Summary:
The standard GNU atomic operations are a very common way to target
hardware atomics on the device. With more heterogenous devices being
introduced, the concept of memory scopes has been in the LLVM language
for awhile via the `syncscope` modifier. For targets, such as the GPU,
this can change code generation depending on whether or not we only need
to be consistent with the memory ordering with the entire system, the
single GPU device, or lower.
Previously these scopes were only exported via the `opencl` and `hip`
variants of these functions. However, this made it difficult to use
outside of those languages and the semantics were different from the
standard GNU versions. This patch introduces a `__scoped_atomic` variant
for the common functions. There was some discussion over whether or not
these should be overloads of the existing ones, or simply new variants.
I leant towards new variants to be less disruptive.
The scope here can be one of the following
```
__MEMORY_SCOPE_SYSTEM // All devices and systems
__MEMORY_SCOPE_DEVICE // Just this device
__MEMORY_SCOPE_WRKGRP // A 'work-group' AKA CUDA block
__MEMORY_SCOPE_WVFRNT // A 'wavefront' AKA CUDA warp
__MEMORY_SCOPE_SINGLE // A single thread.
```
Naming consistency was attempted, but it is difficult to capture to full
spectrum with no many names. Suggestions appreciated.
A new builtin function __builtin_isfpclass is added. It is called as:
__builtin_isfpclass(<floating point value>, <test>)
and returns an integer value, which is non-zero if the floating point
argument falls into one of the classes specified by the second argument,
and zero otherwise. The set of classes is an integer value, where each
value class is represented by a bit. There are ten data classes, as
defined by the IEEE-754 standard, they are represented by bits:
0x0001 (__FPCLASS_SNAN) - Signaling NaN
0x0002 (__FPCLASS_QNAN) - Quiet NaN
0x0004 (__FPCLASS_NEGINF) - Negative infinity
0x0008 (__FPCLASS_NEGNORMAL) - Negative normal
0x0010 (__FPCLASS_NEGSUBNORMAL) - Negative subnormal
0x0020 (__FPCLASS_NEGZERO) - Negative zero
0x0040 (__FPCLASS_POSZERO) - Positive zero
0x0080 (__FPCLASS_POSSUBNORMAL) - Positive subnormal
0x0100 (__FPCLASS_POSNORMAL) - Positive normal
0x0200 (__FPCLASS_POSINF) - Positive infinity
They have corresponding builtin macros to facilitate using the builtin
function:
if (__builtin_isfpclass(x, __FPCLASS_NEGZERO | __FPCLASS_POSZERO) {
// x is any zero.
}
The data class encoding is identical to that used in llvm.is.fpclass
function.
Differential Revision: https://reviews.llvm.org/D152351
These tests should have added -std=c++23 instead of replacing -std=c++2b
in D149553.
Reviewed By: aaron.ballman
Differential Revision: https://reviews.llvm.org/D150063
During the ISO C++ Committee meeting plenary session the C++23 Standard
has been voted as technical complete.
This updates the reference to c++2b to c++23 and updates the __cplusplus
macro.
Drive-by fixes c++1z -> c++17 and c++2a -> c++20 when seen.
Reviewed By: aaron.ballman
Differential Revision: https://reviews.llvm.org/D149553
These directives define per-test lit substitutions. The concept was
discussed at
<https://discourse.llvm.org/t/iterating-lit-run-lines/62596/10>.
For example, the following directives can be inserted into a test file
to define `%{cflags}` and `%{fcflags}` substitutions with empty
initial values, which serve as the parameters of another newly defined
`%{check}` substitution:
```
// DEFINE: %{cflags} =
// DEFINE: %{fcflags} =
// DEFINE: %{check} = %clang_cc1 %{cflags} -emit-llvm -o - %s | \
// DEFINE: FileCheck %{fcflags} %s
```
The following directives then redefine the parameters before each use
of `%{check}`:
```
// REDEFINE: %{cflags} = -foo
// REDEFINE: %{fcflags} = -check-prefix=FOO
// RUN: %{check}
// REDEFINE: %{cflags} = -bar
// REDEFINE: %{fcflags} = -check-prefix=BAR
// RUN: %{check}
```
Of course, `%{check}` would typically be more elaborate, increasing
the benefit of the reuse.
One issue is that the strings `DEFINE:` and `REDEFINE:` already appear
in 5 tests. This patch adjusts those tests not to use those strings.
Our prediction is that, in the vast majority of cases, if a test
author mistakenly uses one of those strings for another purpose, the
text appearing after the string will not happen to have the syntax
required for these directives. Thus, the test author will discover
the mistake immediately when lit reports the syntax error.
This patch also expands the documentation on existing lit substitution
behavior.
Reviewed By: jhenderson, MaskRay, awarzynski
Differential Revision: https://reviews.llvm.org/D132513
Part of the _BitInt feature in C2x
(http://www.open-std.org/jtc1/sc22/wg14/www/docs/n2763.pdf) is a new
macro in limits.h named BITINT_MAXWIDTH that can be used to determine
the maximum width of a bit-precise integer type. This macro must expand
to a value that is at least as large as ULLONG_WIDTH.
This adds an implementation-defined macro named __BITINT_MAXWIDTH__ to
specify that value, which is used by limits.h for the standard macro.
This also limits the maximum bit width to 128 bits because backends do
not currently support all mathematical operations (such as division) on
wider types yet. This maximum is expected to be increased in the future.
This completes the implementation of
WG14 N2412 (http://www.open-std.org/jtc1/sc22/wg14/www/docs/n2412.pdf),
which standardizes C on a twos complement representation for integer
types. The only work that remained there was to define the correct
macros in the standard headers, which this patch does.
This causes modern glibc to unset math_errhandling MATH_ERRNO. gcc 12
also sets some other macros, but most of them are associated with
flags ignored by clang, so without library examples, it is difficult to
determine whether they should be set. I think setting this one macro is
OK for now.
MSVC's libc doesn't provide thread.h, so we should set the macro to
indicate that.
We could just set it in C mode, but I noticed that Darwin sets it
unconditionally, so perhaps we should do the same here.
Differential revision: https://reviews.llvm.org/D112081
The intent of this patch is to add support of -fp-model=[source|double|extended] to allow
the compiler to use a wider type for intermediate floating point calculations. As a side
effect to that, the value of FLT_EVAL_METHOD is changed according to the pragma
float_control.
Unfortunately some issue was uncovered with this change in preprocessing. See details in
https://reviews.llvm.org/D93769 . We are therefore reverting this patch until we find a way
to reconcile the value of FLT_EVAL_METHOD, the pragma and the -E flow.
This reverts commit 66ddac22e2a7f268e91c26d694112970dfa607ae.
The Intel compiler ICC supports the option "-fp-model=(source|double|extended)"
which causes the compiler to use a wider type for intermediate floating point
calculations. Also supported is a way to embed this effect in the source
program with #pragma float_control(source|double|extended).
This patch extends pragma float_control syntax, and also adds support
for a new floating point option "-ffp-eval-method=(source|double|extended)".
source: intermediate results use source precision
double: intermediate results use double precision
extended: intermediate results use extended precision
Reviewed By: Aaron Ballman
Differential Revision: https://reviews.llvm.org/D93769
The Intel compiler ICC supports the option "-fp-model=(source|double|extended)"
which causes the compiler to use a wider type for intermediate floating point
calculations. Also supported is a way to embed this effect in the source
program with #pragma float_control(source|double|extended).
This patch extends pragma float_control syntax, and also adds support
for a new floating point option "-ffp-eval-method=(source|double|extended)".
source: intermediate results use source precision
double: intermediate results use double precision
extended: intermediate results use extended precision
Reviewed By: Aaron Ballman
Differential Revision: https://reviews.llvm.org/D93769
The headers shipped with the XMOS XCore compiler expect __xcore__ to be defined.
The __XS1B__ macro, already defined, is for the default subtarget.
No other targets affected.
Adds the __clang_literal_encoding__ and __clang_wide_literal_encoding__
predefined macros to expose the encoding used for string literals to
the preprocessor.
Currently the emscripten frontend driver injects this when building
with thread support. Moving this into the clang driver itself makes
the emscripten python driver less magical.
Differential Revision: https://reviews.llvm.org/D96171
When the -matomics feature is not enabled, disable POSIXThreads
mode and set the thread model to Single, so that we don't predefine
macros like `__STDCPP_THREADS__`.
Differential Revision: https://reviews.llvm.org/D96091
Android has a handful of API levels relevant to developers described
here: https://developer.android.com/studio/build#module-level.
`__ANDROID_API__` is too vague and confuses a lot of people. Introduce
a new macro name that is explicit about which one it represents. Keep
the old name around because code has been using it for a decade.
Some parts of the test had been extracted into separate files previously.
This patch continues the trend and extracts few more large blocks.
This reduces wall time for the test from a single 14s-long test into a set of
smaller tests that can be run in parallel.
Before/after state of the check-clang tests are here:
https://gist.github.com/Artem-B/d0b05c2e98a49158c02de23f7f4f0279
Differential Revision: https://reviews.llvm.org/D85798
