The validateOutputSize and validateInputSize need to check whether
AVX or AVX512 are enabled. But this can be affected by the
target attribute so we need to factor that in.
This patch copies some of the code from CodeGen to create an
appropriate feature map that we can pass to the function. Probably
need some refactoring here to share more code with Codegen. Is
there a good place to do that? Also need to support the cpu_specific
attribute as well.
Differential Revision: https://reviews.llvm.org/D68627
Most of the functions emitted here should probably be convergent, but
only barriers are currently marked. Introduce this helper before
adding convergent to more functions.
Remove dead virtual functions from vtables with
replaceNonMetadataUsesWith, so that CGProfile metadata gets cleaned up
correctly.
Original commit message:
Currently, it is hard for the compiler to remove unused C++ virtual
functions, because they are all referenced from vtables, which are referenced
by constructors. This means that if the constructor is called from any live
code, then we keep every virtual function in the final link, even if there
are no call sites which can use it.
This patch allows unused virtual functions to be removed during LTO (and
regular compilation in limited circumstances) by using type metadata to match
virtual function call sites to the vtable slots they might load from. This
information can then be used in the global dead code elimination pass instead
of the references from vtables to virtual functions, to more accurately
determine which functions are reachable.
To make this transformation safe, I have changed clang's code-generation to
always load virtual function pointers using the llvm.type.checked.load
intrinsic, instead of regular load instructions. I originally tried writing
this using clang's existing code-generation, which uses the llvm.type.test
and llvm.assume intrinsics after doing a normal load. However, it is possible
for optimisations to obscure the relationship between the GEP, load and
llvm.type.test, causing GlobalDCE to fail to find virtual function call
sites.
The existing linkage and visibility types don't accurately describe the scope
in which a virtual call could be made which uses a given vtable. This is
wider than the visibility of the type itself, because a virtual function call
could be made using a more-visible base class. I've added a new
!vcall_visibility metadata type to represent this, described in
TypeMetadata.rst. The internalization pass and libLTO have been updated to
change this metadata when linking is performed.
This doesn't currently work with ThinLTO, because it needs to see every call
to llvm.type.checked.load in the linkage unit. It might be possible to
extend this optimisation to be able to use the ThinLTO summary, as was done
for devirtualization, but until then that combination is rejected in the
clang driver.
To test this, I've written a fuzzer which generates random C++ programs with
complex class inheritance graphs, and virtual functions called through object
and function pointers of different types. The programs are spread across
multiple translation units and DSOs to test the different visibility
restrictions.
I've also tried doing bootstrap builds of LLVM to test this. This isn't
ideal, because only classes in anonymous namespaces can be optimised with
-fvisibility=default, and some parts of LLVM (plugins and bugpoint) do not
work correctly with -fvisibility=hidden. However, there are only 12 test
failures when building with -fvisibility=hidden (and an unmodified compiler),
and this change does not cause any new failures for either value of
-fvisibility.
On the 7 C++ sub-benchmarks of SPEC2006, this gives a geomean code-size
reduction of ~6%, over a baseline compiled with "-O2 -flto
-fvisibility=hidden -fwhole-program-vtables". The best cases are reductions
of ~14% in 450.soplex and 483.xalancbmk, and there are no code size
increases.
I've also run this on a set of 8 mbed-os examples compiled for Armv7M, which
show a geomean size reduction of ~3%, again with no size increases.
I had hoped that this would have no effect on performance, which would allow
it to awlays be enabled (when using -fwhole-program-vtables). However, the
changes in clang to use the llvm.type.checked.load intrinsic are causing ~1%
performance regression in the C++ parts of SPEC2006. It should be possible to
recover some of this perf loss by teaching optimisations about the
llvm.type.checked.load intrinsic, which would make it worth turning this on
by default (though it's still dependent on -fwhole-program-vtables).
Differential revision: https://reviews.llvm.org/D63932
llvm-svn: 375094
Currently, it is hard for the compiler to remove unused C++ virtual
functions, because they are all referenced from vtables, which are referenced
by constructors. This means that if the constructor is called from any live
code, then we keep every virtual function in the final link, even if there
are no call sites which can use it.
This patch allows unused virtual functions to be removed during LTO (and
regular compilation in limited circumstances) by using type metadata to match
virtual function call sites to the vtable slots they might load from. This
information can then be used in the global dead code elimination pass instead
of the references from vtables to virtual functions, to more accurately
determine which functions are reachable.
To make this transformation safe, I have changed clang's code-generation to
always load virtual function pointers using the llvm.type.checked.load
intrinsic, instead of regular load instructions. I originally tried writing
this using clang's existing code-generation, which uses the llvm.type.test
and llvm.assume intrinsics after doing a normal load. However, it is possible
for optimisations to obscure the relationship between the GEP, load and
llvm.type.test, causing GlobalDCE to fail to find virtual function call
sites.
The existing linkage and visibility types don't accurately describe the scope
in which a virtual call could be made which uses a given vtable. This is
wider than the visibility of the type itself, because a virtual function call
could be made using a more-visible base class. I've added a new
!vcall_visibility metadata type to represent this, described in
TypeMetadata.rst. The internalization pass and libLTO have been updated to
change this metadata when linking is performed.
This doesn't currently work with ThinLTO, because it needs to see every call
to llvm.type.checked.load in the linkage unit. It might be possible to
extend this optimisation to be able to use the ThinLTO summary, as was done
for devirtualization, but until then that combination is rejected in the
clang driver.
To test this, I've written a fuzzer which generates random C++ programs with
complex class inheritance graphs, and virtual functions called through object
and function pointers of different types. The programs are spread across
multiple translation units and DSOs to test the different visibility
restrictions.
I've also tried doing bootstrap builds of LLVM to test this. This isn't
ideal, because only classes in anonymous namespaces can be optimised with
-fvisibility=default, and some parts of LLVM (plugins and bugpoint) do not
work correctly with -fvisibility=hidden. However, there are only 12 test
failures when building with -fvisibility=hidden (and an unmodified compiler),
and this change does not cause any new failures for either value of
-fvisibility.
On the 7 C++ sub-benchmarks of SPEC2006, this gives a geomean code-size
reduction of ~6%, over a baseline compiled with "-O2 -flto
-fvisibility=hidden -fwhole-program-vtables". The best cases are reductions
of ~14% in 450.soplex and 483.xalancbmk, and there are no code size
increases.
I've also run this on a set of 8 mbed-os examples compiled for Armv7M, which
show a geomean size reduction of ~3%, again with no size increases.
I had hoped that this would have no effect on performance, which would allow
it to awlays be enabled (when using -fwhole-program-vtables). However, the
changes in clang to use the llvm.type.checked.load intrinsic are causing ~1%
performance regression in the C++ parts of SPEC2006. It should be possible to
recover some of this perf loss by teaching optimisations about the
llvm.type.checked.load intrinsic, which would make it worth turning this on
by default (though it's still dependent on -fwhole-program-vtables).
Differential revision: https://reviews.llvm.org/D63932
llvm-svn: 374539
Summary:
When a variable is named in a context where we can't directly emit a
reference to it (because we don't know for sure that it's going to be
defined, or it's from an enclosing function and not captured, or the
reference might not "work" for some reason), we emit a copy of the
variable as a global and use that for the known-to-be-read-only access.
This reinstates r363295, reverted in r363352, with a fix for PR42276:
we now produce a proper name for a non-odr-use reference to a static
constexpr data member. The name <mangled-name>.const is used in that
case; such names are reserved to the implementation for cases such as
this and should demangle nicely.
Reviewers: rjmccall
Subscribers: jdoerfert, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D63157
llvm-svn: 363428
Revert 363340 "Remove unused SK_LValueToRValue initialization step."
Revert 363337 "PR23833, DR2140: an lvalue-to-rvalue conversion on a glvalue of type"
Revert 363295 "C++ DR712 and others: handle non-odr-use resulting from an lvalue-to-rvalue conversion applied to a member access or similar not-quite-trivial lvalue expression."
llvm-svn: 363352
Summary:
When a variable is named in a context where we can't directly emit a
reference to it (because we don't know for sure that it's going to be
defined, or it's from an enclosing function and not captured, or the
reference might not "work" for some reason), we emit a copy of the
variable as a global and use that for the known-to-be-read-only access.
Reviewers: rjmccall
Subscribers: jdoerfert, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D63157
llvm-svn: 363295
This patch implements a limited form of autolinking primarily designed to allow
either the --dependent-library compiler option, or "comment lib" pragmas (
https://docs.microsoft.com/en-us/cpp/preprocessor/comment-c-cpp?view=vs-2017) in
C/C++ e.g. #pragma comment(lib, "foo"), to cause an ELF linker to automatically
add the specified library to the link when processing the input file generated
by the compiler.
Currently this extension is unique to LLVM and LLD. However, care has been taken
to design this feature so that it could be supported by other ELF linkers.
The design goals were to provide:
- A simple linking model for developers to reason about.
- The ability to to override autolinking from the linker command line.
- Source code compatibility, where possible, with "comment lib" pragmas in other
environments (MSVC in particular).
Dependent library support is implemented differently for ELF platforms than on
the other platforms. Primarily this difference is that on ELF we pass the
dependent library specifiers directly to the linker without manipulating them.
This is in contrast to other platforms where they are mapped to a specific
linker option by the compiler. This difference is a result of the greater
variety of ELF linkers and the fact that ELF linkers tend to handle libraries in
a more complicated fashion than on other platforms. This forces us to defer
handling the specifiers to the linker.
In order to achieve a level of source code compatibility with other platforms
we have restricted this feature to work with libraries that meet the following
"reasonable" requirements:
1. There are no competing defined symbols in a given set of libraries, or
if they exist, the program owner doesn't care which is linked to their
program.
2. There may be circular dependencies between libraries.
The binary representation is a mergeable string section (SHF_MERGE,
SHF_STRINGS), called .deplibs, with custom type SHT_LLVM_DEPENDENT_LIBRARIES
(0x6fff4c04). The compiler forms this section by concatenating the arguments of
the "comment lib" pragmas and --dependent-library options in the order they are
encountered. Partial (-r, -Ur) links are handled by concatenating .deplibs
sections with the normal mergeable string section rules. As an example, #pragma
comment(lib, "foo") would result in:
.section ".deplibs","MS",@llvm_dependent_libraries,1
.asciz "foo"
For LTO, equivalent information to the contents of a the .deplibs section can be
retrieved by the LLD for bitcode input files.
LLD processes the dependent library specifiers in the following way:
1. Dependent libraries which are found from the specifiers in .deplibs sections
of relocatable object files are added when the linker decides to include that
file (which could itself be in a library) in the link. Dependent libraries
behave as if they were appended to the command line after all other options. As
a consequence the set of dependent libraries are searched last to resolve
symbols.
2. It is an error if a file cannot be found for a given specifier.
3. Any command line options in effect at the end of the command line parsing apply
to the dependent libraries, e.g. --whole-archive.
4. The linker tries to add a library or relocatable object file from each of the
strings in a .deplibs section by; first, handling the string as if it was
specified on the command line; second, by looking for the string in each of the
library search paths in turn; third, by looking for a lib<string>.a or
lib<string>.so (depending on the current mode of the linker) in each of the
library search paths.
5. A new command line option --no-dependent-libraries tells LLD to ignore the
dependent libraries.
Rationale for the above points:
1. Adding the dependent libraries last makes the process simple to understand
from a developers perspective. All linkers are able to implement this scheme.
2. Error-ing for libraries that are not found seems like better behavior than
failing the link during symbol resolution.
3. It seems useful for the user to be able to apply command line options which
will affect all of the dependent libraries. There is a potential problem of
surprise for developers, who might not realize that these options would apply
to these "invisible" input files; however, despite the potential for surprise,
this is easy for developers to reason about and gives developers the control
that they may require.
4. This algorithm takes into account all of the different ways that ELF linkers
find input files. The different search methods are tried by the linker in most
obvious to least obvious order.
5. I considered adding finer grained control over which dependent libraries were
ignored (e.g. MSVC has /nodefaultlib:<library>); however, I concluded that this
is not necessary: if finer control is required developers can fall back to using
the command line directly.
RFC thread: http://lists.llvm.org/pipermail/llvm-dev/2019-March/131004.html.
Differential Revision: https://reviews.llvm.org/D60274
llvm-svn: 360984
We need to be able to enqueue internal function that initializes
global constructors on the host side. Therefore it has to be
converted to a kernel.
This change factors out common logic for adding kernel metadata
and moves it from CodeGenFunction to CodeGenModule in order to
make it accessible for the extra use case.
Differential revision: https://reviews.llvm.org/D61488
llvm-svn: 360342
This provides a code size win on the caller side, since the init
message send is done in the runtime function.
rdar://44987038
Differential revision: https://reviews.llvm.org/D57936
llvm-svn: 354056
Emit{Nounwind,}RuntimeCall{,OrInvoke} have been modified to take a
FunctionCallee as an argument, and CreateRuntimeFunction has been
modified to return a FunctionCallee. All callers have been updated.
Additionally, CreateBuiltinFunction is removed, as it was redundant
with CreateRuntimeFunction after some previous changes.
Differential Revision: https://reviews.llvm.org/D57668
llvm-svn: 353184
This patch implements parsing and sema for "omp declare mapper"
directive. User defined mapper, i.e., declare mapper directive, is a new
feature in OpenMP 5.0. It is introduced to extend existing map clauses
for the purpose of simplifying the copy of complex data structures
between host and device (i.e., deep copy). An example is shown below:
struct S { int len; int *d; };
#pragma omp declare mapper(struct S s) map(s, s.d[0:s.len]) // Memory region that d points to is also mapped using this mapper.
Contributed-by: Lingda Li <lildmh@gmail.com>
Differential Revision: https://reviews.llvm.org/D56326
llvm-svn: 352906
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
It is faster to directly call the ObjC runtime for methods such as retain/release instead of sending a message to those functions.
Differential Revision: https://reviews.llvm.org/D55869
Reviewed By: rjmccall
llvm-svn: 349952
Implement options in clang to enable recording the driver command-line
in an ELF section.
Implement a new special named metadata, llvm.commandline, to support
frontends embedding their command-line options in IR/ASM/ELF.
This differs from the GCC implementation in some key ways:
* In GCC there is only one command-line possible per compilation-unit,
in LLVM it mirrors llvm.ident and multiple are allowed.
* In GCC individual options are separated by NULL bytes, in LLVM entire
command-lines are separated by NULL bytes. The advantage of the GCC
approach is to clearly delineate options in the face of embedded
spaces. The advantage of the LLVM approach is to support merging
multiple command-lines unambiguously, while handling embedded spaces
with escaping.
Differential Revision: https://reviews.llvm.org/D54487
Clang Differential Revision: https://reviews.llvm.org/D54489
llvm-svn: 349155
It is faster to directly call the ObjC runtime for methods such as alloc/allocWithZone instead of sending a message to those functions.
This patch adds support for converting messages to alloc/allocWithZone to their equivalent runtime calls.
Tests included for the positive case of applying this transformation, negative tests that we ensure we only convert "alloc" to objc_alloc, not "alloc2", and also a driver test to ensure we enable this only for supported runtime versions.
Reviewed By: rjmccall
https://reviews.llvm.org/D55349
llvm-svn: 348687
As suggested by Richard Smith, and initially put up for review here:
https://reviews.llvm.org/D53341, this patch removes a hack that was used
to ensure that proper target-feature lists were used when emitting
cpu-dispatch (and eventually, target-clones) implementations. As a part
of this, the GlobalDecl object is proliferated to a bunch more
locations.
Originally, this was put up for review (see above) to get acceptance on
the approach, though discussion with Richard in San Diego showed he
approved of the approach taken here. Thus, I believe this is acceptable
for Review-After-commit
Differential Revision: https://reviews.llvm.org/D53341
Change-Id: I0a0bd673340d334d93feac789d653e03d9f6b1d5
llvm-svn: 346757
Similar to how ICC handles CPU-Dispatch on Windows, this patch uses the
resolver function directly to forward the call to the proper function.
This is not nearly as efficient as IFuncs of course, but is still quite
useful for large functions specifically developed for certain
processors.
This is unfortunately still limited to x86, since it depends on
__builtin_cpu_supports and __builtin_cpu_is, which are x86 builtins.
The naming for the resolver/forwarding function for cpu-dispatch was
taken from ICC's implementation, which uses the unmodified name for this
(no mangling additions). This is possible, since cpu-dispatch uses '.A'
for the 'default' version.
In 'target' multiversioning, this function keeps the '.resolver'
extension in order to keep the default function keeping the default
mangling.
Change-Id: I4731555a39be26c7ad59a2d8fda6fa1a50f73284
Differential Revision: https://reviews.llvm.org/D53586
llvm-svn: 345298
Add support for OMP5.0 requires directive and unified_address clause.
Patches to follow will include support for additional clauses.
Differential Revision: https://reviews.llvm.org/D52359
llvm-svn: 343063
Previously the alignment on the newly created rtti/typeinfo data was largely
not set, meaning that DataLayout::getPreferredAlignment was free to overalign
it to 16 bytes. This causes unnecessary code bloat.
Differential Revision: https://reviews.llvm.org/D51416
llvm-svn: 342053
As documented here: https://software.intel.com/en-us/node/682969 and
https://software.intel.com/en-us/node/523346. cpu_dispatch multiversioning
is an ICC feature that provides for function multiversioning.
This feature is implemented with two attributes: First, cpu_specific,
which specifies the individual function versions. Second, cpu_dispatch,
which specifies the location of the resolver function and the list of
resolvable functions.
This is valuable since it provides a mechanism where the resolver's TU
can be specified in one location, and the individual implementions
each in their own translation units.
The goal of this patch is to be source-compatible with ICC, so this
implementation diverges from the ICC implementation in a few ways:
1- Linux x86/64 only: This implementation uses ifuncs in order to
properly dispatch functions. This is is a valuable performance benefit
over the ICC implementation. A future patch will be provided to enable
this feature on Windows, but it will obviously more closely fit ICC's
implementation.
2- CPU Identification functions: ICC uses a set of custom functions to identify
the feature list of the host processor. This patch uses the cpu_supports
functionality in order to better align with 'target' multiversioning.
1- cpu_dispatch function def/decl: ICC's cpu_dispatch requires that the function
marked cpu_dispatch be an empty definition. This patch supports that as well,
however declarations are also permitted, since the linker will solve the
issue of multiple emissions.
Differential Revision: https://reviews.llvm.org/D47474
llvm-svn: 337552
Similarly to CFI on virtual and indirect calls, this implementation
tries to use program type information to make the checks as precise
as possible. The basic way that it works is as follows, where `C`
is the name of the class being defined or the target of a call and
the function type is assumed to be `void()`.
For virtual calls:
- Attach type metadata to the addresses of function pointers in vtables
(not the functions themselves) of type `void (B::*)()` for each `B`
that is a recursive dynamic base class of `C`, including `C` itself.
This type metadata has an annotation that the type is for virtual
calls (to distinguish it from the non-virtual case).
- At the call site, check that the computed address of the function
pointer in the vtable has type `void (C::*)()`.
For non-virtual calls:
- Attach type metadata to each non-virtual member function whose address
can be taken with a member function pointer. The type of a function
in class `C` of type `void()` is each of the types `void (B::*)()`
where `B` is a most-base class of `C`. A most-base class of `C`
is defined as a recursive base class of `C`, including `C` itself,
that does not have any bases.
- At the call site, check that the function pointer has one of the types
`void (B::*)()` where `B` is a most-base class of `C`.
Differential Revision: https://reviews.llvm.org/D47567
llvm-svn: 335569
This is not only semantically correct but ensures that they will not
be marked as address-significant once D48155 lands.
Differential Revision: https://reviews.llvm.org/D48206
llvm-svn: 334982
Summary:
When requirement imposed by __target__ attributes on functions
are not satisfied, prefer printing those requirements, which
are explicitly mentioned in the attributes.
This makes such messages more useful, e.g. printing avx512f instead of avx2
in the following scenario:
```
$ cat foo.c
static inline void __attribute__((__always_inline__, __target__("avx512f")))
x(void)
{
}
int main(void)
{
x();
}
$ clang foo.c
foo.c:7:2: error: always_inline function 'x' requires target feature 'avx2', but would be inlined into function 'main' that is compiled without support for 'avx2'
x();
^
1 error generated.
```
bugzilla: https://bugs.llvm.org/show_bug.cgi?id=37338
Reviewers: craig.topper, echristo, dblaikie
Reviewed By: craig.topper, echristo
Differential Revision: https://reviews.llvm.org/D46541
llvm-svn: 334174
Some targets have constant address space (e.g. amdgcn). For them string literal should be
emitted in constant address space then casted to default address space.
Differential Revision: https://reviews.llvm.org/D46643
llvm-svn: 332279
This is similar to the LLVM change https://reviews.llvm.org/D46290.
We've been running doxygen with the autobrief option for a couple of
years now. This makes the \brief markers into our comments
redundant. Since they are a visual distraction and we don't want to
encourage more \brief markers in new code either, this patch removes
them all.
Patch produced by
for i in $(git grep -l '\@brief'); do perl -pi -e 's/\@brief //g' $i & done
for i in $(git grep -l '\\brief'); do perl -pi -e 's/\\brief //g' $i & done
Differential Revision: https://reviews.llvm.org/D46320
llvm-svn: 331834
register destructor functions annotated with __attribute__((destructor))
using __cxa_atexit or atexit.
Register destructor functions annotated with __attribute__((destructor))
calling __cxa_atexit in a synthesized constructor function instead of
emitting references to the functions in a special section.
The primary reason for adding this option is that we are planning to
deprecate the __mod_term_funcs section on Darwin in the future. This
feature is enabled by default only on Darwin. Users who do not want this
can use command line option 'fno_register_global_dtors_with_atexit' to
disable it.
rdar://problem/33887655
Differential Revision: https://reviews.llvm.org/D45578
llvm-svn: 330199
Found via codespell -q 3 -I ../clang-whitelist.txt
Where whitelist consists of:
archtype
cas
classs
checkk
compres
definit
frome
iff
inteval
ith
lod
methode
nd
optin
ot
pres
statics
te
thru
Patch by luzpaz! (This is a subset of D44188 that applies cleanly with a few
files that have dubious fixes reverted.)
Differential revision: https://reviews.llvm.org/D44188
llvm-svn: 329399
Summary:
The following class hierarchy requires that we be able to emit a
this-adjusting thunk for B::foo in C's vftable:
struct Incomplete;
struct A {
virtual A* foo(Incomplete p) = 0;
};
struct B : virtual A {
void foo(Incomplete p) override;
};
struct C : B { int c; };
This TU is valid, but lacks a definition of 'Incomplete', which makes it
hard to build a thunk for the final overrider, B::foo.
Before this change, Clang gives up attempting to emit the thunk, because
it assumes that if the parameter types are incomplete, it must be
emitting the thunk for optimization purposes. This is untrue for the MS
ABI, where the implementation of B::foo has no idea what thunks C's
vftable may require. Clang needs to emit the thunk without necessarily
having access to the complete prototype of foo.
This change makes Clang emit a musttail variadic call when it needs such
a thunk. I call these "unprototyped" thunks, because they only prototype
the "this" parameter, which must always come first in the MS C++ ABI.
These thunks work, but they create ugly LLVM IR. If the call to the
thunk is devirtualized, it will be a call to a bitcast of a function
pointer. Today, LLVM cannot inline through such a call, but I want to
address that soon, because we also use this pattern for virtual member
pointer thunks.
This change also implements an old FIXME in the code about reusing the
thunk's computed CGFunctionInfo as much as possible. Now we don't end up
computing the thunk's mangled name and arranging it's prototype up to
around three times.
Fixes PR25641
Reviewers: rjmccall, rsmith, hans
Subscribers: Prazek, cfe-commits
Differential Revision: https://reviews.llvm.org/D45112
llvm-svn: 329009
This is the next step in setting dso_local for COFF.
The patches changes setGVProperties to first set dllimport/dllexport
and changes a few cases that were setting dllimport/dllexport
manually. With this a few more GVs are marked dso_local.
llvm-svn: 326397
Reid Kleckner