Undo change that dropped const-propped information

This PR included a change that only used constant results if the
result was *better*. However, this is not what we want, because
even though the result may have been the same, the IR we computed
will likely still have more precise (and fewer) statements, because
we optimized a bunch of things away based on constant information.

Author: Keno

base/compiler/abstractinterpretation.jl

@@ -67,7 +67,7 @@ function abstract_call_gf_by_type(interp::AbstractInterpreter, @nospecialize(f),
                     result, f, this_arginfo, si, match, sv)
                 const_result = volatile_inf_result
                 if const_call_result !== nothing
-                    if !(rt ⊑ₚ const_call_result.rt)
+                    if const_call_result.rt ⊑ₚ rt
                         rt = const_call_result.rt
                         exct = const_call_result.exct
                         (; effects, const_result, edge) = const_call_result
@@ -76,7 +76,7 @@ function abstract_call_gf_by_type(interp::AbstractInterpreter, @nospecialize(f),
                         (; effects, const_result, edge) = const_call_result
                     elseif !(exct ⊑ₚ const_call_result.exct)
                         exct = const_call_result.exct
-                        (; effects, const_result, edge) = const_call_result
+                        (; const_result, edge) = const_call_result
                     else
                         add_remark!(interp, sv, "[constprop] Discarded because the result was wider than inference")
                     end
@@ -113,9 +113,14 @@ function abstract_call_gf_by_type(interp::AbstractInterpreter, @nospecialize(f),
             if const_call_result !== nothing
                 this_const_conditional = ignorelimited(const_call_result.rt)
                 this_const_rt = widenwrappedconditional(const_call_result.rt)
-                # return type of const-prop' inference can be wider than that of non const-prop' inference
-                # e.g. in cases when there are cycles but cached result is still accurate
-                if !(this_rt ⊑ₚ this_const_rt)
+                if this_const_rt ⊑ₚ this_rt
+                    # As long as the const-prop result we have is not *worse* than
+                    # what we found out on types, we'd like to use it. Even if the
+                    # end result is exactly equivalent, it is likely that the IR
+                    # we produced while constproping is better than that with
+                    # generic types.
+                    # Return type of const-prop' inference can be wider than that of non const-prop' inference
+                    # e.g. in cases when there are cycles but cached result is still accurate
                     this_conditional = this_const_conditional
                     this_rt = this_const_rt
                     this_exct = const_call_result.exct
@@ -125,7 +130,7 @@ function abstract_call_gf_by_type(interp::AbstractInterpreter, @nospecialize(f),
                     (; effects, const_result, edge) = const_call_result
                 elseif !(this_exct ⊑ₚ const_call_result.exct)
                     this_exct = const_call_result.exct
-                    (; effects, const_result, edge) = const_call_result
+                    (; const_result, edge) = const_call_result
                 else
                     add_remark!(interp, sv, "[constprop] Discarded because the result was wider than inference")
                 end