Toatie : functional hardware description with dependent types

Describing correct circuits remains a tall order, despite four decades of evolution in Hardware Description Languages (HDLs). Many enticing circuit architectures require recursive structures or complex compile-time computation — two patterns that prove difficult to capture in traditional HDLs. In a signal processing context, the Fast FIR Algorithm (FFA) structure for efficient parallel filtering proves to be naturally recursive, and most Multiple Constant Multiplication (MCM) blocks decompose multiplications into graphs of simple shifts and adds using demanding compile time computation. Generalised versions of both remain mostly in academic folklore. The implementations which do exist are often ad hoc circuit generators, written in software languages. These pose challenges for verification and are resistant to composition. Embedded functional HDLs, that represent circuits as data, allow for these descriptions at the cost of forcing the designer to work at the gate-level. A promising alternative is to use a stand-alone compiler, representing circuits as plain functions, exemplified by the CλaSH HDL. This, however, raises new challenges in capturing a circuit’s staging — which expressions in the single language should be reduced during compile-time elaboration, and which should remain in the circuit’s run-time? To better reflect the physical separation between circuit phases, this work proposes a new functional HDL (representing circuits as functions) with first-class staging constructs. Orthogonal to this, there are also long-standing challenges in the verification of parameterised circuit families. Industry surveys have consistently reported that only a slim minority of FPGA projects reach production without non-trivial bugs. While a healthy growth in the adoption of automatic formal methods is also reported, the majority of testing remains dynamic — presenting difficulties for testing entire circuit families at once. This research offers an alternative verification methodology via the combination of dependent types and automatic synthesis of user-defined data types. Given precise enough types for synthesisable data, this environment can be used to develop circuit families with full functional verification in a correct-by-construction fashion. This approach allows for verification of entire circuit families (not just one concrete member) and side-steps the state-space explosion of model checking methods. Beyond the existing work, this research offers synthesis of combinatorial circuits — not just a software model of their behaviour. This additional step requires careful consideration of staging, erasure & irrelevance, deriving bit representations of user-defined data types, and a new synthesis scheme. This thesis contributes steps towards HDLs with sufficient expressivity for awkward, combinatorial signal processing structures, allowing for a correct-by-construction approach, and a prototype compiler for netlist synthesis.Describing correct circuits remains a tall order, despite four decades of evolution in Hardware Description Languages (HDLs). Many enticing circuit architectures require recursive structures or complex compile-time computation — two patterns that prove difficult to capture in traditional HDLs. In a signal processing context, the Fast FIR Algorithm (FFA) structure for efficient parallel filtering proves to be naturally recursive, and most Multiple Constant Multiplication (MCM) blocks decompose multiplications into graphs of simple shifts and adds using demanding compile time computation. Generalised versions of both remain mostly in academic folklore. The implementations which do exist are often ad hoc circuit generators, written in software languages. These pose challenges for verification and are resistant to composition. Embedded functional HDLs, that represent circuits as data, allow for these descriptions at the cost of forcing the designer to work at the gate-level. A promising alternative is to use a stand-alone compiler, representing circuits as plain functions, exemplified by the CλaSH HDL. This, however, raises new challenges in capturing a circuit’s staging — which expressions in the single language should be reduced during compile-time elaboration, and which should remain in the circuit’s run-time? To better reflect the physical separation between circuit phases, this work proposes a new functional HDL (representing circuits as functions) with first-class staging constructs. Orthogonal to this, there are also long-standing challenges in the verification of parameterised circuit families. Industry surveys have consistently reported that only a slim minority of FPGA projects reach production without non-trivial bugs. While a healthy growth in the adoption of automatic formal methods is also reported, the majority of testing remains dynamic — presenting difficulties for testing entire circuit families at once. This research offers an alternative verification methodology via the combination of dependent types and automatic synthesis of user-defined data types. Given precise enough types for synthesisable data, this environment can be used to develop circuit families with full functional verification in a correct-by-construction fashion. This approach allows for verification of entire circuit families (not just one concrete member) and side-steps the state-space explosion of model checking methods. Beyond the existing work, this research offers synthesis of combinatorial circuits — not just a software model of their behaviour. This additional step requires careful consideration of staging, erasure & irrelevance, deriving bit representations of user-defined data types, and a new synthesis scheme. This thesis contributes steps towards HDLs with sufficient expressivity for awkward, combinatorial signal processing structures, allowing for a correct-by-construction approach, and a prototype compiler for netlist synthesis