This fixes remaining issues in my previous PR #90959.
Changes:
- Removed dependency on LLVM header in `xray_interface.cpp`
- Fixed XRay patching for some targets due to missing changes in
architecture-specific patching functions
- Addressed some remaining compiler warnings that I missed in the
previous patch
- Formatting
I have tested these changes on `x86_64` (natively), as well as
`ppc64le`, `aarch64` and `arm32` (cross-compiled and emulated using
qemu).
**Original description:**
This PR introduces shared library (DSO) support for XRay based on a
revised version of the implementation outlined in [this
RFC](https://discourse.llvm.org/t/rfc-upstreaming-dso-instrumentation-support-for-xray/73000).
The feature enables the patching and handling of events from DSOs,
supporting both libraries linked at startup or explicitly loaded, e.g.
via `dlopen`.
This patch adds the following:
- The `-fxray-shared` flag to enable the feature (turned off by default)
- A small runtime library that is linked into every instrumented DSO,
providing position-independent trampolines and code to register with the
main XRay runtime
- Changes to the XRay runtime to support management and patching of
multiple objects
These changes are fully backward compatible, i.e. running without
instrumented DSOs will produce identical traces (in terms of recorded
function IDs) to the previous implementation.
Due to my limited ability to test on other architectures, this feature
is only implemented and tested with x86_64. Extending support to other
architectures is fairly straightforward, requiring only a
position-independent implementation of the architecture-specific
trampoline implementation (see
`compiler-rt/lib/xray/xray_trampoline_x86_64.S` for reference).
This patch does not include any functionality to resolve function IDs
from DSOs for the provided logging/tracing modes. These modes still work
and will record calls from DSOs, but symbol resolution for these
functions in not available. Getting this to work properly requires
recording information about the loaded DSOs and should IMO be discussed
in a separate RFC, as there are mulitple feasible approaches.
---------
Co-authored-by: Sebastian Kreutzer <sebastian.kreutzer@tu-darmstadt.de>
Currently when using OpenMP atomics we depend on some symbols from
libatomic. These symbols are provided in a separate library for the
libgcc runtime, so we should link to that when rtlib=libgcc.
For the compiler-rt case, the presence and location of the symbols is
dependent on how compiler-rt itself was built so we cannot make that
decision for the user. As such no extra flags are added in that case.
The underlying issue was caused by a file included in two different
places which resulted in duplicate definition errors when linking
individual shared libraries. This was fixed in c3201ddaeac02a2c86a38b
[#109874].
This PR introduces shared library (DSO) support for XRay based on a
revised version of the implementation outlined in [this
RFC](https://discourse.llvm.org/t/rfc-upstreaming-dso-instrumentation-support-for-xray/73000).
The feature enables the patching and handling of events from DSOs,
supporting both libraries linked at startup or explicitly loaded, e.g.
via `dlopen`.
This patch adds the following:
- The `-fxray-shared` flag to enable the feature (turned off by default)
- A small runtime library that is linked into every instrumented DSO,
providing position-independent trampolines and code to register with the
main XRay runtime
- Changes to the XRay runtime to support management and patching of
multiple objects
These changes are fully backward compatible, i.e. running without
instrumented DSOs will produce identical traces (in terms of recorded
function IDs) to the previous implementation.
Due to my limited ability to test on other architectures, this feature
is only implemented and tested with x86_64. Extending support to other
architectures is fairly straightforward, requiring only a
position-independent implementation of the architecture-specific
trampoline implementation (see
`compiler-rt/lib/xray/xray_trampoline_x86_64.S` for reference).
This patch does not include any functionality to resolve function IDs
from DSOs for the provided logging/tracing modes. These modes still work
and will record calls from DSOs, but symbol resolution for these
functions in not available. Getting this to work properly requires
recording information about the loaded DSOs and should IMO be discussed
in a separate RFC, as there are mulitple feasible approaches.
@petrhosek @jplehr
Add support for -fdiagnostics-color and -fdiagnostics-color=. Add
documentation for -fdiagnostics-color= which should also be visible in
clang.
Partially addresses requests in #89888
Add support for the -frecord-command-line option that will produce the
llvm.commandline metadata which will eventually be saved in the object
file. This behavior is also supported in clang. Some refactoring of the
code in flang to handle these command line options was carried out. The
corresponding -grecord-command-line option which saves the command line
in the debug information has not yet been enabled for flang.
Summary:
I initially thought that it would be convenient to automatically link
these libraries like they are for standard C/C++ targets. However, this
created issues when trying to use C++ as a GPU target. This patch moves
the logic to now implicitly pass it as part of the offloading toolchain
instead, if found. This means that the user needs to set the target
toolchain for the link job for automatic detection, but can still be
done manually via `-Xoffload-linker -lc`.
This adds new Clang flags to support codegen (CG) data:
- `-fcodegen-data-generate{=path}`: This flag passes
`-codegen-data-generate` as a boolean to the LLVM backend, causing the
raw CG data to be emitted into a custom section. Currently, for LLD
MachO only, it also passes `--codegen-data-generate-path=<path>` so that
the indexed CG data file can be automatically produced at link time. For
linkers that do not yet support this feature, `llvm-cgdata` can be used
manually to merge this CG data in object files.
- `-fcodegen-data-use{=path}`: This flag passes
`-codegen-data-use-path=<path>` to the LLVM backend, enabling the use of
specified CG data to optimistically outline functions.
- The default `<path>` is set to `default.cgdata` when not specified.
This depends on https://github.com/llvm/llvm-project/pull/108733.
This is a patch for
https://discourse.llvm.org/t/rfc-enhanced-machine-outliner-part-2-thinlto-nolto/78753.
Introduce the `-fsanitize=realtime` flag in clang driver
Plug in the RealtimeSanitizer PassManager pass in Codegen, and attribute
a function based on if it has the `[[clang::nonblocking]]` function
effect.
-Wa,-mmapsyms=implicit enables the alternative mapping symbol scheme
discussed at #99718.
While not conforming to the current aaelf64 ABI, the option is
invaluable for those with full control over their toolchain, no reliance
on weird relocatable files, and a strong focus on minimizing both
relocatable and executable sizes.
The option is discouraged when portability of the relocatable objects is
a concern.
https://maskray.me/blog/2024-07-21-mapping-symbols-rethinking-for-efficiency
elaborates the risk.
Pull Request: https://github.com/llvm/llvm-project/pull/104542
Through the new `-foffload-via-llvm` flag, CUDA kernels can now be
lowered to the LLVM/Offload API. On the Clang side, this is simply done
by using the OpenMP offload toolchain and emitting calls to `llvm*`
functions to orchestrate the kernel launch rather than `cuda*`
functions. These `llvm*` functions are implemented on top of the
existing LLVM/Offload API.
As we are about to redefine the Offload API, this wil help us in the
design process as a second offload language.
We do not support any CUDA APIs yet, however, we could:
https://www.osti.gov/servlets/purl/1892137
For proper host execution we need to resurrect/rebase
https://tianshilei.me/wp-content/uploads/2021/12/llpp-2021.pdf
(which was designed for debugging).
```
❯❯❯ cat test.cu
extern "C" {
void *llvm_omp_target_alloc_shared(size_t Size, int DeviceNum);
void llvm_omp_target_free_shared(void *DevicePtr, int DeviceNum);
}
__global__ void square(int *A) { *A = 42; }
int main(int argc, char **argv) {
int DevNo = 0;
int *Ptr = reinterpret_cast<int *>(llvm_omp_target_alloc_shared(4, DevNo));
*Ptr = 7;
printf("Ptr %p, *Ptr %i\n", Ptr, *Ptr);
square<<<1, 1>>>(Ptr);
printf("Ptr %p, *Ptr %i\n", Ptr, *Ptr);
llvm_omp_target_free_shared(Ptr, DevNo);
}
❯❯❯ clang++ test.cu -O3 -o test123 -foffload-via-llvm --offload-arch=native
❯❯❯ llvm-objdump --offloading test123
test123: file format elf64-x86-64
OFFLOADING IMAGE [0]:
kind elf
arch gfx90a
triple amdgcn-amd-amdhsa
producer openmp
❯❯❯ LIBOMPTARGET_INFO=16 ./test123
Ptr 0x155448ac8000, *Ptr 7
Ptr 0x155448ac8000, *Ptr 42
```
This patch modifies the flang driver to introduce the `-fopenmp-targets`
option to the frontend compiler invocations corresponding to the OpenMP
host device on offloading-enabled compilations.
This option holds the list of offloading triples associated to the
compilation and is used by clang to determine whether offloading calls
should be generated for the host.
For now only focus on the CPU type, will work on the CPU features part
later.
With the CPU handling in TargetParser, clang and llc/opt are able to
query common interfaces.
So we can set same default CPU and CPU features with same interfaces.
Summary:
The previous patches (The other commits in this chain) allow the
offloading toolchain to directly invoke the device linker. Because of
this, we can now just have the toolchain implicitly include `-lc` and
`-lm` like a standard target does. This removes the old handling that
went through the fat binary `-lcgpu`.
When re-enabling safestack testing on Solaris after the unexplained
b0260c5b1052f8e3ff1ec77dc42a11f42da762cc, all tests `FAIL`ed to link:
```
Undefined first referenced
symbol in file
__safestack_unsafe_stack_ptr buffer-copy-vla.o
__safestack_init (command line)
ld: fatal: symbol referencing errors
```
The problem is that `-u __safestack_init` was passed to the linker after
the corresponding version of `libclang_rt.safestack-*.a`. Since the
Solaris linker (like Unix linkers for decades) respects the command line
argument order (unlike e.g. GNU ld which uses GNU getopt), this cannot
work. Fixed by moving the `-u` arg further to the front. Two affected
testcases were fixed accordingly.
Tested on `amd64-pc-solaris2.11`, `sparcv9-sun-solaris2.11`,
`x86_64-pc-linux-gnu`, and `sparc64-unknown-linux-gnu`.
* `-fsanitize=numerical,undefined`: don't link in the ubsan standalone
runtime.
* `-shared-libsan`: link against `libclang_rt.nsan.so`
The compiler-rt part will be properly fixed by #98415
On Android, we always want frame pointers to make debugging in the field
easier. Since frame pointers are already enabled for AArch64, ARM and
RISCV64, effectively this change further enables frame pointers for X86
and X86_64.
This patch implements the -mcmodel flag from clang, allowing the Code
Model to be changed for the LLVM module. The same set of mcmodel
flags are accepted as in clang and the same Code Model attributes are
added to the LLVM module for those flags.
Also add `-mlarge-data-threshold` for x86-64, which is automatically set
by the shared command-line code (see below). This is also added as an
attribute into the LLVM module and on the target machine.
A function is created for `addMCModel` that is copied out of clang's
argument handling so that it can be shared with flang.
---------
Co-authored-by: Mats Petersson <mats.petersson@arm.com>
When using the -mframe-chain=aapcs or -mframe-chain=aapcs-leaf options,
we cannot use r11 as an allocatable register, even if
-fomit-frame-pointer is also used. This is so that r11 will always point
to a valid frame record, even if we don't create one in every function.
I'm planning to remove StringRef::equals in favor of
StringRef::operator==.
- StringRef::operator==/!= outnumber StringRef::equals by a factor of
13 under clang/ in terms of their usage.
- The elimination of StringRef::equals brings StringRef closer to
std::string_view, which has operator== but not equals.
- S == "foo" is more readable than S.equals("foo"), especially for
!Long.Expression.equals("str") vs Long.Expression != "str".
The implementation only enables when the `-enable-tlsdesc` option is
passed and the TLS model is `dynamic`.
LoongArch's GCC has the same option(-mtls-dialet=) as RISC-V.
Reviewers: heiher, MaskRay, SixWeining
Reviewed By: SixWeining, MaskRay
Pull Request: https://github.com/llvm/llvm-project/pull/90159
This patch changes the behaviour for flang to only create and link to a
`main` entry point when the Fortran code has a program statement in it.
This means that flang-new can be used to link even when the program is
a mixed C/Fortran code with `main` present in C and no entry point
present in Fortran.
This also removes the `-fno-fortran-main` flag as this no longer has any
functionality.
Before this change -gdwarf-5 on AIX will cause backend crash, because
some DWARF5 sections are not defined in XCOFF.
Explicitly statement -gdwarf-5 as unsupported in frontend on AIX.
Summary:
We want to pass these GPU libraries by default if a certain offloading
toolchain is loaded for OpenMP. Previously I parsed this from the
arguments because it's only available in the compilation. This doesn't
really work for `native` and it's extra effort, so this patch just
passes in the `Compilation` as an extr argument and uses that. Tests
should be unaffected.
Summary:
The original intention of the `openmp-add-rpath` option was to add the
rpath to the language runtime directory. However, the current
implementation only adds it to the compiler's resource directory. This
patch adds support for appending the `-rpath` to the compiler's standard
library directory as well. Currently this is `<exe>/../lib/<triple>`.
Added --offload-compression-level= option to clang and
-compression-level=
option to clang-offload-bundler for controlling compression level.
Added support of long distance matching (LDM) for llvm::zstd which is
off
by default. Enable it for clang-offload-bundler by default since it
improves compression rate in general.
Change default compression level to 3 for zstd for clang-offload-bundler
since it works well for bundle entry size from 1KB to 32MB, which should
cover most of the clang-offload-bundler usage. Users can still specify
compression level by -compression-level= option if necessary.
Summary:
We still use this bitcode library in one case, the NVPTX non-LTO build.
The patch updated the search paths to treat it the same as other
libraries, which unintentionally prioritized system paths over
LIBRARY_PATH which is generally not correct. Also we had a test that
relied on system state so remove that.
Summary:
One recurring problem we have with the OpenMP libraries is that they are
potentially conflicting with ones found on the system, this occurs when
there are two copies and one is used for linking that it not attached to
the correspoding clang compiler. LLVM already uses target specific
directories for this, like with libc++, which are always searched first.
This patch changes the install directory to be
`lib/x86_64-unknown-linux-gnu` for example.
Notable changes would be that users will need to change their
LD_LIBRARY_PATH settings optionally, or use default rt-rpath options.
This should fix problems were users are linking the wrong versions of
static libraries
Summary:
This is a massive patch because it reworks the entire build and
everything that depends on it. This is not split up because various bots
would fail otherwise. I will attempt to describe the necessary changes
here.
This patch completely reworks how the GPU build is built and targeted.
Previously, we used a standard runtimes build and handled both NVPTX and
AMDGPU in a single build via multi-targeting. This added a lot of
divergence in the build system and prevented us from doing various
things like building for the CPU / GPU at the same time, or exporting
the startup libraries or running tests without a full rebuild.
The new appraoch is to handle the GPU builds as strict cross-compiling
runtimes. The first step required
https://github.com/llvm/llvm-project/pull/81557 to allow the `LIBC`
target to build for the GPU without touching the other targets. This
means that the GPU uses all the same handling as the other builds in
`libc`.
The new expected way to build the GPU libc is with
`LLVM_LIBC_RUNTIME_TARGETS=amdgcn-amd-amdhsa;nvptx64-nvidia-cuda`.
The second step was reworking how we generated the embedded GPU library
by moving it into the library install step. Where we previously had one
`libcgpu.a` we now have `libcgpu-amdgpu.a` and `libcgpu-nvptx.a`. This
patch includes the necessary clang / OpenMP changes to make that not
break the bots when this lands.
We unfortunately still require that the NVPTX target has an `internal`
target for tests. This is because the NVPTX target needs to do LTO for
the provided version (The offloading toolchain can handle it) but cannot
use it for the native toolchain which is used for making tests.
This approach is vastly superior in every way, allowing us to treat the
GPU as a standard cross-compiling target. We can now install the GPU
utilities to do things like use the offload tests and other fun things.
Some certain utilities need to be built with
`--target=${LLVM_HOST_TRIPLE}` as well. I think this is a fine
workaround as we
will always assume that the GPU `libc` is a cross-build with a
functioning host.
Depends on https://github.com/llvm/llvm-project/pull/81557
Implemented few entry points for REAL(16) math in FortranF128Math
static library. It is a thin wrapper around GNU libquadmath.
Flang driver can always link it, and the dependencies will
be brought in as needed.
The final Fortran program/library that uses any of the entry points
will depend on the underlying third-party library - this dependency
has to be resolved somehow. I added FLANG_RUNTIME_F128_MATH_LIB
CMake control so that the compiler driver and the runtime library
can be built using the same third-party library: this way the linker
knows which dependency to link in (under --as-needed).
The compiler distribution should specify which third-party library
is required for linking/running the apps that use REAL(16).
The compiler package may provide a version of the third-party library
or at least a stub library that can be used for linking, but
the final program execution will still require the actual library.
Introduce Code Object V6 in Clang, LLD, Flang and LLVM. This is the same
as V5 except a new "generic version" flag can be present in EFLAGS. This
is related to new generic targets that'll be added in a follow-up patch.
It's also likely V6 will have new changes (possibly new metadata
entries) added later.
Docs change are part of the follow-up patch #76955
GCC supports -mtls-dialect= for several architectures to select TLSDESC.
This patch supports the following values
* x86: "gnu". "gnu2" (TLSDESC) is not supported yet.
* RISC-V: "trad" (general dynamic), "desc" (TLSDESC, see #66915)
AArch64 toolchains seem to support TLSDESC from the beginning, and the
general dynamic model has poor support. Nobody seems to use the option
-mtls-dialect= at all, so we don't bother with it.
There also seems very little interest in AArch32's TLSDESC support.
TLSDESC does not change IR, but affects object file generation. Without
a backend option the option is a no-op for in-process ThinLTO.
There seems no motivation to have fine-grained control mixing trad/desc
for TLS, so we just pass -mllvm, and don't bother with a modules flag
metadata or function attribute.
Co-authored-by: Paul Kirth <paulkirth@google.com>
Flags `-fveclib=name` were not passed to LTO flags.
This pass fixes that by converting the `-fveclib` flags to their
relevant names for opt's `-vector-lib=name` flags.
For example:
`-fveclib=SLEEF` would become `-vector-library=sleefgnuabi` and passed
through the `-plugin-opt` flag.
