[llvm-branch-commits] [clang] [LifetimeSafety] Track per-field origins for record types (PR #195603)

[Zhijie Wang via llvm-branch-commits]

aeft wrote:

## Performance Analysis

According to [llvm-compile-time-tracker](https://llvm-compile-time-tracker.com/?config=Overview&stat=instructions%3Au&remote=usx95), this PR introduced 0.85% regression on top of baseline-with-lifetime-safety (which itself adds 1.86% over lifetime-safety-disabled).

Many Sema files have significant regression. For example, `SemaExpr.cpp.o` regresses by +11.52%. I profiled clang while compiling `SemaExpr.cpp` on my laptop. Several metrics below.

The following table is produced by instrumenting `FactsGenerator::flow`, sampling the `dst` OriginNode tree shape at every top-level entry (Hope this can give a sense of how big these trees can grow):

- `depth`: longest path from the root to a leaf
- `max_fanout`: largest number of children at any single node
- `nodes`: total number of nodes in the tree

|  | NFC PR | This PR |
| --- | ---:| ---:|
| flows (top-level) | 160,449 | 155,544 |
| max_depth | 3 | 9 |
| max_fanout | 1 | 28 |
| max_nodes | 3 | 100 |

| depth | NFC PR | This PR |
| ---:| ---:| ---:|
| 1 | 147,745 | 138,033 |
| 2 | 12,702 | 12,870 |
| 3 | 2 | 2,399 |
| 4-6 | 0 | 1,539 |
| 7-9 | 0 | 703 |

| nodes | NFC PR | This PR |
| ---:| ---:| ---:|
| 1 | 147,745 | 138,033 |
| 2 | 12,702 | 12,711 |
| 3 | 2 | 796 |
| 4-15 | 0 | 2,315 |
| 31 | 0 | 1 |
| 46 | 0 | 959 |
| 47 | 0 | 12 |
| 48 | 0 | 637 |
| 49-54 | 0 | 76 |
| 99-100 | 0 | 4 |

Here, I report the performance of three core methods on `SemaExpr.cpp`:

| method | NFC PR | This PR |
| --- | ---:| ---:|
| flow (top-level) | 3\.01 ms | 4\.60 ms |
| getOriginNode | 96\.55 ms | 1712\.11 ms |
| hasOrigins | 0\.55 ms | 2\.00 ms |

>From the table, the regression is concentrated in `getOriginNode`; `flow` and `hasOrigins` stay cheap. The regression likely comes from building larger trees in `getOriginNode`. I'll share more analysis later and look into optimizing (or bugfix?) `getOriginNode`.

This also matches the performance change in `FactGenerator` (as profiled with `-ftime-trace`):

|  | NFC PR | This PR | delta |
| --- | ---:| ---:| ---:|
| LifetimeSafetyAnalysis | 0\.269 s | 2\.516 s | \+2.247 s |
| FactGenerator | 0\.123 s | 1\.766 s | \+1.644 s |
| LiveOrigins | 0\.055 s | 0\.386 s | \+0.331 s |



https://github.com/llvm/llvm-project/pull/195603