-O3: load float constants from a data pool, not GPR materialization

• Type: feature (optimization — Track O; Track A files/gate). x86-64 (aarch64 mirror later).
• Opened: 2026-07-18, after [[feature-opt-float-intree-xmm-fusion]] + the residency expansion closed the transfer/reload gap.

The remaining gap

After xmm-fusion + residency, bench/portable/mandelbrot.pas is -O3 0.63s vs FPC 0.32s (was 1.31s at -O2; 4.2x → 2.0x). A chunk of what's left is float constant materialization: the value model carries a Double as its bits in rax, so a float literal in a hot loop (2.0*zre*zim, the 4.0 escape test) emits movabs rax, <bits> + movq xmm, rax every iteration. FPC keeps the constant in memory and folds it into the op: mulsd xmm, [rip+const].

In EmitFloatTree a float-const leaf (IR_CONST_INT with a float IRTk = double bits) currently goes through IREmitNode (movabs) then movq xmm,rax. Replace with a single movsd xmm<dst>, [pool] from a deduped read-only double pool.

Why it is not a trivial slice

The BSS-global float load already uses movsd xmm,[abs32] via EmitGlobRef (4-byte disp32 fixup). But a constant needs INITIALIZED data:

• EmitDataRef emits a pointer-width (8-byte) absolute address — wrong width for the movsd [04 25 disp32] operand (needs 4 bytes).
• So this needs one of: a new 4-byte data-address fixup, or emit the 8 bytes to BSS + a startup initializer (like globals), or RIP-relative addressing (F2 0F 10 /r with a mod=00 rip disp32 + a pc-relative fixup). RIP-relative is the cleanest and smallest code, but is new fixup machinery.

Scope

• A per-compile deduped pool of 8-byte doubles (map bits -> data offset).
• EmitFloatTree float-const leaf -> movsd xmm<dst>, [pool].
• Ideally also fold a const operand straight into the arith op (mulsd xmm,[pool]) — bigger, do the load form first.
• -O3 only (EmitFloatTree is -O3), so -O0/-O2 stay byte-identical for self-host. Numerically identical (same 8 bytes) — mandelbrot checksum MUST stay 74607393270.

Acceptance

• mandelbrot/nbody faster at -O3, checksum/energy byte-identical.
• No integer/-O2 regression; C float differential (O0/O2/O3) stays clean.
• Gate: make test + self-host byte-identical + the bench suite.

REJECTED 2026-07-18 — implemented, correct, but PERF-NEUTRAL

Implemented end to end (added a 4-byte DATA-abs fixup list DataFix32 mirroring GlobFix; EmitDataRef32; writeELF apply-loop; EmitFloatTree float-const leaf -> 8-byte pool + movsd xmm,[abs32]). Result: mandelbrot -O3 0.64s vs 0.63s (neutral/noise), nbody unchanged — checksum/energy byte-identical, self-host byte-identical. The movsd from an L1-hot pool is not cheaper than movabs rax + movq; the constant materialization was never the bottleneck. REVERTED.

Broader finding: THREE float micro-opts now measure perf-neutral on mandelbrot — compare-fusion, dead-store-elim, and const-pool (see [[project_float_intree_xmm_fusion]]). The residual gap vs FPC (0.63 vs 0.32) is therefore STRUCTURAL — instruction scheduling / the arithmetic dependency chain / FPC's cross-loop register allocation — not the remaining GPR<->XMM transfers or constant loads. Further float wins need a real scheduler or a genuine xmm-resident value model across the whole loop, not more per-leaf peepholes. Don't re-attempt const-pooling without a const-materialization-bound benchmark proving a win.