On Wed, Aug 26, 2015 at 3:16 PM, Jakub Jelinek <ja...@redhat.com> wrote: > On Wed, Aug 26, 2015 at 03:02:02PM +0200, Richard Biener wrote: >> > AVX-512 is such target. Current representation forces multiple scalar >> > mask -> vector mask and back transformations which are artificially >> > introduced by current bool patterns and are hard to optimize out. >> >> I dislike the bool patterns anyway and we should try to remove those >> and make the vectorizer handle them in other ways (they have single-use >> issues anyway). I don't remember exactly what caused us to add them >> but one reason was there wasn't a vector type for 'bool' (but I don't see how >> it should be necessary to ask "get me a vector type for 'bool'"). > > That was just one of the reasons. The other reason is that even if we would > choose some vector of integer type as vector of bool, the question is what > type. E.g. if you use vector of chars, you almost always get terrible > vectorized code, except for the AVX-512 you really want an integral type > that has the size of the types you are comparing.
Yeah, but the way STMT_VINFO_VECTYPE is computed is that we always first compute the vector type for the comparison itself (which is "fixed") and thus we can compute the vector type of any bitwise op on it as well. > And I'd say this is very much related to the need to do some type promotions > or demotions on the scalar code meant to be vectorized (but only the copy > for vectorizations), so that we have as few different scalar type sizes in > the loop as possible, because widening / narrowing vector conversions aren't > exactly cheap and a single char operation in a loop otherwise full of long > long operations might unnecessarily turn a vf=2 (or 4 or 8) loop into > vf=16 (or 32 or 64), increasing it a lot. That's true but unrelated. With conditions this gets to optimizing where the promotion/demotion happens (which depends on how the result is used). The current pattern approach has the issue that it doesn't work for multiple uses in the condition bitops which is bad as well. But it couldn't have been _only_ the vector type computation that made us invent the patterns, no? Do you remember anything else? Thanks, Richard. > > Jakub
