Memory operand costs in i486.ad and amd64.ad are currently folded into
instruction pattern costs. Because instruction patterns with a
'memory' operand cannot know what that operand looks like, these
costs are necessarily imprecise. Memory operand costs should be
moved from the instruction patterns into the operand patterns,
leaving instruction pattern costs to reflect actual instruction
issue costs. At the same time, immediate operand costs should be
analyzed and revised to reflect their actual use costs.

This work should be part of a thorough review and improvement of
x86 and amd64 code generation.