necc compiler measurements · 21 November 2026 The compiler was built from necc b786f2bc576048ab0e4a0a4843244296ed0f6e29. Its native frontend and backend were identical in both arms. The comparison applies to the default C compiler, including its SDK preprocessor and assembler. The published SDK build change is necc fa1a144725ab5405fc5dd69262b1272a551613ae. https://github.com/Spaceghost/necc/commit/fa1a144725ab5405fc5dd69262b1272a551613ae Both SDK executables were rebuilt with the same TinyCC seed, configured source, x87 fix and standalone Nelibc archive. TinyCC mob 32d1c0a5614b9979768b3711726db3d601750ed3 was the upstream source. The updated build substitutes fputs_unlocked only while compiling TinyCC's single-threaded command-line executable. C programs built by necc still reference the locking fputs symbol. SHA-256 3cbc4f540bde29660b43afc5194fbff4f749f0c64b471d0d3071f247bbfa60ae necc executable, used in both arms 47d56168275721d017cf1dcb6ede3386053e239a7f4ae4e527dd956c042d9070 TinyCC rebuilt with previous SDK recipe eed8f3bfc9211ca6e47dc5cc63f1f5eede512560c3927edb2044fdcbeb735121 TinyCC rebuilt with updated SDK recipe a40cf80936b5154391e37bbab27dc17e486ebd8fde15edec59e59d119a7838c5 archived 19 November toolchain and workload bundle Primary field An Intel Core i7-11700F ran Debian 12 userspace in a disconnected, disposable container. All benchmark processes inherited affinity to logical CPU 5, physical core 5. Its SMT sibling is CPU 13. The one-minute system load average moved from 2.50 to 3.73. The GCC toolchain was Debian 12.2.0, Clang 14.0.6 and Zig 0.17.0 (Clang 22.1.8). TinyCC, chibicc, lacc, cproc/QBE and PCC used the configurations from the original 19 November field. The external TinyCC row uses the benchmark's archived reference binary. The original compiler harness ran its six C translation units and two generated-program workloads. Qualification checked outputs, followed by one untimed randomized warm-up and 23 rounds. A fixed random seed of 60491 shuffled every qualifying case and configuration within each round. Compile times cover driver launch through writing the object. Each compile received a new unused macro value to avoid object-cache reuse. Execution times cover process launch and execution. Checks of results were outside the timed interval. Quartiles use inclusive interpolation. summary.csv records all 120 medians, quartiles, object sizes and text sizes. samples.csv records the elapsed nanoseconds for each of the 2,760 measured operations. necc-base means the prior SDK and necc-next the rebuilt SDK. In both cases necc is the same executable. GCC -O0/-O2, Clang 14 -O0/-O2, Zig cc -O0/-O2, TinyCC, chibicc, lacc/lacc -O2, cproc/QBE and PCC/PCC -O all qualified for each timed workload. Correctness verified-objects.csv gives the object checksums. Both necc SDKs produced byte-identical objects for every timed compile and the separate Lua VM compile probe. The Lua character and opcode tables matched 257 and 83 bytes from GCC. Arithmetic returned the expected eight bytes, and the bootstrap executable reproduced all 11,661 input bytes. The Lua VM probe required compilation and the exported luaV_execute symbol. cproc/QBE rejected a volatile store in that probe. PCC -O exceeded its 20-second qualification timeout. Neither failure affects a timed workload. The updated SDK passed standalone ABI, SysV, MS x64, runtime and public-driver checks, including the ordinary fputs target reference. Rebuilding TinyCC a second time reproduced its executable byte for byte. The Zig-built seed and both rebuilt preprocessors emitted identical text for the 128-function input. On the same source, a separate syscall trace of the previous SDK's preprocessor counted 17,518 gettid calls and 8,748 futex calls. The updated SDK made 42 and 10 respectively. Both made three write calls. Repeat policy An initial four-configuration run on Bak (Intel Core i5-10400, CPU 4) encountered memory and I/O pressure. Some GCC qualification commands timed out. Both necc variants completed every timed workload. We retained its medians in bak-four-tool-summary.csv. We repeated the same four configurations on the i7-11700F in alienware-four-tool-summary.csv before rerunning that host with the entire field. The page uses the complete field; the two narrower runs are independent measurements. Compiler phases A separate 25-round trial on i7-11700F CPU 5 repeated the original phase commands for both SDKs. Each round shuffled all 24 case, command and SDK combinations with seed 747. The original frontend row invokes necc --cc1 on preprocessed C. The additional public-S rows invoke necc -S on the same preprocessed C file. TinyCC assembly reads assembly on standard input. The two complete compile rows each start from the original C input. Each row includes a separate process launch. The independent phase times therefore do not add up to the full driver time. The one-minute system load average moved from 19.86 to 19.95 during this separate comparison. For each fixture, the two SDKs emitted identical preprocessed C, assembly, native object and TinyCC object bytes. The public -S and --cc1 assembly outputs were also identical, and the separately assembled object matched the full necc compile. The objects linked with the original test driver and returned the expected eight result bytes. Each timed output, including each warm-up, was hashed and compared after the timer stopped. phase-summary.csv gives all 24 medians and quartiles, phase-samples.csv contains all 600 durations, and phase-outputs.csv gives hashes of the output bytes common to both SDKs. Download phase-replay.py alongside these files to repeat the check with the pinned fixture directory. Supply an available logical CPU. python3 phase-replay.py /absolute/path/to/necc previous-sdk updated-sdk fixtures output --cpu 5 --rounds 25 Generated arithmetic control The disassembled necc loop multiplies by 33 and masks on each iteration. The qualified Zig cc -O2 object multiplies by 33⁴ (0x121881), adds an offset that starts at 0x9a18a and advances by 0x24310 for every four input iterations, then masks to 20 bits. Its scalar loop runs 1,500,000 times for the six-million-iteration input. A separate loop handles the remainder for other trip counts. Adding only -fno-unroll-loops to the original Zig cc -O2 command produced a one-step loop. In an independent 23-round randomized repeat on CPU 5 with the original process timer and eight-byte output check, the original Zig object took 3.310 ms, the no-unroll control took 6.607 ms, and necc took 10.223 ms (medians). Both Zig objects and necc also matched the reference on 64 trip-count and seed combinations, including empty loops and all four remainders. All 92 samples from the four-tool control are in arithmetic-control.csv. GCC -O2 served as an additional reference.