4df874f73d
Implements the following simd reduction intrinsics: - simd_reduce_add_ordered - simd_reduce_mul_ordered - simd_reduce_min_nanless - simd_reduce_max_nanless - simd_reduce_xor - simd_reduce_any - simd_reduce_all Also fixes the ordering of simd_reduce_min and simd_reduce_max, which were tested to be flipped. Both simd_reduce_min_nanless and simd_reduce_max_nanless are identical to their non-nanless variants for the time being. An attempt was made at a more optimal codegen solution based on vector_reduce_op. However, this approach ran into masking issues for floating-point vector types, which appears to be broken for the same reason that comparison operations such as simd_lt are broken for floating-point vector types. More investigation is required, however, to determine a root cause and appropriate fix. This should be enough to pass the generic-reduction-pass.rs ui tests with the 'master' feature enabled. Signed-off-by: Andy Sadler <andrewsadler122@gmail.com> |
||
---|---|---|
.github/workflows | ||
build_sysroot | ||
crate_patches | ||
example | ||
patches | ||
rustc_patches | ||
src | ||
tests | ||
tools | ||
.gitignore | ||
.rustfmt.toml | ||
build.sh | ||
Cargo.lock | ||
cargo.sh | ||
Cargo.toml | ||
clean_all.sh | ||
config.sh | ||
failing-ui-tests12.txt | ||
failing-ui-tests.txt | ||
LICENSE-APACHE | ||
LICENSE-MIT | ||
prepare_build.sh | ||
prepare.sh | ||
Readme.md | ||
rust-toolchain | ||
rustup.sh | ||
test.sh |
WIP libgccjit codegen backend for rust
This is a GCC codegen for rustc, which means it can be loaded by the existing rustc frontend, but benefits from GCC: more architectures are supported and GCC's optimizations are used.
Despite its name, libgccjit can be used for ahead-of-time compilation, as is used here.
Motivation
The primary goal of this project is to be able to compile Rust code on platforms unsupported by LLVM. A secondary goal is to check if using the gcc backend will provide any run-time speed improvement for the programs compiled using rustc.
Building
This requires a patched libgccjit in order to work. The patches in this repository need to be applied. (Those patches should work when applied on master, but in case it doesn't work, they are known to work when applied on 079c23cfe079f203d5df83fea8e92a60c7d7e878.) You can also use my fork of gcc which already includes these patches.
Put the path to your custom build of libgccjit in the file gcc_path
.
$ git clone https://github.com/rust-lang/rustc_codegen_gcc.git
$ cd rustc_codegen_gcc
$ git clone https://github.com/llvm/llvm-project llvm --depth 1 --single-branch
$ export RUST_COMPILER_RT_ROOT="$PWD/llvm/compiler-rt"
$ ./prepare_build.sh # download and patch sysroot src
$ ./build.sh --release
To run the tests:
$ ./prepare.sh # download and patch sysroot src and install hyperfine for benchmarking
$ ./test.sh --release
Usage
$cg_gccjit_dir
is the directory you cloned this repo into in the following instructions.
Cargo
$ CHANNEL="release" $cg_gccjit_dir/cargo.sh run
If you compiled cg_gccjit in debug mode (aka you didn't pass --release
to ./test.sh
) you should use CHANNEL="debug"
instead or omit CHANNEL="release"
completely.
Rustc
You should prefer using the Cargo method.
$ rustc +$(cat $cg_gccjit_dir/rust-toolchain) -Cpanic=abort -Zcodegen-backend=$cg_gccjit_dir/target/release/librustc_codegen_gcc.so --sysroot $cg_gccjit_dir/build_sysroot/sysroot my_crate.rs
Env vars
- CG_GCCJIT_INCR_CACHE_DISABLED
- Don't cache object files in the incremental cache. Useful during development of cg_gccjit to make it possible to use incremental mode for all analyses performed by rustc without caching object files when their content should have been changed by a change to cg_gccjit.
- CG_GCCJIT_DISPLAY_CG_TIME
- Display the time it took to perform codegen for a crate
Debugging
Sometimes, libgccjit will crash and output an error like this:
during RTL pass: expand
libgccjit.so: error: in expmed_mode_index, at expmed.h:249
0x7f0da2e61a35 expmed_mode_index
../../../gcc/gcc/expmed.h:249
0x7f0da2e61aa4 expmed_op_cost_ptr
../../../gcc/gcc/expmed.h:271
0x7f0da2e620dc sdiv_cost_ptr
../../../gcc/gcc/expmed.h:540
0x7f0da2e62129 sdiv_cost
../../../gcc/gcc/expmed.h:558
0x7f0da2e73c12 expand_divmod(int, tree_code, machine_mode, rtx_def*, rtx_def*, rtx_def*, int)
../../../gcc/gcc/expmed.c:4335
0x7f0da2ea1423 expand_expr_real_2(separate_ops*, rtx_def*, machine_mode, expand_modifier)
../../../gcc/gcc/expr.c:9240
0x7f0da2cd1a1e expand_gimple_stmt_1
../../../gcc/gcc/cfgexpand.c:3796
0x7f0da2cd1c30 expand_gimple_stmt
../../../gcc/gcc/cfgexpand.c:3857
0x7f0da2cd90a9 expand_gimple_basic_block
../../../gcc/gcc/cfgexpand.c:5898
0x7f0da2cdade8 execute
../../../gcc/gcc/cfgexpand.c:6582
To see the code which causes this error, call the following function:
gcc_jit_context_dump_to_file(ctxt, "/tmp/output.c", 1 /* update_locations */)
This will create a C-like file and add the locations into the IR pointing to this C file. Then, rerun the program and it will output the location in the second line:
libgccjit.so: /tmp/something.c:61322:0: error: in expmed_mode_index, at expmed.h:249
Or add a breakpoint to add_error
in gdb and print the line number using:
p loc->m_line
p loc->m_filename->m_buffer
To print a debug representation of a tree:
debug_tree(expr);
To get the rustc
command to run in gdb
, add the --verbose
flag to cargo build
.
How to use a custom-build rustc
- Build the stage2 compiler (
rustup toolchain link debug-current build/x86_64-unknown-linux-gnu/stage2
). - Clean and rebuild the codegen with
debug-current
in the filerust-toolchain
.
How to use mem-trace
rustc
needs to be built without jemalloc
so that mem-trace
can overload malloc
since jemalloc
is linked statically, so a LD_PRELOAD
-ed library won't a chance to intercept the calls to malloc
.
How to build a cross-compiling libgccjit
Building libgccjit
- Follow these instructions: https://preshing.com/20141119/how-to-build-a-gcc-cross-compiler/ with the following changes:
- Configure gcc with
../gcc/configure --enable-host-shared --disable-multilib --enable-languages=c,jit,c++ --disable-bootstrap --enable-checking=release --prefix=/opt/m68k-gcc/ --target=m68k-linux --without-headers
. - Some shells, like fish, don't define the environment variable
$MACHTYPE
. - Add
CFLAGS="-Wno-error=attributes -g -O2"
at the end of the configure command for building glibc (CFLAGS="-Wno-error=attributes -Wno-error=array-parameter -Wno-error=stringop-overflow -Wno-error=array-bounds -g -O2"
for glibc 2.31, which is useful for Debian).
Configuring rustc_codegen_gcc
- Set
TARGET_TRIPLE="m68k-unknown-linux-gnu"
in config.sh. - Since rustc doesn't support this architecture yet, set it back to
TARGET_TRIPLE="mips-unknown-linux-gnu"
(or another target having the same attributes). Alternatively, create a target specification file (note that thearch
specified in this file must be supported by the rust compiler). - Set
linker='-Clinker=m68k-linux-gcc'
. - Set the path to the cross-compiling libgccjit in
gcc_path
. - Comment the line:
context.add_command_line_option("-masm=intel");
in src/base.rs. - (might not be necessary) Disable the compilation of libstd.so (and possibly libcore.so?).