On Tue, Jul 14, 2015 at 5:29 PM, Ajit Kumar Agarwal <ajit.kumar.agar...@xilinx.com> wrote: > > > -----Original Message----- > From: Richard Biener [mailto:richard.guent...@gmail.com] > Sent: Tuesday, July 14, 2015 6:35 PM > To: Ajit Kumar Agarwal > Cc: Jeff Law; Jan Hubicka; gcc@gcc.gnu.org; Vinod Kathail; Shail Aditya > Gupta; Vidhumouli Hunsigida; Nagaraju Mekala > Subject: Re: Traces on Data Dependency graph. > > On Tue, Jul 14, 2015 at 2:12 PM, Ajit Kumar Agarwal > <ajit.kumar.agar...@xilinx.com> wrote: >> All: >> >> I am wondering how useful to form the traces on Data Dependency Graph. >> On top of the traces in the Control flow graph, I was thinking of forming >> the traces on data Dependency graph(DDG). >> >> Would this helps in further vectorization and parallelization candidates. >> >> Thoughts? > >>>Not sure how you'd do that? Transform data dependences to control >>>dependences? > > Work has done with respect to the control flow traces and the dependency > traces where the control flow traces > Can also explicitly identifies the def and use with respect to the registers > but for the flow information of memory is difficult > To extract from the control flow traces and thus dependent on data dependency > traces. The data dependence traces > Are long in the sense that it requires the execution instance of the > statement along with the data dependence. The DDG > Graph annotates the statement along with the statement which can be further > enhanced with execution instance of the > Statement. There has been work done by Zhang and Gupta etal to form the whole > execution Traces with respect to control flow > edges and the dependency edges at granularity level of statement and the > blocks. > > I was thinking of data dependency traces with respect to the DDG having the > statements and augmented with the execution > Instance of the statement that will help with traces for flow information of > memory. > > Both Dependence(control and data) profiles have been used in Data Speculative > optimizations (Lin etal 2003 - 2004), Speculative > Optimizations (2004) and computation of dynamic slices. > > I am not sure how useful would it be and would like to have your opinion on > this.
I'm still not sure what you are after. The loop distribution pass basically computes DDG "traces" and distributes independent traces to independent loops. But it only works on loops and I wouldn't call this "traces". Richard. > Thanks & Regards > Ajit > >> Thanks & Regards >> Ajit
