<html><body><p><font size="2">--------------------------</font><br><font size="2">Wed, Dec 18, 2019:</font><br><font size="2">--------------------------</font><br><br><font size="2">- Next Meeting will be on Jan 15th, 2020. </font><br><br><font size="2"> ****Please consider the meeting on Jan 1st CANCELED.****</font><br><br><font size="2">- Dependence Analysis - Delinearization of constant size arrays. (Bardia)</font><br><font size="2"> - Support for delinearization of fixed size arrays were removed in </font><br><font size="2"> </font><a href="https://reviews.llvm.org/D35430"><font size="2">https://reviews.llvm.org/D35430</font></a><font size="2">.</font><br><font size="2"> - Some lost coverage in DA and interchange, also makes writing useful </font><br><font size="2"> tests harder.</font><br><font size="2"> - Bardia to post a patch soon to use GEPs for computing subscripts for</font><br><font size="2"> fixed-size arrays similar to how polly does it (and reuse/refactor </font><br><font size="2"> function in polly and SCEV).</font><br><font size="2"> - The patch will initially make use of </font><br><font size="2"> `-da-disable-delinearization-checks` but eventually can be replaced </font><br><font size="2"> with checking for the `inrange` keyword in GEP.</font><br><font size="2"> - `inrange` improvements needed: </font><br><font size="2"> 1. remove restriction of `constant` GEPs </font><br><font size="2"> 2. allow inrange on more than one index.</font><br><br><font size="2">- Implementation of Unroll-and-Jam (Whitney)</font><br><font size="2"> - Currently does not handle loops with more than one inner subloop.</font><br><font size="2"> - It is implemented as loop pass, thus traversing loops from inner to </font><br><font size="2"> outer and making transformation along the way. Unrolling often creates</font><br><font size="2"> residual subloops making subsequent UAJ of the outer loops impossible.</font><br><font size="2"> - Proposal:</font><br><font size="2"> - make it a function pass and walk the nest from outer to inner, which </font><br><font size="2"> should also result in cleaner code after the transform.</font><br><font size="2"> - Whitney to write a small RFC and send it to llvm.dev mailing list.</font><br><br><font size="2">- Status Updates</font><br><font size="2"> - Data Dependence Graph (Bardia)</font><br><font size="2"> - DDG ordinals </font><a href="https://reviews.llvm.org/D70986"><font size="2">https://reviews.llvm.org/D70986</font></a><font size="2"> ready for review/approval</font><br><font size="2"> - Loop Nest Analysis (Ettore)</font><br><font size="2"> - </font><a href="https://reviews.llvm.org/D68789"><font size="2">https://reviews.llvm.org/D68789</font></a><br><font size="2"> - Ettore to consider making it a utility instead of an analysis because </font><br><font size="2"> invalidation and update of the cached analysis results is problematic </font><br><font size="2"> and the compile-time benefit of caching it may not be realized with the</font><br><font size="2"> currently intended use cases. </font><br><font size="2"> - Loop Pragma RFC on llvm.dev (Michael Kruse)</font><br><font size="2"> - </font><a href="http://lists.llvm.org/pipermail/cfe-dev/2019-December/064122.html"><font size="2">http://lists.llvm.org/pipermail/cfe-dev/2019-December/064122.html</font></a><br><font size="2"> - Hal and Michael Kruse are preparing another RFC for loop opt framework </font><br><font size="2"> (to be also discussed in one of the upcoming calls as well).</font><br><br><font size="2">-----------------------------------</font><br><font size="2">Tentative Agenda for Jan 15, 2020</font><br><font size="2">-----------------------------------</font><br><br><font size="2"><Looking for topics></font><br><br><font size="2">- Status Updates</font><br><font size="2"> - Data Dependence Graph (Bardia)</font><br><font size="2"> - Loop Nest Analysis (Ettore)</font><br><br><br><font size="2">Bardia Mahjour<br>Compiler Optimizations<br>IBM Toronto Software Lab</font><BR>
</body></html>