Why does fall through from a cold block to a warm block need to be treated specially?

Unless I'm mistaken:
warm_set = reachable(follow_exception_edges=False)
reachable_set = reachable(follow_exception_edges=True)
cold_set = reachable_set - warm_set
Am I missing something?

Why does fall through from a cold block to a warm block need to be treated specially?

Unless I'm mistaken: warm_set = reachable(follow_exception_edges=False) reachable_set = reachable(follow_exception_edges=True) cold_set = reachable_set - warm_set Am I missing something?

We can't change b->next of a block with a fallthrough.

We can't change b->next of a block with a fallthrough.

You could convert it to a jump and eliminate the fallthrough edge.

We can't change b->next of a block with a fallthrough.

You could convert it to a jump and eliminate the fallthrough edge.

That's the idea. But there could already be a conditional jump there, so if we don't want to have to worry about multiple jumps in the same block for the remaining parts of the assembler, it needs to be a new block.

Would it better to add a warm parameter to mark_reachable?
mark_warm, mark_cold and mark_reachable are all doing the same thing, except for which edges they follow.

My first version did mark_cold/warm at the same time as mark_reachable. But I had a situation where the graph changed between that time and the time when I can do the push_cold_to_end. So now mark cold/warm happens just before push_cold_to_end. I could possibly share the code though.