On Tue, Apr 21, 2020 at 10:42 AM Kewen.Lin <li...@linux.ibm.com> wrote: > > on 2020/4/17 下午7:32, Richard Biener wrote: > > On Fri, Apr 17, 2020 at 1:10 PM Kewen.Lin via Gcc <gcc@gcc.gnu.org> wrote: > >> > >> Hi all, > >> > >> This is one question origining from PR61837, which shows that doloop > >> pass could introduce some loop invariants against its outer loop when > >> doloop performs and prepares the iteration count for hardware > >> count register assignment. > >> > > I suggest to try both approaches and count the number of transforms > > done in each instance. > > > > Richard. > > > > Hi Richi, > > Thanks for the suggestion, some statistics were collected as below. > > A: default > B: move pass_rtl_move_loop_invariants after pass_rtl_doloop > C: rerun pass_rtl_move_loop_invariants after pass_rtl_doloop > D: C + doloop_begin > > Ran by bootstrapping and regression testing on ppc64le Power8 configured > with languages c,c++,fortran,objc,obj-c++,go. > > Counting move #transformations in function move_invariant_reg (before > label fail, probably better with inv == repr to filter out those > replacements with rep, but I think the trend is similar?). > > A: 802650 > B: 841476 > C: 803485 (C1) + 827883 (C2) > D: 802622 (D1) + 841476 (D2) > > Let's call pass_rtl_move_loop_invariants as hoisting. > PS: C1/D1 for 1st time hoisting while C2/D2 for 2nd time hoisting. > The small differences (~0.1%) among A/C1/D1 should be caused by noise. > > The numbers with twice runs (C/D) are almost two times of one time run, > which surprised me. By further investigation, it looks the current > pass_rtl_move_loop_invariants has something to be improved if we want > to rerun it. Taking gcc/testsuite/gfortran.dg/inline_matmul_16.f90 at > -O1 as example. C1 does 178 transforms and C2 does 165, it's unrelated > to unroll/doloop passes, this result isn't changed by disabling them > explicitly. > > Currently, without flag_ira_loop_pressure, the regs_used estimation > isn't good, I'd expect that invs which are hoisted first time from > the loop should be counted as regs_used next time at regs_used > analysis. By checking the regs_used, it's set as 2 for all loops of > case inline_matmul_16, either C1 or C2. I think it leads the 2nd > hoisting optimistically estimate register pressure and hoist more. > By simple hacking by considering 1st hoisting new_reg, I can see the > 2nd hoisting has fewer moves (57). It means the above statistics > comparison is unfair and unreliable.
OK, so that alone argues against doing C or D without better understanding and fixing this. That is, when you invoke invariant motion twice at its current place the second invocation shouldn't really do any more hoisting, definitely not a significant amount. > With flag_ira_loop_pressure, the #transforms become to 255 (1st) and > 68 (2nd), it looks better but might also need more enhancements? > > Since rs6000 sets flag_ira_loop_pressure at O3, I did SPEC2017 > performance evaluation on Power8 (against baseline A) with option > -Ofast -funroll-loops: > * B showed 525.x264_r +1.43%, 538.imagick_r +1.23% speedup > but 503.bwaves_r -2.74% degradation. > * C showed 500.perlbench_r -1.31%, 520.omnetpp_r -2.20% degradation. > > The evaluation shows running hoisting after doloop can give us some > benefits, but to rerun it twice isn't able to give us the similar > gains. It looks regardless of flag_ira_loop_pressure, to rerun > the pass requires more tweaks, probably considering those related > parameters. If go with B, we need to figure out what we miss forbwaves_r. Of course it also requires benchmarking on other archs. > BR, > Kewen >