Hacker News Reader

Seems like the compiler puts the test for the first loop before loading the constant the first time, and therefor needs to load it again before the second loop. So the "tradeoff" is that if neither loop runs it will load the constant zero times. Of course this isn't what a human would do but at least there is some kind of sliver of logic to it. (Like if vpbroadcastd was a 2000 cycle instruction this pattern might have made sense)

View on HN

In this modified version where the for-loops get converted to do-loops (possibly unsafe/incorrect behavior), unnecessary multiple vpbroadcastd's for ymm1 are still done:

https://godbolt.org/z/61jYejsra

With the original, for-loop version, it could be argued that in case no loop iteration gets run at all, the vpbroadcastd's can be totally skipped, and to generate extra code for different cases of whether each loop is empty to avoid unnecessary vpbroadcastd's is not worth it (e.g. the greater icache pressure resulting from longer code).

With the do-loop variant, both loops will get at least one iteration so the compiler really could just unconditionally do the vpbroadcastd once before the first loop. It somehow fails to realize that ymm1 did not get clobbered between the two vpbroadcastd's.

Reply on news.ycombinator.com