Description
Description
A browser-wasm AOT publish (RunAOTCompilation=true, default WasmDedup=true) spends one to two hours in the single-threaded aot-instances.dll pass of mono-aot-cross when a generic method recurses into itself with its struct type arguments wrapped one level deeper:
static void MapInstance<TSource, TTarget>(in TSource source, in TTarget target) // TSource, TTarget : struct
{
foreach (var f in ...)
MapInstance(new StructureSource<TSource>(source, f), new StructureTarget<TTarget>(target, f));
}At run time the depth is bounded by real data (2 in the repro). AOT cannot know that, so the chain is instantiated up to the generic depth limit, and every level needs gsharedvt wrappers. The same app with the recursion made non-generic publishes in about a minute. The output is equivalent, and both builds run correctly.
Reproduction
Standalone repro, no third-party code: https://github.com/LostBeard/mono-wasm-aot-gsharedvt-recursion
- ON:
dotnet publish src/Repro.csproj -c Release - OFF:
dotnet publish src/Repro.csproj -c Release -p:NoPolymorphicRecursion=true(the only difference is the#if NORECblocks insrc/Program.cs)
GitHub Actions measures both arms on .NET 9, 10 and 11 (ubuntu-24.04), and the results page also hosts the built apps: https://lostbeard.github.io/mono-wasm-aot-gsharedvt-recursion/
| Framework | MonoAOTCompiler task, ON |
OFF | Publish wall time, ON | OFF | aot-instances compiled, ON / OFF |
|---|---|---|---|---|---|
| net9.0 | 6,497 s | 5.2 s | 109 min | 82 s | 7,462 / 7,252 |
| net10.0 (10.0.15) | 6,400 s | 5.2 s | 108 min | 54 s | 9,131 / 8,921 |
| net11.0 (11.0.0) | 4,797 s | 6.2 s | 81 min | 64 s | 9,043 / 8,833 |
On Windows with 10.0.14 locally: 8,050 s ON vs 5.15 s OFF. The recursion adds about 210 instances and more than an hour of AOT.
Where the time goes
I profiled the aot-instances pass of a real application with this shape (ILGPU's ArgumentMapper). I used a stack sampler and the public PDB for mono-aot-cross.exe 10.0.14, with 100 ms samples over an 84-minute pass. In every 10-minute window, 94-100% of samples were under:
compile_methods
compile_method
add_gsharedvt_wrappers
mini_get_gsharedvt_in_sig_wrapper
mini_get_underlying_signature
get_wrapper_shared_type_full <-> get_wrapper_shared_vtype (12-17 levels deep per sample, max 22)
mono_class_inflate_generic_class_checked (~68% inclusive)
mono_metadata_inflate_generic_inst / type_dup / malloc / free
get_wrapper_shared_vtype maps every field type through get_wrapper_shared_type_full, which maps every generic argument back through get_wrapper_shared_vtype. Nothing is cached, so a nested struct is re-expanded and re-inflated at every level, for every signature that contains it. The gsharedvt wrapper signatures in that app's aot-instances.dll nest Mono.ValueTuple 8 to 31 levels deep.
Mono's own counters for that pass were 19.2 ms per method in the emit phase (54,728 instances in 1,049 s of JIT). The 105 ordinary per-assembly passes of the same app ran at 0.54 ms per method. The collect phase, which prints no timing, took most of the remaining ~73 minutes.
Relation to #135029 / #135438 / #135334, and a request
#135029 / #135438 (merged to main) add a parent-frame chain with cycle detection and a depth limit of 16 to get_wrapper_shared_vtype, to fix a stack overflow. #135334 backported it to release/11.0 on 2026-11-16, after 11.0.0, so it should ship in 11.0.1. There is no backport to release/10.0.
I could not measure whether that depth limit also bounds this cost:
- 10.0.x and 11.0.0 do not contain it.
- The daily .NET 11 SDKs still bundle runtime 11.0.0-rc.2.26508.107.
- .NET 12 dailies run browser-wasm on CoreCLR, so they do not run Mono AOT.
Requests:
- Could someone who can build
mono-aot-crossfrom release/11.0 or main run the repro's ON arm? That would show whether [release/11.0] [mono][aot] Fix recursive struct wrapper normalization #135334 covers this. - If it does, please consider backporting it to release/10.0. .NET 10 is LTS, and this makes browser AOT publishes take hours. We hit it in production deploys: our ML demo's GitHub Pages deploy took 2 h 40 min, and its
aot-instancespass alone took 89 min. - If it does not, memoizing
get_wrapper_shared_type_full/get_wrapper_shared_vtyperesults per type would remove the repeated expansion.
Workaround for affected users
Make the polymorphic recursion non-generic. For example, keep the root type as a generic argument and carry the nested field path as a plain array, which is the OFF arm of the repro. ILGPU did this in SpawnDev.ILGPU 5.4.38. On the same GitHub runner, that app's whole Mono AOT step, including aot-instances, now takes 1.4 minutes; the aot-instances pass alone used to take 89. The repro README lists how to tell whether a build is affected.
Environment
- Windows 11, .NET SDK 10.0.403, runtime packs 10.0.14 (local).
- GitHub Actions ubuntu-24.04: .NET 9.0 (SDK 9.0.321), 10.0 (cross compiler 10.0.15), 11.0 (11.0.0).
wasm-toolsworkload,RunAOTCompilation=true,WasmDedupdefault (true), trimming default.
Activity
README: link the upstream issue dotnet/runtime#137919 [skip ci]