[mono][mini] Interlocked.CompareExchange and Interlocked.Exchange intrinsics for small types and enums #106660
Conversation
ghost
added
the
lambdageek
changed the title
also add mono_atomic_cas_u8 utility function
lambdageek
requested review from
BrzVlad,
fanyang-mono,
kotlarmilos,
marek-safar and
steveisok
as code owners
lambdageek
changed the title
This comment was marked as outdated.
This comment was marked as outdated.
|
@kg do you want to do the jiterp work here or in a follow up? |
|
Jiterp implementation should be done already, I pushed it into the PR. |
for all i/u 4 variants? |
i/u 4 variants aren't in this PR, I think? They're on the todo list. |
| </Target> | ||
|
|
||
| <ItemGroup> | ||
| <ProjectReference Include="$(RepoTasksDir)AotCompilerTask\MonoAOTCompiler.csproj" ReferenceOutputAssembly="false" Condition="'$(RunAOTCompilation)' == 'true'" AdditionalProperties="Configuration=Debug"/> |
There was a problem hiding this comment.
For my education, why those changes were necessary? I thought we could already run the mono sample with FullAOT
There was a problem hiding this comment.
./build.sh mono+libs+tasks on OSX it doesn't bulid the AOT compiler task anymore because someone added a condition to tasks.proj to skip it on non-mobile platforms. I think the long-term direction for tasks.proj is to get rid of it and add proper project references to all the projects that need tasks
|
| OPDEF(MINT_MONO_EXCHANGE_U1, "mono_interlocked.xchg.u1", 4, 1, 2, MintOpNoArgs) | ||
| OPDEF(MINT_MONO_EXCHANGE_I1, "mono_interlocked.xchg.i1", 4, 1, 2, MintOpNoArgs) | ||
| OPDEF(MINT_MONO_EXCHANGE_U2, "mono_interlocked.xchg.u2", 4, 1, 2, MintOpNoArgs) | ||
| OPDEF(MINT_MONO_EXCHANGE_I2, "mono_interlocked.xchg.i2", 4, 1, 2, MintOpNoArgs) | ||
| OPDEF(MINT_MONO_EXCHANGE_I4, "mono_interlocked.xchg.i4", 4, 1, 2, MintOpNoArgs) | ||
| OPDEF(MINT_MONO_EXCHANGE_I8, "mono_interlocked.xchg.i8", 4, 1, 2, MintOpNoArgs) | ||
| OPDEF(MINT_MONO_CMPXCHG_U1, "mono_interlocked.cmpxchg.u1", 5, 1, 3, MintOpNoArgs) | ||
| OPDEF(MINT_MONO_CMPXCHG_I1, "mono_interlocked.cmpxchg.i1", 5, 1, 3, MintOpNoArgs) | ||
| OPDEF(MINT_MONO_CMPXCHG_U2, "mono_interlocked.cmpxchg.u2", 5, 1, 3, MintOpNoArgs) | ||
| OPDEF(MINT_MONO_CMPXCHG_I2, "mono_interlocked.cmpxchg.i2", 5, 1, 3, MintOpNoArgs) |
There was a problem hiding this comment.
@BrzVlad @kg instead of having a ton of different opcodes, we could have a single MINT_MONO_EXCHANGE_INT (and MINT_MONO_CMPXCHG_INT) opcode with a isSigned << 4 | size byte argument. That would make the big opcode switch in interp.c a little smaller, but make the code a little more complicated. Any opinions about which way is better?
|
/backport to release/9.0 |
|
Started backporting to release/9.0: https://github.com/dotnet/runtime/actions/runs/10531732073 |
|
Thank you, @lamdageek. |
6 participants
Only arm64, amd64 and WebAssembly. Interpreter, JIT and LLVM AOT. No x86 or arm32.
The complication for the 8- and 16-bit operands is that the .NET operand stack is 32-bit. so we have to zero- or sign-extend the inputs and output.
I ended up following the same approach as clang: zero-extend "expected" value and then sign-extend the return value of the CAS. One nice consequence is that the LLVM support is essentially free.
Also this PR makes the intrinsic apply to types that have byte/int/long as the underlying type (bool, enums) not just when the type is literally in the signature
TODO:
Interlocked.Exchange(int)intrinsicsRelated to #105335
Related to #93488