bits and extract big bits from packed bits
Previously I tried to improve the size of .pcm files by introducing
packed bits. And I find we can improve it further by reordering the
bits.
The secret comes from the VBR format. We can find the formal definition
of VBR format in the doc of LLVM. The VBR format will be pretty
efficicent for small numbers.
For example, if we need to pack 8 bits into a value and the stored value
is 0xf0, the actual stored value will be 0b000111'110000, which takes 12
bits actually. However, if we changed the order to be 0x0f, then we
can store it as 0b001111, which takes 6 bits only now.
So we can improve the size by placing bits with lower probability to be
1 in the higher bits and extract bit bigs from the packed bits to make
it possible to be optimized by VBR.
After this patch, the size of std module becomes to 27.7MB from 28.1MB.
This patch tries to pack more bits into a value to reduce the size of
.pcm files. Also, after we introduced BitsPackers, it may slightly
better to adjust the way we use Abbrev.
After this patch, the size of the BMI for std module reduce from 28.94MB
to 28.1 MB.
This was reverted due to it broke the build of lldb. The reason that we
skip the serialization of a source location incorrectly. And this patch
now fixes that.
This patch tries to pack more bits into a value to reduce the size of
.pcm files. Also, after we introduced BitsPackers, it may slightly
better to adjust the way we use Abbrev.
After this patch, the size of the BMI for std module reduce from 28.94MB
to 28.1 MB.
This patch tries to reduce the size of the BMIs by packing more bits
into an unsigned integer.
This patch was reverted due to buildbot failure report. But it should be
irrevelent after I took a double look. So I tried to recommit this NFC
change again.
Previously, the boolean values will occupy spaces that can contain
integers. It wastes the spaces especially if the boolean values are
serialized consecutively. The patch tries to pack such consecutive
boolean values (and enum values) so that we can save more spaces and so
the times.
Before the patch, we need 4.478s (in my machine) to build the std module
(https://libcxx.llvm.org/Modules.html) with 28712 bytes for size of the
BMI. After the patch, the time becomes to 4.374s and the size becomes to
27388 bytes for the size of the BMI.
This is intended to be a NFC patch.
This patch doesn't optimize all such cases. We can do it later after we
have consensus on this.
This patch introduces a new enumerator `Invalid = 0`, shifting other enumerators by +1. Contrary to how it might sound, this actually affirms status quo of how this enum is stored in `clang::Decl`:
```
/// If 0, we have not computed the linkage of this declaration.
/// Otherwise, it is the linkage + 1.
mutable unsigned CacheValidAndLinkage : 3;
```
This patch makes debuggers to not be mistaken about enumerator stored in this bit-field. It also converts `clang::Linkage` to a scoped enum.
During the recent refactoring (b120fe8d3288c4dca1b5427ca34839ce8833f71c) this enum was moved out of `RecordDecl`. During post-commit review it was found out that its association with `RecordDecl` should be expressed in the name.
This patch converts `LinkageSpecDecl::LanguageIDs` into scoped enum, and moves it to namespace scope, so that it can be forward-declared where required.
This patch moves `ObjCMethodDecl::ImplementationControl` to a DeclBase.h so that it's complete at the point where corresponsing bit-field is declared. This patch also converts it to a scoped enum `clang::ObjCImplementationControl`.
This patch moves `OMPDeclareReductionDecl::InitKind` to DeclBase.h, so that it's complete at the point where corresponding bit-field is declared. This patch also converts it to scoped enum named `OMPDeclareReductionInitKind`
This removes the `ClassScopeFunctionSpecializationDecl` `Decl` node, and
instead uses `DependentFunctionTemplateSpecializationInfo` to handle
such declarations. `DependentFunctionTemplateSpecializationInfo` is also
changed to store a `const ASTTemplateArgumentListInfo*` to be more in
line with `FunctionTemplateSpecializationInfo`.
This also changes `FunctionDecl::isFunctionTemplateSpecialization` to
return `true` for dependent specializations, and
`FunctionDecl::getTemplateSpecializationKind`/`FunctionDecl::getTemplateSpecializationKindForInstantiation`
to return `TSK_ExplicitSpecialization` for non-friend dependent
specializations (the same behavior as dependent class scope
`ClassTemplateSepcializationDecl` & `VarTemplateSepcializationDecl`).
While working on #68377 inspecting `Allocate()` calls, I found out that
there are couple of places where we forget to use placement-new to
create objects in the allocated memory.
This patch adds a concept AST node (`ConceptLoc`) and uses it at the corresponding places.
There are three objects that might have constraints via concepts:
`TypeConstraint`, `ConceptSpecializationExpr` and `AutoTypeLoc`.
The first two inherit from `ConceptReference` while the latter has
the information about a possible constraint directly stored in `AutoTypeLocInfo`. It would be nice if the concept information would be stored the same way in all three cases.
Moreover the current structure makes it difficult to deal with these concepts. For example in Clangd accessing the locations of constraints of a `AutoTypeLoc` can only be done with quite ugly hacks.
So we think that it makes sense to create a new AST node for such concepts.
In details we propose the following:
- Rename `ConceptReference` to `ConceptLoc` (or something else what is approriate)
and make it the new AST node.
- `TypeConstraint` and `ConceptSpecializationExpr` do not longer inherit from `ConceptReference` but store a pointer to a `ConceptLoc`.
- `AutoTypeLoc` stores a pointer to `ConceptLoc` instead of storing the concept info in `AutoTypeLocInfo`.
This patch implements a first version of this idea which compiles and where the existing tests pass.
To make this patch as small as possible we keep the existing member functions to access concept data. Later these can be replaced by directly calling the corresponding functions of the `ConceptLoc`s.
Differential Revision: https://reviews.llvm.org/D155858
member.
Previously, we wouldn't do this if the first member loaded is within a
definition that's added to a class via an update record, which happens
when template instantiation adds a class definition to a declaration
that was imported from an AST file.
This would lead to classes having member functions whose getParent
returned a class declaration that wasn't the primary definition, which
in turn caused the vtable builder to build broken vtables.
I don't yet have a reduced testcase for the wrong-code bug here, because
the setup required to get us into the broken state is very subtle, but
have confirmed that this fixes it.
Platform-specific language extensions often want to provide a way of
indicating that certain functions should be called in a different way,
compiled in a different way, or otherwise treated differently from a
“normal” function. Honoring these indications is often required for
correctness, rather being than an optimization/QoI thing.
If a function declaration has a property P that matters for correctness,
it will be ODR-incompatible with a function that does not have property P.
If a function type has a property P that affects the calling convention,
it will not be two-way compatible with a function type that does not
have property P. These properties therefore affect language semantics.
That in turn means that they cannot be treated as standard [[]]
attributes.
Until now, many of these properties have been specified using GNU-style
attributes instead. GNU attributes have traditionally been more lax
than standard attributes, with many of them having semantic meaning.
Examples include calling conventions and the vector_size attribute.
However, there is a big drawback to using GNU attributes for semantic
information: compilers that don't understand the attributes will
(by default) emit a warning rather than an error. They will go on to
compile the code as though the attributes weren't present, which will
inevitably lead to wrong code in most cases. For users who live
dangerously and disable the warning, this wrong code could even be
generated silently.
A more robust approach would be to specify the properties using
keywords, which older compilers would then reject. Some vendor-specific
extensions have already taken this approach. But traditionally, each
such keyword has been treated as a language extension in its own right.
This has three major drawbacks:
(1) The parsing rules need to be kept up-to-date as the language evolves.
(2) There are often corner cases that similar extensions handle differently.
(3) Each extension requires more custom code than a standard attribute.
The underlying problem for all three is that, unlike for true attributes,
there is no established template that extensions can reuse. The purpose
of this patch series is to try to provide such a template.
One option would have been to pick an existing keyword and do whatever
that keyword does. The problem with that is that most keywords only
apply to specific kinds of types, kinds of decls, etc., and so the
parsing rules are (for good reason) not generally applicable to all
types and decls.
Really, the “only” thing wrong with using standard attributes is that
standard attributes cannot affect semantics. In all other respects
they provide exactly what we need: a well-defined grammar that evolves
with the language, clear rules about what an attribute appertains to,
and so on.
This series therefore adds keyword “attributes” that can appear
exactly where a standard attribute can appear and that appertain
to exactly what a standard attribute would appertain to. The link is
mechanical and no opt-outs or variations are allowed. This should
make the keywords predictable for programmers who are already
familiar with standard attributes.
This does mean that these keywords will be accepted for parsing purposes
in many more places than necessary. Inappropriate uses will then be
diagnosed during semantic analysis. However, the compiler would need
to reject the keywords in those positions whatever happens, and treating
them as ostensible attributes shouldn't be any worse than the alternative.
In some cases it might even be better. For example, SME's
__arm_streaming attribute would make conceptual sense as a statement
attribute, so someone who takes a “try-it-and-see” approach might write:
__arm_streaming { …block-of-code…; }
In fact, we did consider supporting this originally. The reason for
rejecting it was that it was too difficult to implement, rather than
because it didn't make conceptual sense.
One slight disadvantage of the keyword-based approach is that it isn't
possible to use #pragma clang attribute with the keywords. Perhaps we
could add support for that in future, if it turns out to be useful.
For want of a better term, I've called the new attributes "regular"
keyword attributes (in the sense that their parsing is regular wrt
standard attributes), as opposed to "custom" keyword attributes that
have their own parsing rules.
This patch adds the Attr.td support for regular keyword attributes.
Adding an attribute with a RegularKeyword spelling causes tablegen
to define the associated tokens and to record that attributes created
with that syntax are regular keyword attributes rather than custom
keyword attributes.
A follow-on patch contains the main Parse and Sema support,
which is enabled automatically by the Attr.td definition.
Other notes:
* The series does not allow regular keyword attributes to take
arguments, but this could be added in future.
* I wondered about trying to use tablegen for
TypePrinter::printAttributedAfter too, but decided against it.
RegularKeyword is really a spelling-level classification rather
than an attribute-level classification, and in general, an attribute
could have both GNU and RegularKeyword spellings. In contrast,
printAttributedAfter is only given the attribute kind and the type
that results from applying the attribute. AFAIK, it doesn't have
access to the original attribute spelling. This means that some
attribute-specific or type-specific knowledge might be needed
to print the attribute in the best way.
* Generating the tokens automatically from Attr.td means that
pseudo's libgrammar does now depend on tablegen.
* The patch uses the SME __arm_streaming attribute as an example
for testing purposes. The attribute does not do anything at this
stage. Later SME-specific patches will add proper semantics for it,
and add other SME-related keyword attributes.
Differential Revision: https://reviews.llvm.org/D148700
Close https://github.com/llvm/llvm-project/issues/62796.
Previously, we didn't serialize the evaluated result for VarDecl. This
caused the compilation of template metaprogramming become slower than
expect. This patch fixes the issue.
This is a recommit tested with asan built clang.
Close https://github.com/llvm/llvm-project/issues/62796.
Previously, we didn't serialize the evaluated result for VarDecl. This
caused the compilation of template metaprogramming become slower than
expect. This patch fixes the issue.
AttributeCommonInfo::isAlignasAttribute() was used in one place:
isCXX11Attribute(). The intention was for isAlignasAttribute()
to return true for the C++ alignas keyword. However, as a FIXME
noted, the function also returned true for the C _Alignas keyword.
This meant that isCXX11Attribute() returned true for _Alignas as
well as for alignas.
AttributeCommonInfos are now always constructed with an
AttributeCommonInfo::Form. We can use that Form to convey whether
a keyword is alignas or not.
The patch uses 1 bit of an 8-bit hole in the current layout
of AttributeCommonInfo. This might not be the best long-term design,
but it should be easy to adapt the layout if necessary (that is,
if other uses are found for the spare bits).
I don't know of a way of testing this (other than grep -c FIXME)
Differential Revision: https://reviews.llvm.org/D148105
When constructing an attribute, the syntactic form was specified
using two arguments: an attribute-independent syntax type and an
attribute-specific spelling index. This patch replaces them with
a single argument.
In most cases, that's done using a new Form class that combines the
syntax and spelling into a single object. This has the minor benefit
of removing a couple of constructors. But the main purpose is to allow
additional information to be stored as well, beyond just the syntax and
spelling enums.
In the case of the attribute-specific Create and CreateImplicit
functions, the patch instead uses the attribute-specific spelling
enum. This helps to ensure that the syntax and spelling are
consistent with each other and with the Attr.td definition.
If a Create or CreateImplicit caller specified a syntax and
a spelling, the patch drops the syntax argument and keeps the
spelling. If the caller instead specified only a syntax
(so that the spelling was SpellingNotCalculated), the patch
simply drops the syntax argument.
There were two cases of the latter: TargetVersion and Weak.
TargetVersionAttrs were created with GNU syntax, which matches
their definition in Attr.td, but which is also the default.
WeakAttrs were created with Pragma syntax, which does not match
their definition in Attr.td. Dropping the argument switches
them to AS_GNU too (to match [GCC<"weak">]).
Differential Revision: https://reviews.llvm.org/D148102
The purpose of this patch and follow-on patches is to ensure that
AttributeCommonInfos always have a syntax that is appropriate for
their kind (i.e. that it matches one of the entries in Attr.td).
The attribute-specific Create and CreateImplicit methods had four
overloads, based on their tail arguments:
(1) no extra arguments
(2) an AttributeCommonInfo
(3) a SourceRange
(4) a SourceRange, a syntax, and (where necessary) a spelling
When (4) had a spelling argument, it defaulted to
SpellingNotCalculated.
One disadvantage of this was that (1) and (3) zero-initialized
the syntax field of the AttributeCommonInfo, which corresponds
to AS_GNU. But AS_GNU isn't always listed as a possibility
in Attr.td.
This patch therefore removes (1) and (3) and instead provides
the same functionality using default arguments on (4) (a bit
like the existing default argument for the spelling).
The default syntax is taken from the attribute's first valid
spelling.
Doing that raises the question: what should happen for attributes
like AlignNatural and CUDAInvalidTarget that are only ever created
implicitly, and so have no source-code manifestation at all?
The patch adds a new AS_Implicit "syntax" for that case.
The patch also removes the syntax argument for these attributes,
since the syntax must always be AS_Implicit.
For similar reasons, the patch removes the syntax argument if
there is exactly one valid spelling.
Doing this means that AttributeCommonInfo no longer needs the
single-argument constructors. It is always given a syntax instead.
Differential Revision: https://reviews.llvm.org/D148101
This is important to break deserialization cycles, where a lambda in a
default member initializer can refer to the field as its context
declaration, and the initializer of the field can refer back to the
lambda.
This is a follow-up to bc73ef0031b5, which applied the same fix to
variable declarations for the same reason.
Previously, distinct lambdas would get merged, and multiple definitions
of the same lambda would not get merged, because we attempted to
identify lambdas by their ordinal position within their lexical
DeclContext. This failed for lambdas within namespace-scope variables
and within variable templates, where the lexical position in the context
containing the variable didn't uniquely identify the lambda.
In this patch, we instead identify lambda closure types by index within
their context declaration, which does uniquely identify them in a way
that's consistent across modules.
This change causes a deserialization cycle between the type of a
variable with deduced type and a lambda appearing as the initializer of
the variable -- reading the variable's type requires reading and merging
the lambda, and reading the lambda requires reading and merging the
variable. This is addressed by deferring loading the deduced type of a
variable until after we finish recursive deserialization.
This also exposes a pre-existing subtle issue where loading a
variable declaration would trigger immediate loading of its initializer,
which could recursively refer back to properties of the variable. This
particularly causes problems if the initializer contains a
lambda-expression, but can be problematic in general. That is addressed
by switching to lazily loading the initializers of variables rather than
always loading them with the variable declaration. As well as fixing a
deserialization cycle, that should improve laziness of deserialization
in general.
LambdaDefinitionData had 63 spare bits in it, presumably caused by an
off-by-one-error in some previous change. This change claims 32 of those bits
as a counter for the lambda within its context. We could probably move the
numbering to separate storage, like we do for the device-side mangling number,
to optimize the likely-common case where all three numbers (host-side mangling
number, device-side mangling number, and index within the context declaration)
are zero, but that's not done in this change.
Fixes#60985.
Reviewed By: #clang-language-wg, aaron.ballman
Differential Revision: https://reviews.llvm.org/D145737
A single class allows multiple categories to be defined for it. But if
two of such categories have the same name, we emit a warning. It's not a
hard error but a good indication of a potential mistake.
With modules, we can end up with the same category in different modules.
Diagnosing such a situation has little value as the categories in
different modules are equivalent and don't reflect the usage of the same
name for different purposes. When we deserialize a duplicate category,
compare it to an existing one and warn only when the new one is
different.
rdar://104582081
Differential Revision: https://reviews.llvm.org/D144149
When two modules contain interfaces with the same name, check the
definitions are equivalent and diagnose if they are not.
Differential Revision: https://reviews.llvm.org/D140073
When two modules contain struct/union with the same name, check the
definitions are equivalent and diagnose if they are not. This is similar
to `CXXRecordDecl` where we already discover and diagnose mismatches.
rdar://problem/56764293
Differential Revision: https://reviews.llvm.org/D71734
Close https://github.com/llvm/llvm-project/issues/59719.
The root cause of the problem is that we forgot to serialize a new
introduced bit to FunctionDeclBits. Maybe we need to find some methods
to work for detecting this.
This reverts commit b8064374b217db061213c561ec8f3376681ff9c8.
Based on the report here:
https://github.com/llvm/llvm-project/issues/59271
this produces a significant increase in memory use of the compiler and a
large compile-time regression. This patch reverts this so that we don't
branch for release with that issue.
Different versions of a lambda will in general refer to different
enclosing variable declarations, because we do not merge most
block-scope declarations, such as local variables. Keep track of all the
declarations that correspond to a lambda's capture fields so that we can
rewrite the name of any of those variables to the lambda capture,
regardless of which copy of the body of `operator()` we look at.
This patch teaches clang to parse statements on the global scope to allow:
```
./bin/clang-repl
clang-repl> int i = 12;
clang-repl> ++i;
clang-repl> extern "C" int printf(const char*,...);
clang-repl> printf("%d\n", i);
13
clang-repl> %quit
```
Generally, disambiguating between statements and declarations is a non-trivial
task for a C++ parser. The challenge is to allow both standard C++ to be
translated as if this patch does not exist and in the cases where the user typed
a statement to be executed as if it were in a function body.
Clang's Parser does pretty well in disambiguating between declarations and
expressions. We have added DisambiguatingWithExpression flag which allows us to
preserve the existing and optimized behavior where needed and implement the
extra rules for disambiguating. Only few cases require additional attention:
* Constructors/destructors -- Parser::isConstructorDeclarator was used in to
disambiguate between ctor-looking declarations and statements on the global
scope(eg. `Ns::f()`).
* The template keyword -- the template keyword can appear in both declarations
and statements. This patch considers the template keyword to be a declaration
starter which breaks a few cases in incremental mode which will be tackled
later.
* The inline (and similar) keyword -- looking at the first token in many cases
allows us to classify what is a declaration.
* Other language keywords and specifiers -- ObjC/ObjC++/OpenCL/OpenMP rely on
pragmas or special tokens which will be handled in subsequent patches.
The patch conceptually models a "top-level" statement into a TopLevelStmtDecl.
The TopLevelStmtDecl is lowered into a void function with no arguments.
We attach this function to the global initializer list to execute the statement
blocks in the correct order.
Differential revision: https://reviews.llvm.org/D127284
Adds support for NamespaceDecl to inform if its part of a nested namespace.
This flag only corresponds to the inner namespaces in a nested namespace declaration.
In this example:
namespace <X>::<Y>::<Z> {}
Only <Y> and <Z> will be classified as nested.
This flag isn't meant for assisting in building the AST, more for static analysis and refactorings.
Reviewed By: aaron.ballman
Differential Revision: https://reviews.llvm.org/D90568
Since we don't unique specializations for concepts, we can just instantiate
them with the sugared template arguments, at negligible cost.
If we don't track their specializations, we can't resugar them later
anyway, and that would be more expensive than just instantiating them
sugared in the first place since it would require an additional pass.
Signed-off-by: Matheus Izvekov <mizvekov@gmail.com>
Differential Revision: https://reviews.llvm.org/D136566
Since we don't unique specializations for concepts, we can just instantiate
them with the sugared template arguments, at negligible cost.
If we don't track their specializations, we can't resugar them later
anyway, and that would be more expensive than just instantiating them
sugared in the first place since it would require an additional pass.
Signed-off-by: Matheus Izvekov <mizvekov@gmail.com>
Differential Revision: https://reviews.llvm.org/D136566
This reverts commit cecc9a92cfca71c1b6c2a35c5e302ab649496d11.
The problem ended up being how we were handling the lambda-context in
code generation: we were assuming any decl context here would be a
named-decl, but that isn't the case. Instead, we just replace it with
the concept's owning context.
Differential Revision: https://reviews.llvm.org/D136451
This reverts commit b876f6e2f28779211a829d7d4e841fe68885ae20.
Still getting build failures on PPC AIX that aren't obvious what is causing
them, so reverting while I try to figure this out.
