showcase C To Assembly Tests (Low-Level Performance)

[Christian]

Tests I've performed converting C to Assembly. Basically testing performance for code I've made in C and whatnot!

 

A small repository to store my findings with converting C code to Assembly code along with measuring performance between the different clang optimization levels. I'm starting to learn more about Assembly because I want to understand how programs work on a very low level so I can optimize it the best I can.

 

I've made the following source files to test with.

 

• Two source files for copying eight bytes of data from one 8-bit array (8 bytes in size) to another. One source file uses a for loop to achieve this while the other uses the native memcpy() function.
• Two source files for comparing a variable to five values. One source file uses if and else if while the other uses a switch statement.
• A source file that copies a string and outputs it to stdout.
• Two source files testing a for loop along with seeing if there's a difference when specifying pragma #unroll x which should unroll the for loop and result in better performance in our case.
• I'll likely be adding more files to this repository as time goes on.

 

Dumping Assembly Code
I used clang to emit LLVM and create the .bc file with no optimizations by the compiler (the -O0 flag). An example may be found below.

 

clang -c -emit-llvm -O0 -o asm/testO2.bc src/test.c

Since we emit LLVM, we may use the llc command to dump the Assembly code under specific optimization levels. I dump both the native architecture's Assembly code and also Intel's Assembly code (these Assembly files are appended with _intel).

 

Here's an example using optimization level 2 (notice the -O=2 flag in the llc command).

 

# Native architecture's Assembly code.
llc -filetype=asm -O=2 -o asm/testO2.s asm/testO2.bc

# Intel Assembly code.
llc -filetype=asm -O=2 -o asm/testO2_intel.s --x86-asm-syntax=intel asm/testO2.bc

NOTE - I'd recommend using the scripts/genassembly.sh Bash script I made to generate Assembly code under optimization levels 0 (None) - 3 and both non-Intel and Intel architectures. The script only requires one argument which is the name of the source file in src/ without the file extension (.c). Also make sure to modify the ROOTDIR variable if you place the script outside of this repository's scripts/ directory. An example may be found below.

 

./genassembly.sh pointer

Optimization Levels
Clang's optimization levels may be found in its manual page (man clang). For reference, here are the levels.

 

   Code Generation Options
       -O0, -O1, -O2, -O3, -Ofast, -Os, -Oz, -Og, -O, -O4
              Specify which optimization level to use:
                 -O0  Means “no optimization”: this level compiles the fastest
                 and generates the most debuggable code.

                 -O1 Somewhere between -O0 and -O2.

                 -O2 Moderate level of optimization which enables  most  opti‐
                 mizations.

                 -O3  Like -O2, except that it enables optimizations that take
                 longer to perform or that may generate  larger  code  (in  an
                 attempt to make the program run faster).

                 -Ofast  Enables  all  the  optimizations  from -O3 along with
                 other aggressive optimizations that may violate  strict  com‐
                 pliance with language standards.

                 -Os Like -O2 with extra optimizations to reduce code size.

                 -Oz Like -Os (and thus -O2), but reduces code size further.

                 -Og  Like  -O1. In future versions, this option might disable
                 different optimizations in order to improve debuggability.

                 -O Equivalent to -O2.

                 -O4 and higher
                     Currently equivalent to -O3

You'll notice a lot of optimizations within the Assembly code from -O1 to -O3.

 

System
This was all tested on my Linux VM running virtio_net drivers and Ubuntu 20.04 Server. The Linux kernel the tests in asm/ were built with was 5.15.2-051502-generic.

 

Credits

• @Christian

 

GitHub Repository & Source Code