<div dir="ltr">+ a few others.<div><br></div><div>After following this rabbit hole a bit, there are a lot of mutually recursive calls, etc, that may or may not do the right thing with vectors of pointers.</div><div>I can fix *this* particular bug with the attached patch.</div><div><br></div><div>However, it's mostly papering over stuff. Nothing seems to know what to do with a memorylocation that is a vector of pointers. They all expect memorylocation to be a single pointer location.</div><div><br>I would chalk it up to "luck" that this patch fixes the bug.</div><div><br></div><div>It's pretty clear that MemoryLocation doesn't fit the needs of a lot of stuff anymore (we hacked AA nodes into it, and lots of stuff now tries to figure out the invariantess of the locations, blah blah blah), but it seems like a big job to figure out what to replace it with that will work for these cases.</div><div><br></div><div>(I'm pretty positive if we just make it MemoryLocations, and have everything loop over the locations, the compiler will get a lot larger and a lot slower)</div></div><div class="gmail_extra"><br><div class="gmail_quote">On Mon, Aug 29, 2016 at 9:10 AM, Daniel Berlin <span dir="ltr"><<a href="mailto:dberlin@dberlin.org" target="_blank">dberlin@dberlin.org</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr">it would also, for that matter, say the same about an array of pointers.<div><br></div><div>It's not clear to me what will break if you change this to isPtrOrPtrVectorTy.</div><div><br></div><div>In fact, i know it doesn't fix this bug.</div><div>It's a pretty deep rabbit hole of things not quite prepared to understand vectors of pointers.</div><div><br></div><div>(we prepare memorylocations of them, but memory locations expect to be one thing, not a group of things, etc).</div><div><br></div><div><br></div><div><br></div></div><div class="HOEnZb"><div class="h5"><div class="gmail_extra"><br><div class="gmail_quote">On Mon, Aug 29, 2016 at 8:58 AM, Daniel Berlin <span dir="ltr"><<a href="mailto:dberlin@dberlin.org" target="_blank">dberlin@dberlin.org</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div dir="ltr">this is definitely a bug in AA.<div><br></div><div><div> 225 <span style="white-space:pre-wrap"> </span> for (auto I = CS2.arg_begin(), E = CS2.arg_end(); I != E; ++I) {</div><div> 226 <span style="white-space:pre-wrap"> </span> const Value *Arg = *I;</div><div> 227 <span style="white-space:pre-wrap"> </span> if (!Arg->getType()->isPointerTy(<wbr>))</div><div>-> 228 <span style="white-space:pre-wrap"> </span> continue;</div><div> 229 <span style="white-space:pre-wrap"> </span> unsigned CS2ArgIdx = std::distance(CS2.arg_begin(), I);</div><div> 230 <span style="white-space:pre-wrap"> </span> auto CS2ArgLoc = MemoryLocation::getForArgument<wbr>(CS2, CS2ArgIdx, TLI);</div></div><div><br></div><div> AliasAnalysis.cpp:228<br></div><div><br></div><div>It ignores every argument because they are vectors of pointers, not pointers.</div><div><br></div><div><br></div><div>I'm surprised this has not broken anything before. It will never say two intrinsics with vectors of pointers mod/ref each other.</div><div><br></div></div><div><div><div class="gmail_extra"><br><div class="gmail_quote">On Mon, Aug 29, 2016 at 7:36 AM, Davide Italiano <span dir="ltr"><<a href="mailto:davide@freebsd.org" target="_blank">davide@freebsd.org</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">+ Daniel Berlin<br>
<div><div><br>
On Mon, Aug 29, 2016 at 3:42 PM, Chris Sakalis via llvm-dev<br>
<<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>> wrote:<br>
> Hello everyone,<br>
><br>
> I think I have found an gvn / alias analysis related bug, but before opening<br>
> an issue on the tracker I wanted to see if I am missing something. I have<br>
> the following testcase:<br>
><br>
>> define spir_kernel void @test(<2 x i32*> %in1, <2 x i32*> %in2, i32* %out)<br>
>> {<br>
>> entry:<br>
>> ; Just some temporary storage<br>
>> %tmp.0 = alloca i32<br>
>> %tmp.1 = alloca i32<br>
>> %tmp.i = insertelement <2 x i32*> undef, i32* %tmp.0, i32 0<br>
>> %tmp = insertelement <2 x i32*> %tmp.i, i32* %tmp.1, i32 1<br>
>> ; Read from in1 and in2<br>
>> %in1.v = call <2 x i32> @llvm.masked.gather.v2i32(<2 x i32*> %in1, i32<br>
>> 1, <2 x i1> <i1 true, i1 true>, <2 x i32> undef) #1<br>
>> %in2.v = call <2 x i32> @llvm.masked.gather.v2i32(<2 x i32*> %in2, i32<br>
>> 1, <2 x i1> <i1 true, i1 true>, <2 x i32> undef) #1<br>
>> ; Store in1 to the allocas<br>
>> call void @llvm.masked.scatter.v2i32(<2 x i32> %in1.v, <2 x i32*> %tmp,<br>
>> i32 1, <2 x i1> <i1 true, i1 true>);<br>
>> ; Read in1 from the allocas<br>
>> %tmp.v.0 = call <2 x i32> @llvm.masked.gather.v2i32(<2 x i32*> %tmp, i32<br>
>> 1, <2 x i1> <i1 true, i1 true>, <2 x i32> undef) #1<br>
>> ; Store in2 to the allocas<br>
>> call void @llvm.masked.scatter.v2i32(<2 x i32> %in2.v, <2 x i32*> %tmp,<br>
>> i32 1, <2 x i1> <i1 true, i1 true>);<br>
>> ; Read in2 from the allocas<br>
>> %tmp.v.1 = call <2 x i32> @llvm.masked.gather.v2i32(<2 x i32*> %tmp, i32<br>
>> 1, <2 x i1> <i1 true, i1 true>, <2 x i32> undef) #1<br>
>> ; Store in2 to out for good measure<br>
>> %tmp.v.1.0 = extractelement <2 x i32> %tmp.v.1, i32 0<br>
>> %tmp.v.1.1 = extractelement <2 x i32> %tmp.v.1, i32 1<br>
>> store i32 %tmp.v.1.0, i32* %out<br>
>> %out.1 = getelementptr i32, i32* %out, i32 1<br>
>> store i32 %tmp.v.1.1, i32* %out.1<br>
>> ret void<br>
>> }<br>
><br>
><br>
> It uses a masked scatter operation to store a value to the two allocas and<br>
> then uses a masked gather operation to read that same value. This is done<br>
> twice in a row, with two different values. If I run this code through the<br>
> GVN pass, the second gather (%tmp.v.1) will be deemed to be the same as the<br>
> first gather (%tmp.v.0) and it will be removed. After some debugging, I<br>
> realized that this is happening because the Memory Dependence Analysis<br>
> returns %tmp.v.0 as the Def dependency for %tmp.v.1, even though the scatter<br>
> call in between changes the value stored at %tmp. This, in turn, is<br>
> happening because the alias analysis is returning NoModRef for the %tmp.v.1<br>
> and second scatter callsites. The resulting IR produces the wrong result:<br>
><br>
>> define spir_kernel void @test(<2 x i32*> %in1, <2 x i32*> %in2, i32* %out)<br>
>> {<br>
>> entry:<br>
>> %tmp.0 = alloca i32<br>
>> %tmp.1 = alloca i32<br>
>> %tmp.i = insertelement <2 x i32*> undef, i32* %tmp.0, i32 0<br>
>> %tmp = insertelement <2 x i32*> %tmp.i, i32* %tmp.1, i32 1<br>
>> %in1.v = call <2 x i32> @llvm.masked.gather.v2i32(<2 x i32*> %in1, i32<br>
>> 1, <2 x i1> <i1 true, i1 true>, <2 x i32> undef) #1<br>
>> %in2.v = call <2 x i32> @llvm.masked.gather.v2i32(<2 x i32*> %in2, i32<br>
>> 1, <2 x i1> <i1 true, i1 true>, <2 x i32> undef) #1<br>
>> call void @llvm.masked.scatter.v2i32(<2 x i32> %in1.v, <2 x i32*> %tmp,<br>
>> i32 1, <2 x i1> <i1 true, i1 true>)<br>
>> %tmp.v.0 = call <2 x i32> @llvm.masked.gather.v2i32(<2 x i32*> %tmp, i32<br>
>> 1, <2 x i1> <i1 true, i1 true>, <2 x i32> undef) #1<br>
>> call void @llvm.masked.scatter.v2i32(<2 x i32> %in2.v, <2 x i32*> %tmp,<br>
>> i32 1, <2 x i1> <i1 true, i1 true>)<br>
>> ; The call to masked.gather is gone and now we are using the old value<br>
>> (%tmp.v.0)<br>
>> %tmp.v.1.0 = extractelement <2 x i32> %tmp.v.0, i32 0<br>
>> %tmp.v.1.1 = extractelement <2 x i32> %tmp.v.0, i32 1<br>
>> store i32 %tmp.v.1.0, i32* %out<br>
>> %out.1 = getelementptr i32, i32* %out, i32 1<br>
>> store i32 %tmp.v.1.1, i32* %out.1<br>
>> ret void<br>
>> }<br>
><br>
><br>
> The old value read from %tmp is used, instead of the new one. I tested this<br>
> code using `opt -gvn`, with LLVM 3.8.1. I also tried tip (84cb7f4) with the<br>
> same result.<br>
><br>
> I should mention that if I replace the second scatter with stores, the issue<br>
> persists. The only way to avoid it is to replace all scatter/gather<br>
> intrinsics with equivalent sets of store/load, in which case the MemDep<br>
> returns the correct dependencies and the GVN pass will not remove the second<br>
> set of loads.<br>
><br>
> So, my question is, is this a bug or am I doing something that I shouldn't<br>
> be in the IR? And if it is a bug, is it the AA analyses that return the<br>
> wrong result (I presume so) or should GVN handle such cases differently?<br>
><br>
> Best regards,<br>
><br>
> Chris<br>
><br>
</div></div>> ______________________________<wbr>_________________<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/<wbr>mailman/listinfo/llvm-dev</a><br>
><br>
<span><font color="#888888"><br>
<br>
<br>
--<br>
Davide<br>
<br>
"There are no solved problems; there are only problems that are more<br>
or less solved" -- Henri Poincare<br>
</font></span></blockquote></div><br></div>
</div></div></blockquote></div><br></div>
</div></div></blockquote></div><br></div>