Change Details

Do some more bzero "optimization". (That's in quotes because "optimization" is rarely straightforward. What "optimizes" one person's use case may well pessimize another's.) Main changes: * Do 8-byte comparisons in a loop prior to doing 1-byte comparisons in a loop. * In the large case, do 1-byte comparisons in a loop (instead of repe cmpsb). In a user-space test of many different memory blocks with different sizes (up to 105 bytes) and different alignments, this seems to reduce the run time by 42%. Additionally, running the same test @mjg ran (pmcstat to capture samples of bcmp), bcmp fell from being in 0.7% of samples to not showing up on the list at all. One nagging worry is whether making this function larger will negatively impact other aspects of the system.

Do some more bcmp "optimization". (That's in quotes because "optimization" is rarely straightforward. What "optimizes" one person's use case may well pessimize another's.) Main changes: * Do 8-byte comparisons in a loop prior to doing 1-byte comparisons in a loop. * In the large case, do 1-byte comparisons in a loop (instead of repe cmpsb). In a user-space test of many different memory blocks with different sizes (up to 105 bytes) and different alignments, this seems to reduce the run time by 42%. Additionally, running the same test @mjg ran (pmcstat to capture samples of bcmp), bcmp fell from being in 0.7% of samples to not showing up on the list at all. One nagging worry is whether making this function larger will negatively impact other aspects of the system.

Do some more bzerocmp "optimization". (That's in quotes because "optimization" is rarely straightforward. What "optimizes" one person's use case may well pessimize another's.) Main changes: * Do 8-byte comparisons in a loop prior to doing 1-byte comparisons in a loop. * In the large case, do 1-byte comparisons in a loop (instead of repe cmpsb). In a user-space test of many different memory blocks with different sizes (up to 105 bytes) and different alignments, this seems to reduce the run time by 42%. Additionally, running the same test @mjg ran (pmcstat to capture samples of bcmp), bcmp fell from being in 0.7% of samples to not showing up on the list at all. One nagging worry is whether making this function larger will negatively impact other aspects of the system.