Recently, Anthropic unveiled Claude Opus 4.6's 100% AI-written C compiler, nicknamed CCC (Claude's C Compiler). The claim that this compiler can compile the Linux kernel has sent shockwaves through the developer community.

How does an AI-created compiler perform compared to GCC, an industry standard that has been optimized for decades? The results are summarized based on Harshanu의 심층 분석.

1. What is CCC?

CCC is a C compiler written in Rust that supports various architectures such as x86-64, AArch64, and RISC-V. The remarkable thing is that all components, from the front-end to the code generator and linker, were written by Claude Opus 4.6. Humans only played a role in writing test cases and providing guidance.

2. Benchmark Environment and Target

For a fair comparison, the tests were conducted in the following environment:

Target: SQLite 3.46 (single source file), Linux Kernel 6.9
Comparison Group: GCC 14.2 vs CCC
Environment: Debian VM (6 vCPU, 16GB RAM)

The Linux kernel is the ultimate in complexity, and SQLite is the most suitable target for verifying standard C compliance and runtime performance.

3. Shocking Results: 158,000x Difference

In conclusion, CCC is at a level where it "works, but is unusable in practice."

✅ Successful Points (Correctness)

Accuracy: Passed all query tests in SQLite. The results were identical to GCC.
Compilation Success: Compiled 2,844 C files of the Linux kernel without errors.

❌ Failed Points (Performance & Linking)

Runtime Speed: In the SQLite benchmark, it was 737 times slower than GCC (-O0) and 1,242 times slower than optimized GCC (-O2).
Worst Case: In complex SQL operations involving subqueries, it showed a performance that was a whopping 158,000 times slower. (GCC: 0.047 seconds vs CCC: 7,432 seconds)
Kernel Linking Failure: Although C file compilation was successful, it failed to generate the final binary due to incorrect generation of relocation information such as __jump_table during the linking stage.

4. Why is it so Slow? (Root Cause)

The cause of CCC's performance degradation is the lack of compiler optimization.

1. Register Allocation Failure

The key to modern CPU performance is the efficient use of registers. However, CCC has almost no register allocation algorithm, so it stores all variables in the stack (RAM) and uses only the rax register as a shuttle to transport data.

GCC: Inter-register operation (cmpl %eax, %edx)
CCC: Repeated memory read -> register -> memory write (thousands of times slower)

2. Optimization Options in Name Only

CCC accepts the -O2 and -O3 flags, but they don't actually do anything. It generates binaries that are identical down to the byte level to -O0. There is no loop unrolling, inlining, vectorization, or other optimizations performed by GCC.

5. Conclusion: The Current State of AI Coding

Claude's C Compiler is a great success as a proof of concept (PoC) that "AI can create complex system software from scratch." Parsing Linux kernel source and converting it to object code is a very difficult task even for human developers.

However, there are still clear limitations to producing "efficient and optimized results." Compiler optimization is the culmination of decades of research and know-how, and it is a different dimension from simply knowing the syntax.

One-line Summary: Claude created an excellent C parser, but a terrible code generator.

Reference: CCC vs GCC by Harshanu

댓글