On 09/01/14 23:17, Bin.Cheng wrote:
For this specific case, I think the reuse of r84 comes from coalescing
during expanding, and this is necessary to remove redundant reg-moves.
Then we need to fix this in coming passes?
Hmmm. Yea, I can see how that might be happening. There's a certain
inherent tension between coalescing aggressively to remove reg moves and
leaving the reg moves to avoid false dependencies like this.
I wonder if there's a less aggressive approach where we somehow say "I
don't care about coalescing when the result is set, used & dies all
within a basic block. But that's probably out of scope for this problem.
jeff
