[AssumeBundle] Add documentation for the operand bundles of an llvm.assume

[Tyker]

Summary

Operand bundles on an llvm.assume allows representing
assumptions that an attribute holds for a certain value at a certain position.
Operand bundles enable assumptions that are either hard or impossible to
represent as a boolean argument of an llvm.assume.

[jdoerfert]

Overall the test looks fine. I added a few comments, some rewording proposals and minor comments. We also need a proper commit message ;)

@fhahn Please take a look :)

llvm/docs/LangRef.rst

2113

Nit: Maybe say: as a boolean argument of an llvm.assume`.

2119

Nit: are no arguments, dot missing at the end.

2145

It's a bit abstract. Maybe something along the lines of:

Encoding attributes as operand bundles removes the need for an instruction sequence that represents the property, e.g., `icmp ne i32* %p, null` for `nonnull`. These sequences can
have negative effects due to the additional uses they introduce and the potential for them
to be changed or combined with surrounding user code.

2148

"make detecting that the value is used in an :ref:llvm.assume <int_assume> easier."
->
"makes it easy to identify uses in an :ref:llvm.assume <int_assume>. This then simplifies and improves heuristics, e.g., for use "use-sensitive" optimizations."

[Tyker]

tried to fix/improve formulations

[jdoerfert]

Looks better, I went over again and commented on minor language stuff. I'm not an expert either ;)

@fhahn what do you think?

llvm/docs/LangRef.rst

2113

allows representing -> allow representing

Operand bundles enable assumptions that are either hard or impossible to represent as a boolean argument of

2119

It's not the path arguably but the (single) location of the llvm.assume

2134

path -> call location

2136

are incorrect -> are violated at runtime
we might also reference the regular assume here.

2139

Even if the assumed property can be encoded as a boolean value like `nonnull`.

has benefits *or* can still have benefits

[lebedev.ri]

llvm/docs/LangRef.rst

2125–2128

It may be good to add a footnote that the arguments needn't be constants (or are they?)

[jdoerfert]

llvm/docs/LangRef.rst

2125–2128

I would prefer not to require that. As long as we come from llvm::Attribute that is the case but we derive information from other places (e.g., the user) eventually.

[jdoerfert]

llvm/docs/LangRef.rst

2125–2128

Didn't read it properly. Good idea to say explicitly they might not be constant! Thx @lebedev.ri

[fhahn]

Added some additional reviewers, I think it would be good to have a few more people taking a look.

llvm/docs/LangRef.rst

2115

I think this list needs a bit of context, like 'An assume operand bundle has the form "<tag>"(<holds for value>, <attribute argument>):

• ....

Maybe also link to the section about attributes and clarify that only valid attributes are supported.

2117

nit: second

also, what for attributes with more than 1 argument?

2119

Clarify which path?

2125

align isn't a string attribute I think. Shouldn't we use a non-string attribute here?

17591

nit: maybe say something like 'More complex assumptions can be encoded as :ref:operand bundles <assume_opbundles>.

17597–17598

Is this change needed?

[Tyker]

took comments in account.

llvm/docs/LangRef.rst

2115

yeah this is easier to read.

2117

they currently don't seem representable in IR by any of the *AttributeImpl.
do they exist ?

2125

align isn't a string attribute but tags in operand bundles need to be a string. so all attributes are encoded with there name as a strings.

2125–2128

i edited an example so it is not always a constant is it clear enough this way ?

2145

the use is mostly for the second bullet point. but i tried to reformulate to be more explicit.

that said maybe the 2 bullet point could be merged under there is not need for extra instructions. and then 2 bullet point on why extra expression are detrimental.

17597–17598

this was to avoid confusion about which argument. but maybe it isn't needed.

[jdoerfert]

I have some more minor nits (below) but other than that I'm still fine with this. I think we have all the relevant parts flushed out by now and minor adjustments can be made as we go.

llvm/docs/LangRef.rst

2119

", <attribute argument>" -> "[, <attribute argument>]" to make it optional

2125–2128

It's especially interesting for attributes that take only a constant in their attribute for, e.g., align. Maybe extend the example to show both non-constant values for those and how tags can occur multiple times:
call void @llvm.assume(i1 true) ["align"(i32* %val, i32 8), "align"(i32* %ptr, i32 %alignment)]

2145

. If -> , if

2149

"value like `nonnull. Using" -> "value , like nonnull`, using"

17597–17598

I think it is better this way, explicit is good.

[Tyker]

addressed comments.

[nlopes]

LGTM overall, just some nitpicks.
Something I would add is a link to the list of attributes (is this function-attributes?).

llvm/docs/LangRef.rst

2126

which call location?

2134

nitpick, maybe at this point -> from this point forward
since the assumptions hold for the rest of the program.

2141

which call location? The one that returns %cond or %val? Should be made explicit.

[jdoerfert]

In D74209#1876692, @nlopes wrote:

LGTM overall, just some nitpicks.
Something I would add is a link to the list of attributes (is this function-attributes?).

LGTM as well with some minor suggestions, and one clarification for @nlopes, below.
I think with the modifications we can go ahead and improve as we go.
[EDIT: Please improve the commit message, e.g., by copying part of the lang ref definition.]

The first paragraph has the word attribute. We can link to the two attribute lang ref sections (as we have both parameter and function attributes). Maybe:

... ​assumptions that a :ref:function attribute <fnattrs> or :ref:parameter attribute <paramattrs> holds ...

llvm/docs/LangRef.rst

2126

of the `llvm.assume` call location.

2134

nitpick, maybe at this point -> from this point forward
since the assumptions hold for the rest of the program.

Arguably not the ones that do not have an associated value, e.g., the cold() example below. All the ones that "bind" to a value do hold for the rest of the program though. But since that is "implied", and we have the other ones, I would not go there for now.

2141

Again, we need to say: the `llvm.assume` call location

[Tyker]

llvm/docs/LangRef.rst

2125–2128

it is possible to represent non-constant arguments using operand bundles. but the current query API doesn't handle that case. it is possible to deal with it but i think this will be at the cost of worse API and some performance.

i think this should be disallowed at least for now.

[Tyker]

addressed comments.

removed non-constant argument support as it isn't actually supported.

[jdoerfert]

In D74209#1879679, @Tyker wrote:

addressed comments.

removed non-constant argument support as it isn't actually supported.

Fine with me for now.

llvm/docs/LangRef.rst

 ``"gc-transition"`` operand bundle tag. These operand bundles mark a
 call as a transition between a function with one GC strategy to a
 function with a different GC strategy. If coordinating the transition
 between GC strategies requires additional code generation at the call
 site, these bundles may contain any values that are needed by the
 generated code.  For more details, see :ref:`GC Transitions
 <gc_transition_args>`.
 
+.. _assume_opbundles:
+
+Assume Operand Bundles
+^^^^^^^^^^^^^^^^^^^^^^
+
+Operand bundles on an :ref:`llvm.assume <int_assume>` allows representing
+assumptions that an attribute holds for a certain value at a certain position.
+Operand bundles enable assumptions that are either hard or impossible to
+represent as a boolean argument of an :ref:`llvm.assume <int_assume>`.
+
+An assume operand bundle has the form:
+
+::
+
+      "<tag>"([ <holds for value> [, <attribute argument>] ])
+
+* The tag of the operand bundle is the name of attribute that can be assumed
+  to hold.
+* The first argument if present is the value for which the attribute hold.
+* The second argument if present is an argument of the attribute.
+
+If there are no arguments the attribute is a property of the call location.
+
+For example:
+
+.. code-block:: llvm
+
+      call void @llvm.assume(i1 true) ["align"(i32* %val, i32 8), "align"(i32* %ptr, i32 %alignment)]
+
+allows the optimizer to assume that at this point ``%val`` has an alignment of
+at least 8 and ``%ptr`` has an alignment of at least ``%alignment``.
+
+.. code-block:: llvm
+
+      call void @llvm.assume(i1 %cond) ["cold"(), "nonnull"(i64* %val)]
+
+allows the optimizer to assume that the call location is cold and that ``%val``
+may not be null.
+
+Just like for the argument of :ref:`llvm.assume <int_assume>`, if any of the
+provided guarantees are are violated at runtime the behavior is undefined.
+
+Even if the assumed property can be encoded as a boolean value, like
+``nonnull``, using operand bundles to express the property can still have
+benefits:
+
+* Attributes that can be expressed via operand bundles are directly the
+  property that the optimizer uses and cares about. Encoding attributes as
+  operand bundles removes the need for an instruction sequence that represents
+  the property (e.g., `icmp ne i32* %p, null` for `nonnull`) and for the
+  optimizer to deduce the property from that instruction sequence.
+* Expressing the property using operand bundles makes it easy to identify the
+  use of the value as a use in an :ref:`llvm.assume <int_assume>`. This then
+  simplifies and improves heuristics, e.g., for use "use-sensitive"
+  optimizations.
+
 .. _moduleasm:
 
 Module-Level Inline Assembly
 ----------------------------
 
 Modules may contain "module-level inline asm" blocks, which corresponds
 to the GCC "file scope inline asm" blocks. These blocks are internally
 concatenated by LLVM and treated as a single unit, but may be separated

 
 Overview:
 """""""""
 
 The ``llvm.assume`` allows the optimizer to assume that the provided
 condition is true. This information can then be used in simplifying other parts
 of the code.
 
+More complex assumptions can be encoded as
+:ref:`assume operand bundles <assume_opbundles>`.
+
 Arguments:
 """"""""""
 
-The condition which the optimizer may assume is always true.
+The argument of the call is the condition which the optimizer may assume is
+always true.
 
 Semantics:
 """"""""""
 
 The intrinsic allows the optimizer to assume that the provided condition is
 always true whenever the control flow reaches the intrinsic call. No code is
 generated for this intrinsic, and instructions that contribute only to the
 provided condition are not used for code generation. If the condition is