Optimize memory bounds checks on narrow targets

[espmaniac]

Summary

• keep the existing 64-bit effective-address check on native 64-bit targets
• use an overflow-safe 32-bit memory32 check on narrower targets
• apply the optimized address type and bounds check to all load/store handlers

Why

WebAssembly memory32 base addresses and instruction offsets are unsigned 32-bit values. The previous implementation promoted every effective address to u64, which makes narrow targets such as AVR emit software 64-bit addition and comparison helpers for memory operations.

The new narrow-target path detects 32-bit addition wraparound before validating the access size. It is selected by native pointer width rather than by any platform-specific AVR condition.

Impact

ATmega1284P build with float, fuel, and snapshots explicitly enabled:

• flash: 109,972 -> 105,706 bytes (-4,266 bytes / -3.88%)
• static RAM: 5,188 -> 5,188 bytes
• 57 load/store handlers: 12,752 -> 8,460 bytes
• software 64-bit helper calls in memory handlers: 132 -> 0

The x86-64 build keeps the previous code path and its measured binary size is unchanged.

Validation

• built the full AVR configuration with fuel and snapshots enabled
• checked valid last-byte loads/stores, zero-length memory, end-crossing accesses, and effective-address wraparound
• compared the narrow check with the original 64-bit reference across 1,000,000 generated inputs
• ran native fuel tests in default, no-float, and no-bounds-check configurations
• ran an AddressSanitizer/UndefinedBehaviorSanitizer boundary check