struct field reordering and optimization
This is work in progress. The goal is to divorce the order of fields in source code from the order of fields in the LLVM IR, then optimize structs (and tuples/enum variants)by always ordering fields from least to most aligned. It does not work yet. I intend to check compiler memory usage as a benchmark, and a crater run will probably be required.
I don't know enough of the compiler to complete this work unaided. If you see places that still need updating, please mention them. The only one I know of currently is debuginfo, which I'm putting off intentionally until a bit later.
r? @eddyb
Currently libraries installed by rustbuild on OSX have an incorrect
`LC_ID_DYLIB` directive located in the dynamic libraries that are
installed. The directive we expect looks like:
@rpath/libstd.dylib
Which means that if you want to find that dynamic library you should
look at the dylib's other `@rpath` directives. Typically our `@rpath`
directives look like `@loader_path/../lib` for the compiler as that's
where the installed libraries will be located. Currently, though,
rustbuild produces dylibs with the directive that looks like:
/Users/rustbuild/src/rust-buildbot/slave/nightly-dist-rustc-mac/build/build/x86_64-apple-darwin/stage1-std/x86_64-apple-darwin/release/deps/libstd-713ad88203512705.dylib
In other words, the build directory is encoded erroneously. The compiler
already [knows how] to change this directive, but it only passes that
argument when `-C rpath` is also passed. The rustbuild system, however,
explicitly [does not pass] this option explicitly and instead bakes its
own. This logic then also erroneously didn't pass `-Wl,-install_name`
like the compiler.
[knows how]: 4a008cccaa/src/librustc_trans/back/linker.rs (L210-L214)
[does not pass]: 4a008cccaa/src/bootstrap/bin/rustc.rs (L133-L158)
To fix this regression this patch introduces a new `-Z` flag, `-Z
osx-rpath-install-name` which basically just forces the compiler to take
the previous `-install_name` branch when creating a dynamic library.
Hopefully we can sort out a better rpath story in the future, but for
now this "hack" should suffice in getting our nightly builds back to the
same state as before.
Closes#38430
The standard implementations of Hasher have architecture-dependent
results when hashing integers. This causes problems when the hashes are
stored within metadata - metadata written by one host architecture can't
be read by another.
To fix that, implement an architecture-independent StableHasher and use
it in all places an architecture-independent hasher is needed.
Fixes#38177.
[LLVM 4.0] Move debuginfo alignment argument
Alignment was removed from createBasicType and moved to
- createGlobalVariable
- createAutoVariable
- createStaticMemberType (unused in Rust)
- createTempGlobalVariableFwdDecl (unused in Rust)
e69c459a6e
Alignment was removed from createBasicType and moved to
- createGlobalVariable
- createAutoVariable
- createStaticMemberType (unused in Rust)
- createTempGlobalVariableFwdDecl (unused in Rust)
e69c459a6e
Consider provided trait methods in middle::reachable
Fixes https://github.com/rust-lang/rust/issues/38226 by also considering trait methods with default implementation instead of just methods provided in an impl.
r? @alexcrichton
cc @panicbit
Don't apply msvc link opts for non-opt build
`/OPT:REF,ICF` sometimes takes lots of time. It makes no sense to apply them when doing debug build. MSVC's linker by default disables these optimizations when `/DEBUG` is specified, unless they are explicitly passed.
Add i686-unknown-openbsd target.
It is a preliminary work. I still have some tests failing, but I have a working rustc binary which is able to rebuild itself.
an update of libc should be required too, but I dunno how to do it with vendor/ layout.
r? @alexcrichton
incr.comp.: Add more output to -Z incremental-info.
Also makes sure that all output from `-Z incremental-info` is prefixed with `incremental:` for better grep-ability.
r? @nikomatsakis
Add new #[target_feature = "..."] attribute.
This commit adds a new attribute that instructs the compiler to emit
target specific code for a single function. For example, the following
function is permitted to use instructions that are part of SSE 4.2:
#[target_feature = "+sse4.2"]
fn foo() { ... }
In particular, use of this attribute does not require setting the
-C target-feature or -C target-cpu options on rustc.
This attribute does not have any protections built into it. For example,
nothing stops one from calling the above `foo` function on hosts without
SSE 4.2 support. Doing so may result in a SIGILL.
I've also expanded the x86 target feature whitelist.
In LLVM 4.0, this enum becomes an actual type-safe enum, which breaks
all of the interfaces. Introduce our own copy of the bitflags that we
can then safely convert to the LLVM one.
[9/n] rustc: move type information out of AdtDef and TraitDef.
_This is part of a series ([prev](https://github.com/rust-lang/rust/pull/37688) | [next]()) of patches designed to rework rustc into an out-of-order on-demand pipeline model for both better feature support (e.g. [MIR-based](https://github.com/solson/miri) early constant evaluation) and incremental execution of compiler passes (e.g. type-checking), with beneficial consequences to IDE support as well.
If any motivation is unclear, please ask for additional PR description clarifications or code comments._
<hr>
Both `AdtDef` and `TraitDef` contained type information (field types, generics and predicates) which was required to create them, preventing their use before that type information exists, or in the case of field types, *mutation* was required, leading to a variance-magicking implementation of `ivar`s.
This PR takes that information out and the resulting cleaner setup could even eventually end up merged with HIR, because, just like `AssociatedItem` before it, there's no dependency on types anymore.
(With one exception, variant discriminants should probably be moved into their own map later.)
[LLVM 4.0] Don't assume llvm::StringRef is null terminated
StringRefs have a length and their contents are not usually null-terminated. The solution is to either copy the string data (in `rustc_llvm::diagnostic`) or take the size into account (in LLVMRustPrintPasses).
I couldn't trigger a bug caused by this (apparently all the strings returned in practice are actually null-terminated) but this is more correct and more future-proof.
cc #37609
rustdoc: link to cross-crate sources directly.
Fixes#37684 by implementing proper support for getting the `Span` of definitions across crates.
In rustdoc this is used to generate direct links to the original source instead of fragile redirects.
This functionality could be expanded further for making error reporting code more uniform and seamless across crates, although at the moment there is no actual source to print, only file/line/column information.
Closes#37870 which is also "fixes" #37684 by throwing away the builtin macro docs from libcore.
After this lands, #37727 could be reverted, although it doesn't matter much either way.
Refactor TraitObject to Slice<ExistentialPredicate>
For reference, the primary types changes in this PR are shown below. They may add in the understanding of what is discussed below, though they should not be required.
We change `TraitObject` into a list of `ExistentialPredicate`s to allow for a couple of things:
- Principal (ExistentialPredicate::Trait) is now optional.
- Region bounds are moved out of `TraitObject` into `TyDynamic`. This permits wrapping only the `ExistentialPredicate` list in `Binder`.
- `BuiltinBounds` and `BuiltinBound` are removed entirely from the codebase, to permit future non-constrained auto traits. These are replaced with `ExistentialPredicate::AutoTrait`, which only requires a `DefId`. For the time being, only `Send` and `Sync` are supported; this constraint can be lifted in a future pull request.
- Binder-related logic is extracted from `ExistentialPredicate` into the parent (`Binder<Slice<EP>>`), so `PolyX`s are inside `TraitObject` are replaced with `X`.
The code requires a sorting order for `ExistentialPredicate`s in the interned `Slice`. The sort order is asserted to be correct during interning, but the slices are not sorted at that point.
1. `ExistentialPredicate::Trait` are defined as always equal; **This may be wrong; should we be comparing them and sorting them in some way?**
1. `ExistentialPredicate::Projection`: Compared by `ExistentialProjection::sort_key`.
1. `ExistentialPredicate::AutoTrait`: Compared by `TraitDef.def_path_hash`.
Construction of `ExistentialPredicate`s is conducted through `TyCtxt::mk_existential_predicates`, which interns a passed iterator as a `Slice`. There are no convenience functions to construct from a set of separate iterators; callers must pass an iterator chain. The lack of convenience functions is primarily due to few uses and the relative difficulty in defining a nice API due to optional parts and difficulty in recognizing which argument goes where. It is also true that the current situation isn't significantly better than 4 arguments to a constructor function; but the extra work is deemed unnecessary as of this time.
```rust
// before this PR
struct TraitObject<'tcx> {
pub principal: PolyExistentialTraitRef<'tcx>,
pub region_bound: &'tcx ty::Region,
pub builtin_bounds: BuiltinBounds,
pub projection_bounds: Vec<PolyExistentialProjection<'tcx>>,
}
// after
pub enum ExistentialPredicate<'tcx> {
// e.g. Iterator
Trait(ExistentialTraitRef<'tcx>),
// e.g. Iterator::Item = T
Projection(ExistentialProjection<'tcx>),
// e.g. Send
AutoTrait(DefId),
}
```
This commit adds a new attribute that instructs the compiler to emit
target specific code for a single function. For example, the following
function is permitted to use instructions that are part of SSE 4.2:
#[target_feature = "+sse4.2"]
fn foo() { ... }
In particular, use of this attribute does not require setting the
-C target-feature or -C target-cpu options on rustc.
This attribute does not have any protections built into it. For example,
nothing stops one from calling the above `foo` function on hosts without
SSE 4.2 support. Doing so may result in a SIGILL.
This commit also expands the target feature whitelist to include lzcnt,
popcnt and sse4a. Namely, lzcnt and popcnt have their own CPUID bits,
but were introduced with SSE4.
StringRefs have a length and their contents are not usually null-terminated.
The solution is to either copy the string data (in rustc_llvm::diagnostic) or take the size into account (in LLVMRustPrintPasses).
I couldn't trigger a bug caused by this (apparently all the strings returned in practice are actually null-terminated) but this is more correct and more future-proof.
don't double-apply variant padding to const enums
`build_const_struct` already returns the struct with padding - don't double-apply it in the `General` case.
This should hopefully be the last time we have this sort of bug.
Fixes#38002.
Beta-nominating because regression.
r? @eddyb
[LLVM 4.0] OptimizationDiagnostic FFI forward compatibility
- getMsg() changed to return std::string by-value. Fix: copy the data to a rust String during unpacking.
- getPassName() changed to return StringRef
cc #37609
Biggest change: Revised print-type-sizes output to include breakdown
of layout.
Includes info about field sizes (and alignment + padding when padding
is injected; the injected padding is derived from the offsets computed
by layout module).
Output format is illustrated in commit that has the ui tests.
Note: there exists (at least) one case of variant w/o name: empty
enums. Namely, empty enums use anonymous univariant repr. So for such
cases, print the number of the variant instead of the name.
----
Also, eddyb suggested of reading from `layout_cache` post-trans.
(For casual readers: the compiler source often uses the word "cache"
for tables that are in fact not periodically purged, and thus are
useful as the basis for data like this.)
Some types that were previously not printed are now included in the
output. (See e.g. the tests `print_type_sizes/generics.rs` and
`print_type_sizes/variants.rs`)
----
Other review feedback:
switch to an exhaustive match when filtering in just structural types.
switch to hashset for layout info and move sort into print method.
----
Driveby change: Factored session::code_stats into its own module
----
incorporate njn feedback re output formatting.
Clean up `ast::Attribute`, `ast::CrateConfig`, and string interning
This PR
- removes `ast::Attribute_` (changing `Attribute` from `Spanned<Attribute_>` to a struct),
- moves a `MetaItem`'s name from the `MetaItemKind` variants to a field of `MetaItem`,
- avoids needlessly wrapping `ast::MetaItem` with `P`,
- moves string interning into `syntax::symbol` (`ast::Name` is a reexport of `symbol::Symbol` for now),
- replaces `InternedString` with `Symbol` in the AST, HIR, and various other places, and
- refactors `ast::CrateConfig` from a `Vec` to a `HashSet`.
r? @eddyb
[LLVM 4.0] Set EH personality when resuming stack unwinding
To resume stack unwinding, the LLVM `resume` instruction must be used.
In order to use this instruction, the calling function must have an
exception handling personality set.
LLVM 4.0 adds a new IR validation check to ensure a personality is
always set in these cases.
This was introduced in [r277360](https://reviews.llvm.org/rL277360).
[LLVM 4.0] Use llvm::Attribute APIs instead of "raw value" APIs
The latter will be removed in LLVM 4.0 (see 4a6fc8bacf).
The librustc_llvm API remains mostly unchanged, except that llvm::Attribute is no longer a bitflag but represents only a *single* attribute.
The ability to store many attributes in a small number of bits and modify them without interacting with LLVM is only used in rustc_trans::abi and closely related modules, and only attributes for function arguments are considered there.
Thus rustc_trans::abi now has its own bit-packed representation of argument attributes, which are translated to rustc_llvm::Attribute when applying the attributes.
cc #37609
fix `extern "aapcs" fn`
to actually use the AAPCS calling convention
closes#37810
This is technically a [breaking-change] because it changes the ABI of
`extern "aapcs"` functions that (a) involve `f32`/`f64` arguments/return
values and (b) are compiled for arm-eabihf targets from
"aapcs-vfp" (wrong) to "aapcs" (correct).
Appendix:
What these ABIs mean?
- In the "aapcs-vfp" ABI or "hard float" calling convention: Floating
point values are passed/returned through FPU registers (s0, s1, d0, etc.)
- Whereas, in the "aapcs" ABI or "soft float" calling convention:
Floating point values are passed/returned through general purpose
registers (r0, r1, etc.)
Mixing these ABIs can cause problems if the caller assumes that the
routine is using one of these ABIs but it's actually using the other
one.
---
r? @alexcrichton We are going this `extern "aapcs" fn` thing to implement some intrinsics (floatundidf) for the eabihf targets in order to comply with LLVM's calling convention of intrinsics.
Oh, and the value of the enum came from [here](http://llvm.org/docs/doxygen/html/namespacellvm_1_1CallingConv.html).
cc @TimNN @parched
To resume stack unwinding, the LLVM `resume` instruction must be used.
In order to use this instruction, the calling function must have an
exception handling personality set.
LLVM 4.0 adds a new IR validation check to ensure a personality is
always set in these cases.
This was introduced in [r277360](https://reviews.llvm.org/rL277360).
Separate impl items from the parent impl
This change separates impl item bodies out of the impl itself. This gives incremental more resolution. In so doing, it refactors how the visitors work, and cleans up a bit of the collect/check logic (mostly by moving things out of collect that didn't really belong there, because they were just checking conditions).
However, this is not as effective as I expected, for a kind of frustrating reason. In particular, when invoking `foo.bar()` you still wind up with dependencies on private items. The problem is that the method resolution code scans that list for methods with the name `bar` -- and this winds up touching *all* the methods, even private ones.
I can imagine two obvious ways to fix this:
- separating fn bodies from fn sigs (#35078, currently being pursued by @flodiebold)
- a more aggressive model of incremental that @michaelwoerister has been advocating, in which we hash the intermediate results (e.g., the outputs of collect) so that we can see that the intermediate result hasn't changed, even if a particular impl item has changed.
So all in all I'm not quite sure whether to land this or not. =) It still seems like it has to be a win in some cases, but not with the test cases we have just now. I can try to gin up some test cases, but I'm not sure if they will be totally realistic. On the other hand, some of the early refactorings to the visitor trait seem worthwhile to me regardless.
cc #36349 -- well, this is basically a fix for that issue, I guess
r? @michaelwoerister
NB: Based atop of @eddyb's PR https://github.com/rust-lang/rust/pull/37402; don't land until that lands.
The librustc_llvm API remains mostly unchanged, except that llvm::Attribute is no longer a bitflag but represents only a *single* attribute.
The ability to store many attributes in a small number of bits and modify them without interacting with LLVM is only used in rustc_trans::abi and closely related modules, and only attributes for function arguments are considered there.
Thus rustc_trans::abi now has its own bit-packed representation of argument attributes, which are translated to rustc_llvm::Attribute when applying the attributes.
to actually use the AAPCS calling convention
closes#37810
This is technically a [breaking-change] because it changes the ABI of
`extern "aapcs"` functions that (a) involve `f32`/`f64` arguments/return
values and (b) are compiled for arm-eabihf targets from
"aapcs-vfp" (wrong) to "aapcs" (correct).
Appendix:
What these ABIs mean?
- In the "aapcs-vfp" ABI or "hard float" calling convention: Floating
point values are passed/returned through FPU registers (s0, s1, d0, etc.)
- Whereas, in the "aapcs" ABI or "soft float" calling convention:
Floating point values are passed/returned through general purpose
registers (r0, r1, etc.)
Mixing these ABIs can cause problems if the caller assumes that the
routine is using one of these ABIs but it's actually using the other
one.
