This commit is an implementation of LLVM's ThinLTO for consumption in rustc
itself. Currently today LTO works by merging all relevant LLVM modules into one
and then running optimization passes. "Thin" LTO operates differently by having
more sharded work and allowing parallelism opportunities between optimizing
codegen units. Further down the road Thin LTO also allows *incremental* LTO
which should enable even faster release builds without compromising on the
performance we have today.
This commit uses a `-Z thinlto` flag to gate whether ThinLTO is enabled. It then
also implements two forms of ThinLTO:
* In one mode we'll *only* perform ThinLTO over the codegen units produced in a
single compilation. That is, we won't load upstream rlibs, but we'll instead
just perform ThinLTO amongst all codegen units produced by the compiler for
the local crate. This is intended to emulate a desired end point where we have
codegen units turned on by default for all crates and ThinLTO allows us to do
this without performance loss.
* In anther mode, like full LTO today, we'll optimize all upstream dependencies
in "thin" mode. Unlike today, however, this LTO step is fully parallelized so
should finish much more quickly.
There's a good bit of comments about what the implementation is doing and where
it came from, but the tl;dr; is that currently most of the support here is
copied from upstream LLVM. This code duplication is done for a number of
reasons:
* Controlling parallelism means we can use the existing jobserver support to
avoid overloading machines.
* We will likely want a slightly different form of incremental caching which
integrates with our own incremental strategy, but this is yet to be
determined.
* This buys us some flexibility about when/where we run ThinLTO, as well as
having it tailored to fit our needs for the time being.
* Finally this allows us to reuse some artifacts such as our `TargetMachine`
creation, where all our options we used today aren't necessarily supported by
upstream LLVM yet.
My hope is that we can get some experience with this copy/paste in tree and then
eventually upstream some work to LLVM itself to avoid the duplication while
still ensuring our needs are met. Otherwise I fear that maintaining these
bindings may be quite costly over the years with LLVM updates!
Commit c4710203c0 in #43492 make `LLVMRustHasFeature` "more robust"
by using `getFeatureTable()`. However, this function is specific to
Rust's own LLVM fork, not upstream LLVM-4.0, so we need to use
`#if LLVM_RUSTLLVM` to guard this call.
The function should accept feature strings that old LLVM might not
support.
Simplify the code using the same approach used by
LLVMRustPrintTargetFeatures.
Dummify the function for non 4.0 LLVM and update the tests accordingly.
Add RWPI/ROPI relocation model support
This PR adds support for using LLVM 4's ROPI and RWPI relocation models for ARM.
ROPI (Read-Only Position Independence) and RWPI (Read-Write Position Independence) are two new relocation models in LLVM for the ARM backend ([LLVM changset](https://reviews.llvm.org/rL278015)). The motivation is that these are the specific strategies we use in userspace [Tock](https://www.tockos.org) apps, so supporting this is an important step (perhaps the final step, but can't confirm yet) in enabling userspace Rust processes.
## Explanation
ROPI makes all code and immutable accesses PC relative, but not assumed to be overriden at runtime (so for example, jumps are always relative).
RWPI uses a base register (`r9`) that stores the addresses of the GOT in memory so the runtime (e.g. a kernel) only adjusts r9 tell running code where the GOT is.
## Complications adding support in Rust
While this landed in LLVM master back in August, the header files in `llvm-c` have not been updated yet to reflect it. Rust replicates that header file's version of the `LLVMRelocMode` enum as the Rust enum `llvm::RelocMode` and uses an implicit cast in the ffi to translate from Rust's notion of the relocation model to the LLVM library's notion.
My workaround for this currently is to replace the `LLVMRelocMode` argument to `LLVMTargetMachineRef` with an int and using the hardcoded int representation of the `RelocMode` enum. This is A Bad Idea(tm), but I think very nearly the right thing.
Would a better alternative be to patch rust-llvm to support these enum variants (also a fairly trivial change)?
Replaces the llvm-c exposed LLVMRelocMode, which does not include all
relocation model variants, with a LLVMRustRelocMode modeled after
LLVMRustCodeMode.
Improve naming style in rustllvm.
As per the LLVM style guide, use CamelCase for all locals and classes,
and camelCase for all non-FFI functions.
Also, make names of variables of commonly used types more consistent.
Fixes#38688.
r? @rkruppe
As per the LLVM style guide, use CamelCase for all locals and classes,
and camelCase for all non-FFI functions.
Also, make names of variables of commonly used types more consistent.
Fixes#38688.
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.
Previously the C type LLVMRelocMode (available as RelocMode in Rust)
was passed as is to the function.
However createTargetMachine expects a Reloc::Model, which is an enum
just one value short.
Additionally, the function was marked as requiring Reloc::Model in the
C code, but RelocMode on the Rust-side.
We now use the correct C type LLVMRelocMode and convert it to an
Optional<Reloc::Model> as expected by the createTargetMachine call the
same the original LLVMCreateTargetMachine function does.
See
c9b262bfbd/lib/Target/TargetMachineC.cpp (L104-L121)
This was found by @eddyb.
LLVM was upgraded to a new version in this commit:
f9d4149c29
which was part of this pull request:
https://github.com/rust-lang/rust/issues/26025
Consider the following two lines from that commit:
f9d4149c29 (diff-a3b24dbe2ea7c1981f9ac79f9745f40aL462)f9d4149c29 (diff-a3b24dbe2ea7c1981f9ac79f9745f40aL469)
The purpose of these lines is to register LLVM passes. Prior to the that
commit, the passes being handled were assumed to be ModulePasses (a
specific type of LLVM pass) since they were being added to a ModulePass
manager. After that commit, both lines were refactored (presumably in an
attempt to DRY out the code), but the ModulePasses were changed to be
registered to a FunctionPass manager. This change resulted in
ModulePasses being run, but a Function object was being passed as a
parameter to the pass instead of a Module, which resulted in
segmentation faults.
In this commit, I changed relevant sections of the code to check the
type of the passes being added and register them to the appropriate pass
manager.
Closes https://github.com/rust-lang/rust/issues/31067
Turns out for OSX our data layout was subtly wrong and the LLVM update must have
exposed this. Instead of fixing this I've removed all data layouts from the
compiler to just use the defaults that LLVM provides for all targets. All data
layouts (and a number of dead modules) are removed from the compiler here.
Custom target specifications can still provide a custom data layout, but it is
now an optional key as the default will be used if one isn't specified.
This commit updates the LLVM submodule in use to the current HEAD of the LLVM
repository. This is primarily being done to start picking up unwinding support
for MSVC, which is currently unimplemented in the revision of LLVM we are using.
Along the way a few changes had to be made:
* As usual, lots of C++ debuginfo bindings in LLVM changed, so there were some
significant changes to our RustWrapper.cpp
* As usual, some pass management changed in LLVM, so clang was re-scrutinized to
ensure that we're doing the same thing as clang.
* Some optimization options are now passed directly into the
`PassManagerBuilder` instead of through CLI switches to LLVM.
* The `NoFramePointerElim` option was removed from LLVM, favoring instead the
`no-frame-pointer-elim` function attribute instead.
Additionally, LLVM has picked up some new optimizations which required fixing an
existing soundness hole in the IR we generate. It appears that the current LLVM
we use does not expose this hole. When an enum is moved, the previous slot in
memory is overwritten with a bit pattern corresponding to "dropped". When the
drop glue for this slot is run, however, the switch on the discriminant can
often start executing the `unreachable` block of the switch due to the
discriminant now being outside the normal range. This was patched over locally
for now by having the `unreachable` block just change to a `ret void`.
Position independent code has fewer requirements in executables, so pass
the appropriate flag to LLVM in order to allow more optimization. At the
moment this means faster thread-local storage.
The core library in theory has 0 dependencies, but in practice it has some in
order for it to be efficient. These dependencies are in the form of the basic
memory operations provided by libc traditionally, such as memset, memcmp, etc.
These functions are trivial to implement and themselves have 0 dependencies.
This commit adds a new crate, librlibc, which will serve the purpose of
providing these dependencies. The crate is never linked to by default, but is
available to be linked to by downstream consumers. Normally these functions are
provided by the system libc, but in other freestanding contexts a libc may not
be available. In these cases, librlibc will suffice for enabling execution with
libcore.
cc #10116
The compiler has previously been producing binaries on the order of 1.8MB for
hello world programs "fn main() {}". This is largely a result of the compilation
model used by compiling entire libraries into a single object file and because
static linking is favored by default.
When linking, linkers will pull in the entire contents of an object file if any
symbol from the object file is used. This means that if any symbol from a rust
library is used, the entire library is pulled in unconditionally, regardless of
whether the library is used or not.
Traditional C/C++ projects do not normally encounter these large executable
problems because their archives (rust's rlibs) are composed of many objects.
Because of this, linkers can eliminate entire objects from being in the final
executable. With rustc, however, the linker does not have the opportunity to
leave out entire object files.
In order to get similar benefits from dead code stripping at link time, this
commit enables the -ffunction-sections and -fdata-sections flags in LLVM, as
well as passing --gc-sections to the linker *by default*. This means that each
function and each global will be placed into its own section, allowing the
linker to GC all unused functions and data symbols.
By enabling these flags, rust is able to generate much smaller binaries default.
On linux, a hello world binary went from 1.8MB to 597K (a 67% reduction in
size). The output size of dynamic libraries remained constant, but the output
size of rlibs increased, as seen below:
libarena - 2.27% bigger ( 292872 => 299508)
libcollections - 0.64% bigger ( 6765884 => 6809076)
libflate - 0.83% bigger ( 186516 => 188060)
libfourcc - 14.71% bigger ( 307290 => 352498)
libgetopts - 4.42% bigger ( 761468 => 795102)
libglob - 2.73% bigger ( 899932 => 924542)
libgreen - 9.63% bigger ( 1281718 => 1405124)
libhexfloat - 13.88% bigger ( 333738 => 380060)
liblibc - 10.79% bigger ( 551280 => 610736)
liblog - 10.93% bigger ( 218208 => 242060)
libnative - 8.26% bigger ( 1362096 => 1474658)
libnum - 2.34% bigger ( 2583400 => 2643916)
librand - 1.72% bigger ( 1608684 => 1636394)
libregex - 6.50% bigger ( 1747768 => 1861398)
librustc - 4.21% bigger (151820192 => 158218924)
librustdoc - 8.96% bigger ( 13142604 => 14320544)
librustuv - 4.13% bigger ( 4366896 => 4547304)
libsemver - 2.66% bigger ( 396166 => 406686)
libserialize - 1.91% bigger ( 6878396 => 7009822)
libstd - 3.59% bigger ( 39485286 => 40902218)
libsync - 3.95% bigger ( 1386390 => 1441204)
libsyntax - 4.96% bigger ( 35757202 => 37530798)
libterm - 13.99% bigger ( 924580 => 1053902)
libtest - 6.04% bigger ( 2455720 => 2604092)
libtime - 2.84% bigger ( 1075708 => 1106242)
liburl - 6.53% bigger ( 590458 => 629004)
libuuid - 4.63% bigger ( 326350 => 341466)
libworkcache - 8.45% bigger ( 1230702 => 1334750)
This increase in size is a result of encoding many more section names into each
object file (rlib). These increases are moderate enough that this change seems
worthwhile to me, due to the drastic improvements seen in the final artifacts.
The overall increase of the stage2 target folder (not the size of an install)
went from 337MB to 348MB (3% increase).
Additionally, linking is generally slower when executed with all these new
sections plus the --gc-sections flag. The stage0 compiler takes 1.4s to link the
`rustc` binary, where the stage1 compiler takes 1.9s to link the binary. Three
megabytes are shaved off the binary. I found this increase in link time to be
acceptable relative to the benefits of code size gained.
This commit only enables --gc-sections for *executables*, not dynamic libraries.
LLVM does all the heavy lifting when producing an object file for a dynamic
library, so there is little else for the linker to do (remember that we only
have one object file).
I conducted similar experiments by putting a *module's* functions and data
symbols into its own section (granularity moved to a module level instead of a
function/static level). The size benefits of a hello world were seen to be on
the order of 400K rather than 1.2MB. It seemed that enough benefit was gained
using ffunction-sections that this route was less desirable, despite the lesser
increases in binary rlib size.
