Often, it is necessary to run a multithreaded application in Pin sequentially,
interleaving threads at fine granularity. For example, a single-threaded
cycle-accurate simulator needs to switch threads every few instructions to
simulate timing faithfully. PIN_SaveContext and PIN_ExecuteAt are too
slow for this purpose, as they save and restore the huge thread context in
full. They only support a few hundred thousand context-switches a second,
limiting execution speed.
libspin solves this problem, allowing about a billion context-switches per second at the expense of making code between context-switches somewhat slower. libspin adds instrumentation to read and write registers from memory on each basic block or at a finer granularity (between every two possible switch-points). By storing registers in memory, a context-switch can be done through a simple jump. libspin includes optimizations to reduce the overheads of register-memory transfers.
Main features:
- Can be used by your pintool without knowing the gory details
- Provides a simple interface to perform context-switches
- Should work transparently to all applications, including handling of syscalls and exceptional conditions
- Has a simple, slow implementation using SaveContext and ExecuteAt
You can build and run the example tool and a test application as follows:
scons
$PINPATH/intel64/bin/pinbin -ifeellucky -t build/opt/tools_fast/interleaver.so -- build/opt/tests/atomic_add 100 1000
Read the source of the simple, slow implementation, and the examples/tests, to learn the interface. At a high level, libspin interposes on all your analysis functions. Some of your analysis functions can be special calls called switchcalls, which must return the next thread to switch to (or the same thread if you do not wish to switch). Switchcalls are always run before normal analysis functions.
libspin was developed as an internal component of the Swarm architecture simulator, so if you are building your own simulator, we recommend checking the Swarm simulator source for usage examples.
Copyright (C) 2015-2020 by Massachusetts Institute of Technology
libspin is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, version 2.
Additionally, if you use this software in your research, we request that you reference the Swarm paper ("A Scalable Architecture for Ordered Parallelism", Jeffrey et al., MICRO-48, 2015) as the source of libspin in any publications that use this software, and that you send us a citation of your work.