gpl: use cached pin count in reportInstanceExtensionByPinDensity

[oharboe]

Summary

Optimizes gpl::NesterovBaseCommon::reportInstanceExtensionByPinDensity() to use the cached stats.pin_count per cell master rather than querying db_inst->getITerms().size() redundantly for every instance across the design.

Rationale

reportInstanceExtensionByPinDensity() collects per-master statistics and calculates area-per-pin metrics across all instances in the placement graph. Previously, db_inst->getITerms().size() was invoked on every instance. Since db_inst->getITerms().size() traverses an unindexed linked list of _dbITerm pointers in OpenDB, calling it on large designs (>2M instances) spent 99.5% of reporting execution time inside linked-list traversals.

Using the stats.pin_count already cached per cell master on line 1981 avoids 2.4 million linked-list traversals while producing identical statistics output.

[gemini-code-assist]

Code Review

This pull request optimizes the reportInstanceExtensionByPinDensity function in nesterovBase.cpp by utilizing the cached stats.pin_count instead of recalculating the pin count for each instance. The reviewer identified an existing logical issue where master statistics are overwritten in each iteration rather than accumulated, resulting in final reports that only reflect the last processed instance. The reviewer provided a suggestion to accumulate these areas and calculate the averages after the loop.

[oharboe]

/gemini review

[gemini-code-assist]

Code Review

This pull request refactors the calculation of per-master derived metrics in reportInstanceExtensionByPinDensity to run post-loop, accumulating areas using +=. Feedback highlights a critical issue where master_stats_map is declared as static, causing accumulated statistics to persist and grow incorrectly across multiple calls. Additionally, there is a risk of signed integer overflow when multiplying pin_count and instance_count on large designs, which can be mitigated by casting to double.

[oharboe]

/gemini review

[gemini-code-assist]

Code Review

This pull request refactors nesterovBase.cpp by removing the static qualifier from master_stats_map and moving the calculation of per-master derived metrics post-loop. Feedback is provided to optimize performance by accumulating raw DBU areas inside the loop and converting them to microns post-loop, which avoids repetitive floating-point divisions.

[gemini-code-assist]

Performance Optimization

Currently, block->dbuAreaToMicrons() is called inside the loop for every single instance (potentially millions of times). Since dbuAreaToMicrons performs a floating-point division, doing this inside the loop adds unnecessary overhead.

We can optimize this by accumulating the raw DBU areas as integers/doubles inside the loop, and then converting the final accumulated totals to microns once per master in the post-loop. This avoids millions of division operations.

    stats.total_original_area += orig_area;
    stats.total_extended_area += ext_area;
  }

  // Calculate per-master derived metrics post-loop
  for (auto& entry : master_stats_map) {
    MasterStats& stats = entry.second;
    stats.total_original_area = block->dbuAreaToMicrons(static_cast<int64_t>(stats.total_original_area));
    stats.total_extended_area = block->dbuAreaToMicrons(static_cast<int64_t>(stats.total_extended_area));
    if (stats.pin_count > 0 && stats.instance_count > 0) {
      stats.original_area_per_pin
          = stats.total_original_area
            / (static_cast<double>(stats.pin_count) * stats.instance_count);
      stats.extended_area_per_pin
          = stats.total_extended_area
            / (static_cast<double>(stats.pin_count) * stats.instance_count);
    }
    if (stats.total_original_area != 0.0) {
      stats.area_diff = static_cast<float>(100.0 * (stats.total_extended_area - stats.total_original_area) / stats.total_original_area);
    } else {
      stats.area_diff = 0.0f;
    }