Diffractor integration interferes with invoke inlining

[Keno]

With this patch:

diff --git a/src/ssair/inlining.jl b/src/ssair/inlining.jl
index c7e052e..0afdfd5 100644
--- a/src/ssair/inlining.jl
+++ b/src/ssair/inlining.jl
@@ -1643,6 +1643,10 @@ function assemble_inline_todo!(ir::IRCode, state::InliningState)
             continue
         end

+        if sig.f === Core.invoke
+            @show info
+        end
+
         # handle special cased builtins
         if isa(info, OpaqueClosureCallInfo)
             handle_opaque_closure_call!(todo, ir, idx, stmt, info, flag, sig, state)

and the testcast from #19, calls to evalpoly fail to inline

info = FRuleCallInfo(Compiler.InvokeCallInfo, Compiler.ConstCallInfo)
info = FRuleCallInfo(Compiler.InvokeCallInfo, Compiler.ConstCallInfo)

We should either the order of the infos (although InvokeCallInfo does not support a nested general info), or make inlining robust to this case.

[serenity4]

I faced inlining issues with a custom CallInfo structure in JuliaDebug/Cthulhu.jl#640, which in the end seemed to be solved by supporting the CallInfo interface.

It might be that the current issue is a corner case I haven't yet observed, though.

[Keno]

Yes, invoke is rare. You can try looking at log with your Cthulhu changes.

[serenity4]

Indeed, it's the same situation. I'd say that it would be convenient and somewhat idiomatic to make inlining robust to CallInfo wrappers. Perhaps with a more explicit API to unwrap them correctly.

With this FRuleCallInfo would you expect to treat it as if it were just the InvokeCallInfo, treating it as some CallInfo wrapper that inlining should ignore?

[Keno]

I think inlining could key on sig.f === invoke rather than the info and then handle_invoke could use the getresult/getsplit interface (potentially asserting nsplit == 1). InvokeCallInfo would have to synthesize the data structure, but I think it should work. Haven't tried though. Be my guest to have a go at it, otherwise I'll try it tomorrow.

[serenity4]

I decided to focus more on invalidation for today. I'd be happy to pick this one up tomorrow in addition to the Phi tainting logic if you'd like to work on other things.

[serenity4]

Fixed by JuliaLang/julia#58328.