t27b: @bitCast between an integer and a float of one size, @intFromBool, and std.math.nan / isNan / isPositiveInf / isNegativeInf #7791

Part of #7746 (t27b passes specs/numeric/gf16.t27), which is part of #6063 (t27b coverage), item C10 of #6488. This is the first PR of #7746's split.

What gf16.t27 needs

t27b refuses these in gf16.t27 today (t27b test --blockers, master 2781ae3):

• ExprCall(@bitCast), 22 sites: const bits : u32 = @bitCast(value); reads an f32's bits, and const out : f32 = @bitCast(out_bits); builds one. All 22 are f32 <-> u32.
• ExprCall(std.*), 7 sites: std.math.nan(f32) (3), std.math.isNan (2), std.math.isPositiveInf, std.math.isNegativeInf. (std.math.pow is a later PR of t27b: specs/numeric/gf16.t27 after the type mismatch family -- @bitCast, f32 rounding and libm, std.math, @setEvalBranchQuota #7746.)
• ExprCall(@intFromBool), 3 sites: @as(u8, @intFromBool(is_fin)).

What the reference does

Measured on the t27c lab (zig 0.16.0, the x86_64 Debug backend t27c test-report uses):

• @bitCast keeps the bits exactly. A signalling NaN 0x7F800001 survives a bitcast, a call and a return. A u32 0xFFFFFFFF is the i32 -1.
• Zig refuses a literal operand ("cannot @bitcast from 'comptime_int'"), two sizes ("@bitcast size mismatch"), and a call with no result type. It accepts a bool to a u1.
• std.math.nan(f32) is 0x7FC00000, and std.math.nan(f64) is 0x7FF8000000000000. -std.math.nan(f32) is 0xFFC00000.
• isNan(x) is x != x. isPositiveInf(x) is x == inf, and isNegativeInf(x) is x == -inf (std/math/isnan.zig, isinf.zig).
• @intFromBool(b) is a u1, 0 or 1.

Plan

• @bitCast and @intFromBool. A new plan, specs/tri/t27b/bit_cast_plan.t27, through t27c gen-rust:
  • an int <-> float of one size goes through a frame slot: a store of the operand, then a load of the result type, with no new IR node;
  • two integers of one size are a Cast with no trap site;
  • a bool to a u1, and @intFromBool, is a Widen;
  • a constant folds;
  • the rest is refused by name, as Zig refuses it.
• The std.math calls. They join specs/tri/t27b/libm_plan.t27. Their routines are t27 in specs/tri/t27b/libm.t27, as @exp and @log are. nan names its type, so its routine takes no argument.
• A conformance spec in gf16's own shapes: specs/tri/t27b/conformance/bit_cast.t27.

Acceptance criteria

• t27b test --blockers specs/numeric/gf16.t27 names no ExprCall(@bitCast), no ExprCall(@intFromBool) and no call to std.math.nan, isNan, isPositiveInf or isNegativeInf.
• The conformance spec passes under t27c test-report with 0 vacuous tests, and the same tests pass under t27b.
• The lab corpus receipt compare gives IMPROVED_ONLY, EQUIVALENT or NEUTRAL.

User Scenarios

• Given a gf16.t27 spec using @bitCast to convert between f32 and u32, when t27b compiles it, then the expression is accepted and the bit pattern is preserved exactly.
• Given a gf16.t27 spec calling std.math.nan(f32), std.math.isNan(x), std.math.isPositiveInf(x), or std.math.isNegativeInf(x), when t27b compiles it, then the calls are resolved to t27 routines that match Zig's semantics.
• Given a gf16.t27 spec using @intFromBool(b) where b is a bool, when t27b compiles it, then the result is a u1 value 0 or 1 matching Zig's behavior.

Requirements

FR-001: The planner in specs/tri/t27b/bit_cast_plan.t27 MUST handle @bitCast between an integer and a float of identical size by emitting a store of the operand and a load of the result type through a frame slot with no new IR node.
FR-002: The planner in specs/tri/t27b/bit_cast_plan.t27 MUST handle @bitCast between two integers of identical size as a Cast with no trap site.
FR-003: The planner in specs/tri/t27b/bit_cast_plan.t27 MUST handle @bitCast from bool to u1 and @intFromBool as a Widen.
FR-004: The planner in specs/tri/t27b/bit_cast_plan.t27 MUST fold constant @bitCast and @intFromBool expressions at compile time.
FR-005: The planner in specs/tri/t27b/bit_cast_plan.t27 MUST refuse all other @bitCast cases by name, matching Zig's refusal semantics.
FR-006: The specs/tri/t27b/libm_plan.t27 planner MUST resolve std.math.nan, std.math.isNan, std.math.isPositiveInf, and std.math.isNegativeInf calls to t27 routines in specs/tri/t27b/libm.t27.
FR-007: The std.math.nan routine in specs/tri/t27b/libm.t27 MUST take no argument and return the canonical quiet NaN bit pattern for the target float type (0x7FC00000 for f32, 0x7FF8000000000000 for f64).
FR-008: The std.math.isNan routine in specs/tri/t27b/libm.t27 MUST implement the semantics x != x.
FR-009: The std.math.isPositiveInf routine in specs/tri/t27b/libm.t27 MUST implement the semantics x == inf.
FR-010: The std.math.isNegativeInf routine in specs/tri/t27b/libm.t27 MUST implement the semantics x == -inf.
FR-011: The conformance spec specs/tri/t27b/conformance/bit_cast.t27 MUST exercise all accepted @bitCast forms, @intFromBool, and the four std.math functions with values matching Zig's reference behavior.

Success Criteria

• t27b test --blockers specs/numeric/gf16.t27 exits 0 and names no ExprCall(@bitCast), no ExprCall(@intFromBool), and no call to std.math.nan, isNan, isPositiveInf, or isNegativeInf.
• t27c test-report specs/tri/t27b/conformance/bit_cast.t27 exits 0 with 0 vacuous tests.
• t27b test specs/tri/t27b/conformance/bit_cast.t27 exits 0 with the same test count as t27c.
• Lab corpus receipt compare yields IMPROVED_ONLY, EQUIVALENT, or NEUTRAL.

Boundary

• specs/numeric/gf16.t27
• specs/tri/t27b/bit_cast_plan.t27
• specs/tri/t27b/libm_plan.t27
• specs/tri/t27b/libm.t27
• specs/tri/t27b/conformance/bit_cast.t27