On 16/11/18 18:19, Pat Haugen wrote:
On 11/8/18 6:10 AM, Kyrill Tkachov wrote: > The attached patch avoids that by making the alap calculation only > look at true dependencies. This shouldn't be too bad, since we use > INSN_PRIORITY as the final tie-breaker than that does take > anti-dependencies into account. > > This reduces the number of spills in the hot function from 436.cactusADM > by 14% on aarch64 at -O3 (and the number of instructions in general). > SPEC2017 shows a minor improvement on Cortex-A72 (about 0.1% overall). > Thanks to Wilco for the benchmarking. I tried the patch on PowerPC since it also uses SCHED_PRESSURE_MODEL algorithm. For CPU2006 only cactusADM had a noticeable difference, but I'm seeing a 5% degradation. Looking at the generated asm for function bench_staggeredleapfrog2_(), I see about a 1% increase in number of loads and stores generated and an extra 100 bytes allocated on the stack.
Thanks for trying it out! Given that, I'll try to polish up the more targeted approach that filters out only the fake dependencies, which should give a more targeted approach. Thanks, Kyrill
-Pat
