Kernel compilation details

Kernel compilation details

Implementation reference checked against f5e786f. Fundamental explanations remain in docs/arch.

TGeneratedKernel and TKernelSlot are declared in generated.h. Generation records the AST, named entrypoints, referenced payload storage, execution-stage identity, and a slot. Slot function order matches Entrypoints; the slot retains the runner and, on the cached path, a separate JIT lifetime owner.

TKernelCompiler::EmitKernel appends to the configured sink. With BindNow it binds immediately; normal lowering sets it false and finalizes the collected query programs later. Stage is diagnostic, while ExecStageId is the identity. Sort key and aggregate output-layout metadata accompany the generated program.

JitFinalizeKernels deduplicates AST/entrypoint specifications, builds a fused program, binds entries, and retains payloads and external runtime ownership. QDB_JIT_CACHE_DIR enables the native object cache. Referenced external modules currently bypass that cache path. A warm cache still leaves frontend and query-specific compilation/linking work.

The browser finalizer in plan_export.cpp exports a query-level wasm64 artifact and stage entrypoint map. It uses Memory64, without a wasm32 ABI fallback. Export does not execute the native graph.

Consumers

• Physical graph JSON attaches kernel artifacts to semantic task groups by ExecStageId. Scheduler-only plumbing has separate monotonically assigned task:<id> groups; debug strings are display-only.
• BuildExecPlan collapses physical lanes and contracts plumbing using stable stage/input IDs, producing an exporter-neutral typed DAG.
• The exec JSON codec indexes kernels by ExecStageId; stage entrypoints and physical graph artifacts therefore refer to kernels emitted by the same scheduler lowering. Sort stages take radixKeys directly from retained kernel metadata. Nothing is lowered or compiled twice.