<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 id="divtagdefaultwrapper" dir="ltr" style="font-size:12pt; color:#000000; font-family:Calibri,Arial,Helvetica,sans-serif">
<p>LLVM makes several assumptions about address space 0. However,</p>
<p>alloca is presently constrained to always return this address space.</p>
<p>There's no real way to avoid using alloca, so without this</p>
<p>there is no way to opt out of these assumptions.</p>
<p><span style="font-size:12pt"><br>
</span></p>
<p><span style="font-size:12pt">The problematic assumptions include:</span></p>
<p></p>
<ul>
<li><span style="font-size:12pt">That the pointer size used for the stack is the same size as </span><span style="font-size:12pt">the code size pointer, which is also the maximum sized pointer.</span></li><li><span style="font-size:12pt">That 0 is an invalid, non-dereferencable pointer value.</span></li></ul>
<div></div>
<p></p>
<span style="font-size:12pt">These are problems for AMDGPU because alloca is used to </span><span style="font-size:12pt">implement the private address space, which uses a 32-bit </span><span style="font-size:12pt">index as the pointer value. Other pointers
are 64-bit </span><span style="font-size:12pt">and behave more like LLVM's notion of generic address </span><span style="font-size:12pt">space. By changing the address space used for allocas, </span><span style="font-size:12pt">we can change our generic pointer
type to be LLVM's generic </span><span style="font-size:12pt">pointer type which does have similar properties.</span>
<div><br>
</div>
<div>The proposal here is to add a -A field to the datalayout string which will specify the address space for allocas. IRBuilder::CreateAlloca and company gain a DataLayout argument, and some intrinsics that currently don't support address spaces need to support
them.</div>
<div><br>
</div>
<div>This has been implemented out of tree before before for CHERI.<br>
<span style="font-size:12pt"></span>
<div>
<div><br>
</div>
<p>This has also been proposed before but for different reasons: <a href="http://lists.llvm.org/pipermail/llvm-dev/2015-August/089706.html" class="OWAAutoLink" id="LPlnk407520" previewremoved="true">http://lists.llvm.org/pipermail/llvm-dev/2015-August/089706.html</a></p>
<p><br>
</p>
<p>My current proposal is more focused than the previous proposal. Instead of allowing specifying address spaces on individual allocas, this restricts it to one chosen address space specified in the datalayout. I think the work to support different address
spaces per-alloca is a strict superset of this proposal, so if people are interested in that I think that is a separate step beyond this.</p>
<p><br>
</p>
<p>-Matt</p>
<p><br>
</p>
<p><br>
</p>
</div>
</div>
</div>
</body>
</html>