Track O: in-tree xmm fusion for float expression temporaries

• Type: feature (optimization — Track O; Track A files/gate). Per-backend value-model work, NOT a shared-IR pass — x86-64 first, aarch64 later, per the O charter (32-bit/ESP keep their models).
• Status: done
• Opened: 2026-07-18, split out of [[feature-opt-float-register-temporaries]] as its actionable first slice.

Why this, split from the big ticket

[[feature-opt-float-register-temporaries]] is the FULL fix: an xmm-resident float value model touching every float consumer (binops, compares, call args/returns, stores, writeln) — a multi-session arc. This ticket is the narrow subset that captures "probably most of the win" with a one-expression-tree blast radius and no cross-statement state. Do this first; the full value-model arc stays the someday version.

The gap (bench oracle)

bench/portable/mandelbrot.pas: pxx -O2 1.33s vs FPC -O2 0.32s = 4.2x, with an IDENTICAL checksum (74607393270 @1600x1200) — same work, worse codegen. Root: the x86-64 value model carries a Double as raw bits in RAX, so each float IR_BINOP emits 3 GPR↔XMM transfers + a stack round-trip:

eval left -> rax, push rax; eval right -> rax, mov rcx <- pop;
movq xmm1, rax; movq xmm0, rcx; <op>; movq rax, xmm0

The mandelbrot inner loop (zr*zr - zi*zi + cr, …) round-trips every temporary through memory + GPR.

Scope of THIS ticket

• Within a single float expression tree, evaluate to xmm2..xmm5 and keep intermediates resident — no push/pop, no movq rax<->xmm transit between sub-operations of the same tree. Spill only when the tree's register budget is exceeded (as MOVSD, not GPR-transit).
• Binop-tree only. Cross-statement float locals staying xmm-resident is the big ticket (regcall-style residency with the same guards) — explicitly OUT here.
• -O0 unchanged — the GPR-transit accumulator stays the simple debuggable baseline; this is -O2+/-O3 behavior (Track O promotion: land behind -O3, promote to -O2 after the full gate).

The -O3 entanglement

-O3 currently makes float WORSE (2.47s) — the W1 operand scheduler ([[project_o3_w1_operand_scheduler]]) pessimizes float chains. So landing this behind -O3 must NOT compose with W1's float pessimization: either make W1 xmm-aware or gate W1 off for float trees as part of this work. May be a quick standalone win on its own.

Acceptance

• mandelbrot-p (and nbody) measurably faster at the target -O level, with the checksum BYTE-IDENTICAL (strict IEEE, no FMA contraction — the bench header documents why; a differing checksum is a correctness regression, not a speedup).
• No regression on the integer benches or the -O2 default.
• Gate: make test + self-host byte-identical (-O0 emission unchanged) + the bench suite.

Non-goals

• Not the full xmm-resident value model (that's [[feature-opt-float-register-temporaries]]).
• Not aarch64 in v1 (x86-64 first; aarch64 mirrors after).
• Not FMA / reassociation — checksums must stay bit-identical.

Log

• 2026-07-18 — resolved, commit c14f35a1.