Paradoxical subregs and !INSN_SCHEDULING. Why?

[Hans-Peter Nilsson]

Why are paradoxical subregs allowed and even generated in
combine (rather, recognized as valid combination attempts by
e.g. register_operand through general_operand) when not
INSN_SCHEDULING?  Is it because of a hope that combine will do
better than with the actual registers and that the pesky
scheduler (boo-hoo) doesn't handle the invalid (IMO)
intermediate-paradoxical-code?

To me it just seems an odd case causing bugs like PR
target/18701.

brgds, H-P