On 06/17/2013 10:00 AM, Iyer, Balaji V wrote: > In hindsight, I could have for __sec_reduce_max and __sec_reduce_min. I was > more familiar with conditional expression. Out of curiosity, is there a big > performance benefit of using max/min expr over conditional?
There can be. The COND->MIN/MAX transformation is not done without the -ffinite-math-only component of -ffast-math. I.e. we don't try the transform when NaNs are a possibility. So, yes, you probably should generate MIN/MAX_EXPR right from the start. r~