On 11/14/2015 5:27 AM, Jeffrey Walton wrote: > I'm trying to ensure I get a conditional move if its available during > a saturating subtract operation. The saturating subtract clamps the > min value at 0. > > Here's what it looks like in high level code: > > // Perform a-b, clamp at 0 > return (a > b) ? (a - b) : 0; > > But based on this bug report: > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22568, I can't tell if > that's what I should be using. > > What pattern should I use to get a CMOV? > > Thanks in advance. Under what I am guessing may be similar but not identical circumstances, I would expect better results with fmaxf(a - b,0.f), if it is possible to set -ffast-math. gfortran apparently has specific support in i386.md for optimization of max(a-b,0.) but it doesn't extend to g++ std::max() (as it does in icc). I'm not certain whether integer data types are treated symmetrically. Sorry if I've taken this too far afield into my personal gripes, but I'm not sure whether you have given enough clues to guess your target in terms of instruction set, data type, and admissible compile flags. -- Tim Prince
