[PATCH] D52685: [LoopVectorizer] Adjust heuristics for a truncated load

[Hideki Saito via Phabricator via llvm-commits]

hsaito added a comment.

I prefer not to add this kind of "heuristics" on the functions that should just return "the fact". Even if the loaded value is truncated immediately, the load itself needs to be vectorized.

These kinds of issues should be handled by enabling TTI.shouldMaximizeVectorBandwidth() and then apply appropriate cost modeling fix on it. If we keep adding those little things
to places that doesn't really belong, we'll never fix the cost model properly to enable TTI.shouldMaximizeVectorBandwidth(). Please use this as an incentive to spend time on the cost
model. Please don't get confused by the name of the function. It just means vectorizer has more VF choices to lower the cost of executing the loop.

Thanks,
Hideki


https://reviews.llvm.org/D52685