Struct CompilePipeline 

pub struct CompilePipeline {
    pub target: FusionTarget,
    pub opts: FusionOptions,
    pub arena_alignment: usize,
    pub assert_fusion_clean: bool,
    pub supported_ops: Option<&'static [OpKind]>,
    pub kernel_dispatch: KernelDispatchConfig,
}

End-to-end compiler pipeline configuration.

Fields

target: FusionTarget

opts: FusionOptions

arena_alignment: usize

assert_fusion_clean: bool

When true, [compile_hir] / [compile_graph] panic if fusion diagnostics report missed block-level patterns.

supported_ops: Option<&'static [OpKind]>

Backend op claim set. When Some and non-empty, fusion passes are gated on these kinds and the optimized graph is legalized afterward. When None, supported_for_target is used.

kernel_dispatch: KernelDispatchConfig

Native vs common IR lowering for logical kernels (see rlx_ir::logical_kernel).

Implementations

impl CompilePipeline

pub fn new(target: FusionTarget) -> CompilePipeline

pub fn with_assert_fusion_clean(self, assert: bool) -> CompilePipeline

pub fn lower_hir(hir: HirModule) -> Result<MirModule, LowerError>

HIR → MIR (block lowering only).

pub fn preprocess_mir(mir: MirModule) -> MirModule

Optional cleanup before fusion (DCE + control-flow lowering).

pub fn with_supported_ops(self, ops: &'static [OpKind]) -> CompilePipeline

pub fn with_kernel_dispatch( self, policy: KernelDispatchPolicy, ) -> CompilePipeline

pub fn with_kernel_dispatch_config( self, config: KernelDispatchConfig, ) -> CompilePipeline

pub fn optimize_with_report(&self, mir: MirModule) -> (MirModule, FusionReport)

Run fusion + cleanup passes on MIR, returning fusion diagnostics.

pub fn optimize(&self, mir: MirModule) -> MirModule

Run fusion + cleanup passes on MIR.

pub fn plan_lir(&self, mir: MirModule) -> LirModule

MIR → LIR (memory plan + schedule + phases + I/O manifest).

pub fn plan_lir_with_options( &self, mir: MirModule, opts: MemoryPlanOptions, ) -> LirModule

MIR → LIR with explicit boundary allocation policy.

pub fn specialize_lir(&self, lir: &LirModule, binding: &DimBinding) -> LirModule

Bind symbolic dims and re-run buffer planning on specialized MIR.

pub fn compile_hir(&self, hir: HirModule) -> Result<CompileResult, LowerError>

HIR → LIR in one call with fusion report.

pub fn compile_mir(&self, mir: MirModule) -> CompileResult

Legacy MIR entry: optimize + plan with fusion report.

pub fn compile_graph(&self, graph: Graph) -> CompileResult

Legacy entry: optimize an existing graph and plan buffers.

pub fn compile_module( &self, module: GraphModule, ) -> Result<CompileResult, LowerError>

Unified entry for GraphModule at any pipeline stage.

impl CompilePipeline

pub fn compile_training( &self, module: GraphModule, wrt: &[NodeId], ) -> Result<TrainingCompileResult, TrainingCompileError>

HIR/MIR → forward LIR (fused) + backward LIR (AD + cleanup), shared weights.

pub fn compile_training_mir( &self, mir: MirModule, wrt: &[NodeId], ) -> TrainingCompileResult

MIR → forward LIR (fused) + backward LIR (AD + cleanup), shared weights.

pub fn optimize_backward(&self, mir: MirModule) -> MirModule

Cleanup passes for backward MIR — no forward-only fusion.

Trait Implementations

impl Clone for CompilePipeline

fn clone(&self) -> CompilePipeline

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for CompilePipeline

fn fmt(&self, f: &mut Formatter<'_>) -> Result<(), Error>

Formats the value using the given formatter.

impl Default for CompilePipeline

fn default() -> CompilePipeline

Returns the “default value” for a type.

impl Copy for CompilePipeline