The `RzFloat` changes fix or improve:
- binary80 explicit-integer-bit, pseudo-value, infinity, and NaN handling;
- binary16 conversions;
- gradual underflow and directed rounding;
- overflow, underflow, invalid-operation, and inexact exception reporting;
- exception propagation through nested conversions and arithmetic operations;
- binary80 fused multiply-add rounding, including reduced-precision and double-rounding edge cases;
- thread-local SoftFloat state, preventing rounding state from leaking between threads.
The `RzIL` changes add scoped binary80 precision support through `RzFloatRPrecision` and `FWITH_RPREC`. The supported precisions are 32, 64, and 80. Precision scopes restore the
previous thread-local SoftFloat state after successful evaluation and evaluation failures.
Runtime rounding modes are represented explicitly by dedicated pure opcodes:
- `FCONVERT_WITH_RMODE`
- `FROUND_WITH_RMODE`
- `FSQRT_WITH_RMODE`
- `FADD_WITH_RMODE`
- `FSUB_WITH_RMODE`
- `FMUL_WITH_RMODE`
- `FDIV_WITH_RMODE`
- `FMOD_WITH_RMODE`
Their rounding-mode operand is a 32-bit IL bitvector whose values correspond to `RzFloatRMode`: RNE, RNA, RTP, RTN, and RTZ. Invalid operand widths are rejected by validation,
while invalid runtime values cause evaluation to fail with an error.
Dedicated opcodes keep runtime-controlled floating-point expressions compact. This is useful for architectures whose rounding mode is selected from register state and avoids the
expression duplication caused by expanding every operation into nested `ITE` branches.
The new operations are supported by:
- construction, duplication, and destruction;
- type and operand validation;
- VM evaluation;
- plain, Unicode, and JSON exporters;
- graph output and opcode stringification.
`FEXCEPT` now emits a VM event only when the queried exception is present, while preserving exceptions raised by nested conversions and arithmetic operations.
