[NewOptimizer] Fix _apply elision (#26821)

The old optimizer had an extra loop outside the inlining pass that would
turn _apply's in to regular calls. When I wrote the new inliner we discussed
that this wasn't actually necessary because we could just keep track of
this information in the inliner (as we do for invoke). However, that of
course also means that if we can't turn something into an :invoke, we
should still at least turn it into a regular call. Do that.

Author: Keno

base/compiler/ssair/inlining2.jl

@@ -309,14 +309,22 @@ function maybe_make_invoke!(ir::IRCode, idx::Int, @nospecialize(etype), atypes::
                     @nospecialize(atype_unlimited), isinvoke::Bool, isapply::Bool, @nospecialize(invoke_data))
     nu = countunionsplit(atypes)
     nu > 1 && return # TODO: The old optimizer did union splitting here. Is this the right place?
-    ex = Expr(:invoke)
     linfo = spec_lambda(atype_unlimited, sv, invoke_data)
-    linfo === nothing && return
-    add_backedge!(linfo, sv)
+    if linfo === nothing
+        if !isapply || isinvoke
+            return
+        end
+        # We might not have an linfo, but we can still rewrite the _apply into a regular call
+        # based on our analysis
+        ex = Expr(:call)
+    else
+        ex = Expr(:invoke)
+        add_backedge!(linfo, sv)
+    end
     argexprs = ir[SSAValue(idx)].args
     argexprs = rewrite_exprargs((node, typ)->insert_node!(ir, idx, typ, node), arg->exprtype(arg, ir, ir.mod),
         isinvoke, isapply, argexprs)
-    pushfirst!(argexprs, linfo)
+    linfo !== nothing && pushfirst!(argexprs, linfo)
     ex.typ = etype
     ex.args = argexprs
     ir[SSAValue(idx)] = ex
@@ -466,7 +474,6 @@ function assemble_inline_todo!(ir::IRCode, linetable::Vector{LineInfoNode}, sv::
             methsp = methsp::SimpleVector
         end
 
-
         methsig = method.sig
         if !(atype <: metharg)
             maybe_make_invoke!(ir, idx, stmt.typ, atypes, sv, atype_unlimited, isinvoke, isapply, invoke_data)