Chapter 13: double (Floating-Point) Support

[johnhringiv]

Chapter 13: double (Floating-Point) Support

Adds double (64-bit IEEE-754) end-to-end, removes the deprecated text-based emitter, adds linker-library / object-file inputs, and fixes a pre-existing tackifier panic.

Language features

• New double type, floating-point constants (decimal/exponent lexing), arithmetic (+ - * /), unary negation, comparisons, and implicit/explicit conversions to and from every integer type.
• Out-of-range floating constants round to ±infinity / zero (C §6.4.4.2 UB given a defined result), flagged by -Woverflow.

Pipeline

• Lexer: ConstantDouble token + maximal-munch float regex; double keyword.
• Parser: Type::Double, Const::ConstDouble(f64), is_integer() helper, float-literal -Woverflow; extracted local_label_prefix().
• Validator: StaticInt → StaticInit with a DoubleInit(f64) variant; double constant folding using the same rules as runtime (cvttsd2si-matching double→int bounds); double rejected in switch/case; warning guards so integer-only diagnostics (div-by-zero, shift-count, constant-conversion) skip doubles; ConstEvalError::InvalidType.
• Tackifier: IntToDouble/UIntToDouble/DoubleToInt/DoubleToUInt conversion instructions (+ inline LLVM-correspondence docs).
• Codegen: XMM registers folded into Reg; AssemblyType::Double; a .rodata constant pool (ConstantPool, deduped by f64::to_bits, per-use 8/16-byte alignment); SSE lowerings — addsd/subsd/mulsd/divsd, xorpd negation (-0.0 mask), comisd, and cvtsi2sd/cvttsd2si with the SSE2 workarounds for the unsigned conversions; System V ABI updated to pass double args in XMM0–7 and return in XMM0 (shared classify_operands for callee/caller); pseudo-replacement and fix_invalid legalize the new SSE forms (XMM14/XMM15 scratch).
• Emitter (iced): encodes movsd/addsd/…/comisd/cvt*, emits the .rodata pool, adds setp/setnp + jp/jnp; -S prints the constant pool as bit-exact .quad 0x… # <decimal>. Fixed the disassembly symbol resolver so immediates (e.g. $0) are no longer mis-symbolized to a code symbol at address 0.

IEEE-754 / NaN semantics

Comparisons use comisd (unsigned-style flags) and handle unordered (NaN) per IEEE: </<= swap operands to the carry-clear A/AE (false on NaN), ==/!= combine the parity flag (NaN == x false, NaN != x true), and NaN is truthy in conditions. CondCode gained P/NP with per-variant docs.

Deterministic / "Safer C" edges

double→integer truncates toward zero; out-of-range or NaN yields the x86 cvttsd2si "integer indefinite" value (target MIN). Constant evaluation matches runtime for all of the above.

Bug fix (pre-existing since ch12)

A non-main function with an unsigned return type that falls off the end panicked the tackifier — the synthetic-return only handled int/long, so unsigned/double hit unreachable!. Now a real synthetic ret is emitted for every scalar return type. Regression test: tests/c_programs/unsigned/implicit_return_falloff.c (returns 2; fall-off values are deliberately unobserved per C §6.9.1p12).

Removed: text-based emitter

Deleted src/emit.rs and the --no-iced flag (the iced path is the sole, fully-featured backend; -S already provides readable disassembly). Removed the now-orphaned CondCode::ins_suffix and StaticInit::data_directive.

New CLI

• -l <lib> (e.g. -lm) — forwarded to the linker (libwild / ld).
• Pre-built .o inputs are accepted and passed straight to the linker (caller-owned, not cleaned up).

Test harness (tests/runner.rs)

• requires_mathlib → -lm (Linux; macOS uses libSystem).
• helper_libs/ excluded from standalone discovery; the libs mapping is honored — helpers are gcc-compiled and linked via NCC (exercising NCC's linker with external objects).
• CHAPTER_COMPLETED/EXTRA_COMPLETED 12 → 13.

Docs

README (Data Types table, grammar, IEEE/NaN note, -l/.o) and CLAUDE.md (type system, Floating Point note) updated; --no-iced/emit.rs references removed.

🤖 Generated with Claude Code

[claude]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.

[codecov]

Codecov Report

❌ Patch coverage is 89.73545% with 97 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/validate.rs	75.88%	41 Missing ⚠️
src/emit_iced.rs	80.00%	29 Missing ⚠️
src/main.rs	64.10%	14 Missing ⚠️
src/parser.rs	88.46%	6 Missing ⚠️
src/codegen.rs	99.15%	4 Missing ⚠️
src/tacky.rs	94.73%	3 Missing ⚠️

📢 Thoughts on this report? Let us know!

[claude]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.