llvm dev - Jan 2016 - Proposal: virtual constant propagation

Hi,

I just thought about another use case: VTable compression.
If you know that an entry in the Vtable is never used, just remove it!
I’d hope we could even eliminate some unused virtual functions from the final
binary.

— 
Mehdi

> On Jan 27, 2016, at 10:29 PM, Mehdi Amini via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
> 
> Hi Peter,
> 
> Pete (Cooper, CC'ed) had a similar idea a few months ago about
devirtualization and LTO: we can know the set of possible overload based on the
knowledge of a closed type hierarchy. He worked on a prototype, and I continued
to evolve it. I planned to submit something on the mailing list but couldn’t fit
it in my planning before April :)
> Our plan was about removing virtual call in general, or maybe in some cases
turning the indirect call into a switch table.
> 
> In both use case (your constant propagation and devirtualization), I think
the interesting part is “how to figure the set of possible callees for a call
site” and how we encode this in the IR.
> 
> I haven’t worked on this for a few weeks, but I’ll try to give a rough
description of where I stopped on my prototype:
> I was storing the full inheritance tree in metadata. There was one metadata
per entry of the VTable, and any method that override one of the base class will
have a metadata attached pointing to the slot in the base class.
> The front-end is modified to emit any load of a VTable through a new
intrinsic llvm.read.vtable that takes a pointer to the VTable, the index of the
slot in the VTable, and a pointer to the metadata that describe the slot for the
current type. Using the metadata you access directly to any override and
construct naturally and very easily the set of possible overloads. The pass that
transform the IR is quite easy to implement since you don’t need to walk the IR
but access directly all the uses of the llvm.read.vtable intrinsic and have all
the needed information in the operands.
> I haven’t give much thought about your representation yet to know how it
compares (and I’m not very familiar with how CFI works), but I’m interested in
your feedback!
> 
> I was limited during my experiment by the lack of black/white list to
define what hierarchy is/isn’t closed, I’m glad if such flag/control would be
added!
> 
> Best,
> 
> — 
> Mehdi
> 
> 
> 
>> On Jan 27, 2016, at 7:57 PM, Peter Collingbourne via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
>> 
>> Hi all,
>> 
>> I'd like to make the following proposal to implement an
optimization called
>> virtual constant propagation.
>> 
>> ==Introduction=>> After enabling control flow integrity
protection in Chromium, we have
>> observed an unacceptable performance regression in certain critical
layout
>> microbenchmarks. Profiling [0] revealed that the cause of the
regression was
>> a large number of virtual calls, each requiring the overhead of a
control
>> flow integrity check.
>> 
>> We took a closer look at the hottest virtual call sites, and found that
they
>> were mostly type checking functions of the following form:
>> 
>> struct Base {
>> enum Type { kBase, kDerived };
>> virtual bool isDerived() const { return false; } // or:
>> virtual bool isOfType(Type t) const { return t == kBase; }
>> };
>> struct Derived : Base {
>> virtual bool isDerived() const { return true; }
>> virtual bool isOfType(Type t) const { return t == kDerived; }
>> };
>> 
>> bool f(Base *b) {
>> return b->isDerived();
>> }
>> bool g(Base *b) {
>> return b->isOfType(Base::kDerived);
>> }
>> 
>> We can make the observation that if the type hierarchy is known to be
closed
>> and the definition of each virtual function and each virtual call is
visible
>> to the compiler, calls to either of these functions can be implemented
more
>> efficiently by loading directly from the vtable (we refer to this as
virtual
>> constant propagation):
>> 
>> struct Base_VTable {
>> bool isDerived : 1, isOfTypeBase : 1, isOfTypeDerived : 1;
>> };
>> Base_VTable base_vtable{false, true, false};
>> Derived_VTable derived_vtable{true, false, true};
>> bool f(Base *b) {
>> return b->vtable->isDerived;
>> }
>> bool g(Base *b) {
>> return b->vtable->isOfTypeDerived;
>> }
>> 
>> Another advantage of implementing the calls this way is that because
there
>> is no indirect call taking place, no control flow integrity check is
necessary.
>> 
>> I implemented a prototype of this idea that specifically targeted the
hottest
>> virtual function in one of Chromium’s layout benchmarks, and observed a
>> performance improvement of median 1.62%, min -4.72%, max 204.52% for
the
>> layout benchmark suite. I’d like to emphasise that these numbers are
from
>> a regular LTO build without CFI, and that we expect to see better
performance
>> from a general implementation that targets all virtual functions.
>> 
>> ==User interface=>> To instruct the compiler to assume that all
type hierarchies are closed, a
>> user can pass the -fwhole-program-vtables flag. The
-fwhole-program-vtables
>> flag requires the -flto flag to also be specified.
>> 
>> Of course, there may be some type hierarchies that are not entirely
closed, but
>> the underlying assumption is that most hierarchies will not be. To
support open
>> hierarchies the user can also specify the path to a blacklist that
lists the
>> names of the open class hierarchies using a
-fwhole-program-vtables-blacklist>> flag. By default, the compiler will
assume that hierarchies in the std
>> namespace are open.
>> 
>> ==High level design=>> Devirtualization will take place at LTO
time. Candidates for virtual constant
>> propagation will be found by searching for call sites with all-constant
>> arguments to virtual functions which in all derived classes are
readnone,
>> which means that it is a pure function of its arguments. For each
(virtual
>> function, arguments) candidate pair, the compiler will compute a offset
in
>> bits (positive or negative) from the class’s vtable address point in
which
>> to store the function result, extend the size of any vtable objects to
cover
>> any such offsets, evaluate the function to compute a result and store
it
>> at the correct offsets. To avoid breaking ABI, the vtable layout will
not
>> be changed in any way. This includes pointers to optimized functions in
the
>> vtable, and (for Itanium ABI) the distance between address points
within a
>> single vtable symbol.
>> 
>> For example, the vtable for Base and Derived will look like this:
>> 
>> struct Base_VTable {
>> bool isDerived : 1, isOfTypeBase : 1, isOfTypeDerived : 1;
>> size_t padding : 61;
>> size_t offset_to_top;
>> size_t rtti;
>> bool (*isDerived)(const Base *this);
>> bool (*isOfType)(const Base *this, Type t);
>> };
>> Base_VTable base_vtable{
>> false, true, false, 0, 0, 0, &Base::isDerived, &Base::IsOfType
>> };
>> Base_VTable derived_vtable{
>> True, false, true, 0, 0, 0, &Derived::isDerived,
&Derived::IsOfType
>> };
>> 
>> The symbol _ZTV4Base, representing the start of Base’s vtable set, will
>> be defined as base_vtable+8; similarly for _ZTV7Derived and
derived_vtable+8.
>> 
>> To avoid bloating virtual tables by too much in degenerate cases, there
will
>> be a limit for each virtual function on the number of bits that the
virtual
>> function may contribute to the size of the vtable, which we may fix at
(say)
>> 2 machine words.
>> 
>> To give an idea of how the generated code would look, here is an asm
snippet
>> from the hottest virtual call from Chromium’s layout benchmarks before
>> the transformation:
>> 
>> 255c9ae:       48 8b 03                mov    (%rbx),%rax
>> 255c9b1:       be 30 00 00 00          mov    $0x30,%esi
>> 255c9b6:       48 89 df                mov    %rbx,%rdi
>> 255c9b9:       ff 90 e0 02 00 00       callq  *0x2e0(%rax)
>> 255c9bf:       84 c0                   test   %al,%al
>> 255c9c1:       74 04                   je     255c9c7
>> 255c9c3:       31 c0                   xor    %eax,%eax
>> 255c9c5:       5b                      pop    %rbx
>> 255c9c6:       c3                      retq   
>> 
>> And after:
>> 
>> 255a8ee:       48 8b 03                mov    (%rbx),%rax
>> 255a8f1:       f6 40 e6 01             testb  $0x1,-0x1a(%rax)
>> 255a8f5:       74 04                   je     255a8fb
>> 255a8f7:       31 c0                   xor    %eax,%eax
>> 255a8f9:       5b                      pop    %rbx
>> 255a8fa:       c3                      retq   
>> 
>> ==LLVM IR-level design=>> Given a class name, how can we
determine the closed set of derived classes and
>> possible function pointers for a call site? As it turns out the IR
already
>> has a way of expressing this information: bitsets [1], which are
already being
>> used to implement CFI. We can encode the information for
devirtualization
>> by combining the @llvm.bitset.test and @llvm.assume intrinsics.
>> 
>> %vtable = load i8**, i8*** %obj
>> %p = call i1 @llvm.bitset.test(%vtable, “Base”)
>> call void @llvm.assume(i1 %p) ; %vtable is assumed to point to
>>                             ; _ZTV4Base+16 or _ZTV7Derived+16.
>> %fptrptr = getelementptr i8* %vtable, 1
>> %fptr = load i8*, i8** %fptrptr ; %fptr must point to Base::isOfType
>>                               ; or Derived::isOfType.
>> %fptr_casted = bitcast i8% %fptr to i1 (i8***, i32)
>> %result = call i1 %fptr_casted(i8*** %obj, i32 0)
>> 
>> This gives the optimizer all the information it needs to implement the
>> optimization: the addresses of the virtual tables, the addresses of the
>> function pointers (within the virtual table initializers) and the
constant
>> arguments. Note that the compiler does not need to do any of the work
>> described in the above link to lay out globals or create bitsets if all
>> calls to llvm.bitset.test are passed to llvm.assume.
>> 
>> If virtual constant propagation succeeds, the transformed IR will look
>> like this:
>> 
>> %vtable = load i8*, i8** %obj
>> %vtable_bit_addr = getelementptr i8* %vtable, i64 -17
>> %vtable_bit = load i8 %vtable_bit_addr
>> %vtable_bit_and = and i8 %vtable_bit, 1
>> %result = icmp ne %vtable_bit_and, 0
>> 
>> The pass that applies this transformation will be placed early in the
LTO
>> pipeline, before most of the regular optimization passes. The pass
could later
>> be extended to do general devirtualization based on the bitset
information:
>> 
>> - The pass can implement single-deriver (i.e. only one derived class
with a
>>  non-pure virtual function definition) and single-implementation
>>  (i.e. multiple derived classes all sharing a virtual function
definition
>>  from a base class) devirtualization by checking that each of the
possible
>>  values loaded from the vtables in the bitset are either identical or
>>  __cxa_pure_virtual (calls to pure virtuals are UB, so we’d be fine
>>  disregarding them), and propagating the identical address into the
>>  function callee.
>> - The pass could also potentially implement speculative
devirtualization
>>  (i.e. testing the vtable entry and guarding direct calls on it) by
>>  generating a comparison against the known address point. The pass will
>>  know exactly how many possible classes there will be for each call
site,
>>  so we could have logic to speculatively devirtualize if that number is
>>  sufficiently small.
>> 
>> The design when CFI is enabled is slightly different, as we need to
give
>> the compiler the ability to eliminate the unnecessary llvm.bitset.test
>> call. Recall that the IR looks like this:
>> 
>> %vtable = load i8**, i8*** %obj
>> %p = call i1 @llvm.bitset.test(%vtable, “Base”)
>> br i1 %p, label %call, label %abort
>> 
>> call:
>> %fptr = load i8*, i8** %vtable
>> %fptr_casted = bitcast i8% %fptr to i1 (i8***, i32)
>> call i1 %fptr_casted(i8*** %obj, i32 0)
>> 
>> abort:
>> call void @llvm.trap()
>> 
>> We cannot simply have our optimization pass transform this IR to
eliminate the
>> llvm.bitset.test call, as the transformation would not be
semantics-preserving
>> (we already have a CFI setting that emits diagnostics in the abort
branch,
>> which would break under such a transformation). Instead, we can
introduce
>> a new intrinsic, @llvm.bitset.check, to be used only in production mode
>> (i.e. with diagnostics disabled):
>> 
>> {i8*, i1} @llvm.bitset.check(i8* %ptr, metadata %name)
>> 
>> If %ptr is in the bitset identified by %name, the function returns
{%ptr,
>> true}. But if %ptr is not in %name, the behaviour is undefined, modulo
that
>> if the second element of the result is non-zero, the program may load
and
>> call a function pointer from an address derived from the first element
of the
>> result without causing an indirect function call to any function other
than
>> one potentially loaded from one of the constant globals of which %name
is
>> a member (i.e. one of the valid vtables for %name). A CFI-protected
virtual
>> call would then look like this (eliding bitcasts for brevity):
>> 
>> %vtable = load i8**, i8*** %obj
>> %pair = call {i8*, i1} @llvm.bitset.check(i8* %vtable, “Base”)
>> %checked_vtable = extractelement %pair, 0
>> %p = extractelement %pair, 1
>> br i1 %p, label %call, label %abort
>> 
>> call:
>> %fptr = load i8*, i8** %checked_vtable
>> %result = call i1 %fptr(i8*** %obj, i32 0)
>> 
>> abort:
>> call void @llvm.trap()
>> 
>> Applying virtual constant propagation to @llvm.bitset.check would be
very
>> similar to applying it to @llvm.bitset.test, but if the constant
propagation
>> succeeds, the function will return true as the second element:
>> 
>> %vtable = load i8**, i8*** %obj
>> br i1 true, label %call, label %abort
>> 
>> call:
>> %vtable_bit_addr = getelementptr i8* %vtable, i64 -17
>> %vtable_bit = load i8 %vtable_bit_addr
>> %vtable_bit_and = and i8 %vtable_bit, 1
>> %result = icmp ne %vtable_bit_and, 0
>> 
>> abort:
>> call void @llvm.trap()
>> 
>> Now the existing SimplifyCFG pass can easily eliminate the unnecessary
branching.
>> 
>> In order to allow for regular constant folding through the
llvm.bitset.check
>> intrinsic, the code in lib/Analysis/ConstantFolding.cpp would be taught
to
>> look through the intrinsic’s argument.
>> 
>> -- 
>> Peter
>> 
>> [0] https://code.google.com/p/chromium/issues/detail?id=580389
>> [1] http://llvm.org/docs/BitSets.html
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> 
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

Hans wrote:
> (and start-up time if we can drop the vtables and
> void the dynamic relocations).
On Thu, Jan 28, 2016 at 09:15:05AM -0800, Mehdi Amini
wrote:
> Hi,
> 
> I just thought about another use case: VTable compression.
> If you know that an entry in the Vtable is never used, just remove it!
> I’d hope we could even eliminate some unused virtual functions from the
final binary.
We could most likely extend the design to do both of these things, but we'd
need to be a little careful as these would both be ABI breaks, so we'd
probably
neeed some separate way of saying "I promise not to call virtual functions
from these classes" as opposed to "I promise not to extend these
classes".
> 
> — 
> Mehdi
> 
> 
> > On Jan 27, 2016, at 10:29 PM, Mehdi Amini via llvm-dev <llvm-dev at
lists.llvm.org> wrote:
> > 
> > Hi Peter,
> > 
> > Pete (Cooper, CC'ed) had a similar idea a few months ago about
devirtualization and LTO: we can know the set of possible overload based on the
knowledge of a closed type hierarchy. He worked on a prototype, and I continued
to evolve it. I planned to submit something on the mailing list but couldn’t fit
it in my planning before April :)
> > Our plan was about removing virtual call in general, or maybe in some
cases turning the indirect call into a switch table.
> > 
> > In both use case (your constant propagation and devirtualization), I
think the interesting part is “how to figure the set of possible callees for a
call site” and how we encode this in the IR.
> > 
> > I haven’t worked on this for a few weeks, but I’ll try to give a rough
description of where I stopped on my prototype:
> > I was storing the full inheritance tree in metadata. There was one
metadata per entry of the VTable, and any method that override one of the base
class will have a metadata attached pointing to the slot in the base class.
> > The front-end is modified to emit any load of a VTable through a new
intrinsic llvm.read.vtable that takes a pointer to the VTable, the index of the
slot in the VTable, and a pointer to the metadata that describe the slot for the
current type. Using the metadata you access directly to any override and
construct naturally and very easily the set of possible overloads. The pass that
transform the IR is quite easy to implement since you don’t need to walk the IR
but access directly all the uses of the llvm.read.vtable intrinsic and have all
the needed information in the operands.
> > I haven’t give much thought about your representation yet to know how
it compares (and I’m not very familiar with how CFI works), but I’m interested
in your feedback!
> > 
> > I was limited during my experiment by the lack of black/white list to
define what hierarchy is/isn’t closed, I’m glad if such flag/control would be
added!
Thanks for sharing that design, but I think it would be best to have a single
representation of the type information in the tree. The bitset representation
has the advantage of being proven with other features like CFI, while I'm
not sure that a design that worked at the virtual function level would work
well with CFI. (It also includes the insight that you don't need to store
an inheritance graph or anything like that, just the valid address points
for each vtable).

I did think about having something like your llvm.load.vtable intrinsic to
load function pointers, but I thought it would be more orthogonal to extend
existing mechanisms to load the vtable than to have a special type of load
that only works with vtables, to avoid having to teach passes about this
new type of load. Did you find this to be an issue in practice?

Now that I think about it, I believe that this sort of intrinsic could make
ABI-breaking vtable transformations like the ones above easier to implement
(there could be some IR module flag that means "I promise not to load from
vtables except via this intrinsic" that could be used to turn on those
transformations).

I had another look at http://llvm.org/docs/BitSets.html and it occurred
to me that it doesn't clearly explain how to use bitsets to encode type
information. I can certainly take another pass over that page to see if
I can improve the documentation there as we start to use bitsets in more
places in the compiler.

Thanks,
-- 
Peter

Hi Peter,

Thanks for your answer!
> On Jan 28, 2016, at 10:17 AM, Peter Collingbourne <peter at
pcc.me.uk> wrote:
> 
> Hans wrote:
>> (and start-up time if we can drop the vtables and
>> void the dynamic relocations).
> 
> On Thu, Jan 28, 2016 at 09:15:05AM -0800, Mehdi Amini wrote:
>> Hi,
>> 
>> I just thought about another use case: VTable compression.
>> If you know that an entry in the Vtable is never used, just remove it!
>> I’d hope we could even eliminate some unused virtual functions from the
final binary.
> 
> We could most likely extend the design to do both of these things, but
we'd
> need to be a little careful as these would both be ABI breaks, so we'd
probably
> neeed some separate way of saying "I promise not to call virtual
functions
> from these classes" as opposed to "I promise not to extend these
classes”.
Mmmm, good point, that’s not at the same level as the other optimization :-/
> 
>> 
>> — 
>> Mehdi
>> 
>> 
>>> On Jan 27, 2016, at 10:29 PM, Mehdi Amini via llvm-dev <llvm-dev
at lists.llvm.org> wrote:
>>> 
>>> Hi Peter,
>>> 
>>> Pete (Cooper, CC'ed) had a similar idea a few months ago about
devirtualization and LTO: we can know the set of possible overload based on the
knowledge of a closed type hierarchy. He worked on a prototype, and I continued
to evolve it. I planned to submit something on the mailing list but couldn’t fit
it in my planning before April :)
>>> Our plan was about removing virtual call in general, or maybe in
some cases turning the indirect call into a switch table.
>>> 
>>> In both use case (your constant propagation and devirtualization),
I think the interesting part is “how to figure the set of possible callees for a
call site” and how we encode this in the IR.
>>> 
>>> I haven’t worked on this for a few weeks, but I’ll try to give a
rough description of where I stopped on my prototype:
>>> I was storing the full inheritance tree in metadata. There was one
metadata per entry of the VTable, and any method that override one of the base
class will have a metadata attached pointing to the slot in the base class.
>>> The front-end is modified to emit any load of a VTable through a
new intrinsic llvm.read.vtable that takes a pointer to the VTable, the index of
the slot in the VTable, and a pointer to the metadata that describe the slot for
the current type. Using the metadata you access directly to any override and
construct naturally and very easily the set of possible overloads. The pass that
transform the IR is quite easy to implement since you don’t need to walk the IR
but access directly all the uses of the llvm.read.vtable intrinsic and have all
the needed information in the operands.
>>> I haven’t give much thought about your representation yet to know
how it compares (and I’m not very familiar with how CFI works), but I’m
interested in your feedback!
>>> 
>>> I was limited during my experiment by the lack of black/white list
to define what hierarchy is/isn’t closed, I’m glad if such flag/control would be
added!
> 
> Thanks for sharing that design, but I think it would be best to have a
single
> representation of the type information in the tree. The bitset
representation
> has the advantage of being proven with other features like CFI, while
I'm
> not sure that a design that worked at the virtual function level would work
> well with CFI.
I agree with all of what you say, and I shared the design FYI only. No NIH
syndrome here, reinventing the wheel is just driven by ignorance :)
(I’m just curious about the pros/cons of both approach for my personal
information)

> (It also includes the insight that you don't need to store
> an inheritance graph or anything like that, just the valid address points
> for each vtable).
I’m not sure what you mean in this sentence.
> 
> I did think about having something like your llvm.load.vtable intrinsic to
> load function pointers, but I thought it would be more orthogonal to extend
> existing mechanisms to load the vtable than to have a special type of load
> that only works with vtables, to avoid having to teach passes about this
> new type of load. Did you find this to be an issue in practice?
Initially Pete had to thread special handling for this intrinsic everywhere till
instruction selection.
I changed the semantic of the intrinsic to be completely opaque, the codegen I
have looks like:

  %vtable = load void (%struct.Derived2*)**, void (%struct.Derived2*)*** %3,
align 8
  %vfn_ptr = getelementptr inbounds void (%struct.Derived2*)*, void
(%struct.Derived2*)** %vtable, i64 1
  %vfn = load void (%struct.Derived2*)*, void (%struct.Derived2*)** %vfn_ptr,
align 8
  call void %vfn(%struct.Derived2* %2)

to:

  %vtable = load void (%struct.Derived2*)**, void (%struct.Derived2*)*** %3,
align 8
  %vfn_ptr = getelementptr inbounds void (%struct.Derived2*)*, void
(%struct.Derived2*)** %vtable, i64 1
  %vfn = load void (%struct.Derived2*)*, void (%struct.Derived2*)** %vfn_ptr,
align 8

  %vfn_i8ptr = bitcast %struct.DerivedDerivedNonVirtual* %vfn to i8*
  %vtable_actual = bitcast [3 x i8*]* @_ZTV8Derived2 to i8*
  %vfn.devirt_i8ptr = call i8* @llvm.read.vtable(i8* %vfn_i8ptr, i32 0, metadata
!5, i8* %vtable_actual)
  %vfn.devirt = bitcast i8* %vfn.devirt_i8ptr to void (%struct.Derived2*)*

  call void %vfn.devirt(%struct.DerivedDerivedNonVirtual* %2)

The call to the llvm.read.vtable() is just here to bundle the %vfn we just read
out of the vtable with the slot number in the vtable, the metadata describing
the entry for this slot, and a reference to the VTable for the actual type (top
of the current hierarchy).

It is always valid to RAUW the intrinsic with its first argument, but the others
carry the information about the possible overload for this VTable slot.

> 
> Now that I think about it, I believe that this sort of intrinsic could make
> ABI-breaking vtable transformations like the ones above easier to implement
> (there could be some IR module flag that means "I promise not to load
from
> vtables except via this intrinsic" that could be used to turn on those
> transformations).
> 
> I had another look at http://llvm.org/docs/BitSets.html and it occurred
> to me that it doesn't clearly explain how to use bitsets to encode type
> information. I can certainly take another pass over that page to see if
> I can improve the documentation there as we start to use bitsets in more
> places in the compiler.
I’d love that if you could improve the docs! :)
The CFI implementation has been on my reading list for too long now…

Thanks,

— 
Mehdi