Sanitizers + New Pass Manager

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.

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?

+some sanitizer/new pass manager folks

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.

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?)

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.

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.

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?

Sounds like this needs pinpointing where the behavior should be different than what’s currently happening.
Inlining in the NPM doesn’t seem the issue, it’s what it triggers afterwards. Is it that the lifetime intrinsics shouldn’t have been stripped, or is it that ASan shouldn’t rely on these to add instrumentation because previous optimizations may have removed them (hence it should not rely on what optimizations are done or not done; as you suggested, maybe changing the pass pipeline to add ASan earlier may be needed?), or is it something else?

Alina

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.

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.

Lifetime markers are an example. Once removed, which is generally legal in IR, you cannot argue accesses after the end are UB.

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?

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.

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?

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.

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)?

While trying newPM with inliner turned off on O1 is fine and probably a right approach on digging into the cause of the failure, actually changing Ox to be that different depending on sanitizer mode seems to be a bad idea to me. regards, Fedor.

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?

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.

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?

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.

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)?

While trying newPM with inliner turned off on O1 is fine and probably a right approach on digging into the cause of the failure,
actually changing Ox to be that different depending on sanitizer mode seems to be a bad idea to me.

Agreed. I don’t want to turn off the inliner at O1. Reduce it down for sure, but not turn it off. If you need help investigating these let me know and we can try to help.

-eric

Sanitizer passes really should not run before the inliner. For example, ASan moves all allocas into a “mega-alloca” to obtain fixed frame layout for reporting purposes. It also inserts a fake stack check in the function prologue which will get duplicated (but will probably still work) after inlining.

MSan removes readnone/readonly from all functions because they all update shadow which is a memory write. Doing this early will suppress all kinds of IPO.

I think ASan (and probably MSan) does the same thing as TSan regarding memory intrinsics because it needs the runtime library to be able to tell which memcpy/memset calls are application logic (and therefore need to check/update shadow) and which are compiler-inserted after sanitizer instrumentation. Doing this too early suppresses some optimizations.

It’s probably OK to run the passes a little earlier (but not before inlining), but I’d be careful to weigh the benefits of that against the loss of performance. If the goal is to catch more bugs (specifically, any undefined behavior that gets optimized out before instrumentation), I’d rather do something in the frontend and in the IR passes suppress those optimizations in sanitized functions.

Sanitizer passes really should not run before the inliner. For example, ASan moves all allocas into a “mega-alloca” to obtain fixed frame layout for reporting purposes. It also inserts a fake stack check in the function prologue which will get duplicated (but will probably still work) after inlining.

MSan removes readnone/readonly from all functions because they all update shadow which is a memory write. Doing this early will suppress all kinds of IPO.

I think ASan (and probably MSan) does the same thing as TSan regarding memory intrinsics because it needs the runtime library to be able to tell which memcpy/memset calls are application logic (and therefore need to check/update shadow) and which are compiler-inserted after sanitizer instrumentation. Doing this too early suppresses some optimizations.

It’s probably OK to run the passes a little earlier (but not before inlining), but I’d be careful to weigh the benefits of that against the loss of performance. If the goal is to catch more bugs (specifically, any undefined behavior that gets optimized out before instrumentation), I’d rather do something in the frontend and in the IR passes suppress those optimizations in sanitized functions.

I think I’ve seen some function attributes added when sanitizers are enabled, maybe we can use those to e.g. not strip lifetime intrinsics. I’ll look into that.

Sanitizer passes really should not run before the inliner. For example,
ASan moves all allocas into a "mega-alloca" to obtain fixed frame layout
for reporting purposes. It also inserts a fake stack check in the function
prologue which will get duplicated (but will probably still work) after
inlining.

MSan removes readnone/readonly from all functions because they all update
shadow which is a memory write. Doing this early will suppress all kinds of
IPO.

I think ASan (and probably MSan) does the same thing as TSan regarding
memory intrinsics because it needs the runtime library to be able to tell
which memcpy/memset calls are application logic (and therefore need to
check/update shadow) and which are compiler-inserted after sanitizer
instrumentation. Doing this too early suppresses some optimizations.

It's probably OK to run the passes a little earlier (but not before
inlining), but I'd be careful to weigh the benefits of that against the
loss of performance. If the goal is to catch more bugs (specifically, any
undefined behavior that gets optimized out before instrumentation), I'd
rather do something in the frontend and in the IR passes suppress those
optimizations in sanitized functions.

I think I've seen some function attributes added when sanitizers are
enabled, maybe we can use those to e.g. not strip lifetime intrinsics. I'll
look into that.

I guess that is an option we have to see how invasive this would be.

I could imagine to use the `isDroppable` interface for this as we wanted

to include lifetime markers and other things in there anyway. Passes can

replace their explicit handling of lifetime markers for an `isDroppable`

check that is influenced by the function attributes.

Would it make sense to have an option that allows the user to modify

the sanitizer position, thus exposing the trade off to them?