Inserts markers in C/C++ code. Markers expose information about the optimizations a compiler has applied, e.g., whether a piece of code was dead code eliminated.
Project description
program-markers
inserts markers (as function calls) in C or C++ source code to enable differential testing for missed optimizations:
- Instrument a program.
- Compile it with two or more compilers (or compiler versions, or optimization levels, etc).
- For each output: the markers whose corresponding calls are still present in the generated assembly are alive, the remaining are dead.
- Differential testing: compare the sets of alive and dead markers across outputs.
There are two kinds of markers supported:
- DCE (Dead Code Elimination) Markers
- VR (Value Range) Markers
A DCEMarker tests if a compiler dead code eliminates a piece of code (basic block). For example,
if (Cond){
DCEMarker0_();
STMT0;
STMT1;
//...
}
If call DCEMarker0_();
is not present in the generated assembly code then the
compiler determined that Cond
must always be false and therefore the body of
this if statement is dead.
A VRMarker tests if a compiler can determine subsets of the value ranges of variables, for example:
if (a <= C)
VRMarker0_();
If call VRMarker0_();
is not present in the generated assembly code then the
compiler determined that a
's value is always a > C
. Currently only integer
variables are instrumented.
To build just the clang tool
Prerequisites: cmake
, make
, clang/llvm
13/14.
mkdir build
cd build
cmake ..
cmake --build . [--parallel]
cmake --install . --prefix=/where/to/install/
Usage
cat test.c
int foo(int a) {
if (a == 0)
return 1;
else {
a = 5;
}
return a;
}
dead-instrument test.c --
cat test.c | clang-format
#if defined DisableDCEMarker0_
#define DCEMARKERMACRO0_ ;
#elif defined UnreachableDCEMarker0_
#define DCEMARKERMACRO0_ __builtin_unreachable();
#else
#define DCEMARKERMACRO0_ DCEMarker0_();
void DCEMarker0_(void);
#endif
#if defined DisableDCEMarker1_
#define DCEMARKERMACRO1_ ;
#elif defined UnreachableDCEMarker1_
#define DCEMARKERMACRO1_ __builtin_unreachable();
#else
#define DCEMARKERMACRO1_ DCEMarker1_();
void DCEMarker1_(void);
#endif
int foo(int a) {
if (a == 0)
{
DCEMARKERMACRO1_
return 1;
}
else {
DCEMARKERMACRO0_
a = 5;
}
return a;
}
Individual markers can be disabled or turned into unreachables (useful for helping the compiler optimize parts known to be dead):
gcc -E -P -DDisableDCEMarker0_ -DUnreachableDCEMarker1_ test.c | clang-format
int foo(int a) {
if (a == 0) {
__builtin_unreachable();
return 1;
} else {
;
a = 5;
}
return a;
}
Passing --ignore-functions-with-macros
to program-markers
will cause it to ignore any functions that contain macro expansions.
Value range markers can be emitted instead by using --mode=vr
:
cat test.c
int foo(int a) {
if (a == 0)
return 1;
return 0;
}
program-markers --mode=vr test.c --
cat test.c | clang-format
#if defined DisableVRMarkerLE0_
#define VRMARKERMACROLE0_(VAR)
#elif defined UnreachableVRMarkerLE0_
#define VRMARKERMACROLE0_(VAR) \
if ((VAR) <= VRMarkerConstantLE0_) \
__builtin_unreachable();
#else
#define VRMARKERMACROLE0_(VAR) \
if ((VAR) <= VRMarkerConstantLE0_) \
VRMarkerLE0_();
void VRMarkerLE0_(void);
#endif
#ifndef VRMarkerConstantLE0_
#define VRMarkerConstantLE0_ 0
#endif
#if defined DisableVRMarkerGE0_
#define VRMARKERMACROGE0_(VAR)
#elif defined UnreachableVRMarkerGE0_
#define VRMARKERMACROGE0_(VAR) \
if ((VAR) >= VRMarkerConstantGE0_) \
__builtin_unreachable();
#else
#define VRMARKERMACROGE0_(VAR) \
if ((VAR) >= VRMarkerConstantGE0_) \
VRMarkerGE0_();
void VRMarkerGE0_(void);
#endif
#ifndef VRMarkerConstantGE0_
#define VRMarkerConstantGE0_ 0
#endif
int foo(int a) {
VRMARKERMACROLE0_(a)
VRMARKERMACROGE0_(a)
if (a == 0)
return 1;
return 0;
}
The ranges that each marker test can be adjucted via macros and individual markers can be disabled or turned into unreachables:
gcc -E -P -DDisableVRMarkerLE0_ -DVRMarkerConstantGE0_=8 test.c | clang-format
void VRMarkerGE0_(void);
int foo(int a) {
if ((a) >= 8)
VRMarkerGE0_();
if (a == 0)
return 1;
return 0;
}
Python wrapper
pip install program-markers
To use the instrumenter in python import from program_markers.instrumenter import instrument_program
: instrument_program(program: diopter.SourceProgram, ignore_functions_with_macros: bool) -> InstrumentedProgram
.
Building the python wrapper
Local build
./build_python_wheel_local.sh #this will build the current branch
pip install .
Docker based build
docker run --rm -e REVISION=REV -v `pwd`:/io theodort/manylinux-with-llvm:latest /io/build_python_wheel_docker.sh
This will build multiple wheels for REV
with multiple python versions.
The output is stored in the wheelhouse
directory.
The docker image is based on https://github.com/thetheodor/manylinux-with-llvm.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distributions
Built Distributions
Hashes for program_markers-0.4.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_28_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 9aa0a5391e0108173efd6b07ba4ef854f5c49474e931863c193fa4081d708f39 |
|
MD5 | 1bc296906e74581a33a10079887e10de |
|
BLAKE2b-256 | e06e5fd08138da9ac98beb39c12830a644ec0ac62e0ac7413bf87ce91dfad2a6 |
Hashes for program_markers-0.4.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_28_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 3618c79148fcf643eb6cafe068770e00f740e02bb9b6389a2f0b26c3b497c282 |
|
MD5 | cb17dbbf6ac0206245a09c9014038f1c |
|
BLAKE2b-256 | aa73e67065ff48d862ff78d50d94d992341a8b744dfd47a8e58655e6ab29f869 |