Effect C names its own definitions with HID, and a module that uses HID must (#1533) by Lorenzobattistela · Pull Request #2017 · bendlang/bend

Lorenzobattistela · 2026-11-02T19:47:24Z

Fixes #1533 for packages that adopt HID; supersedes #1546.

Problem

-o pastes every effect .c into one C file. A program that imports two versions of one hub package pastes the same C twice, so clang refuses the second definition (redefinition of 'wall_raw_run'), while bend main.bend runs fine. In the real bend-kit-time case it is also silently wrong: the helper sits behind #ifndef TIME_EFFS, so version 2 would quietly use version 1's helper.

Fix

HID(name) in an effect's C is a C id of the module that imports it (c_ids, same pass as CID/FID). The key is the module, not the file: bend_tensors shares a guard and a struct across five .c files of one module.
An imported module whose C says HID (outside comments and strings, found in effect_src's existing scan) is held to it. compile_book pastes its C twice, and the second copy is dead: constructor/destructor become __unused__, and -Wtentative-definition-compat is an error inside it. A name defined without HID therefore fails the author's own build, instead of the first program that meets two versions.
Each copy push/pops the macros it defines, so include guards and macros need no HID.
C without HID, Base, and the root module are emitted exactly as before.

Cost: in the two-version bend-kit-time build, the dead copy adds 136 bytes to the binary and no measurable build time. Runtime is unchanged.

Commits

refactor(comp): sheds 386 ttok from comp.ts. The C and JS emitted for every test are byte-identical to main's.
fix(effects): the feature, tests/io/effect_versions (two versions with the same names, each keeping its own counter: 8 71 9 72 in the interpreter, JS and native), and a paragraph in guide/EFFECTS.md.
refactor(effs): Base's own effect C (bend2/effs/*.c, 281 names) uses HID, so the reference packages copy shows the rule, and a package's names cannot clash with Base's. The three tests/gfx C files that reach into window.c internals say HID too. Base and the program's own file share namespace "", so HID(BendWin) there is Base's. Base's macros keep their names. With HID undone, the diff is token-identical to main, including the Linux X11/ALSA and CUDA paths.

comp.ts ends at 65,987 / 66k.

Verified

Gates on d094958: repo 54/54, ping 81/81, perf 124/124 (one contention-only miss, clean on rerun), test 1822/1822, safe 0 false rejects.
Copies of the two real bend-kit-time versions with HID adopted build and run natively together.
Removing any single HID (helper, global, entry, constructor, tentative global) fails the build, even with a single version imported.
Probes from three review rounds pass: macro names containing push_; __constructor__/destructors in the dead copy; HID( in a comment; external HID_* macros; a package's own unused macro.
Base conversion: every emitted test C syntax-checks on macOS, on macOS with -DBEND_METAL=1 (the 44 ! programs), and on Debian clang 14 with X11/XI/ALSA headers. A throwaway compiler that forces the dead copy on Base reports no unwrapped Base definition on either OS, and covers all 281 names.

Limits

Published packages still clash until they adopt HID and republish.
An HID module's constructors cannot take a priority.
The tentative-global check needs clang 21+.
The test gate cannot express an expected native build failure, so the enforcement was verified by hand.
The dead copy does not apply to Base (namespace "" is shared with the program's own file, which must stay free to skip HID), so a future bare name in bend2/effs is not caught by the build.
Separate translation units were considered and dropped: the effects paste into the runtime and call the evaluator, so it would need a runtime header/body split.

🤖 Generated with Claude Code

Behavior-preserving simplifications to make room under comp.ts's 66k cap: term_use takes a term, single-use helpers inlined (arr_leaf, val_brw), switch defaults return after the switch, set "seen" checks use size, empty checks use !length, Seg flags optional, type and table declarations packed. The C and JS emitted for every test are byte-identical to main's. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ses it must (#1533) Two versions of one hub package paste the same C into one file, so `-o` failed with `redefinition` while the interpreter ran. HID(name) is now a C id of the importing module. An imported module whose C says HID (outside comments and strings) pastes it twice, the second copy dead (constructor and destructor attributes become __unused__), so a name defined without HID fails the author's own build instead of the first program that meets two versions: `redefinition`, or, under -Wtentative-definition-compat made an error there, a repeated global without an initializer. Each copy push/pops the macros it defines, so include guards and macros need no HID. C without HID builds as before. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Every name Base's effect sources define at file scope (functions, the effects and their constructors, globals, types and tags, enum constants, the BendView class) is now HID(name), so bend2/effs is the reference a package's C copies, and a package's functions, globals and types cannot clash with Base's. Base's macros keep their global names, as before. The gfx tests that reach into window.c's internals say HID too; Base and the program's own file share a namespace, so HID(BendWin) there is Base's. EFFECTS.md's examples use HID. No runtime or emitter change; window_dev the CUDA kernel is the runtime's and keeps its name. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

nicolas-abril

All issues are in the dead-copy logic in compile_book/effect_src (cf0d171). I found nothing wrong in the refactor(comp) commit or the Base HID conversion. I reproduced the tentative-definition error with Homebrew clang 22 and the unknown-warning message with Apple clang 17. The rest come from reading the code.

nicolas-abril · 2026-11-03T19:50:51Z

+        .join("") + `\n#pragma clang diagnostic ${d}\n`;
+      return at("push") + c_ids(src, m) + at("pop") + at("push")
+        + dead.replace(/\S+/g, "#define $& __unused__\n")
+        + `#pragma clang diagnostic error "-Wtentative-definition-compat"\n`


Making -Wtentative-definition-compat an error in the dead copy also rejects valid HID code: a static global declared first and defined later.

static const struct N HID(a); static const struct N HID(b) = {&HID(a)}; static const struct N HID(a) = {&HID(b)};

Copy 1 compiles. In copy 2, Homebrew clang 22 reports error: duplicate declaration of 'HID2_..._a' is invalid in C++, so a correct package can't be built natively.

Separately, this pragma isn't guarded. Clang before 21, including Apple clang 17 (the macOS default clang, which cc_find picks first), prints warning: unknown warning group '-Wtentative-definition-compat', ignored once per HID module on every -o build. The check then does nothing. Wrapping it in #if __has_warning("-Wtentative-definition-compat") fixes that.

nicolas-abril · 2026-11-03T19:50:51Z

+        ...src.match(/^[ \t]*#[ \t]*(define|undef)[ \t]+\w+/gm) ?? []]
+        .map((x) => `\n#pragma ${d}_macro("${x.split(/\s/).pop()}")`)
+        .join("") + `\n#pragma clang diagnostic ${d}\n`;
+      return at("push") + c_ids(src, m) + at("pop") + at("push")


push_macro saves an include guard's earlier state but never #undefs it. If the guard is already defined, the HID copy skips its own guarded body.

Example: a program imports bend-kit-time v1 (no HID, #ifndef TIME_EFFS guard) and v2 (HID, same guard, since the guide says guards don't need HID). v1 is in the non-HID part, so it's pasted first and defines TIME_EFFS. v2's copy then skips its own helpers and fails with an undeclared HID_..._helper. If the guarded block only has macros, v2 silently uses v1's, which is #1533 again. Emitting #undef NAME after each push_macro fixes it.

nicolas-abril · 2026-11-03T19:50:51Z

    const src = FL.js ? raw : raw.replace(/\\[ \t]*(\r\n?|\n)/g, "");
    FL.effs.set(path, [m, src]);
    if (!FL.js) c_ids(src, m, (p, q) => {
+      if (p === "HID" && m) FL.hids.add(m);


This skips HID enforcement when the root module imports the package's C (m === ""). So the author's own build never catches a bare name, contrary to the PR's claim. A package author who runs bend -o out lib.bend on the file that does import "./lib.c" (like effect_version_a.bend) gets no dead copy, and a bare static int count; passes. The clash only shows up when a user imports two versions.

nicolas-abril · 2026-11-03T19:50:51Z

      return t;
    }
+    if (p === "HID") {
+      return use ? (use(p, k), t) : name_id(h, m + ":" + k);


HID(k) always resolves in the importing module's namespace, so a package's C can't name Base's effect internals, which are now HID-renamed, or another package's.

A hub package whose C extends windows or channels by naming Base's BendWin, WinQueue or a chan.c helper breaks after the Base conversion. The bare name is undeclared, and HID(BendWin) in the package gives HID_pkg_BendWin, not Base's HID__BendWin. Only the root file (namespace "") can reach Base, so there's no supported way to depend on another module's C.

nicolas-abril · 2026-11-03T19:50:51Z

+  const reqs = effs.map(([m, s]) => FL.hids.has(m) ? "" : c_ids(s, m)).join("")
+    + [...FL.hids].map((m) => {
+      const src = effs.filter((e) => e[0] === m).map((e) => e[1]).join("\n");
+      const dead = "constructor destructor __constructor__ __destructor__";


The dead copy only neutralizes constructor and destructor attributes. Objective-C in it still runs or registers at image load: +load methods run and categories attach.

Example: a macOS package with an #ifdef __OBJC__ path defines @implementation NSWindow (HID(Ext)) or a class with +load. The dead copy's +load runs at startup a second time. Its category methods share copy 1's selectors and may win, and they read HID2_ globals whose constructor never ran.

nicolas-abril · 2026-11-03T19:50:51Z

+      return at("push") + c_ids(src, m) + at("pop") + at("push")
+        + dead.replace(/\S+/g, "#define $& __unused__\n")
+        + `#pragma clang diagnostic error "-Wtentative-definition-compat"\n`
+        + c_ids(src, m, undefined, "HID2_") + at("pop");


The dead copy keeps every external-linkage function and the data it reaches, so a package's initialized tables end up twice in every consuming binary. For example, a non-static Term HID(run)(...) that reads a 2 MB static const u8 HID(font)[] gets a second, unused 2 MB HID2_ copy, and -O3 can't drop it because the HID2_ entry point has external linkage. The 136-byte cost in the PR description only holds for tiny packages.

nicolas-abril · 2026-11-03T19:50:51Z

+  // (no constructor or destructor), so a name defined without HID clashes
+  // in its own build; each copy keeps its macros.
+  const effs = [...FL.effs.values()];
+  const reqs = effs.map(([m, s]) => FL.hids.has(m) ? "" : c_ids(s, m)).join("")


Non-HID effect sources are still joined with "" (the HID path uses "\n"), so a .c file without a trailing newline merges into the next one. #endif followed by Term foo(...) becomes #endifTerm foo(...), an invalid directive. If the first file ends in a // comment, the next file's first line becomes part of the comment.

nicolas-abril · 2026-11-03T19:50:51Z

  }).join("\n\n");
-  const reqs = [...FL.effs.values()].map(([m, src]) => c_ids(src, m)).join("");
+  // An imported module whose C says HID pastes it twice, the second dead
+  // (no constructor or destructor), so a name defined without HID clashes


Design: this enforces an author-side lint by compiling every HID package's C twice in every downstream build. That costs compile time and binary size, needs clang 21, and still misses Base and root. It also brings the issues above: ObjC load side effects, no constructor priorities, and false tentative-definition errors. All to catch a mistake only the publisher can fix.

The same check, run once when a package is tested or published on the hub, or as a TS-side scan of top-level definitions, would enforce the rule with no per-build cost.

nicolas-abril · 2026-11-03T19:50:51Z

+    + [...FL.hids].map((m) => {
+      const src = effs.filter((e) => e[0] === m).map((e) => e[1]).join("\n");
+      const dead = "constructor destructor __constructor__ __destructor__";
+      const at = (d: string) => [...dead.split(" "),


at() reruns the #define/#undef regex over the whole module source and re-splits dead on each of its four calls per module. Computing const names = [...dead, ...macros] once per module and mapping it to push and pop avoids that. Making dead an array also removes the split and the replace.

Lorenzobattistela mentioned this pull request Nov 2, 2026

Effect-defined C names use explicit HID markers #1546

Closed

Lorenzobattistela and others added 3 commits November 2, 2026 17:43

Lorenzobattistela force-pushed the fix-1533-effect-names branch from 42ba453 to 1b42dc1 Compare November 2, 2026 20:56

vicmcorrea mentioned this pull request Nov 3, 2026

perf(c): build lists directly from flat arrays #2012

Draft

nicolas-abril reviewed Nov 3, 2026

View reviewed changes

Navigation Menu

Repository navigation

Effect C names its own definitions with HID, and a module that uses HID must (#1533) - #2017

Effect C names its own definitions with HID, and a module that uses HID must (#1533)#2017
Lorenzobattistela wants to merge 3 commits into
mainfrom
fix-1533-effect-names

Lorenzobattistela commented Nov 2, 2026 •

edited

Loading

Uh oh!

nicolas-abril left a comment

Uh oh!

nicolas-abril Nov 3, 2026

Uh oh!

nicolas-abril Nov 3, 2026

Uh oh!

nicolas-abril Nov 3, 2026

Uh oh!

nicolas-abril Nov 3, 2026

Uh oh!

nicolas-abril Nov 3, 2026

Uh oh!

nicolas-abril Nov 3, 2026

Uh oh!

nicolas-abril Nov 3, 2026

Uh oh!

nicolas-abril Nov 3, 2026

Uh oh!

nicolas-abril Nov 3, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

Lorenzobattistela commented Nov 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

Fix

Commits

Verified

Limits

Uh oh!

nicolas-abril left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Lorenzobattistela commented Nov 2, 2026 •

edited

Loading