[clang-doc] Merge data into persistent memory
We have a need for persistent memory for the final info. Since each
group processes a single USR at a time, every USR is only ever processed by
a single thread from the thread pool. This means that we can keep per
thread persistent storage for all the info. There is significant
duplicated data between all the serialized records, so we can just merge
the final/unique items into the persistent arena, and clear out the
scratch/transient arena as we process each record in the bitcode.
The patch adds some APIs to help with managing the data, merging, and
allocation of data in the correct arena. It also safely merges and deep
copies data from the transient arenas into persistent storage that is
never reset until the program completes.
This patch reduces memory by another % over the previous patches,
bringing the total savings over the baseline to 57%. Runtime performance
and benchmarks stay mostly flat with modest improvements.
| Metric | Baseline | Prev | This | Culm% | Seq% |
| :--- | :--- | :--- | :--- | :--- | :--- |
| Time | 920.5s | 991.5s | 987.2s | +7.2% | -0.4% |
| Memory | 86.0G | 40.0G | 36.9G | -57.1% | -8.0% |
| Benchmark | Baseline | Prev | This | Culm% | Seq% |
| :--- | :--- | :--- | :--- | :--- | :--- |
| BM_BitcodeReader_Scale/10 | 67.9us | 72.2us | 72.2us | +6.3% | -0.0% |
| BM_BitcodeReader_Scale/10000 | 70.5ms | 22.5ms | 17.3ms | -75.5% | -23.2% |
| BM_BitcodeReader_Scale/4096 | 23.2ms | 6.6ms | 7.1ms | -69.5% | +7.4% |
| BM_BitcodeReader_Scale/512 | 509.4us | 898.7us | 550.5us | +8.1% | -38.7% |
| BM_BitcodeReader_Scale/64 | 114.8us | 133.7us | 120.8us | +5.2% | -9.6% |
| BM_EmitInfoFunction | 1.6us | 1.9us | 1.8us | +12.9% | -3.3% |
| BM_Index_Insertion/10 | 2.3us | 4.1us | 3.5us | +52.4% | -14.7% |
| BM_Index_Insertion/10000 | 3.1ms | 5.3ms | 4.8ms | +53.5% | -10.0% |
| BM_Index_Insertion/4096 | 1.3ms | 2.1ms | 1.9ms | +50.8% | -9.2% |
| BM_Index_Insertion/512 | 153.6us | 251.8us | 227.0us | +47.8% | -9.9% |
| BM_Index_Insertion/64 | 18.1us | 30.2us | 26.7us | +47.9% | -11.7% |
| BM_JSONGenerator_Scale/10000 | 89.6ms | 81.4ms | 83.4ms | -6.9% | +2.5% |
| BM_JSONGenerator_Scale/4096 | 33.7ms | 31.0ms | 32.4ms | -3.9% | +4.5% |
| BM_Mapper_Scale/10000 | 104.3ms | 112.3ms | 103.5ms | -0.8% | -7.9% |
| BM_Mapper_Scale/4096 | 44.3ms | 45.0ms | 43.8ms | -1.2% | -2.5% |
| BM_Mapper_Scale/512 | 7.6ms | 7.7ms | 7.5ms | -1.3% | -2.4% |
| BM_Mapper_Scale/64 | 3.1ms | 3.0ms | 3.0ms | -1.9% | -0.3% |
| BM_MergeInfos_Scale/10000 | 12.2ms | 575.6us | 500.1us | -95.9% | -13.1% |
| BM_MergeInfos_Scale/2 | 1.9us | 1.8us | 1.8us | -4.4% | -1.7% |
| BM_MergeInfos_Scale/4096 | 2.8ms | 205.3us | 200.4us | -92.8% | -2.4% |
| BM_MergeInfos_Scale/512 | 68.9us | 20.5us | 19.5us | -71.7% | -5.1% |
| BM_MergeInfos_Scale/64 | 10.3us | 3.8us | 4.0us | -60.9% | +4.8% |
| BM_MergeInfos_Scale/8 | 2.8us | 1.9us | 1.9us | -31.7% | -1.8% |
| BM_SerializeFunctionInfo | 25.5us | 25.8us | 26.1us | +2.2% | +1.3% |
