Add L1 block cache and direct-mapped BHT

[jserv]

This introduces two lookup optimizations:

1. L1 Direct-Mapped Block Cache (256 entries, interpreter only):

• Separated tag/pointer arrays: tag array (1KB) checked first, pointer loaded only on hit for cache efficiency
• Invalid tag sentinel (0xFFFFFFFF) for clean miss detection
• EXT_C-aware index shift (1 vs 2 bits) to reduce conflict misses
• New block_lookup_or_find() as primary lookup with hash fallback

2. Direct-Mapped Branch History Table (both JIT and non-JIT):

• O(1) lookup replacing O(n) linear search: index = (PC >> 2) & mask
• Remove idx field (non-JIT) and bht_find_min_idx() - no longer needed
• Update bht_find_max_idx() to scan all entries since zeros can appear at any index with direct-mapped scheme
• Add static assert for power-of-2 HISTORY_SIZE

Benchmarks on Intel Xeon E5-2650:

• CoreMark: 804.66 -> ~887 iterations/sec (+10.2%)
• Dhrystone: 1366 -> ~1382 DMIPS (+1.2%)

---

Summary by cubic

Adds an L1 direct-mapped block cache (interpreter only) and a direct-mapped branch history table to speed up hot-path lookups. On Xeon E5-2650 this yields ~+10% CoreMark and ~+1% Dhrystone.

• New Features
  • L1 block cache: 256 entries, split tag/pointer arrays, invalid tag sentinel, EXT_C-aware index shift; new block_lookup_or_find() with hash fallback; proper init/clear.
  • Direct-mapped BHT (JIT and non-JIT): O(1) index = (PC >> 2) & mask; remove idx and bht_find_min_idx(); update bht_find_max_idx() to scan all entries; require power-of-2 HISTORY_SIZE; update templates to use direct lookup and replacement, with SATP checks for JIT.

^{Written for commit f89a114. Summary will update on new commits.}

[cubic-dev-ai]

1 issue found across 5 files

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them.


<file name="src/riscv_private.h">

<violation number="1" location="src/riscv_private.h:251">
P3: The block_l1 cache comment lists incorrect sizes (128B/256B). With BLOCK_L1_SIZE=256, the tag array is 1KB and the pointer array is 2KB, so the comment should match the actual layout to avoid confusion during tuning or future changes.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[jserv]

Benchmarks

Details

Benchmark suite	Current: f89a114	Previous: a37c812	Ratio
Dhrystone	1656.333 DMIPS	1584 DMIPS	0.96
CoreMark	953.17 iterations/sec	939.552 iterations/sec	0.99

This comment was automatically generated by workflow using github-action-benchmark.