Implementation Notes

This section collects design documents and pipeline traces that describe how TT-Lang lowers operations from Python to hardware code. These are intended for contributors and anyone who needs to understand compiler internals.

Design Documents

• Static Execution Analysis - exact operation cardinality in structured control flow

• Value Origin Analysis - conservative SSA value origins through control flow and tensor updates

• ComputeOp Creation and Fusion - immutable planning for ttl.compute creation, fusion, DFB publication, and intermediate materialization

• DST Register Allocation — how the TTLAssignDST pass assigns destination registers to tile operations

• DST Register Utilization — maximizing tile throughput per DST synchronization cycle

Lowering Pipeline Traces

These documents trace specific operations through the full compiler pipeline, from Python input through MLIR passes to generated C++ kernel code.

• Multi-tile Compute Operations — traces a 2x2 multi-tile add through the pipeline