<div dir="ltr"><div>> Is it the case that with the legacy PM there is no inlining at either -O2 or -O3 and with newPM there is? Or is there something else going on?<br></div><div>Legacy PM inlines at -O2/-O3, new PM inlines at -O1/-O2/-O3. These cases where inlining occurs also coincide with the test failure. I agree that inlining itself isn't the issue, but it seems to contribute to the stripping of lifetime intrinsics and thus the test failure.</div><div><br></div><div>ASan needs lifetime intrinsics to implement use-after-scope. It may just be luck that the legacy PM's -O1 happens to not strip lifetime intrinsics before ASan gets to them?<br></div><div><div><br></div><div>I tried moving the *San passes before inlining and now the tests that I was having trouble with pass, but new tests that specifically test inlining fail, so I think it's probably fairly important to run *San passes after inlining. But I tried moving them right after inlining (before SROA/something gets to optimize things out) and some TSan tests now fail. I found out TSan expects to be run after pretty much all other passes so that it can change any memset/memcpy intrinsics into just the normal function call to memset()/memcpy(). But if we move TSan to be run before InstCombine, then InstCombine will change calls to memset()/memcpy() back to the intrinsics and the tests don't like that. I think it's reasonable to treat all *San passes the same way, so I don't think putting ASan somewhere different from TSan makes sense. But that means there's no good place to put these sanitizers.</div></div><div><br></div><div>I feel like a lot of these sanitizer tests just happened to work at -O1 under the legacy PM. I think I'll try having the new PM not inline when any sanitizers are enabled and see if I can keep the house of cards standing. Or is it a bad idea to change the PM behavior based on the detection of sanitizers, since sanitizers won't be testing the same code you'd get without sanitizing (e.g. a lot of the inlining tests would lose a lot of their value)?</div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Wed, May 13, 2020 at 3:14 PM Johannes Doerfert <<a href="mailto:johannesdoerfert@gmail.com" target="_blank">johannesdoerfert@gmail.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">
<div>
<p><br>
</p>
<div>On 5/13/20 3:31 PM, David Blaikie via
llvm-dev wrote:<br>
</div>
<blockquote type="cite">
<pre>I believe it's meant to run after /some/ optimizations to make it a bit
more efficient, while not so optimized that it misses opportunities to
detect bugs - but I could be wrong there. I'll leave it up to other folks
to chime in.
</pre>
</blockquote>
<p>I think that is right. The more transformations you run the more
UB you can also "loose" as it is defined to something by the
transformation.<br>
</p>
<p>Lifetime markers are an example. Once removed, which is generally
legal in IR, you cannot argue accesses after the end are UB.<br>
</p>
<p><br>
</p>
<blockquote type="cite">
<pre>On Wed, May 13, 2020 at 1:04 PM Arthur Eubanks <a href="mailto:aeubanks@google.com" target="_blank"><aeubanks@google.com></a> wrote:
</pre>
<blockquote type="cite">
<pre>Just tested it out, that test does indeed fail under the old PM at -O3 and
even at -O2.
If the ASan pass runs after optimizations and is designed to detect
undefined behavior at runtime, I don't see how it can be super reliable at
higher optimization levels.
On Wed, May 13, 2020 at 12:39 PM David Blaikie <a href="mailto:dblaikie@gmail.com" target="_blank"><dblaikie@gmail.com></a> wrote:
</pre>
<blockquote type="cite">
<pre>+some sanitizer/new pass manager folks
On Wed, May 13, 2020 at 12:22 PM Arthur Eubanks via llvm-dev <
<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>> wrote:
</pre>
<blockquote type="cite">
<pre>Hi,
I've been trying to burn down the remaining sanitizer failures under the
new pass manager. Right now I'm stuck on some ASan tests.
Some ASan tests run under -O1. There are a couple differences between
the old and new pass managers under -O1, e.g. the old PM doesn't inline
whereas the new PM does. The differences seem to cause some lifetime
intrinsics to get stripped out (e.g. via SROA, InstCombine). It might be
due to ASan specifically testing undefined behavior, and different
optimizations run means different behavior. For a specific example,
use-after-scope-dtor-order.cpp runs under -O1 and fails under the new PM
because SROA strips out the lifetime intrinsics and by the time the ASan
pass runs it doesn't find the lifetime intrinsics to add its own
instrumentation.
</pre>
</blockquote>
<pre>That, to me, sounds like a real bug in the optimizations/asan
implementation if this choice fo optimizations makes the diagnosis go away.
(is ASan able to diagnose the problem at -O3 (where I guess SROA and other
things run) with the legacy pass manager?)
</pre>
<blockquote type="cite">
<pre>What's the proper way to resolve this? Run the tests under -O0? Change
the passes pipeline under the new PM when ASan (and maybe other sanitizers)
is detected?
_______________________________________________
LLVM Developers mailing list
<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>
<a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a>
</pre>
</blockquote>
<pre></pre>
</blockquote>
</blockquote>
<pre></pre>
<br>
<fieldset></fieldset>
<pre>_______________________________________________
LLVM Developers mailing list
<a href="mailto:llvm-dev@lists.llvm.org" target="_blank">llvm-dev@lists.llvm.org</a>
<a href="https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev" target="_blank">https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev</a>
</pre>
</blockquote>
</div>
</blockquote></div>