Recently we observed high memory pressure caused by clang during some parallel builds.
We discovered that we have several projects that have a large number of #define directives
in their TUs (on the order of millions), which caused huge memory consumption in clang due
to a lot of allocations for MacroInfo. We would like to reduce the memory overhead of
clang for a single #define to reduce the memory overhead for these files, to allow us to
reduce the memory pressure on the system during highly parallel builds. This change achieves
that by removing the SmallVector in MacroInfo and instead storing the tokens in an array
allocated using the bump pointer allocator, after all tokens are lexed.
The added unit test with 1000000 #define directives illustrates the problem. Prior to this
change, on arm64 macOS, clang's PP bump pointer allocator allocated 272007616 bytes, and
used roughly 272 bytes per #define. After this change, clang's PP bump pointer allocator
allocates 120002016 bytes, and uses only roughly 120 bytes per #define.
For an example test file that we have internally with 7.8 million #define directives, this
change produces the following improvement on arm64 macOS: Persistent allocation footprint for
this test case file as it's being compiled to LLVM IR went down 22% from 5.28 GB to 4.07 GB
and the total allocations went down 14% from 8.26 GB to 7.05 GB. Furthermore, this change
reduced the total number of allocations made by the system for this clang invocation from
1454853 to 133663, an order of magnitude improvement.
Differential Revision: https://reviews.llvm.org/D117348
Masked reduction intrinsics are specical cases which don't need to have policy
operand. The mask only affects which elements are read. It doesn't effect the
destination register.
The reduction intrinsics have a dedicated destination operand. If it
is undef, we use tail agnostic. If it not undef we use tail
undisturbed.
Co-Authored-by: Craig Topper <craig.topper@sifive.com>
Differential Revision: https://reviews.llvm.org/D117681
OpenCL C 3.0 introduces optionality to some builtins, in particularly
to those which are conditionally supported with pipe, device enqueue
and generic address space features.
The idea is to conditionally support such builtins depending on the language options
being set for a certain feature. This allows users to define functions with names
of those optional builtins in OpenCL (as such names are not reserved).
Reviewed By: Anastasia
Differential Revision: https://reviews.llvm.org/D118605
The introduction and some examples are on this page:
https://devblogs.microsoft.com/cppblog/announcing-jmc-stepping-in-visual-studio/
The `/JMC` flag enables these instrumentations:
- Insert at the beginning of every function immediately after the prologue with
a call to `void __fastcall __CheckForDebuggerJustMyCode(unsigned char *JMC_flag)`.
The argument for `__CheckForDebuggerJustMyCode` is the address of a boolean
global variable (the global variable is initialized to 1) with the name
convention `__<hash>_<filename>`. All such global variables are placed in
the `.msvcjmc` section.
- The `<hash>` part of `__<hash>_<filename>` has a one-to-one mapping
with a directory path. MSVC uses some unknown hashing function. Here I
used DJB.
- Add a dummy/empty COMDAT function `__JustMyCode_Default`.
- Add `/alternatename:__CheckForDebuggerJustMyCode=__JustMyCode_Default` link
option via ".drectve" section. This is to prevent failure in
case `__CheckForDebuggerJustMyCode` is not provided during linking.
Implementation:
All the instrumentations are implemented in an IR codegen pass. The pass is placed immediately before CodeGenPrepare pass. This is to not interfere with mid-end optimizations and make the instrumentation target-independent (I'm still working on an ELF port in a separate patch).
Reviewed By: hans
Differential Revision: https://reviews.llvm.org/D118428
There is a clangd crash at `__memcmp_avx2_movbe`. Short problem description is below.
The method `HeaderIncludes::addExistingInclude` stores `Include` objects by reference at 2 places: `ExistingIncludes` (primary storage) and `IncludesByPriority` (pointer to the object's location at ExistingIncludes). `ExistingIncludes` is a map where value is a `SmallVector`. A new element is inserted by `push_back`. The operation might do resize. As result pointers stored at `IncludesByPriority` might become invalid.
Typical stack trace
```
frame #0: 0x00007f11460dcd94 libc.so.6`__memcmp_avx2_movbe + 308
frame #1: 0x00000000004782b8 clangd`llvm::StringRef::compareMemory(Lhs="
\"t2.h\"", Rhs="", Length=6) at StringRef.h:76:22
frame #2: 0x0000000000701253 clangd`llvm::StringRef::compare(this=0x0000
7f10de7d8610, RHS=(Data = "", Length = 7166742329480737377)) const at String
Ref.h:206:34
* frame #3: 0x00000000007603ab clangd`llvm::operator<(llvm::StringRef, llv
m::StringRef)(LHS=(Data = "\"t2.h\"", Length = 6), RHS=(Data = "", Length =
7166742329480737377)) at StringRef.h:907:23
frame #4: 0x0000000002d0ad9f clangd`clang::tooling::HeaderIncludes::inse
rt(this=0x00007f10de7fb1a0, IncludeName=(Data = "t2.h\"", Length = 4), IsAng
led=false) const at HeaderIncludes.cpp:365:22
frame #5: 0x00000000012ebfdd clangd`clang::clangd::IncludeInserter::inse
rt(this=0x00007f10de7fb148, VerbatimHeader=(Data = "\"t2.h\"", Length = 6))
const at Headers.cpp:262:70
```
A unit test test for the crash was created (`HeaderIncludesTest.RepeatedIncludes`). The proposed solution is to use std::list instead of llvm::SmallVector
Test Plan
```
./tools/clang/unittests/Tooling/ToolingTests --gtest_filter=HeaderIncludesTest.RepeatedIncludes
```
Reviewed By: sammccall
Differential Revision: https://reviews.llvm.org/D118755
we already accept "--target=". No reason to not accept "-target" too
(that's the one I typically use for some reason).
Differential revision: https://reviews.llvm.org/D119446
ASTReader
This is a cleanup to reduce the lines of code to handle default template
argument in ASTReader.
Reviewed By: urnathan
Differential Revision: https://reviews.llvm.org/D118437
LRGraph is the key component of the clang pseudo parser, it is a
deterministic handle-finding finite-state machine, which is used to
generated the LR parsing table.
Separate from https://reviews.llvm.org/D118196.
Differential Revision: https://reviews.llvm.org/D119172
This will allow moving the IncludeCleaner library essentials to Clang
and decoupling them from the majority of clangd.
The patch itself just moves the code, it doesn't change existing
functionality.
Reviewed By: sammccall
Differential Revision: https://reviews.llvm.org/D119130
This patch adds support for generating MLIR files in Flang's frontend
driver (i.e. `flang-new -fc1`). `-emit-fir` is added as an alias for
`-emit-mlir`. We may want to decide to split the two in the future.
A new parent class for code-gen frontend actions is introduced:
`CodeGenAction`. We will be using this class to encapsulate logic shared
between all code-generation actions, but not required otherwise. For
now, it will:
* run prescanning, parsing and semantic checks,
* lower the input to MLIR.
`EmitObjAction` is updated to inherit from this class. This means that
the behaviour of `flang-new -fc1 -emit-obj` is also updated (previously,
it would just exit immediately). This change required
`flang/test/Driver/syntax-only.f90` to be updated.
For `-emit-fir`, a specialisation of `CodeGenAction` is introduced:
`EmitMLIRAction`. The key logic for this class is implemented in
`EmitMLIRAction::ExecuteAction`.
Differential Revision: https://reviews.llvm.org/D118985
code object version determines ABI, therefore should not be mixed.
This patch emits amdgpu_code_object_version module flag in LLVM IR
based on code object version (default 4).
The amdgpu_code_object_version value is code object version times 100.
LLVM IR with different amdgpu_code_object_version module flag cannot
be linked.
The -cc1 option -mcode-object-version=none is for ROCm device library use
only, which supports multiple ABI.
Reviewed by: Artem Belevich
Differential Revision: https://reviews.llvm.org/D119026
The "-fzero-call-used-regs" option tells the compiler to zero out
certain registers before the function returns. It's also available as a
function attribute: zero_call_used_regs.
The two upper categories are:
- "used": Zero out used registers.
- "all": Zero out all registers, whether used or not.
The individual options are:
- "skip": Don't zero out any registers. This is the default.
- "used": Zero out all used registers.
- "used-arg": Zero out used registers that are used for arguments.
- "used-gpr": Zero out used registers that are GPRs.
- "used-gpr-arg": Zero out used GPRs that are used as arguments.
- "all": Zero out all registers.
- "all-arg": Zero out all registers used for arguments.
- "all-gpr": Zero out all GPRs.
- "all-gpr-arg": Zero out all GPRs used for arguments.
This is used to help mitigate Return-Oriented Programming exploits.
Reviewed By: nickdesaulniers
Differential Revision: https://reviews.llvm.org/D110869
The above change assumed that malloc (and friends) would always
allocate memory to getNewAlign(), even for allocations which have a
smaller size. This is not actually required by spec (a 1-byte
allocation may validly have 1-byte alignment).
Some real-world malloc implementations do not provide this guarantee,
and thus this optimization is breaking programs.
Fixes#53540
This reverts commit c2297544c0.
Differential Revision: https://reviews.llvm.org/D118804
The original warning added in D115501 when pacbti is used with an
incompatible architecture was not exactly correct because it was
not really ignored and can affect codegen.
Therefore reword to say that the pacbti option is incompatible with
the given architecture.
Reviewed By: chill
Differential Revision: https://reviews.llvm.org/D119166
These changes make the Clang parser recognize expression parameter pack
expansion and initializer lists in attribute arguments. Because
expression parameter pack expansion requires additional handling while
creating and instantiating templates, the support for them must be
explicitly supported through the AcceptsExprPack flag.
Handling expression pack expansions may require a delay to when the
arguments of an attribute are correctly populated. To this end,
attributes that are set to accept these - through setting the
AcceptsExprPack flag - will automatically have an additional variadic
expression argument member named DelayedArgs. This member is not
exposed the same way other arguments are but is set through the new
CreateWithDelayedArgs creator function generated for applicable
attributes.
To illustrate how to implement support for expression pack expansion
support, clang::annotate is made to support pack expansions. This is
done by making handleAnnotationAttr delay setting the actual attribute
arguments until after template instantiation if it was unable to
populate the arguments due to dependencies in the parsed expressions.
Implement P2128R6 in C++23 mode.
Unlike GCC's implementation, this doesn't try to recover when a user
meant to use a comma expression.
Because the syntax changes meaning in C++23, the patch is *NOT*
implemented as an extension. Instead, declaring an array with not
exactly 1 parameter is an error in older languages modes. There is an
off-by-default extension warning in C++23 mode.
Unlike the standard, we supports default arguments;
Ie, we assume, based on conversations in WG21, that the proposed
resolution to CWG2507 will be accepted.
We allow arrays OpenMP sections and C++23 multidimensional array to
coexist:
[a , b] multi dimensional array
[a : b] open mp section
[a, b: c] // error
The rest of the patch is relatively straight forward: we take care to
support an arbitrary number of arguments everywhere.
Among many FoldingSet users most notable seem to be ASTContext and CodeGenTypes.
The reasons that we spend not-so-tiny amount of time in FoldingSet calls from there, are following:
1. Default FoldingSet capacity for 2^6 items very often is not enough.
For PointerTypes/ElaboratedTypes/ParenTypes it's not unlikely to observe growing it to 256 or 512 items.
FunctionProtoTypes can easily exceed 1k items capacity growing up to 4k or even 8k size.
2. FoldingSetBase::GrowBucketCount cost itself is not very bad (pure reallocations are rather cheap thanks to BumpPtrAllocator).
What matters is high collision rate when lot of items end up in same bucket slowing down FoldingSetBase::FindNodeOrInsertPos and trashing CPU cache
(as items with same hash are organized in intrusive linked list which need to be traversed).
This change address both issues by increasing initial size of FoldingSets used in ASTContext and CodeGenTypes.
Extracted from: https://reviews.llvm.org/D118385
Differential Revision: https://reviews.llvm.org/D118608
D117898 added the generic __builtin_elementwise_add_sat and __builtin_elementwise_sub_sat with the same integer behaviour as the SSE/AVX instructions
This patch removes the __builtin_ia32_padd/psub saturated intrinsics and just uses the generics - the existing tests see no changes:
__m256i test_mm256_adds_epi8(__m256i a, __m256i b) {
// CHECK-LABEL: test_mm256_adds_epi8
// CHECK: call <32 x i8> @llvm.sadd.sat.v32i8(<32 x i8> %{{.*}}, <32 x i8> %{{.*}})
return _mm256_adds_epi8(a, b);
}
D117898 added the generic __builtin_elementwise_add_sat and __builtin_elementwise_sub_sat with the same integer behaviour as the SSE/AVX instructions
This patch removes the __builtin_ia32_padd/psub saturated intrinsics and just uses the generics - the existing tests see no changes:
__m256i test_mm256_adds_epi8(__m256i a, __m256i b) {
// CHECK-LABEL: test_mm256_adds_epi8
// CHECK: call <32 x i8> @llvm.sadd.sat.v32i8(<32 x i8> %{{.*}}, <32 x i8> %{{.*}})
return _mm256_adds_epi8(a, b);
}
Done in manner similar to mutexinoutset
(see https://reviews.llvm.org/D57576)
Runtime support already exists in LLVM OpenMP runtime (see
https://reviews.llvm.org/D97085).
The value used to identify an inoutset dependency type in the LLVM
OpenMP runtime is 8.
Some tests updated due to change in dependency type error messages that
now include new dependency type. Also updated
test/OpenMP/task_codegen.cpp to verify we emit the right code.
This patch implements `__builtin_elementwise_add_sat` and `__builtin_elementwise_sub_sat` builtins.
These map to the add/sub saturated math intrinsics described here:
https://llvm.org/docs/LangRef.html#saturation-arithmetic-intrinsics
With this in place we should then be able to replace the x86 SSE adds/subs intrinsics with these generic variants - it looks like other targets should be able to use these as well (arm/aarch64/webassembly all have similar examples in cgbuiltin).
Differential Revision: https://reviews.llvm.org/D117898
Since the serialization code would recognize modules by names and the
name of all global module fragment is <global>, so that the
serialization code would complain for the same module.
This patch fixes this by using a unique global module fragment in Sema.
Before this patch, the compiler would fail on an assertion complaining
the duplicated modules.
Reviewed By: urnathan, rsmith
Differential Revision: https://reviews.llvm.org/D115610
See the discussion in https://reviews.llvm.org/D100282. The coroutine
marked always inline might not be inlined properly in current compiler
support. Since the coroutine would be splitted into pieces. And the call
to resume() and destroy() functions might be indirect call. Also the
ramp function wouldn't get inlined under O0 due to pipeline ordering
problems. It might be different to what users expects to. Emit a warning
to tell it.
This is what GCC does too: https://godbolt.org/z/7eajb1Gf8
Reviewed By: Quuxplusone
Differential Revision: https://reviews.llvm.org/D115867
AArch32/Armv8A introduced the performance deprecation of certain patterns
of IT instructions. After some debate internal to ARM, this is now being
reverted; i.e. no IT instruction patterns are performance deprecated
anymore, as the perfomance degredation is not significant enough.
This reverts the following:
"ARMv8-A deprecates some uses of the T32 IT instruction. All uses of
IT that apply to instructions other than a single subsequent 16-bit
instruction from a restricted set are deprecated, as are explicit
references to the PC within that single 16-bit instruction. This permits
the non-deprecated forms of IT and subsequent instructions to be treated
as a single 32-bit conditional instruction."
The deprecation no longer applies, but the behaviour may be controlled
by the -arm-restrict-it and -arm-no-restrict-it command-line options,
with the latter being the default. No warnings about complex IT blocks
will be generated.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D118044
Introduced by 23a5090c6, some style option markers indicated 'clang-format 14',
though their respective options were available in earlier releases.
Note: Even though the value type of 'SpacesInAngles' option changed,
this option has been already present since version 3.4.
Differential Revision: https://reviews.llvm.org/D118991
Summary:
The name of the AMDGPU device library was changes. Previously it was
called 'libomptarget-amdgcn'. This patch changes fixes the tests to use
the new name of the library and adds a new flag with the same name.
This reverts commit 852afed5e0.
Changes since D114732:
On PS4, we reverse the expectation that classes whose constructor is deleted are not trivially relocatable. Because, at the moment, only classes which are passed in registers are trivially relocatable, and PS4 allows passing in registers if the copy constructor is deleted, the original assertions were broken on PS4.
(This is kinda similar to DR1734.)
Reviewed By: gribozavr2
Differential Revision: https://reviews.llvm.org/D119017
New device library supporting v4 and v5 has abi_version_400.bc and abi
version_500.bc.
For v5, abi_version_500.bc is linked.
For v2-4, abi_version_400.bc is linked.
For old device library, for v2-4, none of the above is linked. For v5,
error is emitted about unsupported ABI version.
Reviewed by: Artem Belevich
Differential Revision: https://reviews.llvm.org/D118949
Fixes: SWDEV-321313
OpenMP Spec 5.0 [2.12.5, Restrictions]: If a device clause in which the
ancestor device-modifier appears is present on the target construct,
then a requires directive with the reverse_offload clause must be
specified.
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D118887
This patch introduces the Grammar class, which is a critial piece for constructing
a tabled-based parser.
As the first patch, the scope is limited to:
- define base types (symbol, rules) of modeling the grammar
- construct Grammar by parsing the BNF file (annotations are excluded for now)
Differential Revision: https://reviews.llvm.org/D114790
This change enables library code to skip paired move-construction and destruction for `trivial_abi` types, as if they were trivially-movable and trivially-destructible. This offers an extension to the performance fix offered by `trivial_abi`: rather than only offering trivial-type-like performance for pass-by-value, it also offers it for library code that moves values but not as arguments.
For example, if we use `memcpy` for trivially relocatable types inside of vector reallocation, and mark `unique_ptr` as `trivial_abi` (via `_LIBCPP_ABI_ENABLE_UNIQUE_PTR_TRIVIAL_ABI` / `_LIBCPP_ABI_UNSTABLE` / etc.), this would speed up `vector<unique_ptr>::push_back` by 40% on my benchmarks. (Though note that in this case, the compiler could have done this anyway, but happens not to due to the inlining horizon.)
If accepted, I intend to follow up with exactly such changes to library code, including and especially `std::vector`, making them use a trivial relocation operation on trivially relocatable types.
**D50119 and P1144:**
This change is very similar to D50119, which was rejected from Clang. (That change was an implementation of P1144, which is not yet part of the C++ standard.)
The intent of this change, rather than trying to pick a winning proposal for trivial relocation operations, is to extend the behavior of `trivial_abi` in a way that could be made compatible with any such proposal. If P1144 or any similar proposal were accepted, then `trivial_abi`, `__is_trivially_relocatable`, and everything else in this change would be redefined in terms of that.
**Safety:**
It's worth pointing out, specifically, that `trivial_abi` already implies trivial relocatability in a narrow sense: a `trivial_abi` type, when passed by value, has its constructor run in one location, and its destructor run in another, after the type has been trivially relocated (through registers).
Trivial relocatability optimizations could change the number of paired constructor/destructor calls, but this seems unlikely to matter for `trivial_abi` types.
Reviewed By: rsmith
Differential Revision: https://reviews.llvm.org/D114732
Now that VS2017 support has been dropped (D114639), the LLVM_HAS_RVALUE_REFERENCE_THIS define is always true and the LLVM_LVALUE_FUNCTION define is always enabled for ref-qualifiers.
This patch proposes we remove the defines and use the qualifiers directly.
Differential Revision: https://reviews.llvm.org/D118609
This patch extends clang frontend to add metadata that can be used to emit macho files with two build version load commands.
It utilizes "darwin.target_variant.triple" and "darwin.target_variant.SDK Version" metadata names for that.
MachO uses two build version load commands to represent an object file / binary that is targeting both the macOS target,
and the Mac Catalyst target. At runtime, a dynamic library that supports both targets can be loaded from either a native
macOS or a Mac Catalyst app on a macOS system. We want to add support to this to upstream to LLVM to be able to build
compiler-rt for both targets, to finish the complete support for the Mac Catalyst platform, which is right now targetable
by upstream clang, but the compiler-rt bits aren't supported because of the lack of this multiple build version support.
Differential Revision: https://reviews.llvm.org/D115415
Alex Lorenz