Extended: formalize as Double alias (or reject)

• Type: chore (type system honesty)
• Status: backlog
• Owner: Track A
• Opened: 2026-06-20
• Priority: low

Problem

Extended is half-real. tyExtended exists (defs.inc:453, size 10, x87 load/store on x86-64) but all arithmetic runs in SSE2 double — there is no true 80-bit math. On cross targets it's treated as double. So Extended today is effectively a Double that occasionally costs 10 bytes of storage and an x87 round trip, with no extra precision.

Decision needed

Pick one:

• (a) Alias to Double everywhere (drop the 10-byte x87 storage; Extended = Double). Honest, simplest, matches what the arithmetic already does. Cost: a hair of FPC source-compat where code declares Extended expecting 80-bit (we don't target that precision anyway).
• (b) Reject Extended on cross targets (and document x86-64 as double-precision-despite-the-name).
• (c) Real 80-bit — explicitly NOT wanted (no demand, big x87/soft cost).

Recommendation: (a). Low priority; no user demand. Filed so the half-state is tracked, not silently shipped.

Notes

• Related: [[feature-single-first-class]] (the other end of the float-width spectrum — that one we DO want first-class).

Log

• 2026-06-22 — DONE, option (a) (Track A, commit d1a1ea9). The Extended type name (parser.inc tkExtended_T + the CaseEqual(lo,'extended') alias) now maps to tyDouble at parse, so the front end never produces tyExtended; Extended is plain Double on every target. The tyExtended enum + its x87 load/store paths stay but are unreachable (FloatBinopResultTk's tyExtended branch only fired on an already-Extended operand). compiler.pas uses no Extended, so the self-image is unchanged: make test + fpc-check byte-identical, cross-bootstrap (i386/aarch64/arm32) byte-identical self-fixedpoint. Test test/test_extended_is_double.pas.