LLVM’s Attribute APIs need an overhaul.
Current problems
Firstly, thanks for looking at this. I took a look a couple of years ago but quickly fell in to patch hell with way too many patches all over the compiler!
I did manage to find a few of my changes, and I ultimately focused on specific APIs which looked expensive. For example, AttributeSet::addAttributes, or even just code which looks and feels like its examining internal state, like AttributeSet::getNumSlots().
LLVM’s Attribute APIs need an overhaul.
Current problems
First, testing for an attribute on an Argument is slow.
llvm::AttributeSet::getAttributes(int) consumed 2% of cycles while optimizing
llc during LTO. Our mid-level optimizations are constantly asking if a given
argument has some attribute (nonnull, dereferencable, etc), and this is
currently linear in the size of the function prototype. This should be constant
time.
My guess is that arguments having attributes is still a relatively new thing, so they have just outgrown the implementation. Back when this was written i don’t think we were nearly as aggressive about adding or querying things like non null on arguments so the cost of the implementation wasn’t really clear.
Adding and removing individual attributes is also inefficient. Every single
attribute addition or removal convenience method, for example, inserts a new
AttributeSet into a FoldingSet. AttributeFuncs::mergeAttributesForInlining is
written in terms of these APIs, and is therefore quite inefficient. This also
shows up during LTO. I don’t think it’s practical to remove these inherently
expensive APIs, but if we make AttributeBuilder easier to use, it will be a lot
easier to fix these problems as they come up.
I think the worst offender is AttributeSet::addAttributes where we have to build intermediate sets just to add them to a parent set.
If you completely flatten the list so that return, function, and arguments all have their own, as you suggest later, then I think you’ll just happen to fix most of the slow construction APIs. As you’ve mentioned AttrBuilder, that probably will handle most of the remaining construction cases efficiently.
Lastly, the Attribute APIs are hard to use because they do too much information
hiding. In particular, the choice to make AttributeSetNode an internal
implementation detail of lib/IR is problematic. This type describes all of the
attributes on an individual argument, return value, or function, which IPO
transforms often want. Today the getFnAttributes, getRetAttributes, and
getParamAttributes APIs find the relevant AttributeSetNode* and wrap it in a
new, uniqued AttributeListImpl. This requires callers to keep around the index
of the extracted attributes so they can look through the wrapper list. If we
make AttributeSetNode public, we can simplify a lot of IPO transform code.
Do you have a particular IPO spot where this looks bad today? I’d be interested in taking a look to see whether flattening them out just works here too?
Naming
The naming of today’s APIs is confusing. I’ll try to explain what the current
important classes are.
AttributeSet: This is a uniqued, ordered list of sets of attributes, and is
associated with a function or call prototype. It is stored on Function,
CallInst, and InvokeInst. It is a smart pointer value type that wraps
AttributeSetImpl, which contains the actual storage.AttributeSetImpl: The private implementation of AttributeSet. Owned by the
LLVMContext. Today this is a vector of pairs of attribute indices and
AttributeSetNode pointers.AttributeSetNode: This is an ordered, uniqued set of Attributes that might
apply to a single function, callee, return value, parameter, or argument. It
uses TrailingObjects to store the attributes, and until Jan 2016, tested for
attribute presence by a linear scan. Matthias Braum added a bitset to speed up
tests in r259251.AttributeBuilder: A mutable representation of an AttributeSetNode. Used for
efficiently building a collection of attributes before freezing it into an
AttributeSetNode.Attribute: Pointer wrapping an AttributeImpl.
AttributeImpl: Polymorphic base class of StringAttributeImpl,
EnumAttributeImpl, and IntAttributeImpl. Enums have the attribute kind,
integers have a uint64_t value, and strings have two StringRefs for the kind
and value.AttributeSet doesn’t seem like a good name to me. In the past it was called
AttrListPtr and PAListPtr. Today’s AttributeSetImpl was called
“ParameterAttributeList”, which is why we have “PAL” local variables. I’d like
to rename AttributeSet to just “AttributeList”. It’s a list of sets of
attributes that is parallel to some function prototype. I already have a patch
out for this here: https://reviews.llvm.org/D31102
Thats where PAL came from! I always wondered.
The natural second thing to do is to rename AttributeSetNode to AttributeSet, or
create a pointer wrapper type called AttributeSet and rename today’s
AttributeSetNode to AttributeSetImpl. It is inherently an ordered set of
attributes. I also propose to make this type public, as described earlier.
Whether as an intermediate step, or the final naming, this all seems fine to me.
Optimizations
Testing for presence of an attribute on a parameter should be fast. Today it is
linear in the size of the prototype. I propose that we change AttributeList to
make attributes randomly accessible by slot index, so that it stores a trailing
object array of AttributeSetNode*, or some equivalent type that makes it
efficient to test for attribute presence. I’ll do some benchmarking to show that
this change doesn’t measurably impact LTO memory usage.
At the very least here, it seems like we should have Argument’s store their argument number in the Value subclass data or something like that.
uint64_t Argument::getDereferenceableOrNullBytes() const {
assert(getType()->isPointerTy() &&
“Only pointers have dereferenceable bytes”);
return getParent()->getDereferenceableOrNullBytes(getArgNo()+1);
}
This is horrible once you realize that getArgNo() is a linear scan over the arguments then getDereferenceableOrNullBytes() is another linear scan over the slots. Worst case 2n just to check a simple value!
I also want to reduce the number of loads needed to test a single attribute.
One potential optimization is to move the bitset of enum attributes up out of
AttributeSetNode* into a wrapper type, perhaps called AttributeSet. The wrapper
type could be a bitset and pointer pair, or we could try to union those together
in the usual way to save memory. I suspect the memory savings of the union are
unimportant, but I will measure. If we can always store the enumerated
attributes that have no associated size locally, that seems ideal, since it
allows modifying them without copying the immutable data stored in the LLVM
context.Another idea is to eliminate the use of AttributeList completely from Function,
and instead store the attributes of an Argument directly in the Argument. The
function and return attributes would live on the Function. At the end of the
day,Arg->hasNonNullAttr()should compile to a load and a bittest. This is a
more invasive change that would require more API migration, but again it would
avoid more FoldingSet usage.
You may get lucky in terms of this change as most of the queries are already on the argument or the function and can be changed internally to their hasNonNullAttr (or whatever) methods.
Any of the approaches here sounds fine, and its looks like you’re going to measure anyway. I think you’ve already observed that we query these far more than we construct them, so i’d go for the ‘load and check a bit’ approach. We want that query to be as fast as possible. Even if it means putting a pointer on every Argument, I expect we have a high enough proportion of Argument’s with attributes at this point that its likely not too bad.
The last two optimizations taken together are particularly nice, because it
means that changing enum attributes on functions and arguments won’t trigger
copies of immutable objects. It will only involve flipping bits in a locally
stored bitset.Separately, I want to collapse dereferenceable(N), nonnull, and
dereferenceable_or_null(N) for efficiency and orthogonality. I will probably
raise this in a separate RFC, but I think it was a mistake to have to attribute
imply nonnull. I’m mostly looking at this from an efficiency and redundancy
perspective, though, not a semantic one.
Does this all seem reasonable? Speak up now if you think any of this looks like
a step in the wrong direction.
Seems good to me.
Thanks again for doing this!
Pete
My guess is that arguments having attributes is still a relatively new
thing, so they have just outgrown the implementation. Back when this was
written i don’t think we were nearly as aggressive about adding or querying
things like non null on arguments so the cost of the implementation wasn’t
really clear.
Well, we've had argument attributes forever, but yes, they have become much
more common in the last few years now that we use them for more than just
ABI notes.
I think the worst offender is AttributeSet::addAttributes where we have to
build intermediate sets just to add them to a parent set.
If you completely flatten the list so that return, function, and arguments
all have their own, as you suggest later, then I think you’ll just happen
to fix most of the slow construction APIs. As you’ve mentioned
AttrBuilder, that probably will handle most of the remaining construction
cases efficiently.
The addAttributes implementation is obviously very inefficient, but I
suspect it's not on the hot path.
Do you have a particular IPO spot where this looks bad today? I’d be
interested in taking a look to see whether flattening them out just works
here too?
I don't think IPO pass manipulation of attributes is a hotspot, what's hot
is really just instcombine asking if a value is nonnull. This paragraph was
mostly about making IPO code more readable by changing the
AttributeList::get() API to take a list of AttributeSetNodes instead of a
list of pairs of slot numbers and AttributeSetNodes. DeadArgElim in
particular hacks on AttributeLists, and it's pretty ugly.
Thats where PAL came from! I always wondered.
Yep. 
Testing for presence of an attribute on a parameter should be fast. Today
it is
linear in the size of the prototype. I propose that we change
AttributeList to
make attributes randomly accessible by slot index, so that it stores a
trailing
object array of AttributeSetNode*, or some equivalent type that makes it
efficient to test for attribute presence. I'll do some benchmarking to
show that
this change doesn't measurably impact LTO memory usage.At the very least here, it seems like we should have Argument’s store
their argument number in the Value subclass data or something like that.uint64_t Argument::getDereferenceableOrNullBytes() const {
assert(getType()->isPointerTy() &&
"Only pointers have dereferenceable bytes");
return getParent()->getDereferenceableOrNullBytes(getArgNo()+1);
}This is horrible once you realize that getArgNo() is a linear scan over
the arguments then getDereferenceableOrNullBytes() is another linear scan
over the slots. Worst case 2n just to check a simple value!
I actually fixed getArgNo() first, that seemed uncontroversial.
Arguments store their argument number now, so getArgNo() is just a load. We
could go further and eliminate that field and make it two loads and some
pointer arithmetic as Sanjoy pointed out, but one load seems better than
two to me.
You may get lucky in terms of this change as most of the queries are
already on the argument or the function and can be changed internally to
their hasNonNullAttr (or whatever) methods.
Yep. That's the API that users seem to want anyway. All this "PAL" stuff is
just implementation details to most callers.
Any of the approaches here sounds fine, and its looks like you’re going to
measure anyway. I think you’ve already observed that we query these far
more than we construct them, so i’d go for the ‘load and check a bit’
approach. We want that query to be as fast as possible. Even if it means
putting a pointer on every Argument, I expect we have a high enough
proportion of Argument’s with attributes at this point that its likely not
too bad.
Great. I've also already changed Arguments to not be stored in an iplist in
function, so we have some extra memory headroom here.
Reid,
Thank for you taking a close look at this area. Attributes definitely need a good amount of love. As you point out, the naming of some of the code involved is more than a bit confusing and overly abstracted. However, there's also a delicate balance here in terms of having a generic understandable API. In general, I really like your proposal, a few comments inline.
LLVM's Attribute APIs need an overhaul.
Current problems
First, testing for an attribute on an Argument is slow.
llvm::AttributeSet::getAttributes(int) consumed 2% of cycles while optimizing
llc during LTO. Our mid-level optimizations are constantly asking if a given
argument has some attribute (nonnull, dereferencable, etc), and this is
currently linear in the size of the function prototype. This should be constant
time.Adding and removing individual attributes is also inefficient. Every single
attribute addition or removal convenience method, for example, inserts a new
AttributeSet into a FoldingSet. AttributeFuncs::mergeAttributesForInlining is
written in terms of these APIs, and is therefore quite inefficient. This also
shows up during LTO. I don't think it's practical to remove these inherently
expensive APIs, but if we make AttributeBuilder easier to use, it will be a lot
easier to fix these problems as they come up.
One observation I want to highlight is that enum attributes are the most common by far, but that string attributes are also widely used. Despite that, there are very few string attributes actually alive in a module at any one time. On the other hand, the presence of an attribute on one argument makes it much more likely that other arguments will also have that attribute. (Generally, we either attributed a function, or we didn't.) I don't have an exact design to suggest here, but I strongly suspect these observations could lead to an efficient bitset representation in the fastpath with a slower fallback path to handle the general case.
Lastly, the Attribute APIs are hard to use because they do too much information
hiding. In particular, the choice to make AttributeSetNode an internal
implementation detail of lib/IR is problematic. This type describes all of the
attributes on an individual argument, return value, or function, which IPO
transforms often want. Today the getFnAttributes, getRetAttributes, and
getParamAttributes APIs find the relevant AttributeSetNode* and wrap it in a
new, uniqued AttributeListImpl. This requires callers to keep around the index
of the extracted attributes so they can look through the wrapper list. If we
make AttributeSetNode public, we can simplify a lot of IPO transform code.Naming
The naming of today's APIs is confusing. I'll try to explain what the current
important classes are.- AttributeSet: This is a uniqued, ordered list of sets of attributes, and is
associated with a function or call prototype. It is stored on Function,
CallInst, and InvokeInst. It is a smart pointer value type that wraps
AttributeSetImpl, which contains the actual storage.- AttributeSetImpl: The private implementation of AttributeSet. Owned by the
LLVMContext. Today this is a vector of pairs of attribute indices and
AttributeSetNode pointers.- AttributeSetNode: This is an ordered, uniqued set of Attributes that might
apply to a single function, callee, return value, parameter, or argument. It
uses TrailingObjects to store the attributes, and until Jan 2016, tested for
attribute presence by a linear scan. Matthias Braum added a bitset to speed up
tests in r259251.- AttributeBuilder: A mutable representation of an AttributeSetNode. Used for
efficiently building a collection of attributes before freezing it into an
AttributeSetNode.- Attribute: Pointer wrapping an AttributeImpl.
- AttributeImpl: Polymorphic base class of StringAttributeImpl,
EnumAttributeImpl, and IntAttributeImpl. Enums have the attribute kind,
integers have a uint64_t value, and strings have two StringRefs for the kind
and value.AttributeSet doesn't seem like a good name to me. In the past it was called
AttrListPtr and PAListPtr. Today's AttributeSetImpl was called
"ParameterAttributeList", which is why we have "PAL" local variables. I'd like
to rename AttributeSet to just "AttributeList". It's a list of sets of
attributes that is parallel to some function prototype. I already have a patch
out for this here: ⚙ D31102 Rename AttributeSet to AttributeListThe natural second thing to do is to rename AttributeSetNode to AttributeSet, or
create a pointer wrapper type called AttributeSet and rename today's
AttributeSetNode to AttributeSetImpl. It is inherently an ordered set of
attributes. I also propose to make this type public, as described earlier.
SGTM. From an API design perspective, I'd absolutely love to see AttributeSet die. The fact that it represents both the set of attributes on a particular operand and the set of attributes across *all* operands is needlessly confusing. (Note: only in the API, the implementation is a bit more sane.) I really like your suggestion to rename the current AttributeSet to something like AttributeList and publishing the existing AttributeSetNode. Once we've done that, we can update the bulk add/remove APIs to be sane. The current semantics are an utter mess.
Optimizations
Testing for presence of an attribute on a parameter should be fast. Today it is
linear in the size of the prototype. I propose that we change AttributeList to
make attributes randomly accessible by slot index, so that it stores a trailing
object array of AttributeSetNode*, or some equivalent type that makes it
efficient to test for attribute presence. I'll do some benchmarking to show that
this change doesn't measurably impact LTO memory usage.
Sounds like a generally good idea. Maybe we can hide the indexing from the client though? Do something like store the index in the Value itself and have the AttributeSet (current name) interface be in terms of the Use? Not a strong requirement or anything, just thinking that might keep the interface a bit cleaner.
I also want to reduce the number of loads needed to test a single attribute.
One potential optimization is to move the bitset of enum attributes up out of
AttributeSetNode* into a wrapper type, perhaps called AttributeSet. The wrapper
type could be a bitset and pointer pair, or we could try to union those together
in the usual way to save memory. I suspect the memory savings of the union are
unimportant, but I will measure. If we can always store the enumerated
attributes that have no associated size locally, that seems ideal, since it
allows modifying them without copying the immutable data stored in the LLVM
context.
I got lost here. Might be good to land some of the renames and then restate this. I think we may be confusing meanings of AttributeSet?
Another idea is to eliminate the use of AttributeList completely from Function,
and instead store the attributes of an Argument directly *in* the Argument. The
function and return attributes would live on the Function. At the end of the
day, `Arg->hasNonNullAttr()` should compile to a load and a bittest. This is a
more invasive change that would require more API migration, but again it would
avoid more FoldingSet usage.
I particularly like this. It's worth pointing out that we can do the same for return values of Invokes and Calls, but that handling parameter attributes is much harder. This is also the point where I'd point out that metadata and return attributes have grown to heavily overlap in semantic with distinct sets of instructions. Maybe it's time to common the implementation, even if the interfaces stay separate? Once we'd done that, we'd have three cases: return attributes/metadata on all instructions, parameter attributes on Call and Invoke, and Argument attributes on Functions.
Getting back to this after a two week hiatus...
Reid,
Thank for you taking a close look at this area. Attributes definitely
need a good amount of love. As you point out, the naming of some of the
code involved is more than a bit confusing and overly abstracted. However,
there's also a delicate balance here in terms of having a generic
understandable API. In general, I really like your proposal, a few
comments inline.
Thanks!
The natural second thing to do is to rename AttributeSetNode to
AttributeSet, or
create a pointer wrapper type called AttributeSet and rename today's
AttributeSetNode to AttributeSetImpl. It is inherently an ordered set of
attributes. I also propose to make this type public, as described earlier.SGTM. From an API design perspective, I'd absolutely love to see
AttributeSet die. The fact that it represents both the set of attributes
on a particular operand and the set of attributes across *all* operands is
needlessly confusing. (Note: only in the API, the implementation is a bit
more sane.) I really like your suggestion to rename the current
AttributeSet to something like AttributeList and publishing the existing
AttributeSetNode. Once we've done that, we can update the bulk add/remove
APIs to be sane. The current semantics are an utter mess.
Yep. AttributeLists shouldn't be a list of more AttributeLists. I've got a
patch out which tries to untangle that a little bit more:
https://reviews.llvm.org/D31198
With that patch, getParamAttributes(unsigned) returns an AttributeSetNode*
instead of an AttributeList.
Optimizations
Testing for presence of an attribute on a parameter should be fast. Today
it is
linear in the size of the prototype. I propose that we change
AttributeList to
make attributes randomly accessible by slot index, so that it stores a
trailing
object array of AttributeSetNode*, or some equivalent type that makes it
efficient to test for attribute presence. I'll do some benchmarking to
show that
this change doesn't measurably impact LTO memory usage.Sounds like a generally good idea. Maybe we can hide the indexing from
the client though? Do something like store the index in the Value itself
and have the AttributeSet (current name) interface be in terms of the Use?
Not a strong requirement or anything, just thinking that might keep the
interface a bit cleaner.
Huh, I hadn't considered building the API in terms of Use& or Use*. That
would greatly reduce the need for all these extra integer induction
variables and let us use a lot more range-based for loops. I think we can
compute operand # from Use* in constant time, right? Thanks for the idea!
I also want to reduce the number of loads needed to test a single
attribute.
One potential optimization is to move the bitset of enum attributes up
out of
AttributeSetNode* into a wrapper type, perhaps called AttributeSet. The
wrapper
type could be a bitset and pointer pair, or we could try to union those
together
in the usual way to save memory. I suspect the memory savings of the
union are
unimportant, but I will measure. If we can always store the enumerated
attributes that have no associated size locally, that seems ideal, since
it
allows modifying them without copying the immutable data stored in the
LLVM
context.I got lost here. Might be good to land some of the renames and then
restate this. I think we may be confusing meanings of AttributeSet?
Yeah, I was trying to describe a *new* AttributeSet type. It would probably
look something like:
class AttributeSet {
uint64_t EnumAttrs; // bitset of that can answer hasAttr with a load and
test
AttributeSetNode *SlowAttrs; // string and integer attributes
...
};
This would be a trivially copyable value type, and would describe the
attributes of a single function, arg, or ret. AttributeList would become an
array of these. Adding an enum attribute is very cheap now: copy 16 (or 12)
bytes of data and set a bit.
Another idea is to eliminate the use of AttributeList completely from
Function,
and instead store the attributes of an Argument directly *in* the
Argument. The
function and return attributes would live on the Function. At the end of
the
day, `Arg->hasNonNullAttr()` should compile to a load and a bittest. This
is a
more invasive change that would require more API migration, but again it
would
avoid more FoldingSet usage.I particularly like this. It's worth pointing out that we can do the same
for return values of Invokes and Calls, but that handling parameter
attributes is much harder. This is also the point where I'd point out that
metadata and return attributes have grown to heavily overlap in semantic
with distinct sets of instructions. Maybe it's time to common the
implementation, even if the interfaces stay separate? Once we'd done that,
we'd have three cases: return attributes/metadata on all instructions,
parameter attributes on Call and Invoke, and Argument attributes on
Functions.
As nice as this is, I think I've convinced myself to stop short of doing
this. It's nice if Function, CallInst, and InvokeInst all use the same data
structures to store attributes. As long as Arg->hasAttr(Enum) is O(1), I
think we've won. Moving the data out of the parallel array and into
Argument probably isn't necessary.