RFR: 8358781: C2 fails with assert "bad profile data type" when TypeProfileCasts is disabled
Manuel Hässig
mhaessig at openjdk.org
Thu Aug 7 11:41:15 UTC 2025
On Tue, 5 Aug 2025 10:40:19 GMT, Saranya Natarajan <snatarajan at openjdk.org> wrote:
> **Issue**
> An error, `assert(data->is_ReceiverTypeData()) failed: bad profile data type`, is encountered during C2 compilation due to bad profile data. This occurs when the code is compiled with `TypeProfileCasts` option disabled.
>
> **Analysis**
> The assertion failure occurs in `record_profiled_receiver_for_speculation` that analyzes the profiling information in the method data to determine whether a null value has been observed in the `instanceof` operation. This information is encoded in the `BitData` during profiling. When the method identifies that a null has been seen, it proceeds to inspect the associated `ReceiverTypeData` to see if the type check is always performed against null. However, in this scenario, the incoming profiling data is of type `BitData` rather than `ReceiverTypeData`, leading to the assertion failure.
>
> The profiling information for null seen for operations `aastore`, `instanceof`, and `checkcast` is recorded by the method `profile_null_seen `(in` src/hotspot/cpu/x86/templateTable_x86.cpp `). On investigating this method, it can be observed that the method data pointer is not updated for `VirtualCallData` (which is a subclass of `ReceiverTypeData`) when the `TypeProfileCasts` option is disabled.
>
> **Solution**
> My proposal is to inspect the `ReceiverTypeData` in function `record_profiled_receiver_for_speculation` only if `TypeProfileCasts` is enabled (this is based on the fact that the relevant method data pointer is not updated when `TypeProfileCasts` is disabled).
>
> **Question to reviewers**
> Do you think this is a reasonable fix ?
>
> **Testing**
> GitHub Actions
> tier1 to tier3 on windows-x64, linux-x64, linux-aarch64, macosx-x64, and macosx-aarch64.
Thank you for working on this, @sarannat! The fix seems reasonable to me since `GraphKit::maybe_cast_profiled_receiver` has a similar exception.
However, you are missing a regression test or a `noreg-*` label in JBS. However, in this case, I think a small regression test is warranted.
-------------
Changes requested by mhaessig (Committer).
PR Review: https://git.openjdk.org/jdk/pull/26640#pullrequestreview-3096696793
More information about the hotspot-compiler-dev
mailing list