Previous slide Next slide Toggle fullscreen Toggle overview view Open presenter view
A RISC-V SoC written in HardCaml
Started with an RV32E MCU on an iCEstick
Now boots Linux 6.18, including SMP, on a Colorlight i9 (ECP5)
HardCaml: an RTL library embedded in OCaml
The designer chooses the registers and cycles; no HLS scheduling
+-> Cyclesim -> OCaml reference model
OCaml / HardCaml -------+
+-> Verilog -> Yosys / nextpnr -> FPGA
Most code is written by coding agents. Tests must check their work.
github.com/takibi-lang/takibi-hardware
RTL Meetup (19), 2026-11-21
S0: Can we construct the circuit?
RTL counterpart: compilation, elaboration and structural checks
OCaml checks host-language types; HardCaml checks circuit structure
Detect width mismatches, undriven signals and combinational loops
Also prove selected optimized blocks equal simple references
(Hardcaml_verify.Sec + z3)
Example: optional S-mode registers were left undriven (B-0025).
Boundary: well-formed hardware can still implement the wrong function.
Scope: circuit construction; equivalence checks for selected blocks.
RTL Meetup (19), 2026-11-21
S0f: Does a property hold across cycles?
RTL counterpart: property checking with formal tools
Generated Verilog + hand-written SystemVerilog properties
SymbiYosys: bounded model checking and k-induction
Check shared-bus exclusion and atomic-operation ordering
Example: one core could decide that an SC succeeded while the other
core held the bus for an AMO. A 3-cycle counterexample exposed it
before two-core simulation (B-0048).
Boundary: proofs depend on the properties and environment assumptions.
Scope: selected dual-core arbiter properties, not the complete CPU.
RTL Meetup (19), 2026-11-21
S1: Does the circuit do the right thing?
RTL counterpart: cycle simulation with a reference model / scoreboard
Cyclesim runs directly from the HardCaml circuit
Directed tests, random programs, interrupts and memory wait states
Compare architectural effects at every retired instruction
against an OCaml instruction-set model
Example: two redirects let an old fetch response become valid again.
Random lockstep with interrupts and wait states caught it (B-0047).
Boundary: two-state, zero-delay; shared mistakes in models remain a risk.
Scope: IP, cores and system benches; independent ISA tests add another oracle.
RTL Meetup (19), 2026-11-21
S2: Does the generated Verilog behave the same?
RTL counterpart: simulation of the emitted RTL
Compile the generated Verilog with Verilator
Reuse the OCaml benches through a Cyclesim-compatible interface
Apply the same stimulus, reference model and checks
Evidence: MCU firmware and UART benches run on both backends.
No ledger entry was first found at S2 proper in this audit.
Boundary: reusing a bench also reuses its blind spots;
this is not a four-state initialization check.
Scope: shared benches for smaller designs; a separate SoC Verilator bench
also boots Linux in lockstep.
RTL Meetup (19), 2026-11-21
S2b: Does the board top start correctly in RTL?
RTL counterpart: four-state, board-top RTL simulation
Icarus runs the generated board Verilog with firmware
Exercise reset/start-up and UART at the pins
Look for uninitialized state hidden by two-state simulation
Example: a port named time passed Cyclesim but failed Verilog
generation for board-sim: it is a reserved word (B-0031).
The regression now emits Verilog at S1.
Boundary: X behavior depends on the models and testbench setup;
simulation does not establish when real silicon is ready.
Scope: iCEstick and i9 MCU board tops, not a complete board model.
RTL Meetup (19), 2026-11-21
S2n: Does the synthesized circuit still work?
RTL counterpart: post-synthesis functional simulation
Read Yosys iCE40 / ECP5 netlists back into HardCaml
Model LUTs, flip-flops and block RAM; reuse Cyclesim benches
Check firmware in BRAM initialization and cell power-up values
Example: Yosys chose power-up 1 for a counter bit, bypassing the
i9 start-up hold. RTL tests had missed this assumption (B-0023).
Boundary: two-state cell models; no routed delays.
The netlist reader and cell models also need verification.
Scope: MCU benches and board tops for iCE40 / ECP5.
RTL Meetup (19), 2026-11-21
S3: Does it fit, and does it meet timing?
RTL counterpart: synthesis, place-and-route and static timing analysis
Yosys + nextpnr; check logic budget and expected BRAM count
Check target frequency and flip-flop naming
Record tool versions: placement and compiler results depend on them
Examples:
Unwanted read/write collision logic around BRAM (B-0004)
First Linux SoC build: 34.27 MHz vs. a 50 MHz target (B-0043)
Boundary: results depend on correct timing constraints and device models.
There is no post-route delay simulation in this flow.
Scope: board builds, including the Linux SoC.
RTL Meetup (19), 2026-11-21
S4: Do our assumptions match the board?
RTL counterpart: FPGA bring-up and hardware validation
Start small: blink, UART, memory test, then firmware / Linux
Change one thing at a time, or keep an ablation switch
Turn discoveries into cheaper regressions where possible
Example: consecutive SDRAM reads returned the next word's data.
The CAS-latency timing derivation used the wrong edge (B-0046).
The simulation model used our chosen latency, so it agreed with us.
Boundary: passing a board test covers only the exercised conditions.
Scope: iCEstick MCU; i9 MCU, I/O bring-up and Linux SoC, including SMP.
RTL Meetup (19), 2026-11-21
What the defect ledger says -- and what I want to ask
First found
S0
S0f
S1
S2
S2b
S2n
S3
S4
Total
Defects
3
1
28
0
2
3
10
4
51
32/51 first found at S0/S0f/S1. Includes benches, models and tools;
this is not an RTL coverage measurement or a comparison with Verilog.
Where are the most serious gaps in this verification flow?
Where would you spend the next unit of verification effort?
BUGS.md: entries, causes and regressions
/ Workflow
RTL Meetup (19), 2026-11-21
Takibi: the software side
Move runtime failures into compile-time checks
A research systems language and a kernel developed together
OCaml + LLVM compiler; runs Alpine BusyBox on RPi5 and QEMU/AArch64
Future goal: hardware/software co-design on FPGA
Today: feedback on the hardware verification flow is very welcome.
github.com/takibi-lang/takibi
RTL Meetup (19), 2026-11-21
Timing: 15 seconds. These are project-specific stage names, not a standard.
Timing: 30 seconds. No OCaml syntax is needed for this talk.
The stages below have different scopes; not every block runs every stage.
Timing: 25 seconds. Signal widths are checked at elaboration;
do not imply that OCaml's type system proves all widths or functionality.
Timing: 25 seconds. SC = store-conditional; AMO = atomic memory operation.
This is separate from the block equivalence checks on the preceding slide.
Timing: 25 seconds. Expect tests record observable output, but expected
results must be derived independently, not blindly updated after a failure.
Timing: 25 seconds. Zero first discoveries does not establish that a
stage is unnecessary. S2b and S2n are counted separately.
Timing: 25 seconds. The example is a generation failure encountered
while running this stage, not an X-propagation defect.
Timing: 25 seconds. Cell-model tests and intentionally inverted LUTs
check the verification machinery itself. This is simulation, not a full
formal equivalence proof between RTL and the mapped netlist.
Timing: 25 seconds. S3 is run early alongside functional verification,
not postponed until all simulations pass.
Timing: 25 seconds. The point is model validity: formal verification
can also prove a design against an incorrect external-device assumption.
Timing: 35 seconds. Audit: B-0001..B-0051, 2026-11-17.
Counts are Found at, not the cheapest stage that can now reproduce a defect.
B-0016's historical S2 board-sim label is normalized to S2b.
This is a request for expert feedback, not a claim of sufficient coverage.
Timing: 20 seconds. FPGA co-design is a future goal;
the current Takibi kernel is not running on this RISC-V SoC.