← board

i386 refuses a frozen record field write under the byte-prefix mode

Blocks the phase-4 flip on i386. The flip turns -dPXX_SHORTSTRING on globally; today i386 cannot compile a record with a string[N] field being assigned.

type R = record f: string[10]; g: string[4]; end;
var r: R;
begin r.f := 'field'; end.



pascal26:2: error: target i386: write of this operand not yet supported

Compiles and runs correctly in the default mode on i386, and correctly in BOTH modes on x86-64, arm32, aarch64 and riscv32. Also hit by an array of records with a string field.

This is an honest refusal, not a miscompile — the backend says the operand is unimplemented and stops, which is the right failure. It is a missing codegen arm, so it should be cheap next to the wrong-value bugs in this family.

Found by running a 20-probe construct suite in both modes on all five runnable targets; x86-64 alone shows nothing.

The summary was wrong about which statement fails (2026-09-03)

r.f := 'field' is NOT the failing statement and never was. Measured at 482b714d0 on i386 under the flag, the three-line body in the repro above compiles, links and runs with exit 0 — the error comes from the WriteLn(r.f) that the finding suite had beside it. "write" in the diagnostic is IR_WRITE, i.e. Write/WriteLn, not an lvalue store; the original summary read it as the verb and sent the next reader to the store path.

Cause and fix

compiler/ir_codegen386.inc, the IR_WRITE operand dispatch: the frozen arm was spelled else if tk = tyString then. i386 was the last = tyString in an IR_WRITE arm — riscv32, arm32, aarch64, xtensa and wasm32 all say TypeIsFrozenString(tk) and x86-64 does not reach this shape. A frozen operand is tagged tyString only where the walker had a symbol to normalise through StrValTk; a RECORD FIELD load keeps its storage kind, so WriteLn(r.f) arrived as tyShortString, matched no arm, and fell to the catch-all Error. WriteLn(s) for a bare var s: string[10] was correct in the same program, which is why a variable-only suite saw nothing.

Widened to TypeIsFrozenString(tk), matching the six other backends.

Verified

Five targets (x86_64 native; i386/arm32/aarch64/riscv32 under qemu), BOTH modes, value-asserted against a written .expected — not exit codes, not Length(): record field, second field, and an array[0..1] of R element all print their contents, are compared to a literal in both directions, and are indexed. All ten cells PASS. The pre-fix binary refused to compile the i386 cells at all.

The regression test fails on the pre-fix binary

test/test_frozen_arg_and_field_write.pas + .expected, wired for the native sweep and the i386 / arm32 / aarch64 / riscv32 cross batteries, both modes each (10 rows). Run against a rebuilt pre-fix compiler (7f95d3b1c5c2, the same sha this session's first build produced, so it is the tree at 482b714d0 and not a walked seed):

row pre-fix post-fix
i386 -dPXX_SHORTSTRING compile error PASS
i386 default PASS PASS
arm32 / aarch64 / riscv32, both modes PASS PASS

The i386 DEFAULT row passes on the pre-fix binary and is kept deliberately: it is a regression guard for the path that was never broken, not a detector, and the test says so rather than letting a reader count it as coverage.

Log