MoonCompiler 1.0.0 - a Delphi-focused Win64/Linux toolchain built on Unleashed

[Moonbot-Tech]

Hi Unleashed Pascal team,

We have just released MoonCompiler 1.0.0, a Delphi-focused toolchain for Win64 and Linux x86-64 built on Unleashed Pascal.

The project began with a concrete production requirement: running the core of a high-load crypto trading terminal on Linux while keeping its Delphi 12.2 codebase. Unleashed already provided much of the modern Delphi language surface we needed and became the foundation of this work. Thank you - it saved us a great deal of time and gave us a strong place to start.

From there, our requirements led to a deliberately narrower product profile than Unleashed: Delphi and Unicode semantics by default, one tested compiler + RTL + memory-manager configuration, and exactly two supported targets - Win64 and Linux x86-64.

Those choices would not fit Unleashed's opt-in and backward-compatibility contract, so MoonCompiler became a separate product-focused fork rather than a product-specific patch stack on your main branch.

MoonCompiler 1.0.0 now includes:

• more than 120 catalogued compiler and RTL correctness, compatibility, and performance repairs;
• a Unicode product RTL and a bundled multithreaded mORMot-based memory manager;
• a conservative optimizer block covering effect-aware LICM, repeated-address elimination, exception-aware register liveness, and x86-64 code placement;
• reproducible Win64 and Linux toolchains;
• a qualification system built from Delphi 12.2 oracles, semantic matrices, generated combinations, real library corpora, and performance benchmarks.

Unleashed attribution remains in the compiler banner, README, history, and licensing.

Repository: https://github.com/Moonbot-Tech/MoonCompiler
Release: https://github.com/Moonbot-Tech/MoonCompiler/releases/tag/v1.0.0

Some of our repairs are independent of MoonCompiler's fixed product profile. If that would be useful, we can rebase and validate candidates individually against the current Unleashed main, then prepare focused PRs with minimal reproducers and regression tests.

Thank you again for giving us such a strong starting point. We would be very interested in your technical feedback on what grew out of it.

[fibodevy]

Thanks for the note and for keeping the attribution.

Regarding upstreaming, a few constraints first. This fork is kept mergeable with upstream FPC with as little friction as possible, so changes must stay small and additive; anything that rewrites larger parts of upstream code will not be taken regardless of merit. Anything that assumes Delphi/Unicode defaults or your RTL/MM setup will not fit either, so please filter those out before opening PRs.

For the rest there are two paths, depending on what a fix touches:

• Bug in FPC itself (reproducible on stock trunk): report it on the FPC GitLab first. Once it has an issue ID, a PR here is fine: the fix plus a stock-style regression test in tests/webtbs, referencing that ID.
• Bug in an Unleashed feature: no upstream report needed. A PR with the patch and a test in unleashed/tests is enough.

Either way, one PR per fix, against the current main, with a minimal reproducer.

Good luck with the release.

[Moonbot-Tech]

Thanks for the clear rules; we follow them: one fix per PR against current main, an FPC issue first, a tests/webtbs regression, nothing tied to Unicode defaults or our RTL/MM.

We cross-checked our fixes against current main (a359fc1) with a stock ANSI build. Twelve open FPC issues still reproduce on main and are fixed in MoonCompiler; each is now a separate PR here, with the reproducer as tests/webtbs/twNNNNN.pp, red on main and green with the patch, and 405 neighbouring tests from tests/test unchanged:

• inlining: copy inlinelevel with the call node (fixes upstream#41184) #40 (FPC 41184), rtti: skip generic methods in the legacy VMT method table (fixes upstream#41410) #41 (41410), helpers: key distinct type aliases by their own identity (fixes upstream#41564) #42 (41564), init: run Initialize for static arrays of managed records (fixes upstream#41451) #45 (41451), x86: normalize immediates when narrowing folded arithmetic (fixes upstream#41480) #46 (41480), helpers: materialize a non-lvalue instance before calling a helper (fixes upstream#41589) #47 (41589), x86: derive the asm memory width of by-reference formals (fixes upstream#41630) #48 (41630), parser: allow readonly storage and results for untyped constref #49 (41766), generics: defer constraint checks for implementation-header parameters (fixes upstream#41770) #50 (41770), generics: defer the VMT of a partial specialization (fixes upstream#41788) #51 (41788);
• generics: keep the with instance for inline specializations (fixes upstream#41711, upstream#41712) #43 (41711 + 41712): an alternative to your !1435 that also covers the member-scope case of 41711;
• cse: drop short-circuit values before an always-evaluated right side (fixes upstream#41809) #44 (41809): an alternative to CuriousKit's !1509; Pierre's reproducer crashes on main and passes with it.

Held back for now: 41678 and 41808 (Delphi-mode semantics, larger diffs), 41031 (your !1427; our one-line variant in symsym.pas passes Pierre's tb0728, which main still fails), and 41612 b/c (implicit generics, your feature, so we would rather ask first).

Two neighbours of your own fixes also reproduce on main: the AND/CMP -> CMP peephole narrows relational compares the way 41824 did for MOVZX/CMP, and MovMovXX folding widens a narrow constant without normalizing its high bits. We will file both in FPC first, then bring them here.

[fibodevy]

Thanks for splitting them up.

After #44 I have to tighten the rule for stock-only fixes, i.e. anything that does not touch an Unleashed feature. A fix merged here and later merged upstream in a slightly different shape conflicts on every sync from then on, and #44 was exactly that: upstream merged its version five hours after you opened the PR.

So for these PRs: open the MR on the FPC GitLab first and link it here. The PR stays open until upstream decides. Merged there: it arrives here with the next sync and I close the PR. Rejected or stalled: I merge the PR here. Reproducer and tests/webtbs test as you have them are fine either way.