Bring up DartDev isolate from kernel instead of an AppJIT snapshot

[bkonyi]

Using a snapshot to bring up the DartDev isolate can cause a failure if incompatible VM flags are provided. We should change to using a kernel file instead to avoid these incompatibilities.

[mit-mit]

This is a bit hard to understand; can you explain the issue in a bit more detail?

[bkonyi]

An example from #42958 where passing --sound-null-safety causes dartdev to fail to load:

out/ReleaseX64/dart --observe --enable-experiment=non-nullable
--sound-null-safety test_hello.dart 

Snapshot not compatible with the current VM
configuration: the snapshot requires 'release no-dwarf_stack_traces_mode
causal_async_stacks no-lazy_async_stacks lazy_dispatchers use_bare_instructions
no-dedup_instructions no-"asserts" "use_field_guards" "use_osr" no-code-comments
x64-sysv no-null-safety' but the VM has 'release no-dwarf_stack_traces_mode
causal_async_stacks no-lazy_async_stacks lazy_dispatchers use_bare_instructions
no-dedup_instructions no-"asserts" "use_field_guards" "use_osr" no-code-comments
x64-sysv null-safety' 

Observatory listening on
http://127.0.0.1:8181/3JgcHLS6A2Q=/

Basically, AppJIT snapshots require that certain VM flags used to generate the snapshot must also be set identically in the VM the snapshot is being loaded into; if they don't match, the snapshot won't be loaded. We have the same issue when we load the kernel isolate, but we fall back to using the kernel snapshot if the AppJIT snapshot fails to load.

[mit-mit]

So here specifically the VM is running with sound null safety and thus cannot load an AppJit snapshot that has unsound null safety? Why is that?

I understand that other "low level" VM flags can affect the snapshot format, but why does something like null safety impact it?

[bkonyi]

I'm not 100% confident (@a-siva, please correct me if I'm wrong) but I believe that what determines whether or not a flag is compatible with a snapshot is if the flag changes behavior or memory layout in an incompatible way. For the example above we're trying to load a snapshot which was not generated with strong null safety enabled into a VM which expects strong null safety which would break assumptions made by the VM.

[a-siva]

app JIT snapshots contains code.
The code/runtime stubs that are generated for sound null safety is different from unsound null safety, so it would be incorrect to load up an app JIT snapshot generated in unsound null safety mode into a VM that is expected to run in sound null safety mode or vice-versa.