Tested against C++ and Rust.

Differential benchmarks evaluating Prismio against Clang++ C++20 (-O3) and Rustc 2021 (opt-level 3). Every workload is tested against identical algorithms, multi-run medians, and canonical checksum verification—with transparent tracking for capabilities still in development.

Overall vs C++

Clang++ -O3

0.86×14.2% faster

Geometric mean of the time ratio across 62 comparison workloads. Lower is faster; 1.00× is parity.

20 faster39 parity3 slower

Overall vs Rust

rustc opt-level 3

0.83×16.9% faster

Geometric mean of the time ratio across 62 comparison workloads. Lower is faster; 1.00× is parity.

24 faster34 parity4 slower

Dead Code Pruning

LLVM Backend

1/1100% eliminated

Complete compile-time dead branch elimination across adversarial suites. A probe counts as eliminated when it runs in 10 µs or less.

Verified via binary symbol extraction and differential runtime sampling.

Toolchain & Binary Footprint

Wall-clock compilation speed and final stripped binary size across the full 78-workload suite.

Compilation Speed

Suite Clean Build

0.89 s62.7% faster than Clang++

Full compilation through the self-hosted frontend, AIF escape inference, and LLVM backend.

Prismio was compiled by a release build of the compiler.

Prismio0.89 s

C++20 (Clang -O3)2.38 s

Rust (rustc opt-3)3.23 s

Binary Footprint

Stripped Executable

220 KB63.1% smaller than Rust

Self-contained binary containing embedded AIF memory runtime, allocators, and stdlib routines.

C++20 (Clang -O3)150 KB

Prismio220 KB

Rust (rustc opt-3)596 KB

Coverage is part of the result.

63 of 78 cataloged workloads are implemented. The remaining 15 make an absence visible instead of allowing a bespoke benchmark-only substitute.

All Workloads Matrix

Filter, search, and inspect the individual runs. Timings represent elapsed execution medians across 5 runs. Click any row to view raw sample times, memory RSS, and checksum verification.

Outcome:

Showing 62 of 62 workloads

Benchmark workloads with median elapsed time per arm. Column headers sort the table.
						Details
Algorithms	12.06 ms±1.3%	12.08 ms	12.05 ms	1.00× (parity)	1.00× (parity)	Inspect
Algorithms	505.0 µs±0.4%	473.9 µs	645.3 µs	1.07× (+7%)	0.78× (-22%)	Inspect
Algorithms	25.84 ms	23.84 ms	25.79 ms	1.08× (+8%)	1.00× (parity)	Inspect
Algorithms	70.17 ms±0.2%	69.49 ms	134.30 ms	1.01× (+1%)	0.52× (-48%)	Inspect
Algorithms	4.78 ms	4.64 ms	4.63 ms	1.03× (+3%)	1.03× (+3%)	Inspect
Algorithms	3.97 ms	3.84 ms	5.16 ms	1.03× (+3%)	0.77× (-23%)	Inspect
Algorithms	56.0 µs	95.3 µs	172.3 µs	0.59× (-41%)	0.33× (-67%)	Inspect
Algorithms	667.0 µs	667.4 µs	662.3 µs	1.00× (parity)	1.01× (+1%)	Inspect
Algorithms	39.0 µs	149.7 µs	433.1 µs	0.26× (-74%)	0.09× (-91%)	Inspect
Algorithms	312.0 µs±1.6%	359.0 µs	290.8 µs	0.87× (-13%)	1.07× (+7%)	Inspect
Algorithms	515.0 µs	524.8 µs	492.8 µs	0.98× (-2%)	1.05× (+5%)	Inspect
Algorithms	456.0 µs	360.3 µs	394.8 µs	1.27× (+27%)	1.16× (+16%)	Inspect
Algorithms	1.69 ms±1.6%	1.90 ms	2.70 ms	0.89× (-11%)	0.62× (-38%)	Inspect
Algorithms	292.0 µs	475.2 µs	421.0 µs	0.61× (-39%)	0.69× (-31%)	Inspect
Algorithms	4.96 ms	8.62 ms	6.43 ms	0.58× (-42%)	0.77× (-23%)	Inspect
Algorithms	602.0 µs±2.2%	625.5 µs	652.6 µs	0.96× (-4%)	0.92× (-8%)	Inspect
Data structures	4.50 ms±2.3%	6.19 ms	11.63 ms	0.73× (-27%)	0.39× (-61%)	Inspect
Data structures	11.93 ms	11.31 ms	10.45 ms	1.05× (+5%)	1.14× (+14%)	Inspect
Data structures	76.50 ms±0.0%	78.40 ms	76.57 ms	0.98× (-2%)	1.00× (parity)	Inspect
Data structures	4.82 ms±0.3%	5.44 ms	8.25 ms	0.88× (-12%)	0.58× (-42%)	Inspect
Data structures	1.47 ms	1.46 ms	1.42 ms	1.01× (+1%)	1.04× (+4%)	Inspect
Data structures	3.89 ms±1.0%	3.82 ms	4.14 ms	1.02× (+2%)	0.94× (-6%)	Inspect
Data structures	3.55 ms	5.11 ms	6.65 ms	0.69× (-31%)	0.53× (-47%)	Inspect
Compute	4.46 ms±2.3%	4.41 ms	4.59 ms	1.01× (+1%)	0.97× (-3%)	Inspect
Compute	3.24 ms	3.27 ms	4.15 ms	0.99× (-1%)	0.78× (-22%)	Inspect
Compute	2.96 ms±1.1%	2.97 ms	3.27 ms	1.00× (parity)	0.91× (-9%)	Inspect
Compute	2.37 ms	2.34 ms	2.27 ms	1.01× (+1%)	1.04× (+4%)	Inspect
Compute	12.40 ms±0.8%	17.76 ms	12.03 ms	0.70× (-30%)	1.03× (+3%)	Inspect
Compute	3.56 ms±0.3%	3.54 ms	3.45 ms	1.00× (parity)	1.03× (+3%)	Inspect
Compute	44.91 ms±0.3%	44.94 ms	44.78 ms	1.00× (parity)	1.00× (parity)	Inspect
Compute	18.92 ms±0.7%	19.03 ms	23.79 ms	0.99× (-1%)	0.80× (-20%)	Inspect
Compute	3.43 ms	3.42 ms	3.21 ms	1.00× (parity)	1.07× (+7%)	Inspect
Compute	7.52 ms	7.95 ms	7.56 ms	0.95× (-5%)	1.00× (parity)	Inspect
Compute	180.0 µs±1.7%	180.4 µs	190.3 µs	1.00× (parity)	0.95× (-5%)	Inspect
Compute	155.0 µs±0.0%	150.9 µs	151.0 µs	1.03× (+3%)	1.03× (+3%)	Inspect
Compute	332.0 µs±0.9%	363.9 µs	412.3 µs	0.91× (-9%)	0.81× (-19%)	Inspect
Compute	2.32 ms±0.9%	2.97 ms	2.25 ms	0.78× (-22%)	1.03× (+3%)	Inspect
Compute	19.32 ms±0.2%	18.90 ms	19.60 ms	1.02× (+2%)	0.99× (-1%)	Inspect
Memory	574.0 µs±2.4%	1.40 ms	1.89 ms	0.41× (-59%)	0.30× (-70%)	Inspect
Memory	5.18 ms±2.5%	6.51 ms	5.15 ms	0.80× (-20%)	1.00× (parity)	Inspect
Memory	6.73 ms	6.58 ms	6.40 ms	1.02× (+2%)	1.05× (+5%)	Inspect
Memory	2.15 ms±0.4%	2.35 ms	2.43 ms	0.91× (-9%)	0.88× (-12%)	Inspect
Memory	6.76 ms	5.86 ms	5.76 ms	1.15× (+15%)	1.17× (+17%)	Inspect
Memory	305.0 µs	849.5 µs	469.4 µs	0.36× (-64%)	0.65× (-35%)	Inspect
Memory	1.82 ms±1.4%	4.16 ms	5.58 ms	0.44× (-56%)	0.33× (-67%)	Inspect
I/O and parsing	701.0 µs±1.3%	1.26 ms	694.6 µs	0.56× (-44%)	1.01× (+1%)	Inspect
I/O and parsing	828.0 µs	823.8 µs	780.0 µs	1.01× (+1%)	1.06× (+6%)	Inspect
I/O and parsing	317.0 µs±2.2%	930.4 µs	303.8 µs	0.34× (-66%)	1.04× (+4%)	Inspect
I/O and parsing	145.0 µs±0.7%	237.4 µs	141.2 µs	0.61× (-39%)	1.03× (+3%)	Inspect
I/O and parsing	813.0 µs±1.8%	786.4 µs	764.9 µs	1.03× (+3%)	1.06× (+6%)	Inspect
I/O and parsing	1.06 ms	1.13 ms	1.80 ms	0.93× (-7%)	0.59× (-41%)	Inspect
Adversarial	36.68 ms	36.09 ms	35.45 ms	1.02× (+2%)	1.03× (+3%)	Inspect
Adversarial	23.50 ms±0.8%	23.48 ms	23.89 ms	1.00× (parity)	0.98× (-2%)	Inspect
Adversarial	37.33 ms	36.05 ms	41.46 ms	1.04× (+4%)	0.90× (-10%)	Inspect
Adversarial	59.95 ms±2.5%	60.01 ms	60.27 ms	1.00× (parity)	0.99× (-1%)	Inspect
Adversarial	9.51 ms±1.2%	9.46 ms	9.52 ms	1.01× (+1%)	1.00× (parity)	Inspect
Adversarial	11.15 ms	12.03 ms	13.23 ms	0.93× (-7%)	0.84× (-16%)	Inspect
Adversarial	38.74 ms±0.4%	38.72 ms	36.86 ms	1.00× (parity)	1.05× (+5%)	Inspect
Adversarial	18.83 ms±0.9%	18.82 ms	18.76 ms	1.00× (parity)	1.00× (parity)	Inspect
Adversarial	8.38 ms±1.0%	9.77 ms	9.15 ms	0.86× (-14%)	0.92× (-8%)	Inspect
Adversarial	53.28 ms±0.2%	53.72 ms	52.79 ms	0.99× (-1%)	1.01× (+1%)	Inspect
Adversarial	1.63 ms±1.3%	1.59 ms	1.61 ms	1.02× (+2%)	1.01× (+1%)	Inspect

Unsupported means exactly that.

The suite does not award a comparison result when Prismio lacks the ordinary feature the workload is meant to evaluate. Rather than hiding gaps or using benchmark-only shims, missing capabilities are documented as formal roadmap obligations.

15 cataloged gaps across 5 domains

Algorithms

1 gap

regex_matchingcpu-memory

Regular expression engine and NFA/DFA automaton compiler in stdlib

Data Structures

4 gaps

linked_listallocation

Standard LinkedList / Deque container and operations in stdlib

binary_search_treememory-allocation

Ordered balanced tree map/set (std::set / BTreeSet) in stdlib

priority_queuecpu-memory

Binary heap / priority queue data structure and push/pop API in stdlib

lock_free_queueconcurrency

User-space atomic primitives (AtomicInt, CAS, memory orderings)

Compute & Concurrency

4 gaps

async_event_loopconcurrency

Async/await syntax, non-blocking I/O event reactor, cooperative scheduler

mutex_contentionconcurrency

User-space synchronization locks (Mutex, RwLock, Condvar) in stdlib

work_stealing_poolconcurrency

Work-stealing task thread-pool runtime (Rayon / GCD equivalent)

simd_vector_opscpu

Explicit portable SIMD vector types (e.g. f32x8, i32x4) and target intrinsics

Memory & Allocation

1 gap

custom_allocator_churnallocation

Custom allocator trait to pass user-defined bump/arena allocators into containers

I/O & Networking

5 gaps

json_parsecpu-allocation

Standard JSON parser and AST/DOM value model in stdlib

json_serializecpu-allocation

Standard JSON serializer and character escaping engine in stdlib

tcp_echo_serverio-concurrency

Network socket API (std.net, POSIX socket FFI wrappers)

mmap_file_ioio

Memory-mapped file I/O (mmap / MapViewOfFile) in std.fs

generic_serializationcpu-allocation

Compile-time reflection, procedural macros, or derive code generation attributes

Test Environment & Benchmark Methodology

Exact host hardware specifications, compiler toolchains, optimization flags, and harness parameters used for the latest checked-in benchmark run.

Host Hardware & OS

Processor: Apple M5 (10 cores)
Memory: 16 GB
Operating System: macOS 27.0 (26A428)
Target Triple: arm64-apple-darwin27.0.0
Power State: Battery Power

Toolchains & Compilers

Prismio: 0.1.0 (LLVM 23.1.1)
Clang++: 23.1.1 (C++20)
Rustc: 1.97.1 (Edition 2021)
Timing Harness: benchmarks/run.py (schema v3)
LLVM Backend: LLVM 23.1.1
Prismio compiler: release build
Source: a4cc420 (uncommitted changes)

Measurement Parameters

Sampling: 5 independent runs / arm
Reported Metric: Median elapsed nanoseconds
Memory Metric: Peak RSS (task_info)
Correctness: 64-bit checksum verification
DCE Filter: ≤ 10 µs (excluded)

Compiler Build Commands & Flags

# 1. Prismio

prismio build suite.psm -o prismio-suite

# 2. Clang++ (C++20)

clang++ -O3 -flto -std=c++20 -pthread {suite,algorithms,data_structures,compute,memory,io,adversarial}.cpp -o cpp-suite

# 3. Rustc

rustc -C opt-level=3 -C lto=fat -C codegen-units=1 --edition=2021 suite.rs -o rust-suite

Run & Reproduce Locally

The harness (benchmarks/run.py), benchmark sources, and checked-in results are part of the open-source repository. To build and execute the entire differential suite locally:

Terminal

>prismio build

>prismio bench

That's all. The Prismio CLI handles compiling the suite, invoking comparative baselines, sampling iterations, and validating checksums.

Reproduce or challenge a result.

The harness, benchmark sources, unsupported-workload records, and raw checked-in results are fully open-source. Inspect the source arms, audit the methodology contract, or contribute workloads.

Open the suite Read the roadmap

No operating system. The same programs.

Prismio against C (clang -O3) and Rust (no_std, opt-level 3) on bare-metal AArch64: no C library, no allocator, no startup files. One shared boot harness links into all three, so the only code that differs is the benchmarks. Every workload is counted in guest instructions under QEMU, which is exact and repeatable, with identical checksums across all three languages.

Overall vs C

clang -O3

0.95×5.3% fewer instructions

Geometric mean of Prismio's instruction count over the other arm's, across 20 workloads. Below 1.00× is fewer instructions.

5 fewer15 parity0 more

Overall vs Rust

rustc opt-level 3, no_std

0.85×14.7% fewer instructions

Geometric mean of Prismio's instruction count over the other arm's, across 20 workloads. Below 1.00× is fewer instructions.

9 fewer11 parity0 more

Repeatable

QEMU -icount

Exact3 boots agree

The virtual clock advances one nanosecond per instruction, so a timer tick is 16 instructions and a repeat boot returns the same count.

Build & code footprint

Time to build the suite, and the machine code each compiler produced for it.

Compilation speed

compile and link

Prismio166 ms

C (clang -O3)118 ms

Rust (opt-3)390 ms

Each arm is built from source and linked with the shared harness. Prismio compiles the whole suite as one program.

Image size

.text of the image

Prismio15 KB

C (clang -O3)15 KB

Rust (opt-3)16 KB

Executable code, including 9 KB of harness common to all three.

All freestanding workloads

Guest instructions per workload. Click any row to see the checksum all three arms returned and each arm's exact count.

Outcome:

Showing 20 of 20 workloads

Freestanding workloads with guest instructions executed per arm. Column headers sort the table.
						Details
Algorithms	4.96 M	5.28 M	4.96 M	0.94× (-6%)	1.00× (parity)	Inspect
Algorithms	5.50 M	5.50 M	5.63 M	1.00× (parity)	0.98× (-2%)	Inspect
Algorithms	5.06 M	5.06 M	5.06 M	1.00× (parity)	1.00× (parity)	Inspect
Algorithms	11.04 M	11.04 M	13.44 M	1.00× (parity)	0.82× (-18%)	Inspect
Algorithms	9.99 M	10.10 M	12.96 M	0.99× (-1%)	0.77× (-23%)	Inspect
Algorithms	10.91 M	10.70 M	14.55 M	1.02× (+2%)	0.75× (-25%)	Inspect
Algorithms	3.14 M	5.12 M	7.40 M	0.61× (-39%)	0.42× (-58%)	Inspect
Algorithms	5.54 M	5.60 M	10.10 M	0.99× (-1%)	0.55× (-45%)	Inspect
Algorithms	9.96 M	9.96 M	9.99 M	1.00× (parity)	1.00× (parity)	Inspect
Compute	8.59 M	8.59 M	8.62 M	1.00× (parity)	1.00× (parity)	Inspect
Compute	8.63 M	9.55 M	8.78 M	0.90× (-10%)	0.98× (-2%)	Inspect
Compute	9.60 M	9.60 M	9.60 M	1.00× (parity)	1.00× (parity)	Inspect
Compute	9.04 M	8.99 M	8.99 M	1.01× (+1%)	1.01× (+1%)	Inspect
Compute	7.45 M	9.85 M	9.90 M	0.76× (-24%)	0.75× (-25%)	Inspect
Memory	9.63 M	9.63 M	16.85 M	1.00× (parity)	0.57× (-43%)	Inspect
Memory	3.75 M	4.43 M	3.75 M	0.85× (-15%)	1.00× (parity)	Inspect
Memory	9.48 M	9.48 M	9.48 M	1.00× (parity)	1.00× (parity)	Inspect
Hardware	4.50 M	4.50 M	4.50 M	1.00× (parity)	1.00× (parity)	Inspect
Hardware	8.80 M	8.80 M	8.80 M	1.00× (parity)	1.00× (parity)	Inspect
Hardware	9.00 M	9.00 M	9.90 M	1.00× (parity)	0.91× (-9%)	Inspect

How to read this

What the number is, what it leaves out, and what is held equal.

What the number is

Guest instructions executed, read from the generic timer under QEMU's -icount shift=0 (16 instructions per tick). It is exact and repeatable, so there is no median, spread or noise model: a verdict is a fixed ±2% on the ratio.

What it leaves out

It models no cache, pipeline or branch predictor. A workload dominated by memory stalls, such as pointer_chase or mem_copy, costs more on real silicon than its count says, and a count cannot show a vectorised loop's advantage. It shows how much code each compiler asked the machine to run.

Held equal

One harness object for all three arms, built once. No floating point or SIMD and strict alignment for every arm (the MMU is off, so an unaligned access faults). The same algorithm in each language, checked by identical checksums.

Known difference

Rust indexes with bounds checks, as safe Rust does; C and Prismio's fixed arrays are unchecked. That is the language's price for safe indexing, so it stays in the number.

Toolchains

prismio 0.1.0 (bin build) · clang 23.1.1 · rustc 1.97.1 · lld 22.1.8 · qemu 11.1.1