← board

--shared reports an internal error where i386 gets a clean refusal

Measured under pin v403 (c31d03b202da) and again at HEAD ce19e5482 (9bcfd2b4da30) — identical both times, so this is not a recent regression:

$ pascal26 --target=i386    --shared t.pas t.so
error: --shared: shared-library output is x86-64 only
$ pascal26 --target=aarch64 --shared t.pas t.so
error: internal: no init/fini thunk prologue for --target=aarch64
$ pascal26 --target=arm32   --shared t.pas t.so
error: internal: no init/fini thunk prologue for --target=arm32
$ pascal26 --target=riscv32 --shared t.pas t.so
error: internal: no init/fini thunk prologue for --target=riscv32
$ pascal26 --target=xtensa  --shared t.pas t.so
error: internal: no init/fini thunk prologue for --target=xtensa

x86-64 works, so the restriction is real and intended.

Why the wording is the bug and not a cosmetic

The i386 arm is the correct behaviour: a deliberate, documented limitation reported as one. The other four reach a later stage that has no thunk for the target and fail there, so the user is told internal: — the string the compiler reserves for "the compiler is broken". That is a false statement about whose fault it is, and it routes the reader wrong: the reasonable response to internal: is to file a compiler bug, and the correct response here is "this feature is x86-64 only", which the docs already say.

It is the same shape as [[bug-a-the-emit-obj-refusal-names-a-target-set-that-excludes-x86-64]] — the diagnostic being the instrument a reader trusts over the docs, and being the thing that is wrong. That one is fixed; --emit-obj now answers --emit-obj: no object writer for --target=<t> on aarch64 and arm32, which is exactly the shape this one wants.

The fix

The refusal that i386 gets should be reached for every non-x86-64 target, at the same early point, rather than four targets falling through to the thunk stage. One target test where the existing one is, not four new arms.

Aperture

Found by executing the sweep method in [[bug-d-docs-scope-claims-about-a-flag-are-invisible-to-a-flag-existence-sweep]] across docs/**. The documentation is correct here — docs/reference/cli.md, docs/reference/limits.md and docs/reference/objects.md all say --shared is x86-64 only, and all three are right. This is purely the compiler's message for a limitation the docs already state accurately.