yxsamliu wrote:

@YonahGoldberg 

I updated the heuristic to try to keep the useful performance wins from 
vectorizing alloca partitions, while avoiding the cases that looked like they 
were causing regressions, and I triggered a new llvm-opt-benchmark run for that 
latest change here: 

https://github.com/dtcxzyw/llvm-opt-benchmark-nightly/pull/152

The IR report there still shows a few apparent regressions, so I followed up 
locally by checking the corresponding assembly for those changed files, and I 
could not find a clear asm-level performance regression signal. Most of the 
changed cases looked neutral to better in asm, and the one remaining small 
positive asm delta in `cpython/compile` still looked more like control-flow / 
register-allocation noise than a real slowdown. So my current reading is that 
the updated heuristic is keeping the vectorization benefit we want, and that 
the apparent regressions in the IR diff do not look like real performance 
regressions from the assembly side.


https://github.com/llvm/llvm-project/pull/165159
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to