indirect immutable freeze by-value function parameters.
Right now, `rustc` only examines function signatures and the platform ABI when
determining the LLVM attributes to apply to parameters. This results in missed
optimizations, because there are some attributes that can be determined via
analysis of the MIR making up the function body. In particular, `readonly`
could be applied to most indirectly-passed by-value function arguments
(specifically, those that are freeze and are observed not to be mutated), but
it currently is not.
This patch introduces the machinery that allows `rustc` to determine those
attributes. It consists of a query, `deduced_param_attrs`, that, when
evaluated, analyzes the MIR of the function to determine supplementary
attributes. The results of this query for each function are written into the
crate metadata so that the deduced parameter attributes can be applied to
cross-crate functions. In this patch, we simply check the parameter for
mutations to determine whether the `readonly` attribute should be applied to
parameters that are indirect immutable freeze by-value. More attributes could
conceivably be deduced in the future: `nocapture` and `noalias` come to mind.
Adding `readonly` to indirect function parameters where applicable enables some
potential optimizations in LLVM that are discussed in [issue 103103] and [PR
103070] around avoiding stack-to-stack memory copies that appear in functions
like `core::fmt::Write::write_fmt` and `core::panicking::assert_failed`. These
functions pass a large structure unchanged by value to a subfunction that also
doesn't mutate it. Since the structure in this case is passed as an indirect
parameter, it's a pointer from LLVM's perspective. As a result, the
intermediate copy of the structure that our codegen emits could be optimized
away by LLVM's MemCpyOptimizer if it knew that the pointer is `readonly
nocapture noalias` in both the caller and callee. We already pass `nocapture
noalias`, but we're missing `readonly`, as we can't determine whether a
by-value parameter is mutated by examining the signature in Rust. I didn't have
much success with having LLVM infer the `readonly` attribute, even with fat
LTO; it seems that deducing it at the MIR level is necessary.
No large benefits should be expected from this optimization *now*; LLVM needs
some changes (discussed in [PR 103070]) to more aggressively use the `noalias
nocapture readonly` combination in its alias analysis. I have some LLVM patches
for these optimizations and have had them looked over. With all the patches
applied locally, I enabled LLVM to remove all the `memcpy`s from the following
code:
```rust
fn main() {
println!("Hello {}", 3);
}
```
which is a significant codegen improvement over the status quo. I expect that
if this optimization kicks in in multiple places even for such a simple
program, then it will apply to Rust code all over the place.
[issue 103103]: https://github.com/rust-lang/rust/issues/103103
[PR 103070]: https://github.com/rust-lang/rust/pull/103070
This attribute now does more than just place noalias on the return,
and has specific requirements for the signature.
Drop the test entirely, as we already check __rust_alloc attributes
in other codegen tests.
Beyond `&`/`&mut`/`Box`, this covers `char`, discriminants, `NonZero*`, etc.
All such types currently cause a Miri error if left uninitialized,
and an `invalid_value` lint in cases like `mem::uninitialized::<char>()`
Note that this _does not_ change whether or not it is UB for `u64` (or
other integer types with no invalid values) to be undef.
This doesn't handle `char` because it's a bit awkward to distinguish it
from u32 at this point in codegen.
Note that for some types (like `&Struct` and `&mut Struct`),
we already apply `dereferenceable`, which implies `noundef`,
so the IR does not change.
This is step 2 towards fixing #77548.
In the codegen and codegen-units test suites, the `//` comment markers
were kept in order not to affect any source locations. This is because
these tests cannot be automatically `--bless`ed.
This saves us both the Freeze/Unpin queries, and avoids placing
noalias attributes, which have a compile-time impact on LLVM
even in optnone builds (due to always_inline functions).
Instead disable creation of assumptions during inlining using an
LLVM opt flag.
The -Z arg-align-attributes option which previously controlled this
behavior is removed.
We represent `bool` as `i1` in a `ScalarPair`, unlike other aggregates,
to optimize IR for checked operators and the like. With this patch, we
still do so when the pair is an immediate value, but we use the `i8`
memory type when the value is loaded or stored as an LLVM aggregate.
So `(bool, bool)` looks like an `{ i1, i1 }` immediate, but `{ i8, i8 }`
in memory. When a pair is a direct function argument, `PassMode::Pair`,
it is still passed using the immediate `i1` type, but as a return value
it will use the `i8` memory type. Also, `bool`-like` enum tags will now
use scalar pairs when possible, where they were previously excluded due
to optimization issues.
This used to be disabled due to LLVM bugs in the handling of
noalias information in conjunction with unwinding. However,
according to #31681 all known LLVM bugs have been fixed by
LLVM 6.0, so it's probably time to reenable this optimization.
Noalias annotations will not be emitted by default if either
-C panic=abort (as previously) or LLVM >= 6.0 (new).
-Z mutable-noalias=no is left as an escape-hatch to allow
debugging problems suspected to stem from this change.
Initially MIR differentiated between arguments and locals, which
introduced a need to add extra copies assigning the argument to a
local, even for simple bindings. This differentiation no longer exists,
but we're still creating those copies, bloating the MIR and LLVM IR we
emit.
Additionally, the current approach means that we create debug info for
both the incoming argument (marking it as an argument), and then
immediately shadow it a local that goes by the same name. This can be
confusing when using e.g. "info args" in gdb, or when e.g. a debugger
with a GUI displays the function arguments separately from the local
variables, especially when the binding is mutable, because the argument
doesn't change, while the local variable does.
This PR is an implementation of [RFC 1974] which specifies a new method of
defining a global allocator for a program. This obsoletes the old
`#![allocator]` attribute and also removes support for it.
[RFC 1974]: https://github.com/rust-lang/rfcs/pull/197
The new `#[global_allocator]` attribute solves many issues encountered with the
`#![allocator]` attribute such as composition and restrictions on the crate
graph itself. The compiler now has much more control over the ABI of the
allocator and how it's implemented, allowing much more freedom in terms of how
this feature is implemented.
cc #27389
LLVM's memory dependence analysis doesn't properly account for calls
that could unwind and thus effectively act as a branching point. This
can lead to stores that are only visible when the call unwinds being
removed, possibly leading to calls to drop() functions with b0rked
memory contents.
As there is no fix for this in LLVM yet and we want to keep
compatibility to current LLVM versions anyways, we have to workaround
this bug by omitting the noalias attribute on &mut function arguments.
Benchmarks suggest that the performance loss by this change is very
small.
Thanks to @RalfJung for pushing me towards not removing too many
noalias annotations and @alexcrichton for helping out with the test for
this bug.
Fixes#29485
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 current codegen tests only compare IR line counts between similar
rust and C programs, the latter getting compiled with clang. That looked
like a good idea back then, but actually things like lifetime intrinsics
mean that less IR isn't always better, so the metric isn't really
helpful.
Instead, we can start doing tests that check specific aspects of the
generated IR, like attributes or metadata. To do that, we can use LLVM's
FileCheck tool which has a number of useful features for such tests.
To start off, I created some tests for a few things that were recently
added and/or broken.
