diff options
| author | Chia-hung Duan <chiahungduan@google.com> | 2023-08-21 17:26:41 +0000 |
|---|---|---|
| committer | Chia-hung Duan <chiahungduan@google.com> | 2023-08-21 17:26:42 +0000 |
| commit | 5f771c9936b828d5fc017ad558e84702500df99c (patch) | |
| tree | d0e67a762ba47f0f9bb5e8172fd9b9b526044243 /llvm/lib/CodeGen/BranchRelaxation.cpp | |
| parent | 634b2fd2cac251e2d645dadc6b00a2b2b80a08d0 (diff) | |
| download | llvm-5f771c9936b828d5fc017ad558e84702500df99c.zip llvm-5f771c9936b828d5fc017ad558e84702500df99c.tar.gz llvm-5f771c9936b828d5fc017ad558e84702500df99c.tar.bz2 | |
[scudo] Support dumping fragmentation data in SizeClassAllocator64
This tells the number of pages that still have blocks in-used, i.e.,
those pages can't do releaseToOSMaybe(). Along with the information of
getStats() and RSS usage from the system (like smaps), we can tell if
the heuristic in releaseToOSMaybe() works well in certain scenarios.
Here's the sample output:
```
Fragmentation Stats: SizeClassAllocator64: page size = 4096
01 ( 32): inuse/total blocks: 275/ 416 inuse/total pages: 4/ 4 inuse bytes: 16K
02 ( 48): inuse/total blocks: 182/ 312 inuse/total pages: 4/ 4 inuse bytes: 16K
03 ( 64): inuse/total blocks: 169/ 312 inuse/total pages: 5/ 5 inuse bytes: 20K
...
32 ( 1040): inuse/total blocks: 90/ 152 inuse/total pages: 37/ 39 inuse bytes: 148K
33 ( 1168): inuse/total blocks: 136/ 232 inuse/total pages: 64/ 67 inuse bytes: 256K
34 ( 1296): inuse/total blocks: 138/ 226 inuse/total pages: 68/ 72 inuse bytes: 272K
35 ( 1424): inuse/total blocks: 140/ 232 inuse/total pages: 78/ 81 inuse bytes: 312K
36 ( 1552): inuse/total blocks: 157/ 248 inuse/total pages: 89/ 94 inuse bytes: 356K
...
```
SizeClassAllocator32 requires further refactoring to support this.
Reviewed By: cferris
Differential Revision: https://reviews.llvm.org/D158225
Diffstat (limited to 'llvm/lib/CodeGen/BranchRelaxation.cpp')
0 files changed, 0 insertions, 0 deletions