<div dir="ltr"><div class="gmail_quote"><div dir="ltr">On Sat, Aug 1, 2015 at 6:39 AM Hal Finkel <<a href="mailto:hfinkel@anl.gov">hfinkel@anl.gov</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">----- Original Message -----<br>
> From: "Sanjoy Das" <<a href="mailto:sanjoy@playingwithpointers.com" target="_blank">sanjoy@playingwithpointers.com</a>><br>
> To: "Reid Kleckner" <<a href="mailto:rnk@google.com" target="_blank">rnk@google.com</a>><br>
> Cc: "Piotr Padlewski" <<a href="mailto:prazek@google.com" target="_blank">prazek@google.com</a>>, "<a href="mailto:cfe-dev@cs.uiuc.edu" target="_blank">cfe-dev@cs.uiuc.edu</a> Developers" <<a href="mailto:cfe-dev@cs.uiuc.edu" target="_blank">cfe-dev@cs.uiuc.edu</a>>, "LLVM Developers<br>
> Mailing List" <<a href="mailto:llvmdev@cs.uiuc.edu" target="_blank">llvmdev@cs.uiuc.edu</a>><br>
> Sent: Saturday, August 1, 2015 1:22:50 AM<br>
> Subject: Re: [LLVMdev] [cfe-dev] Clang devirtualization proposal<br>
><br>
> On Fri, Jul 31, 2015 at 6:18 PM, Reid Kleckner <<a href="mailto:rnk@google.com" target="_blank">rnk@google.com</a>><br>
> wrote:<br>
> > Consider this pseudo-IR and some possible transforms that I would<br>
> > expect to<br>
> > be semantics preserving:<br>
> ><br>
> > void f(i32* readonly %a, i32* %b) {<br>
> > llvm.assume(%a == %b)<br>
> > store i32 42, i32* %b<br>
> > }<br>
> > ...<br>
> > %p = alloca i32<br>
> > store i32 13, i32* %p<br>
> > call f(i32* readonly %p, i32* %p)<br>
> > %r = load i32, i32* %p<br>
> ><br>
> > ; Propagate llvm.assume info<br>
> > void f(i32* readonly %a, i32* %b) {<br>
> > store i32 42, i32* %a<br>
> > }<br>
> > ...<br>
> > %p = alloca i32<br>
> > store i32 13, i32* %p<br>
> > call f(i32* readonly %p, i32* %p)<br>
> > %r = load i32, i32* %p<br>
><br>
> I'd say this first transformation is incorrect. `readonly` is<br>
> effectively part of `%a`'s "type" as it constrains and affects the<br>
> operations you can do on `%a`. Even if `%b` is bitwise equivalent to<br>
> `%a` at runtime, it is "type incompatible" to replace `%a` with `%b`.<br>
><br>
> This is similar to how you cannot replace `store i32 42, i32<br>
> addrspace(1)* %a` with `store i32 42, i32 addrspace(2)* %b`, even if<br>
> you can prove `ptrtoint %a` == `ptrtoint %b` -- the nature of `store`<br>
> is dependent on the type of the pointer you store through.<br>
><br>
> The glitch in LLVM IR right now is that the `readonly`ness of `%a` is<br>
> not modeled in the type system, when I think it should be. An `i32<br>
> readonly*` should be a different type from `i32*`. In practice this<br>
> may be non-trivial to get right (for instance `phi`s and `selects`<br>
> will either have to do a type merge, or we'd have to have explicit<br>
> type operators at the IR level).<br>
<br>
We could do this, but then we'd need to promote these things to first-class parts of the type system (and I'd need to put further thought about how this interacts with dynamically-true properties at callsites and inlining).<br>
<br>
The alternative way of looking at it, which is true today, is that @llvm.assume is not removed even when its information is 'used'. It appears, given this example, that this is actually required for correctness, and that dead-argument elimination needs to specifically not ignore effectively-ephemeral values/arguments.<br></blockquote><div><br></div><div style="font-size:13.1999998092651px;line-height:19.7999992370605px">What follows is an off-the-cuff response. I'm still thinking through it, but wanted to let others do so as well.</div><div style="font-size:13.1999998092651px;line-height:19.7999992370605px"><br></div><div style="font-size:13.1999998092651px;line-height:19.7999992370605px">There is yet another interpretation that we might use: the final transformation Reid proposed is actually correct and allowed according to the IR.</div><div style="font-size:13.1999998092651px;line-height:19.7999992370605px"><br></div><div style="font-size:13.1999998092651px;line-height:19.7999992370605px">Specifically, we could make 'readonly' a property of the memory much like aliasing is. That would mean that the function promises not to modify the memory pointed to by %a in this example. The optimizer gets to ignore any such modifications while remaining correct.</div><div style="font-size:13.1999998092651px;line-height:19.7999992370605px"><br></div><div><span style="font-size:13.1999998092651px;line-height:19.7999992370605px">This would, in turn, mean that the store in Reid's "@f" function is an unobservable in the case that %a == %b through some dynamic property, whatever that may be. And as a consequence, the store-forwarding is a correct transformation.</span></div><div><br></div><div>-Chandler</div></div></div>