This transformation is not blindly applied to every case. It is invoked when the then-paths of two consecutive if-regions contain loops that can be fused together.
-Mei -----Original Message----- From: Sun Chan [mailto:sun.c...@gmail.com] Sent: Wednesday, May 18, 2011 2:41 PM To: Ye, Mei Cc: open64-devel@lists.sourceforge.net Subject: Re: [Open64-devel] code review - more if-merging transformations why is this xform better? If comp is false, you have added extra compare needed. Sun On Thu, May 19, 2011 at 2:52 AM, Ye, Mei <mei...@amd.com> wrote: > Attached are two patches. Patch1 does a transformation as explained below. > Patch2 changes the interface of a function call returning two values. > > > > Problem: > > In the code shown below, the first if-condition has an explicit cast to > "unsigned long long", since "comp" is an "int", so it will need to be > casted to a "signed 64 bit" followed by an unsigned" cast before its value > is bit-anded. If "comp" is a negative value, and "width - 1" is greater or > equal to "32", the 1st if-condition can be evaluated to "TRUE", and the 2nd > if-condition can be evaluated to "FALSE". Therefore, it is illegal to > if-merging them. > > if (comp & ((unsigned long long) 1 << (width - 1))) { > > .... > > } > > else { > > ... > > } > > if (comp & (1<<(width-1))) { > > .... > > } > > else { > > .... > > } > > Solution: > > If we create a if-condition "if (width <= 32)" and wrap it around the if- > regions, we can remove the conversion in the THEN-path of the first if- > region. Therefore we can if-merge the THEN-paths. > > > > if (width <= 32) { > > if (comp & ( 1 << (width - 1))) { > > } > > else { > > .... > > } > > } > > else { > > if (comp & ((unsigned long long) 1 << (width - 1))) { > > ... > > } > > else { > > ... > > } > > } > > ..... > > if (width <= 32) { > > if (comp & ( 1 << (width - 1))) { > > ... > > } > > else { > > ... > > } > > } > > else { > > if (comp & ( 1 << (width - 1))) { > > ... > > } > > else { > > ... > > } > > } > > > > changes: > > opt_proactive.cxx > > Pattern match and invoke transformations as described above. Also add code > to infer values when possible. > > opt_cfg.[h,cxx] > > Add code to Insert a SC_IF. > > opt_wn.cxx > > Add code to evaluate a multiply operation. > > ------------------------------------------------------------------------------ > What Every C/C++ and Fortran developer Should Know! > Read this article and learn how Intel has extended the reach of its > next-generation tools to help Windows* and Linux* C/C++ and Fortran > developers boost performance applications - including clusters. > http://p.sf.net/sfu/intel-dev2devmay > _______________________________________________ > Open64-devel mailing list > Open64-devel@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/open64-devel > > ------------------------------------------------------------------------------ What Every C/C++ and Fortran developer Should Know! Read this article and learn how Intel has extended the reach of its next-generation tools to help Windows* and Linux* C/C++ and Fortran developers boost performance applications - including clusters. http://p.sf.net/sfu/intel-dev2devmay _______________________________________________ Open64-devel mailing list Open64-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/open64-devel ------------------------------------------------------------------------------ What Every C/C++ and Fortran developer Should Know! Read this article and learn how Intel has extended the reach of its next-generation tools to help Windows* and Linux* C/C++ and Fortran developers boost performance applications - including clusters. http://p.sf.net/sfu/intel-dev2devmay _______________________________________________ Open64-devel mailing list Open64-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/open64-devel
