Homework 5: LLVM - Spinning Up

Author

Qucheng Jiang

Env

Env setup

New dependency packages are required to build targets in this Homework. I choose using conda for clear isolation with the system build toolchain. However, things were much more difficult than ever if the idiot VSCode (and unreliable official plugins) came to the context: VSCode IntelliSense could not easily recognize the header file, although LSP had already found out where to go; copilot continued to ask me to fix the include path to clear these import error; and the corresponding items in setting panel I have specified and pointed to the correct direction is always ignored or overwritten by their default location, which is point to the system default path. After trying a bunch of suggestions posted in the VSCode community forum, I finally gave up on making the autocomplete and type-checking stuff work.

conda install conda-forge::clang
conda install conda-forge::cmake
conda install -c conda-forge llvmdev;  # header file, cmake recipe

After experiencing these kinds of frustrating config stuff, I strongly recommend installing/updating the system’s default LLVM toolchain to make VSCode find the definition of the functions and variables, and not draw red underlines to interrupt your mind.

Env startup

If the LLVM toolchain is installed in the system-level default location, there’s no need to do this extra step.

export LLVM_DIR=$CONDA_PREFIX/lib/cmake/llvm
export LD_LIBRARY_PATH=<PRJ_DIR>/HW5/llvm-passes/show-float-div/build:$LD_LIBRARY_PATH;

Detail Design

There’re 3 passes implemented in this Homework: - loop-unroll: Implemented a straightforward loop unrolling pass, only unrolling once if it finds a valid loop. (An ambitious task as the instruction of this homework suggested.) - show-bin-op: Display when there is a binary operation, log it out when doing compile, and call a stab function in rtlib during runtime to print the value stored in the destination. - show-float-div: As the instruction of this homework suggested, this pass is an example of an unambitious task, which displays only the fdiv operation, and acts the very same as the show-bin-op pass does when encountering fdiv operation.

* Check Passes Source Code for details.

Integration, Testing

Some examples of the real-ish C/C++ program are provided in this homework. - example.c in each pass: Unit Tests to make sure passes are performing the correct results as wished. - matmul.cpp: A square matrix multiply testbench, which would be a great scenario to test the loop unrolling pass, but also can test the other two display passes. - projectile_motion.cpp: Simulates the motion of a projectile launched at a given angle and initial velocity within finite N steps.

To build the tests:

For examples.c and the pass library, go to each pass folder, and run make. All the results will be generated, including: - $(BUILD_DIR)/$(PASS_NAME)/$(PASS_NAME)Pass.so: Pass plugin as a dynamic lib. - $(BUILD_DIR)/lib$(RTLIB_LIB_NAME).so: Additional Runtime functions (i.e. Stab call funcs). - example: Executable of examples.c. To run the example, use make run instead of make.

For matmul.cpp and projectile_motion.cpp, go to the app folder, and run make. Executable will be generated in the same folder. Use make run to test the execution.

Results and Analysis

Matrix Multiply - Loop Unroll

Slightly different in wall clock speed.

❯ make run
Running target with custom LLVM pass...
time ./matmul_with_pass
C[0][0] = 200
0.02user 0.00system 0:00.02elapsed 96%CPU (0avgtext+0avgdata 2772maxresident)k
0inputs+0outputs (0major+149minor)pagefaults 0swaps
Running target without custom LLVM pass...
time ./matmul_without_pass
C[0][0] = 200
0.01user 0.00system 0:00.02elapsed 95%CPU (0avgtext+0avgdata 2896maxresident)k
0inputs+0outputs (0major+153minor)pagefaults 0swaps

Example - Show `FDiv`

# ~/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div    main !8 ?7      ✔  15s   course-eece7398-st-compiler   jiang@builder  04:29:01  

❯ make clean
❯ make run
mkdir -p build
Building RTLib...
make[1]: Entering directory '/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div'
make[1]: 'build' is up to date.
make[1]: Leaving directory '/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div'
clang -fPIC -shared rtlib.c -o build/librtlib.so
Building LLVM Pass...
make[1]: Entering directory '/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div'
make[1]: 'build' is up to date.
make[1]: Leaving directory '/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div'
cd build && cmake .. && make
-- The C compiler identification is GNU 14.2.0
-- The CXX compiler identification is GNU 9.4.0
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Check for working C compiler: /home/jiang/miniconda3/envs/course-eece7398-st-compiler/bin/cc - skipped
-- Detecting C compile features
-- Detecting C compile features - done
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Check for working CXX compiler: /usr/bin/c++ - skipped
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Found ZLIB: /usr/lib/x86_64-linux-gnu/libz.so (found version "1.2.11")
-- Found zstd: /home/jiang/miniconda3/envs/course-eece7398-st-compiler/lib/libzstd.so
-- Found LibXml2: /home/jiang/miniconda3/envs/course-eece7398-st-compiler/lib/libxml2.so (found version "2.13.1")
-- Linker detection: GNU ld
-- Registering ShowFDivPass as a pass plugin (static build: OFF)
-- Configuring done (0.7s)
-- Generating done (0.0s)
-- Build files have been written to: /home/jiang/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div/build
make[1]: Entering directory '/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div/build'
make[2]: Entering directory '/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div/build'
make[3]: Entering directory '/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div/build'
make[3]: Leaving directory '/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div/build'
make[3]: Entering directory '/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div/build'
[ 50%] Building CXX object ShowFDiv/CMakeFiles/ShowFDivPass.dir/ShowFDiv.cpp.o
[100%] Linking CXX shared module ShowFDivPass.so
make[3]: Leaving directory '/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div/build'
[100%] Built target ShowFDivPass
make[2]: Leaving directory '/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div/build'
make[1]: Leaving directory '/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div/build'
Compiling target...
clang -fpass-plugin=build/ShowFDiv/ShowFDivPass.so -c example.c

ShowFDiv.cpp:34: Found FDiv operator:   %38 = fdiv double %37, 2.000000e+00
  %38 = fdiv double %37, 2.000000e+00
OpcodeName: fdiv

clang example.o -Lbuild -lrtlib -o example
rm -f example.o
Running target...
export LD_LIBRARY_PATH=/home/jiang/scratch/work/course/EECE7398_ST_Compiler/HW5/llvm-passes/show-float-div/build:$LD_LIBRARY_PATH && \
./example
Enter a number: 123
125
121
246
61
Enter a float number: 123.456
125.456001
121.456001
246.912003
[logfdiv: 61.728001 = 123.456001 / 2.000000]
61.728001

Env