JIT: Extract all side effects of the index in optRemoveRangeCheck#92116
JIT: Extract all side effects of the index in optRemoveRangeCheck#92116jakobbotsch merged 1 commit intodotnet:mainfrom
Conversation
optRemoveRangeCheck extracts only GTF_ASG from the bounds check. If the BOUNDS_CHECK is complex, that results in silently dropping side effects on the floor (see the example case). The ideal fix is that we should always extract all side effects from the index and length operands, however this has large regressions because the length typically has an ARR_LENGTH that we then extract. This PR instead has a surgical fix for the problem case that can be backported to .NET 8. It extracts all side effects from the index, but keeps extracting only GTF_ASG from the length to get around the issue mentioned above. Fix dotnet#91862
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsoptRemoveRangeCheck extracts only GTF_ASG from the bounds check. If the BOUNDS_CHECK is complex, that results in silently dropping side effects on the floor (see the example case). The ideal fix is that we should always extract all side effects from the index and length operands, however this has large regressions because the length typically has an ARR_LENGTH that we then extract. This PR instead has a surgical fix for the problem case that can be backported to .NET 8. It extracts all side effects from the index, but keeps extracting only GTF_ASG from the length to get around the issue mentioned above. Fix #91862
|
#84248 should help to mitigate a lot of regressions in that case (tried locally a while ago) |
|
/backport to release/8.0 |
|
Started backporting to release/8.0: https://github.com/dotnet/runtime/actions/runs/6220544189 |
optRemoveRangeCheck extracts only GTF_ASG from the bounds check. If the BOUNDS_CHECK is complex, that results in silently dropping side effects on the floor (see the example case).
The ideal fix is that we should always extract all side effects from the index and length operands, however this has large regressions because the length typically has an ARR_LENGTH that we then extract. This PR instead has a surgical fix for the problem case that can be backported to .NET 8. It extracts all side effects from the index, but keeps extracting only GTF_ASG from the length to get around the issue mentioned above.
Fix #91862