vandersonmr <vanderson...@gmail.com> writes:
> We add some of the statistics collected in the TCGProfiler > into the TBStats, having the statistics not only for the whole > emulation but for each TB. Then, we removed these stats > from TCGProfiler and reconstruct the information for the > "info jit" using the sum of all TBStats statistics. > > The goal is to have one unique and better way of collecting > emulation statistics. Moreover, checking dynamiclly if the > profiling is enabled showed to have an insignificant impact > on the performance: > https://wiki.qemu.org/Internships/ProjectIdeas/TCGCodeQuality#Overheads. > > Signed-off-by: Vanderson M. do Rosario <vanderson...@gmail.com> > --- > accel/tcg/tb-stats.c | 95 +++++++++++++++++++++++++++++++++++++++ > accel/tcg/translate-all.c | 8 +--- > include/exec/tb-stats.h | 11 +++++ > tcg/tcg.c | 93 +++++--------------------------------- > tcg/tcg.h | 10 ----- > 5 files changed, 118 insertions(+), 99 deletions(-) > > diff --git a/accel/tcg/tb-stats.c b/accel/tcg/tb-stats.c > index 3489133e9e..9b720d9b86 100644 > --- a/accel/tcg/tb-stats.c > +++ b/accel/tcg/tb-stats.c > @@ -1,9 +1,104 @@ > #include "qemu/osdep.h" > > #include "disas/disas.h" > +#include "exec/exec-all.h" > +#include "tcg.h" > + > +#include "qemu/qemu-print.h" > > #include "exec/tb-stats.h" > > +struct jit_profile_info { > + uint64_t translations; > + uint64_t aborted; > + uint64_t ops; > + unsigned ops_max; > + uint64_t del_ops; > + uint64_t temps; > + unsigned temps_max; > + uint64_t host; > + uint64_t guest; > + uint64_t search_data; > +}; > + > +/* accumulate the statistics from all TBs */ > +static void collect_jit_profile_info(void *p, uint32_t hash, void *userp) > +{ > + struct jit_profile_info *jpi = userp; > + TBStatistics *tbs = p; > + > + jpi->translations += tbs->translations.total; > + jpi->ops += tbs->code.num_tcg_ops; > + if (stat_per_translation(tbs, code.num_tcg_ops) > jpi->ops_max) { > + jpi->ops_max = stat_per_translation(tbs, code.num_tcg_ops); > + } > + jpi->del_ops += tbs->code.deleted_ops; > + jpi->temps += tbs->code.temps; > + if (stat_per_translation(tbs, code.temps) > jpi->temps_max) { > + jpi->temps_max = stat_per_translation(tbs, code.temps); > + } > + jpi->host += tbs->code.out_len; > + jpi->guest += tbs->code.in_len; > + jpi->search_data += tbs->code.search_out_len; > +} > + > +/* dump JIT statisticis using TCGProfile and TBStats */ > +void dump_jit_profile_info(TCGProfile *s) > +{ > + if (!tb_stats_collection_enabled()) { > + return; > + } > + > + struct jit_profile_info *jpi = g_new0(struct jit_profile_info, 1); > + > + qht_iter(&tb_ctx.tb_stats, collect_jit_profile_info, jpi); > + > + if (jpi->translations) { > + qemu_printf("translated TBs %" PRId64 "\n", jpi->translations); > + qemu_printf("avg ops/TB %0.1f max=%d\n", > + jpi->ops / (double) jpi->translations, jpi->ops_max); > + qemu_printf("deleted ops/TB %0.2f\n", > + jpi->del_ops / (double) jpi->translations); > + qemu_printf("avg temps/TB %0.2f max=%d\n", > + jpi->temps / (double) jpi->translations, jpi->temps_max); > + qemu_printf("avg host code/TB %0.1f\n", > + jpi->host / (double) jpi->translations); > + qemu_printf("avg search data/TB %0.1f\n", > + jpi->search_data / (double) jpi->translations); > + > + if (s) { > + int64_t tot = s->interm_time + s->code_time; > + qemu_printf("JIT cycles %" PRId64 " (%0.3f s at 2.4 > GHz)\n", > + tot, tot / 2.4e9); > + qemu_printf("cycles/op %0.1f\n", > + jpi->ops ? (double)tot / jpi->ops : 0); > + qemu_printf("cycles/in byte %0.1f\n", > + jpi->guest ? (double)tot / jpi->guest : 0); > + qemu_printf("cycles/out byte %0.1f\n", > + jpi->host ? (double)tot / jpi->host : 0); > + qemu_printf("cycles/search byte %0.1f\n", > + jpi->search_data ? (double)tot / jpi->search_data : > 0); > + if (tot == 0) { > + tot = 1; > + } > + qemu_printf(" gen_interm time %0.1f%%\n", > + (double)s->interm_time / tot * 100.0); > + qemu_printf(" gen_code time %0.1f%%\n", > + (double)s->code_time / tot * 100.0); > + qemu_printf("optim./code time %0.1f%%\n", > + (double)s->opt_time / (s->code_time ? s->code_time : > 1) > + * 100.0); > + qemu_printf("liveness/code time %0.1f%%\n", > + (double)s->la_time / (s->code_time ? s->code_time : 1) * > 100.0); > + qemu_printf("cpu_restore count %" PRId64 "\n", > + s->restore_count); > + qemu_printf(" avg cycles %0.1f\n", > + s->restore_count ? (double)s->restore_time / > s->restore_count : 0); > + } > + } I think the g_free(jpi) should be moved from the later patches to here. Otherwise: Reviewed-by: Alex Bennée <alex.ben...@linaro.org> -- Alex Bennée
