<div dir="ltr"><div class="gmail_extra"><div class="gmail_quote">On Mon, Jul 18, 2016 at 8:29 PM, David Majnemer <span dir="ltr"><<a href="mailto:david.majnemer@gmail.com" target="_blank">david.majnemer@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><br><div class="gmail_extra"><br><div class="gmail_quote"><span>On Mon, Jul 18, 2016 at 6:02 PM, Peter Collingbourne via llvm-dev <span dir="ltr"><<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr">Hi all,<div><br></div><div>I'd like to propose an IR extension that allows the inbounds keyword to be attached to indices in a getelementptr constantexpr.</div><div><div><div><br>By placing the inbounds keyword on an index, any pointer derived from the getelementptr outside of the bounds of the element referred to by that index, other than the pointer one past the end of the element, shall be treated as a poison value.<br></div></div></div></div></blockquote><div><br></div><div><br></div></span><div>I have read this sentence several times and I am still not quite sure what it means.</div><div>Can you please provide more examples of exactly what you are trying to represent?</div></div></div></div></blockquote><div><br></div><div>Thank you for asking me to provide another example because thinking about it revealed a flaw in my original specification. Here's the example, which involves nested arrays.</div><div><br></div><div>@x = internal global [2 x [2 x [2 x i32]]] [[2 x [2 x i32]] [[2 x i32] [i32 1, i32 2], [2 x i32] [i32 3, i32 4]], [2 x [2 x i32]] [[2 x i32] [i32 5, i32 6], [2 x i32] [i32 7, i32 8]]]</div><div><div>@x0 = constant i32* getelementptr inbounds ([2 x [2 x [2 x i32]]], [2 x [2 x [2 x i32]]]* @x, i32 0, inbounds i32 0, i32 0, i32 0)<br></div><div></div><div>@x1 = constant i32* getelementptr inbounds ([2 x [2 x [2 x i32]]], [2 x [2 x [2 x i32]]]* @x, i32 0, inbounds i32 1, i32 0, i32 0)<br></div><div></div><div><br></div></div><div>The intention is that the pointer that x0 is initialized to may only be used to access the subelement {{1,2},{3,4}} of x (that being the element selected by the second index), and the pointer that x1 is initialized to may only be used to access subelement {{5,6},{7,8}} of x.</div><div><br></div><div><div>Semantically, we would want this function's behavior to be undefined:</div><div><br></div><div>define void @foo() {</div><div> %p = load i32** @x0</div><div> %q = getelementptr i32*, i32** %p, i32 4</div><div> %r = load i32* %q ; UB here</div><div> ret void</div><div>}</div><div><div><br></div><div>while this one would be fine:</div><div><br></div><div>define void @foo() {</div><div> %p = load i32** @x0</div><div> %q = getelementptr i32*, i32** %p, i32 2</div><div> %r = load i32* %q ; loads 3</div><div> ret void</div><div>}</div></div></div><div><br></div><div>As a consequence, if there are no other references to x, the optimizer may split x into two globals:</div><div><br></div><div>@x.0 = internal global [2 x [2 x i32]] [[2 x i32] [i32 1, i32 2], [2 x i32] [i32 3, i32 4]]<br></div><div><div>@x.1 = internal global [2 x [2 x i32]] [[2 x i32] [i32 5, i32 6], [2 x i32] [i32 7, i32 8]]</div></div><div><br></div><div>replace x0's initializer with a reference to x.0 and replace x1's initializer with a reference to x1. This is possible because the program contains no pointer to x that may be adjusted to access the entire global.</div><div><br></div><div>However, it occurred to me that the semantics I previously gave are insufficient for splitting, because this instruction:</div><div><br></div><div><div> %q = getelementptr i32*, i32** %p, i32 4</div></div><div><br></div><div>would satisfy the "one-past-the-end" rule, and as such would not be a poison value, and the program would be able to load from it to produce the value 5.</div><div><br></div><div>I think the solution is to specify index inbounds in terms of loads and stores, by replacing the sentence you referred to with the following:</div><div><br></div><div>"Loading from or storing to any pointer derived from a getelementptr has undefined behavior if any of the indices of the getelementptr are marked as inbounds and the load or store would access memory outside of the bounds of the element selected by the index marked as inbounds."</div><div><br></div><div>To conclude this example, let's look at strengthening or weakening the guarantee by placing the keyword on different indices. Suppose we introduce this constant:</div><div><br></div><div><div>@x00 = constant i32* getelementptr inbounds ([2 x [2 x [2 x i32]]], [2 x [2 x [2 x i32]]]* @x, i32 0, i32 0, inbounds i32 0, i32 0)<br></div></div><div><br></div><div>Now the pointer that x00 is initialized to refers to the subelement {1,2} of x (being the element referred to by the third index), although that doesn't really help us split the constant because the existence of the initializer in x0 requires that we keep {1,2} and {3,4} together. Similarly, if we introduce this constant:</div><div><br></div><div>@x_all = constant i32* getelementptr inbounds ([2 x [2 x [2 x i32]]], [2 x [2 x [2 x i32]]]* @x, inbounds i32 0, i32 0, i32 0, i32 0)<br></div><div><br></div><div>this reference to x does not make any guarantees about which subelements a derived pointer may access. As such, the elements {{1,2},{3,4}} and {{5,6},{7,8}} must be kept together as in the original global if this constant appears in the program.</div><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div><br></div><div>I know what it means for inbounds to be on the GEP but none of the indices: that's the GEP of today.</div><div>What does it mean for the GEP to be marked inbound while only some of the indices are inbounds? What if the GEP isn't marked inbounds?</div><div>What does it mean if all the GEP indices are marked inbounds but the GEP isn't marked inbounds?</div></div></div></div></blockquote><div><br></div><div>Interesting points. Having thought about this further, it seems that we do not need to allow multiple inbounds annotations on indices, as each inbounds annotation will be at least as restrictive as an earlier one. For example, in:</div><div><br></div><div>i32* getelementptr inbounds ([2 x [2 x [2 x i32]]], [2 x [2 x [2 x i32]]]* @x, i32 0, inbounds i32 0, inbounds i32 0, i32 0)<br></div><div><br></div><div>the two inbounds markers would respectively guarantee that the pointer will only access {{1,2},{3,4}} and {1,2}, and the latter implies the former.</div><div><br></div><div>Regarding the relationship with GEP inbounds: it does seem that the two properties are somewhat orthogonal, as GEP inbounds guarantees that while being calculated a pointer is within a dynamic address range (dependent on object size), while index inbounds guarantees that once calculated a pointer (and its derivatives) are within a static address range.</div><div><br></div><div>A possible semantic simplification could be to specify that index inbounds also implies the dynamic address range guarantee, but I don't think we should do that, because a program (or opt pass) may wish to adjust an index inbounds pointer outside of the static range, which would potentially invalidate GEP inbounds without invalidating index inbounds.</div><div><br></div><div>Given that the two properties are orthogonal, perhaps we should use a different keyword for index inbounds...</div><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div><br></div><div>GEPs are folded and optimized, two GEPs can compute the same numeric position with differing indices. What happens when we are giving out new indices for a GEP with one (or more) inbounds index?</div><div><br></div><div>I'm also a little confused when you talk about pointers derived from the GEP being outside of the bounds... The inbounds of present day GEP refers to the base pointer and offsets, it does not directly have semantics on pointers derived from such a GEP. This means it is perfectly OK to have an out of bounds GEP of an inbounds GEP. My reading of your extension says that this is not possible...</div></div></div></div></blockquote><div><br></div><div>Right, this is not possible with index inbounds in that the derived pointer cannot be loaded from or stored to.</div><div><br></div><div>Peter</div><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div class="gmail_extra"><div class="gmail_quote"><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-style:solid;border-left-color:rgb(204,204,204);padding-left:1ex"><span><div dir="ltr"><div><div><div></div></div><div><br></div><div>The main motivation is to allow the optimizer to split vtable groups along vtable boundaries thus reducing code size when certain compiler features [1] are enabled, in a way that avoids breaking ABI. The idea is that it would be safe to split only if the vtable has local linkage and each reference to the global has an inbounds keyword on the correct index. If both of these conditions are satisfied, the address of the entire global is known to not be taken, and therefore a split would not break semantics. The new attribute could also potentially be used with other features such as alias analysis.</div></div><div><div><br>This proposal arises from some concerns raised by Eli at <a href="http://reviews.llvm.org/D22295" target="_blank">http://reviews.llvm.org/D22295</a> regarding an earlier implementation of global splitting, which was based on metadata. I'm posting this new proposal as an RFC as the increased intrusiveness in the IR warrants greater visibility.</div></div><div><br></div><div>Example</div><div><br></div><div>i8** getelementptr inbounds ({[4 x i8*], [4 x i8*]}, {[4 x i8*], [4 x i8*]}* @_ZTVfoo, i32 0, inbounds i32 1, i32 2)</div><div><br></div><div>This is a reference to the address point of the second element of _ZTVfoo, a virtual table group with two virtual tables, such that any pointer extending beyond the bounds of the second virtual table is a poison value.</div><div><br></div><div>Alternatives</div><div><br></div><div>We could consider attaching metadata to globals as proposed and implemented in D22295, which would avoid extending the constant expression IR. However as pointed out by Eli this is problematic because optimization passes may plausibly rebuild globals in non-permitted ways, for example by deriving a pointer to one past the end of a partition (which may cause it to point to another global).</div><div><br></div><div>We could consider using intrinsics instead of constant expressions to represent references to partitions. However this is not enough at least for vtables because other globals (e.g. VTTs) may need to contain pointers to vtables.</div><div><br></div><div>Thanks,</div><span><font color="#888888"><div><div data-smartmail="gmail_signature"><div dir="ltr">-- <div>Peter</div><div><br></div><div>[1] Specifically, if either of the whole-program devirtualization or</div><div>control flow integrity features are being used. In the former case, under</div><div>virtual constant propagation we are able to place propagated constants</div><div>directly in front of virtual tables of classes with multiple bases. In the</div><div>latter case, we can arrange virtual tables with multiple bases in a more</div><div>hierarchical order, which reduces the required amount of runtime data and</div><div>simplifies the required checks.</div></div></div>
</div></font></span></div>
<br></span>_______________________________________________<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" rel="noreferrer" target="_blank">http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a><br>
<br></blockquote></div><br></div></div>
</blockquote></div><br><br clear="all"><div><br></div>-- <br><div data-smartmail="gmail_signature"><div dir="ltr">-- <div>Peter</div></div></div>
</div></div>