<html><head><meta http-equiv="Content-Type" content="text/html charset=us-ascii"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" class=""><br class=""><div><blockquote type="cite" class=""><div class="">On May 10, 2016, at 6:06 PM, Hal Finkel <<a href="mailto:hfinkel@anl.gov" class="">hfinkel@anl.gov</a>> wrote:</div><br class="Apple-interchange-newline"><div class=""><div style="font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; font-family: arial, helvetica, sans-serif; font-size: 10pt;" class=""><br class=""><br class=""><hr id="zwchr" class=""><blockquote style="border-left-width: 2px; border-left-style: solid; border-left-color: rgb(16, 16, 255); margin-left: 5px; padding-left: 5px; font-weight: normal; font-style: normal; text-decoration: none; font-family: Helvetica, Arial, sans-serif; font-size: 12pt;" class=""><b class="">From:<span class="Apple-converted-space"> </span></b>"vivek pandya" <<a href="mailto:vivekvpandya@gmail.com" class="">vivekvpandya@gmail.com</a>><br class=""><b class="">To:<span class="Apple-converted-space"> </span></b>"llvm-dev" <<a href="mailto:llvm-dev@lists.llvm.org" class="">llvm-dev@lists.llvm.org</a>>, "Tim Amini Golling" <<a href="mailto:mehdi.amini@apple.com" class="">mehdi.amini@apple.com</a>>, "Hal Finkel" <<a href="mailto:hfinkel@anl.gov" class="">hfinkel@anl.gov</a>><br class=""><b class="">Cc:<span class="Apple-converted-space"> </span></b>"Quentin Colombet" <<a href="mailto:qcolombet@apple.com" class="">qcolombet@apple.com</a>><br class=""><b class="">Sent:<span class="Apple-converted-space"> </span></b>Tuesday, May 10, 2016 2:59:16 PM<br class=""><b class="">Subject:<span class="Apple-converted-space"> </span></b>[GSoC 2016] Interprocedural Register Allocation - Introduction and Feedback<br class=""><br class=""><div dir="ltr" class=""><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">Hello LLVM Community,<br class=""></div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica; min-height: 14px;" class=""><br class=""></div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">Sorry for delay as I was busy in final exams.</div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica; min-height: 14px;" class=""><br class=""></div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">I am Vivek from India. Thanks for choosing my proposal for Interprocedural Register Allocation (IPRA) in LLVM. Mehdi Amini and Hal Finkel will be mentoring me for this project.</div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica; min-height: 14px;" class=""><br class=""></div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">IPRA can reduce code size and runtime of programs by allocating register across the module and procedure boundaries.</div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica; min-height: 14px;" class=""><br class=""></div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">I have identified some old but effective research work on this area.</div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">I want community's feedback for feasibility of these approach and I am targeting to implement two of them during this project.</div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica; min-height: 14px;" class=""><br class=""></div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">Here is list of the papers, I have read first two papers and I would like to discuss those approach first, I will read other two paper then initiate discussion for them as well. All I want is to find out a concrete implementation plan before 23 May, 2016 and for that I need community's help.</div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica; min-height: 14px;" class=""><br class=""></div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">1) Compile time ----- Minimizing register usage penalty at procedure calls -<span class="Apple-converted-space"> </span><a href="http://dl.acm.org/citation.cfm?id=53999" target="_blank" class="">http://dl.acm.org/citation.cfm?id=53999</a></div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">====================================================================In this approach intra-procedural register allocation is used as base but machine code generation order is bottom up traversal of call graph and inter-procedural effect is achieved by propagating register usage information of callee function to caller (i.e child to parent in CallGraph) so that caller can use different registers than callee and can save load store cost at procedure call, this is not trivial as it seems due to recursive calls, library function usage etc. Also for upper region of the graph in this technique available number of registers might become zero in that case it should fall back to normal load store at procedure call. Apart from these difficulties other difficulties have been identified please follow this mail-chain<span class="Apple-converted-space"> </span><a href="https://groups.google.com/d/topic/llvm-dev/HOYAXv3m1LY/discussion" target="_blank" class="">https://groups.google.com/d/topic/llvm-dev/HOYAXv3m1LY/discussion</a></div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">My mentor has already provided me a patch that alters code generation order as per bottom up call graph traversal, I am working from that point now. Any other help/suggestion is always welcomed.</div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica; min-height: 14px;" class=""><br class=""></div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">2) Link time ----- Global register allocation at link time -<span class="Apple-converted-space"> </span><a href="http://dl.acm.org/citation.cfm?id=989415" target="_blank" class="">http://dl.acm.org/citation.cfm?id=989415</a></div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">====================================================================In this particular approach (sort of true IPRA) registers will be reallocated (this optimization will be optional if turned off still code will be compiled as per intra-procedural allocation) at link time. Here modules are first complied as per normal compilation but the object code is annotated with details so that linker can build call graph and also calculate usage information at link time. Compiler also write hints in object code that if particular variable is allocated in some other register ( due to new allocation) then how the code should be changed? Thus linker can use these information to decide which variables (global) need to be in same register through out the program execution and also according to register usage information in call graph which procedure will not be active simultaneously so that locals for that procedures can be in same registers with out load store at procedure calls. </div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">For these particular method help me to analyze feasibility: </div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">1) Can llvm collects following information at module level in MachineIR? list of procedures in module, list of locals in procedures, list of procedures that a particular procedure can call, and a list of the variables this procedure references. Each entry in the last two lists includes an estimate of the number of times the procedure is called or the variable is referenced in each execution of this procedure </div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">2) Can llvm write informative commands to object files?</div><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">3) Can LTO is capable of leveraging those commands?<span class="Apple-converted-space"> </span><br class=""></div></div></blockquote>In terms of scoping the project for the summer, I definitely recommend that you focus on (1) first. If you finish that, we can certainly move on to other things. </div></div></blockquote><div><br class=""></div><div>I'll add +1 here, but I already wrote the same thing on IRC when discussing with Vivek. True IPRA without a proper MachineModule infrastructure won't be doable in my opinion (even with such infrastructure, it may not be trivial in LLVM in general).</div><br class=""><blockquote type="cite" class=""><div class=""><div style="font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; font-family: arial, helvetica, sans-serif; font-size: 10pt;" class="">Regarding link time, note that any such a design would likely look much different than in David Wall's paper however, because our LTO re-codegens everything anyway. The paper says, "Finally, it keeps us honest as designers of the system; once we postpone anything until link time, the temptation is great to postpone everything, ..." - Well, we've long-since succumb to that temptation when we LTO. C'est la vie.<br class=""></div></div></blockquote><div><br class=""></div><div>+1 as well, our LTO will benefit naturally from the leaf-to-root information propagation. ThinLTO will be more challenging/interesting though!</div><blockquote type="cite" class=""><div class=""><div style="font-style: normal; font-variant-caps: normal; font-weight: normal; letter-spacing: normal; orphans: auto; text-align: start; text-indent: 0px; text-transform: none; white-space: normal; widows: auto; word-spacing: 0px; -webkit-text-stroke-width: 0px; font-family: arial, helvetica, sans-serif; font-size: 10pt;" class=""><blockquote style="border-left-width: 2px; border-left-style: solid; border-left-color: rgb(16, 16, 255); margin-left: 5px; padding-left: 5px; font-weight: normal; font-style: normal; text-decoration: none; font-family: Helvetica, Arial, sans-serif; font-size: 12pt;" class=""><div dir="ltr" class=""><p style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class=""></p><div style="margin: 0px; font-size: 12px; line-height: normal; font-family: Helvetica;" class="">For the first part a mechanism similar to MachineModulePass would be desirable but that may not be possible during this project, but if we can make some sort of smaller version of that to suit our purpose.</div></div></blockquote>I don't think we need to make any kind of MachineModulePass to make this work. Once we alter the visitation order based on the CGSCC iteration scheme, we can keep state in-between functions in the pre-existing hacky way (using static members of the relevant function passes).<br class=""></div></div></blockquote><div><br class=""></div><div>I also don't see where/why we need a MachineModule(Pass) for the CGSCC scheme, that said I'd rather avoid using a function pass with static members, if we can have a ModuleAnalysis that is bookkeeping the results for functions in the module and queries by the register allocator somehow.</div><div>Matthias/Quentin may have other inputs on this aspect.</div><div><br class=""></div><div>-- </div><div>Mehdi</div><div><br class=""></div></div></body></html>