[llvm-dev] GEP with a null pointer base
Peter Lawrence via llvm-dev
llvm-dev at lists.llvm.org
Tue Aug 1 16:17:11 PDT 2017
Dave,
I will try to locate and take a look at the actual llvm logic that deletes based
on UB-presence, one of these days, and report back. In the mean time...
Your “For example:" is a plausibility argument only. It is not meaningful until
you can show this happening in real source code from real applications
that are compiler warning free and static analysis warning free.
Hal already conceded that this is never likely to happen in for example SPEC
benchmarks, I’m willig to bet it never happens even in a highly abstracted application
like llvm itself.
Saying “The C++ language lets me assume that that won’t happen, & optimize
on that basis” is an assumption that that’s what the user wants, but you
haven’t asked you’ve just assumed, and AFAICT it is an incorrect assumption.
What’s worse is that there isn’t any way to opt-out of this assumption
(-fsanitize=undefined isn’t opt’ing out, its opt’ing in to a lot extra that I
might or might not want).
Peter Lawrence.
> On Jul 31, 2017, at 9:48 AM, David Blaikie <dblaikie at gmail.com> wrote:
>
>
>
> On Mon, Jul 31, 2017 at 7:40 AM Peter Lawrence <peterl95124 at sbcglobal.net <mailto:peterl95124 at sbcglobal.net>> wrote:
> Dave,
> Dead code elimination is generally done in a pass called dead code elimination,
> Can you give concrete examples why the same would not be true for UB code elimination ?
>
> I haven't actually looked at how optimizations on the basis of the code being UB-free cause code to be eliminated - I'd hazard a guess that many optimizations would replace certain code with undef or unreachable not only a specific pass intended for that purpose. At least I'm pretty sure that's how it is today (I know we don't have a "UB code elimination" pass - it comes out as a result of many other passes making steps like mutation to undef/unreachable)
>
> Yes, speculatively hoisting code requires it to be UB-free, but that has nothing to do with
> UBCE deleting entire blocks of code because of the existence of UB.
>
> There's more similarity there than might first appear - if we assume the code is UB-free, then when an unconditional store to a null pointer is found we assume that must be dynamically unreachable (& the only reason it came up is due to inlining, etc - not that the user wrote/intended to store to null - maybe this particular call site got inlined and now what was a dynamic property (the user always called that function with a parameter that ensured the program didn't go down that code path and load from the pointer - but other callers do, so that's fine and good and correct) is a static property of this particular version/post-inlining state so the compiler can trivially observe it and replace it with unreachable)
>
> For example:
>
> inline __attribute__((always_inline)) void f1(bool b, int *i) {
> if (b)
> *i = 3;
> }
>
> bool f2(int*);
>
> void f3() {
> int *x = nullptr;
> f1(f2(x), x);
> }
>
> For the sake of argument, you could imagine that this code arose from perfectly reasonable user code - the user knows that 'f2' always returns false for a null pointer because it says so in the name (in the user code, with real names, etc).
>
> & it's actually Jump Threading that makes the code in the 'if' block (once inlined into f3) unreachable:
>
> *** IR Dump Before Jump Threading ***
> ; Function Attrs: uwtable
> define void @_Z2f3v() local_unnamed_addr #1 {
> entry:
> %call = call zeroext i1 @_Z2f2Pi(i32* null)
> br i1 %call, label %if.then.i, label %_Z2f1bPi.exit
>
> if.then.i: ; preds = %entry
> store i32 3, i32* null, align 4, !tbaa !2
> br label %_Z2f1bPi.exit
>
> _Z2f1bPi.exit: ; preds = %entry, %if.then.i
> ret void
> }
> *** IR Dump After Jump Threading ***
> ; Function Attrs: uwtable
> define void @_Z2f3v() local_unnamed_addr #1 {
> entry:
> %call = call zeroext i1 @_Z2f2Pi(i32* null)
> br i1 %call, label %if.then.i, label %_Z2f1bPi.exit
>
> if.then.i: ; preds = %entry
> call void @llvm.trap()
> unreachable
>
> _Z2f1bPi.exit: ; preds = %entry
> ret void
> }
>
> Sure, it didn't remove the block, but it certainly removed the code. Nice enough to replace it with a trap, though. (so, yeah, not a perfect example - but you get the idea, all sorts of optimizations might decide to simplify code on the basis that UB can't actually happen)
>
> The former requires
> an analysis proving UB-absense, the later requires an analysis proving UB-presence. But it
> isn’t logical to say you must delete UB-presence to enable proving UB-absense, because
> in the real world after programs have passed static-analysis there won’t be any UB-presence
> to delete.
>
> I don't understand this last bit, sorry. Without knowledge of the whole program I can't prove that the example above will never execute UB - but the C++ language lets me, the compiler, assume that that won't happen & optimize on that basis. No local static analysis could prove it - but the assumption helps improve the code. (& even if we had the whole program, some invariants would be difficult/impossible to practically prove - programmer indexes into an array based on the number of elementsn in a hash table - pretty darn difficult to statically analyze the whole probing hash table system, and the dynamic logic that populates it, to prove that only N elements will ever end up in the hash table, and thus using its size as an index into an array of N elements is safe - but we can assume, so there's no need to synthesize bounds checks around the array access, etc)
>
>
>
> Peter Lawrence.
>
>
>> On Jul 27, 2017, at 7:45 PM, David Blaikie <dblaikie at gmail.com <mailto:dblaikie at gmail.com>> wrote:
>>
>>
>>
>> On Thu, Jul 27, 2017 at 7:30 PM Peter Lawrence <peterl95124 at sbcglobal.net <mailto:peterl95124 at sbcglobal.net>> wrote:
>> Dave,
>> The way I see it there should be just one pass that implements
>> deleting UB (maybe it would come to be called UBCE), and that one pass
>> should have a command line option simply for the reason than all passes
>> should have one.
>>
>> I would hazard a guess that would be very difficult to implement the same functionality (as the compiler has today) efficiently with that approach - many optimizations would be inhibited from making progress because they couldn't prove the code didn't have UB, etc.
>>
>> Generally, even deleting UB code isn't approached directly, but may come about as a consequence of various other steps the compiler is taking.
>>
>> - Dave
>>
>>
>>
>> Peter Lawrence.
>>
>>
>>> On Jul 26, 2017, at 10:02 PM, David Blaikie <dblaikie at gmail.com <mailto:dblaikie at gmail.com>> wrote:
>>>
>>>
>>>
>>> On Wed, Jul 26, 2017 at 9:23 PM Peter Lawrence <peterl95124 at sbcglobal.net <mailto:peterl95124 at sbcglobal.net>> wrote:
>>> David,
>>> -fsanitize=undefined sounds great, but is not quite what I want.
>>>
>>>
>>> I recently ran into a problem with "CodeGen/MachineSink.cpp” [*], for a target
>>> that has to expand Select into control flow.
>>> The original IR had two select in a row that were based on the same condition,
>>> so the CMP that sets the FLAGS reg in the second select was MCSE’ed to the
>>> earlier CMP in the first select, so here we see the second Select without a CMP:
>>>
>>> BB#10: derived from LLVM BB %for.body.5
>>> Predecessors according to CFG: BB#3 BB#9
>>> %vreg49<def> = PHI %vreg47, <BB#9>, %vreg48, <BB#3>; DataRegs:%vreg49,%vreg47,%vreg48
>>>
>>> //// <=== this SLLI clobbers FLAGS <============
>>> %vreg46<def> = SLLI %vreg5, 1, %FLAGS<imp-def,dead>; DataRegs:%vreg46,%vreg5
>>> BCC 2, <BB#12>, %FLAGS<imp-use>
>>> Successors according to CFG: BB#11 BB#12
>>>
>>>
>>> The problem is that Machine Code Sinking put an “SLLI" instruction, that
>>> modifies the FLAGS registers, in between the CMP and the BCC.
>>>
>>> The way I was able to work around this problem was to add
>>> a command line option to “MachineSink.cpp” that defaults to false,
>>> and add a check in its runOnMachineFunction() to omit this pass.
>>>
>>>
>>> But I should not have had to, every FunctionPass and MachineFunctionPass
>>> should have a name and a command line option to disable it by name.
>>>
>>> Other compilers I’ve worked on have had such options, and I use them to
>>> track down compiler bugs. In this case I instead had to "—debug-after-all"
>>> and very tediously search through thousands of lines of output to locate
>>> this bug.
>>>
>>>
>>> So I hope you can see where I’m coming from, the pass that deletes UB
>>> should be no different, I should be able to disable it from the command line
>>> as a matter of course.
>>>
>>> Thoughts ?
>>>
>>> These things seem like distinct goals to me - debuggability of the compiler and dealing with UB. I think -fsanitize=undefined is a pretty good way to help users deal with, diagnose, and act on UB. It isolates the issue - rather than having all optimizations have ways of switching themselves off when part of their analysis relies on UB - instead the optimizations rely on the IR definitions to make valid transformations, and it's a separate pass in the frontend (clang) that can add extra checks to avoid UB (& diagnose it as such, if that's the best thing to do).
>>>
>>> One could make an IR version of UBSan, that could take some IR and add null checks around every load/store, all the other things UBSan does. But I wouldn't expect that would be a priority for anyone to build (as it's more a compiler developer tool at that point - smaller audience, etc).
>>>
>>> For your debugging situation - bugpoint can slice & dice the pass list to try to reduce both the set of code being optimized, and the particular optimization that seems to be causing the problem. That might be worth a shot for you in cases like this?
>>>
>>> Otherwise I find print-after-all or the like to be pretty handy. Usually I reduce the input code first (with something like creduce or delta) so the IR isn't massive/hard to read. (bugpoint can help with that reduction as well)
>>>
>>> - Dave
>>>
>>>
>>>
>>> Peter.
>>>
>>>
>>> [* I haven’t reported this as a bug yet because I’m on 3.7.1, and haven’t had time
>>> to replicate it in 4.0.1, but should be able to within a month. My target resembles
>>> MSP430, so I’ll try to replicate it for that target in 4.0.1 ]
>>>
>>>
>>>
>>>> On Jul 24, 2017, at 9:08 AM, David Blaikie <dblaikie at gmail.com <mailto:dblaikie at gmail.com>> wrote:
>>>>
>>>>
>>>>
>>>> On Mon, Jul 24, 2017 at 9:02 AM Peter Lawrence <peterl95124 at sbcglobal.net <mailto:peterl95124 at sbcglobal.net>> wrote:
>>>>> On Jul 21, 2017, at 10:55 PM, Mehdi AMINI <joker.eph at gmail.com <mailto:joker.eph at gmail.com>> wrote:
>>>>>
>>>>>
>>>>>
>>>>> 2017-07-21 22:44 GMT-07:00 Peter Lawrence <peterl95124 at sbcglobal.net <mailto:peterl95124 at sbcglobal.net>>:
>>>>> Mehdi,
>>>>> Hal’s transformation only kicks in in the *presence* of UB
>>>>>
>>>>> No, sorry I entirely disagree with this assertion: I believe we optimize program where there is no UB. We delete dead code, code that never runs, so it is code that does not exercise UB.
>>>>>
>>>>
>>>> Mehdi,
>>>> I had to read that sentence several times to figure out what the problem
>>>> is, which is sloppy terminology on my part
>>>>
>>>> Strictly speaking the C standard uses “undefined behavior” to describe what
>>>> happens at runtime when an “illegal” construct is executed. I have been using
>>>> “undefined behavior” and UB to describe the “illegal” construct whether it is
>>>> executed or not.
>>>>
>>>> Hence I say “Hal’s transform is triggered by UB”, when I should be saying
>>>> “Hal’s transformation is triggered by illegal IR”.
>>>>
>>>> All I can say is I’m not the only one being sloppy, what started this entire
>>>> conversation is the paper titled “Taming Undefined Behavior in LLVM”, while
>>>> the correct title would be “Taming Illegal IR in LLVM”. (I think we are all
>>>> pretty confident that LLVM itself is UB-free, or at least we all hope so :-).
>>>> I believe you are being sloppy when you say "we optimize program
>>>> where there is no UB”, because I believe you mean "we optimize program
>>>> under the assumption that there is no UB”. In other words we recognize
>>>> “Illegal” constructs and then assume they are unreachable, and delete
>>>> them, even when we can’t prove by any other means that they are
>>>> unreachable. We don’t know that there is no (runtime) UB, we just assume it.
>>>>
>>>>
>>>>> The example Hal showed does not exhibit UB, it is perfectly valid according to the standard.
>>>>>
>>>>
>>>> Whether it exhibits UB at runtime or not is not the issue, the issue is what
>>>> a static analyzer or compiler can tell before runtime, see below
>>>>
>>>>>
>>>>> , and
>>>>> it does not matter how that UB got there, whether by function inlining
>>>>> or without function inlining.
>>>>>
>>>>> The problem with Hal’s argument is that the compiler does not have
>>>>> a built in ouija board with which it can conjure up the spirit of the
>>>>> author of the source code and find out if the UB was intentional
>>>>> with the expectation of it being deleted, or is simply a bug.
>>>>> Function inlining does not magically turn a bug into not-a-bug, nor
>>>>> does post-inlining simplification magically turn a bug into not-a-bug.
>>>>>
>>>>> Let me say it again: if the compiler can find this UB (after whatever
>>>>> optimizations it takes to get there) then the static analyzer must
>>>>> be able to do the same thing, forcing the programmer to fix it
>>>>> rather than have the compiler optimize it.
>>>>>
>>>>> This is again incorrect: there is no UB in the program, there is nothing the static analyzer should report.
>>>>
>>>>
>>>> Hal’s example starts with this template
>>>>
>>>>> template <typename T>
>>>>> int do_something(T mask, bool cond) {
>>>>> if (mask & 2)
>>>>> return 42;
>>>>>
>>>>> if (cond) {
>>>>> T high_mask = mask >> 48; // UB if sizeof(T) < 8, and cond true
>>>>> if (high_mask > 5)
>>>>> do_something_1(high_mask);
>>>>> else
>>>>> do_something_2();
>>>>> }
>>>>>
>>>>> return 0;
>>>>> }
>>>>
>>>>
>>>> Which is then instantiated with T = char,
>>>> and where it is impossible for either a static analyzer or a
>>>> compiler to figure out and prove that ‘cond’ is always false.
>>>>
>>>> Hence a static analyzer issues a warning about the shift,
>>>> while llvm gives no warning and instead optimizes the entire
>>>> if-statement away on the assumption that it is unreachable.
>>>>
>>>> Yes a static analyzer does issue a warning in this case.
>>>>
>>>>
>>>> This is not the only optimization to be based on assumption
>>>> rather than fact, for example type-based-alias-analysis is
>>>> based on the assumption that the program is free of this sort
>>>> of aliasing. The difference is that a user can disable TBAA
>>>> and only TBAA if a program seems to be running incorrectly
>>>> when optimized and thereby possibly track down a bug, but
>>>> so far there is no command line option to disable UB-based-
>>>> analysis (or ‘illegal-IR-based” :-), but there really needs to be.
>>>>
>>>> Do we at least agree on that last paragraph ?
>>>>
>>>> We likely agree it's good to have tools to help developers identify/diagnose UB in their programs. And we have that: -fsanitize=undefined (not only does it effectively disable many UB-based optimizations (because it makes them not undefined - by conditionalizing the code to check that UB isn't reached, as such) - it even provides pretty diagnostics (of course you can't actually continue running the program - if the line after the diagnostic will dereference a null pointer - there's no non-null pointer we can magic-up, so execution must stop))
>>>>
>>>>
>>>>
>>>> Peter Lawrence.
>>>>
>>>>
>>>>
>>>>
>>>>
>>>>>
>>>>> The compile is still able to delete some code, because of breaking the abstraction through inlining or template instantiation for example (cf Hal example).
>>>>>
>>>>> --
>>>>> Mehdi
>>>>>
>>>>>
>>>>>
>>>>> Or, to put it another way: there is no difference between a compiler
>>>>> and a static analyzer [*]. So regardless of whether it is the compiler or
>>>>> the static analyzer that finds any UB, the only rational thing to do with
>>>>> it is report it as a bug.
>>>>>
>>>>>
>>>>> Peter Lawrence.
>>>>>
>>>>>
>>>>> [* in fact that’s one of the primary reasons Apple adopted llvm, to use
>>>>> It as a base for static analysis]
>>>>>
>>>>>
>>>>>
>>>>>> On Jul 21, 2017, at 10:03 PM, Mehdi AMINI <joker.eph at gmail.com <mailto:joker.eph at gmail.com>> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>> 2017-07-21 21:27 GMT-07:00 Peter Lawrence <peterl95124 at sbcglobal.net <mailto:peterl95124 at sbcglobal.net>>:
>>>>>> Sean,
>>>>>> Let me re-phrase a couple words to make it perfectly clear
>>>>>>
>>>>>>> On Jul 21, 2017, at 6:29 PM, Peter Lawrence <peterl95124 at sbcglobal.net <mailto:peterl95124 at sbcglobal.net>> wrote:
>>>>>>>
>>>>>>> Sean,
>>>>>>>
>>>>>>> Dan Gohman’s “transform” changes a loop induction variable, but does not change the CFG,
>>>>>>>
>>>>>>> Hal’s “transform” deletes blocks out of the CFG, fundamentally altering it.
>>>>>>>
>>>>>>> These are two totally different transforms.
>>>>>>>
>>>>>>>
>>>>>>> And even the analysis is different,
>>>>>>>
>>>>>>> The first is based on an *assumption* of non-UB (actually there is no analysis to perform)
>>>>>> the *absence* of UB
>>>>>>>
>>>>>>> the second Is based on a *proof* of existence of UB (here typically some non-trivial analysis is required)
>>>>>> the *presence* of UB
>>>>>>
>>>>>>> These have, practically speaking, nothing in common.
>>>>>>>
>>>>>>
>>>>>>
>>>>>> In particular, the first is an optimization, while the second is a transformation that
>>>>>> fails to be an optimization because the opportunity for it happening in real world
>>>>>> code that is expected to pass compilation without warnings, static analysis without
>>>>>> warnings, and dynamic sanitizers without warnings, is zero.
>>>>>>
>>>>>> Or to put it another way, if llvm manages to find some UB that no analyzer or
>>>>>> sanitizer does, and then deletes the UB, then the author of that part of llvm
>>>>>> is in the wrong group, and belongs over in the analyzer and/or sanitizer group.
>>>>>>
>>>>>> I don't understand your claim, it does not match at all my understand of what we managed to get on agreement on in the past.
>>>>>>
>>>>>> The second transformation (dead code elimination to simplify) is based on the assumption that there is no UB.
>>>>>>
>>>>>> I.e. after inlining for example, the extra context of the calling function allows us to deduce the value of some conditional branching in the inline body based on the impossibility of one of the path *in the context of this particular caller*.
>>>>>>
>>>>>> This does not mean that the program written by the programmer has any UB inside.
>>>>>>
>>>>>> This is exactly the example that Hal gave.
>>>>>>
>>>>>> This can't be used to expose any meaningful information to the programmer, because it would be full of false positive. Basically a program could be clean of any static analyzer error, of any UBSAN error, and totally UB-free, and still exhibit tons and tons of such issues.
>>>>>>
>>>>>> --
>>>>>> Mehdi
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170801/0bfcc7aa/attachment.html>
More information about the llvm-dev
mailing list