On Fri, Jun 23, 2017 at 11:50 PM, Marc Glisse <marc.gli...@inria.fr> wrote: > On Fri, 23 Jun 2017, Andrew Pinski wrote: > >> Hi, >> I saw this on llvm's review site (https://reviews.llvm.org/D34579) >> and I thought why not add it to GCC. I expanded more than what was >> done on the LLVM patch. >> >> I added the following optimizations: >> Transform X * (X > 0 ? 1 : -1) into ABS(X). >> Transform X * (X >= 0 ? 1 : -1) into ABS(X). >> Transform X * (X > 0.0 ? 1.0 : -1.0) into ABS(X). >> Transform X * (X >= 0.0 ? 1.0 : -1.0) into ABS(X). >> Transform X * (X > 0 ? -1 : 1) into -ABS(X). >> Transform X * (X >= 0 ? -1 : 1) into -ABS(X). >> Transform X * (X > 0.0 ? -1.0 : 1.0) into -ABS(X). >> Transform X * (X >= 0.0 ? -1.0 : 1.0) into -ABS(X). >> Transform X * (X < 0 ? 1 : -1) into -ABS(X). >> Transform X * (X <= 0 ? 1 : -1) into -ABS(X). >> Transform X * (X < 0.0 ? 1.0 : -1.0) into -ABS(X). >> Transform X * (X <= 0.0 ? 1.0 : -1.0) into -ABS(X). >> Transform X * (X < 0 ? -1 : 1) into ABS(X). >> Transform X * (X <= 0 ? -1 : 1) into ABS(X). >> Transform X * (X < 0.0 ? -1.0 : 1.0) into ABS(X). >> Transform X * (X <= 0.0 ? -1.0 : 1.0) into ABS(X). >> >> The floating points ones only happen when not honoring SNANS and not >> honoring signed zeros. > > > Some random comments (not a review): > > * if X is NaN, we may get a qNaN with the wrong sign bit. We probably don't > care much though...
Ok, I changed it to when not honoring NANs. > > * I am surprised (X<0.?-1.:1.) and copysign(1., X) remain different for the > whole optimization pipeline with -ffast-math. X*copysign(1., X) is another > candidate to become fabs(X). This might be a better idea because of ... > > * Whenever you get -ABS(X) for integers, what about the case where X is > INT_MIN? This. Yes this is an issue; I guess I need to rethink the integer patterns. > > * I guess we can't get there with an unsigned type because X>0 would have > become X!=0 . No, unsigned is not an issue. > > * I wonder if we could use something like > > (for cmp (gt ge lt le) > outp (convert convert negate negate) > outn (negate negate convert convert) > [...] > (outp (abs @0)) > > to reduce duplication or if that would be less readable. I did thought of that but I added the lt/le parts latter on. > > * Some of the cases are handled by PRE turning > > # iftmp.0_1 = PHI <1.0e+0(5), -1.0e+0(3)> > _3 = iftmp.0_1 * a_2(D); > > into > > _5 = -a_2(D); > [...] > # iftmp.0_1 = PHI <1.0e+0(2), -1.0e+0(3)> > # prephitmp_6 = PHI <a_2(D)(2), _5(3)> > > which phiopt3 can handle (quite late). > > * With cond, this currently (?) only affects generic, so I am not sure it > will hit very often... But it will be there if someone later writes a > match.pd->phiopt generator ;-) I have a start of this patch but I have not finished it yet. Both phiopt and ifcombine should be moved over to gimple match and simplify. I will submit a new patch which implements some of the above by the end of the day; I might split up the patch into two (one for the integer case and one for the floating point case). Thanks, Andrew > > -- > Marc Glisse
