RFR: 8357060: [premain] assert(left >= right) failed: avoid underflow

Vladimir Kozlov kvn at openjdk.org
Fri May 16 18:05:00 UTC 2025


On Fri, 16 May 2025 15:15:39 GMT, Ashutosh Mehra <asmehra at openjdk.org> wrote:

> This PR fixes a few things in the premain:
> 1. When storing metadata in aot code cache, `AOTCacheAccess::delta_from_shared_address_base` was incorrectly using `SharedBaseAddress` for computing the offset. It should be using `MetaspaceShared::requested_base_address` because we convert the input address to the requested address.
> 2. Fixing the above issue results in crash in C1 and C2 compiled code during production run because CompressedKlassPointer::base() value was hardcoded in the generated code. In mainline we emit relocation for `CompressedKlassPointer::base()`. This patch adds the same changes to premain. In addition to that, it also modifies `MacroAssembler::decode_and_move_klass_not_null` and `MacroAssembler::encode_and_move_klass_not_null` which are used by C2 compiled code.
> 3. Fixing 2 reveals another problem when preload the code. `AOTCodeEntry::_method` can be invalid if the AOT Cache gets mapped to different address than the "requested" address, and can result in crash when accessing `AOTCodeEntry::_method` during preload. Fix is to store the offset of the `AOTCodeEntry::_method` and use the offset on load to get the correct Method pointer.
> 4. While working on this issue, I realized archived `AOTCodeCache::compile_nmethod` is using archived nmethod to print the assembly. This results in crash as archived nmethod has some state cleaned up. Updated `AOTCodeCache::compile_nmethod` to fix this issue.

Can you add new methods to  `AOTCacheAccess` to get addresses based on offsets instead of: `(Method*)((address)SharedBaseAddress + entry->method_offset());` ?

-------------

Changes requested by kvn (Committer).

PR Review: https://git.openjdk.org/leyden/pull/68#pullrequestreview-2847313056


More information about the leyden-dev mailing list