Refactor away RBML from rustc_metadata.
RBML and `ty{en,de}code` have had their long-overdue purge. Summary of changes:
* Metadata is now a tree encoded in post-order and with relative backward references pointing to children nodes. With auto-deriving and type safety, this makes maintenance and adding new information to metadata painless and bug-free by default. It's also more compact and cache-friendly (cache misses should be proportional to the depth of the node being accessed, not the number of siblings as in EBML/RBML).
* Metadata sizes have been reduced, for `libcore` it went down 16% (`8.38MB` -> `7.05MB`) and for `libstd` 14% (`3.53MB` -> `3.03MB`), while encoding more or less the same information
* Specialization is used in the bundled `libserialize` (crates.io `rustc_serialize` remains unaffected) to customize the encoding (and more importantly, decoding) of various types, most notably those interned in the `TyCtxt`. Some of this abuses a soundness hole pending a fix (cc @aturon), but when that fix arrives, we'll move to macros 1.1 `#[derive]` and custom `TyCtxt`-aware serialization traits.
* Enumerating children of modules from other crates is now orthogonal to describing those items via `Def` - this is a step towards bridging crate-local HIR and cross-crate metadata
* `CrateNum` has been moved to `rustc` and both it and `NodeId` are now newtypes instead of `u32` aliases, for specializing their decoding. This is `[syntax-breaking]` (cc @Manishearth ).
cc @rust-lang/compiler
trans: Only instantiate #[inline] functions in codegen units referencing them
This PR changes how `#[inline]` functions are translated. Before, there was one "master instance" of the function with `external` linkage and a number of on-demand instances with `available_externally` linkage in each codegen unit that referenced the function. This had two downsides:
* Public functions marked with `#[inline]` would be present in machine code of libraries unnecessarily (see #36280 for an example)
* LLVM would crash on `i686-pc-windows-msvc` due to what I suspect to be a bug in LLVM's Win32 exception handling code, because it doesn't like `available_externally` there (#36309).
This PR changes the behavior, so that there is no master instance and only on-demand instances with `internal` linkage. The downside of this is potential code-bloat if LLVM does not completely inline away the `internal` instances because then there'd be N instances of the function instead of 1. However, this can only become a problem when using more than one codegen unit per crate.
cc @rust-lang/compiler
trans: Let the collector find drop-glue for all vtables, not just VTableImpl.
This fixes#36260. So far, the collector has only recorded drop-glue for insertion into a vtable if the vtable was for an impl. But there's actually no reason why it shouldn't do just the same for closure vtables, afaict.
r? @eddyb
cc @rust-lang/compiler
invoke drop glue with a ptr to (data, meta)
This is done by creating a little space on the stack. Hokey, but it's the simplest fix I can see, and I am in "kill regressions" mode right now.
Fixes#35546
r? @eddyb
crate-ify compiler-rt into compiler-builtins
libcompiler-rt.a is dead, long live libcompiler-builtins.rlib
This commit moves the logic that used to build libcompiler-rt.a into a
compiler-builtins crate on top of the core crate and below the std crate.
This new crate still compiles the compiler-rt instrinsics using gcc-rs
but produces an .rlib instead of a static library.
Also, with this commit rustc no longer passes -lcompiler-rt to the
linker. This effectively makes the "no-compiler-rt" field of target
specifications a no-op. Users of `no_std` will have to explicitly add
the compiler-builtins crate to their crate dependency graph *if* they
need the compiler-rt intrinsics - this is a [breaking-change]. Users
of the `std` have to do nothing extra as the std crate depends
on compiler-builtins.
Finally, this a step towards lazy compilation of std with Cargo as the
compiler-rt intrinsics can now be built by Cargo instead of having to
be supplied by the user by some other method.
closes#34400
---
r? @alexcrichton
core: add likely and unlikely intrinsics
I'm no good at reading assembly, but I have tried a stage1 compiler with this patch, and it does cause different asm output. Additionally, testing this compiler on my httparse crate with some `likely` usage added in to the branches does affect benchmarks. However, I'm sure a codegen test should be included, if anyone knows what it should look like.
There isn't an entry in `librustc_trans/context.rs` in this diff, because it already exists (`llvm.expect.i1` is used for array indices).
----
Even though this does affect httparse benchmarks, it doesn't seem to affect it the same way GCC's `__builtin_expect` affects picohttpparser. I was confused that the deviation on the benchmarks grew hugely when testing this, especially since I'm absolutely certain that the branchs where I added `likely` were always `true`. I chalk that up to GCC and LLVM handle branch prediction differently.
cc #26179
libcompiler-rt.a is dead, long live libcompiler-builtins.rlib
This commit moves the logic that used to build libcompiler-rt.a into a
compiler-builtins crate on top of the core crate and below the std crate.
This new crate still compiles the compiler-rt instrinsics using gcc-rs
but produces an .rlib instead of a static library.
Also, with this commit rustc no longer passes -lcompiler-rt to the
linker. This effectively makes the "no-compiler-rt" field of target
specifications a no-op. Users of `no_std` will have to explicitly add
the compiler-builtins crate to their crate dependency graph *if* they
need the compiler-rt intrinsics. Users of the `std` have to do nothing
extra as the std crate depends on compiler-builtins.
Finally, this a step towards lazy compilation of std with Cargo as the
compiler-rt intrinsics can now be built by Cargo instead of having to
be supplied by the user by some other method.
closes#34400
Add s390x support
This adds support for building the Rust compiler and standard
library for s390x-linux, allowing a full cross-bootstrap sequence
to complete. This includes:
- Makefile/configure changes to allow native s390x builds
- Full Rust compiler support for the s390x C ABI
(only the non-vector ABI is supported at this point)
- Port of the standard library to s390x
- Update the liblibc submodule to a version including s390x support
- Testsuite fixes to allow clean "make check" on s390x
Caveats:
- Resets base cpu to "z10" to bring support in sync with the default
behaviour of other compilers on the platforms. (Usually, upstream
supports all older processors; a distribution build may then chose
to require a more recent base version.) (Also, using zEC12 causes
failures in the valgrind tests since valgrind doesn't fully support
this CPU yet.)
- z13 vector ABI is not yet supported. To ensure compatible code
generation, the -vector feature is passed to LLVM. Note that this
means that even when compiling for z13, no vector instructions
will be used. In the future, support for the vector ABI should be
added (this will require common code support for different ABIs
that need different data_layout strings on the same platform).
- Two test cases are (temporarily) ignored on s390x to allow passing
the test suite. The underlying issues still need to be fixed:
* debuginfo/simd.rs fails because of incorrect debug information.
This seems to be a LLVM bug (also seen with C code).
* run-pass/union/union-basic.rs simply seems to be incorrect for
all big-endian platforms.
Signed-off-by: Ulrich Weigand <ulrich.weigand@de.ibm.com>
This adds support for building the Rust compiler and standard
library for s390x-linux, allowing a full cross-bootstrap sequence
to complete. This includes:
- Makefile/configure changes to allow native s390x builds
- Full Rust compiler support for the s390x C ABI
(only the non-vector ABI is supported at this point)
- Port of the standard library to s390x
- Update the liblibc submodule to a version including s390x support
- Testsuite fixes to allow clean "make check" on s390x
Caveats:
- Resets base cpu to "z10" to bring support in sync with the default
behaviour of other compilers on the platforms. (Usually, upstream
supports all older processors; a distribution build may then chose
to require a more recent base version.) (Also, using zEC12 causes
failures in the valgrind tests since valgrind doesn't fully support
this CPU yet.)
- z13 vector ABI is not yet supported. To ensure compatible code
generation, the -vector feature is passed to LLVM. Note that this
means that even when compiling for z13, no vector instructions
will be used. In the future, support for the vector ABI should be
added (this will require common code support for different ABIs
that need different data_layout strings on the same platform).
- Two test cases are (temporarily) ignored on s390x to allow passing
the test suite. The underlying issues still need to be fixed:
* debuginfo/simd.rs fails because of incorrect debug information.
This seems to be a LLVM bug (also seen with C code).
* run-pass/union/union-basic.rs simply seems to be incorrect for
all big-endian platforms.
Signed-off-by: Ulrich Weigand <ulrich.weigand@de.ibm.com>
While attempting to port Rust to s390x, I ran into an ABI violation
(that caused rust_eh_personality to be miscompiled, breaking unwinding).
The problem is that this function returns an enum type, which is
supposed to be sign-extended according to the s390x ABI. However,
common code would ignore target sign-/zero-extension rules for any
types that do not satisfy is_integral(), which includes enums.
For the general case of Rust enum types, which map to structure types
with a discriminant, that seems correct. However, in the special case
of simple enums that map directly to C enum types (i.e. LLVM integers),
this is incorrect; we must follow the target extension rules for those.
Signed-off-by: Ulrich Weigand <ulrich.weigand@de.ibm.com>
incr. comp.: Take spans into account for ICH
This PR makes the ICH (incr. comp. hash) take spans into account when debuginfo is enabled.
A side-effect of this is that the SVH (which is based on the ICHs of all items in the crate) becomes sensitive to the tiniest change in a code base if debuginfo is enabled. Since we are not trying to model ABI compatibility via the SVH anymore (this is done via the crate disambiguator now), this should be not be a problem.
Fixes#33888.
Fixes#32753.
Fix incorrect LLVM Linkage enum
Followup of #33994 to actually work.
The `Linkage` enum in librustc_llvm got out of sync with the version in LLVM and it caused two variants of the `#[linkage=""]` attribute to break.
This adds the functions `LLVMRustGetLinkage` and `LLVMRustSetLinkage` which convert between the Rust Linkage enum and the LLVM one, which should stop this from breaking every time LLVM changes it.
Possible remaining concerns:
1. There could be a codegen test to make sure that the attributes are applied correctly (I don't know how to do this).
2. ~~The test does not exercise the `appending` linkage. I can't figure out how to make a global static raw pointer to an array. This might not even be possible? If not we should probably remove appending linkage as its unusable in rust.~~ Appending linkage is not 'emittable' anyway.
3. The test only runs on Linux.
Fixes#33992
r? @alexcrichton
The `Linkage` enum in librustc_llvm got out of sync with the version in LLVM and it caused two variants of the #[linkage=""] attribute to break.
This adds the functions `LLVMRustGetLinkage` and `LLVMRustSetLinkage` which convert between the Rust Linkage enum and the LLVM one, which should stop this from breaking every time LLVM changes it.
Fixes#33992
Fix optimization regressions for operations on [x; n]-initialized arrays.
Fixes#35662 by using `!=` instead of `<` as the stop condition for `[x; n]` initialization loops.
Also included is cc2009f02d, a hack to run the GVN pass twice, another time after InstCombine.
This hack results in removal of redundant `memset` and `memcpy` calls (from loops over arrays).
cc @nrc Can we get performance numbers on this? Not sure if it regresses anything else.
