<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=iso-8859-1">
<style type="text/css" style="display:none;"> P {margin-top:0;margin-bottom:0;} </style>
</head>
<body dir="ltr">
<div style="font-family: Calibri, Arial, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
<span>I am looking for feedback to add support for a new loop pragma to Clang/LLVM.<br>
</span>
<div>With "#pragma tail_predicate" the idea would be to indicate that a loop<br>
</div>
<div>epilogue/tail can, or should be, folded into the main loop. I see two use<br>
</div>
<div>cases for this pragma.<br>
</div>
<div><br>
</div>
<div>First, this could be interesting for the vectorizer. It currently supports tail<br>
</div>
<div>folding by masking all loop instructions/blocks, but does this only when<br>
</div>
<div>optimising for size is enabled. This pragma could override the cost-model/opt-level.<br>
</div>
<div><br>
</div>
<div>Second use case would be the Armv8.1-M MVE vector extension, which supports<br>
</div>
<div>tail-predicated hardware loops. This version of hardware loops sets the vector<br>
</div>
<div>lanes to be masked, and is thus a nice optimisation that avoids generating a<br>
</div>
<div>tail loop when the number of elements processed is not a multiple of the vector<br>
</div>
<div>length.<br>
</div>
<div><br>
</div>
<div>For this use case, the tail predicate pragma could be good user experience<br>
</div>
<div>improvement, as it would for example allow this more compact form without<br>
</div>
<div>any predicated intrinsics:<br>
</div>
<div><br>
</div>
<div> #pragma tail_predicate<br>
</div>
<div> do {<br>
</div>
<div> VLD(..); // some vector load intrinsic<br>
</div>
<div> VST(..); // some vector store intrinsic<br>
</div>
<div> ..<br>
</div>
<div> } while (N);<br>
</div>
<div><br>
</div>
<div>which can then be transformed and predication made explicit through data<br>
</div>
<div>dependencies like so:<br>
</div>
<div><br>
</div>
<div> do {<br>
</div>
<div> mask = vctp(N); // intrinsic that generates the mask of active lanes<br>
</div>
<div> VLD(.., mask);<br>
</div>
<div> VST(.., mask);<br>
</div>
<div> ..<br>
</div>
<div> } while (N);<br>
</div>
<div><br>
</div>
<div>A vector loop in this form can easily be picked up the new hardware loop pass,<br>
</div>
<div>and the corresponding tail-predicated hardware loop can be generated. This is<br>
</div>
<div>only a small example, but we think for more complicated examples we think<br>
</div>
<div>the benefit could be substantial.<br>
</div>
<div><br>
</div>
<div>I have uploaded a patch for the initial Clang plumbing exercise here:<br>
</div>
<span><a href="https://reviews.llvm.org/D64744" id="LPlnk553447">https://reviews.llvm.org/D64744</a></span><br>
</div>
<div style="font-family: Calibri, Arial, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
<br>
</div>
<div style="font-family: Calibri, Arial, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
Cheers,</div>
<div style="font-family: Calibri, Arial, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
Sjoerd.</div>
</body>
</html>