On 9/29/21 8:36 AM, Hao Xu wrote: > 在 2021/9/28 下午7:08, Pavel Begunkov 写道: >> On 9/27/21 7:17 AM, Hao Xu wrote: >>> Remove a local variable. >> >> It's there to help alias analysis, which usually can't do anything >> with pointer heavy logic. Compare ASMs below, before and after >> respectively: >> testq %rax, %rax # next >> >> replaced with >> cmpq $0, (%rdi) #, node_2(D)->next >> >> One extra memory dereference and a bigger binary >> wq_list_add_after: >> # fs/io-wq.h:48: node->next = pos->next; >> movq (%rsi), %rax # pos_3(D)->next, _5 >> # fs/io-wq.h:48: node->next = pos->next; >> movq %rax, (%rdi) # _5, node_2(D)->next >> # fs/io-wq.h:49: pos->next = node; >> movq %rdi, (%rsi) # node, pos_3(D)->next >> # fs/io-wq.h:50: if (!node->next) >> cmpq $0, (%rdi) #, node_2(D)->next > hmm, this is definitely not good, not sure why this is not optimised to > cmpq $0, %rax (haven't touched assembly for a long time..) Nothing strange, alias analysis, it can't infer that the pointers don't point to overlapping memory, and so can do nothing but reload. __restrict__ C keyword would've helped, but it's not used in the kernel. -- Pavel Begunkov
