[cfe-dev] [RFC] automatic variable initialization

Wed Jan 16 17:00:26 PST 2019

On Wed, Jan 16, 2019 at 4:48 PM Kostya Serebryany <kcc at google.com> wrote:

> we've collected the performance and code size data on one of our major
> server-side benchmarks (huge x86_64 binary)
>
> Stack Init Method -O1 -O2 -O3 -O2, ThinLTO -O2, ThinLTO, FDO
> zero-init Size Regression 0.5% 0.6% 0.6% 0.4% 0.3%
> Perf Regression 0.7% -0.9% 1.1% 1.0% 1.7%
> pattern-init Size Regression 1.6% 1.6% 1.6% 1.4% 1.6%
> Perf Regression 1.1% 1.2% 1.2% 0.8% 1.1%
>
> so, already not bad, but we pay too much code size for the pattern-init.
>

Thanks for posting the benchmarks, this is instructive!

It is interesting that FDO adds a non-trivial runtime regression over
non-FDO here (and somehow worse for zero init vs pattern), do you have any
clue about this?

Thanks!

-- 
Mehdi

> The following seems to be the dominating code pattern where we have to
> insert initialization:
>
> struct Bar {  // think std::vector
>   Bar() : a(nullptr), b(nullptr), c(nullptr) {}
>   long *a, *b, *c;
> };
> struct Foo {
>   Foo() *__attribute__((noinline))* {}
>   Bar v1, v2, v3;
> };
> int main() { Foo foo; }
>
> The Foo CTOR does not get inlined and so we insert initialization for all
> Foo's fields.
> What's worse, is that -ftrivial-auto-var-init*=pattern* produces very
> sub-optimal code here:
>
> x86_64:
> movq .L__const.main.foo+64(%rip), %rax
> movq %rax, 64(%rsp)
> movups .L__const.main.foo+48(%rip), %xmm0
> movaps %xmm0, 48(%rsp)
> movups .L__const.main.foo+32(%rip), %xmm0
> movaps %xmm0, 32(%rsp)
> movups .L__const.main.foo+16(%rip), %xmm0
> movaps %xmm0, 16(%rsp)
> movups .L__const.main.foo(%rip), %xmm0
> movaps %xmm0, (%rsp)
>
> aarch64:
> adrp x8, .L__const.main.foo
> add x8, x8, :lo12:.L__const.main.foo
> ldp q0, q3, [x8]
> ldr x9, [x8, #64]
> ldp q1, q2, [x8, #32]
> mov x0, sp
> stp q0, q3, [sp]
> str x9, [sp, #64]
> stp q1, q2, [sp, #32]
>
> I.e. instead of materializing the 16-byte pattern once, it does it many
> times.
>
> Compare to -ftrivial-auto-var-init*=zero*:
>
> xorps %xmm0, %xmm0
> movaps %xmm0, 48(%rsp)
> movaps %xmm0, 32(%rsp)
> movaps %xmm0, 16(%rsp)
> movaps %xmm0, (%rsp)
> movq $0, 64(%rsp)
>
> movi v0.2d, #0000000000000000
> mov x0, sp
> str xzr, [sp, #64]
> stp q0, q0, [sp, #32]
> stp q0, q0, [sp]
>
>
> OTOH, with a relatively simple LTO-based analysis we can get rid of these
> extra auto-inits completely.
>
> CC @Vitaly Buka <vitalybuka at google.com> @Peter Collingbourne
> <pcc at google.com> @Evgeniy Stepanov <eugenis at google.com>
>
>
> --kcc
>
> On Tue, Dec 18, 2018 at 11:38 AM Kostya Serebryany <kcc at google.com> wrote:
>
>> 2019 is going to have one bug class fewer. :)
>> (unless Jann&Co find new bug classes again)
>>
>> Looking forward to the followup patches (e.g. padding gaps)
>> and to doing the measurements.
>>
>> --kcc
>>
>> On Tue, Dec 18, 2018 at 10:53 AM JF Bastien <jfbastien at apple.com> wrote:
>>
>>> Hello fans of the pre-C++11 `auto` keyword!
>>>
>>> Thanks to reviews from Peter and Richard, automatic variable
>>> initialization landed last night:
>>>
>>> https://reviews.llvm.org/rL349442
>>>
>>>
>>> Chandler pointed out that I should have circled back to the RFC in case
>>> any of the finer points needed tuning. Let me know if there’s any follow-up
>>> you’d like! I’ve made zero-init harder / uglier to use as was requested,
>>> while allowing folks like Kostya to test it out.
>>>
>>> Let me know if you find bugs (in your code, or in the implementation),
>>> and please file bugs on any missed optimization opportunities. I know of a
>>> handful of missing optimizations which I intend to tackle in the near
>>> future. I’d also love to hear what kind of performance / size impact you
>>> see from various codebases.
>>>
>>> As pointed out in the patch there’s a few other hardenings that I’ll
>>> want to tackle next. I’ll do so in the new year.
>>>
>>> Thanks,
>>>
>>> JF
>>>
>>>
>>> On Dec 11, 2018, at 3:18 PM, Joe Bialek via cfe-dev <
>>> cfe-dev at lists.llvm.org> wrote:
>>>
>>> We went with zero initialization for production builds in Visual Studio
>>> because we anticipate zero being the fastest, smallest code gen, and safest
>>> value for the following cases:
>>>
>>>    - Pointer: A non-zero pointer could potentially point at valid code.
>>>    On Windows x64, the first 64kb of the virtual address space is guaranteed
>>>    to not be mappable and the last 512GB of the virtual address space (today)
>>>    has the space guarantee. So null pointer or near-null pointer dereferences
>>>    are denial of service at worst.
>>>    - Size: If you forget to set a size as an out parameter we’d rather
>>>    it’s set to 0 so you don’t index out of bounds.
>>>    - Index: Same as size.
>>>
>>>
>>> We are using this mitigation to downgrade vulnerabilities from remote
>>> code execution, elevation of privilege, or information disclosure, down to
>>> denial of service at worst.
>>>
>>> Assuming the denial of service isn’t for a scenario where it matters
>>> (there are many types of DOS we don’t really worry about for security), we
>>> will not bother servicing the vulnerabilities down-level and instead only
>>> fix them for new code.
>>>
>>> It is helpful for us to have deterministic behavior to help accomplish
>>> this goal. If an engineer knows a value is always set to 0, they can
>>> quickly determine if the bug report is actually exploitable or not. If it
>>> is initialized to a random or compile time constant value, we’ll need to do
>>> more work to determine if the bug is exploitable.
>>>
>>> We are not zero initializing CHK builds to prevent folks from taking a
>>> dependency on this feature.
>>>
>>>
>>> Joe
>>>
>>> *From:* Kostya Serebryany <kcc at google.com>
>>> *Sent:* Tuesday, December 11, 2018 3:04 PM
>>> *To:* Mehdi AMINI <joker.eph at gmail.com>; Jann Horn <jannh at google.com>
>>> *Cc:* David Blaikie <dblaikie at gmail.com>; Richard Smith <
>>> richard at metafoo.co.uk>; Clang Dev <cfe-dev at lists.llvm.org>; Joe Bialek <
>>> jobialek at microsoft.com>
>>> *Subject:* Re: [cfe-dev] [RFC] automatic variable initialization
>>>
>>> One more dimension in this discussion (zero-init vs pattern-init) is
>>> what will security get from those options.
>>> My second-hand knowledge here suggests that zero-init may have *on
>>> average* better security guarantees than non-zero init.
>>> For example if the uninitialized data is interpreted as a length of
>>> something, it better be zero than any large number or, worse, a negative
>>> signed number.
>>> +Jann Horn <jannh at google.com> and Joe Bialek who have first-hand
>>> knowledge here.
>>>
>>> --kcc
>>>
>>>
>>>
>>>
>>> On Thu, Dec 6, 2018 at 3:58 PM Kostya Serebryany <kcc at google.com> wrote:
>>>
>>>
>>> On Thu, Dec 6, 2018 at 3:55 PM Mehdi AMINI <joker.eph at gmail.com> wrote:
>>>
>>>
>>> On Thu, Dec 6, 2018 at 3:43 PM Kostya Serebryany <kcc at google.com> wrote:
>>>
>>>
>>> On Thu, Dec 6, 2018 at 3:03 PM Mehdi AMINI <joker.eph at gmail.com> wrote:
>>>
>>>
>>> On Thu, Dec 6, 2018 at 1:01 PM Kostya Serebryany <kcc at google.com> wrote:
>>>
>>>
>>> On Wed, Dec 5, 2018 at 6:07 PM Mehdi AMINI via cfe-dev <
>>> cfe-dev at lists.llvm.org> wrote:
>>>
>>> I have the same question as David: I don't understand why this isn't
>>> just an experimental build option disabled by default? (Google, Apple, etc.
>>> can just enable it in their build script, no need to patch the source).
>>>
>>>
>>> I will need to rebuild half a dozen compiler binaries to do the
>>> measurements I need.
>>> This is going to double the cost of the effort for me because it adds
>>> too many extra moving pieces.
>>>
>>>
>>> Not sure I follow, you have to build the compiler anyway to get it after
>>> the code is patched?
>>> The only thing you would have to do is one CL to enable the build flag
>>> (that expose the command line flag) inside your codebase and then you get
>>> your toolchain as usual?
>>>
>>>
>>> Once the JF's patch is in, I'll have it in most production compilers I
>>> care about in 2-6 weeks, w/o any effort on my side.
>>> None of those builders (easily) support adding custom patches, and even
>>> a build flag is going to be very non-trivial.
>>> I can deal with one or two of those builds relatively easily, but not
>>> with the rest.
>>>
>>>
>>> What builders are you referring to? So far I was assuming we were
>>> talking about your internal infrastructure that produces your production
>>> compilers.
>>>
>>>
>>> I am talking about our internal infra.
>>> But we have multiple independent builders from multiple independent
>>> compiler users (e.g. Chrome and Android are using different compiler
>>> builds).
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> --
>>> Mehdi
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> The only argument I read in the thread was about allowing people to make
>>> performance measurement without rebuilding the compiler, but I have a hard
>>> time reconciliation this with the fact that we're talking about not
>>> shipping this before performing the actual measurements?
>>>
>>> I expect that anyone that cares enough about the perf impact of this to
>>> influence the development of the feature in clang should already be
>>> rebuilding their compiler today.
>>>
>>> --
>>> Mehdi
>>>
>>>
>>> On Mon, Dec 3, 2018 at 5:58 PM David Blaikie via cfe-dev <
>>> cfe-dev at lists.llvm.org> wrote:
>>>
>>> Just out of curiosity - what's the major difference between a build-time
>>> off-by-default feature and a build-time on-by-default-but-off-in-release
>>> branches feature? If we're only targeting groups that build/release the
>>> compiler themselves, then they're likely able to opt-in to a build-time
>>> feature easily enough, I'd think? & then there's no need to make our
>>> releases different from day-to-day builds?
>>>
>>> But sounds like folks are in general agreement of a way forward, so I
>>> don't want to disrupt/delay that.
>>> On Wed, Nov 28, 2018 at 11:14 PM Chandler Carruth <chandlerc at gmail.com>
>>> wrote:
>>>
>>> Suggested compromise technique to at least get an initial set of numbers:
>>>
>>> 1) Require a special, long, ugly flag name.
>>> 2) Make it a CC1 flag, requiring -Xclang ... to use.
>>> 3) Emit a warning by default (that cannot be suppressed with a -Wno-...
>>> flag) when this flag is enabled.
>>> 4) Commit to never including this flag in any upstream release. Either
>>> we remove it before the next release branches or we revert it on the branch.
>>>
>>> Most of the folks we're hoping to get performance data with are willing
>>> to use a not-yet-released build of Clang. They won't have to actually patch
>>> it in any way. They will have strong reminders to not deploy this in any
>>> way due to the warning.
>>>
>>> Thoughts?
>>>
>>> On Wed, Nov 28, 2018 at 4:34 PM Kostya Serebryany via cfe-dev <
>>> cfe-dev at lists.llvm.org> wrote:
>>>
>>> On Wed, Nov 28, 2018 at 3:28 PM David Blaikie <dblaikie at gmail.com>
>>> wrote:
>>>
>>>
>>> On Wed, Nov 28, 2018 at 3:17 PM Kostya Serebryany <kcc at google.com>
>>> wrote:
>>>
>>>
>>> Seems easier to me to separate the two pieces - move ahead with the
>>> non-zero options, and separate the discussion on the zero option. You can
>>> present performance numbers from what you can measure without shipping a
>>> compiler with the feature - and if those numbers are sufficiently
>>> compelling compared to the risks of slicing the language, then perhaps we
>>> go that way.
>>>
>>>
>>> This approach will significantly impair my ability to do the
>>> measurements I need.
>>>
>>>
>>> I'm aware waht I'm proposing would make it more difficult for some
>>> people to take measurements - that's a tradeoff to be sure - one where I
>>> err in this direction.
>>>
>>> Specifically for Google though - would it be that difficult for Google
>>> to opt-in to a certain build configuration of LLVM?
>>>
>>>
>>> Absolutely yes.
>>> Google is not just a single monolithic code base
>>> <https://nam06.safelinks.protection.outlook.com/?url=http%3A%2F%2Fdelivery.acm.org%2F10.1145%2F2860000%2F2854146%2Fp78-potvin.pdf%3Fip%3D104.133.8.94%26id%3D2854146%26acc%3DOA%26key%3D4D4702B0C3E38B35%252E4D4702B0C3E38B35%252E4D4702B0C3E38B35%252E5945DC2EABF3343C%26__acm__%3D1543446999_3aadcd36f657e2297430c38bee93f16c&data=02%7C01%7Cjobialek%40microsoft.com%7C533e0fe1444d4b5a3a6b08d65fbcf479%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636801662510848537&sdata=dd4I1xQ5qfawzLwsnBM3iqsypqZ%2BlFswj5IgKa%2FLHbw%3D&reserved=0>
>>> .
>>>
>>>
>>> Couldn't access this link ("An error occurred while processing your
>>> request" - but yeah, I understand there's a bunch of different pieces of
>>> Google beyond the "stuff that runs in data centers" piece we mostly support.
>>>
>>>
>>> Besides that monolithic thing, we have Android, Chrome, ChromeOS,
>>> Fuchsia, and a bazillion of smaller efforts that use their own toolchains.
>>>
>>>
>>> Still, most/all of these build their own compilers, I think? But yeah,
>>> that adds an opt-in overhead to each project, for sure.
>>>
>>>
>>> In some cases the most reliable and complete way of measuring
>>> performance changes is to submit the changes to revision control,
>>> and let the performance bots shew it for a couple of days. That's how we
>>> iterated with the LLVM's CFI in Chrome.
>>> We will also need to work with the upstream Linux kernel -- it's hard
>>> enough for them to use clang and a modified clang will cost us much more
>>> effort.
>>>
>>>
>>> Yeah, I can imagine that one's a bit trickier - how's performance
>>> evaluation of the kernel done?
>>>
>>>
>>> I don't think anyone knows that. :-|
>>> And requiring a compiler patch will shift the problem from "hard" to
>>> "I'd better do something else".
>>>
>>>
>>> (though, again, I imagine a fair amount of progress could be made
>>> without the zero-init feature - perhaps enough to say "hey, here are all
>>> the places we have run tests & seen the performance tradeoff is worthwhile
>>> for us (& possibly that it's close to the zero-init case, but that's sort
>>> of orthogonal, imho - that it's worthwhile is the main thing) - perhaps
>>> other folks would be willing to test it (non-zero init) & see if it's
>>> worthwhile to them - and if it isn't/they're interested in more
>>> performance, maybe all that evidence we can gain from the places where it's
>>> easy for us to rebuild compilers, etc, would make it interesting enough to
>>> motivate someone to do build the kernel with a custom compiler & do some
>>> performance measurements, etc...
>>>
>>> Sorry that was a bit rambly, anyway.
>>>
>>> - Dave
>>>
>>>
>>>
>>>
>>>
>>>
>>> We/Google do build the compiler from scratch, I assume we pick the
>>> configuration options we build with & some of them probably aren't the
>>> defaults for a release build of LLVM. So if it was important that Google's
>>> production compiler had these features enabled (rather than building a test
>>> compiler for running some experiments), that doesn't seem (at least to me,
>>> at this moment) especially prohibitive, is it?
>>>
>>>
>>> _______________________________________________
>>>
>>>
>>> cfe-dev mailing list
>>> cfe-dev at lists.llvm.org
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>>> <https://nam06.safelinks.protection.outlook.com/?url=http%3A%2F%2Flists.llvm.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Fcfe-dev&data=02%7C01%7Cjobialek%40microsoft.com%7C533e0fe1444d4b5a3a6b08d65fbcf479%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636801662510858549&sdata=J1qIBl35xJOU4LXt2pmJKrl7Jo3RzuG04e7mWmXdULo%3D&reserved=0>
>>>
>>> _______________________________________________
>>> cfe-dev mailing list
>>> cfe-dev at lists.llvm.org
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>>> <https://nam06.safelinks.protection.outlook.com/?url=http%3A%2F%2Flists.llvm.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Fcfe-dev&data=02%7C01%7Cjobialek%40microsoft.com%7C533e0fe1444d4b5a3a6b08d65fbcf479%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636801662510868553&sdata=wC0zWaGYC6YVWA3W7X3PEUjbePjhZdOoKVy5vzD6y2I%3D&reserved=0>
>>>
>>> _______________________________________________
>>> cfe-dev mailing list
>>> cfe-dev at lists.llvm.org
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>>> <https://nam06.safelinks.protection.outlook.com/?url=http%3A%2F%2Flists.llvm.org%2Fcgi-bin%2Fmailman%2Flistinfo%2Fcfe-dev&data=02%7C01%7Cjobialek%40microsoft.com%7C533e0fe1444d4b5a3a6b08d65fbcf479%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C636801662510868553&sdata=wC0zWaGYC6YVWA3W7X3PEUjbePjhZdOoKVy5vzD6y2I%3D&reserved=0>
>>>
>>> _______________________________________________
>>> cfe-dev mailing list
>>> cfe-dev at lists.llvm.org
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>>>
>>>
>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20190116/f78b7732/attachment.html>