I want some constriants to be register constraints when compiling with a particular set of compiler flags, and extra constraints when compiling with another. (In the first case, alternatives that use this constraint can be selected by reload for reloading non-matching insns, whereas in the second case, they will only be used if they already match. The register constraint then yields NO_REG to allow the other definition to have effect). That was straightforward with the old constraint target macros, but when you define this in constraints.md, you get an error.
Are there any objections to implement constraint overloading in genpreds.c ? Is it OK to just accept multiple definitions for disjoint categories, or would you like to see msome extra syntactic sugar to affirm that the overloading is intentional?
