barepy 0.1.0

Python as an elegant coding bridge to bare-metal microcontroller C. A hybrid optimizing transpiler with source-mapped debugging.

barepy

Python as an elegant coding bridge to bare-metal microcontroller C.
A hybrid optimizing transpiler with source-mapped debugging.

---

Write firmware in a clean Python subset. barepy transpiles it to readable C, compiles it for your board (or a native simulator), and — crucially — maps every compiler and runtime error back to the exact Python line and column.

import barepy as bp
from barepy.boards import STM32F4

bp.init(STM32F4, clock="168MHz")

led = bp.Pin(13, bp.OUTPUT)
btn = bp.Pin(0, bp.INPUT_PULLUP)

@bp.setup
def setup():
    led.off()

@bp.loop
def loop():
    led.toggle() if btn.pressed() else led.off()
    bp.delay(200)

barepy sim examples/blink.py          # transpile -> C -> compile -> simulate

---

✨ Highlights

🐍 Python in, C out	A static, zero-overhead subset transpiles to small, fast, readable C — not an on-chip interpreter.
🎯 Source-mapped everything	Compiler, board, and runtime errors surface against your Python with a caret — never the generated C.
⚖️ Hybrid tiers	Most code is zero-cost static. Genuinely dynamic code opts into a small boxed-value runtime via @bp.dynamic. Pay only for what you use.
🔬 Real optimizer	Fix-point const-fold/prop, algebraic identities, branch-fold, DCE — every rewrite auditable with optimization-table.
🛡️ Panic-free	Every pipeline stage runs inside a fault barrier; a 1,500-case fuzzer asserts zero escapes. Runtime faults degrade gracefully.
🔥 Hot-code analysis	Ranks patterns dangerous on an MCU: blocking in an ISR, while True with no exit, div-by-zero, 32-bit overflow, busy loops.
🔌 Everything is a plugin	Boards, backends, passes, tiers, CLI/TUI commands, lints, flashers — all extension points on a plugin fabric.
🖥️ Textual studio	Cockpit TUI around your IDE: diagnostics (incl. C-level errors mapped to Python) · cost · optimizations · hot-code · generated C · sim · debug-C memory profiler · flash · test suite · live activity monitor.

Why it exists

A transpiler, not an on-chip interpreter — the output is native C, so it is fast and small. It is hybrid: most code is a static, zero-overhead subset; genuinely dynamic code (config parsing, protocol work) can opt into a small boxed-value runtime with @bp.dynamic. You pay only for what you use, and the build report tells you the cost.

Install

pip install barepy                 # core (from PyPI)
pip install 'barepy[studio]'       # + Textual TUI

# from source
pip install -e .
pip install -e '.[studio]'

Requires Python 3.11+ and a C compiler (cc for the native sim, arm-none-eabi-gcc for ARM builds).

Commands

command	does
barepy transpile f.py [--show] [--no-opt]	emit C + .map.json, no compile
barepy build f.py [--target sim|arm]	transpile + optimize + compile
barepy sim f.py [--ticks N]	build native + run the simulator
barepy debug f.py	build; render any error with a caret on Python source
barepy optimization-table f.py	table of every optimization applied
barepy hot-code f.py [--strict]	ranked report of fragile / dangerous code
barepy board list / board show <name|file>	inspect boards
barepy plugins list	show discovered plugins + extension points
barepy studio [f.py]	launch the Textual TUI

The compiler

The pipeline is a real optimizing compiler, not a translator:

• Type inference (M1) — flow-forward inference over the static tier assigns a concrete C type to every local (int32_t x, bool, …) and reports type errors against Python spans.

• Optimizer (M2) — fix-point passes: constant folding, constant propagation, algebraic identities (x*1, x+0, x*0), branch folding (if True), and dead-code elimination. barepy optimization-table shows each rewrite:

  line  pass         kind      before        -> after
    14  const_fold   arith     (2 + 12)      -> 14
    15  algebraic    x*1       (x * 1)       -> x
    17  branch_fold  if-true   if <true>     -> <then branch>
    19  const_fold   arith     (14 + 100)    -> 114
  

• SSA layer (M2+): inlining · monomorphization · copy-prop — a whole-program pass inlines single-return @bp.static helpers at their call sites (binding parameters to the actual argument trees), monomorphizes un-annotated helpers per call-site type signature, and propagates single-definition (trivially SSA) copies into their uses. A fully-inlined helper emits no C function at all — it costs nothing:

  line  pass          kind        before      -> after
    24  monomorphize  specialize  dbl(...)    -> dbl<int>
    24  inline        expand      dbl(...)    -> ((21 * 2) + 1)
    25  monomorphize  specialize  dbl(...)    -> dbl<bool>
    26  copy_prop     propagate   a           -> ((21 * 2) + 1)
    26  const_fold    arith       (42 + 1)    -> 43
  

• Hybrid dynamic tier (M4) — @bp.dynamic functions transpile to C over a boxed-value runtime (Obj*, bump arena). dict / list / for / unpack / int() / .split() / .get() all work. The static caller unboxes scalar results and resets the arena at the boundary — you pay only for what you use. Generated dynamic C is portable C11 — container literals are hoisted into ordinary statements, so nothing depends on GCC/Clang statement-expressions.

• Escape analysis → region inference → GC (M4+) — escape analysis classifies every dynamic allocation into a region (arena for non-escaping, gc for escaping/cyclic), surfaced in optimization-table. Opt into the cycle-collecting heap with bp.init(..., gc="mark-sweep"): allocations move onto a precise mark-sweep collector, generated code roots its live locals, and bp.gc_collect() reclaims unreachable objects — including reference cycles the bump arena never could:

  @bp.dynamic
  def churn(n: int) -> int:
      i = 0
      while i < n:
          d = {}
          d[b"self"] = d      # a reference cycle
          bp.gc_collect()     # reclaimed each iteration; live memory stays flat
          i = i + 1
      return bp.gc_live()     # bounded, no matter how large n is
  

Hot-code analysis

barepy hot-code ranks patterns that are dangerous on a microcontroller: blocking or allocating in an ISR, while True with no exit, division by zero, 32-bit overflow, ValueError-prone unpacking, busy loops, and more.

🟥 CRIT  scary.py:22  [isr-blocking] in on_button()
        blocking bp.delay() inside ISR 'on_button' stalls the whole system
        fix: ISRs must return fast; set a flag instead

Fault tolerance (aerospace-grade)

The transpiler is panic-free: every pipeline stage runs inside a fault barrier, so any internal bug becomes a contained ICE diagnostic, never a crash. A 1,500-case fuzzer of malformed / adversarial / binary input asserts zero escapes (tests/test_robustness.py). At runtime, dynamic-tier faults (bad int(), unpack mismatch, KeyError) print a source-mapped error and the program degrades gracefully instead of hard-faulting.

Source-mapped errors

barepy emits #line directives into the generated C and keeps a span sidecar, so a mistake surfaces against your Python, not the generated C:

$ barepy debug examples/bad_blink.py

examples/bad_blink.py:18:5  [static]
        btn.toggle()
        ^^^^^^^^^^^^
error: 'btn.toggle()' needs an OUTPUT pin, but 'btn' is INPUT_PULLUP
  hint: declare btn = bp.Pin(0, bp.OUTPUT)

Board-level constraints map back too:

$ barepy build examples/blink.py --board examples/custom_board.board.toml

examples/blink.py:4:1  [board]
    bp.init(STM32F4, clock="168MHz")
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
error: clock 168MHz exceeds MyTinyBoard max 48MHz

Define your own board

Three ways, all producing the same Board object.

1. Data file (*.board.toml) — shareable, no code:

name = "MyTinyBoard"
[mcu]
core = "cortex-m0"
flash = { base = 0x08000000, size = "64KB" }
ram   = { base = 0x20000000, size = "8KB" }
[clock]
max_hz = 48_000_000
[peripherals]
GPIOA = 0x48000000
[[pins]]
id = 13
port = "GPIOA"
bit = 5

barepy build blink.py --board my.board.toml

2. Python class — full control (see barepy/boards/__init__.py).

3. Plugin package — ship a board (and flasher, backend, passes…) on PyPI.

Hardware bring-up (real clock / RCC / SysTick)

The ARM target no longer assumes a hand-configured clock. Each build emits a barepy_board.h derived from the Board definition, and the runtime runs the real bring-up sequence:

• PLL / RCC — from the board clock tree (HSE, PLL m/n/p) barepy computes the core frequency (e.g. STM32F4: 8 MHz × 336 / (8 × 2) = 168 MHz), sets the flash wait states, programs the PLL, and switches SYSCLK over.
• GPIO clocks — only the ports your pins actually use are enabled, via an RCC AHB1ENR mask computed at build time.
• SysTick timebase — bp.delay(ms) waits on a real millisecond SysTick counter (WFI between ticks), not a tuned busy-loop.

// generated barepy_board.h  (STM32F4 @ clock="168MHz")
#define BP_CORE_HZ        168000000u
#define BP_RCC_BASE       0x40023800u
#define BP_FLASH_WS       5u
#define BP_GPIO_ENR_MASK  0x00000005u   // GPIOA + GPIOC (the pins in use)
#define BP_HAS_PLL        1
#define BP_PLL_M 8u  #define BP_PLL_N 336u  #define BP_PLL_P 2u

Plugins — everything is an extension point

A plugin is a package exposing a Plugin through the barepy.plugins entry-point group. It can register boards, backends, MIR passes, runtime tiers, CLI/TUI commands, lints, and flashers. See examples/plugins/barepy_tinyavr/.

from barepy.ext import Plugin, hookimpl

class plugin(Plugin):
    @hookimpl
    def register_boards(self, reg):  reg.add(MyBoard)
    @hookimpl
    def register_flasher(self, reg): reg.add("avr", avrdude_flash)

Architecture

Python source
  -> frontend      tokenize/AST -> IR (+ spans)   barepy/frontend.py
  -> analysis      effect/tier/pin checks         barepy/frontend.py:analyze
  -> board         resolve + validate             barepy/hal.py
  -> codegen       IR -> C + #line + source map   barepy/codegen.py
  -> driver        write artifacts, compile        barepy/driver.py
  -> runtime       T0 static core (native | arm)  barepy/runtime/

The IR carries a Span on every node — the source-map spine that lets any failure at any layer point back to Python. Later milestones lower the IR through SSA/MIR (optimizer) and LIR (boxing, coroutines, ISR lowering) without changing this pipeline shape.

Known limitations

barepy 0.1.0 is the M0 vertical slice — honest about its edges:

• Inlining is limited to single-return @bp.static helpers. General multi-statement static functions with parameters are inlined only in that expression form; a full MIR call-lowering is future work.
• The GC is a stop-the-world mark-sweep with a fixed object pool. It handles cycles correctly, but there is no generational / incremental collection yet, and collection runs only at safe points (never mid-expression).
• PLL / RCC bring-up targets the STM32F4 register layout. Other families compile against the same generated barepy_board.h, but non-F4 clock trees may need a board-specific recipe (ship one via a plugin backend).
• Supported Python is a subset. Classes, generators, exceptions, and arbitrary imports are out of scope by design — see the frontend for the exact grammar.

Status & roadmap

• frontend → IR with spans (for, bool ops, bytes, dict/list, subscript, unpack)
• static tier → C, #line + .map.json
• effect/pin/board diagnostics with caret rendering
• native simulator + STM32 ARM runtime
• board system: Python class, TOML, validation
• plugin fabric (boards/backends/passes/tiers/commands/lints/flashers)
• Textual studio (diagnostics · C · cost · optimizations · hot-code · sim · debug-C memory profiler · flash · test suite · live activity monitor)
• M1 type inference (flow-forward, C-type assignment, type errors)
• M2 optimizer (const-fold/prop, algebraic, branch-fold, DCE) + optimization-table
• M4 hybrid dynamic tier (boxed Obj* runtime, bump arena, boundary marshalling)
• hot-code risk analysis (hot-code)
• aerospace-grade fault barrier (panic-free, fuzz-verified) + mapped runtime faults
• M2+ SSA layer: inlining, monomorphization, copy-propagation
• M4+ escape analysis → region inference → cycle-collecting mark-sweep GC
• portable dynamic codegen (C11, no statement-expressions)
• real ARM clock/RCC bring-up + SysTick bp.delay (generated barepy_board.h)
• M2++ full MIR call-lowering (general static functions), CSE/GVN
• M4++ generational / incremental GC
• M5 async state machines · M6 DMA ownership · M8 LLVM backend (RISC-V/ESP)

M0 is the working vertical slice this repo ships. Each later milestone is independently useful.