On Fri, May 23, 2025 at 2:39 AM Richard Biener <richard.guent...@gmail.com> wrote: > > On Thu, May 22, 2025 at 3:11 AM Jeff Law <jeffreya...@gmail.com> wrote: > > > > > > > > On 5/18/25 10:38 AM, Andrew Pinski wrote: > > > While working on improving forwprop and removal of > > > forward_propagate_into_gimple_cond/forward_propagate_into_comparison, > > > I came cross a case where we end up with SSA_NAME in the resulting > > > gimple_match_op and one statement in the sequence. This was the result > > > of simplification of: > > > ``` > > > _3 = MIN_EXPR <maxlen_2(D), 264> > 16 > > > if (_3 > 16) ... > > > ``` > > > > > > Which simplifies down to: > > > (maxlen_2(D) > 16) & (264 > 16) > > > into > > > (maxlen_2(D) > 16) & 1 > > > > > > Which `maxlen_2(D) > 16` gets pushed onto the sequence > > > and then the & 1 is removed via the match pattern: > > > ``` > > > /* x & ~0 -> x */ > > > (simplify > > > (bit_and @0 integer_all_onesp) > > > (non_lvalue @0)) > > > ``` > > > > > > So what this patch does is to undo the push extracting the new op > > > from the pushed statement and remove the sequence as it is not used > > > any more. > > > > > > Bootstrapped and tested on x86_64-linux-gnu. > > > > > > gcc/ChangeLog: > > > > > > PR tree-optimization/120331 > > > * gimple-match-exports.cc (maybe_undo_push): New function. > > > (gimple_simplify): Call maybe_undo_push if resimplify was > > > successfull. > > OK, but this isn't my strongest area. So give Richi until Monday AM to > > chime in if he's got concerns. > > It's a bit ugly and IMO it should be in-line at the single user and have an > explaining comment why we still might see such cases. I'll note that > this does not help when match recursively applies simplifications or when > 'seq' is passed as NULL. So addressing this in the genmatch generated > code directly would be superior (but IIRC what the bug said this isn't > possible?)
I am going to try again to see if there is a way to get genmatch generated code to do it. I hope to get to it next week. Thanks, Andrew > > Richard. > > > Thanks, > > jeff > >
