Fix codegen for getfield of homogeneous tuples

[Keno]

This fastpath wasn't checking whether or not the element type
was inline allocated or not.

Fixes #34206
Fixes #34207

[vtjnash]

Much thanks for this! I think this is perhaps just a design issue with typed_load? It's rather awkward that all callers now avoid calling it unless they're fairly confident that they won't accidentally hit the isboxed case. Seems like that should have been an argument to the function instead?

[JeffBezanson]

Yes, it would be nice if we could preserve this fast path for the case where the element type is known not to be inlinealloc.

[cafaxo]

Seems like something went wrong here?

[JeffBezanson]

Should julia_type_to_llvm be checking inlinealloc instead of jl_is_concrete_immutable?

[JeffBezanson]

How about something like this:

--- a/src/cgutils.cpp
+++ b/src/cgutils.cpp
@@ -1321,6 +1321,10 @@ static jl_cgval_t typed_load(jl_codectx_t &ctx, Value *ptr, Value *idx_0
     Type *elty = julia_type_to_llvm(jltype, &isboxed);
     if (type_is_ghost(elty))
         return ghostValue(jltype);
+    if (!isboxed && !jl_datatype_isinlinealloc(jltype)) {
+        elty = T_prjlvalue;
+        isboxed = true;
+    }
     Type *ptrty = PointerType::get(elty, ptr->getType()->getPointerAddressSpace());