This has a number of advantages compared to creating a copy in memory
and passing a pointer. The obvious one is that we don't have to put the
data into memory but can keep it in registers. Since we're currently
passing a pointer anyway (instead of using e.g. a known offset on the
stack, which is what the `byval` attribute would achieve), we only use a
single additional register for each fat pointer, but save at least two
pointers worth of stack in exchange (sometimes more because more than
one copy gets eliminated). On archs that pass arguments on the stack, we
save a pointer worth of stack even without considering the omitted
copies.
Additionally, LLVM can optimize the code a lot better, to a large degree
due to the fact that lots of copies are gone or can be optimized away.
Additionally, we can now emit attributes like nonnull on the data and/or
vtable pointers contained in the fat pointer, potentially allowing for
even more optimizations.
This results in LLVM passes being about 3-7% faster (depending on the
crate), and the resulting code is also a few percent smaller, for
example:
text data filename
5671479 3941461 before/librustc-d8ace771.so
5447663 3905745 after/librustc-d8ace771.so
1944425 2394024 before/libstd-d8ace771.so
1896769 2387610 after/libstd-d8ace771.so
I had to remove a call in the backtrace-debuginfo test, because LLVM can
now merge the tails of some blocks when optimizations are turned on,
which can't correctly preserve line info.
Fixes#22924
Cc #22891 (at least for fat pointers the code is good now)
The former stopped making sense when we started interning substs and made
TraitRef a 2-word copy type, and I'm moving the latter into an arena as
they live as long as the type context.
We provide tools to tell what exact symbols to emit for any fn or static, but
don’t quite check if that won’t cause any issues later on. Some of the issues
include LLVM mangling our names again and our names pointing to wrong locations,
us generating dumb foreign call wrappers, linker errors, extern functions
resolving to different symbols altogether (extern {fn fail();} fail(); in some
cases calling fail1()), etc.
Before the commit we had a function called note_unique_llvm_symbol, so it is
clear somebody was aware of the issue at some point, but the function was barely
used, mostly in irrelevant locations.
Along with working on it I took liberty to start refactoring trans/base into
a few smaller modules. The refactoring is incomplete and I hope I will find some
motivation to carry on with it.
This is possibly a [breaking-change] because it makes dumbly written code
properly invalid.
The primary motivation here is to sidestep #19032 -- for a time, I thought that we should improve coherence or otherwise extend the language, but I now think that any such changes will require more time to bake. In the meantime, inheritance amongst the fn traits is both logically correct *and* a simple solution to that obstacle. This change introduces inheritance and modifies the compiler so that it can properly generate impls for closures and fns.
Things enabled by this PR (but not included in this PR):
1. An impl of `FnMut` for `&mut F` where `F : FnMut` (https://github.com/rust-lang/rust/issues/23015).
2. A better version of `Thunk` I've been calling `FnBox`.
I did not include either of these in the PR because:
1. Adding the impls in 1 currently induces a coherence conflict with the pattern trait. This is interesting and merits some discussion.
2. `FnBox` deserves to be a PR of its own.
The main downside to this design is (a) the need to write impls by hand; (b) the possibility of implementing `FnMut` with different semantics from `Fn`, etc. Point (a) is minor -- in particular, it does not affect normal closure usage -- and could be addressed in the future in many ways (better defaults; convenient macros; specialization; etc). Point (b) is unfortunate but "just a bug" from my POV, and certainly not unique to these traits (c.f. Copy/Clone, PartialEq/Eq, etc). (Until we lift the feature-gate on implementing the Fn traits, in any case, there is room to correct both of these if we find a nice way.)
Note that I believe this change is reversible in the future if we decide on another course of action, due to the feature gate on implementing the `Fn` traits, though I do not (currently) think we should reverse it.
Fixes#18835.
r? @nrc
impls.
This requires:
1. modifying trait selection a bit so that when we synthesize impls for
fn pointers and closures;
2. adding code to trans so that we can synthesize a `FnMut`/`FnOnce`
impl for a `Fn` closure and so forth.
... objects
For method calls through trait objects, we currently generate the llvm
function argument attributes using the non-opaque method signature that
still has the trait object fat pointer for the self pointer. This leads
to attributes that are plain wrong, e.g. noalias. As we don't know
anything about the concrete type of the underlying object, we must
replace the self argument with an opaque i8 pointer before applying the
attributes.
For method calls through trait objects, we currently generate the llvm
function argument attributes using the non-opaque method signature that
still has the trait object fat pointer for the self pointer. This leads
to attributes that are plain wrong, e.g. noalias. As we don't know
anything about the concrete type of the underlying object, we must
replace the self argument with an opaque i8 pointer before applying the
attributes.
This is one more step towards completing #13231
This series of commits add support for default trait implementations. The changes in this PR don't break existing code and they are expected to preserve the existing behavior in the compiler as far as built-in bounds checks go.
The PR adds negative implementations of `Send`/`Sync` for some types and it removes the special cases for `Send`/`Sync` during the trait obligations checks. That is, it now fully relies on the traits check rather than lang items.
Once this patch lands and a new snapshot is created, it'll be possible to add default impls for `Send` and `Sync` and remove entirely the use of `BuiltinBound::{BoundSend,BoundSync}` for positive implementations as well.
This PR also removes the restriction on negative implementations. That is, it is now possible to add negative implementations for traits other than `Send`/`Sync`
We already do this for the function arguments, but miss it for the
retslot pointer, which can lead to LLVM assertions because the retslot
has the wrong type.
Fixes#22663
So far, the source location an LLVM instruction was linked to was controlled by
`debuginfo::set_source_location()` and `debuginfo::clear_source_location()`.
This interface mimicked how LLVM's `IRBuilder` handles debug location
assignment. While this interface has some theoretical performance benefits, it
also makes things terribly unstable: One sets some quasi-global state and then
hopes that it is still correct when a given instruction is emitted---an
assumption that has been proven to not hold a bit too often.
This patch requires the debug source location to be passed to the actual
instruction emitting function. This makes source location assignment explicit
and will prevent future changes to `trans` from accidentally breaking things in
the majority of cases.
This patch does not yet implement the new principle for all instruction kinds
but the stepping experience should have improved significantly nonetheless
already.
