [^weak-mem]: We can also clearly see this approach in most discussions around weak memory concurrency effects, which typically are all about which reorderings the compiler is allowed to perform and how barriers can be used to prevent the reorderings. This is the optimizations-first approach in action.
239 To avoid the problem of incompatible optimizations, I think we need to take compiler IRs more serious as programming languages in their own right, and properly define the Abstract Machine that describes their semantics---including all the UB!
240 I call this "language-first".
241 This language definition does not need to be a formal artifact in a proof assistant, but the description needs to be precise enough such that there are no ambiguities, and such that for each desired optimization we can evaluate whether it is *correct* for this IR or not.
242 One great way to achieve this is to implement a *reference interpreter* for the IR, an interpreter that not only evaluates IR programs but also says if there was any UB triggered by this execution.
243 Writing such an interpreter is a great exercise because it requires explicitly writing out what one could call the "state space" of the Abstract Machine:
244 just what *is* a value, which pieces of data are needed to fully describe it, what kind of information is stored in memory, and so on.
245 Doing so makes it *immediately obvious* that a pointer has provenance, since otherwise it is impossible to correctly check for out-of-bounds accesses.
247 This is really my main motivation for working on the [Miri interpreter](https://github.com/rust-lang/miri/).
248 Of course, practically speaking, its main purpose is to help unsafe code authors avoid UB, but for me personally, I find it equally important that it helps us think about the semantics of Rust and MIR in an *operational* way.
249 This helps shift the discussion around MIR from "optimizations-first" more towards "language-first".
250 Of course, optimizations are still the key motivating factor for defining the language semantics this way or that way, but ultimately it is the language semantics that serves as "ground truth" and keeps it all together, ensuring everything is mutually compatible.
252 I hope this was educational, and thanks for reading. :)
253 As usual, this post can be discussed in the Rust forums.
254 I am curious what your thoughts are on how we can build compilers that do not suffer from these issues.
projects / web.git / blob