Implement coloring::iterate_solution().

It can happen that the currently live registers are blocking a smaller
register class completely, so the only way of solving the allocation
problem is to turn some of the live-through registers into solver
variables.

When the quick_solve attempt fails, try to free up registers in the
critical register class by turning live-through values into solver
variables.
This commit is contained in:
Jakob Stoklund Olesen
2017-09-29 14:38:30 -07:00
parent 86e22e7de5
commit 51a6901a7f
5 changed files with 181 additions and 6 deletions

View File

@@ -590,6 +590,18 @@ impl Solver {
self.find_solution()
}
/// Try harder to find a solution.
///
/// Call this method after `quick_solve()` fails.
///
/// This may return an error with a register class that has run out of registers. If registers
/// can be freed up in the starving class, this method can be called again after adding
/// variables for the freed registers.
pub fn real_solve(&mut self) -> Result<AllocatableSet, RegClass> {
// TODO: Sort variables to assign smallest register classes first.
self.find_solution()
}
/// Search for a solution with the current list of variables.
///
/// If a solution was found, returns `Ok(regs)` with the set of available registers on the
@@ -623,6 +635,11 @@ impl Solver {
pub fn vars(&self) -> &[Variable] {
&self.vars
}
/// Check if `value` can be added as a variable to help find a solution.
pub fn can_add_var(&mut self, _value: Value, constraint: RegClass, from: RegUnit) -> bool {
!self.regs_in.is_avail(constraint, from)
}
}
/// Interface for working with parallel copies once a solution has been found.