← all workloads

ir_builder

One enormous straight-line function: `n` trivial integer SSA ops, almost no semantic work. Stresses IR instruction construction / dedup / hash-cons and the SSA simplifier (the shared IRBuilder layer the v2026.7 release regressed). Scaling null: n scales straight-line op count; ideal cost is O(n) with amortized-constant hash-consing, so super-linear sweep growth means IR dedup/use-list structures scale worse than the code.

bucket: ir_infra  ·  mode: target  ·  flags: -target spirv -emit-spirv-directly

Phase composition vs N (stacked sub-counters)

compileInner split into phase buckets (named leaves + (self) residuals) stacked across the sweep sizes — the top edge is compileInner, so you can see which phase drives the scaling.

ir_builder — phase composition vs N (v2026.5, median ms) ir_builder 20.9× over N 500→4000 0.0 525 1051 500 1000 2000 4000 N ir_builder — parseTranslationUnit ir_builder — SemanticChecking ir_builder — generateIR ir_builder — frontEndExecute (self) ir_builder — specializeModule ir_builder — simplifyIR ir_builder — linkIR ir_builder — unrollLoopsInModule ir_builder — legalizeResourceTypes ir_builder — legalizeExistentialTypeLayout ir_builder — performMandatoryEarlyInlining ir_builder — performForceInlining ir_builder — linkAndOptimizeIR (self) ir_builder — generateOutput (self) ir_builder — compileInner (self) phase buckets parseTranslationUnit SemanticChecking generateIR frontEndExecute (self) specializeModule simplifyIR linkIR unrollLoopsInModule legalizeResourceTypes legalizeExistentialTypeLayout performMandatoryEarlyInlining performForceInlining linkAndOptimizeIR (self) emitEntryPointsSourceFromIR generateOutput (self) compileInner (self)

Scaling analysis

floor-subtracted power-law fit (t − floor) = a·Nk; floor = the minimal workload (fixed per-compile cost), k the global exponent, top-2× the local high-end doubling ratio.

N rangefloor (ms)k (work)fit R²t(Nmin)t(Nmax)top-2×
500–400091.560.990479733.64×

Growth attribution (N=500 → N=4000)

compileInner grows by 926 ms across the sweep; the mutually-exclusive phase buckets below partition that growth exactly (no nested-timer double counting). × lin is the same metric as the top-level panels, per bucket: the end point vs a linear expectation anchored to the bucket's share of the minimal floor and fitted on the low-N half — 1.0 = grew exactly linearly, >1 bends up. The super-linearity lives where × lin (and k) are red.

buckett@N=500t@N=4000Δ msshare× lin∝Nk
generateOutput (self)9306+29732%3.77×1.83
legalizeExistentialTypeLayout5296+29131%4.36×1.96
legalizeResourceTypes5290+28531%4.39×1.94
SemanticChecking2248+263%0.48×0.44

Near-constant (≤2% of growth each): generateIR (1→9 ms), linkAndOptimizeIR (self) (1→8 ms), simplifyIR (1→8 ms), specializeModule (0→3 ms), linkIR (1→2 ms), parseTranslationUnit (0→2 ms), frontEndExecute (self) (0→0 ms), performForceInlining (0→0 ms), compileInner (self) (0→0 ms), performMandatoryEarlyInlining (0→0 ms), unrollLoopsInModule (0→0 ms).

Sweep numbers (median ms)

NcompileInnergenerateIRsimplifyIR
5004711
10009522
200026754
400097398