The original version of this was reverted, and @rjmcall provided some
advice to architect a new solution. This is that solution.
This implements a builtin to provide a unique name that is stable across
compilations of this TU for the purposes of implementing the library
component of the unnamed kernel feature of SYCL. It does this by
running the Itanium mangler with a few modifications.
Because it is somewhat common to wrap non-kernel-related lambdas in
macros that aren't present on the device (such as for logging), this
uniquely generates an ID for all lambdas involved in the naming of a
kernel. It uses the lambda-mangling number to do this, except replaces
this with its own number (starting at 10000 for readabililty reasons)
for lambdas used to name a kernel.
Additionally, this implements itself as constexpr with a slight catch:
if a name would be invalidated by the use of this lambda in a later
kernel invocation, it is diagnosed as an error (see the Sema tests).
Differential Revision: https://reviews.llvm.org/D103112
Allow use of bit-fields as a clang extension
in OpenCL. The extension can be enabled using
pragma directives.
This fixes PR45339!
Differential Revision: https://reviews.llvm.org/D101843
Added __cl_clang_non_portable_kernel_param_types extension that
allows using non-portable types as kernel parameters. This allows
bypassing the portability guarantees from the restrictions specified
in C++ for OpenCL v1.0 s2.4.
Currently this only disables the restrictions related to the data
layout. The programmer should ensure the compiler generates the same
layout for host and device or otherwise the argument should only be
accessed on the device side. This extension could be extended to other
case (e.g. permitting size_t) if desired in the future.
Patch by olestrohm (Ole Strohm)!
https://reviews.llvm.org/D101168
This extension is primarily targeting SPIR-V compilations flow
as the IR translation is the same between 1.x and 2.x atomics.
Differential Revision: https://reviews.llvm.org/D101089
Adds the __clang_literal_encoding__ and __clang_wide_literal_encoding__
predefined macros to expose the encoding used for string literals to
the preprocessor.
Updates __is_unsigned to have the same behavior as the standard
specifies. This is in line with 511dbd8, which applied the same change
to __is_signed.
Refs D67897.
Differential Revision: https://reviews.llvm.org/D98104
This patch adds support for two new variants of the vectorize_width
pragma:
1. vectorize_width(X[, fixed|scalable]) where an optional second
parameter is passed to the vectorize_width pragma, which indicates if
the user wishes to use fixed width or scalable vectorization. For
example the user can now write something like:
#pragma clang loop vectorize_width(4, fixed)
or
#pragma clang loop vectorize_width(4, scalable)
In the absence of a second parameter it is assumed the user wants
fixed width vectorization, in order to maintain compatibility with
existing code.
2. vectorize_width(fixed|scalable) where the width is left unspecified,
but the user hints what type of vectorization they prefer, either
fixed width or scalable.
I have implemented this by making use of the LLVM loop hint attribute:
llvm.loop.vectorize.scalable.enable
Tests were added to
clang/test/CodeGenCXX/pragma-loop.cpp
for both the 'fixed' and 'scalable' optional parameter.
See this thread for context: http://lists.llvm.org/pipermail/cfe-dev/2020-November/067262.html
Differential Revision: https://reviews.llvm.org/D89031
With the internal clang extension '__cl_clang_variadic_functions'
variadic functions are accepted by the frontend.
This is not a fully supported vendor/Khronos extension
as it can only be used on targets with variadic prototype
support or in metaprogramming to represent functions with
generic prototype without calling such functions in the
kernel code.
Tags: #clang
Differential Revision: https://reviews.llvm.org/D94027
The new clang internal extension '__cl_clang_function_pointers'
allows use of function pointers and other features that have
the same functionality:
- Use of member function pointers;
- Unrestricted use of references to functions;
- Virtual member functions.
This not a vendor extension and therefore it doesn't require any
special target support. Exposing this functionality fully
will require vendor or Khronos extension.
Tags: #clang
Differential Revision: https://reviews.llvm.org/D94021
The `assume` attribute is a way to provide additional, arbitrary
information to the optimizer. For now, assumptions are restricted to
strings which will be accumulated for a function and emitted as comma
separated string function attribute. The key of the LLVM-IR function
attribute is `llvm.assume`. Similar to `llvm.assume` and
`__builtin_assume`, the `assume` attribute provides a user defined
assumption to the compiler.
A follow up patch will introduce an LLVM-core API to query the
assumptions attached to a function. We also expect to add more options,
e.g., expression arguments, to the `assume` attribute later on.
The `omp [begin] asssumes` pragma will leverage this attribute and
expose the functionality in the absence of OpenMP.
Reviewed By: aaron.ballman
Differential Revision: https://reviews.llvm.org/D91979
Recently HIP toolchain made a change to use clang instead of opt/llc to do compilation
(https://reviews.llvm.org/D81861). The intention is to make HIP toolchain canonical like
other toolchains.
However, this change introduced an unintentional change regarding backend fp fuse
option, which caused regressions in some HIP applications.
Basically before the change, HIP toolchain used clang to generate bitcode, then use
opt/llc to optimize bitcode and generate ISA. As such, the amdgpu backend takes
the default fp fuse mode which is 'Standard'. This mode respect contract flag of
fmul/fadd instructions and do not fuse fmul/fadd instructions without contract flag.
However, after the change, HIP toolchain now use clang to generate IR, do optimization,
and generate ISA as one process. Now amdgpu backend fp fuse option is determined
by -ffp-contract option, which is 'fast' by default. And this -ffp-contract=fast language option
is translated to 'Fast' fp fuse option in backend. Suddenly backend starts to fuse fmul/fadd
instructions without contract flag.
This causes wrong result for some device library functions, e.g. tan(-1e20), which should
return 0.8446, now returns -0.933. What is worse is that since backend with 'Fast' fp fuse
option does not respect contract flag, there is no way to use #pragma clang fp contract
directive to enforce fp contract requirements.
This patch fixes the regression by introducing a new value 'fast-honor-pragmas' for -ffp-contract
and use it for HIP by default. 'fast-honor-pragmas' is equivalent to 'fast' in frontend but
let the backend to use 'Standard' fp fuse option. 'fast-honor-pragmas' is useful since 'Fast'
fp fuse option in backend does not honor contract flag, it is of little use to HIP
applications since all code with #pragma STDC FP_CONTRACT or any IR from a
source compiled with -ffp-contract=on is broken.
Differential Revision: https://reviews.llvm.org/D90174
Pragma 'clang fp' is extended to support a new option, 'exceptions'. It
allows to specify floating point exception behavior more flexibly.
Differential Revision: https://reviews.llvm.org/D89849
This patch updates the documentation about `__builtin_memcpy_inline` and reorders the sections so it is more consitent and understandable.
Differential Revision: https://reviews.llvm.org/D87458
This enables us to use the __builtin_rotateleft / __builtin_rotateright 8/16/32/64 intrinsics inside constexpr code.
Differential Revision: https://reviews.llvm.org/D86342
Summary:
This patch upstreams support for a new storage only bfloat16 C type.
This type is used to implement primitive support for bfloat16 data, in
line with the Bfloat16 extension of the Armv8.6-a architecture, as
detailed here:
https://community.arm.com/developer/ip-products/processors/b/processors-ip-blog/posts/arm-architecture-developments-armv8-6-a
The bfloat type, and its properties are specified in the Arm Architecture
Reference Manual:
https://developer.arm.com/docs/ddi0487/latest/arm-architecture-reference-manual-armv8-for-armv8-a-architecture-profile
In detail this patch:
- introduces an opaque, storage-only C-type __bf16, which introduces a new bfloat IR type.
This is part of a patch series, starting with command-line and Bfloat16
assembly support. The subsequent patches will upstream intrinsics
support for BFloat16, followed by Matrix Multiplication and the
remaining Virtualization features of the armv8.6-a architecture.
The following people contributed to this patch:
- Luke Cheeseman
- Momchil Velikov
- Alexandros Lamprineas
- Luke Geeson
- Simon Tatham
- Ties Stuij
Reviewers: SjoerdMeijer, rjmccall, rsmith, liutianle, RKSimon, craig.topper, jfb, LukeGeeson, fpetrogalli
Reviewed By: SjoerdMeijer
Subscribers: labrinea, majnemer, asmith, dexonsmith, kristof.beyls, arphaman, danielkiss, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D76077
Extension vectors now can be used in element-wise conditional selector.
For example:
```
R[i] = C[i]? A[i] : B[i]
```
This feature was previously only enabled in OpenCL C. Now it's also
available in C. Not that it has different behaviors than GNU vectors
(i.e. __vector_size__). Extension vectors selects on signdness of the
vector. GNU vectors on the other hand do normal bool conversions. Also,
this feature is not available in C++.
Differential Revision: https://reviews.llvm.org/D80574
test cases
Add support for #pragma float_control
Reviewers: rjmccall, erichkeane, sepavloff
Differential Revision: https://reviews.llvm.org/D72841
This reverts commit 85dc033cac, and makes
corrections to the test cases that failed on buildbots.
Summary:
When using -ftrivial-auto-var-init=* options to initiate automatic
variables in a file, to disable initialization on some variables,
currently we have to manually annotate the variables with uninitialized
attribute, such as
int dont_initialize_me __attribute((uninitialized));
Making pragma clang attribute to support this attribute would make
annotating variables much easier, and could be particular useful for
bisection efforts, e.g.
void use(void*);
void buggy() {
int arr[256];
int boom;
float bam;
struct { int oops; } oops;
union { int oof; float aaaaa; } oof;
use(&arr);
use(&boom);
use(&bam);
use(&oops);
use(&oof);
}
Reviewers: jfb, rjmccall, aaron.ballman
Reviewed By: jfb, aaron.ballman
Subscribers: aaron.ballman, george.burgess.iv, dexonsmith, MaskRay, phosek, hubert.reinterpretcast, gbiv, manojgupta, llozano, srhines, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D78693
This reverts commit 61ba1481e2.
I'm reverting this because it breaks the lldb build with
incomplete switch coverage warnings. I would fix it forward,
but am not familiar enough with lldb to determine the correct
fix.
lldb/source/Plugins/TypeSystem/Clang/TypeSystemClang.cpp:3958:11: error: enumeration values 'DependentExtInt' and 'ExtInt' not handled in switch [-Werror,-Wswitch]
switch (qual_type->getTypeClass()) {
^
lldb/source/Plugins/TypeSystem/Clang/TypeSystemClang.cpp:4633:11: error: enumeration values 'DependentExtInt' and 'ExtInt' not handled in switch [-Werror,-Wswitch]
switch (qual_type->getTypeClass()) {
^
lldb/source/Plugins/TypeSystem/Clang/TypeSystemClang.cpp:4889:11: error: enumeration values 'DependentExtInt' and 'ExtInt' not handled in switch [-Werror,-Wswitch]
switch (qual_type->getTypeClass()) {
Introduction/Motivation:
LLVM-IR supports integers of non-power-of-2 bitwidth, in the iN syntax.
Integers of non-power-of-two aren't particularly interesting or useful
on most hardware, so much so that no language in Clang has been
motivated to expose it before.
However, in the case of FPGA hardware normal integer types where the
full bitwidth isn't used, is extremely wasteful and has severe
performance/space concerns. Because of this, Intel has introduced this
functionality in the High Level Synthesis compiler[0]
under the name "Arbitrary Precision Integer" (ap_int for short). This
has been extremely useful and effective for our users, permitting them
to optimize their storage and operation space on an architecture where
both can be extremely expensive.
We are proposing upstreaming a more palatable version of this to the
community, in the form of this proposal and accompanying patch. We are
proposing the syntax _ExtInt(N). We intend to propose this to the WG14
committee[1], and the underscore-capital seems like the active direction
for a WG14 paper's acceptance. An alternative that Richard Smith
suggested on the initial review was __int(N), however we believe that
is much less acceptable by WG14. We considered _Int, however _Int is
used as an identifier in libstdc++ and there is no good way to fall
back to an identifier (since _Int(5) is indistinguishable from an
unnamed initializer of a template type named _Int).
[0]https://www.intel.com/content/www/us/en/software/programmable/quartus-prime/hls-compiler.html)
[1]http://www.open-std.org/jtc1/sc22/wg14/www/docs/n2472.pdf
Differential Revision: https://reviews.llvm.org/D73967
LanguageExtensions.rst:2191: WARNING: Title underline too short.
llvm-symbolizer.rst:157: Error in "code-block" directive: maximum 1 argument(s) allowed, 30 supplied.
memchr consistent and comprehensible, and document them.
We previously allowed evaluation of memcmp on arrays of integers of any
size, so long as the call evaluated to 0, and allowed evaluation of
memchr on any array of integral type of size 1 (including enums). The
purpose of constant-evaluating these builtins is only to support
constexpr std::char_traits, so we now consistently allow them on arrays
of (possibly signed or unsigned) char only.
In order to support non-user-named kernels, SYCL needs some way in the
integration headers to name the kernel object themselves. Initially, the
design considered just RTTI naming of the lambdas, this results in a
quite unstable situation in light of some device/host macros.
Additionally, this ends up needing to use RTTI, which is a burden on the
implementation and typically unsupported.
Instead, we've introduced a builtin, __builtin_unique_stable_name, which
takes a type or expression, and results in a constexpr constant
character array that uniquely represents the type (or type of the
expression) being passed to it.
The implementation accomplishes that simply by using a slightly modified
version of the Itanium Mangling. The one exception is when mangling
lambdas, instead of appending the index of the lambda in the function,
it appends the macro-expansion back-trace of the lambda itself in the
form LINE->COL[~LINE->COL...].
Differential Revision: https://reviews.llvm.org/D76620
There are a few places with unexpected indents that trip over sphinx and
other syntax errors.
Also, the C++ syntax highlighting does not work for
class [[gsl::Owner(int)]] IntOwner {
Use a regular code:: block instead.
There are a few other warnings errors remaining, of the form
'Duplicate explicit target name: "cmdoption-clang--prefix"'. They seem
to be caused by the following
.. option:: -B<dir>, --prefix <arg>, --prefix=<arg>
I am no Restructured Text expert, but it seems like sphinx 1.8.5
tries to generate the same target for the --prefix <arg> and
--prefix=<arg>. This pops up in a lot of places and I am not sure how to
best resolve it
Reviewers: jfb, Bigcheese, dexonsmith, rjmccall
Reviewed By: rjmccall
Differential Revision: https://reviews.llvm.org/D76534
Summary:
Clang's "asm goto" feature didn't initially support outputs constraints. That
was the same behavior as gcc's implementation. The decision by gcc not to
support outputs was based on a restriction in their IR regarding terminators.
LLVM doesn't restrict terminators from returning values (e.g. 'invoke'), so
it made sense to support this feature.
Output values are valid only on the 'fallthrough' path. If an output value's used
on an indirect branch, then it's 'poisoned'.
In theory, outputs *could* be valid on the 'indirect' paths, but it's very
difficult to guarantee that the original semantics would be retained. E.g.
because indirect labels could be used as data, we wouldn't be able to split
critical edges in situations where two 'callbr' instructions have the same
indirect label, because the indirect branch's destination would no longer be
the same.
Reviewers: jyknight, nickdesaulniers, hfinkel
Reviewed By: jyknight, nickdesaulniers
Subscribers: MaskRay, rsmith, hiraditya, llvm-commits, cfe-commits, craig.topper, rnk
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D69876
user interface and documentation, and update __cplusplus for C++20.
WG21 considers the C++20 standard to be finished (even though it still
has some more steps to pass through in the ISO process).
The old flag names are accepted for compatibility, as usual, and we
still have lots of references to C++2a in comments and identifiers;
those can be cleaned up separately.
GCC supports the conditional operator on VectorTypes that acts as a
'select' in C++ mode. This patch implements the support. Types are
converted as closely to GCC's behavior as possible, though in a few
places consistency with our existing vector type support was preferred.
Note that this implementation is different from the OpenCL version in a
number of ways, so it unfortunately required a different implementation.
First, the SEMA rules and promotion rules are significantly different.
Secondly, GCC implements COND[i] != 0 ? LHS[i] : RHS[i] (where i is in
the range 0- VectorSize, for each element). In OpenCL, the condition is
COND[i] < 0 ? LHS[i]: RHS[i].
In the process of implementing this, it was also required to make the
expression COND ? LHS : RHS type dependent if COND is type dependent,
since the type is now dependent on the condition. For example:
T ? 1 : 2;
Is not typically type dependent, since the result can be deduced from
the operands. HOWEVER, if T is a VectorType now, it could change this
to a 'select' (basically a swizzle with a non-constant mask) with the 1
and 2 being promoted to vectors themselves.
While this is a change, it is NOT a standards incompatible change. Based
on my (and D. Gregor's, at the time of writing the code) reading of the
standard, the expression is supposed to be type dependent if ANY
sub-expression is type dependent.
Differential Revision: https://reviews.llvm.org/D71463
This change introduces three new builtins (which work on both pointers
and integers) that can be used instead of common bitwise arithmetic:
__builtin_align_up(x, alignment), __builtin_align_down(x, alignment) and
__builtin_is_aligned(x, alignment).
I originally added these builtins to the CHERI fork of LLVM a few years ago
to handle the slightly different C semantics that we use for CHERI [1].
Until recently these builtins (or sequences of other builtins) were
required to generate correct code. I have since made changes to the default
C semantics so that they are no longer strictly necessary (but using them
does generate slightly more efficient code). However, based on our experience
using them in various projects over the past few years, I believe that adding
these builtins to clang would be useful.
These builtins have the following benefit over bit-manipulation and casts
via uintptr_t:
- The named builtins clearly convey the semantics of the operation. While
checking alignment using __builtin_is_aligned(x, 16) versus
((x & 15) == 0) is probably not a huge win in readably, I personally find
__builtin_align_up(x, N) a lot easier to read than (x+(N-1))&~(N-1).
- They preserve the type of the argument (including const qualifiers). When
using casts via uintptr_t, it is easy to cast to the wrong type or strip
qualifiers such as const.
- If the alignment argument is a constant value, clang can check that it is
a power-of-two and within the range of the type. Since the semantics of
these builtins is well defined compared to arbitrary bit-manipulation,
it is possible to add a UBSAN checker that the run-time value is a valid
power-of-two. I intend to add this as a follow-up to this change.
- The builtins avoids int-to-pointer casts both in C and LLVM IR.
In the future (i.e. once most optimizations handle it), we could use the new
llvm.ptrmask intrinsic to avoid the ptrtoint instruction that would normally
be generated.
- They can be used to round up/down to the next aligned value for both
integers and pointers without requiring two separate macros.
- In many projects the alignment operations are already wrapped in macros (e.g.
roundup2 and rounddown2 in FreeBSD), so by replacing the macro implementation
with a builtin call, we get improved diagnostics for many call-sites while
only having to change a few lines.
- Finally, the builtins also emit assume_aligned metadata when used on pointers.
This can improve code generation compared to the uintptr_t casts.
[1] In our CHERI compiler we have compilation mode where all pointers are
implemented as capabilities (essentially unforgeable 128-bit fat pointers).
In our original model, casts from uintptr_t (which is a 128-bit capability)
to an integer value returned the "offset" of the capability (i.e. the
difference between the virtual address and the base of the allocation).
This causes problems for cases such as checking the alignment: for example, the
expression `if ((uintptr_t)ptr & 63) == 0` is generally used to check if the
pointer is aligned to a multiple of 64 bytes. The problem with offsets is that
any pointer to the beginning of an allocation will have an offset of zero, so
this check always succeeds in that case (even if the address is not correctly
aligned). The same issues also exist when aligning up or down. Using the
alignment builtins ensures that the address is used instead of the offset. While
I have since changed the default C semantics to return the address instead of
the offset when casting, this offset compilation mode can still be used by
passing a command-line flag.
Reviewers: rsmith, aaron.ballman, theraven, fhahn, lebedev.ri, nlopes, aqjune
Reviewed By: aaron.ballman, lebedev.ri
Differential Revision: https://reviews.llvm.org/D71499
Remove description of language mode from the language
extensions and add a link to pdf document.
Tags: #clang
Differential Revision: https://reviews.llvm.org/D72076
We seem to have been gradually growing support for atomic min/max operations
(exposing longstanding IR atomicrmw instructions). But until now there have
been gaps in the expected intrinsics. This adds support for the C11-style
intrinsics (i.e. taking _Atomic, rather than individually blessed by C11
standard), and the variants that return the new value instead of the original
one.
That way, people won't be misled by trying one form and it not working, and the
front-end is more friendly to people using _Atomic types, as we recommend.
Summary: This patch fixes the __is_signed builtin type trait to work with floating point types and enums. Now, the builtin will return true if it is passed a floating point type and false for an enum type.
Reviewers: EricWF, rsmith, erichkeane, craig.topper, efriedma
Subscribers: cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D67897
llvm-svn: 372621
Commit c15aa241f8 ("[CLANG][BPF] change __builtin_preserve_access_index()
signature") changed the builtin function signature to
PointerT __builtin_preserve_access_index(PointerT ptr)
with a pointer type as the argument/return type, where argument and
return types must be the same.
There is really no reason for this constraint. The builtin just
presented a code region so that IR builtins
__builtin_{array, struct, union}_preserve_access_index
can be applied.
This patch removed the pointer type restriction to permit any
argument type as long as it is permitted by the compiler.
Differential Revision: https://reviews.llvm.org/D67883
llvm-svn: 372516
The clang intrinsic __builtin_preserve_access_index() currently
has signature:
const void * __builtin_preserve_access_index(const void * ptr)
This may cause compiler warning when:
- parameter type is "volatile void *" or "const volatile void *", or
- the assign-to type of the intrinsic does not have "const" qualifier.
Further, this signature does not allow dereference of the
builtin result pointer as it is a "const void *" type, which
adds extra step for the user to do type casting.
Let us change the signature to:
PointerT __builtin_preserve_access_index(PointerT ptr)
such that the result and argument types are the same.
With this, directly dereferencing the builtin return value
becomes possible.
Differential Revision: https://reviews.llvm.org/D67734
llvm-svn: 372294
This fixes some minor grammatical issues I noticed when reading the docs, and changes the recommended feature testing approach to use __has_attribute instead of __has_extension.
llvm-svn: 369687
Summary:
Previously __has_builtin(__builtin_*) would return false for
__builtin_*s that we modeled as keywords rather than as functions
(because they take type arguments). With this patch, all builtins
that are called with function-call-like syntax return true from
__has_builtin (covering __builtin_* and also the __is_* and __has_* type
traits and the handful of similar builtins without such a prefix).
Update the documentation on __has_builtin and on type traits to match.
While doing this I noticed the type trait documentation was out of date
and incomplete; that's fixed here too.
Reviewers: aaron.ballman
Subscribers: jfb, kristina, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D66100
llvm-svn: 368785
Rename lang mode flag to -cl-std=clc++/-cl-std=CLC++
or -std=clc++/-std=CLC++.
This aligns with OpenCL C conversion and removes ambiguity
with OpenCL C++.
Differential Revision: https://reviews.llvm.org/D65102
llvm-svn: 367008
This adds a new vectorize predication loop hint:
#pragma clang loop vectorize_predicate(enable)
that can be used to indicate to the vectoriser that all (load/store)
instructions should be predicated (masked). This allows, for example, folding
of the remainder loop into the main loop.
This patch will be followed up with D64916 and D65197. The former is a
refactoring in the loopvectorizer and the groundwork to make tail loop folding
a more general concept, and in the latter the actual tail loop folding
transformation will be implemented.
Differential Revision: https://reviews.llvm.org/D64744
llvm-svn: 366989
Added documentation of C++ for OpenCL mode into Clang
User Manual and Language Extensions document.
Differential Revision: https://reviews.llvm.org/D64418
llvm-svn: 366351
Some targets such as Python 2.7.16 still use VERSION in
their builds. Without VERSION defined, the source code
has syntax errors.
Reverting as it will probably break many other things.
Noticed by Sterling Augustine
llvm-svn: 365992
Summary:
It has been introduced in 2011 for gcc compat:
ad1a4c6e89
it is probably time to remove it
Reviewers: rnk, dexonsmith
Reviewed By: rnk
Subscribers: dschuff, aheejin, fedor.sergeev, arphaman, cfe-commits
Tags: #clang
Differential Revision: https://reviews.llvm.org/D64062
llvm-svn: 365962
For background of BPF CO-RE project, please refer to
http://vger.kernel.org/bpfconf2019.html
In summary, BPF CO-RE intends to compile bpf programs
adjustable on struct/union layout change so the same
program can run on multiple kernels with adjustment
before loading based on native kernel structures.
In order to do this, we need keep track of GEP(getelementptr)
instruction base and result debuginfo types, so we
can adjust on the host based on kernel BTF info.
Capturing such information as an IR optimization is hard
as various optimization may have tweaked GEP and also
union is replaced by structure it is impossible to track
fieldindex for union member accesses.
Three intrinsic functions, preserve_{array,union,struct}_access_index,
are introducted.
addr = preserve_array_access_index(base, index, dimension)
addr = preserve_union_access_index(base, di_index)
addr = preserve_struct_access_index(base, gep_index, di_index)
here,
base: the base pointer for the array/union/struct access.
index: the last access index for array, the same for IR/DebugInfo layout.
dimension: the array dimension.
gep_index: the access index based on IR layout.
di_index: the access index based on user/debuginfo types.
If using these intrinsics blindly, i.e., transforming all GEPs
to these intrinsics and later on reducing them to GEPs, we have
seen up to 7% more instructions generated. To avoid such an overhead,
a clang builtin is proposed:
base = __builtin_preserve_access_index(base)
such that user wraps to-be-relocated GEPs in this builtin
and preserve_*_access_index intrinsics only apply to
those GEPs. Such a buyin will prevent performance degradation
if people do not use CO-RE, even for programs which use
bpf_probe_read().
For example, for the following example,
$ cat test.c
struct sk_buff {
int i;
int b1:1;
int b2:2;
union {
struct {
int o1;
int o2;
} o;
struct {
char flags;
char dev_id;
} dev;
int netid;
} u[10];
};
static int (*bpf_probe_read)(void *dst, int size, const void *unsafe_ptr)
= (void *) 4;
#define _(x) (__builtin_preserve_access_index(x))
int bpf_prog(struct sk_buff *ctx) {
char dev_id;
bpf_probe_read(&dev_id, sizeof(char), _(&ctx->u[5].dev.dev_id));
return dev_id;
}
$ clang -target bpf -O2 -g -emit-llvm -S -mllvm -print-before-all \
test.c >& log
The generated IR looks like below:
...
define dso_local i32 @bpf_prog(%struct.sk_buff*) #0 !dbg !15 {
%2 = alloca %struct.sk_buff*, align 8
%3 = alloca i8, align 1
store %struct.sk_buff* %0, %struct.sk_buff** %2, align 8, !tbaa !45
call void @llvm.dbg.declare(metadata %struct.sk_buff** %2, metadata !43, metadata !DIExpression()), !dbg !49
call void @llvm.lifetime.start.p0i8(i64 1, i8* %3) #4, !dbg !50
call void @llvm.dbg.declare(metadata i8* %3, metadata !44, metadata !DIExpression()), !dbg !51
%4 = load i32 (i8*, i32, i8*)*, i32 (i8*, i32, i8*)** @bpf_probe_read, align 8, !dbg !52, !tbaa !45
%5 = load %struct.sk_buff*, %struct.sk_buff** %2, align 8, !dbg !53, !tbaa !45
%6 = call [10 x %union.anon]* @llvm.preserve.struct.access.index.p0a10s_union.anons.p0s_struct.sk_buffs(
%struct.sk_buff* %5, i32 2, i32 3), !dbg !53, !llvm.preserve.access.index !19
%7 = call %union.anon* @llvm.preserve.array.access.index.p0s_union.anons.p0a10s_union.anons(
[10 x %union.anon]* %6, i32 1, i32 5), !dbg !53
%8 = call %union.anon* @llvm.preserve.union.access.index.p0s_union.anons.p0s_union.anons(
%union.anon* %7, i32 1), !dbg !53, !llvm.preserve.access.index !26
%9 = bitcast %union.anon* %8 to %struct.anon.0*, !dbg !53
%10 = call i8* @llvm.preserve.struct.access.index.p0i8.p0s_struct.anon.0s(
%struct.anon.0* %9, i32 1, i32 1), !dbg !53, !llvm.preserve.access.index !34
%11 = call i32 %4(i8* %3, i32 1, i8* %10), !dbg !52
%12 = load i8, i8* %3, align 1, !dbg !54, !tbaa !55
%13 = sext i8 %12 to i32, !dbg !54
call void @llvm.lifetime.end.p0i8(i64 1, i8* %3) #4, !dbg !56
ret i32 %13, !dbg !57
}
!19 = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "sk_buff", file: !3, line: 1, size: 704, elements: !20)
!26 = distinct !DICompositeType(tag: DW_TAG_union_type, scope: !19, file: !3, line: 5, size: 64, elements: !27)
!34 = distinct !DICompositeType(tag: DW_TAG_structure_type, scope: !26, file: !3, line: 10, size: 16, elements: !35)
Note that @llvm.preserve.{struct,union}.access.index calls have metadata llvm.preserve.access.index
attached to instructions to provide struct/union debuginfo type information.
For &ctx->u[5].dev.dev_id,
. The "%6 = ..." represents struct member "u" with index 2 for IR layout and index 3 for DI layout.
. The "%7 = ..." represents array subscript "5".
. The "%8 = ..." represents union member "dev" with index 1 for DI layout.
. The "%10 = ..." represents struct member "dev_id" with index 1 for both IR and DI layout.
Basically, traversing the use-def chain recursively for the 3rd argument of bpf_probe_read() and
examining all preserve_*_access_index calls, the debuginfo struct/union/array access index
can be achieved.
The intrinsics also contain enough information to regenerate codes for IR layout.
For array and structure intrinsics, the proper GEP can be constructed.
For union intrinsics, replacing all uses of "addr" with "base" should be enough.
Signed-off-by: Yonghong Song <yhs@fb.com>
Differential Revision: https://reviews.llvm.org/D61809
llvm-svn: 365438
For background of BPF CO-RE project, please refer to
http://vger.kernel.org/bpfconf2019.html
In summary, BPF CO-RE intends to compile bpf programs
adjustable on struct/union layout change so the same
program can run on multiple kernels with adjustment
before loading based on native kernel structures.
In order to do this, we need keep track of GEP(getelementptr)
instruction base and result debuginfo types, so we
can adjust on the host based on kernel BTF info.
Capturing such information as an IR optimization is hard
as various optimization may have tweaked GEP and also
union is replaced by structure it is impossible to track
fieldindex for union member accesses.
Three intrinsic functions, preserve_{array,union,struct}_access_index,
are introducted.
addr = preserve_array_access_index(base, index, dimension)
addr = preserve_union_access_index(base, di_index)
addr = preserve_struct_access_index(base, gep_index, di_index)
here,
base: the base pointer for the array/union/struct access.
index: the last access index for array, the same for IR/DebugInfo layout.
dimension: the array dimension.
gep_index: the access index based on IR layout.
di_index: the access index based on user/debuginfo types.
If using these intrinsics blindly, i.e., transforming all GEPs
to these intrinsics and later on reducing them to GEPs, we have
seen up to 7% more instructions generated. To avoid such an overhead,
a clang builtin is proposed:
base = __builtin_preserve_access_index(base)
such that user wraps to-be-relocated GEPs in this builtin
and preserve_*_access_index intrinsics only apply to
those GEPs. Such a buyin will prevent performance degradation
if people do not use CO-RE, even for programs which use
bpf_probe_read().
For example, for the following example,
$ cat test.c
struct sk_buff {
int i;
int b1:1;
int b2:2;
union {
struct {
int o1;
int o2;
} o;
struct {
char flags;
char dev_id;
} dev;
int netid;
} u[10];
};
static int (*bpf_probe_read)(void *dst, int size, const void *unsafe_ptr)
= (void *) 4;
#define _(x) (__builtin_preserve_access_index(x))
int bpf_prog(struct sk_buff *ctx) {
char dev_id;
bpf_probe_read(&dev_id, sizeof(char), _(&ctx->u[5].dev.dev_id));
return dev_id;
}
$ clang -target bpf -O2 -g -emit-llvm -S -mllvm -print-before-all \
test.c >& log
The generated IR looks like below:
...
define dso_local i32 @bpf_prog(%struct.sk_buff*) #0 !dbg !15 {
%2 = alloca %struct.sk_buff*, align 8
%3 = alloca i8, align 1
store %struct.sk_buff* %0, %struct.sk_buff** %2, align 8, !tbaa !45
call void @llvm.dbg.declare(metadata %struct.sk_buff** %2, metadata !43, metadata !DIExpression()), !dbg !49
call void @llvm.lifetime.start.p0i8(i64 1, i8* %3) #4, !dbg !50
call void @llvm.dbg.declare(metadata i8* %3, metadata !44, metadata !DIExpression()), !dbg !51
%4 = load i32 (i8*, i32, i8*)*, i32 (i8*, i32, i8*)** @bpf_probe_read, align 8, !dbg !52, !tbaa !45
%5 = load %struct.sk_buff*, %struct.sk_buff** %2, align 8, !dbg !53, !tbaa !45
%6 = call [10 x %union.anon]* @llvm.preserve.struct.access.index.p0a10s_union.anons.p0s_struct.sk_buffs(
%struct.sk_buff* %5, i32 2, i32 3), !dbg !53, !llvm.preserve.access.index !19
%7 = call %union.anon* @llvm.preserve.array.access.index.p0s_union.anons.p0a10s_union.anons(
[10 x %union.anon]* %6, i32 1, i32 5), !dbg !53
%8 = call %union.anon* @llvm.preserve.union.access.index.p0s_union.anons.p0s_union.anons(
%union.anon* %7, i32 1), !dbg !53, !llvm.preserve.access.index !26
%9 = bitcast %union.anon* %8 to %struct.anon.0*, !dbg !53
%10 = call i8* @llvm.preserve.struct.access.index.p0i8.p0s_struct.anon.0s(
%struct.anon.0* %9, i32 1, i32 1), !dbg !53, !llvm.preserve.access.index !34
%11 = call i32 %4(i8* %3, i32 1, i8* %10), !dbg !52
%12 = load i8, i8* %3, align 1, !dbg !54, !tbaa !55
%13 = sext i8 %12 to i32, !dbg !54
call void @llvm.lifetime.end.p0i8(i64 1, i8* %3) #4, !dbg !56
ret i32 %13, !dbg !57
}
!19 = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "sk_buff", file: !3, line: 1, size: 704, elements: !20)
!26 = distinct !DICompositeType(tag: DW_TAG_union_type, scope: !19, file: !3, line: 5, size: 64, elements: !27)
!34 = distinct !DICompositeType(tag: DW_TAG_structure_type, scope: !26, file: !3, line: 10, size: 16, elements: !35)
Note that @llvm.preserve.{struct,union}.access.index calls have metadata llvm.preserve.access.index
attached to instructions to provide struct/union debuginfo type information.
For &ctx->u[5].dev.dev_id,
. The "%6 = ..." represents struct member "u" with index 2 for IR layout and index 3 for DI layout.
. The "%7 = ..." represents array subscript "5".
. The "%8 = ..." represents union member "dev" with index 1 for DI layout.
. The "%10 = ..." represents struct member "dev_id" with index 1 for both IR and DI layout.
Basically, traversing the use-def chain recursively for the 3rd argument of bpf_probe_read() and
examining all preserve_*_access_index calls, the debuginfo struct/union/array access index
can be achieved.
The intrinsics also contain enough information to regenerate codes for IR layout.
For array and structure intrinsics, the proper GEP can be constructed.
For union intrinsics, replacing all uses of "addr" with "base" should be enough.
Signed-off-by: Yonghong Song <yhs@fb.com>
llvm-svn: 365435
Summary:
This patch implements the source location builtins `__builtin_LINE(), `__builtin_FUNCTION()`, `__builtin_FILE()` and `__builtin_COLUMN()`. These builtins are needed to implement [`std::experimental::source_location`](https://rawgit.com/cplusplus/fundamentals-ts/v2/main.html#reflection.src_loc.creation).
With the exception of `__builtin_COLUMN`, GCC also implements these builtins, and Clangs behavior is intended to match as closely as possible.
Reviewers: rsmith, joerg, aaron.ballman, bogner, majnemer, shafik, martong
Reviewed By: rsmith
Subscribers: rnkovacs, loskutov, riccibruno, mgorny, kunitoki, alexr, majnemer, hfinkel, cfe-commits
Differential Revision: https://reviews.llvm.org/D37035
llvm-svn: 360937
Summary:
This patch adds support for __builtin_dcbf for PPC.
__builtin_dcbf copies the contents of a modified block from the data cache
to main memory and flushes the copy from the data cache.
Differential revision: https://reviews.llvm.org/D59843
llvm-svn: 359517
Summary:
PowerPC64/PowerPC64le supports the builtin function __builtin_setrnd to set the floating point rounding mode. This function will use the least significant two bits of integer argument to set the floating point rounding mode.
double __builtin_setrnd(int mode);
The effective values for mode are:
0 - round to nearest
1 - round to zero
2 - round to +infinity
3 - round to -infinity
Note that the mode argument will modulo 4, so if the int argument is greater than 3, it will only use the least significant two bits of the mode. Namely, builtin_setrnd(102)) is equal to builtin_setrnd(2).
Reviewed By: jsji
Differential Revision: https://reviews.llvm.org/D59403
llvm-svn: 357242
This builtin has the same UI as __builtin_object_size, but has the
potential to be evaluated dynamically. It is meant to be used as a
drop-in replacement for libraries that use __builtin_object_size when
a dynamic checking mode is enabled. For instance,
__builtin_object_size fails to provide any extra checking in the
following function:
void f(size_t alloc) {
char* p = malloc(alloc);
strcpy(p, "foobar"); // expands to __builtin___strcpy_chk(p, "foobar", __builtin_object_size(p, 0))
}
This is an overflow if alloc < 7, but because LLVM can't fold the
object size intrinsic statically, it folds __builtin_object_size to
-1. With __builtin_dynamic_object_size, alloc is passed through to
__builtin___strcpy_chk.
rdar://32212419
Differential revision: https://reviews.llvm.org/D56760
llvm-svn: 352665
As Discussed here:
http://lists.llvm.org/pipermail/llvm-dev/2019-January/129543.html
There are problems exposing the _Float16 type on architectures that
haven't defined the ABI/ISel for the type yet, so we're temporarily
disabling the type and making it opt-in.
Differential Revision: https://reviews.llvm.org/D57188
Change-Id: I5db7366dedf1deb9485adb8948b1deb7e612a736
llvm-svn: 352221
Namespaces are introduced by adding an "identifier." before a
push/pop directive. Pop directives with namespaces can only pop a
attribute group that was pushed with the same namespace. Push and pop
directives that don't opt into namespaces have the same semantics.
This is necessary to prevent a pitfall of using multiple #pragma
clang attribute directives spread out in a large file, particularly
when macros are involved. It isn't easy to see which pop corripsonds
to which push, so its easy to inadvertently pop the wrong group.
Differential revision: https://reviews.llvm.org/D55628
llvm-svn: 349845
This clarifies that __has_cpp_attribute is no longer always an extension since it's now available in C++2a. Also, Both __has_cpp_attribute and __has_c_attribute can accept attribute scope tokens with alternative spelling (clang vs _Clang and gnu vs __gnu__).
llvm-svn: 347312
This reverts commit r345487, which reverted r345486. I think the crashes were
caused by an OOM on the builder, trying again to confirm...
llvm-svn: 345517
This commit enables pushing an empty #pragma clang attribute push, then adding
multiple attributes to it, then popping them all with #pragma clang attribute
pop, just like #pragma clang diagnostic. We still support the current way of
adding these, #pragma clang attribute push(__attribute__((...))), by treating it
like a combined push/attribute. This is needed to create macros like:
DO_SOMETHING_BEGIN(attr1, attr2, attr3)
// ...
DO_SOMETHING_END
rdar://45496947
Differential revision: https://reviews.llvm.org/D53621
llvm-svn: 345486
This is a partial retry of rL340137 (reverted at rL340138 because of gcc host compiler crashing)
with 1 change:
Remove the changes to make microsoft builtins also use the LLVM intrinsics.
This exposes the LLVM funnel shift intrinsics as more familiar bit rotation functions in clang
(when both halves of a funnel shift are the same value, it's a rotate).
We're free to name these as we want because we're not copying gcc, but if there's some other
existing art (eg, the microsoft ops) that we want to replicate, we can change the names.
The funnel shift intrinsics were added here:
https://reviews.llvm.org/D49242
With improved codegen in:
https://reviews.llvm.org/rL337966https://reviews.llvm.org/rL339359
And basic IR optimization added in:
https://reviews.llvm.org/rL338218https://reviews.llvm.org/rL340022
...so these are expected to produce asm output that's equal or better to the multi-instruction
alternatives using primitive C/IR ops.
In the motivating loop example from PR37387:
https://bugs.llvm.org/show_bug.cgi?id=37387#c7
...we get the expected 'rolq' x86 instructions if we substitute the rotate builtin into the source.
Differential Revision: https://reviews.llvm.org/D50924
llvm-svn: 340141
Nico Weber