We were previously reading metadata via `ar p`, but as learned from rustdoc
awhile back, spawning a process to do something is pretty slow. Turns out LLVM
has an Archive class to read archives, but it cannot write archives.
This commits adds bindings to the read-only version of the LLVM archive class
(with a new type that only has a read() method), and then it uses this class
when reading the metadata out of rlibs. When you put this in tandem of not
compressing the metadata, reading the metadata is 4x faster than it used to be
The timings I got for reading metadata from the respective libraries was:
libstd-04ff901e-0.9-pre.dylib => 100ms
libstd-04ff901e-0.9-pre.rlib => 23ms
librustuv-7945354c-0.9-pre.dylib => 4ms
librustuv-7945354c-0.9-pre.rlib => 1ms
librustc-5b94a16f-0.9-pre.dylib => 87ms
librustc-5b94a16f-0.9-pre.rlib => 35ms
libextra-a6ebb16f-0.9-pre.dylib => 63ms
libextra-a6ebb16f-0.9-pre.rlib => 15ms
libsyntax-2e4c0458-0.9-pre.dylib => 86ms
libsyntax-2e4c0458-0.9-pre.rlib => 22ms
In order to always take advantage of these faster metadata read-times, I sort
the files in filesearch based on whether they have an rlib extension or not
(prefer all rlib files first).
Overall, this halved the compile time for a `fn main() {}` crate from 0.185s to
0.095s on my system (when preferring dynamic linking). Reading metadata is still
the slowest pass of the compiler at 0.035s, but it's getting pretty close to
linking at 0.021s! The next best optimization is to just not copy the metadata
from LLVM because that's the most expensive part of reading metadata right now.
llvm supports both win32 native threads and pthread,
but configure tries to find pthread first.
This manually disables pthread to use native api.
This removes libpthreads-2.dll dependency on librustc.
When performing LTO, the rust compiler has an opportunity to completely strip
all landing pads in all dependent libraries. I've modified the LTO pass to
recognize the -Z no-landing-pads option when also running an LTO pass to flag
everything in LLVM as nothrow. I've verified that this prevents any and all
invoke instructions from being emitted.
I believe that this is one of our best options for moving forward with
accomodating use-cases where unwinding doesn't really make sense. This will
allow libraries to be built with landing pads by default but allow usage of them
in contexts where landing pads aren't necessary.
cc #10780
This commit implements LTO for rust leveraging LLVM's passes. What this means
is:
* When compiling an rlib, in addition to insdering foo.o into the archive, also
insert foo.bc (the LLVM bytecode) of the optimized module.
* When the compiler detects the -Z lto option, it will attempt to perform LTO on
a staticlib or binary output. The compiler will emit an error if a dylib or
rlib output is being generated.
* The actual act of performing LTO is as follows:
1. Force all upstream libraries to have an rlib version available.
2. Load the bytecode of each upstream library from the rlib.
3. Link all this bytecode into the current LLVM module (just using llvm
apis)
4. Run an internalization pass which internalizes all symbols except those
found reachable for the local crate of compilation.
5. Run the LLVM LTO pass manager over this entire module
6a. If assembling an archive, then add all upstream rlibs into the output
archive. This ignores all of the object/bitcode/metadata files rust
generated and placed inside the rlibs.
6b. If linking a binary, create copies of all upstream rlibs, remove the
rust-generated object-file, and then link everything as usual.
As I have explained in #10741, this process is excruciatingly slow, so this is
*not* turned on by default, and it is also why I have decided to hide it behind
a -Z flag for now. The good news is that the binary sizes are about as small as
they can be as a result of LTO, so it's definitely working.
Closes#10741Closes#10740
The main one removed is rust_upcall_reset_stack_limit (continuation of #10156),
and this also removes the upcall_trace function. The was hidden behind a
`-Z trace` flag, but if you attempt to use this now you'll get a linker error
because there is no implementation of the 'upcall_trace' function. Due to this
no longer working, I decided to remove it entirely from the compiler (I'm also a
little unsure on what it did in the first place).
LLVM's JIT has been updated numerous times, and we haven't been tracking it at
all. The existing LLVM glue code no longer compiles, and the JIT isn't used for
anything currently.
This also rebases out the FixedStackSegment support which we have added to LLVM.
None of this is still in use by the compiler, and there's no need to keep this
functionality around inside of LLVM.
This is needed to unblock #10708 (where we're tripping an LLVM assertion).
This commit implements the support necessary for generating both intermediate
and result static rust libraries. This is an implementation of my thoughts in
https://mail.mozilla.org/pipermail/rust-dev/2013-November/006686.html.
When compiling a library, we still retain the "lib" option, although now there
are "rlib", "staticlib", and "dylib" as options for crate_type (and these are
stackable). The idea of "lib" is to generate the "compiler default" instead of
having too choose (although all are interchangeable). For now I have left the
"complier default" to be a dynamic library for size reasons.
Of the rust libraries, lib{std,extra,rustuv} will bootstrap with an
rlib/dylib pair, but lib{rustc,syntax,rustdoc,rustpkg} will only be built as a
dynamic object. I chose this for size reasons, but also because you're probably
not going to be embedding the rustc compiler anywhere any time soon.
Other than the options outlined above, there are a few defaults/preferences that
are now opinionated in the compiler:
* If both a .dylib and .rlib are found for a rust library, the compiler will
prefer the .rlib variant. This is overridable via the -Z prefer-dynamic option
* If generating a "lib", the compiler will generate a dynamic library. This is
overridable by explicitly saying what flavor you'd like (rlib, staticlib,
dylib).
* If no options are passed to the command line, and no crate_type is found in
the destination crate, then an executable is generated
With this change, you can successfully build a rust program with 0 dynamic
dependencies on rust libraries. There is still a dynamic dependency on
librustrt, but I plan on removing that in a subsequent commit.
This change includes no tests just yet. Our current testing
infrastructure/harnesses aren't very amenable to doing flavorful things with
linking, so I'm planning on adding a new mode of testing which I believe belongs
as a separate commit.
Closes#552
Example:
void ({ i64, %tydesc*, i8*, i8*, i8 }*, i64*, %"struct.std::fmt::Formatter[#1]"*)*
Before, we would print 20 levels deep due to recursion in the type
definition.
This change adds -Z soft-float option for generating
software floating point library calls.
It also implies using soft float ABI, that is the same as llc.
It is useful for targets that have no FPU.
Also fixed nasty bug caused by calling LLVMDIBuilderCreateStructType() with a null pointer where an empty array was expected (which would trigger an unintelligable assertion somewhere down the line).
The only changes to the default passes is that O1 now doesn't run the inline
pass, just always-inline with lifetime intrinsics. O2 also now has a threshold
of 225 instead of 275. Otherwise the default passes being run is the same.
I've also added a few more options for configuring the pass pipeline. Namely you
can now specify arguments to LLVM directly via the `--llvm-args` command line
option which operates similarly to `--passes`. I also added the ability to turn
off pre-population of the pass manager in case you want to run *only* your own
passes.
Beforehand, it was unclear whether rust was performing the "recommended set" of
optimizations provided by LLVM for code. This commit changes the way we run
passes to closely mirror that of clang, which in theory does it correctly. The
notable changes include:
* Passes are no longer explicitly added one by one. This would be difficult to
keep up with as LLVM changes and we don't guaranteed always know the best
order in which to run passes
* Passes are now managed by LLVM's PassManagerBuilder object. This is then used
to populate the various pass managers run.
* We now run both a FunctionPassManager and a module-wide PassManager. This is
what clang does, and I presume that we *may* see a speed boost from the
module-wide passes just having to do less work. I have no measured this.
* The codegen pass manager has been extracted to its own separate pass manager
to not get mixed up with the other passes
* All pass managers now include passes for target-specific data layout and
analysis passes
Some new features include:
* You can now print all passes being run with `-Z print-llvm-passes`
* When specifying passes via `--passes`, the passes are now appended to the
default list of passes instead of overwriting them.
* The output of `--passes list` is now generated by LLVM instead of maintaining
a list of passes ourselves
* Loop vectorization is turned on by default as an optimization pass and can be
disabled with `-Z no-vectorize-loops`
We currently have no need for the frame pointers on any platform. They
may eventually be needed on platforms without an equivalent to the DWARF
call frame information to walk the stack in the garbage collector.
Closes#7477
The first commit message is pretty good, but whomever reviews this should probably also at least glance at the changes I made in LLVM. I basically reorganized our pending patch queue to be a bit more organized and clearer in what needs to go where. After this, our queue would be:
* Add the `no-split-stack` attribute
* Add the `fixedstacksegment` attribute
* Add split-stacks for arm android
* Add split-stacks for arm linux
* Add split stacks for mips
Then there's a patch which I added to get rust to build at all on LLVM-head, and I'm not quite sure why it's there, but nothing seems to be crashing for now! (famous last words).
Otherwise, I just updated code to reflect the changes I made in LLVM with the only major change being the advent of the new `no_split_stack` attribute. This is work towards #1226, but someone more familiar with the code should probably actually assign the attribute to the appropriate functions.
Also as a bonus, I've verified that this closes#5774
* This has one workaround patch (everything's testing just fine...)
* I reworked the fixedstacksegment attribute to be specified with a string
rather than using a keyword and an integer and modifying the parser
* I added a "no-split-stack" attribute along the same lines as the
"fixedstacksegment" attribute for #1226
Adds `--target-cpu` flag which lets you choose a more specific target cpu instead of just passing the default, `generic`. It's more or less akin to `-mcpu`/`-mtune` in clang/gcc.
This can be applied to statics and it will indicate that LLVM will attempt to
merge the constant in .data with other statics.
I have preliminarily applied this to all of the statics generated by the new
`ifmt!` syntax extension. I compiled a file with 1000 calls to `ifmt!` and a
separate file with 1000 calls to `fmt!` to compare the sizes, and the results
were:
fmt 310k
ifmt (before) 529k
ifmt (after) 202k
This now means that ifmt! is both faster and smaller than fmt!, yay!
* LLVM now has a C interface to LLVMBuildAtomicRMW
* The exception handling support for the JIT seems to have been dropped
* Various interfaces have been added or headers have changed
These blocks were required because previously we could only insert
instructions at the end of blocks, but we wanted to have all allocas in
one place, so they can be collapse. But now we have "direct" access the
the LLVM IR builder and can position it freely. This allows us to use
the same trick that clang uses, which means that we insert a dummy
"marker" instruction to identify the spot at which we want to insert
allocas. We can then later position the IR builder at that spot and
insert the alloca instruction, without any dedicated block.
The block for loading the closure environment can now also go away,
because the function context now provides the toplevel block, and the
translation of the loading happens first, so that's good enough.
Makes the LLVM IR a bit more readable, saving a bunch of branches in the
unoptimized code, which benefits unoptimized builds.
