STOKE is a stochastic optimizer and program synthesizer for the x86-64 instruction set. STOKE uses random search to explore the extremely high-dimensional space of all possible program transformations. Although any one random transformation is unlikely to produce a code sequence that is desirable, the repeated application of millions of transformations is sufficient to produce novel and non-obvious code sequences. STOKE can be used in many different scenarios, such as optimizing code for performance or size, synthesizing an implementation from scratch or to trade accuracy of floating point computations for performance. As a superoptimizer, STOKE has been shown to outperform the code produced by general-purpose and domain-specific compilers, and in some cases expert hand-written code.
In addition to searching over programs, STOKE contains verification infrastructure to show the equivalence between x86-64 programs. STOKE can consider test-cases, perform bounded verification all the way to fully formal verification that shows the equivalence for all possible inputs.
STOKE has appeared in a number of publications. For a thorough introduction to the design of STOKE, see:
- Stochastic Superoptimization -- ASPLOS 2013
- Data-Driven Equivalence Checking -- OOPSLA 2013
- Stochastic Optimization of Floating-Point Programs with Tunable Precision -- PLDI 2014
- Conditionally Correct Superoptimization -- OOPSLA 2015
- Stochastic Program Optimization -- CACM 2016
- Stratified Synthesis: Automatically Learning the x86-64 Instruction Set -- PLDI 2016
- Sound Loop Superoptimization for Google Native Client -- ASPLOS 2017
STOKE isn't production ready. It's a research prototype that demonstrates the viability of superoptimization techniques in various domains. It's not a general-purpose tool. The papers above describe specific areas where successes have been shown beyond the state of the art: in optimizing straight line code, code where correctness can be relaxed (e.g. floating point), synthesizing sematic specifications for an instruction set, and in optimizing code containing loops with special compilation requirements (e.g. Google Native Client). We're not quite at the point where we can take a generic loop and expect to improve gcc/llvm -O3 code. In part, this is because these compilers have decades of work behind them to make them really great; moreover, some (but not all!) outputs from optimizing compilers don't leave much room for improvement. If you're willing to get your hands dirty, we very much appreciate contributions!
- Prerequisites
- Choosing a STOKE version
- Building STOKE
- Using the formal validator
- Using STOKE
- Additional Features
- User FAQ
- Developer FAQ
- Extending STOKE
- Code Organization
- Gadgets
- Initial Search State
- Search Transformations
- Cost Function
- Live-out Error
- Verification Strategy
- Command Line Args
- Contact
STOKE will run on modern 64-bit x86 processors. We officially support Haswell
processors with AVX2 extensions. STOKE should also run on Sandy Bridge
systems (with AVX, but not AVX2). It might run on newer architectures, but we currently don't test these.
Running ./configure.sh as described next will automatically configure the build for the correct architecture.
STOKE is supported on the latest Ubuntu LTS release; in practice, it will also run on Ubuntu 13.10+ and on debian testing. If you're trying to get STOKE to work on another linux distribution, having the right version of g++ is key. STOKE is supported on 4.9 only. It should also work on later versions, but it will not work with g++ 4.8.x (missing regular expression support), and g++ 4.7.x and older definitely will not work.
IMPORTANT Ubuntu 16.04 shipped with a new gcc that changed the ABI. This is causing us some problems and we're still playing catch-up. If you're on Ubuntu 16.04, the best thing at the present-time is to install gcc-4.9 and use the ubuntu16.04 branch of this repository. After running ./configure.sh you'll want to be sure that CXX=/usr/bin/g++-4.9 and CC=/usr/bin/gcc-4.9 are in the .stoke_config file. Hopefully we'll get everything working for gcc-5.3 soon.
Most of STOKE's software dependencies are available through apt. These can be satisfied by typing:
$ sudo apt-get install bison ccache cmake doxygen exuberant-ctags flex g++-4.9 g++-multilib gcc-4.9 ghc git libantlr3c-dev libboost-dev libboost-filesystem-dev libboost-thread-dev libcln-dev libghc-regex-compat-dev libghc-regex-tdfa-dev libghc-split-dev libjsoncpp-dev python subversion libiml-dev libgmp-dev libboost-regex-dev
Note that your distribution might not have g++-4.9 by default. You may consider installing a PPA as described here: https://askubuntu.com/questions/466651/how-do-i-use-the-latest-gcc-4-9-on-ubuntu-14-04
The rest of the dependencies will be fetched automatically as part of the build process.
The entire STOKE code base is available on GitHub under the Apache Software License version 2.0 at github.com/StanfordPL/stoke.
To check it out, type:
$ git clone https://github.com/StanfordPL/stoke
This will check out the default develop branch. Unless you are looking for a specific version or modification of STOKE, this is the branch to use. It contains all the latest changes and is reasonably stable. This branch is supposed to always pass all tests.
See the previous sections on how to download STOKE, a list of dependencies, and to check your hardware support level. The remainder of STOKE's software dependencies are available on GitHub and will be downloaded automatically the first time that STOKE is built. When you build STOKE the first time, run
$ ./configure.sh
This will figure out the correct build parameters (such as the platform). To build STOKE, run
$ make
To add STOKE and its related components to your path, type:
$ export PATH=$PATH:/<path_to_stoke>/bin
To run the tests, choose the appropriate command:
$ make test
The files generated during the build process can be deleted by typing:
$ make clean
To delete STOKE's dependencies as well (this is useful if an error occurs during the first build), type:
$ make dist_clean
The following toy example shows a typical workflow for using STOKE. All of the
following code can be found in the examples/tutorial/ directory. As this
code is tested using our continuous integration system, the code there will
always be up-to-date, but this README can fall behind.
Consider a
C++ program that repeatedly counts the number of bits (population count) in the
64-bit representation of an integer. (Keeping track of a running sum prevents
g++ from eliminating the calls to popcnt() altogether.)
// main.cc
#include <cstdlib>
#include <stddef.h>
#include <stdint.h>
using namespace std;
size_t popcnt(uint64_t x) {
int res = 0;
for ( ; x > 0; x >>= 1 ) {
res += x & 0x1ull;
}
return res;
}
int main(int argc, char** argv) {
const auto itr = atoi(argv[1]);
auto ret = 0;
for ( auto i = 0; i < itr; ++i ) {
ret += popcnt(i);
}
return ret;
}STOKE is a compiler and programming language agnostic optimization tool. It can be applied to any x86-64 ELF binary. Although this example uses the GNU toolchain, nothing prevents the use of other tools. To build this code with full optimizations, type:
$ g++ -std=c++11 -O3 -fno-inline main.cc
To measure runtime, type:
$ time ./a.out 100000000
real 0m1.046s
user 0m1.047s
sys 0m0.000s
A profiler will reveal that the runtime of ./a.out is dominated by calls to
the popcnt() function. STOKE can be used to improve the implementation of
this function as follows. The first step is to disassemble the program by
typing:
$ stoke extract -i ./a.out -o bins
This will produce a directory named bins that contains the text of every
function contained in the binary ./a.out.
Help for stoke or any of its subcommands can be obtained by typing:
$ stoke -h
$ stoke <subcommand> -h
STOKE can accept arguments either through the command line or through a
configuration file. The invocation of stoke extract shown above is equivalent
to the following:
$ stoke extract --config extract.conf
Where extract.conf contains:
##### stoke extract config file
-i ./a.out # Path to the elf binary to disassemble
-o bins # Path to the directory to store disassembled text in
Every STOKE subcommand can be used to generate example configuration files by typing:
$ stoke <subcommand> --example_config <path/to/file.conf>
Because main.cc was compiled using g++, the text of the popcnt() function
will appear under the mangled name _Z6popcntm in bins/_Z6popcntm.s.
.text
.globl _Z6popcntm
.type _Z6popcntm, @function
_Z6popcntm:
xorl %eax,%eax
testq %rdi,%rdi
je .L_4005b0
nop
.L_4005a0:
movq %rdi,%rdx
andl $0x1,%edx
addq %rdx,%rax
shrq $0x1,%rdi
jne .L_4005a0
retq
.L_4005b0:
retq
nop
nop
.size _Z6popcntm, .-_Z6popcntmThe next step is to generate a set of testcases for guiding STOKE's search procedure. These can be obtained by typing:
$ stoke testcase --config testcase.conf
where testcase.conf contains:
##### stoke testcase config file
--bin ./a.out # The name of the binary to use to generate testcases
--args 10000000 # Command line arguments that should be passed to ./a.out
--functions bins # Disassembly directory created by stoke extract
-o popcnt.tc # Path to file to write testcases to
--fxn _Z6popcntm # The name of the function to generate testcases for
--max_testcases 1024 # The maximum number of testcases to generate.
The resulting file will contain 1024 entires, all of the form:
Testcase 0:
%rax 00 00 00 00 00 98 96 80
%rcx 00 00 00 00 00 00 00 00
%rdx 00 00 00 00 00 00 00 0a
%rbx 00 00 00 00 00 00 00 01
%rsp 00 00 7f ff 97 44 36 28
%rbp 00 00 00 00 00 00 00 00
%rsi 19 99 99 99 99 99 99 99
%rdi 00 00 00 00 00 00 00 00
%r8 00 00 2a c9 68 1a 50 40
%r9 00 00 7f ff 97 44 46 01
%r10 00 00 00 00 00 98 96 80
%r11 00 00 00 00 00 00 00 0a
%r12 00 00 00 00 00 98 96 80
%r13 00 00 7f ff 97 44 37 20
%r14 00 00 00 00 00 00 00 00
%r15 00 00 00 00 00 00 00 00
%ymm0 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ff 00 00
%ymm1 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 2f 2f 2f 2f 2f 2f 2f 2f 2f 2f 2f 2f 2f 2f 2f 2f
%ymm2 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
%ymm3 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 ff 00 00 00 00 00 00 00 ff
%ymm4 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
%ymm5 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
%ymm6 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
%ymm7 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
%ymm8 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
%ymm9 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
%ymm10 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
%ymm11 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
%ymm12 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
%ymm13 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
%ymm14 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
%ymm15 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
%cf 0
%1 1
%pf 1
%0 0
%af 0
%0 0
%zf 0
%sf 0
%tf 0
%if 1
%df 0
%of 0
%iopl[0] 0
%iopl[1] 0
%nt 0
%0 0
%rf 0
%vm 0
%ac 0
%vif 0
%vip 0
%id 0
[ 00007fff 97443630 - 00007fff 97443620 ]
[ 1 valid rows shown ]
00007fff 97443628 d d d d d d d d 00 00 00 00 00 40 04 6c
[ 00000000 00000000 - 00000000 00000000 ]
[ 0 valid rows shown ]
[ 00000000 00000000 - 00000000 00000000 ]
[ 0 valid rows shown ]
0 more segment(s)
Each entry corresponds to the hardware state that was observed just prior to an
execution of the popcnt() function. The first 60 rows represent the contents
of general purpose, sse, and eflags registers, and the remaining rows represent
the contents of memory, both on the stack and the heap. Memory is shown eight
bytes at a time, where a block of eight bytes appears only if the target
dereferenced at least one of those bytes. Each row contains values and state
flags. Bytes are flagged as either (v)alid (the target dereferenced this byte),
or (.)invalid (the target did not dereference this byte).
Each of the random transformations performed by STOKE are evaluated with
respect to the contents of this file. Rewrites are compiled into a sandbox and
executed beginning from the machine state represented by each entry. Rewrites
are only permitted to dereference defined locations. This includes registers
that are flagged as def_in (see synthesize.conf, below), memory locations that
are flagged as 'v', or locations that were written previously. Rewrites are
permitted to write values to all registers and to any memory location that is
flagged as valid.
STOKE will produce optimal code that works on the testcases. The testcases
need to be selected to help ensure that STOKE doesn't produce an incorrect
rewrite. In our main.cc file in examples/tutorial we choose arguments to the
popcnt function to make sure that it sometimes provides arguments that use
more than 32 bits. Otherwise, STOKE will sometimes produce a rewrite using the
popcntl instruction, which only operates on the bottom half of the register,
instead of the popcntq instruction, which operates on the whole thing.
Alternatively you can use the formal validator in bounded mode with a large
bound (over 32). This large bound is tractable because this example doesn't
has a small number of cases for memory aliasing (namely, none at all!). If a
counterexample is found it can be automatically added to the search so STOKE
won't make this mistake again.
The STOKE sandbox will safely halt the execution of rewrites that perform
undefined behavior. This includes leaving registers in a state that violates
the x86-64 callee-save ABI, dereferencing invalid memory, performing a
computation that results in a floating-point exception, or becoming trapped in
a loop that performs more than max_jumps (see synthesize.conf, below).
The final step is to use these testcases and the target code contained in
bins/_Z6popcntm.s to run STOKE search in synthesis mode (i.e., trying to find a program starting from the empty program) by typing:
$ stoke synthesize --config synthesize.conf
where synthesize.conf contains:
##### stoke search config file
--out result.s # Path to write results to
--target bins/_Z6popcntm.s # Path to the function to optimize
--def_in "{ %rdi }" # The registers that are defined on entry to the target
--live_out "{ %rax }" # The registers that are live on exit from the target
--testcases popcnt.tc # Path to testcase file
--training_set "{ 0 ... 7 }" # Testcases to use for measuring correctness during search
--test_set "{ 8 ... 1023 }" # Testcases to use as holdout set for checking correctness
--distance hamming # Metric for measuring error between live-outs
--misalign_penalty 1 # Penalty for results that appear in the wrong location
--reduction sum # Method for summing errors across testcases
--sig_penalty 9999 # Score to assign to rewrites that produce non-zero signals
--cost "correctness + latency" # Measure performance by summing instruction latencies
--global_swap_mass 0 # Proposal mass
--instruction_mass 1 # Proposal mass
--local_swap_mass 1 # Proposal mass
--opcode_mass 1 # Proposal mass
--operand_mass 1 # Proposal mass
--rotate_mass 0 # Proposal mass
--beta 1 # Search annealing constant
--initial_instruction_number 5 # The number of nops to start with
--statistics_interval 100000 # Print statistics every 100k proposals
--timeout_iterations 16000000 # Propose 16m modifications total before giving up
--cycle_timeout 1000000 # Try 1m modifications before restarting
--strategy hold_out # Verify results using a larger hold out testcase set
STOKE search will produce two types of status messages. Progress update messages will be printed whenever STOKE discovers a new lowest cost verified or unverified rewrite. The code shown on the left is not equivalent to the target code; the code shown on the right is with respect to the current set of testcases.
Progress Update:
Lowest Cost Discovered (9) Lowest Known Correct Cost (15)
btrq $0xffffffffffffffc0, %rdi testq %rdi, %rdi
retq je .L_X64ASM_0
xorl %eax, %eax
.L_X64ASM_1:
movl %edi, %edx
andl $0x1, %edx
addl %edx, %eax
shrq $0x1, %rdi
jne .L_X64ASM_1
cltq
retq
.L_X64ASM_0:
xorl %eax, %eax
retq
Statistics updates will be printed every statistics_interval proposals.
Statistics are shown for the number of proposals that have taken place, elapsed
time, proposal throughput, and for each of the transformations specified to
have non-zero mass in synthesize.conf.
Statistics Update:
Iterations: 100000
Elapsed Time: 0.0836948s
Iterations/s: 1.19482e+06
Move Type Proposed Succeeded Accepted
Instruction 16.791% 5.83% 2.009%
Opcode 16.646% 8.857% 4.013%
Operand 16.593% 10.444% 6.864%
Rotate 16.611% 0.791% 0.789%
Local Swap 16.597% 1.556% 1.128%
Global Swap 16.762% 7.066% 6.08%
Extension 0% 0% 0%
Total 100% 34.544% 20.883%
When search has run to completion, STOKE will write the lowest cost verified
rewrite that it discovered to result.s. Because this is a particularly simple
example, STOKE is almost guaranteed to produce the optimal rewrite:
.text
.globl _Z6popcntm
.type _Z6popcntm @function
_Z6popcntm:
popcnt %rdi, %rax
retq
.size _Z6popcntm, .-_Z6popcntmThis result can then be patched back into the original binary by typing:
$ stoke replace --config replace.conf
where result.conf contains:
##### stoke replace config file
-i ./a.out # Path to the elf binary to patch
--rewrite result.s # Path to the replacement function
And runtime can once again be measured by typing:
$ time ./a.out 100000000
real 0m0.133s
user 0m0.109s
sys 0m0.000s
As expected, the results are close to an order of magnitude faster than the original.
STOKE includes a formal validator. It's design and interface
are detailed in the src/validator/README.md file. To use the formal
validator instead of hold out testing, specify --strategy bounded for any
STOKE binary that you use. For code with loops, all paths will be explored up
to a certain depth, specified using the --bound argument, which defaults to 2. There's also --strategy ddec which attempts to run the data-driven equivalence checking algorithm; however, the current implementation isn't very robust -- please file bug reports with (target, rewrite) pairs that fail to validate but should.
An example of using the validator can be found in the examples/pairity
folder; this example has a Makefile much like the tutorial's and should be easy
to follow. The key difference is that the pairity example does not use
testcases to guide search. Instead, after producing a candidate rewrite,
the validator checks for equivalence. If the codes are not equivalent,
a counterexample is found, and this is used as a new testcase to help guide search.
There are some important limitations to keep in mind while using the validator:
- Only some instructions are supported. The
--validator_must_supportflag can be used to only propose instructions that can be validated. - Only the general purpose registers, SSE registers (
ymm0-ymm15) and five of the status flags (CF,SF,PF,ZF,OF) are supported. - Memory is now fully supported, even in the presence of complex aliasing.
In addition to the subcommands described above, STOKE has facilities for
debugging and benchmarking the performance of each of its core components. See stoke --help for an up-to-date list.
stoke debug cfg: Generate a pdf of a control flow graph.stoke debug cost: Compute the cost of a rewrite.stoke debug diff: Diff the resulting state of two functions.stoke debug effect: Show the effect of a function on the state.stoke debug formula: Show the SMT formula for a straight-line piece of code.stoke debug sandbox: Step through the execution of a rewrite.stoke debug search: View the changes produced by performing and undoing a program transformation.stoke debug simplify: Take an x86 program and simplify it (by removing redundant instructions).stoke debug state: Check the behavior of operators that manipulate hardware machine states.stoke debug tunit: Show the instruction sizes and RIP-offsets for a code.stoke debug verify: Check the equivalence of two programs.stoke benchmark cfg: Measure the time required to recompute a control flow graph.stoke benchmark cost: Measure the time required to compute a cost function.stoke benchmark sandbox: Measure the time required to execute a program in a STOKE sandbox.stoke benchmark search: Measure the time required to perform and undo a transformation to a program.stoke benchmark state: Measure the time required to reset the memory of a hardware machine state.stoke benchmark verify: Measure the time required to check the equivalence of two programs.
STOKE also comes with support for bash and zsh completion. To enable either, type:
$ make bash_completion $ make zsh_completion
STOKE can not only propose instructions when searching for programs, but also propose calls to a list of known functions using the --functions command-line argument. To decide whether these functions read any undefined state (before proposing them), we use a dataflow analysis. Sometimes, the dataflow analysis can be too imprecise, which is why STOKE allows the user to annotate dataflow information in comments. Here is an example of a function that clears the overflow flag. STOKEs dataflow analysis is too imprecise for this code.
.text
.globl clear_of
.type clear_of, @function
#! maybe-read { }
#! maybe-write { %of %r15 }
#! maybe-undef { }
.clear_of:
pushfq
popq %r15
andq $0xfffff7ff, %r15
pushq %r15
popfq
retq
.size clear_of, .-clear_of
Note that it is enough to specify the maybe sets, as STOKE will automatically realize that the must sets need to be contained in the maybe set.
Both use the same core search algorithm, but in synthesis mode, STOKE starts from the empty program and tries to find a rewrite from scratch. This is great for finding implementations that are very different than the target. In optimization mode however, STOKE starts from an initial program, usually the target. This allows STOKE to work on much longer programs (because it already starts with a correct program) and apply optimizations to that program.
Right now, stoke replace has a limitation where it can only replace a function if the old implementation has at least the size (in bytes) of the new implementation.
If you still want to use stoke replace, and if you control the compilation of the binary, a workaround is to make the old implementation artificially larger by using the compiler flag -falign-functions=N for some large enough N, say 64. In this case, the compiler will align functions at N bytes, which typically requires padding the functions with nops. This increases the chance of stoke replace to succeed.
There is a good explanation in the issue tracker. We also have a script to compare how gcc and the x64asm assembler assemble an instruction.
STOKEs sandbox catches SIGFPEs, and thus running STOKEs search in the sandbox causes gdb to pause very often. To not have it stop for SIGFPEs (they are almost never a problem for STOKE), run this inside gdb:
handle SIGFPE nostop noprint
You can enable this by default by running the following command:
echo "handle SIGFPE nostop noprint" > .gdbinit
This repository contains a minimal implementation of STOKE as described in the academic papers about STOKE. Most, but not all of the features described in those papers appear here. Developers who are interested in refining these features or adding their own extensions are encouraged to try modifying this implementation as described below.
The STOKE source is organized into modules, each of which correspond to a
subdirectory of the src/ directory:
src/analysis: An aliasing analysis used by the validator.src/cfg: Control flow graph representation and program analysis.src/cost: Different cost functions that can be used in the search.src/disassembler: Runs objdump and parses the results into STOKE's format.src/expr: A helper used to parse arithmetic expressions in config files.src/ext: External dependencies.src/sandbox: A sandbox for testing proposed code on the hardware.src/search: An implementation of an MCMC-like algorithm for search.src/state: Data structures to represent concrete machine states (testcases).src/stategen: Generates concrete machine states (testcaes) for a piece of code.src/symstate: Models the symbolic state of the hardware; only used by the formal validator.src/target: Code to find which instruction sets the CPU supports.src/transform: Transforms used during search to mutate the code.src/tunit: Classes for representing a function (x86-64 code along with a name and other metadata).src/verifier: Wrappers around verification techniques such as testing for formal validation.src/validator: The formal validator for proving two codes equivalent.
The tools/ directory has the code that performs application logic and reads
command line arguments.
tools/apps: The application logic for stoke binariestools/args: Lists of command line arguments used by a gadget (see below).tools/gadgets: Modules used by applications to configure internal APIs with command line arguments.tools/io: Code to read/write certain kinds of command line arguments.tools/scripts: Where we put stuff when we don't have a better place. Nothing to see here.tools/target: Arbitrarily named directory with code to read CPU features from cpuinfo.
The stoke codebase is setup in a very modular way. We have components like the
Sandbox, which emulates execution of a rewrite on hardware. Or, we have
subclasses of CostFunction which evaluate the quality of a rewrite. Or, we
have an SMTSolver which is used by the formal validator to query a backend
like Z3 or CVC4.
Often, several stoke applications will wish to configure one of these modules
in the same way, depending on command line arguments. Thus, we have "Gadgets".
A "Gadget" is a subclass of the class you wish to configure which takes care of
extracting all the appropriate command line arguments. Some Gadgets, like
SandboxGadget just define a constructor so that modifies the object's
configuration. More involved ones, like CostFunctionGadget actually do work
to create a new CostFunction object and define methods that act as a wrapper.
Therefore, if you want to add a command line option to an existing component of
stoke, you normally are going to want to modify the gadget for that component
in tools/gadgets and add the argument in tools/args. Once you do that, it
should show up uniformly in all of the stoke tools that use that module.
Initial state types are defined in src/search/init.h along with an additional
type for user-defined extensions.
enum class Init {
EMPTY,
ZERO,
TARGET,
PREVIOUS,
// Add user-defined extensions here ...
EXTENSION
};Initial state is specified using the --init command line argument which controls the initial values
given to the current, lowest cost, and lowest cost correct search states. This value
affects the behavior of the Search::configure() const method, which
dispatches to the family of Search::configure_xxxxx() const methods. User-defined
extensions should be placed in the Search::configure_extension() const method,
which can be triggered by specifying --init extension.
void Search::configure_extension(const Cfg& target, SearchState& state) const {
// Add user-defined logic here ...
// Invariant 1: Search state should agree with target on boundary conditions.
assert(state.current.def_ins() == target.def_ins());
assert(state.current.live_outs() == target.live_outs());
assert(state.best_yet.def_ins() == target.def_ins());
assert(state.best_yet.live_outs() == target.live_outs());
assert(state.best_correct.def_ins() == target.def_ins());
assert(state.best_correct.live_outs() == target.live_outs());
// Invariant 2: Search state must be in a valid state. This function isn't on
// a critical path, so this can safely be accomplished by calling
state.current.recompute();
state.best_yet.recompute();
state.best_correct.recompute();
// Invariant 3: Search state must agree on first instruction. This instruction
// must be the label definition that appears in the target.
assert(state.current.get_code()[0] == target.get_code()[0]);
assert(state.best_yet.get_code()[0] == target.get_code()[0]);
assert(state.best_correct.get_code()[0] == target.get_code()[0]);
// See Search::configure for additional invariants
}Transformation types are defined in the src/transform directory. Each
transform is a subclass of the abstract class Transform. Existing transforms are,
| Name | Description |
|---|---|
| add_nops | Adds one extra nop instruction into the rewrite. |
| delete | Deletes one instruction at random. |
| instruction | Replaces an instruction with another one chosen at random. |
| opcode | Replaces an instruction's opcode with a new one that takes operands of the same type. |
| operand | Replaces an operand of one instruction with another. |
| rotate | Formerly "resize". Moves an instruction from one basic block to another, and shifts all the instructions in between. |
| local_swap | Takes two instructions in the same basic block and swaps them. |
| global_swap | Takes two instructions in the entire program and swaps them. |
| weighted | Selects from among several other transforms at random. |
These subclasses each implement operator()(Cfg& cfg) to mutate a Cfg. This
function returns an object, TransformInfo that contains all the information
needed to undo this transformation, and also whether the transform succeeded
(transforms are allowed to fail). It's common for this object to be set with
indexes of instructions in the code that were modified, for example. The
subclass also implements undo(Cfg& cfg, TransformInfo ti).
Transforms will often want to select from a collection of operands and opcodes,
and for this purpose they can access the pools_ protected variable of the
Transformsuperclass. This is of typeTransformPoolsand allows access to these collections. This makes it possible to configure the collection of available opcodes and operands independently of the transforms. Also, theTransformsuperclass has agen_` member which is used to produce random
numbers with a known seed.
Transformation weights are specified using the family of --xxxxx_mass command
line arguments. These values control the distribution of proposals that are
made by the WeightedTransform, which is the transform used by the
search.
A simple example of how to impelement a transform is in
src/transform/global_swap.cc. Note that all transforms must appropriately
make a call to recompute any Cfg information that needs to be updated and
ensure that cfg.check_invariants() returns true when done (you can assume it
returns true at the beginning of the function).
A cost function is specified using the --cost command line argument. It's an
expression composed using standard unsigned arithmetic operators. As
variables, you can use several measurements of the current rewrite. The most
important of these is correctness. The value correctness is (by default)
the number of bits that differ in the outputs of the target versus the rewrite
summed across all testcases. There are some tunable options for this, for
example, for floating point computations. In all cases, lower cost is better.
Some other important cost-variables you can use are:
| Name | Description |
|---|---|
| binsize | The size (in bytes) of the assembled rewrite using the x64asm library. |
| correctness | How "correct" the rewrite's output appears. Very configurable. |
| size | The number of instructions in the assembled rewrite. |
| latency | A poor-man's estimate of the rewrite latency, in clock cycles, based on the per-opcode latency table in src/cost/tables. |
| measured | An estimate of running time by counting the number of instructions actually executed on the testcases. Good for loops and algorithmic improvements. |
| sseavx | Returns '1' if both avx and sse instructions are used (this is usually bad!), and '0' otherwise. Often used with a multiplier like correctness + 1000*sseavx |
| nongoal | Returns '1' if the code (after minimization) is found to be equivalent to one in --non_goal. Can also be used with a multiplier. |
In typical usage, you will combine the value of correctness with other values
you want to optimize for. A good starting point is correctness + measured or
correctness + latency (the latter being default). Improvements might assign
an SSE-AVX penalty, like correctness + latency + 10000*sseavx.
To add a new cost function, drop a file into src/cost that subclasses
stoke::CostFunction. Look at src/cost/sseavx.h for a simple example. It
comes down to overloading the operator() function to return the value you
want. Look at measured.h for an example of how to use runtime data from the
sandbox to generate values. Then, add an entry to the map in
tools/gadgets/cost_function.h so that your new function can be found on the
command line.
Live-out error measurement types are defined in src/cost/distance.h along with an additional type for user-defined extensions.
enum class Distance {
HAMMING,
ULP,
// Add user-defined extensions here ...
EXTENSION
};Measurement type is specified using the --distance command line argument.
This value controls the behavior of the CorrectnessCost::evaluate_distance() const method, which dispatches to the family of
CorrectnessCost::xxxxx_distance() const methods, each of which represent a
method computing the distance between 64-bit values. User-defined extensions
should be placed in the CostFunction::extension_distance() const method,
which can be triggered by specifying --distance extension.
Cost CostFunction::extension_distance(uint64_t x, uint64_t y) const {
Cost res = 0;
// Add user-defined implementation here ...
// Invariant 1: Return value should not exceed max_error_cost
assert(res <= max_error_cost);
return res;
}The verification strategy specifies what kind of verification to do on the
rewrite. It's controlled using the --strategy command line argument. Right
now, the options are 'hold_out', 'straight_line' or 'bounded'.
Command line arguments can be added to any of the STOKE subcommands using the following syntax. Argument separators which are printed as part of help messages are specified by defining a heading variable:
auto& heading = Heading::create("Heading Description:");Command line flags are specified by declaring a FlagArg.
auto& flag = FlagArg::create("flag_name")
.alternate("alternate_flag_name")
.description("What this flag does");
Any of the built-in c++ primitive types are specified by declaring a ValueArg.
auto& val = ValueArg<int>::create("value_name")
.alternate("alternate_value_name")
.usage("<int>")
.description("What this value represents")
.default_val(0);User-defined types are specified by additionally providing function objects that define I/O methods.
struct T {
int x, y, z;
};
struct Reader {
void operator()(istream& is, T& t) const {
is >> t.x >> t.y >> t.z;
}
};
struct Writer {
void operator()(ostream& os, const T& t) const {
os << t.x << " " << t.y << " " << t.z;
}
};
auto& val = ValueArg<T, Reader, Writer>::create("value_name")
.alternate("alternate_value_name")
.usage("<int> <int> <int>")
.description("What this value represents")
.default_val({0,0,0});For complex values that are better suited to being read from files, a FileArg
may be more appropriate than a ValueArg. The syntax is identical.
auto& val = FileArg<Complex, ComplexReader, ComplexWriter>::create("value_name")
.alternate("alternate_value_name")
.usage("<complex representation>")
.description("What this value represents")
.default_val(Complex());Questions and comments are encouraged. Please reach us through the GitHub issue tracker, or alternatively at [email protected].