<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 style="color:black;font-size:12pt;font-family:Arial,Helvetica,sans-serif;background-color:white"><span style="font-family:Calibri,Arial,Helvetica,sans-serif">Hi,</span></span></div>
<div style="font-family: Calibri, Arial, Helvetica, sans-serif; font-size: 12pt; color: rgb(0, 0, 0);">
<div style="color:black;font-size:12pt;font-family:Arial,Helvetica,sans-serif;background-color:white">
<div style="font-family:Calibri,Arial,Helvetica,sans-serif"><br>
</div>
<div style="font-family:Calibri,Arial,Helvetica,sans-serif">We've noticed some codegen regressions due to spurious peeling in the simple loop unrolling pass. This happens when the incoming values of the phi nodes causing the peeling are equivalent, since GVN
only runs after the loop peeling. Ironically, this prevents vectorization on our target due to the loss of alignment in the loop body caused by the peeling. The spurious peeling seems to affect AArch64 as well, so it's possible other targets may be affected.<br>
</div>
<div style="font-family:Calibri,Arial,Helvetica,sans-serif"><br>
</div>
<div style="font-family:Calibri,Arial,Helvetica,sans-serif">Attached is an example test case and the corresponding codegen as of 2632ba6a358a62c5cbaddc141de81b756b68698f (with and without 35b3989a30eefa66cd6edca4c6e1ec061c05ad96 reverted). Compiling with -O3
-S -target aarch64-linux-gnu gives the attached 3 logs. Prerotate is just before loop rotation. The postrotate logfile shows the IR after the rotate loop pass and a few more passes have run (the logs after the rotate loop pass only show the loop body). Loop
rotate introduces two elidable phis:</div>
<div style="font-family:Calibri,Arial,Helvetica,sans-serif"><br>
</div>
<div style="font-family:Calibri,Arial,Helvetica,sans-serif"> %33 = phi i8* [ %9, %22 ], [ %14, %32 ]<br>
</div>
<div style="font-family:Calibri,Arial,Helvetica,sans-serif"> %34 = phi %class.HomemadeVector.0* [ %8, %22 ], [ %13, %32 ]</div>
<div style="font-family:Calibri,Arial,Helvetica,sans-serif"><br>
</div>
<div style="font-family:Calibri,Arial,Helvetica,sans-serif">This leads to loop peeling inside simple loop unrolling to peel the first iteration of the loop. Prior to 35b3989a30eefa66cd6edca4c6e1ec061c05ad96 the phi would remain until GVN runs and would remove
them since %8 and %13 are the same values. The same would happen to %9 and %14.</div>
<div style="font-family:Calibri,Arial,Helvetica,sans-serif"><br>
</div>
<div style="font-family:Calibri,Arial,Helvetica,sans-serif">Running GVN earlier would fix this issue, but I suspect it would have other regressions. Does anyone know how to address this?</div>
<div style="font-family:Calibri,Arial,Helvetica,sans-serif"><br>
</div>
<div style="font-family:Calibri,Arial,Helvetica,sans-serif">Best regards,</div>
<div style="font-family:Calibri,Arial,Helvetica,sans-serif"><br>
</div>
<span style="font-family:Calibri,Arial,Helvetica,sans-serif">Thomas</span></div>
<br>
</div>
</body>
</html>