Exception for specific cases like linting, additional passes should
be going into the module pass manager (even if they are function
passes). The separate function pass manager is only used for very
early optimization passes.
Rather than apparending passes to the MPM, use the OptimizerLast
and EnabledOnOptLevel0 pass manager builder extension hooks, which
allow adding passes directly before finalization (alias
canonicalization and name-anon-globals).
The main effect and purpose of this change is to add sanitizer
passes at the end of the pipeline, which is where they belong.
In LLVM 9 the address sanitizer can't be used as a pass in the
early function pass manager, because it has a dependence on a
module-level analysis pass.
Prepare for LLVM 9 update
Main changes:
* In preparation for opaque pointer types, the `byval` attribute now takes a type. As such, the argument type needs to be threaded through to the function/callsite attribute application logic.
* On ARM the `+fp-only-sp` and `+d16` features have become `-fp64` and `-d32`. I've switched the target definitions to use the new names, but also added bidirectional emulation so either can be used on any LLVM version for backwards compatibility.
* The datalayout can now specify function pointer alignment. In particular on ARM `Fi8` is specified, which means that function pointer alignment is independent of function alignment. I've added this to our datalayouts to match LLVM (which is something we check) and strip the fnptr alignment for older LLVM versions.
* The fmul/fadd reductions now always respect the accumulator (including for unordered reductions), so we should pass the identity instead of undef.
Open issues:
* https://reviews.llvm.org/D62106 causes linker errors with ld.bdf due to https://sourceware.org/bugzilla/show_bug.cgi?id=24784. To avoid this I've enabled `RelaxELFRelocations`, which results in a GOTPCRELX relocation for `__tls_get_addr` and avoids the issue. However, this is likely not acceptable because relax relocations are not supported by older linker versions. We may need an LLVM option to keep using PLT for `__tls_get_addr` despite `RtLibUseGOT`.
The corresponding llvm-project PR is https://github.com/rust-lang/llvm-project/pull/19.
r? @ghost
In order to make sure that Rust's bitflags types are passed the same
way in the Rust ABI as they are in the C ABI, we need to use the attribute
repr(transparent) over the repr(C) attribute for the single-field bitflags
structs in in order to prevent ABI mismatches. Thanks to Michael Karcher
for finding this bug.
rustc: Start implementing compat with LLVM 9
This commit doesn't actually migrate to LLVM 9, but it brings our own
C++ bindings in line with LLVM 9 and able to compile against tip of
tree. The changes made were:
* The `MainSubprogram` flag for debuginfo moved between flag types.
* Iteration of archive members was tweaked slightly and we have to
construct the two iterators before constructing the returned
`RustArchiveIterator` value.
* The `getOrInsertFunction` binding now returns a wrapper which we use
`getCallee()` on to get the value we're interested in.
I was trying to output LLVM IR directly to the console:
$ rustc hello.rs --emit=llvm-ir -o /dev/stdout
LLVM ERROR: IO failure on output stream: Bad file descriptor
Now `LLVMRustPrintModule` returns an error, and we print:
error: failed to write LLVM IR to /dev/stdout.hello.7rcbfp3g-cgu.0.rcgu.ll: Permission denied
... which is more informative.
This commit doesn't actually migrate to LLVM 9, but it brings our own
C++ bindings in line with LLVM 9 and able to compile against tip of
tree. The changes made were:
* The `MainSubprogram` flag for debuginfo moved between flag types.
* Iteration of archive members was tweaked slightly and we have to
construct the two iterators before constructing the returned
`RustArchiveIterator` value.
* The `getOrInsertFunction` binding now returns a wrapper which we use
`getCallee()` on to get the value we're interested in.
rustc: Implement incremental "fat" LTO
Currently the compiler will produce an error if both incremental
compilation and full fat LTO is requested. With recent changes and the
advent of incremental ThinLTO, however, all the hard work is already
done for us and it's actually not too bad to remove this error!
This commit updates the codegen backend to allow incremental full fat
LTO. The semantics are that the input modules to LTO are all produce
incrementally, but the final LTO step is always done unconditionally
regardless of whether the inputs changed or not. The only real
incremental win we could have here is if zero of the input modules
changed, but that's so rare it's unlikely to be worthwhile to implement
such a code path.
cc #57968
cc rust-lang/cargo#6643
Currently the compiler will produce an error if both incremental
compilation and full fat LTO is requested. With recent changes and the
advent of incremental ThinLTO, however, all the hard work is already
done for us and it's actually not too bad to remove this error!
This commit updates the codegen backend to allow incremental full fat
LTO. The semantics are that the input modules to LTO are all produce
incrementally, but the final LTO step is always done unconditionally
regardless of whether the inputs changed or not. The only real
incremental win we could have here is if zero of the input modules
changed, but that's so rare it's unlikely to be worthwhile to implement
such a code path.
cc #57968
cc rust-lang/cargo#6643
Implement optimize(size) and optimize(speed) attributes
This PR implements both `optimize(size)` and `optimize(speed)` attributes.
While the functionality itself works fine now, this PR is not yet complete: the code might be messy in places and, most importantly, the compiletest must be improved with functionality to run tests with custom optimization levels. Otherwise the new attribute cannot be tested properly. Oh, and not all of the RFC is implemented – attribute propagation is not implemented for example.
# TODO
* [x] Improve compiletest so that tests can be written;
* [x] Assign a proper error number (E9999 currently, no idea how to allocate a number properly);
* [ ] Perhaps reduce the duplication in LLVM attribute assignment code…
The new git submodule src/llvm-project is a monorepo replacing src/llvm
and src/tools/{clang,lld,lldb}. This also serves as a rebase for these
projects to the new 8.x branch from trunk.
The src/llvm-emscripten fork is unchanged for now.
Enable -mergefunc-use-aliases
If the Rust LLVM fork is used, enable the -mergefunc-use-aliases
flag, which will create aliases for merged functions, rather than
inserting a call from one to the other.
A number of codegen tests needed to be adjusted, because functions
that previously fell below the thunk limit are now being merged.
Merging is prevented in various ways now.
I expect that this is going to break something, somewhere, because
it isn't able to deal with aliases properly, but we won't find out
until we try :)
This fixes#52651.
r? @rkruppe
If the Rust LLVM fork is used, enable the -mergefunc-use-aliases
flag, which will create aliases for merged functions, rather than
inserting a call from one to the other.
A number of codegen tests needed to be adjusted, because functions
that previously fell below the thunk limit are now being merged.
Merging is prevented either using -C no-prepopulate-passes, or by
making the functions non-identical.
I expect that this is going to break something, somewhere, because
it isn't able to deal with aliases properly, but we won't find out
until we try :)
This fixes#52651.
