We should adopt a convention of using eax as the final result register. This is pretty close to what we actually do in practice: only one case of op_add and some resolve opcodes returning two results differ.
Created attachment 25164 [details] Proposed patch This patch is no performance change when I test with networking on (but Safari and Mail closed). I will test again with networking off before landing.
Comment on attachment 25164 [details] Proposed patch r=me
Landed in r38419.