[mono][llvm] Disable simd intrinsics when we might be interoping between jit and llvmaot #74797
Conversation
|
Note that @fanyang-mono's previous change (enabled on amd64 where I tested) was enough to run the test suite where I was encountering simd related issues. On a debug build it was still failing however due to more explicit debug asserts in the bcl. I'm not sure if there is a justification to risk it, since behavior can keep changing with aot profiles AFAIK, and we can come up with a more detailed plan for .net 8, in order to facilitate support of all intrinsics on aot (probably still with some compromises). This change might still be overly conservative. What do you guys think @fanyang-mono @vargaz @lambdageek |
|
@BrzVlad @vargaz I started a discussion about some different options for net8 for the SIMD intrinsics over in #74587 (comment) |
…en jit and llvmaot Revert previous partial disabling of simd intrinsics. If we have a method that checkes if some hardware functionality `IsSupported` and then calls another method making use of this functionality if so, we have a problem if the method using the actual intrinsic is not also aot compiled and the intrisic is not supported by the jit. This issue is exposed inside bcl code, where this pattern is very common, by use of profiled aot on android. This change goes for the most conservative approach, by disabling all simd if we are not fullaot-ing and we use llvm.
BrzVlad
force-pushed
the
fix-disable-simd-llvmaot
branch
from
bfaec4f
to
3d743ee
Compare
|
/azp run runtime-extra-platforms |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
/backport to release/7.0 |
|
Started backporting to release/7.0: https://github.com/dotnet/runtime/actions/runs/3015062050 |
…ing betwen jit and llvmaot (dotnet#74797)" This reverts commit c3ccb8a.
ghost
locked as resolved and limited conversation to collaborators
Revert previous partial disabling of simd intrinsics.
If we have a method that checkes if some hardware functionality
IsSupportedand then calls another method making use of this functionality if so, we have a problem if the method using the actual intrinsic is not also aot compiled and the intrisic is not supported by the jit. This issue is exposed inside bcl code, where this pattern is very common, by use of profiled aot on android. This change goes for the most conservative approach, by disabling all simd if we are not fullaot-ing and we use llvm.