X-Git-Url: https://git.ralfj.de/web.git/blobdiff_plain/1afcf299355caf931b823f9473da59e1835d9f8e..45a06433e7b9f45bccf29e4f37dac99020551604:/personal/_posts/2020-07-15-unused-data.md diff --git a/personal/_posts/2020-07-15-unused-data.md b/personal/_posts/2020-07-15-unused-data.md index 16068e8..dcd6b78 100644 --- a/personal/_posts/2020-07-15-unused-data.md +++ b/personal/_posts/2020-07-15-unused-data.md @@ -1,6 +1,7 @@ --- title: "Why even unused data needs to be valid" categories: rust +forum: https://internals.rust-lang.org/t/why-even-unused-data-needs-to-be-valid/12734 --- The Rust compiler has a few assumptions that it makes about the behavior of all code. @@ -9,7 +10,7 @@ Since Rust is a safe-by-default language, programmers usually do not have to wor but authors of `unsafe` code are themselves responsible for upholding these requirements. Those assumptions are [listed in the Rust reference](https://doc.rust-lang.org/reference/behavior-considered-undefined.html). -The one that seems to be most surprising to many people is the clause which says that unsafe code may not *produce* "[...] an invalid value, even in private fields and locals". +The one that seems to be most surprising to many people is the clause which says that Rust code may not *produce* "[...] an invalid value, even in private fields and locals". The reference goes on to explain that "*producing* a value happens any time a value is assigned to or read from a place, passed to a function/primitive operation or returned from a function/primitive operation". In other words, even just *constructing*, for example, an invalid `bool`, is Undefined Behavior---no matter whether that `bool` is ever actually "used" by the program. The purpose of this post is to explain why that rule is so strict. @@ -29,6 +30,9 @@ fn example(b: bool) -> i32 { I hope it is not very surprising that calling `example` on, e.g., `3` transmuted to `bool` is Undefined Behavior (UB). When compiling `if`, the compiler assumes that `0` and `1` are the only possible values; there is no saying what could go wrong when that assumption is violated. +For example, the compiler might use a [jump table](https://en.wikipedia.org/wiki/Branch_table); an out-of-bounds index in that table could literally execute any code, so there is no way to bound the behavior in that case. +(This is a compiler-understood *validity invariant* that is fixed in the language specification, which is very different from a user-defined *safety invariant*. +See [this earlier post]({% post_url 2018-08-22-two-kinds-of-invariants %}) for more details on that distinction.) What is less obvious is why calling `example` on `3` is UB even when there is no such `if` being executed. To understand why that is important, let us consider the following example: @@ -52,7 +56,7 @@ That transformation can be used to turn our `example` function into the followin {% highlight rust %} fn example(b: bool, num: u32) -> i32 { let mut acc = 0; - let incr = if b { 42 } else { 23 } + let incr = if b { 42 } else { 23 }; for _i in 0..num { acc += incr; } @@ -71,7 +75,9 @@ Before the optimization, that call was fine. After the optimization, that call is UB because we are doing `if b` where `b` is `3`. This is because loop-invariant code motion makes dead code live when the loop is not actually executed. -To fix this, we could require the compiler to prove that the loop runs at least once, but in general that will be hard (and in `example` it is impossible, since `num` can indeed be `0`). +(We can think of this as a form of "speculative execution", though entirely unrelated to CPU-level speculative execution.) +To fix the optimization, we could require the compiler to prove that the loop runs at least once (i.e., we could avoid speculative execution), but in general that will be hard (and in `example` it is impossible, since `num` can indeed be `0`). +Another option is to restructure the code to only compute `incr` if `num > 0`, but again this can be hard to do in general. So the alternative that Rust uses is to require that even "unused" data satisfies some basic validity. That makes `example(transmute(3u8), 0)` UB in *both* versions of `example`, and as such the optimization is correct for this input (and indeed for all possible inputs). @@ -85,4 +91,4 @@ But this means our hands are pretty much tied: since we cannot take into account To support inlining and outlining, we also do not want the function boundary to be relevant, which ultimately leads us to the rule that Rust requires today: whenever data of a given type is *produced* anywhere, the data needs to be valid for that type. I hope this post was helpful in explaining why Undefined Behavior in Rust is defined the way it is. -As usual, if you have any comments or questions, let me know in the forums. +As usual, if you have any comments or questions, let me know in the [forums](https://internals.rust-lang.org/t/why-even-unused-data-needs-to-be-valid/12734).