T4: Compiling Sequential Code for Effective Speculative Parallelization in Hardware

This repository holds the code released to accompany the paper by Victor A. Ying, Mark C. Jeffrey, and Daniel Sanchez presented at the 47th Annual International Symposium on Computer Architecture (ISCA), June 2020.

This code release centers around T4, the first compiler released as part of SCC, a project to develop compiler toolchains for the Swarm architecture. T4 parallelizes C and C++ programs while maintaining purely sequential semantics.

Getting started with the T4 compiler

When cloning this repository, pass the --recurse-submodules option to git clone. Or, after cloning this repository, run the following command in this repository:

git submodule update --init --remote --recursive

to obtain all the code for the compiler and simulator.

Read and follow simulator/README.md to build the Swarm hardware simulator and run a small example Swarm program to verify if the simulator works. Then, see compiler/README.md and compiler/setup.sh to build the compiler. After you build the compiler and set your PATH using compiler/env.sh, you can compile a parallelized version of all the test programs in the microbenchmarks directory by running the following command in the root directory of this repository:

scons

You may want to use the -j option of scons to parallelize the compilation. You can pass --scc-remarks to scons to receive some information about the code transformations the compiler is performing for parallelization.

Then, you can run one of the test programs as follows:

simulator/build/opt/sim/sim -config configs/64c.cfg -- build/scc/microbenchmarks/single_loop_auto

The directory configs holds configuration files for the simulator. These configuration files will simulate systems with 4-wide superscalar out-of-order cores, as described in the T4 paper.

Running T4 tasks without simulation

When developing and debugging compiler optimizations, it is often useful to run the output of the compiler natively, without simulation. This lets you verify if your compiler is working correctly without the simulator. It also allows for easily examining the execution of the compiled application using debuggers and tracing tools. To facilitate this, the compiler can generate code that uses a software priority queue to schedule tasks, executing them serially, instead of using hardware features implemented by the simulator. You can build a version of the applications that is compiled into parallelizable tasks but executes serially and natively by running the command:

scons runtime=scc_serial --allocator=native

Then, you can run any of the resulting binaries without using the simulator. For example:

build/scc_serial/microbenchmarks/native_single_loop_auto

License and copyright

These repositories hold free software; you can redistribute the software and/or modify it under the terms of the license included in the root directory of each particular git repository or submodule. The license for particular files is also referenced at the top of most files. In particular, the Swarm simulator is available under the GNU General Public License as published by the Free Software Foundation, version 2. Meanwhile, the T4 compiler is based on LLVM/Clang 5.0, so it adopts the UIUC/NCSA license from that project.

If you use SCC in your research, we ask that you cite the T4 paper in your publications and that you send a citation of your work to [email protected].