<div dir="ltr">Right now the mir-cannon Canonicalization pass is limited within a basic block. I don't know the full details of how the outliner works, but I suspect if outlining success rate is limited by factors like vreg names and/or instruction ordering then it could be helpful. Of course, I think if you used some canonicalization pass to reorder the schedule prior to outlining, then you'd definitely want to run the outlined code through the scheduler.<div><br></div><div>PL</div></div><div class="gmail_extra"><br><div class="gmail_quote">On Thu, Aug 31, 2017 at 5:12 PM, Jessica Paquette <span dir="ltr"><<a href="mailto:jpaquette@apple.com" target="_blank">jpaquette@apple.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div style="word-wrap:break-word;line-break:after-white-space">Hey Puyan,<div><br></div><div>Do you think that mir-canon could be used with, say, the MachineOutliner/other outlining/code size technologies?</div><div><br></div><div>If mir-canon is inter-procedural, I imagine that reducing the differences between semantically equivalent sequences of instructions might reveal more opportunities for outlining.</div><span class="HOEnZb"><font color="#888888"><div><br></div><div>- Jessica</div></font></span><div><div><br><blockquote type="cite"><div><div class="h5"><div>On Aug 15, 2017, at 12:06 PM, Puyan Lotfi via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>> wrote:</div><br class="m_-572906451126838774Apple-interchange-newline"></div></div><div><div><div class="h5"><div dir="ltr"><span id="m_-572906451126838774gmail-docs-internal-guid-44256f1b-e740-950e-75a8-4282e27328f5"><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:14pt;font-family:Arial;background-color:transparent;font-weight:700;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap"><span style="font-size:14.666666984558105px;font-weight:normal">Hi,</span></span></div><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:14pt;font-family:Arial;background-color:transparent;font-weight:700;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap"><span style="font-size:14.666666984558105px;font-weight:normal"><br></span></span></div><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:14pt;font-family:Arial;background-color:transparent;font-weight:700;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap"><span style="font-size:14.666666984558105px;font-weight:normal">My name is Puyan and </span></span><span style="font-family:Arial;font-size:14.666666984558105px;white-space:pre-wrap;background-color:transparent">I've been exploring ways to improve the state of instruction level diffing using llvm and MIR. Below is a proposal for a new llvm tool meant to address issues encountered when diffing at the machine level. I'm eager to hear the community's feedback.</span></div><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-family:Arial;font-size:14.666666984558105px;white-space:pre-wrap;background-color:transparent"><br></span></div><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-family:Arial;font-size:14.666666984558105px;white-space:pre-wrap;background-color:transparent">Thanks</span></div><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-family:Arial;font-size:14.666666984558105px;white-space:pre-wrap;background-color:transparent"><br></span></div><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-family:Arial;font-size:14.666666984558105px;white-space:pre-wrap;background-color:transparent">PL</span></div><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:14pt;font-family:Arial;background-color:transparent;font-weight:700;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap"><br></span></div><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:14pt;font-family:Arial;background-color:transparent;font-weight:700;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">mir-canon: A new tool for canonicalizing MIR for cleaner diffing.</span></div><br><br><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;font-family:Arial;background-color:transparent;font-weight:700;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Problem Statement and Context:</span></div><br><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;font-family:Arial;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Diff-tools are regularly used for comparing IR and assembly files. This often involves reasoning through differences that are semantically equivalent and it can be very time consuming for a person to do said reasoning.</span></div><br><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;font-family:Arial;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Specifically in the context of GlobalISel development there are problems of correctness verification. There is a need to compare two programs, compiled from identical IR by two different instruction selectors in a way where the true differences stand out.</span></div><br><br><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;font-family:Arial;background-color:transparent;font-weight:700;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Proposal:</span></div><br><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;font-family:Arial;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">We propose a new tool that we have tentatively named 'mir-canon' that performs canonical transformations on MIR. The goal is for MIR pre-processed with mir-canon to show fewer differences than if it were not pre-processed.</span></div><br><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;font-family:Arial;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">At the time of this writing we have a prototype canonicalization tool. We’ve come up with some techniques that show promise and would like to open discussion with the community to get feedback and suggestions on refining said techniques. Currently we think of this as an open ended project.</span></div><br><br><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;font-family:Arial;background-color:transparent;font-weight:700;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Techniques:</span></div><br><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;font-family:Arial;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Our prototype does the following for each basic block in a Reverse Post Ordering:</span></div><br><ul style="margin-top:0pt;margin-bottom:0pt"><li dir="ltr" style="list-style-type:disc;font-size:11pt;font-family:Arial;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline"><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Canonical instruction ordering is done by moving a given instruction as close to the nearest use of its def as possible.</span></div></li></ul><br><ul style="margin-top:0pt;margin-bottom:0pt"><li dir="ltr" style="list-style-type:disc;font-size:11pt;font-family:Arial;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline"><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Next, canonical VReg renaming is done by building a collection of candidate instructions that can be thought of as sinks in the def-use graph: they are typically instructions that write to physical registers or store to memory. These candidates are used as the root of a breadth first walk over the vreg operand def-use graph that determines a canonical vreg ordering.</span></div></li></ul><br><ul style="margin-top:0pt;margin-bottom:0pt"><li dir="ltr" style="list-style-type:disc;font-size:11pt;font-family:Arial;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline"><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Using said canonical vreg ordering we rename monotonically, but before we do this we skip several vreg values in order to increase the chance that we land on the same vreg number for two different input MIR files. We also do this to reduce the chances that a difference in previously def-use walks will affect the vreg renaming for subsequent walks. This skipping step could be thought of as a kind of vreg number reckoning: we skip modulo n vregs so that we are likely to land on the same vreg for two different files.</span></div></li></ul><br><br><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;font-family:Arial;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">This approach is completely agnostic of ISA specific semantics so it should work for any target.</span></div><br><br><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;font-family:Arial;background-color:transparent;font-weight:700;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Current status:</span></div><br><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;font-family:Arial;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">At the moment we have a standalone llvm tool that uses a single pass to do the above described transformations. We have test inputs that show promise but we still need a wider set of tests as well as hard metrics.</span></div><br><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;font-family:Arial;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">Our approach processes a single file at a time. The primary benefit to this approach is lower complexity in initial implementation and exploration of building such a tool. We are open to other approaches such as an llvm-diff like (two file at a time) approach, but we have not explored that avenue fully yet.</span></div><br><div style="line-height:1.38;margin-top:0pt;margin-bottom:0pt"><span style="font-size:11pt;font-family:Arial;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap">We’re eager to hear community feedback and will be ready to share patches for review soon.</span></div><div><span style="font-size:11pt;font-family:Arial;background-color:transparent;font-variant-ligatures:normal;font-variant-east-asian:normal;vertical-align:baseline;white-space:pre-wrap"><br></span></div></span></div></div></div><span class="">
______________________________<wbr>_________________<br>LLVM Developers mailing list<br><a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a><br><a href="http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" target="_blank">http://lists.llvm.org/cgi-bin/<wbr>mailman/listinfo/llvm-dev</a><br></span></div></blockquote></div><br></div></div></blockquote></div><br></div>