On 3 February 2017 at 14:58, Maxim Kuvyrkov <maxim.kuvyr...@linaro.org> wrote: >> On Jan 30, 2017, at 5:50 PM, Maxim Kuvyrkov <maxim.kuvyr...@linaro.org> >> wrote: >> >>> On Jan 30, 2017, at 3:23 PM, Kyrill Tkachov <kyrylo.tkac...@foss.arm.com> >>> wrote: >>> >>> Hi Maxim, >>> >>> On 30/01/17 12:06, Maxim Kuvyrkov wrote: >>>> This patch enables prefetching at -O3 for aarch64 cores that set >>>> "simultaneous prefetches" parameter above 0. There are currently no such >>>> settings, so this patch doesn't change default code generation. >>>> >>>> I'm now working on improvements to -fprefetch-loop-arrays pass to make it >>>> suitable for -O2. I'll post this work in the next month. >>>> >>>> Bootstrapped and regtested on x86_64-linux-gnu and aarch64-linux-gnu. >>>> >>> >>> Are you aiming to get this in for GCC 8? >>> I have one small comment on this patch: >>> >>> + /* Enable sw prefetching at -O3 for CPUS that have prefetch, and we >>> + have deemed it beneficial (signified by setting >>> + prefetch.num_slots to 1 or more). */ >>> + if (flag_prefetch_loop_arrays < 0 >>> + && HAVE_prefetch >>> >>> HAVE_prefetch will always be true on aarch64. >>> I imagine midend code that had logic like this would need this check, but >>> aarch64-specific code shouldn't need it. >> >> Agree, I'll remove HAVE_prefetch. >> >> This pattern was copied from other backends, and HAVE_prefetch is most >> likely a historical artifact. > > Andrew raised a good point in the review of his patch that it is a bad idea > to use one of prefetching parameters (simultaneous_prefetches) as indicator > for whether to enable prefetching pass by default. Indeed there are cases > when we want to set simultaneous_prefetch according to HW documentation (or > experimental results), but not enable prefetching pass by default. > > This update to the patch addresses it. The patch adds a new explicit field > to prefetch tuning structure "default_opt_level" that sets optimization level > from which prefetching should be enabled by default. The current value is to > enable prefetching at -O3; additionally, this parameter will come handy for > enabling prefetching at -O2 [when it is ready]. > > -- > Maxim Kuvyrkov > www.linaro.org > >
AArch64 maintainers, ping? Here is a rebased patch against current trunk. OK to commit, assuming bootstrap and reg-test passes? -- Maxim Kuvyrkov www.linaro.org
0004-Enable-fprefetch-loop-arrays-at-O3-for-cores-that-be.patch
Description: Binary data
