CalcKernel v0.15.2

Changelog

All notable user-visible changes to CalcKernel are recorded here.

0.15.2 — WebAssembly Performance and Correctness Update

This patch release improves eligible WebAssembly kernels while preserving the
existing language, CLI, public ABI, strict floating-point behavior, and
baseline/simd128 feature boundaries. Unsupported cases retain their safe
scalar paths.

• Adds independently verified vector and scalar lowering for eligible matrix, stencil, and
  other numerical kernels. The optimizations preserve input order, bounds and
  alias guards, integer wraparound, and strict floating-point evaluation.
• Development-candidate evidence from the strict 11-kernel suite records two
  independent seven-round runs. Each run contains 66 backend/profile/workload
  results with all correctness checks passing and all profile comparisons
  eligible. Against the faster equivalent Clang/Rust WebAssembly
  implementation, baseline throughput geometric means were 1.117× and 1.118×
  (slowest cases 0.920× and 0.914×); SIMD128 means were 1.016× and 1.008×
  (slowest cases 0.923× and 0.920×). The reports declare a clean source checkout at commit
  08292f18b6374f9636bae5ab07fbfcb8e9e30091 and ckc binary SHA-256
  5db7c4db4146dda00020941f1047a1ade794741374e9446af4b2669de8979e3b, and
  report the candidate as ckc 0.15.1; report SHA-256 values are
  ed56abd762d3d02404ae92b09ce213f8e8ef95f378020dea0f4b7355f302351e and
  4384dc2262a4a967a40d1b6f04096235def2b51408c12ae9c363c7946243378b. The
  binary-to-source relation is not independently verifiable; the exact binary
  hash is authoritative. These are development-candidate results, not measurements of the v0.15.2 release
  archives. See the performance guide
  for methodology and currently published measurements.
• The hot-call comparison excludes compilation, process startup, fixture I/O,
  WebAssembly instantiation, memory growth, and report generation. In the same
  candidate runs, baseline and SIMD128 modules were 28,325 and 34,941 bytes;
  measured module compilation was about 0.16–0.18 ms. First-call costs remain
  workload-dependent (for example, baseline matmul was about 43.95 ms and
  SIMD128 matmul about 20.58 ms). These cold costs and module sizes are reported
  separately from hot throughput and are not claimed as parity results.

0.15.1

The first published 0.15 release carries forward the compiler and WebAssembly
features from the immutable v0.15.0 tagged source snapshot. It also fixes the
release workflow's validation of annotated tags while retaining the exact
tagged source and all existing release gates.

• Includes the opt-in WebAssembly simd128 profile and O3 Bulk Memory paths
  for eligible copies and byte fills. Unsupported cases retain scalar code.
• Includes direct verified-KIR lowering to WAT and binary modules, while
  preserving exported function signatures and caller-owned memory. The
  ck.wasm.target custom-section metadata uses schema 2.
• In local Node.js 24.14.0/V8 tests on Apple M5 Max, O3 simd128 maps
  measured 3.1–6.0× Node.js for i32 and 2.2–2.8× for f64 across
  4,096–65,536 elements. These hot-kernel measurements exclude compilation,
  instantiation, and data preparation; they describe one machine and runtime,
  not general Wasm performance.

0.15.0 — Tagged source snapshot; not released

The immutable v0.15.0 tag records this source history. Its workflow did not
create an official GitHub Release or publish compiler archives or checksums;
the source is carried forward in 0.15.1.

• Added an opt-in WebAssembly simd128 profile. At O3, eligible contiguous
  slice maps and modulo-2^32 i32/u32 sum and product reductions can use SIMD;
  unsupported cases and scalar remainders retain their original behavior.
• Enabled Bulk Memory in the default baseline and opt-in simd128 profiles.
  O3 can use guarded memory.copy and memory.fill paths for eligible copies
  and byte fills, with the original scalar loop as fallback.
• WebAssembly emission now lowers verified KIR directly to WAT and binary
  modules, with structured control flow, direct binary encoding, and improved
  loop-address lowering.
• Preserved exported function signatures and caller-owned memory. The
  ck.wasm.target custom-section metadata advances to schema 2 to identify the
  selected Wasm feature profile.
• In local Node.js 24.14.0/V8 tests on Apple M5 Max, O3 simd128 maps
  measured 3.1–6.0× Node.js for i32 and 2.2–2.8× for f64 across
  4,096–65,536 elements. These
  hot-kernel measurements exclude compilation, instantiation, and data
  preparation; they describe one machine and runtime, not general Wasm
  performance.

0.14.0

• Corrected the PGO-generation library flush status when its collection
  directory becomes unwritable after build: it now reports directory failure
  (43) instead of retrying sixteen false name collisions and reporting write
  failure (44). Successful shards and genuine collision retries remain intact;
  public Native C ABI 1, Runtime ABI 2, and profile wire formats do not change.
• Retained 0.13.0 ordinary compilation, PGO, and multiversion behavior on the
  same six native platforms. Offline Auto-Tuning is deferred: ckc tune and
  ckc build --tune-use fail explicitly without output side effects. This
  version does not claim a new optimizer speedup or passing tuning gates.

0.13.0

• Added deterministic CK-owned CKPART01 shards and CKPROF01 workload
  profiles, with directory-safe collection, canonical merge/inspection, and the
  transactional ckc pgo build convenience workflow.
• Added non-proof profile analysis and independently checked O2 late machine
  layout plus O3 guarded specialization, inlining, unrolling, SLP, and Loop SIMD
  decisions. Profiles may affect profitability but never establish safety.
• Added explicit --cpu multiversion Native builds with one portable baseline,
  bounded verified feature variants, a baseline-safe process-local detector,
  stable public thunks, and executable/dynamic/static named-object assembly.
• Added real library-generation workflows with the full-identity
  ck_profile_flush_* control symbol. Final profile-use artifacts contain no
  counters, profile paths, writer, or generation runtime.
• Advanced the private LLVM bridge to ABI 4, KIR to v3, and the Native object
  cache to CKCOBJ03 key/manifest schema 4. Public Native C ABI 1 and Runtime
  ABI 2 remain unchanged; 0.12 source and observable semantics remain accepted.
• Added closed profile/target/dispatch/cache identities, corruption and mutation
  tests, transactional multi-file output, and schema-8 performance and exact-SHA
  release audits.
• Auto-Tuning, indirect-call promotion, scalable KIR vectors, and adaptive JIT
  PGO remained future work.

0.12.0 - Unreleased

• Added KIR v2 fixed-vector and mask instructions plus deterministic Native
  KirTargetProfile capability/cost data derived from pinned LLVM 22.1.8.
• Added transactional, independently checked O3 specialization, controlled
  full/partial unrolling, SLP, and Loop SIMD frontiers with monotonic analysis
  budgets and stable optimization explanations.
• Added unit-stride Loop SIMD for integer and strict element-wise f64 arithmetic,
  supported casts, pure compare/select diamonds, splats, and contiguous memory;
  strict f64 keeps per-element ordering and never enables fast math.
• Added total runtime alias versioning with the unchanged scalar loop as the
  fallback, scalar epilogues, and exact unchecked modular u32 add/multiply
  reductions. Checked failures, effects, unsupported recurrences, scans, C,
  and WebAssembly remain scalar.
• Advanced the private LLVM bridge to ABI 3, KIR identity to kir-v2, and the
  Native object cache to CKCOBJ02 key/manifest schema 3 while retaining public
  Native C ABI 1, Runtime ABI 2, source syntax, diagnostics, and checked
  first-error behavior.
• Added KIR/pre-LLVM/object structural evidence, fixed-seed O0/O3 differential
  coverage, mutation tests, target-feature containment, and schema-7 performance
  gate inputs. PGO/multiversioning and Auto-Tuning remain future work.

0.11.0 - Unreleased

• Added explicit unsafe fn contracts for affine range requirements,
  multiple_of, noalias, alignment, and slice memory-effect ceilings. Unsafe
  calls require an unsafe { ... } statement and executable main remains safe.
• Added deterministic emit-kir inspection, verified facts/effect summaries,
  proof-carrying guard elimination explanations, and opt-in Native contract
  sanitization with CKR0007.
• Replaced the former target-neutral MIR optimizer with one verified KIR
  pipeline shared by C, WebAssembly, and Native LLVM. Semantic MIR and stable
  emit-mir output remain the source-order and first-error boundary.
• Added scalar/path, region alias and Memory SSA, interprocedural effect, loop,
  GVN/load-forwarding/dead-store, LICM, and evidence-audited backend facts.
• Kept Native C ABI 1 while advancing the private LLVM bridge and runtime ABI
  to 2 and using the KIR v1 native cache/code-generation identity.
• Added fixed-seed differential and mutation suites, pre-LLVM fact audits, and
  performance gates against both pinned Clang and exact CalcKernel 0.10.

0.10.0 - 2026-08-27

• Added parameterless internal main, ckc run, and Native executable output.
• Added deterministic Native output builtins for signed/unsigned integers,
  f64, booleans, and newline; library, C, and WASM roots reject reachable print.
• Replaced product Clang subprocesses with pinned LLVM 22.1.8 structural code
  generation, ORC execution, archive writing, and in-process LLD linking.
• Expanded ckc build --kind to executable, dynamic, static, and object outputs;
  dynamic remains the default and build-llvm is a deprecated compatibility alias.
• Unified Native object/static/dynamic exports under one generated-header Native
  C ABI with ...

CalcKernel v0.15.1

Changelog

All notable user-visible changes to CalcKernel are recorded here.

0.15.1

The first published 0.15 release carries forward the compiler and WebAssembly
features from the immutable v0.15.0 tagged source snapshot. It also fixes the
release workflow's validation of annotated tags while retaining the exact
tagged source and all existing release gates.

• Includes the opt-in WebAssembly simd128 profile and O3 Bulk Memory paths
  for eligible copies and byte fills. Unsupported cases retain scalar code.
• Includes direct verified-KIR lowering to WAT and binary modules, while
  preserving exported function signatures and caller-owned memory. The
  ck.wasm.target custom-section metadata uses schema 2.
• In local Node.js 24.14.0/V8 tests on Apple M5 Max, O3 simd128 maps
  measured 3.1–6.0× Node.js for i32 and 2.2–2.8× for f64 across
  4,096–65,536 elements. These hot-kernel measurements exclude compilation,
  instantiation, and data preparation; they describe one machine and runtime,
  not general Wasm performance.

0.15.0 — Tagged source snapshot; not released

The immutable v0.15.0 tag records this source history. Its workflow did not
create an official GitHub Release or publish compiler archives or checksums;
the source is carried forward in 0.15.1.

• Added an opt-in WebAssembly simd128 profile. At O3, eligible contiguous
  slice maps and modulo-2^32 i32/u32 sum and product reductions can use SIMD;
  unsupported cases and scalar remainders retain their original behavior.
• Enabled Bulk Memory in the default baseline and opt-in simd128 profiles.
  O3 can use guarded memory.copy and memory.fill paths for eligible copies
  and byte fills, with the original scalar loop as fallback.
• WebAssembly emission now lowers verified KIR directly to WAT and binary
  modules, with structured control flow, direct binary encoding, and improved
  loop-address lowering.
• Preserved exported function signatures and caller-owned memory. The
  ck.wasm.target custom-section metadata advances to schema 2 to identify the
  selected Wasm feature profile.
• In local Node.js 24.14.0/V8 tests on Apple M5 Max, O3 simd128 maps
  measured 3.1–6.0× Node.js for i32 and 2.2–2.8× for f64 across
  4,096–65,536 elements. These
  hot-kernel measurements exclude compilation, instantiation, and data
  preparation; they describe one machine and runtime, not general Wasm
  performance.

0.14.0

• Corrected the PGO-generation library flush status when its collection
  directory becomes unwritable after build: it now reports directory failure
  (43) instead of retrying sixteen false name collisions and reporting write
  failure (44). Successful shards and genuine collision retries remain intact;
  public Native C ABI 1, Runtime ABI 2, and profile wire formats do not change.
• Retained 0.13.0 ordinary compilation, PGO, and multiversion behavior on the
  same six native platforms. Offline Auto-Tuning is deferred: ckc tune and
  ckc build --tune-use fail explicitly without output side effects. This
  version does not claim a new optimizer speedup or passing tuning gates.

0.13.0

• Added deterministic CK-owned CKPART01 shards and CKPROF01 workload
  profiles, with directory-safe collection, canonical merge/inspection, and the
  transactional ckc pgo build convenience workflow.
• Added non-proof profile analysis and independently checked O2 late machine
  layout plus O3 guarded specialization, inlining, unrolling, SLP, and Loop SIMD
  decisions. Profiles may affect profitability but never establish safety.
• Added explicit --cpu multiversion Native builds with one portable baseline,
  bounded verified feature variants, a baseline-safe process-local detector,
  stable public thunks, and executable/dynamic/static named-object assembly.
• Added real library-generation workflows with the full-identity
  ck_profile_flush_* control symbol. Final profile-use artifacts contain no
  counters, profile paths, writer, or generation runtime.
• Advanced the private LLVM bridge to ABI 4, KIR to v3, and the Native object
  cache to CKCOBJ03 key/manifest schema 4. Public Native C ABI 1 and Runtime
  ABI 2 remain unchanged; 0.12 source and observable semantics remain accepted.
• Added closed profile/target/dispatch/cache identities, corruption and mutation
  tests, transactional multi-file output, and schema-8 performance and exact-SHA
  release audits.
• Auto-Tuning, indirect-call promotion, scalable KIR vectors, and adaptive JIT
  PGO remained future work.

0.12.0 - Unreleased

• Added KIR v2 fixed-vector and mask instructions plus deterministic Native
  KirTargetProfile capability/cost data derived from pinned LLVM 22.1.8.
• Added transactional, independently checked O3 specialization, controlled
  full/partial unrolling, SLP, and Loop SIMD frontiers with monotonic analysis
  budgets and stable optimization explanations.
• Added unit-stride Loop SIMD for integer and strict element-wise f64 arithmetic,
  supported casts, pure compare/select diamonds, splats, and contiguous memory;
  strict f64 keeps per-element ordering and never enables fast math.
• Added total runtime alias versioning with the unchanged scalar loop as the
  fallback, scalar epilogues, and exact unchecked modular u32 add/multiply
  reductions. Checked failures, effects, unsupported recurrences, scans, C,
  and WebAssembly remain scalar.
• Advanced the private LLVM bridge to ABI 3, KIR identity to kir-v2, and the
  Native object cache to CKCOBJ02 key/manifest schema 3 while retaining public
  Native C ABI 1, Runtime ABI 2, source syntax, diagnostics, and checked
  first-error behavior.
• Added KIR/pre-LLVM/object structural evidence, fixed-seed O0/O3 differential
  coverage, mutation tests, target-feature containment, and schema-7 performance
  gate inputs. PGO/multiversioning and Auto-Tuning remain future work.

0.11.0 - Unreleased

• Added explicit unsafe fn contracts for affine range requirements,
  multiple_of, noalias, alignment, and slice memory-effect ceilings. Unsafe
  calls require an unsafe { ... } statement and executable main remains safe.
• Added deterministic emit-kir inspection, verified facts/effect summaries,
  proof-carrying guard elimination explanations, and opt-in Native contract
  sanitization with CKR0007.
• Replaced the former target-neutral MIR optimizer with one verified KIR
  pipeline shared by C, WebAssembly, and Native LLVM. Semantic MIR and stable
  emit-mir output remain the source-order and first-error boundary.
• Added scalar/path, region alias and Memory SSA, interprocedural effect, loop,
  GVN/load-forwarding/dead-store, LICM, and evidence-audited backend facts.
• Kept Native C ABI 1 while advancing the private LLVM bridge and runtime ABI
  to 2 and using the KIR v1 native cache/code-generation identity.
• Added fixed-seed differential and mutation suites, pre-LLVM fact audits, and
  performance gates against both pinned Clang and exact CalcKernel 0.10.

0.10.0 - 2026-08-27

• Added parameterless internal main, ckc run, and Native executable output.
• Added deterministic Native output builtins for signed/unsigned integers,
  f64, booleans, and newline; library, C, and WASM roots reject reachable print.
• Replaced product Clang subprocesses with pinned LLVM 22.1.8 structural code
  generation, ORC execution, archive writing, and in-process LLD linking.
• Expanded ckc build --kind to executable, dynamic, static, and object outputs;
  dynamic remains the default and build-llvm is a deprecated compatibility alias.
• Unified Native object/static/dynamic exports under one generated-header Native
  C ABI with target ABI classification and checked status thunks.
• Added Native checked overflow and slice bounds, preserving the C CK_Status
  meanings and first-error order.
• Added an isolated run child, secure persistent object cache, fixed runtime
  diagnostics/statuses, eager symbol resolution, and audited JIT page permissions.
• Added checked/unchecked C-oracle performance gates and six-host functional,
  artifact, dependency, provenance, and immutable release gates.
• Reserved main and the seven print builtin names, limited Native target output
  to the host, retired the standalone LLVM exported-shape promise, and kept
  emit-c source-only. See the compatibility policy for migration guidance.

0.9.0 - 2026-08-26

• Added break and continue for structured control inside while loops.
• Added explicit void procedures, empty return;, and procedure-call statements.
• Added non-owning slice<T> values, slice(data, len), indexing, .data / .len,
  and half-open sub-slices written as items[start..end].
• Added optional checked slice bounds to the C backend through --bounds checked;
  unchecked bounds remain the default, while WASM and LLVM reject checked bounds.
• Stabilized native C, WebAssembly, and LLVM output paths and their V0.9 ABIs.
• Reorganized the repository around durable compiler, contract, example, benchmark,
  and test responsibilities without changing the compiler's public behavior.
• Froze the V0.9 compatibility boundary: patch releases in the 0.9.x line preserve
  accepted source, diagnostic identifiers, CLI behavior, textual MIR, and documented
  ABI contracts. A later 0.10.0 may make documented breaking changes with migration
  guidance; long-term compatibility begins with a future 1.0.0 release.
• Added signed-off native ckc release archives and SHA-256 checksums for macOS,
  Linux, and Windows on both arm64 and x64.

CalcKernel v0.14.0

Changelog

All notable user-visible changes to CalcKernel are recorded here.

0.14.0

• Corrected the PGO-generation library flush status when its collection
  directory becomes unwritable after build: it now reports directory failure
  (43) instead of retrying sixteen false name collisions and reporting write
  failure (44). Successful shards and genuine collision retries remain intact;
  public Native C ABI 1, Runtime ABI 2, and profile wire formats do not change.
• Retained 0.13.0 ordinary compilation, PGO, and multiversion behavior on the
  same six native platforms. Offline Auto-Tuning is deferred: ckc tune and
  ckc build --tune-use fail explicitly without output side effects. This
  version does not claim a new optimizer speedup or passing tuning gates.

0.13.0

• Added deterministic CK-owned CKPART01 shards and CKPROF01 workload
  profiles, with directory-safe collection, canonical merge/inspection, and the
  transactional ckc pgo build convenience workflow.
• Added non-proof profile analysis and independently checked O2 late machine
  layout plus O3 guarded specialization, inlining, unrolling, SLP, and Loop SIMD
  decisions. Profiles may affect profitability but never establish safety.
• Added explicit --cpu multiversion Native builds with one portable baseline,
  bounded verified feature variants, a baseline-safe process-local detector,
  stable public thunks, and executable/dynamic/static named-object assembly.
• Added real library-generation workflows with the full-identity
  ck_profile_flush_* control symbol. Final profile-use artifacts contain no
  counters, profile paths, writer, or generation runtime.
• Advanced the private LLVM bridge to ABI 4, KIR to v3, and the Native object
  cache to CKCOBJ03 key/manifest schema 4. Public Native C ABI 1 and Runtime
  ABI 2 remain unchanged; 0.12 source and observable semantics remain accepted.
• Added closed profile/target/dispatch/cache identities, corruption and mutation
  tests, transactional multi-file output, and schema-8 performance and exact-SHA
  release audits.
• Auto-Tuning, indirect-call promotion, scalable KIR vectors, and adaptive JIT
  PGO remained future work.

0.12.0 - Unreleased

• Added KIR v2 fixed-vector and mask instructions plus deterministic Native
  KirTargetProfile capability/cost data derived from pinned LLVM 22.1.8.
• Added transactional, independently checked O3 specialization, controlled
  full/partial unrolling, SLP, and Loop SIMD frontiers with monotonic analysis
  budgets and stable optimization explanations.
• Added unit-stride Loop SIMD for integer and strict element-wise f64 arithmetic,
  supported casts, pure compare/select diamonds, splats, and contiguous memory;
  strict f64 keeps per-element ordering and never enables fast math.
• Added total runtime alias versioning with the unchanged scalar loop as the
  fallback, scalar epilogues, and exact unchecked modular u32 add/multiply
  reductions. Checked failures, effects, unsupported recurrences, scans, C,
  and WebAssembly remain scalar.
• Advanced the private LLVM bridge to ABI 3, KIR identity to kir-v2, and the
  Native object cache to CKCOBJ02 key/manifest schema 3 while retaining public
  Native C ABI 1, Runtime ABI 2, source syntax, diagnostics, and checked
  first-error behavior.
• Added KIR/pre-LLVM/object structural evidence, fixed-seed O0/O3 differential
  coverage, mutation tests, target-feature containment, and schema-7 performance
  gate inputs. PGO/multiversioning and Auto-Tuning remain future work.

0.11.0 - Unreleased

• Added explicit unsafe fn contracts for affine range requirements,
  multiple_of, noalias, alignment, and slice memory-effect ceilings. Unsafe
  calls require an unsafe { ... } statement and executable main remains safe.
• Added deterministic emit-kir inspection, verified facts/effect summaries,
  proof-carrying guard elimination explanations, and opt-in Native contract
  sanitization with CKR0007.
• Replaced the former target-neutral MIR optimizer with one verified KIR
  pipeline shared by C, WebAssembly, and Native LLVM. Semantic MIR and stable
  emit-mir output remain the source-order and first-error boundary.
• Added scalar/path, region alias and Memory SSA, interprocedural effect, loop,
  GVN/load-forwarding/dead-store, LICM, and evidence-audited backend facts.
• Kept Native C ABI 1 while advancing the private LLVM bridge and runtime ABI
  to 2 and using the KIR v1 native cache/code-generation identity.
• Added fixed-seed differential and mutation suites, pre-LLVM fact audits, and
  performance gates against both pinned Clang and exact CalcKernel 0.10.

0.10.0 - 2026-08-27

• Added parameterless internal main, ckc run, and Native executable output.
• Added deterministic Native output builtins for signed/unsigned integers,
  f64, booleans, and newline; library, C, and WASM roots reject reachable print.
• Replaced product Clang subprocesses with pinned LLVM 22.1.8 structural code
  generation, ORC execution, archive writing, and in-process LLD linking.
• Expanded ckc build --kind to executable, dynamic, static, and object outputs;
  dynamic remains the default and build-llvm is a deprecated compatibility alias.
• Unified Native object/static/dynamic exports under one generated-header Native
  C ABI with target ABI classification and checked status thunks.
• Added Native checked overflow and slice bounds, preserving the C CK_Status
  meanings and first-error order.
• Added an isolated run child, secure persistent object cache, fixed runtime
  diagnostics/statuses, eager symbol resolution, and audited JIT page permissions.
• Added checked/unchecked C-oracle performance gates and six-host functional,
  artifact, dependency, provenance, and immutable release gates.
• Reserved main and the seven print builtin names, limited Native target output
  to the host, retired the standalone LLVM exported-shape promise, and kept
  emit-c source-only. See the compatibility policy for migration guidance.

0.9.0 - 2026-08-26

• Added break and continue for structured control inside while loops.
• Added explicit void procedures, empty return;, and procedure-call statements.
• Added non-owning slice<T> values, slice(data, len), indexing, .data / .len,
  and half-open sub-slices written as items[start..end].
• Added optional checked slice bounds to the C backend through --bounds checked;
  unchecked bounds remain the default, while WASM and LLVM reject checked bounds.
• Stabilized native C, WebAssembly, and LLVM output paths and their V0.9 ABIs.
• Reorganized the repository around durable compiler, contract, example, benchmark,
  and test responsibilities without changing the compiler's public behavior.
• Froze the V0.9 compatibility boundary: patch releases in the 0.9.x line preserve
  accepted source, diagnostic identifiers, CLI behavior, textual MIR, and documented
  ABI contracts. A later 0.10.0 may make documented breaking changes with migration
  guidance; long-term compatibility begins with a future 1.0.0 release.
• Added signed-off native ckc release archives and SHA-256 checksums for macOS,
  Linux, and Windows on both arm64 and x64.