On Thu, 21 Jul 2011, Tobias Grosser wrote: > On 07/21/2011 08:31 PM, Sebastian Pop wrote: > > Hi, > > > > This patch-set addresses the comments from Tobias on fixing PR47654. > > > > The first patches are cleanups: > > > > - Start counting nesting level from 0 and use the standard "Polyhedral > > SCattering Transformed" psct_* interface. > > - Do not compute twice type, lb, and ub. > > - Record the loop level that defines a clast_name. > > - Cleanup function params using a struct. > > - Add fixme comment. > > - Rename gcc_type_for_clast_* into type_for_clast_* > > - Remove max_signed_precision_type. > > > > The following patch is an optimization: > > > > - Compute once and cache the LB and UB for each clast_name. > > > > that uses the hashtable of clast_names to also attach the information > > on bounds of the name: this would avoid asking "polyhedral questions" > > every time we see the name of an induction variable in an expression. > > > > The following patch is also an optimization: > > > > - Generate signed types whenever possible. > > > > this makes us generate signed types instead of unsigned, whenever the > > interval would still fit in a signed type of the same precision. > > > > Finally the fix for PR47654: > > > > - Infer types based on lb and ub. > > > > I adapted the previous patch such that we now compute both the type > > and the bounds during the same traversal of the CLAST tree. The > > computation of the type now also integrates the information of the > > bounds of the sub-expressions and the result of the expression. > > > > The patch-set bootstrapped and tested on amd64-linux. Ok for trunk? > > Hi Sebastian, > > this looks nice. I tried, but was not able to find any bug. :-( > I think this is definitely a big improvement. From my side it's OK to commit.
Looks ok to me as well. Thanks, Richard.
