On Oct 22, 2014, at 4:53 AM, Vladimir Makarov <vmaka...@redhat.com> wrote:
> On 10/20/2014 11:16 PM, Maxim Kuvyrkov wrote: >> Hi, >> >> This patch cleans up (removes) cached_first_cycle_multipass_dfa_lookahead >> and cached_issue_rate. >> >> These seem to be an artifact from the scheduler refactoring 10+ years ago. >> They assume that dfa_lookahead and issue_rate can change mid-way through >> scheduling, which is never the case. All backends currently treat >> dfa_lookahead and issue_rate as constants for the duration of scheduling >> passes. >> >> Bootstrapped on x86_64-linux-gnu. Regression testing is in progress. OK to >> commit if no regressions? >> > Yes. the patch for issue rate itself is ok but you should have modified > doc/tm.texi too for dfa lookahead hook saying that it should be a > constant (issue rate already has such clause). > > On the other hand I'd not assume that dfa look ahead is a constant. In > future we could make it non-constant to differentiate non-hot and hot > functions to speed up the scheduler as dfa look ahead scheduling is > pretty expensive. > > So issue rate change is ok but I'd not rush to change dfa look ahead > related code. It is trivial to prove that currently cached_first_cycle_multipass_dfa_lookahead always has the same value as dfa_lookahead. And, even should targetm.sched.first_cycle_multipass_dfa_lookahead start returning different values, max_issue will happily continue to use the value that the hook returned in sched_init(). Also, your suggestion to use different dfa_lookahead values for hot/cold functions is not affected by this patch. The values of dfa_lookahead variable have scope of the scheduling pass, which is invoked separately for every function. I guess, you could, potentially, start differentiating values of dfa_lookahead based on whether a basic_block is hot or cold, but handling this scenario would require significant changes throughout the scheduler. To summarize, dfa_lookahead is currently a pass-time invariant, that is free to change between invocations of the scheduler pass. Cached_first_cycle_multipass_dfa_lookahead always holds the same value as dfa_lookahead, and, as such, is extraneous. Thank you, -- Maxim Kuvyrkov www.linaro.org
