Skip to content

EGCL v0.0.4

Choose a tag to compare

@github-actions github-actions released this 10 Oct 13:41
· 662 commits to main since this release
6559fc9

0.0.4 - 2026-10-10

Performance and native compilation

  • Reduce native call overhead on Linux x86-64: eligible T1 and T2 functions use the new native ABI by default, removing exceptional-status checks after successful returns while preserving moving GC roots, multiple values, speculative arithmetic, recursion and guard recovery. Unsupported bodies use an interpreter adapter; existing native T1 code remains when T2 declines. Temporary limitation: current-frame OSR is disabled on Linux x86-64, so unsupported bodies and long loops in cold functions may run slower (#204, #207, #210).
  • Optimize named calls, builtin calls, GETHASH, nested FUNCALL callbacks, captured-variable access, simple-string access and missed hash-table lookups on supported compiled paths. Retain replacement checks, multiple values, GC safety and error recovery (#142, #145, #149, #150, #151, #152, #153, #155).
  • Extend compiler support for captured lexical reads and writes, shared compilation across closures, nested callbacks and retained-body leaf inlining, preserving lexical RETURN-FROM and captured state during deoptimization. Some of these paths still require the legacy compiler; Linux x86-64 uses interpreter fallback where mapped native support is absent (#146, #147, #148, #175).
  • Improve numeric guards and recovery on supported compiled paths, including nullable values and bignums. Remove redundant multiple-value resets and loop tag checks, retain only live deoptimization locals, and preserve pending recursive operands. Repair x86-64 OSR recovery metadata for future use; Linux x86-64 OSR remains disabled in this release. Decline unsupported checked RISC-V OSR entries safely (#100, #103, #106, #131, #135, #144).

Compatibility

  • Saved bytecode now uses version 2.0 and core images use format 9. Recompile older FASLs and rebuild older images. Saved functions retain their definitions and lexical captures across redefinition, aliases, unbinding and restoration (#174, #202, #204).

Calling and control-flow correctness

  • Keep each function's selected definition, captures and compiled code together across concurrent redefinition and tier promotion, including callable-instance replacements and saved source-backed functions. Preserve closure function and control scopes through installation and image restoration, and resolve symbol FUNCALL/APPLY targets in the global function namespace (#148, #192, #193, #204).
  • Honor replacement builtins, readers and memory fences in interpreted, compiled and saved code, including lexical VALUES and BYTE, LDB, TYPEP, FIRST-CHAR and returning ERROR replacements. Preserve replacement functions' multiple values and nonlocal exits (#194, #195, #196, #197, #202).
  • Prevent stale secondary values from leaking through optimized primitives, predicates, accessors and numeric operations. Preserve single-float T2 comparisons and correct spilled memory-barrier results (#132, #198, #199).
  • Preserve the exact restart selected by CERROR, even when an outer restart is also named CONTINUE; restore special bindings before cleanup and destination clauses, including transfers that replace an earlier unwind (#201).
  • Fix EQUALP hash-table operations across integer and signed single-float zero, and signal TYPE-ERROR for improper property lists passed to GETF (#89, #154).

Native exceptional transfers

These transfers are enabled by default on x86-64 Linux. Full compiled-shape coverage and the rollout on other platforms remain incomplete.

  • Keep eligible recursive calls, calls between T2 definitions, FUNCALL, APPLY and protected functions in native segments. Preserve live linkage, callable identity, cleanup, restarts, multiple values and moving arguments across legacy calls, Rust reentry and fallback (#157, #164, #165, #166, #171, #172).
  • Compile fixnum arithmetic and recover speculative guards without replaying effects or losing callers. Reduce root bookkeeping and unnecessary fallback; support rooted heap literals and debugger arguments. Keep rejection caches and deoptimization feedback tied to the correct definition (#158, #162, #168, #187, #189, #194).
  • Preserve continuations and roots across fiber suspension and migration, validate destination-worker compatibility, and release retained execution state when it stops. Preserve all ordinary and interactive restart return values. DISASSEMBLE distinguishes cached transfer code and the actual ABI of installed code (#156, #159, #160, #168).

Applications and deployment

  • Embed read-only build-host files into saved images with egcl-ext:embed-file and ASDF components from egcl-embed-asdf, so applications can read them after deployment without the original files (#173).

  • Rename the tree-shaking command to --shake, its manifest to egcl-shake-manifest, and its ASDF extension to egcl-shake-asdf (#163).

  • Describe shaken executables in ASDF system definitions with egcl-shake-asdf; asdf:make writes the specification, saves the core and runs the tree shaker to produce the executable (#161).

Platform support

  • Restore RISC-V loop promotion into T2 by checking imported fixnum values when entering optimized loops, with recovery for failed checks (#140).
  • Add opt-in published native callable entries for Rust-to-Lisp calls on AArch64 Unix with EGCL_NATIVE_TRANSFER=1 (#206).
  • Decode RISC-V native instructions in process for DISASSEMBLE, including compressed instructions, ABI register names and branch targets (#136).
  • Decode POWER native instructions in process for DISASSEMBLE, with branch targets, OSR entries, runtime helper names and tagged constants (#139).
  • Windows x86-64: add opt-in native callable entry boundaries with rooted arguments and Windows unwind metadata. Enable with EGCL_NATIVE_TRANSFER=1; unsupported process mitigation policies produce an explicit error. Compiled bodies retain their existing ABI behind the boundary (#205).
  • s390x: support ELF-ABI foreign calls and Lisp callbacks, including floating point, narrow integers, large argument lists and supported variadic calls (#119, #133, #141).
  • s390x: reduce T1/T2 direct-call, recursive-call and builtin-call overhead; improve fixnum guards, constants and register saving. Direct recursive deoptimization preserves arguments and backtraces; wrong-arity T2 self-calls signal PROGRAM-ERROR on every target (#77, #80, #88, #94, #97, #107, #108, #111, #113, #115, #122).
  • s390x: compile more predicates, simple-string operations, closure factories, fixnum multiplication and memory barriers, including saved bytecode. Improve compilation diagnostics and increase the default CL stack to 4 MiB (512 KiB elsewhere) (#91, #99, #101, #102, #118, #126, #127).
  • ppc64le: correct T2 single-float arithmetic; inline more baseline numeric operations and optimize helper calls. Enable T2 loops, more predicates and bit operations, timely loop polling, and live T1-to-T2 loop replacement (#95, #96, #98, #104, #105, #110, #114).
  • riscv64: add baseline and optimizing native compilers with live OSR and deoptimization, native fiber switching, scalar LP64D foreign calls including variadics, recoverable null/stack-guard faults, and an opt-in native segment boundary that preserves integer and floating-point registers (#93, #112, #120, #121, #123, #128).

Quicklisp and library compatibility

  • Add an x86-64 Linux Quicklisp client port with pinned setup instructions, native networking and filesystem adapters, fresh installation and offline reload. Honor custom *MACROEXPAND-HOOK* functions, accept file streams in PROBE-FILE, and fix wildcard-directory traversal for distribution discovery (#86, #87, #90, #92).