Fixes #398: Apple Clang Optimization Breaking Bytecode Mapping

[M-SE0K]

Problem

| #398

When built on macOS (Apple Clang), the opcode labels of the computed goto interpreter are merged into a single address by tail merging. ByteCode::opcode() performs a reverse address → opcode lookup through m_addressToOpcodeTable, so once merging occurs, different opcodes end up sharing the same address, which corrupts this table and makes it return the wrong opcode.

Fix

• Insert an __asm__ volatile barrier after each opcode label so that the labels are not merged even when their bodies are identical. (Since an extra label is introduced, -Wno-unused-label is added as well.)
• Remove the existing workaround that intentionally reversed the field order in MoveFloat and others for the same purpose, restoring src/dst to their original order.
• Assert m_addressToOpcodeTable.size() == OpcodeKindEnd at initialization time so that any recurrence of the merging is detected immediately.

Benchmark Results

aarch64 macOS, test/wasmBenchmarker, average of 10 runs for each of 31 benchmarks (unit: seconds)

• tools/run-tests.py

	Before	After	Ratio
Interpreter	4.536	4.406	0.971
JIT	0.363	0.364	1.005

[M-SE0K]

Closed in favor of PR #472. Reopened in a new PR to fix a few test failures I initially missed.