fix : local mem in simx

[talubik]

Fix local memory address aliasing in simx simulator

When I tried to run OpenCL kernels that use __local buffers larger than 2048 bytes, I found that writes to addresses beyond the 2048-byte boundary silently overwrote data at lower addresses. This Pull Request fixes a bug in sim/simx/local_mem.cpp where the address mask used to index the local memory RAM was too narrow.

Bug: to_local_addr() uses line count instead of capacity for address bits

The function to_local_addr() extracts the lower N bits of the incoming address to produce a RAM index. N was computed as log2ceil(capacity / line_size) (the number of cache lines) instead of log2ceil(capacity) (the number of addressable bytes).

With the default configuration (capacity = 16384, line_size = 8):

• Buggy: total_lines = 16384 / 8 = 2048, line_bits_ = log2ceil(2048) = 11, mask = 0x7FF
• Fixed: addr_bits_ = log2ceil(16384) = 14, mask = 0x3FFF

The 11-bit mask means any two byte addresses that differ by exactly 2048 map to the same RAM location. For example, A[0] (offset 0) and A[512] (offset 2048) alias each other, so writing A[512] silently overwrites A[0].

Fix

Compute addr_bits_ directly from config.capacity instead of from the number of lines:

// Before
uint32_t total_lines = config.capacity / config.line_size;
line_bits_ = log2ceil(total_lines);

// After
addr_bits_ = log2ceil(config.capacity);

I also added a test which reproduces the aliasing:
each thread t writes A[t] = t+1, then A[t+512] = -(t+1).
With the bug, A[t+512] at byte offset 2048 + t*4 aliases A[t], so reading A[t] returns -(t+1) instead of t+1.
With the fix, all reads return the correct values.