On 12/12/2017 01:28 AM, Markus Trippelsdorf wrote: > As the testcase shows, trunk currently generates horrible code for > divisions used in tight loops. This happens because the algorithm > expanding div/mod doesn't take parallelism into account and this makes > the cost model unrealistic. > Fix the issue by increasing the estimated latencies a bit. > > Tested on X86_64. > OK for trunk? > (I will keep an eye on the periodic SPEC testers.) > > PR target/83358 > * x86-tune-costs.h (skylake_cost, core_cost): Increase div/mod > latencies a bit. THanks. I added a ChangeLog for the testsuite and installed this onto the trunk.
We can iterate on other micro-architecture costs if necessary. jeff
