Plain (non-const) by-value record param >8B rejects a temporary argument

• Type: bug (compiler — call-arg lowering)
• Track: A — compiler/**
• Status: backlog (filed by Track B)
• Owner: — (Track A)
• Opened: 2026-06-25
• Found-by: [[feature-demo-raytracer]] — vector algebra over a 24-byte Vec3 = record x, y, z: Double end, composing VAdd(VScale(V(…), s), …).
• Relation: the sibling [[bug-const-byref-record-param-temp]] (done, 2026-06-19) fixed exactly this for const record params. This is the remaining, symmetric gap for plain by-value record params.

Symptom

A function-result temporary passed to a plain by-value record param fails when the record is larger than 8 bytes:

type Vec3 = record x, y, z: Double end;          { 24 bytes }
function V(x,y,z: Double): Vec3;
function VScale(a: Vec3; s: Double): Vec3;        { plain by-value }
function VAdd(a, b: Vec3): Vec3;

z := VAdd(VScale(V(1,1,1), 0.5), V(2,2,2));
        { pascal26: error: by-reference argument must be a variable () }

The exact same code compiles and runs correctly if the params are declared const:

function VScale(const a: Vec3; s: Double): Vec3;  { OK }
function VAdd(const a, b: Vec3): Vec3;            { OK -> z.x = 2.500 }

Named-variable arguments also work for the plain by-value form; only a temporary (function result) is rejected.

Size dependence

So it is the same root cause as the truncation work in [[bug-record-byvalue-arg-truncation]] (done): records >8 bytes are passed by-reference internally (with a callee copy for by-value semantics), and the AST-path call-arg check only binds a true lvalue — so a temporary is refused.

Why the const fix didn't cover this

bug-const-byref-record-param-temp's fix relaxed the call-arg check only for params flagged const (ProcParamIsConst[...], parser.inc ~3328 / ~5862), deliberately keeping var/out strict. A plain by-value param is neither const nor var, so it falls through to the strict path even though, for a

8-byte record, it is lowered by-ref exactly like a const param (the needTemp materialization in IRLowerCallArg already handles it).

Fix

Extend the temporary-materialization allowance to plain by-value record params (records passed by-ref for ABI reasons), not just const ones. The callee already copies for by-value semantics, so a hidden-local temp is sound. var/out must still require a real lvalue.

Done when

• VAdd(VScale(V(1,1,1), 0.5), V(2,2,2)) with plain by-value params compiles and gives z.x = 2.5.
• Nested/mixed temp+named args correct; var param still errors on a temp.
• examples/raytracer no longer needs const solely to satisfy the compiler.
• Regression test under make test; self-host fixedpoint byte-identical.

Workaround (in use)

examples/raytracer declares its vector-input params const — which is also the idiomatic, more efficient style for non-mutated record inputs, so the demo stays platonic. This ticket tracks making the bare by-value form work too.

Resolution (2026-06-25, v63)

The stated Fix — extend the temporary-materialization allowance to plain by-value record params — is implemented:

• defs.inc: new ProcParamExplicitByRef[] records whether a param was declared by-ref at the source level (var/out/const), captured before the >8-byte record ABI forces Params[].IsRef := True.
• symtab.inc/parser.inc: persist it alongside ProcParamIsConst.
• parser.inc (both call-arg checks): a tyRecord argument temporary is accepted when the param is const or plain by-value (not ProcParamExplicitByRef); a genuine var/out record still requires a true lvalue (verified: var/out
  • temp still error). IRLowerCallArg's needTemp path already copies the temp into a hidden local.

Verified end-to-end with integer records (Add(Mk(1,2,3), Mk(10,20,30)) → 11 22 33; nested/mixed temp+named). Regression test/test_byval_record_temp.pas under make test. Self-host byte-identical; pinned v63.

Caveat — float records still wrong: while testing the Vec3 (Double) chain I found the values come back zero regardless of const/by-value/named — because a function returning a record with float fields (≥3 float fields) loses them. That is a separate, pre-existing codegen ABI bug, filed as [[bug-float-field-record-function-return]]. So this ticket's done-when #1 (the z.x = 2.5 Vec3 value) and the raytracer rendering depend on THAT fix; the parse-allowance this ticket asked for is done and correct for working (integer/≤2-float-field) records.