RFC: Add additional inline intents

Author: scottmcm

text/0000-inline-intents.md

@@ -0,0 +1,189 @@
+- Feature Name: `inline_intents`
+- Start Date: 2025-02-22
+- RFC PR: [rust-lang/rfcs#0000](https://github.com/rust-lang/rfcs/pull/0000)
+- Rust Issue: [rust-lang/rust#0000](https://github.com/rust-lang/rust/issues/0000)
+
+# Summary
+[summary]: #summary
+
+Add `#[inline(trampoline)]` and `#[inline(rarely)]` to give additional control
+over the inlining decisions made by the compiler.
+
+# Motivation
+[motivation]: #motivation
+
+Right now it's pretty common to just slap `#[inline]` on things without thinking
+about it all that hard about it, and the existing controls aren't great.
+
+Often we'll get PRs using `inline(always)` "because it's just calling something
+else so of course it should be inlined", for example.  But because of the
+bottom-up nature of inlining, that's a bad thing to do because if the callee
+happens to get inlined, then it'll "always" inline that callee too, which might
+not be what was actually desired.
+
+At the same time, sometimes it's useful to put `inline` on things to make the
+definition available to to LLVM, but where it probably shouldn't actually be
+inlined in general, only in particular special cases (perhaps when one of the
+arguments is a small constant, for example).
+
+It would thus be nice to give additional options to the user to let them
+describe *why* they wanted a function inlined, and hopefully be able to make
+better decisions in the backend as a result.
+
+
+# Guide-level explanation
+[guide-level-explanation]: #guide-level-explanation
+
+In most cases, plain `#[inline]` is fine, especially with PGO and LTO.
+However, if you've measured and things are making a poor choice, there are some
+options you can use to hint the compiler towards what you want.
+
+## `inline(trampoline)`
+
+This is intended for functions which quickly "bounce" the caller off to some
+other implementation, after doing some initial checks or transformations.
+
+For example, this is useful in a safe function which does some safety checks,
+then calls an `unsafe` version of the function to do the actual work.  Or maybe
+it's a function with a common trivial path, but which sometimes needs to call
+out to a more complicated version, like how `Vec::push` is usually trivial but
+occasionally needs to reallocate.
+
+## `inline(rarely)`
+
+This is intended for functions which normally shouldn't be inlined, but where
+exceptions exist so you don't want full `inline(never)`.
+
+For example, maybe this is a vectorized loop that you wouldn't want copied into
+every caller, but you know that using it for short arrays is common, and thus
+want the back-end to be able to inline and fully unroll it in those cases.
+
+## In combination
+
+These can work particularly well together.
+
+For example, perhaps the public function is `inline(trampoline)` so that the
+initial `NonZero::new` check can be inlined into the caller (where it's more
+easily optimized out) but that calls a private `inline(rarely)` function which
+takes `NonZero` to avoid needing extra checks internally.
+
+Or perhaps you write an `inline(trampoline)` function that picks a strategy
+based on the types and arguments, then dispatches to one of many possible
+`inline(rarely)` implementations.
+
+
+# Reference-level explanation
+[reference-level-explanation]: #reference-level-explanation
+
+Like `inline`, these "do nothing" in a spec sense.
+
+The only language change is thus in allowing the two new tokens in the attribute
+(in addition to `never` and `always`).
+
+## Implementation options
+
+⚠ These are not normative, just to illustrate possibilities. ⚠
+
+In LLVM, `#[inline]` sets the [`inlinehint` function attribute](https://llvm.org/docs/LangRef.html#function-attributes),
+so `inline(rarely)` could skip doing that, and thus comparatively slightly
+discourage inlining it.
+
+In MIR inlining, we already attempt to deduce trampolines as of [#127113].
+This would let people opt-in to that behaviour, even in places it's not obvious.
+Then we could allow inlining trampolines into trampolines, but avoid inlining
+non-trampolines into a trampoline.  And we have an internal attribute
+`rustc_no_mir_inline` which blocks MIR inlining, so `inline(rarely)` could also
+do that (or maybe just make the threshold very restrictive).
+
+[#127113]: https://github.com/rust-lang/rust/pull/127113
+
+
+# Drawbacks
+[drawbacks]: #drawbacks
+
+These are still up to the programmer to get right, so
+- they might just make analysis paralysis worse
+- they might be worse than having something like a PGO-based system instead
+- they might turn out to not actually help as hoped
+- they might lead to more bugs in the tracker when they don't do what people thought
+- they might be the wrong pivots and we'll just end up needing more
+
+However, at worst we just make them allowed but not do anything, so at worst the
+cost of having them is very small: a compiler just parses-then-ignores them.
+
+
+# Rationale and alternatives
+[rationale-and-alternatives]: #rationale-and-alternatives
+
+The hope here is that by giving *intent* we can tune how these work in a better
+way than is possible with the existing knobs.
+
+The existing options are insufficient because `always` and `never` are too strong.
+With LLVM's bottom-up inlining logic, `always` isn't the right answer the vast
+majority of the time, and if LLVM really doesn't want to inline it, most of the
+time there's a good reason for that.  Similarly if it really thinks that
+something would be good to inline, the majority of the time blocking that with
+`never` isn't really want you want.
+
+For example, the library wants to *discourage* inlining of the UB check
+functions, but doesn't want `never` because if all the arguments are constants
+inlining it to const-fold away all the checks is good.  (It'd be silly to force
+`NonNull::new(ptr::without_provenance(2))` to have an actual call to
+`precondition_check(2_ptr)`.)
+
+And the blanket impl of `Into::into` that just calls `From::from` wants to
+"always" be inlined, but because of how inlining works it'd be wrong for it to
+use `always` since it doesn't necessarily want to inline the whole subtree.
+
+Thus today usually all that one can say is `#[inline]`, and hope that the
+compiler does the right thing.  Often it does, but sometimes it also makes poor
+decisions that can result in binary bloat or poor performance.  With these to
+nudge it one way or the other, hopefully that will avoid the downsides of the
+existing really big hammers, while also not over-promising for what will always
+just be a hint, not a promise.
+
+`inline(rarely)` could instead be done with something in the body to increase
+the cost, perhaps marking it `#[inline]` but then calling a hypothetical magic
+`core::hint::discourage_inlining()`.  But to make it an inlining candidate at
+all means it still needs the attribute, so it seems nicer to let it just be
+mentioned in the attribute without needing to inspect the body, given that
+the attribute is already parameterized.
+
+
+# Prior art
+[prior-art]: #prior-art
+
+Various languages have inlining controls, but I'm unaware of any with either of
+these specific intents.
+
+[LLVM](https://llvm.org/docs/LangRef.html#function-attributes)
+has `alwaysinline` vs `inlinehint` vs `noduplicate` vs `noinline`.
+
+[GCC](https://gcc.gnu.org/onlinedocs/gcc/Inline.html)
+has `inline` and `__attribute__((always_inline))` and `extern inline`.
+
+[DotNet](https://learn.microsoft.com/en-us/dotnet/api/system.runtime.compilerservices.methodimploptions)
+has `NoInlining` vs `AggressiveInlining` hints.
+
+
+# Unresolved questions
+[unresolved-questions]: #unresolved-questions
+
+The exact behaviour of how these affect program compilation will likely continue
+to be tweaked even after they stabilize, assuming this were to be accepted.
+
+The goal of the RFC process is to pick tokens that are sufficiently evocative
+of what intent the coder is expressing by using them; the details happen later.
+
+
+# Future possibilities
+[future-possibilities]: #future-possibilities
+
+There are always more possible intents that one could imagine.  For example, one
+for "this is small and uninteresting" could make sense, which could even have
+more non-semantic effects like emitting less debug info.  But for now it's less
+obvious that that's worth distinguishing, since just things being small (like
+basic accessor functions often are) is enough to already trigger making it
+cross-crate inlinable even without the attribute, and things that are small
+and simple are reliably inlined without extra hinting anyway.
+