<html xmlns:v="urn:schemas-microsoft-com:vml" xmlns:o="urn:schemas-microsoft-com:office:office" xmlns:w="urn:schemas-microsoft-com:office:word" xmlns:m="http://schemas.microsoft.com/office/2004/12/omml" xmlns="http://www.w3.org/TR/REC-html40">
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
<meta name="Generator" content="Microsoft Word 15 (filtered medium)">
<style><!--
/* Font Definitions */
@font-face
{font-family:"Cambria Math";
panose-1:2 4 5 3 5 4 6 3 2 4;}
@font-face
{font-family:Calibri;
panose-1:2 15 5 2 2 2 4 3 2 4;}
@font-face
{font-family:"Ericsson Hilda";
panose-1:0 0 5 0 0 0 0 0 0 0;}
@font-face
{font-family:Consolas;
panose-1:2 11 6 9 2 2 4 3 2 4;}
@font-face
{font-family:Menlo;
panose-1:0 0 0 0 0 0 0 0 0 0;}
/* Style Definitions */
p.MsoNormal, li.MsoNormal, div.MsoNormal
{margin:0cm;
margin-bottom:.0001pt;
font-size:11.0pt;
font-family:"Calibri",sans-serif;}
a:link, span.MsoHyperlink
{mso-style-priority:99;
color:blue;
text-decoration:underline;}
pre
{mso-style-priority:99;
mso-style-link:"HTML Preformatted Char";
margin:0cm;
margin-bottom:.0001pt;
font-size:10.0pt;
font-family:"Courier New";}
p.MsoListParagraph, li.MsoListParagraph, div.MsoListParagraph
{mso-style-priority:34;
margin-top:0cm;
margin-right:0cm;
margin-bottom:0cm;
margin-left:36.0pt;
margin-bottom:.0001pt;
font-size:11.0pt;
font-family:"Calibri",sans-serif;}
span.HTMLPreformattedChar
{mso-style-name:"HTML Preformatted Char";
mso-style-priority:99;
mso-style-link:"HTML Preformatted";
font-family:Consolas;}
span.EmailStyle22
{mso-style-type:personal-compose;
font-family:"Ericsson Hilda",serif;
color:windowtext;}
.MsoChpDefault
{mso-style-type:export-only;
font-size:10.0pt;}
@page WordSection1
{size:612.0pt 792.0pt;
margin:72.0pt 72.0pt 72.0pt 72.0pt;}
div.WordSection1
{page:WordSection1;}
/* List Definitions */
@list l0
{mso-list-id:1771461442;
mso-list-type:hybrid;
mso-list-template-ids:-2108259978 69009423 69009433 69009435 69009423 69009433 69009435 69009423 69009433 69009435;}
@list l0:level1
{mso-level-tab-stop:none;
mso-level-number-position:left;
text-indent:-18.0pt;}
@list l0:level2
{mso-level-number-format:alpha-lower;
mso-level-tab-stop:none;
mso-level-number-position:left;
text-indent:-18.0pt;}
@list l0:level3
{mso-level-number-format:roman-lower;
mso-level-tab-stop:none;
mso-level-number-position:right;
text-indent:-9.0pt;}
@list l0:level4
{mso-level-tab-stop:none;
mso-level-number-position:left;
text-indent:-18.0pt;}
@list l0:level5
{mso-level-number-format:alpha-lower;
mso-level-tab-stop:none;
mso-level-number-position:left;
text-indent:-18.0pt;}
@list l0:level6
{mso-level-number-format:roman-lower;
mso-level-tab-stop:none;
mso-level-number-position:right;
text-indent:-9.0pt;}
@list l0:level7
{mso-level-tab-stop:none;
mso-level-number-position:left;
text-indent:-18.0pt;}
@list l0:level8
{mso-level-number-format:alpha-lower;
mso-level-tab-stop:none;
mso-level-number-position:left;
text-indent:-18.0pt;}
@list l0:level9
{mso-level-number-format:roman-lower;
mso-level-tab-stop:none;
mso-level-number-position:right;
text-indent:-9.0pt;}
ol
{margin-bottom:0cm;}
ul
{margin-bottom:0cm;}
--></style><!--[if gte mso 9]><xml>
<o:shapedefaults v:ext="edit" spidmax="1026" />
</xml><![endif]--><!--[if gte mso 9]><xml>
<o:shapelayout v:ext="edit">
<o:idmap v:ext="edit" data="1" />
</o:shapelayout></xml><![endif]-->
</head>
<body lang="SV" link="blue" vlink="purple">
<div class="WordSection1">
<p class="MsoNormal"><span style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US">Hi all,<o:p></o:p></span></p>
<p class="MsoNormal"><span style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US" style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US">I was pondering on how regbank-select could find the optimal assignment, and came up with the following algorithm:<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US" style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US"><o:p> </o:p></span></p>
<ol style="margin-top:0cm" start="1" type="1">
<li class="MsoListParagraph" style="margin-left:0cm;mso-list:l0 level1 lfo1"><span lang="EN-US" style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US">Traverse top-down and compute cost of putting the given result in a particular register bank.
The cost equals the cost of the InstructionMapping itself plus the cost of putting each operand in the register bank as required by the operand’s ValueMapping. Hence the cost definition is recursive and increases monotonically. If some operand does not have
an InstructionMapping that puts it in the required register bank, add such an option and compute the cost as before plus the cost of converting the register bank. Hence such options would only be computed on demand.
<o:p></o:p></span></li><li class="MsoListParagraph" style="margin-left:0cm;mso-list:l0 level1 lfo1"><span lang="EN-US" style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US">Traverse bottom-up and mark the InstructionMapping that puts the result in the appropriate
register bank at lowest cost.<o:p></o:p></span></li><li class="MsoListParagraph" style="margin-left:0cm;mso-list:l0 level1 lfo1"><span lang="EN-US" style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US">Traverse top-down and select the register banks, also emitting repair code if needed.<o:p></o:p></span></li></ol>
<p class="MsoNormal"><span lang="EN-US" style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US" style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US">I started to implement this to see whether it would work, but I encountered problems with AMDGPU in the sense that its RegisterBankInfo seems to assume
that the register banks are assigned progressively, meaning that when regbankselect queries AMDGPU’s RegisterBankInfo for the possible InstructionMappings for a given MachineInstr, then it only returns the smallest set necessary instead of all possible combinations.
Meaning, for a MachineInstr where its input operands have not been assigned, it does not return the whole set of interesting InstructionMappings which are needed for the algorithm above.<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US" style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US" style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US">I might finish the implementation just for fun and learning, but I would appreciate if anyone had some comments on whether this is a viable approach
or will lead to naught.<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US" style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US"><o:p> </o:p></span></p>
<p class="MsoNormal"><span lang="EN-US" style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US">Cheers,<o:p></o:p></span></p>
<p class="MsoNormal"><span lang="EN-US" style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US">Gabriel<sup><o:p></o:p></sup></span></p>
<p class="MsoNormal"><span lang="EN-US" style="font-family:"Ericsson Hilda",serif;mso-fareast-language:EN-US"><o:p> </o:p></span></p>
<div>
<div style="border:none;border-top:solid #E1E1E1 1.0pt;padding:3.0pt 0cm 0cm 0cm">
<p class="MsoNormal"><b><span lang="EN-US">From:</span></b><span lang="EN-US"> llvm-dev <llvm-dev-bounces@lists.llvm.org>
<b>On Behalf Of </b>Quentin Colombet via llvm-dev<br>
<b>Sent:</b> den 9 oktober 2020 19:16<br>
<b>To:</b> Dominik Montada <dominik.montada@hightec-rt.com><br>
<b>Cc:</b> llvm-de >> LLVM Developers' List <llvm-dev@lists.llvm.org><br>
<b>Subject:</b> Re: [llvm-dev] GlobalISel round table follow up: register bank select<o:p></o:p></span></p>
</div>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<p class="MsoNormal">On Oct 9, 2020, at 12:07 AM, Dominik Montada <<a href="mailto:dominik.montada@hightec-rt.com">dominik.montada@hightec-rt.com</a>> wrote:<o:p></o:p></p>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<div>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">Hi Quentin,<o:p></o:p></p>
<div>
<p class="MsoNormal">Am 08.10.20 um 21:17 schrieb Quentin Colombet:<o:p></o:p></p>
</div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<p class="MsoNormal">Hi Dominik,<o:p></o:p></p>
<div>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<p class="MsoNormal">On Oct 8, 2020, at 5:03 AM, Dominik Montada <<a href="mailto:dominik.montada@hightec-rt.com">dominik.montada@hightec-rt.com</a>> wrote:<o:p></o:p></p>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<div>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">Hi Quentin,<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">thanks for picking up the conversation!<o:p></o:p></p>
<div>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">> I think we should step back and check what we want before investing any time in some rewrite.<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">That is a very fair point and I might have been getting ahead of myself in my last email.<br>
What I would like to see from RegBankSelect is to produce the mapping with the overall lowest cost. Keeping track of all different combinations of mappings will certainly be non-trivial however, so I wonder if there is a smart way to do this without spending
too much compilation time. <o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">Ideally for instructions with no operands (like G_CONSTANT) it could also check whether a cross-bank copy is actually worth it or if it would be more beneficial to simply rematerialize
the instruction on the required bank. For such instructions this information should already be available as part of the cost-modelling in RegBankSelect: we could simply compare the cost of a mapping on the required bank vs. the cost of a cross-bank copy.<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">Would you see this as a valid direction for RegBankSelect?<o:p></o:p></p>
</div>
</div>
</div>
</blockquote>
<p class="MsoNormal">Yes, I think this is a valid direction. I actually think we shouldn’t restrict ourselves to instructions with no operands. We could always duplicate next to the original instruction, i.e., we wouldn’t have any issue materializing the arguments.<o:p></o:p></p>
</div>
</blockquote>
<p class="MsoNormal">Sure, I guess that would also work as long we don't have to introduce copies for the operands. Theoretically this could even try to duplicate the operands if overall it's still cheaper than a cross-bank copy but that is probably some pandoras
box just waiting to be opened :) <o:p></o:p></p>
</div>
</div>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Hehe!<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">Ideally we would compute the costs globally and do the code transformation only once. But I am guessing this is wishful thinking.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<p class="MsoNormal">The one thing that needs work is the cost model.<o:p></o:p></p>
</div>
</blockquote>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">What would you say does the cost model need in its current state?<o:p></o:p></p>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
</div>
</div>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<p class="MsoNormal">I don’t remember how everything works so take it with a grain of salt.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">I’d say we would need to take into account how much it would cost to duplicate the definitions and compute the repairing cost with those duplications.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">The thing is it may become compile time intensive pretty quickly if we want to consider all the possibilities.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">As a starter we could focus on the best mapping for the current instruction and try to match the desired regbanks from each operand and only evaluate that cost against just plain repairing.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">In any case, I think it would require a non-trivial amount of work.<o:p></o:p></p>
</div>
<div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">By the way, does this have anything to do with what Matt was talking about during the round table? Unfortunately I don't remember exactly what his issues with RegBankSelect were
but I'd be interested to know whether what we talked about would also benefit him.<o:p></o:p></p>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
</div>
</div>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<p class="MsoNormal">Partially. IIRC Matt’s problem was also how we apply the mapping and how, right now, we have to abuse the observer mechanism to do what we want.<br>
<br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">I do remember that AMDGPU is apparently doing something very different compared to AArch64 for example and I'd be interested to know the reasoning behind the different approaches.<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">Cheers,<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">Dominik<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Cheers,<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">-Quentin<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<div>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">Best regards,<o:p></o:p></p>
<p class="MsoNormal" style="mso-margin-top-alt:auto;mso-margin-bottom-alt:auto">Dominik<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Am 07.10.20 um 19:47 schrieb Quentin Colombet:<o:p></o:p></p>
</div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<p class="MsoNormal">Hi Dominik, <o:p></o:p></p>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Thanks for sending this!<o:p></o:p></p>
<div>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<p class="MsoNormal">On Oct 7, 2020, at 5:21 AM, Dominik Montada via llvm-dev <<a href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a>> wrote:<o:p></o:p></p>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
<div>
<div>
<p class="MsoNormal">Hi all,<br>
<br>
this is the second email for the round table follow-up, this time regarding the issues around the greedy RegBankSelect and alternative mappings.<br>
<br>
The issue I brought up was that because RegBankSelect goes top-down, it never looks at all available mappings for the operands when considering which of the mappings to apply to the current instruction. In our architecture we have one register bank dedicated
to pointers and another one for anything else. We often see code where we have a G_PTR_ADD with a constant. Since the constant is not a pointer, we put it on the other register bank. We could put it on the address regbank and do provide alternative mappings
for that, but since the greedy algorithm doesn't actually check the usage of the constant, it is always put on the other bank.<o:p></o:p></p>
</div>
</div>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">The intent behind the greedy algorithm was for the mapping to look at X instructions ahead, depending on the optimization level, when assigning one instruction.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">Right now the window is simply 1 instruction and the code is not structured to allow to have more than one.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">As we gain more insights on what we want RegBankSelect to do, it makes sense to redesign it. <o:p></o:p></p>
</div>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal"><br>
When RegBankSelect then sees the G_PTR_ADD and sees that one of its inputs is on the other register bank already, it then inserts a costly cross-bank copy instead of checking if that operand has any alternative mappings which would make the overall mapping
for the current instruction cheaper.<o:p></o:p></p>
</div>
</div>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">The idea is when a mapping is done, that means it was the best one at the time of the decision (greedy), thus we don’t challenge that.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">“Reverting” an already set mapping is not that simple because yes for this particular use we are inserting costly cross copies, but there are no guarantees that replacing the mapping of the definition will not insert even more costly copies
for the other uses.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">E.g., consider:<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">```<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">A = def <— RBS starts here</span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">= useFP A</span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">= useInt A</span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">= useInt A</span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">```<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Let’s assume that greedy works the way it is intended. I.e., it assigns A to the int register bank because there are 2 such uses vs. only 1 fp bank use:<o:p></o:p></p>
</div>
<div>
<div>
<p class="MsoNormal">```<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">A<int> = def</span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">= useFP A<int> <— Next, RBS looks at this one</span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">= useInt A<int></span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">= useInt A<int></span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">```<o:p></o:p></p>
</div>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<p class="MsoNormal">Now the regbank for useFP is not right and has to be repaired. Right now, we will insert the costly cross copy for that use:<o:p></o:p></p>
</div>
<div>
<div>
<div>
<p class="MsoNormal">```<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">A<int> = def</span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">AFP<fp> = cross_copy A<int></span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">= useFP AFP<fp></span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">= useInt A<int></span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">= useInt A<int></span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">```<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Now, if we were to change the definition of A to avoid this copy we would create two costly copy for the useInt. Actually, another question is what would we do when we look at the first useInt? Is this use allowed to change the definition
of the instruction again? Do we duplicate the definition? Etc..<o:p></o:p></p>
</div>
<div>
<div>
<p class="MsoNormal">```<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">A<fp> = def <— reassign</span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">= useFP AFP<fp></span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">AInt1<int> = cross_copy A<fp></span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">= useInt AInt1<int></span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">AInt2<int> = cross_copy A<fp></span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><span style="font-family:"Menlo",serif">= useInt AInt2<int></span><o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">```<o:p></o:p></p>
</div>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Bottom line, if we allow to modify the assignments of the definition, the decision making is not local anymore and in particular may require to add repairing code all over the place. As a result the cost model becomes much more complicated.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
</div>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal"><br>
Matt suggested that RegBankSelect should probably go bottom-up instead and I agree with him. I don't think there is a particular reason why RegBankSelect necessarily has to go top-down.<o:p></o:p></p>
</div>
</div>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">The rationale for going top-down is that when you reach an instruction, all the operands are assigned so you know what would be the cost of repairing.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">You could said that the problem is the same for definitions when going bottom-up. However, this is likely to be more problematic because usually you have fewer definitions than arguments on each individual instruction, therefore there is
more guess work going on (e.g., top-down, you assume a cost for 1 definition, going bottom-up you have to assume a cost for 2 arguments or more precisely you would have to track a window of X instructions for 2 arguments instead of 1 definition.)<o:p></o:p></p>
</div>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal"><br>
I'm not too familiar with the implementation of RegBankSelect. Would it be a big effort to make it work bottom-up instead? I'm guessing one of the biggest areas would be the check whether a cross-bank copy is needed as well as calculating the overall cost for
alternative mappings as now all usages of the current instruction would have to be checked instead of the much more limited operands. How big of an impact would this have?<o:p></o:p></p>
</div>
</div>
</blockquote>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">It’s been a while since I looked at the implementation but I would expect this to be significant.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">I think we should step back and check what we want before investing any time in some rewrite. For instance, I don’t see what bottom-up fundamentally gives us. It seems like a workaround to me.<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">That said, it would work either way!<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
<div>
<p class="MsoNormal">Cheers,<o:p></o:p></p>
</div>
<div>
<p class="MsoNormal">-Quentin<o:p></o:p></p>
</div>
<p class="MsoNormal"><br>
<br>
<o:p></o:p></p>
<blockquote style="margin-top:5.0pt;margin-bottom:5.0pt">
<div>
<div>
<p class="MsoNormal"><br>
Cheers,<br>
<br>
Dominik<br>
<br>
_______________________________________________<br>
LLVM Developers mailing list<br>
<a href="mailto:llvm-dev@lists.llvm.org">llvm-dev@lists.llvm.org</a><br>
<a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><o:p></o:p></p>
</div>
</div>
</blockquote>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
</blockquote>
<pre>-- <o:p></o:p></pre>
<pre>----------------------------------------------------------------------<o:p></o:p></pre>
<pre>Dominik Montada Email: <a href="mailto:dominik.montada@hightec-rt.com">dominik.montada@hightec-rt.com</a><o:p></o:p></pre>
<pre>HighTec EDV-Systeme GmbH Phone: +49 681 92613 19<o:p></o:p></pre>
<pre>Europaallee 19 Fax: +49-681-92613-26<o:p></o:p></pre>
<pre>D-66113 Saarbrücken WWW: <a href="https://protect2.fireeye.com/v1/url?k=10181272-4eb8bcbf-101852e9-86073b36ea28-3b408a428d93631b&q=1&e=a2c4ed0a-0e5c-4bef-9460-752197daba21&u=http%3A%2F%2Fwww.hightec-rt.com%2F">http://www.hightec-rt.com</a><o:p></o:p></pre>
<pre><o:p> </o:p></pre>
<pre>Managing Director: Vera Strothmann<o:p></o:p></pre>
<pre>Register Court: Saarbrücken, HRB 10445, VAT ID: DE 138344222<o:p></o:p></pre>
<pre><o:p> </o:p></pre>
<pre>This e-mail may contain confidential and/or privileged information. If<o:p></o:p></pre>
<pre>you are not the intended recipient please notify the sender immediately<o:p></o:p></pre>
<pre>and destroy this e-mail. Any unauthorised copying, disclosure or<o:p></o:p></pre>
<pre>distribution of the material in this e-mail is strictly forbidden.<o:p></o:p></pre>
<pre>--- <o:p></o:p></pre>
</div>
</div>
</blockquote>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
</blockquote>
<pre>-- <o:p></o:p></pre>
<pre>----------------------------------------------------------------------<o:p></o:p></pre>
<pre>Dominik Montada Email: <a href="mailto:dominik.montada@hightec-rt.com">dominik.montada@hightec-rt.com</a><o:p></o:p></pre>
<pre>HighTec EDV-Systeme GmbH Phone: +49 681 92613 19<o:p></o:p></pre>
<pre>Europaallee 19 Fax: +49-681-92613-26<o:p></o:p></pre>
<pre>D-66113 Saarbrücken WWW: <a href="https://protect2.fireeye.com/v1/url?k=09ae2d90-570e835d-09ae6d0b-86073b36ea28-251999af69af6c7f&q=1&e=a2c4ed0a-0e5c-4bef-9460-752197daba21&u=http%3A%2F%2Fwww.hightec-rt.com%2F">http://www.hightec-rt.com</a><o:p></o:p></pre>
<pre><o:p> </o:p></pre>
<pre>Managing Director: Vera Strothmann<o:p></o:p></pre>
<pre>Register Court: Saarbrücken, HRB 10445, VAT ID: DE 138344222<o:p></o:p></pre>
<pre><o:p> </o:p></pre>
<pre>This e-mail may contain confidential and/or privileged information. If<o:p></o:p></pre>
<pre>you are not the intended recipient please notify the sender immediately<o:p></o:p></pre>
<pre>and destroy this e-mail. Any unauthorised copying, disclosure or<o:p></o:p></pre>
<pre>distribution of the material in this e-mail is strictly forbidden.<o:p></o:p></pre>
<pre>--- <o:p></o:p></pre>
</div>
</div>
</blockquote>
</div>
<p class="MsoNormal"><o:p> </o:p></p>
</div>
</body>
</html>